summaryrefslogtreecommitdiff
path: root/Build/source/texk/web2c/mf.web
blob: 065dd2f802183d79f1d5bb8bf748da7bfb2a228f (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
2507
2508
2509
2510
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574
2575
2576
2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782
2783
2784
2785
2786
2787
2788
2789
2790
2791
2792
2793
2794
2795
2796
2797
2798
2799
2800
2801
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843
2844
2845
2846
2847
2848
2849
2850
2851
2852
2853
2854
2855
2856
2857
2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
2875
2876
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
2964
2965
2966
2967
2968
2969
2970
2971
2972
2973
2974
2975
2976
2977
2978
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
3074
3075
3076
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
3099
3100
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161
3162
3163
3164
3165
3166
3167
3168
3169
3170
3171
3172
3173
3174
3175
3176
3177
3178
3179
3180
3181
3182
3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
3203
3204
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
3298
3299
3300
3301
3302
3303
3304
3305
3306
3307
3308
3309
3310
3311
3312
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376
3377
3378
3379
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
3409
3410
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
3432
3433
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
3473
3474
3475
3476
3477
3478
3479
3480
3481
3482
3483
3484
3485
3486
3487
3488
3489
3490
3491
3492
3493
3494
3495
3496
3497
3498
3499
3500
3501
3502
3503
3504
3505
3506
3507
3508
3509
3510
3511
3512
3513
3514
3515
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
3526
3527
3528
3529
3530
3531
3532
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
3557
3558
3559
3560
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
3624
3625
3626
3627
3628
3629
3630
3631
3632
3633
3634
3635
3636
3637
3638
3639
3640
3641
3642
3643
3644
3645
3646
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
3711
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721
3722
3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
3736
3737
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
3751
3752
3753
3754
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
3767
3768
3769
3770
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781
3782
3783
3784
3785
3786
3787
3788
3789
3790
3791
3792
3793
3794
3795
3796
3797
3798
3799
3800
3801
3802
3803
3804
3805
3806
3807
3808
3809
3810
3811
3812
3813
3814
3815
3816
3817
3818
3819
3820
3821
3822
3823
3824
3825
3826
3827
3828
3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
3855
3856
3857
3858
3859
3860
3861
3862
3863
3864
3865
3866
3867
3868
3869
3870
3871
3872
3873
3874
3875
3876
3877
3878
3879
3880
3881
3882
3883
3884
3885
3886
3887
3888
3889
3890
3891
3892
3893
3894
3895
3896
3897
3898
3899
3900
3901
3902
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
3917
3918
3919
3920
3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
3952
3953
3954
3955
3956
3957
3958
3959
3960
3961
3962
3963
3964
3965
3966
3967
3968
3969
3970
3971
3972
3973
3974
3975
3976
3977
3978
3979
3980
3981
3982
3983
3984
3985
3986
3987
3988
3989
3990
3991
3992
3993
3994
3995
3996
3997
3998
3999
4000
4001
4002
4003
4004
4005
4006
4007
4008
4009
4010
4011
4012
4013
4014
4015
4016
4017
4018
4019
4020
4021
4022
4023
4024
4025
4026
4027
4028
4029
4030
4031
4032
4033
4034
4035
4036
4037
4038
4039
4040
4041
4042
4043
4044
4045
4046
4047
4048
4049
4050
4051
4052
4053
4054
4055
4056
4057
4058
4059
4060
4061
4062
4063
4064
4065
4066
4067
4068
4069
4070
4071
4072
4073
4074
4075
4076
4077
4078
4079
4080
4081
4082
4083
4084
4085
4086
4087
4088
4089
4090
4091
4092
4093
4094
4095
4096
4097
4098
4099
4100
4101
4102
4103
4104
4105
4106
4107
4108
4109
4110
4111
4112
4113
4114
4115
4116
4117
4118
4119
4120
4121
4122
4123
4124
4125
4126
4127
4128
4129
4130
4131
4132
4133
4134
4135
4136
4137
4138
4139
4140
4141
4142
4143
4144
4145
4146
4147
4148
4149
4150
4151
4152
4153
4154
4155
4156
4157
4158
4159
4160
4161
4162
4163
4164
4165
4166
4167
4168
4169
4170
4171
4172
4173
4174
4175
4176
4177
4178
4179
4180
4181
4182
4183
4184
4185
4186
4187
4188
4189
4190
4191
4192
4193
4194
4195
4196
4197
4198
4199
4200
4201
4202
4203
4204
4205
4206
4207
4208
4209
4210
4211
4212
4213
4214
4215
4216
4217
4218
4219
4220
4221
4222
4223
4224
4225
4226
4227
4228
4229
4230
4231
4232
4233
4234
4235
4236
4237
4238
4239
4240
4241
4242
4243
4244
4245
4246
4247
4248
4249
4250
4251
4252
4253
4254
4255
4256
4257
4258
4259
4260
4261
4262
4263
4264
4265
4266
4267
4268
4269
4270
4271
4272
4273
4274
4275
4276
4277
4278
4279
4280
4281
4282
4283
4284
4285
4286
4287
4288
4289
4290
4291
4292
4293
4294
4295
4296
4297
4298
4299
4300
4301
4302
4303
4304
4305
4306
4307
4308
4309
4310
4311
4312
4313
4314
4315
4316
4317
4318
4319
4320
4321
4322
4323
4324
4325
4326
4327
4328
4329
4330
4331
4332
4333
4334
4335
4336
4337
4338
4339
4340
4341
4342
4343
4344
4345
4346
4347
4348
4349
4350
4351
4352
4353
4354
4355
4356
4357
4358
4359
4360
4361
4362
4363
4364
4365
4366
4367
4368
4369
4370
4371
4372
4373
4374
4375
4376
4377
4378
4379
4380
4381
4382
4383
4384
4385
4386
4387
4388
4389
4390
4391
4392
4393
4394
4395
4396
4397
4398
4399
4400
4401
4402
4403
4404
4405
4406
4407
4408
4409
4410
4411
4412
4413
4414
4415
4416
4417
4418
4419
4420
4421
4422
4423
4424
4425
4426
4427
4428
4429
4430
4431
4432
4433
4434
4435
4436
4437
4438
4439
4440
4441
4442
4443
4444
4445
4446
4447
4448
4449
4450
4451
4452
4453
4454
4455
4456
4457
4458
4459
4460
4461
4462
4463
4464
4465
4466
4467
4468
4469
4470
4471
4472
4473
4474
4475
4476
4477
4478
4479
4480
4481
4482
4483
4484
4485
4486
4487
4488
4489
4490
4491
4492
4493
4494
4495
4496
4497
4498
4499
4500
4501
4502
4503
4504
4505
4506
4507
4508
4509
4510
4511
4512
4513
4514
4515
4516
4517
4518
4519
4520
4521
4522
4523
4524
4525
4526
4527
4528
4529
4530
4531
4532
4533
4534
4535
4536
4537
4538
4539
4540
4541
4542
4543
4544
4545
4546
4547
4548
4549
4550
4551
4552
4553
4554
4555
4556
4557
4558
4559
4560
4561
4562
4563
4564
4565
4566
4567
4568
4569
4570
4571
4572
4573
4574
4575
4576
4577
4578
4579
4580
4581
4582
4583
4584
4585
4586
4587
4588
4589
4590
4591
4592
4593
4594
4595
4596
4597
4598
4599
4600
4601
4602
4603
4604
4605
4606
4607
4608
4609
4610
4611
4612
4613
4614
4615
4616
4617
4618
4619
4620
4621
4622
4623
4624
4625
4626
4627
4628
4629
4630
4631
4632
4633
4634
4635
4636
4637
4638
4639
4640
4641
4642
4643
4644
4645
4646
4647
4648
4649
4650
4651
4652
4653
4654
4655
4656
4657
4658
4659
4660
4661
4662
4663
4664
4665
4666
4667
4668
4669
4670
4671
4672
4673
4674
4675
4676
4677
4678
4679
4680
4681
4682
4683
4684
4685
4686
4687
4688
4689
4690
4691
4692
4693
4694
4695
4696
4697
4698
4699
4700
4701
4702
4703
4704
4705
4706
4707
4708
4709
4710
4711
4712
4713
4714
4715
4716
4717
4718
4719
4720
4721
4722
4723
4724
4725
4726
4727
4728
4729
4730
4731
4732
4733
4734
4735
4736
4737
4738
4739
4740
4741
4742
4743
4744
4745
4746
4747
4748
4749
4750
4751
4752
4753
4754
4755
4756
4757
4758
4759
4760
4761
4762
4763
4764
4765
4766
4767
4768
4769
4770
4771
4772
4773
4774
4775
4776
4777
4778
4779
4780
4781
4782
4783
4784
4785
4786
4787
4788
4789
4790
4791
4792
4793
4794
4795
4796
4797
4798
4799
4800
4801
4802
4803
4804
4805
4806
4807
4808
4809
4810
4811
4812
4813
4814
4815
4816
4817
4818
4819
4820
4821
4822
4823
4824
4825
4826
4827
4828
4829
4830
4831
4832
4833
4834
4835
4836
4837
4838
4839
4840
4841
4842
4843
4844
4845
4846
4847
4848
4849
4850
4851
4852
4853
4854
4855
4856
4857
4858
4859
4860
4861
4862
4863
4864
4865
4866
4867
4868
4869
4870
4871
4872
4873
4874
4875
4876
4877
4878
4879
4880
4881
4882
4883
4884
4885
4886
4887
4888
4889
4890
4891
4892
4893
4894
4895
4896
4897
4898
4899
4900
4901
4902
4903
4904
4905
4906
4907
4908
4909
4910
4911
4912
4913
4914
4915
4916
4917
4918
4919
4920
4921
4922
4923
4924
4925
4926
4927
4928
4929
4930
4931
4932
4933
4934
4935
4936
4937
4938
4939
4940
4941
4942
4943
4944
4945
4946
4947
4948
4949
4950
4951
4952
4953
4954
4955
4956
4957
4958
4959
4960
4961
4962
4963
4964
4965
4966
4967
4968
4969
4970
4971
4972
4973
4974
4975
4976
4977
4978
4979
4980
4981
4982
4983
4984
4985
4986
4987
4988
4989
4990
4991
4992
4993
4994
4995
4996
4997
4998
4999
5000
5001
5002
5003
5004
5005
5006
5007
5008
5009
5010
5011
5012
5013
5014
5015
5016
5017
5018
5019
5020
5021
5022
5023
5024
5025
5026
5027
5028
5029
5030
5031
5032
5033
5034
5035
5036
5037
5038
5039
5040
5041
5042
5043
5044
5045
5046
5047
5048
5049
5050
5051
5052
5053
5054
5055
5056
5057
5058
5059
5060
5061
5062
5063
5064
5065
5066
5067
5068
5069
5070
5071
5072
5073
5074
5075
5076
5077
5078
5079
5080
5081
5082
5083
5084
5085
5086
5087
5088
5089
5090
5091
5092
5093
5094
5095
5096
5097
5098
5099
5100
5101
5102
5103
5104
5105
5106
5107
5108
5109
5110
5111
5112
5113
5114
5115
5116
5117
5118
5119
5120
5121
5122
5123
5124
5125
5126
5127
5128
5129
5130
5131
5132
5133
5134
5135
5136
5137
5138
5139
5140
5141
5142
5143
5144
5145
5146
5147
5148
5149
5150
5151
5152
5153
5154
5155
5156
5157
5158
5159
5160
5161
5162
5163
5164
5165
5166
5167
5168
5169
5170
5171
5172
5173
5174
5175
5176
5177
5178
5179
5180
5181
5182
5183
5184
5185
5186
5187
5188
5189
5190
5191
5192
5193
5194
5195
5196
5197
5198
5199
5200
5201
5202
5203
5204
5205
5206
5207
5208
5209
5210
5211
5212
5213
5214
5215
5216
5217
5218
5219
5220
5221
5222
5223
5224
5225
5226
5227
5228
5229
5230
5231
5232
5233
5234
5235
5236
5237
5238
5239
5240
5241
5242
5243
5244
5245
5246
5247
5248
5249
5250
5251
5252
5253
5254
5255
5256
5257
5258
5259
5260
5261
5262
5263
5264
5265
5266
5267
5268
5269
5270
5271
5272
5273
5274
5275
5276
5277
5278
5279
5280
5281
5282
5283
5284
5285
5286
5287
5288
5289
5290
5291
5292
5293
5294
5295
5296
5297
5298
5299
5300
5301
5302
5303
5304
5305
5306
5307
5308
5309
5310
5311
5312
5313
5314
5315
5316
5317
5318
5319
5320
5321
5322
5323
5324
5325
5326
5327
5328
5329
5330
5331
5332
5333
5334
5335
5336
5337
5338
5339
5340
5341
5342
5343
5344
5345
5346
5347
5348
5349
5350
5351
5352
5353
5354
5355
5356
5357
5358
5359
5360
5361
5362
5363
5364
5365
5366
5367
5368
5369
5370
5371
5372
5373
5374
5375
5376
5377
5378
5379
5380
5381
5382
5383
5384
5385
5386
5387
5388
5389
5390
5391
5392
5393
5394
5395
5396
5397
5398
5399
5400
5401
5402
5403
5404
5405
5406
5407
5408
5409
5410
5411
5412
5413
5414
5415
5416
5417
5418
5419
5420
5421
5422
5423
5424
5425
5426
5427
5428
5429
5430
5431
5432
5433
5434
5435
5436
5437
5438
5439
5440
5441
5442
5443
5444
5445
5446
5447
5448
5449
5450
5451
5452
5453
5454
5455
5456
5457
5458
5459
5460
5461
5462
5463
5464
5465
5466
5467
5468
5469
5470
5471
5472
5473
5474
5475
5476
5477
5478
5479
5480
5481
5482
5483
5484
5485
5486
5487
5488
5489
5490
5491
5492
5493
5494
5495
5496
5497
5498
5499
5500
5501
5502
5503
5504
5505
5506
5507
5508
5509
5510
5511
5512
5513
5514
5515
5516
5517
5518
5519
5520
5521
5522
5523
5524
5525
5526
5527
5528
5529
5530
5531
5532
5533
5534
5535
5536
5537
5538
5539
5540
5541
5542
5543
5544
5545
5546
5547
5548
5549
5550
5551
5552
5553
5554
5555
5556
5557
5558
5559
5560
5561
5562
5563
5564
5565
5566
5567
5568
5569
5570
5571
5572
5573
5574
5575
5576
5577
5578
5579
5580
5581
5582
5583
5584
5585
5586
5587
5588
5589
5590
5591
5592
5593
5594
5595
5596
5597
5598
5599
5600
5601
5602
5603
5604
5605
5606
5607
5608
5609
5610
5611
5612
5613
5614
5615
5616
5617
5618
5619
5620
5621
5622
5623
5624
5625
5626
5627
5628
5629
5630
5631
5632
5633
5634
5635
5636
5637
5638
5639
5640
5641
5642
5643
5644
5645
5646
5647
5648
5649
5650
5651
5652
5653
5654
5655
5656
5657
5658
5659
5660
5661
5662
5663
5664
5665
5666
5667
5668
5669
5670
5671
5672
5673
5674
5675
5676
5677
5678
5679
5680
5681
5682
5683
5684
5685
5686
5687
5688
5689
5690
5691
5692
5693
5694
5695
5696
5697
5698
5699
5700
5701
5702
5703
5704
5705
5706
5707
5708
5709
5710
5711
5712
5713
5714
5715
5716
5717
5718
5719
5720
5721
5722
5723
5724
5725
5726
5727
5728
5729
5730
5731
5732
5733
5734
5735
5736
5737
5738
5739
5740
5741
5742
5743
5744
5745
5746
5747
5748
5749
5750
5751
5752
5753
5754
5755
5756
5757
5758
5759
5760
5761
5762
5763
5764
5765
5766
5767
5768
5769
5770
5771
5772
5773
5774
5775
5776
5777
5778
5779
5780
5781
5782
5783
5784
5785
5786
5787
5788
5789
5790
5791
5792
5793
5794
5795
5796
5797
5798
5799
5800
5801
5802
5803
5804
5805
5806
5807
5808
5809
5810
5811
5812
5813
5814
5815
5816
5817
5818
5819
5820
5821
5822
5823
5824
5825
5826
5827
5828
5829
5830
5831
5832
5833
5834
5835
5836
5837
5838
5839
5840
5841
5842
5843
5844
5845
5846
5847
5848
5849
5850
5851
5852
5853
5854
5855
5856
5857
5858
5859
5860
5861
5862
5863
5864
5865
5866
5867
5868
5869
5870
5871
5872
5873
5874
5875
5876
5877
5878
5879
5880
5881
5882
5883
5884
5885
5886
5887
5888
5889
5890
5891
5892
5893
5894
5895
5896
5897
5898
5899
5900
5901
5902
5903
5904
5905
5906
5907
5908
5909
5910
5911
5912
5913
5914
5915
5916
5917
5918
5919
5920
5921
5922
5923
5924
5925
5926
5927
5928
5929
5930
5931
5932
5933
5934
5935
5936
5937
5938
5939
5940
5941
5942
5943
5944
5945
5946
5947
5948
5949
5950
5951
5952
5953
5954
5955
5956
5957
5958
5959
5960
5961
5962
5963
5964
5965
5966
5967
5968
5969
5970
5971
5972
5973
5974
5975
5976
5977
5978
5979
5980
5981
5982
5983
5984
5985
5986
5987
5988
5989
5990
5991
5992
5993
5994
5995
5996
5997
5998
5999
6000
6001
6002
6003
6004
6005
6006
6007
6008
6009
6010
6011
6012
6013
6014
6015
6016
6017
6018
6019
6020
6021
6022
6023
6024
6025
6026
6027
6028
6029
6030
6031
6032
6033
6034
6035
6036
6037
6038
6039
6040
6041
6042
6043
6044
6045
6046
6047
6048
6049
6050
6051
6052
6053
6054
6055
6056
6057
6058
6059
6060
6061
6062
6063
6064
6065
6066
6067
6068
6069
6070
6071
6072
6073
6074
6075
6076
6077
6078
6079
6080
6081
6082
6083
6084
6085
6086
6087
6088
6089
6090
6091
6092
6093
6094
6095
6096
6097
6098
6099
6100
6101
6102
6103
6104
6105
6106
6107
6108
6109
6110
6111
6112
6113
6114
6115
6116
6117
6118
6119
6120
6121
6122
6123
6124
6125
6126
6127
6128
6129
6130
6131
6132
6133
6134
6135
6136
6137
6138
6139
6140
6141
6142
6143
6144
6145
6146
6147
6148
6149
6150
6151
6152
6153
6154
6155
6156
6157
6158
6159
6160
6161
6162
6163
6164
6165
6166
6167
6168
6169
6170
6171
6172
6173
6174
6175
6176
6177
6178
6179
6180
6181
6182
6183
6184
6185
6186
6187
6188
6189
6190
6191
6192
6193
6194
6195
6196
6197
6198
6199
6200
6201
6202
6203
6204
6205
6206
6207
6208
6209
6210
6211
6212
6213
6214
6215
6216
6217
6218
6219
6220
6221
6222
6223
6224
6225
6226
6227
6228
6229
6230
6231
6232
6233
6234
6235
6236
6237
6238
6239
6240
6241
6242
6243
6244
6245
6246
6247
6248
6249
6250
6251
6252
6253
6254
6255
6256
6257
6258
6259
6260
6261
6262
6263
6264
6265
6266
6267
6268
6269
6270
6271
6272
6273
6274
6275
6276
6277
6278
6279
6280
6281
6282
6283
6284
6285
6286
6287
6288
6289
6290
6291
6292
6293
6294
6295
6296
6297
6298
6299
6300
6301
6302
6303
6304
6305
6306
6307
6308
6309
6310
6311
6312
6313
6314
6315
6316
6317
6318
6319
6320
6321
6322
6323
6324
6325
6326
6327
6328
6329
6330
6331
6332
6333
6334
6335
6336
6337
6338
6339
6340
6341
6342
6343
6344
6345
6346
6347
6348
6349
6350
6351
6352
6353
6354
6355
6356
6357
6358
6359
6360
6361
6362
6363
6364
6365
6366
6367
6368
6369
6370
6371
6372
6373
6374
6375
6376
6377
6378
6379
6380
6381
6382
6383
6384
6385
6386
6387
6388
6389
6390
6391
6392
6393
6394
6395
6396
6397
6398
6399
6400
6401
6402
6403
6404
6405
6406
6407
6408
6409
6410
6411
6412
6413
6414
6415
6416
6417
6418
6419
6420
6421
6422
6423
6424
6425
6426
6427
6428
6429
6430
6431
6432
6433
6434
6435
6436
6437
6438
6439
6440
6441
6442
6443
6444
6445
6446
6447
6448
6449
6450
6451
6452
6453
6454
6455
6456
6457
6458
6459
6460
6461
6462
6463
6464
6465
6466
6467
6468
6469
6470
6471
6472
6473
6474
6475
6476
6477
6478
6479
6480
6481
6482
6483
6484
6485
6486
6487
6488
6489
6490
6491
6492
6493
6494
6495
6496
6497
6498
6499
6500
6501
6502
6503
6504
6505
6506
6507
6508
6509
6510
6511
6512
6513
6514
6515
6516
6517
6518
6519
6520
6521
6522
6523
6524
6525
6526
6527
6528
6529
6530
6531
6532
6533
6534
6535
6536
6537
6538
6539
6540
6541
6542
6543
6544
6545
6546
6547
6548
6549
6550
6551
6552
6553
6554
6555
6556
6557
6558
6559
6560
6561
6562
6563
6564
6565
6566
6567
6568
6569
6570
6571
6572
6573
6574
6575
6576
6577
6578
6579
6580
6581
6582
6583
6584
6585
6586
6587
6588
6589
6590
6591
6592
6593
6594
6595
6596
6597
6598
6599
6600
6601
6602
6603
6604
6605
6606
6607
6608
6609
6610
6611
6612
6613
6614
6615
6616
6617
6618
6619
6620
6621
6622
6623
6624
6625
6626
6627
6628
6629
6630
6631
6632
6633
6634
6635
6636
6637
6638
6639
6640
6641
6642
6643
6644
6645
6646
6647
6648
6649
6650
6651
6652
6653
6654
6655
6656
6657
6658
6659
6660
6661
6662
6663
6664
6665
6666
6667
6668
6669
6670
6671
6672
6673
6674
6675
6676
6677
6678
6679
6680
6681
6682
6683
6684
6685
6686
6687
6688
6689
6690
6691
6692
6693
6694
6695
6696
6697
6698
6699
6700
6701
6702
6703
6704
6705
6706
6707
6708
6709
6710
6711
6712
6713
6714
6715
6716
6717
6718
6719
6720
6721
6722
6723
6724
6725
6726
6727
6728
6729
6730
6731
6732
6733
6734
6735
6736
6737
6738
6739
6740
6741
6742
6743
6744
6745
6746
6747
6748
6749
6750
6751
6752
6753
6754
6755
6756
6757
6758
6759
6760
6761
6762
6763
6764
6765
6766
6767
6768
6769
6770
6771
6772
6773
6774
6775
6776
6777
6778
6779
6780
6781
6782
6783
6784
6785
6786
6787
6788
6789
6790
6791
6792
6793
6794
6795
6796
6797
6798
6799
6800
6801
6802
6803
6804
6805
6806
6807
6808
6809
6810
6811
6812
6813
6814
6815
6816
6817
6818
6819
6820
6821
6822
6823
6824
6825
6826
6827
6828
6829
6830
6831
6832
6833
6834
6835
6836
6837
6838
6839
6840
6841
6842
6843
6844
6845
6846
6847
6848
6849
6850
6851
6852
6853
6854
6855
6856
6857
6858
6859
6860
6861
6862
6863
6864
6865
6866
6867
6868
6869
6870
6871
6872
6873
6874
6875
6876
6877
6878
6879
6880
6881
6882
6883
6884
6885
6886
6887
6888
6889
6890
6891
6892
6893
6894
6895
6896
6897
6898
6899
6900
6901
6902
6903
6904
6905
6906
6907
6908
6909
6910
6911
6912
6913
6914
6915
6916
6917
6918
6919
6920
6921
6922
6923
6924
6925
6926
6927
6928
6929
6930
6931
6932
6933
6934
6935
6936
6937
6938
6939
6940
6941
6942
6943
6944
6945
6946
6947
6948
6949
6950
6951
6952
6953
6954
6955
6956
6957
6958
6959
6960
6961
6962
6963
6964
6965
6966
6967
6968
6969
6970
6971
6972
6973
6974
6975
6976
6977
6978
6979
6980
6981
6982
6983
6984
6985
6986
6987
6988
6989
6990
6991
6992
6993
6994
6995
6996
6997
6998
6999
7000
7001
7002
7003
7004
7005
7006
7007
7008
7009
7010
7011
7012
7013
7014
7015
7016
7017
7018
7019
7020
7021
7022
7023
7024
7025
7026
7027
7028
7029
7030
7031
7032
7033
7034
7035
7036
7037
7038
7039
7040
7041
7042
7043
7044
7045
7046
7047
7048
7049
7050
7051
7052
7053
7054
7055
7056
7057
7058
7059
7060
7061
7062
7063
7064
7065
7066
7067
7068
7069
7070
7071
7072
7073
7074
7075
7076
7077
7078
7079
7080
7081
7082
7083
7084
7085
7086
7087
7088
7089
7090
7091
7092
7093
7094
7095
7096
7097
7098
7099
7100
7101
7102
7103
7104
7105
7106
7107
7108
7109
7110
7111
7112
7113
7114
7115
7116
7117
7118
7119
7120
7121
7122
7123
7124
7125
7126
7127
7128
7129
7130
7131
7132
7133
7134
7135
7136
7137
7138
7139
7140
7141
7142
7143
7144
7145
7146
7147
7148
7149
7150
7151
7152
7153
7154
7155
7156
7157
7158
7159
7160
7161
7162
7163
7164
7165
7166
7167
7168
7169
7170
7171
7172
7173
7174
7175
7176
7177
7178
7179
7180
7181
7182
7183
7184
7185
7186
7187
7188
7189
7190
7191
7192
7193
7194
7195
7196
7197
7198
7199
7200
7201
7202
7203
7204
7205
7206
7207
7208
7209
7210
7211
7212
7213
7214
7215
7216
7217
7218
7219
7220
7221
7222
7223
7224
7225
7226
7227
7228
7229
7230
7231
7232
7233
7234
7235
7236
7237
7238
7239
7240
7241
7242
7243
7244
7245
7246
7247
7248
7249
7250
7251
7252
7253
7254
7255
7256
7257
7258
7259
7260
7261
7262
7263
7264
7265
7266
7267
7268
7269
7270
7271
7272
7273
7274
7275
7276
7277
7278
7279
7280
7281
7282
7283
7284
7285
7286
7287
7288
7289
7290
7291
7292
7293
7294
7295
7296
7297
7298
7299
7300
7301
7302
7303
7304
7305
7306
7307
7308
7309
7310
7311
7312
7313
7314
7315
7316
7317
7318
7319
7320
7321
7322
7323
7324
7325
7326
7327
7328
7329
7330
7331
7332
7333
7334
7335
7336
7337
7338
7339
7340
7341
7342
7343
7344
7345
7346
7347
7348
7349
7350
7351
7352
7353
7354
7355
7356
7357
7358
7359
7360
7361
7362
7363
7364
7365
7366
7367
7368
7369
7370
7371
7372
7373
7374
7375
7376
7377
7378
7379
7380
7381
7382
7383
7384
7385
7386
7387
7388
7389
7390
7391
7392
7393
7394
7395
7396
7397
7398
7399
7400
7401
7402
7403
7404
7405
7406
7407
7408
7409
7410
7411
7412
7413
7414
7415
7416
7417
7418
7419
7420
7421
7422
7423
7424
7425
7426
7427
7428
7429
7430
7431
7432
7433
7434
7435
7436
7437
7438
7439
7440
7441
7442
7443
7444
7445
7446
7447
7448
7449
7450
7451
7452
7453
7454
7455
7456
7457
7458
7459
7460
7461
7462
7463
7464
7465
7466
7467
7468
7469
7470
7471
7472
7473
7474
7475
7476
7477
7478
7479
7480
7481
7482
7483
7484
7485
7486
7487
7488
7489
7490
7491
7492
7493
7494
7495
7496
7497
7498
7499
7500
7501
7502
7503
7504
7505
7506
7507
7508
7509
7510
7511
7512
7513
7514
7515
7516
7517
7518
7519
7520
7521
7522
7523
7524
7525
7526
7527
7528
7529
7530
7531
7532
7533
7534
7535
7536
7537
7538
7539
7540
7541
7542
7543
7544
7545
7546
7547
7548
7549
7550
7551
7552
7553
7554
7555
7556
7557
7558
7559
7560
7561
7562
7563
7564
7565
7566
7567
7568
7569
7570
7571
7572
7573
7574
7575
7576
7577
7578
7579
7580
7581
7582
7583
7584
7585
7586
7587
7588
7589
7590
7591
7592
7593
7594
7595
7596
7597
7598
7599
7600
7601
7602
7603
7604
7605
7606
7607
7608
7609
7610
7611
7612
7613
7614
7615
7616
7617
7618
7619
7620
7621
7622
7623
7624
7625
7626
7627
7628
7629
7630
7631
7632
7633
7634
7635
7636
7637
7638
7639
7640
7641
7642
7643
7644
7645
7646
7647
7648
7649
7650
7651
7652
7653
7654
7655
7656
7657
7658
7659
7660
7661
7662
7663
7664
7665
7666
7667
7668
7669
7670
7671
7672
7673
7674
7675
7676
7677
7678
7679
7680
7681
7682
7683
7684
7685
7686
7687
7688
7689
7690
7691
7692
7693
7694
7695
7696
7697
7698
7699
7700
7701
7702
7703
7704
7705
7706
7707
7708
7709
7710
7711
7712
7713
7714
7715
7716
7717
7718
7719
7720
7721
7722
7723
7724
7725
7726
7727
7728
7729
7730
7731
7732
7733
7734
7735
7736
7737
7738
7739
7740
7741
7742
7743
7744
7745
7746
7747
7748
7749
7750
7751
7752
7753
7754
7755
7756
7757
7758
7759
7760
7761
7762
7763
7764
7765
7766
7767
7768
7769
7770
7771
7772
7773
7774
7775
7776
7777
7778
7779
7780
7781
7782
7783
7784
7785
7786
7787
7788
7789
7790
7791
7792
7793
7794
7795
7796
7797
7798
7799
7800
7801
7802
7803
7804
7805
7806
7807
7808
7809
7810
7811
7812
7813
7814
7815
7816
7817
7818
7819
7820
7821
7822
7823
7824
7825
7826
7827
7828
7829
7830
7831
7832
7833
7834
7835
7836
7837
7838
7839
7840
7841
7842
7843
7844
7845
7846
7847
7848
7849
7850
7851
7852
7853
7854
7855
7856
7857
7858
7859
7860
7861
7862
7863
7864
7865
7866
7867
7868
7869
7870
7871
7872
7873
7874
7875
7876
7877
7878
7879
7880
7881
7882
7883
7884
7885
7886
7887
7888
7889
7890
7891
7892
7893
7894
7895
7896
7897
7898
7899
7900
7901
7902
7903
7904
7905
7906
7907
7908
7909
7910
7911
7912
7913
7914
7915
7916
7917
7918
7919
7920
7921
7922
7923
7924
7925
7926
7927
7928
7929
7930
7931
7932
7933
7934
7935
7936
7937
7938
7939
7940
7941
7942
7943
7944
7945
7946
7947
7948
7949
7950
7951
7952
7953
7954
7955
7956
7957
7958
7959
7960
7961
7962
7963
7964
7965
7966
7967
7968
7969
7970
7971
7972
7973
7974
7975
7976
7977
7978
7979
7980
7981
7982
7983
7984
7985
7986
7987
7988
7989
7990
7991
7992
7993
7994
7995
7996
7997
7998
7999
8000
8001
8002
8003
8004
8005
8006
8007
8008
8009
8010
8011
8012
8013
8014
8015
8016
8017
8018
8019
8020
8021
8022
8023
8024
8025
8026
8027
8028
8029
8030
8031
8032
8033
8034
8035
8036
8037
8038
8039
8040
8041
8042
8043
8044
8045
8046
8047
8048
8049
8050
8051
8052
8053
8054
8055
8056
8057
8058
8059
8060
8061
8062
8063
8064
8065
8066
8067
8068
8069
8070
8071
8072
8073
8074
8075
8076
8077
8078
8079
8080
8081
8082
8083
8084
8085
8086
8087
8088
8089
8090
8091
8092
8093
8094
8095
8096
8097
8098
8099
8100
8101
8102
8103
8104
8105
8106
8107
8108
8109
8110
8111
8112
8113
8114
8115
8116
8117
8118
8119
8120
8121
8122
8123
8124
8125
8126
8127
8128
8129
8130
8131
8132
8133
8134
8135
8136
8137
8138
8139
8140
8141
8142
8143
8144
8145
8146
8147
8148
8149
8150
8151
8152
8153
8154
8155
8156
8157
8158
8159
8160
8161
8162
8163
8164
8165
8166
8167
8168
8169
8170
8171
8172
8173
8174
8175
8176
8177
8178
8179
8180
8181
8182
8183
8184
8185
8186
8187
8188
8189
8190
8191
8192
8193
8194
8195
8196
8197
8198
8199
8200
8201
8202
8203
8204
8205
8206
8207
8208
8209
8210
8211
8212
8213
8214
8215
8216
8217
8218
8219
8220
8221
8222
8223
8224
8225
8226
8227
8228
8229
8230
8231
8232
8233
8234
8235
8236
8237
8238
8239
8240
8241
8242
8243
8244
8245
8246
8247
8248
8249
8250
8251
8252
8253
8254
8255
8256
8257
8258
8259
8260
8261
8262
8263
8264
8265
8266
8267
8268
8269
8270
8271
8272
8273
8274
8275
8276
8277
8278
8279
8280
8281
8282
8283
8284
8285
8286
8287
8288
8289
8290
8291
8292
8293
8294
8295
8296
8297
8298
8299
8300
8301
8302
8303
8304
8305
8306
8307
8308
8309
8310
8311
8312
8313
8314
8315
8316
8317
8318
8319
8320
8321
8322
8323
8324
8325
8326
8327
8328
8329
8330
8331
8332
8333
8334
8335
8336
8337
8338
8339
8340
8341
8342
8343
8344
8345
8346
8347
8348
8349
8350
8351
8352
8353
8354
8355
8356
8357
8358
8359
8360
8361
8362
8363
8364
8365
8366
8367
8368
8369
8370
8371
8372
8373
8374
8375
8376
8377
8378
8379
8380
8381
8382
8383
8384
8385
8386
8387
8388
8389
8390
8391
8392
8393
8394
8395
8396
8397
8398
8399
8400
8401
8402
8403
8404
8405
8406
8407
8408
8409
8410
8411
8412
8413
8414
8415
8416
8417
8418
8419
8420
8421
8422
8423
8424
8425
8426
8427
8428
8429
8430
8431
8432
8433
8434
8435
8436
8437
8438
8439
8440
8441
8442
8443
8444
8445
8446
8447
8448
8449
8450
8451
8452
8453
8454
8455
8456
8457
8458
8459
8460
8461
8462
8463
8464
8465
8466
8467
8468
8469
8470
8471
8472
8473
8474
8475
8476
8477
8478
8479
8480
8481
8482
8483
8484
8485
8486
8487
8488
8489
8490
8491
8492
8493
8494
8495
8496
8497
8498
8499
8500
8501
8502
8503
8504
8505
8506
8507
8508
8509
8510
8511
8512
8513
8514
8515
8516
8517
8518
8519
8520
8521
8522
8523
8524
8525
8526
8527
8528
8529
8530
8531
8532
8533
8534
8535
8536
8537
8538
8539
8540
8541
8542
8543
8544
8545
8546
8547
8548
8549
8550
8551
8552
8553
8554
8555
8556
8557
8558
8559
8560
8561
8562
8563
8564
8565
8566
8567
8568
8569
8570
8571
8572
8573
8574
8575
8576
8577
8578
8579
8580
8581
8582
8583
8584
8585
8586
8587
8588
8589
8590
8591
8592
8593
8594
8595
8596
8597
8598
8599
8600
8601
8602
8603
8604
8605
8606
8607
8608
8609
8610
8611
8612
8613
8614
8615
8616
8617
8618
8619
8620
8621
8622
8623
8624
8625
8626
8627
8628
8629
8630
8631
8632
8633
8634
8635
8636
8637
8638
8639
8640
8641
8642
8643
8644
8645
8646
8647
8648
8649
8650
8651
8652
8653
8654
8655
8656
8657
8658
8659
8660
8661
8662
8663
8664
8665
8666
8667
8668
8669
8670
8671
8672
8673
8674
8675
8676
8677
8678
8679
8680
8681
8682
8683
8684
8685
8686
8687
8688
8689
8690
8691
8692
8693
8694
8695
8696
8697
8698
8699
8700
8701
8702
8703
8704
8705
8706
8707
8708
8709
8710
8711
8712
8713
8714
8715
8716
8717
8718
8719
8720
8721
8722
8723
8724
8725
8726
8727
8728
8729
8730
8731
8732
8733
8734
8735
8736
8737
8738
8739
8740
8741
8742
8743
8744
8745
8746
8747
8748
8749
8750
8751
8752
8753
8754
8755
8756
8757
8758
8759
8760
8761
8762
8763
8764
8765
8766
8767
8768
8769
8770
8771
8772
8773
8774
8775
8776
8777
8778
8779
8780
8781
8782
8783
8784
8785
8786
8787
8788
8789
8790
8791
8792
8793
8794
8795
8796
8797
8798
8799
8800
8801
8802
8803
8804
8805
8806
8807
8808
8809
8810
8811
8812
8813
8814
8815
8816
8817
8818
8819
8820
8821
8822
8823
8824
8825
8826
8827
8828
8829
8830
8831
8832
8833
8834
8835
8836
8837
8838
8839
8840
8841
8842
8843
8844
8845
8846
8847
8848
8849
8850
8851
8852
8853
8854
8855
8856
8857
8858
8859
8860
8861
8862
8863
8864
8865
8866
8867
8868
8869
8870
8871
8872
8873
8874
8875
8876
8877
8878
8879
8880
8881
8882
8883
8884
8885
8886
8887
8888
8889
8890
8891
8892
8893
8894
8895
8896
8897
8898
8899
8900
8901
8902
8903
8904
8905
8906
8907
8908
8909
8910
8911
8912
8913
8914
8915
8916
8917
8918
8919
8920
8921
8922
8923
8924
8925
8926
8927
8928
8929
8930
8931
8932
8933
8934
8935
8936
8937
8938
8939
8940
8941
8942
8943
8944
8945
8946
8947
8948
8949
8950
8951
8952
8953
8954
8955
8956
8957
8958
8959
8960
8961
8962
8963
8964
8965
8966
8967
8968
8969
8970
8971
8972
8973
8974
8975
8976
8977
8978
8979
8980
8981
8982
8983
8984
8985
8986
8987
8988
8989
8990
8991
8992
8993
8994
8995
8996
8997
8998
8999
9000
9001
9002
9003
9004
9005
9006
9007
9008
9009
9010
9011
9012
9013
9014
9015
9016
9017
9018
9019
9020
9021
9022
9023
9024
9025
9026
9027
9028
9029
9030
9031
9032
9033
9034
9035
9036
9037
9038
9039
9040
9041
9042
9043
9044
9045
9046
9047
9048
9049
9050
9051
9052
9053
9054
9055
9056
9057
9058
9059
9060
9061
9062
9063
9064
9065
9066
9067
9068
9069
9070
9071
9072
9073
9074
9075
9076
9077
9078
9079
9080
9081
9082
9083
9084
9085
9086
9087
9088
9089
9090
9091
9092
9093
9094
9095
9096
9097
9098
9099
9100
9101
9102
9103
9104
9105
9106
9107
9108
9109
9110
9111
9112
9113
9114
9115
9116
9117
9118
9119
9120
9121
9122
9123
9124
9125
9126
9127
9128
9129
9130
9131
9132
9133
9134
9135
9136
9137
9138
9139
9140
9141
9142
9143
9144
9145
9146
9147
9148
9149
9150
9151
9152
9153
9154
9155
9156
9157
9158
9159
9160
9161
9162
9163
9164
9165
9166
9167
9168
9169
9170
9171
9172
9173
9174
9175
9176
9177
9178
9179
9180
9181
9182
9183
9184
9185
9186
9187
9188
9189
9190
9191
9192
9193
9194
9195
9196
9197
9198
9199
9200
9201
9202
9203
9204
9205
9206
9207
9208
9209
9210
9211
9212
9213
9214
9215
9216
9217
9218
9219
9220
9221
9222
9223
9224
9225
9226
9227
9228
9229
9230
9231
9232
9233
9234
9235
9236
9237
9238
9239
9240
9241
9242
9243
9244
9245
9246
9247
9248
9249
9250
9251
9252
9253
9254
9255
9256
9257
9258
9259
9260
9261
9262
9263
9264
9265
9266
9267
9268
9269
9270
9271
9272
9273
9274
9275
9276
9277
9278
9279
9280
9281
9282
9283
9284
9285
9286
9287
9288
9289
9290
9291
9292
9293
9294
9295
9296
9297
9298
9299
9300
9301
9302
9303
9304
9305
9306
9307
9308
9309
9310
9311
9312
9313
9314
9315
9316
9317
9318
9319
9320
9321
9322
9323
9324
9325
9326
9327
9328
9329
9330
9331
9332
9333
9334
9335
9336
9337
9338
9339
9340
9341
9342
9343
9344
9345
9346
9347
9348
9349
9350
9351
9352
9353
9354
9355
9356
9357
9358
9359
9360
9361
9362
9363
9364
9365
9366
9367
9368
9369
9370
9371
9372
9373
9374
9375
9376
9377
9378
9379
9380
9381
9382
9383
9384
9385
9386
9387
9388
9389
9390
9391
9392
9393
9394
9395
9396
9397
9398
9399
9400
9401
9402
9403
9404
9405
9406
9407
9408
9409
9410
9411
9412
9413
9414
9415
9416
9417
9418
9419
9420
9421
9422
9423
9424
9425
9426
9427
9428
9429
9430
9431
9432
9433
9434
9435
9436
9437
9438
9439
9440
9441
9442
9443
9444
9445
9446
9447
9448
9449
9450
9451
9452
9453
9454
9455
9456
9457
9458
9459
9460
9461
9462
9463
9464
9465
9466
9467
9468
9469
9470
9471
9472
9473
9474
9475
9476
9477
9478
9479
9480
9481
9482
9483
9484
9485
9486
9487
9488
9489
9490
9491
9492
9493
9494
9495
9496
9497
9498
9499
9500
9501
9502
9503
9504
9505
9506
9507
9508
9509
9510
9511
9512
9513
9514
9515
9516
9517
9518
9519
9520
9521
9522
9523
9524
9525
9526
9527
9528
9529
9530
9531
9532
9533
9534
9535
9536
9537
9538
9539
9540
9541
9542
9543
9544
9545
9546
9547
9548
9549
9550
9551
9552
9553
9554
9555
9556
9557
9558
9559
9560
9561
9562
9563
9564
9565
9566
9567
9568
9569
9570
9571
9572
9573
9574
9575
9576
9577
9578
9579
9580
9581
9582
9583
9584
9585
9586
9587
9588
9589
9590
9591
9592
9593
9594
9595
9596
9597
9598
9599
9600
9601
9602
9603
9604
9605
9606
9607
9608
9609
9610
9611
9612
9613
9614
9615
9616
9617
9618
9619
9620
9621
9622
9623
9624
9625
9626
9627
9628
9629
9630
9631
9632
9633
9634
9635
9636
9637
9638
9639
9640
9641
9642
9643
9644
9645
9646
9647
9648
9649
9650
9651
9652
9653
9654
9655
9656
9657
9658
9659
9660
9661
9662
9663
9664
9665
9666
9667
9668
9669
9670
9671
9672
9673
9674
9675
9676
9677
9678
9679
9680
9681
9682
9683
9684
9685
9686
9687
9688
9689
9690
9691
9692
9693
9694
9695
9696
9697
9698
9699
9700
9701
9702
9703
9704
9705
9706
9707
9708
9709
9710
9711
9712
9713
9714
9715
9716
9717
9718
9719
9720
9721
9722
9723
9724
9725
9726
9727
9728
9729
9730
9731
9732
9733
9734
9735
9736
9737
9738
9739
9740
9741
9742
9743
9744
9745
9746
9747
9748
9749
9750
9751
9752
9753
9754
9755
9756
9757
9758
9759
9760
9761
9762
9763
9764
9765
9766
9767
9768
9769
9770
9771
9772
9773
9774
9775
9776
9777
9778
9779
9780
9781
9782
9783
9784
9785
9786
9787
9788
9789
9790
9791
9792
9793
9794
9795
9796
9797
9798
9799
9800
9801
9802
9803
9804
9805
9806
9807
9808
9809
9810
9811
9812
9813
9814
9815
9816
9817
9818
9819
9820
9821
9822
9823
9824
9825
9826
9827
9828
9829
9830
9831
9832
9833
9834
9835
9836
9837
9838
9839
9840
9841
9842
9843
9844
9845
9846
9847
9848
9849
9850
9851
9852
9853
9854
9855
9856
9857
9858
9859
9860
9861
9862
9863
9864
9865
9866
9867
9868
9869
9870
9871
9872
9873
9874
9875
9876
9877
9878
9879
9880
9881
9882
9883
9884
9885
9886
9887
9888
9889
9890
9891
9892
9893
9894
9895
9896
9897
9898
9899
9900
9901
9902
9903
9904
9905
9906
9907
9908
9909
9910
9911
9912
9913
9914
9915
9916
9917
9918
9919
9920
9921
9922
9923
9924
9925
9926
9927
9928
9929
9930
9931
9932
9933
9934
9935
9936
9937
9938
9939
9940
9941
9942
9943
9944
9945
9946
9947
9948
9949
9950
9951
9952
9953
9954
9955
9956
9957
9958
9959
9960
9961
9962
9963
9964
9965
9966
9967
9968
9969
9970
9971
9972
9973
9974
9975
9976
9977
9978
9979
9980
9981
9982
9983
9984
9985
9986
9987
9988
9989
9990
9991
9992
9993
9994
9995
9996
9997
9998
9999
10000
10001
10002
10003
10004
10005
10006
10007
10008
10009
10010
10011
10012
10013
10014
10015
10016
10017
10018
10019
10020
10021
10022
10023
10024
10025
10026
10027
10028
10029
10030
10031
10032
10033
10034
10035
10036
10037
10038
10039
10040
10041
10042
10043
10044
10045
10046
10047
10048
10049
10050
10051
10052
10053
10054
10055
10056
10057
10058
10059
10060
10061
10062
10063
10064
10065
10066
10067
10068
10069
10070
10071
10072
10073
10074
10075
10076
10077
10078
10079
10080
10081
10082
10083
10084
10085
10086
10087
10088
10089
10090
10091
10092
10093
10094
10095
10096
10097
10098
10099
10100
10101
10102
10103
10104
10105
10106
10107
10108
10109
10110
10111
10112
10113
10114
10115
10116
10117
10118
10119
10120
10121
10122
10123
10124
10125
10126
10127
10128
10129
10130
10131
10132
10133
10134
10135
10136
10137
10138
10139
10140
10141
10142
10143
10144
10145
10146
10147
10148
10149
10150
10151
10152
10153
10154
10155
10156
10157
10158
10159
10160
10161
10162
10163
10164
10165
10166
10167
10168
10169
10170
10171
10172
10173
10174
10175
10176
10177
10178
10179
10180
10181
10182
10183
10184
10185
10186
10187
10188
10189
10190
10191
10192
10193
10194
10195
10196
10197
10198
10199
10200
10201
10202
10203
10204
10205
10206
10207
10208
10209
10210
10211
10212
10213
10214
10215
10216
10217
10218
10219
10220
10221
10222
10223
10224
10225
10226
10227
10228
10229
10230
10231
10232
10233
10234
10235
10236
10237
10238
10239
10240
10241
10242
10243
10244
10245
10246
10247
10248
10249
10250
10251
10252
10253
10254
10255
10256
10257
10258
10259
10260
10261
10262
10263
10264
10265
10266
10267
10268
10269
10270
10271
10272
10273
10274
10275
10276
10277
10278
10279
10280
10281
10282
10283
10284
10285
10286
10287
10288
10289
10290
10291
10292
10293
10294
10295
10296
10297
10298
10299
10300
10301
10302
10303
10304
10305
10306
10307
10308
10309
10310
10311
10312
10313
10314
10315
10316
10317
10318
10319
10320
10321
10322
10323
10324
10325
10326
10327
10328
10329
10330
10331
10332
10333
10334
10335
10336
10337
10338
10339
10340
10341
10342
10343
10344
10345
10346
10347
10348
10349
10350
10351
10352
10353
10354
10355
10356
10357
10358
10359
10360
10361
10362
10363
10364
10365
10366
10367
10368
10369
10370
10371
10372
10373
10374
10375
10376
10377
10378
10379
10380
10381
10382
10383
10384
10385
10386
10387
10388
10389
10390
10391
10392
10393
10394
10395
10396
10397
10398
10399
10400
10401
10402
10403
10404
10405
10406
10407
10408
10409
10410
10411
10412
10413
10414
10415
10416
10417
10418
10419
10420
10421
10422
10423
10424
10425
10426
10427
10428
10429
10430
10431
10432
10433
10434
10435
10436
10437
10438
10439
10440
10441
10442
10443
10444
10445
10446
10447
10448
10449
10450
10451
10452
10453
10454
10455
10456
10457
10458
10459
10460
10461
10462
10463
10464
10465
10466
10467
10468
10469
10470
10471
10472
10473
10474
10475
10476
10477
10478
10479
10480
10481
10482
10483
10484
10485
10486
10487
10488
10489
10490
10491
10492
10493
10494
10495
10496
10497
10498
10499
10500
10501
10502
10503
10504
10505
10506
10507
10508
10509
10510
10511
10512
10513
10514
10515
10516
10517
10518
10519
10520
10521
10522
10523
10524
10525
10526
10527
10528
10529
10530
10531
10532
10533
10534
10535
10536
10537
10538
10539
10540
10541
10542
10543
10544
10545
10546
10547
10548
10549
10550
10551
10552
10553
10554
10555
10556
10557
10558
10559
10560
10561
10562
10563
10564
10565
10566
10567
10568
10569
10570
10571
10572
10573
10574
10575
10576
10577
10578
10579
10580
10581
10582
10583
10584
10585
10586
10587
10588
10589
10590
10591
10592
10593
10594
10595
10596
10597
10598
10599
10600
10601
10602
10603
10604
10605
10606
10607
10608
10609
10610
10611
10612
10613
10614
10615
10616
10617
10618
10619
10620
10621
10622
10623
10624
10625
10626
10627
10628
10629
10630
10631
10632
10633
10634
10635
10636
10637
10638
10639
10640
10641
10642
10643
10644
10645
10646
10647
10648
10649
10650
10651
10652
10653
10654
10655
10656
10657
10658
10659
10660
10661
10662
10663
10664
10665
10666
10667
10668
10669
10670
10671
10672
10673
10674
10675
10676
10677
10678
10679
10680
10681
10682
10683
10684
10685
10686
10687
10688
10689
10690
10691
10692
10693
10694
10695
10696
10697
10698
10699
10700
10701
10702
10703
10704
10705
10706
10707
10708
10709
10710
10711
10712
10713
10714
10715
10716
10717
10718
10719
10720
10721
10722
10723
10724
10725
10726
10727
10728
10729
10730
10731
10732
10733
10734
10735
10736
10737
10738
10739
10740
10741
10742
10743
10744
10745
10746
10747
10748
10749
10750
10751
10752
10753
10754
10755
10756
10757
10758
10759
10760
10761
10762
10763
10764
10765
10766
10767
10768
10769
10770
10771
10772
10773
10774
10775
10776
10777
10778
10779
10780
10781
10782
10783
10784
10785
10786
10787
10788
10789
10790
10791
10792
10793
10794
10795
10796
10797
10798
10799
10800
10801
10802
10803
10804
10805
10806
10807
10808
10809
10810
10811
10812
10813
10814
10815
10816
10817
10818
10819
10820
10821
10822
10823
10824
10825
10826
10827
10828
10829
10830
10831
10832
10833
10834
10835
10836
10837
10838
10839
10840
10841
10842
10843
10844
10845
10846
10847
10848
10849
10850
10851
10852
10853
10854
10855
10856
10857
10858
10859
10860
10861
10862
10863
10864
10865
10866
10867
10868
10869
10870
10871
10872
10873
10874
10875
10876
10877
10878
10879
10880
10881
10882
10883
10884
10885
10886
10887
10888
10889
10890
10891
10892
10893
10894
10895
10896
10897
10898
10899
10900
10901
10902
10903
10904
10905
10906
10907
10908
10909
10910
10911
10912
10913
10914
10915
10916
10917
10918
10919
10920
10921
10922
10923
10924
10925
10926
10927
10928
10929
10930
10931
10932
10933
10934
10935
10936
10937
10938
10939
10940
10941
10942
10943
10944
10945
10946
10947
10948
10949
10950
10951
10952
10953
10954
10955
10956
10957
10958
10959
10960
10961
10962
10963
10964
10965
10966
10967
10968
10969
10970
10971
10972
10973
10974
10975
10976
10977
10978
10979
10980
10981
10982
10983
10984
10985
10986
10987
10988
10989
10990
10991
10992
10993
10994
10995
10996
10997
10998
10999
11000
11001
11002
11003
11004
11005
11006
11007
11008
11009
11010
11011
11012
11013
11014
11015
11016
11017
11018
11019
11020
11021
11022
11023
11024
11025
11026
11027
11028
11029
11030
11031
11032
11033
11034
11035
11036
11037
11038
11039
11040
11041
11042
11043
11044
11045
11046
11047
11048
11049
11050
11051
11052
11053
11054
11055
11056
11057
11058
11059
11060
11061
11062
11063
11064
11065
11066
11067
11068
11069
11070
11071
11072
11073
11074
11075
11076
11077
11078
11079
11080
11081
11082
11083
11084
11085
11086
11087
11088
11089
11090
11091
11092
11093
11094
11095
11096
11097
11098
11099
11100
11101
11102
11103
11104
11105
11106
11107
11108
11109
11110
11111
11112
11113
11114
11115
11116
11117
11118
11119
11120
11121
11122
11123
11124
11125
11126
11127
11128
11129
11130
11131
11132
11133
11134
11135
11136
11137
11138
11139
11140
11141
11142
11143
11144
11145
11146
11147
11148
11149
11150
11151
11152
11153
11154
11155
11156
11157
11158
11159
11160
11161
11162
11163
11164
11165
11166
11167
11168
11169
11170
11171
11172
11173
11174
11175
11176
11177
11178
11179
11180
11181
11182
11183
11184
11185
11186
11187
11188
11189
11190
11191
11192
11193
11194
11195
11196
11197
11198
11199
11200
11201
11202
11203
11204
11205
11206
11207
11208
11209
11210
11211
11212
11213
11214
11215
11216
11217
11218
11219
11220
11221
11222
11223
11224
11225
11226
11227
11228
11229
11230
11231
11232
11233
11234
11235
11236
11237
11238
11239
11240
11241
11242
11243
11244
11245
11246
11247
11248
11249
11250
11251
11252
11253
11254
11255
11256
11257
11258
11259
11260
11261
11262
11263
11264
11265
11266
11267
11268
11269
11270
11271
11272
11273
11274
11275
11276
11277
11278
11279
11280
11281
11282
11283
11284
11285
11286
11287
11288
11289
11290
11291
11292
11293
11294
11295
11296
11297
11298
11299
11300
11301
11302
11303
11304
11305
11306
11307
11308
11309
11310
11311
11312
11313
11314
11315
11316
11317
11318
11319
11320
11321
11322
11323
11324
11325
11326
11327
11328
11329
11330
11331
11332
11333
11334
11335
11336
11337
11338
11339
11340
11341
11342
11343
11344
11345
11346
11347
11348
11349
11350
11351
11352
11353
11354
11355
11356
11357
11358
11359
11360
11361
11362
11363
11364
11365
11366
11367
11368
11369
11370
11371
11372
11373
11374
11375
11376
11377
11378
11379
11380
11381
11382
11383
11384
11385
11386
11387
11388
11389
11390
11391
11392
11393
11394
11395
11396
11397
11398
11399
11400
11401
11402
11403
11404
11405
11406
11407
11408
11409
11410
11411
11412
11413
11414
11415
11416
11417
11418
11419
11420
11421
11422
11423
11424
11425
11426
11427
11428
11429
11430
11431
11432
11433
11434
11435
11436
11437
11438
11439
11440
11441
11442
11443
11444
11445
11446
11447
11448
11449
11450
11451
11452
11453
11454
11455
11456
11457
11458
11459
11460
11461
11462
11463
11464
11465
11466
11467
11468
11469
11470
11471
11472
11473
11474
11475
11476
11477
11478
11479
11480
11481
11482
11483
11484
11485
11486
11487
11488
11489
11490
11491
11492
11493
11494
11495
11496
11497
11498
11499
11500
11501
11502
11503
11504
11505
11506
11507
11508
11509
11510
11511
11512
11513
11514
11515
11516
11517
11518
11519
11520
11521
11522
11523
11524
11525
11526
11527
11528
11529
11530
11531
11532
11533
11534
11535
11536
11537
11538
11539
11540
11541
11542
11543
11544
11545
11546
11547
11548
11549
11550
11551
11552
11553
11554
11555
11556
11557
11558
11559
11560
11561
11562
11563
11564
11565
11566
11567
11568
11569
11570
11571
11572
11573
11574
11575
11576
11577
11578
11579
11580
11581
11582
11583
11584
11585
11586
11587
11588
11589
11590
11591
11592
11593
11594
11595
11596
11597
11598
11599
11600
11601
11602
11603
11604
11605
11606
11607
11608
11609
11610
11611
11612
11613
11614
11615
11616
11617
11618
11619
11620
11621
11622
11623
11624
11625
11626
11627
11628
11629
11630
11631
11632
11633
11634
11635
11636
11637
11638
11639
11640
11641
11642
11643
11644
11645
11646
11647
11648
11649
11650
11651
11652
11653
11654
11655
11656
11657
11658
11659
11660
11661
11662
11663
11664
11665
11666
11667
11668
11669
11670
11671
11672
11673
11674
11675
11676
11677
11678
11679
11680
11681
11682
11683
11684
11685
11686
11687
11688
11689
11690
11691
11692
11693
11694
11695
11696
11697
11698
11699
11700
11701
11702
11703
11704
11705
11706
11707
11708
11709
11710
11711
11712
11713
11714
11715
11716
11717
11718
11719
11720
11721
11722
11723
11724
11725
11726
11727
11728
11729
11730
11731
11732
11733
11734
11735
11736
11737
11738
11739
11740
11741
11742
11743
11744
11745
11746
11747
11748
11749
11750
11751
11752
11753
11754
11755
11756
11757
11758
11759
11760
11761
11762
11763
11764
11765
11766
11767
11768
11769
11770
11771
11772
11773
11774
11775
11776
11777
11778
11779
11780
11781
11782
11783
11784
11785
11786
11787
11788
11789
11790
11791
11792
11793
11794
11795
11796
11797
11798
11799
11800
11801
11802
11803
11804
11805
11806
11807
11808
11809
11810
11811
11812
11813
11814
11815
11816
11817
11818
11819
11820
11821
11822
11823
11824
11825
11826
11827
11828
11829
11830
11831
11832
11833
11834
11835
11836
11837
11838
11839
11840
11841
11842
11843
11844
11845
11846
11847
11848
11849
11850
11851
11852
11853
11854
11855
11856
11857
11858
11859
11860
11861
11862
11863
11864
11865
11866
11867
11868
11869
11870
11871
11872
11873
11874
11875
11876
11877
11878
11879
11880
11881
11882
11883
11884
11885
11886
11887
11888
11889
11890
11891
11892
11893
11894
11895
11896
11897
11898
11899
11900
11901
11902
11903
11904
11905
11906
11907
11908
11909
11910
11911
11912
11913
11914
11915
11916
11917
11918
11919
11920
11921
11922
11923
11924
11925
11926
11927
11928
11929
11930
11931
11932
11933
11934
11935
11936
11937
11938
11939
11940
11941
11942
11943
11944
11945
11946
11947
11948
11949
11950
11951
11952
11953
11954
11955
11956
11957
11958
11959
11960
11961
11962
11963
11964
11965
11966
11967
11968
11969
11970
11971
11972
11973
11974
11975
11976
11977
11978
11979
11980
11981
11982
11983
11984
11985
11986
11987
11988
11989
11990
11991
11992
11993
11994
11995
11996
11997
11998
11999
12000
12001
12002
12003
12004
12005
12006
12007
12008
12009
12010
12011
12012
12013
12014
12015
12016
12017
12018
12019
12020
12021
12022
12023
12024
12025
12026
12027
12028
12029
12030
12031
12032
12033
12034
12035
12036
12037
12038
12039
12040
12041
12042
12043
12044
12045
12046
12047
12048
12049
12050
12051
12052
12053
12054
12055
12056
12057
12058
12059
12060
12061
12062
12063
12064
12065
12066
12067
12068
12069
12070
12071
12072
12073
12074
12075
12076
12077
12078
12079
12080
12081
12082
12083
12084
12085
12086
12087
12088
12089
12090
12091
12092
12093
12094
12095
12096
12097
12098
12099
12100
12101
12102
12103
12104
12105
12106
12107
12108
12109
12110
12111
12112
12113
12114
12115
12116
12117
12118
12119
12120
12121
12122
12123
12124
12125
12126
12127
12128
12129
12130
12131
12132
12133
12134
12135
12136
12137
12138
12139
12140
12141
12142
12143
12144
12145
12146
12147
12148
12149
12150
12151
12152
12153
12154
12155
12156
12157
12158
12159
12160
12161
12162
12163
12164
12165
12166
12167
12168
12169
12170
12171
12172
12173
12174
12175
12176
12177
12178
12179
12180
12181
12182
12183
12184
12185
12186
12187
12188
12189
12190
12191
12192
12193
12194
12195
12196
12197
12198
12199
12200
12201
12202
12203
12204
12205
12206
12207
12208
12209
12210
12211
12212
12213
12214
12215
12216
12217
12218
12219
12220
12221
12222
12223
12224
12225
12226
12227
12228
12229
12230
12231
12232
12233
12234
12235
12236
12237
12238
12239
12240
12241
12242
12243
12244
12245
12246
12247
12248
12249
12250
12251
12252
12253
12254
12255
12256
12257
12258
12259
12260
12261
12262
12263
12264
12265
12266
12267
12268
12269
12270
12271
12272
12273
12274
12275
12276
12277
12278
12279
12280
12281
12282
12283
12284
12285
12286
12287
12288
12289
12290
12291
12292
12293
12294
12295
12296
12297
12298
12299
12300
12301
12302
12303
12304
12305
12306
12307
12308
12309
12310
12311
12312
12313
12314
12315
12316
12317
12318
12319
12320
12321
12322
12323
12324
12325
12326
12327
12328
12329
12330
12331
12332
12333
12334
12335
12336
12337
12338
12339
12340
12341
12342
12343
12344
12345
12346
12347
12348
12349
12350
12351
12352
12353
12354
12355
12356
12357
12358
12359
12360
12361
12362
12363
12364
12365
12366
12367
12368
12369
12370
12371
12372
12373
12374
12375
12376
12377
12378
12379
12380
12381
12382
12383
12384
12385
12386
12387
12388
12389
12390
12391
12392
12393
12394
12395
12396
12397
12398
12399
12400
12401
12402
12403
12404
12405
12406
12407
12408
12409
12410
12411
12412
12413
12414
12415
12416
12417
12418
12419
12420
12421
12422
12423
12424
12425
12426
12427
12428
12429
12430
12431
12432
12433
12434
12435
12436
12437
12438
12439
12440
12441
12442
12443
12444
12445
12446
12447
12448
12449
12450
12451
12452
12453
12454
12455
12456
12457
12458
12459
12460
12461
12462
12463
12464
12465
12466
12467
12468
12469
12470
12471
12472
12473
12474
12475
12476
12477
12478
12479
12480
12481
12482
12483
12484
12485
12486
12487
12488
12489
12490
12491
12492
12493
12494
12495
12496
12497
12498
12499
12500
12501
12502
12503
12504
12505
12506
12507
12508
12509
12510
12511
12512
12513
12514
12515
12516
12517
12518
12519
12520
12521
12522
12523
12524
12525
12526
12527
12528
12529
12530
12531
12532
12533
12534
12535
12536
12537
12538
12539
12540
12541
12542
12543
12544
12545
12546
12547
12548
12549
12550
12551
12552
12553
12554
12555
12556
12557
12558
12559
12560
12561
12562
12563
12564
12565
12566
12567
12568
12569
12570
12571
12572
12573
12574
12575
12576
12577
12578
12579
12580
12581
12582
12583
12584
12585
12586
12587
12588
12589
12590
12591
12592
12593
12594
12595
12596
12597
12598
12599
12600
12601
12602
12603
12604
12605
12606
12607
12608
12609
12610
12611
12612
12613
12614
12615
12616
12617
12618
12619
12620
12621
12622
12623
12624
12625
12626
12627
12628
12629
12630
12631
12632
12633
12634
12635
12636
12637
12638
12639
12640
12641
12642
12643
12644
12645
12646
12647
12648
12649
12650
12651
12652
12653
12654
12655
12656
12657
12658
12659
12660
12661
12662
12663
12664
12665
12666
12667
12668
12669
12670
12671
12672
12673
12674
12675
12676
12677
12678
12679
12680
12681
12682
12683
12684
12685
12686
12687
12688
12689
12690
12691
12692
12693
12694
12695
12696
12697
12698
12699
12700
12701
12702
12703
12704
12705
12706
12707
12708
12709
12710
12711
12712
12713
12714
12715
12716
12717
12718
12719
12720
12721
12722
12723
12724
12725
12726
12727
12728
12729
12730
12731
12732
12733
12734
12735
12736
12737
12738
12739
12740
12741
12742
12743
12744
12745
12746
12747
12748
12749
12750
12751
12752
12753
12754
12755
12756
12757
12758
12759
12760
12761
12762
12763
12764
12765
12766
12767
12768
12769
12770
12771
12772
12773
12774
12775
12776
12777
12778
12779
12780
12781
12782
12783
12784
12785
12786
12787
12788
12789
12790
12791
12792
12793
12794
12795
12796
12797
12798
12799
12800
12801
12802
12803
12804
12805
12806
12807
12808
12809
12810
12811
12812
12813
12814
12815
12816
12817
12818
12819
12820
12821
12822
12823
12824
12825
12826
12827
12828
12829
12830
12831
12832
12833
12834
12835
12836
12837
12838
12839
12840
12841
12842
12843
12844
12845
12846
12847
12848
12849
12850
12851
12852
12853
12854
12855
12856
12857
12858
12859
12860
12861
12862
12863
12864
12865
12866
12867
12868
12869
12870
12871
12872
12873
12874
12875
12876
12877
12878
12879
12880
12881
12882
12883
12884
12885
12886
12887
12888
12889
12890
12891
12892
12893
12894
12895
12896
12897
12898
12899
12900
12901
12902
12903
12904
12905
12906
12907
12908
12909
12910
12911
12912
12913
12914
12915
12916
12917
12918
12919
12920
12921
12922
12923
12924
12925
12926
12927
12928
12929
12930
12931
12932
12933
12934
12935
12936
12937
12938
12939
12940
12941
12942
12943
12944
12945
12946
12947
12948
12949
12950
12951
12952
12953
12954
12955
12956
12957
12958
12959
12960
12961
12962
12963
12964
12965
12966
12967
12968
12969
12970
12971
12972
12973
12974
12975
12976
12977
12978
12979
12980
12981
12982
12983
12984
12985
12986
12987
12988
12989
12990
12991
12992
12993
12994
12995
12996
12997
12998
12999
13000
13001
13002
13003
13004
13005
13006
13007
13008
13009
13010
13011
13012
13013
13014
13015
13016
13017
13018
13019
13020
13021
13022
13023
13024
13025
13026
13027
13028
13029
13030
13031
13032
13033
13034
13035
13036
13037
13038
13039
13040
13041
13042
13043
13044
13045
13046
13047
13048
13049
13050
13051
13052
13053
13054
13055
13056
13057
13058
13059
13060
13061
13062
13063
13064
13065
13066
13067
13068
13069
13070
13071
13072
13073
13074
13075
13076
13077
13078
13079
13080
13081
13082
13083
13084
13085
13086
13087
13088
13089
13090
13091
13092
13093
13094
13095
13096
13097
13098
13099
13100
13101
13102
13103
13104
13105
13106
13107
13108
13109
13110
13111
13112
13113
13114
13115
13116
13117
13118
13119
13120
13121
13122
13123
13124
13125
13126
13127
13128
13129
13130
13131
13132
13133
13134
13135
13136
13137
13138
13139
13140
13141
13142
13143
13144
13145
13146
13147
13148
13149
13150
13151
13152
13153
13154
13155
13156
13157
13158
13159
13160
13161
13162
13163
13164
13165
13166
13167
13168
13169
13170
13171
13172
13173
13174
13175
13176
13177
13178
13179
13180
13181
13182
13183
13184
13185
13186
13187
13188
13189
13190
13191
13192
13193
13194
13195
13196
13197
13198
13199
13200
13201
13202
13203
13204
13205
13206
13207
13208
13209
13210
13211
13212
13213
13214
13215
13216
13217
13218
13219
13220
13221
13222
13223
13224
13225
13226
13227
13228
13229
13230
13231
13232
13233
13234
13235
13236
13237
13238
13239
13240
13241
13242
13243
13244
13245
13246
13247
13248
13249
13250
13251
13252
13253
13254
13255
13256
13257
13258
13259
13260
13261
13262
13263
13264
13265
13266
13267
13268
13269
13270
13271
13272
13273
13274
13275
13276
13277
13278
13279
13280
13281
13282
13283
13284
13285
13286
13287
13288
13289
13290
13291
13292
13293
13294
13295
13296
13297
13298
13299
13300
13301
13302
13303
13304
13305
13306
13307
13308
13309
13310
13311
13312
13313
13314
13315
13316
13317
13318
13319
13320
13321
13322
13323
13324
13325
13326
13327
13328
13329
13330
13331
13332
13333
13334
13335
13336
13337
13338
13339
13340
13341
13342
13343
13344
13345
13346
13347
13348
13349
13350
13351
13352
13353
13354
13355
13356
13357
13358
13359
13360
13361
13362
13363
13364
13365
13366
13367
13368
13369
13370
13371
13372
13373
13374
13375
13376
13377
13378
13379
13380
13381
13382
13383
13384
13385
13386
13387
13388
13389
13390
13391
13392
13393
13394
13395
13396
13397
13398
13399
13400
13401
13402
13403
13404
13405
13406
13407
13408
13409
13410
13411
13412
13413
13414
13415
13416
13417
13418
13419
13420
13421
13422
13423
13424
13425
13426
13427
13428
13429
13430
13431
13432
13433
13434
13435
13436
13437
13438
13439
13440
13441
13442
13443
13444
13445
13446
13447
13448
13449
13450
13451
13452
13453
13454
13455
13456
13457
13458
13459
13460
13461
13462
13463
13464
13465
13466
13467
13468
13469
13470
13471
13472
13473
13474
13475
13476
13477
13478
13479
13480
13481
13482
13483
13484
13485
13486
13487
13488
13489
13490
13491
13492
13493
13494
13495
13496
13497
13498
13499
13500
13501
13502
13503
13504
13505
13506
13507
13508
13509
13510
13511
13512
13513
13514
13515
13516
13517
13518
13519
13520
13521
13522
13523
13524
13525
13526
13527
13528
13529
13530
13531
13532
13533
13534
13535
13536
13537
13538
13539
13540
13541
13542
13543
13544
13545
13546
13547
13548
13549
13550
13551
13552
13553
13554
13555
13556
13557
13558
13559
13560
13561
13562
13563
13564
13565
13566
13567
13568
13569
13570
13571
13572
13573
13574
13575
13576
13577
13578
13579
13580
13581
13582
13583
13584
13585
13586
13587
13588
13589
13590
13591
13592
13593
13594
13595
13596
13597
13598
13599
13600
13601
13602
13603
13604
13605
13606
13607
13608
13609
13610
13611
13612
13613
13614
13615
13616
13617
13618
13619
13620
13621
13622
13623
13624
13625
13626
13627
13628
13629
13630
13631
13632
13633
13634
13635
13636
13637
13638
13639
13640
13641
13642
13643
13644
13645
13646
13647
13648
13649
13650
13651
13652
13653
13654
13655
13656
13657
13658
13659
13660
13661
13662
13663
13664
13665
13666
13667
13668
13669
13670
13671
13672
13673
13674
13675
13676
13677
13678
13679
13680
13681
13682
13683
13684
13685
13686
13687
13688
13689
13690
13691
13692
13693
13694
13695
13696
13697
13698
13699
13700
13701
13702
13703
13704
13705
13706
13707
13708
13709
13710
13711
13712
13713
13714
13715
13716
13717
13718
13719
13720
13721
13722
13723
13724
13725
13726
13727
13728
13729
13730
13731
13732
13733
13734
13735
13736
13737
13738
13739
13740
13741
13742
13743
13744
13745
13746
13747
13748
13749
13750
13751
13752
13753
13754
13755
13756
13757
13758
13759
13760
13761
13762
13763
13764
13765
13766
13767
13768
13769
13770
13771
13772
13773
13774
13775
13776
13777
13778
13779
13780
13781
13782
13783
13784
13785
13786
13787
13788
13789
13790
13791
13792
13793
13794
13795
13796
13797
13798
13799
13800
13801
13802
13803
13804
13805
13806
13807
13808
13809
13810
13811
13812
13813
13814
13815
13816
13817
13818
13819
13820
13821
13822
13823
13824
13825
13826
13827
13828
13829
13830
13831
13832
13833
13834
13835
13836
13837
13838
13839
13840
13841
13842
13843
13844
13845
13846
13847
13848
13849
13850
13851
13852
13853
13854
13855
13856
13857
13858
13859
13860
13861
13862
13863
13864
13865
13866
13867
13868
13869
13870
13871
13872
13873
13874
13875
13876
13877
13878
13879
13880
13881
13882
13883
13884
13885
13886
13887
13888
13889
13890
13891
13892
13893
13894
13895
13896
13897
13898
13899
13900
13901
13902
13903
13904
13905
13906
13907
13908
13909
13910
13911
13912
13913
13914
13915
13916
13917
13918
13919
13920
13921
13922
13923
13924
13925
13926
13927
13928
13929
13930
13931
13932
13933
13934
13935
13936
13937
13938
13939
13940
13941
13942
13943
13944
13945
13946
13947
13948
13949
13950
13951
13952
13953
13954
13955
13956
13957
13958
13959
13960
13961
13962
13963
13964
13965
13966
13967
13968
13969
13970
13971
13972
13973
13974
13975
13976
13977
13978
13979
13980
13981
13982
13983
13984
13985
13986
13987
13988
13989
13990
13991
13992
13993
13994
13995
13996
13997
13998
13999
14000
14001
14002
14003
14004
14005
14006
14007
14008
14009
14010
14011
14012
14013
14014
14015
14016
14017
14018
14019
14020
14021
14022
14023
14024
14025
14026
14027
14028
14029
14030
14031
14032
14033
14034
14035
14036
14037
14038
14039
14040
14041
14042
14043
14044
14045
14046
14047
14048
14049
14050
14051
14052
14053
14054
14055
14056
14057
14058
14059
14060
14061
14062
14063
14064
14065
14066
14067
14068
14069
14070
14071
14072
14073
14074
14075
14076
14077
14078
14079
14080
14081
14082
14083
14084
14085
14086
14087
14088
14089
14090
14091
14092
14093
14094
14095
14096
14097
14098
14099
14100
14101
14102
14103
14104
14105
14106
14107
14108
14109
14110
14111
14112
14113
14114
14115
14116
14117
14118
14119
14120
14121
14122
14123
14124
14125
14126
14127
14128
14129
14130
14131
14132
14133
14134
14135
14136
14137
14138
14139
14140
14141
14142
14143
14144
14145
14146
14147
14148
14149
14150
14151
14152
14153
14154
14155
14156
14157
14158
14159
14160
14161
14162
14163
14164
14165
14166
14167
14168
14169
14170
14171
14172
14173
14174
14175
14176
14177
14178
14179
14180
14181
14182
14183
14184
14185
14186
14187
14188
14189
14190
14191
14192
14193
14194
14195
14196
14197
14198
14199
14200
14201
14202
14203
14204
14205
14206
14207
14208
14209
14210
14211
14212
14213
14214
14215
14216
14217
14218
14219
14220
14221
14222
14223
14224
14225
14226
14227
14228
14229
14230
14231
14232
14233
14234
14235
14236
14237
14238
14239
14240
14241
14242
14243
14244
14245
14246
14247
14248
14249
14250
14251
14252
14253
14254
14255
14256
14257
14258
14259
14260
14261
14262
14263
14264
14265
14266
14267
14268
14269
14270
14271
14272
14273
14274
14275
14276
14277
14278
14279
14280
14281
14282
14283
14284
14285
14286
14287
14288
14289
14290
14291
14292
14293
14294
14295
14296
14297
14298
14299
14300
14301
14302
14303
14304
14305
14306
14307
14308
14309
14310
14311
14312
14313
14314
14315
14316
14317
14318
14319
14320
14321
14322
14323
14324
14325
14326
14327
14328
14329
14330
14331
14332
14333
14334
14335
14336
14337
14338
14339
14340
14341
14342
14343
14344
14345
14346
14347
14348
14349
14350
14351
14352
14353
14354
14355
14356
14357
14358
14359
14360
14361
14362
14363
14364
14365
14366
14367
14368
14369
14370
14371
14372
14373
14374
14375
14376
14377
14378
14379
14380
14381
14382
14383
14384
14385
14386
14387
14388
14389
14390
14391
14392
14393
14394
14395
14396
14397
14398
14399
14400
14401
14402
14403
14404
14405
14406
14407
14408
14409
14410
14411
14412
14413
14414
14415
14416
14417
14418
14419
14420
14421
14422
14423
14424
14425
14426
14427
14428
14429
14430
14431
14432
14433
14434
14435
14436
14437
14438
14439
14440
14441
14442
14443
14444
14445
14446
14447
14448
14449
14450
14451
14452
14453
14454
14455
14456
14457
14458
14459
14460
14461
14462
14463
14464
14465
14466
14467
14468
14469
14470
14471
14472
14473
14474
14475
14476
14477
14478
14479
14480
14481
14482
14483
14484
14485
14486
14487
14488
14489
14490
14491
14492
14493
14494
14495
14496
14497
14498
14499
14500
14501
14502
14503
14504
14505
14506
14507
14508
14509
14510
14511
14512
14513
14514
14515
14516
14517
14518
14519
14520
14521
14522
14523
14524
14525
14526
14527
14528
14529
14530
14531
14532
14533
14534
14535
14536
14537
14538
14539
14540
14541
14542
14543
14544
14545
14546
14547
14548
14549
14550
14551
14552
14553
14554
14555
14556
14557
14558
14559
14560
14561
14562
14563
14564
14565
14566
14567
14568
14569
14570
14571
14572
14573
14574
14575
14576
14577
14578
14579
14580
14581
14582
14583
14584
14585
14586
14587
14588
14589
14590
14591
14592
14593
14594
14595
14596
14597
14598
14599
14600
14601
14602
14603
14604
14605
14606
14607
14608
14609
14610
14611
14612
14613
14614
14615
14616
14617
14618
14619
14620
14621
14622
14623
14624
14625
14626
14627
14628
14629
14630
14631
14632
14633
14634
14635
14636
14637
14638
14639
14640
14641
14642
14643
14644
14645
14646
14647
14648
14649
14650
14651
14652
14653
14654
14655
14656
14657
14658
14659
14660
14661
14662
14663
14664
14665
14666
14667
14668
14669
14670
14671
14672
14673
14674
14675
14676
14677
14678
14679
14680
14681
14682
14683
14684
14685
14686
14687
14688
14689
14690
14691
14692
14693
14694
14695
14696
14697
14698
14699
14700
14701
14702
14703
14704
14705
14706
14707
14708
14709
14710
14711
14712
14713
14714
14715
14716
14717
14718
14719
14720
14721
14722
14723
14724
14725
14726
14727
14728
14729
14730
14731
14732
14733
14734
14735
14736
14737
14738
14739
14740
14741
14742
14743
14744
14745
14746
14747
14748
14749
14750
14751
14752
14753
14754
14755
14756
14757
14758
14759
14760
14761
14762
14763
14764
14765
14766
14767
14768
14769
14770
14771
14772
14773
14774
14775
14776
14777
14778
14779
14780
14781
14782
14783
14784
14785
14786
14787
14788
14789
14790
14791
14792
14793
14794
14795
14796
14797
14798
14799
14800
14801
14802
14803
14804
14805
14806
14807
14808
14809
14810
14811
14812
14813
14814
14815
14816
14817
14818
14819
14820
14821
14822
14823
14824
14825
14826
14827
14828
14829
14830
14831
14832
14833
14834
14835
14836
14837
14838
14839
14840
14841
14842
14843
14844
14845
14846
14847
14848
14849
14850
14851
14852
14853
14854
14855
14856
14857
14858
14859
14860
14861
14862
14863
14864
14865
14866
14867
14868
14869
14870
14871
14872
14873
14874
14875
14876
14877
14878
14879
14880
14881
14882
14883
14884
14885
14886
14887
14888
14889
14890
14891
14892
14893
14894
14895
14896
14897
14898
14899
14900
14901
14902
14903
14904
14905
14906
14907
14908
14909
14910
14911
14912
14913
14914
14915
14916
14917
14918
14919
14920
14921
14922
14923
14924
14925
14926
14927
14928
14929
14930
14931
14932
14933
14934
14935
14936
14937
14938
14939
14940
14941
14942
14943
14944
14945
14946
14947
14948
14949
14950
14951
14952
14953
14954
14955
14956
14957
14958
14959
14960
14961
14962
14963
14964
14965
14966
14967
14968
14969
14970
14971
14972
14973
14974
14975
14976
14977
14978
14979
14980
14981
14982
14983
14984
14985
14986
14987
14988
14989
14990
14991
14992
14993
14994
14995
14996
14997
14998
14999
15000
15001
15002
15003
15004
15005
15006
15007
15008
15009
15010
15011
15012
15013
15014
15015
15016
15017
15018
15019
15020
15021
15022
15023
15024
15025
15026
15027
15028
15029
15030
15031
15032
15033
15034
15035
15036
15037
15038
15039
15040
15041
15042
15043
15044
15045
15046
15047
15048
15049
15050
15051
15052
15053
15054
15055
15056
15057
15058
15059
15060
15061
15062
15063
15064
15065
15066
15067
15068
15069
15070
15071
15072
15073
15074
15075
15076
15077
15078
15079
15080
15081
15082
15083
15084
15085
15086
15087
15088
15089
15090
15091
15092
15093
15094
15095
15096
15097
15098
15099
15100
15101
15102
15103
15104
15105
15106
15107
15108
15109
15110
15111
15112
15113
15114
15115
15116
15117
15118
15119
15120
15121
15122
15123
15124
15125
15126
15127
15128
15129
15130
15131
15132
15133
15134
15135
15136
15137
15138
15139
15140
15141
15142
15143
15144
15145
15146
15147
15148
15149
15150
15151
15152
15153
15154
15155
15156
15157
15158
15159
15160
15161
15162
15163
15164
15165
15166
15167
15168
15169
15170
15171
15172
15173
15174
15175
15176
15177
15178
15179
15180
15181
15182
15183
15184
15185
15186
15187
15188
15189
15190
15191
15192
15193
15194
15195
15196
15197
15198
15199
15200
15201
15202
15203
15204
15205
15206
15207
15208
15209
15210
15211
15212
15213
15214
15215
15216
15217
15218
15219
15220
15221
15222
15223
15224
15225
15226
15227
15228
15229
15230
15231
15232
15233
15234
15235
15236
15237
15238
15239
15240
15241
15242
15243
15244
15245
15246
15247
15248
15249
15250
15251
15252
15253
15254
15255
15256
15257
15258
15259
15260
15261
15262
15263
15264
15265
15266
15267
15268
15269
15270
15271
15272
15273
15274
15275
15276
15277
15278
15279
15280
15281
15282
15283
15284
15285
15286
15287
15288
15289
15290
15291
15292
15293
15294
15295
15296
15297
15298
15299
15300
15301
15302
15303
15304
15305
15306
15307
15308
15309
15310
15311
15312
15313
15314
15315
15316
15317
15318
15319
15320
15321
15322
15323
15324
15325
15326
15327
15328
15329
15330
15331
15332
15333
15334
15335
15336
15337
15338
15339
15340
15341
15342
15343
15344
15345
15346
15347
15348
15349
15350
15351
15352
15353
15354
15355
15356
15357
15358
15359
15360
15361
15362
15363
15364
15365
15366
15367
15368
15369
15370
15371
15372
15373
15374
15375
15376
15377
15378
15379
15380
15381
15382
15383
15384
15385
15386
15387
15388
15389
15390
15391
15392
15393
15394
15395
15396
15397
15398
15399
15400
15401
15402
15403
15404
15405
15406
15407
15408
15409
15410
15411
15412
15413
15414
15415
15416
15417
15418
15419
15420
15421
15422
15423
15424
15425
15426
15427
15428
15429
15430
15431
15432
15433
15434
15435
15436
15437
15438
15439
15440
15441
15442
15443
15444
15445
15446
15447
15448
15449
15450
15451
15452
15453
15454
15455
15456
15457
15458
15459
15460
15461
15462
15463
15464
15465
15466
15467
15468
15469
15470
15471
15472
15473
15474
15475
15476
15477
15478
15479
15480
15481
15482
15483
15484
15485
15486
15487
15488
15489
15490
15491
15492
15493
15494
15495
15496
15497
15498
15499
15500
15501
15502
15503
15504
15505
15506
15507
15508
15509
15510
15511
15512
15513
15514
15515
15516
15517
15518
15519
15520
15521
15522
15523
15524
15525
15526
15527
15528
15529
15530
15531
15532
15533
15534
15535
15536
15537
15538
15539
15540
15541
15542
15543
15544
15545
15546
15547
15548
15549
15550
15551
15552
15553
15554
15555
15556
15557
15558
15559
15560
15561
15562
15563
15564
15565
15566
15567
15568
15569
15570
15571
15572
15573
15574
15575
15576
15577
15578
15579
15580
15581
15582
15583
15584
15585
15586
15587
15588
15589
15590
15591
15592
15593
15594
15595
15596
15597
15598
15599
15600
15601
15602
15603
15604
15605
15606
15607
15608
15609
15610
15611
15612
15613
15614
15615
15616
15617
15618
15619
15620
15621
15622
15623
15624
15625
15626
15627
15628
15629
15630
15631
15632
15633
15634
15635
15636
15637
15638
15639
15640
15641
15642
15643
15644
15645
15646
15647
15648
15649
15650
15651
15652
15653
15654
15655
15656
15657
15658
15659
15660
15661
15662
15663
15664
15665
15666
15667
15668
15669
15670
15671
15672
15673
15674
15675
15676
15677
15678
15679
15680
15681
15682
15683
15684
15685
15686
15687
15688
15689
15690
15691
15692
15693
15694
15695
15696
15697
15698
15699
15700
15701
15702
15703
15704
15705
15706
15707
15708
15709
15710
15711
15712
15713
15714
15715
15716
15717
15718
15719
15720
15721
15722
15723
15724
15725
15726
15727
15728
15729
15730
15731
15732
15733
15734
15735
15736
15737
15738
15739
15740
15741
15742
15743
15744
15745
15746
15747
15748
15749
15750
15751
15752
15753
15754
15755
15756
15757
15758
15759
15760
15761
15762
15763
15764
15765
15766
15767
15768
15769
15770
15771
15772
15773
15774
15775
15776
15777
15778
15779
15780
15781
15782
15783
15784
15785
15786
15787
15788
15789
15790
15791
15792
15793
15794
15795
15796
15797
15798
15799
15800
15801
15802
15803
15804
15805
15806
15807
15808
15809
15810
15811
15812
15813
15814
15815
15816
15817
15818
15819
15820
15821
15822
15823
15824
15825
15826
15827
15828
15829
15830
15831
15832
15833
15834
15835
15836
15837
15838
15839
15840
15841
15842
15843
15844
15845
15846
15847
15848
15849
15850
15851
15852
15853
15854
15855
15856
15857
15858
15859
15860
15861
15862
15863
15864
15865
15866
15867
15868
15869
15870
15871
15872
15873
15874
15875
15876
15877
15878
15879
15880
15881
15882
15883
15884
15885
15886
15887
15888
15889
15890
15891
15892
15893
15894
15895
15896
15897
15898
15899
15900
15901
15902
15903
15904
15905
15906
15907
15908
15909
15910
15911
15912
15913
15914
15915
15916
15917
15918
15919
15920
15921
15922
15923
15924
15925
15926
15927
15928
15929
15930
15931
15932
15933
15934
15935
15936
15937
15938
15939
15940
15941
15942
15943
15944
15945
15946
15947
15948
15949
15950
15951
15952
15953
15954
15955
15956
15957
15958
15959
15960
15961
15962
15963
15964
15965
15966
15967
15968
15969
15970
15971
15972
15973
15974
15975
15976
15977
15978
15979
15980
15981
15982
15983
15984
15985
15986
15987
15988
15989
15990
15991
15992
15993
15994
15995
15996
15997
15998
15999
16000
16001
16002
16003
16004
16005
16006
16007
16008
16009
16010
16011
16012
16013
16014
16015
16016
16017
16018
16019
16020
16021
16022
16023
16024
16025
16026
16027
16028
16029
16030
16031
16032
16033
16034
16035
16036
16037
16038
16039
16040
16041
16042
16043
16044
16045
16046
16047
16048
16049
16050
16051
16052
16053
16054
16055
16056
16057
16058
16059
16060
16061
16062
16063
16064
16065
16066
16067
16068
16069
16070
16071
16072
16073
16074
16075
16076
16077
16078
16079
16080
16081
16082
16083
16084
16085
16086
16087
16088
16089
16090
16091
16092
16093
16094
16095
16096
16097
16098
16099
16100
16101
16102
16103
16104
16105
16106
16107
16108
16109
16110
16111
16112
16113
16114
16115
16116
16117
16118
16119
16120
16121
16122
16123
16124
16125
16126
16127
16128
16129
16130
16131
16132
16133
16134
16135
16136
16137
16138
16139
16140
16141
16142
16143
16144
16145
16146
16147
16148
16149
16150
16151
16152
16153
16154
16155
16156
16157
16158
16159
16160
16161
16162
16163
16164
16165
16166
16167
16168
16169
16170
16171
16172
16173
16174
16175
16176
16177
16178
16179
16180
16181
16182
16183
16184
16185
16186
16187
16188
16189
16190
16191
16192
16193
16194
16195
16196
16197
16198
16199
16200
16201
16202
16203
16204
16205
16206
16207
16208
16209
16210
16211
16212
16213
16214
16215
16216
16217
16218
16219
16220
16221
16222
16223
16224
16225
16226
16227
16228
16229
16230
16231
16232
16233
16234
16235
16236
16237
16238
16239
16240
16241
16242
16243
16244
16245
16246
16247
16248
16249
16250
16251
16252
16253
16254
16255
16256
16257
16258
16259
16260
16261
16262
16263
16264
16265
16266
16267
16268
16269
16270
16271
16272
16273
16274
16275
16276
16277
16278
16279
16280
16281
16282
16283
16284
16285
16286
16287
16288
16289
16290
16291
16292
16293
16294
16295
16296
16297
16298
16299
16300
16301
16302
16303
16304
16305
16306
16307
16308
16309
16310
16311
16312
16313
16314
16315
16316
16317
16318
16319
16320
16321
16322
16323
16324
16325
16326
16327
16328
16329
16330
16331
16332
16333
16334
16335
16336
16337
16338
16339
16340
16341
16342
16343
16344
16345
16346
16347
16348
16349
16350
16351
16352
16353
16354
16355
16356
16357
16358
16359
16360
16361
16362
16363
16364
16365
16366
16367
16368
16369
16370
16371
16372
16373
16374
16375
16376
16377
16378
16379
16380
16381
16382
16383
16384
16385
16386
16387
16388
16389
16390
16391
16392
16393
16394
16395
16396
16397
16398
16399
16400
16401
16402
16403
16404
16405
16406
16407
16408
16409
16410
16411
16412
16413
16414
16415
16416
16417
16418
16419
16420
16421
16422
16423
16424
16425
16426
16427
16428
16429
16430
16431
16432
16433
16434
16435
16436
16437
16438
16439
16440
16441
16442
16443
16444
16445
16446
16447
16448
16449
16450
16451
16452
16453
16454
16455
16456
16457
16458
16459
16460
16461
16462
16463
16464
16465
16466
16467
16468
16469
16470
16471
16472
16473
16474
16475
16476
16477
16478
16479
16480
16481
16482
16483
16484
16485
16486
16487
16488
16489
16490
16491
16492
16493
16494
16495
16496
16497
16498
16499
16500
16501
16502
16503
16504
16505
16506
16507
16508
16509
16510
16511
16512
16513
16514
16515
16516
16517
16518
16519
16520
16521
16522
16523
16524
16525
16526
16527
16528
16529
16530
16531
16532
16533
16534
16535
16536
16537
16538
16539
16540
16541
16542
16543
16544
16545
16546
16547
16548
16549
16550
16551
16552
16553
16554
16555
16556
16557
16558
16559
16560
16561
16562
16563
16564
16565
16566
16567
16568
16569
16570
16571
16572
16573
16574
16575
16576
16577
16578
16579
16580
16581
16582
16583
16584
16585
16586
16587
16588
16589
16590
16591
16592
16593
16594
16595
16596
16597
16598
16599
16600
16601
16602
16603
16604
16605
16606
16607
16608
16609
16610
16611
16612
16613
16614
16615
16616
16617
16618
16619
16620
16621
16622
16623
16624
16625
16626
16627
16628
16629
16630
16631
16632
16633
16634
16635
16636
16637
16638
16639
16640
16641
16642
16643
16644
16645
16646
16647
16648
16649
16650
16651
16652
16653
16654
16655
16656
16657
16658
16659
16660
16661
16662
16663
16664
16665
16666
16667
16668
16669
16670
16671
16672
16673
16674
16675
16676
16677
16678
16679
16680
16681
16682
16683
16684
16685
16686
16687
16688
16689
16690
16691
16692
16693
16694
16695
16696
16697
16698
16699
16700
16701
16702
16703
16704
16705
16706
16707
16708
16709
16710
16711
16712
16713
16714
16715
16716
16717
16718
16719
16720
16721
16722
16723
16724
16725
16726
16727
16728
16729
16730
16731
16732
16733
16734
16735
16736
16737
16738
16739
16740
16741
16742
16743
16744
16745
16746
16747
16748
16749
16750
16751
16752
16753
16754
16755
16756
16757
16758
16759
16760
16761
16762
16763
16764
16765
16766
16767
16768
16769
16770
16771
16772
16773
16774
16775
16776
16777
16778
16779
16780
16781
16782
16783
16784
16785
16786
16787
16788
16789
16790
16791
16792
16793
16794
16795
16796
16797
16798
16799
16800
16801
16802
16803
16804
16805
16806
16807
16808
16809
16810
16811
16812
16813
16814
16815
16816
16817
16818
16819
16820
16821
16822
16823
16824
16825
16826
16827
16828
16829
16830
16831
16832
16833
16834
16835
16836
16837
16838
16839
16840
16841
16842
16843
16844
16845
16846
16847
16848
16849
16850
16851
16852
16853
16854
16855
16856
16857
16858
16859
16860
16861
16862
16863
16864
16865
16866
16867
16868
16869
16870
16871
16872
16873
16874
16875
16876
16877
16878
16879
16880
16881
16882
16883
16884
16885
16886
16887
16888
16889
16890
16891
16892
16893
16894
16895
16896
16897
16898
16899
16900
16901
16902
16903
16904
16905
16906
16907
16908
16909
16910
16911
16912
16913
16914
16915
16916
16917
16918
16919
16920
16921
16922
16923
16924
16925
16926
16927
16928
16929
16930
16931
16932
16933
16934
16935
16936
16937
16938
16939
16940
16941
16942
16943
16944
16945
16946
16947
16948
16949
16950
16951
16952
16953
16954
16955
16956
16957
16958
16959
16960
16961
16962
16963
16964
16965
16966
16967
16968
16969
16970
16971
16972
16973
16974
16975
16976
16977
16978
16979
16980
16981
16982
16983
16984
16985
16986
16987
16988
16989
16990
16991
16992
16993
16994
16995
16996
16997
16998
16999
17000
17001
17002
17003
17004
17005
17006
17007
17008
17009
17010
17011
17012
17013
17014
17015
17016
17017
17018
17019
17020
17021
17022
17023
17024
17025
17026
17027
17028
17029
17030
17031
17032
17033
17034
17035
17036
17037
17038
17039
17040
17041
17042
17043
17044
17045
17046
17047
17048
17049
17050
17051
17052
17053
17054
17055
17056
17057
17058
17059
17060
17061
17062
17063
17064
17065
17066
17067
17068
17069
17070
17071
17072
17073
17074
17075
17076
17077
17078
17079
17080
17081
17082
17083
17084
17085
17086
17087
17088
17089
17090
17091
17092
17093
17094
17095
17096
17097
17098
17099
17100
17101
17102
17103
17104
17105
17106
17107
17108
17109
17110
17111
17112
17113
17114
17115
17116
17117
17118
17119
17120
17121
17122
17123
17124
17125
17126
17127
17128
17129
17130
17131
17132
17133
17134
17135
17136
17137
17138
17139
17140
17141
17142
17143
17144
17145
17146
17147
17148
17149
17150
17151
17152
17153
17154
17155
17156
17157
17158
17159
17160
17161
17162
17163
17164
17165
17166
17167
17168
17169
17170
17171
17172
17173
17174
17175
17176
17177
17178
17179
17180
17181
17182
17183
17184
17185
17186
17187
17188
17189
17190
17191
17192
17193
17194
17195
17196
17197
17198
17199
17200
17201
17202
17203
17204
17205
17206
17207
17208
17209
17210
17211
17212
17213
17214
17215
17216
17217
17218
17219
17220
17221
17222
17223
17224
17225
17226
17227
17228
17229
17230
17231
17232
17233
17234
17235
17236
17237
17238
17239
17240
17241
17242
17243
17244
17245
17246
17247
17248
17249
17250
17251
17252
17253
17254
17255
17256
17257
17258
17259
17260
17261
17262
17263
17264
17265
17266
17267
17268
17269
17270
17271
17272
17273
17274
17275
17276
17277
17278
17279
17280
17281
17282
17283
17284
17285
17286
17287
17288
17289
17290
17291
17292
17293
17294
17295
17296
17297
17298
17299
17300
17301
17302
17303
17304
17305
17306
17307
17308
17309
17310
17311
17312
17313
17314
17315
17316
17317
17318
17319
17320
17321
17322
17323
17324
17325
17326
17327
17328
17329
17330
17331
17332
17333
17334
17335
17336
17337
17338
17339
17340
17341
17342
17343
17344
17345
17346
17347
17348
17349
17350
17351
17352
17353
17354
17355
17356
17357
17358
17359
17360
17361
17362
17363
17364
17365
17366
17367
17368
17369
17370
17371
17372
17373
17374
17375
17376
17377
17378
17379
17380
17381
17382
17383
17384
17385
17386
17387
17388
17389
17390
17391
17392
17393
17394
17395
17396
17397
17398
17399
17400
17401
17402
17403
17404
17405
17406
17407
17408
17409
17410
17411
17412
17413
17414
17415
17416
17417
17418
17419
17420
17421
17422
17423
17424
17425
17426
17427
17428
17429
17430
17431
17432
17433
17434
17435
17436
17437
17438
17439
17440
17441
17442
17443
17444
17445
17446
17447
17448
17449
17450
17451
17452
17453
17454
17455
17456
17457
17458
17459
17460
17461
17462
17463
17464
17465
17466
17467
17468
17469
17470
17471
17472
17473
17474
17475
17476
17477
17478
17479
17480
17481
17482
17483
17484
17485
17486
17487
17488
17489
17490
17491
17492
17493
17494
17495
17496
17497
17498
17499
17500
17501
17502
17503
17504
17505
17506
17507
17508
17509
17510
17511
17512
17513
17514
17515
17516
17517
17518
17519
17520
17521
17522
17523
17524
17525
17526
17527
17528
17529
17530
17531
17532
17533
17534
17535
17536
17537
17538
17539
17540
17541
17542
17543
17544
17545
17546
17547
17548
17549
17550
17551
17552
17553
17554
17555
17556
17557
17558
17559
17560
17561
17562
17563
17564
17565
17566
17567
17568
17569
17570
17571
17572
17573
17574
17575
17576
17577
17578
17579
17580
17581
17582
17583
17584
17585
17586
17587
17588
17589
17590
17591
17592
17593
17594
17595
17596
17597
17598
17599
17600
17601
17602
17603
17604
17605
17606
17607
17608
17609
17610
17611
17612
17613
17614
17615
17616
17617
17618
17619
17620
17621
17622
17623
17624
17625
17626
17627
17628
17629
17630
17631
17632
17633
17634
17635
17636
17637
17638
17639
17640
17641
17642
17643
17644
17645
17646
17647
17648
17649
17650
17651
17652
17653
17654
17655
17656
17657
17658
17659
17660
17661
17662
17663
17664
17665
17666
17667
17668
17669
17670
17671
17672
17673
17674
17675
17676
17677
17678
17679
17680
17681
17682
17683
17684
17685
17686
17687
17688
17689
17690
17691
17692
17693
17694
17695
17696
17697
17698
17699
17700
17701
17702
17703
17704
17705
17706
17707
17708
17709
17710
17711
17712
17713
17714
17715
17716
17717
17718
17719
17720
17721
17722
17723
17724
17725
17726
17727
17728
17729
17730
17731
17732
17733
17734
17735
17736
17737
17738
17739
17740
17741
17742
17743
17744
17745
17746
17747
17748
17749
17750
17751
17752
17753
17754
17755
17756
17757
17758
17759
17760
17761
17762
17763
17764
17765
17766
17767
17768
17769
17770
17771
17772
17773
17774
17775
17776
17777
17778
17779
17780
17781
17782
17783
17784
17785
17786
17787
17788
17789
17790
17791
17792
17793
17794
17795
17796
17797
17798
17799
17800
17801
17802
17803
17804
17805
17806
17807
17808
17809
17810
17811
17812
17813
17814
17815
17816
17817
17818
17819
17820
17821
17822
17823
17824
17825
17826
17827
17828
17829
17830
17831
17832
17833
17834
17835
17836
17837
17838
17839
17840
17841
17842
17843
17844
17845
17846
17847
17848
17849
17850
17851
17852
17853
17854
17855
17856
17857
17858
17859
17860
17861
17862
17863
17864
17865
17866
17867
17868
17869
17870
17871
17872
17873
17874
17875
17876
17877
17878
17879
17880
17881
17882
17883
17884
17885
17886
17887
17888
17889
17890
17891
17892
17893
17894
17895
17896
17897
17898
17899
17900
17901
17902
17903
17904
17905
17906
17907
17908
17909
17910
17911
17912
17913
17914
17915
17916
17917
17918
17919
17920
17921
17922
17923
17924
17925
17926
17927
17928
17929
17930
17931
17932
17933
17934
17935
17936
17937
17938
17939
17940
17941
17942
17943
17944
17945
17946
17947
17948
17949
17950
17951
17952
17953
17954
17955
17956
17957
17958
17959
17960
17961
17962
17963
17964
17965
17966
17967
17968
17969
17970
17971
17972
17973
17974
17975
17976
17977
17978
17979
17980
17981
17982
17983
17984
17985
17986
17987
17988
17989
17990
17991
17992
17993
17994
17995
17996
17997
17998
17999
18000
18001
18002
18003
18004
18005
18006
18007
18008
18009
18010
18011
18012
18013
18014
18015
18016
18017
18018
18019
18020
18021
18022
18023
18024
18025
18026
18027
18028
18029
18030
18031
18032
18033
18034
18035
18036
18037
18038
18039
18040
18041
18042
18043
18044
18045
18046
18047
18048
18049
18050
18051
18052
18053
18054
18055
18056
18057
18058
18059
18060
18061
18062
18063
18064
18065
18066
18067
18068
18069
18070
18071
18072
18073
18074
18075
18076
18077
18078
18079
18080
18081
18082
18083
18084
18085
18086
18087
18088
18089
18090
18091
18092
18093
18094
18095
18096
18097
18098
18099
18100
18101
18102
18103
18104
18105
18106
18107
18108
18109
18110
18111
18112
18113
18114
18115
18116
18117
18118
18119
18120
18121
18122
18123
18124
18125
18126
18127
18128
18129
18130
18131
18132
18133
18134
18135
18136
18137
18138
18139
18140
18141
18142
18143
18144
18145
18146
18147
18148
18149
18150
18151
18152
18153
18154
18155
18156
18157
18158
18159
18160
18161
18162
18163
18164
18165
18166
18167
18168
18169
18170
18171
18172
18173
18174
18175
18176
18177
18178
18179
18180
18181
18182
18183
18184
18185
18186
18187
18188
18189
18190
18191
18192
18193
18194
18195
18196
18197
18198
18199
18200
18201
18202
18203
18204
18205
18206
18207
18208
18209
18210
18211
18212
18213
18214
18215
18216
18217
18218
18219
18220
18221
18222
18223
18224
18225
18226
18227
18228
18229
18230
18231
18232
18233
18234
18235
18236
18237
18238
18239
18240
18241
18242
18243
18244
18245
18246
18247
18248
18249
18250
18251
18252
18253
18254
18255
18256
18257
18258
18259
18260
18261
18262
18263
18264
18265
18266
18267
18268
18269
18270
18271
18272
18273
18274
18275
18276
18277
18278
18279
18280
18281
18282
18283
18284
18285
18286
18287
18288
18289
18290
18291
18292
18293
18294
18295
18296
18297
18298
18299
18300
18301
18302
18303
18304
18305
18306
18307
18308
18309
18310
18311
18312
18313
18314
18315
18316
18317
18318
18319
18320
18321
18322
18323
18324
18325
18326
18327
18328
18329
18330
18331
18332
18333
18334
18335
18336
18337
18338
18339
18340
18341
18342
18343
18344
18345
18346
18347
18348
18349
18350
18351
18352
18353
18354
18355
18356
18357
18358
18359
18360
18361
18362
18363
18364
18365
18366
18367
18368
18369
18370
18371
18372
18373
18374
18375
18376
18377
18378
18379
18380
18381
18382
18383
18384
18385
18386
18387
18388
18389
18390
18391
18392
18393
18394
18395
18396
18397
18398
18399
18400
18401
18402
18403
18404
18405
18406
18407
18408
18409
18410
18411
18412
18413
18414
18415
18416
18417
18418
18419
18420
18421
18422
18423
18424
18425
18426
18427
18428
18429
18430
18431
18432
18433
18434
18435
18436
18437
18438
18439
18440
18441
18442
18443
18444
18445
18446
18447
18448
18449
18450
18451
18452
18453
18454
18455
18456
18457
18458
18459
18460
18461
18462
18463
18464
18465
18466
18467
18468
18469
18470
18471
18472
18473
18474
18475
18476
18477
18478
18479
18480
18481
18482
18483
18484
18485
18486
18487
18488
18489
18490
18491
18492
18493
18494
18495
18496
18497
18498
18499
18500
18501
18502
18503
18504
18505
18506
18507
18508
18509
18510
18511
18512
18513
18514
18515
18516
18517
18518
18519
18520
18521
18522
18523
18524
18525
18526
18527
18528
18529
18530
18531
18532
18533
18534
18535
18536
18537
18538
18539
18540
18541
18542
18543
18544
18545
18546
18547
18548
18549
18550
18551
18552
18553
18554
18555
18556
18557
18558
18559
18560
18561
18562
18563
18564
18565
18566
18567
18568
18569
18570
18571
18572
18573
18574
18575
18576
18577
18578
18579
18580
18581
18582
18583
18584
18585
18586
18587
18588
18589
18590
18591
18592
18593
18594
18595
18596
18597
18598
18599
18600
18601
18602
18603
18604
18605
18606
18607
18608
18609
18610
18611
18612
18613
18614
18615
18616
18617
18618
18619
18620
18621
18622
18623
18624
18625
18626
18627
18628
18629
18630
18631
18632
18633
18634
18635
18636
18637
18638
18639
18640
18641
18642
18643
18644
18645
18646
18647
18648
18649
18650
18651
18652
18653
18654
18655
18656
18657
18658
18659
18660
18661
18662
18663
18664
18665
18666
18667
18668
18669
18670
18671
18672
18673
18674
18675
18676
18677
18678
18679
18680
18681
18682
18683
18684
18685
18686
18687
18688
18689
18690
18691
18692
18693
18694
18695
18696
18697
18698
18699
18700
18701
18702
18703
18704
18705
18706
18707
18708
18709
18710
18711
18712
18713
18714
18715
18716
18717
18718
18719
18720
18721
18722
18723
18724
18725
18726
18727
18728
18729
18730
18731
18732
18733
18734
18735
18736
18737
18738
18739
18740
18741
18742
18743
18744
18745
18746
18747
18748
18749
18750
18751
18752
18753
18754
18755
18756
18757
18758
18759
18760
18761
18762
18763
18764
18765
18766
18767
18768
18769
18770
18771
18772
18773
18774
18775
18776
18777
18778
18779
18780
18781
18782
18783
18784
18785
18786
18787
18788
18789
18790
18791
18792
18793
18794
18795
18796
18797
18798
18799
18800
18801
18802
18803
18804
18805
18806
18807
18808
18809
18810
18811
18812
18813
18814
18815
18816
18817
18818
18819
18820
18821
18822
18823
18824
18825
18826
18827
18828
18829
18830
18831
18832
18833
18834
18835
18836
18837
18838
18839
18840
18841
18842
18843
18844
18845
18846
18847
18848
18849
18850
18851
18852
18853
18854
18855
18856
18857
18858
18859
18860
18861
18862
18863
18864
18865
18866
18867
18868
18869
18870
18871
18872
18873
18874
18875
18876
18877
18878
18879
18880
18881
18882
18883
18884
18885
18886
18887
18888
18889
18890
18891
18892
18893
18894
18895
18896
18897
18898
18899
18900
18901
18902
18903
18904
18905
18906
18907
18908
18909
18910
18911
18912
18913
18914
18915
18916
18917
18918
18919
18920
18921
18922
18923
18924
18925
18926
18927
18928
18929
18930
18931
18932
18933
18934
18935
18936
18937
18938
18939
18940
18941
18942
18943
18944
18945
18946
18947
18948
18949
18950
18951
18952
18953
18954
18955
18956
18957
18958
18959
18960
18961
18962
18963
18964
18965
18966
18967
18968
18969
18970
18971
18972
18973
18974
18975
18976
18977
18978
18979
18980
18981
18982
18983
18984
18985
18986
18987
18988
18989
18990
18991
18992
18993
18994
18995
18996
18997
18998
18999
19000
19001
19002
19003
19004
19005
19006
19007
19008
19009
19010
19011
19012
19013
19014
19015
19016
19017
19018
19019
19020
19021
19022
19023
19024
19025
19026
19027
19028
19029
19030
19031
19032
19033
19034
19035
19036
19037
19038
19039
19040
19041
19042
19043
19044
19045
19046
19047
19048
19049
19050
19051
19052
19053
19054
19055
19056
19057
19058
19059
19060
19061
19062
19063
19064
19065
19066
19067
19068
19069
19070
19071
19072
19073
19074
19075
19076
19077
19078
19079
19080
19081
19082
19083
19084
19085
19086
19087
19088
19089
19090
19091
19092
19093
19094
19095
19096
19097
19098
19099
19100
19101
19102
19103
19104
19105
19106
19107
19108
19109
19110
19111
19112
19113
19114
19115
19116
19117
19118
19119
19120
19121
19122
19123
19124
19125
19126
19127
19128
19129
19130
19131
19132
19133
19134
19135
19136
19137
19138
19139
19140
19141
19142
19143
19144
19145
19146
19147
19148
19149
19150
19151
19152
19153
19154
19155
19156
19157
19158
19159
19160
19161
19162
19163
19164
19165
19166
19167
19168
19169
19170
19171
19172
19173
19174
19175
19176
19177
19178
19179
19180
19181
19182
19183
19184
19185
19186
19187
19188
19189
19190
19191
19192
19193
19194
19195
19196
19197
19198
19199
19200
19201
19202
19203
19204
19205
19206
19207
19208
19209
19210
19211
19212
19213
19214
19215
19216
19217
19218
19219
19220
19221
19222
19223
19224
19225
19226
19227
19228
19229
19230
19231
19232
19233
19234
19235
19236
19237
19238
19239
19240
19241
19242
19243
19244
19245
19246
19247
19248
19249
19250
19251
19252
19253
19254
19255
19256
19257
19258
19259
19260
19261
19262
19263
19264
19265
19266
19267
19268
19269
19270
19271
19272
19273
19274
19275
19276
19277
19278
19279
19280
19281
19282
19283
19284
19285
19286
19287
19288
19289
19290
19291
19292
19293
19294
19295
19296
19297
19298
19299
19300
19301
19302
19303
19304
19305
19306
19307
19308
19309
19310
19311
19312
19313
19314
19315
19316
19317
19318
19319
19320
19321
19322
19323
19324
19325
19326
19327
19328
19329
19330
19331
19332
19333
19334
19335
19336
19337
19338
19339
19340
19341
19342
19343
19344
19345
19346
19347
19348
19349
19350
19351
19352
19353
19354
19355
19356
19357
19358
19359
19360
19361
19362
19363
19364
19365
19366
19367
19368
19369
19370
19371
19372
19373
19374
19375
19376
19377
19378
19379
19380
19381
19382
19383
19384
19385
19386
19387
19388
19389
19390
19391
19392
19393
19394
19395
19396
19397
19398
19399
19400
19401
19402
19403
19404
19405
19406
19407
19408
19409
19410
19411
19412
19413
19414
19415
19416
19417
19418
19419
19420
19421
19422
19423
19424
19425
19426
19427
19428
19429
19430
19431
19432
19433
19434
19435
19436
19437
19438
19439
19440
19441
19442
19443
19444
19445
19446
19447
19448
19449
19450
19451
19452
19453
19454
19455
19456
19457
19458
19459
19460
19461
19462
19463
19464
19465
19466
19467
19468
19469
19470
19471
19472
19473
19474
19475
19476
19477
19478
19479
19480
19481
19482
19483
19484
19485
19486
19487
19488
19489
19490
19491
19492
19493
19494
19495
19496
19497
19498
19499
19500
19501
19502
19503
19504
19505
19506
19507
19508
19509
19510
19511
19512
19513
19514
19515
19516
19517
19518
19519
19520
19521
19522
19523
19524
19525
19526
19527
19528
19529
19530
19531
19532
19533
19534
19535
19536
19537
19538
19539
19540
19541
19542
19543
19544
19545
19546
19547
19548
19549
19550
19551
19552
19553
19554
19555
19556
19557
19558
19559
19560
19561
19562
19563
19564
19565
19566
19567
19568
19569
19570
19571
19572
19573
19574
19575
19576
19577
19578
19579
19580
19581
19582
19583
19584
19585
19586
19587
19588
19589
19590
19591
19592
19593
19594
19595
19596
19597
19598
19599
19600
19601
19602
19603
19604
19605
19606
19607
19608
19609
19610
19611
19612
19613
19614
19615
19616
19617
19618
19619
19620
19621
19622
19623
19624
19625
19626
19627
19628
19629
19630
19631
19632
19633
19634
19635
19636
19637
19638
19639
19640
19641
19642
19643
19644
19645
19646
19647
19648
19649
19650
19651
19652
19653
19654
19655
19656
19657
19658
19659
19660
19661
19662
19663
19664
19665
19666
19667
19668
19669
19670
19671
19672
19673
19674
19675
19676
19677
19678
19679
19680
19681
19682
19683
19684
19685
19686
19687
19688
19689
19690
19691
19692
19693
19694
19695
19696
19697
19698
19699
19700
19701
19702
19703
19704
19705
19706
19707
19708
19709
19710
19711
19712
19713
19714
19715
19716
19717
19718
19719
19720
19721
19722
19723
19724
19725
19726
19727
19728
19729
19730
19731
19732
19733
19734
19735
19736
19737
19738
19739
19740
19741
19742
19743
19744
19745
19746
19747
19748
19749
19750
19751
19752
19753
19754
19755
19756
19757
19758
19759
19760
19761
19762
19763
19764
19765
19766
19767
19768
19769
19770
19771
19772
19773
19774
19775
19776
19777
19778
19779
19780
19781
19782
19783
19784
19785
19786
19787
19788
19789
19790
19791
19792
19793
19794
19795
19796
19797
19798
19799
19800
19801
19802
19803
19804
19805
19806
19807
19808
19809
19810
19811
19812
19813
19814
19815
19816
19817
19818
19819
19820
19821
19822
19823
19824
19825
19826
19827
19828
19829
19830
19831
19832
19833
19834
19835
19836
19837
19838
19839
19840
19841
19842
19843
19844
19845
19846
19847
19848
19849
19850
19851
19852
19853
19854
19855
19856
19857
19858
19859
19860
19861
19862
19863
19864
19865
19866
19867
19868
19869
19870
19871
19872
19873
19874
19875
19876
19877
19878
19879
19880
19881
19882
19883
19884
19885
19886
19887
19888
19889
19890
19891
19892
19893
19894
19895
19896
19897
19898
19899
19900
19901
19902
19903
19904
19905
19906
19907
19908
19909
19910
19911
19912
19913
19914
19915
19916
19917
19918
19919
19920
19921
19922
19923
19924
19925
19926
19927
19928
19929
19930
19931
19932
19933
19934
19935
19936
19937
19938
19939
19940
19941
19942
19943
19944
19945
19946
19947
19948
19949
19950
19951
19952
19953
19954
19955
19956
19957
19958
19959
19960
19961
19962
19963
19964
19965
19966
19967
19968
19969
19970
19971
19972
19973
19974
19975
19976
19977
19978
19979
19980
19981
19982
19983
19984
19985
19986
19987
19988
19989
19990
19991
19992
19993
19994
19995
19996
19997
19998
19999
20000
20001
20002
20003
20004
20005
20006
20007
20008
20009
20010
20011
20012
20013
20014
20015
20016
20017
20018
20019
20020
20021
20022
20023
20024
20025
20026
20027
20028
20029
20030
20031
20032
20033
20034
20035
20036
20037
20038
20039
20040
20041
20042
20043
20044
20045
20046
20047
20048
20049
20050
20051
20052
20053
20054
20055
20056
20057
20058
20059
20060
20061
20062
20063
20064
20065
20066
20067
20068
20069
20070
20071
20072
20073
20074
20075
20076
20077
20078
20079
20080
20081
20082
20083
20084
20085
20086
20087
20088
20089
20090
20091
20092
20093
20094
20095
20096
20097
20098
20099
20100
20101
20102
20103
20104
20105
20106
20107
20108
20109
20110
20111
20112
20113
20114
20115
20116
20117
20118
20119
20120
20121
20122
20123
20124
20125
20126
20127
20128
20129
20130
20131
20132
20133
20134
20135
20136
20137
20138
20139
20140
20141
20142
20143
20144
20145
20146
20147
20148
20149
20150
20151
20152
20153
20154
20155
20156
20157
20158
20159
20160
20161
20162
20163
20164
20165
20166
20167
20168
20169
20170
20171
20172
20173
20174
20175
20176
20177
20178
20179
20180
20181
20182
20183
20184
20185
20186
20187
20188
20189
20190
20191
20192
20193
20194
20195
20196
20197
20198
20199
20200
20201
20202
20203
20204
20205
20206
20207
20208
20209
20210
20211
20212
20213
20214
20215
20216
20217
20218
20219
20220
20221
20222
20223
20224
20225
20226
20227
20228
20229
20230
20231
20232
20233
20234
20235
20236
20237
20238
20239
20240
20241
20242
20243
20244
20245
20246
20247
20248
20249
20250
20251
20252
20253
20254
20255
20256
20257
20258
20259
20260
20261
20262
20263
20264
20265
20266
20267
20268
20269
20270
20271
20272
20273
20274
20275
20276
20277
20278
20279
20280
20281
20282
20283
20284
20285
20286
20287
20288
20289
20290
20291
20292
20293
20294
20295
20296
20297
20298
20299
20300
20301
20302
20303
20304
20305
20306
20307
20308
20309
20310
20311
20312
20313
20314
20315
20316
20317
20318
20319
20320
20321
20322
20323
20324
20325
20326
20327
20328
20329
20330
20331
20332
20333
20334
20335
20336
20337
20338
20339
20340
20341
20342
20343
20344
20345
20346
20347
20348
20349
20350
20351
20352
20353
20354
20355
20356
20357
20358
20359
20360
20361
20362
20363
20364
20365
20366
20367
20368
20369
20370
20371
20372
20373
20374
20375
20376
20377
20378
20379
20380
20381
20382
20383
20384
20385
20386
20387
20388
20389
20390
20391
20392
20393
20394
20395
20396
20397
20398
20399
20400
20401
20402
20403
20404
20405
20406
20407
20408
20409
20410
20411
20412
20413
20414
20415
20416
20417
20418
20419
20420
20421
20422
20423
20424
20425
20426
20427
20428
20429
20430
20431
20432
20433
20434
20435
20436
20437
20438
20439
20440
20441
20442
20443
20444
20445
20446
20447
20448
20449
20450
20451
20452
20453
20454
20455
20456
20457
20458
20459
20460
20461
20462
20463
20464
20465
20466
20467
20468
20469
20470
20471
20472
20473
20474
20475
20476
20477
20478
20479
20480
20481
20482
20483
20484
20485
20486
20487
20488
20489
20490
20491
20492
20493
20494
20495
20496
20497
20498
20499
20500
20501
20502
20503
20504
20505
20506
20507
20508
20509
20510
20511
20512
20513
20514
20515
20516
20517
20518
20519
20520
20521
20522
20523
20524
20525
20526
20527
20528
20529
20530
20531
20532
20533
20534
20535
20536
20537
20538
20539
20540
20541
20542
20543
20544
20545
20546
20547
20548
20549
20550
20551
20552
20553
20554
20555
20556
20557
20558
20559
20560
20561
20562
20563
20564
20565
20566
20567
20568
20569
20570
20571
20572
20573
20574
20575
20576
20577
20578
20579
20580
20581
20582
20583
20584
20585
20586
20587
20588
20589
20590
20591
20592
20593
20594
20595
20596
20597
20598
20599
20600
20601
20602
20603
20604
20605
20606
20607
20608
20609
20610
20611
20612
20613
20614
20615
20616
20617
20618
20619
20620
20621
20622
20623
20624
20625
20626
20627
20628
20629
20630
20631
20632
20633
20634
20635
20636
20637
20638
20639
20640
20641
20642
20643
20644
20645
20646
20647
20648
20649
20650
20651
20652
20653
20654
20655
20656
20657
20658
20659
20660
20661
20662
20663
20664
20665
20666
20667
20668
20669
20670
20671
20672
20673
20674
20675
20676
20677
20678
20679
20680
20681
20682
20683
20684
20685
20686
20687
20688
20689
20690
20691
20692
20693
20694
20695
20696
20697
20698
20699
20700
20701
20702
20703
20704
20705
20706
20707
20708
20709
20710
20711
20712
20713
20714
20715
20716
20717
20718
20719
20720
20721
20722
20723
20724
20725
20726
20727
20728
20729
20730
20731
20732
20733
20734
20735
20736
20737
20738
20739
20740
20741
20742
20743
20744
20745
20746
20747
20748
20749
20750
20751
20752
20753
20754
20755
20756
20757
20758
20759
20760
20761
20762
20763
20764
20765
20766
20767
20768
20769
20770
20771
20772
20773
20774
20775
20776
20777
20778
20779
20780
20781
20782
20783
20784
20785
20786
20787
20788
20789
20790
20791
20792
20793
20794
20795
20796
20797
20798
20799
20800
20801
20802
20803
20804
20805
20806
20807
20808
20809
20810
20811
20812
20813
20814
20815
20816
20817
20818
20819
20820
20821
20822
20823
20824
20825
20826
20827
20828
20829
20830
20831
20832
20833
20834
20835
20836
20837
20838
20839
20840
20841
20842
20843
20844
20845
20846
20847
20848
20849
20850
20851
20852
20853
20854
20855
20856
20857
20858
20859
20860
20861
20862
20863
20864
20865
20866
20867
20868
20869
20870
20871
20872
20873
20874
20875
20876
20877
20878
20879
20880
20881
20882
20883
20884
20885
20886
20887
20888
20889
20890
20891
20892
20893
20894
20895
20896
20897
20898
20899
20900
20901
20902
20903
20904
20905
20906
20907
20908
20909
20910
20911
20912
20913
20914
20915
20916
20917
20918
20919
20920
20921
20922
20923
20924
20925
20926
20927
20928
20929
20930
20931
20932
20933
20934
20935
20936
20937
20938
20939
20940
20941
20942
20943
20944
20945
20946
20947
20948
20949
20950
20951
20952
20953
20954
20955
20956
20957
20958
20959
20960
20961
20962
20963
20964
20965
20966
20967
20968
20969
20970
20971
20972
20973
20974
20975
20976
20977
20978
20979
20980
20981
20982
20983
20984
20985
20986
20987
20988
20989
20990
20991
20992
20993
20994
20995
20996
20997
20998
20999
21000
21001
21002
21003
21004
21005
21006
21007
21008
21009
21010
21011
21012
21013
21014
21015
21016
21017
21018
21019
21020
21021
21022
21023
21024
21025
21026
21027
21028
21029
21030
21031
21032
21033
21034
21035
21036
21037
21038
21039
21040
21041
21042
21043
21044
21045
21046
21047
21048
21049
21050
21051
21052
21053
21054
21055
21056
21057
21058
21059
21060
21061
21062
21063
21064
21065
21066
21067
21068
21069
21070
21071
21072
21073
21074
21075
21076
21077
21078
21079
21080
21081
21082
21083
21084
21085
21086
21087
21088
21089
21090
21091
21092
21093
21094
21095
21096
21097
21098
21099
21100
21101
21102
21103
21104
21105
21106
21107
21108
21109
21110
21111
21112
21113
21114
21115
21116
21117
21118
21119
21120
21121
21122
21123
21124
21125
21126
21127
21128
21129
21130
21131
21132
21133
21134
21135
21136
21137
21138
21139
21140
21141
21142
21143
21144
21145
21146
21147
21148
21149
21150
21151
21152
21153
21154
21155
21156
21157
21158
21159
21160
21161
21162
21163
21164
21165
21166
21167
21168
21169
21170
21171
21172
21173
21174
21175
21176
21177
21178
21179
21180
21181
21182
21183
21184
21185
21186
21187
21188
21189
21190
21191
21192
21193
21194
21195
21196
21197
21198
21199
21200
21201
21202
21203
21204
21205
21206
21207
21208
21209
21210
21211
21212
21213
21214
21215
21216
21217
21218
21219
21220
21221
21222
21223
21224
21225
21226
21227
21228
21229
21230
21231
21232
21233
21234
21235
21236
21237
21238
21239
21240
21241
21242
21243
21244
21245
21246
21247
21248
21249
21250
21251
21252
21253
21254
21255
21256
21257
21258
21259
21260
21261
21262
21263
21264
21265
21266
21267
21268
21269
21270
21271
21272
21273
21274
21275
21276
21277
21278
21279
21280
21281
21282
21283
21284
21285
21286
21287
21288
21289
21290
21291
21292
21293
21294
21295
21296
21297
21298
21299
21300
21301
21302
21303
21304
21305
21306
21307
21308
21309
21310
21311
21312
21313
21314
21315
21316
21317
21318
21319
21320
21321
21322
21323
21324
21325
21326
21327
21328
21329
21330
21331
21332
21333
21334
21335
21336
21337
21338
21339
21340
21341
21342
21343
21344
21345
21346
21347
21348
21349
21350
21351
21352
21353
21354
21355
21356
21357
21358
21359
21360
21361
21362
21363
21364
21365
21366
21367
21368
21369
21370
21371
21372
21373
21374
21375
21376
21377
21378
21379
21380
21381
21382
21383
21384
21385
21386
21387
21388
21389
21390
21391
21392
21393
21394
21395
21396
21397
21398
21399
21400
21401
21402
21403
21404
21405
21406
21407
21408
21409
21410
21411
21412
21413
21414
21415
21416
21417
21418
21419
21420
21421
21422
21423
21424
21425
21426
21427
21428
21429
21430
21431
21432
21433
21434
21435
21436
21437
21438
21439
21440
21441
21442
21443
21444
21445
21446
21447
21448
21449
21450
21451
21452
21453
21454
21455
21456
21457
21458
21459
21460
21461
21462
21463
21464
21465
21466
21467
21468
21469
21470
21471
21472
21473
21474
21475
21476
21477
21478
21479
21480
21481
21482
21483
21484
21485
21486
21487
21488
21489
21490
21491
21492
21493
21494
21495
21496
21497
21498
21499
21500
21501
21502
21503
21504
21505
21506
21507
21508
21509
21510
21511
21512
21513
21514
21515
21516
21517
21518
21519
21520
21521
21522
21523
21524
21525
21526
21527
21528
21529
21530
21531
21532
21533
21534
21535
21536
21537
21538
21539
21540
21541
21542
21543
21544
21545
21546
21547
21548
21549
21550
21551
21552
21553
21554
21555
21556
21557
21558
21559
21560
21561
21562
21563
21564
21565
21566
21567
21568
21569
21570
21571
21572
21573
21574
21575
21576
21577
21578
21579
21580
21581
21582
21583
21584
21585
21586
21587
21588
21589
21590
21591
21592
21593
21594
21595
21596
21597
21598
21599
21600
21601
21602
21603
21604
21605
21606
21607
21608
21609
21610
21611
21612
21613
21614
21615
21616
21617
21618
21619
21620
21621
21622
21623
21624
21625
21626
21627
21628
21629
21630
21631
21632
21633
21634
21635
21636
21637
21638
21639
21640
21641
21642
21643
21644
21645
21646
21647
21648
21649
21650
21651
21652
21653
21654
21655
21656
21657
21658
21659
21660
21661
21662
21663
21664
21665
21666
21667
21668
21669
21670
21671
21672
21673
21674
21675
21676
21677
21678
21679
21680
21681
21682
21683
21684
21685
21686
21687
21688
21689
21690
21691
21692
21693
21694
21695
21696
21697
21698
21699
21700
21701
21702
21703
21704
21705
21706
21707
21708
21709
21710
21711
21712
21713
21714
21715
21716
21717
21718
21719
21720
21721
21722
21723
21724
21725
21726
21727
21728
21729
21730
21731
21732
21733
21734
21735
21736
21737
21738
21739
21740
21741
21742
21743
21744
21745
21746
21747
21748
21749
21750
21751
21752
21753
21754
21755
21756
21757
21758
21759
21760
21761
21762
21763
21764
21765
21766
21767
21768
21769
21770
21771
21772
21773
21774
21775
21776
21777
21778
21779
21780
21781
21782
21783
21784
21785
21786
21787
21788
21789
21790
21791
21792
21793
21794
21795
21796
21797
21798
21799
21800
21801
21802
21803
21804
21805
21806
21807
21808
21809
21810
21811
21812
21813
21814
21815
21816
21817
21818
21819
21820
21821
21822
21823
21824
21825
21826
21827
21828
21829
21830
21831
21832
21833
21834
21835
21836
21837
21838
21839
21840
21841
21842
21843
21844
21845
21846
21847
21848
21849
21850
21851
21852
21853
21854
21855
21856
21857
21858
21859
21860
21861
21862
21863
21864
21865
21866
21867
21868
21869
21870
21871
21872
21873
21874
21875
21876
21877
21878
21879
21880
21881
21882
21883
21884
21885
21886
21887
21888
21889
21890
21891
21892
21893
21894
21895
21896
21897
21898
21899
21900
21901
21902
21903
21904
21905
21906
21907
21908
21909
21910
21911
21912
21913
21914
21915
21916
21917
21918
21919
21920
21921
21922
21923
21924
21925
21926
21927
21928
21929
21930
21931
21932
21933
21934
21935
21936
21937
21938
21939
21940
21941
21942
21943
21944
21945
21946
21947
21948
21949
21950
21951
21952
21953
21954
21955
21956
21957
21958
21959
21960
21961
21962
21963
21964
21965
21966
21967
21968
21969
21970
21971
21972
21973
21974
21975
21976
21977
21978
21979
21980
21981
21982
21983
21984
21985
21986
21987
21988
21989
21990
21991
21992
21993
21994
21995
21996
21997
21998
21999
22000
22001
22002
22003
22004
22005
22006
22007
22008
22009
22010
22011
22012
22013
22014
22015
22016
22017
22018
22019
22020
22021
22022
22023
22024
22025
22026
22027
22028
22029
22030
22031
22032
22033
22034
22035
22036
22037
22038
22039
22040
22041
22042
22043
22044
22045
22046
22047
22048
22049
22050
22051
22052
22053
22054
22055
22056
22057
22058
22059
22060
22061
22062
22063
22064
22065
22066
22067
22068
22069
22070
22071
22072
22073
22074
22075
22076
22077
22078
22079
22080
22081
22082
22083
22084
22085
22086
22087
22088
22089
22090
22091
22092
22093
22094
22095
22096
22097
22098
22099
22100
22101
22102
22103
22104
22105
22106
22107
22108
22109
22110
22111
22112
22113
22114
22115
22116
22117
22118
22119
22120
22121
22122
22123
22124
22125
22126
22127
22128
22129
22130
22131
22132
22133
22134
22135
22136
22137
22138
22139
22140
22141
22142
22143
22144
22145
22146
22147
22148
22149
22150
22151
22152
22153
22154
22155
22156
22157
22158
22159
22160
22161
22162
22163
22164
22165
22166
22167
22168
22169
22170
22171
22172
22173
22174
22175
22176
22177
22178
22179
22180
22181
22182
22183
22184
22185
22186
22187
22188
22189
22190
22191
22192
22193
22194
22195
22196
22197
22198
22199
22200
22201
22202
22203
22204
22205
22206
22207
22208
22209
22210
22211
22212
22213
22214
22215
22216
22217
22218
22219
22220
22221
22222
22223
22224
22225
22226
22227
22228
22229
22230
22231
22232
22233
22234
22235
22236
22237
22238
22239
22240
22241
22242
22243
22244
22245
22246
22247
22248
22249
22250
22251
22252
22253
22254
22255
22256
22257
22258
22259
22260
22261
22262
22263
22264
22265
22266
22267
22268
22269
22270
22271
22272
22273
22274
22275
22276
22277
22278
22279
22280
22281
22282
22283
22284
22285
22286
22287
22288
22289
22290
22291
22292
22293
22294
22295
22296
22297
22298
22299
22300
22301
22302
22303
22304
22305
22306
22307
22308
22309
22310
22311
22312
22313
22314
22315
22316
22317
22318
22319
22320
22321
22322
22323
22324
22325
22326
22327
22328
22329
22330
22331
22332
22333
22334
22335
22336
22337
22338
22339
22340
22341
22342
22343
22344
22345
22346
22347
22348
22349
22350
22351
22352
22353
22354
22355
22356
22357
22358
22359
22360
22361
22362
22363
22364
22365
22366
22367
22368
22369
22370
22371
22372
22373
22374
22375
22376
22377
22378
22379
22380
22381
22382
22383
22384
22385
22386
22387
22388
22389
22390
22391
22392
22393
22394
22395
22396
22397
22398
22399
22400
22401
22402
22403
22404
22405
22406
22407
22408
22409
22410
22411
22412
22413
22414
22415
22416
22417
22418
22419
22420
22421
22422
22423
22424
22425
22426
22427
22428
22429
22430
22431
22432
22433
22434
22435
22436
22437
22438
22439
22440
22441
22442
22443
22444
22445
22446
22447
22448
22449
22450
22451
22452
22453
22454
22455
22456
22457
22458
22459
22460
22461
22462
22463
22464
22465
22466
22467
22468
22469
22470
22471
22472
22473
22474
22475
22476
22477
22478
22479
22480
22481
22482
22483
22484
22485
22486
22487
22488
22489
22490
22491
22492
22493
22494
22495
22496
22497
22498
22499
22500
22501
22502
22503
22504
22505
22506
22507
22508
22509
22510
22511
22512
22513
22514
22515
22516
22517
22518
22519
22520
22521
22522
22523
22524
22525
22526
22527
22528
22529
22530
22531
22532
22533
22534
22535
22536
22537
22538
22539
22540
22541
22542
22543
22544
22545
22546
22547
22548
22549
22550
22551
22552
22553
22554
22555
22556
22557
22558
22559
22560
22561
22562
22563
22564
22565
22566
22567
22568
22569
22570
22571
22572
22573
22574
22575
22576
22577
22578
22579
22580
22581
22582
22583
22584
22585
22586
22587
22588
22589
22590
22591
22592
22593
22594
22595
22596
22597
22598
22599
22600
22601
22602
22603
22604
22605
22606
22607
22608
22609
22610
22611
22612
22613
22614
22615
22616
22617
22618
22619
22620
22621
22622
22623
22624
22625
22626
22627
22628
22629
22630
22631
22632
22633
22634
22635
22636
22637
22638
22639
22640
22641
22642
22643
22644
22645
22646
22647
22648
22649
22650
22651
22652
22653
22654
22655
22656
22657
22658
22659
22660
22661
22662
22663
22664
22665
22666
22667
22668
22669
22670
22671
22672
22673
22674
22675
22676
22677
22678
22679
22680
22681
22682
22683
22684
22685
22686
22687
22688
22689
22690
22691
22692
22693
22694
22695
22696
22697
22698
22699
22700
22701
22702
22703
22704
22705
22706
22707
22708
22709
22710
22711
22712
22713
22714
22715
22716
22717
22718
22719
22720
22721
22722
22723
22724
22725
22726
22727
22728
22729
22730
22731
22732
22733
22734
22735
22736
22737
22738
22739
22740
22741
22742
22743
22744
22745
22746
22747
22748
22749
22750
22751
22752
22753
22754
22755
22756
22757
22758
22759
22760
22761
22762
22763
22764
22765
22766
22767
22768
22769
22770
22771
22772
22773
22774
22775
22776
22777
22778
22779
22780
22781
22782
22783
22784
22785
22786
22787
22788
22789
22790
22791
22792
22793
22794
22795
22796
22797
22798
22799
22800
22801
22802
22803
22804
22805
22806
22807
22808
22809
22810
22811
22812
22813
22814
22815
22816
22817
22818
22819
22820
22821
22822
22823
22824
22825
22826
22827
22828
22829
22830
22831
22832
22833
22834
22835
22836
22837
22838
22839
22840
22841
22842
22843
22844
22845
22846
22847
22848
22849
22850
22851
22852
22853
22854
22855
22856
22857
22858
22859
22860
22861
22862
22863
22864
22865
22866
22867
22868
22869
22870
22871
22872
22873
22874
22875
22876
22877
22878
22879
22880
22881
22882
22883
22884
22885
22886
22887
22888
22889
22890
22891
22892
22893
22894
22895
22896
22897
22898
22899
22900
22901
22902
22903
22904
22905
22906
22907
22908
22909
22910
22911
22912
22913
22914
22915
22916
22917
22918
22919
22920
22921
22922
22923
22924
22925
22926
22927
22928
22929
22930
22931
22932
22933
22934
22935
22936
22937
22938
22939
22940
22941
22942
22943
22944
22945
22946
22947
22948
22949
22950
22951
22952
22953
22954
22955
22956
22957
22958
22959
22960
22961
22962
22963
22964
22965
22966
22967
22968
22969
22970
22971
22972
22973
22974
22975
22976
22977
22978
22979
22980
22981
22982
22983
22984
22985
22986
22987
22988
22989
22990
22991
22992
22993
22994
22995
22996
22997
22998
22999
23000
23001
23002
23003
23004
23005
23006
23007
23008
23009
23010
23011
23012
23013
23014
23015
23016
23017
23018
23019
23020
23021
23022
23023
23024
23025
23026
23027
23028
23029
23030
23031
23032
23033
23034
23035
23036
23037
23038
23039
23040
23041
23042
23043
23044
23045
23046
23047
23048
23049
23050
23051
23052
23053
23054
23055
23056
23057
23058
23059
23060
23061
23062
23063
23064
23065
23066
23067
23068
23069
23070
23071
23072
23073
23074
23075
23076
23077
23078
23079
23080
23081
23082
23083
23084
23085
23086
23087
23088
23089
23090
23091
23092
23093
23094
23095
23096
23097
23098
23099
23100
23101
23102
23103
23104
23105
23106
23107
23108
23109
23110
23111
23112
23113
23114
23115
23116
23117
23118
23119
23120
23121
23122
23123
23124
23125
23126
23127
23128
23129
23130
23131
23132
23133
23134
23135
23136
23137
23138
23139
23140
23141
23142
23143
23144
23145
23146
23147
23148
23149
23150
% This program is copyright (C) 1984 by D. E. Knuth; all rights are reserved.
% Unlimited copying and redistribution of this file are permitted as long
% as this file is not modified. Modifications are permitted, but only if
% the resulting file is not named mf.web. (The WEB system provides
% for alterations via an auxiliary file; the master file should stay intact.)
% In other words, METAFONT is under essentially the same ground rules as TeX.

% TeX is a trademark of the American Mathematical Society.
% METAFONT is a trademark of Addison-Wesley Publishing Company.

% Version 0 was completed on July 28, 1984.
% Version 1 was completed on January 4, 1986; it corresponds to "Volume D".
% Version 1.1 trivially corrected the punctuation in one message (June 1986).
% Version 1.2 corrected an arithmetic overflow problem (July 1986).
% Version 1.3 improved rounding when elliptical pens are made (November 1986).
% Version 1.4 corrected scan_declared_variable timing (May 1988).
% Version 1.5 fixed negative halving in allocator when mem_min<0 (June 1988).
% Version 1.6 kept open_log_file from calling fatal_error (November 1988).
% Version 1.7 solved that problem a better way (December 1988).
% Version 1.8 introduced major changes for 8-bit extensions (September 1989).
% Version 1.9 improved skimping and was edited for style (December 1989).
% Version 2.0 fixed bug in addto; released with TeX version 3.0 (March 1990).
% Version 2.7 made consistent with TeX version 3.1 (September 1990).
% Version 2.71 fixed bug in draw, allowed unprintable filenames (March 1992).
% Version 2.718 fixed bug in <Choose a dependent...> (March 1995).
% Version 2.7182 fixed bugs related to "<unprintable char>" (August 1996).
% Version 2.71828 suppressed autorounding in dangerous cases (June 2003).
% Version 2.718281 was a general cleanup with minor fixes (February 2008).
% Version 2.7182818 was similar (January 2014).
% Version 2.71828182 was similar (January 2021).

% A reward of $327.68 will be paid to the first finder of any remaining bug.

% Although considerable effort has been expended to make the METAFONT program
% correct and reliable, no warranty is implied; the author disclaims any
% obligation or liability for damages, including but not limited to
% special, indirect, or consequential damages arising out of or in
% connection with the use or performance of this software. This work has
% been a ``labor of love'' and the author hopes that users enjoy it.

% Here is TeX material that gets inserted after \input webmac
\def\hang{\hangindent 3em\noindent\ignorespaces}
\def\textindent#1{\hangindent2.5em\noindent\hbox to2.5em{\hss#1 }\ignorespaces}
\font\ninerm=cmr9
\let\mc=\ninerm % medium caps for names like SAIL
\def\PASCAL{Pascal}
\def\ph{\hbox{Pascal-H}}
\def\psqrt#1{\sqrt{\mathstrut#1}}
\def\k{_{k+1}}
\def\pct!{{\char`\%}} % percent sign in ordinary text
\font\tenlogo=logo10 % font used for the METAFONT logo
\font\logos=logosl10
\font\eightlogo=logo8
\def\MF{{\tenlogo META}\-{\tenlogo FONT}}
\def\<#1>{$\langle#1\rangle$}
\def\section{\mathhexbox278}
\let\swap=\leftrightarrow
\def\round{\mathop{\rm round}\nolimits}

\def\(#1){} % this is used to make section names sort themselves better
\def\9#1{} % this is used for sort keys in the index via @@:sort key}{entry@@>

\outer\def\N#1. \[#2]#3.{\MN#1.\vfil\eject % begin starred section
  \def\rhead{PART #2:\uppercase{#3}} % define running headline
  \message{*\modno} % progress report
  \edef\next{\write\cont{\Z{\?#2]#3}{\modno}{\the\pageno}}}\next
  \ifon\startsection{\bf\ignorespaces#3.\quad}\ignorespaces}
\let\?=\relax % we want to be able to \write a \?

\def\title{{\eightlogo METAFONT}}
\def\topofcontents{\hsize 5.5in
  \vglue -30pt plus 1fil minus 1.5in
  \def\?##1]{\hbox to 1in{\hfil##1.\ }}
  }
\def\botofcontents{\vskip 0pt plus 1fil minus 1.5in}
\pageno=3
\def\glob{13} % this should be the section number of "<Global...>"
\def\gglob{20, 26} % this should be the next two sections of "<Global...>"

@* \[1] Introduction.
This is \MF, a font compiler intended to produce typefaces of high quality.
The \PASCAL\ program that follows is the definition of \MF84, a standard
@:PASCAL}{\PASCAL@>
@!@:METAFONT84}{\MF84@>
version of \MF\ that is designed to be highly portable so that identical output
will be obtainable on a great variety of computers. The conventions
of \MF84 are the same as those of \TeX82.

The main purpose of the following program is to explain the algorithms of \MF\
as clearly as possible. As a result, the program will not necessarily be very
efficient when a particular \PASCAL\ compiler has translated it into a
particular machine language. However, the program has been written so that it
can be tuned to run efficiently in a wide variety of operating environments
by making comparatively few changes. Such flexibility is possible because
the documentation that follows is written in the \.{WEB} language, which is
at a higher level than \PASCAL; the preprocessing step that converts \.{WEB}
to \PASCAL\ is able to introduce most of the necessary refinements.
Semi-automatic translation to other languages is also feasible, because the
program below does not make extensive use of features that are peculiar to
\PASCAL.

A large piece of software like \MF\ has inherent complexity that cannot
be reduced below a certain level of difficulty, although each individual
part is fairly simple by itself. The \.{WEB} language is intended to make
the algorithms as readable as possible, by reflecting the way the
individual program pieces fit together and by providing the
cross-references that connect different parts. Detailed comments about
what is going on, and about why things were done in certain ways, have
been liberally sprinkled throughout the program.  These comments explain
features of the implementation, but they rarely attempt to explain the
\MF\ language itself, since the reader is supposed to be familiar with
{\sl The {\logos METAFONT\/}book}.
@.WEB@>
@:METAFONTbook}{\sl The {\logos METAFONT\/}book@>

@ The present implementation has a long ancestry, beginning in the spring
of~1977, when its author wrote a prototype set of subroutines and macros
@^Knuth, Donald Ervin@>
that were used to develop the first Computer Modern fonts.
This original proto-\MF\ required the user to recompile a {\mc SAIL} program
whenever any character was changed, because it was not a ``language'' for
font design; the language was {\mc SAIL}. After several hundred characters
had been designed in that way, the author developed an interpretable language
called \MF, in which it was possible to express the Computer Modern programs
less cryptically. A complete \MF\ processor was designed and coded by the
author in 1979. This program, written in {\mc SAIL}, was adapted for use
with a variety of typesetting equipment and display terminals by Leo Guibas,
Lyle Ramshaw, and David Fuchs.
@^Guibas, Leonidas Ioannis@>
@^Ramshaw, Lyle Harold@>
@^Fuchs, David Raymond@>
Major improvements to the design of Computer Modern fonts were made in the
spring of 1982, after which it became clear that a new language would
better express the needs of letterform designers. Therefore an entirely
new \MF\ language and system were developed in 1984; the present system
retains the name and some of the spirit of \MF79, but all of the details
have changed.

No doubt there still is plenty of room for improvement, but the author
is firmly committed to keeping \MF84 ``frozen'' from now on; stability
and reliability are to be its main virtues.

On the other hand, the \.{WEB} description can be extended without changing
the core of \MF84 itself, and the program has been designed so that such
extensions are not extremely difficult to make.
The |banner| string defined here should be changed whenever \MF\
undergoes any modifications, so that it will be clear which version of
\MF\ might be the guilty party when a problem arises.
@^extensions to \MF@>
@^system dependencies@>

If this program is changed, the resulting system should not be called
`\MF\kern.5pt'; the official name `\MF\kern.5pt' by itself is reserved
for software systems that are fully compatible with each other.
A special test suite called the ``\.{TRAP} test'' is available for
helping to determine whether an implementation deserves to be
known as `\MF\kern.5pt' [cf.~Stanford Computer Science report CS1095,
January 1986].

@d banner=='This is METAFONT, Version 2.71828182' {printed when \MF\ starts}

@ Different \PASCAL s have slightly different conventions, and the present
@!@:PASCAL H}{\ph@>
program expresses \MF\ in terms of the \PASCAL\ that was
available to the author in 1984. Constructions that apply to
this particular compiler, which we shall call \ph, should help the
reader see how to make an appropriate interface for other systems
if necessary. (\ph\ is Charles Hedrick's modification of a compiler
@^Hedrick, Charles Locke@>
for the DECsystem-10 that was originally developed at the University of
Hamburg; cf.\ {\sl Software---Practice and Experience \bf6} (1976),
29--42. The \MF\ program below is intended to be adaptable, without
extensive changes, to most other versions of \PASCAL, so it does not fully
use the admirable features of \ph. Indeed, a conscious effort has been
made here to avoid using several idiosyncratic features of standard
\PASCAL\ itself, so that most of the code can be translated mechanically
into other high-level languages. For example, the `\&{with}' and `\\{new}'
features are not used, nor are pointer types, set types, or enumerated
scalar types; there are no `\&{var}' parameters, except in the case of files
or in the system-dependent |paint_row| procedure;
there are no tag fields on variant records; there are no |real| variables;
no procedures are declared local to other procedures.)

The portions of this program that involve system-dependent code, where
changes might be necessary because of differences between \PASCAL\ compilers
and/or differences between
operating systems, can be identified by looking at the sections whose
numbers are listed under `system dependencies' in the index. Furthermore,
the index entries for `dirty \PASCAL' list all places where the restrictions
of \PASCAL\ have not been followed perfectly, for one reason or another.
@!@^system dependencies@>
@!@^dirty \PASCAL@>

@ The program begins with a normal \PASCAL\ program heading, whose
components will be filled in later, using the conventions of \.{WEB}.
@.WEB@>
For example, the portion of the program called `\X\glob:Global
variables\X' below will be replaced by a sequence of variable declarations
that starts in $\section\glob$ of this documentation. In this way, we are able
to define each individual global variable when we are prepared to
understand what it means; we do not have to define all of the globals at
once.  Cross references in $\section\glob$, where it says ``See also
sections \gglob, \dots,'' also make it possible to look at the set of
all global variables, if desired.  Similar remarks apply to the other
portions of the program heading.

Actually the heading shown here is not quite normal: The |program| line
does not mention any |output| file, because \ph\ would ask the \MF\ user
to specify a file name if |output| were specified here.
@:PASCAL H}{\ph@>
@^system dependencies@>

@d mtype==t@&y@&p@&e {this is a \.{WEB} coding trick:}
@f mtype==type {`\&{mtype}' will be equivalent to `\&{type}'}
@f type==true {but `|type|' will not be treated as a reserved word}

@p @t\4@>@<Compiler directives@>@/
program MF; {all file names are defined dynamically}
label @<Labels in the outer block@>@/
const @<Constants in the outer block@>@/
mtype @<Types in the outer block@>@/
var @<Global variables@>@/
@#
procedure initialize; {this procedure gets things started properly}
  var @<Local variables for initialization@>@/
  begin @<Set initial values of key variables@>@/
  end;@#
@t\4@>@<Basic printing procedures@>@/
@t\4@>@<Error handling procedures@>@/

@ The overall \MF\ program begins with the heading just shown, after which
comes a bunch of procedure declarations and function declarations.
Finally we will get to the main program, which begins with the
comment `|start_here|'. If you want to skip down to the
main program now, you can look up `|start_here|' in the index.
But the author suggests that the best way to understand this program
is to follow pretty much the order of \MF's components as they appear in the
\.{WEB} description you are now reading, since the present ordering is
intended to combine the advantages of the ``bottom up'' and ``top down''
approaches to the problem of understanding a somewhat complicated system.

@ Three labels must be declared in the main program, so we give them
symbolic names.

@d start_of_MF=1 {go here when \MF's variables are initialized}
@d end_of_MF=9998 {go here to close files and terminate gracefully}
@d final_end=9999 {this label marks the ending of the program}

@<Labels in the out...@>=
start_of_MF@t\hskip-2pt@>, end_of_MF@t\hskip-2pt@>,@,final_end;
  {key control points}

@ Some of the code below is intended to be used only when diagnosing the
strange behavior that sometimes occurs when \MF\ is being installed or
when system wizards are fooling around with \MF\ without quite knowing
what they are doing. Such code will not normally be compiled; it is
delimited by the codewords `$|debug|\ldots|gubed|$', with apologies
to people who wish to preserve the purity of English.

Similarly, there is some conditional code delimited by
`$|stat|\ldots|tats|$' that is intended for use when statistics are to be
kept about \MF's memory usage.  The |stat| $\ldots$ |tats| code also
implements special diagnostic information that is printed when
$\\{tracingedges}>1$.
@^debugging@>

@d debug==@{ {change this to `$\\{debug}\equiv\null$' when debugging}
@d gubed==@t@>@} {change this to `$\\{gubed}\equiv\null$' when debugging}
@f debug==begin
@f gubed==end
@#
@d stat==@{ {change this to `$\\{stat}\equiv\null$' when gathering
  usage statistics}
@d tats==@t@>@} {change this to `$\\{tats}\equiv\null$' when gathering
  usage statistics}
@f stat==begin
@f tats==end

@ This program has two important variations: (1) There is a long and slow
version called \.{INIMF}, which does the extra calculations needed to
@.INIMF@>
initialize \MF's internal tables; and (2)~there is a shorter and faster
production version, which cuts the initialization to a bare minimum.
Parts of the program that are needed in (1) but not in (2) are delimited by
the codewords `$|init|\ldots|tini|$'.

@d init== {change this to `$\\{init}\equiv\.{@@\{}$' in the production version}
@d tini== {change this to `$\\{tini}\equiv\.{@@\}}$' in the production version}
@f init==begin
@f tini==end

@ If the first character of a \PASCAL\ comment is a dollar sign,
\ph\ treats the comment as a list of ``compiler directives'' that will
affect the translation of this program into machine language.  The
directives shown below specify full checking and inclusion of the \PASCAL\
debugger when \MF\ is being debugged, but they cause range checking and other
redundant code to be eliminated when the production system is being generated.
Arithmetic overflow will be detected in all cases.
@:PASCAL H}{\ph@>
@^system dependencies@>
@^overflow in arithmetic@>

@<Compiler directives@>=
@{@&$C-,A+,D-@} {no range check, catch arithmetic overflow, no debug overhead}
@!debug @{@&$C+,D+@}@+ gubed {but turn everything on when debugging}

@ This \MF\ implementation conforms to the rules of the {\sl Pascal User
@:PASCAL}{\PASCAL@>
@^system dependencies@>
Manual} published by Jensen and Wirth in 1975, except where system-dependent
@^Wirth, Niklaus@>
@^Jensen, Kathleen@>
code is necessary to make a useful system program, and except in another
respect where such conformity would unnecessarily obscure the meaning
and clutter up the code: We assume that |case| statements may include a
default case that applies if no matching label is found. Thus, we shall use
constructions like
$$\vbox{\halign{\ignorespaces#\hfil\cr
|case x of|\cr
1: $\langle\,$code for $x=1\,\rangle$;\cr
3: $\langle\,$code for $x=3\,\rangle$;\cr
|othercases| $\langle\,$code for |x<>1| and |x<>3|$\,\rangle$\cr
|endcases|\cr}}$$
since most \PASCAL\ compilers have plugged this hole in the language by
incorporating some sort of default mechanism. For example, the \ph\
compiler allows `|others|:' as a default label, and other \PASCAL s allow
syntaxes like `\&{else}' or `\&{otherwise}' or `\\{otherwise}:', etc. The
definitions of |othercases| and |endcases| should be changed to agree with
local conventions.  Note that no semicolon appears before |endcases| in
this program, so the definition of |endcases| should include a semicolon
if the compiler wants one. (Of course, if no default mechanism is
available, the |case| statements of \MF\ will have to be laboriously
extended by listing all remaining cases. People who are stuck with such
\PASCAL s have, in fact, done this, successfully but not happily!)
@:PASCAL H}{\ph@>

@d othercases == others: {default for cases not listed explicitly}
@d endcases == @+end {follows the default case in an extended |case| statement}
@f othercases == else
@f endcases == end

@ The following parameters can be changed at compile time to extend or
reduce \MF's capacity. They may have different values in \.{INIMF} and
in production versions of \MF.
@.INIMF@>
@^system dependencies@>

@<Constants...@>=
@!mem_max=30000; {greatest index in \MF's internal |mem| array;
  must be strictly less than |max_halfword|;
  must be equal to |mem_top| in \.{INIMF}, otherwise |>=mem_top|}
@!max_internal=100; {maximum number of internal quantities}
@!buf_size=500; {maximum number of characters simultaneously present in
  current lines of open files; must not exceed |max_halfword|}
@!error_line=72; {width of context lines on terminal error messages}
@!half_error_line=42; {width of first lines of contexts in terminal
  error messages; should be between 30 and |error_line-15|}
@!max_print_line=79; {width of longest text lines output; should be at least 60}
@!screen_width=768; {number of pixels in each row of screen display}
@!screen_depth=1024; {number of pixels in each column of screen display}
@!stack_size=30; {maximum number of simultaneous input sources}
@!max_strings=2000; {maximum number of strings; must not exceed |max_halfword|}
@!string_vacancies=8000; {the minimum number of characters that should be
  available for the user's identifier names and strings,
  after \MF's own error messages are stored}
@!pool_size=32000; {maximum number of characters in strings, including all
  error messages and help texts, and the names of all identifiers;
  must exceed |string_vacancies| by the total
  length of \MF's own strings, which is currently about 22000}
@!move_size=5000; {space for storing moves in a single octant}
@!max_wiggle=300; {number of autorounded points per cycle}
@!gf_buf_size=800; {size of the output buffer, must be a multiple of 8}
@!file_name_size=40; {file names shouldn't be longer than this}
@!pool_name='MFbases:MF.POOL                         ';
  {string of length |file_name_size|; tells where the string pool appears}
@.MFbases@>
@!path_size=300; {maximum number of knots between breakpoints of a path}
@!bistack_size=785; {size of stack for bisection algorithms;
  should probably be left at this value}
@!header_size=100; {maximum number of \.{TFM} header words, times~4}
@!lig_table_size=5000; {maximum number of ligature/kern steps, must be
  at least 255 and at most 32510}
@!max_kerns=500; {maximum number of distinct kern amounts}
@!max_font_dimen=50; {maximum number of \&{fontdimen} parameters}

@ Like the preceding parameters, the following quantities can be changed
at compile time to extend or reduce \MF's capacity. But if they are changed,
it is necessary to rerun the initialization program \.{INIMF}
@.INIMF@>
to generate new tables for the production \MF\ program.
One can't simply make helter-skelter changes to the following constants,
since certain rather complex initialization
numbers are computed from them. They are defined here using
\.{WEB} macros, instead of being put into \PASCAL's |const| list, in order to
emphasize this distinction.

@d mem_min=0 {smallest index in the |mem| array, must not be less
  than |min_halfword|}
@d mem_top==30000 {largest index in the |mem| array dumped by \.{INIMF};
  must be substantially larger than |mem_min|
  and not greater than |mem_max|}
@d hash_size=2100 {maximum number of symbolic tokens,
  must be less than |max_halfword-3*param_size|}
@d hash_prime=1777 {a prime number equal to about 85\pct! of |hash_size|}
@d max_in_open=6 {maximum number of input files and error insertions that
  can be going on simultaneously}
@d param_size=150 {maximum number of simultaneous macro parameters}
@^system dependencies@>

@ In case somebody has inadvertently made bad settings of the ``constants,''
\MF\ checks them using a global variable called |bad|.

This is the first of many sections of \MF\ where global variables are
defined.

@<Glob...@>=
@!bad:integer; {is some ``constant'' wrong?}

@ Later on we will say `\ignorespaces|if mem_max>=max_halfword then bad:=10|',
or something similar. (We can't do that until |max_halfword| has been defined.)

@<Check the ``constant'' values for consistency@>=
bad:=0;
if (half_error_line<30)or(half_error_line>error_line-15) then bad:=1;
if max_print_line<60 then bad:=2;
if gf_buf_size mod 8<>0 then bad:=3;
if mem_min+1100>mem_top then bad:=4;
if hash_prime>hash_size then bad:=5;
if header_size mod 4 <> 0 then bad:=6;
if(lig_table_size<255)or(lig_table_size>32510)then bad:=7;

@ Labels are given symbolic names by the following definitions, so that
occasional |goto| statements will be meaningful. We insert the label
`|exit|' just before the `\ignorespaces|end|\unskip' of a procedure in
which we have used the `|return|' statement defined below; the label
`|restart|' is occasionally used at the very beginning of a procedure; and
the label `|reswitch|' is occasionally used just prior to a |case|
statement in which some cases change the conditions and we wish to branch
to the newly applicable case.  Loops that are set up with the |loop|
construction defined below are commonly exited by going to `|done|' or to
`|found|' or to `|not_found|', and they are sometimes repeated by going to
`|continue|'.  If two or more parts of a subroutine start differently but
end up the same, the shared code may be gathered together at
`|common_ending|'.

Incidentally, this program never declares a label that isn't actually used,
because some fussy \PASCAL\ compilers will complain about redundant labels.

@d exit=10 {go here to leave a procedure}
@d restart=20 {go here to start a procedure again}
@d reswitch=21 {go here to start a case statement again}
@d continue=22 {go here to resume a loop}
@d done=30 {go here to exit a loop}
@d done1=31 {like |done|, when there is more than one loop}
@d done2=32 {for exiting the second loop in a long block}
@d done3=33 {for exiting the third loop in a very long block}
@d done4=34 {for exiting the fourth loop in an extremely long block}
@d done5=35 {for exiting the fifth loop in an immense block}
@d done6=36 {for exiting the sixth loop in a block}
@d found=40 {go here when you've found it}
@d found1=41 {like |found|, when there's more than one per routine}
@d found2=42 {like |found|, when there's more than two per routine}
@d not_found=45 {go here when you've found nothing}
@d common_ending=50 {go here when you want to merge with another branch}

@ Here are some macros for common programming idioms.

@d incr(#) == #:=#+1 {increase a variable by unity}
@d decr(#) == #:=#-1 {decrease a variable by unity}
@d negate(#) == #:=-# {change the sign of a variable}
@d double(#) == #:=#+# {multiply a variable by two}
@d loop == @+ while true do@+ {repeat over and over until a |goto| happens}
@f loop == xclause
  {\.{WEB}'s |xclause| acts like `\ignorespaces|while true do|\unskip'}
@d do_nothing == {empty statement}
@d return == goto exit {terminate a procedure call}
@f return == nil {\.{WEB} will henceforth say |return| instead of \\{return}}

@* \[2] The character set.
In order to make \MF\ readily portable to a wide variety of
computers, all of its input text is converted to an internal eight-bit
code that includes standard ASCII, the ``American Standard Code for
Information Interchange.''  This conversion is done immediately when each
character is read in. Conversely, characters are converted from ASCII to
the user's external representation just before they are output to a
text file.
@^ASCII code@>

Such an internal code is relevant to users of \MF\ only with respect to
the \&{char} and \&{ASCII} operations, and the comparison of strings.

@ Characters of text that have been converted to \MF's internal form
are said to be of type |ASCII_code|, which is a subrange of the integers.

@<Types...@>=
@!ASCII_code=0..255; {eight-bit numbers}

@ The original \PASCAL\ compiler was designed in the late 60s, when six-bit
character sets were common, so it did not make provision for lowercase
letters. Nowadays, of course, we need to deal with both capital and small
letters in a convenient way, especially in a program for font design;
so the present specification of \MF\ has been written under the assumption
that the \PASCAL\ compiler and run-time system permit the use of text files
with more than 64 distinguishable characters. More precisely, we assume that
the character set contains at least the letters and symbols associated
with ASCII codes @'40 through @'176; all of these characters are now
available on most computer terminals.

Since we are dealing with more characters than were present in the first
\PASCAL\ compilers, we have to decide what to call the associated data
type. Some \PASCAL s use the original name |char| for the
characters in text files, even though there now are more than 64 such
characters, while other \PASCAL s consider |char| to be a 64-element
subrange of a larger data type that has some other name.

In order to accommodate this difference, we shall use the name |text_char|
to stand for the data type of the characters that are converted to and
from |ASCII_code| when they are input and output. We shall also assume
that |text_char| consists of the elements |chr(first_text_char)| through
|chr(last_text_char)|, inclusive. The following definitions should be
adjusted if necessary.
@^system dependencies@>

@d text_char == char {the data type of characters in text files}
@d first_text_char=0 {ordinal number of the smallest element of |text_char|}
@d last_text_char=255 {ordinal number of the largest element of |text_char|}

@<Local variables for init...@>=
@!i:integer;

@ The \MF\ processor converts between ASCII code and
the user's external character set by means of arrays |xord| and |xchr|
that are analogous to \PASCAL's |ord| and |chr| functions.

@<Glob...@>=
@!xord: array [text_char] of ASCII_code;
  {specifies conversion of input characters}
@!xchr: array [ASCII_code] of text_char;
  {specifies conversion of output characters}

@ Since we are assuming that our \PASCAL\ system is able to read and
write the visible characters of standard ASCII (although not
necessarily using the ASCII codes to represent them), the following
assignment statements initialize the standard part of the |xchr| array
properly, without needing any system-dependent changes. On the other
hand, it is possible to implement \MF\ with less complete character
sets, and in such cases it will be necessary to change something here.
@^system dependencies@>

@<Set init...@>=
xchr[@'40]:=' ';
xchr[@'41]:='!';
xchr[@'42]:='"';
xchr[@'43]:='#';
xchr[@'44]:='$';
xchr[@'45]:='%';
xchr[@'46]:='&';
xchr[@'47]:='''';@/
xchr[@'50]:='(';
xchr[@'51]:=')';
xchr[@'52]:='*';
xchr[@'53]:='+';
xchr[@'54]:=',';
xchr[@'55]:='-';
xchr[@'56]:='.';
xchr[@'57]:='/';@/
xchr[@'60]:='0';
xchr[@'61]:='1';
xchr[@'62]:='2';
xchr[@'63]:='3';
xchr[@'64]:='4';
xchr[@'65]:='5';
xchr[@'66]:='6';
xchr[@'67]:='7';@/
xchr[@'70]:='8';
xchr[@'71]:='9';
xchr[@'72]:=':';
xchr[@'73]:=';';
xchr[@'74]:='<';
xchr[@'75]:='=';
xchr[@'76]:='>';
xchr[@'77]:='?';@/
xchr[@'100]:='@@';
xchr[@'101]:='A';
xchr[@'102]:='B';
xchr[@'103]:='C';
xchr[@'104]:='D';
xchr[@'105]:='E';
xchr[@'106]:='F';
xchr[@'107]:='G';@/
xchr[@'110]:='H';
xchr[@'111]:='I';
xchr[@'112]:='J';
xchr[@'113]:='K';
xchr[@'114]:='L';
xchr[@'115]:='M';
xchr[@'116]:='N';
xchr[@'117]:='O';@/
xchr[@'120]:='P';
xchr[@'121]:='Q';
xchr[@'122]:='R';
xchr[@'123]:='S';
xchr[@'124]:='T';
xchr[@'125]:='U';
xchr[@'126]:='V';
xchr[@'127]:='W';@/
xchr[@'130]:='X';
xchr[@'131]:='Y';
xchr[@'132]:='Z';
xchr[@'133]:='[';
xchr[@'134]:='\';
xchr[@'135]:=']';
xchr[@'136]:='^';
xchr[@'137]:='_';@/
xchr[@'140]:='`';
xchr[@'141]:='a';
xchr[@'142]:='b';
xchr[@'143]:='c';
xchr[@'144]:='d';
xchr[@'145]:='e';
xchr[@'146]:='f';
xchr[@'147]:='g';@/
xchr[@'150]:='h';
xchr[@'151]:='i';
xchr[@'152]:='j';
xchr[@'153]:='k';
xchr[@'154]:='l';
xchr[@'155]:='m';
xchr[@'156]:='n';
xchr[@'157]:='o';@/
xchr[@'160]:='p';
xchr[@'161]:='q';
xchr[@'162]:='r';
xchr[@'163]:='s';
xchr[@'164]:='t';
xchr[@'165]:='u';
xchr[@'166]:='v';
xchr[@'167]:='w';@/
xchr[@'170]:='x';
xchr[@'171]:='y';
xchr[@'172]:='z';
xchr[@'173]:='{';
xchr[@'174]:='|';
xchr[@'175]:='}';
xchr[@'176]:='~';@/

@ The ASCII code is ``standard'' only to a certain extent, since many
computer installations have found it advantageous to have ready access
to more than 94 printing characters.  If \MF\ is being used
on a garden-variety \PASCAL\ for which only standard ASCII
codes will appear in the input and output files, it doesn't really matter
what codes are specified in |xchr[0..@'37]|, but the safest policy is to
blank everything out by using the code shown below.

However, other settings of |xchr| will make \MF\ more friendly on
computers that have an extended character set, so that users can type things
like `\.^^Z' instead of `\.{<>}'.
People with extended character sets can
assign codes arbitrarily, giving an |xchr| equivalent to whatever
characters the users of \MF\ are allowed to have in their input files.
Appropriate changes to \MF's |char_class| table should then be made.
(Unlike \TeX, each installation of \MF\ has a fixed assignment of category
codes, called the |char_class|.) Such changes make portability of programs
more difficult, so they should be introduced cautiously if at all.
@^character set dependencies@>
@^system dependencies@>

@<Set init...@>=
for i:=0 to @'37 do xchr[i]:=' ';
for i:=@'177 to @'377 do xchr[i]:=' ';

@ The following system-independent code makes the |xord| array contain a
suitable inverse to the information in |xchr|. Note that if |xchr[i]=xchr[j]|
where |i<j<@'177|, the value of |xord[xchr[i]]| will turn out to be
|j| or more; hence, standard ASCII code numbers will be used instead of
codes below @'40 in case there is a coincidence.

@<Set init...@>=
for i:=first_text_char to last_text_char do xord[chr(i)]:=@'177;
for i:=@'200 to @'377 do xord[xchr[i]]:=i;
for i:=0 to @'176 do xord[xchr[i]]:=i;

@* \[3] Input and output.
The bane of portability is the fact that different operating systems treat
input and output quite differently, perhaps because computer scientists
have not given sufficient attention to this problem. People have felt somehow
that input and output are not part of ``real'' programming. Well, it is true
that some kinds of programming are more fun than others. With existing
input/output conventions being so diverse and so messy, the only sources of
joy in such parts of the code are the rare occasions when one can find a
way to make the program a little less bad than it might have been. We have
two choices, either to attack I/O now and get it over with, or to postpone
I/O until near the end. Neither prospect is very attractive, so let's
get it over with.

The basic operations we need to do are (1)~inputting and outputting of
text, to or from a file or the user's terminal; (2)~inputting and
outputting of eight-bit bytes, to or from a file; (3)~instructing the
operating system to initiate (``open'') or to terminate (``close'') input or
output from a specified file; (4)~testing whether the end of an input
file has been reached; (5)~display of bits on the user's screen.
The bit-display operation will be discussed in a later section; we shall
deal here only with more traditional kinds of I/O.

\MF\ needs to deal with two kinds of files.
We shall use the term |alpha_file| for a file that contains textual data,
and the term |byte_file| for a file that contains eight-bit binary information.
These two types turn out to be the same on many computers, but
sometimes there is a significant distinction, so we shall be careful to
distinguish between them. Standard protocols for transferring
such files from computer to computer, via high-speed networks, are
now becoming available to more and more communities of users.

The program actually makes use also of a third kind of file, called a
|word_file|, when dumping and reloading base information for its own
initialization.  We shall define a word file later; but it will be possible
for us to specify simple operations on word files before they are defined.

@<Types...@>=
@!eight_bits=0..255; {unsigned one-byte quantity}
@!alpha_file=packed file of text_char; {files that contain textual data}
@!byte_file=packed file of eight_bits; {files that contain binary data}

@ Most of what we need to do with respect to input and output can be handled
by the I/O facilities that are standard in \PASCAL, i.e., the routines
called |get|, |put|, |eof|, and so on. But
standard \PASCAL\ does not allow file variables to be associated with file
names that are determined at run time, so it cannot be used to implement
\MF; some sort of extension to \PASCAL's ordinary |reset| and |rewrite|
is crucial for our purposes. We shall assume that |name_of_file| is a variable
of an appropriate type such that the \PASCAL\ run-time system being used to
implement \MF\ can open a file whose external name is specified by
|name_of_file|.
@^system dependencies@>

@<Glob...@>=
@!name_of_file:packed array[1..file_name_size] of char;@;@/
  {on some systems this may be a \&{record} variable}
@!name_length:0..file_name_size;@/{this many characters are actually
  relevant in |name_of_file| (the rest are blank)}

@ The \ph\ compiler with which the present version of \MF\ was prepared has
extended the rules of \PASCAL\ in a very convenient way. To open file~|f|,
we can write
$$\vbox{\halign{#\hfil\qquad&#\hfil\cr
|reset(f,@t\\{name}@>,'/O')|&for input;\cr
|rewrite(f,@t\\{name}@>,'/O')|&for output.\cr}}$$
The `\\{name}' parameter, which is of type `\ignorespaces|packed
array[@t\<\\{any}>@>] of text_char|', stands for the name of
the external file that is being opened for input or output.
Blank spaces that might appear in \\{name} are ignored.

The `\.{/O}' parameter tells the operating system not to issue its own
error messages if something goes wrong. If a file of the specified name
cannot be found, or if such a file cannot be opened for some other reason
(e.g., someone may already be trying to write the same file), we will have
|@!erstat(f)<>0| after an unsuccessful |reset| or |rewrite|.  This allows
\MF\ to undertake appropriate corrective action.
@:PASCAL H}{\ph@>
@^system dependencies@>

\MF's file-opening procedures return |false| if no file identified by
|name_of_file| could be opened.

@d reset_OK(#)==erstat(#)=0
@d rewrite_OK(#)==erstat(#)=0

@p function a_open_in(var @!f:alpha_file):boolean;
  {open a text file for input}
begin reset(f,name_of_file,'/O'); a_open_in:=reset_OK(f);
end;
@#
function a_open_out(var @!f:alpha_file):boolean;
  {open a text file for output}
begin rewrite(f,name_of_file,'/O'); a_open_out:=rewrite_OK(f);
end;
@#
function b_open_out(var @!f:byte_file):boolean;
  {open a binary file for output}
begin rewrite(f,name_of_file,'/O'); b_open_out:=rewrite_OK(f);
end;
@#
function w_open_in(var @!f:word_file):boolean;
  {open a word file for input}
begin reset(f,name_of_file,'/O'); w_open_in:=reset_OK(f);
end;
@#
function w_open_out(var @!f:word_file):boolean;
  {open a word file for output}
begin rewrite(f,name_of_file,'/O'); w_open_out:=rewrite_OK(f);
end;

@ Files can be closed with the \ph\ routine `|close(f)|', which
@:PASCAL H}{\ph@>
@^system dependencies@>
should be used when all input or output with respect to |f| has been completed.
This makes |f| available to be opened again, if desired; and if |f| was used for
output, the |close| operation makes the corresponding external file appear
on the user's area, ready to be read.

@p procedure a_close(var @!f:alpha_file); {close a text file}
begin close(f);
end;
@#
procedure b_close(var @!f:byte_file); {close a binary file}
begin close(f);
end;
@#
procedure w_close(var @!f:word_file); {close a word file}
begin close(f);
end;

@ Binary input and output are done with \PASCAL's ordinary |get| and |put|
procedures, so we don't have to make any other special arrangements for
binary~I/O. Text output is also easy to do with standard \PASCAL\ routines.
The treatment of text input is more difficult, however, because
of the necessary translation to |ASCII_code| values.
\MF's conventions should be efficient, and they should
blend nicely with the user's operating environment.

@ Input from text files is read one line at a time, using a routine called
|input_ln|. This function is defined in terms of global variables called
|buffer|, |first|, and |last| that will be described in detail later; for
now, it suffices for us to know that |buffer| is an array of |ASCII_code|
values, and that |first| and |last| are indices into this array
representing the beginning and ending of a line of text.

@<Glob...@>=
@!buffer:array[0..buf_size] of ASCII_code; {lines of characters being read}
@!first:0..buf_size; {the first unused position in |buffer|}
@!last:0..buf_size; {end of the line just input to |buffer|}
@!max_buf_stack:0..buf_size; {largest index used in |buffer|}

@ The |input_ln| function brings the next line of input from the specified
file into available positions of the buffer array and returns the value
|true|, unless the file has already been entirely read, in which case it
returns |false| and sets |last:=first|.  In general, the |ASCII_code|
numbers that represent the next line of the file are input into
|buffer[first]|, |buffer[first+1]|, \dots, |buffer[last-1]|; and the
global variable |last| is set equal to |first| plus the length of the
line. Trailing blanks are removed from the line; thus, either |last=first|
(in which case the line was entirely blank) or |buffer[last-1]<>" "|.
@^inner loop@>

An overflow error is given, however, if the normal actions of |input_ln|
would make |last>=buf_size|; this is done so that other parts of \MF\
can safely look at the contents of |buffer[last+1]| without overstepping
the bounds of the |buffer| array. Upon entry to |input_ln|, the condition
|first<buf_size| will always hold, so that there is always room for an
``empty'' line.

The variable |max_buf_stack|, which is used to keep track of how large
the |buf_size| parameter must be to accommodate the present job, is
also kept up to date by |input_ln|.

If the |bypass_eoln| parameter is |true|, |input_ln| will do a |get|
before looking at the first character of the line; this skips over
an |eoln| that was in |f^|. The procedure does not do a |get| when it
reaches the end of the line; therefore it can be used to acquire input
from the user's terminal as well as from ordinary text files.

Standard \PASCAL\ says that a file should have |eoln| immediately
before |eof|, but \MF\ needs only a weaker restriction: If |eof|
occurs in the middle of a line, the system function |eoln| should return
a |true| result (even though |f^| will be undefined).

@p function input_ln(var @!f:alpha_file;@!bypass_eoln:boolean):boolean;
  {inputs the next line or returns |false|}
var @!last_nonblank:0..buf_size; {|last| with trailing blanks removed}
begin if bypass_eoln then if not eof(f) then get(f);
  {input the first character of the line into |f^|}
last:=first; {cf.\ Matthew 19\thinspace:\thinspace30}
if eof(f) then input_ln:=false
else  begin last_nonblank:=first;
  while not eoln(f) do
    begin if last>=max_buf_stack then
      begin max_buf_stack:=last+1;
      if max_buf_stack=buf_size then
        @<Report overflow of the input buffer, and abort@>;
      end;
    buffer[last]:=xord[f^]; get(f); incr(last);
    if buffer[last-1]<>" " then last_nonblank:=last;
    end;
  last:=last_nonblank; input_ln:=true;
  end;
end;

@ The user's terminal acts essentially like other files of text, except
that it is used both for input and for output. When the terminal is
considered an input file, the file variable is called |term_in|, and when it
is considered an output file the file variable is |term_out|.
@^system dependencies@>

@<Glob...@>=
@!term_in:alpha_file; {the terminal as an input file}
@!term_out:alpha_file; {the terminal as an output file}

@ Here is how to open the terminal files
in \ph. The `\.{/I}' switch suppresses the first |get|.
@:PASCAL H}{\ph@>
@^system dependencies@>

@d t_open_in==reset(term_in,'TTY:','/O/I') {open the terminal for text input}
@d t_open_out==rewrite(term_out,'TTY:','/O')
 {open the terminal for text output}

@ Sometimes it is necessary to synchronize the input/output mixture that
happens on the user's terminal, and three system-dependent
procedures are used for this
purpose. The first of these, |update_terminal|, is called when we want
to make sure that everything we have output to the terminal so far has
actually left the computer's internal buffers and been sent.
The second, |clear_terminal|, is called when we wish to cancel any
input that the user may have typed ahead (since we are about to
issue an unexpected error message). The third, |wake_up_terminal|,
is supposed to revive the terminal if the user has disabled it by
some instruction to the operating system.  The following macros show how
these operations can be specified in \ph:
@:PASCAL H}{\ph@>
@^system dependencies@>

@d update_terminal == break(term_out) {empty the terminal output buffer}
@d clear_terminal == break_in(term_in,true) {clear the terminal input buffer}
@d wake_up_terminal == do_nothing {cancel the user's cancellation of output}

@ We need a special routine to read the first line of \MF\ input from
the user's terminal. This line is different because it is read before we
have opened the transcript file; there is sort of a ``chicken and
egg'' problem here. If the user types `\.{input cmr10}' on the first
line, or if some macro invoked by that line does such an \.{input},
the transcript file will be named `\.{cmr10.log}'; but if no \.{input}
commands are performed during the first line of terminal input, the transcript
file will acquire its default name `\.{mfput.log}'. (The transcript file
will not contain error messages generated by the first line before the
first \.{input} command.)
@.mfput@>

The first line is even more special if we are lucky enough to have an operating
system that treats \MF\ differently from a run-of-the-mill \PASCAL\ object
program. It's nice to let the user start running a \MF\ job by typing
a command line like `\.{MF cmr10}'; in such a case, \MF\ will operate
as if the first line of input were `\.{cmr10}', i.e., the first line will
consist of the remainder of the command line, after the part that invoked \MF.

The first line is special also because it may be read before \MF\ has
input a base file. In such cases, normal error messages cannot yet
be given. The following code uses concepts that will be explained later.
(If the \PASCAL\ compiler does not support non-local |@!goto|\unskip, the
@^system dependencies@>
statement `|goto final_end|' should be replaced by something that
quietly terminates the program.)

@<Report overflow of the input buffer, and abort@>=
if base_ident=0 then
  begin write_ln(term_out,'Buffer size exceeded!'); goto final_end;
@.Buffer size exceeded@>
  end
else begin cur_input.loc_field:=first; cur_input.limit_field:=last-1;
  overflow("buffer size",buf_size);
@:METAFONT capacity exceeded buffer size}{\quad buffer size@>
  end

@ Different systems have different ways to get started. But regardless of
what conventions are adopted, the routine that initializes the terminal
should satisfy the following specifications:

\yskip\textindent{1)}It should open file |term_in| for input from the
  terminal. (The file |term_out| will already be open for output to the
  terminal.)

\textindent{2)}If the user has given a command line, this line should be
  considered the first line of terminal input. Otherwise the
  user should be prompted with `\.{**}', and the first line of input
  should be whatever is typed in response.

\textindent{3)}The first line of input, which might or might not be a
  command line, should appear in locations |first| to |last-1| of the
  |buffer| array.

\textindent{4)}The global variable |loc| should be set so that the
  character to be read next by \MF\ is in |buffer[loc]|. This
  character should not be blank, and we should have |loc<last|.

\yskip\noindent(It may be necessary to prompt the user several times
before a non-blank line comes in. The prompt is `\.{**}' instead of the
later `\.*' because the meaning is slightly different: `\.{input}' need
not be typed immediately after~`\.{**}'.)

@d loc==cur_input.loc_field {location of first unread character in |buffer|}

@ The following program does the required initialization
without retrieving a possible command line.
It should be clear how to modify this routine to deal with command lines,
if the system permits them.
@^system dependencies@>

@p function init_terminal:boolean; {gets the terminal input started}
label exit;
begin t_open_in;
loop@+begin wake_up_terminal; write(term_out,'**'); update_terminal;
@.**@>
  if not input_ln(term_in,true) then {this shouldn't happen}
    begin write_ln(term_out);
    write(term_out,'! End of file on the terminal... why?');
@.End of file on the terminal@>
    init_terminal:=false; return;
    end;
  loc:=first;
  while (loc<last)and(buffer[loc]=" ") do incr(loc);
  if loc<last then
    begin init_terminal:=true;
    return; {return unless the line was all blank}
    end;
  write_ln(term_out,'Please type the name of your input file.');
  end;
exit:end;

@* \[4] String handling.
Symbolic token names and diagnostic messages are variable-length strings
of eight-bit characters. Since \PASCAL\ does not have a well-developed string
mechanism, \MF\ does all of its string processing by homegrown methods.

Elaborate facilities for dynamic strings are not needed, so all of the
necessary operations can be handled with a simple data structure.
The array |str_pool| contains all of the (eight-bit) ASCII codes in all
of the strings, and the array |str_start| contains indices of the starting
points of each string. Strings are referred to by integer numbers, so that
string number |s| comprises the characters |str_pool[j]| for
|str_start[s]<=j<str_start[s+1]|. Additional integer variables
|pool_ptr| and |str_ptr| indicate the number of entries used so far
in |str_pool| and |str_start|, respectively; locations
|str_pool[pool_ptr]| and |str_start[str_ptr]| are
ready for the next string to be allocated.

String numbers 0 to 255 are reserved for strings that correspond to single
ASCII characters. This is in accordance with the conventions of \.{WEB},
@.WEB@>
which converts single-character strings into the ASCII code number of the
single character involved, while it converts other strings into integers
and builds a string pool file. Thus, when the string constant \.{"."} appears
in the program below, \.{WEB} converts it into the integer 46, which is the
ASCII code for a period, while \.{WEB} will convert a string like \.{"hello"}
into some integer greater than~255. String number 46 will presumably be the
single character `\..'\thinspace; but some ASCII codes have no standard visible
representation, and \MF\ may need to be able to print an arbitrary
ASCII character, so the first 256 strings are used to specify exactly what
should be printed for each of the 256 possibilities.

Elements of the |str_pool| array must be ASCII codes that can actually be
printed; i.e., they must have an |xchr| equivalent in the local
character set. (This restriction applies only to preloaded strings,
not to those generated dynamically by the user.)

Some \PASCAL\ compilers won't pack integers into a single byte unless the
integers lie in the range |-128..127|. To accommodate such systems
we access the string pool only via macros that can easily be redefined.
@^system dependencies@>

@d si(#) == # {convert from |ASCII_code| to |packed_ASCII_code|}
@d so(#) == # {convert from |packed_ASCII_code| to |ASCII_code|}

@<Types...@>=
@!pool_pointer = 0..pool_size; {for variables that point into |str_pool|}
@!str_number = 0..max_strings; {for variables that point into |str_start|}
@!packed_ASCII_code = 0..255; {elements of |str_pool| array}

@ @<Glob...@>=
@!str_pool:packed array[pool_pointer] of packed_ASCII_code; {the characters}
@!str_start : array[str_number] of pool_pointer; {the starting pointers}
@!pool_ptr : pool_pointer; {first unused position in |str_pool|}
@!str_ptr : str_number; {number of the current string being created}
@!init_pool_ptr : pool_pointer; {the starting value of |pool_ptr|}
@!init_str_ptr : str_number; {the starting value of |str_ptr|}
@!max_pool_ptr : pool_pointer; {the maximum so far of |pool_ptr|}
@!max_str_ptr : str_number; {the maximum so far of |str_ptr|}

@ Several of the elementary string operations are performed using \.{WEB}
macros instead of \PASCAL\ procedures, because many of the
operations are done quite frequently and we want to avoid the
overhead of procedure calls. For example, here is
a simple macro that computes the length of a string.
@.WEB@>

@d length(#)==(str_start[#+1]-str_start[#]) {the number of characters
  in string number \#}

@ The length of the current string is called |cur_length|:

@d cur_length == (pool_ptr - str_start[str_ptr])

@ Strings are created by appending character codes to |str_pool|.
The |append_char| macro, defined here, does not check to see if the
value of |pool_ptr| has gotten too high; this test is supposed to be
made before |append_char| is used.

To test if there is room to append |l| more characters to |str_pool|,
we shall write |str_room(l)|, which aborts \MF\ and gives an
apologetic error message if there isn't enough room.

@d append_char(#) == {put |ASCII_code| \# at the end of |str_pool|}
begin str_pool[pool_ptr]:=si(#); incr(pool_ptr);
end
@d str_room(#) == {make sure that the pool hasn't overflowed}
  begin if pool_ptr+# > max_pool_ptr then
    begin if pool_ptr+# > pool_size then
      overflow("pool size",pool_size-init_pool_ptr);
@:METAFONT capacity exceeded pool size}{\quad pool size@>
    max_pool_ptr:=pool_ptr+#;
    end;
  end

@ \MF's string expressions are implemented in a brute-force way: Every
new string or substring that is needed is simply copied into the string pool.

Such a scheme can be justified because string expressions aren't a big
deal in \MF\ applications; strings rarely need to be saved from one
statement to the next. But it would waste space needlessly if we didn't
try to reclaim the space of strings that are going to be used only once.

Therefore a simple reference count mechanism is provided: If there are
@^reference counts@>
no references to a certain string from elsewhere in the program, and
if there are no references to any strings created subsequent to it,
then the string space will be reclaimed.

The number of references to string number |s| will be |str_ref[s]|. The
special value |str_ref[s]=max_str_ref=127| is used to denote an unknown
positive number of references; such strings will never be recycled. If
a string is ever referred to more than 126 times, simultaneously, we
put it in this category. Hence a single byte suffices to store each |str_ref|.

@d max_str_ref=127 {``infinite'' number of references}
@d add_str_ref(#)==begin if str_ref[#]<max_str_ref then incr(str_ref[#]);
  end

@<Glob...@>=
@!str_ref:array[str_number] of 0..max_str_ref;

@ Here's what we do when a string reference disappears:

@d delete_str_ref(#)== begin if str_ref[#]<max_str_ref then
    if str_ref[#]>1 then decr(str_ref[#])@+else flush_string(#);
    end

@<Declare the procedure called |flush_string|@>=
procedure flush_string(@!s:str_number);
begin if s<str_ptr-1 then str_ref[s]:=0
else  repeat decr(str_ptr);
  until str_ref[str_ptr-1]<>0;
pool_ptr:=str_start[str_ptr];
end;

@ Once a sequence of characters has been appended to |str_pool|, it
officially becomes a string when the function |make_string| is called.
This function returns the identification number of the new string as its
value.

@p function make_string : str_number; {current string enters the pool}
begin if str_ptr=max_str_ptr then
  begin if str_ptr=max_strings then
    overflow("number of strings",max_strings-init_str_ptr);
@:METAFONT capacity exceeded number of strings}{\quad number of strings@>
  incr(max_str_ptr);
  end;
str_ref[str_ptr]:=1; incr(str_ptr); str_start[str_ptr]:=pool_ptr;
make_string:=str_ptr-1;
end;

@ The following subroutine compares string |s| with another string of the
same length that appears in |buffer| starting at position |k|;
the result is |true| if and only if the strings are equal.

@p function str_eq_buf(@!s:str_number;@!k:integer):boolean;
  {test equality of strings}
label not_found; {loop exit}
var @!j: pool_pointer; {running index}
@!result: boolean; {result of comparison}
begin j:=str_start[s];
while j<str_start[s+1] do
  begin if so(str_pool[j])<>buffer[k] then
    begin result:=false; goto not_found;
    end;
  incr(j); incr(k);
  end;
result:=true;
not_found: str_eq_buf:=result;
end;

@ Here is a similar routine, but it compares two strings in the string pool,
and it does not assume that they have the same length. If the first string
is lexicographically greater than, less than, or equal to the second,
the result is respectively positive, negative, or zero.

@p function str_vs_str(@!s,@!t:str_number):integer;
  {test equality of strings}
label exit;
var @!j,@!k: pool_pointer; {running indices}
@!ls,@!lt:integer; {lengths}
@!l:integer; {length remaining to test}
begin ls:=length(s); lt:=length(t);
if ls<=lt then l:=ls@+else l:=lt;
j:=str_start[s]; k:=str_start[t];
while l>0 do
  begin if str_pool[j]<>str_pool[k] then
    begin str_vs_str:=str_pool[j]-str_pool[k]; return;
    end;
  incr(j); incr(k); decr(l);
  end;
str_vs_str:=ls-lt;
exit:end;

@ The initial values of |str_pool|, |str_start|, |pool_ptr|,
and |str_ptr| are computed by the \.{INIMF} program, based in part
on the information that \.{WEB} has output while processing \MF.
@.INIMF@>
@^string pool@>

@p @!init function get_strings_started:boolean; {initializes the string pool,
  but returns |false| if something goes wrong}
label done,exit;
var @!k,@!l:0..255; {small indices or counters}
@!m,@!n:text_char; {characters input from |pool_file|}
@!g:str_number; {the string just created}
@!a:integer; {accumulator for check sum}
@!c:boolean; {check sum has been checked}
begin pool_ptr:=0; str_ptr:=0; max_pool_ptr:=0; max_str_ptr:=0; str_start[0]:=0;
@<Make the first 256 strings@>;
@<Read the other strings from the \.{MF.POOL} file and return |true|,
  or give an error message and return |false|@>;
exit:end;
tini

@ @d app_lc_hex(#)==l:=#;
  if l<10 then append_char(l+"0")@+else append_char(l-10+"a")

@<Make the first 256...@>=
for k:=0 to 255 do
  begin if (@<Character |k| cannot be printed@>) then
    begin append_char("^"); append_char("^");
    if k<@'100 then append_char(k+@'100)
    else if k<@'200 then append_char(k-@'100)
    else begin app_lc_hex(k div 16); app_lc_hex(k mod 16);
      end;
    end
  else append_char(k);
  g:=make_string; str_ref[g]:=max_str_ref;
  end

@ The first 128 strings will contain 95 standard ASCII characters, and the
other 33 characters will be printed in three-symbol form like `\.{\^\^A}'
unless a system-dependent change is made here. Installations that have
an extended character set, where for example |xchr[@'32]=@t\.{\'^^Z\'}@>|,
would like string @'32 to be the single character @'32 instead of the
three characters @'136, @'136, @'132 (\.{\^\^Z}). On the other hand,
even people with an extended character set will want to represent string
@'15 by \.{\^\^M}, since @'15 is ASCII's ``carriage return'' code; the idea is
to produce visible strings instead of tabs or line-feeds or carriage-returns
or bell-rings or characters that are treated anomalously in text files.

Unprintable characters of codes 128--255 are, similarly, rendered
\.{\^\^80}--\.{\^\^ff}.

The boolean expression defined here should be |true| unless \MF\ internal
code number~|k| corresponds to a non-troublesome visible symbol in the
local character set.
If character |k| cannot be printed, and |k<@'200|, then character |k+@'100| or
|k-@'100| must be printable; moreover, ASCII codes
|[@'60..@'71, @'136, @'141..@'146]|
must be printable.
@^character set dependencies@>
@^system dependencies@>

@<Character |k| cannot be printed@>=
  (k<" ")or(k>"~")

@ When the \.{WEB} system program called \.{TANGLE} processes the \.{MF.WEB}
description that you are now reading, it outputs the \PASCAL\ program
\.{MF.PAS} and also a string pool file called \.{MF.POOL}. The \.{INIMF}
@.WEB@>@.INIMF@>
program reads the latter file, where each string appears as a two-digit decimal
length followed by the string itself, and the information is recorded in
\MF's string memory.

@<Glob...@>=
@!init @!pool_file:alpha_file; {the string-pool file output by \.{TANGLE}}
tini

@ @d bad_pool(#)==begin wake_up_terminal; write_ln(term_out,#);
  a_close(pool_file); get_strings_started:=false; return;
  end
@<Read the other strings...@>=
name_of_file:=pool_name; {we needn't set |name_length|}
if a_open_in(pool_file) then
  begin c:=false;
  repeat @<Read one string, but return |false| if the
    string memory space is getting too tight for comfort@>;
  until c;
  a_close(pool_file); get_strings_started:=true;
  end
else  bad_pool('! I can''t read MF.POOL.')
@.I can't read MF.POOL@>

@ @<Read one string...@>=
begin if eof(pool_file) then bad_pool('! MF.POOL has no check sum.');
@.MF.POOL has no check sum@>
read(pool_file,m,n); {read two digits of string length}
if m='*' then @<Check the pool check sum@>
else  begin if (xord[m]<"0")or(xord[m]>"9")or@|
      (xord[n]<"0")or(xord[n]>"9") then
    bad_pool('! MF.POOL line doesn''t begin with two digits.');
@.MF.POOL line doesn't...@>
  l:=xord[m]*10+xord[n]-"0"*11; {compute the length}
  if pool_ptr+l+string_vacancies>pool_size then
    bad_pool('! You have to increase POOLSIZE.');
@.You have to increase POOLSIZE@>
  for k:=1 to l do
    begin if eoln(pool_file) then m:=' '@+else read(pool_file,m);
    append_char(xord[m]);
    end;
  read_ln(pool_file); g:=make_string; str_ref[g]:=max_str_ref;
  end;
end

@ The \.{WEB} operation \.{@@\$} denotes the value that should be at the
end of this \.{MF.POOL} file; any other value means that the wrong pool
file has been loaded.
@^check sum@>

@<Check the pool check sum@>=
begin a:=0; k:=1;
loop@+  begin if (xord[n]<"0")or(xord[n]>"9") then
  bad_pool('! MF.POOL check sum doesn''t have nine digits.');
@.MF.POOL check sum...@>
  a:=10*a+xord[n]-"0";
  if k=9 then goto done;
  incr(k); read(pool_file,n);
  end;
done: if a<>@$ then bad_pool('! MF.POOL doesn''t match; TANGLE me again.');
@.MF.POOL doesn't match@>
c:=true;
end

@* \[5] On-line and off-line printing.
Messages that are sent to a user's terminal and to the transcript-log file
are produced by several `|print|' procedures. These procedures will
direct their output to a variety of places, based on the setting of
the global variable |selector|, which has the following possible
values:

\yskip
\hang |term_and_log|, the normal setting, prints on the terminal and on the
  transcript file.

\hang |log_only|, prints only on the transcript file.

\hang |term_only|, prints only on the terminal.

\hang |no_print|, doesn't print at all. This is used only in rare cases
  before the transcript file is open.

\hang |pseudo|, puts output into a cyclic buffer that is used
  by the |show_context| routine; when we get to that routine we shall discuss
  the reasoning behind this curious mode.

\hang |new_string|, appends the output to the current string in the
  string pool.

\yskip
\noindent The symbolic names `|term_and_log|', etc., have been assigned
numeric codes that satisfy the convenient relations |no_print+1=term_only|,
|no_print+2=log_only|, |term_only+2=log_only+1=term_and_log|.

Three additional global variables, |tally| and |term_offset| and
|file_offset|, record the number of characters that have been printed
since they were most recently cleared to zero. We use |tally| to record
the length of (possibly very long) stretches of printing; |term_offset|
and |file_offset|, on the other hand, keep track of how many characters
have appeared so far on the current line that has been output to the
terminal or to the transcript file, respectively.

@d no_print=0 {|selector| setting that makes data disappear}
@d term_only=1 {printing is destined for the terminal only}
@d log_only=2 {printing is destined for the transcript file only}
@d term_and_log=3 {normal |selector| setting}
@d pseudo=4 {special |selector| setting for |show_context|}
@d new_string=5 {printing is deflected to the string pool}
@d max_selector=5 {highest selector setting}

@<Glob...@>=
@!log_file : alpha_file; {transcript of \MF\ session}
@!selector : 0..max_selector; {where to print a message}
@!dig : array[0..22] of 0..15; {digits in a number being output}
@!tally : integer; {the number of characters recently printed}
@!term_offset : 0..max_print_line;
  {the number of characters on the current terminal line}
@!file_offset : 0..max_print_line;
  {the number of characters on the current file line}
@!trick_buf:array[0..error_line] of ASCII_code; {circular buffer for
  pseudoprinting}
@!trick_count: integer; {threshold for pseudoprinting, explained later}
@!first_count: integer; {another variable for pseudoprinting}

@ @<Initialize the output routines@>=
selector:=term_only; tally:=0; term_offset:=0; file_offset:=0;

@ Macro abbreviations for output to the terminal and to the log file are
defined here for convenience. Some systems need special conventions
for terminal output, and it is possible to adhere to those conventions
by changing |wterm|, |wterm_ln|, and |wterm_cr| here.
@^system dependencies@>

@d wterm(#)==write(term_out,#)
@d wterm_ln(#)==write_ln(term_out,#)
@d wterm_cr==write_ln(term_out)
@d wlog(#)==write(log_file,#)
@d wlog_ln(#)==write_ln(log_file,#)
@d wlog_cr==write_ln(log_file)

@ To end a line of text output, we call |print_ln|.

@<Basic print...@>=
procedure print_ln; {prints an end-of-line}
begin case selector of
term_and_log: begin wterm_cr; wlog_cr;
  term_offset:=0; file_offset:=0;
  end;
log_only: begin wlog_cr; file_offset:=0;
  end;
term_only: begin wterm_cr; term_offset:=0;
  end;
no_print,pseudo,new_string: do_nothing;
end; {there are no other cases}
end; {note that |tally| is not affected}

@ The |print_char| procedure sends one character to the desired destination,
using the |xchr| array to map it into an external character compatible with
|input_ln|. All printing comes through |print_ln| or |print_char|.

@<Basic printing...@>=
procedure print_char(@!s:ASCII_code); {prints a single character}
begin case selector of
term_and_log: begin wterm(xchr[s]); wlog(xchr[s]);
  incr(term_offset); incr(file_offset);
  if term_offset=max_print_line then
    begin wterm_cr; term_offset:=0;
    end;
  if file_offset=max_print_line then
    begin wlog_cr; file_offset:=0;
    end;
  end;
log_only: begin wlog(xchr[s]); incr(file_offset);
  if file_offset=max_print_line then print_ln;
  end;
term_only: begin wterm(xchr[s]); incr(term_offset);
  if term_offset=max_print_line then print_ln;
  end;
no_print: do_nothing;
pseudo: if tally<trick_count then trick_buf[tally mod error_line]:=s;
new_string: begin if pool_ptr<pool_size then append_char(s);
  end; {we drop characters if the string space is full}
end; {there are no other cases}
incr(tally);
end;

@ An entire string is output by calling |print|. Note that if we are outputting
the single standard ASCII character \.c, we could call |print("c")|, since
|"c"=99| is the number of a single-character string, as explained above. But
|print_char("c")| is quicker, so \MF\ goes directly to the |print_char|
routine when it knows that this is safe. (The present implementation
assumes that it is always safe to print a visible ASCII character.)
@^system dependencies@>

@<Basic print...@>=
procedure print(@!s:integer); {prints string |s|}
var @!j:pool_pointer; {current character code position}
begin if (s<0)or(s>=str_ptr) then s:="???"; {this can't happen}
@.???@>
if (s<256)and(selector>pseudo) then print_char(s)
else begin j:=str_start[s];
  while j<str_start[s+1] do
    begin print_char(so(str_pool[j])); incr(j);
    end;
  end;
end;

@ Sometimes it's necessary to print a string whose characters
may not be visible ASCII codes. In that case |slow_print| is used.

@<Basic print...@>=
procedure slow_print(@!s:integer); {prints string |s|}
var @!j:pool_pointer; {current character code position}
begin if (s<0)or(s>=str_ptr) then s:="???"; {this can't happen}
@.???@>
if (s<256)and(selector>pseudo) then print_char(s)
else begin j:=str_start[s];
  while j<str_start[s+1] do
    begin print(so(str_pool[j])); incr(j);
    end;
  end;
end;

@ Here is the very first thing that \MF\ prints: a headline that identifies
the version number and base name. The |term_offset| variable is temporarily
incorrect, but the discrepancy is not serious since we assume that this
part of the program is system dependent.
@^system dependencies@>

@<Initialize the output...@>=
wterm(banner);
if base_ident=0 then wterm_ln(' (no base preloaded)')
else  begin slow_print(base_ident); print_ln;
  end;
update_terminal;

@ The procedure |print_nl| is like |print|, but it makes sure that the
string appears at the beginning of a new line.

@<Basic print...@>=
procedure print_nl(@!s:str_number); {prints string |s| at beginning of line}
begin if ((term_offset>0)and(odd(selector)))or@|
  ((file_offset>0)and(selector>=log_only)) then print_ln;
print(s);
end;

@ An array of digits in the range |0..9| is printed by |print_the_digs|.

@<Basic print...@>=
procedure print_the_digs(@!k:eight_bits);
  {prints |dig[k-1]|$\,\ldots\,$|dig[0]|}
begin while k>0 do
  begin decr(k); print_char("0"+dig[k]);
  end;
end;

@ The following procedure, which prints out the decimal representation of a
given integer |n|, has been written carefully so that it works properly
if |n=0| or if |(-n)| would cause overflow. It does not apply |mod| or |div|
to negative arguments, since such operations are not implemented consistently
by all \PASCAL\ compilers.

@<Basic print...@>=
procedure print_int(@!n:integer); {prints an integer in decimal form}
var k:0..23; {index to current digit; we assume that $\vert n\vert<10^{23}$}
@!m:integer; {used to negate |n| in possibly dangerous cases}
begin k:=0;
if n<0 then
  begin print_char("-");
  if n>-100000000 then negate(n)
  else  begin m:=-1-n; n:=m div 10; m:=(m mod 10)+1; k:=1;
    if m<10 then dig[0]:=m
    else  begin dig[0]:=0; incr(n);
      end;
    end;
  end;
repeat dig[k]:=n mod 10; n:=n div 10; incr(k);
until n=0;
print_the_digs(k);
end;

@ \MF\ also makes use of a trivial procedure to print two digits. The
following subroutine is usually called with a parameter in the range |0<=n<=99|.

@p procedure print_dd(@!n:integer); {prints two least significant digits}
begin n:=abs(n) mod 100; print_char("0"+(n div 10));
print_char("0"+(n mod 10));
end;

@ Here is a procedure that asks the user to type a line of input,
assuming that the |selector| setting is either |term_only| or |term_and_log|.
The input is placed into locations |first| through |last-1| of the
|buffer| array, and echoed on the transcript file if appropriate.

This procedure is never called when |interaction<scroll_mode|.

@d prompt_input(#)==begin wake_up_terminal; print(#); term_input;
    end {prints a string and gets a line of input}

@p procedure term_input; {gets a line from the terminal}
var @!k:0..buf_size; {index into |buffer|}
begin update_terminal; {now the user sees the prompt for sure}
if not input_ln(term_in,true) then fatal_error("End of file on the terminal!");
@.End of file on the terminal@>
term_offset:=0; {the user's line ended with \<\rm return>}
decr(selector); {prepare to echo the input}
if last<>first then for k:=first to last-1 do print(buffer[k]);
print_ln; buffer[last]:="%"; incr(selector); {restore previous status}
end;

@* \[6] Reporting errors.
When something anomalous is detected, \MF\ typically does something like this:
$$\vbox{\halign{#\hfil\cr
|print_err("Something anomalous has been detected");|\cr
|help3("This is the first line of my offer to help.")|\cr
|("This is the second line. I'm trying to")|\cr
|("explain the best way for you to proceed.");|\cr
|error;|\cr}}$$
A two-line help message would be given using |help2|, etc.; these informal
helps should use simple vocabulary that complements the words used in the
official error message that was printed. (Outside the U.S.A., the help
messages should preferably be translated into the local vernacular. Each
line of help is at most 60 characters long, in the present implementation,
so that |max_print_line| will not be exceeded.)

The |print_err| procedure supplies a `\.!' before the official message,
and makes sure that the terminal is awake if a stop is going to occur.
The |error| procedure supplies a `\..' after the official message, then it
shows the location of the error; and if |interaction=error_stop_mode|,
it also enters into a dialog with the user, during which time the help
message may be printed.
@^system dependencies@>

@ The global variable |interaction| has four settings, representing increasing
amounts of user interaction:

@d batch_mode=0 {omits all stops and omits terminal output}
@d nonstop_mode=1 {omits all stops}
@d scroll_mode=2 {omits error stops}
@d error_stop_mode=3 {stops at every opportunity to interact}
@d print_err(#)==begin if interaction=error_stop_mode then wake_up_terminal;
  print_nl("! "); print(#);
@.!\relax@>
  end

@<Glob...@>=
@!interaction:batch_mode..error_stop_mode; {current level of interaction}

@ @<Set init...@>=interaction:=error_stop_mode;

@ \MF\ is careful not to call |error| when the print |selector| setting
might be unusual. The only possible values of |selector| at the time of
error messages are

\yskip\hang|no_print| (when |interaction=batch_mode|
  and |log_file| not yet open);

\hang|term_only| (when |interaction>batch_mode| and |log_file| not yet open);

\hang|log_only| (when |interaction=batch_mode| and |log_file| is open);

\hang|term_and_log| (when |interaction>batch_mode| and |log_file| is open).

@<Initialize the print |selector| based on |interaction|@>=
if interaction=batch_mode then selector:=no_print@+else selector:=term_only

@ A global variable |deletions_allowed| is set |false| if the |get_next|
routine is active when |error| is called; this ensures that |get_next|
will never be called recursively.
@^recursion@>

The global variable |history| records the worst level of error that
has been detected. It has four possible values: |spotless|, |warning_issued|,
|error_message_issued|, and |fatal_error_stop|.

Another global variable, |error_count|, is increased by one when an
|error| occurs without an interactive dialog, and it is reset to zero at
the end of every statement.  If |error_count| reaches 100, \MF\ decides
that there is no point in continuing further.

@d spotless=0 {|history| value when nothing has been amiss yet}
@d warning_issued=1 {|history| value when |begin_diagnostic| has been called}
@d error_message_issued=2 {|history| value when |error| has been called}
@d fatal_error_stop=3 {|history| value when termination was premature}

@<Glob...@>=
@!deletions_allowed:boolean; {is it safe for |error| to call |get_next|?}
@!history:spotless..fatal_error_stop; {has the source input been clean so far?}
@!error_count:-1..100; {the number of scrolled errors since the
  last statement ended}

@ The value of |history| is initially |fatal_error_stop|, but it will
be changed to |spotless| if \MF\ survives the initialization process.

@<Set init...@>=
deletions_allowed:=true; error_count:=0; {|history| is initialized elsewhere}

@ Since errors can be detected almost anywhere in \MF, we want to declare the
error procedures near the beginning of the program. But the error procedures
in turn use some other procedures, which need to be declared |forward|
before we get to |error| itself.

It is possible for |error| to be called recursively if some error arises
when |get_next| is being used to delete a token, and/or if some fatal error
occurs while \MF\ is trying to fix a non-fatal one. But such recursion
@^recursion@>
is never more than two levels deep.

@<Error handling...@>=
procedure@?normalize_selector; forward;@t\2@>@/
procedure@?get_next; forward;@t\2@>@/
procedure@?term_input; forward;@t\2@>@/
procedure@?show_context; forward;@t\2@>@/
procedure@?begin_file_reading; forward;@t\2@>@/
procedure@?open_log_file; forward;@t\2@>@/
procedure@?close_files_and_terminate; forward;@t\2@>@/
procedure@?clear_for_error_prompt; forward;@t\2@>@/
@t\4\hskip-\fontdimen2\font@>@;@+@!debug@+procedure@?debug_help;
  forward;@;@+gubed@;@/
@t\4@>@<Declare the procedure called |flush_string|@>

@ Individual lines of help are recorded in the array |help_line|, which
contains entries in positions |0..(help_ptr-1)|. They should be printed
in reverse order, i.e., with |help_line[0]| appearing last.

@d hlp1(#)==help_line[0]:=#;@+end
@d hlp2(#)==help_line[1]:=#; hlp1
@d hlp3(#)==help_line[2]:=#; hlp2
@d hlp4(#)==help_line[3]:=#; hlp3
@d hlp5(#)==help_line[4]:=#; hlp4
@d hlp6(#)==help_line[5]:=#; hlp5
@d help0==help_ptr:=0 {sometimes there might be no help}
@d help1==@+begin help_ptr:=1; hlp1 {use this with one help line}
@d help2==@+begin help_ptr:=2; hlp2 {use this with two help lines}
@d help3==@+begin help_ptr:=3; hlp3 {use this with three help lines}
@d help4==@+begin help_ptr:=4; hlp4 {use this with four help lines}
@d help5==@+begin help_ptr:=5; hlp5 {use this with five help lines}
@d help6==@+begin help_ptr:=6; hlp6 {use this with six help lines}

@<Glob...@>=
@!help_line:array[0..5] of str_number; {helps for the next |error|}
@!help_ptr:0..6; {the number of help lines present}
@!use_err_help:boolean; {should the |err_help| string be shown?}
@!err_help:str_number; {a string set up by \&{errhelp}}

@ @<Set init...@>=
help_ptr:=0; use_err_help:=false; err_help:=0;

@ The |jump_out| procedure just cuts across all active procedure levels and
goes to |end_of_MF|. This is the only nontrivial |@!goto| statement in the
whole program. It is used when there is no recovery from a particular error.

Some \PASCAL\ compilers do not implement non-local |goto| statements.
@^system dependencies@>
In such cases the body of |jump_out| should simply be
`|close_files_and_terminate|;\thinspace' followed by a call on some system
procedure that quietly terminates the program.

@<Error hand...@>=
procedure jump_out;
begin goto end_of_MF;
end;

@ Here now is the general |error| routine.

@<Error hand...@>=
procedure error; {completes the job of error reporting}
label continue,exit;
var @!c:ASCII_code; {what the user types}
@!s1,@!s2,@!s3:integer; {used to save global variables when deleting tokens}
@!j:pool_pointer; {character position being printed}
begin if history<error_message_issued then history:=error_message_issued;
print_char("."); show_context;
if interaction=error_stop_mode then @<Get user's advice and |return|@>;
incr(error_count);
if error_count=100 then
  begin print_nl("(That makes 100 errors; please try again.)");
@.That makes 100 errors...@>
  history:=fatal_error_stop; jump_out;
  end;
@<Put help message on the transcript file@>;
exit:end;

@ @<Get user's advice...@>=
loop@+begin continue: if interaction<>error_stop_mode then return;
  clear_for_error_prompt; prompt_input("? ");
@.?\relax@>
  if last=first then return;
  c:=buffer[first];
  if c>="a" then c:=c+"A"-"a"; {convert to uppercase}
  @<Interpret code |c| and |return| if done@>;
  end

@ It is desirable to provide an `\.E' option here that gives the user
an easy way to return from \MF\ to the system editor, with the offending
line ready to be edited. But such an extension requires some system
wizardry, so the present implementation simply types out the name of the
file that should be
edited and the relevant line number.
@^system dependencies@>

There is a secret `\.D' option available when the debugging routines haven't
been commented~out.
@^debugging@>

@<Interpret code |c| and |return| if done@>=
case c of
"0","1","2","3","4","5","6","7","8","9": if deletions_allowed then
  @<Delete |c-"0"| tokens and |goto continue|@>;
@t\4\4@>@;@+@!debug "D":begin debug_help;goto continue;@+end;@+gubed@/
"E": if file_ptr>0 then if input_stack[file_ptr].name_field>=256 then
  begin print_nl("You want to edit file ");
@.You want to edit file x@>
  slow_print(input_stack[file_ptr].name_field);
  print(" at line "); print_int(line);@/
  interaction:=scroll_mode; jump_out;
  end;
"H": @<Print the help information and |goto continue|@>;
"I":@<Introduce new material from the terminal and |return|@>;
"Q","R","S":@<Change the interaction level and |return|@>;
"X":begin interaction:=scroll_mode; jump_out;
  end;
othercases do_nothing
endcases;@/
@<Print the menu of available options@>

@ @<Print the menu...@>=
begin print("Type <return> to proceed, S to scroll future error messages,");@/
@.Type <return> to proceed...@>
print_nl("R to run without stopping, Q to run quietly,");@/
print_nl("I to insert something, ");
if file_ptr>0 then if input_stack[file_ptr].name_field>=256 then
  print("E to edit your file,");
if deletions_allowed then
  print_nl("1 or ... or 9 to ignore the next 1 to 9 tokens of input,");
print_nl("H for help, X to quit.");
end

@ Here the author of \MF\ apologizes for making use of the numerical
relation between |"Q"|, |"R"|, |"S"|, and the desired interaction settings
|batch_mode|, |nonstop_mode|, |scroll_mode|.
@^Knuth, Donald Ervin@>

@<Change the interaction...@>=
begin error_count:=0; interaction:=batch_mode+c-"Q";
print("OK, entering ");
case c of
"Q":begin print("batchmode"); decr(selector);
  end;
"R":print("nonstopmode");
"S":print("scrollmode");
end; {there are no other cases}
print("..."); print_ln; update_terminal; return;
end

@ When the following code is executed, |buffer[(first+1)..(last-1)]| may
contain the material inserted by the user; otherwise another prompt will
be given. In order to understand this part of the program fully, you need
to be familiar with \MF's input stacks.

@<Introduce new material...@>=
begin begin_file_reading; {enter a new syntactic level for terminal input}
if last>first+1 then
  begin loc:=first+1; buffer[first]:=" ";
  end
else  begin prompt_input("insert>"); loc:=first;
@.insert>@>
  end;
first:=last+1; cur_input.limit_field:=last; return;
end

@ We allow deletion of up to 99 tokens at a time.

@<Delete |c-"0"| tokens...@>=
begin s1:=cur_cmd; s2:=cur_mod; s3:=cur_sym; OK_to_interrupt:=false;
if (last>first+1) and (buffer[first+1]>="0")and(buffer[first+1]<="9") then
  c:=c*10+buffer[first+1]-"0"*11
else c:=c-"0";
while c>0 do
  begin get_next; {one-level recursive call of |error| is possible}
  @<Decrease the string reference count, if the current token is a string@>;
  decr(c);
  end;
cur_cmd:=s1; cur_mod:=s2; cur_sym:=s3; OK_to_interrupt:=true;
help2("I have just deleted some text, as you asked.")@/
("You can now delete more, or insert, or whatever.");
show_context; goto continue;
end

@ @<Print the help info...@>=
begin if use_err_help then
  begin @<Print the string |err_help|, possibly on several lines@>;
  use_err_help:=false;
  end
else  begin if help_ptr=0 then
    help2("Sorry, I don't know how to help in this situation.")@/
    @t\kern1em@>("Maybe you should try asking a human?");
  repeat decr(help_ptr); print(help_line[help_ptr]); print_ln;
  until help_ptr=0;
  end;
help4("Sorry, I already gave what help I could...")@/
  ("Maybe you should try asking a human?")@/
  ("An error might have occurred before I noticed any problems.")@/
  ("``If all else fails, read the instructions.''");@/
goto continue;
end

@ @<Print the string |err_help|, possibly on several lines@>=
j:=str_start[err_help];
while j<str_start[err_help+1] do
  begin if str_pool[j]<>si("%") then print(so(str_pool[j]))
  else if j+1=str_start[err_help+1] then print_ln
  else if str_pool[j+1]<>si("%") then print_ln
  else  begin incr(j); print_char("%");
    end;
  incr(j);
  end

@ @<Put help message on the transcript file@>=
if interaction>batch_mode then decr(selector); {avoid terminal output}
if use_err_help then
  begin print_nl("");
  @<Print the string |err_help|, possibly on several lines@>;
  end
else while help_ptr>0 do
  begin decr(help_ptr); print_nl(help_line[help_ptr]);
  end;
print_ln;
if interaction>batch_mode then incr(selector); {re-enable terminal output}
print_ln

@ In anomalous cases, the print selector might be in an unknown state;
the following subroutine is called to fix things just enough to keep
running a bit longer.

@p procedure normalize_selector;
begin if log_opened then selector:=term_and_log
else selector:=term_only;
if job_name=0 then open_log_file;
if interaction=batch_mode then decr(selector);
end;

@ The following procedure prints \MF's last words before dying.

@d succumb==begin if interaction=error_stop_mode then
    interaction:=scroll_mode; {no more interaction}
  if log_opened then error;
  @!debug if interaction>batch_mode then debug_help;@;@+gubed@;@/
  history:=fatal_error_stop; jump_out; {irrecoverable error}
  end

@<Error hand...@>=
procedure fatal_error(@!s:str_number); {prints |s|, and that's it}
begin normalize_selector;@/
print_err("Emergency stop"); help1(s); succumb;
@.Emergency stop@>
end;

@ Here is the most dreaded error message.

@<Error hand...@>=
procedure overflow(@!s:str_number;@!n:integer); {stop due to finiteness}
begin normalize_selector;
print_err("METAFONT capacity exceeded, sorry [");
@.METAFONT capacity exceeded ...@>
print(s); print_char("="); print_int(n); print_char("]");
help2("If you really absolutely need more capacity,")@/
  ("you can ask a wizard to enlarge me.");
succumb;
end;

@ The program might sometime run completely amok, at which point there is
no choice but to stop. If no previous error has been detected, that's bad
news; a message is printed that is really intended for the \MF\
maintenance person instead of the user (unless the user has been
particularly diabolical).  The index entries for `this can't happen' may
help to pinpoint the problem.
@^dry rot@>

@<Error hand...@>=
procedure confusion(@!s:str_number);
  {consistency check violated; |s| tells where}
begin normalize_selector;
if history<error_message_issued then
  begin print_err("This can't happen ("); print(s); print_char(")");
@.This can't happen@>
  help1("I'm broken. Please show this to someone who can fix can fix");
  end
else  begin print_err("I can't go on meeting you like this");
@.I can't go on...@>
  help2("One of your faux pas seems to have wounded me deeply...")@/
    ("in fact, I'm barely conscious. Please fix it and try again.");
  end;
succumb;
end;

@ Users occasionally want to interrupt \MF\ while it's running.
If the \PASCAL\ runtime system allows this, one can implement
a routine that sets the global variable |interrupt| to some nonzero value
when such an interrupt is signalled. Otherwise there is probably at least
a way to make |interrupt| nonzero using the \PASCAL\ debugger.
@^system dependencies@>
@^debugging@>

@d check_interrupt==begin if interrupt<>0 then pause_for_instructions;
  end

@<Global...@>=
@!interrupt:integer; {should \MF\ pause for instructions?}
@!OK_to_interrupt:boolean; {should interrupts be observed?}

@ @<Set init...@>=
interrupt:=0; OK_to_interrupt:=true;

@ When an interrupt has been detected, the program goes into its
highest interaction level and lets the user have the full flexibility of
the |error| routine.  \MF\ checks for interrupts only at times when it is
safe to do this.

@p procedure pause_for_instructions;
begin if OK_to_interrupt then
  begin interaction:=error_stop_mode;
  if (selector=log_only)or(selector=no_print) then
    incr(selector);
  print_err("Interruption");
@.Interruption@>
  help3("You rang?")@/
  ("Try to insert an instruction for me (e.g., `I show x;'),")@/
  ("unless you just want to quit by typing `X'.");
  deletions_allowed:=false; error; deletions_allowed:=true;
  interrupt:=0;
  end;
end;

@ Many of \MF's error messages state that a missing token has been
inserted behind the scenes. We can save string space and program space
by putting this common code into a subroutine.

@p procedure missing_err(@!s:str_number);
begin print_err("Missing `"); print(s); print("' has been inserted");
@.Missing...inserted@>
end;

@* \[7] Arithmetic with scaled numbers.
The principal computations performed by \MF\ are done entirely in terms of
integers less than $2^{31}$ in magnitude; thus, the arithmetic specified in this
program can be carried out in exactly the same way on a wide variety of
computers, including some small ones.
@^small computers@>

But \PASCAL\ does not define the @!|div|
operation in the case of negative dividends; for example, the result of
|(-2*n-1) div 2| is |-(n+1)| on some computers and |-n| on others.
There are two principal types of arithmetic: ``translation-preserving,''
in which the identity |(a+q*b)div b=(a div b)+q| is valid; and
``negation-preserving,'' in which |(-a)div b=-(a div b)|. This leads to
two \MF s, which can produce different results, although the differences
should be negligible when the language is being used properly.
The \TeX\ processor has been defined carefully so that both varieties
of arithmetic will produce identical output, but it would be too
inefficient to constrain \MF\ in a similar way.

@d el_gordo == @'17777777777 {$2^{31}-1$, the largest value that \MF\ likes}

@ One of \MF's most common operations is the calculation of
$\lfloor{a+b\over2}\rfloor$,
the midpoint of two given integers |a| and~|b|. The only decent way to do
this in \PASCAL\ is to write `|(a+b) div 2|'; but on most machines it is
far more efficient to calculate `|(a+b)| right shifted one bit'.

Therefore the midpoint operation will always be denoted by `|half(a+b)|'
in this program. If \MF\ is being implemented with languages that permit
binary shifting, the |half| macro should be changed to make this operation
as efficient as possible.

@d half(#)==(#) div 2

@ A single computation might use several subroutine calls, and it is
desirable to avoid producing multiple error messages in case of arithmetic
overflow. So the routines below set the global variable |arith_error| to |true|
instead of reporting errors directly to the user.
@^overflow in arithmetic@>

@<Glob...@>=
@!arith_error:boolean; {has arithmetic overflow occurred recently?}

@ @<Set init...@>=
arith_error:=false;

@ At crucial points the program will say |check_arith|, to test if
an arithmetic error has been detected.

@d check_arith==begin if arith_error then clear_arith;@+end

@p procedure clear_arith;
begin print_err("Arithmetic overflow");
@.Arithmetic overflow@>
help4("Uh, oh. A little while ago one of the quantities that I was")@/
  ("computing got too large, so I'm afraid your answers will be")@/
  ("somewhat askew. You'll probably have to adopt different")@/
  ("tactics next time. But I shall try to carry on anyway.");
error; arith_error:=false;
end;

@ Addition is not always checked to make sure that it doesn't overflow,
but in places where overflow isn't too unlikely the |slow_add| routine
is used.

@p function slow_add(@!x,@!y:integer):integer;
begin if x>=0 then
  if y<=el_gordo-x then slow_add:=x+y
  else  begin arith_error:=true; slow_add:=el_gordo;
    end
else  if -y<=el_gordo+x then slow_add:=x+y
  else  begin arith_error:=true; slow_add:=-el_gordo;
    end;
end;

@ Fixed-point arithmetic is done on {\sl scaled integers\/} that are multiples
of $2^{-16}$. In other words, a binary point is assumed to be sixteen bit
positions from the right end of a binary computer word.

@d quarter_unit == @'40000 {$2^{14}$, represents 0.250000}
@d half_unit == @'100000 {$2^{15}$, represents 0.50000}
@d three_quarter_unit == @'140000 {$3\cdot2^{14}$, represents 0.75000}
@d unity == @'200000 {$2^{16}$, represents 1.00000}
@d two == @'400000 {$2^{17}$, represents 2.00000}
@d three == @'600000 {$2^{17}+2^{16}$, represents 3.00000}

@<Types...@>=
@!scaled = integer; {this type is used for scaled integers}
@!small_number=0..63; {this type is self-explanatory}

@ The following function is used to create a scaled integer from a given decimal
fraction $(.d_0d_1\ldots d_{k-1})$, where |0<=k<=17|. The digit $d_i$ is
given in |dig[i]|, and the calculation produces a correctly rounded result.

@p function round_decimals(@!k:small_number) : scaled;
  {converts a decimal fraction}
var @!a:integer; {the accumulator}
begin a:=0;
while k>0 do
  begin decr(k); a:=(a+dig[k]*two) div 10;
  end;
round_decimals:=half(a+1);
end;

@ Conversely, here is a procedure analogous to |print_int|. If the output
of this procedure is subsequently read by \MF\ and converted by the
|round_decimals| routine above, it turns out that the original value will
be reproduced exactly. A decimal point is printed only if the value is
not an integer. If there is more than one way to print the result with
the optimum number of digits following the decimal point, the closest
possible value is given.

The invariant relation in the \&{repeat} loop is that a sequence of
decimal digits yet to be printed will yield the original number if and only if
they form a fraction~$f$ in the range $s-\delta\L10\cdot2^{16}f<s$.
We can stop if and only if $f=0$ satisfies this condition; the loop will
terminate before $s$ can possibly become zero.

@<Basic printing...@>=
procedure print_scaled(@!s:scaled); {prints scaled real, rounded to five
  digits}
var @!delta:scaled; {amount of allowable inaccuracy}
begin if s<0 then
  begin print_char("-"); negate(s); {print the sign, if negative}
  end;
print_int(s div unity); {print the integer part}
s:=10*(s mod unity)+5;
if s<>5 then
  begin delta:=10; print_char(".");
  repeat if delta>unity then
    s:=s+@'100000-(delta div 2); {round the final digit}
  print_char("0"+(s div unity)); s:=10*(s mod unity); delta:=delta*10;
  until s<=delta;
  end;
end;

@ We often want to print two scaled quantities in parentheses,
separated by a comma.

@<Basic printing...@>=
procedure print_two(@!x,@!y:scaled); {prints `|(x,y)|'}
begin print_char("("); print_scaled(x); print_char(","); print_scaled(y);
print_char(")");
end;

@ The |scaled| quantities in \MF\ programs are generally supposed to be
less than $2^{12}$ in absolute value, so \MF\ does much of its internal
arithmetic with 28~significant bits of precision. A |fraction| denotes
a scaled integer whose binary point is assumed to be 28 bit positions
from the right.

@d fraction_half==@'1000000000 {$2^{27}$, represents 0.50000000}
@d fraction_one==@'2000000000 {$2^{28}$, represents 1.00000000}
@d fraction_two==@'4000000000 {$2^{29}$, represents 2.00000000}
@d fraction_three==@'6000000000 {$3\cdot2^{28}$, represents 3.00000000}
@d fraction_four==@'10000000000 {$2^{30}$, represents 4.00000000}

@<Types...@>=
@!fraction=integer; {this type is used for scaled fractions}

@ In fact, the two sorts of scaling discussed above aren't quite
sufficient; \MF\ has yet another, used internally to keep track of angles
in units of $2^{-20}$ degrees.

@d forty_five_deg==@'264000000 {$45\cdot2^{20}$, represents $45^\circ$}
@d ninety_deg==@'550000000 {$90\cdot2^{20}$, represents $90^\circ$}
@d one_eighty_deg==@'1320000000 {$180\cdot2^{20}$, represents $180^\circ$}
@d three_sixty_deg==@'2640000000 {$360\cdot2^{20}$, represents $360^\circ$}

@<Types...@>=
@!angle=integer; {this type is used for scaled angles}

@ The |make_fraction| routine produces the |fraction| equivalent of
|p/q|, given integers |p| and~|q|; it computes the integer
$f=\lfloor2^{28}p/q+{1\over2}\rfloor$, when $p$ and $q$ are
positive. If |p| and |q| are both of the same scaled type |t|,
the ``type relation'' |make_fraction(t,t)=fraction| is valid;
and it's also possible to use the subroutine ``backwards,'' using
the relation |make_fraction(t,fraction)=t| between scaled types.

If the result would have magnitude $2^{31}$ or more, |make_fraction|
sets |arith_error:=true|. Most of \MF's internal computations have
been designed to avoid this sort of error.

Notice that if 64-bit integer arithmetic were available,
we could simply compute |@t$(2^{29}$@>*p+q)div (2*q)|.
But when we are restricted to \PASCAL's 32-bit arithmetic we
must either resort to multiple-precision maneuvering
or use a simple but slow iteration. The multiple-precision technique
would be about three times faster than the code adopted here, but it
would be comparatively long and tricky, involving about sixteen
additional multiplications and divisions.

This operation is part of \MF's ``inner loop''; indeed, it will
consume nearly 10\pct! of the running time (exclusive of input and output)
if the code below is left unchanged. A machine-dependent recoding
will therefore make \MF\ run faster. The present implementation
is highly portable, but slow; it avoids multiplication and division
except in the initial stage. System wizards should be careful to
replace it with a routine that is guaranteed to produce identical
results in all cases.
@^system dependencies@>

As noted below, a few more routines should also be replaced by machine-dependent
code, for efficiency. But when a procedure is not part of the ``inner loop,''
such changes aren't advisable; simplicity and robustness are
preferable to trickery, unless the cost is too high.
@^inner loop@>

@p function make_fraction(@!p,@!q:integer):fraction;
var @!f:integer; {the fraction bits, with a leading 1 bit}
@!n:integer; {the integer part of $\vert p/q\vert$}
@!negative:boolean; {should the result be negated?}
@!be_careful:integer; {disables certain compiler optimizations}
begin if p>=0 then negative:=false
else  begin negate(p); negative:=true;
  end;
if q<=0 then
  begin debug if q=0 then confusion("/");@;@+gubed@;@/
@:this can't happen /}{\quad \./@>
  negate(q); negative:=not negative;
  end;
n:=p div q; p:=p mod q;
if n>=8 then
  begin arith_error:=true;
  if negative then make_fraction:=-el_gordo@+else make_fraction:=el_gordo;
  end
else  begin n:=(n-1)*fraction_one;
  @<Compute $f=\lfloor 2^{28}(1+p/q)+{1\over2}\rfloor$@>;
  if negative then make_fraction:=-(f+n)@+else make_fraction:=f+n;
  end;
end;

@ The |repeat| loop here preserves the following invariant relations
between |f|, |p|, and~|q|:
(i)~|0<=p<q|; (ii)~$fq+p=2^k(q+p_0)$, where $k$ is an integer and
$p_0$ is the original value of~$p$.

Notice that the computation specifies
|(p-q)+p| instead of |(p+p)-q|, because the latter could overflow.
Let us hope that optimizing compilers do not miss this point; a
special variable |be_careful| is used to emphasize the necessary
order of computation. Optimizing compilers should keep |be_careful|
in a register, not store it in memory.
@^inner loop@>

@<Compute $f=\lfloor 2^{28}(1+p/q)+{1\over2}\rfloor$@>=
f:=1;
repeat be_careful:=p-q; p:=be_careful+p;
if p>=0 then f:=f+f+1
else  begin double(f); p:=p+q;
  end;
until f>=fraction_one;
be_careful:=p-q;
if be_careful+p>=0 then incr(f)

@ The dual of |make_fraction| is |take_fraction|, which multiplies a
given integer~|q| by a fraction~|f|. When the operands are positive, it
computes $p=\lfloor qf/2^{28}+{1\over2}\rfloor$, a symmetric function
of |q| and~|f|.

This routine is even more ``inner loopy'' than |make_fraction|;
the present implementation consumes almost 20\pct! of \MF's computation
time during typical jobs, so a machine-language or 64-bit
substitute is advisable.
@^inner loop@> @^system dependencies@>

@p function take_fraction(@!q:integer;@!f:fraction):integer;
var @!p:integer; {the fraction so far}
@!negative:boolean; {should the result be negated?}
@!n:integer; {additional multiple of $q$}
@!be_careful:integer; {disables certain compiler optimizations}
begin @<Reduce to the case that |f>=0| and |q>=0|@>;
if f<fraction_one then n:=0
else  begin n:=f div fraction_one; f:=f mod fraction_one;
  if q<=el_gordo div n then n:=n*q
  else  begin arith_error:=true; n:=el_gordo;
    end;
  end;
f:=f+fraction_one;
@<Compute $p=\lfloor qf/2^{28}+{1\over2}\rfloor-q$@>;
be_careful:=n-el_gordo;
if be_careful+p>0 then
  begin arith_error:=true; n:=el_gordo-p;
  end;
if negative then take_fraction:=-(n+p)
else take_fraction:=n+p;
end;

@ @<Reduce to the case that |f>=0| and |q>=0|@>=
if f>=0 then negative:=false
else  begin negate(f); negative:=true;
  end;
if q<0 then
  begin negate(q); negative:=not negative;
  end;

@ The invariant relations in this case are (i)~$\lfloor(qf+p)/2^k\rfloor
=\lfloor qf_0/2^{28}+{1\over2}\rfloor$, where $k$ is an integer and
$f_0$ is the original value of~$f$; (ii)~$2^k\L f<2^{k+1}$.
@^inner loop@>

@<Compute $p=\lfloor qf/2^{28}+{1\over2}\rfloor-q$@>=
p:=fraction_half; {that's $2^{27}$; the invariants hold now with $k=28$}
if q<fraction_four then
  repeat if odd(f) then p:=half(p+q)@+else p:=half(p);
  f:=half(f);
  until f=1
else  repeat if odd(f) then p:=p+half(q-p)@+else p:=half(p);
  f:=half(f);
  until f=1


@ When we want to multiply something by a |scaled| quantity, we use a scheme
analogous to |take_fraction| but with a different scaling.
Given positive operands, |take_scaled|
computes the quantity $p=\lfloor qf/2^{16}+{1\over2}\rfloor$.

Once again it is a good idea to use 64-bit arithmetic if
possible; otherwise |take_scaled| will use more than 2\pct! of the running time
when the Computer Modern fonts are being generated.
@^inner loop@>

@p function take_scaled(@!q:integer;@!f:scaled):integer;
var @!p:integer; {the fraction so far}
@!negative:boolean; {should the result be negated?}
@!n:integer; {additional multiple of $q$}
@!be_careful:integer; {disables certain compiler optimizations}
begin @<Reduce to the case that |f>=0| and |q>=0|@>;
if f<unity then n:=0
else  begin n:=f div unity; f:=f mod unity;
  if q<=el_gordo div n then n:=n*q
  else  begin arith_error:=true; n:=el_gordo;
    end;
  end;
f:=f+unity;
@<Compute $p=\lfloor qf/2^{16}+{1\over2}\rfloor-q$@>;
be_careful:=n-el_gordo;
if be_careful+p>0 then
  begin arith_error:=true; n:=el_gordo-p;
  end;
if negative then take_scaled:=-(n+p)
else take_scaled:=n+p;
end;

@ @<Compute $p=\lfloor qf/2^{16}+{1\over2}\rfloor-q$@>=
p:=half_unit; {that's $2^{15}$; the invariants hold now with $k=16$}
@^inner loop@>
if q<fraction_four then
  repeat if odd(f) then p:=half(p+q)@+else p:=half(p);
  f:=half(f);
  until f=1
else  repeat if odd(f) then p:=p+half(q-p)@+else p:=half(p);
  f:=half(f);
  until f=1

@ For completeness, there's also |make_scaled|, which computes a
quotient as a |scaled| number instead of as a |fraction|.
In other words, the result is $\lfloor2^{16}p/q+{1\over2}\rfloor$, if the
operands are positive. \ (This procedure is not used especially often,
so it is not part of \MF's inner loop.)

@p function make_scaled(@!p,@!q:integer):scaled;
var @!f:integer; {the fraction bits, with a leading 1 bit}
@!n:integer; {the integer part of $\vert p/q\vert$}
@!negative:boolean; {should the result be negated?}
@!be_careful:integer; {disables certain compiler optimizations}
begin if p>=0 then negative:=false
else  begin negate(p); negative:=true;
  end;
if q<=0 then
  begin debug if q=0 then confusion("/");@+gubed@;@/
@:this can't happen /}{\quad \./@>
  negate(q); negative:=not negative;
  end;
n:=p div q; p:=p mod q;
if n>=@'100000 then
  begin arith_error:=true;
  if negative then make_scaled:=-el_gordo@+else make_scaled:=el_gordo;
  end
else  begin n:=(n-1)*unity;
  @<Compute $f=\lfloor 2^{16}(1+p/q)+{1\over2}\rfloor$@>;
  if negative then make_scaled:=-(f+n)@+else make_scaled:=f+n;
  end;
end;

@ @<Compute $f=\lfloor 2^{16}(1+p/q)+{1\over2}\rfloor$@>=
f:=1;
repeat be_careful:=p-q; p:=be_careful+p;
if p>=0 then f:=f+f+1
else  begin double(f); p:=p+q;
  end;
until f>=unity;
be_careful:=p-q;
if be_careful+p>=0 then incr(f)

@ Here is a typical example of how the routines above can be used.
It computes the function
$${1\over3\tau}f(\theta,\phi)=
{\tau^{-1}\bigl(2+\sqrt2\,(\sin\theta-{1\over16}\sin\phi)
 (\sin\phi-{1\over16}\sin\theta)(\cos\theta-\cos\phi)\bigr)\over
3\,\bigl(1+{1\over2}(\sqrt5-1)\cos\theta+{1\over2}(3-\sqrt5\,)\cos\phi\bigr)},$$
where $\tau$ is a |scaled| ``tension'' parameter. This is \MF's magic
fudge factor for placing the first control point of a curve that starts
at an angle $\theta$ and ends at an angle $\phi$ from the straight path.
(Actually, if the stated quantity exceeds 4, \MF\ reduces it to~4.)

The trigonometric quantity to be multiplied by $\sqrt2$ is less than $\sqrt2$.
(It's a sum of eight terms whose absolute values can be bounded using
relations such as $\sin\theta\cos\theta\L{1\over2}$.) Thus the numerator
is positive; and since the tension $\tau$ is constrained to be at least
$3\over4$, the numerator is less than $16\over3$. The denominator is
nonnegative and at most~6.  Hence the fixed-point calculations below
are guaranteed to stay within the bounds of a 32-bit computer word.

The angles $\theta$ and $\phi$ are given implicitly in terms of |fraction|
arguments |st|, |ct|, |sf|, and |cf|, representing $\sin\theta$, $\cos\theta$,
$\sin\phi$, and $\cos\phi$, respectively.

@p function velocity(@!st,@!ct,@!sf,@!cf:fraction;@!t:scaled):fraction;
var @!acc,@!num,@!denom:integer; {registers for intermediate calculations}
begin acc:=take_fraction(st-(sf div 16), sf-(st div 16));
acc:=take_fraction(acc,ct-cf);
num:=fraction_two+take_fraction(acc,379625062);
  {$2^{28}\sqrt2\approx379625062.497$}
denom:=fraction_three+take_fraction(ct,497706707)+take_fraction(cf,307599661);
  {$3\cdot2^{27}\cdot(\sqrt5-1)\approx497706706.78$ and
    $3\cdot2^{27}\cdot(3-\sqrt5\,)\approx307599661.22$}
if t<>unity then num:=make_scaled(num,t);
  {|make_scaled(fraction,scaled)=fraction|}
if num div 4>=denom then velocity:=fraction_four
else velocity:=make_fraction(num,denom);
end;

@ The following somewhat different subroutine tests rigorously if $ab$ is
greater than, equal to, or less than~$cd$,
given integers $(a,b,c,d)$. In most cases a quick decision is reached.
The result is $+1$, 0, or~$-1$ in the three respective cases.

@d return_sign(#)==begin ab_vs_cd:=#; return;
  end

@p function ab_vs_cd(@!a,b,c,d:integer):integer;
label exit;
var @!q,@!r:integer; {temporary registers}
begin @<Reduce to the case that |a,c>=0|, |b,d>0|@>;
loop@+  begin q := a div d; r := c div b;
  if q<>r then
    if q>r then return_sign(1)@+else return_sign(-1);
  q := a mod d; r := c mod b;
  if r=0 then
    if q=0 then return_sign(0)@+else return_sign(1);
  if q=0 then return_sign(-1);
  a:=b; b:=q; c:=d; d:=r;
  end; {now |a>d>0| and |c>b>0|}
exit:end;

@ @<Reduce to the case that |a...@>=
if a<0 then
  begin negate(a); negate(b);
  end;
if c<0 then
  begin negate(c); negate(d);
  end;
if d<=0 then
  begin if b>=0 then
    if ((a=0)or(b=0))and((c=0)or(d=0)) then return_sign(0)
    else return_sign(1);
  if d=0 then
    if a=0 then return_sign(0)@+else return_sign(-1);
  q:=a; a:=c; c:=q; q:=-b; b:=-d; d:=q;
  end
else if b<=0 then
  begin if b<0 then if a>0 then return_sign(-1);
  if c=0 then return_sign(0) else return_sign(-1);
  end

@ We conclude this set of elementary routines with some simple rounding
and truncation operations that are coded in a machine-independent fashion.
The routines are slightly complicated because we want them to work
without overflow whenever $-2^{31}\L x<2^{31}$.

@p function floor_scaled(@!x:scaled):scaled;
  {$2^{16}\lfloor x/2^{16}\rfloor$}
var @!be_careful:integer; {temporary register}
begin if x>=0 then floor_scaled:=x-(x mod unity)
else  begin be_careful:=x+1;
  floor_scaled:=x+((-be_careful) mod unity)+1-unity;
  end;
end;
@#
function floor_unscaled(@!x:scaled):integer;
  {$\lfloor x/2^{16}\rfloor$}
var @!be_careful:integer; {temporary register}
begin if x>=0 then floor_unscaled:=x div unity
else  begin be_careful:=x+1; floor_unscaled:=-(1+((-be_careful) div unity));
  end;
end;
@#
function round_unscaled(@!x:scaled):integer;
  {$\lfloor x/2^{16}+.5\rfloor$}
var @!be_careful:integer; {temporary register}
begin if x>=half_unit then round_unscaled:=1+((x-half_unit) div unity)
else if x>=-half_unit then round_unscaled:=0
else  begin be_careful:=x+1;
  round_unscaled:=-(1+((-be_careful-half_unit) div unity));
  end;
end;
@#
function round_fraction(@!x:fraction):scaled;
  {$\lfloor x/2^{12}+.5\rfloor$}
var @!be_careful:integer; {temporary register}
begin if x>=2048 then round_fraction:=1+((x-2048) div 4096)
else if x>=-2048 then round_fraction:=0
else  begin be_careful:=x+1;
  round_fraction:=-(1+((-be_careful-2048) div 4096));
  end;
end;

@* \[8] Algebraic and transcendental functions.
\MF\ computes all of the necessary special functions from scratch, without
relying on |real| arithmetic or system subroutines for sines, cosines, etc.

@ To get the square root of a |scaled| number |x|, we want to calculate
$s=\lfloor 2^8\!\sqrt x +{1\over2}\rfloor$. If $x>0$, this is the unique
integer such that $2^{16}x-s\L s^2<2^{16}x+s$. The following subroutine
determines $s$ by an iterative method that maintains the invariant
relations $x=2^{46-2k}x_0\bmod 2^{30}$, $0<y=\lfloor 2^{16-2k}x_0\rfloor
-s^2+s\L q=2s$, where $x_0$ is the initial value of $x$. The value of~$y$
might, however, be zero at the start of the first iteration.

@p function square_rt(@!x:scaled):scaled;
var @!k:small_number; {iteration control counter}
@!y,@!q:integer; {registers for intermediate calculations}
begin if x<=0 then @<Handle square root of zero or negative argument@>
else  begin k:=23; q:=2;
  while x<fraction_two do {i.e., |while x<@t$2^{29}$@>|\unskip}
    begin decr(k); x:=x+x+x+x;
    end;
  if x<fraction_four then y:=0
  else  begin x:=x-fraction_four; y:=1;
    end;
  repeat @<Decrease |k| by 1, maintaining the invariant
    relations between |x|, |y|, and~|q|@>;
  until k=0;
  square_rt:=half(q);
  end;
end;

@ @<Handle square root of zero...@>=
begin if x<0 then
  begin print_err("Square root of ");
@.Square root...replaced by 0@>
  print_scaled(x); print(" has been replaced by 0");
  help2("Since I don't take square roots of negative numbers,")@/
    ("I'm zeroing this one. Proceed, with fingers crossed.");
  error;
  end;
square_rt:=0;
end

@ @<Decrease |k| by 1, maintaining...@>=
double(x); double(y);
if x>=fraction_four then {note that |fraction_four=@t$2^{30}$@>|}
  begin x:=x-fraction_four; incr(y);
  end;
double(x); y:=y+y-q; double(q);
if x>=fraction_four then
  begin x:=x-fraction_four; incr(y);
  end;
if y>q then
  begin y:=y-q; q:=q+2;
  end
else if y<=0 then
  begin q:=q-2; y:=y+q;
  end;
decr(k)

@ Pythagorean addition $\psqrt{a^2+b^2}$ is implemented by an elegant
iterative scheme due to Cleve Moler and Donald Morrison [{\sl IBM Journal
@^Moler, Cleve Barry@>
@^Morrison, Donald Ross@>
of Research and Development\/ \bf27} (1983), 577--581]. It modifies |a| and~|b|
in such a way that their Pythagorean sum remains invariant, while the
smaller argument decreases.

@p function pyth_add(@!a,@!b:integer):integer;
label done;
var @!r:fraction; {register used to transform |a| and |b|}
@!big:boolean; {is the result dangerously near $2^{31}$?}
begin a:=abs(a); b:=abs(b);
if a<b then
  begin r:=b; b:=a; a:=r;
  end; {now |0<=b<=a|}
if b>0 then
  begin if a<fraction_two then big:=false
  else  begin a:=a div 4; b:=b div 4; big:=true;
    end; {we reduced the precision to avoid arithmetic overflow}
  @<Replace |a| by an approximation to $\psqrt{a^2+b^2}$@>;
  if big then
    if a<fraction_two then a:=a+a+a+a
    else  begin arith_error:=true; a:=el_gordo;
      end;
  end;
pyth_add:=a;
end;

@ The key idea here is to reflect the vector $(a,b)$ about the
line through $(a,b/2)$.

@<Replace |a| by an approximation to $\psqrt{a^2+b^2}$@>=
loop@+  begin r:=make_fraction(b,a);
  r:=take_fraction(r,r); {now $r\approx b^2/a^2$}
  if r=0 then goto done;
  r:=make_fraction(r,fraction_four+r);
  a:=a+take_fraction(a+a,r); b:=take_fraction(b,r);
  end;
done:

@ Here is a similar algorithm for $\psqrt{a^2-b^2}$.
It converges slowly when $b$ is near $a$, but otherwise it works fine.

@p function pyth_sub(@!a,@!b:integer):integer;
label done;
var @!r:fraction; {register used to transform |a| and |b|}
@!big:boolean; {is the input dangerously near $2^{31}$?}
begin a:=abs(a); b:=abs(b);
if a<=b then @<Handle erroneous |pyth_sub| and set |a:=0|@>
else  begin if a<fraction_four then big:=false
  else  begin a:=half(a); b:=half(b); big:=true;
    end;
  @<Replace |a| by an approximation to $\psqrt{a^2-b^2}$@>;
  if big then a:=a+a;
  end;
pyth_sub:=a;
end;

@ @<Replace |a| by an approximation to $\psqrt{a^2-b^2}$@>=
loop@+  begin r:=make_fraction(b,a);
  r:=take_fraction(r,r); {now $r\approx b^2/a^2$}
  if r=0 then goto done;
  r:=make_fraction(r,fraction_four-r);
  a:=a-take_fraction(a+a,r); b:=take_fraction(b,r);
  end;
done:

@ @<Handle erroneous |pyth_sub| and set |a:=0|@>=
begin if a<b then
  begin print_err("Pythagorean subtraction "); print_scaled(a);
  print("+-+"); print_scaled(b); print(" has been replaced by 0");
@.Pythagorean...@>
  help2("Since I don't take square roots of negative numbers,")@/
    ("I'm zeroing this one. Proceed, with fingers crossed.");
  error;
  end;
a:=0;
end

@ The subroutines for logarithm and exponential involve two tables.
The first is simple: |two_to_the[k]| equals $2^k$. The second involves
a bit more calculation, which the author claims to have done correctly:
|spec_log[k]| is $2^{27}$ times $\ln\bigl(1/(1-2^{-k})\bigr)=
2^{-k}+{1\over2}2^{-2k}+{1\over3}2^{-3k}+\cdots\,$, rounded to the
nearest integer.

@<Glob...@>=
@!two_to_the:array[0..30] of integer; {powers of two}
@!spec_log:array[1..28] of integer; {special logarithms}

@ @<Local variables for initialization@>=
@!k:integer; {all-purpose loop index}

@ @<Set init...@>=
two_to_the[0]:=1;
for k:=1 to 30 do two_to_the[k]:=2*two_to_the[k-1];
spec_log[1]:=93032640;
spec_log[2]:=38612034;
spec_log[3]:=17922280;
spec_log[4]:=8662214;
spec_log[5]:=4261238;
spec_log[6]:=2113709;
spec_log[7]:=1052693;
spec_log[8]:=525315;
spec_log[9]:=262400;
spec_log[10]:=131136;
spec_log[11]:=65552;
spec_log[12]:=32772;
spec_log[13]:=16385;
for k:=14 to 27 do spec_log[k]:=two_to_the[27-k];
spec_log[28]:=1;

@ Here is the routine that calculates $2^8$ times the natural logarithm
of a |scaled| quantity; it is an integer approximation to $2^{24}\ln(x/2^{16})$,
when |x| is a given positive integer.

The method is based on exercise 1.2.2--25 in {\sl The Art of Computer
Programming\/}: During the main iteration we have $1\L 2^{-30}x<1/(1-2^{1-k})$,
and the logarithm of $2^{30}x$ remains to be added to an accumulator
register called~$y$. Three auxiliary bits of accuracy are retained in~$y$
during the calculation, and sixteen auxiliary bits to extend |y| are
kept in~|z| during the initial argument reduction. (We add
$100\cdot2^{16}=6553600$ to~|z| and subtract 100 from~|y| so that |z| will
not become negative; also, the actual amount subtracted from~|y| is~96,
not~100, because we want to add~4 for rounding before the final division by~8.)

@p function m_log(@!x:scaled):scaled;
var @!y,@!z:integer; {auxiliary registers}
@!k:integer; {iteration counter}
begin if x<=0 then @<Handle non-positive logarithm@>
else  begin y:=1302456956+4-100; {$14\times2^{27}\ln2\approx1302456956.421063$}
  z:=27595+6553600; {and $2^{16}\times .421063\approx 27595$}
  while x<fraction_four do
    begin double(x); y:=y-93032639; z:=z-48782;
    end; {$2^{27}\ln2\approx 93032639.74436163$
      and $2^{16}\times.74436163\approx 48782$}
  y:=y+(z div unity); k:=2;
  while x>fraction_four+4 do
    @<Increase |k| until |x| can be multiplied by a
      factor of $2^{-k}$, and adjust $y$ accordingly@>;
  m_log:=y div 8;
  end;
end;

@ @<Increase |k| until |x| can...@>=
begin z:=((x-1) div two_to_the[k])+1; {$z=\lceil x/2^k\rceil$}
while x<fraction_four+z do
  begin z:=half(z+1); k:=k+1;
  end;
y:=y+spec_log[k]; x:=x-z;
end

@ @<Handle non-positive logarithm@>=
begin print_err("Logarithm of ");
@.Logarithm...replaced by 0@>
print_scaled(x); print(" has been replaced by 0");
help2("Since I don't take logs of non-positive numbers,")@/
  ("I'm zeroing this one. Proceed, with fingers crossed.");
error; m_log:=0;
end

@ Conversely, the exponential routine calculates $\exp(x/2^8)$,
when |x| is |scaled|. The result is an integer approximation to
$2^{16}\exp(x/2^{24})$, when |x| is regarded as an integer.

@p function m_exp(@!x:scaled):scaled;
var @!k:small_number; {loop control index}
@!y,@!z:integer; {auxiliary registers}
begin if x>174436200 then
    {$2^{24}\ln((2^{31}-1)/2^{16})\approx 174436199.51$}
  begin arith_error:=true; m_exp:=el_gordo;
  end
else if x<-197694359 then m_exp:=0
    {$2^{24}\ln(2^{-1}/2^{16})\approx-197694359.45$}
else  begin if x<=0 then
    begin z:=-8*x; y:=@'4000000; {$y=2^{20}$}
    end
  else  begin if x<=127919879 then z:=1023359037-8*x
      {$2^{27}\ln((2^{31}-1)/2^{20})\approx 1023359037.125$}
    else z:=8*(174436200-x); {|z| is always nonnegative}
    y:=el_gordo;
    end;
  @<Multiply |y| by $\exp(-z/2^{27})$@>;
  if x<=127919879 then m_exp:=(y+8) div 16@+else m_exp:=y;
  end;
end;

@ The idea here is that subtracting |spec_log[k]| from |z| corresponds
to multiplying |y| by $1-2^{-k}$.

A subtle point (which had to be checked) was that if $x=127919879$, the
value of~|y| will decrease so that |y+8| doesn't overflow. In fact,
$z$ will be 5 in this case, and |y| will decrease by~64 when |k=25|
and by~16 when |k=27|.

@<Multiply |y| by...@>=
k:=1;
while z>0 do
  begin while z>=spec_log[k] do
    begin z:=z-spec_log[k];
    y:=y-1-((y-two_to_the[k-1]) div two_to_the[k]);
    end;
  incr(k);
  end

@ The trigonometric subroutines use an auxiliary table such that
|spec_atan[k]| contains an approximation to the |angle| whose tangent
is~$1/2^k$.

@<Glob...@>=
@!spec_atan:array[1..26] of angle; {$\arctan2^{-k}$ times $2^{20}\cdot180/\pi$}

@ @<Set init...@>=
spec_atan[1]:=27855475;
spec_atan[2]:=14718068;
spec_atan[3]:=7471121;
spec_atan[4]:=3750058;
spec_atan[5]:=1876857;
spec_atan[6]:=938658;
spec_atan[7]:=469357;
spec_atan[8]:=234682;
spec_atan[9]:=117342;
spec_atan[10]:=58671;
spec_atan[11]:=29335;
spec_atan[12]:=14668;
spec_atan[13]:=7334;
spec_atan[14]:=3667;
spec_atan[15]:=1833;
spec_atan[16]:=917;
spec_atan[17]:=458;
spec_atan[18]:=229;
spec_atan[19]:=115;
spec_atan[20]:=57;
spec_atan[21]:=29;
spec_atan[22]:=14;
spec_atan[23]:=7;
spec_atan[24]:=4;
spec_atan[25]:=2;
spec_atan[26]:=1;

@ Given integers |x| and |y|, not both zero, the |n_arg| function
returns the |angle| whose tangent points in the direction $(x,y)$.
This subroutine first determines the correct octant, then solves the
problem for |0<=y<=x|, then converts the result appropriately to
return an answer in the range |-one_eighty_deg<=@t$\theta$@><=one_eighty_deg|.
(The answer is |+one_eighty_deg| if |y=0| and |x<0|, but an answer of
|-one_eighty_deg| is possible if, for example, |y=-1| and $x=-2^{30}$.)

The octants are represented in a ``Gray code,'' since that turns out
to be computationally simplest.

@d negate_x=1
@d negate_y=2
@d switch_x_and_y=4
@d first_octant=1
@d second_octant=first_octant+switch_x_and_y
@d third_octant=first_octant+switch_x_and_y+negate_x
@d fourth_octant=first_octant+negate_x
@d fifth_octant=first_octant+negate_x+negate_y
@d sixth_octant=first_octant+switch_x_and_y+negate_x+negate_y
@d seventh_octant=first_octant+switch_x_and_y+negate_y
@d eighth_octant=first_octant+negate_y

@p function n_arg(@!x,@!y:integer):angle;
var @!z:angle; {auxiliary register}
@!t:integer; {temporary storage}
@!k:small_number; {loop counter}
@!octant:first_octant..sixth_octant; {octant code}
begin if x>=0 then octant:=first_octant
else  begin negate(x); octant:=first_octant+negate_x;
  end;
if y<0 then
  begin negate(y); octant:=octant+negate_y;
  end;
if x<y then
  begin t:=y; y:=x; x:=t; octant:=octant+switch_x_and_y;
  end;
if x=0 then @<Handle undefined arg@>
else  begin @<Set variable |z| to the arg of $(x,y)$@>;
  @<Return an appropriate answer based on |z| and |octant|@>;
  end;
end;

@ @<Handle undefined arg@>=
begin print_err("angle(0,0) is taken as zero");
@.angle(0,0)...zero@>
help2("The `angle' between two identical points is undefined.")@/
  ("I'm zeroing this one. Proceed, with fingers crossed.");
error; n_arg:=0;
end

@ @<Return an appropriate answer...@>=
case octant of
first_octant:n_arg:=z;
second_octant:n_arg:=ninety_deg-z;
third_octant:n_arg:=ninety_deg+z;
fourth_octant:n_arg:=one_eighty_deg-z;
fifth_octant:n_arg:=z-one_eighty_deg;
sixth_octant:n_arg:=-z-ninety_deg;
seventh_octant:n_arg:=z-ninety_deg;
eighth_octant:n_arg:=-z;
end {there are no other cases}

@ At this point we have |x>=y>=0|, and |x>0|. The numbers are scaled up
or down until $2^{28}\L x<2^{29}$, so that accurate fixed-point calculations
will be made.

@<Set variable |z| to the arg...@>=
while x>=fraction_two do
  begin x:=half(x); y:=half(y);
  end;
z:=0;
if y>0 then
  begin while x<fraction_one do
    begin double(x); double(y);
    end;
  @<Increase |z| to the arg of $(x,y)$@>;
  end

@ During the calculations of this section, variables |x| and~|y|
represent actual coordinates $(x,2^{-k}y)$. We will maintain the
condition |x>=y|, so that the tangent will be at most $2^{-k}$.
If $x<2y$, the tangent is greater than $2^{-k-1}$. The transformation
$(a,b)\mapsto(a+b\tan\phi,b-a\tan\phi)$ replaces $(a,b)$ by
coordinates whose angle has decreased by~$\phi$; in the special case
$a=x$, $b=2^{-k}y$, and $\tan\phi=2^{-k-1}$, this operation reduces
to the particularly simple iteration shown here. [Cf.~John E. Meggitt,
@^Meggitt, John E.@>
{\sl IBM Journal of Research and Development\/ \bf6} (1962), 210--226.]

The initial value of |x| will be multiplied by at most
$(1+{1\over2})(1+{1\over8})(1+{1\over32})\cdots\approx 1.7584$; hence
there is no chance of integer overflow.

@<Increase |z|...@>=
k:=0;
repeat double(y); incr(k);
if y>x then
  begin z:=z+spec_atan[k]; t:=x; x:=x+(y div two_to_the[k+k]); y:=y-t;
  end;
until k=15;
repeat double(y); incr(k);
if y>x then
  begin z:=z+spec_atan[k]; y:=y-x;
  end;
until k=26

@ Conversely, the |n_sin_cos| routine takes an |angle| and produces the sine
and cosine of that angle. The results of this routine are
stored in global integer variables |n_sin| and |n_cos|.

@<Glob...@>=
@!n_sin,@!n_cos:fraction; {results computed by |n_sin_cos|}

@ Given an integer |z| that is $2^{20}$ times an angle $\theta$ in degrees,
the purpose of |n_sin_cos(z)| is to set
|x=@t$r\cos\theta$@>| and |y=@t$r\sin\theta$@>| (approximately),
for some rather large number~|r|. The maximum of |x| and |y|
will be between $2^{28}$ and $2^{30}$, so that there will be hardly
any loss of accuracy. Then |x| and~|y| are divided by~|r|.

@p procedure n_sin_cos(@!z:angle); {computes a multiple of the sine and cosine}
var @!k:small_number; {loop control variable}
@!q:0..7; {specifies the quadrant}
@!r:fraction; {magnitude of |(x,y)|}
@!x,@!y,@!t:integer; {temporary registers}
begin while z<0 do z:=z+three_sixty_deg;
z:=z mod three_sixty_deg; {now |0<=z<three_sixty_deg|}
q:=z div forty_five_deg; z:=z mod forty_five_deg;
x:=fraction_one; y:=x;
if not odd(q) then z:=forty_five_deg-z;
@<Subtract angle |z| from |(x,y)|@>;
@<Convert |(x,y)| to the octant determined by~|q|@>;
r:=pyth_add(x,y); n_cos:=make_fraction(x,r); n_sin:=make_fraction(y,r);
end;

@ In this case the octants are numbered sequentially.

@<Convert |(x,...@>=
case q of
0:do_nothing;
1:begin t:=x; x:=y; y:=t;
  end;
2:begin t:=x; x:=-y; y:=t;
  end;
3:negate(x);
4:begin negate(x); negate(y);
  end;
5:begin t:=x; x:=-y; y:=-t;
  end;
6:begin t:=x; x:=y; y:=-t;
  end;
7:negate(y);
end {there are no other cases}

@ The main iteration of |n_sin_cos| is similar to that of |n_arg| but
applied in reverse. The values of |spec_atan[k]| decrease slowly enough
that this loop is guaranteed to terminate before the (nonexistent) value
|spec_atan[27]| would be required.

@<Subtract angle |z|...@>=
k:=1;
while z>0 do
  begin if z>=spec_atan[k] then
    begin z:=z-spec_atan[k]; t:=x;@/
    x:=t+y div two_to_the[k];
    y:=y-t div two_to_the[k];
    end;
  incr(k);
  end;
if y<0 then y:=0 {this precaution may never be needed}

@ And now let's complete our collection of numeric utility routines
by considering random number generation.
\MF\ generates pseudo-random numbers with the additive scheme recommended
in Section 3.6 of {\sl The Art of Computer Programming}; however, the
results are random fractions between 0 and |fraction_one-1|, inclusive.

There's an auxiliary array |randoms| that contains 55 pseudo-random
fractions. Using the recurrence $x_n=(x_{n-55}-x_{n-24})\bmod 2^{28}$,
we generate batches of 55 new $x_n$'s at a time by calling |new_randoms|.
The global variable |j_random| tells which element has most recently
been consumed.

@<Glob...@>=
@!randoms:array[0..54] of fraction; {the last 55 random values generated}
@!j_random:0..54; {the number of unused |randoms|}

@ To consume a random fraction, the program below will say `|next_random|'
and then it will fetch |randoms[j_random]|. The |next_random| macro
actually accesses the numbers backwards; blocks of 55~$x$'s are
essentially being ``flipped.'' But that doesn't make them less random.

@d next_random==if j_random=0 then new_randoms
  else decr(j_random)

@p procedure new_randoms;
var @!k:0..54; {index into |randoms|}
@!x:fraction; {accumulator}
begin for k:=0 to 23 do
  begin x:=randoms[k]-randoms[k+31];
  if x<0 then x:=x+fraction_one;
  randoms[k]:=x;
  end;
for k:=24 to 54 do
  begin x:=randoms[k]-randoms[k-24];
  if x<0 then x:=x+fraction_one;
  randoms[k]:=x;
  end;
j_random:=54;
end;

@ To initialize the |randoms| table, we call the following routine.

@p procedure init_randoms(@!seed:scaled);
var @!j,@!jj,@!k:fraction; {more or less random integers}
@!i:0..54; {index into |randoms|}
begin j:=abs(seed);
while j>=fraction_one do j:=half(j);
k:=1;
for i:=0 to 54 do
  begin jj:=k; k:=j-k; j:=jj;
  if k<0 then k:=k+fraction_one;
  randoms[(i*21)mod 55]:=j;
  end;
new_randoms; new_randoms; new_randoms; {``warm up'' the array}
end;

@ To produce a uniform random number in the range |0<=u<x| or |0>=u>x|
or |0=u=x|, given a |scaled| value~|x|, we proceed as shown here.

Note that the call of |take_fraction| will produce the values 0 and~|x|
with about half the probability that it will produce any other particular
values between 0 and~|x|, because it rounds its answers.

@p function unif_rand(@!x:scaled):scaled;
var @!y:scaled; {trial value}
begin next_random; y:=take_fraction(abs(x),randoms[j_random]);
if y=abs(x) then unif_rand:=0
else if x>0 then unif_rand:=y
else unif_rand:=-y;
end;

@ Finally, a normal deviate with mean zero and unit standard deviation
can readily be obtained with the ratio method (Algorithm 3.4.1R in
{\sl The Art of Computer Programming\/}).

@p function norm_rand:scaled;
var @!x,@!u,@!l:integer; {what the book would call $2^{16}X$, $2^{28}U$,
  and $-2^{24}\ln U$}
begin repeat
  repeat next_random;
  x:=take_fraction(112429,randoms[j_random]-fraction_half);
    {$2^{16}\sqrt{8/e}\approx 112428.82793$}
  next_random; u:=randoms[j_random];
  until abs(x)<u;
x:=make_fraction(x,u);
l:=139548960-m_log(u); {$2^{24}\cdot12\ln2\approx139548959.6165$}
until ab_vs_cd(1024,l,x,x)>=0;
norm_rand:=x;
end;

@* \[9] Packed data.
In order to make efficient use of storage space, \MF\ bases its major data
structures on a |memory_word|, which contains either a (signed) integer,
possibly scaled, or a small number of fields that are one half or one
quarter of the size used for storing integers.

If |x| is a variable of type |memory_word|, it contains up to four
fields that can be referred to as follows:
$$\vbox{\halign{\hfil#&#\hfil&#\hfil\cr
|x|&.|int|&(an |integer|)\cr
|x|&.|sc|\qquad&(a |scaled| integer)\cr
|x.hh.lh|, |x.hh|&.|rh|&(two halfword fields)\cr
|x.hh.b0|, |x.hh.b1|, |x.hh|&.|rh|&(two quarterword fields, one halfword
  field)\cr
|x.qqqq.b0|, |x.qqqq.b1|, |x.qqqq|&.|b2|, |x.qqqq.b3|\hskip-100pt
  &\qquad\qquad\qquad(four quarterword fields)\cr}}$$
This is somewhat cumbersome to write, and not very readable either, but
macros will be used to make the notation shorter and more transparent.
The \PASCAL\ code below gives a formal definition of |memory_word| and
its subsidiary types, using packed variant records. \MF\ makes no
assumptions about the relative positions of the fields within a word.

Since we are assuming 32-bit integers, a halfword must contain at least
16 bits, and a quarterword must contain at least 8 bits.
@^system dependencies@>
But it doesn't hurt to have more bits; for example, with enough 36-bit
words you might be able to have |mem_max| as large as 262142.

N.B.: Valuable memory space will be dreadfully wasted unless \MF\ is compiled
by a \PASCAL\ that packs all of the |memory_word| variants into
the space of a single integer. Some \PASCAL\ compilers will pack an
integer whose subrange is `|0..255|' into an eight-bit field, but others
insist on allocating space for an additional sign bit; on such systems you
can get 256 values into a quarterword only if the subrange is `|-128..127|'.

The present implementation tries to accommodate as many variations as possible,
so it makes few assumptions. If integers having the subrange
`|min_quarterword..max_quarterword|' can be packed into a quarterword,
and if integers having the subrange `|min_halfword..max_halfword|'
can be packed into a halfword, everything should work satisfactorily.

It is usually most efficient to have |min_quarterword=min_halfword=0|,
so one should try to achieve this unless it causes a severe problem.
The values defined here are recommended for most 32-bit computers.

@d min_quarterword=0 {smallest allowable value in a |quarterword|}
@d max_quarterword=255 {largest allowable value in a |quarterword|}
@d min_halfword==0 {smallest allowable value in a |halfword|}
@d max_halfword==65535 {largest allowable value in a |halfword|}

@ Here are the inequalities that the quarterword and halfword values
must satisfy (or rather, the inequalities that they mustn't satisfy):

@<Check the ``constant''...@>=
init if mem_max<>mem_top then bad:=10;@+tini@;@/
if mem_max<mem_top then bad:=10;
if (min_quarterword>0)or(max_quarterword<127) then bad:=11;
if (min_halfword>0)or(max_halfword<32767) then bad:=12;
if (min_quarterword<min_halfword)or@|
  (max_quarterword>max_halfword) then bad:=13;
if (mem_min<min_halfword)or(mem_max>=max_halfword) then bad:=14;
if max_strings>max_halfword then bad:=15;
if buf_size>max_halfword then bad:=16;
if (max_quarterword-min_quarterword<255)or@|
  (max_halfword-min_halfword<65535) then bad:=17;

@ The operation of subtracting |min_halfword| occurs rather frequently in
\MF, so it is convenient to abbreviate this operation by using the macro
|ho| defined here.  \MF\ will run faster with respect to compilers that
don't optimize the expression `|x-0|', if this macro is simplified in the
obvious way when |min_halfword=0|. Similarly, |qi| and |qo| are used for
input to and output from quarterwords.
@^system dependencies@>

@d ho(#)==#-min_halfword
  {to take a sixteen-bit item from a halfword}
@d qo(#)==#-min_quarterword {to read eight bits from a quarterword}
@d qi(#)==#+min_quarterword {to store eight bits in a quarterword}

@ The reader should study the following definitions closely:
@^system dependencies@>

@d sc==int {|scaled| data is equivalent to |integer|}

@<Types...@>=
@!quarterword = min_quarterword..max_quarterword; {1/4 of a word}
@!halfword=min_halfword..max_halfword; {1/2 of a word}
@!two_choices = 1..2; {used when there are two variants in a record}
@!three_choices = 1..3; {used when there are three variants in a record}
@!two_halves = packed record@;@/
  @!rh:halfword;
  case two_choices of
  1: (@!lh:halfword);
  2: (@!b0:quarterword; @!b1:quarterword);
  end;
@!four_quarters = packed record@;@/
  @!b0:quarterword;
  @!b1:quarterword;
  @!b2:quarterword;
  @!b3:quarterword;
  end;
@!memory_word = record@;@/
  case three_choices of
  1: (@!int:integer);
  2: (@!hh:two_halves);
  3: (@!qqqq:four_quarters);
  end;
@!word_file = file of memory_word;

@ When debugging, we may want to print a |memory_word| without knowing
what type it is; so we print it in all modes.
@^dirty \PASCAL@>@^debugging@>

@p @!debug procedure print_word(@!w:memory_word);
  {prints |w| in all ways}
begin print_int(w.int); print_char(" ");@/
print_scaled(w.sc); print_char(" "); print_scaled(w.sc div @'10000); print_ln;@/
print_int(w.hh.lh); print_char("="); print_int(w.hh.b0); print_char(":");
print_int(w.hh.b1); print_char(";"); print_int(w.hh.rh); print_char(" ");@/
print_int(w.qqqq.b0); print_char(":"); print_int(w.qqqq.b1); print_char(":");
print_int(w.qqqq.b2); print_char(":"); print_int(w.qqqq.b3);
end;
gubed

@* \[10] Dynamic memory allocation.
The \MF\ system does nearly all of its own memory allocation, so that it
can readily be transported into environments that do not have automatic
facilities for strings, garbage collection, etc., and so that it can be in
control of what error messages the user receives. The dynamic storage
requirements of \MF\ are handled by providing a large array |mem| in
which consecutive blocks of words are used as nodes by the \MF\ routines.

Pointer variables are indices into this array, or into another array
called |eqtb| that will be explained later. A pointer variable might
also be a special flag that lies outside the bounds of |mem|, so we
allow pointers to assume any |halfword| value. The minimum memory
index represents a null pointer.

@d pointer==halfword {a flag or a location in |mem| or |eqtb|}
@d null==mem_min {the null pointer}

@ The |mem| array is divided into two regions that are allocated separately,
but the dividing line between these two regions is not fixed; they grow
together until finding their ``natural'' size in a particular job.
Locations less than or equal to |lo_mem_max| are used for storing
variable-length records consisting of two or more words each. This region
is maintained using an algorithm similar to the one described in exercise
2.5--19 of {\sl The Art of Computer Programming}. However, no size field
appears in the allocated nodes; the program is responsible for knowing the
relevant size when a node is freed. Locations greater than or equal to
|hi_mem_min| are used for storing one-word records; a conventional
\.{AVAIL} stack is used for allocation in this region.

Locations of |mem| between |mem_min| and |mem_top| may be dumped as part
of preloaded base files, by the \.{INIMF} preprocessor.
@.INIMF@>
Production versions of \MF\ may extend the memory at the top end in order to
provide more space; these locations, between |mem_top| and |mem_max|,
are always used for single-word nodes.

The key pointers that govern |mem| allocation have a prescribed order:
$$\hbox{|null=mem_min<lo_mem_max<hi_mem_min<mem_top<=mem_end<=mem_max|.}$$

@<Glob...@>=
@!mem : array[mem_min..mem_max] of memory_word; {the big dynamic storage area}
@!lo_mem_max : pointer; {the largest location of variable-size memory in use}
@!hi_mem_min : pointer; {the smallest location of one-word memory in use}

@ Users who wish to study the memory requirements of specific applications can
use optional special features that keep track of current and
maximum memory usage. When code between the delimiters |@!stat| $\ldots$
|tats| is not ``commented out,'' \MF\ will run a bit slower but it will
report these statistics when |tracing_stats| is positive.

@<Glob...@>=
@!var_used, @!dyn_used : integer; {how much memory is in use}

@ Let's consider the one-word memory region first, since it's the
simplest. The pointer variable |mem_end| holds the highest-numbered location
of |mem| that has ever been used. The free locations of |mem| that
occur between |hi_mem_min| and |mem_end|, inclusive, are of type
|two_halves|, and we write |info(p)| and |link(p)| for the |lh|
and |rh| fields of |mem[p]| when it is of this type. The single-word
free locations form a linked list
$$|avail|,\;\hbox{|link(avail)|},\;\hbox{|link(link(avail))|},\;\ldots$$
terminated by |null|.

@d link(#) == mem[#].hh.rh {the |link| field of a memory word}
@d info(#) == mem[#].hh.lh {the |info| field of a memory word}

@<Glob...@>=
@!avail : pointer; {head of the list of available one-word nodes}
@!mem_end : pointer; {the last one-word node used in |mem|}

@ If one-word memory is exhausted, it might mean that the user has forgotten
a token like `\&{enddef}' or `\&{endfor}'. We will define some procedures
later that try to help pinpoint the trouble.

@p @t\4@>@<Declare the procedure called |show_token_list|@>@;
@t\4@>@<Declare the procedure called |runaway|@>

@ The function |get_avail| returns a pointer to a new one-word node whose
|link| field is null. However, \MF\ will halt if there is no more room left.
@^inner loop@>

@p function get_avail : pointer; {single-word node allocation}
var @!p:pointer; {the new node being got}
begin p:=avail; {get top location in the |avail| stack}
if p<>null then avail:=link(avail) {and pop it off}
else if mem_end<mem_max then {or go into virgin territory}
  begin incr(mem_end); p:=mem_end;
  end
else   begin decr(hi_mem_min); p:=hi_mem_min;
  if hi_mem_min<=lo_mem_max then
    begin runaway; {if memory is exhausted, display possible runaway text}
    overflow("main memory size",mem_max+1-mem_min);
      {quit; all one-word nodes are busy}
@:METAFONT capacity exceeded main memory size}{\quad main memory size@>
    end;
  end;
link(p):=null; {provide an oft-desired initialization of the new node}
@!stat incr(dyn_used);@+tats@;{maintain statistics}
get_avail:=p;
end;

@ Conversely, a one-word node is recycled by calling |free_avail|.

@d free_avail(#)== {single-word node liberation}
  begin link(#):=avail; avail:=#;
  @!stat decr(dyn_used);@+tats@/
  end

@ There's also a |fast_get_avail| routine, which saves the procedure-call
overhead at the expense of extra programming. This macro is used in
the places that would otherwise account for the most calls of |get_avail|.
@^inner loop@>

@d fast_get_avail(#)==@t@>@;@/
  begin #:=avail; {avoid |get_avail| if possible, to save time}
  if #=null then #:=get_avail
  else  begin avail:=link(#); link(#):=null;
    @!stat incr(dyn_used);@+tats@/
    end;
  end

@ The available-space list that keeps track of the variable-size portion
of |mem| is a nonempty, doubly-linked circular list of empty nodes,
pointed to by the roving pointer |rover|.

Each empty node has size 2 or more; the first word contains the special
value |max_halfword| in its |link| field and the size in its |info| field;
the second word contains the two pointers for double linking.

Each nonempty node also has size 2 or more. Its first word is of type
|two_halves|\kern-1pt, and its |link| field is never equal to |max_halfword|.
Otherwise there is complete flexibility with respect to the contents
of its other fields and its other words.

(We require |mem_max<max_halfword| because terrible things can happen
when |max_halfword| appears in the |link| field of a nonempty node.)

@d empty_flag == max_halfword {the |link| of an empty variable-size node}
@d is_empty(#) == (link(#)=empty_flag) {tests for empty node}
@d node_size == info {the size field in empty variable-size nodes}
@d llink(#) == info(#+1) {left link in doubly-linked list of empty nodes}
@d rlink(#) == link(#+1) {right link in doubly-linked list of empty nodes}

@<Glob...@>=
@!rover : pointer; {points to some node in the list of empties}

@ A call to |get_node| with argument |s| returns a pointer to a new node
of size~|s|, which must be 2~or more. The |link| field of the first word
of this new node is set to null. An overflow stop occurs if no suitable
space exists.

If |get_node| is called with $s=2^{30}$, it simply merges adjacent free
areas and returns the value |max_halfword|.

@p function get_node(@!s:integer):pointer; {variable-size node allocation}
label found,exit,restart;
var @!p:pointer; {the node currently under inspection}
@!q:pointer; {the node physically after node |p|}
@!r:integer; {the newly allocated node, or a candidate for this honor}
@!t,@!tt:integer; {temporary registers}
@^inner loop@>
begin restart: p:=rover; {start at some free node in the ring}
repeat @<Try to allocate within node |p| and its physical successors,
  and |goto found| if allocation was possible@>;
p:=rlink(p); {move to the next node in the ring}
until p=rover; {repeat until the whole list has been traversed}
if s=@'10000000000 then
  begin get_node:=max_halfword; return;
  end;
if lo_mem_max+2<hi_mem_min then if lo_mem_max+2<=mem_min+max_halfword then
  @<Grow more variable-size memory and |goto restart|@>;
overflow("main memory size",mem_max+1-mem_min);
  {sorry, nothing satisfactory is left}
@:METAFONT capacity exceeded main memory size}{\quad main memory size@>
found: link(r):=null; {this node is now nonempty}
@!stat var_used:=var_used+s; {maintain usage statistics}
tats@;@/
get_node:=r;
exit:end;

@ The lower part of |mem| grows by 1000 words at a time, unless
we are very close to going under. When it grows, we simply link
a new node into the available-space list. This method of controlled
growth helps to keep the |mem| usage consecutive when \MF\ is
implemented on ``virtual memory'' systems.
@^virtual memory@>

@<Grow more variable-size memory and |goto restart|@>=
begin if hi_mem_min-lo_mem_max>=1998 then t:=lo_mem_max+1000
else t:=lo_mem_max+1+(hi_mem_min-lo_mem_max) div 2;
  {|lo_mem_max+2<=t<hi_mem_min|}
if t>mem_min+max_halfword then t:=mem_min+max_halfword;
p:=llink(rover); q:=lo_mem_max; rlink(p):=q; llink(rover):=q;@/
rlink(q):=rover; llink(q):=p; link(q):=empty_flag; node_size(q):=t-lo_mem_max;@/
lo_mem_max:=t; link(lo_mem_max):=null; info(lo_mem_max):=null;
rover:=q; goto restart;
end

@ @<Try to allocate...@>=
q:=p+node_size(p); {find the physical successor}
while is_empty(q) do {merge node |p| with node |q|}
  begin t:=rlink(q); tt:=llink(q);
@^inner loop@>
  if q=rover then rover:=t;
  llink(t):=tt; rlink(tt):=t;@/
  q:=q+node_size(q);
  end;
r:=q-s;
if r>p+1 then @<Allocate from the top of node |p| and |goto found|@>;
if r=p then if rlink(p)<>p then
  @<Allocate entire node |p| and |goto found|@>;
node_size(p):=q-p {reset the size in case it grew}

@ @<Allocate from the top...@>=
begin node_size(p):=r-p; {store the remaining size}
rover:=p; {start searching here next time}
goto found;
end

@ Here we delete node |p| from the ring, and let |rover| rove around.

@<Allocate entire...@>=
begin rover:=rlink(p); t:=llink(p);
llink(rover):=t; rlink(t):=rover;
goto found;
end

@ Conversely, when some variable-size node |p| of size |s| is no longer needed,
the operation |free_node(p,s)| will make its words available, by inserting
|p| as a new empty node just before where |rover| now points.

@p procedure free_node(@!p:pointer; @!s:halfword); {variable-size node
  liberation}
var @!q:pointer; {|llink(rover)|}
begin node_size(p):=s; link(p):=empty_flag;
@^inner loop@>
q:=llink(rover); llink(p):=q; rlink(p):=rover; {set both links}
llink(rover):=p; rlink(q):=p; {insert |p| into the ring}
@!stat var_used:=var_used-s;@+tats@;{maintain statistics}
end;

@ Just before \.{INIMF} writes out the memory, it sorts the doubly linked
available space list. The list is probably very short at such times, so a
simple insertion sort is used. The smallest available location will be
pointed to by |rover|, the next-smallest by |rlink(rover)|, etc.

@p @!init procedure sort_avail; {sorts the available variable-size nodes
  by location}
var @!p,@!q,@!r: pointer; {indices into |mem|}
@!old_rover:pointer; {initial |rover| setting}
begin p:=get_node(@'10000000000); {merge adjacent free areas}
p:=rlink(rover); rlink(rover):=max_halfword; old_rover:=rover;
while p<>old_rover do @<Sort |p| into the list starting at |rover|
  and advance |p| to |rlink(p)|@>;
p:=rover;
while rlink(p)<>max_halfword do
  begin llink(rlink(p)):=p; p:=rlink(p);
  end;
rlink(p):=rover; llink(rover):=p;
end;
tini

@ The following |while| loop is guaranteed to
terminate, since the list that starts at
|rover| ends with |max_halfword| during the sorting procedure.

@<Sort |p|...@>=
if p<rover then
  begin q:=p; p:=rlink(q); rlink(q):=rover; rover:=q;
  end
else  begin q:=rover;
  while rlink(q)<p do q:=rlink(q);
  r:=rlink(p); rlink(p):=rlink(q); rlink(q):=p; p:=r;
  end

@* \[11] Memory layout.
Some areas of |mem| are dedicated to fixed usage, since static allocation is
more efficient than dynamic allocation when we can get away with it. For
example, locations |mem_min| to |mem_min+2| are always used to store the
specification for null pen coordinates that are `$(0,0)$'. The
following macro definitions accomplish the static allocation by giving
symbolic names to the fixed positions. Static variable-size nodes appear
in locations |mem_min| through |lo_mem_stat_max|, and static single-word nodes
appear in locations |hi_mem_stat_min| through |mem_top|, inclusive.

@d null_coords==mem_min {specification for pen offsets of $(0,0)$}
@d null_pen==null_coords+3 {we will define |coord_node_size=3|}
@d dep_head==null_pen+10 {and |pen_node_size=10|}
@d zero_val==dep_head+2 {two words for a permanently zero value}
@d temp_val==zero_val+2 {two words for a temporary value node}
@d end_attr==temp_val {we use |end_attr+2| only}
@d inf_val==end_attr+2 {and |inf_val+1| only}
@d bad_vardef==inf_val+2 {two words for \&{vardef} error recovery}
@d lo_mem_stat_max==bad_vardef+1  {largest statically
  allocated word in the variable-size |mem|}
@#
@d sentinel==mem_top {end of sorted lists}
@d temp_head==mem_top-1 {head of a temporary list of some kind}
@d hold_head==mem_top-2 {head of a temporary list of another kind}
@d hi_mem_stat_min==mem_top-2 {smallest statically allocated word in
  the one-word |mem|}

@ The following code gets the dynamic part of |mem| off to a good start,
when \MF\ is initializing itself the slow way.

@<Initialize table entries (done by \.{INIMF} only)@>=
rover:=lo_mem_stat_max+1; {initialize the dynamic memory}
link(rover):=empty_flag;
node_size(rover):=1000; {which is a 1000-word available node}
llink(rover):=rover; rlink(rover):=rover;@/
lo_mem_max:=rover+1000; link(lo_mem_max):=null; info(lo_mem_max):=null;@/
for k:=hi_mem_stat_min to mem_top do
  mem[k]:=mem[lo_mem_max]; {clear list heads}
avail:=null; mem_end:=mem_top;
hi_mem_min:=hi_mem_stat_min; {initialize the one-word memory}
var_used:=lo_mem_stat_max+1-mem_min; dyn_used:=mem_top+1-hi_mem_min;
  {initialize statistics}

@ The procedure |flush_list(p)| frees an entire linked list of one-word
nodes that starts at a given position, until coming to |sentinel| or a
pointer that is not in the one-word region. Another procedure,
|flush_node_list|, frees an entire linked list of one-word and two-word
nodes, until coming to a |null| pointer.
@^inner loop@>

@p procedure flush_list(@!p:pointer); {makes list of single-word nodes
  available}
label done;
var @!q,@!r:pointer; {list traversers}
begin if p>=hi_mem_min then if p<>sentinel then
  begin r:=p;
  repeat q:=r; r:=link(r); @!stat decr(dyn_used);@+tats@/
  if r<hi_mem_min then goto done;
  until r=sentinel;
  done: {now |q| is the last node on the list}
  link(q):=avail; avail:=p;
  end;
end;
@#
procedure flush_node_list(@!p:pointer);
var @!q:pointer; {the node being recycled}
begin while p<>null do
  begin q:=p; p:=link(p);
  if q<hi_mem_min then free_node(q,2)@+else free_avail(q);
  end;
end;

@ If \MF\ is extended improperly, the |mem| array might get screwed up.
For example, some pointers might be wrong, or some ``dead'' nodes might not
have been freed when the last reference to them disappeared. Procedures
|check_mem| and |search_mem| are available to help diagnose such
problems. These procedures make use of two arrays called |free| and
|was_free| that are present only if \MF's debugging routines have
been included. (You may want to decrease the size of |mem| while you
@^debugging@>
are debugging.)

@<Glob...@>=
@!debug @!free: packed array [mem_min..mem_max] of boolean; {free cells}
@t\hskip1em@>@!was_free: packed array [mem_min..mem_max] of boolean;
  {previously free cells}
@t\hskip1em@>@!was_mem_end,@!was_lo_max,@!was_hi_min: pointer;
  {previous |mem_end|, |lo_mem_max|, and |hi_mem_min|}
@t\hskip1em@>@!panicking:boolean; {do we want to check memory constantly?}
gubed

@ @<Set initial...@>=
@!debug was_mem_end:=mem_min; {indicate that everything was previously free}
was_lo_max:=mem_min; was_hi_min:=mem_max;
panicking:=false;
gubed

@ Procedure |check_mem| makes sure that the available space lists of
|mem| are well formed, and it optionally prints out all locations
that are reserved now but were free the last time this procedure was called.

@p @!debug procedure check_mem(@!print_locs : boolean);
label done1,done2; {loop exits}
var @!p,@!q,@!r:pointer; {current locations of interest in |mem|}
@!clobbered:boolean; {is something amiss?}
begin for p:=mem_min to lo_mem_max do free[p]:=false; {you can probably
  do this faster}
for p:=hi_mem_min to mem_end do free[p]:=false; {ditto}
@<Check single-word |avail| list@>;
@<Check variable-size |avail| list@>;
@<Check flags of unavailable nodes@>;
@<Check the list of linear dependencies@>;
if print_locs then @<Print newly busy locations@>;
for p:=mem_min to lo_mem_max do was_free[p]:=free[p];
for p:=hi_mem_min to mem_end do was_free[p]:=free[p];
  {|was_free:=free| might be faster}
was_mem_end:=mem_end; was_lo_max:=lo_mem_max; was_hi_min:=hi_mem_min;
end;
gubed

@ @<Check single-word...@>=
p:=avail; q:=null; clobbered:=false;
while p<>null do
  begin if (p>mem_end)or(p<hi_mem_min) then clobbered:=true
  else if free[p] then clobbered:=true;
  if clobbered then
    begin print_nl("AVAIL list clobbered at ");
@.AVAIL list clobbered...@>
    print_int(q); goto done1;
    end;
  free[p]:=true; q:=p; p:=link(q);
  end;
done1:

@ @<Check variable-size...@>=
p:=rover; q:=null; clobbered:=false;
repeat if (p>=lo_mem_max)or(p<mem_min) then clobbered:=true
  else if (rlink(p)>=lo_mem_max)or(rlink(p)<mem_min) then clobbered:=true
  else if  not(is_empty(p))or(node_size(p)<2)or@|
   (p+node_size(p)>lo_mem_max)or@| (llink(rlink(p))<>p) then clobbered:=true;
  if clobbered then
  begin print_nl("Double-AVAIL list clobbered at ");
@.Double-AVAIL list clobbered...@>
  print_int(q); goto done2;
  end;
for q:=p to p+node_size(p)-1 do {mark all locations free}
  begin if free[q] then
    begin print_nl("Doubly free location at ");
@.Doubly free location...@>
    print_int(q); goto done2;
    end;
  free[q]:=true;
  end;
q:=p; p:=rlink(p);
until p=rover;
done2:

@ @<Check flags...@>=
p:=mem_min;
while p<=lo_mem_max do {node |p| should not be empty}
  begin if is_empty(p) then
    begin print_nl("Bad flag at "); print_int(p);
@.Bad flag...@>
    end;
  while (p<=lo_mem_max) and not free[p] do incr(p);
  while (p<=lo_mem_max) and free[p] do incr(p);
  end

@ @<Print newly busy...@>=
begin print_nl("New busy locs:");
@.New busy locs@>
for p:=mem_min to lo_mem_max do
  if not free[p] and ((p>was_lo_max) or was_free[p]) then
    begin print_char(" "); print_int(p);
    end;
for p:=hi_mem_min to mem_end do
  if not free[p] and
   ((p<was_hi_min) or (p>was_mem_end) or was_free[p]) then
    begin print_char(" "); print_int(p);
    end;
end

@ The |search_mem| procedure attempts to answer the question ``Who points
to node~|p|?'' In doing so, it fetches |link| and |info| fields of |mem|
that might not be of type |two_halves|. Strictly speaking, this is
@^dirty \PASCAL@>
undefined in \PASCAL, and it can lead to ``false drops'' (words that seem to
point to |p| purely by coincidence). But for debugging purposes, we want
to rule out the places that do {\sl not\/} point to |p|, so a few false
drops are tolerable.

@p @!debug procedure search_mem(@!p:pointer); {look for pointers to |p|}
var @!q:integer; {current position being searched}
begin for q:=mem_min to lo_mem_max do
  begin if link(q)=p then
    begin print_nl("LINK("); print_int(q); print_char(")");
    end;
  if info(q)=p then
    begin print_nl("INFO("); print_int(q); print_char(")");
    end;
  end;
for q:=hi_mem_min to mem_end do
  begin if link(q)=p then
    begin print_nl("LINK("); print_int(q); print_char(")");
    end;
  if info(q)=p then
    begin print_nl("INFO("); print_int(q); print_char(")");
    end;
  end;
@<Search |eqtb| for equivalents equal to |p|@>;
end;
gubed

@* \[12] The command codes.
Before we can go much further, we need to define symbolic names for the internal
code numbers that represent the various commands obeyed by \MF. These codes
are somewhat arbitrary, but not completely so. For example,
some codes have been made adjacent so that |case| statements in the
program need not consider cases that are widely spaced, or so that |case|
statements can be replaced by |if| statements. A command can begin an
expression if and only if its code lies between |min_primary_command| and
|max_primary_command|, inclusive. The first token of a statement that doesn't
begin with an expression has a command code between |min_command| and
|max_statement_command|, inclusive. The ordering of the highest-numbered
commands (|comma<semicolon<end_group<stop|) is crucial for the parsing
and error-recovery methods of this program.

At any rate, here is the list, for future reference.

@d if_test=1 {conditional text (\&{if})}
@d fi_or_else=2 {delimiters for conditionals (\&{elseif}, \&{else}, \&{fi})}
@d input=3 {input a source file (\&{input}, \&{endinput})}
@d iteration=4 {iterate (\&{for}, \&{forsuffixes}, \&{forever}, \&{endfor})}
@d repeat_loop=5 {special command substituted for \&{endfor}}
@d exit_test=6 {premature exit from a loop (\&{exitif})}
@d relax=7 {do nothing (\.{\char`\\})}
@d scan_tokens=8 {put a string into the input buffer}
@d expand_after=9 {look ahead one token}
@d defined_macro=10 {a macro defined by the user}
@d min_command=defined_macro+1
@d display_command=11 {online graphic output (\&{display})}
@d save_command=12 {save a list of tokens (\&{save})}
@d interim_command=13 {save an internal quantity (\&{interim})}
@d let_command=14 {redefine a symbolic token (\&{let})}
@d new_internal=15 {define a new internal quantity (\&{newinternal})}
@d macro_def=16 {define a macro (\&{def}, \&{vardef}, etc.)}
@d ship_out_command=17 {output a character (\&{shipout})}
@d add_to_command=18 {add to edges (\&{addto})}
@d cull_command=19 {cull and normalize edges (\&{cull})}
@d tfm_command=20 {command for font metric info (\&{ligtable}, etc.)}
@d protection_command=21 {set protection flag (\&{outer}, \&{inner})}
@d show_command=22 {diagnostic output (\&{show}, \&{showvariable}, etc.)}
@d mode_command=23 {set interaction level (\&{batchmode}, etc.)}
@d random_seed=24 {initialize random number generator (\&{randomseed})}
@d message_command=25 {communicate to user (\&{message}, \&{errmessage})}
@d every_job_command=26 {designate a starting token (\&{everyjob})}
@d delimiters=27 {define a pair of delimiters (\&{delimiters})}
@d open_window=28 {define a window on the screen (\&{openwindow})}
@d special_command=29 {output special info (\&{special}, \&{numspecial})}
@d type_name=30 {declare a type (\&{numeric}, \&{pair}, etc.)}
@d max_statement_command=type_name
@d min_primary_command=type_name
@d left_delimiter=31 {the left delimiter of a matching pair}
@d begin_group=32 {beginning of a group (\&{begingroup})}
@d nullary=33 {an operator without arguments (e.g., \&{normaldeviate})}
@d unary=34 {an operator with one argument (e.g., \&{sqrt})}
@d str_op=35 {convert a suffix to a string (\&{str})}
@d cycle=36 {close a cyclic path (\&{cycle})}
@d primary_binary=37 {binary operation taking `\&{of}' (e.g., \&{point})}
@d capsule_token=38 {a value that has been put into a token list}
@d string_token=39 {a string constant (e.g., |"hello"|)}
@d internal_quantity=40 {internal numeric parameter (e.g., \&{pausing})}
@d min_suffix_token=internal_quantity
@d tag_token=41 {a symbolic token without a primitive meaning}
@d numeric_token=42 {a numeric constant (e.g., \.{3.14159})}
@d max_suffix_token=numeric_token
@d plus_or_minus=43 {either `\.+' or `\.-'}
@d max_primary_command=plus_or_minus {should also be |numeric_token+1|}
@d min_tertiary_command=plus_or_minus
@d tertiary_secondary_macro=44 {a macro defined by \&{secondarydef}}
@d tertiary_binary=45 {an operator at the tertiary level (e.g., `\.{++}')}
@d max_tertiary_command=tertiary_binary
@d left_brace=46 {the operator `\.{\char`\{}'}
@d min_expression_command=left_brace
@d path_join=47 {the operator `\.{..}'}
@d ampersand=48 {the operator `\.\&'}
@d expression_tertiary_macro=49 {a macro defined by \&{tertiarydef}}
@d expression_binary=50 {an operator at the expression level (e.g., `\.<')}
@d equals=51 {the operator `\.='}
@d max_expression_command=equals
@d and_command=52 {the operator `\&{and}'}
@d min_secondary_command=and_command
@d secondary_primary_macro=53 {a macro defined by \&{primarydef}}
@d slash=54 {the operator `\./'}
@d secondary_binary=55 {an operator at the binary level (e.g., \&{shifted})}
@d max_secondary_command=secondary_binary
@d param_type=56 {type of parameter (\&{primary}, \&{expr}, \&{suffix}, etc.)}
@d controls=57 {specify control points explicitly (\&{controls})}
@d tension=58 {specify tension between knots (\&{tension})}
@d at_least=59 {bounded tension value (\&{atleast})}
@d curl_command=60 {specify curl at an end knot (\&{curl})}
@d macro_special=61 {special macro operators (\&{quote}, \.{\#\AT!}, etc.)}
@d right_delimiter=62 {the right delimiter of a matching pair}
@d left_bracket=63 {the operator `\.['}
@d right_bracket=64 {the operator `\.]'}
@d right_brace=65 {the operator `\.{\char`\}}'}
@d with_option=66 {option for filling (\&{withpen}, \&{withweight})}
@d cull_op=67 {the operator `\&{keeping}' or `\&{dropping}'}
@d thing_to_add=68
  {variant of \&{addto} (\&{contour}, \&{doublepath}, \&{also})}
@d of_token=69 {the operator `\&{of}'}
@d from_token=70 {the operator `\&{from}'}
@d to_token=71 {the operator `\&{to}'}
@d at_token=72 {the operator `\&{at}'}
@d in_window=73 {the operator `\&{inwindow}'}
@d step_token=74 {the operator `\&{step}'}
@d until_token=75 {the operator `\&{until}'}
@d lig_kern_token=76
  {the operators `\&{kern}' and `\.{=:}' and `\.{=:\char'174}', etc.}
@d assignment=77 {the operator `\.{:=}'}
@d skip_to=78 {the operation `\&{skipto}'}
@d bchar_label=79 {the operator `\.{\char'174\char'174:}'}
@d double_colon=80 {the operator `\.{::}'}
@d colon=81 {the operator `\.:'}
@#
@d comma=82 {the operator `\.,', must be |colon+1|}
@d end_of_statement==cur_cmd>comma
@d semicolon=83 {the operator `\.;', must be |comma+1|}
@d end_group=84 {end a group (\&{endgroup}), must be |semicolon+1|}
@d stop=85 {end a job (\&{end}, \&{dump}), must be |end_group+1|}
@d max_command_code=stop
@d outer_tag=max_command_code+1 {protection code added to command code}

@<Types...@>=
@!command_code=1..max_command_code;

@ Variables and capsules in \MF\ have a variety of ``types,''
distinguished by the following code numbers:

@d undefined=0 {no type has been declared}
@d unknown_tag=1 {this constant is added to certain type codes below}
@d vacuous=1 {no expression was present}
@d boolean_type=2 {\&{boolean} with a known value}
@d unknown_boolean=boolean_type+unknown_tag
@d string_type=4 {\&{string} with a known value}
@d unknown_string=string_type+unknown_tag
@d pen_type=6 {\&{pen} with a known value}
@d unknown_pen=pen_type+unknown_tag
@d future_pen=8 {subexpression that will become a \&{pen} at a higher level}
@d path_type=9 {\&{path} with a known value}
@d unknown_path=path_type+unknown_tag
@d picture_type=11 {\&{picture} with a known value}
@d unknown_picture=picture_type+unknown_tag
@d transform_type=13 {\&{transform} variable or capsule}
@d pair_type=14 {\&{pair} variable or capsule}
@d numeric_type=15 {variable that has been declared \&{numeric} but not used}
@d known=16 {\&{numeric} with a known value}
@d dependent=17 {a linear combination with |fraction| coefficients}
@d proto_dependent=18 {a linear combination with |scaled| coefficients}
@d independent=19 {\&{numeric} with unknown value}
@d token_list=20 {variable name or suffix argument or text argument}
@d structured=21 {variable with subscripts and attributes}
@d unsuffixed_macro=22 {variable defined with \&{vardef} but no \.{\AT!\#}}
@d suffixed_macro=23 {variable defined with \&{vardef} and \.{\AT!\#}}
@#
@d unknown_types==unknown_boolean,unknown_string,
  unknown_pen,unknown_picture,unknown_path

@<Basic printing procedures@>=
procedure print_type(@!t:small_number);
begin case t of
vacuous:print("vacuous");
boolean_type:print("boolean");
unknown_boolean:print("unknown boolean");
string_type:print("string");
unknown_string:print("unknown string");
pen_type:print("pen");
unknown_pen:print("unknown pen");
future_pen:print("future pen");
path_type:print("path");
unknown_path:print("unknown path");
picture_type:print("picture");
unknown_picture:print("unknown picture");
transform_type:print("transform");
pair_type:print("pair");
known:print("known numeric");
dependent:print("dependent");
proto_dependent:print("proto-dependent");
numeric_type:print("numeric");
independent:print("independent");
token_list:print("token list");
structured:print("structured");
unsuffixed_macro:print("unsuffixed macro");
suffixed_macro:print("suffixed macro");
othercases print("undefined")
endcases;
end;

@ Values inside \MF\ are stored in two-word nodes that have a |name_type|
as well as a |type|. The possibilities for |name_type| are defined
here; they will be explained in more detail later.

@d root=0 {|name_type| at the top level of a variable}
@d saved_root=1 {same, when the variable has been saved}
@d structured_root=2 {|name_type| where a |structured| branch occurs}
@d subscr=3 {|name_type| in a subscript node}
@d attr=4 {|name_type| in an attribute node}
@d x_part_sector=5 {|name_type| in the \&{xpart} of a node}
@d y_part_sector=6 {|name_type| in the \&{ypart} of a node}
@d xx_part_sector=7 {|name_type| in the \&{xxpart} of a node}
@d xy_part_sector=8 {|name_type| in the \&{xypart} of a node}
@d yx_part_sector=9 {|name_type| in the \&{yxpart} of a node}
@d yy_part_sector=10 {|name_type| in the \&{yypart} of a node}
@d capsule=11 {|name_type| in stashed-away subexpressions}
@d token=12 {|name_type| in a numeric token or string token}

@ Primitive operations that produce values have a secondary identification
code in addition to their command code; it's something like genera and species.
For example, `\.*' has the command code |primary_binary|, and its
secondary identification is |times|. The secondary codes start at 30 so that
they don't overlap with the type codes; some type codes (e.g., |string_type|)
are used as operators as well as type identifications.

@d true_code=30 {operation code for \.{true}}
@d false_code=31 {operation code for \.{false}}
@d null_picture_code=32 {operation code for \.{nullpicture}}
@d null_pen_code=33 {operation code for \.{nullpen}}
@d job_name_op=34 {operation code for \.{jobname}}
@d read_string_op=35 {operation code for \.{readstring}}
@d pen_circle=36 {operation code for \.{pencircle}}
@d normal_deviate=37 {operation code for \.{normaldeviate}}
@d odd_op=38 {operation code for \.{odd}}
@d known_op=39 {operation code for \.{known}}
@d unknown_op=40 {operation code for \.{unknown}}
@d not_op=41 {operation code for \.{not}}
@d decimal=42 {operation code for \.{decimal}}
@d reverse=43 {operation code for \.{reverse}}
@d make_path_op=44 {operation code for \.{makepath}}
@d make_pen_op=45 {operation code for \.{makepen}}
@d total_weight_op=46 {operation code for \.{totalweight}}
@d oct_op=47 {operation code for \.{oct}}
@d hex_op=48 {operation code for \.{hex}}
@d ASCII_op=49 {operation code for \.{ASCII}}
@d char_op=50 {operation code for \.{char}}
@d length_op=51 {operation code for \.{length}}
@d turning_op=52 {operation code for \.{turningnumber}}
@d x_part=53 {operation code for \.{xpart}}
@d y_part=54 {operation code for \.{ypart}}
@d xx_part=55 {operation code for \.{xxpart}}
@d xy_part=56 {operation code for \.{xypart}}
@d yx_part=57 {operation code for \.{yxpart}}
@d yy_part=58 {operation code for \.{yypart}}
@d sqrt_op=59 {operation code for \.{sqrt}}
@d m_exp_op=60 {operation code for \.{mexp}}
@d m_log_op=61 {operation code for \.{mlog}}
@d sin_d_op=62 {operation code for \.{sind}}
@d cos_d_op=63 {operation code for \.{cosd}}
@d floor_op=64 {operation code for \.{floor}}
@d uniform_deviate=65 {operation code for \.{uniformdeviate}}
@d char_exists_op=66 {operation code for \.{charexists}}
@d angle_op=67 {operation code for \.{angle}}
@d cycle_op=68 {operation code for \.{cycle}}
@d plus=69 {operation code for \.+}
@d minus=70 {operation code for \.-}
@d times=71 {operation code for \.*}
@d over=72 {operation code for \./}
@d pythag_add=73 {operation code for \.{++}}
@d pythag_sub=74 {operation code for \.{+-+}}
@d or_op=75 {operation code for \.{or}}
@d and_op=76 {operation code for \.{and}}
@d less_than=77 {operation code for \.<}
@d less_or_equal=78 {operation code for \.{<=}}
@d greater_than=79 {operation code for \.>}
@d greater_or_equal=80 {operation code for \.{>=}}
@d equal_to=81 {operation code for \.=}
@d unequal_to=82 {operation code for \.{<>}}
@d concatenate=83 {operation code for \.\&}
@d rotated_by=84 {operation code for \.{rotated}}
@d slanted_by=85 {operation code for \.{slanted}}
@d scaled_by=86 {operation code for \.{scaled}}
@d shifted_by=87 {operation code for \.{shifted}}
@d transformed_by=88 {operation code for \.{transformed}}
@d x_scaled=89 {operation code for \.{xscaled}}
@d y_scaled=90 {operation code for \.{yscaled}}
@d z_scaled=91 {operation code for \.{zscaled}}
@d intersect=92 {operation code for \.{intersectiontimes}}
@d double_dot=93 {operation code for improper \.{..}}
@d substring_of=94 {operation code for \.{substring}}
@d min_of=substring_of
@d subpath_of=95 {operation code for \.{subpath}}
@d direction_time_of=96 {operation code for \.{directiontime}}
@d point_of=97 {operation code for \.{point}}
@d precontrol_of=98 {operation code for \.{precontrol}}
@d postcontrol_of=99 {operation code for \.{postcontrol}}
@d pen_offset_of=100 {operation code for \.{penoffset}}

@p procedure print_op(@!c:quarterword);
begin if c<=numeric_type then print_type(c)
else case c of
true_code:print("true");
false_code:print("false");
null_picture_code:print("nullpicture");
null_pen_code:print("nullpen");
job_name_op:print("jobname");
read_string_op:print("readstring");
pen_circle:print("pencircle");
normal_deviate:print("normaldeviate");
odd_op:print("odd");
known_op:print("known");
unknown_op:print("unknown");
not_op:print("not");
decimal:print("decimal");
reverse:print("reverse");
make_path_op:print("makepath");
make_pen_op:print("makepen");
total_weight_op:print("totalweight");
oct_op:print("oct");
hex_op:print("hex");
ASCII_op:print("ASCII");
char_op:print("char");
length_op:print("length");
turning_op:print("turningnumber");
x_part:print("xpart");
y_part:print("ypart");
xx_part:print("xxpart");
xy_part:print("xypart");
yx_part:print("yxpart");
yy_part:print("yypart");
sqrt_op:print("sqrt");
m_exp_op:print("mexp");
m_log_op:print("mlog");
sin_d_op:print("sind");
cos_d_op:print("cosd");
floor_op:print("floor");
uniform_deviate:print("uniformdeviate");
char_exists_op:print("charexists");
angle_op:print("angle");
cycle_op:print("cycle");
plus:print_char("+");
minus:print_char("-");
times:print_char("*");
over:print_char("/");
pythag_add:print("++");
pythag_sub:print("+-+");
or_op:print("or");
and_op:print("and");
less_than:print_char("<");
less_or_equal:print("<=");
greater_than:print_char(">");
greater_or_equal:print(">=");
equal_to:print_char("=");
unequal_to:print("<>");
concatenate:print("&");
rotated_by:print("rotated");
slanted_by:print("slanted");
scaled_by:print("scaled");
shifted_by:print("shifted");
transformed_by:print("transformed");
x_scaled:print("xscaled");
y_scaled:print("yscaled");
z_scaled:print("zscaled");
intersect:print("intersectiontimes");
substring_of:print("substring");
subpath_of:print("subpath");
direction_time_of:print("directiontime");
point_of:print("point");
precontrol_of:print("precontrol");
postcontrol_of:print("postcontrol");
pen_offset_of:print("penoffset");
othercases print("..")
endcases;
end;

@ \MF\ also has a bunch of internal parameters that a user might want to
fuss with. Every such parameter has an identifying code number, defined here.

@d tracing_titles=1 {show titles online when they appear}
@d tracing_equations=2 {show each variable when it becomes known}
@d tracing_capsules=3 {show capsules too}
@d tracing_choices=4 {show the control points chosen for paths}
@d tracing_specs=5 {show subdivision of paths into octants before digitizing}
@d tracing_pens=6 {show details of pens that are made}
@d tracing_commands=7 {show commands and operations before they are performed}
@d tracing_restores=8 {show when a variable or internal is restored}
@d tracing_macros=9 {show macros before they are expanded}
@d tracing_edges=10 {show digitized edges as they are computed}
@d tracing_output=11 {show digitized edges as they are output}
@d tracing_stats=12 {show memory usage at end of job}
@d tracing_online=13 {show long diagnostics on terminal and in the log file}
@d year=14 {the current year (e.g., 1984)}
@d month=15 {the current month (e.g., 3 $\equiv$ March)}
@d day=16 {the current day of the month}
@d time=17 {the number of minutes past midnight when this job started}
@d char_code=18 {the number of the next character to be output}
@d char_ext=19 {the extension code of the next character to be output}
@d char_wd=20 {the width of the next character to be output}
@d char_ht=21 {the height of the next character to be output}
@d char_dp=22 {the depth of the next character to be output}
@d char_ic=23 {the italic correction of the next character to be output}
@d char_dx=24 {the device's $x$ movement for the next character, in pixels}
@d char_dy=25 {the device's $y$ movement for the next character, in pixels}
@d design_size=26 {the unit of measure used for |char_wd..char_ic|, in points}
@d hppp=27 {the number of horizontal pixels per point}
@d vppp=28 {the number of vertical pixels per point}
@d x_offset=29 {horizontal displacement of shipped-out characters}
@d y_offset=30 {vertical displacement of shipped-out characters}
@d pausing=31 {positive to display lines on the terminal before they are read}
@d showstopping=32 {positive to stop after each \&{show} command}
@d fontmaking=33 {positive if font metric output is to be produced}
@d proofing=34 {positive for proof mode, negative to suppress output}
@d smoothing=35 {positive if moves are to be ``smoothed''}
@d autorounding=36 {controls path modification to ``good'' points}
@d granularity=37 {autorounding uses this pixel size}
@d fillin=38 {extra darkness of diagonal lines}
@d turning_check=39 {controls reorientation of clockwise paths}
@d warning_check=40 {controls error message when variable value is large}
@d boundary_char=41 {the boundary character for ligatures}
@d max_given_internal=41

@<Glob...@>=
@!internal:array[1..max_internal] of scaled;
  {the values of internal quantities}
@!int_name:array[1..max_internal] of str_number;
  {their names}
@!int_ptr:max_given_internal..max_internal;
  {the maximum internal quantity defined so far}

@ @<Set init...@>=
for k:=1 to max_given_internal do internal[k]:=0;
int_ptr:=max_given_internal;

@ The symbolic names for internal quantities are put into \MF's hash table
by using a routine called |primitive|, which will be defined later. Let us
enter them now, so that we don't have to list all those names again
anywhere else.

@<Put each of \MF's primitives into the hash table@>=
primitive("tracingtitles",internal_quantity,tracing_titles);@/
@!@:tracingtitles_}{\&{tracingtitles} primitive@>
primitive("tracingequations",internal_quantity,tracing_equations);@/
@!@:tracing_equations_}{\&{tracingequations} primitive@>
primitive("tracingcapsules",internal_quantity,tracing_capsules);@/
@!@:tracing_capsules_}{\&{tracingcapsules} primitive@>
primitive("tracingchoices",internal_quantity,tracing_choices);@/
@!@:tracing_choices_}{\&{tracingchoices} primitive@>
primitive("tracingspecs",internal_quantity,tracing_specs);@/
@!@:tracing_specs_}{\&{tracingspecs} primitive@>
primitive("tracingpens",internal_quantity,tracing_pens);@/
@!@:tracing_pens_}{\&{tracingpens} primitive@>
primitive("tracingcommands",internal_quantity,tracing_commands);@/
@!@:tracing_commands_}{\&{tracingcommands} primitive@>
primitive("tracingrestores",internal_quantity,tracing_restores);@/
@!@:tracing_restores_}{\&{tracingrestores} primitive@>
primitive("tracingmacros",internal_quantity,tracing_macros);@/
@!@:tracing_macros_}{\&{tracingmacros} primitive@>
primitive("tracingedges",internal_quantity,tracing_edges);@/
@!@:tracing_edges_}{\&{tracingedges} primitive@>
primitive("tracingoutput",internal_quantity,tracing_output);@/
@!@:tracing_output_}{\&{tracingoutput} primitive@>
primitive("tracingstats",internal_quantity,tracing_stats);@/
@!@:tracing_stats_}{\&{tracingstats} primitive@>
primitive("tracingonline",internal_quantity,tracing_online);@/
@!@:tracing_online_}{\&{tracingonline} primitive@>
primitive("year",internal_quantity,year);@/
@!@:year_}{\&{year} primitive@>
primitive("month",internal_quantity,month);@/
@!@:month_}{\&{month} primitive@>
primitive("day",internal_quantity,day);@/
@!@:day_}{\&{day} primitive@>
primitive("time",internal_quantity,time);@/
@!@:time_}{\&{time} primitive@>
primitive("charcode",internal_quantity,char_code);@/
@!@:char_code_}{\&{charcode} primitive@>
primitive("charext",internal_quantity,char_ext);@/
@!@:char_ext_}{\&{charext} primitive@>
primitive("charwd",internal_quantity,char_wd);@/
@!@:char_wd_}{\&{charwd} primitive@>
primitive("charht",internal_quantity,char_ht);@/
@!@:char_ht_}{\&{charht} primitive@>
primitive("chardp",internal_quantity,char_dp);@/
@!@:char_dp_}{\&{chardp} primitive@>
primitive("charic",internal_quantity,char_ic);@/
@!@:char_ic_}{\&{charic} primitive@>
primitive("chardx",internal_quantity,char_dx);@/
@!@:char_dx_}{\&{chardx} primitive@>
primitive("chardy",internal_quantity,char_dy);@/
@!@:char_dy_}{\&{chardy} primitive@>
primitive("designsize",internal_quantity,design_size);@/
@!@:design_size_}{\&{designsize} primitive@>
primitive("hppp",internal_quantity,hppp);@/
@!@:hppp_}{\&{hppp} primitive@>
primitive("vppp",internal_quantity,vppp);@/
@!@:vppp_}{\&{vppp} primitive@>
primitive("xoffset",internal_quantity,x_offset);@/
@!@:x_offset_}{\&{xoffset} primitive@>
primitive("yoffset",internal_quantity,y_offset);@/
@!@:y_offset_}{\&{yoffset} primitive@>
primitive("pausing",internal_quantity,pausing);@/
@!@:pausing_}{\&{pausing} primitive@>
primitive("showstopping",internal_quantity,showstopping);@/
@!@:showstopping_}{\&{showstopping} primitive@>
primitive("fontmaking",internal_quantity,fontmaking);@/
@!@:fontmaking_}{\&{fontmaking} primitive@>
primitive("proofing",internal_quantity,proofing);@/
@!@:proofing_}{\&{proofing} primitive@>
primitive("smoothing",internal_quantity,smoothing);@/
@!@:smoothing_}{\&{smoothing} primitive@>
primitive("autorounding",internal_quantity,autorounding);@/
@!@:autorounding_}{\&{autorounding} primitive@>
primitive("granularity",internal_quantity,granularity);@/
@!@:granularity_}{\&{granularity} primitive@>
primitive("fillin",internal_quantity,fillin);@/
@!@:fillin_}{\&{fillin} primitive@>
primitive("turningcheck",internal_quantity,turning_check);@/
@!@:turning_check_}{\&{turningcheck} primitive@>
primitive("warningcheck",internal_quantity,warning_check);@/
@!@:warning_check_}{\&{warningcheck} primitive@>
primitive("boundarychar",internal_quantity,boundary_char);@/
@!@:boundary_char_}{\&{boundarychar} primitive@>

@ Well, we do have to list the names one more time, for use in symbolic
printouts.

@<Initialize table...@>=
int_name[tracing_titles]:="tracingtitles";
int_name[tracing_equations]:="tracingequations";
int_name[tracing_capsules]:="tracingcapsules";
int_name[tracing_choices]:="tracingchoices";
int_name[tracing_specs]:="tracingspecs";
int_name[tracing_pens]:="tracingpens";
int_name[tracing_commands]:="tracingcommands";
int_name[tracing_restores]:="tracingrestores";
int_name[tracing_macros]:="tracingmacros";
int_name[tracing_edges]:="tracingedges";
int_name[tracing_output]:="tracingoutput";
int_name[tracing_stats]:="tracingstats";
int_name[tracing_online]:="tracingonline";
int_name[year]:="year";
int_name[month]:="month";
int_name[day]:="day";
int_name[time]:="time";
int_name[char_code]:="charcode";
int_name[char_ext]:="charext";
int_name[char_wd]:="charwd";
int_name[char_ht]:="charht";
int_name[char_dp]:="chardp";
int_name[char_ic]:="charic";
int_name[char_dx]:="chardx";
int_name[char_dy]:="chardy";
int_name[design_size]:="designsize";
int_name[hppp]:="hppp";
int_name[vppp]:="vppp";
int_name[x_offset]:="xoffset";
int_name[y_offset]:="yoffset";
int_name[pausing]:="pausing";
int_name[showstopping]:="showstopping";
int_name[fontmaking]:="fontmaking";
int_name[proofing]:="proofing";
int_name[smoothing]:="smoothing";
int_name[autorounding]:="autorounding";
int_name[granularity]:="granularity";
int_name[fillin]:="fillin";
int_name[turning_check]:="turningcheck";
int_name[warning_check]:="warningcheck";
int_name[boundary_char]:="boundarychar";

@ The following procedure, which is called just before \MF\ initializes its
input and output, establishes the initial values of the date and time.
@^system dependencies@>
Since standard \PASCAL\ cannot provide such information, something special
is needed. The program here simply assumes that suitable values appear in
the global variables \\{sys\_time}, \\{sys\_day}, \\{sys\_month}, and
\\{sys\_year} (which are initialized to noon on 4 July 1776,
in case the implementor is careless).

Note that the values are |scaled| integers. Hence \MF\ can no longer
be used after the year 32767.

@p procedure fix_date_and_time;
begin sys_time:=12*60;
sys_day:=4; sys_month:=7; sys_year:=1776;  {self-evident truths}
internal[time]:=sys_time*unity; {minutes since midnight}
internal[day]:=sys_day*unity; {day of the month}
internal[month]:=sys_month*unity; {month of the year}
internal[year]:=sys_year*unity; {Anno Domini}
end;

@ \MF\ is occasionally supposed to print diagnostic information that
goes only into the transcript file, unless |tracing_online| is positive.
Now that we have defined |tracing_online| we can define
two routines that adjust the destination of print commands:

@<Basic printing...@>=
procedure begin_diagnostic; {prepare to do some tracing}
begin old_setting:=selector;
if(internal[tracing_online]<=0)and(selector=term_and_log) then
  begin decr(selector);
  if history=spotless then history:=warning_issued;
  end;
end;
@#
procedure end_diagnostic(@!blank_line:boolean);
  {restore proper conditions after tracing}
begin print_nl("");
if blank_line then print_ln;
selector:=old_setting;
end;

@ Of course we had better declare a few more global variables, if the previous
routines are going to work.

@<Glob...@>=
@!old_setting:0..max_selector;
@!sys_time,@!sys_day,@!sys_month,@!sys_year:integer;
    {date and time supplied by external system}

@ We will occasionally use |begin_diagnostic| in connection with line-number
printing, as follows. (The parameter |s| is typically |"Path"| or
|"Cycle spec"|, etc.)

@<Basic printing...@>=
procedure print_diagnostic(@!s,@!t:str_number;@!nuline:boolean);
begin begin_diagnostic;
if nuline then print_nl(s)@+else print(s);
print(" at line "); print_int(line);
print(t); print_char(":");
end;

@ The 256 |ASCII_code| characters are grouped into classes by means of
the |char_class| table. Individual class numbers have no semantic
or syntactic significance, except in a few instances defined here.
There's also |max_class|, which can be used as a basis for additional
class numbers in nonstandard extensions of \MF.

@d digit_class=0 {the class number of \.{0123456789}}
@d period_class=1 {the class number of `\..'}
@d space_class=2 {the class number of spaces and nonstandard characters}
@d percent_class=3 {the class number of `\.\%'}
@d string_class=4 {the class number of `\."'}
@d right_paren_class=8 {the class number of `\.)'}
@d isolated_classes==5,6,7,8 {characters that make length-one tokens only}
@d letter_class=9 {letters and the underline character}
@d left_bracket_class=17 {`\.['}
@d right_bracket_class=18 {`\.]'}
@d invalid_class=20 {bad character in the input}
@d max_class=20 {the largest class number}

@<Glob...@>=
@!char_class:array[ASCII_code] of 0..max_class; {the class numbers}

@ If changes are made to accommodate non-ASCII character sets, they should
follow the guidelines in Appendix~C of {\sl The {\logos METAFONT\/}book}.
@:METAFONTbook}{\sl The {\logos METAFONT\/}book@>
@^system dependencies@>

@<Set init...@>=
for k:="0" to "9" do char_class[k]:=digit_class;
char_class["."]:=period_class;
char_class[" "]:=space_class;
char_class["%"]:=percent_class;
char_class[""""]:=string_class;@/
char_class[","]:=5;
char_class[";"]:=6;
char_class["("]:=7;
char_class[")"]:=right_paren_class;
for k:="A" to "Z" do char_class[k]:=letter_class;
for k:="a" to "z" do char_class[k]:=letter_class;
char_class["_"]:=letter_class;@/
char_class["<"]:=10;
char_class["="]:=10;
char_class[">"]:=10;
char_class[":"]:=10;
char_class["|"]:=10;@/
char_class["`"]:=11;
char_class["'"]:=11;@/
char_class["+"]:=12;
char_class["-"]:=12;@/
char_class["/"]:=13;
char_class["*"]:=13;
char_class["\"]:=13;@/
char_class["!"]:=14;
char_class["?"]:=14;@/
char_class["#"]:=15;
char_class["&"]:=15;
char_class["@@"]:=15;
char_class["$"]:=15;@/
char_class["^"]:=16;
char_class["~"]:=16;@/
char_class["["]:=left_bracket_class;
char_class["]"]:=right_bracket_class;@/
char_class["{"]:=19;
char_class["}"]:=19;@/
for k:=0 to " "-1 do char_class[k]:=invalid_class;
for k:=127 to 255 do char_class[k]:=invalid_class;

@* \[13] The hash table.
Symbolic tokens are stored and retrieved by means of a fairly standard hash
table algorithm called the method of ``coalescing lists'' (cf.\ Algorithm 6.4C
in {\sl The Art of Computer Programming\/}). Once a symbolic token enters the
table, it is never removed.

The actual sequence of characters forming a symbolic token is
stored in the |str_pool| array together with all the other strings. An
auxiliary array |hash| consists of items with two halfword fields per
word. The first of these, called |next(p)|, points to the next identifier
belonging to the same coalesced list as the identifier corresponding to~|p|;
and the other, called |text(p)|, points to the |str_start| entry for
|p|'s identifier. If position~|p| of the hash table is empty, we have
|text(p)=0|; if position |p| is either empty or the end of a coalesced
hash list, we have |next(p)=0|.

An auxiliary pointer variable called |hash_used| is maintained in such a
way that all locations |p>=hash_used| are nonempty. The global variable
|st_count| tells how many symbolic tokens have been defined, if statistics
are being kept.

The first 256 locations of |hash| are reserved for symbols of length one.

There's a parallel array called |eqtb| that contains the current equivalent
values of each symbolic token. The entries of this array consist of
two halfwords called |eq_type| (a command code) and |equiv| (a secondary
piece of information that qualifies the |eq_type|).

@d next(#) == hash[#].lh {link for coalesced lists}
@d text(#) == hash[#].rh {string number for symbolic token name}
@d eq_type(#) == eqtb[#].lh {the current ``meaning'' of a symbolic token}
@d equiv(#) == eqtb[#].rh {parametric part of a token's meaning}
@d hash_base=257 {hashing actually starts here}
@d hash_is_full == (hash_used=hash_base) {are all positions occupied?}

@<Glob...@>=
@!hash_used:pointer; {allocation pointer for |hash|}
@!st_count:integer; {total number of known identifiers}

@ Certain entries in the hash table are ``frozen'' and not redefinable,
since they are used in error recovery.

@d hash_top==hash_base+hash_size {the first location of the frozen area}
@d frozen_inaccessible==hash_top {|hash| location to protect the frozen area}
@d frozen_repeat_loop==hash_top+1 {|hash| location of a loop-repeat token}
@d frozen_right_delimiter==hash_top+2 {|hash| location of a permanent `\.)'}
@d frozen_left_bracket==hash_top+3 {|hash| location of a permanent `\.['}
@d frozen_slash==hash_top+4 {|hash| location of a permanent `\./'}
@d frozen_colon==hash_top+5 {|hash| location of a permanent `\.:'}
@d frozen_semicolon==hash_top+6 {|hash| location of a permanent `\.;'}
@d frozen_end_for==hash_top+7 {|hash| location of a permanent \&{endfor}}
@d frozen_end_def==hash_top+8 {|hash| location of a permanent \&{enddef}}
@d frozen_fi==hash_top+9 {|hash| location of a permanent \&{fi}}
@d frozen_end_group==hash_top+10
  {|hash| location of a permanent `\.{endgroup}'}
@d frozen_bad_vardef==hash_top+11 {|hash| location of `\.{a bad variable}'}
@d frozen_undefined==hash_top+12 {|hash| location that never gets defined}
@d hash_end==hash_top+12 {the actual size of the |hash| and |eqtb| arrays}

@<Glob...@>=
@!hash: array[1..hash_end] of two_halves; {the hash table}
@!eqtb: array[1..hash_end] of two_halves; {the equivalents}

@ @<Set init...@>=
next(1):=0; text(1):=0; eq_type(1):=tag_token; equiv(1):=null;
for k:=2 to hash_end do
  begin hash[k]:=hash[1]; eqtb[k]:=eqtb[1];
  end;

@ @<Initialize table entries...@>=
hash_used:=frozen_inaccessible; {nothing is used}
st_count:=0;@/
text(frozen_bad_vardef):="a bad variable";
text(frozen_fi):="fi";
text(frozen_end_group):="endgroup";
text(frozen_end_def):="enddef";
text(frozen_end_for):="endfor";@/
text(frozen_semicolon):=";";
text(frozen_colon):=":";
text(frozen_slash):="/";
text(frozen_left_bracket):="[";
text(frozen_right_delimiter):=")";@/
text(frozen_inaccessible):=" INACCESSIBLE";@/
eq_type(frozen_right_delimiter):=right_delimiter;

@ @<Check the ``constant'' values...@>=
if hash_end+max_internal>max_halfword then bad:=21;

@ Here is the subroutine that searches the hash table for an identifier
that matches a given string of length~|l| appearing in |buffer[j..
(j+l-1)]|. If the identifier is not found, it is inserted; hence it
will always be found, and the corresponding hash table address
will be returned.

@p function id_lookup(@!j,@!l:integer):pointer; {search the hash table}
label found; {go here when you've found it}
var @!h:integer; {hash code}
@!p:pointer; {index in |hash| array}
@!k:pointer; {index in |buffer| array}
begin if l=1 then @<Treat special case of length 1 and |goto found|@>;
@<Compute the hash code |h|@>;
p:=h+hash_base; {we start searching here; note that |0<=h<hash_prime|}
loop@+  begin if text(p)>0 then if length(text(p))=l then
    if str_eq_buf(text(p),j) then goto found;
  if next(p)=0 then
    @<Insert a new symbolic token after |p|, then
      make |p| point to it and |goto found|@>;
  p:=next(p);
  end;
found: id_lookup:=p;
end;

@ @<Treat special case of length 1...@>=
begin p:=buffer[j]+1; text(p):=p-1; goto found;
end

@ @<Insert a new symbolic...@>=
begin if text(p)>0 then
  begin repeat if hash_is_full then
    overflow("hash size",hash_size);
@:METAFONT capacity exceeded hash size}{\quad hash size@>
  decr(hash_used);
  until text(hash_used)=0; {search for an empty location in |hash|}
  next(p):=hash_used; p:=hash_used;
  end;
str_room(l);
for k:=j to j+l-1 do append_char(buffer[k]);
text(p):=make_string; str_ref[text(p)]:=max_str_ref;
@!stat incr(st_count);@+tats@;@/
goto found;
end

@ The value of |hash_prime| should be roughly 85\pct! of |hash_size|, and it
should be a prime number.  The theory of hashing tells us to expect fewer
than two table probes, on the average, when the search is successful.
[See J.~S. Vitter, {\sl Journal of the ACM\/ \bf30} (1983), 231--258.]
@^Vitter, Jeffrey Scott@>

@<Compute the hash code |h|@>=
h:=buffer[j];
for k:=j+1 to j+l-1 do
  begin h:=h+h+buffer[k];
  while h>=hash_prime do h:=h-hash_prime;
  end

@ @<Search |eqtb| for equivalents equal to |p|@>=
for q:=1 to hash_end do
  begin if equiv(q)=p then
    begin print_nl("EQUIV("); print_int(q); print_char(")");
    end;
  end

@ We need to put \MF's ``primitive'' symbolic tokens into the hash
table, together with their command code (which will be the |eq_type|)
and an operand (which will be the |equiv|). The |primitive| procedure
does this, in a way that no \MF\ user can. The global value |cur_sym|
contains the new |eqtb| pointer after |primitive| has acted.

@p @!init procedure primitive(@!s:str_number;@!c:halfword;@!o:halfword);
var @!k:pool_pointer; {index into |str_pool|}
@!j:small_number; {index into |buffer|}
@!l:small_number; {length of the string}
begin k:=str_start[s]; l:=str_start[s+1]-k;
  {we will move |s| into the (empty) |buffer|}
for j:=0 to l-1 do buffer[j]:=so(str_pool[k+j]);
cur_sym:=id_lookup(0,l);@/
if s>=256 then {we don't want to have the string twice}
  begin flush_string(str_ptr-1); text(cur_sym):=s;
  end;
eq_type(cur_sym):=c; equiv(cur_sym):=o;
end;
tini

@ Many of \MF's primitives need no |equiv|, since they are identifiable
by their |eq_type| alone. These primitives are loaded into the hash table
as follows:

@<Put each of \MF's primitives into the hash table@>=
primitive("..",path_join,0);@/
@!@:.._}{\.{..} primitive@>
primitive("[",left_bracket,0); eqtb[frozen_left_bracket]:=eqtb[cur_sym];@/
@!@:[ }{\.{[} primitive@>
primitive("]",right_bracket,0);@/
@!@:] }{\.{]} primitive@>
primitive("}",right_brace,0);@/
@!@:]]}{\.{\char`\}} primitive@>
primitive("{",left_brace,0);@/
@!@:][}{\.{\char`\{} primitive@>
primitive(":",colon,0); eqtb[frozen_colon]:=eqtb[cur_sym];@/
@!@:: }{\.{:} primitive@>
primitive("::",double_colon,0);@/
@!@::: }{\.{::} primitive@>
primitive("||:",bchar_label,0);@/
@!@:::: }{\.{\char'174\char'174:} primitive@>
primitive(":=",assignment,0);@/
@!@::=_}{\.{:=} primitive@>
primitive(",",comma,0);@/
@!@:, }{\., primitive@>
primitive(";",semicolon,0); eqtb[frozen_semicolon]:=eqtb[cur_sym];@/
@!@:; }{\.; primitive@>
primitive("\",relax,0);@/
@!@:]]\\}{\.{\char`\\} primitive@>
@#
primitive("addto",add_to_command,0);@/
@!@:add_to_}{\&{addto} primitive@>
primitive("at",at_token,0);@/
@!@:at_}{\&{at} primitive@>
primitive("atleast",at_least,0);@/
@!@:at_least_}{\&{atleast} primitive@>
primitive("begingroup",begin_group,0); bg_loc:=cur_sym;@/
@!@:begin_group_}{\&{begingroup} primitive@>
primitive("controls",controls,0);@/
@!@:controls_}{\&{controls} primitive@>
primitive("cull",cull_command,0);@/
@!@:cull_}{\&{cull} primitive@>
primitive("curl",curl_command,0);@/
@!@:curl_}{\&{curl} primitive@>
primitive("delimiters",delimiters,0);@/
@!@:delimiters_}{\&{delimiters} primitive@>
primitive("display",display_command,0);@/
@!@:display_}{\&{display} primitive@>
primitive("endgroup",end_group,0);
 eqtb[frozen_end_group]:=eqtb[cur_sym]; eg_loc:=cur_sym;@/
@!@:endgroup_}{\&{endgroup} primitive@>
primitive("everyjob",every_job_command,0);@/
@!@:every_job_}{\&{everyjob} primitive@>
primitive("exitif",exit_test,0);@/
@!@:exit_if_}{\&{exitif} primitive@>
primitive("expandafter",expand_after,0);@/
@!@:expand_after_}{\&{expandafter} primitive@>
primitive("from",from_token,0);@/
@!@:from_}{\&{from} primitive@>
primitive("inwindow",in_window,0);@/
@!@:in_window_}{\&{inwindow} primitive@>
primitive("interim",interim_command,0);@/
@!@:interim_}{\&{interim} primitive@>
primitive("let",let_command,0);@/
@!@:let_}{\&{let} primitive@>
primitive("newinternal",new_internal,0);@/
@!@:new_internal_}{\&{newinternal} primitive@>
primitive("of",of_token,0);@/
@!@:of_}{\&{of} primitive@>
primitive("openwindow",open_window,0);@/
@!@:open_window_}{\&{openwindow} primitive@>
primitive("randomseed",random_seed,0);@/
@!@:random_seed_}{\&{randomseed} primitive@>
primitive("save",save_command,0);@/
@!@:save_}{\&{save} primitive@>
primitive("scantokens",scan_tokens,0);@/
@!@:scan_tokens_}{\&{scantokens} primitive@>
primitive("shipout",ship_out_command,0);@/
@!@:ship_out_}{\&{shipout} primitive@>
primitive("skipto",skip_to,0);@/
@!@:skip_to_}{\&{skipto} primitive@>
primitive("step",step_token,0);@/
@!@:step_}{\&{step} primitive@>
primitive("str",str_op,0);@/
@!@:str_}{\&{str} primitive@>
primitive("tension",tension,0);@/
@!@:tension_}{\&{tension} primitive@>
primitive("to",to_token,0);@/
@!@:to_}{\&{to} primitive@>
primitive("until",until_token,0);@/
@!@:until_}{\&{until} primitive@>

@ Each primitive has a corresponding inverse, so that it is possible to
display the cryptic numeric contents of |eqtb| in symbolic form.
Every call of |primitive| in this program is therefore accompanied by some
straightforward code that forms part of the |print_cmd_mod| routine
explained below.

@<Cases of |print_cmd_mod| for symbolic printing of primitives@>=
add_to_command:print("addto");
assignment:print(":=");
at_least:print("atleast");
at_token:print("at");
bchar_label:print("||:");
begin_group:print("begingroup");
colon:print(":");
comma:print(",");
controls:print("controls");
cull_command:print("cull");
curl_command:print("curl");
delimiters:print("delimiters");
display_command:print("display");
double_colon:print("::");
end_group:print("endgroup");
every_job_command:print("everyjob");
exit_test:print("exitif");
expand_after:print("expandafter");
from_token:print("from");
in_window:print("inwindow");
interim_command:print("interim");
left_brace:print("{");
left_bracket:print("[");
let_command:print("let");
new_internal:print("newinternal");
of_token:print("of");
open_window:print("openwindow");
path_join:print("..");
random_seed:print("randomseed");
relax:print_char("\");
right_brace:print("}");
right_bracket:print("]");
save_command:print("save");
scan_tokens:print("scantokens");
semicolon:print(";");
ship_out_command:print("shipout");
skip_to:print("skipto");
step_token:print("step");
str_op:print("str");
tension:print("tension");
to_token:print("to");
until_token:print("until");

@ We will deal with the other primitives later, at some point in the program
where their |eq_type| and |equiv| values are more meaningful.  For example,
the primitives for macro definitions will be loaded when we consider the
routines that define macros.
It is easy to find where each particular
primitive was treated by looking in the index at the end; for example, the
section where |"def"| entered |eqtb| is listed under `\&{def} primitive'.

@* \[14] Token lists.
A \MF\ token is either symbolic or numeric or a string, or it denotes
a macro parameter or capsule; so there are five corresponding ways to encode it
@^token@>
internally: (1)~A symbolic token whose hash code is~|p|
is represented by the number |p|, in the |info| field of a single-word
node in~|mem|. (2)~A numeric token whose |scaled| value is~|v| is
represented in a two-word node of~|mem|; the |type| field is |known|,
the |name_type| field is |token|, and the |value| field holds~|v|.
The fact that this token appears in a two-word node rather than a
one-word node is, of course, clear from the node address.
(3)~A string token is also represented in a two-word node; the |type|
field is |string_type|, the |name_type| field is |token|, and the
|value| field holds the corresponding |str_number|.  (4)~Capsules have
|name_type=capsule|, and their |type| and |value| fields represent
arbitrary values (in ways to be explained later).  (5)~Macro parameters
are like symbolic tokens in that they appear in |info| fields of
one-word nodes. The $k$th parameter is represented by |expr_base+k| if it
is of type \&{expr}, or by |suffix_base+k| if it is of type \&{suffix}, or
by |text_base+k| if it is of type \&{text}.  (Here |0<=k<param_size|.)
Actual values of these parameters are kept in a separate stack, as we will
see later.  The constants |expr_base|, |suffix_base|, and |text_base| are,
of course, chosen so that there will be no confusion between symbolic
tokens and parameters of various types.

It turns out that |value(null)=0|, because |null=null_coords|;
we will make use of this coincidence later.

Incidentally, while we're speaking of coincidences, we might note that
the `\\{type}' field of a node has nothing to do with ``type'' in a
printer's sense. It's curious that the same word is used in such different ways.

@d type(#) == mem[#].hh.b0 {identifies what kind of value this is}
@d name_type(#) == mem[#].hh.b1 {a clue to the name of this value}
@d token_node_size=2 {the number of words in a large token node}
@d value_loc(#)==#+1 {the word that contains the |value| field}
@d value(#)==mem[value_loc(#)].int {the value stored in a large token node}
@d expr_base==hash_end+1 {code for the zeroth \&{expr} parameter}
@d suffix_base==expr_base+param_size {code for the zeroth \&{suffix} parameter}
@d text_base==suffix_base+param_size {code for the zeroth \&{text} parameter}

@<Check the ``constant''...@>=
if text_base+param_size>max_halfword then bad:=22;

@ A numeric token is created by the following trivial routine.

@p function new_num_tok(@!v:scaled):pointer;
var @!p:pointer; {the new node}
begin p:=get_node(token_node_size); value(p):=v;
type(p):=known; name_type(p):=token; new_num_tok:=p;
end;

@ A token list is a singly linked list of nodes in |mem|, where
each node contains a token and a link.  Here's a subroutine that gets rid
of a token list when it is no longer needed.

@p procedure@?token_recycle; forward;@t\2@>@;@/
procedure flush_token_list(@!p:pointer);
var @!q:pointer; {the node being recycled}
begin while p<>null do
  begin q:=p; p:=link(p);
  if q>=hi_mem_min then free_avail(q)
  else  begin case type(q) of
    vacuous,boolean_type,known:do_nothing;
    string_type:delete_str_ref(value(q));
    unknown_types,pen_type,path_type,future_pen,picture_type,
     pair_type,transform_type,dependent,proto_dependent,independent:
      begin g_pointer:=q; token_recycle;
      end;
    othercases confusion("token")
@:this can't happen token}{\quad token@>
    endcases;@/
    free_node(q,token_node_size);
    end;
  end;
end;

@ The procedure |show_token_list|, which prints a symbolic form of
the token list that starts at a given node |p|, illustrates these
conventions. The token list being displayed should not begin with a reference
count. However, the procedure is intended to be fairly robust, so that if the
memory links are awry or if |p| is not really a pointer to a token list,
almost nothing catastrophic can happen.

An additional parameter |q| is also given; this parameter is either null
or it points to a node in the token list where a certain magic computation
takes place that will be explained later. (Basically, |q| is non-null when
we are printing the two-line context information at the time of an error
message; |q| marks the place corresponding to where the second line
should begin.)

The generation will stop, and `\.{\char`\ ETC.}' will be printed, if the length
of printing exceeds a given limit~|l|; the length of printing upon entry is
assumed to be a given amount called |null_tally|. (Note that
|show_token_list| sometimes uses itself recursively to print
variable names within a capsule.)
@^recursion@>

Unusual entries are printed in the form of all-caps tokens
preceded by a space, e.g., `\.{\char`\ BAD}'.

@<Declare the procedure called |show_token_list|@>=
procedure@?print_capsule; forward; @t\2@>@;@/
procedure show_token_list(@!p,@!q:integer;@!l,@!null_tally:integer);
label exit;
var @!class,@!c:small_number; {the |char_class| of previous and new tokens}
@!r,@!v:integer; {temporary registers}
begin class:=percent_class;
tally:=null_tally;
while (p<>null) and (tally<l) do
  begin if p=q then @<Do magic computation@>;
  @<Display token |p| and set |c| to its class;
    but |return| if there are problems@>;
  class:=c; p:=link(p);
  end;
if p<>null then print(" ETC.");
@.ETC@>
exit:
end;

@ @<Display token |p| and set |c| to its class...@>=
c:=letter_class; {the default}
if (p<mem_min)or(p>mem_end) then
  begin print(" CLOBBERED"); return;
@.CLOBBERED@>
  end;
if p<hi_mem_min then @<Display two-word token@>
else  begin r:=info(p);
  if r>=expr_base then @<Display a parameter token@>
  else if r<1 then
    if r=0 then @<Display a collective subscript@>
    else print(" IMPOSSIBLE")
@.IMPOSSIBLE@>
  else  begin r:=text(r);
    if (r<0)or(r>=str_ptr) then print(" NONEXISTENT")
@.NONEXISTENT@>
    else @<Print string |r| as a symbolic token
      and set |c| to its class@>;
    end;
  end

@ @<Display two-word token@>=
if name_type(p)=token then
  if type(p)=known then @<Display a numeric token@>
  else if type(p)<>string_type then print(" BAD")
@.BAD@>
  else  begin print_char(""""); slow_print(value(p)); print_char("""");
    c:=string_class;
    end
else if (name_type(p)<>capsule)or(type(p)<vacuous)or(type(p)>independent) then
  print(" BAD")
else  begin g_pointer:=p; print_capsule; c:=right_paren_class;
  end

@ @<Display a numeric token@>=
begin if class=digit_class then print_char(" ");
v:=value(p);
if v<0 then
  begin if class=left_bracket_class then print_char(" ");
  print_char("["); print_scaled(v); print_char("]");
  c:=right_bracket_class;
  end
else  begin print_scaled(v); c:=digit_class;
  end;
end

@ Strictly speaking, a genuine token will never have |info(p)=0|.
But we will see later (in the definition of attribute nodes) that
it is convenient to let |info(p)=0| stand for `\.{[]}'.

@<Display a collective subscript@>=
begin if class=left_bracket_class then print_char(" ");
print("[]"); c:=right_bracket_class;
end

@ @<Display a parameter token@>=
begin if r<suffix_base then
  begin print("(EXPR"); r:=r-(expr_base);
@.EXPR@>
  end
else if r<text_base then
  begin print("(SUFFIX"); r:=r-(suffix_base);
@.SUFFIX@>
  end
else  begin print("(TEXT"); r:=r-(text_base);
@.TEXT@>
  end;
print_int(r); print_char(")"); c:=right_paren_class;
end

@ @<Print string |r| as a symbolic token...@>=
begin c:=char_class[so(str_pool[str_start[r]])];
if c=class then
  case c of
  letter_class:print_char(".");
  isolated_classes:do_nothing;
  othercases print_char(" ")
  endcases;
slow_print(r);
end

@ The following procedures have been declared |forward| with no parameters,
because the author dislikes \PASCAL's convention about |forward| procedures
with parameters. It was necessary to do something, because |show_token_list|
is recursive (although the recursion is limited to one level), and because
|flush_token_list| is syntactically (but not semantically) recursive.
@^recursion@>

@<Declare miscellaneous procedures that were declared |forward|@>=
procedure print_capsule;
begin print_char("("); print_exp(g_pointer,0); print_char(")");
end;
@#
procedure token_recycle;
begin recycle_value(g_pointer);
end;

@ @<Glob...@>=
@!g_pointer:pointer; {(global) parameter to the |forward| procedures}

@ Macro definitions are kept in \MF's memory in the form of token lists
that have a few extra one-word nodes at the beginning.

The first node contains a reference count that is used to tell when the
list is no longer needed. To emphasize the fact that a reference count is
present, we shall refer to the |info| field of this special node as the
|ref_count| field.
@^reference counts@>

The next node or nodes after the reference count serve to describe the
formal parameters. They consist of zero or more parameter tokens followed
by a code for the type of macro.

@d ref_count==info {reference count preceding a macro definition or pen header}
@d add_mac_ref(#)==incr(ref_count(#)) {make a new reference to a macro list}
@d general_macro=0 {preface to a macro defined with a parameter list}
@d primary_macro=1 {preface to a macro with a \&{primary} parameter}
@d secondary_macro=2 {preface to a macro with a \&{secondary} parameter}
@d tertiary_macro=3 {preface to a macro with a \&{tertiary} parameter}
@d expr_macro=4 {preface to a macro with an undelimited \&{expr} parameter}
@d of_macro=5 {preface to a macro with
  undelimited `\&{expr} |x| \&{of}~|y|' parameters}
@d suffix_macro=6 {preface to a macro with an undelimited \&{suffix} parameter}
@d text_macro=7 {preface to a macro with an undelimited \&{text} parameter}

@p procedure delete_mac_ref(@!p:pointer);
  {|p| points to the reference count of a macro list that is
    losing one reference}
begin if ref_count(p)=null then flush_token_list(p)
else decr(ref_count(p));
end;

@ The following subroutine displays a macro, given a pointer to its
reference count.

@p @t\4@>@<Declare the procedure called |print_cmd_mod|@>@;
procedure show_macro(@!p:pointer;@!q,@!l:integer);
label exit;
var @!r:pointer; {temporary storage}
begin p:=link(p); {bypass the reference count}
while info(p)>text_macro do
  begin r:=link(p); link(p):=null;
  show_token_list(p,null,l,0); link(p):=r; p:=r;
  if l>0 then l:=l-tally@+else return;
  end; {control printing of `\.{ETC.}'}
@.ETC@>
tally:=0;
case info(p) of
general_macro:print("->");
@.->@>
primary_macro,secondary_macro,tertiary_macro:begin print_char("<");
  print_cmd_mod(param_type,info(p)); print(">->");
  end;
expr_macro:print("<expr>->");
of_macro:print("<expr>of<primary>->");
suffix_macro:print("<suffix>->");
text_macro:print("<text>->");
end; {there are no other cases}
show_token_list(link(p),q,l-tally,0);
exit:end;

@* \[15] Data structures for variables.
The variables of \MF\ programs can be simple, like `\.x', or they can
combine the structural properties of arrays and records, like `\.{x20a.b}'.
A \MF\ user assigns a type to a variable like \.{x20a.b} by saying, for
example, `\.{boolean} \.{x[]a.b}'. It's time for us to study how such
things are represented inside of the computer.

Each variable value occupies two consecutive words, either in a two-word
node called a value node, or as a two-word subfield of a larger node.  One
of those two words is called the |value| field; it is an integer,
containing either a |scaled| numeric value or the representation of some
other type of quantity. (It might also be subdivided into halfwords, in
which case it is referred to by other names instead of |value|.) The other
word is broken into subfields called |type|, |name_type|, and |link|.  The
|type| field is a quarterword that specifies the variable's type, and
|name_type| is a quarterword from which \MF\ can reconstruct the
variable's name (sometimes by using the |link| field as well).  Thus, only
1.25 words are actually devoted to the value itself; the other
three-quarters of a word are overhead, but they aren't wasted because they
allow \MF\ to deal with sparse arrays and to provide meaningful diagnostics.

In this section we shall be concerned only with the structural aspects of
variables, not their values. Later parts of the program will change the
|type| and |value| fields, but we shall treat those fields as black boxes
whose contents should not be touched.

However, if the |type| field is |structured|, there is no |value| field,
and the second word is broken into two pointer fields called |attr_head|
and |subscr_head|. Those fields point to additional nodes that
contain structural information, as we shall see.

@d subscr_head_loc(#) == #+1 {where |value|, |subscr_head|, and |attr_head| are}
@d attr_head(#) == info(subscr_head_loc(#)) {pointer to attribute info}
@d subscr_head(#) == link(subscr_head_loc(#)) {pointer to subscript info}
@d value_node_size=2 {the number of words in a value node}

@ An attribute node is three words long. Two of these words contain |type|
and |value| fields as described above, and the third word contains
additional information:  There is an |attr_loc| field, which contains the
hash address of the token that names this attribute; and there's also a
|parent| field, which points to the value node of |structured| type at the
next higher level (i.e., at the level to which this attribute is
subsidiary).  The |name_type| in an attribute node is `|attr|'.  The
|link| field points to the next attribute with the same parent; these are
arranged in increasing order, so that |attr_loc(link(p))>attr_loc(p)|. The
final attribute node links to the constant |end_attr|, whose |attr_loc|
field is greater than any legal hash address. The |attr_head| in the
parent points to a node whose |name_type| is |structured_root|; this
node represents the null attribute, i.e., the variable that is relevant
when no attributes are attached to the parent. The |attr_head| node
has the fields of either
a value node, a subscript node, or an attribute node, depending on what
the parent would be if it were not structured; but the subscript and
attribute fields are ignored, so it effectively contains only the data of
a value node. The |link| field in this special node points to an attribute
node whose |attr_loc| field is zero; the latter node represents a collective
subscript `\.{[]}' attached to the parent, and its |link| field points to
the first non-special attribute node (or to |end_attr| if there are none).

A subscript node likewise occupies three words, with |type| and |value| fields
plus extra information; its |name_type| is |subscr|. In this case the
third word is called the |subscript| field, which is a |scaled| integer.
The |link| field points to the subscript node with the next larger
subscript, if any; otherwise the |link| points to the attribute node
for collective subscripts at this level. We have seen that the latter node
contains an upward pointer, so that the parent can be deduced.

The |name_type| in a parent-less value node is |root|, and the |link|
is the hash address of the token that names this value.

In other words, variables have a hierarchical structure that includes
enough threads running around so that the program is able to move easily
between siblings, parents, and children. An example should be helpful:
(The reader is advised to draw a picture while reading the following
description, since that will help to firm up the ideas.)
Suppose that `\.x' and `\.{x.a}' and `\.{x[]b}' and `\.{x5}'
and `\.{x20b}' have been mentioned in a user's program, where
\.{x[]b} has been declared to be of \&{boolean} type. Let |h(x)|, |h(a)|,
and |h(b)| be the hash addresses of \.x, \.a, and~\.b. Then
|eq_type(h(x))=tag_token| and |equiv(h(x))=p|, where |p|~is a two-word value
node with |name_type(p)=root| and |link(p)=h(x)|. We have |type(p)=structured|,
|attr_head(p)=q|, and |subscr_head(p)=r|, where |q| points to a value
node and |r| to a subscript node. (Are you still following this? Use
a pencil to draw a diagram.) The lone variable `\.x' is represented by
|type(q)| and |value(q)|; furthermore
|name_type(q)=structured_root| and |link(q)=q1|, where |q1| points
to an attribute node representing `\.{x[]}'. Thus |name_type(q1)=attr|,
|attr_loc(q1)=collective_subscript=0|, |parent(q1)=p|,
|type(q1)=structured|, |attr_head(q1)=qq|, and |subscr_head(q1)=qq1|;
|qq| is a three-word ``attribute-as-value'' node with |type(qq)=numeric_type|
(assuming that \.{x5} is numeric, because |qq| represents `\.{x[]}'
with no further attributes), |name_type(qq)=structured_root|,
|attr_loc(qq)=0|, |parent(qq)=p|, and
|link(qq)=qq1|. (Now pay attention to the next part.) Node |qq1| is
an attribute node representing `\.{x[][]}', which has never yet
occurred; its |type| field is |undefined|, and its |value| field is
undefined. We have |name_type(qq1)=attr|, |attr_loc(qq1)=collective_subscript|,
|parent(qq1)=q1|, and |link(qq1)=qq2|. Since |qq2| represents
`\.{x[]b}', |type(qq2)=unknown_boolean|; also |attr_loc(qq2)=h(b)|,
|parent(qq2)=q1|, |name_type(qq2)=attr|, |link(qq2)=end_attr|.
(Maybe colored lines will help untangle your picture.)
 Node |r| is a subscript node with |type| and |value|
representing `\.{x5}'; |name_type(r)=subscr|, |subscript(r)=5.0|,
and |link(r)=r1| is another subscript node. To complete the picture,
see if you can guess what |link(r1)| is; give up? It's~|q1|.
Furthermore |subscript(r1)=20.0|, |name_type(r1)=subscr|,
|type(r1)=structured|, |attr_head(r1)=qqq|, |subscr_head(r1)=qqq1|,
and we finish things off with three more nodes
|qqq|, |qqq1|, and |qqq2| hung onto~|r1|. (Perhaps you should start again
with a larger sheet of paper.) The value of variable `\.{x20b}'
appears in node~|qqq2=link(qqq1)|, as you can well imagine.
Similarly, the value of `\.{x.a}' appears in node |q2=link(q1)|, where
|attr_loc(q2)=h(a)| and |parent(q2)=p|.

If the example in the previous paragraph doesn't make things crystal
clear, a glance at some of the simpler subroutines below will reveal how
things work out in practice.

The only really unusual thing about these conventions is the use of
collective subscript attributes. The idea is to avoid repeating a lot of
type information when many elements of an array are identical macros
(for which distinct values need not be stored) or when they don't have
all of the possible attributes. Branches of the structure below collective
subscript attributes do not carry actual values except for macro identifiers;
branches of the structure below subscript nodes do not carry significant
information in their collective subscript attributes.

@d attr_loc_loc(#)==#+2 {where the |attr_loc| and |parent| fields are}
@d attr_loc(#)==info(attr_loc_loc(#)) {hash address of this attribute}
@d parent(#)==link(attr_loc_loc(#)) {pointer to |structured| variable}
@d subscript_loc(#)==#+2 {where the |subscript| field lives}
@d subscript(#)==mem[subscript_loc(#)].sc {subscript of this variable}
@d attr_node_size=3 {the number of words in an attribute node}
@d subscr_node_size=3 {the number of words in a subscript node}
@d collective_subscript=0 {code for the attribute `\.{[]}'}

@<Initialize table...@>=
attr_loc(end_attr):=hash_end+1; parent(end_attr):=null;

@ Variables of type \&{pair} will have values that point to four-word
nodes containing two numeric values. The first of these values has
|name_type=x_part_sector| and the second has |name_type=y_part_sector|;
the |link| in the first points back to the node whose |value| points
to this four-word node.

Variables of type \&{transform} are similar, but in this case their
|value| points to a 12-word node containing six values, identified by
|x_part_sector|, |y_part_sector|, |xx_part_sector|, |xy_part_sector|,
|yx_part_sector|, and |yy_part_sector|.

When an entire structured variable is saved, the |root| indication
is temporarily replaced by |saved_root|.

Some variables have no name; they just are used for temporary storage
while expressions are being evaluated. We call them {\sl capsules}.

@d x_part_loc(#)==# {where the \&{xpart} is found in a pair or transform node}
@d y_part_loc(#)==#+2 {where the \&{ypart} is found in a pair or transform node}
@d xx_part_loc(#)==#+4 {where the \&{xxpart} is found in a transform node}
@d xy_part_loc(#)==#+6 {where the \&{xypart} is found in a transform node}
@d yx_part_loc(#)==#+8 {where the \&{yxpart} is found in a transform node}
@d yy_part_loc(#)==#+10 {where the \&{yypart} is found in a transform node}
@#
@d pair_node_size=4 {the number of words in a pair node}
@d transform_node_size=12 {the number of words in a transform node}

@<Glob...@>=
@!big_node_size:array[transform_type..pair_type] of small_number;

@ The |big_node_size| array simply contains two constants that \MF\
occasionally needs to know.

@<Set init...@>=
big_node_size[transform_type]:=transform_node_size;
big_node_size[pair_type]:=pair_node_size;

@ If |type(p)=pair_type| or |transform_type| and if |value(p)=null|, the
procedure call |init_big_node(p)| will allocate a pair or transform node
for~|p|.  The individual parts of such nodes are initially of type
|independent|.

@p procedure init_big_node(@!p:pointer);
var @!q:pointer; {the new node}
@!s:small_number; {its size}
begin s:=big_node_size[type(p)]; q:=get_node(s);
repeat s:=s-2; @<Make variable |q+s| newly independent@>;
name_type(q+s):=half(s)+x_part_sector; link(q+s):=null;
until s=0;
link(q):=p; value(p):=q;
end;

@ The |id_transform| function creates a capsule for the
identity transformation.

@p function id_transform:pointer;
var @!p,@!q,@!r:pointer; {list manipulation registers}
begin p:=get_node(value_node_size); type(p):=transform_type;
name_type(p):=capsule; value(p):=null; init_big_node(p); q:=value(p);
r:=q+transform_node_size;
repeat r:=r-2;
type(r):=known; value(r):=0;
until r=q;
value(xx_part_loc(q)):=unity; value(yy_part_loc(q)):=unity;
id_transform:=p;
end;

@ Tokens are of type |tag_token| when they first appear, but they point
to |null| until they are first used as the root of a variable.
The following subroutine establishes the root node on such grand occasions.

@p procedure new_root(@!x:pointer);
var @!p:pointer; {the new node}
begin p:=get_node(value_node_size); type(p):=undefined; name_type(p):=root;
link(p):=x; equiv(x):=p;
end;

@ These conventions for variable representation are illustrated by the
|print_variable_name| routine, which displays the full name of a
variable given only a pointer to its two-word value packet.

@p procedure print_variable_name(@!p:pointer);
label found,exit;
var @!q:pointer; {a token list that will name the variable's suffix}
@!r:pointer; {temporary for token list creation}
begin while name_type(p)>=x_part_sector do
  @<Preface the output with a part specifier; |return| in the
    case of a capsule@>;
q:=null;
while name_type(p)>saved_root do
  @<Ascend one level, pushing a token onto list |q|
   and replacing |p| by its parent@>;
r:=get_avail; info(r):=link(p); link(r):=q;
if name_type(p)=saved_root then print("(SAVED)");
@.SAVED@>
show_token_list(r,null,el_gordo,tally); flush_token_list(r);
exit:end;

@ @<Ascend one level, pushing a token onto list |q|...@>=
begin if name_type(p)=subscr then
  begin r:=new_num_tok(subscript(p));
  repeat p:=link(p);
  until name_type(p)=attr;
  end
else if name_type(p)=structured_root then
    begin p:=link(p); goto found;
    end
else  begin if name_type(p)<>attr then confusion("var");
@:this can't happen var}{\quad var@>
  r:=get_avail; info(r):=attr_loc(p);
  end;
link(r):=q; q:=r;
found:  p:=parent(p);
end

@ @<Preface the output with a part specifier...@>=
begin case name_type(p) of
x_part_sector: print_char("x");
y_part_sector: print_char("y");
xx_part_sector: print("xx");
xy_part_sector: print("xy");
yx_part_sector: print("yx");
yy_part_sector: print("yy");
capsule: begin print("%CAPSULE"); print_int(p-null); return;
@.CAPSULE@>
  end;
end; {there are no other cases}
print("part "); p:=link(p-2*(name_type(p)-x_part_sector));
end

@ The |interesting| function returns |true| if a given variable is not
in a capsule, or if the user wants to trace capsules.

@p function interesting(@!p:pointer):boolean;
var @!t:small_number; {a |name_type|}
begin if internal[tracing_capsules]>0 then interesting:=true
else  begin t:=name_type(p);
  if t>=x_part_sector then if t<>capsule then
    t:=name_type(link(p-2*(t-x_part_sector)));
  interesting:=(t<>capsule);
  end;
end;

@ Now here is a subroutine that converts an unstructured type into an
equivalent structured type, by inserting a |structured| node that is
capable of growing. This operation is done only when |name_type(p)=root|,
|subscr|, or |attr|.

The procedure returns a pointer to the new node that has taken node~|p|'s
place in the structure. Node~|p| itself does not move, nor are its
|value| or |type| fields changed in any way.

@p function new_structure(@!p:pointer):pointer;
var @!q,@!r:pointer; {list manipulation registers}
begin case name_type(p) of
root: begin q:=link(p); r:=get_node(value_node_size); equiv(q):=r;
  end;
subscr: @<Link a new subscript node |r| in place of node |p|@>;
attr: @<Link a new attribute node |r| in place of node |p|@>;
othercases confusion("struct")
@:this can't happen struct}{\quad struct@>
endcases;@/
link(r):=link(p); type(r):=structured; name_type(r):=name_type(p);
attr_head(r):=p; name_type(p):=structured_root;@/
q:=get_node(attr_node_size); link(p):=q; subscr_head(r):=q;
parent(q):=r; type(q):=undefined; name_type(q):=attr; link(q):=end_attr;
attr_loc(q):=collective_subscript; new_structure:=r;
end;

@ @<Link a new subscript node |r| in place of node |p|@>=
begin q:=p;
repeat q:=link(q);
until name_type(q)=attr;
q:=parent(q); r:=subscr_head_loc(q); {|link(r)=subscr_head(q)|}
repeat q:=r; r:=link(r);
until r=p;
r:=get_node(subscr_node_size);
link(q):=r; subscript(r):=subscript(p);
end

@ If the attribute is |collective_subscript|, there are two pointers to
node~|p|, so we must change both of them.

@<Link a new attribute node |r| in place of node |p|@>=
begin q:=parent(p); r:=attr_head(q);
repeat q:=r; r:=link(r);
until r=p;
r:=get_node(attr_node_size); link(q):=r;@/
mem[attr_loc_loc(r)]:=mem[attr_loc_loc(p)]; {copy |attr_loc| and |parent|}
if attr_loc(p)=collective_subscript then
  begin q:=subscr_head_loc(parent(p));
  while link(q)<>p do q:=link(q);
  link(q):=r;
  end;
end

@ The |find_variable| routine is given a pointer~|t| to a nonempty token
list of suffixes; it returns a pointer to the corresponding two-word
value. For example, if |t| points to token \.x followed by a numeric
token containing the value~7, |find_variable| finds where the value of
\.{x7} is stored in memory. This may seem a simple task, and it
usually is, except when \.{x7} has never been referenced before.
Indeed, \.x may never have even been subscripted before; complexities
arise with respect to updating the collective subscript information.

If a macro type is detected anywhere along path~|t|, or if the first
item on |t| isn't a |tag_token|, the value |null| is returned.
Otherwise |p| will be a non-null pointer to a node such that
|undefined<type(p)<structured|.

@d abort_find==begin find_variable:=null; return;@+end

@p function find_variable(@!t:pointer):pointer;
label exit;
var @!p,@!q,@!r,@!s:pointer; {nodes in the ``value'' line}
@!pp,@!qq,@!rr,@!ss:pointer; {nodes in the ``collective'' line}
@!n:integer; {subscript or attribute}
@!save_word:memory_word; {temporary storage for a word of |mem|}
@^inner loop@>
begin p:=info(t); t:=link(t);
if eq_type(p) mod outer_tag<>tag_token then abort_find;
if equiv(p)=null then new_root(p);
p:=equiv(p); pp:=p;
while t<>null do
  begin @<Make sure that both nodes |p| and |pp| are of |structured| type@>;
  if t<hi_mem_min then
    @<Descend one level for the subscript |value(t)|@>
  else @<Descend one level for the attribute |info(t)|@>;
  t:=link(t);
  end;
if type(pp)>=structured then
  if type(pp)=structured then pp:=attr_head(pp)@+else abort_find;
if type(p)=structured then p:=attr_head(p);
if type(p)=undefined then
  begin if type(pp)=undefined then
    begin type(pp):=numeric_type; value(pp):=null;
    end;
  type(p):=type(pp); value(p):=null;
  end;
find_variable:=p;
exit:end;

@ Although |pp| and |p| begin together, they diverge when a subscript occurs;
|pp|~stays in the collective line while |p|~goes through actual subscript
values.

@<Make sure that both nodes |p| and |pp|...@>=
if type(pp)<>structured then
  begin if type(pp)>structured then abort_find;
  ss:=new_structure(pp);
  if p=pp then p:=ss;
  pp:=ss;
  end; {now |type(pp)=structured|}
if type(p)<>structured then {it cannot be |>structured|}
  p:=new_structure(p) {now |type(p)=structured|}

@ We want this part of the program to be reasonably fast, in case there are
@^inner loop@>
lots of subscripts at the same level of the data structure. Therefore
we store an ``infinite'' value in the word that appears at the end of the
subscript list, even though that word isn't part of a subscript node.

@<Descend one level for the subscript |value(t)|@>=
begin n:=value(t);
pp:=link(attr_head(pp)); {now |attr_loc(pp)=collective_subscript|}
q:=link(attr_head(p)); save_word:=mem[subscript_loc(q)];
subscript(q):=el_gordo; s:=subscr_head_loc(p); {|link(s)=subscr_head(p)|}
repeat r:=s; s:=link(s);
until n<=subscript(s);
if n=subscript(s) then p:=s
else  begin p:=get_node(subscr_node_size); link(r):=p; link(p):=s;
  subscript(p):=n; name_type(p):=subscr; type(p):=undefined;
  end;
mem[subscript_loc(q)]:=save_word;
end

@ @<Descend one level for the attribute |info(t)|@>=
begin n:=info(t);
ss:=attr_head(pp);
repeat rr:=ss; ss:=link(ss);
until n<=attr_loc(ss);
if n<attr_loc(ss) then
  begin qq:=get_node(attr_node_size); link(rr):=qq; link(qq):=ss;
  attr_loc(qq):=n; name_type(qq):=attr; type(qq):=undefined;
  parent(qq):=pp; ss:=qq;
  end;
if p=pp then
  begin p:=ss; pp:=ss;
  end
else  begin pp:=ss; s:=attr_head(p);
  repeat r:=s; s:=link(s);
  until n<=attr_loc(s);
  if n=attr_loc(s) then p:=s
  else  begin q:=get_node(attr_node_size); link(r):=q; link(q):=s;
    attr_loc(q):=n; name_type(q):=attr; type(q):=undefined;
    parent(q):=p; p:=q;
    end;
  end;
end

@ Variables lose their former values when they appear in a type declaration,
or when they are defined to be macros or \&{let} equal to something else.
A subroutine will be defined later that recycles the storage associated
with any particular |type| or |value|; our goal now is to study a higher
level process called |flush_variable|, which selectively frees parts of a
variable structure.

This routine has some complexity because of examples such as
`\hbox{\tt numeric x[]a[]b}',
which recycles all variables of the form \.{x[i]a[j]b} (and no others), while
`\hbox{\tt vardef x[]a[]=...}'
discards all variables of the form \.{x[i]a[j]} followed by an arbitrary
suffix, except for the collective node \.{x[]a[]} itself. The obvious way
to handle such examples is to use recursion; so that's what we~do.
@^recursion@>

Parameter |p| points to the root information of the variable;
parameter |t| points to a list of one-word nodes that represent
suffixes, with |info=collective_subscript| for subscripts.

@p @t\4@>@<Declare subroutines for printing expressions@>@;@/
@t\4@>@<Declare basic dependency-list subroutines@>@;
@t\4@>@<Declare the recycling subroutines@>@;
@t\4@>@<Declare the procedure called |flush_cur_exp|@>@;
@t\4@>@<Declare the procedure called |flush_below_variable|@>@;
procedure flush_variable(@!p,@!t:pointer;@!discard_suffixes:boolean);
label exit;
var @!q,@!r:pointer; {list manipulation}
@!n:halfword; {attribute to match}
begin while t<>null do
  begin if type(p)<>structured then return;
  n:=info(t); t:=link(t);
  if n=collective_subscript then
    begin r:=subscr_head_loc(p); q:=link(r); {|q=subscr_head(p)|}
    while name_type(q)=subscr do
      begin flush_variable(q,t,discard_suffixes);
      if t=null then
        if type(q)=structured then r:=q
        else  begin link(r):=link(q); free_node(q,subscr_node_size);
          end
      else r:=q;
      q:=link(r);
      end;
    end;
  p:=attr_head(p);
  repeat r:=p; p:=link(p);
  until attr_loc(p)>=n;
  if attr_loc(p)<>n then return;
  end;
if discard_suffixes then flush_below_variable(p)
else  begin if type(p)=structured then p:=attr_head(p);
  recycle_value(p);
  end;
exit:end;

@ The next procedure is simpler; it wipes out everything but |p| itself,
which becomes undefined.

@<Declare the procedure called |flush_below_variable|@>=
procedure flush_below_variable(@!p:pointer);
var @!q,@!r:pointer; {list manipulation registers}
begin if type(p)<>structured then
  recycle_value(p) {this sets |type(p)=undefined|}
else  begin q:=subscr_head(p);
  while name_type(q)=subscr do
    begin flush_below_variable(q); r:=q; q:=link(q);
    free_node(r,subscr_node_size);
    end;
  r:=attr_head(p); q:=link(r); recycle_value(r);
  if name_type(p)<=saved_root then free_node(r,value_node_size)
  else free_node(r,subscr_node_size);
    {we assume that |subscr_node_size=attr_node_size|}
  repeat flush_below_variable(q); r:=q; q:=link(q); free_node(r,attr_node_size);
  until q=end_attr;
  type(p):=undefined;
  end;
end;

@ Just before assigning a new value to a variable, we will recycle the
old value and make the old value undefined. The |und_type| routine
determines what type of undefined value should be given, based on
the current type before recycling.

@p function und_type(@!p:pointer):small_number;
begin case type(p) of
undefined,vacuous:und_type:=undefined;
boolean_type,unknown_boolean:und_type:=unknown_boolean;
string_type,unknown_string:und_type:=unknown_string;
pen_type,unknown_pen,future_pen:und_type:=unknown_pen;
path_type,unknown_path:und_type:=unknown_path;
picture_type,unknown_picture:und_type:=unknown_picture;
transform_type,pair_type,numeric_type:und_type:=type(p);
known,dependent,proto_dependent,independent:und_type:=numeric_type;
end; {there are no other cases}
end;

@ The |clear_symbol| routine is used when we want to redefine the equivalent
of a symbolic token. It must remove any variable structure or macro
definition that is currently attached to that symbol. If the |saving|
parameter is true, a subsidiary structure is saved instead of destroyed.

@p procedure clear_symbol(@!p:pointer;@!saving:boolean);
var @!q:pointer; {|equiv(p)|}
begin q:=equiv(p);
case eq_type(p) mod outer_tag of
defined_macro,secondary_primary_macro,tertiary_secondary_macro,
 expression_tertiary_macro: if not saving then delete_mac_ref(q);
tag_token:if q<>null then
  if saving then name_type(q):=saved_root
  else  begin flush_below_variable(q); free_node(q,value_node_size);
    end;@;
othercases do_nothing
endcases;@/
eqtb[p]:=eqtb[frozen_undefined];
end;

@* \[16] Saving and restoring equivalents.
The nested structure provided by \&{begingroup} and \&{endgroup}
allows |eqtb| entries to be saved and restored, so that temporary changes
can be made without difficulty.  When the user requests a current value to
be saved, \MF\ puts that value into its ``save stack.'' An appearance of
\&{endgroup} ultimately causes the old values to be removed from the save
stack and put back in their former places.

The save stack is a linked list containing three kinds of entries,
distinguished by their |info| fields. If |p| points to a saved item,
then

\smallskip\hang
|info(p)=0| stands for a group boundary; each \&{begingroup} contributes
such an item to the save stack and each \&{endgroup} cuts back the stack
until the most recent such entry has been removed.

\smallskip\hang
|info(p)=q|, where |1<=q<=hash_end|, means that |mem[p+1]| holds the former
contents of |eqtb[q]|. Such save stack entries are generated by \&{save}
commands.

\smallskip\hang
|info(p)=hash_end+q|, where |q>0|, means that |value(p)| is a |scaled|
integer to be restored to internal parameter number~|q|. Such entries
are generated by \&{interim} commands.

\smallskip\noindent
The global variable |save_ptr| points to the top item on the save stack.

@d save_node_size=2 {number of words per non-boundary save-stack node}
@d saved_equiv(#)==mem[#+1].hh {where an |eqtb| entry gets saved}
@d save_boundary_item(#)==begin #:=get_avail; info(#):=0;
  link(#):=save_ptr; save_ptr:=#;
  end

@<Glob...@>=@!save_ptr:pointer; {the most recently saved item}

@ @<Set init...@>=save_ptr:=null;

@ The |save_variable| routine is given a hash address |q|; it salts this
address away in the save stack, together with its current equivalent,
then makes token~|q| behave as though it were brand new.

Nothing is stacked when |save_ptr=null|, however; there's no way to remove
things from the stack when the program is not inside a group, so there's
no point in wasting the space.

@p procedure save_variable(@!q:pointer);
var @!p:pointer; {temporary register}
begin if save_ptr<>null then
  begin p:=get_node(save_node_size); info(p):=q; link(p):=save_ptr;
  saved_equiv(p):=eqtb[q]; save_ptr:=p;
  end;
clear_symbol(q,(save_ptr<>null));
end;

@ Similarly, |save_internal| is given the location |q| of an internal
quantity like |tracing_pens|. It creates a save stack entry of the
third kind.

@p procedure save_internal(@!q:halfword);
var @!p:pointer; {new item for the save stack}
begin if save_ptr<>null then
  begin p:=get_node(save_node_size); info(p):=hash_end+q;
  link(p):=save_ptr; value(p):=internal[q]; save_ptr:=p;
  end;
end;

@ At the end of a group, the |unsave| routine restores all of the saved
equivalents in reverse order. This routine will be called only when there
is at least one boundary item on the save stack.

@p procedure unsave;
var @!q:pointer; {index to saved item}
@!p:pointer; {temporary register}
begin while info(save_ptr)<>0 do
  begin q:=info(save_ptr);
  if q>hash_end then
    begin if internal[tracing_restores]>0 then
      begin begin_diagnostic; print_nl("{restoring ");
      slow_print(int_name[q-(hash_end)]); print_char("=");
      print_scaled(value(save_ptr)); print_char("}");
      end_diagnostic(false);
      end;
    internal[q-(hash_end)]:=value(save_ptr);
    end
  else  begin if internal[tracing_restores]>0 then
      begin begin_diagnostic; print_nl("{restoring ");
      slow_print(text(q)); print_char("}");
      end_diagnostic(false);
      end;
    clear_symbol(q,false);
    eqtb[q]:=saved_equiv(save_ptr);
    if eq_type(q) mod outer_tag=tag_token then
      begin p:=equiv(q);
      if p<>null then name_type(p):=root;
      end;
    end;
  p:=link(save_ptr); free_node(save_ptr,save_node_size); save_ptr:=p;
  end;
p:=link(save_ptr); free_avail(save_ptr); save_ptr:=p;
end;

@* \[17] Data structures for paths.
When a \MF\ user specifies a path, \MF\ will create a list of knots
and control points for the associated cubic spline curves. If the
knots are $z_0$, $z_1$, \dots, $z_n$, there are control points
$z_k^+$ and $z_{k+1}^-$ such that the cubic splines between knots
$z_k$ and $z_{k+1}$ are defined by B\'ezier's formula
@:Bezier}{B\'ezier, Pierre Etienne@>
$$\eqalign{z(t)&=B(z_k,z_k^+,z_{k+1}^-,z_{k+1};t)\cr
&=(1-t)^3z_k+3(1-t)^2tz_k^++3(1-t)t^2z_{k+1}^-+t^3z_{k+1}\cr}$$
for |0<=t<=1|.

There is a 7-word node for each knot $z_k$, containing one word of
control information and six words for the |x| and |y| coordinates
of $z_k^-$ and $z_k$ and~$z_k^+$. The control information appears
in the |left_type| and |right_type| fields, which each occupy
a quarter of the first word in the node; they specify properties
of the curve as it enters and leaves the knot. There's also a
halfword |link| field, which points to the following knot.

If the path is a closed contour, knots 0 and |n| are identical;
i.e., the |link| in knot |n-1| points to knot~0. But if the path
is not closed, the |left_type| of knot~0 and the |right_type| of knot~|n|
are equal to |endpoint|. In the latter case the |link| in knot~|n| points
to knot~0, and the control points $z_0^-$ and $z_n^+$ are not used.

@d left_type(#) == mem[#].hh.b0 {characterizes the path entering this knot}
@d right_type(#) == mem[#].hh.b1 {characterizes the path leaving this knot}
@d endpoint=0 {|left_type| at path beginning and |right_type| at path end}
@d x_coord(#) == mem[#+1].sc {the |x| coordinate of this knot}
@d y_coord(#) == mem[#+2].sc {the |y| coordinate of this knot}
@d left_x(#) == mem[#+3].sc {the |x| coordinate of previous control point}
@d left_y(#) == mem[#+4].sc {the |y| coordinate of previous control point}
@d right_x(#) == mem[#+5].sc {the |x| coordinate of next control point}
@d right_y(#) == mem[#+6].sc {the |y| coordinate of next control point}
@d knot_node_size=7 {number of words in a knot node}

@ Before the B\'ezier control points have been calculated, the memory
space they will ultimately occupy is taken up by information that can be
used to compute them. There are four cases:

\yskip
\textindent{$\bullet$} If |right_type=open|, the curve should leave
the knot in the same direction it entered; \MF\ will figure out a
suitable direction.

\yskip
\textindent{$\bullet$} If |right_type=curl|, the curve should leave the
knot in a direction depending on the angle at which it enters the next
knot and on the curl parameter stored in |right_curl|.

\yskip
\textindent{$\bullet$} If |right_type=given|, the curve should leave the
knot in a nonzero direction stored as an |angle| in |right_given|.

\yskip
\textindent{$\bullet$} If |right_type=explicit|, the B\'ezier control
point for leaving this knot has already been computed; it is in the
|right_x| and |right_y| fields.

\yskip\noindent
The rules for |left_type| are similar, but they refer to the curve entering
the knot, and to \\{left} fields instead of \\{right} fields.

Non-|explicit| control points will be chosen based on ``tension'' parameters
in the |left_tension| and |right_tension| fields. The
`\&{atleast}' option is represented by negative tension values.
@:at_least_}{\&{atleast} primitive@>

For example, the \MF\ path specification
$$\.{z0..z1..tension atleast 1..\{curl 2\}z2..z3\{-1,-2\}..tension
  3 and 4..p},$$
where \.p is the path `\.{z4..controls z45 and z54..z5}', will be represented
by the six knots
\def\lodash{\hbox to 1.1em{\thinspace\hrulefill\thinspace}}
$$\vbox{\halign{#\hfil&&\qquad#\hfil\cr
|left_type|&\\{left} info&|x_coord,y_coord|&|right_type|&\\{right} info\cr
\noalign{\yskip}
|endpoint|&\lodash$,\,$\lodash&$x_0,y_0$&|curl|&$1.0,1.0$\cr
|open|&\lodash$,1.0$&$x_1,y_1$&|open|&\lodash$,-1.0$\cr
|curl|&$2.0,-1.0$&$x_2,y_2$&|curl|&$2.0,1.0$\cr
|given|&$d,1.0$&$x_3,y_3$&|given|&$d,3.0$\cr
|open|&\lodash$,4.0$&$x_4,y_4$&|explicit|&$x_{45},y_{45}$\cr
|explicit|&$x_{54},y_{54}$&$x_5,y_5$&|endpoint|&\lodash$,\,$\lodash\cr}}$$
Here |d| is the |angle| obtained by calling |n_arg(-unity,-two)|.
Of course, this example is more complicated than anything a normal user
would ever write.

These types must satisfy certain restrictions because of the form of \MF's
path syntax:
(i)~|open| type never appears in the same node together with |endpoint|,
|given|, or |curl|.
(ii)~The |right_type| of a node is |explicit| if and only if the
|left_type| of the following node is |explicit|.
(iii)~|endpoint| types occur only at the ends, as mentioned above.

@d left_curl==left_x {curl information when entering this knot}
@d left_given==left_x {given direction when entering this knot}
@d left_tension==left_y {tension information when entering this knot}
@d right_curl==right_x {curl information when leaving this knot}
@d right_given==right_x {given direction when leaving this knot}
@d right_tension==right_y {tension information when leaving this knot}
@d explicit=1 {|left_type| or |right_type| when control points are known}
@d given=2 {|left_type| or |right_type| when a direction is given}
@d curl=3 {|left_type| or |right_type| when a curl is desired}
@d open=4 {|left_type| or |right_type| when \MF\ should choose the direction}

@ Here is a diagnostic routine that prints a given knot list
in symbolic form. It illustrates the conventions discussed above,
and checks for anomalies that might arise while \MF\ is being debugged.

@<Declare subroutines for printing expressions@>=
procedure print_path(@!h:pointer;@!s:str_number;@!nuline:boolean);
label done,done1;
var @!p,@!q:pointer; {for list traversal}
begin print_diagnostic("Path",s,nuline); print_ln;
@.Path at line...@>
p:=h;
repeat q:=link(p);
if (p=null)or(q=null) then
  begin print_nl("???"); goto done; {this won't happen}
@.???@>
  end;
@<Print information for adjacent knots |p| and |q|@>;
p:=q;
if (p<>h)or(left_type(h)<>endpoint) then
  @<Print two dots, followed by |given| or |curl| if present@>;
until p=h;
if left_type(h)<>endpoint then print("cycle");
done:end_diagnostic(true);
end;

@ @<Print information for adjacent knots...@>=
print_two(x_coord(p),y_coord(p));
case right_type(p) of
endpoint: begin if left_type(p)=open then print("{open?}"); {can't happen}
@.open?@>
  if (left_type(q)<>endpoint)or(q<>h) then q:=null; {force an error}
  goto done1;
  end;
explicit: @<Print control points between |p| and |q|, then |goto done1|@>;
open: @<Print information for a curve that begins |open|@>;
curl,given: @<Print information for a curve that begins |curl| or |given|@>;
othercases print("???") {can't happen}
@.???@>
endcases;@/
if left_type(q)<=explicit then print("..control?") {can't happen}
@.control?@>
else if (right_tension(p)<>unity)or(left_tension(q)<>unity) then
  @<Print tension between |p| and |q|@>;
done1:

@ Since |n_sin_cos| produces |fraction| results, which we will print as if they
were |scaled|, the magnitude of a |given| direction vector will be~4096.

@<Print two dots...@>=
begin print_nl(" ..");
if left_type(p)=given then
  begin n_sin_cos(left_given(p)); print_char("{");
  print_scaled(n_cos); print_char(",");
  print_scaled(n_sin); print_char("}");
  end
else if left_type(p)=curl then
  begin print("{curl "); print_scaled(left_curl(p)); print_char("}");
  end;
end

@ @<Print tension between |p| and |q|@>=
begin print("..tension ");
if right_tension(p)<0 then print("atleast");
print_scaled(abs(right_tension(p)));
if right_tension(p)<>left_tension(q) then
  begin print(" and ");
  if left_tension(q)<0 then print("atleast");
  print_scaled(abs(left_tension(q)));
  end;
end

@ @<Print control points between |p| and |q|, then |goto done1|@>=
begin print("..controls "); print_two(right_x(p),right_y(p)); print(" and ");
if left_type(q)<>explicit then print("??") {can't happen}
@.??@>
else print_two(left_x(q),left_y(q));
goto done1;
end

@ @<Print information for a curve that begins |open|@>=
if (left_type(p)<>explicit)and(left_type(p)<>open) then
  print("{open?}") {can't happen}
@.open?@>

@ A curl of 1 is shown explicitly, so that the user sees clearly that
\MF's default curl is present.

@<Print information for a curve that begins |curl|...@>=
begin if left_type(p)=open then print("??"); {can't happen}
@.??@>
if right_type(p)=curl then
  begin print("{curl "); print_scaled(right_curl(p));
  end
else  begin n_sin_cos(right_given(p)); print_char("{");
  print_scaled(n_cos); print_char(","); print_scaled(n_sin);
  end;
print_char("}");
end

@ If we want to duplicate a knot node, we can say |copy_knot|:

@p function copy_knot(@!p:pointer):pointer;
var @!q:pointer; {the copy}
@!k:0..knot_node_size-1; {runs through the words of a knot node}
begin q:=get_node(knot_node_size);
for k:=0 to knot_node_size-1 do mem[q+k]:=mem[p+k];
copy_knot:=q;
end;

@ The |copy_path| routine makes a clone of a given path.

@p function copy_path(@!p:pointer):pointer;
label exit;
var @!q,@!pp,@!qq:pointer; {for list manipulation}
begin q:=get_node(knot_node_size); {this will correspond to |p|}
qq:=q; pp:=p;
loop@+  begin left_type(qq):=left_type(pp);
  right_type(qq):=right_type(pp);@/
  x_coord(qq):=x_coord(pp); y_coord(qq):=y_coord(pp);@/
  left_x(qq):=left_x(pp); left_y(qq):=left_y(pp);@/
  right_x(qq):=right_x(pp); right_y(qq):=right_y(pp);@/
  if link(pp)=p then
    begin link(qq):=q; copy_path:=q; return;
    end;
  link(qq):=get_node(knot_node_size); qq:=link(qq); pp:=link(pp);
  end;
exit:end;

@ Similarly, there's a way to copy the {\sl reverse\/} of a path. This procedure
returns a pointer to the first node of the copy, if the path is a cycle,
but to the final node of a non-cyclic copy. The global
variable |path_tail| will point to the final node of the original path;
this trick makes it easier to implement `\&{doublepath}'.

All node types are assumed to be |endpoint| or |explicit| only.

@p function htap_ypoc(@!p:pointer):pointer;
label exit;
var @!q,@!pp,@!qq,@!rr:pointer; {for list manipulation}
begin q:=get_node(knot_node_size); {this will correspond to |p|}
qq:=q; pp:=p;
loop@+  begin right_type(qq):=left_type(pp); left_type(qq):=right_type(pp);@/
  x_coord(qq):=x_coord(pp); y_coord(qq):=y_coord(pp);@/
  right_x(qq):=left_x(pp); right_y(qq):=left_y(pp);@/
  left_x(qq):=right_x(pp); left_y(qq):=right_y(pp);@/
  if link(pp)=p then
    begin link(q):=qq; path_tail:=pp; htap_ypoc:=q; return;
    end;
  rr:=get_node(knot_node_size); link(rr):=qq; qq:=rr; pp:=link(pp);
  end;
exit:end;

@ @<Glob...@>=
@!path_tail:pointer; {the node that links to the beginning of a path}

@ When a cyclic list of knot nodes is no longer needed, it can be recycled by
calling the following subroutine.

@<Declare the recycling subroutines@>=
procedure toss_knot_list(@!p:pointer);
var @!q:pointer; {the node being freed}
@!r:pointer; {the next node}
begin q:=p;
repeat r:=link(q); free_node(q,knot_node_size); q:=r;
until q=p;
end;

@* \[18] Choosing control points.
Now we must actually delve into one of \MF's more difficult routines,
the |make_choices| procedure that chooses angles and control points for
the splines of a curve when the user has not specified them explicitly.
The parameter to |make_choices| points to a list of knots and
path information, as described above.

A path decomposes into independent segments at ``breakpoint'' knots,
which are knots whose left and right angles are both prespecified in
some way (i.e., their |left_type| and |right_type| aren't both open).

@p @t\4@>@<Declare the procedure called |solve_choices|@>@;
procedure make_choices(@!knots:pointer);
label done;
var @!h:pointer; {the first breakpoint}
@!p,@!q:pointer; {consecutive breakpoints being processed}
@<Other local variables for |make_choices|@>@;
begin check_arith; {make sure that |arith_error=false|}
if internal[tracing_choices]>0 then
  print_path(knots,", before choices",true);
@<If consecutive knots are equal, join them explicitly@>;
@<Find the first breakpoint, |h|, on the path;
  insert an artificial breakpoint if the path is an unbroken cycle@>;
p:=h;
repeat @<Fill in the control points between |p| and the next breakpoint,
  then advance |p| to that breakpoint@>;
until p=h;
if internal[tracing_choices]>0 then
  print_path(knots,", after choices",true);
if arith_error then @<Report an unexpected problem during the choice-making@>;
end;

@ @<Report an unexpected problem during the choice...@>=
begin print_err("Some number got too big");
@.Some number got too big@>
help2("The path that I just computed is out of range.")@/
  ("So it will probably look funny. Proceed, for a laugh.");
put_get_error; arith_error:=false;
end

@ Two knots in a row with the same coordinates will always be joined
by an explicit ``curve'' whose control points are identical with the
knots.

@<If consecutive knots are equal, join them explicitly@>=
p:=knots;
repeat q:=link(p);
if x_coord(p)=x_coord(q) then if y_coord(p)=y_coord(q) then
 if right_type(p)>explicit then
  begin right_type(p):=explicit;
  if left_type(p)=open then
    begin left_type(p):=curl; left_curl(p):=unity;
    end;
  left_type(q):=explicit;
  if right_type(q)=open then
    begin right_type(q):=curl; right_curl(q):=unity;
    end;
  right_x(p):=x_coord(p); left_x(q):=x_coord(p);@/
  right_y(p):=y_coord(p); left_y(q):=y_coord(p);
  end;
p:=q;
until p=knots

@ If there are no breakpoints, it is necessary to compute the direction
angles around an entire cycle. In this case the |left_type| of the first
node is temporarily changed to |end_cycle|.

@d end_cycle=open+1

@<Find the first breakpoint, |h|, on the path...@>=
h:=knots;
loop@+  begin if left_type(h)<>open then goto done;
  if right_type(h)<>open then goto done;
  h:=link(h);
  if h=knots then
    begin left_type(h):=end_cycle; goto done;
    end;
  end;
done:

@ If |right_type(p)<given| and |q=link(p)|, we must have
|right_type(p)=left_type(q)=explicit| or |endpoint|.

@<Fill in the control points between |p| and the next breakpoint...@>=
q:=link(p);
if right_type(p)>=given then
  begin while (left_type(q)=open)and(right_type(q)=open) do q:=link(q);
  @<Fill in the control information between
    consecutive breakpoints |p| and |q|@>;
  end;
p:=q

@ Before we can go further into the way choices are made, we need to
consider the underlying theory. The basic ideas implemented in |make_choices|
are due to John Hobby, who introduced the notion of ``mock curvature''
@^Hobby, John Douglas@>
at a knot. Angles are chosen so that they preserve mock curvature when
a knot is passed, and this has been found to produce excellent results.

It is convenient to introduce some notations that simplify the necessary
formulas. Let $d_{k,k+1}=\vert z\k-z_k\vert$ be the (nonzero) distance
between knots |k| and |k+1|; and let
$${z\k-z_k\over z_k-z_{k-1}}={d_{k,k+1}\over d_{k-1,k}}e^{i\psi_k}$$
so that a polygonal line from $z_{k-1}$ to $z_k$ to $z\k$ turns left
through an angle of~$\psi_k$. We assume that $\vert\psi_k\vert\L180^\circ$.
The control points for the spline from $z_k$ to $z\k$ will be denoted by
$$\eqalign{z_k^+&=z_k+
  \textstyle{1\over3}\rho_k e^{i\theta_k}(z\k-z_k),\cr
 z\k^-&=z\k-
  \textstyle{1\over3}\sigma\k e^{-i\phi\k}(z\k-z_k),\cr}$$
where $\rho_k$ and $\sigma\k$ are nonnegative ``velocity ratios'' at the
beginning and end of the curve, while $\theta_k$ and $\phi\k$ are the
corresponding ``offset angles.'' These angles satisfy the condition
$$\theta_k+\phi_k+\psi_k=0,\eqno(*)$$
whenever the curve leaves an intermediate knot~|k| in the direction that
it enters.

@ Let $\alpha_k$ and $\beta\k$ be the reciprocals of the ``tension'' of
the curve at its beginning and ending points. This means that
$\rho_k=\alpha_k f(\theta_k,\phi\k)$ and $\sigma\k=\beta\k f(\phi\k,\theta_k)$,
where $f(\theta,\phi)$ is \MF's standard velocity function defined in
the |velocity| subroutine. The cubic spline $B(z_k^{\phantom+},z_k^+,
z\k^-,z\k^{\phantom+};t)$
has curvature
@^curvature@>
$${2\sigma\k\sin(\theta_k+\phi\k)-6\sin\theta_k\over\rho_k^2d_{k,k+1}}
\qquad{\rm and}\qquad
{2\rho_k\sin(\theta_k+\phi\k)-6\sin\phi\k\over\sigma\k^2d_{k,k+1}}$$
at |t=0| and |t=1|, respectively. The mock curvature is the linear
@^mock curvature@>
approximation to this true curvature that arises in the limit for
small $\theta_k$ and~$\phi\k$, if second-order terms are discarded.
The standard velocity function satisfies
$$f(\theta,\phi)=1+O(\theta^2+\theta\phi+\phi^2);$$
hence the mock curvatures are respectively
$${2\beta\k(\theta_k+\phi\k)-6\theta_k\over\alpha_k^2d_{k,k+1}}
\qquad{\rm and}\qquad
{2\alpha_k(\theta_k+\phi\k)-6\phi\k\over\beta\k^2d_{k,k+1}}.\eqno(**)$$

@ The turning angles $\psi_k$ are given, and equation $(*)$ above
determines $\phi_k$ when $\theta_k$ is known, so the task of
angle selection is essentially to choose appropriate values for each
$\theta_k$. When equation~$(*)$ is used to eliminate $\phi$~variables
from $(**)$, we obtain a system of linear equations of the form
$$A_k\theta_{k-1}+(B_k+C_k)\theta_k+D_k\theta\k=-B_k\psi_k-D_k\psi\k,$$
where
$$A_k={\alpha_{k-1}\over\beta_k^2d_{k-1,k}},
\qquad B_k={3-\alpha_{k-1}\over\beta_k^2d_{k-1,k}},
\qquad C_k={3-\beta\k\over\alpha_k^2d_{k,k+1}},
\qquad D_k={\beta\k\over\alpha_k^2d_{k,k+1}}.$$
The tensions are always $3\over4$ or more, hence each $\alpha$ and~$\beta$
will be at most $4\over3$. It follows that $B_k\G{5\over4}A_k$ and
$C_k\G{5\over4}D_k$; hence the equations are diagonally dominant;
hence they have a unique solution. Moreover, in most cases the tensions
are equal to~1, so that $B_k=2A_k$ and $C_k=2D_k$. This makes the
solution numerically stable, and there is an exponential damping
effect: The data at knot $k\pm j$ affects the angle at knot~$k$ by
a factor of~$O(2^{-j})$.

@ However, we still must consider the angles at the starting and ending
knots of a non-cyclic path. These angles might be given explicitly, or
they might be specified implicitly in terms of an amount of ``curl.''

Let's assume that angles need to be determined for a non-cyclic path
starting at $z_0$ and ending at~$z_n$. Then equations of the form
$$A_k\theta_{k-1}+(B_k+C_k)\theta_k+D_k\theta_{k+1}=R_k$$
have been given for $0<k<n$, and it will be convenient to introduce
equations of the same form for $k=0$ and $k=n$, where
$$A_0=B_0=C_n=D_n=0.$$
If $\theta_0$ is supposed to have a given value $E_0$, we simply
define $C_0=1$, $D_0=0$, and $R_0=E_0$. Otherwise a curl
parameter, $\gamma_0$, has been specified at~$z_0$; this means
that the mock curvature at $z_0$ should be $\gamma_0$ times the
mock curvature at $z_1$; i.e.,
$${2\beta_1(\theta_0+\phi_1)-6\theta_0\over\alpha_0^2d_{01}}
=\gamma_0{2\alpha_0(\theta_0+\phi_1)-6\phi_1\over\beta_1^2d_{01}}.$$
This equation simplifies to
$$(\alpha_0\chi_0+3-\beta_1)\theta_0+
 \bigl((3-\alpha_0)\chi_0+\beta_1\bigr)\theta_1=
 -\bigl((3-\alpha_0)\chi_0+\beta_1\bigr)\psi_1,$$
where $\chi_0=\alpha_0^2\gamma_0/\beta_1^2$; so we can set $C_0=
\chi_0\alpha_0+3-\beta_1$, $D_0=(3-\alpha_0)\chi_0+\beta_1$, $R_0=-D_0\psi_1$.
It can be shown that $C_0>0$ and $C_0B_1-A_1D_0>0$ when $\gamma_0\G0$,
hence the linear equations remain nonsingular.

Similar considerations apply at the right end, when the final angle $\phi_n$
may or may not need to be determined. It is convenient to let $\psi_n=0$,
hence $\theta_n=-\phi_n$. We either have an explicit equation $\theta_n=E_n$,
or we have
$$\bigl((3-\beta_n)\chi_n+\alpha_{n-1}\bigr)\theta_{n-1}+
(\beta_n\chi_n+3-\alpha_{n-1})\theta_n=0,\qquad
  \chi_n={\beta_n^2\gamma_n\over\alpha_{n-1}^2}.$$

When |make_choices| chooses angles, it must compute the coefficients of
these linear equations, then solve the equations. To compute the coefficients,
it is necessary to compute arctangents of the given turning angles~$\psi_k$.
When the equations are solved, the chosen directions $\theta_k$ are put
back into the form of control points by essentially computing sines and
cosines.

@ OK, we are ready to make the hard choices of |make_choices|.
Most of the work is relegated to an auxiliary procedure
called |solve_choices|, which has been introduced to keep
|make_choices| from being extremely long.

@<Fill in the control information between...@>=
@<Calculate the turning angles $\psi_k$ and the distances $d_{k,k+1}$;
  set $n$ to the length of the path@>;
@<Remove |open| types at the breakpoints@>;
solve_choices(p,q,n)

@ It's convenient to precompute quantities that will be needed several
times later. The values of |delta_x[k]| and |delta_y[k]| will be the
coordinates of $z\k-z_k$, and the magnitude of this vector will be
|delta[k]=@t$d_{k,k+1}$@>|. The path angle $\psi_k$ between $z_k-z_{k-1}$
and $z\k-z_k$ will be stored in |psi[k]|.

@<Glob...@>=
@!delta_x,@!delta_y,@!delta:array[0..path_size] of scaled; {knot differences}
@!psi:array[1..path_size] of angle; {turning angles}

@ @<Other local variables for |make_choices|@>=
@!k,@!n:0..path_size; {current and final knot numbers}
@!s,@!t:pointer; {registers for list traversal}
@!delx,@!dely:scaled; {directions where |open| meets |explicit|}
@!sine,@!cosine:fraction; {trig functions of various angles}

@ @<Calculate the turning angles...@>=
k:=0; s:=p; n:=path_size;
repeat t:=link(s);
delta_x[k]:=x_coord(t)-x_coord(s);
delta_y[k]:=y_coord(t)-y_coord(s);
delta[k]:=pyth_add(delta_x[k],delta_y[k]);
if k>0 then
  begin sine:=make_fraction(delta_y[k-1],delta[k-1]);
  cosine:=make_fraction(delta_x[k-1],delta[k-1]);
  psi[k]:=n_arg(take_fraction(delta_x[k],cosine)+
      take_fraction(delta_y[k],sine),
    take_fraction(delta_y[k],cosine)-
      take_fraction(delta_x[k],sine));
  end;
@:METAFONT capacity exceeded path size}{\quad path size@>
incr(k); s:=t;
if k=path_size then overflow("path size",path_size);
if s=q then n:=k;
until (k>=n)and(left_type(s)<>end_cycle);
if k=n then psi[n]:=0@+else psi[k]:=psi[1]

@ When we get to this point of the code, |right_type(p)| is either
|given| or |curl| or |open|. If it is |open|, we must have
|left_type(p)=end_cycle| or |left_type(p)=explicit|. In the latter
case, the |open| type is converted to |given|; however, if the
velocity coming into this knot is zero, the |open| type is
converted to a |curl|, since we don't know the incoming direction.

Similarly, |left_type(q)| is either |given| or |curl| or |open| or
|end_cycle|. The |open| possibility is reduced either to |given| or to |curl|.

@<Remove |open| types at the breakpoints@>=
if left_type(q)=open then
  begin delx:=right_x(q)-x_coord(q); dely:=right_y(q)-y_coord(q);
  if (delx=0)and(dely=0) then
    begin left_type(q):=curl; left_curl(q):=unity;
    end
  else  begin left_type(q):=given; left_given(q):=n_arg(delx,dely);
    end;
  end;
if (right_type(p)=open)and(left_type(p)=explicit) then
  begin delx:=x_coord(p)-left_x(p); dely:=y_coord(p)-left_y(p);
  if (delx=0)and(dely=0) then
    begin right_type(p):=curl; right_curl(p):=unity;
    end
  else  begin right_type(p):=given; right_given(p):=n_arg(delx,dely);
    end;
  end

@ Linear equations need to be solved whenever |n>1|; and also when |n=1|
and exactly one of the breakpoints involves a curl. The simplest case occurs
when |n=1| and there is a curl at both breakpoints; then we simply draw
a straight line.

But before coding up the simple cases, we might as well face the general case,
since we must deal with it sooner or later, and since the general case
is likely to give some insight into the way simple cases can be handled best.

When there is no cycle, the linear equations to be solved form a tri-diagonal
system, and we can apply the standard technique of Gaussian elimination
to convert that system to a sequence of equations of the form
$$\theta_0+u_0\theta_1=v_0,\quad
\theta_1+u_1\theta_2=v_1,\quad\ldots,\quad
\theta_{n-1}+u_{n-1}\theta_n=v_{n-1},\quad
\theta_n=v_n.$$
It is possible to do this diagonalization while generating the equations.
Once $\theta_n$ is known, it is easy to determine $\theta_{n-1}$, \dots,
$\theta_1$, $\theta_0$; thus, the equations will be solved.

The procedure is slightly more complex when there is a cycle, but the
basic idea will be nearly the same. In the cyclic case the right-hand
sides will be $v_k+w_k\theta_0$ instead of simply $v_k$, and we will start
the process off with $u_0=v_0=0$, $w_0=1$. The final equation will be not
$\theta_n=v_n$ but $\theta_n+u_n\theta_1=v_n+w_n\theta_0$; an appropriate
ending routine will take account of the fact that $\theta_n=\theta_0$ and
eliminate the $w$'s from the system, after which the solution can be
obtained as before.

When $u_k$, $v_k$, and $w_k$ are being computed, the three pointer
variables |r|, |s|,~|t| will point respectively to knots |k-1|, |k|,
and~|k+1|. The $u$'s and $w$'s are scaled by $2^{28}$, i.e., they are
of type |fraction|; the $\theta$'s and $v$'s are of type |angle|.

@<Glob...@>=
@!theta:array[0..path_size] of angle; {values of $\theta_k$}
@!uu:array[0..path_size] of fraction; {values of $u_k$}
@!vv:array[0..path_size] of angle; {values of $v_k$}
@!ww:array[0..path_size] of fraction; {values of $w_k$}

@ Our immediate problem is to get the ball rolling by setting up the
first equation or by realizing that no equations are needed, and to fit
this initialization into a framework suitable for the overall computation.

@<Declare the procedure called |solve_choices|@>=
@t\4@>@<Declare subroutines needed by |solve_choices|@>@;
procedure solve_choices(@!p,@!q:pointer;@!n:halfword);
label found,exit;
var @!k:0..path_size; {current knot number}
@!r,@!s,@!t:pointer; {registers for list traversal}
@<Other local variables for |solve_choices|@>@;
begin k:=0; s:=p;
loop@+  begin t:=link(s);
  if k=0 then @<Get the linear equations started; or |return|
    with the control points in place, if linear equations
    needn't be solved@>
  else  case left_type(s) of
    end_cycle,open:@<Set up equation to match mock curvatures
      at $z_k$; then |goto found| with $\theta_n$
      adjusted to equal $\theta_0$, if a cycle has ended@>;
    curl:@<Set up equation for a curl at $\theta_n$
      and |goto found|@>;
    given:@<Calculate the given value of $\theta_n$
      and |goto found|@>;
    end; {there are no other cases}
  r:=s; s:=t; incr(k);
  end;
found:@<Finish choosing angles and assigning control points@>;
exit:end;

@ On the first time through the loop, we have |k=0| and |r| is not yet
defined. The first linear equation, if any, will have $A_0=B_0=0$.

@<Get the linear equations started...@>=
case right_type(s) of
given: if left_type(t)=given then @<Reduce to simple case of two givens
    and |return|@>
  else @<Set up the equation for a given value of $\theta_0$@>;
curl: if left_type(t)=curl then @<Reduce to simple case of straight line
    and |return|@>
  else @<Set up the equation for a curl at $\theta_0$@>;
open: begin uu[0]:=0; vv[0]:=0; ww[0]:=fraction_one;
  end; {this begins a cycle}
end {there are no other cases}

@ The general equation that specifies equality of mock curvature at $z_k$ is
$$A_k\theta_{k-1}+(B_k+C_k)\theta_k+D_k\theta\k=-B_k\psi_k-D_k\psi\k,$$
as derived above. We want to combine this with the already-derived equation
$\theta_{k-1}+u_{k-1}\theta_k=v_{k-1}+w_{k-1}\theta_0$ in order to obtain
a new equation
$\theta_k+u_k\theta\k=v_k+w_k\theta_0$. This can be done by dividing the
equation
$$(B_k-u_{k-1}A_k+C_k)\theta_k+D_k\theta\k=-B_k\psi_k-D_k\psi\k-A_kv_{k-1}
    -A_kw_{k-1}\theta_0$$
by $B_k-u_{k-1}A_k+C_k$. The trick is to do this carefully with
fixed-point arithmetic, avoiding the chance of overflow while retaining
suitable precision.

The calculations will be performed in several registers that
provide temporary storage for intermediate quantities.

@<Other local variables for |solve_choices|@>=
@!aa,@!bb,@!cc,@!ff,@!acc:fraction; {temporary registers}
@!dd,@!ee:scaled; {likewise, but |scaled|}
@!lt,@!rt:scaled; {tension values}

@ @<Set up equation to match mock curvatures...@>=
begin @<Calculate the values $\\{aa}=A_k/B_k$, $\\{bb}=D_k/C_k$,
  $\\{dd}=(3-\alpha_{k-1})d_{k,k+1}$, $\\{ee}=(3-\beta\k)d_{k-1,k}$,
  and $\\{cc}=(B_k-u_{k-1}A_k)/B_k$@>;
@<Calculate the ratio $\\{ff}=C_k/(C_k+B_k-u_{k-1}A_k)$@>;
uu[k]:=take_fraction(ff,bb);
@<Calculate the values of $v_k$ and $w_k$@>;
if left_type(s)=end_cycle then
  @<Adjust $\theta_n$ to equal $\theta_0$ and |goto found|@>;
end

@ Since tension values are never less than 3/4, the values |aa| and
|bb| computed here are never more than 4/5.

@<Calculate the values $\\{aa}=...@>=
if abs(right_tension(r))=unity then
  begin aa:=fraction_half; dd:=2*delta[k];
  end
else  begin aa:=make_fraction(unity,3*abs(right_tension(r))-unity);
  dd:=take_fraction(delta[k],
    fraction_three-make_fraction(unity,abs(right_tension(r))));
  end;
if abs(left_tension(t))=unity then
  begin bb:=fraction_half; ee:=2*delta[k-1];
  end
else  begin bb:=make_fraction(unity,3*abs(left_tension(t))-unity);
  ee:=take_fraction(delta[k-1],
    fraction_three-make_fraction(unity,abs(left_tension(t))));
  end;
cc:=fraction_one-take_fraction(uu[k-1],aa)

@ The ratio to be calculated in this step can be written in the form
$$\beta_k^2\cdot\\{ee}\over\beta_k^2\cdot\\{ee}+\alpha_k^2\cdot
  \\{cc}\cdot\\{dd},$$
because of the quantities just calculated. The values of |dd| and |ee|
will not be needed after this step has been performed.

@<Calculate the ratio $\\{ff}=C_k/(C_k+B_k-u_{k-1}A_k)$@>=
dd:=take_fraction(dd,cc); lt:=abs(left_tension(s)); rt:=abs(right_tension(s));
if lt<>rt then {$\beta_k^{-1}\ne\alpha_k^{-1}$}
  if lt<rt then
    begin ff:=make_fraction(lt,rt);
    ff:=take_fraction(ff,ff); {$\alpha_k^2/\beta_k^2$}
    dd:=take_fraction(dd,ff);
    end
  else  begin ff:=make_fraction(rt,lt);
    ff:=take_fraction(ff,ff); {$\beta_k^2/\alpha_k^2$}
    ee:=take_fraction(ee,ff);
    end;
ff:=make_fraction(ee,ee+dd)

@ The value of $u_{k-1}$ will be |<=1| except when $k=1$ and the previous
equation was specified by a curl. In that case we must use a special
method of computation to prevent overflow.

Fortunately, the calculations turn out to be even simpler in this ``hard''
case. The curl equation makes $w_0=0$ and $v_0=-u_0\psi_1$, hence
$-B_1\psi_1-A_1v_0=-(B_1-u_0A_1)\psi_1=-\\{cc}\cdot B_1\psi_1$.

@<Calculate the values of $v_k$ and $w_k$@>=
acc:=-take_fraction(psi[k+1],uu[k]);
if right_type(r)=curl then
  begin ww[k]:=0;
  vv[k]:=acc-take_fraction(psi[1],fraction_one-ff);
  end
else  begin ff:=make_fraction(fraction_one-ff,cc); {this is
    $B_k/(C_k+B_k-u_{k-1}A_k)<5$}
  acc:=acc-take_fraction(psi[k],ff);
  ff:=take_fraction(ff,aa); {this is $A_k/(C_k+B_k-u_{k-1}A_k)$}
  vv[k]:=acc-take_fraction(vv[k-1],ff);
  if ww[k-1]=0 then ww[k]:=0
  else ww[k]:=-take_fraction(ww[k-1],ff);
  end

@ When a complete cycle has been traversed, we have $\theta_k+u_k\theta\k=
v_k+w_k\theta_0$, for |1<=k<=n|. We would like to determine the value of
$\theta_n$ and reduce the system to the form $\theta_k+u_k\theta\k=v_k$
for |0<=k<n|, so that the cyclic case can be finished up just as if there
were no cycle.

The idea in the following code is to observe that
$$\eqalign{\theta_n&=v_n+w_n\theta_0-u_n\theta_1=\cdots\cr
&=v_n+w_n\theta_0-u_n\bigl(v_1+w_1\theta_0-u_1(v_2+\cdots
  -u_{n-2}(v_{n-1}+w_{n-1}\theta_0-u_{n-1}\theta_0)\ldots{})\bigr),\cr}$$
so we can solve for $\theta_n=\theta_0$.

@<Adjust $\theta_n$ to equal $\theta_0$ and |goto found|@>=
begin aa:=0; bb:=fraction_one; {we have |k=n|}
repeat decr(k);
if k=0 then k:=n;
aa:=vv[k]-take_fraction(aa,uu[k]);
bb:=ww[k]-take_fraction(bb,uu[k]);
until k=n; {now $\theta_n=\\{aa}+\\{bb}\cdot\theta_n$}
aa:=make_fraction(aa,fraction_one-bb);
theta[n]:=aa; vv[0]:=aa;
for k:=1 to n-1 do vv[k]:=vv[k]+take_fraction(aa,ww[k]);
goto found;
end

@ @d reduce_angle(#)==if abs(#)>one_eighty_deg then
  if #>0 then #:=#-three_sixty_deg@+else #:=#+three_sixty_deg

@<Calculate the given value of $\theta_n$...@>=
begin theta[n]:=left_given(s)-n_arg(delta_x[n-1],delta_y[n-1]);
reduce_angle(theta[n]);
goto found;
end

@ @<Set up the equation for a given value of $\theta_0$@>=
begin vv[0]:=right_given(s)-n_arg(delta_x[0],delta_y[0]);
reduce_angle(vv[0]);
uu[0]:=0; ww[0]:=0;
end

@ @<Set up the equation for a curl at $\theta_0$@>=
begin cc:=right_curl(s); lt:=abs(left_tension(t)); rt:=abs(right_tension(s));
if (rt=unity)and(lt=unity) then
  uu[0]:=make_fraction(cc+cc+unity,cc+two)
else uu[0]:=curl_ratio(cc,rt,lt);
vv[0]:=-take_fraction(psi[1],uu[0]); ww[0]:=0;
end

@ @<Set up equation for a curl at $\theta_n$...@>=
begin cc:=left_curl(s); lt:=abs(left_tension(s)); rt:=abs(right_tension(r));
if (rt=unity)and(lt=unity) then
  ff:=make_fraction(cc+cc+unity,cc+two)
else ff:=curl_ratio(cc,lt,rt);
theta[n]:=-make_fraction(take_fraction(vv[n-1],ff),
    fraction_one-take_fraction(ff,uu[n-1]));
goto found;
end

@ The |curl_ratio| subroutine has three arguments, which our previous notation
encourages us to call $\gamma$, $\alpha^{-1}$, and $\beta^{-1}$. It is
a somewhat tedious program to calculate
$${(3-\alpha)\alpha^2\gamma+\beta^3\over
  \alpha^3\gamma+(3-\beta)\beta^2},$$
with the result reduced to 4 if it exceeds 4. (This reduction of curl
is necessary only if the curl and tension are both large.)
The values of $\alpha$ and $\beta$ will be at most~4/3.

@<Declare subroutines needed by |solve_choices|@>=
function curl_ratio(@!gamma,@!a_tension,@!b_tension:scaled):fraction;
var @!alpha,@!beta,@!num,@!denom,@!ff:fraction; {registers}
begin alpha:=make_fraction(unity,a_tension);
beta:=make_fraction(unity,b_tension);@/
if alpha<=beta then
  begin ff:=make_fraction(alpha,beta); ff:=take_fraction(ff,ff);
  gamma:=take_fraction(gamma,ff);@/
  beta:=beta div @'10000; {convert |fraction| to |scaled|}
  denom:=take_fraction(gamma,alpha)+three-beta;
  num:=take_fraction(gamma,fraction_three-alpha)+beta;
  end
else  begin ff:=make_fraction(beta,alpha); ff:=take_fraction(ff,ff);
  beta:=take_fraction(beta,ff) div @'10000; {convert |fraction| to |scaled|}
  denom:=take_fraction(gamma,alpha)+(ff div 1365)-beta;
    {$1365\approx 2^{12}/3$}
  num:=take_fraction(gamma,fraction_three-alpha)+beta;
  end;
if num>=denom+denom+denom+denom then curl_ratio:=fraction_four
else curl_ratio:=make_fraction(num,denom);
end;

@ We're in the home stretch now.

@<Finish choosing angles and assigning control points@>=
for k:=n-1 downto 0 do theta[k]:=vv[k]-take_fraction(theta[k+1],uu[k]);
s:=p; k:=0;
repeat t:=link(s);@/
n_sin_cos(theta[k]); st:=n_sin; ct:=n_cos;@/
n_sin_cos(-psi[k+1]-theta[k+1]); sf:=n_sin; cf:=n_cos;@/
set_controls(s,t,k);@/
incr(k); s:=t;
until k=n

@ The |set_controls| routine actually puts the control points into
a pair of consecutive nodes |p| and~|q|. Global variables are used to
record the values of $\sin\theta$, $\cos\theta$, $\sin\phi$, and
$\cos\phi$ needed in this calculation.

@<Glob...@>=
@!st,@!ct,@!sf,@!cf:fraction; {sines and cosines}

@ @<Declare subroutines needed by |solve_choices|@>=
procedure set_controls(@!p,@!q:pointer;@!k:integer);
var @!rr,@!ss:fraction; {velocities, divided by thrice the tension}
@!lt,@!rt:scaled; {tensions}
@!sine:fraction; {$\sin(\theta+\phi)$}
begin lt:=abs(left_tension(q)); rt:=abs(right_tension(p));
rr:=velocity(st,ct,sf,cf,rt);
ss:=velocity(sf,cf,st,ct,lt);
if (right_tension(p)<0)or(left_tension(q)<0) then @<Decrease the velocities,
  if necessary, to stay inside the bounding triangle@>;
right_x(p):=x_coord(p)+take_fraction(
  take_fraction(delta_x[k],ct)-take_fraction(delta_y[k],st),rr);
right_y(p):=y_coord(p)+take_fraction(
  take_fraction(delta_y[k],ct)+take_fraction(delta_x[k],st),rr);
left_x(q):=x_coord(q)-take_fraction(
  take_fraction(delta_x[k],cf)+take_fraction(delta_y[k],sf),ss);
left_y(q):=y_coord(q)-take_fraction(
  take_fraction(delta_y[k],cf)-take_fraction(delta_x[k],sf),ss);
right_type(p):=explicit; left_type(q):=explicit;
end;

@ The boundedness conditions $\\{rr}\L\sin\phi\,/\sin(\theta+\phi)$ and
$\\{ss}\L\sin\theta\,/\sin(\theta+\phi)$ are to be enforced if $\sin\theta$,
$\sin\phi$, and $\sin(\theta+\phi)$ all have the same sign. Otherwise
there is no ``bounding triangle.''

@<Decrease the velocities, if necessary...@>=
if((st>=0)and(sf>=0))or((st<=0)and(sf<=0)) then
  begin sine:=take_fraction(abs(st),cf)+take_fraction(abs(sf),ct);
  if sine>0 then
    begin sine:=take_fraction(sine,fraction_one+unity); {safety factor}
    if right_tension(p)<0 then
     if ab_vs_cd(abs(sf),fraction_one,rr,sine)<0 then
      rr:=make_fraction(abs(sf),sine);
    if left_tension(q)<0 then
     if ab_vs_cd(abs(st),fraction_one,ss,sine)<0 then
      ss:=make_fraction(abs(st),sine);
    end;
  end

@ Only the simple cases remain to be handled.

@<Reduce to simple case of two givens and |return|@>=
begin aa:=n_arg(delta_x[0],delta_y[0]);@/
n_sin_cos(right_given(p)-aa); ct:=n_cos; st:=n_sin;@/
n_sin_cos(left_given(q)-aa); cf:=n_cos; sf:=-n_sin;@/
set_controls(p,q,0); return;
end

@ @<Reduce to simple case of straight line and |return|@>=
begin right_type(p):=explicit; left_type(q):=explicit;
lt:=abs(left_tension(q)); rt:=abs(right_tension(p));
if rt=unity then
  begin if delta_x[0]>=0 then right_x(p):=x_coord(p)+((delta_x[0]+1) div 3)
  else right_x(p):=x_coord(p)+((delta_x[0]-1) div 3);
  if delta_y[0]>=0 then right_y(p):=y_coord(p)+((delta_y[0]+1) div 3)
  else right_y(p):=y_coord(p)+((delta_y[0]-1) div 3);
  end
else  begin ff:=make_fraction(unity,3*rt); {$\alpha/3$}
  right_x(p):=x_coord(p)+take_fraction(delta_x[0],ff);
  right_y(p):=y_coord(p)+take_fraction(delta_y[0],ff);
  end;
if lt=unity then
  begin if delta_x[0]>=0 then left_x(q):=x_coord(q)-((delta_x[0]+1) div 3)
  else left_x(q):=x_coord(q)-((delta_x[0]-1) div 3);
  if delta_y[0]>=0 then left_y(q):=y_coord(q)-((delta_y[0]+1) div 3)
  else left_y(q):=y_coord(q)-((delta_y[0]-1) div 3);
  end
else  begin ff:=make_fraction(unity,3*lt); {$\beta/3$}
  left_x(q):=x_coord(q)-take_fraction(delta_x[0],ff);
  left_y(q):=y_coord(q)-take_fraction(delta_y[0],ff);
  end;
return;
end

@* \[19] Generating discrete moves.
The purpose of the next part of \MF\ is to compute discrete approximations
to curves described as parametric polynomial functions $z(t)$.
We shall start with the low level first, because an efficient ``engine''
is needed to support the high-level constructions.

Most of the subroutines are based on variations of a single theme,
namely the idea of {\sl bisection}. Given a Bernshte{\u\i}n polynomial
@^Bernshte{\u\i}n, Serge{\u\i} Natanovich@>
$$B(z_0,z_1,\ldots,z_n;t)=\sum_k{n\choose k}t^k(1-t)^{n-k}z_k,$$
we can conveniently bisect its range as follows:

\smallskip
\textindent{1)} Let $z_k^{(0)}=z_k$, for |0<=k<=n|.

\smallskip
\textindent{2)} Let $z_k^{(j+1)}={1\over2}(z_k^{(j)}+z\k^{(j)})$, for
|0<=k<n-j|, for |0<=j<n|.

\smallskip\noindent
Then
$$B(z_0,z_1,\ldots,z_n;t)=B(z_0^{(0)},z_0^{(1)},\ldots,z_0^{(n)};2t)
 =B(z_0^{(n)},z_1^{(n-1)},\ldots,z_n^{(0)};2t-1).$$
This formula gives us the coefficients of polynomials to use over the ranges
$0\L t\L{1\over2}$ and ${1\over2}\L t\L1$.

In our applications it will usually be possible to work indirectly with
numbers that allow us to deduce relevant properties of the polynomials
without actually computing the polynomial values. We will deal with
coefficients $Z_k=2^l(z_k-z_{k-1})$ for |1<=k<=n|, instead of
the actual numbers $z_0$, $z_1$, \dots,~$z_n$, and the value of~|l| will
increase by~1 at each bisection step. This technique reduces the
amount of calculation needed for bisection and also increases the
accuracy of evaluation (since one bit of precision is gained at each
bisection). Indeed, the bisection process now becomes one level shorter:

\smallskip
\textindent{$1'$)} Let $Z_k^{(1)}=Z_k$, for |1<=k<=n|.

\smallskip
\textindent{$2'$)} Let $Z_k^{(j+1)}={1\over2}(Z_k^{(j)}+Z\k^{(j)})$, for
|1<=k<=n-j|, for |1<=j<n|.

\smallskip\noindent
The relevant coefficients $(Z'_1,\ldots,Z'_n)$ and $(Z''_1,\ldots,Z''_n)$
for the two subintervals after bisection are respectively
$(Z_1^{(1)},Z_1^{(2)},\ldots,Z_1^{(n)})$ and
$(Z_1^{(n)},Z_2^{(n-1)},\ldots,Z_n^{(1)})$.
And the values of $z_0$ appropriate for the bisected interval are $z'_0=z_0$
and $z''_0=z_0+(Z'_1+Z'_2+\cdots+Z'_n)/2^{l+1}$.

Step $2'$ involves division by~2, which introduces computational errors
of at most $1\over2$ at each step; thus after $l$~levels of bisection the
integers $Z_k$ will differ from their true values by at most $(n-1)l/2$.
This error rate is quite acceptable, considering that we have $l$~more
bits of precision in the $Z$'s by comparison with the~$z$'s.  Note also
that the $Z$'s remain bounded; there's no danger of integer overflow, even
though we have the identity $Z_k=2^l(z_k-z_{k-1})$ for arbitrarily large~$l$.

In fact, we can show not only that the $Z$'s remain bounded, but also that
they become nearly equal, since they are control points for a polynomial
of one less degree. If $\vert Z\k-Z_k\vert\L M$ initially, it is possible
to prove that $\vert Z\k-Z_k\vert\L\lceil M/2^l\rceil$ after $l$~levels
of bisection, even in the presence of rounding errors. Here's the
proof [cf.~Lane and Riesenfeld, {\sl IEEE Trans.\ on Pattern Analysis
@^Lane, Jeffrey Michael@>
@^Riesenfeld, Richard Franklin@>
and Machine Intelligence\/ \bf PAMI-2} (1980), 35--46]: Assuming that
$\vert Z\k-Z_k\vert\L M$ before bisection, we want to prove that
$\vert Z\k-Z_k\vert\L\lceil M/2\rceil$ afterward. First we show that
$\vert Z\k^{(j)}-Z_k^{(j)}\vert\L M$ for all $j$ and~$k$, by induction
on~$j$; this follows from the fact that
$$\bigl\vert\\{half}(a+b)-\\{half}(b+c)\bigr\vert\L
 \max\bigl(\vert a-b\vert,\vert b-c\vert\bigr)$$
holds for both of the rounding rules $\\{half}(x)=\lfloor x/2\rfloor$
and $\\{half}(x)={\rm sign}(x)\lfloor\vert x/2\vert\rfloor$.
(If $\vert a-b\vert$ and $\vert b-c\vert$ are equal, then
$a+b$ and $b+c$ are both even or both odd. The rounding errors either
cancel or round the numbers toward each other; hence
$$\eqalign{\bigl\vert\\{half}(a+b)-\\{half}(b+c)\bigr\vert
&\L\textstyle\bigl\vert{1\over2}(a+b)-{1\over2}(b+c)\bigr\vert\cr
&=\textstyle\bigl\vert{1\over2}(a-b)+{1\over2}(b-c)\bigr\vert
\L\max\bigl(\vert a-b\vert,\vert b-c\vert\bigr),\cr}$$
as required. A simpler argument applies if $\vert a-b\vert$ and
$\vert b-c\vert$ are unequal.)  Now it is easy to see that
$\vert Z_1^{(j+1)}-Z_1^{(j)}\vert\L\bigl\lfloor{1\over2}
\vert Z_2^{(j)}-Z_1^{(j)}\vert+{1\over2}\bigr\rfloor
\L\bigl\lfloor{1\over2}(M+1)\bigr\rfloor=\lceil M/2\rceil$.

Another interesting fact about bisection is the identity
$$Z_1'+\cdots+Z_n'+Z_1''+\cdots+Z_n''=2(Z_1+\cdots+Z_n+E),$$
where $E$ is the sum of the rounding errors in all of the halving
operations ($\vert E\vert\L n(n-1)/4$).

@ We will later reduce the problem of digitizing a complex cubic
$z(t)=B(z_0,z_1,z_2,z_3;t)$ to the following simpler problem:
Given two real cubics
$x(t)=B(x_0,x_1,x_2,x_3;t)$
and $y(t)=B(y_0,y_1,y_2,y_3;t)$ that are monotone nondecreasing,
determine the set of integer points
$$P=\bigl\{\bigl(\lfloor x(t)\rfloor,\lfloor y(t)\rfloor\bigr)
\bigm\vert 0\L t\L 1\bigr\}.$$
Well, the problem isn't actually quite so clean as this; when the path
goes very near an integer point $(a,b)$, computational errors may
make us think that $P$ contains $(a-1,b)$ while in reality it should
contain $(a,b-1)$. Furthermore, if the path goes {\sl exactly\/}
through the integer points $(a-1,b-1)$ and
$(a,b)$, we will want $P$ to contain one
of the two points $(a-1,b)$ or $(a,b-1)$, so that $P$ can be described
entirely by ``rook moves'' upwards or to the right; no diagonal
moves from $(a-1,b-1)$ to~$(a,b)$ will be allowed.

Thus, the set $P$ we wish to compute will merely be an approximation
to the set described in the formula above. It will consist of
$\lfloor x(1)\rfloor-\lfloor x(0)\rfloor$ rightward moves and
$\lfloor y(1)\rfloor-\lfloor y(0)\rfloor$ upward moves, intermixed
in some order. Our job will be to figure out a suitable order.

The following recursive strategy suggests itself, when we recall that
$x(0)=x_0$, $x(1)=x_3$, $y(0)=y_0$, and $y(1)=y_3$:

\smallskip
If $\lfloor x_0\rfloor=\lfloor x_3\rfloor$ then take
$\lfloor y_3\rfloor-\lfloor y_0\rfloor$ steps up.

Otherwise if $\lfloor y_0\rfloor=\lfloor y_3\rfloor$ then take
$\lfloor x_3\rfloor-\lfloor x_0\rfloor$ steps to the right.

Otherwise bisect the current cubics and repeat the process on both halves.

\yskip\noindent
This intuitively appealing formulation does not quite solve the problem,
because it may never terminate. For example, it's not hard to see that
no steps will {\sl ever\/} be taken if $(x_0,x_1,x_2,x_3)=(y_0,y_1,y_2,y_3)$!
However, we can surmount this difficulty with a bit of care; so let's
proceed to flesh out the algorithm as stated, before worrying about
such details.

The bisect-and-double strategy discussed above suggests that we represent
$(x_0,x_1,x_2,x_3)$ by $(X_1,X_2,X_3)$, where $X_k=2^l(x_k-x_{k-1})$
for some~$l$. Initially $l=16$, since the $x$'s are |scaled|.
In order to deal with other aspects of the algorithm we will want to
maintain also the quantities $m=\lfloor x_3\rfloor-\lfloor x_0\rfloor$
and $R=2^l(x_0\bmod 1)$. Similarly,
$(y_0,y_1,y_2,y_3)$ will be represented by $(Y_1,Y_2,Y_3)$,
$n=\lfloor y_3\rfloor-\lfloor y_0\rfloor$,
and $S=2^l(y_0\bmod 1)$. The algorithm now takes the following form:

\smallskip
If $m=0$ then take $n$ steps up.

Otherwise if $n=0$ then take $m$ steps to the right.

Otherwise bisect the current cubics and repeat the process on both halves.

\smallskip\noindent
The bisection process for $(X_1,X_2,X_3,m,R,l)$ reduces, in essence,
to the following formulas:
$$\vbox{\halign{$#\hfil$\cr
X_2'=\\{half}(X_1+X_2),\quad
X_2''=\\{half}(X_2+X_3),\quad
X_3'=\\{half}(X_2'+X_2''),\cr
X_1'=X_1,\quad
X_1''=X_3',\quad
X_3''=X_3,\cr
R'=2R,\quad
T=X_1'+X_2'+X_3'+R',\quad
R''=T\bmod 2^{l+1},\cr
m'=\lfloor T/2^{l+1}\rfloor,\quad
m''=m-m'.\cr}}$$

@ When $m=n=1$, the computation can be speeded up because we simply
need to decide between two alternatives, (up,\thinspace right)
versus (right,\thinspace up). There appears to be no simple, direct
way to make the correct decision by looking at the values of
$(X_1,X_2,X_3,R)$ and
$(Y_1,Y_2,Y_3,S)$; but we can streamline the bisection process, and
we can use the fact that only one of the two descendants needs to
be examined after each bisection. Furthermore, we observed earlier
that after several levels of bisection the $X$'s and $Y$'s will be nearly
equal; so we will be justified in assuming that the curve is essentially a
straight line. (This, incidentally, solves the problem of infinite
recursion mentioned earlier.)

It is possible to show that
$$m=\bigl\lfloor(X_1+X_2+X_3+R+E)\,/\,2^l\bigr\rfloor,$$
where $E$ is an accumulated rounding error that is at most
$3\cdot(2^{l-16}-1)$ in absolute value. We will make sure that
the $X$'s are less than $2^{28}$; hence when $l=30$ we must
have |m<=1|. This proves that the special case $m=n=1$ is
bound to be reached by the time $l=30$. Furthermore $l=30$ is
a suitable time to make the straight line approximation,
if the recursion hasn't already died out, because the maximum
difference between $X$'s will then be $<2^{14}$; this corresponds
to an error of $<1$ with respect to the original scaling.
(Stating this another way, each bisection makes the curve two bits
closer to a straight line, hence 14 bisections are sufficient for
28-bit accuracy.)

In the case of a straight line, the curve goes first right, then up,
if and only if $(T-2^l)(2^l-S)>(U-2^l)(2^l-R)$, where
$T=X_1+X_2+X_3+R$ and $U=Y_1+Y_2+Y_3+S$. For the actual curve
essentially runs from $(R/2^l,S/2^l)$ to $(T/2^l,U/2^l)$, and
we are testing whether or not $(1,1)$ is above the straight
line connecting these two points. (This formula assumes that $(1,1)$
is not exactly on the line.)

@ We have glossed over the problem of tie-breaking in ambiguous
cases when the cubic curve passes exactly through integer points.
\MF\ finesses this problem by assuming that coordinates
$(x,y)$ actually stand for slightly perturbed values $(x+\xi,y+\eta)$,
where $\xi$ and~$\eta$ are infinitesimals whose signs will determine
what to do when $x$ and/or~$y$ are exact integers. The quantities
$\lfloor x\rfloor$ and~$\lfloor y\rfloor$ in the formulas above
should actually read $\lfloor x+\xi\rfloor$ and $\lfloor y+\eta\rfloor$.

If $x$ is a |scaled| value, we have $\lfloor x+\xi\rfloor=\lfloor x\rfloor$
if $\xi>0$, and $\lfloor x+\xi\rfloor=\lfloor x-2^{-16}\rfloor$ if
$\xi<0$. It is convenient to represent $\xi$ by the integer |xi_corr|,
defined to be 0~if $\xi>0$ and 1~if $\xi<0$; then, for example, the
integer $\lfloor x+\xi\rfloor$ can be computed as
|floor_unscaled(x-xi_corr)|. Similarly, $\eta$ is conveniently
represented by~|eta_corr|.

In our applications the sign of $\xi-\eta$ will always be the same as
the sign of $\xi$. Therefore it turns out that the rule for straight
lines, as stated above, should be modified as follows in the case of
ties: The line goes first right, then up, if and only if
$(T-2^l)(2^l-S)+\xi>(U-2^l)(2^l-R)$. And this relation holds iff
$|ab_vs_cd|(T-2^l,2^l-S,U-2^l,2^l-R)-|xi_corr|\ge0$.

These conventions for rounding are symmetrical, in the sense that the
digitized moves obtained from $(x_0,x_1,x_2,x_3,y_0,y_1,y_2,y_3,\xi,\eta)$
will be exactly complementary to the moves that would be obtained from
$(-x_3,-x_2,-x_1,-x_0,-y_3,-y_2,-y_1,-y_0,-\xi,-\eta)$, if arithmetic
is exact. However, truncation errors in the bisection process might
upset the symmetry. We can restore much of the lost symmetry by adding
|xi_corr| or |eta_corr| when halving the data.

@ One further possibility needs to be mentioned: The algorithm
will be applied only to cubic polynomials $B(x_0,x_1,x_2,x_3;t)$ that
are nondecreasing as $t$~varies from 0 to~1; this condition turns
out to hold if and only if $x_0\L x_1$ and $x_2\L x_3$, and either
$x_1\L x_2$ or $(x_1-x_2)^2\L(x_1-x_0)(x_3-x_2)$. If bisection were
carried out with perfect accuracy, these relations would remain
invariant. But rounding errors can creep in, hence the bisection
algorithm can produce non-monotonic subproblems from monotonic
initial conditions. This leads to the potential danger that $m$ or~$n$
could become negative in the algorithm described above.

For example, if we start with $(x_1-x_0,x_2-x_1,x_3-x_2)=
(X_1,X_2,X_3)=(7,-16,39)$, the corresponding polynomial is
monotonic, because $16^2<7\cdot39$. But the bisection algorithm
produces the left descendant $(7,-5,3)$, which is nonmonotonic;
its right descendant is~$(0,-1,3)$.

\def\xt{{\tilde x}}
Fortunately we can prove that such rounding errors will never cause
the algorithm to make a tragic mistake. At every stage we are working
with numbers corresponding to a cubic polynomial $B(\xt_0,
\xt_1,\xt_2,\xt_3)$ that approximates some
monotonic polynomial $B(x_0,x_1,x_2,x_3)$. The accumulated errors are
controlled so that $\vert x_k-\xt_k\vert<\epsilon=3\cdot2^{-16}$.
If bisection is done at some stage of the recursion, we have
$m=\lfloor\xt_3\rfloor-\lfloor\xt_0\rfloor>0$, and the algorithm
computes a bisection value $\bar x$ such that $m'=\lfloor\bar x\rfloor-
\lfloor\xt_0\rfloor$
and $m''=\lfloor\xt_3\rfloor-\lfloor\bar x\rfloor$. We want to prove
that neither $m'$ nor $m''$ can be negative. Since $\bar x$ is an
approximation to a value in the interval $[x_0,x_3]$, we have
$\bar x>x_0-\epsilon$ and $\bar x<x_3+\epsilon$, hence $\bar x>
\xt_0-2\epsilon$ and $\bar x<\xt_3+2\epsilon$.
If $m'$ is negative we must have $\xt_0\bmod 1<2\epsilon$;
if $m''$ is negative we must have $\xt_3\bmod 1>1-2\epsilon$.
In either case the condition $\lfloor\xt_3\rfloor-\lfloor\xt_0\rfloor>0$
implies that $\xt_3-\xt_0>1-2\epsilon$, hence $x_3-x_0>1-4\epsilon$.
But it can be shown that if $B(x_0,x_1,x_2,x_3;t)$ is a monotonic
cubic, then $B(x_0,x_1,x_2,x_3;{1\over2})$ is always between
$.06[x_0,x_3]$ and $.94[x_0,x_3]$; and it is impossible for $\bar x$
to be within~$\epsilon$ of such a number. Contradiction!
(The constant .06 is actually $(2-\sqrt3\,)/4$; the worst case
occurs for polynomials like $B(0,2-\sqrt3,1-\sqrt3,3;t)$.)

@ OK, now that a long theoretical preamble has justified the
bisection-and-doubling algorithm, we are ready to proceed with
its actual coding. But we still haven't discussed the
form of the output.

For reasons to be discussed later, we shall find it convenient to
record the output as follows: Moving one step up is represented by
appending a `1' to a list; moving one step right is represented by
adding unity to the element at the end of the list. Thus, for example,
the net effect of ``(up, right, right, up, right)'' is to append
$(3,2)$.

The list is kept in a global array called |move|. Before starting the
algorithm, \MF\ should check that $\\{move\_ptr}+\lfloor y_3\rfloor
-\lfloor y_0\rfloor\L\\{move\_size}$, so that the list won't exceed
the bounds of this array.

@<Glob...@>=
@!move:array[0..move_size] of integer; {the recorded moves}
@!move_ptr:0..move_size; {the number of items in the |move| list}

@ When bisection occurs, we ``push'' the subproblem corresponding
to the right-hand subinterval onto the |bisect_stack| while
we continue to work on the left-hand subinterval. Thus, the |bisect_stack|
will hold $(X_1,X_2,X_3,R,m,Y_1,Y_2,Y_3,S,n,l)$ values for
subproblems yet to be tackled.

At most 15 subproblems will be on the stack at once (namely, for
$l=15$,~16, \dots,~29); but the stack is bigger than this, because
it is used also for more complicated bisection algorithms.

@d stack_x1==bisect_stack[bisect_ptr] {stacked value of $X_1$}
@d stack_x2==bisect_stack[bisect_ptr+1] {stacked value of $X_2$}
@d stack_x3==bisect_stack[bisect_ptr+2] {stacked value of $X_3$}
@d stack_r==bisect_stack[bisect_ptr+3] {stacked value of $R$}
@d stack_m==bisect_stack[bisect_ptr+4] {stacked value of $m$}
@d stack_y1==bisect_stack[bisect_ptr+5] {stacked value of $Y_1$}
@d stack_y2==bisect_stack[bisect_ptr+6] {stacked value of $Y_2$}
@d stack_y3==bisect_stack[bisect_ptr+7] {stacked value of $Y_3$}
@d stack_s==bisect_stack[bisect_ptr+8] {stacked value of $S$}
@d stack_n==bisect_stack[bisect_ptr+9] {stacked value of $n$}
@d stack_l==bisect_stack[bisect_ptr+10] {stacked value of $l$}
@d move_increment=11 {number of items pushed by |make_moves|}

@<Glob...@>=
@!bisect_stack:array[0..bistack_size] of integer;
@!bisect_ptr:0..bistack_size;

@ @<Check the ``constant'' values...@>=
if 15*move_increment>bistack_size then bad:=31;

@ The |make_moves| subroutine is given |scaled| values $(x_0,x_1,x_2,x_3)$
and $(y_0,y_1,y_2,y_3)$ that represent monotone-nondecreasing polynomials;
it makes $\lfloor x_3+\xi\rfloor-\lfloor x_0+\xi\rfloor$ rightward moves
and $\lfloor y_3+\eta\rfloor-\lfloor y_0+\eta\rfloor$ upward moves, as
explained earlier.  (Here $\lfloor x+\xi\rfloor$ actually stands for
$\lfloor x/2^{16}-|xi_corr|\rfloor$, if $x$ is regarded as an integer
without scaling.) The unscaled integers $x_k$ and~$y_k$ should be less
than $2^{28}$ in magnitude.

It is assumed that $|move_ptr| + \lfloor y_3+\eta\rfloor -
\lfloor y_0+\eta\rfloor < |move_size|$ when this procedure is called,
so that the capacity of the |move| array will not be exceeded.

The variables |r| and |s| in this procedure stand respectively for
$R-|xi_corr|$ and $S-|eta_corr|$ in the theory discussed above.

@p procedure make_moves(@!xx0,@!xx1,@!xx2,@!xx3,@!yy0,@!yy1,@!yy2,@!yy3:
  scaled;@!xi_corr,@!eta_corr:small_number);
label continue, done, exit;
var @!x1,@!x2,@!x3,@!m,@!r,@!y1,@!y2,@!y3,@!n,@!s,@!l:integer;
  {bisection variables explained above}
@!q,@!t,@!u,@!x2a,@!x3a,@!y2a,@!y3a:integer; {additional temporary registers}
begin if (xx3<xx0)or(yy3<yy0) then confusion("m");
@:this can't happen m}{\quad m@>
l:=16; bisect_ptr:=0;@/
x1:=xx1-xx0; x2:=xx2-xx1; x3:=xx3-xx2;
if xx0>=xi_corr then r:=(xx0-xi_corr) mod unity
else r:=unity-1-((-xx0+xi_corr-1) mod unity);
m:=(xx3-xx0+r) div unity;@/
y1:=yy1-yy0; y2:=yy2-yy1; y3:=yy3-yy2;
if yy0>=eta_corr then s:=(yy0-eta_corr) mod unity
else s:=unity-1-((-yy0+eta_corr-1) mod unity);
n:=(yy3-yy0+s) div unity;@/
if (xx3-xx0>=fraction_one)or(yy3-yy0>=fraction_one) then
  @<Divide the variables by two, to avoid overflow problems@>;
loop@+  begin continue:@<Make moves for current subinterval;
    if bisection is necessary, push the second subinterval
    onto the stack, and |goto continue| in order to handle
    the first subinterval@>;
  if bisect_ptr=0 then return;
  @<Remove a subproblem for |make_moves| from the stack@>;
  end;
exit: end;

@ @<Remove a subproblem for |make_moves| from the stack@>=
bisect_ptr:=bisect_ptr-move_increment;@/
x1:=stack_x1; x2:=stack_x2; x3:=stack_x3; r:=stack_r; m:=stack_m;@/
y1:=stack_y1; y2:=stack_y2; y3:=stack_y3; s:=stack_s; n:=stack_n;@/
l:=stack_l

@ Our variables |(x1,x2,x3)| correspond to $(X_1,X_2,X_3)$ in the notation
of the theory developed above. We need to keep them less than $2^{28}$
in order to avoid integer overflow in weird circumstances.
For example, data like $x_0=-2^{28}+2^{16}-1$ and $x_1=x_2=x_3=2^{28}-1$
would otherwise be problematical. Hence this part of the code is
needed, if only to thwart malicious users.

@<Divide the variables by two, to avoid overflow problems@>=
begin x1:=half(x1+xi_corr); x2:=half(x2+xi_corr); x3:=half(x3+xi_corr);
r:=half(r+xi_corr);@/
y1:=half(y1+eta_corr); y2:=half(y2+eta_corr); y3:=half(y3+eta_corr);
s:=half(s+eta_corr);@/
l:=15;
end

@ @<Make moves...@>=
if m=0 then @<Move upward |n| steps@>
else if n=0 then @<Move to the right |m| steps@>
else if m+n=2 then @<Make one move of each kind@>
else  begin incr(l); stack_l:=l;@/
  stack_x3:=x3; stack_x2:=half(x2+x3+xi_corr); x2:=half(x1+x2+xi_corr);
  x3:=half(x2+stack_x2+xi_corr); stack_x1:=x3;@/
  r:=r+r+xi_corr; t:=x1+x2+x3+r;@/
  q:=t div two_to_the[l]; stack_r:=t mod two_to_the[l];@/
  stack_m:=m-q; m:=q;@/
  stack_y3:=y3; stack_y2:=half(y2+y3+eta_corr); y2:=half(y1+y2+eta_corr);
  y3:=half(y2+stack_y2+eta_corr); stack_y1:=y3;@/
  s:=s+s+eta_corr; u:=y1+y2+y3+s;@/
  q:=u div two_to_the[l]; stack_s:=u mod two_to_the[l];@/
  stack_n:=n-q; n:=q;@/
  bisect_ptr:=bisect_ptr+move_increment; goto continue;
  end

@ @<Move upward |n| steps@>=
while n>0 do
  begin incr(move_ptr); move[move_ptr]:=1; decr(n);
  end

@ @<Move to the right |m| steps@>=
move[move_ptr]:=move[move_ptr]+m

@ @<Make one move of each kind@>=
begin r:=two_to_the[l]-r; s:=two_to_the[l]-s;@/
while l<30 do
  begin x3a:=x3; x2a:=half(x2+x3+xi_corr); x2:=half(x1+x2+xi_corr);
  x3:=half(x2+x2a+xi_corr);
  t:=x1+x2+x3; r:=r+r-xi_corr;@/
  y3a:=y3; y2a:=half(y2+y3+eta_corr); y2:=half(y1+y2+eta_corr);
  y3:=half(y2+y2a+eta_corr);
  u:=y1+y2+y3; s:=s+s-eta_corr;@/
  if t<r then if u<s then @<Switch to the right subinterval@>
    else  begin @<Move up then right@>; goto done;
      end
  else if u<s then
    begin @<Move right then up@>; goto done;
    end;
  incr(l);
  end;
r:=r-xi_corr; s:=s-eta_corr;
if ab_vs_cd(x1+x2+x3,s,y1+y2+y3,r)-xi_corr>=0 then @<Move right then up@>
  else @<Move up then right@>;
done:
end

@ @<Switch to the right subinterval@>=
begin x1:=x3; x2:=x2a; x3:=x3a; r:=r-t;
y1:=y3; y2:=y2a; y3:=y3a; s:=s-u;
end

@ @<Move right then up@>=
begin incr(move[move_ptr]); incr(move_ptr); move[move_ptr]:=1;
end

@ @<Move up then right@>=
begin incr(move_ptr); move[move_ptr]:=2;
end

@ After |make_moves| has acted, possibly for several curves that move toward
the same octant, a ``smoothing'' operation might be done on the |move| array.
This removes optical glitches that can arise even when the curve has been
digitized without rounding errors.

The smoothing process replaces the integers $a_0\ldots a_n$ in
|move[b..t]| by ``smoothed'' integers $a_0'\ldots a_n'$ defined as
follows:
$$a_k'=a_k+\delta\k-\delta_k;\qquad
\delta_k=\cases{+1,&if $1<k<n$ and $a_{k-2}\G a_{k-1}\ll a_k\G a\k$;\cr
-1,&if $1<k<n$ and $a_{k-2}\L a_{k-1}\gg a_k\L a\k$;\cr
0,&otherwise.\cr}$$
Here $a\ll b$ means that $a\L b-2$, and $a\gg b$ means that $a\G b+2$.

The smoothing operation is symmetric in the sense that, if $a_0\ldots a_n$
smooths to $a_0'\ldots a_n'$, then the reverse sequence $a_n\ldots a_0$
smooths to $a_n'\ldots a_0'$; also the complementary sequence
$(m-a_0)\ldots(m-a_n)$ smooths to $(m-a_0')\ldots(m-a_n')$.
We have $a_0'+\cdots+a_n'=a_0+\cdots+a_n$ because $\delta_0=\delta_{n+1}=0$.

@p procedure smooth_moves(@!b,@!t:integer);
var@!k:1..move_size; {index into |move|}
@!a,@!aa,@!aaa:integer; {original values of |move[k],move[k-1],move[k-2]|}
begin if t-b>=3 then
  begin k:=b+2; aa:=move[k-1]; aaa:=move[k-2];
  repeat a:=move[k];
  if abs(a-aa)>1 then
    @<Increase and decrease |move[k-1]| and |move[k]| by $\delta_k$@>;
  incr(k); aaa:=aa; aa:=a;
  until k=t;
  end;
end;

@ @<Increase and decrease |move[k-1]| and |move[k]| by $\delta_k$@>=
if a>aa then
  begin if aaa>=aa then if a>=move[k+1] then
    begin incr(move[k-1]); move[k]:=a-1;
    end;
  end
else  begin if aaa<=aa then if a<=move[k+1] then
    begin decr(move[k-1]); move[k]:=a+1;
    end;
  end

@* \[20] Edge structures.
Now we come to \MF's internal scheme for representing what the user can
actually ``see,'' the edges between pixels. Each pixel has an integer
weight, obtained by summing the weights on all edges to its left. \MF\
represents only the nonzero edge weights, since most of the edges are
weightless; in this way, the data storage requirements grow only linearly
with respect to the number of pixels per point, even though two-dimensional
data is being represented. (Well, the actual dependence on the underlying
resolution is order $n\log n$, but the $\log n$ factor is buried in our
implicit restriction on the maximum raster size.) The sum of all edge
weights in each row should be zero.

The data structure for edge weights must be compact and flexible,
yet it should support efficient updating and display operations. We
want to be able to have many different edge structures in memory at
once, and we want the computer to be able to translate them, reflect them,
and/or merge them together with relative ease.

\MF's solution to this problem requires one single-word node per
nonzero edge weight, plus one two-word node for each row in a contiguous
set of rows. There's also a header node that provides global information
about the entire structure.

@ Let's consider the edge-weight nodes first. The |info| field of such
nodes contains both an $m$~value and a weight~$w$, in the form
$8m+w+c$, where $c$ is a constant that depends on data found in the header.
We shall consider $c$ in detail later; for now, it's best just to think
of it as a way to compensate for the fact that $m$ and~$w$ can be negative,
together with the fact that an |info| field must have a value between
|min_halfword| and |max_halfword|. The $m$ value is an unscaled $x$~coordinate,
so it satisfies $\vert m\vert<
4096$; the $w$ value is always in the range $1\L\vert w\vert\L3$. We can
unpack the data in the |info| field by fetching |ho(info(p))=
info(p)-min_halfword| and dividing this nonnegative number by~8;
the constant~$c$ will be chosen so that the remainder of this division
is $4+w$. Thus, for example, a remainder of~3 will correspond to
the edge weight $w=-1$.

Every row of an edge structure contains two lists of such edge-weight
nodes, called the |sorted| and |unsorted| lists, linked together by their
|link| fields in the normal way. The difference between them is that we
always have |info(p)<=info(link(p))| in the |sorted| list, but there's no
such restriction on the elements of the |unsorted| list. The reason for
this distinction is that it would take unnecessarily long to maintain
edge-weight lists in sorted order while they're being updated; but when we
need to process an entire row from left to right in order of the
$m$~values, it's fairly easy and quick to sort a short list of unsorted
elements and to merge them into place among their sorted cohorts.
Furthermore, the fact that the |unsorted| list is empty can sometimes be
used to good advantage, because it allows us to conclude that a particular
row has not changed since the last time we sorted it.

The final |link| of the |sorted| list will be |sentinel|, which points to
a special one-word node whose |info| field is essentially infinite; this
facilitates the sorting and merging operations. The final |link| of the
|unsorted| list will be either |null| or |void|, where |void=null+1|
is used to avoid redisplaying data that has not changed:
A |void| value is stored at the head of the
unsorted list whenever the corresponding row has been displayed.

@d zero_w=4
@d void==null+1

@<Initialize table entries...@>=
info(sentinel):=max_halfword; {|link(sentinel)=null|}

@ The rows themselves are represented by row header nodes that
contain four link fields. Two of these four, |sorted| and |unsorted|,
point to the first items of the edge-weight lists just mentioned.
The other two, |link| and |knil|, point to the headers of the two
adjacent rows. If |p| points to the header for row number~|n|, then
|link(p)| points up to the header for row~|n+1|, and |knil(p)| points
down to the header for row~|n-1|. This double linking makes it
convenient to move through consecutive rows either upward or downward;
as usual, we have |link(knil(p))=knil(link(p))=p| for all row headers~|p|.

The row associated with a given value of |n| contains weights for
edges that run between the lattice points |(m,n)| and |(m,n+1)|.

@d knil==info {inverse of the |link| field, in a doubly linked list}
@d sorted_loc(#)==#+1 {where the |sorted| link field resides}
@d sorted(#)==link(sorted_loc(#)) {beginning of the list of sorted edge weights}
@d unsorted(#)==info(#+1) {beginning of the list of unsorted edge weights}
@d row_node_size=2 {number of words in a row header node}

@ The main header node |h| for an edge structure has |link| and |knil|
fields that link it above the topmost row and below the bottommost row.
It also has fields called |m_min|, |m_max|, |n_min|, and |n_max| that
bound the current extent of the edge data: All |m| values in edge-weight
nodes should lie between |m_min(h)-4096| and |m_max(h)-4096|, inclusive.
Furthermore the topmost row header, pointed to by |knil(h)|,
is for row number |n_max(h)-4096|; the bottommost row header, pointed to by
|link(h)|, is for row number |n_min(h)-4096|.

The offset constant |c| that's used in all of the edge-weight data is
represented implicitly in |m_offset(h)|; its actual value is
$$\hbox{|c=min_halfword+zero_w+8*m_offset(h)|.}$$
Notice that it's possible to shift an entire edge structure by an
amount $(\Delta m,\Delta n)$ by adding $\Delta n$ to |n_min(h)| and |n_max(h)|,
adding $\Delta m$ to |m_min(h)| and |m_max(h)|, and subtracting
$\Delta m$ from |m_offset(h)|;
none of the other edge data needs to be modified. Initially the |m_offset|
field is~4096, but it will change if the user requests such a shift.
The contents of these five fields should always be positive and less than
8192; |n_max| should, in fact, be less than 8191.  Furthermore
|m_min+m_offset-4096| and |m_max+m_offset-4096| must also lie strictly
between 0 and 8192, so that the |info| fields of edge-weight nodes will
fit in a halfword.

The header node of an edge structure also contains two somewhat unusual
fields that are called |last_window(h)| and |last_window_time(h)|. When this
structure is displayed in window~|k| of the user's screen, after that
window has been updated |t| times, \MF\ sets |last_window(h):=k| and
|last_window_time(h):=t|; it also sets |unsorted(p):=void| for all row
headers~|p|, after merging any existing unsorted weights with the sorted
ones.  A subsequent display in the same window will be able to avoid
redisplaying rows whose |unsorted| list is still |void|, if the window
hasn't been used for something else in the meantime.

A pointer to the row header of row |n_pos(h)-4096| is provided in
|n_rover(h)|. Most of the algorithms that update an edge structure
are able to get by without random row references; they usually
access rows that are neighbors of each other or of the current |n_pos| row.
Exception: If |link(h)=h| (so that the edge structure contains
no rows), we have |n_rover(h)=h|, and |n_pos(h)| is irrelevant.

@d zero_field=4096 {amount added to coordinates to make them positive}
@d n_min(#)==info(#+1) {minimum row number present, plus |zero_field|}
@d n_max(#)==link(#+1) {maximum row number present, plus |zero_field|}
@d m_min(#)==info(#+2) {minimum column number present, plus |zero_field|}
@d m_max(#)==link(#+2) {maximum column number present, plus |zero_field|}
@d m_offset(#)==info(#+3) {translation of $m$ data in edge-weight nodes}
@d last_window(#)==link(#+3) {the last display went into this window}
@d last_window_time(#)==mem[#+4].int {after this many window updates}
@d n_pos(#)==info(#+5) {the row currently in |n_rover|, plus |zero_field|}
@d n_rover(#)==link(#+5) {a row recently referenced}
@d edge_header_size=6 {number of words in an edge-structure header}
@d valid_range(#)==(abs(#-4096)<4096) {is |#| strictly between 0 and 8192?}
@d empty_edges(#)==link(#)=# {are there no rows in this edge header?}

@p procedure init_edges(@!h:pointer); {initialize an edge header to null values}
begin knil(h):=h; link(h):=h;@/
n_min(h):=zero_field+4095; n_max(h):=zero_field-4095;
m_min(h):=zero_field+4095; m_max(h):=zero_field-4095;
m_offset(h):=zero_field;@/
last_window(h):=0; last_window_time(h):=0;@/
n_rover(h):=h; n_pos(h):=0;@/
end;

@ When a lot of work is being done on a particular edge structure, we plant
a pointer to its main header in the global variable |cur_edges|.
This saves us from having to pass this pointer as a parameter over and
over again between subroutines.

Similarly, |cur_wt| is a global weight that is being used by several
procedures at once.

@<Glob...@>=
@!cur_edges:pointer; {the edge structure of current interest}
@!cur_wt:integer; {the edge weight of current interest}

@ The |fix_offset| routine goes through all the edge-weight nodes of
|cur_edges| and adds a constant to their |info| fields, so that
|m_offset(cur_edges)| can be brought back to |zero_field|. (This
is necessary only in unusual cases when the offset has gotten too
large or too small.)

@p procedure fix_offset;
var @!p,@!q:pointer; {list traversers}
@!delta:integer; {the amount of change}
begin delta:=8*(m_offset(cur_edges)-zero_field);
m_offset(cur_edges):=zero_field;
q:=link(cur_edges);
while q<>cur_edges do
  begin p:=sorted(q);
  while p<>sentinel do
    begin info(p):=info(p)-delta; p:=link(p);
    end;
  p:=unsorted(q);
  while p>void do
    begin info(p):=info(p)-delta; p:=link(p);
    end;
  q:=link(q);
  end;
end;

@ The |edge_prep| routine makes the |cur_edges| structure ready to
accept new data whose coordinates satisfy |ml<=m<=mr| and |nl<=n<=nr-1|,
assuming that |-4096<ml<=mr<4096| and |-4096<nl<=nr<4096|. It makes
appropriate adjustments to |m_min|, |m_max|, |n_min|, and |n_max|,
adding new empty rows if necessary.

@p procedure edge_prep(@!ml,@!mr,@!nl,@!nr:integer);
var @!delta:halfword; {amount of change}
@!p,@!q:pointer; {for list manipulation}
begin ml:=ml+zero_field; mr:=mr+zero_field;
nl:=nl+zero_field; nr:=nr-1+zero_field;@/
if ml<m_min(cur_edges) then m_min(cur_edges):=ml;
if mr>m_max(cur_edges) then m_max(cur_edges):=mr;
if not valid_range(m_min(cur_edges)+m_offset(cur_edges)-zero_field) or@|
 not valid_range(m_max(cur_edges)+m_offset(cur_edges)-zero_field) then
  fix_offset;
if empty_edges(cur_edges) then {there are no rows}
  begin n_min(cur_edges):=nr+1; n_max(cur_edges):=nr;
  end;
if nl<n_min(cur_edges) then
  @<Insert exactly |n_min(cur_edges)-nl| empty rows at the bottom@>;
if nr>n_max(cur_edges) then
  @<Insert exactly |nr-n_max(cur_edges)| empty rows at the top@>;
end;

@ @<Insert exactly |n_min(cur_edges)-nl| empty rows at the bottom@>=
begin delta:=n_min(cur_edges)-nl; n_min(cur_edges):=nl;
p:=link(cur_edges);
repeat q:=get_node(row_node_size); sorted(q):=sentinel; unsorted(q):=void;
knil(p):=q; link(q):=p; p:=q; decr(delta);
until delta=0;
knil(p):=cur_edges; link(cur_edges):=p;
if n_rover(cur_edges)=cur_edges then n_pos(cur_edges):=nl-1;
end

@ @<Insert exactly |nr-n_max(cur_edges)| empty rows at the top@>=
begin delta:=nr-n_max(cur_edges); n_max(cur_edges):=nr;
p:=knil(cur_edges);
repeat q:=get_node(row_node_size); sorted(q):=sentinel; unsorted(q):=void;
link(p):=q; knil(q):=p; p:=q; decr(delta);
until delta=0;
link(p):=cur_edges; knil(cur_edges):=p;
if n_rover(cur_edges)=cur_edges then n_pos(cur_edges):=nr+1;
end

@ The |print_edges| subroutine gives a symbolic rendition of an edge
structure, for use in `\&{show}' commands. A rather terse output
format has been chosen since edge structures can grow quite large.

@<Declare subroutines for printing expressions@>=
@t\4@>@<Declare the procedure called |print_weight|@>@;@/
procedure print_edges(@!s:str_number;@!nuline:boolean;@!x_off,@!y_off:integer);
var @!p,@!q,@!r:pointer; {for list traversal}
@!n:integer; {row number}
begin print_diagnostic("Edge structure",s,nuline);
p:=knil(cur_edges); n:=n_max(cur_edges)-zero_field;
while p<>cur_edges do
  begin q:=unsorted(p); r:=sorted(p);
  if(q>void)or(r<>sentinel) then
    begin print_nl("row "); print_int(n+y_off); print_char(":");
    while q>void do
      begin print_weight(q,x_off); q:=link(q);
      end;
    print(" |");
    while r<>sentinel do
      begin print_weight(r,x_off); r:=link(r);
      end;
    end;
  p:=knil(p); decr(n);
  end;
end_diagnostic(true);
end;

@ @<Declare the procedure called |print_weight|@>=
procedure print_weight(@!q:pointer;@!x_off:integer);
var @!w,@!m:integer; {unpacked weight and coordinate}
@!d:integer; {temporary data register}
begin d:=ho(info(q)); w:=d mod 8; m:=(d div 8)-m_offset(cur_edges);
if file_offset>max_print_line-9 then print_nl(" ")
else print_char(" ");
print_int(m+x_off);
while w>zero_w do
  begin print_char("+"); decr(w);
  end;
while w<zero_w do
  begin print_char("-"); incr(w);
  end;
end;

@ Here's a trivial subroutine that copies an edge structure. (Let's hope
that the given structure isn't too gigantic.)

@p function copy_edges(@!h:pointer):pointer;
var @!p,@!r:pointer; {variables that traverse the given structure}
@!hh,@!pp,@!qq,@!rr,@!ss:pointer; {variables that traverse the new structure}
begin hh:=get_node(edge_header_size);
mem[hh+1]:=mem[h+1]; mem[hh+2]:=mem[h+2];
mem[hh+3]:=mem[h+3]; mem[hh+4]:=mem[h+4]; {we've now copied |n_min|, |n_max|,
  |m_min|, |m_max|, |m_offset|, |last_window|, and |last_window_time|}
n_pos(hh):=n_max(hh)+1;n_rover(hh):=hh;@/
p:=link(h); qq:=hh;
while p<>h do
  begin pp:=get_node(row_node_size); link(qq):=pp; knil(pp):=qq;
  @<Copy both |sorted| and |unsorted| lists of |p| to |pp|@>;
  p:=link(p); qq:=pp;
  end;
link(qq):=hh; knil(hh):=qq;
copy_edges:=hh;
end;

@ @<Copy both |sorted| and |unsorted|...@>=
r:=sorted(p); rr:=sorted_loc(pp); {|link(rr)=sorted(pp)|}
while r<>sentinel do
  begin ss:=get_avail; link(rr):=ss; rr:=ss; info(rr):=info(r);@/
  r:=link(r);
  end;
link(rr):=sentinel;@/
r:=unsorted(p); rr:=temp_head;
while r>void do
  begin ss:=get_avail; link(rr):=ss; rr:=ss; info(rr):=info(r);@/
  r:=link(r);
  end;
link(rr):=r; unsorted(pp):=link(temp_head)

@ Another trivial routine flips |cur_edges| about the |x|-axis
(i.e., negates all the |y| coordinates), assuming that at least
one row is present.

@p procedure y_reflect_edges;
var @!p,@!q,@!r:pointer; {list manipulation registers}
begin p:=n_min(cur_edges);
n_min(cur_edges):=zero_field+zero_field-1-n_max(cur_edges);
n_max(cur_edges):=zero_field+zero_field-1-p;
n_pos(cur_edges):=zero_field+zero_field-1-n_pos(cur_edges);@/
p:=link(cur_edges); q:=cur_edges; {we assume that |p<>q|}
repeat r:=link(p); link(p):=q; knil(q):=p; q:=p; p:=r;
until q=cur_edges;
last_window_time(cur_edges):=0;
end;

@ It's somewhat more difficult, yet not too hard, to reflect about the |y|-axis.

@p procedure x_reflect_edges;
var @!p,@!q,@!r,@!s:pointer; {list manipulation registers}
@!m:integer; {|info| fields will be reflected with respect to this number}
begin p:=m_min(cur_edges);
m_min(cur_edges):=zero_field+zero_field-m_max(cur_edges);
m_max(cur_edges):=zero_field+zero_field-p;
m:=(zero_field+m_offset(cur_edges))*8+zero_w+min_halfword+zero_w+min_halfword;
m_offset(cur_edges):=zero_field;
p:=link(cur_edges);
repeat @<Reflect the edge-and-weight data in |sorted(p)|@>;
@<Reflect the edge-and-weight data in |unsorted(p)|@>;
p:=link(p);
until p=cur_edges;
last_window_time(cur_edges):=0;
end;

@ We want to change the sign of the weight as we change the sign of the
|x|~coordinate. Fortunately, it's easier to do this than to negate
one without the other.

@<Reflect the edge-and-weight data in |unsorted(p)|@>=
q:=unsorted(p);
while q>void do
  begin info(q):=m-info(q); q:=link(q);
  end

@ Reversing the order of a linked list is best thought of as the process of
popping nodes off one stack and pushing them on another. In this case we
pop from stack~|q| and push to stack~|r|.

@<Reflect the edge-and-weight data in |sorted(p)|@>=
q:=sorted(p); r:=sentinel;
while q<>sentinel do
  begin s:=link(q); link(q):=r; r:=q; info(r):=m-info(q); q:=s;
  end;
sorted(p):=r

@ Now let's multiply all the $y$~coordinates of a nonempty edge structure
by a small integer $s>1$:

@p procedure y_scale_edges(@!s:integer);
var @!p,@!q,@!pp,@!r,@!rr,@!ss:pointer; {list manipulation registers}
@!t:integer; {replication counter}
begin if (s*(n_max(cur_edges)+1-zero_field)>=4096) or@|
 (s*(n_min(cur_edges)-zero_field)<=-4096) then
  begin print_err("Scaled picture would be too big");
@.Scaled picture...big@>
  help3("I can't yscale the picture as requested---it would")@/
    ("make some coordinates too large or too small.")@/
    ("Proceed, and I'll omit the transformation.");
  put_get_error;
  end
else  begin n_max(cur_edges):=s*(n_max(cur_edges)+1-zero_field)-1+zero_field;
  n_min(cur_edges):=s*(n_min(cur_edges)-zero_field)+zero_field;
  @<Replicate every row exactly $s$ times@>;
  last_window_time(cur_edges):=0;
  end;
end;

@ @<Replicate...@>=
p:=cur_edges;
repeat q:=p; p:=link(p);
for t:=2 to s do
  begin pp:=get_node(row_node_size); link(q):=pp; knil(p):=pp;
  link(pp):=p; knil(pp):=q; q:=pp;
  @<Copy both |sorted| and |unsorted|...@>;
  end;
until link(p)=cur_edges

@ Scaling the $x$~coordinates is, of course, our next task.

@p procedure x_scale_edges(@!s:integer);
var @!p,@!q:pointer; {list manipulation registers}
@!t:0..65535; {unpacked |info| field}
@!w:0..7; {unpacked weight}
@!delta:integer; {amount added to scaled |info|}
begin if (s*(m_max(cur_edges)-zero_field)>=4096) or@|
 (s*(m_min(cur_edges)-zero_field)<=-4096) then
  begin print_err("Scaled picture would be too big");
@.Scaled picture...big@>
  help3("I can't xscale the picture as requested---it would")@/
    ("make some coordinates too large or too small.")@/
    ("Proceed, and I'll omit the transformation.");
  put_get_error;
  end
else if (m_max(cur_edges)<>zero_field)or(m_min(cur_edges)<>zero_field) then
  begin m_max(cur_edges):=s*(m_max(cur_edges)-zero_field)+zero_field;
  m_min(cur_edges):=s*(m_min(cur_edges)-zero_field)+zero_field;
  delta:=8*(zero_field-s*m_offset(cur_edges))+min_halfword;
  m_offset(cur_edges):=zero_field;@/
  @<Scale the $x$~coordinates of each row by $s$@>;
  last_window_time(cur_edges):=0;
  end;
end;

@ The multiplications cannot overflow because we know that |s<4096|.

@<Scale the $x$~coordinates of each row by $s$@>=
q:=link(cur_edges);
repeat p:=sorted(q);
while p<>sentinel do
  begin t:=ho(info(p)); w:=t mod 8; info(p):=(t-w)*s+w+delta; p:=link(p);
  end;
p:=unsorted(q);
while p>void do
  begin t:=ho(info(p)); w:=t mod 8; info(p):=(t-w)*s+w+delta; p:=link(p);
  end;
q:=link(q);
until q=cur_edges

@ Here is a routine that changes the signs of all the weights, without
changing anything else.

@p procedure negate_edges(@!h:pointer);
label done;
var @!p,@!q,@!r,@!s,@!t,@!u:pointer; {structure traversers}
begin p:=link(h);
while p<>h do
  begin q:=unsorted(p);
  while q>void do
    begin info(q):=8-2*((ho(info(q))) mod 8)+info(q); q:=link(q);
    end;
  q:=sorted(p);
  if q<>sentinel then
    begin repeat info(q):=8-2*((ho(info(q))) mod 8)+info(q); q:=link(q);
    until q=sentinel;
    @<Put the list |sorted(p)| back into sort@>;
    end;
  p:=link(p);
  end;
last_window_time(h):=0;
end;

@ \MF\ would work even if the code in this section were omitted, because
a list of edge-and-weight data that is sorted only by
|m| but not~|w| turns out to be good enough for correct operation.
However, the author decided not to make the program even trickier than
it is already, since |negate_edges| isn't needed very often.
The simpler-to-state condition, ``keep the |sorted| list fully sorted,''
is therefore being preserved at the cost of extra computation.

@<Put the list |sorted(p)|...@>=
u:=sorted_loc(p); q:=link(u); r:=q; s:=link(r); {|q=sorted(p)|}
loop@+  if info(s)>info(r) then
    begin link(u):=q;
    if s=sentinel then goto done;
    u:=r; q:=s; r:=q; s:=link(r);
    end
  else  begin t:=s; s:=link(t); link(t):=q; q:=t;
    end;
done: link(r):=sentinel

@ The |unsorted| edges of a row are merged into the |sorted| ones by
a subroutine called |sort_edges|. It uses simple insertion sort,
followed by a merge, because the unsorted list is supposedly quite short.
However, the unsorted list is assumed to be nonempty.

@p procedure sort_edges(@!h:pointer); {|h| is a row header}
label done;
var @!k:halfword; {key register that we compare to |info(q)|}
@!p,@!q,@!r,@!s:pointer;
begin r:=unsorted(h); unsorted(h):=null;
p:=link(r); link(r):=sentinel; link(temp_head):=r;
while p>void do {sort node |p| into the list that starts at |temp_head|}
  begin k:=info(p); q:=temp_head;
  repeat r:=q; q:=link(r);
  until k<=info(q);
  link(r):=p; r:=link(p); link(p):=q; p:=r;
  end;
@<Merge the |temp_head| list into |sorted(h)|@>;
end;

@ In this step we use the fact that |sorted(h)=link(sorted_loc(h))|.

@<Merge the |temp_head| list into |sorted(h)|@>=
begin r:=sorted_loc(h); q:=link(r); p:=link(temp_head);
loop@+  begin k:=info(p);
  while k>info(q) do
    begin r:=q; q:=link(r);
    end;
  link(r):=p; s:=link(p); link(p):=q;
  if s=sentinel then goto done;
  r:=p; p:=s;
  end;
done:end

@ The |cull_edges| procedure ``optimizes'' an edge structure by making all
the pixel weights either |w_out| or~|w_in|. The weight will be~|w_in| after the
operation if and only if it was in the closed interval |[w_lo,w_hi]|
before, where |w_lo<=w_hi|. Either |w_out| or |w_in| is zero, while the other is
$\pm1$, $\pm2$, or $\pm3$. The parameters will be such that zero-weight
pixels will remain of weight zero.  (This is fortunate,
because there are infinitely many of them.)

The procedure also computes the tightest possible bounds on the resulting
data, by updating |m_min|, |m_max|, |n_min|, and~|n_max|.

@p procedure cull_edges(@!w_lo,@!w_hi,@!w_out,@!w_in:integer);
label done;
var @!p,@!q,@!r,@!s:pointer; {for list manipulation}
@!w:integer; {new weight after culling}
@!d:integer; {data register for unpacking}
@!m:integer; {the previous column number, including |m_offset|}
@!mm:integer; {the next column number, including |m_offset|}
@!ww:integer; {accumulated weight before culling}
@!prev_w:integer; {value of |w| before column |m|}
@!n,@!min_n,@!max_n:pointer; {current and extreme row numbers}
@!min_d,@!max_d:pointer; {extremes of the new edge-and-weight data}
begin min_d:=max_halfword; max_d:=min_halfword;
min_n:=max_halfword; max_n:=min_halfword;@/
p:=link(cur_edges); n:=n_min(cur_edges);
while p<>cur_edges do
  begin if unsorted(p)>void then sort_edges(p);
  if sorted(p)<>sentinel then
    @<Cull superfluous edge-weight entries from |sorted(p)|@>;
  p:=link(p); incr(n);
  end;
@<Delete empty rows at the top and/or bottom;
  update the boundary values in the header@>;
last_window_time(cur_edges):=0;
end;

@ The entire |sorted| list is returned to available memory in this step;
a new list is built starting (temporarily) at |temp_head|.
Since several edges can occur at the same column, we need to be looking
ahead of where the actual culling takes place. This means that it's
slightly tricky to get the iteration started and stopped.

@<Cull superfluous...@>=
begin r:=temp_head; q:=sorted(p); ww:=0; m:=1000000; prev_w:=0;
loop@+  begin if q=sentinel then mm:=1000000
  else  begin d:=ho(info(q)); mm:=d div 8; ww:=ww+(d mod 8)-zero_w;
    end;
  if mm>m then
    begin @<Insert an edge-weight for edge |m|, if the new pixel
      weight has changed@>;
    if q=sentinel then goto done;
    end;
  m:=mm;
  if ww>=w_lo then if ww<=w_hi then w:=w_in
    else w:=w_out
  else w:=w_out;
  s:=link(q); free_avail(q); q:=s;
  end;
done: link(r):=sentinel; sorted(p):=link(temp_head);
if r<>temp_head then @<Update the max/min amounts@>;
end

@ @<Insert an edge-weight for edge |m|, if...@>=
if w<>prev_w then
  begin s:=get_avail; link(r):=s;
  info(s):=8*m+min_halfword+zero_w+w-prev_w;
  r:=s; prev_w:=w;
  end

@ @<Update the max/min amounts@>=
begin if min_n=max_halfword then min_n:=n;
max_n:=n;
if min_d>info(link(temp_head)) then min_d:=info(link(temp_head));
if max_d<info(r) then max_d:=info(r);
end

@ @<Delete empty rows at the top and/or bottom...@>=
if min_n>max_n then @<Delete all the row headers@>
else  begin n:=n_min(cur_edges); n_min(cur_edges):=min_n;
  while min_n>n do
    begin p:=link(cur_edges); link(cur_edges):=link(p);
    knil(link(p)):=cur_edges;
    free_node(p,row_node_size); incr(n);
    end;
  n:=n_max(cur_edges); n_max(cur_edges):=max_n;
  n_pos(cur_edges):=max_n+1; n_rover(cur_edges):=cur_edges;
  while max_n<n do
    begin p:=knil(cur_edges); knil(cur_edges):=knil(p);
    link(knil(p)):=cur_edges;
    free_node(p,row_node_size); decr(n);
    end;
  m_min(cur_edges):=((ho(min_d)) div 8)-m_offset(cur_edges)+zero_field;
  m_max(cur_edges):=((ho(max_d)) div 8)-m_offset(cur_edges)+zero_field;
  end

@ We get here if the edges have been entirely culled away.

@<Delete all the row headers@>=
begin p:=link(cur_edges);
while p<>cur_edges do
  begin q:=link(p); free_node(p,row_node_size); p:=q;
  end;
init_edges(cur_edges);
end


@ The last and most difficult routine for transforming an edge structure---and
the most interesting one!---is |xy_swap_edges|, which interchanges the
r\^^Doles of rows and columns. Its task can be viewed as the job of
creating an edge structure that contains only horizontal edges, linked
together in columns, given an edge structure that contains only
vertical edges linked together in rows; we must do this without changing
the implied pixel weights.

Given any two adjacent rows of an edge structure, it is not difficult to
determine the horizontal edges that lie ``between'' them: We simply look
for vertically adjacent pixels that have different weight, and insert
a horizontal edge containing the difference in weights. Every horizontal
edge determined in this way should be put into an appropriate linked
list. Since random access to these linked lists is desirable, we use
the |move| array to hold the list heads. If we work through the given
edge structure from top to bottom, the constructed lists will not need
to be sorted, since they will already be in order.

The following algorithm makes use of some ideas suggested by John Hobby.
@^Hobby, John Douglas@>
It assumes that the edge structure is non-null, i.e., that |link(cur_edges)
<>cur_edges|, hence |m_max(cur_edges)>=m_min(cur_edges)|.

@p procedure xy_swap_edges; {interchange |x| and |y| in |cur_edges|}
label done;
var @!m_magic,@!n_magic:integer; {special values that account for offsets}
@!p,@!q,@!r,@!s:pointer; {pointers that traverse the given structure}
@<Other local variables for |xy_swap_edges|@>@;
begin @<Initialize the array of new edge list heads@>;
@<Insert blank rows at the top and bottom, and set |p| to the new top row@>;
@<Compute the magic offset values@>;
repeat q:=knil(p);@+if unsorted(q)>void then sort_edges(q);
@<Insert the horizontal edges defined by adjacent rows |p,q|,
  and destroy row~|p|@>;
p:=q; n_magic:=n_magic-8;
until knil(p)=cur_edges;
free_node(p,row_node_size); {now all original rows have been recycled}
@<Adjust the header to reflect the new edges@>;
end;

@ Here we don't bother to keep the |link| entries up to date, since the
procedure looks only at the |knil| fields as it destroys the former
edge structure.

@<Insert blank rows at the top and bottom...@>=
p:=get_node(row_node_size); sorted(p):=sentinel; unsorted(p):=null;@/
knil(p):=cur_edges; knil(link(cur_edges)):=p; {the new bottom row}
p:=get_node(row_node_size); sorted(p):=sentinel;
knil(p):=knil(cur_edges); {the new top row}

@ The new lists will become |sorted| lists later, so we initialize
empty lists to |sentinel|.

@<Initialize the array of new edge list heads@>=
m_spread:=m_max(cur_edges)-m_min(cur_edges); {this is |>=0| by assumption}
if m_spread>move_size then overflow("move table size",move_size);
@:METAFONT capacity exceeded move table size}{\quad move table size@>
for j:=0 to m_spread do move[j]:=sentinel

@ @<Other local variables for |xy_swap_edges|@>=
@!m_spread:integer; {the difference between |m_max| and |m_min|}
@!j,@!jj:0..move_size; {indices into |move|}
@!m,@!mm:integer; {|m| values at vertical edges}
@!pd,@!rd:integer; {data fields from edge-and-weight nodes}
@!pm,@!rm:integer; {|m| values from edge-and-weight nodes}
@!w:integer; {the difference in accumulated weight}
@!ww:integer; {as much of |w| that can be stored in a single node}
@!dw:integer; {an increment to be added to |w|}

@ At the point where we test |w<>0|, variable |w| contains
the accumulated weight from edges already passed in
row~|p| minus the accumulated weight from edges already passed in row~|q|.

@<Insert the horizontal edges defined by adjacent rows |p,q|...@>=
r:=sorted(p); free_node(p,row_node_size); p:=r;@/
pd:=ho(info(p)); pm:=pd div 8;@/
r:=sorted(q); rd:=ho(info(r)); rm:=rd div 8; w:=0;
loop@+  begin if pm<rm then mm:=pm@+else mm:=rm;
  if w<>0 then
    @<Insert horizontal edges of weight |w| between |m| and~|mm|@>;
  if pd<rd then
    begin dw:=(pd mod 8)-zero_w;
    @<Advance pointer |p| to the next vertical edge,
      after destroying the previous one@>;
    end
  else  begin if r=sentinel then goto done; {|rd=pd=ho(max_halfword)|}
    dw:=-((rd mod 8)-zero_w);
    @<Advance pointer |r| to the next vertical edge@>;
    end;
  m:=mm; w:=w+dw;
  end;
done:

@ @<Advance pointer |r| to the next vertical edge@>=
r:=link(r); rd:=ho(info(r)); rm:=rd div 8

@ @<Advance pointer |p| to the next vertical edge...@>=
s:=link(p); free_avail(p); p:=s; pd:=ho(info(p)); pm:=pd div 8

@ Certain ``magic'' values are needed to make the following code work,
because of the various offsets in our data structure. For now, let's not
worry about their precise values; we shall compute |m_magic| and |n_magic|
later, after we see what the code looks like.

@ @<Insert horizontal edges of weight |w| between |m| and~|mm|@>=
if m<>mm then
  begin if mm-m_magic>=move_size then confusion("xy");
@:this can't happen xy}{\quad xy@>
  extras:=(abs(w)-1) div 3;
  if extras>0 then
    begin if w>0 then xw:=+3@+else xw:=-3;
    ww:=w-extras*xw;
    end
  else ww:=w;
  repeat j:=m-m_magic;
  for k:=1 to extras do
    begin s:=get_avail; info(s):=n_magic+xw;
    link(s):=move[j]; move[j]:=s;
    end;
  s:=get_avail; info(s):=n_magic+ww;
  link(s):=move[j]; move[j]:=s;@/
  incr(m);
  until m=mm;
  end

@ @<Other local variables for |xy...@>=
@!extras:integer; {the number of additional nodes to make weights |>3|}
@!xw:-3..3; {the additional weight in extra nodes}
@!k:integer; {loop counter for inserting extra nodes}

@ At the beginning of this step, |move[m_spread]=sentinel|, because no
horizontal edges will extend to the right of column |m_max(cur_edges)|.

@<Adjust the header to reflect the new edges@>=
move[m_spread]:=0; j:=0;
while move[j]=sentinel do incr(j);
if j=m_spread then init_edges(cur_edges) {all edge weights are zero}
else  begin mm:=m_min(cur_edges);
  m_min(cur_edges):=n_min(cur_edges);
  m_max(cur_edges):=n_max(cur_edges)+1;
  m_offset(cur_edges):=zero_field;
  jj:=m_spread-1;
  while move[jj]=sentinel do decr(jj);
  n_min(cur_edges):=j+mm; n_max(cur_edges):=jj+mm; q:=cur_edges;
  repeat p:=get_node(row_node_size); link(q):=p; knil(p):=q;
  sorted(p):=move[j]; unsorted(p):=null; incr(j); q:=p;
  until j>jj;
  link(q):=cur_edges; knil(cur_edges):=q;
  n_pos(cur_edges):=n_max(cur_edges)+1; n_rover(cur_edges):=cur_edges;
  last_window_time(cur_edges):=0;
  end;

@ The values of |m_magic| and |n_magic| can be worked out by trying the
code above on a small example; if they work correctly in simple cases,
they should work in general.

@<Compute the magic offset values@>=
m_magic:=m_min(cur_edges)+m_offset(cur_edges)-zero_field;
n_magic:=8*n_max(cur_edges)+8+zero_w+min_halfword

@ Now let's look at the subroutine that merges the edges from a given
edge structure into |cur_edges|. The given edge structure loses all its
edges.

@p procedure merge_edges(@!h:pointer);
label done;
var @!p,@!q,@!r,@!pp,@!qq,@!rr:pointer; {list manipulation registers}
@!n:integer; {row number}
@!k:halfword; {key register that we compare to |info(q)|}
@!delta:integer; {change to the edge/weight data}
begin if link(h)<>h then
  begin if (m_min(h)<m_min(cur_edges))or(m_max(h)>m_max(cur_edges))or@|
    (n_min(h)<n_min(cur_edges))or(n_max(h)>n_max(cur_edges)) then
    edge_prep(m_min(h)-zero_field,m_max(h)-zero_field,
      n_min(h)-zero_field,n_max(h)-zero_field+1);
  if m_offset(h)<>m_offset(cur_edges) then
    @<Adjust the data of |h| to account for a difference of offsets@>;
  n:=n_min(cur_edges); p:=link(cur_edges); pp:=link(h);
  while n<n_min(h) do
    begin incr(n); p:=link(p);
    end;
  repeat @<Merge row |pp| into row |p|@>;
  pp:=link(pp); p:=link(p);
  until pp=h;
  end;
end;

@ @<Adjust the data of |h| to account for a difference of offsets@>=
begin pp:=link(h); delta:=8*(m_offset(cur_edges)-m_offset(h));
repeat qq:=sorted(pp);
while qq<>sentinel do
  begin info(qq):=info(qq)+delta; qq:=link(qq);
  end;
qq:=unsorted(pp);
while qq>void do
  begin info(qq):=info(qq)+delta; qq:=link(qq);
  end;
pp:=link(pp);
until pp=h;
end

@ The |sorted| and |unsorted| lists are merged separately. After this
step, row~|pp| will have no edges remaining, since they will all have
been merged into row~|p|.

@<Merge row |pp|...@>=
qq:=unsorted(pp);
if qq>void then
  if unsorted(p)<=void then unsorted(p):=qq
  else  begin while link(qq)>void do qq:=link(qq);
    link(qq):=unsorted(p); unsorted(p):=unsorted(pp);
    end;
unsorted(pp):=null; qq:=sorted(pp);
if qq<>sentinel then
  begin if unsorted(p)=void then unsorted(p):=null;
  sorted(pp):=sentinel; r:=sorted_loc(p); q:=link(r); {|q=sorted(p)|}
  if q=sentinel then sorted(p):=qq
  else loop@+begin k:=info(qq);
    while k>info(q) do
      begin r:=q; q:=link(r);
      end;
    link(r):=qq; rr:=link(qq); link(qq):=q;
    if rr=sentinel then goto done;
    r:=qq; qq:=rr;
    end;
  end;
done:

@ The |total_weight| routine computes the total of all pixel weights
in a given edge structure. It's not difficult to prove that this is
the sum of $(-w)$ times $x$ taken over all edges,
where $w$ and~$x$ are the weight and $x$~coordinates stored in an edge.
It's not necessary to worry that this quantity will overflow the
size of an |integer| register, because it will be less than~$2^{31}$
unless the edge structure has more than 174,762 edges. However, we had
better not try to compute it as a |scaled| integer, because a total
weight of almost $12\times 2^{12}$ can be produced by only four edges.

@p function total_weight(@!h:pointer):integer; {|h| is an edge header}
var @!p,@!q:pointer; {variables that traverse the given structure}
@!n:integer; {accumulated total so far}
@!m:0..65535; {packed $x$ and $w$ values, including offsets}
begin n:=0; p:=link(h);
while p<>h do
  begin q:=sorted(p);
  while q<>sentinel do
    @<Add the contribution of node |q| to the total weight,
      and set |q:=link(q)|@>;
  q:=unsorted(p);
  while q>void do
    @<Add the contribution of node |q| to the total weight,
      and set |q:=link(q)|@>;
  p:=link(p);
  end;
total_weight:=n;
end;

@ It's not necessary to add the offsets to the $x$ coordinates, because
an entire edge structure can be shifted without affecting its total weight.
Similarly, we don't need to subtract |zero_field|.

@<Add the contribution of node |q| to the total weight...@>=
begin m:=ho(info(q)); n:=n-((m mod 8)-zero_w)*(m div 8);
q:=link(q);
end

@ So far we've done lots of things to edge structures assuming that
edges are actually present, but we haven't seen how edges get created
in the first place. Let's turn now to the problem of generating new edges.

\MF\ will display new edges as they are being computed, if |tracing_edges|
is positive. In order to keep such data reasonably compact, only the
points at which the path makes a $90^\circ$ or $180^\circ$ turn are listed.

The tracing algorithm must remember some past history in order to suppress
unnecessary data. Three variables |trace_x|, |trace_y|, and |trace_yy|
provide this history: The last coordinates printed were |(trace_x,trace_y)|,
and the previous edge traced ended at |(trace_x,trace_yy)|. Before anything
at all has been traced, |trace_x=-4096|.

@<Glob...@>=
@!trace_x:integer; {$x$~coordinate most recently shown in a trace}
@!trace_y:integer; {$y$~coordinate most recently shown in a trace}
@!trace_yy:integer; {$y$~coordinate most recently encountered}

@ Edge tracing is initiated by the |begin_edge_tracing| routine,
continued by the |trace_a_corner| routine, and terminated by the
|end_edge_tracing| routine.

@p procedure begin_edge_tracing;
begin print_diagnostic("Tracing edges","",true);
print(" (weight "); print_int(cur_wt); print_char(")"); trace_x:=-4096;
end;
@#
procedure trace_a_corner;
begin if file_offset>max_print_line-13 then print_nl("");
print_char("("); print_int(trace_x); print_char(","); print_int(trace_yy);
print_char(")"); trace_y:=trace_yy;
end;
@#
procedure end_edge_tracing;
begin if trace_x=-4096 then print_nl("(No new edges added.)")
@.No new edges added@>
else  begin trace_a_corner; print_char(".");
  end;
end_diagnostic(true);
end;

@ Just after a new edge weight has been put into the |info| field of
node~|r|, in row~|n|, the following routine continues an ongoing trace.

@p procedure trace_new_edge(@!r:pointer;@!n:integer);
var @!d:integer; {temporary data register}
@!w:-3..3; {weight associated with an edge transition}
@!m,@!n0,@!n1:integer; {column and row numbers}
begin d:=ho(info(r)); w:=(d mod 8)-zero_w; m:=(d div 8)-m_offset(cur_edges);
if w=cur_wt then
  begin n0:=n+1; n1:=n;
  end
else  begin n0:=n; n1:=n+1;
  end; {the edges run from |(m,n0)| to |(m,n1)|}
if m<>trace_x then
  begin if trace_x=-4096 then
    begin print_nl(""); trace_yy:=n0;
    end
  else if trace_yy<>n0 then print_char("?") {shouldn't happen}
  else trace_a_corner;
  trace_x:=m; trace_a_corner;
  end
else  begin if n0<>trace_yy then print_char("!"); {shouldn't happen}
  if ((n0<n1)and(trace_y>trace_yy))or((n0>n1)and(trace_y<trace_yy)) then
    trace_a_corner;
  end;
trace_yy:=n1;
end;

@ One way to put new edge weights into an edge structure is to use the
following routine, which simply draws a straight line from |(x0,y0)| to
|(x1,y1)|. More precisely, it introduces weights for the edges of the
discrete path $\bigl(\lfloor t[x_0,x_1]+{1\over2}+\epsilon\rfloor,
\lfloor t[y_0,y_1]+{1\over2}+\epsilon\delta\rfloor\bigr)$,
as $t$ varies from 0 to~1, where $\epsilon$ and $\delta$ are extremely small
positive numbers.

The structure header is assumed to be |cur_edges|; downward edge weights
will be |cur_wt|, while upward ones will be |-cur_wt|.

Of course, this subroutine will be called only in connection with others
that eventually draw a complete cycle, so that the sum of the edge weights
in each row will be zero whenever the row is displayed.

@p procedure line_edges(@!x0,@!y0,@!x1,@!y1:scaled);
label done,done1;
var @!m0,@!n0,@!m1,@!n1:integer; {rounded and unscaled coordinates}
@!delx,@!dely:scaled; {the coordinate differences of the line}
@!yt:scaled; {smallest |y| coordinate that rounds the same as |y0|}
@!tx:scaled; {tentative change in |x|}
@!p,@!r:pointer; {list manipulation registers}
@!base:integer; {amount added to edge-and-weight data}
@!n:integer; {current row number}
begin n0:=round_unscaled(y0);
n1:=round_unscaled(y1);
if n0<>n1 then
  begin m0:=round_unscaled(x0); m1:=round_unscaled(x1);
  delx:=x1-x0; dely:=y1-y0;
  yt:=n0*unity-half_unit; y0:=y0-yt; y1:=y1-yt;
  if n0<n1 then @<Insert upward edges for a line@>
  else @<Insert downward edges for a line@>;
  n_rover(cur_edges):=p; n_pos(cur_edges):=n+zero_field;
  end;
end;

@ Here we are careful to cancel any effect of rounding error.

@<Insert upward edges for a line@>=
begin base:=8*m_offset(cur_edges)+min_halfword+zero_w-cur_wt;
if m0<=m1 then edge_prep(m0,m1,n0,n1)@+else edge_prep(m1,m0,n0,n1);
@<Move to row |n0|, pointed to by |p|@>;
y0:=unity-y0;
loop@+  begin r:=get_avail; link(r):=unsorted(p); unsorted(p):=r;@/
  tx:=take_fraction(delx,make_fraction(y0,dely));
  if ab_vs_cd(delx,y0,dely,tx)<0 then decr(tx);
    {now $|tx|=\lfloor|y0|\cdot|delx|/|dely|\rfloor$}
  info(r):=8*round_unscaled(x0+tx)+base;@/
  y1:=y1-unity;
  if internal[tracing_edges]>0 then trace_new_edge(r,n);
  if y1<unity then goto done;
  p:=link(p); y0:=y0+unity; incr(n);
  end;
done: end

@ @<Insert downward edges for a line@>=
begin base:=8*m_offset(cur_edges)+min_halfword+zero_w+cur_wt;
if m0<=m1 then edge_prep(m0,m1,n1,n0)@+else edge_prep(m1,m0,n1,n0);
decr(n0); @<Move to row |n0|, pointed to by |p|@>;
loop@+  begin r:=get_avail; link(r):=unsorted(p); unsorted(p):=r;@/
  tx:=take_fraction(delx,make_fraction(y0,dely));
  if ab_vs_cd(delx,y0,dely,tx)<0 then incr(tx);
    {now $|tx|=\lceil|y0|\cdot|delx|/|dely|\rceil$, since |dely<0|}
  info(r):=8*round_unscaled(x0-tx)+base;@/
  y1:=y1+unity;
  if internal[tracing_edges]>0 then trace_new_edge(r,n);
  if y1>=0 then goto done1;
  p:=knil(p); y0:=y0+unity; decr(n);
  end;
done1: end

@ @<Move to row |n0|, pointed to by |p|@>=
n:=n_pos(cur_edges)-zero_field; p:=n_rover(cur_edges);
if n<>n0 then
  if n<n0 then
    repeat incr(n); p:=link(p);
    until n=n0
  else  repeat decr(n); p:=knil(p);
    until n=n0

@ \MF\ inserts most of its edges into edge structures via the
|move_to_edges| subroutine, which uses the data stored in the |move| array
to specify a sequence of ``rook moves.'' The starting point |(m0,n0)|
and finishing point |(m1,n1)| of these moves, as seen from the standpoint
of the first octant, are supplied as parameters; the moves should, however,
be rotated into a given octant.  (We're going to study octant
transformations in great detail later; the reader may wish to come back to
this part of the program after mastering the mysteries of octants.)

The rook moves themselves are defined as follows, from a |first_octant|
point of view: ``Go right |move[k]| steps, then go up one, for |0<=k<n1-n0|;
then go right |move[n1-n0]| steps and stop.'' The sum of |move[k]|
for |0<=k<=n1-n0| will be equal to |m1-m0|.

As in the |line_edges| routine, we use |+cur_wt| as the weight of
all downward edges and |-cur_wt| as the weight of all upward edges,
after the moves have been rotated to the proper octant direction.

There are two main cases to consider: \\{fast\_case} is for moves that
travel in the direction of octants 1, 4, 5, and~8, while \\{slow\_case}
is for moves that travel toward octants 2, 3, 6, and~7. The latter directions
are comparatively cumbersome because they generate more upward or downward
edges; a curve that travels horizontally doesn't produce any edges at all,
but a curve that travels vertically touches lots of rows.

@d fast_case_up=60 {for octants 1 and 4}
@d fast_case_down=61 {for octants 5 and 8}
@d slow_case_up=62 {for octants 2 and 3}
@d slow_case_down=63 {for octants 6 and 7}

@p procedure move_to_edges(@!m0,@!n0,@!m1,@!n1:integer);
label fast_case_up,fast_case_down,slow_case_up,slow_case_down,done;
var @!delta:0..move_size; {extent of |move| data}
@!k:0..move_size; {index into |move|}
@!p,@!r:pointer; {list manipulation registers}
@!dx:integer; {change in edge-weight |info| when |x| changes by 1}
@!edge_and_weight:integer; {|info| to insert}
@!j:integer; {number of consecutive vertical moves}
@!n:integer; {the current row pointed to by |p|}
debug @!sum:integer;@+gubed@;@/
begin delta:=n1-n0;
debug sum:=move[0]; for k:=1 to delta do sum:=sum+abs(move[k]);
if sum<>m1-m0 then confusion("0");@+gubed@;@/
@:this can't happen 0}{\quad 0@>
@<Prepare for and switch to the appropriate case, based on |octant|@>;
fast_case_up:@<Add edges for first or fourth octants, then |goto done|@>;
fast_case_down:@<Add edges for fifth or eighth octants, then |goto done|@>;
slow_case_up:@<Add edges for second or third octants, then |goto done|@>;
slow_case_down:@<Add edges for sixth or seventh octants, then |goto done|@>;
done: n_pos(cur_edges):=n+zero_field; n_rover(cur_edges):=p;
end;

@ The current octant code appears in a global variable. If, for example,
we have |octant=third_octant|, it means that a curve traveling in a north to
north-westerly direction has been rotated for the purposes of internal
calculations so that the |move| data travels in an east to north-easterly
direction. We want to unrotate as we update the edge structure.

@<Glob...@>=
@!octant:first_octant..sixth_octant; {the current octant of interest}

@ @<Prepare for and switch to the appropriate case, based on |octant|@>=
case octant of
first_octant:begin dx:=8; edge_prep(m0,m1,n0,n1); goto fast_case_up;
  end;
second_octant:begin dx:=8; edge_prep(n0,n1,m0,m1); goto slow_case_up;
  end;
third_octant:begin dx:=-8; edge_prep(-n1,-n0,m0,m1); negate(n0);
  goto slow_case_up;
  end;
fourth_octant:begin dx:=-8; edge_prep(-m1,-m0,n0,n1); negate(m0);
  goto fast_case_up;
  end;
fifth_octant:begin dx:=-8; edge_prep(-m1,-m0,-n1,-n0); negate(m0);
  goto fast_case_down;
  end;
sixth_octant:begin dx:=-8; edge_prep(-n1,-n0,-m1,-m0); negate(n0);
  goto slow_case_down;
  end;
seventh_octant:begin dx:=8; edge_prep(n0,n1,-m1,-m0); goto slow_case_down;
  end;
eighth_octant:begin dx:=8; edge_prep(m0,m1,-n1,-n0); goto fast_case_down;
  end;
end; {there are only eight octants}

@ @<Add edges for first or fourth octants, then |goto done|@>=
@<Move to row |n0|, pointed to by |p|@>;
if delta>0 then
  begin k:=0;
  edge_and_weight:=8*(m0+m_offset(cur_edges))+min_halfword+zero_w-cur_wt;
  repeat edge_and_weight:=edge_and_weight+dx*move[k];
  fast_get_avail(r); link(r):=unsorted(p); info(r):=edge_and_weight;
  if internal[tracing_edges]>0 then trace_new_edge(r,n);
  unsorted(p):=r; p:=link(p); incr(k); incr(n);
  until k=delta;
  end;
goto done

@ @<Add edges for fifth or eighth octants, then |goto done|@>=
n0:=-n0-1; @<Move to row |n0|, pointed to by |p|@>;
if delta>0 then
  begin k:=0;
  edge_and_weight:=8*(m0+m_offset(cur_edges))+min_halfword+zero_w+cur_wt;
  repeat edge_and_weight:=edge_and_weight+dx*move[k];
  fast_get_avail(r); link(r):=unsorted(p); info(r):=edge_and_weight;
  if internal[tracing_edges]>0 then trace_new_edge(r,n);
  unsorted(p):=r; p:=knil(p); incr(k); decr(n);
  until k=delta;
  end;
goto done

@ @<Add edges for second or third octants, then |goto done|@>=
edge_and_weight:=8*(n0+m_offset(cur_edges))+min_halfword+zero_w-cur_wt;
n0:=m0; k:=0; @<Move to row |n0|, pointed to by |p|@>;
repeat j:=move[k];
while j>0 do
  begin fast_get_avail(r); link(r):=unsorted(p); info(r):=edge_and_weight;
  if internal[tracing_edges]>0 then trace_new_edge(r,n);
  unsorted(p):=r; p:=link(p); decr(j); incr(n);
  end;
edge_and_weight:=edge_and_weight+dx; incr(k);
until k>delta;
goto done

@ @<Add edges for sixth or seventh octants, then |goto done|@>=
edge_and_weight:=8*(n0+m_offset(cur_edges))+min_halfword+zero_w+cur_wt;
n0:=-m0-1; k:=0; @<Move to row |n0|, pointed to by |p|@>;
repeat j:=move[k];
while j>0 do
  begin fast_get_avail(r); link(r):=unsorted(p); info(r):=edge_and_weight;
  if internal[tracing_edges]>0 then trace_new_edge(r,n);
  unsorted(p):=r; p:=knil(p); decr(j); decr(n);
  end;
edge_and_weight:=edge_and_weight+dx; incr(k);
until k>delta;
goto done

@ All the hard work of building an edge structure is undone by the following
subroutine.

@<Declare the recycling subroutines@>=
procedure toss_edges(@!h:pointer);
var @!p,@!q:pointer; {for list manipulation}
begin q:=link(h);
while q<>h do
  begin flush_list(sorted(q));
  if unsorted(q)>void then flush_list(unsorted(q));
  p:=q; q:=link(q); free_node(p,row_node_size);
  end;
free_node(h,edge_header_size);
end;

@* \[21] Subdivision into octants.
When \MF\ digitizes a path, it reduces the problem to the special
case of paths that travel in ``first octant'' directions; i.e.,
each cubic $z(t)=\bigl(x(t),y(t)\bigr)$ being digitized will have the property
that $0\L y'(t)\L x'(t)$. This assumption makes digitizing simpler
and faster than if the direction of motion has to be tested repeatedly.

When $z(t)$ is cubic, $x'(t)$ and $y'(t)$ are quadratic, hence the four
polynomials $x'(t)$, $y'(t)$, $x'(t)-y'(t)$, and $x'(t)+y'(t)$ cross
through~0 at most twice each. If we subdivide the given cubic at these
places, we get at most nine subintervals in each of which
$x'(t)$, $y'(t)$, $x'(t)-y'(t)$, and $x'(t)+y'(t)$ all have a constant
sign. The curve can be transformed in each of these subintervals so that
it travels entirely in first octant directions, if we reflect $x\swap-x$,
$y\swap-y$, and/or $x\swap y$ as necessary. (Incidentally, it can be
shown that a cubic such that $x'(t)=16(2t-1)^2+2(2t-1)-1$ and
$y'(t)=8(2t-1)^2+4(2t-1)$ does indeed split into nine subintervals.)

@ The transformation that rotates coordinates, so that first octant motion
can be assumed, is defined by the |skew| subroutine, which sets global
variables |cur_x| and |cur_y| to the values that are appropriate in a
given octant.  (Octants are encoded as they were in the |n_arg| subroutine.)

This transformation is ``skewed'' by replacing |(x,y)| by |(x-y,y)|,
once first octant motion has been established. It turns out that
skewed coordinates are somewhat better to work with when curves are
actually digitized.

@d set_two_end(#)==cur_y:=#;@+end
@d set_two(#)==begin cur_x:=#; set_two_end

@p procedure skew(@!x,@!y:scaled;@!octant:small_number);
begin case octant of
first_octant: set_two(x-y)(y);
second_octant: set_two(y-x)(x);
third_octant: set_two(y+x)(-x);
fourth_octant: set_two(-x-y)(y);
fifth_octant: set_two(-x+y)(-y);
sixth_octant: set_two(-y+x)(-x);
seventh_octant: set_two(-y-x)(x);
eighth_octant: set_two(x+y)(-y);
end; {there are no other cases}
end;

@ Conversely, the following subroutine sets |cur_x| and
|cur_y| to the original coordinate values of a point, given an octant
code and the point's coordinates |(x,y)| after they have been mapped into
the first octant and skewed.

@<Declare subroutines for printing expressions@>=
procedure unskew(@!x,@!y:scaled;@!octant:small_number);
begin case octant of
first_octant: set_two(x+y)(y);
second_octant: set_two(y)(x+y);
third_octant: set_two(-y)(x+y);
fourth_octant: set_two(-x-y)(y);
fifth_octant: set_two(-x-y)(-y);
sixth_octant: set_two(-y)(-x-y);
seventh_octant: set_two(y)(-x-y);
eighth_octant: set_two(x+y)(-y);
end; {there are no other cases}
end;

@ @<Glob...@>=
@!cur_x,@!cur_y:scaled;
  {outputs of |skew|, |unskew|, and a few other routines}

@ The conversion to skewed and rotated coordinates takes place in
stages, and at one point in the transformation we will have negated the
$x$ and/or $y$ coordinates so as to make curves travel in the first
{\sl quadrant}. At this point the relevant ``octant'' code will be
either |first_octant| (when no transformation has been done),
or |fourth_octant=first_octant+negate_x| (when $x$ has been negated),
or |fifth_octant=first_octant+negate_x+negate_y| (when both have been
negated), or |eighth_octant=first_octant+negate_y| (when $y$ has been
negated). The |abnegate| routine is sometimes needed to convert
from one of these transformations to another.

@p procedure abnegate(@!x,@!y:scaled;
  @!octant_before,@!octant_after:small_number);
begin if odd(octant_before)=odd(octant_after) then cur_x:=x
  else cur_x:=-x;
if (octant_before>negate_y)=(octant_after>negate_y) then cur_y:=y
  else cur_y:=-y;
end;

@ Now here's a subroutine that's handy for subdivision: Given a
quadratic polynomial $B(a,b,c;t)$, the |crossing_point| function
returns the unique |fraction| value |t| between 0 and~1 at which
$B(a,b,c;t)$ changes from positive to negative, or returns
|t=fraction_one+1| if no such value exists. If |a<0| (so that $B(a,b,c;t)$
is already negative at |t=0|), |crossing_point| returns the value zero.

@d no_crossing==begin crossing_point:=fraction_one+1; return;
  end
@d one_crossing==begin crossing_point:=fraction_one; return;
  end
@d zero_crossing==begin crossing_point:=0; return;
  end

@p function crossing_point(@!a,@!b,@!c:integer):fraction;
label exit;
var @!d:integer; {recursive counter}
@!x,@!xx,@!x0,@!x1,@!x2:integer; {temporary registers for bisection}
begin if a<0 then zero_crossing;
if c>=0 then
  begin if b>=0 then
    if c>0 then no_crossing
    else if (a=0)and(b=0) then no_crossing
    else one_crossing;
  if a=0 then zero_crossing;
  end
else if a=0 then if b<=0 then zero_crossing;
@<Use bisection to find the crossing point, if one exists@>;
exit:end;

@ The general bisection method is quite simple when $n=2$, hence
|crossing_point| does not take much time. At each stage in the
recursion we have a subinterval defined by |l| and~|j| such that
$B(a,b,c;2^{-l}(j+t))=B(x_0,x_1,x_2;t)$, and we want to ``zero in'' on
the subinterval where $x_0\G0$ and $\min(x_1,x_2)<0$.

It is convenient for purposes of calculation to combine the values
of |l| and~|j| in a single variable $d=2^l+j$, because the operation
of bisection then corresponds simply to doubling $d$ and possibly
adding~1. Furthermore it proves to be convenient to modify
our previous conventions for bisection slightly, maintaining the
variables $X_0=2^lx_0$, $X_1=2^l(x_0-x_1)$, and $X_2=2^l(x_1-x_2)$.
With these variables the conditions $x_0\ge0$ and $\min(x_1,x_2)<0$ are
equivalent to $\max(X_1,X_1+X_2)>X_0\ge0$.

The following code maintains the invariant relations
$0\L|x0|<\max(|x1|,|x1|+|x2|)$,
$\vert|x1|\vert<2^{30}$, $\vert|x2|\vert<2^{30}$;
it has been constructed in such a way that no arithmetic overflow
will occur if the inputs satisfy
$a<2^{30}$, $\vert a-b\vert<2^{30}$, and $\vert b-c\vert<2^{30}$.

@<Use bisection to find the crossing point...@>=
d:=1; x0:=a; x1:=a-b; x2:=b-c;
repeat x:=half(x1+x2);
if x1-x0>x0 then
  begin x2:=x; double(x0); double(d);
  end
else  begin xx:=x1+x-x0;
  if xx>x0 then
    begin x2:=x; double(x0); double(d);
    end
  else  begin x0:=x0-xx;
    if x<=x0 then if x+x2<=x0 then no_crossing;
    x1:=x; d:=d+d+1;
    end;
  end;
until d>=fraction_one;
crossing_point:=d-fraction_one

@ Octant subdivision is applied only to cycles, i.e., to closed paths.
A ``cycle spec'' is a data structure that contains specifications of
@!@^cycle spec@>
cubic curves and octant mappings for the cycle that has been subdivided
into segments belonging to single octants. It is composed entirely of
knot nodes, similar to those in the representation of paths; but the
|explicit| type indications have been replaced by positive numbers
that give further information. Additional |endpoint| data is also
inserted at the octant boundaries.

Recall that a cubic polynomial is represented by four control points
that appear in adjacent nodes |p| and~|q| of a knot list. The |x|~coordinates
are |x_coord(p)|, |right_x(p)|, |left_x(q)|, and |x_coord(q)|; the
|y|~coordinates are similar. We shall call this ``the cubic following~|p|''
or ``the cubic between |p| and~|q|'' or ``the cubic preceding~|q|.''

Cycle specs are circular lists of cubic curves mixed with octant
boundaries. Like cubics, the octant boundaries are represented in
consecutive knot nodes |p| and~|q|. In such cases |right_type(p)=
left_type(q)=endpoint|, and the fields |right_x(p)|, |right_y(p)|,
|left_x(q)|, and |left_y(q)| are replaced by other fields called
|right_octant(p)|, |right_transition(p)|, |left_octant(q)|, and
|left_transition(q)|, respectively. For example, when the curve direction
moves from the third octant to the fourth octant, the boundary nodes say
|right_octant(p)=third_octant|, |left_octant(q)=fourth_octant|,
and |right_transition(p)=left_transition(q)=diagonal|. A |diagonal|
transition occurs when moving between octants 1~\AM~2, 3~\AM~4, 5~\AM~6, or
7~\AM~8; an |axis| transition occurs when moving between octants 8~\AM~1,
2~\AM~3, 4~\AM~5, 6~\AM~7. (Such transition information is redundant
but convenient.) Fields |x_coord(p)| and |y_coord(p)| will contain
coordinates of the transition point after rotation from third octant
to first octant; i.e., if the true coordinates are $(x,y)$, the
coordinates $(y,-x)$ will appear in node~|p|. Similarly, a fourth-octant
transformation will have been applied after the transition, so
we will have |x_coord(q)=@t$-x$@>| and |y_coord(q)=y|.

The cubic between |p| and |q| will contain positive numbers in the
fields |right_type(p)| and |left_type(q)|; this makes cubics
distinguishable from octant boundaries, because |endpoint=0|.
The value of |right_type(p)| will be the current octant code,
during the time that cycle specs are being constructed; it will
refer later to a pen offset position, if the envelope of a cycle is
being computed. A cubic that comes from some subinterval of the $k$th
step in the original cyclic path will have |left_type(q)=k|.

@d right_octant==right_x {the octant code before a transition}
@d left_octant==left_x {the octant after a transition}
@d right_transition==right_y {the type of transition}
@d left_transition==left_y {ditto, either |axis| or |diagonal|}
@d axis=0 {a transition across the $x'$- or $y'$-axis}
@d diagonal=1 {a transition where $y'=\pm x'$}

@ Here's a routine that prints a cycle spec in symbolic form, so that it
is possible to see what subdivision has been made.  The point coordinates
are converted back from \MF's internal ``rotated'' form to the external
``true'' form. The global variable~|cur_spec| should point to a knot just
after the beginning of an octant boundary, i.e., such that
|left_type(cur_spec)=endpoint|.

@d print_two_true(#)==unskew(#,octant); print_two(cur_x,cur_y)

@p procedure print_spec(@!s:str_number);
label not_found,done;
var @!p,@!q:pointer; {for list traversal}
@!octant:small_number; {the current octant code}
begin print_diagnostic("Cycle spec",s,true);
@.Cycle spec at line...@>
p:=cur_spec; octant:=left_octant(p); print_ln;
print_two_true(x_coord(cur_spec),y_coord(cur_spec));
print(" % beginning in octant `");
loop@+  begin print(octant_dir[octant]); print_char("'");
  loop@+  begin q:=link(p);
    if right_type(p)=endpoint then goto not_found;
    @<Print the cubic between |p| and |q|@>;
    p:=q;
    end;
not_found: if q=cur_spec then goto done;
  p:=q; octant:=left_octant(p); print_nl("% entering octant `");
  end;
@.entering the nth octant@>
done: print_nl(" & cycle"); end_diagnostic(true);
end;

@ Symbolic octant direction names are kept in the |octant_dir| array.

@<Glob...@>=
@!octant_dir:array[first_octant..sixth_octant] of str_number;

@ @<Set init...@>=
octant_dir[first_octant]:="ENE";
octant_dir[second_octant]:="NNE";
octant_dir[third_octant]:="NNW";
octant_dir[fourth_octant]:="WNW";
octant_dir[fifth_octant]:="WSW";
octant_dir[sixth_octant]:="SSW";
octant_dir[seventh_octant]:="SSE";
octant_dir[eighth_octant]:="ESE";

@ @<Print the cubic between...@>=
begin print_nl("   ..controls ");
print_two_true(right_x(p),right_y(p));
print(" and ");
print_two_true(left_x(q),left_y(q));
print_nl(" ..");
print_two_true(x_coord(q),y_coord(q));
print(" % segment "); print_int(left_type(q)-1);
end

@ A much more compact version of a spec is printed to help users identify
``strange paths.''

@p procedure print_strange(@!s:str_number);
var @!p:pointer; {for list traversal}
@!f:pointer; {starting point in the cycle}
@!q:pointer; {octant boundary to be printed}
@!t:integer; {segment number, plus 1}
begin if interaction=error_stop_mode then wake_up_terminal;
print_nl(">");
@.>\relax@>
@<Find the starting point, |f|@>;
@<Determine the octant boundary |q| that precedes |f|@>;
t:=0;
repeat if left_type(p)<>endpoint then
  begin if left_type(p)<>t then
    begin t:=left_type(p); print_char(" "); print_int(t-1);
    end;
  if q<>null then
    begin @<Print the turns, if any, that start at |q|, and advance |q|@>;
    print_char(" "); print(octant_dir[left_octant(q)]); q:=null;
    end;
  end
else if q=null then q:=p;
p:=link(p);
until p=f;
print_char(" "); print_int(left_type(p)-1);
if q<>null then @<Print the turns...@>;
print_err(s);
end;

@ If the segment numbers on the cycle are $t_1$, $t_2$, \dots, $t_m$,
and if |m<=max_quarterword|,
we have $t_{k-1}\L t_k$ except for at most one value of~$k$. If there are
no exceptions, $f$ will point to $t_1$; otherwise it will point to the
exceptional~$t_k$.

There is at least one segment number (i.e., we always have $m>0$), because
|print_strange| is never called upon to display an entirely ``dead'' cycle.

@<Find the starting point, |f|@>=
p:=cur_spec; t:=max_quarterword+1;
repeat p:=link(p);
if left_type(p)<>endpoint then
  begin if left_type(p)<t then f:=p;
  t:=left_type(p);
  end;
until p=cur_spec

@ @<Determine the octant boundary...@>=
p:=cur_spec; q:=p;
repeat p:=link(p);
if left_type(p)=endpoint then q:=p;
until p=f

@ When two octant boundaries are adjacent, the path is simply changing direction
without moving. Such octant directions are shown in parentheses.

@<Print the turns...@>=
if left_type(link(q))=endpoint then
  begin print(" ("); print(octant_dir[left_octant(q)]); q:=link(q);
  while left_type(link(q))=endpoint do
    begin print_char(" "); print(octant_dir[left_octant(q)]); q:=link(q);
    end;
  print_char(")");
  end

@ The |make_spec| routine is what subdivides paths into octants:
Given a pointer |cur_spec| to a cyclic path, |make_spec| mungs the path data
and returns a pointer to the corresponding cyclic spec.
All ``dead'' cubics (i.e., cubics that don't move at all from
their starting points) will have been removed from the result.
@!@^dead cubics@>

The idea of |make_spec| is fairly simple: Each cubic is first
subdivided, if necessary, into pieces belonging to single octants;
then the octant boundaries are inserted. But some of the details of
this transformation are not quite obvious.

If |autorounding>0|, the path will be adjusted so that critical tangent
directions occur at ``good'' points with respect to the pen called |cur_pen|.

The resulting spec will have all |x| and |y| coordinates at most
$2^{28}-|half_unit|-1-|safety_margin|$ in absolute value.  The pointer
that is returned will start some octant, as required by |print_spec|.

@p @t\4@>@<Declare subroutines needed by |make_spec|@>@;
function make_spec(@!h:pointer;
  @!safety_margin:scaled;@!tracing:integer):pointer;
  {converts a path to a cycle spec}
label continue,done;
var @!p,@!q,@!r,@!s:pointer; {for traversing the lists}
@!k:integer; {serial number of path segment, or octant code}
@!chopped:integer; {positive if data truncated,
          negative if data dangerously large}
@<Other local variables for |make_spec|@>@;
begin cur_spec:=h;
if tracing>0 then
  print_path(cur_spec,", before subdivision into octants",true);
max_allowed:=fraction_one-half_unit-1-safety_margin;
@<Truncate the values of all coordinates that exceed |max_allowed|, and stamp
  segment numbers in each |left_type| field@>;
quadrant_subdivide; {subdivide each cubic into pieces belonging to quadrants}
if (internal[autorounding]>0)and(chopped=0) then xy_round;
octant_subdivide; {complete the subdivision}
if (internal[autorounding]>unity)and(chopped=0) then diag_round;
@<Remove dead cubics@>;
@<Insert octant boundaries and compute the turning number@>;
while left_type(cur_spec)<>endpoint do cur_spec:=link(cur_spec);
if tracing>0 then
  if (internal[autorounding]<=0)or(chopped<>0) then
    print_spec(", after subdivision")
  else if internal[autorounding]>unity then
    print_spec(", after subdivision and double autorounding")
  else print_spec(", after subdivision and autorounding");
make_spec:=cur_spec;
end;

@ The |make_spec| routine has an interesting side effect, namely to set
the global variable |turning_number| to the number of times the tangent
vector of the given cyclic path winds around the origin.

Another global variable |cur_spec| points to the specification as it is
being made, since several subroutines must go to work on it.

And there are two global variables that affect the rounding
decisions, as we'll see later; they are called |cur_pen| and |cur_path_type|.
The latter will be |double_path_code| if |make_spec| is being
applied to a double path.

@d double_path_code=0 {command modifier for `\&{doublepath}'}
@d contour_code=1 {command modifier for `\&{contour}'}
@d also_code=2 {command modifier for `\&{also}'}

@<Glob...@>=
@!cur_spec:pointer; {the principal output of |make_spec|}
@!turning_number:integer; {another output of |make_spec|}
@!cur_pen:pointer; {an implicit input of |make_spec|, used in autorounding}
@!cur_path_type:double_path_code..contour_code; {likewise}
@!max_allowed:scaled; {coordinates must be at most this big}

@ First we do a simple preprocessing step. The segment numbers inserted
here will propagate to all descendants of cubics that are split into
subintervals. These numbers must be nonzero, but otherwise they are
present merely for diagnostic purposes. The cubic from |p| to~|q|
that represents ``time interval'' |(t-1)..t| usually has |left_type(q)=t|,
except when |t| is too large to be stored in a quarterword.

@d procrustes(#)==@+if abs(#)>=dmax then
  if abs(#)>max_allowed then
    begin chopped:=1;
    if #>0 then #:=max_allowed@+else #:=-max_allowed;
    end
  else if chopped=0 then chopped:=-1

@<Truncate the values of all coordinates that exceed...@>=
p:=cur_spec; k:=1; chopped:=0; dmax:=half(max_allowed);
repeat procrustes(left_x(p)); procrustes(left_y(p));
procrustes(x_coord(p)); procrustes(y_coord(p));
procrustes(right_x(p)); procrustes(right_y(p));@/
p:=link(p); left_type(p):=k;
if k<max_quarterword then incr(k)@+else k:=1;
until p=cur_spec;
if chopped>0 then
  begin print_err("Curve out of range");
@.Curve out of range@>
  help4("At least one of the coordinates in the path I'm about to")@/
    ("digitize was really huge (potentially bigger than 4095).")@/
    ("So I've cut it back to the maximum size.")@/
    ("The results will probably be pretty wild.");
  put_get_error;
  end

@ We may need to get rid of constant ``dead'' cubics that clutter up
the data structure and interfere with autorounding.

@<Declare subroutines needed by |make_spec|@>=
procedure remove_cubic(@!p:pointer); {removes the cubic following~|p|}
var @!q:pointer; {the node that disappears}
begin q:=link(p); right_type(p):=right_type(q); link(p):=link(q);@/
x_coord(p):=x_coord(q); y_coord(p):=y_coord(q);@/
right_x(p):=right_x(q); right_y(p):=right_y(q);@/
free_node(q,knot_node_size);
end;

@ The subdivision process proceeds by first swapping $x\swap-x$, if
necessary, to ensure that $x'\G0$; then swapping $y\swap-y$, if necessary,
to ensure that $y'\G0$; and finally swapping $x\swap y$, if necessary,
to ensure that $x'\G y'$.

Recall that the octant codes have been defined in such a way that, for
example, |third_octant=first_octant+negate_x+switch_x_and_y|. The program
uses the fact that |negate_x<negate_y<switch_x_and_y| to handle ``double
negation'': If |c| is an octant code that possibly involves |negate_x|
and/or |negate_y|, but not |switch_x_and_y|, then negating~|y| changes~|c|
either to |c+negate_y| or |c-negate_y|, depending on whether
|c<=negate_y| or |c>negate_y|. Octant codes are always greater than zero.

The first step is to subdivide on |x| and |y| only, so that horizontal
and vertical autorounding can be done before we compare $x'$ to $y'$.

@<Declare subroutines needed by |make_spec|@>=
@t\4@>@<Declare the procedure called |split_cubic|@>@;
procedure quadrant_subdivide;
label continue,exit;
var @!p,@!q,@!r,@!s,@!pp,@!qq:pointer; {for traversing the lists}
@!first_x,@!first_y:scaled; {unnegated coordinates of node |cur_spec|}
@!del1,@!del2,@!del3,@!del,@!dmax:scaled; {proportional to the control
  points of a quadratic derived from a cubic}
@!t:fraction; {where a quadratic crosses zero}
@!dest_x,@!dest_y:scaled; {final values of |x| and |y| in the current cubic}
@!constant_x:boolean; {is |x| constant between |p| and |q|?}
begin p:=cur_spec; first_x:=x_coord(cur_spec); first_y:=y_coord(cur_spec);
repeat continue: q:=link(p);
@<Subdivide the cubic between |p| and |q| so that the results travel
  toward the right halfplane@>;
@<Subdivide all cubics between |p| and |q| so that the results travel
  toward the first quadrant; but |return| or |goto continue| if the
  cubic from |p| to |q| was dead@>;
p:=q;
until p=cur_spec;
exit:end;

@ All three subdivision processes are similar, so it's possible to
get the general idea by studying the first one (which is the simplest).
The calculation makes use of the fact that the derivatives of
Bernshte{\u\i}n polynomials satisfy
$B'(z_0,z_1,\ldots,z_n;t)=nB(z_1-z_0,\ldots,z_n-z_{n-1};t)$.

When this routine begins, |right_type(p)| is |explicit|; we should
set |right_type(p):=first_octant|. However, no assignment is made,
because |explicit=first_octant|. The author apologizes for using
such trickery here; it is really hard to do redundant computations
just for the sake of purity.

@<Subdivide the cubic between |p| and |q| so that the results travel
  toward the right halfplane...@>=
if q=cur_spec then
  begin dest_x:=first_x; dest_y:=first_y;
  end
else  begin dest_x:=x_coord(q); dest_y:=y_coord(q);
  end;
del1:=right_x(p)-x_coord(p); del2:=left_x(q)-right_x(p);
del3:=dest_x-left_x(q);
@<Scale up |del1|, |del2|, and |del3| for greater accuracy;
  also set |del| to the first nonzero element of |(del1,del2,del3)|@>;
if del=0 then constant_x:=true
else  begin constant_x:=false;
  if del<0 then @<Complement the |x| coordinates of the
    cubic between |p| and~|q|@>;
  t:=crossing_point(del1,del2,del3);
  if t<fraction_one then
    @<Subdivide the cubic with respect to $x'$, possibly twice@>;
  end

@ If |del1=del2=del3=0|, it's impossible to obey the title of this
section. We just set |del=0| in that case.
@^inner loop@>

@<Scale up |del1|, |del2|, and |del3| for greater accuracy...@>=
if del1<>0 then del:=del1
else if del2<>0 then del:=del2
else del:=del3;
if del<>0 then
  begin dmax:=abs(del1);
  if abs(del2)>dmax then dmax:=abs(del2);
  if abs(del3)>dmax then dmax:=abs(del3);
  while dmax<fraction_half do
    begin double(dmax); double(del1); double(del2); double(del3);
    end;
  end

@ During the subdivision phases of |make_spec|, the |x_coord| and |y_coord|
fields of node~|q| are not transformed to agree with the octant
stated in |right_type(p)|; they remain consistent with |right_type(q)|.
But |left_x(q)| and |left_y(q)| are governed by |right_type(p)|.

@<Complement the |x| coordinates...@>=
begin negate(x_coord(p)); negate(right_x(p));
negate(left_x(q));@/
negate(del1); negate(del2); negate(del3);@/
negate(dest_x);
right_type(p):=first_octant+negate_x;
end

@ When a cubic is split at a |fraction| value |t|, we obtain two cubics
whose B\'ezier control points are obtained by a generalization of the
bisection process: The formula
`$z_k^{(j+1)}={1\over2}(z_k^{(j)}+z\k^{(j)})$' becomes
`$z_k^{(j+1)}=t[z_k^{(j)},z\k^{(j)}]$'.

It is convenient to define a \.{WEB} macro |t_of_the_way| such that
|t_of_the_way(a)(b)| expands to |a-(a-b)*t|, i.e., to |t[a,b]|.

If |0<=t<=1|, the quantity |t[a,b]| is always between |a| and~|b|, even in
the presence of rounding errors. Our subroutines
also obey the identity |t[a,b]+t[b,a]=a+b|.

@d t_of_the_way_end(#)==#,t@=)@>
@d t_of_the_way(#)==#-take_fraction@=(@>#-t_of_the_way_end

@<Declare the procedure called |split_cubic|@>=
procedure split_cubic(@!p:pointer;@!t:fraction;
  @!xq,@!yq:scaled); {splits the cubic after |p|}
var @!v:scaled; {an intermediate value}
@!q,@!r:pointer; {for list manipulation}
begin q:=link(p); r:=get_node(knot_node_size); link(p):=r; link(r):=q;@/
left_type(r):=left_type(q); right_type(r):=right_type(p);@#
v:=t_of_the_way(right_x(p))(left_x(q));
right_x(p):=t_of_the_way(x_coord(p))(right_x(p));
left_x(q):=t_of_the_way(left_x(q))(xq);
left_x(r):=t_of_the_way(right_x(p))(v);
right_x(r):=t_of_the_way(v)(left_x(q));
x_coord(r):=t_of_the_way(left_x(r))(right_x(r));@#
v:=t_of_the_way(right_y(p))(left_y(q));
right_y(p):=t_of_the_way(y_coord(p))(right_y(p));
left_y(q):=t_of_the_way(left_y(q))(yq);
left_y(r):=t_of_the_way(right_y(p))(v);
right_y(r):=t_of_the_way(v)(left_y(q));
y_coord(r):=t_of_the_way(left_y(r))(right_y(r));
end;

@ Since $x'(t)$ is a quadratic equation, it can cross through zero
at~most twice. When it does cross zero, we make doubly sure that the
derivative is really zero at the splitting point, in case rounding errors
have caused the split cubic to have an apparently nonzero derivative.
We also make sure that the split cubic is monotonic.

@<Subdivide the cubic with respect to $x'$, possibly twice@>=
begin split_cubic(p,t,dest_x,dest_y); r:=link(p);
if right_type(r)>negate_x then right_type(r):=first_octant
else right_type(r):=first_octant+negate_x;
if x_coord(r)<x_coord(p) then x_coord(r):=x_coord(p);
left_x(r):=x_coord(r);
if right_x(p)>x_coord(r) then right_x(p):=x_coord(r);
 {we always have |x_coord(p)<=right_x(p)|}
negate(x_coord(r)); right_x(r):=x_coord(r);
negate(left_x(q)); negate(dest_x);@/
del2:=t_of_the_way(del2)(del3);
  {now |0,del2,del3| represent $x'$ on the remaining interval}
if del2>0 then del2:=0;
t:=crossing_point(0,-del2,-del3);
if t<fraction_one then @<Subdivide the cubic a second time
  with respect to $x'$@>
else begin if x_coord(r)>dest_x then
    begin x_coord(r):=dest_x; left_x(r):=-x_coord(r); right_x(r):=x_coord(r);
    end;
  if left_x(q)>dest_x then left_x(q):=dest_x
  else if left_x(q)<x_coord(r) then left_x(q):=x_coord(r);
  end;
end

@ @<Subdivide the cubic a second time with respect to $x'$@>=
begin split_cubic(r,t,dest_x,dest_y); s:=link(r);
if x_coord(s)<dest_x then x_coord(s):=dest_x;
if x_coord(s)<x_coord(r) then x_coord(s):=x_coord(r);
right_type(s):=right_type(p);
left_x(s):=x_coord(s); {now |x_coord(r)=right_x(r)<=left_x(s)|}
if left_x(q)<dest_x then left_x(q):=-dest_x
else if left_x(q)>x_coord(s) then left_x(q):=-x_coord(s)
else negate(left_x(q));
negate(x_coord(s)); right_x(s):=x_coord(s);
end

@ The process of subdivision with respect to $y'$ is like that with respect
to~$x'$, with the slight additional complication that two or three cubics
might now appear between |p| and~|q|.

@<Subdivide all cubics between |p| and |q| so that the results travel
  toward the first quadrant...@>=
pp:=p;
repeat qq:=link(pp);
abnegate(x_coord(qq),y_coord(qq),right_type(qq),right_type(pp));
dest_x:=cur_x; dest_y:=cur_y;@/
del1:=right_y(pp)-y_coord(pp); del2:=left_y(qq)-right_y(pp);
del3:=dest_y-left_y(qq);
@<Scale up |del1|, |del2|, and |del3| for greater accuracy;
  also set |del| to the first nonzero element of |(del1,del2,del3)|@>;
if del<>0 then {they weren't all zero}
  begin if del<0 then @<Complement the |y| coordinates of the
    cubic between |pp| and~|qq|@>;
  t:=crossing_point(del1,del2,del3);
  if t<fraction_one then
    @<Subdivide the cubic with respect to $y'$, possibly twice@>;
  end
else @<Do any special actions needed when |y| is constant;
  |return| or |goto continue| if a dead cubic from |p| to |q| is removed@>;
pp:=qq;
until pp=q;
if constant_x then @<Correct the octant code in segments with decreasing |y|@>

@ @<Complement the |y| coordinates...@>=
begin negate(y_coord(pp)); negate(right_y(pp));
negate(left_y(qq));@/
negate(del1); negate(del2); negate(del3);@/
negate(dest_y);
right_type(pp):=right_type(pp)+negate_y;
end

@ @<Subdivide the cubic with respect to $y'$, possibly twice@>=
begin split_cubic(pp,t,dest_x,dest_y); r:=link(pp);
if right_type(r)>negate_y then right_type(r):=right_type(r)-negate_y
else right_type(r):=right_type(r)+negate_y;
if y_coord(r)<y_coord(pp) then y_coord(r):=y_coord(pp);
left_y(r):=y_coord(r);
if right_y(pp)>y_coord(r) then right_y(pp):=y_coord(r);
 {we always have |y_coord(pp)<=right_y(pp)|}
negate(y_coord(r)); right_y(r):=y_coord(r);
negate(left_y(qq)); negate(dest_y);@/
if x_coord(r)<x_coord(pp) then x_coord(r):=x_coord(pp)
else if x_coord(r)>dest_x then x_coord(r):=dest_x;
if left_x(r)>x_coord(r) then
  begin left_x(r):=x_coord(r);
  if right_x(pp)>x_coord(r) then right_x(pp):=x_coord(r);
  end;
if right_x(r)<x_coord(r) then
  begin right_x(r):=x_coord(r);
  if left_x(qq)<x_coord(r) then left_x(qq):=x_coord(r);
  end;
del2:=t_of_the_way(del2)(del3);
  {now |0,del2,del3| represent $y'$ on the remaining interval}
if del2>0 then del2:=0;
t:=crossing_point(0,-del2,-del3);
if t<fraction_one then @<Subdivide the cubic a second time
  with respect to $y'$@>
else begin if y_coord(r)>dest_y then
    begin y_coord(r):=dest_y; left_y(r):=-y_coord(r); right_y(r):=y_coord(r);
    end;
  if left_y(qq)>dest_y then left_y(qq):=dest_y
  else if left_y(qq)<y_coord(r) then left_y(qq):=y_coord(r);
  end;
end

@ @<Subdivide the cubic a second time with respect to $y'$@>=
begin split_cubic(r,t,dest_x,dest_y); s:=link(r);@/
if y_coord(s)<dest_y then y_coord(s):=dest_y;
if y_coord(s)<y_coord(r) then y_coord(s):=y_coord(r);
right_type(s):=right_type(pp);
left_y(s):=y_coord(s); {now |y_coord(r)=right_y(r)<=left_y(s)|}
if left_y(qq)<dest_y then left_y(qq):=-dest_y
else if left_y(qq)>y_coord(s) then left_y(qq):=-y_coord(s)
else negate(left_y(qq));
negate(y_coord(s)); right_y(s):=y_coord(s);
if x_coord(s)<x_coord(r) then x_coord(s):=x_coord(r)
else if x_coord(s)>dest_x then x_coord(s):=dest_x;
if left_x(s)>x_coord(s) then
  begin left_x(s):=x_coord(s);
  if right_x(r)>x_coord(s) then right_x(r):=x_coord(s);
  end;
if right_x(s)<x_coord(s) then
  begin right_x(s):=x_coord(s);
  if left_x(qq)<x_coord(s) then left_x(qq):=x_coord(s);
  end;
end

@ If the cubic is constant in $y$ and increasing in $x$, we have classified
it as traveling in the first octant. If the cubic is constant
in~$y$ and decreasing in~$x$, it is desirable to classify it as traveling
in the fifth octant (not the fourth), because autorounding will be consistent
with respect to doublepaths only if the octant number changes by four when
the path is reversed. Therefore we negate the $y$~coordinates
when they are constant but the curve is decreasing in~$x$; this gives
the desired result except in pathological paths.

If the cubic is ``dead,'' i.e., constant in both |x| and |y|, we remove
it unless it is the only cubic in the entire path. We |goto continue|
if it wasn't the final cubic, so that the test |p=cur_spec| does not
falsely imply that all cubics have been processed.

@<Do any special actions needed when |y| is constant...@>=
if constant_x then {|p=pp|, |q=qq|, and the cubic is dead}
  begin if q<>p then
    begin remove_cubic(p); {remove the dead cycle and recycle node |q|}
    if cur_spec<>q then goto continue
    else  begin cur_spec:=p; return;
      end; {the final cubic was dead and is gone}
    end;
  end
else if not odd(right_type(pp)) then {the $x$ coordinates were negated}
  @<Complement the |y| coordinates...@>

@ A similar correction to octant codes deserves to be made when |x| is
constant and |y| is decreasing.

@<Correct the octant code in segments with decreasing |y|@>=
begin pp:=p;
repeat qq:=link(pp);
if right_type(pp)>negate_y then {the $y$ coordinates were negated}
  begin right_type(pp):=right_type(pp)+negate_x;
  negate(x_coord(pp)); negate(right_x(pp)); negate(left_x(qq));
  end;
pp:=qq;
until pp=q;
end

@ Finally, the process of subdividing to make $x'\G y'$ is like the other
two subdivisions, with a few new twists. We skew the coordinates at this time.

@<Declare subroutines needed by |make_spec|@>=
procedure octant_subdivide;
var @!p,@!q,@!r,@!s:pointer; {for traversing the lists}
@!del1,@!del2,@!del3,@!del,@!dmax:scaled; {proportional to the control
  points of a quadratic derived from a cubic}
@!t:fraction; {where a quadratic crosses zero}
@!dest_x,@!dest_y:scaled; {final values of |x| and |y| in the current cubic}
begin p:=cur_spec;
repeat q:=link(p);@/
x_coord(p):=x_coord(p)-y_coord(p);
right_x(p):=right_x(p)-right_y(p);
left_x(q):=left_x(q)-left_y(q);@/
@<Subdivide the cubic between |p| and |q| so that the results travel
  toward the first octant@>;
p:=q;
until p=cur_spec;
end;

@ @<Subdivide the cubic between |p| and |q| so that the results travel
  toward the first octant@>=
@<Set up the variables |(del1,del2,del3)| to represent $x'-y'$@>;
@<Scale up |del1|, |del2|, and |del3| for greater accuracy;
  also set |del| to the first nonzero element of |(del1,del2,del3)|@>;
if del<>0 then {they weren't all zero}
  begin if del<0 then @<Swap the |x| and |y| coordinates of the
    cubic between |p| and~|q|@>;
  t:=crossing_point(del1,del2,del3);
  if t<fraction_one then
    @<Subdivide the cubic with respect to $x'-y'$, possibly twice@>;
  end

@ @<Set up the variables |(del1,del2,del3)| to represent $x'-y'$@>=
if q=cur_spec then
  begin unskew(x_coord(q),y_coord(q),right_type(q));
  skew(cur_x,cur_y,right_type(p)); dest_x:=cur_x; dest_y:=cur_y;
  end
else  begin abnegate(x_coord(q),y_coord(q),right_type(q),right_type(p));
  dest_x:=cur_x-cur_y; dest_y:=cur_y;
  end;
del1:=right_x(p)-x_coord(p); del2:=left_x(q)-right_x(p);
del3:=dest_x-left_x(q)

@ The swapping here doesn't simply interchange |x| and |y| values,
because the coordinates are skewed. It turns out that this is easier
than ordinary swapping, because it can be done in two assignment statements
rather than three.

@ @<Swap the |x| and |y| coordinates...@>=
begin y_coord(p):=x_coord(p)+y_coord(p); negate(x_coord(p));@/
right_y(p):=right_x(p)+right_y(p); negate(right_x(p));@/
left_y(q):=left_x(q)+left_y(q); negate(left_x(q));@/
negate(del1); negate(del2); negate(del3);@/
dest_y:=dest_x+dest_y; negate(dest_x);@/
right_type(p):=right_type(p)+switch_x_and_y;
end

@ A somewhat tedious case analysis is carried out here to make sure that
nasty rounding errors don't destroy our assumptions of monotonicity.

@<Subdivide the cubic with respect to $x'-y'$, possibly twice@>=
begin split_cubic(p,t,dest_x,dest_y); r:=link(p);
if right_type(r)>switch_x_and_y then right_type(r):=right_type(r)-switch_x_and_y
else right_type(r):=right_type(r)+switch_x_and_y;
if y_coord(r)<y_coord(p) then y_coord(r):=y_coord(p)
else if y_coord(r)>dest_y then y_coord(r):=dest_y;
if x_coord(p)+y_coord(r)>dest_x+dest_y then
  y_coord(r):=dest_x+dest_y-x_coord(p);
if left_y(r)>y_coord(r) then
  begin left_y(r):=y_coord(r);
  if right_y(p)>y_coord(r) then right_y(p):=y_coord(r);
  end;
if right_y(r)<y_coord(r) then
  begin right_y(r):=y_coord(r);
  if left_y(q)<y_coord(r) then left_y(q):=y_coord(r);
  end;
if x_coord(r)<x_coord(p) then x_coord(r):=x_coord(p)
else if x_coord(r)+y_coord(r)>dest_x+dest_y then
  x_coord(r):=dest_x+dest_y-y_coord(r);
left_x(r):=x_coord(r);
if right_x(p)>x_coord(r) then right_x(p):=x_coord(r);
 {we always have |x_coord(p)<=right_x(p)|}
y_coord(r):=y_coord(r)+x_coord(r); right_y(r):=right_y(r)+x_coord(r);@/
negate(x_coord(r)); right_x(r):=x_coord(r);@/
left_y(q):=left_y(q)+left_x(q); negate(left_x(q));@/
dest_y:=dest_y+dest_x; negate(dest_x);
if right_y(r)<y_coord(r) then
  begin right_y(r):=y_coord(r);
  if left_y(q)<y_coord(r) then left_y(q):=y_coord(r);
  end;
del2:=t_of_the_way(del2)(del3);
  {now |0,del2,del3| represent $x'-y'$ on the remaining interval}
if del2>0 then del2:=0;
t:=crossing_point(0,-del2,-del3);
if t<fraction_one then
  @<Subdivide the cubic a second time with respect to $x'-y'$@>
else begin if x_coord(r)>dest_x then
    begin x_coord(r):=dest_x; left_x(r):=-x_coord(r); right_x(r):=x_coord(r);
    end;
  if left_x(q)>dest_x then left_x(q):=dest_x
  else if left_x(q)<x_coord(r) then left_x(q):=x_coord(r);
  end;
end

@ @<Subdivide the cubic a second time with respect to $x'-y'$@>=
begin split_cubic(r,t,dest_x,dest_y); s:=link(r);@/
if y_coord(s)<y_coord(r) then y_coord(s):=y_coord(r)
else if y_coord(s)>dest_y then y_coord(s):=dest_y;
if x_coord(r)+y_coord(s)>dest_x+dest_y then
  y_coord(s):=dest_x+dest_y-x_coord(r);
if left_y(s)>y_coord(s) then
  begin left_y(s):=y_coord(s);
  if right_y(r)>y_coord(s) then right_y(r):=y_coord(s);
  end;
if right_y(s)<y_coord(s) then
  begin right_y(s):=y_coord(s);
  if left_y(q)<y_coord(s) then left_y(q):=y_coord(s);
  end;
if x_coord(s)+y_coord(s)>dest_x+dest_y then x_coord(s):=dest_x+dest_y-y_coord(s)
else begin if x_coord(s)<dest_x then x_coord(s):=dest_x;
  if x_coord(s)<x_coord(r) then x_coord(s):=x_coord(r);
  end;
right_type(s):=right_type(p);
left_x(s):=x_coord(s); {now |x_coord(r)=right_x(r)<=left_x(s)|}
if left_x(q)<dest_x then
  begin left_y(q):=left_y(q)+dest_x; left_x(q):=-dest_x;@+end
else if left_x(q)>x_coord(s) then
  begin left_y(q):=left_y(q)+x_coord(s); left_x(q):=-x_coord(s);@+end
else begin left_y(q):=left_y(q)+left_x(q); negate(left_x(q));@+end;
y_coord(s):=y_coord(s)+x_coord(s); right_y(s):=right_y(s)+x_coord(s);@/
negate(x_coord(s)); right_x(s):=x_coord(s);@/
if right_y(s)<y_coord(s) then
  begin right_y(s):=y_coord(s);
  if left_y(q)<y_coord(s) then left_y(q):=y_coord(s);
  end;
end

@ It's time now to consider ``autorounding,'' which tries to make horizontal,
vertical, and diagonal tangents occur at places that will produce appropriate
images after the curve is digitized.

The first job is to fix things so that |x(t)| plus the horizontal pen offset
is an integer multiple of the
current ``granularity'' when the derivative $x'(t)$ crosses through zero.
The given cyclic path contains regions where $x'(t)\G0$ and regions
where $x'(t)\L0$. The |quadrant_subdivide| routine is called into action
before any of the path coordinates have been skewed, but some of them
may have been negated. In regions where $x'(t)\G0$ we have |right_type=
first_octant| or |right_type=eighth_octant|; in regions where $x'(t)\L0$,
we have |right_type=fifth_octant| or |right_type=fourth_octant|.

Within any such region the transformed $x$ values increase monotonically
from, say, $x_0$ to~$x_1$. We want to modify things by applying a linear
transformation to all $x$ coordinates in the region, after which
the $x$ values will increase monotonically from round$(x_0)$ to round$(x_1)$.

This rounding scheme sounds quite simple, and it usually is. But several
complications can arise that might make the task more difficult. In the
first place, autorounding is inappropriate at cusps where $x'$ jumps
discontinuously past zero without ever being zero. In the second place,
the current pen might be unsymmetric in such a way that $x$ coordinates
should round differently in different parts of the curve.
These considerations imply that round$(x_0)$ might be greater
than round$(x_1)$, even though $x_0\L x_1$; in such cases we do not want
to carry out the linear transformation. Furthermore, it's possible to have
round$(x_1)-\hbox{round} (x_0)$ positive but much greater than $x_1-x_0$;
then the transformation might distort the curve drastically, and again we
want to avoid it. Finally, the rounded points must be consistent between
adjacent regions, hence we can't transform one region without knowing
about its neighbors.

To handle all these complications, we must first look at the whole
cycle and choose rounded $x$ values that are ``safe.'' The following
procedure does this: Given $m$~values $(b_0,b_1,\ldots,b_{m-1})$ before
rounding and $m$~corresponding values $(a_0,a_1,\ldots,a_{m-1})$ that would
be desirable after rounding, the |make_safe| routine sets $a$'s to $b$'s
if necessary so that $0\L(a\k-a_k)/(b\k-b_k)\L2$ afterwards. It is
symmetric under cyclic permutation, reversal, and/or negation of the inputs.
(Instead of |a|, |b|, and~|m|, the program uses the names |after|,
|before|, and |cur_rounding_ptr|.)

@<Declare subroutines needed by |make_spec|@>=
procedure make_safe;
var @!k:0..max_wiggle; {runs through the list of inputs}
@!all_safe:boolean; {does everything look OK so far?}
@!next_a:scaled; {|after[k]| before it might have changed}
@!delta_a,@!delta_b:scaled; {|after[k+1]-after[k]| and |before[k+1]-before[k]|}
begin before[cur_rounding_ptr]:=before[0]; {wrap around}
node_to_round[cur_rounding_ptr]:=node_to_round[0];
repeat after[cur_rounding_ptr]:=after[0]; all_safe:=true; next_a:=after[0];
for k:=0 to cur_rounding_ptr-1 do
  begin delta_b:=before[k+1]-before[k];
  if delta_b>=0 then delta_a:=after[k+1]-next_a
  else delta_a:=next_a-after[k+1];
  next_a:=after[k+1];
  if (delta_a<0)or(delta_a>abs(delta_b+delta_b)) then
    begin all_safe:=false; after[k]:=before[k];
    if k=cur_rounding_ptr-1 then after[0]:=before[0]
    else after[k+1]:=before[k+1];
    end;
  end;
until all_safe;
end;

@ The global arrays used by |make_safe| are accompanied by an array of
pointers into the current knot list.

@<Glob...@>=
@!before,@!after:array[0..max_wiggle] of scaled; {data for |make_safe|}
@!node_to_round:array[0..max_wiggle] of pointer; {reference back to the path}
@!cur_rounding_ptr:0..max_wiggle; {how many are being used}
@!max_rounding_ptr:0..max_wiggle; {how many have been used}

@ @<Set init...@>=
max_rounding_ptr:=0;

@ New entries go into the tables via the |before_and_after| routine:

@<Declare subroutines needed by |make_spec|@>=
procedure before_and_after(@!b,@!a:scaled;@!p:pointer);
begin if cur_rounding_ptr=max_rounding_ptr then
  if max_rounding_ptr<max_wiggle then incr(max_rounding_ptr)
  else overflow("rounding table size",max_wiggle);
@:METAFONT capacity exceeded rounding table size}{\quad rounding table size@>
after[cur_rounding_ptr]:=a; before[cur_rounding_ptr]:=b;
node_to_round[cur_rounding_ptr]:=p; incr(cur_rounding_ptr);
end;

@ A global variable called |cur_gran| is used instead of |internal[
granularity]|, because we want to work with a number that's guaranteed to
be positive.

@<Glob...@>=
@!cur_gran:scaled; {the current granularity (which normally is |unity|)}

@ The |good_val| function computes a number |a| that's as close as
possible to~|b|, with the property that |a+o| is a multiple of
|cur_gran|.

If we assume that |cur_gran| is even (since it will in fact be a multiple
of |unity| in all reasonable applications), we have the identity
|good_val(-b-1,-o)=-good_val(b,o)|.

@<Declare subroutines needed by |make_spec|@>=
function good_val(@!b,@!o:scaled):scaled;
var @!a:scaled; {accumulator}
begin a:=b+o;
if a>=0 then a:=a-(a mod cur_gran)-o
else a:=a+((-(a+1)) mod cur_gran)-cur_gran+1-o;
if b-a<a+cur_gran-b then good_val:=a
else good_val:=a+cur_gran;
end;

@ When we're rounding a doublepath, we might need to compromise between
two opposing tendencies, if the pen thickness is not a multiple of the
granularity. The following ``compromise'' adjustment, suggested by
John Hobby, finds the best way out of the dilemma. (Only the value
@^Hobby, John Douglas@>
modulo |cur_gran| is relevant in our applications, so the result turns
out to be essentially symmetric in |u| and~|v|.)

@<Declare subroutines needed by |make_spec|@>=
function compromise(@!u,@!v:scaled):scaled;
begin compromise:=half(good_val(u+u,-u-v));
end;

@ Here, then, is the procedure that rounds $x$ coordinates as described;
it does the same for $y$ coordinates too, independently.

@<Declare subroutines needed by |make_spec|@>=
procedure xy_round;
var @!p,@!q:pointer; {list manipulation registers}
@!b,@!a:scaled; {before and after values}
@!pen_edge:scaled; {offset that governs rounding}
@!alpha:fraction; {coefficient of linear transformation}
begin cur_gran:=abs(internal[granularity]);
if cur_gran=0 then cur_gran:=unity;
p:=cur_spec; cur_rounding_ptr:=0;
repeat q:=link(p);
@<If node |q| is a transition point for |x| coordinates,
  compute and save its before-and-after coordinates@>;
p:=q;
until p=cur_spec;
if cur_rounding_ptr>0 then @<Transform the |x| coordinates@>;
p:=cur_spec; cur_rounding_ptr:=0;
repeat q:=link(p);
@<If node |q| is a transition point for |y| coordinates,
  compute and save its before-and-after coordinates@>;
p:=q;
until p=cur_spec;
if cur_rounding_ptr>0 then @<Transform the |y| coordinates@>;
end;

@ When |x| has been negated, the |octant| codes are even. We allow
for an error of up to .01 pixel (i.e., 655 |scaled| units) in the
derivative calculations at transition nodes.

@<If node |q| is a transition point for |x| coordinates...@>=
if odd(right_type(p))<>odd(right_type(q)) then
  begin if odd(right_type(q)) then b:=x_coord(q)@+else b:=-x_coord(q);
  if (abs(x_coord(q)-right_x(q))<655)or@|
    (abs(x_coord(q)+left_x(q))<655) then
    @<Compute before-and-after |x| values based on the current pen@>
  else a:=b;
  if abs(a)>max_allowed then
    if a>0 then a:=max_allowed@+else a:=-max_allowed;
  before_and_after(b,a,q);
  end

@ When we study the data representation for pens, we'll learn that the
|x|~coordinate of the current pen's west edge is
$$\hbox{|y_coord(link(cur_pen+seventh_octant))|},$$
and that there are similar ways to address other important offsets.

@d north_edge(#)==y_coord(link(#+fourth_octant))
@d south_edge(#)==y_coord(link(#+first_octant))
@d east_edge(#)==y_coord(link(#+second_octant))
@d west_edge(#)==y_coord(link(#+seventh_octant))

@<Compute before-and-after |x| values based on the current pen@>=
begin if cur_pen=null_pen then pen_edge:=0
else if cur_path_type=double_path_code then
  pen_edge:=compromise(east_edge(cur_pen),west_edge(cur_pen))
else if odd(right_type(q)) then pen_edge:=west_edge(cur_pen)
else pen_edge:=east_edge(cur_pen);
a:=good_val(b,pen_edge);
end

@  The monotone transformation computed here with fixed-point arithmetic is
guaranteed to take consecutive |before| values $(b,b')$ into consecutive
|after| values $(a,a')$, even in the presence of rounding errors,
as long as $\vert b-b'\vert<2^{28}$.

@<Transform the |x| coordinates@>=
begin make_safe;
repeat decr(cur_rounding_ptr);
if (after[cur_rounding_ptr]<>before[cur_rounding_ptr])or@|
 (after[cur_rounding_ptr+1]<>before[cur_rounding_ptr+1]) then
  begin p:=node_to_round[cur_rounding_ptr];
  if odd(right_type(p)) then
    begin b:=before[cur_rounding_ptr]; a:=after[cur_rounding_ptr];
    end
  else  begin b:=-before[cur_rounding_ptr]; a:=-after[cur_rounding_ptr];
    end;
  if before[cur_rounding_ptr]=before[cur_rounding_ptr+1] then
    alpha:=fraction_one
  else alpha:=make_fraction(after[cur_rounding_ptr+1]-after[cur_rounding_ptr],@|
    before[cur_rounding_ptr+1]-before[cur_rounding_ptr]);
  repeat x_coord(p):=take_fraction(alpha,x_coord(p)-b)+a;
  right_x(p):=take_fraction(alpha,right_x(p)-b)+a;
  p:=link(p); left_x(p):=take_fraction(alpha,left_x(p)-b)+a;
  until p=node_to_round[cur_rounding_ptr+1];
  end;
until cur_rounding_ptr=0;
end

@ When |y| has been negated, the |octant| codes are |>negate_y|. Otherwise
these routines are essentially identical to the routines for |x| coordinates
that we have just seen.

@<If node |q| is a transition point for |y| coordinates...@>=
if (right_type(p)>negate_y)<>(right_type(q)>negate_y) then
  begin if right_type(q)<=negate_y then b:=y_coord(q)@+else b:=-y_coord(q);
  if (abs(y_coord(q)-right_y(q))<655)or@|
    (abs(y_coord(q)+left_y(q))<655) then
    @<Compute before-and-after |y| values based on the current pen@>
  else a:=b;
  if abs(a)>max_allowed then
    if a>0 then a:=max_allowed@+else a:=-max_allowed;
  before_and_after(b,a,q);
  end

@ @<Compute before-and-after |y| values based on the current pen@>=
begin if cur_pen=null_pen then pen_edge:=0
else if cur_path_type=double_path_code then
  pen_edge:=compromise(north_edge(cur_pen),south_edge(cur_pen))
else if right_type(q)<=negate_y then pen_edge:=south_edge(cur_pen)
else pen_edge:=north_edge(cur_pen);
a:=good_val(b,pen_edge);
end

@ @<Transform the |y| coordinates@>=
begin make_safe;
repeat decr(cur_rounding_ptr);
if (after[cur_rounding_ptr]<>before[cur_rounding_ptr])or@|
 (after[cur_rounding_ptr+1]<>before[cur_rounding_ptr+1]) then
  begin p:=node_to_round[cur_rounding_ptr];
  if right_type(p)<=negate_y then
    begin b:=before[cur_rounding_ptr]; a:=after[cur_rounding_ptr];
    end
  else  begin b:=-before[cur_rounding_ptr]; a:=-after[cur_rounding_ptr];
    end;
  if before[cur_rounding_ptr]=before[cur_rounding_ptr+1] then
    alpha:=fraction_one
  else alpha:=make_fraction(after[cur_rounding_ptr+1]-after[cur_rounding_ptr],@|
    before[cur_rounding_ptr+1]-before[cur_rounding_ptr]);
  repeat y_coord(p):=take_fraction(alpha,y_coord(p)-b)+a;
  right_y(p):=take_fraction(alpha,right_y(p)-b)+a;
  p:=link(p); left_y(p):=take_fraction(alpha,left_y(p)-b)+a;
  until p=node_to_round[cur_rounding_ptr+1];
  end;
until cur_rounding_ptr=0;
end

@ Rounding at diagonal tangents takes place after the subdivision into
octants is complete, hence after the coordinates have been skewed.
The details are somewhat tricky, because we want to round to points
whose skewed coordinates are halfway between integer multiples of
the granularity. Furthermore, both coordinates change when they are
rounded; this means we need a generalization of the |make_safe| routine,
ensuring safety in both |x| and |y|.

In spite of these extra complications, we can take comfort in the fact
that the basic structure of the routine is the same as before.

@<Declare subroutines needed by |make_spec|@>=
procedure diag_round;
var @!p,@!q,@!pp:pointer; {list manipulation registers}
@!b,@!a,@!bb,@!aa,@!d,@!c,@!dd,@!cc:scaled; {before and after values}
@!pen_edge:scaled; {offset that governs rounding}
@!alpha,@!beta:fraction; {coefficients of linear transformation}
@!next_a:scaled; {|after[k]| before it might have changed}
@!all_safe:boolean; {does everything look OK so far?}
@!k:0..max_wiggle; {runs through before-and-after values}
@!first_x,@!first_y:scaled; {coordinates before rounding}
begin p:=cur_spec; cur_rounding_ptr:=0;
repeat q:=link(p);
@<If node |q| is a transition point between octants,
  compute and save its before-and-after coordinates@>;
p:=q;
until p=cur_spec;
if cur_rounding_ptr>0 then @<Transform the skewed coordinates@>;
end;

@ We negate the skewed |x| coordinates in the before-and-after table when
the octant code is greater than |switch_x_and_y|.

@<If node |q| is a transition point between octants...@>=
if right_type(p)<>right_type(q) then
  begin if right_type(q)>switch_x_and_y then b:=-x_coord(q)
  else b:=x_coord(q);
  if abs(right_type(q)-right_type(p))=switch_x_and_y then
    if (abs(x_coord(q)-right_x(q))<655)or(abs(x_coord(q)+left_x(q))<655) then
      @<Compute a good coordinate at a diagonal transition@>
    else a:=b
  else a:=b;
  before_and_after(b,a,q);
  end

@ In octants whose code number is even, $x$~has been
negated; we want to round ambiguous cases downward instead of upward,
so that the rounding will be consistent with octants whose code
number is odd. This downward bias can be achieved by
subtracting~1 from the first argument of |good_val|.

@d diag_offset(#)==x_coord(knil(link(cur_pen+#)))

@<Compute a good coordinate at a diagonal transition@>=
begin if cur_pen=null_pen then pen_edge:=0
else if cur_path_type=double_path_code then @<Compute a compromise |pen_edge|@>
else if right_type(q)<=switch_x_and_y then pen_edge:=diag_offset(right_type(q))
else pen_edge:=-diag_offset(right_type(q));
if odd(right_type(q)) then a:=good_val(b,pen_edge+half(cur_gran))
else a:=good_val(b-1,pen_edge+half(cur_gran));
end

@ (It seems a shame to compute these compromise offsets repeatedly. The
author would have stored them directly in the pen data structure, if the
granularity had been constant.)

@<Compute a compromise...@>=
case right_type(q) of
first_octant,second_octant:pen_edge:=compromise(diag_offset(first_octant),@|
    -diag_offset(fifth_octant));
fifth_octant,sixth_octant:pen_edge:=-compromise(diag_offset(first_octant),@|
    -diag_offset(fifth_octant));
third_octant,fourth_octant:pen_edge:=compromise(diag_offset(fourth_octant),@|
    -diag_offset(eighth_octant));
seventh_octant,eighth_octant:pen_edge:=-compromise(diag_offset(fourth_octant),@|
    -diag_offset(eighth_octant));
end {there are no other cases}

@ @<Transform the skewed coordinates@>=
begin p:=node_to_round[0]; first_x:=x_coord(p); first_y:=y_coord(p);
@<Make sure that all the diagonal roundings are safe@>;
for k:=0 to cur_rounding_ptr-1 do
  begin a:=after[k]; b:=before[k];
  aa:=after[k+1]; bb:=before[k+1];
  if (a<>b)or(aa<>bb) then
    begin p:=node_to_round[k]; pp:=node_to_round[k+1];
    @<Determine the before-and-after values of both coordinates@>;
    if b=bb then alpha:=fraction_one
    else alpha:=make_fraction(aa-a,bb-b);
    if d=dd then beta:=fraction_one
    else beta:=make_fraction(cc-c,dd-d);
    repeat x_coord(p):=take_fraction(alpha,x_coord(p)-b)+a;
    y_coord(p):=take_fraction(beta,y_coord(p)-d)+c;
    right_x(p):=take_fraction(alpha,right_x(p)-b)+a;
    right_y(p):=take_fraction(beta,right_y(p)-d)+c;
    p:=link(p); left_x(p):=take_fraction(alpha,left_x(p)-b)+a;
    left_y(p):=take_fraction(beta,left_y(p)-d)+c;
    until p=pp;
    end;
  end;
end

@ In node |p|, the coordinates |(b,d)| will be rounded to |(a,c)|;
in node |pp|, the coordinates |(bb,dd)| will be rounded to |(aa,cc)|.
(We transform the values from node |pp| so that they agree with the
conventions of node |p|.)

If |aa<>bb|, we know that |abs(right_type(p)-right_type(pp))=switch_x_and_y|.

@<Determine the before-and-after values of both coordinates@>=
if aa=bb then
  begin if pp=node_to_round[0] then
    unskew(first_x,first_y,right_type(pp))
  else unskew(x_coord(pp),y_coord(pp),right_type(pp));
  skew(cur_x,cur_y,right_type(p));
  bb:=cur_x; aa:=bb; dd:=cur_y; cc:=dd;
  if right_type(p)>switch_x_and_y then
    begin b:=-b; a:=-a;
    end;
  end
else  begin if right_type(p)>switch_x_and_y then
    begin bb:=-bb; aa:=-aa; b:=-b; a:=-a;
    end;
  if pp=node_to_round[0] then dd:=first_y-bb@+else dd:=y_coord(pp)-bb;
  if odd(aa-bb) then
    if right_type(p)>switch_x_and_y then cc:=dd-half(aa-bb+1)
    else cc:=dd-half(aa-bb-1)
  else cc:=dd-half(aa-bb);
  end;
d:=y_coord(p);
if odd(a-b) then
  if right_type(p)>switch_x_and_y then c:=d-half(a-b-1)
  else c:=d-half(a-b+1)
else c:=d-half(a-b)

@ @<Make sure that all the diagonal roundings are safe@>=
before[cur_rounding_ptr]:=before[0]; {cf.~|make_safe|}
node_to_round[cur_rounding_ptr]:=node_to_round[0];
repeat after[cur_rounding_ptr]:=after[0]; all_safe:=true; next_a:=after[0];
for k:=0 to cur_rounding_ptr-1 do
  begin a:=next_a; b:=before[k]; next_a:=after[k+1];
  aa:=next_a; bb:=before[k+1];
  if (a<>b)or(aa<>bb) then
    begin p:=node_to_round[k]; pp:=node_to_round[k+1];
    @<Determine the before-and-after values of both coordinates@>;
    if (aa<a)or(cc<c)or(aa-a>2*(bb-b))or(cc-c>2*(dd-d)) then
      begin all_safe:=false; after[k]:=before[k];
      if k=cur_rounding_ptr-1 then after[0]:=before[0]
      else after[k+1]:=before[k+1];
      end;
    end;
  end;
until all_safe

@ Here we get rid of ``dead'' cubics, i.e., polynomials that don't move at
all when |t|~changes, since the subdivision process might have introduced
such things.  If the cycle reduces to a single point, however, we are left
with a single dead cubic that will not be removed until later.

@<Remove dead cubics@>=
p:=cur_spec;
repeat continue: q:=link(p);
if p<>q then
  begin if x_coord(p)=right_x(p) then
   if y_coord(p)=right_y(p) then
    if x_coord(p)=left_x(q) then
     if y_coord(p)=left_y(q) then
    begin unskew(x_coord(q),y_coord(q),right_type(q));
    skew(cur_x,cur_y,right_type(p));
    if x_coord(p)=cur_x then if y_coord(p)=cur_y then
      begin remove_cubic(p); {remove the cubic following |p|}
      if q<>cur_spec then goto continue;
      cur_spec:=p; q:=p;
      end;
    end;
  end;
p:=q;
until p=cur_spec;

@ Finally we come to the last steps of |make_spec|, when boundary nodes
are inserted between cubics that move in different octants. The main
complication remaining arises from consecutive cubics whose octants
are not adjacent; we should insert more than one octant boundary
at such sharp turns, so that the envelope-forming routine will work.

For this purpose, conversion tables between numeric and Gray codes for
octants are desirable.

@<Glob...@>=
@!octant_number:array[first_octant..sixth_octant] of 1..8;
@!octant_code:array[1..8] of first_octant..sixth_octant;

@ @<Set init...@>=
octant_code[1]:=first_octant;
octant_code[2]:=second_octant;
octant_code[3]:=third_octant;
octant_code[4]:=fourth_octant;
octant_code[5]:=fifth_octant;
octant_code[6]:=sixth_octant;
octant_code[7]:=seventh_octant;
octant_code[8]:=eighth_octant;
for k:=1 to 8 do octant_number[octant_code[k]]:=k;

@ The main loop for boundary insertion deals with three consecutive
nodes |p,q,r|.

@<Insert octant boundaries and compute the turning number@>=
turning_number:=0;
p:=cur_spec; q:=link(p);
repeat r:=link(q);
if (right_type(p)<>right_type(q))or(q=r) then
  @<Insert one or more octant boundary nodes just before~|q|@>;
p:=q; q:=r;
until p=cur_spec;

@ The |new_boundary| subroutine comes in handy at this point. It inserts
a new boundary node just after a given node |p|, using a given octant code
to transform the new node's coordinates. The ``transition'' fields are
not computed here.

@<Declare subroutines needed by |make_spec|@>=
procedure new_boundary(@!p:pointer;@!octant:small_number);
var @!q,@!r:pointer; {for list manipulation}
begin q:=link(p); {we assume that |right_type(q)<>endpoint|}
r:=get_node(knot_node_size); link(r):=q; link(p):=r;
left_type(r):=left_type(q); {but possibly |left_type(q)=endpoint|}
left_x(r):=left_x(q); left_y(r):=left_y(q);
right_type(r):=endpoint; left_type(q):=endpoint;
right_octant(r):=octant; left_octant(q):=right_type(q);
unskew(x_coord(q),y_coord(q),right_type(q));
skew(cur_x,cur_y,octant); x_coord(r):=cur_x; y_coord(r):=cur_y;
end;

@ The case |q=r| occurs if and only if |p=q=r=cur_spec|, when we want to turn
$360^\circ$ in eight steps and then remove a solitary dead cubic.
The program below happens to work in that case, but the reader isn't
expected to understand why.

@<Insert one or more octant boundary nodes just before~|q|@>=
begin new_boundary(p,right_type(p)); s:=link(p);
o1:=octant_number[right_type(p)]; o2:=octant_number[right_type(q)];
case o2-o1 of
1,-7,7,-1: goto done;
2,-6: clockwise:=false;
3,-5,4,-4,5,-3: @<Decide whether or not to go clockwise@>;
6,-2: clockwise:=true;
0:clockwise:=rev_turns;
end; {there are no other cases}
@<Insert additional boundary nodes, then |goto done|@>;
done: if q=r then
  begin q:=link(q); r:=q; p:=s; link(s):=q; left_octant(q):=right_octant(q);
  left_type(q):=endpoint; free_node(cur_spec,knot_node_size); cur_spec:=q;
  end;
@<Fix up the transition fields and adjust the turning number@>;
end

@ @<Other local variables for |make_spec|@>=
@!o1,@!o2:small_number; {octant numbers}
@!clockwise:boolean; {should we turn clockwise?}
@!dx1,@!dy1,@!dx2,@!dy2:integer; {directions of travel at a cusp}
@!dmax,@!del:integer; {temporary registers}

@ A tricky question arises when a path jumps four octants. We want the
direction of turning to be counterclockwise if the curve has changed
direction by $180^\circ$, or by something so close to $180^\circ$ that
the difference is probably due to rounding errors; otherwise we want to
turn through an angle of less than $180^\circ$. This decision needs to
be made even when a curve seems to have jumped only three octants, since
a curve may approach direction $(-1,0)$ from the fourth octant, then
it might leave from direction $(+1,0)$ into the first.

The following code solves the problem by analyzing the incoming
direction |(dx1,dy1)| and the outgoing direction |(dx2,dy2)|.

@<Decide whether or not to go clockwise@>=
begin @<Compute the incoming and outgoing directions@>;
unskew(dx1,dy1,right_type(p)); del:=pyth_add(cur_x,cur_y);@/
dx1:=make_fraction(cur_x,del); dy1:=make_fraction(cur_y,del);
  {$\cos\theta_1$ and $\sin\theta_1$}
unskew(dx2,dy2,right_type(q)); del:=pyth_add(cur_x,cur_y);@/
dx2:=make_fraction(cur_x,del); dy2:=make_fraction(cur_y,del);
  {$\cos\theta_2$ and $\sin\theta_2$}
del:=take_fraction(dx1,dy2)-take_fraction(dx2,dy1); {$\sin(\theta_2-\theta_1)$}
if del>4684844 then clockwise:=false
else if del<-4684844 then clockwise:=true
  {$2^{28}\cdot\sin 1^\circ\approx4684844.68$}
else clockwise:=rev_turns;
end

@ Actually the turnarounds just computed will be clockwise,
not counterclockwise, if
the global variable |rev_turns| is |true|; it is usually |false|.

@<Glob...@>=
@!rev_turns:boolean; {should we make U-turns in the English manner?}

@ @<Set init...@>=
rev_turns:=false;

@ @<Compute the incoming and outgoing directions@>=
dx1:=x_coord(s)-left_x(s); dy1:=y_coord(s)-left_y(s);
if dx1=0 then if dy1=0 then
  begin dx1:=x_coord(s)-right_x(p); dy1:=y_coord(s)-right_y(p);
  if dx1=0 then if dy1=0 then
    begin dx1:=x_coord(s)-x_coord(p); dy1:=y_coord(s)-y_coord(p);
    end;  {and they {\sl can't} both be zero}
  end;
dmax:=abs(dx1);@+if abs(dy1)>dmax then dmax:=abs(dy1);
while dmax<fraction_one do
  begin double(dmax); double(dx1); double(dy1);
  end;
dx2:=right_x(q)-x_coord(q); dy2:=right_y(q)-y_coord(q);
if dx2=0 then if dy2=0 then
  begin dx2:=left_x(r)-x_coord(q); dy2:=left_y(r)-y_coord(q);
  if dx2=0 then if dy2=0 then
    begin if right_type(r)=endpoint then
      begin cur_x:=x_coord(r); cur_y:=y_coord(r);
      end
    else  begin unskew(x_coord(r),y_coord(r),right_type(r));
      skew(cur_x,cur_y,right_type(q));
      end;
    dx2:=cur_x-x_coord(q); dy2:=cur_y-y_coord(q);
    end;  {and they {\sl can't} both be zero}
  end;
dmax:=abs(dx2);@+if abs(dy2)>dmax then dmax:=abs(dy2);
while dmax<fraction_one do
  begin double(dmax); double(dx2); double(dy2);
  end

@ @<Insert additional boundary nodes...@>=
loop@+  begin if clockwise then
    if o1=1 then o1:=8@+else decr(o1)
  else if o1=8 then o1:=1@+else incr(o1);
  if o1=o2 then goto done;
  new_boundary(s,octant_code[o1]);
  s:=link(s); left_octant(s):=right_octant(s);
  end

@ Now it remains to insert the redundant
transition information into the |left_transition|
and |right_transition| fields between adjacent octants, in the octant
boundary nodes that have just been inserted between |link(p)| and~|q|.
The turning number is easily computed from these transitions.

@<Fix up the transition fields and adjust the turning number@>=
p:=link(p);
repeat s:=link(p);
o1:=octant_number[right_octant(p)]; o2:=octant_number[left_octant(s)];
if abs(o1-o2)=1 then
  begin if o2<o1 then o2:=o1;
  if odd(o2) then right_transition(p):=axis
  else right_transition(p):=diagonal;
  end
else  begin if o1=8 then incr(turning_number)@+else decr(turning_number);
  right_transition(p):=axis;
  end;
left_transition(s):=right_transition(p);
p:=s;
until p=q

@* \[22] Filling a contour.
Given the low-level machinery for making moves and for transforming a
cyclic path into a cycle spec, we're almost able to fill a digitized path.
All we need is a high-level routine that walks through the cycle spec and
controls the overall process.

Our overall goal is to plot the integer points $\bigl(\round(x(t)),
\round(y(t))\bigr)$ and to connect them by rook moves, assuming that
$\round(x(t))$ and $\round(y(t))$ don't both jump simultaneously from
one integer to another as $t$~varies; these rook moves will be the edge
of the contour that will be filled. We have reduced this problem to the
case of curves that travel in first octant directions, i.e., curves
such that $0\L y'(t)\L x'(t)$, by transforming the original coordinates.

\def\xtilde{{\tilde x}} \def\ytilde{{\tilde y}}
Another transformation makes the problem still simpler. We shall say that
we are working with {\sl biased coordinates\/} when $(x,y)$ has been
replaced by $(\xtilde,\ytilde)=(x-y,y+{1\over2})$. When a curve travels
in first octant directions, the corresponding curve with biased
coordinates travels in first {\sl quadrant\/} directions; the latter
condition is symmetric in $x$ and~$y$, so it has advantages for the
design of algorithms. The |make_spec| routine gives us skewed coordinates
$(x-y,y)$, hence we obtain biased coordinates by simply adding $1\over2$
to the second component.

The most important fact about biased coordinates is that we can determine the
rounded unbiased path $\bigl(\round(x(t)),\round(y(t))\bigr)$ from the
truncated biased path $\bigl(\lfloor\xtilde(t)\rfloor,\lfloor\ytilde(t)\rfloor
\bigr)$ and information about the initial and final endpoints. If the
unrounded and unbiased
path begins at $(x_0,y_0)$ and ends at $(x_1,y_1)$, it's possible to
prove (by induction on the length of the truncated biased path) that the
rounded unbiased path is obtained by the following construction:

\yskip\textindent{1)} Start at $\bigl(\round(x_0),\round(y_0)\bigr)$.

\yskip\textindent{2)} If $(x_0+{1\over2})\bmod1\G(y_0+{1\over2})\bmod1$,
move one step right.

\yskip\textindent{3)} Whenever the path
$\bigl(\lfloor\xtilde(t)\rfloor,\lfloor\ytilde(t)\rfloor\bigr)$
takes an upward step (i.e., when
$\lfloor\xtilde(t+\epsilon)\rfloor=\lfloor\xtilde(t)\rfloor$ and
$\lfloor\ytilde(t+\epsilon)\rfloor=\lfloor\ytilde(t)\rfloor+1$),
move one step up and then one step right.

\yskip\textindent{4)} Whenever the path
$\bigl(\lfloor\xtilde(t)\rfloor,\lfloor\ytilde(t)\rfloor\bigr)$
takes a rightward step (i.e., when
$\lfloor\xtilde(t+\epsilon)\rfloor=\lfloor\xtilde(t)\rfloor+1$ and
$\lfloor\ytilde(t+\epsilon)\rfloor=\lfloor\ytilde(t)\rfloor$),
move one step right.

\yskip\textindent{5)} Finally, if
$(x_1+{1\over2})\bmod1\G(y_1+{1\over2})\bmod1$, move one step left (thereby
cancelling the previous move, which was one step right). You will now be
at the point $\bigl(\round(x_1),\round(y_1)\bigr)$.

@ In order to validate the assumption that $\round(x(t))$ and $\round(y(t))$
don't both jump simultaneously, we shall consider that a coordinate pair
$(x,y)$ actually represents $(x+\epsilon,y+\epsilon\delta)$, where
$\epsilon$ and $\delta$ are extremely small positive numbers---so small
that their precise values never matter.  This convention makes rounding
unambiguous, since there is always a unique integer point nearest to any
given scaled numbers~$(x,y)$.

When coordinates are transformed so that \MF\ needs to work only in ``first
octant'' directions, the transformations involve negating~$x$, negating~$y$,
and/or interchanging $x$ with~$y$. Corresponding adjustments to the
rounding conventions must be made so that consistent values will be
obtained. For example, suppose that we're working with coordinates that
have been transformed so that a third-octant curve travels in first-octant
directions. The skewed coordinates $(x,y)$ in our data structure represent
unskewed coordinates $(-y,x+y)$, which are actually $(-y+\epsilon,
x+y+\epsilon\delta)$. We should therefore round as if our skewed coordinates
were $(x+\epsilon+\epsilon\delta,y-\epsilon)$ instead of $(x,y)$. The following
table shows how the skewed coordinates should be perturbed when rounding
decisions are made:
$$\vcenter{\halign{#\hfil&&\quad$#$\hfil&\hskip4em#\hfil\cr
|first_octant|&(x+\epsilon-\epsilon\delta,y+\epsilon\delta)&
 |fifth_octant|&(x-\epsilon+\epsilon\delta,y-\epsilon\delta)\cr
|second_octant|&(x-\epsilon+\epsilon\delta,y+\epsilon)&
 |sixth_octant|&(x+\epsilon-\epsilon\delta,y-\epsilon)\cr
|third_octant|&(x+\epsilon+\epsilon\delta,y-\epsilon)&
 |seventh_octant|&(x-\epsilon-\epsilon\delta,y+\epsilon)\cr
|fourth_octant|&(x-\epsilon-\epsilon\delta,y+\epsilon\delta)&
 |eighth_octant|&(x+\epsilon+\epsilon\delta,y-\epsilon\delta)\cr}}$$

Four small arrays are set up so that the rounding operations will be
fairly easy in any given octant.

@<Glob...@>=
@!y_corr,@!xy_corr,@!z_corr:array[first_octant..sixth_octant] of 0..1;
@!x_corr:array[first_octant..sixth_octant] of -1..1;

@ Here |xy_corr| is 1 if and only if the $x$ component of a skewed coordinate
is to be decreased by an infinitesimal amount; |y_corr| is similar, but for
the $y$ components. The other tables are set up so that the condition
$$(x+y+|half_unit|)\bmod|unity|\G(y+|half_unit|)\bmod|unity|$$
is properly perturbed to the condition
$$(x+y+|half_unit|-|x_corr|-|y_corr|)\bmod|unity|\G
  (y+|half_unit|-|y_corr|)\bmod|unity|+|z_corr|.$$

@<Set init...@>=
x_corr[first_octant]:=0; y_corr[first_octant]:=0;
xy_corr[first_octant]:=0;@/
x_corr[second_octant]:=0; y_corr[second_octant]:=0;
xy_corr[second_octant]:=1;@/
x_corr[third_octant]:=-1; y_corr[third_octant]:=1;
xy_corr[third_octant]:=0;@/
x_corr[fourth_octant]:=1; y_corr[fourth_octant]:=0;
xy_corr[fourth_octant]:=1;@/
x_corr[fifth_octant]:=0; y_corr[fifth_octant]:=1;
xy_corr[fifth_octant]:=1;@/
x_corr[sixth_octant]:=0; y_corr[sixth_octant]:=1;
xy_corr[sixth_octant]:=0;@/
x_corr[seventh_octant]:=1; y_corr[seventh_octant]:=0;
xy_corr[seventh_octant]:=1;@/
x_corr[eighth_octant]:=-1; y_corr[eighth_octant]:=1;
xy_corr[eighth_octant]:=0;@/
for k:=1 to 8 do z_corr[k]:=xy_corr[k]-x_corr[k];

@ Here's a procedure that handles the details of rounding at the
endpoints: Given skewed coordinates |(x,y)|, it sets |(m1,n1)|
to the corresponding rounded lattice points, taking the current
|octant| into account. Global variable |d1| is also set to 1 if
$(x+y+{1\over2})\bmod1\G(y+{1\over2})\bmod1$.

@p procedure end_round(@!x,@!y:scaled);
begin y:=y+half_unit-y_corr[octant];
x:=x+y-x_corr[octant];
m1:=floor_unscaled(x); n1:=floor_unscaled(y);
if x-unity*m1>=y-unity*n1+z_corr[octant] then d1:=1@+else d1:=0;
end;

@ The outputs |(m1,n1,d1)| of |end_round| will sometimes be moved
to |(m0,n0,d0)|.

@<Glob...@>=
@!m0,@!n0,@!m1,@!n1:integer; {lattice point coordinates}
@!d0,@!d1:0..1; {displacement corrections}

@ We're ready now to fill the pixels enclosed by a given cycle spec~|h|;
the knot list that represents the cycle is destroyed in the process.
The edge structure that gets all the resulting data is |cur_edges|,
and the edges are weighted by |cur_wt|.

@p procedure fill_spec(@!h:pointer);
var @!p,@!q,@!r,@!s:pointer; {for list traversal}
begin if internal[tracing_edges]>0 then begin_edge_tracing;
p:=h; {we assume that |left_type(h)=endpoint|}
repeat octant:=left_octant(p);
@<Set variable |q| to the node at the end of the current octant@>;
if q<>p then
  begin @<Determine the starting and ending
    lattice points |(m0,n0)| and |(m1,n1)|@>;
  @<Make the moves for the current octant@>;
  move_to_edges(m0,n0,m1,n1);
  end;
p:=link(q);
until p=h;
toss_knot_list(h);
if internal[tracing_edges]>0 then end_edge_tracing;
end;

@ @<Set variable |q| to the node at the end of the current octant@>=
q:=p;
while right_type(q)<>endpoint do q:=link(q)

@ @<Determine the starting and ending lattice points |(m0,n0)| and |(m1,n1)|@>=
end_round(x_coord(p),y_coord(p)); m0:=m1; n0:=n1; d0:=d1;@/
end_round(x_coord(q),y_coord(q))

@ Finally we perform the five-step process that was explained at
the very beginning of this part of the program.

@<Make the moves for the current octant@>=
if n1-n0>=move_size then overflow("move table size",move_size);
@:METAFONT capacity exceeded move table size}{\quad move table size@>
move[0]:=d0; move_ptr:=0; r:=p;
repeat s:=link(r);@/
make_moves(x_coord(r),right_x(r),left_x(s),x_coord(s),@|
  y_coord(r)+half_unit,right_y(r)+half_unit,left_y(s)+half_unit,
  y_coord(s)+half_unit,@| xy_corr[octant],y_corr[octant]);
r:=s;
until r=q;
move[move_ptr]:=move[move_ptr]-d1;
if internal[smoothing]>0 then smooth_moves(0,move_ptr)

@* \[23] Polygonal pens.
The next few parts of the program deal with the additional complications
associated with ``envelopes,'' leading up to an algorithm that fills a
contour with respect to a pen whose boundary is a convex polygon. The
mathematics underlying this algorithm is based on simple aspects of the
theory of tracings developed by Leo Guibas, Lyle Ramshaw, and Jorge
Stolfi [``A kinetic framework for computational geometry,''
{\sl Proc.\ IEEE Symp.\ Foundations of Computer Science\/ \bf24} (1983),
100--111].
@^Guibas, Leonidas Ioannis@>
@^Ramshaw, Lyle Harold@>
@^Stolfi, Jorge@>

If the vertices of the polygon are $w_0$, $w_1$, \dots, $w_{n-1}$, $w_n=w_0$,
in counterclockwise order, the convexity condition requires that ``left
turns'' are made at each vertex when a person proceeds from $w_0$ to
$w_1$ to $\cdots$ to~$w_n$. The envelope is obtained if we offset a given
curve $z(t)$ by $w_k$ when that curve is traveling in a direction
$z'(t)$ lying between the directions $w_k-w_{k-1}$ and $w\k-w_k$.
At times~$t$ when the curve direction $z'(t)$ increases past
$w\k-w_k$, we temporarily stop plotting the offset curve and we insert
a straight line from $z(t)+w_k$ to $z(t)+w\k$; notice that this straight
line is tangent to the offset curve. Similarly, when the curve direction
decreases past $w_k-w_{k-1}$, we stop plotting and insert a straight
line from $z(t)+w_k$ to $z(t)+w_{k-1}$; the latter line is actually a
``retrograde'' step, which won't be part of the final envelope under
\MF's assumptions. The result of this construction is a continuous path
that consists of alternating curves and straight line segments. The
segments are usually so short, in practice, that they blend with the
curves; after all, it's possible to represent any digitized path as
a sequence of digitized straight lines.

The nicest feature of this approach to envelopes is that it blends
perfectly with the octant subdivision process we have already developed.
The envelope travels in the same direction as the curve itself, as we
plot it, and we need merely be careful what offset is being added.
Retrograde motion presents a problem, but we will see that there is
a decent way to handle it.

@ We shall represent pens by maintaining eight lists of offsets,
one for each octant direction. The offsets at the boundary points
where a curve turns into a new octant will appear in the lists for
both octants. This means that we can restrict consideration to
segments of the original polygon whose directions aim in the first
octant, as we have done in the simpler case when envelopes were not
required.

An example should help to clarify this situation: Consider the
quadrilateral whose vertices are $w_0=(0,-1)$, $w_1=(3,-1)$,
$w_2=(6,1)$, and $w_3=(1,2)$. A curve that travels in the first octant
will be offset by $w_1$ or $w_2$, unless its slope drops to zero
en route to the eighth octant; in the latter case we should switch to $w_0$ as
we cross the octant boundary. Our list for the first octant will
contain the three offsets $w_0$, $w_1$,~$w_2$. By convention we will
duplicate a boundary offset if the angle between octants doesn't
explicitly appear; in this case there is no explicit line of slope~1
at the end of the list, so the full list is
$$w_0\;w_1\;w_2\;w_2\;=\;(0,-1)\;(3,-1)\;(6,1)\;(6,1).$$
With skewed coordinates $(u-v,v)$ instead of $(u,v)$ we obtain the list
$$w_0\;w_1\;w_2\;w_2\;\mapsto\;(1,-1)\;(4,-1)\;(5,1)\;(5,1),$$
which is what actually appears in the data structure. In the second
octant there's only one offset; we list it twice (with coordinates
interchanged, so as to make the second octant look like the first),
and skew those coordinates, obtaining
$$\tabskip\centering
\halign to\hsize{$\hfil#\;\mapsto\;{}$\tabskip=0pt&
  $#\hfil$&\quad in the #\hfil\tabskip\centering\cr
w_2\;w_2&(-5,6)\;(-5,6)\cr
\noalign{\vskip\belowdisplayskip
\vbox{\noindent\strut as the list of transformed and skewed offsets to use
when curves travel in the second octant. Similarly, we will have\strut}
\vskip\abovedisplayskip}
w_2\;w_2&(7,-6)\;(7,-6)&third;\cr
w_2\;w_2\;w_3\;w_3&(-7,1)\;(-7,1)\;(-3,2)\;(-3,2)&fourth;\cr
w_3\;w_3&(1,-2)\;(1,-2)&fifth;\cr
w_3\;w_3\;w_0\;w_0&(-1,1)\;(-1,1)\;(1,0)\;(1,0)&sixth;\cr
w_0\;w_0&(1,0)\;(1,0)&seventh;\cr
w_0\;w_0&(-1,1)\;(-1,1)&eighth.\cr}$$
Notice that $w_1$ is considered here to be internal to the first octant;
it's not part of the eighth. We could equally well have taken $w_0$ out
of the first octant list and put it into the eighth; then the first octant
list would have been
$$w_1\;w_1\;w_2\;w_2\;\mapsto\;(4,-1)\;(4,-1)\;(5,1)\;(5,1)$$
and the eighth octant list would have been
$$w_0\;w_0\;w_1\;\mapsto\;(-1,1)\;(-1,1)\;(2,1).$$

Actually, there's one more complication: The order of offsets is reversed
in even-numbered octants, because the transformation of coordinates has
reversed counterclockwise and clockwise orientations in those octants.
The offsets in the fourth octant, for example, are really $w_3$, $w_3$,
$w_2$,~$w_2$, not $w_2$, $w_2$, $w_3$,~$w_3$.

@ In general, the list of offsets for an octant will have the form
$$w_0\;\;w_1\;\;\ldots\;\;w_n\;\;w_{n+1}$$
(if we renumber the subscripts in each list), where $w_0$ and $w_{n+1}$
are offsets common to the neighboring lists. We'll often have $w_0=w_1$
and/or $w_n=w_{n+1}$, but the other $w$'s will be distinct. Curves
that travel between slope~0 and direction $w_2-w_1$ will use offset~$w_1$;
curves that travel between directions $w_k-w_{k-1}$ and $w\k-w_k$ will
use offset~$w_k$, for $1<k<n$; curves between direction $w_n-w_{n-1}$
and slope~1 (actually slope~$\infty$ after skewing) will use offset~$w_n$.
In even-numbered octants, the directions are actually $w_k-w\k$ instead
of $w\k-w_k$, because the offsets have been listed in reverse order.

Each offset $w_k$ is represented by skewed coordinates $(u_k-v_k,v_k)$,
where $(u_k,v_k)$ is the representation of $w_k$ after it has been rotated
into a first-octant disguise.

@ The top-level data structure of a pen polygon is a 10-word node containing
a reference count followed by pointers to the eight offset lists, followed
by an indication of the pen's range of values.
@^reference counts@>

If |p|~points to such a node, and if the
offset list for, say, the fourth octant has entries $w_0$, $w_1$, \dots,
$w_n$,~$w_{n+1}$, then |info(p+fourth_octant)| will equal~$n$, and
|link(p+fourth_octant)| will point to the offset node containing~$w_0$.
Memory location |p+fourth_octant| is said to be the {\sl header\/} of
the pen-offset list for the fourth octant. Since this is an even-numbered
octant, $w_0$ is the offset that goes with the fifth octant, and
$w_{n+1}$ goes with the third.

The elements of the offset list themselves are doubly linked 3-word nodes,
containing coordinates in their |x_coord| and |y_coord| fields.
The two link fields are called |link| and |knil|; if |w|~points to
the node for~$w_k$, then |link(w)| and |knil(w)| point respectively
to the nodes for $w\k$ and~$w_{k-1}$. If |h| is the list header,
|link(h)| points to the node for~$w_0$ and |knil(link(h))| to the
node for~$w_{n+1}$.

The tenth word of a pen header node contains the maximum absolute value of
an $x$ or $y$ coordinate among all of the unskewed pen offsets.

The |link| field of a pen header node should be |null| if and only if
the pen is a single point.

@d pen_node_size=10
@d coord_node_size=3
@d max_offset(#)==mem[#+9].sc

@ The |print_pen| subroutine illustrates these conventions by
reconstructing the vertices of a polygon from \MF's complicated
internal offset representation.

@<Declare subroutines for printing expressions@>=
procedure print_pen(@!p:pointer;@!s:str_number;@!nuline:boolean);
var @!nothing_printed:boolean; {has there been any action yet?}
@!k:1..8; {octant number}
@!h:pointer; {offset list head}
@!m,@!n:integer; {offset indices}
@!w,@!ww:pointer; {pointers that traverse the offset list}
begin print_diagnostic("Pen polygon",s,nuline);
nothing_printed:=true; print_ln;
for k:=1 to 8 do
  begin octant:=octant_code[k]; h:=p+octant; n:=info(h); w:=link(h);
  if not odd(k) then w:=knil(w); {in even octants, start at $w_{n+1}$}
  for m:=1 to n+1 do
    begin if odd(k) then ww:=link(w)@+else ww:=knil(w);
    if (x_coord(ww)<>x_coord(w))or(y_coord(ww)<>y_coord(w)) then
      @<Print the unskewed and unrotated coordinates of node |ww|@>;
    w:=ww;
    end;
  end;
if nothing_printed then
  begin w:=link(p+first_octant); print_two(x_coord(w)+y_coord(w),y_coord(w));
  end;
print_nl(" .. cycle"); end_diagnostic(true);
end;

@ @<Print the unskewed and unrotated coordinates of node |ww|@>=
begin if nothing_printed then nothing_printed:=false
else print_nl(" .. ");
print_two_true(x_coord(ww),y_coord(ww));
end

@ A null pen polygon, which has just one vertex $(0,0)$, is
predeclared for error recovery. It doesn't need a proper
reference count, because the |toss_pen| procedure below
will never delete it from memory.
@^reference counts@>

@<Initialize table entries...@>=
ref_count(null_pen):=null; link(null_pen):=null;@/
info(null_pen+1):=1; link(null_pen+1):=null_coords;
for k:=null_pen+2 to null_pen+8 do mem[k]:=mem[null_pen+1];
max_offset(null_pen):=0;@/
link(null_coords):=null_coords;
knil(null_coords):=null_coords;@/
x_coord(null_coords):=0;
y_coord(null_coords):=0;

@ Here's a trivial subroutine that inserts a copy of an offset
on the |link| side of its clone in the doubly linked list.

@p procedure dup_offset(@!w:pointer);
var @!r:pointer; {the new node}
begin r:=get_node(coord_node_size);
x_coord(r):=x_coord(w);
y_coord(r):=y_coord(w);
link(r):=link(w); knil(link(w)):=r;
knil(r):=w; link(w):=r;
end;

@ The following algorithm is somewhat more interesting: It converts a
knot list for a cyclic path into a pen polygon, ignoring everything
but the |x_coord|, |y_coord|, and |link| fields. If the given path
vertices do not define a convex polygon, an error message is issued
and the null pen is returned.

@p function make_pen(@!h:pointer):pointer;
label done,done1,not_found,found;
var @!o,@!oo,@!k:small_number; {octant numbers---old, new, and current}
@!p:pointer; {top-level node for the new pen}
@!q,@!r,@!s,@!w,@!hh:pointer; {for list manipulation}
@!n:integer; {offset counter}
@!dx,@!dy:scaled; {polygon direction}
@!mc:scaled; {the largest coordinate}
begin @<Stamp all nodes with an octant code, compute the maximum offset,
  and set |hh| to the node that begins the first octant;
  |goto not_found| if there's a problem@>;
if mc>=fraction_one-half_unit then goto not_found;
p:=get_node(pen_node_size); q:=hh; max_offset(p):=mc; ref_count(p):=null;
if link(q)<>q then link(p):=null+1;
for k:=1 to 8 do @<Construct the offset list for the |k|th octant@>;
goto found;
not_found:p:=null_pen; @<Complain about a bad pen path@>;
found: if internal[tracing_pens]>0 then print_pen(p," (newly created)",true);
make_pen:=p;
end;

@ @<Complain about a bad pen path@>=
if mc>=fraction_one-half_unit then
  begin print_err("Pen too large");
@.Pen too large@>
  help2("The cycle you specified has a coordinate of 4095.5 or more.")@/
  ("So I've replaced it by the trivial path `(0,0)..cycle'.");@/
  end
else  begin print_err("Pen cycle must be convex");
@.Pen cycle must be convex@>
  help3("The cycle you specified either has consecutive equal points")@/
    ("or turns right or turns through more than 360 degrees.")@/
  ("So I've replaced it by the trivial path `(0,0)..cycle'.");@/
  end;
put_get_error

@ There should be exactly one node whose octant number is less than its
predecessor in the cycle; that is node~|hh|.

The loop here will terminate in all cases, but the proof is somewhat tricky:
If there are at least two distinct $y$~coordinates in the cycle, we will have
|o>4| and |o<=4| at different points of the cycle. Otherwise there are
at least two distinct $x$~coordinates, and we will have |o>2| somewhere,
|o<=2| somewhere.

@<Stamp all nodes...@>=
q:=h; r:=link(q); mc:=abs(x_coord(h));
if q=r then
  begin hh:=h; right_type(h):=0; {this trick is explained below}
  if mc<abs(y_coord(h)) then mc:=abs(y_coord(h));
  end
else  begin o:=0; hh:=null;
  loop@+  begin s:=link(r);
    if mc<abs(x_coord(r)) then mc:=abs(x_coord(r));
    if mc<abs(y_coord(r)) then mc:=abs(y_coord(r));
    dx:=x_coord(r)-x_coord(q); dy:=y_coord(r)-y_coord(q);
    if dx=0 then if dy=0 then goto not_found; {double point}
    if ab_vs_cd(dx,y_coord(s)-y_coord(r),dy,x_coord(s)-x_coord(r))<0 then
      goto not_found; {right turn}
    @<Determine the octant code for direction |(dx,dy)|@>;
    right_type(q):=octant; oo:=octant_number[octant];
    if o>oo then
      begin if hh<>null then goto not_found; {$>360^\circ$}
      hh:=q;
      end;
    o:=oo;
    if (q=h)and(hh<>null) then goto done;
    q:=r; r:=s;
    end;
  done:end


@ We want the octant for |(-dx,-dy)| to be
exactly opposite the octant for |(dx,dy)|.

@<Determine the octant code for direction |(dx,dy)|@>=
if dx>0 then octant:=first_octant
else if dx=0 then
  if dy>0 then octant:=first_octant@+else octant:=first_octant+negate_x
else  begin negate(dx); octant:=first_octant+negate_x;
  end;
if dy<0 then
  begin negate(dy); octant:=octant+negate_y;
  end
else if dy=0 then
  if octant>first_octant then octant:=first_octant+negate_x+negate_y;
if dx<dy then octant:=octant+switch_x_and_y

@ Now |q| points to the node that the present octant shares with the previous
octant, and |right_type(q)| is the octant code during which |q|~should advance.
We have set |right_type(q)=0| in the special case that |q| should never advance
(because the pen is degenerate).

The number of offsets |n| must be smaller than |max_quarterword|, because
the |fill_envelope| routine stores |n+1| in the |right_type| field
of a knot node.

@<Construct the offset list...@>=
begin octant:=octant_code[k]; n:=0; h:=p+octant;
loop@+  begin r:=get_node(coord_node_size);
  skew(x_coord(q),y_coord(q),octant); x_coord(r):=cur_x; y_coord(r):=cur_y;
  if n=0 then link(h):=r
  else  @<Link node |r| to the previous node@>;
  w:=r;
  if right_type(q)<>octant then goto done1;
  q:=link(q); incr(n);
  end;
done1: @<Finish linking the offset nodes, and duplicate the
  borderline offset nodes if necessary@>;
if n>=max_quarterword then overflow("pen polygon size",max_quarterword);
@:METAFONT capacity exceeded pen polygon size}{\quad pen polygon size@>
info(h):=n;
end

@ Now |w| points to the node that was inserted most recently, and
|k| is the current octant number.

@<Link node |r| to the previous node@>=
if odd(k) then
  begin link(w):=r; knil(r):=w;
  end
else  begin knil(w):=r; link(r):=w;
  end

@ We have inserted |n+1| nodes; it remains to duplicate the nodes at the
ends, if slopes 0 and~$\infty$ aren't already represented. At the end of
this section the total number of offset nodes should be |n+2|
(since we call them $w_0$, $w_1$, \dots,~$w_{n+1}$).

@<Finish linking the offset nodes, and duplicate...@>=
r:=link(h);
if odd(k) then
  begin link(w):=r; knil(r):=w;
  end
else  begin knil(w):=r; link(r):=w; link(h):=w; r:=w;
  end;
if (y_coord(r)<>y_coord(link(r)))or(n=0) then
  begin dup_offset(r); incr(n);
  end;
r:=knil(r);
if x_coord(r)<>x_coord(knil(r)) then dup_offset(r)
else decr(n)

@ Conversely, |make_path| goes back from a pen to a cyclic path that
might have generated it. The structure of this subroutine is essentially
the same as |print_pen|.

@p @t\4@>@<Declare the function called |trivial_knot|@>@;
function make_path(@!pen_head:pointer):pointer;
var @!p:pointer; {the most recently copied knot}
@!k:1..8; {octant number}
@!h:pointer; {offset list head}
@!m,@!n:integer; {offset indices}
@!w,@!ww:pointer; {pointers that traverse the offset list}
begin p:=temp_head;
for k:=1 to 8 do
  begin octant:=octant_code[k]; h:=pen_head+octant; n:=info(h); w:=link(h);
  if not odd(k) then w:=knil(w); {in even octants, start at $w_{n+1}$}
  for m:=1 to n+1 do
    begin if odd(k) then ww:=link(w)@+else ww:=knil(w);
    if (x_coord(ww)<>x_coord(w))or(y_coord(ww)<>y_coord(w)) then
      @<Copy the unskewed and unrotated coordinates of node |ww|@>;
    w:=ww;
    end;
  end;
if p=temp_head then
  begin w:=link(pen_head+first_octant);
  p:=trivial_knot(x_coord(w)+y_coord(w),y_coord(w)); link(temp_head):=p;
  end;
link(p):=link(temp_head); make_path:=link(temp_head);
end;

@ @<Copy the unskewed and unrotated coordinates of node |ww|@>=
begin unskew(x_coord(ww),y_coord(ww),octant);
link(p):=trivial_knot(cur_x,cur_y); p:=link(p);
end

@ @<Declare the function called |trivial_knot|@>=
function trivial_knot(@!x,@!y:scaled):pointer;
var @!p:pointer; {a new knot for explicit coordinates |x| and |y|}
begin p:=get_node(knot_node_size);
left_type(p):=explicit; right_type(p):=explicit;@/
x_coord(p):=x; left_x(p):=x; right_x(p):=x;@/
y_coord(p):=y; left_y(p):=y; right_y(p):=y;@/
trivial_knot:=p;
end;

@ That which can be created can be destroyed.

@d add_pen_ref(#)==incr(ref_count(#))
@d delete_pen_ref(#)==if ref_count(#)=null then toss_pen(#)
  else decr(ref_count(#))

@<Declare the recycling subroutines@>=
procedure toss_pen(@!p:pointer);
var @!k:1..8; {relative header locations}
@!w,@!ww:pointer; {pointers to offset nodes}
begin if p<>null_pen then
  begin for k:=1 to 8 do
    begin w:=link(p+k);
    repeat ww:=link(w); free_node(w,coord_node_size); w:=ww;
    until w=link(p+k);
    end;
  free_node(p,pen_node_size);
  end;
end;

@ The |find_offset| procedure sets |(cur_x,cur_y)| to the offset associated
with a given direction~|(x,y)| and a given pen~|p|. If |x=y=0|, the
result is |(0,0)|. If two different offsets apply, one of them is
chosen arbitrarily.

@p procedure find_offset(@!x,@!y:scaled; @!p:pointer);
label done,exit;
var @!octant:first_octant..sixth_octant; {octant code for |(x,y)|}
@!s:-1..+1; {sign of the octant}
@!n:integer; {number of offsets remaining}
@!h,@!w,@!ww:pointer; {list traversal registers}
begin @<Compute the octant code; skew and rotate the coordinates |(x,y)|@>;
if odd(octant_number[octant]) then s:=-1@+else s:=+1;
h:=p+octant; w:=link(link(h)); ww:=link(w); n:=info(h);
while n>1 do
  begin if ab_vs_cd(x,y_coord(ww)-y_coord(w),@|
    y,x_coord(ww)-x_coord(w))<>s then goto done;
  w:=ww; ww:=link(w); decr(n);
  end;
done:unskew(x_coord(w),y_coord(w),octant);
exit:end;

@ @<Compute the octant code; skew and rotate the coordinates |(x,y)|@>=
if x>0 then octant:=first_octant
else if x=0 then
  if y<=0 then
    if y=0 then
      begin cur_x:=0; cur_y:=0; return;
      end
    else octant:=first_octant+negate_x
  else octant:=first_octant
else  begin x:=-x;
  if y=0 then octant:=first_octant+negate_x+negate_y
  else octant:=first_octant+negate_x;
  end;
if y<0 then
  begin octant:=octant+negate_y; y:=-y;
  end;
if x>=y then x:=x-y
else  begin octant:=octant+switch_x_and_y; x:=y-x; y:=y-x;
  end

@* \[24] Filling an envelope.
We are about to reach the culmination of \MF's digital plotting routines:
Almost all of the previous algorithms will be brought to bear on \MF's
most difficult task, which is to fill the envelope of a given cyclic path
with respect to a given pen polygon.

But we still must complete some of the preparatory work before taking such
a big plunge.

@ Given a pointer |c| to a nonempty list of cubics,
and a pointer~|h| to the header information of a pen polygon segment,
the |offset_prep| routine changes the list into cubics that are
associated with particular pen offsets. Namely, the cubic between |p|
and~|q| should be associated with the |k|th offset when |right_type(p)=k|.

List |c| is actually part of a cycle spec, so it terminates at the
first node whose |right_type| is |endpoint|. The cubics all have
monotone-nondecreasing $x(t)$ and $y(t)$.

@p @t\4@>@<Declare subroutines needed by |offset_prep|@>@;
procedure offset_prep(@!c,@!h:pointer);
label done,not_found;
var @!n:halfword; {the number of pen offsets}
@!p,@!q,@!r,@!lh,@!ww:pointer; {for list manipulation}
@!k:halfword; {the current offset index}
@!w:pointer; {a pointer to offset $w_k$}
@<Other local variables for |offset_prep|@>@;
begin p:=c; n:=info(h); lh:=link(h); {now |lh| points to $w_0$}
while right_type(p)<>endpoint do
  begin q:=link(p);
  @<Split the cubic between |p| and |q|, if necessary, into cubics
    associated with single offsets, after which |q| should
    point to the end of the final such cubic@>;
  @<Advance |p| to node |q|, removing any ``dead'' cubics that
    might have been introduced by the splitting process@>;
  end;
end;

@ @<Advance |p| to node |q|, removing any ``dead'' cubics...@>=
repeat r:=link(p);
if x_coord(p)=right_x(p) then if y_coord(p)=right_y(p) then
 if x_coord(p)=left_x(r) then if y_coord(p)=left_y(r) then
  if x_coord(p)=x_coord(r) then if y_coord(p)=y_coord(r) then
  begin remove_cubic(p);
  if r=q then q:=p;
  r:=p;
  end;
p:=r;
until p=q

@ The splitting process uses a subroutine like |split_cubic|, but
(for ``bulletproof'' operation) we check to make sure that the
resulting (skewed) coordinates satisfy $\Delta x\G0$ and $\Delta y\G0$
after splitting; |make_spec| has made sure that these relations hold
before splitting. (This precaution is surely unnecessary, now that
|make_spec| is so much more careful than it used to be. But who
wants to take a chance? Maybe the hardware will fail or something.)

@<Declare subroutines needed by |offset_prep|@>=
procedure split_for_offset(@!p:pointer;@!t:fraction);
var @!q:pointer; {the successor of |p|}
@!r:pointer; {the new node}
begin q:=link(p); split_cubic(p,t,x_coord(q),y_coord(q)); r:=link(p);
if y_coord(r)<y_coord(p) then y_coord(r):=y_coord(p)
else if y_coord(r)>y_coord(q) then y_coord(r):=y_coord(q);
if x_coord(r)<x_coord(p) then x_coord(r):=x_coord(p)
else if x_coord(r)>x_coord(q) then x_coord(r):=x_coord(q);
end;

@ If the pen polygon has |n| offsets, and if $w_k=(u_k,v_k)$ is the $k$th
of these, the $k$th pen slope is defined by the formula
$$s_k={v\k-v_k\over u\k-u_k},\qquad\hbox{for $0<k<n$}.$$
In odd-numbered octants, the numerator and denominator of this fraction
will be nonnegative; in even-numbered octants they will both be nonpositive.
Furthermore we always have $0=s_0\le s_1\le\cdots\le s_n=\infty$. The goal of
|offset_prep| is to find an offset index~|k| to associate with
each cubic, such that the slope $s(t)$ of the cubic satisfies
$$s_{k-1}\le s(t)\le s_k\qquad\hbox{for $0\le t\le 1$.}\eqno(*)$$
We may have to split a cubic into as many as $2n-1$ pieces before each
piece corresponds to a unique offset.

@<Split the cubic between |p| and |q|, if necessary, into cubics...@>=
if n<=1 then right_type(p):=1 {this case is easy}
else  begin @<Prepare for derivative computations;
    |goto not_found| if the current cubic is dead@>;
  @<Find the initial slope, |dy/dx|@>;
  if dx=0 then @<Handle the special case of infinite slope@>
  else  begin @<Find the index |k| such that $s_{k-1}\L\\{dy}/\\{dx}<s_k$@>;
    @<Complete the offset splitting process@>;
    end;
not_found: end

@ The slope of a cubic $B(z_0,z_1,z_2,z_3;t)=\bigl(x(t),y(t)\bigr)$ can be
calculated from the quadratic polynomials
${1\over3}x'(t)=B(x_1-x_0,x_2-x_1,x_3-x_2;t)$ and
${1\over3}y'(t)=B(y_1-y_0,y_2-y_1,y_3-y_2;t)$.
Since we may be calculating slopes from several cubics
split from the current one, it is desirable to do these calculations
without losing too much precision. ``Scaled up'' values of the
derivatives, which will be less tainted by accumulated errors than
derivatives found from the cubics themselves, are maintained in
local variables |x0|, |x1|, and |x2|, representing $X_0=2^l(x_1-x_0)$,
$X_1=2^l(x_2-x_1)$, and $X_2=2^l(x_3-x_2)$; similarly |y0|, |y1|, and~|y2|
represent $Y_0=2^l(y_1-y_0)$, $Y_1=2^l(y_2-y_1)$, and $Y_2=2^l(y_3-y_2)$.
To test whether the slope of the cubic is $\ge s$ or $\le s$, we will test
the sign of the quadratic ${1\over3}2^l\bigl(y'(t)-sx'(t)\bigr)$ if $s\le1$,
or ${1\over3}2^l\bigl(y'(t)/s-x'(t)\bigr)$ if $s>1$.

@<Other local variables for |offset_prep|@>=
@!x0,@!x1,@!x2,@!y0,@!y1,@!y2:integer; {representatives of derivatives}
@!t0,@!t1,@!t2:integer; {coefficients of polynomial for slope testing}
@!du,@!dv,@!dx,@!dy:integer; {for slopes of the pen and the curve}
@!max_coef:integer; {used while scaling}
@!x0a,@!x1a,@!x2a,@!y0a,@!y1a,@!y2a:integer; {intermediate values}
@!t:fraction; {where the derivative passes through zero}
@!s:fraction; {slope or reciprocal slope}

@ @<Prepare for derivative computations...@>=
x0:=right_x(p)-x_coord(p); {should be |>=0|}
x2:=x_coord(q)-left_x(q); {likewise}
x1:=left_x(q)-right_x(p); {but this might be negative}
y0:=right_y(p)-y_coord(p); y2:=y_coord(q)-left_y(q);
y1:=left_y(q)-right_y(p);
max_coef:=abs(x0); {we take |abs| just to make sure}
if abs(x1)>max_coef then max_coef:=abs(x1);
if abs(x2)>max_coef then max_coef:=abs(x2);
if abs(y0)>max_coef then max_coef:=abs(y0);
if abs(y1)>max_coef then max_coef:=abs(y1);
if abs(y2)>max_coef then max_coef:=abs(y2);
if max_coef=0 then goto not_found;
while max_coef<fraction_half do
  begin double(max_coef);
  double(x0); double(x1); double(x2);
  double(y0); double(y1); double(y2);
  end

@ Let us first solve a special case of the problem: Suppose we
know an index~$k$ such that either (i)~$s(t)\G s_{k-1}$ for all~$t$
and $s(0)<s_k$, or (ii)~$s(t)\L s_k$ for all~$t$ and $s(0)>s_{k-1}$.
Then, in a sense, we're halfway done, since one of the two inequalities
in $(*)$ is satisfied, and the other couldn't be satisfied for
any other value of~|k|.

The |fin_offset_prep| subroutine solves the stated subproblem.
It has a boolean parameter called |rising| that is |true| in
case~(i), |false| in case~(ii). When |rising=false|, parameters
|x0| through |y2| represent the negative of the derivative of
the cubic following |p|; otherwise they represent the actual derivative.
The |w| parameter should point to offset~$w_k$.

@<Declare subroutines needed by |offset_prep|@>=
procedure fin_offset_prep(@!p:pointer;@!k:halfword;@!w:pointer;
  @!x0,@!x1,@!x2,@!y0,@!y1,@!y2:integer;@!rising:boolean;@!n:integer);
label exit;
var @!ww:pointer; {for list manipulation}
@!du,@!dv:scaled; {for slope calculation}
@!t0,@!t1,@!t2:integer; {test coefficients}
@!t:fraction; {place where the derivative passes a critical slope}
@!s:fraction; {slope or reciprocal slope}
@!v:integer; {intermediate value for updating |x0..y2|}
begin loop
  begin right_type(p):=k;
  if rising then
    if k=n then return
    else ww:=link(w) {a pointer to $w\k$}
  else  if k=1 then return
    else ww:=knil(w); {a pointer to $w_{k-1}$}
  @<Compute test coefficients |(t0,t1,t2)|
    for $s(t)$ versus $s_k$ or $s_{k-1}$@>;
  t:=crossing_point(t0,t1,t2);
  if t>=fraction_one then return;
  @<Split the cubic at $t$,
    and split off another cubic if the derivative crosses back@>;
  if rising then incr(k)@+else decr(k);
  w:=ww;
  end;
exit:end;

@ @<Compute test coefficients |(t0,t1,t2)| for $s(t)$ versus...@>=
du:=x_coord(ww)-x_coord(w); dv:=y_coord(ww)-y_coord(w);
if abs(du)>=abs(dv) then {$s_{k-1}\le1$ or $s_k\le1$}
  begin s:=make_fraction(dv,du);
  t0:=take_fraction(x0,s)-y0;
  t1:=take_fraction(x1,s)-y1;
  t2:=take_fraction(x2,s)-y2;
  end
else  begin s:=make_fraction(du,dv);
  t0:=x0-take_fraction(y0,s);
  t1:=x1-take_fraction(y1,s);
  t2:=x2-take_fraction(y2,s);
  end

@ The curve has crossed $s_k$ or $s_{k-1}$; its initial segment satisfies
$(*)$, and it might cross again and return towards $s_{k-1}$ or $s_k$,
respectively, yielding another solution of $(*)$.

@<Split the cubic at $t$, and split off another...@>=
begin split_for_offset(p,t); right_type(p):=k; p:=link(p);@/
v:=t_of_the_way(x0)(x1); x1:=t_of_the_way(x1)(x2);
x0:=t_of_the_way(v)(x1);@/
v:=t_of_the_way(y0)(y1); y1:=t_of_the_way(y1)(y2);
y0:=t_of_the_way(v)(y1);@/
t1:=t_of_the_way(t1)(t2);
if t1>0 then t1:=0; {without rounding error, |t1| would be |<=0|}
t:=crossing_point(0,-t1,-t2);
if t<fraction_one then
  begin split_for_offset(p,t); right_type(link(p)):=k;@/
  v:=t_of_the_way(x1)(x2); x1:=t_of_the_way(x0)(x1);
  x2:=t_of_the_way(x1)(v);@/
  v:=t_of_the_way(y1)(y2); y1:=t_of_the_way(y0)(y1);
  y2:=t_of_the_way(y1)(v);
  end;
end

@ Now we must consider the general problem of |offset_prep|, when
nothing is known about a given cubic. We start by finding its
slope $s(0)$ in the vicinity of |t=0|.

If $z'(t)=0$, the given cubic is numerically unstable, since the
slope direction is probably being influenced primarily by rounding
errors. A user who specifies such cuspy curves should expect to generate
rather wild results. The present code tries its best to believe the
existing data, as if no rounding errors were present.

@ @<Find the initial slope, |dy/dx|@>=
dx:=x0; dy:=y0;
if dx=0 then if dy=0 then
  begin dx:=x1; dy:=y1;
  if dx=0 then if dy=0 then
    begin dx:=x2; dy:=y2;
    end;
  end

@ The next step is to bracket the initial slope between consecutive
slopes of the pen polygon. The most important invariant relation in the
following loop is that |dy/dx>=@t$s_{k-1}$@>|.

@<Find the index |k| such that $s_{k-1}\L\\{dy}/\\{dx}<s_k$@>=
k:=1; w:=link(lh);
loop@+  begin if k=n then goto done;
  ww:=link(w);
  if ab_vs_cd(dy,abs(x_coord(ww)-x_coord(w)),@|
   dx,abs(y_coord(ww)-y_coord(w)))>=0 then
    begin incr(k); w:=ww;
    end
  else goto done;
  end;
done:

@ Finally we want to reduce the general problem to situations that
|fin_offset_prep| can handle. If |k=1|, we already are in the desired
situation. Otherwise we can split the cubic into at most three parts
with respect to $s_{k-1}$, and apply |fin_offset_prep| to each part.

@<Complete the offset splitting process@>=
if k=1 then t:=fraction_one+1
else  begin ww:=knil(w); @<Compute test coeff...@>;
  t:=crossing_point(-t0,-t1,-t2);
  end;
if t>=fraction_one then fin_offset_prep(p,k,w,x0,x1,x2,y0,y1,y2,true,n)
else  begin split_for_offset(p,t); r:=link(p);@/
  x1a:=t_of_the_way(x0)(x1); x1:=t_of_the_way(x1)(x2);
  x2a:=t_of_the_way(x1a)(x1);@/
  y1a:=t_of_the_way(y0)(y1); y1:=t_of_the_way(y1)(y2);
  y2a:=t_of_the_way(y1a)(y1);@/
  fin_offset_prep(p,k,w,x0,x1a,x2a,y0,y1a,y2a,true,n); x0:=x2a; y0:=y2a;
  t1:=t_of_the_way(t1)(t2);
  if t1<0 then t1:=0;
  t:=crossing_point(0,t1,t2);
  if t<fraction_one then
    @<Split off another |rising| cubic for |fin_offset_prep|@>;
  fin_offset_prep(r,k-1,ww,-x0,-x1,-x2,-y0,-y1,-y2,false,n);
  end

@ @<Split off another |rising| cubic for |fin_offset_prep|@>=
begin split_for_offset(r,t);@/
x1a:=t_of_the_way(x1)(x2); x1:=t_of_the_way(x0)(x1);
x0a:=t_of_the_way(x1)(x1a);@/
y1a:=t_of_the_way(y1)(y2); y1:=t_of_the_way(y0)(y1);
y0a:=t_of_the_way(y1)(y1a);@/
fin_offset_prep(link(r),k,w,x0a,x1a,x2,y0a,y1a,y2,true,n);
x2:=x0a; y2:=y0a;
end

@ @<Handle the special case of infinite slope@>=
fin_offset_prep(p,n,knil(knil(lh)),-x0,-x1,-x2,-y0,-y1,-y2,false,n)

@ OK, it's time now for the biggie. The |fill_envelope| routine generalizes
|fill_spec| to polygonal envelopes. Its outer structure is essentially the
same as before, except that octants with no cubics do contribute to
the envelope.

@p @t\4@>@<Declare the procedure called |skew_line_edges|@>@;
@t\4@>@<Declare the procedure called |dual_moves|@>@;
procedure fill_envelope(@!spec_head:pointer);
label done, done1;
var @!p,@!q,@!r,@!s:pointer; {for list traversal}
@!h:pointer; {head of pen offset list for current octant}
@!www:pointer; {a pen offset of temporary interest}
@<Other local variables for |fill_envelope|@>@;
begin if internal[tracing_edges]>0 then begin_edge_tracing;
p:=spec_head; {we assume that |left_type(spec_head)=endpoint|}
repeat octant:=left_octant(p); h:=cur_pen+octant;
@<Set variable |q| to the node at the end of the current octant@>;
@<Determine the envelope's starting and ending
    lattice points |(m0,n0)| and |(m1,n1)|@>;
offset_prep(p,h); {this may clobber node~|q|, if it becomes ``dead''}
@<Set variable |q| to the node at the end of the current octant@>;
@<Make the envelope moves for the current octant and insert them
  in the pixel data@>;
p:=link(q);
until p=spec_head;
if internal[tracing_edges]>0 then end_edge_tracing;
toss_knot_list(spec_head);
end;

@ In even-numbered octants we have reflected the coordinates an odd number
of times, hence clockwise and counterclockwise are reversed; this means that
the envelope is being formed in a ``dual'' manner. For the time being, let's
concentrate on odd-numbered octants, since they're easier to understand.
After we have coded the program for odd-numbered octants, the changes needed
to dualize it will not be so mysterious.

It is convenient to assume that we enter an odd-numbered octant with
an |axis| transition (where the skewed slope is zero) and leave at a
|diagonal| one (where the skewed slope is infinite). Then all of the
offset points $z(t)+w(t)$ will lie in a rectangle whose lower left and
upper right corners are the initial and final offset points. If this
assumption doesn't hold we can implicitly change the curve so that it does.
For example, if the entering transition is diagonal, we can draw a
straight line from $z_0+w_{n+1}$ to $z_0+w_0$ and continue as if the
curve were moving rightward. The effect of this on the envelope is simply
to ``doubly color'' the region enveloped by a section of the pen that
goes from $w_0$ to $w_1$ to $\cdots$ to $w_{n+1}$ to~$w_0$. The additional
straight line at the beginning (and a similar one at the end, where it
may be necessary to go from $z_1+w_{n+1}$ to $z_1+w_0$) can be drawn by
the |line_edges| routine; we are thereby saved from the embarrassment that
these lines travel backwards from the current octant direction.

Once we have established the assumption that the curve goes from
$z_0+w_0$ to $z_1+w_{n+1}$, any further retrograde moves that might
occur within the octant can be essentially ignored; we merely need to
keep track of the rightmost edge in each row, in order to compute
the envelope.

Envelope moves consist of offset cubics intermixed with straight line
segments. We record them in a separate |env_move| array, which is
something like |move| but it keeps track of the rightmost position of the
envelope in each row.

@<Glob...@>=
@!env_move:array[0..move_size] of integer;

@ @<Determine the envelope's starting and ending...@>=
w:=link(h);@+if left_transition(p)=diagonal then w:=knil(w);
@!stat if internal[tracing_edges]>unity then
  @<Print a line of diagnostic info to introduce this octant@>;
tats@;@/
ww:=link(h); www:=ww; {starting and ending offsets}
if odd(octant_number[octant]) then www:=knil(www)@+else ww:=knil(ww);
if w<>ww then skew_line_edges(p,w,ww);
end_round(x_coord(p)+x_coord(ww),y_coord(p)+y_coord(ww));
m0:=m1; n0:=n1; d0:=d1;@/
end_round(x_coord(q)+x_coord(www),y_coord(q)+y_coord(www));
if n1-n0>=move_size then overflow("move table size",move_size)
@:METAFONT capacity exceeded move table size}{\quad move table size@>

@ @<Print a line of diagnostic info to introduce this octant@>=
begin print_nl("@@ Octant "); print(octant_dir[octant]);
@:]]]\AT!_Octant}{\.{\AT! Octant...}@>
print(" ("); print_int(info(h)); print(" offset");
if info(h)<>1 then print_char("s");
print("), from ");
print_two_true(x_coord(p)+x_coord(w),y_coord(p)+y_coord(w));@/
ww:=link(h);@+if right_transition(q)=diagonal then ww:=knil(ww);
print(" to ");
print_two_true(x_coord(q)+x_coord(ww),y_coord(q)+y_coord(ww));
end

@ A slight variation of the |line_edges| procedure comes in handy
when we must draw the retrograde lines for nonstandard entry and exit
conditions.

@<Declare the procedure called |skew_line_edges|@>=
procedure skew_line_edges(@!p,@!w,@!ww:pointer);
var @!x0,@!y0,@!x1,@!y1:scaled; {from and to}
begin if (x_coord(w)<>x_coord(ww))or(y_coord(w)<>y_coord(ww)) then
  begin x0:=x_coord(p)+x_coord(w); y0:=y_coord(p)+y_coord(w);@/
  x1:=x_coord(p)+x_coord(ww); y1:=y_coord(p)+y_coord(ww);@/
  unskew(x0,y0,octant); {unskew and unrotate the coordinates}
  x0:=cur_x; y0:=cur_y;@/
  unskew(x1,y1,octant);@/
  @!stat if internal[tracing_edges]>unity then
    begin print_nl("@@ retrograde line from ");
@:]]]\AT!_retro_}{\.{\AT! retrograde line...}@>
  @.retrograde line...@>
    print_two(x0,y0); print(" to "); print_two(cur_x,cur_y); print_nl("");
    end;@+tats@;@/
  line_edges(x0,y0,cur_x,cur_y); {then draw a straight line}
  end;
end;

@ The envelope calculations require more local variables than we needed
in the simpler case of |fill_spec|. At critical points in the computation,
|w| will point to offset $w_k$; |m| and |n| will record the current
lattice positions.  The values of |move_ptr| after the initial and before
the final offset adjustments are stored in |smooth_bot| and |smooth_top|,
respectively.

@<Other local variables for |fill_envelope|@>=
@!m,@!n:integer; {current lattice position}
@!mm0,@!mm1:integer; {skewed equivalents of |m0| and |m1|}
@!k:integer; {current offset number}
@!w,@!ww:pointer; {pointers to the current offset and its neighbor}
@!smooth_bot,@!smooth_top:0..move_size; {boundaries of smoothing}
@!xx,@!yy,@!xp,@!yp,@!delx,@!dely,@!tx,@!ty:scaled;
  {registers for coordinate calculations}

@ @<Make the envelope moves for the current octant...@>=
if odd(octant_number[octant]) then
  begin @<Initialize for ordinary envelope moves@>;
  r:=p; right_type(q):=info(h)+1;
  loop@+  begin if r=q then smooth_top:=move_ptr;
    while right_type(r)<>k do
      @<Insert a line segment to approach the correct offset@>;
    if r=p then smooth_bot:=move_ptr;
    if r=q then goto done;
    move[move_ptr]:=1; n:=move_ptr; s:=link(r);@/
    make_moves(x_coord(r)+x_coord(w),right_x(r)+x_coord(w),
      left_x(s)+x_coord(w),x_coord(s)+x_coord(w),@|
      y_coord(r)+y_coord(w)+half_unit,right_y(r)+y_coord(w)+half_unit,
      left_y(s)+y_coord(w)+half_unit,y_coord(s)+y_coord(w)+half_unit,@|
      xy_corr[octant],y_corr[octant]);@/
    @<Transfer moves from the |move| array to |env_move|@>;
    r:=s;
    end;
done:  @<Insert the new envelope moves in the pixel data@>;
  end
else dual_moves(h,p,q);
right_type(q):=endpoint

@ @<Initialize for ordinary envelope moves@>=
k:=0; w:=link(h); ww:=knil(w);
mm0:=floor_unscaled(x_coord(p)+x_coord(w)-xy_corr[octant]);
mm1:=floor_unscaled(x_coord(q)+x_coord(ww)-xy_corr[octant]);
for n:=0 to n1-n0-1 do env_move[n]:=mm0;
env_move[n1-n0]:=mm1; move_ptr:=0; m:=mm0

@ At this point |n| holds the value of |move_ptr| that was current
when |make_moves| began to record its moves.

@<Transfer moves from the |move| array to |env_move|@>=
repeat m:=m+move[n]-1;
if m>env_move[n] then env_move[n]:=m;
incr(n);
until n>move_ptr

@ Retrograde lines (when |k| decreases) do not need to be recorded in
|env_move| because their edges are not the furthest right in any row.

@<Insert a line segment to approach the correct offset@>=
begin xx:=x_coord(r)+x_coord(w); yy:=y_coord(r)+y_coord(w)+half_unit;
@!stat if internal[tracing_edges]>unity then
  begin print_nl("@@ transition line "); print_int(k); print(", from ");
@:]]]\AT!_trans_}{\.{\AT! transition line...}@>
@.transition line...@>
  print_two_true(xx,yy-half_unit);
  end;@+tats@;@/
if right_type(r)>k then
  begin incr(k); w:=link(w);
  xp:=x_coord(r)+x_coord(w); yp:=y_coord(r)+y_coord(w)+half_unit;
  if yp<>yy then
    @<Record a line segment from |(xx,yy)| to |(xp,yp)| in |env_move|@>;
  end
else  begin decr(k); w:=knil(w);
  xp:=x_coord(r)+x_coord(w); yp:=y_coord(r)+y_coord(w)+half_unit;
  end;
stat if internal[tracing_edges]>unity then
  begin print(" to ");
  print_two_true(xp,yp-half_unit);
  print_nl("");
  end;@+tats@;@/
m:=floor_unscaled(xp-xy_corr[octant]);
move_ptr:=floor_unscaled(yp-y_corr[octant])-n0;
if m>env_move[move_ptr] then env_move[move_ptr]:=m;
end

@ In this step we have |xp>=xx| and |yp>=yy|.

@<Record a line segment from |(xx,yy)| to |(xp,yp)| in |env_move|@>=
begin ty:=floor_scaled(yy-y_corr[octant]); dely:=yp-yy; yy:=yy-ty;
ty:=yp-y_corr[octant]-ty;
if ty>=unity then
  begin delx:=xp-xx; yy:=unity-yy;
  loop@+  begin tx:=take_fraction(delx,make_fraction(yy,dely));
    if ab_vs_cd(tx,dely,delx,yy)+xy_corr[octant]>0 then decr(tx);
    m:=floor_unscaled(xx+tx);
    if m>env_move[move_ptr] then env_move[move_ptr]:=m;
    ty:=ty-unity;
    if ty<unity then goto done1;
    yy:=yy+unity; incr(move_ptr);
    end;
  done1:end;
end

@ @<Insert the new envelope moves in the pixel data@>=
debug if (m<>mm1)or(move_ptr<>n1-n0) then confusion("1");@+gubed@;@/
@:this can't happen /}{\quad 1@>
move[0]:=d0+env_move[0]-mm0;
for n:=1 to move_ptr do
  move[n]:=env_move[n]-env_move[n-1]+1;
move[move_ptr]:=move[move_ptr]-d1;
if internal[smoothing]>0 then smooth_moves(smooth_bot,smooth_top);
move_to_edges(m0,n0,m1,n1);
if right_transition(q)=axis then
  begin w:=link(h); skew_line_edges(q,knil(w),w);
  end

@ We've done it all in the odd-octant case; the only thing remaining
is to repeat the same ideas, upside down and/or backwards.

The following code has been split off as a subprocedure of |fill_envelope|,
because some \PASCAL\ compilers cannot handle procedures as large as
|fill_envelope| would otherwise be.

@<Declare the procedure called |dual_moves|@>=
procedure dual_moves(@!h,@!p,@!q:pointer);
label done,done1;
var @!r,@!s:pointer; {for list traversal}
@<Other local variables for |fill_envelope|@>@;
begin @<Initialize for dual envelope moves@>;
r:=p; {recall that |right_type(q)=endpoint=0| now}
loop@+  begin if r=q then smooth_top:=move_ptr;
  while right_type(r)<>k do
    @<Insert a line segment dually to approach the correct offset@>;
  if r=p then smooth_bot:=move_ptr;
  if r=q then goto done;
  move[move_ptr]:=1; n:=move_ptr; s:=link(r);@/
  make_moves(x_coord(r)+x_coord(w),right_x(r)+x_coord(w),
    left_x(s)+x_coord(w),x_coord(s)+x_coord(w),@|
    y_coord(r)+y_coord(w)+half_unit,right_y(r)+y_coord(w)+half_unit,
    left_y(s)+y_coord(w)+half_unit,y_coord(s)+y_coord(w)+half_unit,@|
    xy_corr[octant],y_corr[octant]);
  @<Transfer moves dually from the |move| array to |env_move|@>;
  r:=s;
  end;
done:@<Insert the new envelope moves dually in the pixel data@>;
end;

@ In the dual case the normal situation is to arrive with a |diagonal|
transition and to leave at the |axis|. The leftmost edge in each row
is relevant instead of the rightmost one.

@<Initialize for dual envelope moves@>=
k:=info(h)+1; ww:=link(h); w:=knil(ww);@/
mm0:=floor_unscaled(x_coord(p)+x_coord(w)-xy_corr[octant]);
mm1:=floor_unscaled(x_coord(q)+x_coord(ww)-xy_corr[octant]);
for n:=1 to n1-n0+1 do env_move[n]:=mm1;
env_move[0]:=mm0; move_ptr:=0; m:=mm0

@ @<Transfer moves dually from the |move| array to |env_move|@>=
repeat if m<env_move[n] then env_move[n]:=m;
m:=m+move[n]-1;
incr(n);
until n>move_ptr

@ Dual retrograde lines occur when |k| increases; the edges of such lines
are not the furthest left in any row.

@<Insert a line segment dually to approach the correct offset@>=
begin xx:=x_coord(r)+x_coord(w); yy:=y_coord(r)+y_coord(w)+half_unit;
@!stat if internal[tracing_edges]>unity then
  begin print_nl("@@ transition line "); print_int(k); print(", from ");
@:]]]\AT!_trans_}{\.{\AT! transition line...}@>
@.transition line...@>
  print_two_true(xx,yy-half_unit);
  end;@+tats@;@/
if right_type(r)<k then
  begin decr(k); w:=knil(w);
  xp:=x_coord(r)+x_coord(w); yp:=y_coord(r)+y_coord(w)+half_unit;
  if yp<>yy then
    @<Record a line segment from |(xx,yy)| to |(xp,yp)| dually in |env_move|@>;
  end
else  begin incr(k); w:=link(w);
  xp:=x_coord(r)+x_coord(w); yp:=y_coord(r)+y_coord(w)+half_unit;
  end;
stat if internal[tracing_edges]>unity then
  begin print(" to ");
  print_two_true(xp,yp-half_unit);
  print_nl("");
  end;@+tats@;@/
m:=floor_unscaled(xp-xy_corr[octant]);
move_ptr:=floor_unscaled(yp-y_corr[octant])-n0;
if m<env_move[move_ptr] then env_move[move_ptr]:=m;
end

@ Again, |xp>=xx| and |yp>=yy|; but this time we are interested in the {\sl
smallest\/} |m| that belongs to a given |move_ptr| position, instead of
the largest~|m|.

@<Record a line segment from |(xx,yy)| to |(xp,yp)| dually in |env_move|@>=
begin ty:=floor_scaled(yy-y_corr[octant]); dely:=yp-yy; yy:=yy-ty;
ty:=yp-y_corr[octant]-ty;
if ty>=unity then
  begin delx:=xp-xx; yy:=unity-yy;
  loop@+  begin if m<env_move[move_ptr] then env_move[move_ptr]:=m;
    tx:=take_fraction(delx,make_fraction(yy,dely));
    if ab_vs_cd(tx,dely,delx,yy)+xy_corr[octant]>0 then decr(tx);
    m:=floor_unscaled(xx+tx);
    ty:=ty-unity; incr(move_ptr);
    if ty<unity then goto done1;
    yy:=yy+unity;
    end;
done1:  if m<env_move[move_ptr] then env_move[move_ptr]:=m;
  end;
end

@ Since |env_move| contains minimum values instead of maximum values, the
finishing-up process is slightly different in the dual case.

@<Insert the new envelope moves dually in the pixel data@>=
debug if (m<>mm1)or(move_ptr<>n1-n0) then confusion("2");@+gubed@;@/
@:this can't happen /}{\quad 2@>
move[0]:=d0+env_move[1]-mm0;
for n:=1 to move_ptr do
  move[n]:=env_move[n+1]-env_move[n]+1;
move[move_ptr]:=move[move_ptr]-d1;
if internal[smoothing]>0 then smooth_moves(smooth_bot,smooth_top);
move_to_edges(m0,n0,m1,n1);
if right_transition(q)=diagonal then
  begin w:=link(h); skew_line_edges(q,w,knil(w));
  end

@* \[25] Elliptical pens.
To get the envelope of a cyclic path with respect to an ellipse, \MF\
calculates the envelope with respect to a polygonal approximation to
the ellipse, using an approach due to John Hobby (Ph.D. thesis,
Stanford University, 1985).
@^Hobby, John Douglas@>
This has two important advantages over trying to obtain the ``exact''
envelope:

\yskip\textindent{1)}It gives better results, because the polygon has been
designed to counteract problems that arise from digitization; the
polygon includes sub-pixel corrections to an exact ellipse that make
the results essentially independent of where the path falls on the raster.
For example, the exact envelope with respect to a pen of diameter~1
blackens a pixel if and only if the path intersects a circle of diameter~1
inscribed in that pixel; the resulting pattern has ``blots'' when the path
is traveling diagonally in unfortunate raster positions. A much better
result is obtained when pixels are blackened only when the path intersects
an inscribed {\sl diamond\/} of diameter~1. Such a diamond is precisely
the polygon that \MF\ uses in the special case of a circle whose diameter is~1.

\yskip\textindent{2)}Polygonal envelopes of cubic splines are cubic
splines, hence it isn't necessary to introduce completely different
routines. By contrast, exact envelopes of cubic splines with respect
to circles are complicated curves, more difficult to plot than cubics.

@ Hobby's construction involves some interesting number theory.
If $u$ and~$v$ are relatively prime integers, we divide the
set of integer points $(m,n)$ into equivalence classes by saying
that $(m,n)$ belongs to class $um+vn$. Then any two integer points
that lie on a line of slope $-u/v$ belong to the same class, because
such points have the form $(m+tv,n-tu)$. Neighboring lines of slope $-u/v$
that go through integer points are separated by distance $1/\psqrt{u^2+v^2}$
from each other, and these lines are perpendicular to lines of slope~$v/u$.
If we start at the origin and travel a distance $k/\psqrt{u^2+v^2}$ in
direction $(u,v)$, we reach the line of slope~$-u/v$ whose points
belong to class~$k$.

For example, let $u=2$ and $v=3$. Then the points $(0,0)$, $(3,-2)$,
$\ldots$ belong to class~0; the points $(-1,1)$, $(2,-1)$, $\ldots$ belong
to class~1; and the distance between these two lines is $1/\sqrt{13}$.
The point $(2,3)$ itself belongs to class~13, hence its distance from
the origin is $13/\sqrt{13}=\sqrt{13}$ (which we already knew).

Suppose we wish to plot envelopes with respect to polygons with
integer vertices. Then the best polygon for curves that travel in
direction $(v,-u)$ will contain the points of class~$k$ such that
$k/\psqrt{u^2+v^2}$ is as close as possible to~$d$, where $d$ is the
maximum distance of the given ellipse from the line $ux+vy=0$.

The |fillin| correction assumes that a diagonal line has an
apparent thickness $$2f\cdot\min(\vert u\vert,\vert v\vert)/\psqrt{u^2+v^2}$$
greater than would be obtained with truly square pixels. (If a
white pixel at an exterior corner is assumed to have apparent
darkness $f_1$ and a black pixel at an interior corner is assumed
to have apparent darkness $1-f_2$, then $f=f_1-f_2$ is the |fillin|
parameter.) Under this assumption we want to choose $k$ so that
$\bigl(k+2f\cdot\min(\vert u\vert,\vert v\vert)\bigr)\big/\psqrt{u^2+v^2}$
is as close as possible to $d$.

Integer coordinates for the vertices work nicely because the thickness of
the envelope at any given slope is independent of the position of the
path with respect to the raster. It turns out, in fact, that the same
property holds for polygons whose vertices have coordinates that are
integer multiples of~$1\over2$, because ellipses are symmetric about
the origin. It's convenient to double all dimensions and require the
resulting polygon to have vertices with integer coordinates. For example,
to get a circle of {\sl diameter}~$r$, we shall compute integer
coordinates for a circle of {\sl radius}~$r$. The circle of radius~$r$
will want to be represented by a polygon that contains the boundary
points $(0,\pm r)$ and~$(\pm r,0)$; later we will divide everything
by~2 and get a polygon with $(0,\pm{1\over2}r)$ and $(\pm{1\over2}r,0)$
on its boundary.

@ In practice the important slopes are those having small values of
$u$ and~$v$; these make regular patterns in which our eyes quickly
spot irregularities. For example, horizontal and vertical lines
(when $u=0$ and $\vert v\vert=1$, or $\vert u\vert=1$ and $v=0$)
are the most important; diagonal lines (when $\vert u\vert=\vert v\vert=1$)
are next; and then come lines with slope $\pm2$ or $\pm1/2$.

The nicest way to generate all rational directions having small
numerators and denominators is to generalize the Stern--Brocot tree
[cf.~{\sl Concrete Mathematics}, section 4.5]
@^Brocot, Achille@>
@^Stern, Moritz Abraham@>
to a ``Stern--Brocot wreath'' as follows: Begin with four nodes
arranged in a circle, containing the respective directions
$(u,v)=(1,0)$, $(0,1)$, $(-1,0)$, and~$(0,-1)$. Then between pairs of
consecutive terms $(u,v)$ and $(u',v')$ of the wreath, insert the
direction $(u+u',v+v')$; continue doing this until some stopping
criterion is fulfilled.

It is not difficult to verify that, regardless of the stopping
criterion, consecutive directions $(u,v)$ and $(u',v')$ of this
wreath will always satisfy the relation $uv'-u'v=1$. Such pairs
of directions have a nice property with respect to the equivalence
classes described above. Let $l$ be a line of equivalent integer points
$(m+tv,n-tu)$ with respect to~$(u,v)$, and let $l'$ be a line of
equivalent integer points $(m'+tv',n'-tu')$ with respect to~$(u',v')$.
Then $l$ and~$l'$ intersect in an integer point $(m'',n'')$, because
the determinant of the linear equations for intersection is $uv'-u'v=1$.
Notice that the class number of $(m'',n'')$ with respect to $(u+u',v+v')$
is the sum of its class numbers with respect to $(u,v)$ and~$(u',v')$.
Moreover, consecutive points on~$l$ and~$l'$ belong to classes that
differ by exactly~1 with respect to $(u+u',v+v')$.

This leads to a nice algorithm in which we construct a polygon having
``correct'' class numbers for as many small-integer directions $(u,v)$
as possible: Assuming that lines $l$ and~$l'$ contain points of the
correct class for $(u,v)$ and~$(u',v')$, respectively, we determine
the intersection $(m'',n'')$ and compute its class with respect to
$(u+u',v+v')$. If the class is too large to be the best approximation,
we move back the proper number of steps from $(m'',n'')$ toward smaller
class numbers on both $l$ and~$l'$, unless this requires moving to points
that are no longer in the polygon; in this way we arrive at two points that
determine a line~$l''$ having the appropriate class. The process continues
recursively, until it cannot proceed without removing the last remaining
point from the class for $(u,v)$ or the class for $(u',v')$.

@ The |make_ellipse| subroutine produces a pointer to a cyclic path
whose vertices define a polygon suitable for envelopes. The control
points on this path will be ignored; in fact, the fields in knot nodes
that are usually reserved for control points are occupied by other
data that helps |make_ellipse| compute the desired polygon.

Parameters |major_axis| and |minor_axis| define the axes of the ellipse;
and parameter |theta| is an angle by which the ellipse is rotated
counterclockwise. If |theta=0|, the ellipse has the equation
$(x/a)^2+(y/b)^2=1$, where |a=major_axis/2| and |b=minor_axis/2|.
In general, the points of the ellipse are generated in the complex plane
by the formula $e^{i\theta}(a\cos t+ib\sin t)$, as $t$~ranges over all
angles. Notice that if |major_axis=minor_axis=d|, we obtain a circle
of diameter~|d|, regardless of the value of |theta|.

The method sketched above is used to produce the elliptical polygon,
except that the main work is done only in the halfplane obtained from
the three starting directions $(0,-1)$, $(1,0)$,~$(0,1)$. Since the ellipse
has circular symmetry, we use the fact that the last half of the polygon
is simply the negative of the first half. Furthermore, we need to compute only
one quarter of the polygon if the ellipse has axis symmetry.

@p function make_ellipse(@!major_axis,@!minor_axis:scaled;
  @!theta:angle):pointer;
label done,done1,found;
var @!p,@!q,@!r,@!s:pointer; {for list manipulation}
@!h:pointer; {head of the constructed knot list}
@!alpha,@!beta,@!gamma,@!delta:integer; {special points}
@!c,@!d:integer; {class numbers}
@!u,@!v:integer; {directions}
@!symmetric:boolean; {should the result be symmetric about the axes?}
begin @<Initialize the ellipse data structure by beginning with
  directions $(0,-1)$, $(1,0)$, $(0,1)$@>;
@<Interpolate new vertices in the ellipse data structure until
  improvement is impossible@>;
if symmetric then
  @<Complete the half ellipse by reflecting the quarter already computed@>;
@<Complete the ellipse by copying the negative of the half already computed@>;
make_ellipse:=h;
end;

@ A special data structure is used only with |make_ellipse|: The
|right_x|, |left_x|, |right_y|, and |left_y| fields of knot nodes
are renamed |right_u|, |left_v|, |right_class|, and |left_length|,
in order to store information that simplifies the necessary computations.

If |p| and |q| are consecutive knots in this data structure, the
|x_coord| and |y_coord| fields of |p| and~|q| contain current vertices
of the polygon; their values are integer multiples
of |half_unit|. Both of these vertices belong to equivalence class
|right_class(p)| with respect to the direction
$\bigl($|right_u(p),left_v(q)|$\bigr)$. The number of points of this class
on the line from vertex~|p| to vertex~|q| is |1+left_length(q)|.
In particular, |left_length(q)=0| means that |x_coord(p)=x_coord(q)|
and |y_coord(p)=y_coord(q)|; such duplicate vertices will be
discarded during the course of the algorithm.

The contents of |right_u(p)| and |left_v(q)| are integer multiples
of |half_unit|, just like the coordinate fields. Hence, for example,
the point $\bigl($|x_coord(p)-left_v(q),y_coord(p)+right_u(p)|$\bigr)$
also belongs to class number |right_class(p)|. This point is one
step closer to the vertex in node~|q|; it equals that vertex
if and only if |left_length(q)=1|.

The |left_type| and |right_type| fields are not used, but |link|
has its normal meaning.

To start the process, we create four nodes for the three directions
$(0,-1)$, $(1,0)$, and $(0,1)$. The corresponding vertices are
$(-\alpha,-\beta)$, $(\gamma,-\beta)$, $(\gamma,\beta)$, and
$(\alpha,\beta)$, where $(\alpha,\beta)$ is a half-integer approximation
to where the ellipse rises highest above the $x$-axis, and where
$\gamma$ is a half-integer approximation to the maximum $x$~coordinate
of the ellipse. The fourth of these nodes is not actually calculated
if the ellipse has axis symmetry.

@d right_u==right_x {|u| value for a pen edge}
@d left_v==left_x {|v| value for a pen edge}
@d right_class==right_y {equivalence class number of a pen edge}
@d left_length==left_y {length of a pen edge}

@<Initialize the ellipse data structure...@>=
@<Calculate integers $\alpha$, $\beta$, $\gamma$ for the vertex
  coordinates@>;
p:=get_node(knot_node_size); q:=get_node(knot_node_size);
r:=get_node(knot_node_size);
if symmetric then s:=null@+else s:=get_node(knot_node_size);
h:=p; link(p):=q; link(q):=r; link(r):=s; {|s=null| or |link(s)=null|}
@<Revise the values of $\alpha$, $\beta$, $\gamma$, if necessary,
  so that degenerate lines of length zero will not be obtained@>;
x_coord(p):=-alpha*half_unit;
y_coord(p):=-beta*half_unit;
x_coord(q):=gamma*half_unit;@/
y_coord(q):=y_coord(p); x_coord(r):=x_coord(q);@/
right_u(p):=0; left_v(q):=-half_unit;@/
right_u(q):=half_unit; left_v(r):=0;@/
right_u(r):=0;
right_class(p):=beta; right_class(q):=gamma; right_class(r):=beta;@/
left_length(q):=gamma+alpha;
if symmetric then
  begin y_coord(r):=0; left_length(r):=beta;
  end
else  begin y_coord(r):=-y_coord(p); left_length(r):=beta+beta;@/
  x_coord(s):=-x_coord(p); y_coord(s):=y_coord(r);@/
  left_v(s):=half_unit; left_length(s):=gamma-alpha;
  end

@ One of the important invariants of the pen data structure is that
the points are distinct. We may need to correct the pen specification
in order to avoid this. (The result of \&{pencircle} will always be at
least one pixel wide and one pixel tall, although \&{makepen} is
capable of producing smaller pens.)

@<Revise the values of $\alpha$, $\beta$, $\gamma$, if necessary...@>=
if beta=0 then beta:=1;
if gamma=0 then gamma:=1;
if gamma<=abs(alpha) then
  if alpha>0 then alpha:=gamma-1
  else alpha:=1-gamma

@ If $a$ and $b$ are the semi-major and semi-minor axes,
the given ellipse rises highest above the $x$-axis at the point
$\bigl((a^2-b^2)\sin\theta\cos\theta/\rho\bigr)+i\rho$, where
$\rho=\sqrt{(a\sin\theta)^2+(b\cos\theta)^2}$. It reaches
furthest to the right of~the $y$-axis at the point
$\sigma+i(a^2-b^2)\sin\theta\cos\theta/\sigma$, where
$\sigma=\sqrt{(a\cos\theta)^2+(b\sin\theta)^2}$.

@<Calculate integers $\alpha$, $\beta$, $\gamma$...@>=
if (major_axis=minor_axis)or(theta mod ninety_deg=0) then
  begin symmetric:=true; alpha:=0;
  if odd(theta div ninety_deg) then
    begin beta:=major_axis; gamma:=minor_axis;
    n_sin:=fraction_one; n_cos:=0; {|n_sin| and |n_cos| are used later}
    end
  else  begin beta:=minor_axis; gamma:=major_axis; theta:=0;
    end; {|n_sin| and |n_cos| aren't needed in this case}
  end
else  begin symmetric:=false;
  n_sin_cos(theta); {set up $|n_sin|=\sin\theta$ and $|n_cos|=\cos\theta$}
  gamma:=take_fraction(major_axis,n_sin);
  delta:=take_fraction(minor_axis,n_cos);
  beta:=pyth_add(gamma,delta);
  alpha:=take_fraction(take_fraction(major_axis,
      make_fraction(gamma,beta)),n_cos)@|
    -take_fraction(take_fraction(minor_axis,
      make_fraction(delta,beta)),n_sin);
  alpha:=(alpha+half_unit) div unity;
  gamma:=pyth_add(take_fraction(major_axis,n_cos),
    take_fraction(minor_axis,n_sin));
  end;
beta:=(beta+half_unit) div unity;
gamma:=(gamma+half_unit) div unity

@ Now |p|, |q|, and |r| march through the list, always representing
three consecutive vertices and two consecutive slope directions.
When a new slope is interpolated, we back up slightly, until
further refinement is impossible; then we march forward again.
The somewhat magical operations performed in this part of the
algorithm are justified by the theory sketched earlier.
Complications arise only from the need to keep zero-length lines
out of the final data structure.

@<Interpolate new vertices in the ellipse data structure...@>=
loop@+  begin u:=right_u(p)+right_u(q); v:=left_v(q)+left_v(r);
  c:=right_class(p)+right_class(q);@/
  @<Compute the distance |d| from class~0 to the edge of the ellipse
    in direction |(u,v)|, times $\psqrt{u^2+v^2}$,
    rounded to the nearest integer@>;
  delta:=c-d; {we want to move |delta| steps back
      from the intersection vertex~|q|}
  if delta>0 then
    begin if delta>left_length(r) then delta:=left_length(r);
    if delta>=left_length(q) then
      @<Remove the line from |p| to |q|,
        and adjust vertex~|q| to introduce a new line@>
    else @<Insert a new line for direction |(u,v)| between |p| and~|q|@>;
    end
  else p:=q;
  @<Move to the next remaining triple |(p,q,r)|, removing and skipping past
    zero-length lines that might be present; |goto done| if all
    triples have been processed@>;
  end;
done:

@ The appearance of a zero-length line means that we should advance |p|
past it. We must not try to straddle a missing direction, because the
algorithm works only on consecutive pairs of directions.

@<Move to the next remaining triple |(p,q,r)|...@>=
loop@+  begin q:=link(p);
  if q=null then goto done;
  if left_length(q)=0 then
    begin link(p):=link(q); right_class(p):=right_class(q);
    right_u(p):=right_u(q); free_node(q,knot_node_size);
    end
  else  begin r:=link(q);
    if r=null then goto done;
    if left_length(r)=0 then
      begin link(p):=r; free_node(q,knot_node_size); p:=r;
      end
    else goto found;
    end;
  end;
found:

@ The `\&{div} 8' near the end of this step comes from
the fact that |delta| is scaled by~$2^{15}$ and $d$~by~$2^{16}$,
while |take_fraction| removes a scale factor of~$2^{28}$.
We also make sure that $d\G\max(\vert u\vert,\vert v\vert)$, so that
the pen will always include a circular pen of diameter~1 as a subset;
then it won't be possible to get disconnected path envelopes.

@<Compute the distance |d| from class~0 to the edge of the ellipse...@>=
delta:=pyth_add(u,v);
if major_axis=minor_axis then d:=major_axis {circles are easy}
else  begin if theta=0 then
    begin alpha:=u; beta:=v;
    end
  else  begin alpha:=take_fraction(u,n_cos)+take_fraction(v,n_sin);
    beta:=take_fraction(v,n_cos)-take_fraction(u,n_sin);
    end;
  alpha:=make_fraction(alpha,delta);
  beta:=make_fraction(beta,delta);
  d:=pyth_add(take_fraction(major_axis,alpha),
    take_fraction(minor_axis,beta));
  end;
alpha:=abs(u); beta:=abs(v);
if alpha<beta then
  begin alpha:=abs(v); beta:=abs(u);
  end; {now $\alpha=\max(\vert u\vert,\vert v\vert)$,
      $\beta=\min(\vert u\vert,\vert v\vert)$}
if internal[fillin]<>0 then
  d:=d-take_fraction(internal[fillin],make_fraction(beta+beta,delta));
d:=take_fraction((d+4) div 8,delta); alpha:=alpha div half_unit;
if d<alpha then d:=alpha

@ At this point there's a line of length |<=delta| from vertex~|p|
to vertex~|q|, orthogonal to direction $\bigl($|right_u(p),left_v(q)|$\bigr)$;
and there's a line of length |>=delta| from vertex~|q|
to vertex~|r|, orthogonal to direction $\bigl($|right_u(q),left_v(r)|$\bigr)$.
The best line to direction $(u,v)$ should replace the line from
|p| to~|q|; this new line will have the same length as the old.

@<Remove the line from |p| to |q|...@>=
begin delta:=left_length(q);@/
right_class(p):=c-delta; right_u(p):=u; left_v(q):=v;@/
x_coord(q):=x_coord(q)-delta*left_v(r);
y_coord(q):=y_coord(q)+delta*right_u(q);@/
left_length(r):=left_length(r)-delta;
end

@ Here is the main case, now that we have dealt with the exception:
We insert a new line of length |delta| for direction |(u,v)|, decreasing
each of the adjacent lines by |delta| steps.

@<Insert a new line for direction |(u,v)| between |p| and~|q|@>=
begin s:=get_node(knot_node_size); link(p):=s; link(s):=q;@/
x_coord(s):=x_coord(q)+delta*left_v(q);
y_coord(s):=y_coord(q)-delta*right_u(p);@/
x_coord(q):=x_coord(q)-delta*left_v(r);
y_coord(q):=y_coord(q)+delta*right_u(q);@/
left_v(s):=left_v(q); right_u(s):=u; left_v(q):=v;@/
right_class(s):=c-delta;@/
left_length(s):=left_length(q)-delta; left_length(q):=delta;
left_length(r):=left_length(r)-delta;
end

@ Only the coordinates need to be copied, not the class numbers and other stuff.
At this point either |link(p)| or |link(link(p))| is |null|.

@<Complete the half ellipse...@>=
begin s:=null; q:=h;
loop@+  begin r:=get_node(knot_node_size); link(r):=s; s:=r;@/
  x_coord(s):=x_coord(q); y_coord(s):=-y_coord(q);
  if q=p then goto done1;
  q:=link(q);
  if y_coord(q)=0 then goto done1;
  end;
done1: if (link(p)<>null) then free_node(link(p),knot_node_size);
link(p):=s; beta:=-y_coord(h);
while y_coord(p)<>beta do p:=link(p);
q:=link(p);
end

@ Now we use a somewhat tricky fact: The pointer |q| will be null if and
only if the line for the final direction $(0,1)$ has been removed. If
that line still survives, it should be combined with a possibly
surviving line in the initial direction $(0,-1)$.

@<Complete the ellipse by copying...@>=
if q<>null then
  begin if right_u(h)=0 then
    begin p:=h; h:=link(h); free_node(p,knot_node_size);@/
    x_coord(q):=-x_coord(h);
    end;
  p:=q;
  end
else q:=p;
r:=link(h); {now |p=q|, |x_coord(p)=-x_coord(h)|, |y_coord(p)=-y_coord(h)|}
repeat s:=get_node(knot_node_size); link(p):=s; p:=s;@/
x_coord(p):=-x_coord(r); y_coord(p):=-y_coord(r); r:=link(r);
until r=q;
link(p):=h

@* \[26] Direction and intersection times.
A path of length $n$ is defined parametrically by functions $x(t)$ and
$y(t)$, for |0<=t<=n|; we can regard $t$ as the ``time'' at which the path
reaches the point $\bigl(x(t),y(t)\bigr)$.  In this section of the program
we shall consider operations that determine special times associated with
given paths: the first time that a path travels in a given direction, and
a pair of times at which two paths cross each other.

@ Let's start with the easier task. The function |find_direction_time| is
given a direction |(x,y)| and a path starting at~|h|. If the path never
travels in direction |(x,y)|, the direction time will be~|-1|; otherwise
it will be nonnegative.

Certain anomalous cases can arise: If |(x,y)=(0,0)|, so that the given
direction is undefined, the direction time will be~0. If $\bigl(x'(t),
y'(t)\bigr)=(0,0)$, so that the path direction is undefined, it will be
assumed to match any given direction at time~|t|.

The routine solves this problem in nondegenerate cases by rotating the path
and the given direction so that |(x,y)=(1,0)|; i.e., the main task will be
to find when a given path first travels ``due east.''

@p function find_direction_time(@!x,@!y:scaled;@!h:pointer):scaled;
label exit,found,not_found,done;
var @!max:scaled; {$\max\bigl(\vert x\vert,\vert y\vert\bigr)$}
@!p,@!q:pointer; {for list traversal}
@!n:scaled; {the direction time at knot |p|}
@!tt:scaled; {the direction time within a cubic}
@<Other local variables for |find_direction_time|@>@;
begin @<Normalize the given direction for better accuracy;
  but |return| with zero result if it's zero@>;
n:=0; p:=h;
loop@+  begin if right_type(p)=endpoint then goto not_found;
  q:=link(p);
  @<Rotate the cubic between |p| and |q|; then
    |goto found| if the rotated cubic travels due east at some time |tt|;
    but |goto not_found| if an entire cyclic path has been traversed@>;
  p:=q; n:=n+unity;
  end;
not_found: find_direction_time:=-unity; return;
found: find_direction_time:=n+tt;
exit:end;

@ @<Normalize the given direction for better accuracy...@>=
if abs(x)<abs(y) then
  begin x:=make_fraction(x,abs(y));
  if y>0 then y:=fraction_one@+else y:=-fraction_one;
  end
else if x=0 then
  begin find_direction_time:=0; return;
  end
else  begin y:=make_fraction(y,abs(x));
  if x>0 then x:=fraction_one@+else x:=-fraction_one;
  end

@ Since we're interested in the tangent directions, we work with the
derivative $${1\over3}B'(x_0,x_1,x_2,x_3;t)=
B(x_1-x_0,x_2-x_1,x_3-x_2;t)$$ instead of
$B(x_0,x_1,x_2,x_3;t)$ itself. The derived coefficients are also scaled up
in order to achieve better accuracy.

The given path may turn abruptly at a knot, and it might pass the critical
tangent direction at such a time. Therefore we remember the direction |phi|
in which the previous rotated cubic was traveling. (The value of |phi| will be
undefined on the first cubic, i.e., when |n=0|.)

@<Rotate the cubic between |p| and |q|; then...@>=
tt:=0;
@<Set local variables |x1,x2,x3| and |y1,y2,y3| to multiples of the control
  points of the rotated derivatives@>;
if y1=0 then if x1>=0 then goto found;
if n>0 then
  begin @<Exit to |found| if an eastward direction occurs at knot |p|@>;
  if p=h then goto not_found;
  end;
if (x3<>0)or(y3<>0) then phi:=n_arg(x3,y3);
@<Exit to |found| if the curve whose derivatives are specified by
  |x1,x2,x3,y1,y2,y3| travels eastward at some time~|tt|@>

@ @<Other local variables for |find_direction_time|@>=
@!x1,@!x2,@!x3,@!y1,@!y2,@!y3:scaled; {multiples of rotated derivatives}
@!theta,@!phi:angle; {angles of exit and entry at a knot}
@!t:fraction; {temp storage}

@ @<Set local variables |x1,x2,x3| and |y1,y2,y3| to multiples...@>=
x1:=right_x(p)-x_coord(p); x2:=left_x(q)-right_x(p);
x3:=x_coord(q)-left_x(q);@/
y1:=right_y(p)-y_coord(p); y2:=left_y(q)-right_y(p);
y3:=y_coord(q)-left_y(q);@/
max:=abs(x1);
if abs(x2)>max then max:=abs(x2);
if abs(x3)>max then max:=abs(x3);
if abs(y1)>max then max:=abs(y1);
if abs(y2)>max then max:=abs(y2);
if abs(y3)>max then max:=abs(y3);
if max=0 then goto found;
while max<fraction_half do
  begin double(max); double(x1); double(x2); double(x3);
  double(y1); double(y2); double(y3);
  end;
t:=x1; x1:=take_fraction(x1,x)+take_fraction(y1,y);
y1:=take_fraction(y1,x)-take_fraction(t,y);@/
t:=x2; x2:=take_fraction(x2,x)+take_fraction(y2,y);
y2:=take_fraction(y2,x)-take_fraction(t,y);@/
t:=x3; x3:=take_fraction(x3,x)+take_fraction(y3,y);
y3:=take_fraction(y3,x)-take_fraction(t,y)

@ @<Exit to |found| if an eastward direction occurs at knot |p|@>=
theta:=n_arg(x1,y1);
if theta>=0 then if phi<=0 then if phi>=theta-one_eighty_deg then goto found;
if theta<=0 then if phi>=0 then if phi<=theta+one_eighty_deg then goto found

@ In this step we want to use the |crossing_point| routine to find the
roots of the quadratic equation $B(y_1,y_2,y_3;t)=0$.
Several complications arise: If the quadratic equation has a double root,
the curve never crosses zero, and |crossing_point| will find nothing;
this case occurs iff $y_1y_3=y_2^2$ and $y_1y_2<0$. If the quadratic
equation has simple roots, or only one root, we may have to negate it
so that $B(y_1,y_2,y_3;t)$ crosses from positive to negative at its first root.
And finally, we need to do special things if $B(y_1,y_2,y_3;t)$ is
identically zero.

@ @<Exit to |found| if the curve whose derivatives are specified by...@>=
if x1<0 then if x2<0 then if x3<0 then goto done;
if ab_vs_cd(y1,y3,y2,y2)=0 then
  @<Handle the test for eastward directions when $y_1y_3=y_2^2$;
    either |goto found| or |goto done|@>;
if y1<=0 then
  if y1<0 then
    begin y1:=-y1; y2:=-y2; y3:=-y3;
    end
  else if y2>0 then
    begin y2:=-y2; y3:=-y3;
    end;
@<Check the places where $B(y_1,y_2,y_3;t)=0$ to see if
  $B(x_1,x_2,x_3;t)\ge0$@>;
done:

@ The quadratic polynomial $B(y_1,y_2,y_3;t)$ begins |>=0| and has at most
two roots, because we know that it isn't identically zero.

It must be admitted that the |crossing_point| routine is not perfectly accurate;
rounding errors might cause it to find a root when $y_1y_3>y_2^2$, or to
miss the roots when $y_1y_3<y_2^2$. The rotation process is itself
subject to rounding errors. Yet this code optimistically tries to
do the right thing.

@d we_found_it==begin tt:=(t+@'4000) div @'10000; goto found;
  end

@<Check the places where $B(y_1,y_2,y_3;t)=0$...@>=
t:=crossing_point(y1,y2,y3);
if t>fraction_one then goto done;
y2:=t_of_the_way(y2)(y3);
x1:=t_of_the_way(x1)(x2);
x2:=t_of_the_way(x2)(x3);
x1:=t_of_the_way(x1)(x2);
if x1>=0 then we_found_it;
if y2>0 then y2:=0;
tt:=t; t:=crossing_point(0,-y2,-y3);
if t>fraction_one then goto done;
x1:=t_of_the_way(x1)(x2);
x2:=t_of_the_way(x2)(x3);
if t_of_the_way(x1)(x2)>=0 then
  begin t:=t_of_the_way(tt)(fraction_one); we_found_it;
  end

@ @<Handle the test for eastward directions when $y_1y_3=y_2^2$;
    either |goto found| or |goto done|@>=
begin if ab_vs_cd(y1,y2,0,0)<0 then
  begin t:=make_fraction(y1,y1-y2);
  x1:=t_of_the_way(x1)(x2);
  x2:=t_of_the_way(x2)(x3);
  if t_of_the_way(x1)(x2)>=0 then we_found_it;
  end
else if y3=0 then
  if y1=0 then
    @<Exit to |found| if the derivative $B(x_1,x_2,x_3;t)$ becomes |>=0|@>
  else if x3>=0 then
    begin tt:=unity; goto found;
    end;
goto done;
end

@ At this point we know that the derivative of |y(t)| is identically zero,
and that |x1<0|; but either |x2>=0| or |x3>=0|, so there's some hope of
traveling east.

@<Exit to |found| if the derivative $B(x_1,x_2,x_3;t)$ becomes |>=0|...@>=
begin t:=crossing_point(-x1,-x2,-x3);
if t<=fraction_one then we_found_it;
if ab_vs_cd(x1,x3,x2,x2)<=0 then
  begin t:=make_fraction(x1,x1-x2); we_found_it;
  end;
end

@ The intersection of two cubics can be found by an interesting variant
of the general bisection scheme described in the introduction to |make_moves|.\
Given $w(t)=B(w_0,w_1,w_2,w_3;t)$ and $z(t)=B(z_0,z_1,z_2,z_3;t)$,
we wish to find a pair of times $(t_1,t_2)$ such that $w(t_1)=z(t_2)$,
if an intersection exists. First we find the smallest rectangle that
encloses the points $\{w_0,w_1,w_2,w_3\}$ and check that it overlaps
the smallest rectangle that encloses
$\{z_0,z_1,z_2,z_3\}$; if not, the cubics certainly don't intersect.
But if the rectangles do overlap, we bisect the intervals, getting
new cubics $w'$ and~$w''$, $z'$~and~$z''$; the intersection routine first
tries for an intersection between $w'$ and~$z'$, then (if unsuccessful)
between $w'$ and~$z''$, then (if still unsuccessful) between $w''$ and~$z'$,
finally (if thrice unsuccessful) between $w''$ and~$z''$. After $l$~successful
levels of bisection we will have determined the intersection times $t_1$
and~$t_2$ to $l$~bits of accuracy.

\def\submin{_{\rm min}} \def\submax{_{\rm max}}
As before, it is better to work with the numbers $W_k=2^l(w_k-w_{k-1})$
and $Z_k=2^l(z_k-z_{k-1})$ rather than the coefficients $w_k$ and $z_k$
themselves. We also need one other quantity, $\Delta=2^l(w_0-z_0)$,
to determine when the enclosing rectangles overlap. Here's why:
The $x$~coordinates of~$w(t)$ are between $u\submin$ and $u\submax$,
and the $x$~coordinates of~$z(t)$ are between $x\submin$ and $x\submax$,
if we write $w_k=(u_k,v_k)$ and $z_k=(x_k,y_k)$ and $u\submin=
\min(u_0,u_1,u_2,u_3)$, etc. These intervals of $x$~coordinates
overlap if and only if $u\submin\L x\submax$ and
$x\submin\L u\submax$. Letting
$$U\submin=\min(0,U_1,U_1+U_2,U_1+U_2+U_3),\;
  U\submax=\max(0,U_1,U_1+U_2,U_1+U_2+U_3),$$
we have $2^lu\submin=2^lu_0+U\submin$, etc.; the condition for overlap
reduces to
$$X\submin-U\submax\L 2^l(u_0-x_0)\L X\submax-U\submin.$$
Thus we want to maintain the quantity $2^l(u_0-x_0)$; similarly,
the quantity $2^l(v_0-y_0)$ accounts for the $y$~coordinates. The
coordinates of $\Delta=2^l(w_0-z_0)$ must stay bounded as $l$ increases,
because of the overlap condition; i.e., we know that $X\submin$,
$X\submax$, and their relatives are bounded, hence $X\submax-
U\submin$ and $X\submin-U\submax$ are bounded.

@ Incidentally, if the given cubics intersect more than once, the process
just sketched will not necessarily find the lexicographically smallest pair
$(t_1,t_2)$. The solution actually obtained will be smallest in ``shuffled
order''; i.e., if $t_1=(.a_1a_2\ldots a_{16})_2$ and
$t_2=(.b_1b_2\ldots b_{16})_2$, then we will minimize
$a_1b_1a_2b_2\ldots a_{16}b_{16}$, not
$a_1a_2\ldots a_{16}b_1b_2\ldots b_{16}$.
Shuffled order agrees with lexicographic order if all pairs of solutions
$(t_1,t_2)$ and $(t_1',t_2')$ have the property that $t_1<t_1'$ iff
$t_2<t_2'$; but in general, lexicographic order can be quite different,
and the bisection algorithm would be substantially less efficient if it were
constrained by lexicographic order.

For example, suppose that an overlap has been found for $l=3$ and
$(t_1,t_2)= (.101,.011)$ in binary, but that no overlap is produced by
either of the alternatives $(.1010,.0110)$, $(.1010,.0111)$ at level~4.
Then there is probably an intersection in one of the subintervals
$(.1011,.011x)$; but lexicographic order would require us to explore
$(.1010,.1xxx)$ and $(.1011,.00xx)$ and $(.1011,.010x)$ first. We wouldn't
want to store all of the subdivision data for the second path, so the
subdivisions would have to be regenerated many times. Such inefficiencies
would be associated with every `1' in the binary representation of~$t_1$.

@ The subdivision process introduces rounding errors, hence we need to
make a more liberal test for overlap. It is not hard to show that the
computed values of $U_i$ differ from the truth by at most~$l$, on
level~$l$, hence $U\submin$ and $U\submax$ will be at most $3l$ in error.
If $\beta$ is an upper bound on the absolute error in the computed
components of $\Delta=(|delx|,|dely|)$ on level~$l$, we will replace
the test `$X\submin-U\submax\L|delx|$' by the more liberal test
`$X\submin-U\submax\L|delx|+|tol|$', where $|tol|=6l+\beta$.

More accuracy is obtained if we try the algorithm first with |tol=0|;
the more liberal tolerance is used only if an exact approach fails.
It is convenient to do this double-take by letting `3' in the preceding
paragraph be a parameter, which is first 0, then 3.

@<Glob...@>=
@!tol_step:0..6; {either 0 or 3, usually}

@ We shall use an explicit stack to implement the recursive bisection
method described above. In fact, the |bisect_stack| array is available for
this purpose. It will contain numerous 5-word packets like
$(U_1,U_2,U_3,U\submin,U\submax)$, as well as 20-word packets comprising
the 5-word packets for $U$, $V$, $X$, and~$Y$.

The following macros define the allocation of stack positions to
the quantities needed for bisection-intersection.

@d stack_1(#)==bisect_stack[#] {$U_1$, $V_1$, $X_1$, or $Y_1$}
@d stack_2(#)==bisect_stack[#+1] {$U_2$, $V_2$, $X_2$, or $Y_2$}
@d stack_3(#)==bisect_stack[#+2] {$U_3$, $V_3$, $X_3$, or $Y_3$}
@d stack_min(#)==bisect_stack[#+3]
  {$U\submin$, $V\submin$, $X\submin$, or $Y\submin$}
@d stack_max(#)==bisect_stack[#+4]
  {$U\submax$, $V\submax$, $X\submax$, or $Y\submax$}
@d int_packets=20 {number of words to represent $U_k$, $V_k$, $X_k$, and $Y_k$}
@#
@d u_packet(#)==#-5
@d v_packet(#)==#-10
@d x_packet(#)==#-15
@d y_packet(#)==#-20
@d l_packets==bisect_ptr-int_packets
@d r_packets==bisect_ptr
@d ul_packet==u_packet(l_packets) {base of $U'_k$ variables}
@d vl_packet==v_packet(l_packets) {base of $V'_k$ variables}
@d xl_packet==x_packet(l_packets) {base of $X'_k$ variables}
@d yl_packet==y_packet(l_packets) {base of $Y'_k$ variables}
@d ur_packet==u_packet(r_packets) {base of $U''_k$ variables}
@d vr_packet==v_packet(r_packets) {base of $V''_k$ variables}
@d xr_packet==x_packet(r_packets) {base of $X''_k$ variables}
@d yr_packet==y_packet(r_packets) {base of $Y''_k$ variables}
@#
@d u1l==stack_1(ul_packet) {$U'_1$}
@d u2l==stack_2(ul_packet) {$U'_2$}
@d u3l==stack_3(ul_packet) {$U'_3$}
@d v1l==stack_1(vl_packet) {$V'_1$}
@d v2l==stack_2(vl_packet) {$V'_2$}
@d v3l==stack_3(vl_packet) {$V'_3$}
@d x1l==stack_1(xl_packet) {$X'_1$}
@d x2l==stack_2(xl_packet) {$X'_2$}
@d x3l==stack_3(xl_packet) {$X'_3$}
@d y1l==stack_1(yl_packet) {$Y'_1$}
@d y2l==stack_2(yl_packet) {$Y'_2$}
@d y3l==stack_3(yl_packet) {$Y'_3$}
@d u1r==stack_1(ur_packet) {$U''_1$}
@d u2r==stack_2(ur_packet) {$U''_2$}
@d u3r==stack_3(ur_packet) {$U''_3$}
@d v1r==stack_1(vr_packet) {$V''_1$}
@d v2r==stack_2(vr_packet) {$V''_2$}
@d v3r==stack_3(vr_packet) {$V''_3$}
@d x1r==stack_1(xr_packet) {$X''_1$}
@d x2r==stack_2(xr_packet) {$X''_2$}
@d x3r==stack_3(xr_packet) {$X''_3$}
@d y1r==stack_1(yr_packet) {$Y''_1$}
@d y2r==stack_2(yr_packet) {$Y''_2$}
@d y3r==stack_3(yr_packet) {$Y''_3$}
@#
@d stack_dx==bisect_stack[bisect_ptr] {stacked value of |delx|}
@d stack_dy==bisect_stack[bisect_ptr+1] {stacked value of |dely|}
@d stack_tol==bisect_stack[bisect_ptr+2] {stacked value of |tol|}
@d stack_uv==bisect_stack[bisect_ptr+3] {stacked value of |uv|}
@d stack_xy==bisect_stack[bisect_ptr+4] {stacked value of |xy|}
@d int_increment=int_packets+int_packets+5 {number of stack words per level}

@<Check the ``constant''...@>=
if int_packets+17*int_increment>bistack_size then bad:=32;

@ Computation of the min and max is a tedious but fairly fast sequence of
instructions; exactly four comparisons are made in each branch.

@d set_min_max(#)==
  if stack_1(#)<0 then
    if stack_3(#)>=0 then
      begin if stack_2(#)<0 then stack_min(#):=stack_1(#)+stack_2(#)
        else stack_min(#):=stack_1(#);
      stack_max(#):=stack_1(#)+stack_2(#)+stack_3(#);
      if stack_max(#)<0 then stack_max(#):=0;
      end
    else  begin stack_min(#):=stack_1(#)+stack_2(#)+stack_3(#);
      if stack_min(#)>stack_1(#) then stack_min(#):=stack_1(#);
      stack_max(#):=stack_1(#)+stack_2(#);
      if stack_max(#)<0 then stack_max(#):=0;
      end
  else if stack_3(#)<=0 then
    begin if stack_2(#)>0 then stack_max(#):=stack_1(#)+stack_2(#)
      else stack_max(#):=stack_1(#);
    stack_min(#):=stack_1(#)+stack_2(#)+stack_3(#);
    if stack_min(#)>0 then stack_min(#):=0;
    end
  else  begin stack_max(#):=stack_1(#)+stack_2(#)+stack_3(#);
    if stack_max(#)<stack_1(#) then stack_max(#):=stack_1(#);
    stack_min(#):=stack_1(#)+stack_2(#);
    if stack_min(#)>0 then stack_min(#):=0;
    end

@ It's convenient to keep the current values of $l$, $t_1$, and $t_2$ in
the integer form $2^l+2^lt_1$ and $2^l+2^lt_2$. The |cubic_intersection|
routine uses global variables |cur_t| and |cur_tt| for this purpose;
after successful completion, |cur_t| and |cur_tt| will contain |unity|
plus the |scaled| values of $t_1$ and~$t_2$.

The values of |cur_t| and |cur_tt| will be set to zero if |cubic_intersection|
finds no intersection. The routine gives up and gives an approximate answer
if it has backtracked
more than 5000 times (otherwise there are cases where several minutes
of fruitless computation would be possible).

@d max_patience=5000

@<Glob...@>=
@!cur_t,@!cur_tt:integer; {controls and results of |cubic_intersection|}
@!time_to_go:integer; {this many backtracks before giving up}
@!max_t:integer; {maximum of $2^{l+1}$ so far achieved}

@ The given cubics $B(w_0,w_1,w_2,w_3;t)$ and
$B(z_0,z_1,z_2,z_3;t)$ are specified in adjacent knot nodes |(p,link(p))|
and |(pp,link(pp))|, respectively.

@p procedure cubic_intersection(@!p,@!pp:pointer);
label continue, not_found, exit;
var @!q,@!qq:pointer; {|link(p)|, |link(pp)|}
begin time_to_go:=max_patience; max_t:=2;
@<Initialize for intersections at level zero@>;
loop@+  begin continue:
  if delx-tol<=stack_max(x_packet(xy))-stack_min(u_packet(uv)) then
   if delx+tol>=stack_min(x_packet(xy))-stack_max(u_packet(uv)) then
   if dely-tol<=stack_max(y_packet(xy))-stack_min(v_packet(uv)) then
   if dely+tol>=stack_min(y_packet(xy))-stack_max(v_packet(uv)) then
    begin if cur_t>=max_t then
      begin if max_t=two then {we've done 17 bisections}
        begin cur_t:=half(cur_t+1); cur_tt:=half(cur_tt+1); return;
        end;
      double(max_t); appr_t:=cur_t; appr_tt:=cur_tt;
      end;
    @<Subdivide for a new level of intersection@>;
    goto continue;
    end;
  if time_to_go>0 then decr(time_to_go)
  else  begin while appr_t<unity do
      begin double(appr_t); double(appr_tt);
      end;
    cur_t:=appr_t; cur_tt:=appr_tt; return;
    end;
  @<Advance to the next pair |(cur_t,cur_tt)|@>;
  end;
exit:end;

@ The following variables are global, although they are used only by
|cubic_intersection|, because it is necessary on some machines to
split |cubic_intersection| up into two procedures.

@<Glob...@>=
@!delx,@!dely:integer; {the components of $\Delta=2^l(w_0-z_0)$}
@!tol:integer; {bound on the uncertainty in the overlap test}
@!uv,@!xy:0..bistack_size; {pointers to the current packets of interest}
@!three_l:integer; {|tol_step| times the bisection level}
@!appr_t,@!appr_tt:integer; {best approximations known to the answers}

@ We shall assume that the coordinates are sufficiently non-extreme that
integer overflow will not occur.
@^overflow in arithmetic@>

@<Initialize for intersections at level zero@>=
q:=link(p); qq:=link(pp); bisect_ptr:=int_packets;@/
u1r:=right_x(p)-x_coord(p); u2r:=left_x(q)-right_x(p);
u3r:=x_coord(q)-left_x(q); set_min_max(ur_packet);@/
v1r:=right_y(p)-y_coord(p); v2r:=left_y(q)-right_y(p);
v3r:=y_coord(q)-left_y(q); set_min_max(vr_packet);@/
x1r:=right_x(pp)-x_coord(pp); x2r:=left_x(qq)-right_x(pp);
x3r:=x_coord(qq)-left_x(qq); set_min_max(xr_packet);@/
y1r:=right_y(pp)-y_coord(pp); y2r:=left_y(qq)-right_y(pp);
y3r:=y_coord(qq)-left_y(qq); set_min_max(yr_packet);@/
delx:=x_coord(p)-x_coord(pp); dely:=y_coord(p)-y_coord(pp);@/
tol:=0; uv:=r_packets; xy:=r_packets; three_l:=0; cur_t:=1; cur_tt:=1

@ @<Subdivide for a new level of intersection@>=
stack_dx:=delx; stack_dy:=dely; stack_tol:=tol; stack_uv:=uv; stack_xy:=xy;
bisect_ptr:=bisect_ptr+int_increment;@/
double(cur_t); double(cur_tt);@/
u1l:=stack_1(u_packet(uv)); u3r:=stack_3(u_packet(uv));
u2l:=half(u1l+stack_2(u_packet(uv)));
u2r:=half(u3r+stack_2(u_packet(uv)));
u3l:=half(u2l+u2r); u1r:=u3l;
set_min_max(ul_packet); set_min_max(ur_packet);@/
v1l:=stack_1(v_packet(uv)); v3r:=stack_3(v_packet(uv));
v2l:=half(v1l+stack_2(v_packet(uv)));
v2r:=half(v3r+stack_2(v_packet(uv)));
v3l:=half(v2l+v2r); v1r:=v3l;
set_min_max(vl_packet); set_min_max(vr_packet);@/
x1l:=stack_1(x_packet(xy)); x3r:=stack_3(x_packet(xy));
x2l:=half(x1l+stack_2(x_packet(xy)));
x2r:=half(x3r+stack_2(x_packet(xy)));
x3l:=half(x2l+x2r); x1r:=x3l;
set_min_max(xl_packet); set_min_max(xr_packet);@/
y1l:=stack_1(y_packet(xy)); y3r:=stack_3(y_packet(xy));
y2l:=half(y1l+stack_2(y_packet(xy)));
y2r:=half(y3r+stack_2(y_packet(xy)));
y3l:=half(y2l+y2r); y1r:=y3l;
set_min_max(yl_packet); set_min_max(yr_packet);@/
uv:=l_packets; xy:=l_packets;
double(delx); double(dely);@/
tol:=tol-three_l+tol_step; double(tol); three_l:=three_l+tol_step

@ @<Advance to the next pair |(cur_t,cur_tt)|@>=
not_found: if odd(cur_tt) then
  if odd(cur_t) then @<Descend to the previous level and |goto not_found|@>
  else  begin incr(cur_t);
    delx:=delx+stack_1(u_packet(uv))+stack_2(u_packet(uv))
      +stack_3(u_packet(uv));
    dely:=dely+stack_1(v_packet(uv))+stack_2(v_packet(uv))
      +stack_3(v_packet(uv));
    uv:=uv+int_packets; {switch from |l_packets| to |r_packets|}
    decr(cur_tt); xy:=xy-int_packets; {switch from |r_packets| to |l_packets|}
    delx:=delx+stack_1(x_packet(xy))+stack_2(x_packet(xy))
      +stack_3(x_packet(xy));
    dely:=dely+stack_1(y_packet(xy))+stack_2(y_packet(xy))
      +stack_3(y_packet(xy));
    end
else  begin incr(cur_tt); tol:=tol+three_l;
  delx:=delx-stack_1(x_packet(xy))-stack_2(x_packet(xy))
    -stack_3(x_packet(xy));
  dely:=dely-stack_1(y_packet(xy))-stack_2(y_packet(xy))
    -stack_3(y_packet(xy));
  xy:=xy+int_packets; {switch from |l_packets| to |r_packets|}
  end

@ @<Descend to the previous level...@>=
begin cur_t:=half(cur_t); cur_tt:=half(cur_tt);
if cur_t=0 then return;
bisect_ptr:=bisect_ptr-int_increment; three_l:=three_l-tol_step;
delx:=stack_dx; dely:=stack_dy; tol:=stack_tol; uv:=stack_uv; xy:=stack_xy;@/
goto not_found;
end

@ The |path_intersection| procedure is much simpler.
It invokes |cubic_intersection| in lexicographic order until finding a
pair of cubics that intersect. The final intersection times are placed in
|cur_t| and~|cur_tt|.

@p procedure path_intersection(@!h,@!hh:pointer);
label exit;
var @!p,@!pp:pointer; {link registers that traverse the given paths}
@!n,@!nn:integer; {integer parts of intersection times, minus |unity|}
begin @<Change one-point paths into dead cycles@>;
tol_step:=0;
repeat n:=-unity; p:=h;
  repeat if right_type(p)<>endpoint then
    begin nn:=-unity; pp:=hh;
    repeat if right_type(pp)<>endpoint then
      begin cubic_intersection(p,pp);
      if cur_t>0 then
        begin cur_t:=cur_t+n; cur_tt:=cur_tt+nn; return;
        end;
      end;
    nn:=nn+unity; pp:=link(pp);
    until pp=hh;
    end;
  n:=n+unity; p:=link(p);
  until p=h;
tol_step:=tol_step+3;
until tol_step>3;
cur_t:=-unity; cur_tt:=-unity;
exit:end;

@ @<Change one-point paths...@>=
if right_type(h)=endpoint then
  begin right_x(h):=x_coord(h); left_x(h):=x_coord(h);
  right_y(h):=y_coord(h); left_y(h):=y_coord(h); right_type(h):=explicit;
  end;
if right_type(hh)=endpoint then
  begin right_x(hh):=x_coord(hh); left_x(hh):=x_coord(hh);
  right_y(hh):=y_coord(hh); left_y(hh):=y_coord(hh); right_type(hh):=explicit;
  end;

@* \[27] Online graphic output.
\MF\ displays images on the user's screen by means of a few primitive
operations that are defined below. These operations have deliberately been
kept simple so that they can be implemented without great difficulty on a
wide variety of machines. Since \PASCAL\ has no traditional standards for
graphic output, some system-dependent code needs to be written in order to
support this aspect of \MF; but the necessary routines are usually quite
easy to write.
@^system dependencies@>

In fact, there are exactly four such routines:

\yskip\hang
|init_screen| does whatever initialization is necessary to
support the other operations; it is a boolean function that returns
|false| if graphic output cannot be supported (e.g., if the other three
routines have not been written, or if the user doesn't have the
right kind of terminal).

\yskip\hang
|blank_rectangle| updates a buffer area in memory so that
all pixels in a specified rectangle will be set to the background color.

\yskip\hang
|paint_row| assigns values to specified pixels in a row of
the buffer just mentioned, based on ``transition'' indices explained below.

\yskip\hang
|update_screen| displays the current screen buffer; the
effects of |blank_rectangle| and |paint_row| commands may or may not
become visible until the next |update_screen| operation is performed.
(Thus, |update_screen| is analogous to |update_terminal|.)

\yskip\noindent
The \PASCAL\ code here is a minimum version of |init_screen| and
|update_screen|, usable on \MF\ installations that don't
support screen output. If |init_screen| is changed to return |true|
instead of |false|, the other routines will simply log the fact
that they have been called; they won't really display anything.
The standard test routines for \MF\ use this log information to check
that \MF\ is working properly, but the |wlog| instructions should be
removed from production versions of \MF.

@p function init_screen:boolean;
begin init_screen:=false;
end;
@#
procedure update_screen; {will be called only if |init_screen| returns |true|}
begin @!init wlog_ln('Calling UPDATESCREEN');@+tini {for testing only}
end;

@ The user's screen is assumed to be a rectangular area, |screen_width|
pixels wide and |screen_depth| pixels deep. The pixel in the upper left
corner is said to be in column~0 of row~0; the pixel in the lower right
corner is said to be in column |screen_width-1| of row |screen_depth-1|.
Notice that row numbers increase from top to bottom, contrary to \MF's
other coordinates.

Each pixel is assumed to have two states, referred to in this documentation
as |black| and |white|. The background color is called |white| and the
other color is called |black|; but any two distinct pixel values
can actually be used. For example, the author developed \MF\ on a
system for which |white| was black and |black| was bright green.

@d white=0 {background pixels}
@d black=1 {visible pixels}

@<Types...@>=
@!screen_row=0..screen_depth; {a row number on the screen}
@!screen_col=0..screen_width; {a column number on the screen}
@!trans_spec=array[screen_col] of screen_col; {a transition spec, see below}
@!pixel_color=white..black; {specifies one of the two pixel values}

@ We'll illustrate the |blank_rectangle| and |paint_row| operations by
pretending to declare a screen buffer called |screen_pixel|. This code
is actually commented out, but it does specify the intended effects.

@<Glob...@>=
@{@+@!screen_pixel:array[screen_row,screen_col] of pixel_color@t; @>@}

@ The |blank_rectangle| routine simply whitens all pixels that lie in
columns |left_col| through |right_col-1|, inclusive, of rows
|top_row| through |bot_row-1|, inclusive, given four parameters that satisfy
the relations
$$\hbox{|0<=left_col<=right_col<=screen_width|,\quad
  |0<=top_row<=bot_row<=screen_depth|.}$$
If |left_col=right_col| or |top_row=bot_row|, nothing happens.

The commented-out code in the following procedure is for illustrative
purposes only.
@^system dependencies@>

@p procedure blank_rectangle(@!left_col,@!right_col:screen_col;
  @!top_row,@!bot_row:screen_row);
var @!r:screen_row;
@!c:screen_col;
begin @{@+for r:=top_row to bot_row-1 do
  for c:=left_col to right_col-1 do
    screen_pixel[r,c]:=white;@+@}@/
@!init wlog_cr; {this will be done only after |init_screen=true|}
wlog_ln('Calling BLANKRECTANGLE(',left_col:1,',',
  right_col:1,',',top_row:1,',',bot_row:1,')');@+tini
end;

@ The real work of screen display is done by |paint_row|. But it's not
hard work, because the operation affects only
one of the screen rows, and it affects only a contiguous set of columns
in that row. There are four parameters: |r|~(the row),
|b|~(the initial color),
|a|~(the array of transition specifications),
and |n|~(the number of transitions). The elements of~|a| will satisfy
$$0\L a[0]<a[1]<\cdots<a[n]\L |screen_width|;$$
the value of |r| will satisfy |0<=r<screen_depth|; and |n| will be positive.

The general idea is to paint blocks of pixels in alternate colors;
the precise details are best conveyed by means of a \PASCAL\
program (see the commented-out code below).
@^system dependencies@>

@p procedure paint_row(@!r:screen_row;@!b:pixel_color;var @!a:trans_spec;
  @!n:screen_col);
var @!k:screen_col; {an index into |a|}
@!c:screen_col; {an index into |screen_pixel|}
begin @{@+k:=0; c:=a[0];
repeat incr(k);
  repeat screen_pixel[r,c]:=b; incr(c);
  until c=a[k];
  b:=black-b; {$|black|\swap|white|$}
  until k=n;@+@}@/
@!init wlog('Calling PAINTROW(',r:1,',',b:1,';');
  {this is done only after |init_screen=true|}
for k:=0 to n do
  begin wlog(a[k]:1); if k<>n then wlog(',');
  end;
wlog_ln(')');@+tini
end;

@ The remainder of \MF's screen routines are system-independent calls
on the four primitives just defined.

First we have a global boolean variable that tells if |init_screen|
has been called, and another one that tells if |init_screen| has
given a |true| response.

@<Glob...@>=
@!screen_started:boolean; {have the screen primitives been initialized?}
@!screen_OK:boolean; {is it legitimate to call |blank_rectangle|,
  |paint_row|, and |update_screen|?}

@ @d start_screen==begin if not screen_started then
    begin screen_OK:=init_screen; screen_started:=true;
    end;
  end

@<Set init...@>=
screen_started:=false; screen_OK:=false;

@ \MF\ provides the user with 16 ``window'' areas on the screen, in each
of which it is possible to produce independent displays.

It should be noted that \MF's windows aren't really independent
``clickable'' entities in the sense of multi-window graphic workstations;
\MF\ simply maps them into subsets of a single screen image that is
controlled by |init_screen|, |blank_rectangle|, |paint_row|, and
|update_screen| as described above. Implementations of \MF\ on a
multi-window workstation probably therefore make use of only two
windows in the other sense: one for the terminal output and another
for the screen with \MF's 16 areas. Henceforth we shall
use the term window only in \MF's sense.

@<Types...@>=
@!window_number=0..15;

@ A user doesn't have to use any of the 16 windows. But when a window is
``opened,'' it is allocated to a specific rectangular portion of the screen
and to a specific rectangle with respect to \MF's coordinates. The relevant
data is stored in global arrays |window_open|, |left_col|, |right_col|,
|top_row|, |bot_row|, |m_window|, and |n_window|.

The |window_open| array is boolean, and its significance is obvious. The
|left_col|, \dots, |bot_row| arrays contain screen coordinates that
can be used to blank the entire window with |blank_rectangle|. And the
other two arrays just mentioned handle the conversion between
actual coordinates and screen coordinates: \MF's pixel in column~$m$
of row~$n$ will appear in screen column |m_window+m| and in screen row
|n_window-n|, provided that these lie inside the boundaries of the window.

Another array |window_time| holds the number of times this window has
been updated.

@<Glob...@>=
@!window_open:array[window_number] of boolean;
  {has this window been opened?}
@!left_col:array[window_number] of screen_col;
  {leftmost column position on screen}
@!right_col:array[window_number] of screen_col;
  {rightmost column position, plus~1}
@!top_row:array[window_number] of screen_row;
  {topmost row position on screen}
@!bot_row:array[window_number] of screen_row;
  {bottommost row position, plus~1}
@!m_window:array[window_number] of integer;
  {offset between user and screen columns}
@!n_window:array[window_number] of integer;
  {offset between user and screen rows}
@!window_time:array[window_number] of integer;
  {it has been updated this often}

@ @<Set init...@>=
for k:=0 to 15 do
  begin window_open[k]:=false; window_time[k]:=0;
  end;

@ Opening a window isn't like opening a file, because you can open it
as often as you like, and you never have to close it again. The idea is
simply to define special points on the current screen display.

Overlapping window specifications may cause complex effects that can
be understood only by scrutinizing \MF's display algorithms; thus it
has been left undefined in the \MF\ user manual, although the behavior
@:METAFONTbook}{\sl The {\logos METAFONT\/}book@>
is in fact predictable.

Here is a subroutine that implements the command `\&{openwindow}~|k|
\&{from}~$(\\{r0},\\{c0})$ \&{to}~$(\\{r1},\\{c1})$ \&{at}~$(x,y)$'.

@p procedure open_a_window(@!k:window_number;@!r0,@!c0,@!r1,@!c1:scaled;
    @!x,@!y:scaled);
var @!m,@!n:integer; {pixel coordinates}
begin @<Adjust the coordinates |(r0,c0)| and |(r1,c1)| so that
  they lie in the proper range@>;
window_open[k]:=true; incr(window_time[k]);@/
left_col[k]:=c0; right_col[k]:=c1; top_row[k]:=r0; bot_row[k]:=r1;@/
@<Compute the offsets between screen coordinates and actual coordinates@>;
start_screen;
if screen_OK then
  begin blank_rectangle(c0,c1,r0,r1); update_screen;
  end;
end;

@ A window whose coordinates don't fit the existing screen size will be
truncated until they do.

@<Adjust the coordinates |(r0,c0)| and |(r1,c1)|...@>=
if r0<0 then r0:=0@+else r0:=round_unscaled(r0);
r1:=round_unscaled(r1);
if r1>screen_depth then r1:=screen_depth;
if r1<r0 then
  if r0>screen_depth then r0:=r1@+else r1:=r0;
if c0<0 then c0:=0@+else c0:=round_unscaled(c0);
c1:=round_unscaled(c1);
if c1>screen_width then c1:=screen_width;
if c1<c0 then
  if c0>screen_width then c0:=c1@+else c1:=c0

@ Three sets of coordinates are rampant, and they must be kept straight!
(i)~\MF's main coordinates refer to the edges between pixels. (ii)~\MF's
pixel coordinates (within edge structures) say that the pixel bounded by
$(m,n)$, $(m,n+1)$, $(m+1,n)$, and~$(m+1,n+1)$ is in pixel row number~$n$
and pixel column number~$m$. (iii)~Screen coordinates, on the other hand,
have rows numbered in increasing order from top to bottom, as mentioned
above.
@^coordinates, explained@>

The program here first computes integers $m$ and $n$ such that
pixel column~$m$ of pixel row~$n$ will be at the upper left corner
of the window. Hence pixel column |m-c0| of pixel row |n+r0|
will be at the upper left corner of the screen.

@<Compute the offsets between screen coordinates and actual coordinates@>=
m:=round_unscaled(x); n:=round_unscaled(y)-1;@/
m_window[k]:=c0-m; n_window[k]:=r0+n

@ Now here comes \MF's most complicated operation related to window
display: Given the number~|k| of an open window, the pixels of positive
weight in |cur_edges| will be shown as |black| in the window; all other
pixels will be shown as |white|.

@p procedure disp_edges(@!k:window_number);
label done,found;
var @!p,@!q:pointer; {for list manipulation}
@!already_there:boolean; {is a previous incarnation in the window?}
@!r:integer; {row number}
@<Other local variables for |disp_edges|@>@;
begin if screen_OK then
 if left_col[k]<right_col[k] then if top_row[k]<bot_row[k] then
  begin already_there:=false;
  if last_window(cur_edges)=k then
   if last_window_time(cur_edges)=window_time[k] then
    already_there:=true;
  if not already_there then
    blank_rectangle(left_col[k],right_col[k],top_row[k],bot_row[k]);
  @<Initialize for the display computations@>;
  p:=link(cur_edges); r:=n_window[k]-(n_min(cur_edges)-zero_field);
  while (p<>cur_edges)and(r>=top_row[k]) do
    begin if r<bot_row[k] then
      @<Display the pixels of edge row |p| in screen row |r|@>;
    p:=link(p); decr(r);
    end;
  update_screen;
  incr(window_time[k]);
  last_window(cur_edges):=k; last_window_time(cur_edges):=window_time[k];
  end;
end;

@ Since it takes some work to display a row, we try to avoid recomputation
whenever we can.

@<Display the pixels of edge row |p| in screen row |r|@>=
begin if unsorted(p)>void then sort_edges(p)
else if unsorted(p)=void then if already_there then goto done;
unsorted(p):=void; {this time we'll paint, but maybe not next time}
@<Set up the parameters needed for |paint_row|;
  but |goto done| if no painting is needed after all@>;
paint_row(r,b,row_transition,n);
done: end

@ The transition-specification parameter to |paint_row| is always the same
array.

@<Glob...@>=
@!row_transition:trans_spec; {an array of |black|/|white| transitions}

@ The job remaining is to go through the list |sorted(p)|, unpacking the
|info| fields into |m| and weight, then making |black| the pixels whose
accumulated weight~|w| is positive.

@<Other local variables for |disp_edges|@>=
@!n:screen_col; {the highest active index in |row_transition|}
@!w,@!ww:integer; {old and new accumulated weights}
@!b:pixel_color; {status of first pixel in the row transitions}
@!m,@!mm:integer; {old and new screen column positions}
@!d:integer; {edge-and-weight without |min_halfword| compensation}
@!m_adjustment:integer; {conversion between edge and screen coordinates}
@!right_edge:integer; {largest edge-and-weight that could affect the window}
@!min_col:screen_col; {the smallest screen column number in the window}

@ Some precomputed constants make the display calculations faster.

@<Initialize for the display computations@>=
m_adjustment:=m_window[k]-m_offset(cur_edges);@/
right_edge:=8*(right_col[k]-m_adjustment);@/
min_col:=left_col[k]

@ @<Set up the parameters needed for |paint_row|...@>=
n:=0; ww:=0; m:=-1; w:=0;
q:=sorted(p); row_transition[0]:=min_col;
loop@+  begin if q=sentinel then d:=right_edge
  else d:=ho(info(q));
  mm:=(d div 8)+m_adjustment;
  if mm<>m then
    begin @<Record a possible transition in column |m|@>;
    m:=mm; w:=ww;
    end;
  if d>=right_edge then goto found;
  ww:=ww+(d mod 8)-zero_w;
  q:=link(q);
  end;
found:@<Wind up the |paint_row| parameter calculation by inserting the
  final transition; |goto done| if no painting is needed@>;

@ Now |m| is a screen column |<right_col[k]|.

@<Record a possible transition in column |m|@>=
if w<=0 then
  begin if ww>0 then if m>min_col then
    begin if n=0 then
      if already_there then
        begin b:=white; incr(n);
        end
      else b:=black
    else incr(n);
    row_transition[n]:=m;
    end;
  end
else if ww<=0 then if m>min_col then
  begin if n=0 then b:=black;
  incr(n); row_transition[n]:=m;
  end

@ If the entire row is |white| in the window area, we can omit painting it
when |already_there| is false, since it has already been blanked out in
that case.

When the following code is invoked, |row_transition[n]| will be
strictly less than |right_col[k]|.

@<Wind up the |paint_row|...@>=
if already_there or(ww>0) then
  begin if n=0 then
    if ww>0 then b:=black
    else b:=white;
  incr(n); row_transition[n]:=right_col[k];
  end
else if n=0 then goto done

@* \[28] Dynamic linear equations.
\MF\ users define variables implicitly by stating equations that should be
satisfied; the computer is supposed to be smart enough to solve those equations.
And indeed, the computer tries valiantly to do so, by distinguishing five
different types of numeric values:

\smallskip\hang
|type(p)=known| is the nice case, when |value(p)| is the |scaled| value
of the variable whose address is~|p|.

\smallskip\hang
|type(p)=dependent| means that |value(p)| is not present, but |dep_list(p)|
points to a {\sl dependency list\/} that expresses the value of variable~|p|
as a |scaled| number plus a sum of independent variables with |fraction|
coefficients.

\smallskip\hang
|type(p)=independent| means that |value(p)=64s+m|, where |s>0| is a ``serial
number'' reflecting the time this variable was first used in an equation;
also |0<=m<64|, and each dependent variable
that refers to this one is actually referring to the future value of
this variable times~$2^m$. (Usually |m=0|, but higher degrees of
scaling are sometimes needed to keep the coefficients in dependency lists
from getting too large. The value of~|m| will always be even.)

\smallskip\hang
|type(p)=numeric_type| means that variable |p| hasn't appeared in an
equation before, but it has been explicitly declared to be numeric.

\smallskip\hang
|type(p)=undefined| means that variable |p| hasn't appeared before.

\smallskip\noindent
We have actually discussed these five types in the reverse order of their
history during a computation: Once |known|, a variable never again
becomes |dependent|; once |dependent|, it almost never again becomes
|independent|; once |independent|, it never again becomes |numeric_type|;
and once |numeric_type|, it never again becomes |undefined| (except
of course when the user specifically decides to scrap the old value
and start again). A backward step may, however, take place: Sometimes
a |dependent| variable becomes |independent| again, when one of the
independent variables it depends on is reverting to |undefined|.

@d s_scale=64 {the serial numbers are multiplied by this factor}
@d new_indep(#)== {create a new independent variable}
  begin if serial_no>el_gordo-s_scale then
      overflow("independent variables",serial_no div s_scale);
@:METAFONT capacity exceeded independent variables}{\quad independent variables@>
  type(#):=independent; serial_no:=serial_no+s_scale;
  value(#):=serial_no;
  end

@<Glob...@>=
@!serial_no:integer; {the most recent serial number, times |s_scale|}

@ @<Make variable |q+s| newly independent@>=new_indep(q+s)

@ But how are dependency lists represented? It's simple: The linear combination
$\alpha_1v_1+\cdots+\alpha_kv_k+\beta$ appears in |k+1| value nodes. If
|q=dep_list(p)| points to this list, and if |k>0|, then |value(q)=
@t$\alpha_1$@>| (which is a |fraction|); |info(q)| points to the location
of $v_1$; and |link(p)| points to the dependency list
$\alpha_2v_2+\cdots+\alpha_kv_k+\beta$. On the other hand if |k=0|,
then |value(q)=@t$\beta$@>| (which is |scaled|) and |info(q)=null|.
The independent variables $v_1$, \dots,~$v_k$ have been sorted so that
they appear in decreasing order of their |value| fields (i.e., of
their serial numbers). \ (It is convenient to use decreasing order,
since |value(null)=0|. If the independent variables were not sorted by
serial number but by some other criterion, such as their location in |mem|,
the equation-solving mechanism would be too system-dependent, because
the ordering can affect the computed results.)

The |link| field in the node that contains the constant term $\beta$ is
called the {\sl final link\/} of the dependency list. \MF\ maintains
a doubly-linked master list of all dependency lists, in terms of a permanently
allocated node
in |mem| called |dep_head|. If there are no dependencies, we have
|link(dep_head)=dep_head| and |prev_dep(dep_head)=dep_head|;
otherwise |link(dep_head)| points to the first dependent variable, say~|p|,
and |prev_dep(p)=dep_head|. We have |type(p)=dependent|, and |dep_list(p)|
points to its dependency list. If the final link of that dependency list
occurs in location~|q|, then |link(q)| points to the next dependent
variable (say~|r|); and we have |prev_dep(r)=q|, etc.

@d dep_list(#)==link(value_loc(#))
  {half of the |value| field in a |dependent| variable}
@d prev_dep(#)==info(value_loc(#))
  {the other half; makes a doubly linked list}
@d dep_node_size=2 {the number of words per dependency node}

@<Initialize table entries...@>= serial_no:=0;
link(dep_head):=dep_head; prev_dep(dep_head):=dep_head;
info(dep_head):=null; dep_list(dep_head):=null;

@ Actually the description above contains a little white lie. There's
another kind of variable called |proto_dependent|, which is
just like a |dependent| one except that the $\alpha$ coefficients
in its dependency list are |scaled| instead of being fractions.
Proto-dependency lists are mixed with dependency lists in the
nodes reachable from |dep_head|.

@ Here is a procedure that prints a dependency list in symbolic form.
The second parameter should be either |dependent| or |proto_dependent|,
to indicate the scaling of the coefficients.

@<Declare subroutines for printing expressions@>=
procedure print_dependency(@!p:pointer;@!t:small_number);
label exit;
var @!v:integer; {a coefficient}
@!pp,@!q:pointer; {for list manipulation}
begin pp:=p;
loop@+  begin v:=abs(value(p)); q:=info(p);
  if q=null then {the constant term}
    begin if (v<>0)or(p=pp) then
      begin if value(p)>0 then if p<>pp then print_char("+");
      print_scaled(value(p));
      end;
    return;
    end;
  @<Print the coefficient, unless it's $\pm1.0$@>;
  if type(q)<>independent then confusion("dep");
@:this can't happen dep}{\quad dep@>
  print_variable_name(q); v:=value(q) mod s_scale;
  while v>0 do
    begin print("*4"); v:=v-2;
    end;
  p:=link(p);
  end;
exit:end;

@ @<Print the coefficient, unless it's $\pm1.0$@>=
if value(p)<0 then print_char("-")
else if p<>pp then print_char("+");
if t=dependent then v:=round_fraction(v);
if v<>unity then print_scaled(v)

@ The maximum absolute value of a coefficient in a given dependency list
is returned by the following simple function.

@p function max_coef(@!p:pointer):fraction;
var @!x:fraction; {the maximum so far}
begin x:=0;
while info(p)<>null do
  begin if abs(value(p))>x then x:=abs(value(p));
  p:=link(p);
  end;
max_coef:=x;
end;

@ One of the main operations needed on dependency lists is to add a multiple
of one list to the other; we call this |p_plus_fq|, where |p| and~|q| point
to dependency lists and |f| is a fraction.

If the coefficient of any independent variable becomes |coef_bound| or
more, in absolute value, this procedure changes the type of that variable
to `|independent_needing_fix|', and sets the global variable |fix_needed|
to~|true|. The value of $|coef_bound|=\mu$ is chosen so that
$\mu^2+\mu<8$; this means that the numbers we deal with won't
get too large. (Instead of the ``optimum'' $\mu=(\sqrt{33}-1)/2\approx
2.3723$, the safer value 7/3 is taken as the threshold.)

The changes mentioned in the preceding paragraph are actually done only if
the global variable |watch_coefs| is |true|. But it usually is; in fact,
it is |false| only when \MF\ is making a dependency list that will soon
be equated to zero.

Several procedures that act on dependency lists, including |p_plus_fq|,
set the global variable |dep_final| to the final (constant term) node of
the dependency list that they produce.

@d coef_bound==@'4525252525 {|fraction| approximation to 7/3}
@d independent_needing_fix=0

@<Glob...@>=
@!fix_needed:boolean; {does at least one |independent| variable need scaling?}
@!watch_coefs:boolean; {should we scale coefficients that exceed |coef_bound|?}
@!dep_final:pointer; {location of the constant term and final link}

@ @<Set init...@>=
fix_needed:=false; watch_coefs:=true;

@ The |p_plus_fq| procedure has a fourth parameter, |t|, that should be
set to |proto_dependent| if |p| is a proto-dependency list. In this
case |f| will be |scaled|, not a |fraction|. Similarly, the fifth parameter~|tt|
should be |proto_dependent| if |q| is a proto-dependency list.

List |q| is unchanged by the operation; but list |p| is totally destroyed.

The final link of the dependency list or proto-dependency list returned
by |p_plus_fq| is the same as the original final link of~|p|. Indeed, the
constant term of the result will be located in the same |mem| location
as the original constant term of~|p|.

Coefficients of the result are assumed to be zero if they are less than
a certain threshold. This compensates for inevitable rounding errors,
and tends to make more variables `|known|'. The threshold is approximately
$10^{-5}$ in the case of normal dependency lists, $10^{-4}$ for
proto-dependencies.

@d fraction_threshold=2685 {a |fraction| coefficient less than this is zeroed}
@d half_fraction_threshold=1342 {half of |fraction_threshold|}
@d scaled_threshold=8 {a |scaled| coefficient less than this is zeroed}
@d half_scaled_threshold=4 {half of |scaled_threshold|}

@<Declare basic dependency-list subroutines@>=
function p_plus_fq(@!p:pointer;@!f:integer;@!q:pointer;
  @!t,@!tt:small_number):pointer;
label done;
var @!pp,@!qq:pointer; {|info(p)| and |info(q)|, respectively}
@!r,@!s:pointer; {for list manipulation}
@!threshold:integer; {defines a neighborhood of zero}
@!v:integer; {temporary register}
begin if t=dependent then threshold:=fraction_threshold
else threshold:=scaled_threshold;
r:=temp_head; pp:=info(p); qq:=info(q);
loop@+  if pp=qq then
    if pp=null then goto done
    else @<Contribute a term from |p|, plus |f| times the
      corresponding term from |q|@>
  else if value(pp)<value(qq) then
    @<Contribute a term from |q|, multiplied by~|f|@>
  else  begin link(r):=p; r:=p; p:=link(p); pp:=info(p);
    end;
done: if t=dependent then
  value(p):=slow_add(value(p),take_fraction(value(q),f))
else  value(p):=slow_add(value(p),take_scaled(value(q),f));
link(r):=p; dep_final:=p; p_plus_fq:=link(temp_head);
end;

@ @<Contribute a term from |p|, plus |f|...@>=
begin if tt=dependent then v:=value(p)+take_fraction(f,value(q))
else v:=value(p)+take_scaled(f,value(q));
value(p):=v; s:=p; p:=link(p);
if abs(v)<threshold then free_node(s,dep_node_size)
else  begin if abs(v)>=coef_bound then if watch_coefs then
    begin type(qq):=independent_needing_fix; fix_needed:=true;
    end;
  link(r):=s; r:=s;
  end;
pp:=info(p); q:=link(q); qq:=info(q);
end

@ @<Contribute a term from |q|, multiplied by~|f|@>=
begin if tt=dependent then v:=take_fraction(f,value(q))
else v:=take_scaled(f,value(q));
if abs(v)>half(threshold) then
  begin s:=get_node(dep_node_size); info(s):=qq; value(s):=v;
  if abs(v)>=coef_bound then if watch_coefs then
    begin type(qq):=independent_needing_fix; fix_needed:=true;
    end;
  link(r):=s; r:=s;
  end;
q:=link(q); qq:=info(q);
end

@ It is convenient to have another subroutine for the special case
of |p_plus_fq| when |f=1.0|. In this routine lists |p| and |q| are
both of the same type~|t| (either |dependent| or |proto_dependent|).

@p function p_plus_q(@!p:pointer;@!q:pointer;@!t:small_number):pointer;
label done;
var @!pp,@!qq:pointer; {|info(p)| and |info(q)|, respectively}
@!r,@!s:pointer; {for list manipulation}
@!threshold:integer; {defines a neighborhood of zero}
@!v:integer; {temporary register}
begin if t=dependent then threshold:=fraction_threshold
else threshold:=scaled_threshold;
r:=temp_head; pp:=info(p); qq:=info(q);
loop@+  if pp=qq then
    if pp=null then goto done
    else @<Contribute a term from |p|, plus the
      corresponding term from |q|@>
  else if value(pp)<value(qq) then
    begin s:=get_node(dep_node_size); info(s):=qq; value(s):=value(q);
    q:=link(q); qq:=info(q); link(r):=s; r:=s;
    end
  else  begin link(r):=p; r:=p; p:=link(p); pp:=info(p);
    end;
done: value(p):=slow_add(value(p),value(q));
link(r):=p; dep_final:=p; p_plus_q:=link(temp_head);
end;

@ @<Contribute a term from |p|, plus the...@>=
begin v:=value(p)+value(q);
value(p):=v; s:=p; p:=link(p); pp:=info(p);
if abs(v)<threshold then free_node(s,dep_node_size)
else  begin if abs(v)>=coef_bound then if watch_coefs then
    begin type(qq):=independent_needing_fix; fix_needed:=true;
    end;
  link(r):=s; r:=s;
  end;
q:=link(q); qq:=info(q);
end

@ A somewhat simpler routine will multiply a dependency list
by a given constant~|v|. The constant is either a |fraction| less than
|fraction_one|, or it is |scaled|. In the latter case we might be forced to
convert a dependency list to a proto-dependency list.
Parameters |t0| and |t1| are the list types before and after;
they should agree unless |t0=dependent| and |t1=proto_dependent|
and |v_is_scaled=true|.

@p function p_times_v(@!p:pointer;@!v:integer;
  @!t0,@!t1:small_number;@!v_is_scaled:boolean):pointer;
var @!r,@!s:pointer; {for list manipulation}
@!w:integer; {tentative coefficient}
@!threshold:integer;
@!scaling_down:boolean;
begin if t0<>t1 then scaling_down:=true@+else scaling_down:=not v_is_scaled;
if t1=dependent then threshold:=half_fraction_threshold
else threshold:=half_scaled_threshold;
r:=temp_head;
while info(p)<>null do
  begin if scaling_down then w:=take_fraction(v,value(p))
  else w:=take_scaled(v,value(p));
  if abs(w)<=threshold then
    begin s:=link(p); free_node(p,dep_node_size); p:=s;
    end
  else  begin if abs(w)>=coef_bound then
      begin fix_needed:=true; type(info(p)):=independent_needing_fix;
      end;
    link(r):=p; r:=p; value(p):=w; p:=link(p);
    end;
  end;
link(r):=p;
if v_is_scaled then value(p):=take_scaled(value(p),v)
else value(p):=take_fraction(value(p),v);
p_times_v:=link(temp_head);
end;

@ Similarly, we sometimes need to divide a dependency list
by a given |scaled| constant.

@<Declare basic dependency-list subroutines@>=
function p_over_v(@!p:pointer;@!v:scaled;
  @!t0,@!t1:small_number):pointer;
var @!r,@!s:pointer; {for list manipulation}
@!w:integer; {tentative coefficient}
@!threshold:integer;
@!scaling_down:boolean;
begin if t0<>t1 then scaling_down:=true@+else scaling_down:=false;
if t1=dependent then threshold:=half_fraction_threshold
else threshold:=half_scaled_threshold;
r:=temp_head;
while info(p)<>null do
  begin if scaling_down then
    if abs(v)<@'2000000 then w:=make_scaled(value(p),v*@'10000)
    else w:=make_scaled(round_fraction(value(p)),v)
  else w:=make_scaled(value(p),v);
  if abs(w)<=threshold then
    begin s:=link(p); free_node(p,dep_node_size); p:=s;
    end
  else  begin if abs(w)>=coef_bound then
      begin fix_needed:=true; type(info(p)):=independent_needing_fix;
      end;
    link(r):=p; r:=p; value(p):=w; p:=link(p);
    end;
  end;
link(r):=p; value(p):=make_scaled(value(p),v);
p_over_v:=link(temp_head);
end;

@ Here's another utility routine for dependency lists. When an independent
variable becomes dependent, we want to remove it from all existing
dependencies. The |p_with_x_becoming_q| function computes the
dependency list of~|p| after variable~|x| has been replaced by~|q|.

This procedure has basically the same calling conventions as |p_plus_fq|:
List~|q| is unchanged; list~|p| is destroyed; the constant node and the
final link are inherited from~|p|; and the fourth parameter tells whether
or not |p| is |proto_dependent|. However, the global variable |dep_final|
is not altered if |x| does not occur in list~|p|.

@p function p_with_x_becoming_q(@!p,@!x,@!q:pointer;@!t:small_number):pointer;
var @!r,@!s:pointer; {for list manipulation}
@!v:integer; {coefficient of |x|}
@!sx:integer; {serial number of |x|}
begin s:=p; r:=temp_head; sx:=value(x);
while value(info(s))>sx do
  begin r:=s; s:=link(s);
  end;
if info(s)<>x then p_with_x_becoming_q:=p
else  begin link(temp_head):=p; link(r):=link(s); v:=value(s);
  free_node(s,dep_node_size);
  p_with_x_becoming_q:=p_plus_fq(link(temp_head),v,q,t,dependent);
  end;
end;

@ Here's a simple procedure that reports an error when a variable
has just received a known value that's out of the required range.

@<Declare basic dependency-list subroutines@>=
procedure val_too_big(@!x:scaled);
begin if internal[warning_check]>0 then
  begin print_err("Value is too large ("); print_scaled(x); print_char(")");
@.Value is too large@>
  help4("The equation I just processed has given some variable")@/
    ("a value of 4096 or more. Continue and I'll try to cope")@/
    ("with that big value; but it might be dangerous.")@/
    ("(Set warningcheck:=0 to suppress this message.)");
  error;
  end;
end;

@ When a dependent variable becomes known, the following routine
removes its dependency list. Here |p| points to the variable, and
|q| points to the dependency list (which is one node long).

@<Declare basic dependency-list subroutines@>=
procedure make_known(@!p,@!q:pointer);
var @!t:dependent..proto_dependent; {the previous type}
begin prev_dep(link(q)):=prev_dep(p);
link(prev_dep(p)):=link(q); t:=type(p);
type(p):=known; value(p):=value(q); free_node(q,dep_node_size);
if abs(value(p))>=fraction_one then val_too_big(value(p));
if internal[tracing_equations]>0 then if interesting(p) then
  begin begin_diagnostic; print_nl("#### ");
@:]]]\#\#\#\#_}{\.{\#\#\#\#}@>
  print_variable_name(p); print_char("="); print_scaled(value(p));
  end_diagnostic(false);
  end;
if cur_exp=p then if cur_type=t then
  begin cur_type:=known; cur_exp:=value(p);
  free_node(p,value_node_size);
  end;
end;

@ The |fix_dependencies| routine is called into action when |fix_needed|
has been triggered. The program keeps a list~|s| of independent variables
whose coefficients must be divided by~4.

In unusual cases, this fixup process might reduce one or more coefficients
to zero, so that a variable will become known more or less by default.

@<Declare basic dependency-list subroutines@>=
procedure fix_dependencies;
label done;
var @!p,@!q,@!r,@!s,@!t:pointer; {list manipulation registers}
@!x:pointer; {an independent variable}
begin r:=link(dep_head); s:=null;
while r<>dep_head do
  begin t:=r;
  @<Run through the dependency list for variable |t|, fixing
    all nodes, and ending with final link~|q|@>;
  r:=link(q);
  if q=dep_list(t) then make_known(t,q);
  end;
while s<>null do
  begin p:=link(s); x:=info(s); free_avail(s); s:=p;
  type(x):=independent; value(x):=value(x)+2;
  end;
fix_needed:=false;
end;

@ @d independent_being_fixed=1 {this variable already appears in |s|}

@<Run through the dependency list for variable |t|...@>=
r:=value_loc(t); {|link(r)=dep_list(t)|}
loop@+  begin q:=link(r); x:=info(q);
  if x=null then goto done;
  if type(x)<=independent_being_fixed then
    begin if type(x)<independent_being_fixed then
      begin p:=get_avail; link(p):=s; s:=p;
      info(s):=x; type(x):=independent_being_fixed;
      end;
    value(q):=value(q) div 4;
    if value(q)=0 then
      begin link(r):=link(q); free_node(q,dep_node_size); q:=r;
      end;
    end;
  r:=q;
  end;
done:

@ The |new_dep| routine installs a dependency list~|p| into the value node~|q|,
linking it into the list of all known dependencies. We assume that
|dep_final| points to the final node of list~|p|.

@p procedure new_dep(@!q,@!p:pointer);
var @!r:pointer; {what used to be the first dependency}
begin dep_list(q):=p; prev_dep(q):=dep_head;
r:=link(dep_head); link(dep_final):=r; prev_dep(r):=dep_final;
link(dep_head):=q;
end;

@ Here is one of the ways a dependency list gets started.
The |const_dependency| routine produces a list that has nothing but
a constant term.

@p function const_dependency(@!v:scaled):pointer;
begin dep_final:=get_node(dep_node_size);
value(dep_final):=v; info(dep_final):=null;
const_dependency:=dep_final;
end;

@ And here's a more interesting way to start a dependency list from scratch:
The parameter to |single_dependency| is the location of an
independent variable~|x|, and the result is the simple dependency list
`|x+0|'.

In the unlikely event that the given independent variable has been doubled so
often that we can't refer to it with a nonzero coefficient,
|single_dependency| returns the simple list `0'.  This case can be
recognized by testing that the returned list pointer is equal to
|dep_final|.

@p function single_dependency(@!p:pointer):pointer;
var @!q:pointer; {the new dependency list}
@!m:integer; {the number of doublings}
begin m:=value(p) mod s_scale;
if m>28 then single_dependency:=const_dependency(0)
else  begin q:=get_node(dep_node_size);
  value(q):=two_to_the[28-m]; info(q):=p;@/
  link(q):=const_dependency(0); single_dependency:=q;
  end;
end;

@ We sometimes need to make an exact copy of a dependency list.

@p function copy_dep_list(@!p:pointer):pointer;
label done;
var @!q:pointer; {the new dependency list}
begin q:=get_node(dep_node_size); dep_final:=q;
loop@+  begin info(dep_final):=info(p); value(dep_final):=value(p);
  if info(dep_final)=null then goto done;
  link(dep_final):=get_node(dep_node_size);
  dep_final:=link(dep_final); p:=link(p);
  end;
done:copy_dep_list:=q;
end;

@ But how do variables normally become known? Ah, now we get to the heart of the
equation-solving mechanism. The |linear_eq| procedure is given a |dependent|
or |proto_dependent| list,~|p|, in which at least one independent variable
appears. It equates this list to zero, by choosing an independent variable
with the largest coefficient and making it dependent on the others. The
newly dependent variable is eliminated from all current dependencies,
thereby possibly making other dependent variables known.

The given list |p| is, of course, totally destroyed by all this processing.

@p procedure linear_eq(@!p:pointer;@!t:small_number);
var @!q,@!r,@!s:pointer; {for link manipulation}
@!x:pointer; {the variable that loses its independence}
@!n:integer; {the number of times |x| had been halved}
@!v:integer; {the coefficient of |x| in list |p|}
@!prev_r:pointer; {lags one step behind |r|}
@!final_node:pointer; {the constant term of the new dependency list}
@!w:integer; {a tentative coefficient}
begin @<Find a node |q| in list |p| whose coefficient |v| is largest@>;
x:=info(q); n:=value(x) mod s_scale;@/
@<Divide list |p| by |-v|, removing node |q|@>;
if internal[tracing_equations]>0 then @<Display the new dependency@>;
@<Simplify all existing dependencies by substituting for |x|@>;
@<Change variable |x| from |independent| to |dependent| or |known|@>;
if fix_needed then fix_dependencies;
end;

@ @<Find a node |q| in list |p| whose coefficient |v| is largest@>=
q:=p; r:=link(p); v:=value(q);
while info(r)<>null do
  begin if abs(value(r))>abs(v) then
    begin q:=r; v:=value(r);
    end;
  r:=link(r);
  end

@ Here we want to change the coefficients from |scaled| to |fraction|,
except in the constant term. In the common case of a trivial equation
like `\.{x=3.14}', we will have |v=-fraction_one|, |q=p|, and |t=dependent|.

@<Divide list |p| by |-v|, removing node |q|@>=
s:=temp_head; link(s):=p; r:=p;
repeat if r=q then
  begin link(s):=link(r); free_node(r,dep_node_size);
  end
else  begin w:=make_fraction(value(r),v);
  if abs(w)<=half_fraction_threshold then
    begin link(s):=link(r); free_node(r,dep_node_size);
    end
  else  begin value(r):=-w; s:=r;
    end;
  end;
r:=link(s);
until info(r)=null;
if t=proto_dependent then value(r):=-make_scaled(value(r),v)
else if v<>-fraction_one then value(r):=-make_fraction(value(r),v);
final_node:=r; p:=link(temp_head)

@ @<Display the new dependency@>=
if interesting(x) then
  begin begin_diagnostic; print_nl("## "); print_variable_name(x);
@:]]]\#\#_}{\.{\#\#}@>
  w:=n;
  while w>0 do
    begin print("*4"); w:=w-2;
    end;
  print_char("="); print_dependency(p,dependent); end_diagnostic(false);
  end

@ @<Simplify all existing dependencies by substituting for |x|@>=
prev_r:=dep_head; r:=link(dep_head);
while r<>dep_head do
  begin s:=dep_list(r); q:=p_with_x_becoming_q(s,x,p,type(r));
  if info(q)=null then make_known(r,q)
  else  begin dep_list(r):=q;
    repeat q:=link(q);
    until info(q)=null;
    prev_r:=q;
    end;
  r:=link(prev_r);
  end

@ @<Change variable |x| from |independent| to |dependent| or |known|@>=
if n>0 then @<Divide list |p| by $2^n$@>;
if info(p)=null then
  begin type(x):=known;
  value(x):=value(p);
  if abs(value(x))>=fraction_one then val_too_big(value(x));
  free_node(p,dep_node_size);
  if cur_exp=x then if cur_type=independent then
    begin cur_exp:=value(x); cur_type:=known;
    free_node(x,value_node_size);
    end;
  end
else  begin type(x):=dependent; dep_final:=final_node; new_dep(x,p);
  if cur_exp=x then if cur_type=independent then cur_type:=dependent;
  end

@ @<Divide list |p| by $2^n$@>=
begin s:=temp_head; link(temp_head):=p; r:=p;
repeat if n>30 then w:=0
else w:=value(r) div two_to_the[n];
if (abs(w)<=half_fraction_threshold)and(info(r)<>null) then
  begin link(s):=link(r);
  free_node(r,dep_node_size);
  end
else  begin value(r):=w; s:=r;
  end;
r:=link(s);
until info(s)=null;
p:=link(temp_head);
end

@ The |check_mem| procedure, which is used only when \MF\ is being
debugged, makes sure that the current dependency lists are well formed.

@<Check the list of linear dependencies@>=
q:=dep_head; p:=link(q);
while p<>dep_head do
  begin if prev_dep(p)<>q then
    begin print_nl("Bad PREVDEP at "); print_int(p);
@.Bad PREVDEP...@>
    end;
  p:=dep_list(p); r:=inf_val;
  repeat if value(info(p))>=value(r) then
    begin print_nl("Out of order at "); print_int(p);
@.Out of order...@>
    end;
  r:=info(p); q:=p; p:=link(q);
  until r=null;
  end

@* \[29] Dynamic nonlinear equations.
Variables of numeric type are maintained by the general scheme of
independent, dependent, and known values that we have just studied;
and the components of pair and transform variables are handled in the
same way. But \MF\ also has five other types of values: \&{boolean},
\&{string}, \&{pen}, \&{path}, and \&{picture}; what about them?

Equations are allowed between nonlinear quantities, but only in a
simple form. Two variables that haven't yet been assigned values are
either equal to each other, or they're not.

Before a boolean variable has received a value, its type is |unknown_boolean|;
similarly, there are variables whose type is |unknown_string|, |unknown_pen|,
|unknown_path|, and |unknown_picture|. In such cases the value is either
|null| (which means that no other variables are equivalent to this one), or
it points to another variable of the same undefined type. The pointers in the
latter case form a cycle of nodes, which we shall call a ``ring.''
Rings of undefined variables may include capsules, which arise as
intermediate results within expressions or as \&{expr} parameters to macros.

When one member of a ring receives a value, the same value is given to
all the other members. In the case of paths and pictures, this implies
making separate copies of a potentially large data structure; users should
restrain their enthusiasm for such generality, unless they have lots and
lots of memory space.

@ The following procedure is called when a capsule node is being
added to a ring (e.g., when an unknown variable is mentioned in an expression).

@p function new_ring_entry(@!p:pointer):pointer;
var q:pointer; {the new capsule node}
begin q:=get_node(value_node_size); name_type(q):=capsule;
type(q):=type(p);
if value(p)=null then value(q):=p@+else value(q):=value(p);
value(p):=q;
new_ring_entry:=q;
end;

@ Conversely, we might delete a capsule or a variable before it becomes known.
The following procedure simply detaches a quantity from its ring,
without recycling the storage.

@<Declare the recycling subroutines@>=
procedure ring_delete(@!p:pointer);
var @!q:pointer;
begin q:=value(p);
if q<>null then if q<>p then
  begin while value(q)<>p do q:=value(q);
  value(q):=value(p);
  end;
end;

@ Eventually there might be an equation that assigns values to all of the
variables in a ring. The |nonlinear_eq| subroutine does the necessary
propagation of values.

If the parameter |flush_p| is |true|, node |p| itself needn't receive a
value; it will soon be recycled.

@p procedure nonlinear_eq(@!v:integer;@!p:pointer;@!flush_p:boolean);
var @!t:small_number; {the type of ring |p|}
@!q,@!r:pointer; {link manipulation registers}
begin t:=type(p)-unknown_tag; q:=value(p);
if flush_p then type(p):=vacuous@+else p:=q;
repeat r:=value(q); type(q):=t;
case t of
boolean_type: value(q):=v;
string_type: begin value(q):=v; add_str_ref(v);
  end;
pen_type: begin value(q):=v; add_pen_ref(v);
  end;
path_type: value(q):=copy_path(v);
picture_type: value(q):=copy_edges(v);
end; {there ain't no more cases}
q:=r;
until q=p;
end;

@ If two members of rings are equated, and if they have the same type,
the |ring_merge| procedure is called on to make them equivalent.

@p procedure ring_merge(@!p,@!q:pointer);
label exit;
var @!r:pointer; {traverses one list}
begin r:=value(p);
while r<>p do
  begin if r=q then
    begin @<Exclaim about a redundant equation@>;
    return;
    end;
  r:=value(r);
  end;
r:=value(p); value(p):=value(q); value(q):=r;
exit:end;

@ @<Exclaim about a redundant equation@>=
begin print_err("Redundant equation");@/
@.Redundant equation@>
help2("I already knew that this equation was true.")@/
  ("But perhaps no harm has been done; let's continue.");@/
put_get_error;
end

@* \[30] Introduction to the syntactic routines.
Let's pause a moment now and try to look at the Big Picture.
The \MF\ program consists of three main parts: syntactic routines,
semantic routines, and output routines. The chief purpose of the
syntactic routines is to deliver the user's input to the semantic routines,
while parsing expressions and locating operators and operands. The
semantic routines act as an interpreter responding to these operators,
which may be regarded as commands. And the output routines are
periodically called on to produce compact font descriptions that can be
used for typesetting or for making interim proof drawings. We have
discussed the basic data structures and many of the details of semantic
operations, so we are good and ready to plunge into the part of \MF\ that
actually controls the activities.

Our current goal is to come to grips with the |get_next| procedure,
which is the keystone of \MF's input mechanism. Each call of |get_next|
sets the value of three variables |cur_cmd|, |cur_mod|, and |cur_sym|,
representing the next input token.
$$\vbox{\halign{#\hfil\cr
  \hbox{|cur_cmd| denotes a command code from the long list of codes
   given earlier;}\cr
  \hbox{|cur_mod| denotes a modifier of the command code;}\cr
  \hbox{|cur_sym| is the hash address of the symbolic token that was
   just scanned,}\cr
  \hbox{\qquad or zero in the case of a numeric or string
   or capsule token.}\cr}}$$
Underlying this external behavior of |get_next| is all the machinery
necessary to convert from character files to tokens. At a given time we
may be only partially finished with the reading of several files (for
which \&{input} was specified), and partially finished with the expansion
of some user-defined macros and/or some macro parameters, and partially
finished reading some text that the user has inserted online,
and so on. When reading a character file, the characters must be
converted to tokens; comments and blank spaces must
be removed, numeric and string tokens must be evaluated.

To handle these situations, which might all be present simultaneously,
\MF\ uses various stacks that hold information about the incomplete
activities, and there is a finite state control for each level of the
input mechanism. These stacks record the current state of an implicitly
recursive process, but the |get_next| procedure is not recursive.

@<Glob...@>=
@!cur_cmd: eight_bits; {current command set by |get_next|}
@!cur_mod: integer; {operand of current command}
@!cur_sym: halfword; {hash address of current symbol}

@ The |print_cmd_mod| routine prints a symbolic interpretation of a
command code and its modifier.
It consists of a rather tedious sequence of print
commands, and most of it is essentially an inverse to the |primitive|
routine that enters a \MF\ primitive into |hash| and |eqtb|. Therefore almost
all of this procedure appears elsewhere in the program, together with the
corresponding |primitive| calls.

@<Declare the procedure called |print_cmd_mod|@>=
procedure print_cmd_mod(@!c,@!m:integer);
begin case c of
@t\4@>@<Cases of |print_cmd_mod| for symbolic printing of primitives@>@/
othercases print("[unknown command code!]")
endcases;
end;

@ Here is a procedure that displays a given command in braces, in the
user's transcript file.

@d show_cur_cmd_mod==show_cmd_mod(cur_cmd,cur_mod)

@p procedure show_cmd_mod(@!c,@!m:integer);
begin begin_diagnostic; print_nl("{");
print_cmd_mod(c,m); print_char("}");
end_diagnostic(false);
end;

@* \[31] Input stacks and states.
The state of \MF's input mechanism appears in the input stack, whose
entries are records with five fields, called |index|, |start|, |loc|,
|limit|, and |name|. The top element of this stack is maintained in a
global variable for which no subscripting needs to be done; the other
elements of the stack appear in an array. Hence the stack is declared thus:

@<Types...@>=
@!in_state_record = record
  @!index_field: quarterword;
  @!start_field,@!loc_field, @!limit_field, @!name_field: halfword;
  end;

@ @<Glob...@>=
@!input_stack : array[0..stack_size] of in_state_record;
@!input_ptr : 0..stack_size; {first unused location of |input_stack|}
@!max_in_stack: 0..stack_size; {largest value of |input_ptr| when pushing}
@!cur_input : in_state_record; {the ``top'' input state}

@ We've already defined the special variable |@!loc==cur_input.loc_field|
in our discussion of basic input-output routines. The other components of
|cur_input| are defined in the same way:

@d index==cur_input.index_field {reference for buffer information}
@d start==cur_input.start_field {starting position in |buffer|}
@d limit==cur_input.limit_field {end of current line in |buffer|}
@d name==cur_input.name_field {name of the current file}

@ Let's look more closely now at the five control variables
(|index|,~|start|,~|loc|,~|limit|,~|name|),
assuming that \MF\ is reading a line of characters that have been input
from some file or from the user's terminal. There is an array called
|buffer| that acts as a stack of all lines of characters that are
currently being read from files, including all lines on subsidiary
levels of the input stack that are not yet completed. \MF\ will return to
the other lines when it is finished with the present input file.

(Incidentally, on a machine with byte-oriented addressing, it would be
appropriate to combine |buffer| with the |str_pool| array,
letting the buffer entries grow downward from the top of the string pool
and checking that these two tables don't bump into each other.)

The line we are currently working on begins in position |start| of the
buffer; the next character we are about to read is |buffer[loc]|; and
|limit| is the location of the last character present. We always have
|loc<=limit|. For convenience, |buffer[limit]| has been set to |"%"|, so
that the end of a line is easily sensed.

The |name| variable is a string number that designates the name of
the current file, if we are reading a text file. It is 0 if we
are reading from the terminal for normal input, or 1 if we are executing a
\&{readstring} command, or 2 if we are reading a string that was
moved into the buffer by \&{scantokens}.

@ Additional information about the current line is available via the
|index| variable, which counts how many lines of characters are present
in the buffer below the current level. We have |index=0| when reading
from the terminal and prompting the user for each line; then if the user types,
e.g., `\.{input font}', we will have |index=1| while reading
the file \.{font.mf}. However, it does not follow that |index| is the
same as the input stack pointer, since many of the levels on the input
stack may come from token lists.

The global variable |in_open| is equal to the |index|
value of the highest non-token-list level. Thus, the number of partially read
lines in the buffer is |in_open+1|, and we have |in_open=index|
when we are not reading a token list.

If we are not currently reading from the terminal,
we are reading from the file variable |input_file[index]|. We use
the notation |terminal_input| as a convenient abbreviation for |name=0|,
and |cur_file| as an abbreviation for |input_file[index]|.

The global variable |line| contains the line number in the topmost
open file, for use in error messages. If we are not reading from
the terminal, |line_stack[index]| holds the line number for the
enclosing level, so that |line| can be restored when the current
file has been read.

If more information about the input state is needed, it can be
included in small arrays like those shown here. For example,
the current page or segment number in the input file might be
put into a variable |@!page|, maintained for enclosing levels in
`\ignorespaces|@!page_stack:array[1..max_in_open] of integer|\unskip'
by analogy with |line_stack|.
@^system dependencies@>

@d terminal_input==(name=0) {are we reading from the terminal?}
@d cur_file==input_file[index] {the current |alpha_file| variable}

@<Glob...@>=
@!in_open : 0..max_in_open; {the number of lines in the buffer, less one}
@!open_parens : 0..max_in_open; {the number of open text files}
@!input_file : array[1..max_in_open] of alpha_file;
@!line : integer; {current line number in the current source file}
@!line_stack : array[1..max_in_open] of integer;

@ However, all this discussion about input state really applies only to the
case that we are inputting from a file. There is another important case,
namely when we are currently getting input from a token list. In this case
|index>max_in_open|, and the conventions about the other state variables
are different:

\yskip\hang|loc| is a pointer to the current node in the token list, i.e.,
the node that will be read next. If |loc=null|, the token list has been
fully read.

\yskip\hang|start| points to the first node of the token list; this node
may or may not contain a reference count, depending on the type of token
list involved.

\yskip\hang|token_type|, which takes the place of |index| in the
discussion above, is a code number that explains what kind of token list
is being scanned.

\yskip\hang|name| points to the |eqtb| address of the macro
being expanded, if the current token list is a macro not defined by
\&{vardef}. Macros defined by \&{vardef} have |name=null|; their name
can be deduced by looking at their first two parameters.

\yskip\hang|param_start|, which takes the place of |limit|, tells where
the parameters of the current macro or loop text begin in the |param_stack|.

\yskip\noindent The |token_type| can take several values, depending on
where the current token list came from:

\yskip
\indent|forever_text|, if the token list being scanned is the body of
a \&{forever} loop;

\indent|loop_text|, if the token list being scanned is the body of
a \&{for} or \&{forsuffixes} loop;

\indent|parameter|, if a \&{text} or \&{suffix} parameter is being scanned;

\indent|backed_up|, if the token list being scanned has been inserted as
`to be read again';

\indent|inserted|, if the token list being scanned has been inserted as
part of error recovery;

\indent|macro|, if the expansion of a user-defined symbolic token is being
scanned.

\yskip\noindent
The token list begins with a reference count if and only if |token_type=
macro|.
@^reference counts@>

@d token_type==index {type of current token list}
@d token_state==(index>max_in_open) {are we scanning a token list?}
@d file_state==(index<=max_in_open) {are we scanning a file line?}
@d param_start==limit {base of macro parameters in |param_stack|}
@d forever_text=max_in_open+1 {|token_type| code for loop texts}
@d loop_text=max_in_open+2 {|token_type| code for loop texts}
@d parameter=max_in_open+3 {|token_type| code for parameter texts}
@d backed_up=max_in_open+4 {|token_type| code for texts to be reread}
@d inserted=max_in_open+5 {|token_type| code for inserted texts}
@d macro=max_in_open+6 {|token_type| code for macro replacement texts}

@ The |param_stack| is an auxiliary array used to hold pointers to the token
lists for parameters at the current level and subsidiary levels of input.
This stack grows at a different rate from the others.

@<Glob...@>=
@!param_stack:array [0..param_size] of pointer;
  {token list pointers for parameters}
@!param_ptr:0..param_size; {first unused entry in |param_stack|}
@!max_param_stack:integer;
  {largest value of |param_ptr|}

@ Thus, the ``current input state'' can be very complicated indeed; there
can be many levels and each level can arise in a variety of ways. The
|show_context| procedure, which is used by \MF's error-reporting routine to
print out the current input state on all levels down to the most recent
line of characters from an input file, illustrates most of these conventions.
The global variable |file_ptr| contains the lowest level that was
displayed by this procedure.

@<Glob...@>=
@!file_ptr:0..stack_size; {shallowest level shown by |show_context|}

@ The status at each level is indicated by printing two lines, where the first
line indicates what was read so far and the second line shows what remains
to be read. The context is cropped, if necessary, so that the first line
contains at most |half_error_line| characters, and the second contains
at most |error_line|. Non-current input levels whose |token_type| is
`|backed_up|' are shown only if they have not been fully read.

@p procedure show_context; {prints where the scanner is}
label done;
var @!old_setting:0..max_selector; {saved |selector| setting}
@<Local variables for formatting calculations@>@/
begin file_ptr:=input_ptr; input_stack[file_ptr]:=cur_input;
  {store current state}
loop@+begin cur_input:=input_stack[file_ptr]; {enter into the context}
  @<Display the current context@>;
  if file_state then
    if (name>2) or (file_ptr=0) then goto done;
  decr(file_ptr);
  end;
done: cur_input:=input_stack[input_ptr]; {restore original state}
end;

@ @<Display the current context@>=
if (file_ptr=input_ptr) or file_state or
   (token_type<>backed_up) or (loc<>null) then
    {we omit backed-up token lists that have already been read}
  begin tally:=0; {get ready to count characters}
  old_setting:=selector;
  if file_state then
    begin @<Print location of current line@>;
    @<Pseudoprint the line@>;
    end
  else  begin @<Print type of token list@>;
    @<Pseudoprint the token list@>;
    end;
  selector:=old_setting; {stop pseudoprinting}
  @<Print two lines using the tricky pseudoprinted information@>;
  end

@ This routine should be changed, if necessary, to give the best possible
indication of where the current line resides in the input file.
For example, on some systems it is best to print both a page and line number.
@^system dependencies@>

@<Print location of current line@>=
if name<=1 then
  if terminal_input and(file_ptr=0) then print_nl("<*>")
  else print_nl("<insert>")
else if name=2 then print_nl("<scantokens>")
else  begin print_nl("l."); print_int(line);
  end;
print_char(" ")

@ @<Print type of token list@>=
case token_type of
forever_text: print_nl("<forever> ");
loop_text: @<Print the current loop value@>;
parameter: print_nl("<argument> ");
backed_up: if loc=null then print_nl("<recently read> ")
  else print_nl("<to be read again> ");
inserted: print_nl("<inserted text> ");
macro: begin print_ln;
  if name<>null then slow_print(text(name))
  else @<Print the name of a \&{vardef}'d macro@>;
  print("->");
  end;
othercases print_nl("?") {this should never happen}
@.?\relax@>
endcases

@ The parameter that corresponds to a loop text is either a token list
(in the case of \&{forsuffixes}) or a ``capsule'' (in the case of \&{for}).
We'll discuss capsules later; for now, all we need to know is that
the |link| field in a capsule parameter is |void| and that
|print_exp(p,0)| displays the value of capsule~|p| in abbreviated form.

@<Print the current loop value@>=
begin print_nl("<for("); p:=param_stack[param_start];
if p<>null then
  if link(p)=void then print_exp(p,0) {we're in a \&{for} loop}
  else show_token_list(p,null,20,tally);
print(")> ");
end

@ The first two parameters of a macro defined by \&{vardef} will be token
lists representing the macro's prefix and ``at point.'' By putting these
together, we get the macro's full name.

@<Print the name of a \&{vardef}'d macro@>=
begin p:=param_stack[param_start];
if p=null then show_token_list(param_stack[param_start+1],null,20,tally)
else  begin q:=p;
  while link(q)<>null do q:=link(q);
  link(q):=param_stack[param_start+1];
  show_token_list(p,null,20,tally);
  link(q):=null;
  end;
end

@ Now it is necessary to explain a little trick. We don't want to store a long
string that corresponds to a token list, because that string might take up
lots of memory; and we are printing during a time when an error message is
being given, so we dare not do anything that might overflow one of \MF's
tables. So `pseudoprinting' is the answer: We enter a mode of printing
that stores characters into a buffer of length |error_line|, where character
$k+1$ is placed into \hbox{|trick_buf[k mod error_line]|} if
|k<trick_count|, otherwise character |k| is dropped. Initially we set
|tally:=0| and |trick_count:=1000000|; then when we reach the
point where transition from line 1 to line 2 should occur, we
set |first_count:=tally| and |trick_count:=@tmax@>(error_line,
tally+1+error_line-half_error_line)|. At the end of the
pseudoprinting, the values of |first_count|, |tally|, and
|trick_count| give us all the information we need to print the two lines,
and all of the necessary text is in |trick_buf|.

Namely, let |l| be the length of the descriptive information that appears
on the first line. The length of the context information gathered for that
line is |k=first_count|, and the length of the context information
gathered for line~2 is $m=\min(|tally|, |trick_count|)-k$. If |l+k<=h|,
where |h=half_error_line|, we print |trick_buf[0..k-1]| after the
descriptive information on line~1, and set |n:=l+k|; here |n| is the
length of line~1. If $l+k>h$, some cropping is necessary, so we set |n:=h|
and print `\.{...}' followed by
$$\hbox{|trick_buf[(l+k-h+3)..k-1]|,}$$
where subscripts of |trick_buf| are circular modulo |error_line|. The
second line consists of |n|~spaces followed by |trick_buf[k..(k+m-1)]|,
unless |n+m>error_line|; in the latter case, further cropping is done.
This is easier to program than to explain.

@<Local variables for formatting...@>=
@!i:0..buf_size; {index into |buffer|}
@!l:integer; {length of descriptive information on line 1}
@!m:integer; {context information gathered for line 2}
@!n:0..error_line; {length of line 1}
@!p: integer; {starting or ending place in |trick_buf|}
@!q: integer; {temporary index}

@ The following code tells the print routines to gather
the desired information.

@d begin_pseudoprint==
  begin l:=tally; tally:=0; selector:=pseudo;
  trick_count:=1000000;
  end
@d set_trick_count==
  begin first_count:=tally;
  trick_count:=tally+1+error_line-half_error_line;
  if trick_count<error_line then trick_count:=error_line;
  end

@ And the following code uses the information after it has been gathered.

@<Print two lines using the tricky pseudoprinted information@>=
if trick_count=1000000 then set_trick_count;
  {|set_trick_count| must be performed}
if tally<trick_count then m:=tally-first_count
else m:=trick_count-first_count; {context on line 2}
if l+first_count<=half_error_line then
  begin p:=0; n:=l+first_count;
  end
else  begin print("..."); p:=l+first_count-half_error_line+3;
  n:=half_error_line;
  end;
for q:=p to first_count-1 do print_char(trick_buf[q mod error_line]);
print_ln;
for q:=1 to n do print_char(" "); {print |n| spaces to begin line~2}
if m+n<=error_line then p:=first_count+m else p:=first_count+(error_line-n-3);
for q:=first_count to p-1 do print_char(trick_buf[q mod error_line]);
if m+n>error_line then print("...")

@ But the trick is distracting us from our current goal, which is to
understand the input state. So let's concentrate on the data structures that
are being pseudoprinted as we finish up the |show_context| procedure.

@<Pseudoprint the line@>=
begin_pseudoprint;
if limit>0 then for i:=start to limit-1 do
  begin if i=loc then set_trick_count;
  print(buffer[i]);
  end

@ @<Pseudoprint the token list@>=
begin_pseudoprint;
if token_type<>macro then show_token_list(start,loc,100000,0)
else show_macro(start,loc,100000)

@ Here is the missing piece of |show_token_list| that is activated when the
token beginning line~2 is about to be shown:

@<Do magic computation@>=set_trick_count

@* \[32] Maintaining the input stacks.
The following subroutines change the input status in commonly needed ways.

First comes |push_input|, which stores the current state and creates a
new level (having, initially, the same properties as the old).

@d push_input==@t@> {enter a new input level, save the old}
  begin if input_ptr>max_in_stack then
    begin max_in_stack:=input_ptr;
    if input_ptr=stack_size then overflow("input stack size",stack_size);
@:METAFONT capacity exceeded input stack size}{\quad input stack size@>
    end;
  input_stack[input_ptr]:=cur_input; {stack the record}
  incr(input_ptr);
  end

@ And of course what goes up must come down.

@d pop_input==@t@> {leave an input level, re-enter the old}
  begin decr(input_ptr); cur_input:=input_stack[input_ptr];
  end

@ Here is a procedure that starts a new level of token-list input, given
a token list |p| and its type |t|. If |t=macro|, the calling routine should
set |name|, reset~|loc|, and increase the macro's reference count.

@d back_list(#)==begin_token_list(#,backed_up) {backs up a simple token list}

@p procedure begin_token_list(@!p:pointer;@!t:quarterword);
begin push_input; start:=p; token_type:=t;
param_start:=param_ptr; loc:=p;
end;

@ When a token list has been fully scanned, the following computations
should be done as we leave that level of input.
@^inner loop@>

@p procedure end_token_list; {leave a token-list input level}
label done;
var @!p:pointer; {temporary register}
begin if token_type>=backed_up then {token list to be deleted}
  if token_type<=inserted then
    begin flush_token_list(start); goto done;
    end
  else delete_mac_ref(start); {update reference count}
while param_ptr>param_start do {parameters must be flushed}
  begin decr(param_ptr);
  p:=param_stack[param_ptr];
  if p<>null then
    if link(p)=void then {it's an \&{expr} parameter}
      begin recycle_value(p); free_node(p,value_node_size);
      end
    else flush_token_list(p); {it's a \&{suffix} or \&{text} parameter}
  end;
done: pop_input; check_interrupt;
end;

@ The contents of |cur_cmd,cur_mod,cur_sym| are placed into an equivalent
token by the |cur_tok| routine.
@^inner loop@>

@p @t\4@>@<Declare the procedure called |make_exp_copy|@>@;@/
function cur_tok:pointer;
var @!p:pointer; {a new token node}
@!save_type:small_number; {|cur_type| to be restored}
@!save_exp:integer; {|cur_exp| to be restored}
begin if cur_sym=0 then
  if cur_cmd=capsule_token then
    begin save_type:=cur_type; save_exp:=cur_exp;
    make_exp_copy(cur_mod); p:=stash_cur_exp; link(p):=null;
    cur_type:=save_type; cur_exp:=save_exp;
    end
  else  begin p:=get_node(token_node_size);
    value(p):=cur_mod; name_type(p):=token;
    if cur_cmd=numeric_token then type(p):=known
    else type(p):=string_type;
    end
else  begin fast_get_avail(p); info(p):=cur_sym;
  end;
cur_tok:=p;
end;

@ Sometimes \MF\ has read too far and wants to ``unscan'' what it has
seen. The |back_input| procedure takes care of this by putting the token
just scanned back into the input stream, ready to be read again.
If |cur_sym<>0|, the values of |cur_cmd| and |cur_mod| are irrelevant.

@p procedure back_input; {undoes one token of input}
var @!p:pointer; {a token list of length one}
begin p:=cur_tok;
while token_state and(loc=null) do end_token_list; {conserve stack space}
back_list(p);
end;

@ The |back_error| routine is used when we want to restore or replace an
offending token just before issuing an error message.  We disable interrupts
during the call of |back_input| so that the help message won't be lost.

@p procedure back_error; {back up one token and call |error|}
begin OK_to_interrupt:=false; back_input; OK_to_interrupt:=true; error;
end;
@#
procedure ins_error; {back up one inserted token and call |error|}
begin OK_to_interrupt:=false; back_input; token_type:=inserted;
OK_to_interrupt:=true; error;
end;

@ The |begin_file_reading| procedure starts a new level of input for lines
of characters to be read from a file, or as an insertion from the
terminal. It does not take care of opening the file, nor does it set |loc|
or |limit| or |line|.
@^system dependencies@>

@p procedure begin_file_reading;
begin if in_open=max_in_open then overflow("text input levels",max_in_open);
@:METAFONT capacity exceeded text input levels}{\quad text input levels@>
if first=buf_size then overflow("buffer size",buf_size);
@:METAFONT capacity exceeded buffer size}{\quad buffer size@>
incr(in_open); push_input; index:=in_open;
line_stack[index]:=line; start:=first;
name:=0; {|terminal_input| is now |true|}
end;

@ Conversely, the variables must be downdated when such a level of input
is finished:

@p procedure end_file_reading;
begin first:=start; line:=line_stack[index];
if index<>in_open then confusion("endinput");
@:this can't happen endinput}{\quad endinput@>
if name>2 then a_close(cur_file); {forget it}
pop_input; decr(in_open);
end;

@ In order to keep the stack from overflowing during a long sequence of
inserted `\.{show}' commands, the following routine removes completed
error-inserted lines from memory.

@p procedure clear_for_error_prompt;
begin while file_state and terminal_input and@|
  (input_ptr>0)and(loc=limit) do end_file_reading;
print_ln; clear_terminal;
end;

@ To get \MF's whole input mechanism going, we perform the following
actions.

@<Initialize the input routines@>=
begin input_ptr:=0; max_in_stack:=0;
in_open:=0; open_parens:=0; max_buf_stack:=0;
param_ptr:=0; max_param_stack:=0;
first:=1;
start:=1; index:=0; line:=0; name:=0;
force_eof:=false;
if not init_terminal then goto final_end;
limit:=last; first:=last+1; {|init_terminal| has set |loc| and |last|}
end;

@* \[33] Getting the next token.
The heart of \MF's input mechanism is the |get_next| procedure, which
we shall develop in the next few sections of the program. Perhaps we
shouldn't actually call it the ``heart,'' however; it really acts as \MF's
eyes and mouth, reading the source files and gobbling them up. And it also
helps \MF\ to regurgitate stored token lists that are to be processed again.

The main duty of |get_next| is to input one token and to set |cur_cmd|
and |cur_mod| to that token's command code and modifier. Furthermore, if
the input token is a symbolic token, that token's |hash| address
is stored in |cur_sym|; otherwise |cur_sym| is set to zero.

Underlying this simple description is a certain amount of complexity
because of all the cases that need to be handled.
However, the inner loop of |get_next| is reasonably short and fast.

@ Before getting into |get_next|, we need to consider a mechanism by which
\MF\ helps keep errors from propagating too far. Whenever the program goes
into a mode where it keeps calling |get_next| repeatedly until a certain
condition is met, it sets |scanner_status| to some value other than |normal|.
Then if an input file ends, or if an `\&{outer}' symbol appears,
an appropriate error recovery will be possible.

The global variable |warning_info| helps in this error recovery by providing
additional information. For example, |warning_info| might indicate the
name of a macro whose replacement text is being scanned.

@d normal=0 {|scanner_status| at ``quiet times''}
@d skipping=1 {|scanner_status| when false conditional text is being skipped}
@d flushing=2 {|scanner_status| when junk after a statement is being ignored}
@d absorbing=3 {|scanner_status| when a \&{text} parameter is being scanned}
@d var_defining=4 {|scanner_status| when a \&{vardef} is being scanned}
@d op_defining=5 {|scanner_status| when a macro \&{def} is being scanned}
@d loop_defining=6 {|scanner_status| when a \&{for} loop is being scanned}

@<Glob...@>=
@!scanner_status:normal..loop_defining; {are we scanning at high speed?}
@!warning_info:integer; {if so, what else do we need to know,
    in case an error occurs?}

@ @<Initialize the input routines@>=
scanner_status:=normal;

@ The following subroutine
is called when an `\&{outer}' symbolic token has been scanned or
when the end of a file has been reached. These two cases are distinguished
by |cur_sym|, which is zero at the end of a file.

@p function check_outer_validity:boolean;
var @!p:pointer; {points to inserted token list}
begin if scanner_status=normal then check_outer_validity:=true
else  begin deletions_allowed:=false;
  @<Back up an outer symbolic token so that it can be reread@>;
  if scanner_status>skipping then
    @<Tell the user what has run away and try to recover@>
  else  begin print_err("Incomplete if; all text was ignored after line ");
@.Incomplete if...@>
    print_int(warning_info);@/
    help3("A forbidden `outer' token occurred in skipped text.")@/
    ("This kind of error happens when you say `if...' and forget")@/
    ("the matching `fi'. I've inserted a `fi'; this might work.");
    if cur_sym=0 then help_line[2]:=@|
      "The file ended while I was skipping conditional text.";
    cur_sym:=frozen_fi; ins_error;
    end;
  deletions_allowed:=true; check_outer_validity:=false;
  end;
end;

@ @<Back up an outer symbolic token so that it can be reread@>=
if cur_sym<>0 then
  begin p:=get_avail; info(p):=cur_sym;
  back_list(p); {prepare to read the symbolic token again}
  end

@ @<Tell the user what has run away...@>=
begin runaway; {print the definition-so-far}
if cur_sym=0 then print_err("File ended")
@.File ended while scanning...@>
else  begin print_err("Forbidden token found");
@.Forbidden token found...@>
  end;
print(" while scanning ");
help4("I suspect you have forgotten an `enddef',")@/
("causing me to read past where you wanted me to stop.")@/
("I'll try to recover; but if the error is serious,")@/
("you'd better type `E' or `X' now and fix your file.");@/
case scanner_status of
@t\4@>@<Complete the error message,
  and set |cur_sym| to a token that might help recover from the error@>@;
end; {there are no other cases}
ins_error;
end

@ As we consider various kinds of errors, it is also appropriate to
change the first line of the help message just given; |help_line[3]|
points to the string that might be changed.

@<Complete the error message,...@>=
flushing: begin print("to the end of the statement");
  help_line[3]:="A previous error seems to have propagated,";
  cur_sym:=frozen_semicolon;
  end;
absorbing: begin print("a text argument");
  help_line[3]:="It seems that a right delimiter was left out,";
  if warning_info=0 then cur_sym:=frozen_end_group
  else  begin cur_sym:=frozen_right_delimiter;
    equiv(frozen_right_delimiter):=warning_info;
    end;
  end;
var_defining, op_defining: begin print("the definition of ");
  if scanner_status=op_defining then slow_print(text(warning_info))
  else print_variable_name(warning_info);
  cur_sym:=frozen_end_def;
  end;
loop_defining: begin print("the text of a "); slow_print(text(warning_info));
  print(" loop");
  help_line[3]:="I suspect you have forgotten an `endfor',";
  cur_sym:=frozen_end_for;
  end;

@ The |runaway| procedure displays the first part of the text that occurred
when \MF\ began its special |scanner_status|, if that text has been saved.

@<Declare the procedure called |runaway|@>=
procedure runaway;
begin if scanner_status>flushing then
  begin print_nl("Runaway ");
  case scanner_status of
  absorbing: print("text?");
  var_defining,op_defining: print("definition?");
  loop_defining: print("loop?");
  end; {there are no other cases}
  print_ln; show_token_list(link(hold_head),null,error_line-10,0);
  end;
end;

@ We need to mention a procedure that may be called by |get_next|.

@p procedure@?firm_up_the_line; forward;

@ And now we're ready to take the plunge into |get_next| itself.

@d switch=25 {a label in |get_next|}
@d start_numeric_token=85 {another}
@d start_decimal_token=86 {and another}
@d fin_numeric_token=87
  {and still another, although |goto| is considered harmful}

@p procedure get_next; {sets |cur_cmd|, |cur_mod|, |cur_sym| to next token}
@^inner loop@>
label restart, {go here to get the next input token}
  exit, {go here when the next input token has been got}
  found, {go here when the end of a symbolic token has been found}
  switch, {go here to branch on the class of an input character}
  start_numeric_token,start_decimal_token,fin_numeric_token,done;
    {go here at crucial stages when scanning a number}
var @!k:0..buf_size; {an index into |buffer|}
@!c:ASCII_code; {the current character in the buffer}
@!class:ASCII_code; {its class number}
@!n,@!f:integer; {registers for decimal-to-binary conversion}
begin restart: cur_sym:=0;
if file_state then
@<Input from external file; |goto restart| if no input found,
  or |return| if a non-symbolic token is found@>
else @<Input from token list; |goto restart| if end of list or
  if a parameter needs to be expanded,
  or |return| if a non-symbolic token is found@>;
@<Finish getting the symbolic token in |cur_sym|;
  |goto restart| if it is illegal@>;
exit:end;

@ When a symbolic token is declared to be `\&{outer}', its command code
is increased by |outer_tag|.
@^inner loop@>

@<Finish getting the symbolic token in |cur_sym|...@>=
cur_cmd:=eq_type(cur_sym); cur_mod:=equiv(cur_sym);
if cur_cmd>=outer_tag then
  if check_outer_validity then cur_cmd:=cur_cmd-outer_tag
  else goto restart

@ A percent sign appears in |buffer[limit]|; this makes it unnecessary
to have a special test for end-of-line.
@^inner loop@>

@<Input from external file;...@>=
begin switch: c:=buffer[loc]; incr(loc); class:=char_class[c];
case class of
digit_class: goto start_numeric_token;
period_class: begin class:=char_class[buffer[loc]];
  if class>period_class then goto switch
  else if class<period_class then {|class=digit_class|}
    begin n:=0; goto start_decimal_token;
    end;
@:. }{\..\ token@>
  end;
space_class: goto switch;
percent_class: begin @<Move to next line of file,
    or |goto restart| if there is no next line@>;
  check_interrupt;
  goto switch;
  end;
string_class: @<Get a string token and |return|@>;
isolated_classes: begin k:=loc-1; goto found;
  end;
invalid_class: @<Decry the invalid character and |goto restart|@>;
othercases do_nothing {letters, etc.}
endcases;@/
k:=loc-1;
while char_class[buffer[loc]]=class do incr(loc);
goto found;
start_numeric_token:@<Get the integer part |n| of a numeric token;
  set |f:=0| and |goto fin_numeric_token| if there is no decimal point@>;
start_decimal_token:@<Get the fraction part |f| of a numeric token@>;
fin_numeric_token:@<Pack the numeric and fraction parts of a numeric token
  and |return|@>;
found: cur_sym:=id_lookup(k,loc-k);
end

@ We go to |restart| instead of to |switch|, because we might enter
|token_state| after the error has been dealt with
(cf.\ |clear_for_error_prompt|).

@<Decry the invalid...@>=
begin print_err("Text line contains an invalid character");
@.Text line contains...@>
help2("A funny symbol that I can't read has just been input.")@/
("Continue, and I'll forget that it ever happened.");@/
deletions_allowed:=false; error; deletions_allowed:=true;
goto restart;
end

@ @<Get a string token and |return|@>=
begin if buffer[loc]="""" then cur_mod:=""
else  begin k:=loc; buffer[limit+1]:="""";
  repeat incr(loc);
  until buffer[loc]="""";
  if loc>limit then @<Decry the missing string delimiter and |goto restart|@>;
  if (loc=k+1) and (length(buffer[k])=1) then cur_mod:=buffer[k]
  else  begin str_room(loc-k);
    repeat append_char(buffer[k]); incr(k);
    until k=loc;
    cur_mod:=make_string;
    end;
  end;
incr(loc); cur_cmd:=string_token; return;
end

@ We go to |restart| after this error message, not to |switch|,
because the |clear_for_error_prompt| routine might have reinstated
|token_state| after |error| has finished.

@<Decry the missing string delimiter and |goto restart|@>=
begin loc:=limit; {the next character to be read on this line will be |"%"|}
print_err("Incomplete string token has been flushed");
@.Incomplete string token...@>
help3("Strings should finish on the same line as they began.")@/
  ("I've deleted the partial string; you might want to")@/
  ("insert another by typing, e.g., `I""new string""'.");@/
deletions_allowed:=false; error; deletions_allowed:=true; goto restart;
end

@ @<Get the integer part |n| of a numeric token...@>=
n:=c-"0";
while char_class[buffer[loc]]=digit_class do
  begin if n<4096 then n:=10*n+buffer[loc]-"0";
  incr(loc);
  end;
if buffer[loc]="." then if char_class[buffer[loc+1]]=digit_class then goto done;
f:=0; goto fin_numeric_token;
done: incr(loc)

@ @<Get the fraction part |f| of a numeric token@>=
k:=0;
repeat if k<17 then {digits for |k>=17| cannot affect the result}
  begin dig[k]:=buffer[loc]-"0"; incr(k);
  end;
incr(loc);
until char_class[buffer[loc]]<>digit_class;
f:=round_decimals(k);
if f=unity then
  begin incr(n); f:=0;
  end

@ @<Pack the numeric and fraction parts of a numeric token and |return|@>=
if n<4096 then cur_mod:=n*unity+f
else  begin print_err("Enormous number has been reduced");
@.Enormous number...@>
  help2("I can't handle numbers bigger than about 4095.99998;")@/
  ("so I've changed your constant to that maximum amount.");@/
  deletions_allowed:=false; error; deletions_allowed:=true;
  cur_mod:=@'1777777777;
  end;
cur_cmd:=numeric_token; return

@ Let's consider now what happens when |get_next| is looking at a token list.
@^inner loop@>

@<Input from token list;...@>=
if loc>=hi_mem_min then {one-word token}
  begin cur_sym:=info(loc); loc:=link(loc); {move to next}
  if cur_sym>=expr_base then
    if cur_sym>=suffix_base then
      @<Insert a suffix or text parameter and |goto restart|@>
    else  begin cur_cmd:=capsule_token;
      cur_mod:=param_stack[param_start+cur_sym-(expr_base)];
      cur_sym:=0; return;
      end;
  end
else if loc>null then
  @<Get a stored numeric or string or capsule token and |return|@>
else  begin {we are done with this token list}
  end_token_list; goto restart; {resume previous level}
  end

@ @<Insert a suffix or text parameter...@>=
begin if cur_sym>=text_base then cur_sym:=cur_sym-param_size;
  {|param_size=text_base-suffix_base|}
begin_token_list(param_stack[param_start+cur_sym-(suffix_base)],parameter);
goto restart;
end

@ @<Get a stored numeric or string or capsule token...@>=
begin if name_type(loc)=token then
  begin cur_mod:=value(loc);
  if type(loc)=known then cur_cmd:=numeric_token
  else  begin cur_cmd:=string_token; add_str_ref(cur_mod);
    end;
  end
else  begin cur_mod:=loc; cur_cmd:=capsule_token;
  end;
loc:=link(loc); return;
end

@ All of the easy branches of |get_next| have now been taken care of.
There is one more branch.

@<Move to next line of file, or |goto restart|...@>=
if name>2 then @<Read next line of file into |buffer|, or
  |goto restart| if the file has ended@>
else  begin if input_ptr>0 then
     {text was inserted during error recovery or by \&{scantokens}}
    begin end_file_reading; goto restart; {resume previous level}
    end;
  if selector<log_only then open_log_file;
  if interaction>nonstop_mode then
    begin if limit=start then {previous line was empty}
      print_nl("(Please type a command or say `end')");
@.Please type...@>
    print_ln; first:=start;
    prompt_input("*"); {input on-line into |buffer|}
@.*\relax@>
    limit:=last; buffer[limit]:="%";
    first:=limit+1; loc:=start;
    end
  else fatal_error("*** (job aborted, no legal end found)");
@.job aborted@>
    {nonstop mode, which is intended for overnight batch processing,
    never waits for on-line input}
  end

@ The global variable |force_eof| is normally |false|; it is set |true|
by an \&{endinput} command.

@<Glob...@>=
@!force_eof:boolean; {should the next \&{input} be aborted early?}

@ @<Read next line of file into |buffer|, or
  |goto restart| if the file has ended@>=
begin incr(line); first:=start;
if not force_eof then
  begin if input_ln(cur_file,true) then {not end of file}
    firm_up_the_line {this sets |limit|}
  else force_eof:=true;
  end;
if force_eof then
  begin print_char(")"); decr(open_parens);
  update_terminal; {show user that file has been read}
  force_eof:=false;
  end_file_reading; {resume previous level}
  if check_outer_validity then goto restart@+else goto restart;
  end;
buffer[limit]:="%"; first:=limit+1; loc:=start; {ready to read}
end

@ If the user has set the |pausing| parameter to some positive value,
and if nonstop mode has not been selected, each line of input is displayed
on the terminal and the transcript file, followed by `\.{=>}'.
\MF\ waits for a response. If the response is null (i.e., if nothing is
typed except perhaps a few blank spaces), the original
line is accepted as it stands; otherwise the line typed is
used instead of the line in the file.

@p procedure firm_up_the_line;
var @!k:0..buf_size; {an index into |buffer|}
begin limit:=last;
if internal[pausing]>0 then if interaction>nonstop_mode then
  begin wake_up_terminal; print_ln;
  if start<limit then for k:=start to limit-1 do print(buffer[k]);
  first:=limit; prompt_input("=>"); {wait for user response}
@.=>@>
  if last>first then
    begin for k:=first to last-1 do {move line down in buffer}
      buffer[k+start-first]:=buffer[k];
    limit:=start+last-first;
    end;
  end;
end;

@* \[34] Scanning macro definitions.
\MF\ has a variety of ways to tuck tokens away into token lists for later
use: Macros can be defined with \&{def}, \&{vardef}, \&{primarydef}, etc.;
repeatable code can be defined with \&{for}, \&{forever}, \&{forsuffixes}.
All such operations are handled by the routines in this part of the program.

The modifier part of each command code is zero for the ``ending delimiters''
like \&{enddef} and \&{endfor}.

@d start_def=1 {command modifier for \&{def}}
@d var_def=2 {command modifier for \&{vardef}}
@d end_def=0 {command modifier for \&{enddef}}
@d start_forever=1 {command modifier for \&{forever}}
@d end_for=0 {command modifier for \&{endfor}}

@<Put each...@>=
primitive("def",macro_def,start_def);@/
@!@:def_}{\&{def} primitive@>
primitive("vardef",macro_def,var_def);@/
@!@:var_def_}{\&{vardef} primitive@>
primitive("primarydef",macro_def,secondary_primary_macro);@/
@!@:primary_def_}{\&{primarydef} primitive@>
primitive("secondarydef",macro_def,tertiary_secondary_macro);@/
@!@:secondary_def_}{\&{secondarydef} primitive@>
primitive("tertiarydef",macro_def,expression_tertiary_macro);@/
@!@:tertiary_def_}{\&{tertiarydef} primitive@>
primitive("enddef",macro_def,end_def); eqtb[frozen_end_def]:=eqtb[cur_sym];@/
@!@:end_def_}{\&{enddef} primitive@>
@#
primitive("for",iteration,expr_base);@/
@!@:for_}{\&{for} primitive@>
primitive("forsuffixes",iteration,suffix_base);@/
@!@:for_suffixes_}{\&{forsuffixes} primitive@>
primitive("forever",iteration,start_forever);@/
@!@:forever_}{\&{forever} primitive@>
primitive("endfor",iteration,end_for); eqtb[frozen_end_for]:=eqtb[cur_sym];@/
@!@:end_for_}{\&{endfor} primitive@>

@ @<Cases of |print_cmd...@>=
macro_def:if m<=var_def then
    if m=start_def then print("def")
    else if m<start_def then print("enddef")
    else print("vardef")
  else if m=secondary_primary_macro then print("primarydef")
  else if m=tertiary_secondary_macro then print("secondarydef")
  else print("tertiarydef");
iteration: if m<=start_forever then
    if m=start_forever then print("forever")@+else print("endfor")
  else if m=expr_base then print("for")@+else print("forsuffixes");

@ Different macro-absorbing operations have different syntaxes, but they
also have a lot in common. There is a list of special symbols that are to
be replaced by parameter tokens; there is a special command code that
ends the definition; the quotation conventions are identical.  Therefore
it makes sense to have most of the work done by a single subroutine. That
subroutine is called |scan_toks|.

The first parameter to |scan_toks| is the command code that will
terminate scanning (either |macro_def| or |iteration|).

The second parameter, |subst_list|, points to a (possibly empty) list
of two-word nodes whose |info| and |value| fields specify symbol tokens
before and after replacement. The list will be returned to free storage
by |scan_toks|.

The third parameter is simply appended to the token list that is built.
And the final parameter tells how many of the special operations
\.{\#\AT!}, \.{\AT!}, and \.{\AT!\#} are to be replaced by suffix parameters.
When such parameters are present, they are called \.{(SUFFIX0)},
\.{(SUFFIX1)}, and \.{(SUFFIX2)}.

@p function scan_toks(@!terminator:command_code;
  @!subst_list,@!tail_end:pointer;@!suffix_count:small_number):pointer;
label done,found;
var @!p:pointer; {tail of the token list being built}
@!q:pointer; {temporary for link management}
@!balance:integer; {left delimiters minus right delimiters}
begin p:=hold_head; balance:=1; link(hold_head):=null;
loop@+  begin get_next;
  if cur_sym>0 then
    begin @<Substitute for |cur_sym|, if it's on the |subst_list|@>;
    if cur_cmd=terminator then
      @<Adjust the balance; |goto done| if it's zero@>
    else if cur_cmd=macro_special then
      @<Handle quoted symbols, \.{\#\AT!}, \.{\AT!}, or \.{\AT!\#}@>;
    end;
  link(p):=cur_tok; p:=link(p);
  end;
done: link(p):=tail_end; flush_node_list(subst_list);
scan_toks:=link(hold_head);
end;

@ @<Substitute for |cur_sym|...@>=
begin q:=subst_list;
while q<>null do
  begin if info(q)=cur_sym then
    begin cur_sym:=value(q); cur_cmd:=relax; goto found;
    end;
  q:=link(q);
  end;
found:end

@ @<Adjust the balance; |goto done| if it's zero@>=
if cur_mod>0 then incr(balance)
else  begin decr(balance);
  if balance=0 then goto done;
  end

@ Four commands are intended to be used only within macro texts: \&{quote},
\.{\#\AT!}, \.{\AT!}, and \.{\AT!\#}. They are variants of a single command
code called |macro_special|.

@d quote=0 {|macro_special| modifier for \&{quote}}
@d macro_prefix=1 {|macro_special| modifier for \.{\#\AT!}}
@d macro_at=2 {|macro_special| modifier for \.{\AT!}}
@d macro_suffix=3 {|macro_special| modifier for \.{\AT!\#}}

@<Put each...@>=
primitive("quote",macro_special,quote);@/
@!@:quote_}{\&{quote} primitive@>
primitive("#@@",macro_special,macro_prefix);@/
@!@:]]]\#\AT!_}{\.{\#\AT!} primitive@>
primitive("@@",macro_special,macro_at);@/
@!@:]]]\AT!_}{\.{\AT!} primitive@>
primitive("@@#",macro_special,macro_suffix);@/
@!@:]]]\AT!\#_}{\.{\AT!\#} primitive@>

@ @<Cases of |print_cmd...@>=
macro_special: case m of
  macro_prefix: print("#@@");
  macro_at: print_char("@@");
  macro_suffix: print("@@#");
  othercases print("quote")
  endcases;

@ @<Handle quoted...@>=
begin if cur_mod=quote then get_next
else if cur_mod<=suffix_count then cur_sym:=suffix_base-1+cur_mod;
end

@ Here is a routine that's used whenever a token will be redefined. If
the user's token is unredefinable, the `|frozen_inaccessible|' token is
substituted; the latter is redefinable but essentially impossible to use,
hence \MF's tables won't get fouled up.

@p procedure get_symbol; {sets |cur_sym| to a safe symbol}
label restart;
begin restart: get_next;
if (cur_sym=0)or(cur_sym>frozen_inaccessible) then
  begin print_err("Missing symbolic token inserted");
@.Missing symbolic token...@>
  help3("Sorry: You can't redefine a number, string, or expr.")@/
    ("I've inserted an inaccessible symbol so that your")@/
    ("definition will be completed without mixing me up too badly.");
  if cur_sym>0 then
    help_line[2]:="Sorry: You can't redefine my error-recovery tokens."
  else if cur_cmd=string_token then delete_str_ref(cur_mod);
  cur_sym:=frozen_inaccessible; ins_error; goto restart;
  end;
end;

@ Before we actually redefine a symbolic token, we need to clear away its
former value, if it was a variable. The following stronger version of
|get_symbol| does that.

@p procedure get_clear_symbol;
begin get_symbol; clear_symbol(cur_sym,false);
end;

@ Here's another little subroutine; it checks that an equals sign
or assignment sign comes along at the proper place in a macro definition.

@p procedure check_equals;
begin if cur_cmd<>equals then if cur_cmd<>assignment then
  begin missing_err("=");@/
@.Missing `='@>
  help5("The next thing in this `def' should have been `=',")@/
    ("because I've already looked at the definition heading.")@/
    ("But don't worry; I'll pretend that an equals sign")@/
    ("was present. Everything from here to `enddef'")@/
    ("will be the replacement text of this macro.");
  back_error;
  end;
end;

@ A \&{primarydef}, \&{secondarydef}, or \&{tertiarydef} is rather easily
handled now that we have |scan_toks|.  In this case there are
two parameters, which will be \.{EXPR0} and \.{EXPR1} (i.e.,
|expr_base| and |expr_base+1|).

@p procedure make_op_def;
var @!m:command_code; {the type of definition}
@!p,@!q,@!r:pointer; {for list manipulation}
begin m:=cur_mod;@/
get_symbol; q:=get_node(token_node_size);
info(q):=cur_sym; value(q):=expr_base;@/
get_clear_symbol; warning_info:=cur_sym;@/
get_symbol; p:=get_node(token_node_size);
info(p):=cur_sym; value(p):=expr_base+1; link(p):=q;@/
get_next; check_equals;@/
scanner_status:=op_defining; q:=get_avail; ref_count(q):=null;
r:=get_avail; link(q):=r; info(r):=general_macro;
link(r):=scan_toks(macro_def,p,null,0);
scanner_status:=normal; eq_type(warning_info):=m;
equiv(warning_info):=q; get_x_next;
end;

@ Parameters to macros are introduced by the keywords \&{expr},
\&{suffix}, \&{text}, \&{primary}, \&{secondary}, and \&{tertiary}.

@<Put each...@>=
primitive("expr",param_type,expr_base);@/
@!@:expr_}{\&{expr} primitive@>
primitive("suffix",param_type,suffix_base);@/
@!@:suffix_}{\&{suffix} primitive@>
primitive("text",param_type,text_base);@/
@!@:text_}{\&{text} primitive@>
primitive("primary",param_type,primary_macro);@/
@!@:primary_}{\&{primary} primitive@>
primitive("secondary",param_type,secondary_macro);@/
@!@:secondary_}{\&{secondary} primitive@>
primitive("tertiary",param_type,tertiary_macro);@/
@!@:tertiary_}{\&{tertiary} primitive@>

@ @<Cases of |print_cmd...@>=
param_type:if m>=expr_base then
    if m=expr_base then print("expr")
    else if m=suffix_base then print("suffix")
    else print("text")
  else if m<secondary_macro then print("primary")
  else if m=secondary_macro then print("secondary")
  else print("tertiary");

@ Let's turn next to the more complex processing associated with \&{def}
and \&{vardef}. When the following procedure is called, |cur_mod|
should be either |start_def| or |var_def|.

@p @t\4@>@<Declare the procedure called |check_delimiter|@>@;
@t\4@>@<Declare the function called |scan_declared_variable|@>@;
procedure scan_def;
var @!m:start_def..var_def; {the type of definition}
@!n:0..3; {the number of special suffix parameters}
@!k:0..param_size; {the total number of parameters}
@!c:general_macro..text_macro; {the kind of macro we're defining}
@!r:pointer; {parameter-substitution list}
@!q:pointer; {tail of the macro token list}
@!p:pointer; {temporary storage}
@!base:halfword; {|expr_base|, |suffix_base|, or |text_base|}
@!l_delim,@!r_delim:pointer; {matching delimiters}
begin m:=cur_mod; c:=general_macro; link(hold_head):=null;@/
q:=get_avail; ref_count(q):=null; r:=null;@/
@<Scan the token or variable to be defined;
  set |n|, |scanner_status|, and |warning_info|@>;
k:=n;
if cur_cmd=left_delimiter then
  @<Absorb delimited parameters, putting them into lists |q| and |r|@>;
if cur_cmd=param_type then
  @<Absorb undelimited parameters, putting them into list |r|@>;
check_equals;
p:=get_avail; info(p):=c; link(q):=p;
@<Attach the replacement text to the tail of node |p|@>;
scanner_status:=normal; get_x_next;
end;

@ We don't put `|frozen_end_group|' into the replacement text of
a \&{vardef}, because the user may want to redefine `\.{endgroup}'.

@<Attach the replacement text to the tail of node |p|@>=
if m=start_def then link(p):=scan_toks(macro_def,r,null,n)
else  begin q:=get_avail; info(q):=bg_loc; link(p):=q;
  p:=get_avail; info(p):=eg_loc;
  link(q):=scan_toks(macro_def,r,p,n);
  end;
if warning_info=bad_vardef then flush_token_list(value(bad_vardef))

@ @<Glob...@>=
@!bg_loc,@!eg_loc:1..hash_end;
  {hash addresses of `\.{begingroup}' and `\.{endgroup}'}

@ @<Scan the token or variable to be defined;...@>=
if m=start_def then
  begin get_clear_symbol; warning_info:=cur_sym; get_next;
  scanner_status:=op_defining; n:=0;
  eq_type(warning_info):=defined_macro; equiv(warning_info):=q;
  end
else  begin p:=scan_declared_variable;
  flush_variable(equiv(info(p)),link(p),true);
  warning_info:=find_variable(p); flush_list(p);
  if warning_info=null then @<Change to `\.{a bad variable}'@>;
  scanner_status:=var_defining; n:=2;
  if cur_cmd=macro_special then if cur_mod=macro_suffix then {\.{\AT!\#}}
    begin n:=3; get_next;
    end;
  type(warning_info):=unsuffixed_macro-2+n; value(warning_info):=q;
  end {|suffixed_macro=unsuffixed_macro+1|}

@ @<Change to `\.{a bad variable}'@>=
begin print_err("This variable already starts with a macro");
@.This variable already...@>
help2("After `vardef a' you can't say `vardef a.b'.")@/
  ("So I'll have to discard this definition.");
error; warning_info:=bad_vardef;
end

@ @<Initialize table entries...@>=
name_type(bad_vardef):=root; link(bad_vardef):=frozen_bad_vardef;
equiv(frozen_bad_vardef):=bad_vardef; eq_type(frozen_bad_vardef):=tag_token;

@ @<Absorb delimited parameters, putting them into lists |q| and |r|@>=
repeat l_delim:=cur_sym; r_delim:=cur_mod; get_next;
if (cur_cmd=param_type)and(cur_mod>=expr_base) then base:=cur_mod
else  begin print_err("Missing parameter type; `expr' will be assumed");
@.Missing parameter type@>
  help1("You should've had `expr' or `suffix' or `text' here.");
  back_error; base:=expr_base;
  end;
@<Absorb parameter tokens for type |base|@>;
check_delimiter(l_delim,r_delim);
get_next;
until cur_cmd<>left_delimiter

@ @<Absorb parameter tokens for type |base|@>=
repeat link(q):=get_avail; q:=link(q); info(q):=base+k;@/
get_symbol; p:=get_node(token_node_size); value(p):=base+k; info(p):=cur_sym;
if k=param_size then overflow("parameter stack size",param_size);
@:METAFONT capacity exceeded parameter stack size}{\quad parameter stack size@>
incr(k); link(p):=r; r:=p; get_next;
until cur_cmd<>comma

@ @<Absorb undelimited parameters, putting them into list |r|@>=
begin p:=get_node(token_node_size);
if cur_mod<expr_base then
  begin c:=cur_mod; value(p):=expr_base+k;
  end
else  begin value(p):=cur_mod+k;
  if cur_mod=expr_base then c:=expr_macro
  else if cur_mod=suffix_base then c:=suffix_macro
  else c:=text_macro;
  end;
if k=param_size then overflow("parameter stack size",param_size);
incr(k); get_symbol; info(p):=cur_sym; link(p):=r; r:=p; get_next;
if c=expr_macro then if cur_cmd=of_token then
  begin c:=of_macro; p:=get_node(token_node_size);
  if k=param_size then overflow("parameter stack size",param_size);
  value(p):=expr_base+k; get_symbol; info(p):=cur_sym;
  link(p):=r; r:=p; get_next;
  end;
end

@* \[35] Expanding the next token.
Only a few command codes |<min_command| can possibly be returned by
|get_next|; in increasing order, they are
|if_test|, |fi_or_else|, |input|, |iteration|, |repeat_loop|,
|exit_test|, |relax|, |scan_tokens|, |expand_after|, and |defined_macro|.

\MF\ usually gets the next token of input by saying |get_x_next|. This is
like |get_next| except that it keeps getting more tokens until
finding |cur_cmd>=min_command|. In other words, |get_x_next| expands
macros and removes conditionals or iterations or input instructions that
might be present.

It follows that |get_x_next| might invoke itself recursively. In fact,
there is massive recursion, since macro expansion can involve the
scanning of arbitrarily complex expressions, which in turn involve
macro expansion and conditionals, etc.
@^recursion@>

Therefore it's necessary to declare a whole bunch of |forward|
procedures at this point, and to insert some other procedures
that will be invoked by |get_x_next|.

@p procedure@?scan_primary; forward;@t\2@>
procedure@?scan_secondary; forward;@t\2@>
procedure@?scan_tertiary; forward;@t\2@>
procedure@?scan_expression; forward;@t\2@>
procedure@?scan_suffix; forward;@t\2@>@/
@t\4@>@<Declare the procedure called |macro_call|@>@;@/
procedure@?get_boolean; forward;@t\2@>
procedure@?pass_text; forward;@t\2@>
procedure@?conditional; forward;@t\2@>
procedure@?start_input; forward;@t\2@>
procedure@?begin_iteration; forward;@t\2@>
procedure@?resume_iteration; forward;@t\2@>
procedure@?stop_iteration; forward;@t\2@>

@ An auxiliary subroutine called |expand| is used by |get_x_next|
when it has to do exotic expansion commands.

@p procedure expand;
var @!p:pointer; {for list manipulation}
@!k:integer; {something that we hope is |<=buf_size|}
@!j:pool_pointer; {index into |str_pool|}
begin if internal[tracing_commands]>unity then if cur_cmd<>defined_macro then
  show_cur_cmd_mod;
case cur_cmd of
if_test:conditional; {this procedure is discussed in Part 36 below}
fi_or_else:@<Terminate the current conditional and skip to \&{fi}@>;
input:@<Initiate or terminate input from a file@>;
iteration:if cur_mod=end_for then
    @<Scold the user for having an extra \&{endfor}@>
  else begin_iteration; {this procedure is discussed in Part 37 below}
repeat_loop: @<Repeat a loop@>;
exit_test: @<Exit a loop if the proper time has come@>;
relax: do_nothing;
expand_after: @<Expand the token after the next token@>;
scan_tokens: @<Put a string into the input buffer@>;
defined_macro:macro_call(cur_mod,null,cur_sym);
end; {there are no other cases}
end;

@ @<Scold the user...@>=
begin print_err("Extra `endfor'");
@.Extra `endfor'@>
help2("I'm not currently working on a for loop,")@/
  ("so I had better not try to end anything.");@/
error;
end

@ The processing of \&{input} involves the |start_input| subroutine,
which will be declared later; the processing of \&{endinput} is trivial.

@<Put each...@>=
primitive("input",input,0);@/
@!@:input_}{\&{input} primitive@>
primitive("endinput",input,1);@/
@!@:end_input_}{\&{endinput} primitive@>

@ @<Cases of |print_cmd_mod|...@>=
input: if m=0 then print("input")@+else print("endinput");

@ @<Initiate or terminate input...@>=
if cur_mod>0 then force_eof:=true
else start_input

@ We'll discuss the complicated parts of loop operations later. For now
it suffices to know that there's a global variable called |loop_ptr|
that will be |null| if no loop is in progress.

@<Repeat a loop@>=
begin while token_state and(loc=null) do end_token_list; {conserve stack space}
if loop_ptr=null then
  begin print_err("Lost loop");
@.Lost loop@>
  help2("I'm confused; after exiting from a loop, I still seem")@/
    ("to want to repeat it. I'll try to forget the problem.");@/
  error;
  end
else resume_iteration; {this procedure is in Part 37 below}
end

@ @<Exit a loop if the proper time has come@>=
begin get_boolean;
if internal[tracing_commands]>unity then show_cmd_mod(nullary,cur_exp);
if cur_exp=true_code then
  if loop_ptr=null then
    begin print_err("No loop is in progress");
@.No loop is in progress@>
    help1("Why say `exitif' when there's nothing to exit from?");
    if cur_cmd=semicolon then error@+else back_error;
    end
  else @<Exit prematurely from an iteration@>
else if cur_cmd<>semicolon then
  begin missing_err(";");@/
@.Missing `;'@>
  help2("After `exitif <boolean expr>' I expect to see a semicolon.")@/
  ("I shall pretend that one was there."); back_error;
  end;
end

@ Here we use the fact that |forever_text| is the only |token_type| that
is less than |loop_text|.

@<Exit prematurely...@>=
begin p:=null;
repeat if file_state then end_file_reading
else  begin if token_type<=loop_text then p:=start;
  end_token_list;
  end;
until p<>null;
if p<>info(loop_ptr) then fatal_error("*** (loop confusion)");
@.loop confusion@>
stop_iteration; {this procedure is in Part 37 below}
end

@ @<Expand the token after the next token@>=
begin get_next;
p:=cur_tok; get_next;
if cur_cmd<min_command then expand else back_input;
back_list(p);
end

@ @<Put a string into the input buffer@>=
begin get_x_next; scan_primary;
if cur_type<>string_type then
  begin disp_err(null,"Not a string");
@.Not a string@>
  help2("I'm going to flush this expression, since")@/
    ("scantokens should be followed by a known string.");
  put_get_flush_error(0);
  end
else  begin back_input;
  if length(cur_exp)>0 then @<Pretend we're reading a new one-line file@>;
  end;
end

@ @<Pretend we're reading a new one-line file@>=
begin begin_file_reading; name:=2;
k:=first+length(cur_exp);
if k>=max_buf_stack then
  begin if k>=buf_size then
    begin max_buf_stack:=buf_size;
    overflow("buffer size",buf_size);
@:METAFONT capacity exceeded buffer size}{\quad buffer size@>
    end;
  max_buf_stack:=k+1;
  end;
j:=str_start[cur_exp]; limit:=k;
while first<limit do
  begin buffer[first]:=so(str_pool[j]); incr(j); incr(first);
  end;
buffer[limit]:="%"; first:=limit+1; loc:=start; flush_cur_exp(0);
end

@ Here finally is |get_x_next|.

The expression scanning routines to be considered later
communicate via the global quantities |cur_type| and |cur_exp|;
we must be very careful to save and restore these quantities while
macros are being expanded.
@^inner loop@>

@p procedure get_x_next;
var @!save_exp:pointer; {a capsule to save |cur_type| and |cur_exp|}
begin get_next;
if cur_cmd<min_command then
  begin save_exp:=stash_cur_exp;
  repeat if cur_cmd=defined_macro then macro_call(cur_mod,null,cur_sym)
  else expand;
  get_next;
  until cur_cmd>=min_command;
  unstash_cur_exp(save_exp); {that restores |cur_type| and |cur_exp|}
  end;
end;

@ Now let's consider the |macro_call| procedure, which is used to start up
all user-defined macros. Since the arguments to a macro might be expressions,
|macro_call| is recursive.
@^recursion@>

The first parameter to |macro_call| points to the reference count of the
token list that defines the macro. The second parameter contains any
arguments that have already been parsed (see below).  The third parameter
points to the symbolic token that names the macro. If the third parameter
is |null|, the macro was defined by \&{vardef}, so its name can be
reconstructed from the prefix and ``at'' arguments found within the
second parameter.

What is this second parameter? It's simply a linked list of one-word items,
whose |info| fields point to the arguments. In other words, if |arg_list=null|,
no arguments have been scanned yet; otherwise |info(arg_list)| points to
the first scanned argument, and |link(arg_list)| points to the list of
further arguments (if any).

Arguments of type \&{expr} are so-called capsules, which we will
discuss later when we concentrate on expressions; they can be
recognized easily because their |link| field is |void|. Arguments of type
\&{suffix} and \&{text} are token lists without reference counts.

@ After argument scanning is complete, the arguments are moved to the
|param_stack|. (They can't be put on that stack any sooner, because
the stack is growing and shrinking in unpredictable ways as more arguments
are being acquired.)  Then the macro body is fed to the scanner; i.e.,
the replacement text of the macro is placed at the top of the \MF's
input stack, so that |get_next| will proceed to read it next.

@<Declare the procedure called |macro_call|@>=
@t\4@>@<Declare the procedure called |print_macro_name|@>@;
@t\4@>@<Declare the procedure called |print_arg|@>@;
@t\4@>@<Declare the procedure called |scan_text_arg|@>@;
procedure macro_call(@!def_ref,@!arg_list,@!macro_name:pointer);
  {invokes a user-defined sequence of commands}
label found;
var @!r:pointer; {current node in the macro's token list}
@!p,@!q:pointer; {for list manipulation}
@!n:integer; {the number of arguments}
@!l_delim,@!r_delim:pointer; {a delimiter pair}
@!tail:pointer; {tail of the argument list}
begin r:=link(def_ref); add_mac_ref(def_ref);
if arg_list=null then n:=0
else @<Determine the number |n| of arguments already supplied,
  and set |tail| to the tail of |arg_list|@>;
if internal[tracing_macros]>0 then
  @<Show the text of the macro being expanded, and the existing arguments@>;
@<Scan the remaining arguments, if any; set |r| to the first token
  of the replacement text@>;
@<Feed the arguments and replacement text to the scanner@>;
end;

@ @<Show the text of the macro...@>=
begin begin_diagnostic; print_ln; print_macro_name(arg_list,macro_name);
if n=3 then print("@@#"); {indicate a suffixed macro}
show_macro(def_ref,null,100000);
if arg_list<>null then
  begin n:=0; p:=arg_list;
  repeat q:=info(p);
  print_arg(q,n,0);
  incr(n); p:=link(p);
  until p=null;
  end;
end_diagnostic(false);
end

@ @<Declare the procedure called |print_macro_name|@>=
procedure print_macro_name(@!a,@!n:pointer);
var @!p,@!q:pointer; {they traverse the first part of |a|}
begin if n<>null then slow_print(text(n))
else  begin p:=info(a);
  if p=null then slow_print(text(info(info(link(a)))))
  else  begin q:=p;
    while link(q)<>null do q:=link(q);
    link(q):=info(link(a));
    show_token_list(p,null,1000,0);
    link(q):=null;
    end;
  end;
end;

@ @<Declare the procedure called |print_arg|@>=
procedure print_arg(@!q:pointer;@!n:integer;@!b:pointer);
begin if link(q)=void then print_nl("(EXPR")
else if (b<text_base)and(b<>text_macro) then print_nl("(SUFFIX")
else print_nl("(TEXT");
print_int(n); print(")<-");
if link(q)=void then print_exp(q,1)
else show_token_list(q,null,1000,0);
end;

@ @<Determine the number |n| of arguments already supplied...@>=
begin n:=1; tail:=arg_list;
while link(tail)<>null do
  begin incr(n); tail:=link(tail);
  end;
end

@ @<Scan the remaining arguments, if any; set |r|...@>=
cur_cmd:=comma+1; {anything |<>comma| will do}
while info(r)>=expr_base do
  begin @<Scan the delimited argument represented by |info(r)|@>;
  r:=link(r);
  end;
if cur_cmd=comma then
  begin print_err("Too many arguments to ");
@.Too many arguments...@>
  print_macro_name(arg_list,macro_name); print_char(";");
  print_nl("  Missing `"); slow_print(text(r_delim));
@.Missing `)'...@>
  print("' has been inserted");
  help3("I'm going to assume that the comma I just read was a")@/
   ("right delimiter, and then I'll begin expanding the macro.")@/
   ("You might want to delete some tokens before continuing.");
  error;
  end;
if info(r)<>general_macro then @<Scan undelimited argument(s)@>;
r:=link(r)

@ At this point, the reader will find it advisable to review the explanation
of token list format that was presented earlier, paying special attention to
the conventions that apply only at the beginning of a macro's token list.

On the other hand, the reader will have to take the expression-parsing
aspects of the following program on faith; we will explain |cur_type|
and |cur_exp| later. (Several things in this program depend on each other,
and it's necessary to jump into the circle somewhere.)

@<Scan the delimited argument represented by |info(r)|@>=
if cur_cmd<>comma then
  begin get_x_next;
  if cur_cmd<>left_delimiter then
    begin print_err("Missing argument to ");
@.Missing argument...@>
    print_macro_name(arg_list,macro_name);
    help3("That macro has more parameters than you thought.")@/
     ("I'll continue by pretending that each missing argument")@/
     ("is either zero or null.");
    if info(r)>=suffix_base then
      begin cur_exp:=null; cur_type:=token_list;
      end
    else  begin cur_exp:=0; cur_type:=known;
      end;
    back_error; cur_cmd:=right_delimiter; goto found;
    end;
  l_delim:=cur_sym; r_delim:=cur_mod;
  end;
@<Scan the argument represented by |info(r)|@>;
if cur_cmd<>comma then @<Check that the proper right delimiter was present@>;
found:  @<Append the current expression to |arg_list|@>

@ @<Check that the proper right delim...@>=
if (cur_cmd<>right_delimiter)or(cur_mod<>l_delim) then
  if info(link(r))>=expr_base then
    begin missing_err(",");
@.Missing `,'@>
    help3("I've finished reading a macro argument and am about to")@/
      ("read another; the arguments weren't delimited correctly.")@/
       ("You might want to delete some tokens before continuing.");
    back_error; cur_cmd:=comma;
    end
  else  begin missing_err(text(r_delim));
@.Missing `)'@>
    help2("I've gotten to the end of the macro parameter list.")@/
       ("You might want to delete some tokens before continuing.");
    back_error;
    end

@ A \&{suffix} or \&{text} parameter will have been scanned as
a token list pointed to by |cur_exp|, in which case we will have
|cur_type=token_list|.

@<Append the current expression to |arg_list|@>=
begin p:=get_avail;
if cur_type=token_list then info(p):=cur_exp
else info(p):=stash_cur_exp;
if internal[tracing_macros]>0 then
  begin begin_diagnostic; print_arg(info(p),n,info(r)); end_diagnostic(false);
  end;
if arg_list=null then arg_list:=p
else link(tail):=p;
tail:=p; incr(n);
end

@ @<Scan the argument represented by |info(r)|@>=
if info(r)>=text_base then scan_text_arg(l_delim,r_delim)
else  begin get_x_next;
  if info(r)>=suffix_base then scan_suffix
  else scan_expression;
  end

@ The parameters to |scan_text_arg| are either a pair of delimiters
or zero; the latter case is for undelimited text arguments, which
end with the first semicolon or \&{endgroup} or \&{end} that is not
contained in a group.

@<Declare the procedure called |scan_text_arg|@>=
procedure scan_text_arg(@!l_delim,@!r_delim:pointer);
label done;
var @!balance:integer; {excess of |l_delim| over |r_delim|}
@!p:pointer; {list tail}
begin warning_info:=l_delim; scanner_status:=absorbing;
p:=hold_head; balance:=1; link(hold_head):=null;
loop@+  begin get_next;
  if l_delim=0 then @<Adjust the balance for an undelimited argument;
    |goto done| if done@>
  else @<Adjust the balance for a delimited argument;
    |goto done| if done@>;
  link(p):=cur_tok; p:=link(p);
  end;
done: cur_exp:=link(hold_head); cur_type:=token_list;
scanner_status:=normal;
end;

@ @<Adjust the balance for a delimited argument...@>=
begin if cur_cmd=right_delimiter then
  begin if cur_mod=l_delim then
    begin decr(balance);
    if balance=0 then goto done;
    end;
  end
else if cur_cmd=left_delimiter then if cur_mod=r_delim then incr(balance);
end

@ @<Adjust the balance for an undelimited...@>=
begin if end_of_statement then {|cur_cmd=semicolon|, |end_group|, or |stop|}
  begin if balance=1 then goto done
  else if cur_cmd=end_group then decr(balance);
  end
else if cur_cmd=begin_group then incr(balance);
end

@ @<Scan undelimited argument(s)@>=
begin if info(r)<text_macro then
  begin get_x_next;
  if info(r)<>suffix_macro then
    if (cur_cmd=equals)or(cur_cmd=assignment) then get_x_next;
  end;
case info(r) of
primary_macro:scan_primary;
secondary_macro:scan_secondary;
tertiary_macro:scan_tertiary;
expr_macro:scan_expression;
of_macro:@<Scan an expression followed by `\&{of} $\langle$primary$\rangle$'@>;
suffix_macro:@<Scan a suffix with optional delimiters@>;
text_macro:scan_text_arg(0,0);
end; {there are no other cases}
back_input; @<Append the current expression to |arg_list|@>;
end

@ @<Scan an expression followed by `\&{of} $\langle$primary$\rangle$'@>=
begin scan_expression; p:=get_avail; info(p):=stash_cur_exp;
if internal[tracing_macros]>0 then
  begin begin_diagnostic; print_arg(info(p),n,0); end_diagnostic(false);
  end;
if arg_list=null then arg_list:=p@+else link(tail):=p;
tail:=p;incr(n);
if cur_cmd<>of_token then
  begin missing_err("of"); print(" for ");
@.Missing `of'@>
  print_macro_name(arg_list,macro_name);
  help1("I've got the first argument; will look now for the other.");
  back_error;
  end;
get_x_next; scan_primary;
end

@ @<Scan a suffix with optional delimiters@>=
begin if cur_cmd<>left_delimiter then l_delim:=null
else  begin l_delim:=cur_sym; r_delim:=cur_mod; get_x_next;
  end;
scan_suffix;
if l_delim<>null then
  begin if(cur_cmd<>right_delimiter)or(cur_mod<>l_delim) then
    begin missing_err(text(r_delim));
@.Missing `)'@>
    help2("I've gotten to the end of the macro parameter list.")@/
       ("You might want to delete some tokens before continuing.");
    back_error;
    end;
  get_x_next;
  end;
end

@ Before we put a new token list on the input stack, it is wise to clean off
all token lists that have recently been depleted. Then a user macro that ends
with a call to itself will not require unbounded stack space.

@<Feed the arguments and replacement text to the scanner@>=
while token_state and(loc=null) do end_token_list; {conserve stack space}
if param_ptr+n>max_param_stack then
  begin max_param_stack:=param_ptr+n;
  if max_param_stack>param_size then
    overflow("parameter stack size",param_size);
@:METAFONT capacity exceeded parameter stack size}{\quad parameter stack size@>
  end;
begin_token_list(def_ref,macro); name:=macro_name; loc:=r;
if n>0 then
  begin p:=arg_list;
  repeat param_stack[param_ptr]:=info(p); incr(param_ptr); p:=link(p);
  until p=null;
  flush_list(arg_list);
  end

@ It's sometimes necessary to put a single argument onto |param_stack|.
The |stack_argument| subroutine does this.

@p procedure stack_argument(@!p:pointer);
begin if param_ptr=max_param_stack then
  begin incr(max_param_stack);
  if max_param_stack>param_size then
    overflow("parameter stack size",param_size);
@:METAFONT capacity exceeded parameter stack size}{\quad parameter stack size@>
  end;
param_stack[param_ptr]:=p; incr(param_ptr);
end;

@* \[36] Conditional processing.
Let's consider now the way \&{if} commands are handled.

Conditions can be inside conditions, and this nesting has a stack
that is independent of other stacks.
Four global variables represent the top of the condition stack:
|cond_ptr| points to pushed-down entries, if~any; |cur_if| tells whether
we are processing \&{if} or \&{elseif}; |if_limit| specifies
the largest code of a |fi_or_else| command that is syntactically legal;
and |if_line| is the line number at which the current conditional began.

If no conditions are currently in progress, the condition stack has the
special state |cond_ptr=null|, |if_limit=normal|, |cur_if=0|, |if_line=0|.
Otherwise |cond_ptr| points to a two-word node; the |type|, |name_type|, and
|link| fields of the first word contain |if_limit|, |cur_if|, and
|cond_ptr| at the next level, and the second word contains the
corresponding |if_line|.

@d if_node_size=2 {number of words in stack entry for conditionals}
@d if_line_field(#)==mem[#+1].int
@d if_code=1 {code for \&{if} being evaluated}
@d fi_code=2 {code for \&{fi}}
@d else_code=3 {code for \&{else}}
@d else_if_code=4 {code for \&{elseif}}

@<Glob...@>=
@!cond_ptr:pointer; {top of the condition stack}
@!if_limit:normal..else_if_code; {upper bound on |fi_or_else| codes}
@!cur_if:small_number; {type of conditional being worked on}
@!if_line:integer; {line where that conditional began}

@ @<Set init...@>=
cond_ptr:=null; if_limit:=normal; cur_if:=0; if_line:=0;

@ @<Put each...@>=
primitive("if",if_test,if_code);@/
@!@:if_}{\&{if} primitive@>
primitive("fi",fi_or_else,fi_code); eqtb[frozen_fi]:=eqtb[cur_sym];@/
@!@:fi_}{\&{fi} primitive@>
primitive("else",fi_or_else,else_code);@/
@!@:else_}{\&{else} primitive@>
primitive("elseif",fi_or_else,else_if_code);@/
@!@:else_if_}{\&{elseif} primitive@>

@ @<Cases of |print_cmd_mod|...@>=
if_test,fi_or_else: case m of
  if_code:print("if");
  fi_code:print("fi");
  else_code:print("else");
  othercases print("elseif")
  endcases;

@ Here is a procedure that ignores text until coming to an \&{elseif},
\&{else}, or \&{fi} at the current level of $\&{if}\ldots\&{fi}$
nesting. After it has acted, |cur_mod| will indicate the token that
was found.

\MF's smallest two command codes are |if_test| and |fi_or_else|; this
makes the skipping process a bit simpler.

@p procedure pass_text;
label done;
var l:integer;
begin scanner_status:=skipping; l:=0; warning_info:=line;
loop@+  begin get_next;
  if cur_cmd<=fi_or_else then
    if cur_cmd<fi_or_else then incr(l)
    else  begin if l=0 then goto done;
      if cur_mod=fi_code then decr(l);
      end
  else @<Decrease the string reference count,
    if the current token is a string@>;
  end;
done: scanner_status:=normal;
end;

@ @<Decrease the string reference count...@>=
if cur_cmd=string_token then delete_str_ref(cur_mod)

@ When we begin to process a new \&{if}, we set |if_limit:=if_code|; then
if \&{elseif} or \&{else} or \&{fi} occurs before the current \&{if}
condition has been evaluated, a colon will be inserted.
A construction like `\.{if fi}' would otherwise get \MF\ confused.

@<Push the condition stack@>=
begin p:=get_node(if_node_size); link(p):=cond_ptr; type(p):=if_limit;
name_type(p):=cur_if; if_line_field(p):=if_line;
cond_ptr:=p; if_limit:=if_code; if_line:=line; cur_if:=if_code;
end

@ @<Pop the condition stack@>=
begin p:=cond_ptr; if_line:=if_line_field(p);
cur_if:=name_type(p); if_limit:=type(p); cond_ptr:=link(p);
free_node(p,if_node_size);
end

@ Here's a procedure that changes the |if_limit| code corresponding to
a given value of |cond_ptr|.

@p procedure change_if_limit(@!l:small_number;@!p:pointer);
label exit;
var q:pointer;
begin if p=cond_ptr then if_limit:=l {that's the easy case}
else  begin q:=cond_ptr;
  loop@+  begin if q=null then confusion("if");
@:this can't happen if}{\quad if@>
    if link(q)=p then
      begin type(q):=l; return;
      end;
    q:=link(q);
    end;
  end;
exit:end;

@ The user is supposed to put colons into the proper parts of conditional
statements. Therefore, \MF\ has to check for their presence.

@p procedure check_colon;
begin if cur_cmd<>colon then
  begin missing_err(":");@/
@.Missing `:'@>
  help2("There should've been a colon after the condition.")@/
    ("I shall pretend that one was there.");@;
  back_error;
  end;
end;

@ A condition is started when the |get_x_next| procedure encounters
an |if_test| command; in that case |get_x_next| calls |conditional|,
which is a recursive procedure.
@^recursion@>

@p procedure conditional;
label exit,done,reswitch,found;
var @!save_cond_ptr:pointer; {|cond_ptr| corresponding to this conditional}
@!new_if_limit:fi_code..else_if_code; {future value of |if_limit|}
@!p:pointer; {temporary register}
begin @<Push the condition stack@>;@+save_cond_ptr:=cond_ptr;
reswitch: get_boolean; new_if_limit:=else_if_code;
if internal[tracing_commands]>unity then
  @<Display the boolean value of |cur_exp|@>;
found: check_colon;
if cur_exp=true_code then
  begin change_if_limit(new_if_limit,save_cond_ptr);
  return; {wait for \&{elseif}, \&{else}, or \&{fi}}
  end;
@<Skip to \&{elseif} or \&{else} or \&{fi}, then |goto done|@>;
done: cur_if:=cur_mod; if_line:=line;
if cur_mod=fi_code then @<Pop the condition stack@>
else if cur_mod=else_if_code then goto reswitch
else  begin cur_exp:=true_code; new_if_limit:=fi_code; get_x_next; goto found;
  end;
exit:end;

@ In a construction like `\&{if} \&{if} \&{true}: $0=1$: \\{foo}
\&{else}: \\{bar} \&{fi}', the first \&{else}
that we come to after learning that the \&{if} is false is not the
\&{else} we're looking for. Hence the following curious logic is needed.

@<Skip to \&{elseif}...@>=
loop@+  begin pass_text;
  if cond_ptr=save_cond_ptr then goto done
  else if cur_mod=fi_code then @<Pop the condition stack@>;
  end


@ @<Display the boolean value...@>=
begin begin_diagnostic;
if cur_exp=true_code then print("{true}")@+else print("{false}");
end_diagnostic(false);
end

@ The processing of conditionals is complete except for the following
code, which is actually part of |get_x_next|. It comes into play when
\&{elseif}, \&{else}, or \&{fi} is scanned.

@<Terminate the current conditional and skip to \&{fi}@>=
if cur_mod>if_limit then
  if if_limit=if_code then {condition not yet evaluated}
    begin missing_err(":");
@.Missing `:'@>
    back_input; cur_sym:=frozen_colon; ins_error;
    end
  else  begin print_err("Extra "); print_cmd_mod(fi_or_else,cur_mod);
@.Extra else@>
@.Extra elseif@>
@.Extra fi@>
    help1("I'm ignoring this; it doesn't match any if.");
    error;
    end
else  begin while cur_mod<>fi_code do pass_text; {skip to \&{fi}}
  @<Pop the condition stack@>;
  end

@* \[37] Iterations.
To bring our treatment of |get_x_next| to a close, we need to consider what
\MF\ does when it sees \&{for}, \&{forsuffixes}, and \&{forever}.

There's a global variable |loop_ptr| that keeps track of the \&{for} loops
that are currently active. If |loop_ptr=null|, no loops are in progress;
otherwise |info(loop_ptr)| points to the iterative text of the current
(innermost) loop, and |link(loop_ptr)| points to the data for any other
loops that enclose the current one.

A loop-control node also has two other fields, called |loop_type| and
|loop_list|, whose contents depend on the type of loop:

\yskip\indent|loop_type(loop_ptr)=null| means that |loop_list(loop_ptr)|
points to a list of one-word nodes whose |info| fields point to the
remaining argument values of a suffix list and expression list.

\yskip\indent|loop_type(loop_ptr)=void| means that the current loop is
`\&{forever}'.

\yskip\indent|loop_type(loop_ptr)=p>void| means that |value(p)|,
|step_size(p)|, and |final_value(p)| contain the data for an arithmetic
progression.

\yskip\noindent In the latter case, |p| points to a ``progression node''
whose first word is not used. (No value could be stored there because the
link field of words in the dynamic memory area cannot be arbitrary.)

@d loop_list_loc(#)==#+1 {where the |loop_list| field resides}
@d loop_type(#)==info(loop_list_loc(#)) {the type of \&{for} loop}
@d loop_list(#)==link(loop_list_loc(#)) {the remaining list elements}
@d loop_node_size=2 {the number of words in a loop control node}
@d progression_node_size=4 {the number of words in a progression node}
@d step_size(#)==mem[#+2].sc {the step size in an arithmetic progression}
@d final_value(#)==mem[#+3].sc {the final value in an arithmetic progression}

@<Glob...@>=
@!loop_ptr:pointer; {top of the loop-control-node stack}

@ @<Set init...@>=
loop_ptr:=null;

@ If the expressions that define an arithmetic progression in
a \&{for} loop don't have known numeric values, the |bad_for|
subroutine screams at the user.

@p procedure bad_for(@!s:str_number);
begin disp_err(null,"Improper "); {show the bad expression above the message}
@.Improper...replaced by 0@>
print(s); print(" has been replaced by 0");
help4("When you say `for x=a step b until c',")@/
  ("the initial value `a' and the step size `b'")@/
  ("and the final value `c' must have known numeric values.")@/
  ("I'm zeroing this one. Proceed, with fingers crossed.");
put_get_flush_error(0);
end;

@ Here's what \MF\ does when \&{for}, \&{forsuffixes}, or \&{forever}
has just been scanned. (This code requires slight familiarity with
expression-parsing routines that we have not yet discussed; but it seems
to belong in the present part of the program, even though the author
didn't write it until later. The reader may wish to come back to it.)

@p procedure begin_iteration;
label continue,done,found;
var @!m:halfword; {|expr_base| (\&{for}) or |suffix_base| (\&{forsuffixes})}
@!n:halfword; {hash address of the current symbol}
@!p,@!q,@!s,@!pp:pointer; {link manipulation registers}
begin m:=cur_mod; n:=cur_sym; s:=get_node(loop_node_size);
if m=start_forever then
  begin loop_type(s):=void; p:=null; get_x_next; goto found;
  end;
get_symbol; p:=get_node(token_node_size); info(p):=cur_sym; value(p):=m;@/
get_x_next;
if (cur_cmd<>equals)and(cur_cmd<>assignment) then
  begin missing_err("=");@/
@.Missing `='@>
  help3("The next thing in this loop should have been `=' or `:='.")@/
    ("But don't worry; I'll pretend that an equals sign")@/
    ("was present, and I'll look for the values next.");@/
  back_error;
  end;
@<Scan the values to be used in the loop@>;
found:@<Check for the presence of a colon@>;
@<Scan the loop text and put it on the loop control stack@>;
resume_iteration;
end;

@ @<Check for the presence of a colon@>=
if cur_cmd<>colon then
  begin missing_err(":");@/
@.Missing `:'@>
  help3("The next thing in this loop should have been a `:'.")@/
    ("So I'll pretend that a colon was present;")@/
    ("everything from here to `endfor' will be iterated.");
  back_error;
  end

@ We append a special |frozen_repeat_loop| token in place of the
`\&{endfor}' at the end of the loop. This will come through \MF's scanner
at the proper time to cause the loop to be repeated.

(A user who tries some shenanigan like `\&{for} $\ldots$ \&{let} \&{endfor}'
will be foiled by the |get_symbol| routine, which keeps frozen
tokens unchanged. Furthermore the |frozen_repeat_loop| is an \&{outer}
token, so it won't be lost accidentally.)

@ @<Scan the loop text...@>=
q:=get_avail; info(q):=frozen_repeat_loop;
scanner_status:=loop_defining; warning_info:=n;
info(s):=scan_toks(iteration,p,q,0); scanner_status:=normal;@/
link(s):=loop_ptr; loop_ptr:=s

@ @<Initialize table...@>=
eq_type(frozen_repeat_loop):=repeat_loop+outer_tag;
text(frozen_repeat_loop):=" ENDFOR";

@ The loop text is inserted into \MF's scanning apparatus by the
|resume_iteration| routine.

@p procedure resume_iteration;
label not_found,exit;
var @!p,@!q:pointer; {link registers}
begin p:=loop_type(loop_ptr);
if p>void then {|p| points to a progression node}
  begin cur_exp:=value(p);
  if @<The arithmetic progression has ended@> then goto not_found;
  cur_type:=known; q:=stash_cur_exp; {make |q| an \&{expr} argument}
  value(p):=cur_exp+step_size(p); {set |value(p)| for the next iteration}
  end
else if p<void then
  begin p:=loop_list(loop_ptr);
  if p=null then goto not_found;
  loop_list(loop_ptr):=link(p); q:=info(p); free_avail(p);
  end
else  begin begin_token_list(info(loop_ptr),forever_text); return;
  end;
begin_token_list(info(loop_ptr),loop_text);
stack_argument(q);
if internal[tracing_commands]>unity then @<Trace the start of a loop@>;
return;
not_found:stop_iteration;
exit:end;

@ @<The arithmetic progression has ended@>=
((step_size(p)>0)and(cur_exp>final_value(p)))or@|
 ((step_size(p)<0)and(cur_exp<final_value(p)))

@ @<Trace the start of a loop@>=
begin begin_diagnostic; print_nl("{loop value=");
@.loop value=n@>
if (q<>null)and(link(q)=void) then print_exp(q,1)
else show_token_list(q,null,50,0);
print_char("}"); end_diagnostic(false);
end

@ A level of loop control disappears when |resume_iteration| has decided
not to resume, or when an \&{exitif} construction has removed the loop text
from the input stack.

@p procedure stop_iteration;
var @!p,@!q:pointer; {the usual}
begin p:=loop_type(loop_ptr);
if p>void then free_node(p,progression_node_size)
else if p<void then
  begin q:=loop_list(loop_ptr);
  while q<>null do
    begin p:=info(q);
    if p<>null then
      if link(p)=void then {it's an \&{expr} parameter}
        begin recycle_value(p); free_node(p,value_node_size);
        end
      else flush_token_list(p); {it's a \&{suffix} or \&{text} parameter}
    p:=q; q:=link(q); free_avail(p);
    end;
  end;
p:=loop_ptr; loop_ptr:=link(p); flush_token_list(info(p));
free_node(p,loop_node_size);
end;

@ Now that we know all about loop control, we can finish up
the missing portion of |begin_iteration| and we'll be done.

The following code is performed after the `\.=' has been scanned in
a \&{for} construction (if |m=expr_base|) or a \&{forsuffixes} construction
(if |m=suffix_base|).

@<Scan the values to be used in the loop@>=
loop_type(s):=null; q:=loop_list_loc(s); link(q):=null; {|link(q)=loop_list(s)|}
repeat get_x_next;
if m<>expr_base then scan_suffix
else  begin if cur_cmd>=colon then if cur_cmd<=comma then goto continue;
  scan_expression;
  if cur_cmd=step_token then if q=loop_list_loc(s) then
    @<Prepare for step-until construction and |goto done|@>;
  cur_exp:=stash_cur_exp;
  end;
link(q):=get_avail; q:=link(q); info(q):=cur_exp; cur_type:=vacuous;
continue: until cur_cmd<>comma;
done:

@ @<Prepare for step-until construction and |goto done|@>=
begin if cur_type<>known then bad_for("initial value");
pp:=get_node(progression_node_size); value(pp):=cur_exp;@/
get_x_next; scan_expression;
if cur_type<>known then bad_for("step size");
step_size(pp):=cur_exp;
if cur_cmd<>until_token then
  begin missing_err("until");@/
@.Missing `until'@>
  help2("I assume you meant to say `until' after `step'.")@/
    ("So I'll look for the final value and colon next.");
  back_error;
  end;
get_x_next; scan_expression;
if cur_type<>known then bad_for("final value");
final_value(pp):=cur_exp; loop_type(s):=pp; goto done;
end

@* \[38] File names.
It's time now to fret about file names.  Besides the fact that different
operating systems treat files in different ways, we must cope with the
fact that completely different naming conventions are used by different
groups of people. The following programs show what is required for one
particular operating system; similar routines for other systems are not
difficult to devise.
@^system dependencies@>

\MF\ assumes that a file name has three parts: the name proper; its
``extension''; and a ``file area'' where it is found in an external file
system.  The extension of an input file is assumed to be
`\.{.mf}' unless otherwise specified; it is `\.{.log}' on the
transcript file that records each run of \MF; it is `\.{.tfm}' on the font
metric files that describe characters in the fonts \MF\ creates; it is
`\.{.gf}' on the output files that specify generic font information; and it
is `\.{.base}' on the base files written by \.{INIMF} to initialize \MF.
The file area can be arbitrary on input files, but files are usually
output to the user's current area.  If an input file cannot be
found on the specified area, \MF\ will look for it on a special system
area; this special area is intended for commonly used input files.

Simple uses of \MF\ refer only to file names that have no explicit
extension or area. For example, a person usually says `\.{input} \.{cmr10}'
instead of `\.{input} \.{cmr10.new}'. Simple file
names are best, because they make the \MF\ source files portable;
whenever a file name consists entirely of letters and digits, it should be
treated in the same way by all implementations of \MF. However, users
need the ability to refer to other files in their environment, especially
when responding to error messages concerning unopenable files; therefore
we want to let them use the syntax that appears in their favorite
operating system.

@ \MF\ uses the same conventions that have proved to be satisfactory for
\TeX. In order to isolate the system-dependent aspects of file names, the
@^system dependencies@>
system-independent parts of \MF\ are expressed in terms
of three system-dependent
procedures called |begin_name|, |more_name|, and |end_name|. In
essence, if the user-specified characters of the file name are $c_1\ldots c_n$,
the system-independent driver program does the operations
$$|begin_name|;\,|more_name|(c_1);\,\ldots\,;\,|more_name|(c_n);
\,|end_name|.$$
These three procedures communicate with each other via global variables.
Afterwards the file name will appear in the string pool as three strings
called |cur_name|\penalty10000\hskip-.05em,
|cur_area|, and |cur_ext|; the latter two are null (i.e.,
|""|), unless they were explicitly specified by the user.

Actually the situation is slightly more complicated, because \MF\ needs
to know when the file name ends. The |more_name| routine is a function
(with side effects) that returns |true| on the calls |more_name|$(c_1)$,
\dots, |more_name|$(c_{n-1})$. The final call |more_name|$(c_n)$
returns |false|; or, it returns |true| and $c_n$ is the last character
on the current input line. In other words,
|more_name| is supposed to return |true| unless it is sure that the
file name has been completely scanned; and |end_name| is supposed to be able
to finish the assembly of |cur_name|, |cur_area|, and |cur_ext| regardless of
whether $|more_name|(c_n)$ returned |true| or |false|.

@<Glob...@>=
@!cur_name:str_number; {name of file just scanned}
@!cur_area:str_number; {file area just scanned, or \.{""}}
@!cur_ext:str_number; {file extension just scanned, or \.{""}}

@ The file names we shall deal with for illustrative purposes have the
following structure:  If the name contains `\.>' or `\.:', the file area
consists of all characters up to and including the final such character;
otherwise the file area is null.  If the remaining file name contains
`\..', the file extension consists of all such characters from the first
remaining `\..' to the end, otherwise the file extension is null.
@^system dependencies@>

We can scan such file names easily by using two global variables that keep track
of the occurrences of area and extension delimiters:

@<Glob...@>=
@!area_delimiter:pool_pointer; {the most recent `\.>' or `\.:', if any}
@!ext_delimiter:pool_pointer; {the relevant `\..', if any}

@ Input files that can't be found in the user's area may appear in a standard
system area called |MF_area|.
This system area name will, of course, vary from place to place.
@^system dependencies@>

@d MF_area=="MFinputs:"
@.MFinputs@>

@ Here now is the first of the system-dependent routines for file name scanning.
@^system dependencies@>

@p procedure begin_name;
begin area_delimiter:=0; ext_delimiter:=0;
end;

@ And here's the second.
@^system dependencies@>

@p function more_name(@!c:ASCII_code):boolean;
begin if c=" " then more_name:=false
else  begin if (c=">")or(c=":") then
    begin area_delimiter:=pool_ptr; ext_delimiter:=0;
    end
  else if (c=".")and(ext_delimiter=0) then ext_delimiter:=pool_ptr;
  str_room(1); append_char(c); {contribute |c| to the current string}
  more_name:=true;
  end;
end;

@ The third.
@^system dependencies@>

@p procedure end_name;
begin if str_ptr+3>max_str_ptr then
  begin if str_ptr+3>max_strings then
    overflow("number of strings",max_strings-init_str_ptr);
@:METAFONT capacity exceeded number of strings}{\quad number of strings@>
  max_str_ptr:=str_ptr+3;
  end;
if area_delimiter=0 then cur_area:=""
else  begin cur_area:=str_ptr; incr(str_ptr);
  str_start[str_ptr]:=area_delimiter+1;
  end;
if ext_delimiter=0 then
  begin cur_ext:=""; cur_name:=make_string;
  end
else  begin cur_name:=str_ptr; incr(str_ptr);
  str_start[str_ptr]:=ext_delimiter; cur_ext:=make_string;
  end;
end;

@ Conversely, here is a routine that takes three strings and prints a file
name that might have produced them. (The routine is system dependent, because
some operating systems put the file area last instead of first.)
@^system dependencies@>

@<Basic printing...@>=
procedure print_file_name(@!n,@!a,@!e:integer);
begin slow_print(a); slow_print(n); slow_print(e);
end;

@ Another system-dependent routine is needed to convert three internal
\MF\ strings
to the |name_of_file| value that is used to open files. The present code
allows both lowercase and uppercase letters in the file name.
@^system dependencies@>

@d append_to_name(#)==begin c:=#; incr(k);
  if k<=file_name_size then name_of_file[k]:=xchr[c];
  end

@p procedure pack_file_name(@!n,@!a,@!e:str_number);
var @!k:integer; {number of positions filled in |name_of_file|}
@!c: ASCII_code; {character being packed}
@!j:pool_pointer; {index into |str_pool|}
begin k:=0;
for j:=str_start[a] to str_start[a+1]-1 do append_to_name(so(str_pool[j]));
for j:=str_start[n] to str_start[n+1]-1 do append_to_name(so(str_pool[j]));
for j:=str_start[e] to str_start[e+1]-1 do append_to_name(so(str_pool[j]));
if k<=file_name_size then name_length:=k@+else name_length:=file_name_size;
for k:=name_length+1 to file_name_size do name_of_file[k]:=' ';
end;

@ A messier routine is also needed, since base file names must be scanned
before \MF's string mechanism has been initialized. We shall use the
global variable |MF_base_default| to supply the text for default system areas
and extensions related to base files.
@^system dependencies@>

@d base_default_length=18 {length of the |MF_base_default| string}
@d base_area_length=8 {length of its area part}
@d base_ext_length=5 {length of its `\.{.base}' part}
@d base_extension=".base" {the extension, as a \.{WEB} constant}

@<Glob...@>=
@!MF_base_default:packed array[1..base_default_length] of char;

@ @<Set init...@>=
MF_base_default:='MFbases:plain.base';
@.MFbases@>
@.plain@>
@^system dependencies@>

@ @<Check the ``constant'' values for consistency@>=
if base_default_length>file_name_size then bad:=41;

@ Here is the messy routine that was just mentioned. It sets |name_of_file|
from the first |n| characters of |MF_base_default|, followed by
|buffer[a..b]|, followed by the last |base_ext_length| characters of
|MF_base_default|.

We dare not give error messages here, since \MF\ calls this routine before
the |error| routine is ready to roll. Instead, we simply drop excess characters,
since the error will be detected in another way when a strange file name
isn't found.
@^system dependencies@>

@p procedure pack_buffered_name(@!n:small_number;@!a,@!b:integer);
var @!k:integer; {number of positions filled in |name_of_file|}
@!c: ASCII_code; {character being packed}
@!j:integer; {index into |buffer| or |MF_base_default|}
begin if n+b-a+1+base_ext_length>file_name_size then
  b:=a+file_name_size-n-1-base_ext_length;
k:=0;
for j:=1 to n do append_to_name(xord[MF_base_default[j]]);
for j:=a to b do append_to_name(buffer[j]);
for j:=base_default_length-base_ext_length+1 to base_default_length do
  append_to_name(xord[MF_base_default[j]]);
if k<=file_name_size then name_length:=k@+else name_length:=file_name_size;
for k:=name_length+1 to file_name_size do name_of_file[k]:=' ';
end;

@ Here is the only place we use |pack_buffered_name|. This part of the program
becomes active when a ``virgin'' \MF\ is trying to get going, just after
the preliminary initialization, or when the user is substituting another
base file by typing `\.\&' after the initial `\.{**}' prompt.  The buffer
contains the first line of input in |buffer[loc..(last-1)]|, where
|loc<last| and |buffer[loc]<>" "|.

@<Declare the function called |open_base_file|@>=
function open_base_file:boolean;
label found,exit;
var @!j:0..buf_size; {the first space after the file name}
begin j:=loc;
if buffer[loc]="&" then
  begin incr(loc); j:=loc; buffer[last]:=" ";
  while buffer[j]<>" " do incr(j);
  pack_buffered_name(0,loc,j-1); {try first without the system file area}
  if w_open_in(base_file) then goto found;
  pack_buffered_name(base_area_length,loc,j-1);
    {now try the system base file area}
  if w_open_in(base_file) then goto found;
  wake_up_terminal;
  wterm_ln('Sorry, I can''t find that base;',' will try PLAIN.');
@.Sorry, I can't find...@>
  update_terminal;
  end;
  {now pull out all the stops: try for the system \.{plain} file}
pack_buffered_name(base_default_length-base_ext_length,1,0);
if not w_open_in(base_file) then
  begin wake_up_terminal;
  wterm_ln('I can''t find the PLAIN base file!');
@.I can't find PLAIN...@>
@.plain@>
  open_base_file:=false; return;
  end;
found:loc:=j; open_base_file:=true;
exit:end;

@ Operating systems often make it possible to determine the exact name (and
possible version number) of a file that has been opened. The following routine,
which simply makes a \MF\ string from the value of |name_of_file|, should
ideally be changed to deduce the full name of file~|f|, which is the file
most recently opened, if it is possible to do this in a \PASCAL\ program.
@^system dependencies@>

This routine might be called after string memory has overflowed, hence
we dare not use `|str_room|'.

@p function make_name_string:str_number;
var @!k:1..file_name_size; {index into |name_of_file|}
begin if (pool_ptr+name_length>pool_size)or(str_ptr=max_strings) then
  make_name_string:="?"
else  begin for k:=1 to name_length do append_char(xord[name_of_file[k]]);
  make_name_string:=make_string;
  end;
end;
function a_make_name_string(var @!f:alpha_file):str_number;
begin a_make_name_string:=make_name_string;
end;
function b_make_name_string(var @!f:byte_file):str_number;
begin b_make_name_string:=make_name_string;
end;
function w_make_name_string(var @!f:word_file):str_number;
begin w_make_name_string:=make_name_string;
end;

@ Now let's consider the ``driver''
routines by which \MF\ deals with file names
in a system-independent manner.  First comes a procedure that looks for a
file name in the input by taking the information from the input buffer.
(We can't use |get_next|, because the conversion to tokens would
destroy necessary information.)

This procedure doesn't allow semicolons or percent signs to be part of
file names, because of other conventions of \MF. The manual doesn't
use semicolons or percents immediately after file names, but some users
no doubt will find it natural to do so; therefore system-dependent
changes to allow such characters in file names should probably
be made with reluctance, and only when an entire file name that
includes special characters is ``quoted'' somehow.
@^system dependencies@>

@p procedure scan_file_name;
label done;
begin begin_name;
while buffer[loc]=" " do incr(loc);
loop@+begin if (buffer[loc]=";")or(buffer[loc]="%") then goto done;
  if not more_name(buffer[loc]) then goto done;
  incr(loc);
  end;
done: end_name;
end;

@ The global variable |job_name| contains the file name that was first
\&{input} by the user. This name is extended by `\.{.log}' and `\.{.gf}' and
`\.{.base}' and `\.{.tfm}' in the names of \MF's output files.

@<Glob...@>=
@!job_name:str_number; {principal file name}
@!log_opened:boolean; {has the transcript file been opened?}
@!log_name:str_number; {full name of the log file}

@ Initially |job_name=0|; it becomes nonzero as soon as the true name is known.
We have |job_name=0| if and only if the `\.{log}' file has not been opened,
except of course for a short time just after |job_name| has become nonzero.

@<Initialize the output...@>=job_name:=0; log_opened:=false;

@ Here is a routine that manufactures the output file names, assuming that
|job_name<>0|. It ignores and changes the current settings of |cur_area|
and |cur_ext|.

@d pack_cur_name==pack_file_name(cur_name,cur_area,cur_ext)

@p procedure pack_job_name(@!s:str_number); {|s = ".log"|, |".gf"|,
  |".tfm"|, or |base_extension|}
begin cur_area:=""; cur_ext:=s;
cur_name:=job_name; pack_cur_name;
end;

@ Actually the main output file extension is usually something like
|".300gf"| instead of just |".gf"|; the additional number indicates the
resolution in pixels per inch, based on the setting of |hppp| when
the file is opened.

@<Glob...@>=
@!gf_ext:str_number; {default extension for the output file}

@ If some trouble arises when \MF\ tries to open a file, the following
routine calls upon the user to supply another file name. Parameter~|s|
is used in the error message to identify the type of file; parameter~|e|
is the default extension if none is given. Upon exit from the routine,
variables |cur_name|, |cur_area|, |cur_ext|, and |name_of_file| are
ready for another attempt at file opening.

@p procedure prompt_file_name(@!s,@!e:str_number);
label done;
var @!k:0..buf_size; {index into |buffer|}
begin if interaction=scroll_mode then wake_up_terminal;
if s="input file name" then print_err("I can't find file `")
@.I can't find file x@>
else print_err("I can't write on file `");
@.I can't write on file x@>
print_file_name(cur_name,cur_area,cur_ext); print("'.");
if e=".mf" then show_context;
print_nl("Please type another "); print(s);
@.Please type...@>
if interaction<scroll_mode then
  fatal_error("*** (job aborted, file error in nonstop mode)");
@.job aborted, file error...@>
clear_terminal; prompt_input(": "); @<Scan file name in the buffer@>;
if cur_ext="" then cur_ext:=e;
pack_cur_name;
end;

@ @<Scan file name in the buffer@>=
begin begin_name; k:=first;
while (buffer[k]=" ")and(k<last) do incr(k);
loop@+  begin if k=last then goto done;
  if not more_name(buffer[k]) then goto done;
  incr(k);
  end;
done:end_name;
end

@ The |open_log_file| routine is used to open the transcript file and to help
it catch up to what has previously been printed on the terminal.

@p procedure open_log_file;
var @!old_setting:0..max_selector; {previous |selector| setting}
@!k:0..buf_size; {index into |months| and |buffer|}
@!l:0..buf_size; {end of first input line}
@!m:integer; {the current month}
@!months:packed array [1..36] of char; {abbreviations of month names}
begin old_setting:=selector;
if job_name=0 then job_name:="mfput";
@.mfput@>
pack_job_name(".log");
while not a_open_out(log_file) do @<Try to get a different log file name@>;
log_name:=a_make_name_string(log_file);
selector:=log_only; log_opened:=true;
@<Print the banner line, including the date and time@>;
input_stack[input_ptr]:=cur_input; {make sure bottom level is in memory}
print_nl("**");
@.**@>
l:=input_stack[0].limit_field-1; {last position of first line}
for k:=1 to l do print(buffer[k]);
print_ln; {now the transcript file contains the first line of input}
selector:=old_setting+2; {|log_only| or |term_and_log|}
end;

@ Sometimes |open_log_file| is called at awkward moments when \MF\ is
unable to print error messages or even to |show_context|.
The |prompt_file_name| routine can result in a |fatal_error|, but the |error|
routine will not be invoked because |log_opened| will be false.

The normal idea of |batch_mode| is that nothing at all should be written
on the terminal. However, in the unusual case that
no log file could be opened, we make an exception and allow
an explanatory message to be seen.

Incidentally, the program always refers to the log file as a `\.{transcript
file}', because some systems cannot use the extension `\.{.log}' for
this file.

@<Try to get a different log file name@>=
begin selector:=term_only;
prompt_file_name("transcript file name",".log");
end

@ @<Print the banner...@>=
begin wlog(banner);
slow_print(base_ident); print("  ");
print_int(sys_day); print_char(" ");
months:='JANFEBMARAPRMAYJUNJULAUGSEPOCTNOVDEC';
for k:=3*sys_month-2 to 3*sys_month do wlog(months[k]);
print_char(" "); print_int(sys_year); print_char(" ");
print_dd(sys_time div 60); print_char(":"); print_dd(sys_time mod 60);
end

@ Here's an example of how these file-name-parsing routines work in practice.
We shall use the macro |set_output_file_name| when it is time to
crank up the output file.

@d set_output_file_name==
  begin if job_name=0 then open_log_file;
  pack_job_name(gf_ext);
  while not b_open_out(gf_file) do
    prompt_file_name("file name for output",gf_ext);
  output_file_name:=b_make_name_string(gf_file);
  end

@<Glob...@>=
@!gf_file: byte_file; {the generic font output goes here}
@!output_file_name: str_number; {full name of the output file}

@ @<Initialize the output...@>=output_file_name:=0;

@ Let's turn now to the procedure that is used to initiate file reading
when an `\.{input}' command is being processed.
Beware: For historic reasons, this code foolishly conserves a tiny bit
of string pool space; but that can confuse the interactive `\.E' option.
@^system dependencies@>

@p procedure start_input; {\MF\ will \.{input} something}
label done;
begin @<Put the desired file name in |(cur_name,cur_ext,cur_area)|@>;
if cur_ext="" then cur_ext:=".mf";
pack_cur_name;
loop@+  begin begin_file_reading; {set up |cur_file| and new level of input}
  if a_open_in(cur_file) then goto done;
  if cur_area="" then
    begin pack_file_name(cur_name,MF_area,cur_ext);
    if a_open_in(cur_file) then goto done;
    end;
  end_file_reading; {remove the level that didn't work}
  prompt_file_name("input file name",".mf");
  end;
done: name:=a_make_name_string(cur_file); str_ref[cur_name]:=max_str_ref;
if job_name=0 then
  begin job_name:=cur_name; open_log_file;
  end; {|open_log_file| doesn't |show_context|, so |limit|
    and |loc| needn't be set to meaningful values yet}
if term_offset+length(name)>max_print_line-2 then print_ln
else if (term_offset>0)or(file_offset>0) then print_char(" ");
print_char("("); incr(open_parens); slow_print(name); update_terminal;
if name=str_ptr-1 then {conserve string pool space (but see note above)}
  begin flush_string(name); name:=cur_name;
  end;
@<Read the first line of the new file@>;
end;

@ Here we have to remember to tell the |input_ln| routine not to
start with a |get|. If the file is empty, it is considered to
contain a single blank line.
@^system dependencies@>

@<Read the first line...@>=
begin line:=1;
if input_ln(cur_file,false) then do_nothing;
firm_up_the_line;
buffer[limit]:="%"; first:=limit+1; loc:=start;
end

@ @<Put the desired file name in |(cur_name,cur_ext,cur_area)|@>=
while token_state and(loc=null) do end_token_list;
if token_state then
  begin print_err("File names can't appear within macros");
@.File names can't...@>
  help3("Sorry...I've converted what follows to tokens,")@/
    ("possibly garbaging the name you gave.")@/
    ("Please delete the tokens and insert the name again.");@/
  error;
  end;
if file_state then scan_file_name
else  begin cur_name:=""; cur_ext:=""; cur_area:="";
  end

@* \[39] Introduction to the parsing routines.
We come now to the central nervous system that sparks many of \MF's activities.
By evaluating expressions, from their primary constituents to ever larger
subexpressions, \MF\ builds the structures that ultimately define fonts of type.

Four mutually recursive subroutines are involved in this process: We call them
$$\hbox{|scan_primary|, |scan_secondary|, |scan_tertiary|,
and |scan_expression|.}$$
@^recursion@>
Each of them is parameterless and begins with the first token to be scanned
already represented in |cur_cmd|, |cur_mod|, and |cur_sym|. After execution,
the value of the primary or secondary or tertiary or expression that was
found will appear in the global variables |cur_type| and |cur_exp|. The
token following the expression will be represented in |cur_cmd|, |cur_mod|,
and |cur_sym|.

Technically speaking, the parsing algorithms are ``LL(1),'' more or less;
backup mechanisms have been added in order to provide reasonable error
recovery.

@<Glob...@>=
@!cur_type:small_number; {the type of the expression just found}
@!cur_exp:integer; {the value of the expression just found}

@ @<Set init...@>=
cur_exp:=0;

@ Many different kinds of expressions are possible, so it is wise to have
precise descriptions of what |cur_type| and |cur_exp| mean in all cases:

\smallskip\hang
|cur_type=vacuous| means that this expression didn't turn out to have a
value at all, because it arose from a \&{begingroup}$\,\ldots\,$\&{endgroup}
construction in which there was no expression before the \&{endgroup}.
In this case |cur_exp| has some irrelevant value.

\smallskip\hang
|cur_type=boolean_type| means that |cur_exp| is either |true_code|
or |false_code|.

\smallskip\hang
|cur_type=unknown_boolean| means that |cur_exp| points to a capsule
node that is in
a ring of equivalent booleans whose value has not yet been defined.

\smallskip\hang
|cur_type=string_type| means that |cur_exp| is a string number (i.e., an
integer in the range |0<=cur_exp<str_ptr|). That string's reference count
includes this particular reference.

\smallskip\hang
|cur_type=unknown_string| means that |cur_exp| points to a capsule
node that is in
a ring of equivalent strings whose value has not yet been defined.

\smallskip\hang
|cur_type=pen_type| means that |cur_exp| points to a pen header node. This
node contains a reference count, which takes account of this particular
reference.

\smallskip\hang
|cur_type=unknown_pen| means that |cur_exp| points to a capsule
node that is in
a ring of equivalent pens whose value has not yet been defined.

\smallskip\hang
|cur_type=future_pen| means that |cur_exp| points to a knot list that
should eventually be made into a pen. Nobody else points to this particular
knot list. The |future_pen| option occurs only as an output of |scan_primary|
and |scan_secondary|, not as an output of |scan_tertiary| or |scan_expression|.

\smallskip\hang
|cur_type=path_type| means that |cur_exp| points to the first node of
a path; nobody else points to this particular path. The control points of
the path will have been chosen.

\smallskip\hang
|cur_type=unknown_path| means that |cur_exp| points to a capsule
node that is in
a ring of equivalent paths whose value has not yet been defined.

\smallskip\hang
|cur_type=picture_type| means that |cur_exp| points to an edges header node.
Nobody else points to this particular set of edges.

\smallskip\hang
|cur_type=unknown_picture| means that |cur_exp| points to a capsule
node that is in
a ring of equivalent pictures whose value has not yet been defined.

\smallskip\hang
|cur_type=transform_type| means that |cur_exp| points to a |transform_type|
capsule node. The |value| part of this capsule
points to a transform node that contains six numeric values,
each of which is |independent|, |dependent|, |proto_dependent|, or |known|.

\smallskip\hang
|cur_type=pair_type| means that |cur_exp| points to a capsule
node whose type is |pair_type|. The |value| part of this capsule
points to a pair node that contains two numeric values,
each of which is |independent|, |dependent|, |proto_dependent|, or |known|.

\smallskip\hang
|cur_type=known| means that |cur_exp| is a |scaled| value.

\smallskip\hang
|cur_type=dependent| means that |cur_exp| points to a capsule node whose type
is |dependent|. The |dep_list| field in this capsule points to the associated
dependency list.

\smallskip\hang
|cur_type=proto_dependent| means that |cur_exp| points to a |proto_dependent|
capsule node. The |dep_list| field in this capsule
points to the associated dependency list.

\smallskip\hang
|cur_type=independent| means that |cur_exp| points to a capsule node
whose type is |independent|. This somewhat unusual case can arise, for
example, in the expression
`$x+\&{begingroup}\penalty0\,\&{string}\,x; 0\,\&{endgroup}$'.

\smallskip\hang
|cur_type=token_list| means that |cur_exp| points to a linked list of
tokens.

\smallskip\noindent
The possible settings of |cur_type| have been listed here in increasing
numerical order. Notice that |cur_type| will never be |numeric_type| or
|suffixed_macro| or |unsuffixed_macro|, although variables of those types
are allowed.  Conversely, \MF\ has no variables of type |vacuous| or
|token_list|.

@ Capsules are two-word nodes that have a similar meaning
to |cur_type| and |cur_exp|. Such nodes have |name_type=capsule|,
and their |type| field is one of the possibilities for |cur_type| listed above.
Also |link<=void| in capsules that aren't part of a token list.

The |value| field of a capsule is, in most cases, the value that
corresponds to its |type|, as |cur_exp| corresponds to |cur_type|.
However, when |cur_exp| would point to a capsule,
no extra layer of indirection is present; the |value|
field is what would have been called |value(cur_exp)| if it had not been
encapsulated.  Furthermore, if the type is |dependent| or
|proto_dependent|, the |value| field of a capsule is replaced by
|dep_list| and |prev_dep| fields, since dependency lists in capsules are
always part of the general |dep_list| structure.

The |get_x_next| routine is careful not to change the values of |cur_type|
and |cur_exp| when it gets an expanded token. However, |get_x_next| might
call a macro, which might parse an expression, which might execute lots of
commands in a group; hence it's possible that |cur_type| might change
from, say, |unknown_boolean| to |boolean_type|, or from |dependent| to
|known| or |independent|, during the time |get_x_next| is called. The
programs below are careful to stash sensitive intermediate results in
capsules, so that \MF's generality doesn't cause trouble.

Here's a procedure that illustrates these conventions. It takes
the contents of $(|cur_type|\kern-.3pt,|cur_exp|\kern-.3pt)$
and stashes them away in a
capsule. It is not used when |cur_type=token_list|.
After the operation, |cur_type=vacuous|; hence there is no need to
copy path lists or to update reference counts, etc.

The special link |void| is put on the capsule returned by
|stash_cur_exp|, because this procedure is used to store macro parameters
that must be easily distinguishable from token lists.

@<Declare the stashing/unstashing routines@>=
function stash_cur_exp:pointer;
var @!p:pointer; {the capsule that will be returned}
begin case cur_type of
unknown_types,transform_type,pair_type,dependent,proto_dependent,
  independent:p:=cur_exp;
othercases begin  p:=get_node(value_node_size); name_type(p):=capsule;
  type(p):=cur_type; value(p):=cur_exp;
  end
endcases;@/
cur_type:=vacuous; link(p):=void; stash_cur_exp:=p;
end;

@ The inverse of |stash_cur_exp| is the following procedure, which
deletes an unnecessary capsule and puts its contents into |cur_type|
and |cur_exp|.

The program steps of \MF\ can be divided into two categories: those in
which |cur_type| and |cur_exp| are ``alive'' and those in which they are
``dead,'' in the sense that |cur_type| and |cur_exp| contain relevant
information or not. It's important not to ignore them when they're alive,
and it's important not to pay attention to them when they're dead.

There's also an intermediate category: If |cur_type=vacuous|, then
|cur_exp| is irrelevant, hence we can proceed without caring if |cur_type|
and |cur_exp| are alive or dead. In such cases we say that |cur_type|
and |cur_exp| are {\sl dormant}. It is permissible to call |get_x_next|
only when they are alive or dormant.

The \\{stash} procedure above assumes that |cur_type| and |cur_exp|
are alive or dormant. The \\{unstash} procedure assumes that they are
dead or dormant; it resuscitates them.

@<Declare the stashing/unstashing...@>=
procedure unstash_cur_exp(@!p:pointer);
begin cur_type:=type(p);
case cur_type of
unknown_types,transform_type,pair_type,dependent,proto_dependent,
  independent: cur_exp:=p;
othercases begin cur_exp:=value(p);
  free_node(p,value_node_size);
  end
endcases;@/
end;

@ The following procedure prints the values of expressions in an
abbreviated format. If its first parameter |p| is null, the value of
|(cur_type,cur_exp)| is displayed; otherwise |p| should be a capsule
containing the desired value. The second parameter controls the amount of
output. If it is~0, dependency lists will be abbreviated to
`\.{linearform}' unless they consist of a single term.  If it is greater
than~1, complicated structures (pens, pictures, and paths) will be displayed
in full.
@.linearform@>

@<Declare subroutines for printing expressions@>=
@t\4@>@<Declare the procedure called |print_dp|@>@;
@t\4@>@<Declare the stashing/unstashing routines@>@;
procedure print_exp(@!p:pointer;@!verbosity:small_number);
var @!restore_cur_exp:boolean; {should |cur_exp| be restored?}
@!t:small_number; {the type of the expression}
@!v:integer; {the value of the expression}
@!q:pointer; {a big node being displayed}
begin if p<>null then restore_cur_exp:=false
else  begin p:=stash_cur_exp; restore_cur_exp:=true;
  end;
t:=type(p);
if t<dependent then v:=value(p)@+else if t<independent then v:=dep_list(p);
@<Print an abbreviated value of |v| with format depending on |t|@>;
if restore_cur_exp then unstash_cur_exp(p);
end;

@ @<Print an abbreviated value of |v| with format depending on |t|@>=
case t of
vacuous:print("vacuous");
boolean_type:if v=true_code then print("true")@+else print("false");
unknown_types,numeric_type:@<Display a variable
  that's been declared but not defined@>;
string_type:begin print_char(""""); slow_print(v); print_char("""");
  end;
pen_type,future_pen,path_type,picture_type:@<Display a complex type@>;
transform_type,pair_type:if v=null then print_type(t)
  else @<Display a big node@>;
known:print_scaled(v);
dependent,proto_dependent:print_dp(t,v,verbosity);
independent:print_variable_name(p);
othercases confusion("exp")
@:this can't happen exp}{\quad exp@>
endcases

@ @<Display a big node@>=
begin print_char("("); q:=v+big_node_size[t];
repeat if type(v)=known then print_scaled(value(v))
else if type(v)=independent then print_variable_name(v)
else print_dp(type(v),dep_list(v),verbosity);
v:=v+2;
if v<>q then print_char(",");
until v=q;
print_char(")");
end

@ Values of type \&{picture}, \&{path}, and \&{pen} are displayed verbosely
in the log file only, unless the user has given a positive value to
\\{tracingonline}.

@<Display a complex type@>=
if verbosity<=1 then print_type(t)
else  begin if selector=term_and_log then
   if internal[tracing_online]<=0 then
    begin selector:=term_only;
    print_type(t); print(" (see the transcript file)");
    selector:=term_and_log;
    end;
  case t of
  pen_type:print_pen(v,"",false);
  future_pen:print_path(v," (future pen)",false);
  path_type:print_path(v,"",false);
  picture_type:begin cur_edges:=v; print_edges("",false,0,0);
    end;
  end; {there are no other cases}
  end

@ @<Declare the procedure called |print_dp|@>=
procedure print_dp(@!t:small_number;@!p:pointer;@!verbosity:small_number);
var @!q:pointer; {the node following |p|}
begin q:=link(p);
if (info(q)=null) or (verbosity>0) then print_dependency(p,t)
else print("linearform");
@.linearform@>
end;

@ The displayed name of a variable in a ring will not be a capsule unless
the ring consists entirely of capsules.

@<Display a variable that's been declared but not defined@>=
begin print_type(t);
if v<>null then
  begin print_char(" ");
  while (name_type(v)=capsule) and (v<>p) do v:=value(v);
  print_variable_name(v);
  end;
end

@ When errors are detected during parsing, it is often helpful to
display an expression just above the error message, using |exp_err|
or |disp_err| instead of |print_err|.

@d exp_err(#)==disp_err(null,#) {displays the current expression}

@<Declare subroutines for printing expressions@>=
procedure disp_err(@!p:pointer;@!s:str_number);
begin if interaction=error_stop_mode then wake_up_terminal;
print_nl(">> ");
@.>>@>
print_exp(p,1); {``medium verbose'' printing of the expression}
if s<>"" then
  begin print_nl("! "); print(s);
@.!\relax@>
  end;
end;

@ If |cur_type| and |cur_exp| contain relevant information that should
be recycled, we will use the following procedure, which changes |cur_type|
to |known| and stores a given value in |cur_exp|. We can think of |cur_type|
and |cur_exp| as either alive or dormant after this has been done,
because |cur_exp| will not contain a pointer value.

@<Declare the procedure called |flush_cur_exp|@>=
procedure flush_cur_exp(@!v:scaled);
begin case cur_type of
unknown_types,transform_type,pair_type,@|dependent,proto_dependent,independent:
  begin recycle_value(cur_exp); free_node(cur_exp,value_node_size);
  end;
pen_type: delete_pen_ref(cur_exp);
string_type:delete_str_ref(cur_exp);
future_pen,path_type: toss_knot_list(cur_exp);
picture_type:toss_edges(cur_exp);
othercases do_nothing
endcases;@/
cur_type:=known; cur_exp:=v;
end;

@ There's a much more general procedure that is capable of releasing
the storage associated with any two-word value packet.

@<Declare the recycling subroutines@>=
procedure recycle_value(@!p:pointer);
label done;
var @!t:small_number; {a type code}
@!v:integer; {a value}
@!vv:integer; {another value}
@!q,@!r,@!s,@!pp:pointer; {link manipulation registers}
begin t:=type(p);
if t<dependent then v:=value(p);
case t of
undefined,vacuous,boolean_type,known,numeric_type:do_nothing;
unknown_types:ring_delete(p);
string_type:delete_str_ref(v);
pen_type:delete_pen_ref(v);
path_type,future_pen:toss_knot_list(v);
picture_type:toss_edges(v);
pair_type,transform_type:@<Recycle a big node@>;
dependent,proto_dependent:@<Recycle a dependency list@>;
independent:@<Recycle an independent variable@>;
token_list,structured:confusion("recycle");
@:this can't happen recycle}{\quad recycle@>
unsuffixed_macro,suffixed_macro:delete_mac_ref(value(p));
end; {there are no other cases}
type(p):=undefined;
end;

@ @<Recycle a big node@>=
if v<>null then
  begin q:=v+big_node_size[t];
  repeat q:=q-2; recycle_value(q);
  until q=v;
  free_node(v,big_node_size[t]);
  end

@ @<Recycle a dependency list@>=
begin q:=dep_list(p);
while info(q)<>null do q:=link(q);
link(prev_dep(p)):=link(q);
prev_dep(link(q)):=prev_dep(p);
link(q):=null; flush_node_list(dep_list(p));
end

@ When an independent variable disappears, it simply fades away, unless
something depends on it. In the latter case, a dependent variable whose
coefficient of dependence is maximal will take its place.
The relevant algorithm is due to Ignacio~A. Zabala, who implemented it
as part of his Ph.D. thesis (Stanford University, December 1982).
@^Zabala Salelles, Ignacio Andr\'es@>

For example, suppose that variable $x$ is being recycled, and that the
only variables depending on~$x$ are $y=2x+a$ and $z=x+b$. In this case
we want to make $y$ independent and $z=.5y-.5a+b$; no other variables
will depend on~$y$. If $\\{tracingequations}>0$ in this situation,
we will print `\.{\#\#\# -2x=-y+a}'.

There's a slight complication, however: An independent variable $x$
can occur both in dependency lists and in proto-dependency lists.
This makes it necessary to be careful when deciding which coefficient
is maximal.

Furthermore, this complication is not so slight when
a proto-dependent variable is chosen to become independent. For example,
suppose that $y=2x+100a$ is proto-dependent while $z=x+b$ is dependent;
then we must change $z=.5y-50a+b$ to a proto-dependency, because of the
large coefficient `50'.

In order to deal with these complications without wasting too much time,
we shall link together the occurrences of~$x$ among all the linear
dependencies, maintaining separate lists for the dependent and
proto-dependent cases.

@<Recycle an independent variable@>=
begin max_c[dependent]:=0; max_c[proto_dependent]:=0;@/
max_link[dependent]:=null; max_link[proto_dependent]:=null;@/
q:=link(dep_head);
while q<>dep_head do
  begin s:=value_loc(q); {now |link(s)=dep_list(q)|}
  loop@+  begin r:=link(s);
    if info(r)=null then goto done;
    if info(r)<>p then s:=r
    else  begin t:=type(q); link(s):=link(r); info(r):=q;
      if abs(value(r))>max_c[t] then
        @<Record a new maximum coefficient of type |t|@>
      else  begin link(r):=max_link[t]; max_link[t]:=r;
        end;
      end;
    end;
done:  q:=link(r);
  end;
if (max_c[dependent]>0)or(max_c[proto_dependent]>0) then
  @<Choose a dependent variable to take the place of the disappearing
    independent variable, and change all remaining dependencies
    accordingly@>;
end

@ The code for independency removal makes use of three two-word arrays.

@<Glob...@>=
@!max_c:array[dependent..proto_dependent] of integer;
  {max coefficient magnitude}
@!max_ptr:array[dependent..proto_dependent] of pointer;
  {where |p| occurs with |max_c|}
@!max_link:array[dependent..proto_dependent] of pointer;
  {other occurrences of |p|}

@ @<Record a new maximum coefficient...@>=
begin if max_c[t]>0 then
  begin link(max_ptr[t]):=max_link[t]; max_link[t]:=max_ptr[t];
  end;
max_c[t]:=abs(value(r)); max_ptr[t]:=r;
end

@ @<Choose a dependent...@>=
begin if (max_c[dependent] div @'10000 >=
          max_c[proto_dependent]) then
  t:=dependent
else t:=proto_dependent;
@<Determine the dependency list |s| to substitute for the independent
  variable~|p|@>;
t:=dependent+proto_dependent-t; {complement |t|}
if max_c[t]>0 then {we need to pick up an unchosen dependency}
  begin link(max_ptr[t]):=max_link[t]; max_link[t]:=max_ptr[t];
  end;
if t<>dependent then @<Substitute new dependencies in place of |p|@>
else @<Substitute new proto-dependencies in place of |p|@>;
flush_node_list(s);
if fix_needed then fix_dependencies;
check_arith;
end

@ Let |s=max_ptr[t]|. At this point we have $|value|(s)=\pm|max_c|[t]$,
and |info(s)| points to the dependent variable~|pp| of type~|t| from
whose dependency list we have removed node~|s|. We must reinsert
node~|s| into the dependency list, with coefficient $-1.0$, and with
|pp| as the new independent variable. Since |pp| will have a larger serial
number than any other variable, we can put node |s| at the head of the
list.

@<Determine the dep...@>=
s:=max_ptr[t]; pp:=info(s); v:=value(s);
if t=dependent then value(s):=-fraction_one@+else value(s):=-unity;
r:=dep_list(pp); link(s):=r;
while info(r)<>null do r:=link(r);
q:=link(r); link(r):=null;
prev_dep(q):=prev_dep(pp); link(prev_dep(pp)):=q;
new_indep(pp);
if cur_exp=pp then if cur_type=t then cur_type:=independent;
if internal[tracing_equations]>0 then @<Show the transformed dependency@>

@ Now $(-v)$ times the formerly independent variable~|p| is being replaced
by the dependency list~|s|.

@<Show the transformed...@>=
if interesting(p) then
  begin begin_diagnostic; print_nl("### ");
@:]]]\#\#\#_}{\.{\#\#\#}@>
  if v>0 then print_char("-");
  if t=dependent then vv:=round_fraction(max_c[dependent])
  else vv:=max_c[proto_dependent];
  if vv<>unity then print_scaled(vv);
  print_variable_name(p);
  while value(p) mod s_scale>0 do
    begin print("*4"); value(p):=value(p)-2;
    end;
  if t=dependent then print_char("=")@+else print(" = ");
  print_dependency(s,t);
  end_diagnostic(false);
  end

@ Finally, there are dependent and proto-dependent variables whose
dependency lists must be brought up to date.

@<Substitute new dependencies...@>=
for t:=dependent to proto_dependent do
  begin r:=max_link[t];
  while r<>null do
    begin q:=info(r);
    dep_list(q):=p_plus_fq(dep_list(q),@|
     make_fraction(value(r),-v),s,t,dependent);
    if dep_list(q)=dep_final then make_known(q,dep_final);
    q:=r; r:=link(r); free_node(q,dep_node_size);
    end;
  end

@ @<Substitute new proto...@>=
for t:=dependent to proto_dependent do
  begin r:=max_link[t];
  while r<>null do
    begin q:=info(r);
    if t=dependent then {for safety's sake, we change |q| to |proto_dependent|}
      begin if cur_exp=q then if cur_type=dependent then
        cur_type:=proto_dependent;
      dep_list(q):=p_over_v(dep_list(q),unity,dependent,proto_dependent);
      type(q):=proto_dependent; value(r):=round_fraction(value(r));
      end;
    dep_list(q):=p_plus_fq(dep_list(q),@|
     make_scaled(value(r),-v),s,proto_dependent,proto_dependent);
    if dep_list(q)=dep_final then make_known(q,dep_final);
    q:=r; r:=link(r); free_node(q,dep_node_size);
    end;
  end

@ Here are some routines that provide handy combinations of actions
that are often needed during error recovery. For example,
`|flush_error|' flushes the current expression, replaces it by
a given value, and calls |error|.

Errors often are detected after an extra token has already been scanned.
The `\\{put\_get}' routines put that token back before calling |error|;
then they get it back again. (Or perhaps they get another token, if
the user has changed things.)

@<Declare the procedure called |flush_cur_exp|@>=
procedure flush_error(@!v:scaled);@+begin error; flush_cur_exp(v);@+end;
@#
procedure@?back_error; forward;@t\2@>@/
procedure@?get_x_next; forward;@t\2@>@/
@#
procedure put_get_error;@+begin back_error; get_x_next;@+end;
@#
procedure put_get_flush_error(@!v:scaled);@+begin put_get_error;
 flush_cur_exp(v);@+end;

@ A global variable called |var_flag| is set to a special command code
just before \MF\ calls |scan_expression|, if the expression should be
treated as a variable when this command code immediately follows. For
example, |var_flag| is set to |assignment| at the beginning of a
statement, because we want to know the {\sl location\/} of a variable at
the left of `\.{:=}', not the {\sl value\/} of that variable.

The |scan_expression| subroutine calls |scan_tertiary|,
which calls |scan_secondary|, which calls |scan_primary|, which sets
|var_flag:=0|. In this way each of the scanning routines ``knows''
when it has been called with a special |var_flag|, but |var_flag| is
usually zero.

A variable preceding a command that equals |var_flag| is converted to a
token list rather than a value. Furthermore, an `\.{=}' sign following an
expression with |var_flag=assignment| is not considered to be a relation
that produces boolean expressions.


@<Glob...@>=
@!var_flag:0..max_command_code; {command that wants a variable}

@ @<Set init...@>=
var_flag:=0;

@* \[40] Parsing primary expressions.
The first parsing routine, |scan_primary|, is also the most complicated one,
since it involves so many different cases. But each case---with one
exception---is fairly simple by itself.

When |scan_primary| begins, the first token of the primary to be scanned
should already appear in |cur_cmd|, |cur_mod|, and |cur_sym|. The values
of |cur_type| and |cur_exp| should be either dead or dormant, as explained
earlier. If |cur_cmd| is not between |min_primary_command| and
|max_primary_command|, inclusive, a syntax error will be signalled.

@<Declare the basic parsing subroutines@>=
procedure scan_primary;
label restart, done, done1, done2;
var @!p,@!q,@!r:pointer; {for list manipulation}
@!c:quarterword; {a primitive operation code}
@!my_var_flag:0..max_command_code; {initial value of |var_flag|}
@!l_delim,@!r_delim:pointer; {hash addresses of a delimiter pair}
@<Other local variables for |scan_primary|@>@;
begin my_var_flag:=var_flag; var_flag:=0;
restart:check_arith;
@<Supply diagnostic information, if requested@>;
case cur_cmd of
left_delimiter:@<Scan a delimited primary@>;
begin_group:@<Scan a grouped primary@>;
string_token:@<Scan a string constant@>;
numeric_token:@<Scan a primary that starts with a numeric token@>;
nullary:@<Scan a nullary operation@>;
unary,type_name,cycle,plus_or_minus:@<Scan a unary operation@>;
primary_binary:@<Scan a binary operation with `\&{of}' between its operands@>;
str_op:@<Convert a suffix to a string@>;
internal_quantity:@<Scan an internal numeric quantity@>;
capsule_token:make_exp_copy(cur_mod);
tag_token:@<Scan a variable primary;
  |goto restart| if it turns out to be a macro@>;
othercases begin bad_exp("A primary"); goto restart;
@.A primary expression...@>
  end
endcases;@/
get_x_next; {the routines |goto done| if they don't want this}
done: if cur_cmd=left_bracket then
  if cur_type>=known then @<Scan a mediation construction@>;
end;

@ Errors at the beginning of expressions are flagged by |bad_exp|.

@p procedure bad_exp(@!s:str_number);
var save_flag:0..max_command_code;
begin print_err(s); print(" expression can't begin with `");
print_cmd_mod(cur_cmd,cur_mod); print_char("'");
help4("I'm afraid I need some sort of value in order to continue,")@/
  ("so I've tentatively inserted `0'. You may want to")@/
  ("delete this zero and insert something else;")@/
  ("see Chapter 27 of The METAFONTbook for an example.");
@:METAFONTbook}{\sl The {\logos METAFONT\/}book@>
back_input; cur_sym:=0; cur_cmd:=numeric_token; cur_mod:=0; ins_error;@/
save_flag:=var_flag; var_flag:=0; get_x_next;
var_flag:=save_flag;
end;

@ @<Supply diagnostic information, if requested@>=
debug if panicking then check_mem(false);@+gubed@;@/
if interrupt<>0 then if OK_to_interrupt then
  begin back_input; check_interrupt; get_x_next;
  end

@ @<Scan a delimited primary@>=
begin l_delim:=cur_sym; r_delim:=cur_mod; get_x_next; scan_expression;
if (cur_cmd=comma) and (cur_type>=known) then
  @<Scan the second of a pair of numerics@>
else check_delimiter(l_delim,r_delim);
end

@ The |stash_in| subroutine puts the current (numeric) expression into a field
within a ``big node.''

@p procedure stash_in(@!p:pointer);
var @!q:pointer; {temporary register}
begin type(p):=cur_type;
if cur_type=known then value(p):=cur_exp
else  begin if cur_type=independent then
    @<Stash an independent |cur_exp| into a big node@>
  else  begin mem[value_loc(p)]:=mem[value_loc(cur_exp)];
     {|dep_list(p):=dep_list(cur_exp)| and |prev_dep(p):=prev_dep(cur_exp)|}
    link(prev_dep(p)):=p;
    end;
  free_node(cur_exp,value_node_size);
  end;
cur_type:=vacuous;
end;

@ In rare cases the current expression can become |independent|. There
may be many dependency lists pointing to such an independent capsule,
so we can't simply move it into place within a big node. Instead,
we copy it, then recycle it.

@ @<Stash an independent |cur_exp|...@>=
begin q:=single_dependency(cur_exp);
if q=dep_final then
  begin type(p):=known; value(p):=0; free_node(q,dep_node_size);
  end
else  begin type(p):=dependent; new_dep(p,q);
  end;
recycle_value(cur_exp);
end

@ @<Scan the second of a pair of numerics@>=
begin p:=get_node(value_node_size); type(p):=pair_type; name_type(p):=capsule;
init_big_node(p); q:=value(p); stash_in(x_part_loc(q));@/
get_x_next; scan_expression;
if cur_type<known then
  begin exp_err("Nonnumeric ypart has been replaced by 0");
@.Nonnumeric...replaced by 0@>
  help4("I thought you were giving me a pair `(x,y)'; but")@/
    ("after finding a nice xpart `x' I found a ypart `y'")@/
    ("that isn't of numeric type. So I've changed y to zero.")@/
    ("(The y that I didn't like appears above the error message.)");
  put_get_flush_error(0);
  end;
stash_in(y_part_loc(q));
check_delimiter(l_delim,r_delim);
cur_type:=pair_type; cur_exp:=p;
end

@ The local variable |group_line| keeps track of the line
where a \&{begingroup} command occurred; this will be useful
in an error message if the group doesn't actually end.

@<Other local variables for |scan_primary|@>=
@!group_line:integer; {where a group began}

@ @<Scan a grouped primary@>=
begin group_line:=line;
if internal[tracing_commands]>0 then show_cur_cmd_mod;
save_boundary_item(p);
repeat do_statement; {ends with |cur_cmd>=semicolon|}
until cur_cmd<>semicolon;
if cur_cmd<>end_group then
  begin print_err("A group begun on line ");
@.A group...never ended@>
  print_int(group_line);
  print(" never ended");
  help2("I saw a `begingroup' back there that hasn't been matched")@/
    ("by `endgroup'. So I've inserted `endgroup' now.");
  back_error; cur_cmd:=end_group;
  end;
unsave; {this might change |cur_type|, if independent variables are recycled}
if internal[tracing_commands]>0 then show_cur_cmd_mod;
end

@ @<Scan a string constant@>=
begin cur_type:=string_type; cur_exp:=cur_mod;
end

@ Later we'll come to procedures that perform actual operations like
addition, square root, and so on; our purpose now is to do the parsing.
But we might as well mention those future procedures now, so that the
suspense won't be too bad:

\smallskip
|do_nullary(c)| does primitive operations that have no operands (e.g.,
`\&{true}' or `\&{pencircle}');

\smallskip
|do_unary(c)| applies a primitive operation to the current expression;

\smallskip
|do_binary(p,c)| applies a primitive operation to the capsule~|p|
and the current expression.

@<Scan a nullary operation@>=do_nullary(cur_mod)

@ @<Scan a unary operation@>=
begin c:=cur_mod; get_x_next; scan_primary; do_unary(c); goto done;
end

@ A numeric token might be a primary by itself, or it might be the
numerator of a fraction composed solely of numeric tokens, or it might
multiply the primary that follows (provided that the primary doesn't begin
with a plus sign or a minus sign). The code here uses the facts that
|max_primary_command=plus_or_minus| and
|max_primary_command-1=numeric_token|. If a fraction is found that is less
than unity, we try to retain higher precision when we use it in scalar
multiplication.

@<Other local variables for |scan_primary|@>=
@!num,@!denom:scaled; {for primaries that are fractions, like `1/2'}

@ @<Scan a primary that starts with a numeric token@>=
begin cur_exp:=cur_mod; cur_type:=known; get_x_next;
if cur_cmd<>slash then
  begin num:=0; denom:=0;
  end
else  begin get_x_next;
  if cur_cmd<>numeric_token then
    begin back_input;
    cur_cmd:=slash; cur_mod:=over; cur_sym:=frozen_slash;
    goto done;
    end;
  num:=cur_exp; denom:=cur_mod;
  if denom=0 then @<Protest division by zero@>
  else cur_exp:=make_scaled(num,denom);
  check_arith; get_x_next;
  end;
if cur_cmd>=min_primary_command then
 if cur_cmd<numeric_token then {in particular, |cur_cmd<>plus_or_minus|}
  begin p:=stash_cur_exp; scan_primary;
  if (abs(num)>=abs(denom))or(cur_type<pair_type) then do_binary(p,times)
  else  begin frac_mult(num,denom);
    free_node(p,value_node_size);
    end;
  end;
goto done;
end

@ @<Protest division...@>=
begin print_err("Division by zero");
@.Division by zero@>
help1("I'll pretend that you meant to divide by 1."); error;
end

@ @<Scan a binary operation with `\&{of}' between its operands@>=
begin c:=cur_mod; get_x_next; scan_expression;
if cur_cmd<>of_token then
  begin missing_err("of"); print(" for "); print_cmd_mod(primary_binary,c);
@.Missing `of'@>
  help1("I've got the first argument; will look now for the other.");
  back_error;
  end;
p:=stash_cur_exp; get_x_next; scan_primary; do_binary(p,c); goto done;
end

@ @<Convert a suffix to a string@>=
begin get_x_next; scan_suffix; old_setting:=selector; selector:=new_string;
show_token_list(cur_exp,null,100000,0); flush_token_list(cur_exp);
cur_exp:=make_string; selector:=old_setting; cur_type:=string_type;
goto done;
end

@ If an internal quantity appears all by itself on the left of an
assignment, we return a token list of length one, containing the address
of the internal quantity plus |hash_end|. (This accords with the conventions
of the save stack, as described earlier.)

@<Scan an internal...@>=
begin q:=cur_mod;
if my_var_flag=assignment then
  begin get_x_next;
  if cur_cmd=assignment then
    begin cur_exp:=get_avail;
    info(cur_exp):=q+hash_end; cur_type:=token_list; goto done;
    end;
  back_input;
  end;
cur_type:=known; cur_exp:=internal[q];
end

@ The most difficult part of |scan_primary| has been saved for last, since
it was necessary to build up some confidence first. We can now face the task
of scanning a variable.

As we scan a variable, we build a token list containing the relevant
names and subscript values, simultaneously following along in the
``collective'' structure to see if we are actually dealing with a macro
instead of a value.

The local variables |pre_head| and |post_head| will point to the beginning
of the prefix and suffix lists; |tail| will point to the end of the list
that is currently growing.

Another local variable, |tt|, contains partial information about the
declared type of the variable-so-far. If |tt>=unsuffixed_macro|, the
relation |tt=type(q)| will always hold. If |tt=undefined|, the routine
doesn't bother to update its information about type. And if
|undefined<tt<unsuffixed_macro|, the precise value of |tt| isn't critical.

@ @<Other local variables for |scan_primary|@>=
@!pre_head,@!post_head,@!tail:pointer;
  {prefix and suffix list variables}
@!tt:small_number; {approximation to the type of the variable-so-far}
@!t:pointer; {a token}
@!macro_ref:pointer; {reference count for a suffixed macro}

@ @<Scan a variable primary...@>=
begin fast_get_avail(pre_head); tail:=pre_head; post_head:=null; tt:=vacuous;
loop@+  begin t:=cur_tok; link(tail):=t;
  if tt<>undefined then
    begin @<Find the approximate type |tt| and corresponding~|q|@>;
    if tt>=unsuffixed_macro then
      @<Either begin an unsuffixed macro call or
        prepare for a suffixed one@>;
    end;
  get_x_next; tail:=t;
  if cur_cmd=left_bracket then
    @<Scan for a subscript; replace |cur_cmd| by |numeric_token| if found@>;
  if cur_cmd>max_suffix_token then goto done1;
  if cur_cmd<min_suffix_token then goto done1;
  end; {now |cur_cmd| is |internal_quantity|, |tag_token|, or |numeric_token|}
done1:@<Handle unusual cases that masquerade as variables, and |goto restart|
  or |goto done| if appropriate;
  otherwise make a copy of the variable and |goto done|@>;
end

@ @<Either begin an unsuffixed macro call or...@>=
begin link(tail):=null;
if tt>unsuffixed_macro then {|tt=suffixed_macro|}
  begin post_head:=get_avail; tail:=post_head; link(tail):=t;@/
  tt:=undefined; macro_ref:=value(q); add_mac_ref(macro_ref);
  end
else @<Set up unsuffixed macro call and |goto restart|@>;
end

@ @<Scan for a subscript; replace |cur_cmd| by |numeric_token| if found@>=
begin get_x_next; scan_expression;
if cur_cmd<>right_bracket then
  @<Put the left bracket and the expression back to be rescanned@>
else  begin if cur_type<>known then bad_subscript;
  cur_cmd:=numeric_token; cur_mod:=cur_exp; cur_sym:=0;
  end;
end

@ The left bracket that we thought was introducing a subscript might have
actually been the left bracket in a mediation construction like `\.{x[a,b]}'.
So we don't issue an error message at this point; but we do want to back up
so as to avoid any embarrassment about our incorrect assumption.

@<Put the left bracket and the expression back to be rescanned@>=
begin back_input; {that was the token following the current expression}
back_expr; cur_cmd:=left_bracket; cur_mod:=0; cur_sym:=frozen_left_bracket;
end

@ Here's a routine that puts the current expression back to be read again.

@p procedure back_expr;
var @!p:pointer; {capsule token}
begin p:=stash_cur_exp; link(p):=null; back_list(p);
end;

@ Unknown subscripts lead to the following error message.

@p procedure bad_subscript;
begin exp_err("Improper subscript has been replaced by zero");
@.Improper subscript...@>
help3("A bracketed subscript must have a known numeric value;")@/
  ("unfortunately, what I found was the value that appears just")@/
  ("above this error message. So I'll try a zero subscript.");
flush_error(0);
end;

@ Every time we call |get_x_next|, there's a chance that the variable we've
been looking at will disappear. Thus, we cannot safely keep |q| pointing
into the variable structure; we need to start searching from the root each time.

@<Find the approximate type |tt| and corresponding~|q|@>=
@^inner loop@>
begin p:=link(pre_head); q:=info(p); tt:=undefined;
if eq_type(q) mod outer_tag=tag_token then
  begin q:=equiv(q);
  if q=null then goto done2;
  loop@+  begin p:=link(p);
    if p=null then
      begin tt:=type(q); goto done2;
      end;
    if type(q)<>structured then goto done2;
    q:=link(attr_head(q)); {the |collective_subscript| attribute}
    if p>=hi_mem_min then {it's not a subscript}
      begin repeat q:=link(q);
      until attr_loc(q)>=info(p);
      if attr_loc(q)>info(p) then goto done2;
      end;
    end;
  end;
done2:end

@ How do things stand now? Well, we have scanned an entire variable name,
including possible subscripts and/or attributes; |cur_cmd|, |cur_mod|, and
|cur_sym| represent the token that follows. If |post_head=null|, a
token list for this variable name starts at |link(pre_head)|, with all
subscripts evaluated. But if |post_head<>null|, the variable turned out
to be a suffixed macro; |pre_head| is the head of the prefix list, while
|post_head| is the head of a token list containing both `\.{\AT!}' and
the suffix.

Our immediate problem is to see if this variable still exists. (Variable
structures can change drastically whenever we call |get_x_next|; users
aren't supposed to do this, but the fact that it is possible means that
we must be cautious.)

The following procedure prints an error message when a variable
unexpectedly disappears. Its help message isn't quite right for
our present purposes, but we'll be able to fix that up.

@p procedure obliterated(@!q:pointer);
begin print_err("Variable "); show_token_list(q,null,1000,0);
print(" has been obliterated");
@.Variable...obliterated@>
help5("It seems you did a nasty thing---probably by accident,")@/
  ("but nevertheless you nearly hornswoggled me...")@/
  ("While I was evaluating the right-hand side of this")@/
  ("command, something happened, and the left-hand side")@/
  ("is no longer a variable! So I won't change anything.");
end;

@ If the variable does exist, we also need to check
for a few other special cases before deciding that a plain old ordinary
variable has, indeed, been scanned.

@<Handle unusual cases that masquerade as variables...@>=
if post_head<>null then @<Set up suffixed macro call and |goto restart|@>;
q:=link(pre_head); free_avail(pre_head);
if cur_cmd=my_var_flag then
  begin cur_type:=token_list; cur_exp:=q; goto done;
  end;
p:=find_variable(q);
if p<>null then make_exp_copy(p)
else  begin obliterated(q);@/
  help_line[2]:="While I was evaluating the suffix of this variable,";
  help_line[1]:="something was redefined, and it's no longer a variable!";
  help_line[0]:="In order to get back on my feet, I've inserted `0' instead.";
  put_get_flush_error(0);
  end;
flush_node_list(q); goto done

@ The only complication associated with macro calling is that the prefix
and ``at'' parameters must be packaged in an appropriate list of lists.

@<Set up unsuffixed macro call and |goto restart|@>=
begin p:=get_avail; info(pre_head):=link(pre_head); link(pre_head):=p;
info(p):=t; macro_call(value(q),pre_head,null); get_x_next; goto restart;
end

@ If the ``variable'' that turned out to be a suffixed macro no longer exists,
we don't care, because we have reserved a pointer (|macro_ref|) to its
token list.

@<Set up suffixed macro call and |goto restart|@>=
begin back_input; p:=get_avail; q:=link(post_head);
info(pre_head):=link(pre_head); link(pre_head):=post_head;
info(post_head):=q; link(post_head):=p; info(p):=link(q); link(q):=null;
macro_call(macro_ref,pre_head,null); decr(ref_count(macro_ref));
get_x_next; goto restart;
end

@ Our remaining job is simply to make a copy of the value that has been
found. Some cases are harder than others, but complexity arises solely
because of the multiplicity of possible cases.

@<Declare the procedure called |make_exp_copy|@>=
@t\4@>@<Declare subroutines needed by |make_exp_copy|@>@;
procedure make_exp_copy(@!p:pointer);
label restart;
var @!q,@!r,@!t:pointer; {registers for list manipulation}
begin restart: cur_type:=type(p);
case cur_type of
vacuous,boolean_type,known:cur_exp:=value(p);
unknown_types:cur_exp:=new_ring_entry(p);
string_type:begin cur_exp:=value(p); add_str_ref(cur_exp);
  end;
pen_type:begin cur_exp:=value(p); add_pen_ref(cur_exp);
  end;
picture_type:cur_exp:=copy_edges(value(p));
path_type,future_pen:cur_exp:=copy_path(value(p));
transform_type,pair_type:@<Copy the big node |p|@>;
dependent,proto_dependent:encapsulate(copy_dep_list(dep_list(p)));
numeric_type:begin new_indep(p); goto restart;
  end;
independent: begin q:=single_dependency(p);
  if q=dep_final then
    begin cur_type:=known; cur_exp:=0; free_node(q,dep_node_size);
    end
  else  begin cur_type:=dependent; encapsulate(q);
    end;
  end;
othercases confusion("copy")
@:this can't happen copy}{\quad copy@>
endcases;
end;

@ The |encapsulate| subroutine assumes that |dep_final| is the
tail of dependency list~|p|.

@<Declare subroutines needed by |make_exp_copy|@>=
procedure encapsulate(@!p:pointer);
begin cur_exp:=get_node(value_node_size); type(cur_exp):=cur_type;
name_type(cur_exp):=capsule; new_dep(cur_exp,p);
end;

@ The most tedious case arises when the user refers to a
\&{pair} or \&{transform} variable; we must copy several fields,
each of which can be |independent|, |dependent|, |proto_dependent|,
or |known|.

@<Copy the big node |p|@>=
begin if value(p)=null then init_big_node(p);
t:=get_node(value_node_size); name_type(t):=capsule; type(t):=cur_type;
init_big_node(t);@/
q:=value(p)+big_node_size[cur_type]; r:=value(t)+big_node_size[cur_type];
repeat q:=q-2; r:=r-2; install(r,q);
until q=value(p);
cur_exp:=t;
end

@ The |install| procedure copies a numeric field~|q| into field~|r| of
a big node that will be part of a capsule.

@<Declare subroutines needed by |make_exp_copy|@>=
procedure install(@!r,@!q:pointer);
var p:pointer; {temporary register}
begin if type(q)=known then
  begin value(r):=value(q); type(r):=known;
  end
else  if type(q)=independent then
    begin p:=single_dependency(q);
    if p=dep_final then
      begin type(r):=known; value(r):=0; free_node(p,dep_node_size);
      end
    else  begin type(r):=dependent; new_dep(r,p);
      end;
    end
  else  begin type(r):=type(q); new_dep(r,copy_dep_list(dep_list(q)));
    end;
end;

@ Expressions of the form `\.{a[b,c]}' are converted into
`\.{b+a*(c-b)}', without checking the types of \.b~or~\.c,
provided that \.a is numeric.

@<Scan a mediation...@>=
begin p:=stash_cur_exp; get_x_next; scan_expression;
if cur_cmd<>comma then
  begin @<Put the left bracket and the expression back...@>;
  unstash_cur_exp(p);
  end
else  begin q:=stash_cur_exp; get_x_next; scan_expression;
  if cur_cmd<>right_bracket then
    begin missing_err("]");@/
@.Missing `]'@>
    help3("I've scanned an expression of the form `a[b,c',")@/
      ("so a right bracket should have come next.")@/
      ("I shall pretend that one was there.");@/
    back_error;
    end;
  r:=stash_cur_exp; make_exp_copy(q);@/
  do_binary(r,minus); do_binary(p,times); do_binary(q,plus); get_x_next;
  end;
end

@ Here is a comparatively simple routine that is used to scan the
\&{suffix} parameters of a macro.

@<Declare the basic parsing subroutines@>=
procedure scan_suffix;
label done;
var @!h,@!t:pointer; {head and tail of the list being built}
@!p:pointer; {temporary register}
begin h:=get_avail; t:=h;
loop@+  begin if cur_cmd=left_bracket then
    @<Scan a bracketed subscript and set |cur_cmd:=numeric_token|@>;
  if cur_cmd=numeric_token then p:=new_num_tok(cur_mod)
  else if (cur_cmd=tag_token)or(cur_cmd=internal_quantity) then
    begin p:=get_avail; info(p):=cur_sym;
    end
  else goto done;
  link(t):=p; t:=p; get_x_next;
  end;
done: cur_exp:=link(h); free_avail(h); cur_type:=token_list;
end;

@ @<Scan a bracketed subscript and set |cur_cmd:=numeric_token|@>=
begin get_x_next; scan_expression;
if cur_type<>known then bad_subscript;
if cur_cmd<>right_bracket then
  begin missing_err("]");@/
@.Missing `]'@>
  help3("I've seen a `[' and a subscript value, in a suffix,")@/
    ("so a right bracket should have come next.")@/
    ("I shall pretend that one was there.");@/
  back_error;
  end;
cur_cmd:=numeric_token; cur_mod:=cur_exp;
end

@* \[41] Parsing secondary and higher expressions.
After the intricacies of |scan_primary|\kern-1pt,
the |scan_secondary| routine is
refreshingly simple. It's not trivial, but the operations are relatively
straightforward; the main difficulty is, again, that expressions and data
structures might change drastically every time we call |get_x_next|, so a
cautious approach is mandatory. For example, a macro defined by
\&{primarydef} might have disappeared by the time its second argument has
been scanned; we solve this by increasing the reference count of its token
list, so that the macro can be called even after it has been clobbered.

@<Declare the basic parsing subroutines@>=
procedure scan_secondary;
label restart,continue;
var @!p:pointer; {for list manipulation}
@!c,@!d:halfword; {operation codes or modifiers}
@!mac_name:pointer; {token defined with \&{primarydef}}
begin restart:if(cur_cmd<min_primary_command)or@|
 (cur_cmd>max_primary_command) then
  bad_exp("A secondary");
@.A secondary expression...@>
scan_primary;
continue: if cur_cmd<=max_secondary_command then
 if cur_cmd>=min_secondary_command then
  begin p:=stash_cur_exp; c:=cur_mod; d:=cur_cmd;
  if d=secondary_primary_macro then
    begin mac_name:=cur_sym; add_mac_ref(c);
    end;
  get_x_next; scan_primary;
  if d<>secondary_primary_macro then do_binary(p,c)
  else  begin back_input; binary_mac(p,c,mac_name);
    decr(ref_count(c)); get_x_next; goto restart;
    end;
  goto continue;
  end;
end;

@ The following procedure calls a macro that has two parameters,
|p| and |cur_exp|.

@p procedure binary_mac(@!p,@!c,@!n:pointer);
var @!q,@!r:pointer; {nodes in the parameter list}
begin q:=get_avail; r:=get_avail; link(q):=r;@/
info(q):=p; info(r):=stash_cur_exp;@/
macro_call(c,q,n);
end;

@ The next procedure, |scan_tertiary|, is pretty much the same deal.

@<Declare the basic parsing subroutines@>=
procedure scan_tertiary;
label restart,continue;
var @!p:pointer; {for list manipulation}
@!c,@!d:halfword; {operation codes or modifiers}
@!mac_name:pointer; {token defined with \&{secondarydef}}
begin restart:if(cur_cmd<min_primary_command)or@|
 (cur_cmd>max_primary_command) then
  bad_exp("A tertiary");
@.A tertiary expression...@>
scan_secondary;
if cur_type=future_pen then materialize_pen;
continue: if cur_cmd<=max_tertiary_command then
 if cur_cmd>=min_tertiary_command then
  begin p:=stash_cur_exp; c:=cur_mod; d:=cur_cmd;
  if d=tertiary_secondary_macro then
    begin mac_name:=cur_sym; add_mac_ref(c);
    end;
  get_x_next; scan_secondary;
  if d<>tertiary_secondary_macro then do_binary(p,c)
  else  begin back_input; binary_mac(p,c,mac_name);
    decr(ref_count(c)); get_x_next; goto restart;
    end;
  goto continue;
  end;
end;

@ A |future_pen| becomes a full-fledged pen here.

@p procedure materialize_pen;
label common_ending;
var @!a_minus_b,@!a_plus_b,@!major_axis,@!minor_axis:scaled; {ellipse variables}
@!theta:angle; {amount by which the ellipse has been rotated}
@!p:pointer; {path traverser}
@!q:pointer; {the knot list to be made into a pen}
begin q:=cur_exp;
if left_type(q)=endpoint then
  begin print_err("Pen path must be a cycle");
@.Pen path must be a cycle@>
  help2("I can't make a pen from the given path.")@/
  ("So I've replaced it by the trivial path `(0,0)..cycle'.");
  put_get_error; cur_exp:=null_pen; goto common_ending;
  end
else if left_type(q)=open then
  @<Change node |q| to a path for an elliptical pen@>;
cur_exp:=make_pen(q);
common_ending: toss_knot_list(q); cur_type:=pen_type;
end;

@ We placed the three points $(0,0)$, $(1,0)$, $(0,1)$ into a \&{pencircle},
and they have now been transformed to $(u,v)$, $(A+u,B+v)$, $(C+u,D+v)$;
this gives us enough information to deduce the transformation
$(x,y)\mapsto(Ax+Cy+u,Bx+Dy+v)$.

Given ($A,B,C,D)$ we can always find $(a,b,\theta,\phi)$ such that
$$\eqalign{A&=a\cos\phi\cos\theta-b\sin\phi\sin\theta;\cr
B&=a\cos\phi\sin\theta+b\sin\phi\cos\theta;\cr
C&=-a\sin\phi\cos\theta-b\cos\phi\sin\theta;\cr
D&=-a\sin\phi\sin\theta+b\cos\phi\cos\theta.\cr}$$
In this notation, the unit circle $(\cos t,\sin t)$ is transformed into
$$\bigl(a\cos(\phi+t)\cos\theta-b\sin(\phi+t)\sin\theta,\;
a\cos(\phi+t)\sin\theta+b\sin(\phi+t)\cos\theta\bigr)\;+\;(u,v),$$
which is an ellipse with semi-axes~$(a,b)$, rotated by~$\theta$ and
shifted by~$(u,v)$. To solve the stated equations, we note that it is
necessary and sufficient to solve
$$\eqalign{A-D&=(a-b)\cos(\theta-\phi),\cr
B+C&=(a-b)\sin(\theta-\phi),\cr}
\qquad
\eqalign{A+D&=(a+b)\cos(\theta+\phi),\cr
B-C&=(a+b)\sin(\theta+\phi);\cr}$$
and it is easy to find $a-b$, $a+b$, $\theta-\phi$, and $\theta+\phi$
from these formulas.

The code below uses |(txx,tyx,txy,tyy,tx,ty)| to stand for
$(A,B,C,D,u,v)$.

@<Change node |q|...@>=
begin tx:=x_coord(q); ty:=y_coord(q);
txx:=left_x(q)-tx; tyx:=left_y(q)-ty;
txy:=right_x(q)-tx; tyy:=right_y(q)-ty;
a_minus_b:=pyth_add(txx-tyy,tyx+txy); a_plus_b:=pyth_add(txx+tyy,tyx-txy);
major_axis:=half(a_minus_b+a_plus_b); minor_axis:=half(abs(a_plus_b-a_minus_b));
if major_axis=minor_axis then theta:=0 {circle}
else theta:=half(n_arg(txx-tyy,tyx+txy)+n_arg(txx+tyy,tyx-txy));
free_node(q,knot_node_size);
q:=make_ellipse(major_axis,minor_axis,theta);
if (tx<>0)or(ty<>0) then @<Shift the coordinates of path |q|@>;
end

@ @<Shift the coordinates of path |q|@>=
begin p:=q;
repeat x_coord(p):=x_coord(p)+tx; y_coord(p):=y_coord(p)+ty; p:=link(p);
until p=q;
end

@ Finally we reach the deepest level in our quartet of parsing routines.
This one is much like the others; but it has an extra complication from
paths, which materialize here.

@d continue_path=25 {a label inside of |scan_expression|}
@d finish_path=26 {another}

@<Declare the basic parsing subroutines@>=
procedure scan_expression;
label restart,done,continue,continue_path,finish_path,exit;
var @!p,@!q,@!r,@!pp,@!qq:pointer; {for list manipulation}
@!c,@!d:halfword; {operation codes or modifiers}
@!my_var_flag:0..max_command_code; {initial value of |var_flag|}
@!mac_name:pointer; {token defined with \&{tertiarydef}}
@!cycle_hit:boolean; {did a path expression just end with `\&{cycle}'?}
@!x,@!y:scaled; {explicit coordinates or tension at a path join}
@!t:endpoint..open; {knot type following a path join}
begin my_var_flag:=var_flag;
restart:if(cur_cmd<min_primary_command)or@|
 (cur_cmd>max_primary_command) then
  bad_exp("An");
@.An expression...@>
scan_tertiary;
continue: if cur_cmd<=max_expression_command then
 if cur_cmd>=min_expression_command then
  if (cur_cmd<>equals)or(my_var_flag<>assignment) then
  begin p:=stash_cur_exp; c:=cur_mod; d:=cur_cmd;
  if d=expression_tertiary_macro then
    begin mac_name:=cur_sym; add_mac_ref(c);
    end;
  if (d<ampersand)or((d=ampersand)and@|
   ((type(p)=pair_type)or(type(p)=path_type))) then
    @<Scan a path construction operation;
      but |return| if |p| has the wrong type@>
  else  begin get_x_next; scan_tertiary;
    if d<>expression_tertiary_macro then do_binary(p,c)
    else  begin back_input; binary_mac(p,c,mac_name);
      decr(ref_count(c)); get_x_next; goto restart;
      end;
    end;
  goto continue;
  end;
exit:end;

@ The reader should review the data structure conventions for paths before
hoping to understand the next part of this code.

@<Scan a path construction operation...@>=
begin cycle_hit:=false;
@<Convert the left operand, |p|, into a partial path ending at~|q|;
  but |return| if |p| doesn't have a suitable type@>;
continue_path: @<Determine the path join parameters;
  but |goto finish_path| if there's only a direction specifier@>;
if cur_cmd=cycle then @<Get ready to close a cycle@>
else  begin scan_tertiary;
  @<Convert the right operand, |cur_exp|,
    into a partial path from |pp| to~|qq|@>;
  end;
@<Join the partial paths and reset |p| and |q| to the head and tail
  of the result@>;
if cur_cmd>=min_expression_command then
 if cur_cmd<=ampersand then if not cycle_hit then goto continue_path;
finish_path:
@<Choose control points for the path and put the result into |cur_exp|@>;
end

@ @<Convert the left operand, |p|, into a partial path ending at~|q|...@>=
begin unstash_cur_exp(p);
if cur_type=pair_type then p:=new_knot
else if cur_type=path_type then p:=cur_exp
else return;
q:=p;
while link(q)<>p do q:=link(q);
if left_type(p)<>endpoint then {open up a cycle}
  begin r:=copy_knot(p); link(q):=r; q:=r;
  end;
left_type(p):=open; right_type(q):=open;
end

@ A pair of numeric values is changed into a knot node for a one-point path
when \MF\ discovers that the pair is part of a path.

@p@t\4@>@<Declare the procedure called |known_pair|@>@;
function new_knot:pointer; {convert a pair to a knot with two endpoints}
var @!q:pointer; {the new node}
begin q:=get_node(knot_node_size); left_type(q):=endpoint;
right_type(q):=endpoint; link(q):=q;@/
known_pair; x_coord(q):=cur_x; y_coord(q):=cur_y;
new_knot:=q;
end;

@ The |known_pair| subroutine sets |cur_x| and |cur_y| to the components
of the current expression, assuming that the current expression is a
pair of known numerics. Unknown components are zeroed, and the
current expression is flushed.

@<Declare the procedure called |known_pair|@>=
procedure known_pair;
var @!p:pointer; {the pair node}
begin if cur_type<>pair_type then
  begin exp_err("Undefined coordinates have been replaced by (0,0)");
@.Undefined coordinates...@>
  help5("I need x and y numbers for this part of the path.")@/
    ("The value I found (see above) was no good;")@/
    ("so I'll try to keep going by using zero instead.")@/
    ("(Chapter 27 of The METAFONTbook explains that")@/
@:METAFONTbook}{\sl The {\logos METAFONT\/}book@>
    ("you might want to type `I ???' now.)");
  put_get_flush_error(0); cur_x:=0; cur_y:=0;
  end
else  begin p:=value(cur_exp);
  @<Make sure that both |x| and |y| parts of |p| are known;
    copy them into |cur_x| and |cur_y|@>;
  flush_cur_exp(0);
  end;
end;

@ @<Make sure that both |x| and |y| parts of |p| are known...@>=
if type(x_part_loc(p))=known then cur_x:=value(x_part_loc(p))
else  begin disp_err(x_part_loc(p),
    "Undefined x coordinate has been replaced by 0");
@.Undefined coordinates...@>
  help5("I need a `known' x value for this part of the path.")@/
    ("The value I found (see above) was no good;")@/
    ("so I'll try to keep going by using zero instead.")@/
    ("(Chapter 27 of The METAFONTbook explains that")@/
@:METAFONTbook}{\sl The {\logos METAFONT\/}book@>
    ("you might want to type `I ???' now.)");
  put_get_error; recycle_value(x_part_loc(p)); cur_x:=0;
  end;
if type(y_part_loc(p))=known then cur_y:=value(y_part_loc(p))
else  begin disp_err(y_part_loc(p),
    "Undefined y coordinate has been replaced by 0");
  help5("I need a `known' y value for this part of the path.")@/
    ("The value I found (see above) was no good;")@/
    ("so I'll try to keep going by using zero instead.")@/
    ("(Chapter 27 of The METAFONTbook explains that")@/
    ("you might want to type `I ???' now.)");
  put_get_error; recycle_value(y_part_loc(p)); cur_y:=0;
  end

@ At this point |cur_cmd| is either |ampersand|, |left_brace|, or |path_join|.

@<Determine the path join parameters...@>=
if cur_cmd=left_brace then
  @<Put the pre-join direction information into node |q|@>;
d:=cur_cmd;
if d=path_join then @<Determine the tension and/or control points@>
else if d<>ampersand then goto finish_path;
get_x_next;
if cur_cmd=left_brace then
  @<Put the post-join direction information into |x| and |t|@>
else if right_type(q)<>explicit then
  begin t:=open; x:=0;
  end

@ The |scan_direction| subroutine looks at the directional information
that is enclosed in braces, and also scans ahead to the following character.
A type code is returned, either |open| (if the direction was $(0,0)$),
or |curl| (if the direction was a curl of known value |cur_exp|), or
|given| (if the direction is given by the |angle| value that now
appears in |cur_exp|).

There's nothing difficult about this subroutine, but the program is rather
lengthy because a variety of potential errors need to be nipped in the bud.

@p function scan_direction:small_number;
var @!t:given..open; {the type of information found}
@!x:scaled; {an |x| coordinate}
begin get_x_next;
if cur_cmd=curl_command then @<Scan a curl specification@>
else @<Scan a given direction@>;
if cur_cmd<>right_brace then
  begin missing_err("}");@/
@.Missing `\char`\}'@>
  help3("I've scanned a direction spec for part of a path,")@/
    ("so a right brace should have come next.")@/
    ("I shall pretend that one was there.");@/
  back_error;
  end;
get_x_next; scan_direction:=t;
end;

@ @<Scan a curl specification@>=
begin get_x_next; scan_expression;
if (cur_type<>known)or(cur_exp<0) then
  begin exp_err("Improper curl has been replaced by 1");
@.Improper curl@>
  help1("A curl must be a known, nonnegative number.");
  put_get_flush_error(unity);
  end;
t:=curl;
end

@ @<Scan a given direction@>=
begin scan_expression;
if cur_type>pair_type then @<Get given directions separated by commas@>
else known_pair;
if (cur_x=0)and(cur_y=0) then t:=open
else  begin t:=given; cur_exp:=n_arg(cur_x,cur_y);
  end;
end

@ @<Get given directions separated by commas@>=
begin if cur_type<>known then
  begin exp_err("Undefined x coordinate has been replaced by 0");
@.Undefined coordinates...@>
  help5("I need a `known' x value for this part of the path.")@/
    ("The value I found (see above) was no good;")@/
    ("so I'll try to keep going by using zero instead.")@/
    ("(Chapter 27 of The METAFONTbook explains that")@/
@:METAFONTbook}{\sl The {\logos METAFONT\/}book@>
    ("you might want to type `I ???' now.)");
  put_get_flush_error(0);
  end;
x:=cur_exp;
if cur_cmd<>comma then
  begin missing_err(",");@/
@.Missing `,'@>
  help2("I've got the x coordinate of a path direction;")@/
    ("will look for the y coordinate next.");
  back_error;
  end;
get_x_next; scan_expression;
if cur_type<>known then
  begin exp_err("Undefined y coordinate has been replaced by 0");
  help5("I need a `known' y value for this part of the path.")@/
    ("The value I found (see above) was no good;")@/
    ("so I'll try to keep going by using zero instead.")@/
    ("(Chapter 27 of The METAFONTbook explains that")@/
    ("you might want to type `I ???' now.)");
  put_get_flush_error(0);
  end;
cur_y:=cur_exp; cur_x:=x;
end

@ At this point |right_type(q)| is usually |open|, but it may have been
set to some other value by a previous operation. We must maintain
the value of |right_type(q)| in cases such as
`\.{..\{curl2\}z\{0,0\}..}'.

@<Put the pre-join...@>=
begin t:=scan_direction;
if t<>open then
  begin right_type(q):=t; right_given(q):=cur_exp;
  if left_type(q)=open then
    begin left_type(q):=t; left_given(q):=cur_exp;
    end; {note that |left_given(q)=left_curl(q)|}
  end;
end

@ Since |left_tension| and |left_y| share the same position in knot nodes,
and since |left_given| is similarly equivalent to |left_x|, we use
|x| and |y| to hold the given direction and tension information when
there are no explicit control points.

@<Put the post-join...@>=
begin t:=scan_direction;
if right_type(q)<>explicit then x:=cur_exp
else t:=explicit; {the direction information is superfluous}
end

@ @<Determine the tension and/or...@>=
begin get_x_next;
if cur_cmd=tension then @<Set explicit tensions@>
else if cur_cmd=controls then @<Set explicit control points@>
else  begin right_tension(q):=unity; y:=unity; back_input; {default tension}
  goto done;
  end;
if cur_cmd<>path_join then
  begin missing_err("..");@/
@.Missing `..'@>
  help1("A path join command should end with two dots.");
  back_error;
  end;
done:end

@ @<Set explicit tensions@>=
begin get_x_next; y:=cur_cmd;
if cur_cmd=at_least then get_x_next;
scan_primary;
@<Make sure that the current expression is a valid tension setting@>;
if y=at_least then negate(cur_exp);
right_tension(q):=cur_exp;
if cur_cmd=and_command then
  begin get_x_next; y:=cur_cmd;
  if cur_cmd=at_least then get_x_next;
  scan_primary;
  @<Make sure that the current expression is a valid tension setting@>;
  if y=at_least then negate(cur_exp);
  end;
y:=cur_exp;
end

@ @d min_tension==three_quarter_unit

@<Make sure that the current expression is a valid tension setting@>=
if (cur_type<>known)or(cur_exp<min_tension) then
  begin exp_err("Improper tension has been set to 1");
@.Improper tension@>
  help1("The expression above should have been a number >=3/4.");
  put_get_flush_error(unity);
  end

@ @<Set explicit control points@>=
begin right_type(q):=explicit; t:=explicit; get_x_next; scan_primary;@/
known_pair; right_x(q):=cur_x; right_y(q):=cur_y;
if cur_cmd<>and_command then
  begin x:=right_x(q); y:=right_y(q);
  end
else  begin get_x_next; scan_primary;@/
  known_pair; x:=cur_x; y:=cur_y;
  end;
end

@ @<Convert the right operand, |cur_exp|, into a partial path...@>=
begin if cur_type<>path_type then pp:=new_knot
else pp:=cur_exp;
qq:=pp;
while link(qq)<>pp do qq:=link(qq);
if left_type(pp)<>endpoint then {open up a cycle}
  begin r:=copy_knot(pp); link(qq):=r; qq:=r;
  end;
left_type(pp):=open; right_type(qq):=open;
end

@ If a person tries to define an entire path by saying `\.{(x,y)\&cycle}',
we silently change the specification to `\.{(x,y)..cycle}', since a cycle
shouldn't have length zero.

@<Get ready to close a cycle@>=
begin cycle_hit:=true; get_x_next; pp:=p; qq:=p;
if d=ampersand then if p=q then
  begin d:=path_join; right_tension(q):=unity; y:=unity;
  end;
end

@ @<Join the partial paths and reset |p| and |q|...@>=
begin if d=ampersand then
 if (x_coord(q)<>x_coord(pp))or(y_coord(q)<>y_coord(pp)) then
  begin print_err("Paths don't touch; `&' will be changed to `..'");
@.Paths don't touch@>
  help3("When you join paths `p&q', the ending point of p")@/
    ("must be exactly equal to the starting point of q.")@/
    ("So I'm going to pretend that you said `p..q' instead.");
  put_get_error; d:=path_join; right_tension(q):=unity; y:=unity;
  end;
@<Plug an opening in |right_type(pp)|, if possible@>;
if d=ampersand then @<Splice independent paths together@>
else  begin @<Plug an opening in |right_type(q)|, if possible@>;
  link(q):=pp; left_y(pp):=y;
  if t<>open then
    begin left_x(pp):=x; left_type(pp):=t;
    end;
  end;
q:=qq;
end

@ @<Plug an opening in |right_type(q)|...@>=
if right_type(q)=open then
  if (left_type(q)=curl)or(left_type(q)=given) then
    begin right_type(q):=left_type(q); right_given(q):=left_given(q);
    end

@ @<Plug an opening in |right_type(pp)|...@>=
if right_type(pp)=open then
  if (t=curl)or(t=given) then
    begin right_type(pp):=t; right_given(pp):=x;
    end

@ @<Splice independent paths together@>=
begin if left_type(q)=open then if right_type(q)=open then
    begin left_type(q):=curl; left_curl(q):=unity;
    end;
if right_type(pp)=open then if t=open then
  begin right_type(pp):=curl; right_curl(pp):=unity;
  end;
right_type(q):=right_type(pp); link(q):=link(pp);@/
right_x(q):=right_x(pp); right_y(q):=right_y(pp);
free_node(pp,knot_node_size);
if qq=pp then qq:=q;
end

@ @<Choose control points for the path...@>=
if cycle_hit then
  begin if d=ampersand then p:=q;
  end
else  begin left_type(p):=endpoint;
  if right_type(p)=open then
    begin right_type(p):=curl; right_curl(p):=unity;
    end;
  right_type(q):=endpoint;
  if left_type(q)=open then
    begin left_type(q):=curl; left_curl(q):=unity;
    end;
  link(q):=p;
  end;
make_choices(p);
cur_type:=path_type; cur_exp:=p

@ Finally, we sometimes need to scan an expression whose value is
supposed to be either |true_code| or |false_code|.

@<Declare the basic parsing subroutines@>=
procedure get_boolean;
begin get_x_next; scan_expression;
if cur_type<>boolean_type then
  begin exp_err("Undefined condition will be treated as `false'");
@.Undefined condition...@>
  help2("The expression shown above should have had a definite")@/
    ("true-or-false value. I'm changing it to `false'.");@/
  put_get_flush_error(false_code); cur_type:=boolean_type;
  end;
end;

@* \[42] Doing the operations.
The purpose of parsing is primarily to permit people to avoid piles of
parentheses. But the real work is done after the structure of an expression
has been recognized; that's when new expressions are generated. We
turn now to the guts of \MF, which handles individual operators that
have come through the parsing mechanism.

We'll start with the easy ones that take no operands, then work our way
up to operators with one and ultimately two arguments. In other words,
we will write the three procedures |do_nullary|, |do_unary|, and |do_binary|
that are invoked periodically by the expression scanners.

First let's make sure that all of the primitive operators are in the
hash table. Although |scan_primary| and its relatives made use of the
\\{cmd} code for these operators, the \\{do} routines base everything
on the \\{mod} code. For example, |do_binary| doesn't care whether the
operation it performs is a |primary_binary| or |secondary_binary|, etc.

@<Put each...@>=
primitive("true",nullary,true_code);@/
@!@:true_}{\&{true} primitive@>
primitive("false",nullary,false_code);@/
@!@:false_}{\&{false} primitive@>
primitive("nullpicture",nullary,null_picture_code);@/
@!@:null_picture_}{\&{nullpicture} primitive@>
primitive("nullpen",nullary,null_pen_code);@/
@!@:null_pen_}{\&{nullpen} primitive@>
primitive("jobname",nullary,job_name_op);@/
@!@:job_name_}{\&{jobname} primitive@>
primitive("readstring",nullary,read_string_op);@/
@!@:read_string_}{\&{readstring} primitive@>
primitive("pencircle",nullary,pen_circle);@/
@!@:pen_circle_}{\&{pencircle} primitive@>
primitive("normaldeviate",nullary,normal_deviate);@/
@!@:normal_deviate_}{\&{normaldeviate} primitive@>
primitive("odd",unary,odd_op);@/
@!@:odd_}{\&{odd} primitive@>
primitive("known",unary,known_op);@/
@!@:known_}{\&{known} primitive@>
primitive("unknown",unary,unknown_op);@/
@!@:unknown_}{\&{unknown} primitive@>
primitive("not",unary,not_op);@/
@!@:not_}{\&{not} primitive@>
primitive("decimal",unary,decimal);@/
@!@:decimal_}{\&{decimal} primitive@>
primitive("reverse",unary,reverse);@/
@!@:reverse_}{\&{reverse} primitive@>
primitive("makepath",unary,make_path_op);@/
@!@:make_path_}{\&{makepath} primitive@>
primitive("makepen",unary,make_pen_op);@/
@!@:make_pen_}{\&{makepen} primitive@>
primitive("totalweight",unary,total_weight_op);@/
@!@:total_weight_}{\&{totalweight} primitive@>
primitive("oct",unary,oct_op);@/
@!@:oct_}{\&{oct} primitive@>
primitive("hex",unary,hex_op);@/
@!@:hex_}{\&{hex} primitive@>
primitive("ASCII",unary,ASCII_op);@/
@!@:ASCII_}{\&{ASCII} primitive@>
primitive("char",unary,char_op);@/
@!@:char_}{\&{char} primitive@>
primitive("length",unary,length_op);@/
@!@:length_}{\&{length} primitive@>
primitive("turningnumber",unary,turning_op);@/
@!@:turning_number_}{\&{turningnumber} primitive@>
primitive("xpart",unary,x_part);@/
@!@:x_part_}{\&{xpart} primitive@>
primitive("ypart",unary,y_part);@/
@!@:y_part_}{\&{ypart} primitive@>
primitive("xxpart",unary,xx_part);@/
@!@:xx_part_}{\&{xxpart} primitive@>
primitive("xypart",unary,xy_part);@/
@!@:xy_part_}{\&{xypart} primitive@>
primitive("yxpart",unary,yx_part);@/
@!@:yx_part_}{\&{yxpart} primitive@>
primitive("yypart",unary,yy_part);@/
@!@:yy_part_}{\&{yypart} primitive@>
primitive("sqrt",unary,sqrt_op);@/
@!@:sqrt_}{\&{sqrt} primitive@>
primitive("mexp",unary,m_exp_op);@/
@!@:m_exp_}{\&{mexp} primitive@>
primitive("mlog",unary,m_log_op);@/
@!@:m_log_}{\&{mlog} primitive@>
primitive("sind",unary,sin_d_op);@/
@!@:sin_d_}{\&{sind} primitive@>
primitive("cosd",unary,cos_d_op);@/
@!@:cos_d_}{\&{cosd} primitive@>
primitive("floor",unary,floor_op);@/
@!@:floor_}{\&{floor} primitive@>
primitive("uniformdeviate",unary,uniform_deviate);@/
@!@:uniform_deviate_}{\&{uniformdeviate} primitive@>
primitive("charexists",unary,char_exists_op);@/
@!@:char_exists_}{\&{charexists} primitive@>
primitive("angle",unary,angle_op);@/
@!@:angle_}{\&{angle} primitive@>
primitive("cycle",cycle,cycle_op);@/
@!@:cycle_}{\&{cycle} primitive@>
primitive("+",plus_or_minus,plus);@/
@!@:+ }{\.{+} primitive@>
primitive("-",plus_or_minus,minus);@/
@!@:- }{\.{-} primitive@>
primitive("*",secondary_binary,times);@/
@!@:* }{\.{*} primitive@>
primitive("/",slash,over); eqtb[frozen_slash]:=eqtb[cur_sym];@/
@!@:/ }{\.{/} primitive@>
primitive("++",tertiary_binary,pythag_add);@/
@!@:++_}{\.{++} primitive@>
primitive("+-+",tertiary_binary,pythag_sub);@/
@!@:+-+_}{\.{+-+} primitive@>
primitive("and",and_command,and_op);@/
@!@:and_}{\&{and} primitive@>
primitive("or",tertiary_binary,or_op);@/
@!@:or_}{\&{or} primitive@>
primitive("<",expression_binary,less_than);@/
@!@:< }{\.{<} primitive@>
primitive("<=",expression_binary,less_or_equal);@/
@!@:<=_}{\.{<=} primitive@>
primitive(">",expression_binary,greater_than);@/
@!@:> }{\.{>} primitive@>
primitive(">=",expression_binary,greater_or_equal);@/
@!@:>=_}{\.{>=} primitive@>
primitive("=",equals,equal_to);@/
@!@:= }{\.{=} primitive@>
primitive("<>",expression_binary,unequal_to);@/
@!@:<>_}{\.{<>} primitive@>
primitive("substring",primary_binary,substring_of);@/
@!@:substring_}{\&{substring} primitive@>
primitive("subpath",primary_binary,subpath_of);@/
@!@:subpath_}{\&{subpath} primitive@>
primitive("directiontime",primary_binary,direction_time_of);@/
@!@:direction_time_}{\&{directiontime} primitive@>
primitive("point",primary_binary,point_of);@/
@!@:point_}{\&{point} primitive@>
primitive("precontrol",primary_binary,precontrol_of);@/
@!@:precontrol_}{\&{precontrol} primitive@>
primitive("postcontrol",primary_binary,postcontrol_of);@/
@!@:postcontrol_}{\&{postcontrol} primitive@>
primitive("penoffset",primary_binary,pen_offset_of);@/
@!@:pen_offset_}{\&{penoffset} primitive@>
primitive("&",ampersand,concatenate);@/
@!@:!!!}{\.{\&} primitive@>
primitive("rotated",secondary_binary,rotated_by);@/
@!@:rotated_}{\&{rotated} primitive@>
primitive("slanted",secondary_binary,slanted_by);@/
@!@:slanted_}{\&{slanted} primitive@>
primitive("scaled",secondary_binary,scaled_by);@/
@!@:scaled_}{\&{scaled} primitive@>
primitive("shifted",secondary_binary,shifted_by);@/
@!@:shifted_}{\&{shifted} primitive@>
primitive("transformed",secondary_binary,transformed_by);@/
@!@:transformed_}{\&{transformed} primitive@>
primitive("xscaled",secondary_binary,x_scaled);@/
@!@:x_scaled_}{\&{xscaled} primitive@>
primitive("yscaled",secondary_binary,y_scaled);@/
@!@:y_scaled_}{\&{yscaled} primitive@>
primitive("zscaled",secondary_binary,z_scaled);@/
@!@:z_scaled_}{\&{zscaled} primitive@>
primitive("intersectiontimes",tertiary_binary,intersect);@/
@!@:intersection_times_}{\&{intersectiontimes} primitive@>

@ @<Cases of |print_cmd...@>=
nullary,unary,primary_binary,secondary_binary,tertiary_binary,
 expression_binary,cycle,plus_or_minus,slash,ampersand,equals,and_command:
  print_op(m);

@ OK, let's look at the simplest \\{do} procedure first.

@p procedure do_nullary(@!c:quarterword);
var @!k:integer; {all-purpose loop index}
begin check_arith;
if internal[tracing_commands]>two then
  show_cmd_mod(nullary,c);
case c of
true_code,false_code:begin cur_type:=boolean_type; cur_exp:=c;
  end;
null_picture_code:begin cur_type:=picture_type;
  cur_exp:=get_node(edge_header_size); init_edges(cur_exp);
  end;
null_pen_code:begin cur_type:=pen_type; cur_exp:=null_pen;
  end;
normal_deviate:begin cur_type:=known; cur_exp:=norm_rand;
  end;
pen_circle:@<Make a special knot node for \&{pencircle}@>;
job_name_op: begin if job_name=0 then open_log_file;
  cur_type:=string_type; cur_exp:=job_name;
  end;
read_string_op:@<Read a string from the terminal@>;
end; {there are no other cases}
check_arith;
end;

@ @<Make a special knot node for \&{pencircle}@>=
begin cur_type:=future_pen; cur_exp:=get_node(knot_node_size);
left_type(cur_exp):=open; right_type(cur_exp):=open;
link(cur_exp):=cur_exp;@/
x_coord(cur_exp):=0; y_coord(cur_exp):=0;@/
left_x(cur_exp):=unity; left_y(cur_exp):=0;@/
right_x(cur_exp):=0; right_y(cur_exp):=unity;@/
end

@ @<Read a string...@>=
begin if interaction<=nonstop_mode then
  fatal_error("*** (cannot readstring in nonstop modes)");
begin_file_reading; name:=1; prompt_input("");
str_room(last-start);
for k:=start to last-1 do append_char(buffer[k]);
end_file_reading; cur_type:=string_type; cur_exp:=make_string;
end

@ Things get a bit more interesting when there's an operand. The
operand to |do_unary| appears in |cur_type| and |cur_exp|.

@p @t\4@>@<Declare unary action procedures@>@;
procedure do_unary(@!c:quarterword);
var @!p,@!q:pointer; {for list manipulation}
@!x:integer; {a temporary register}
begin check_arith;
if internal[tracing_commands]>two then
  @<Trace the current unary operation@>;
case c of
plus:if cur_type<pair_type then
  if cur_type<>picture_type then bad_unary(plus);
minus:@<Negate the current expression@>;
@t\4@>@<Additional cases of unary operators@>@;
end; {there are no other cases}
check_arith;
end;

@ The |nice_pair| function returns |true| if both components of a pair
are known.

@<Declare unary action procedures@>=
function nice_pair(@!p:integer;@!t:quarterword):boolean;
label exit;
begin if t=pair_type then
  begin p:=value(p);
  if type(x_part_loc(p))=known then
   if type(y_part_loc(p))=known then
    begin nice_pair:=true; return;
    end;
  end;
nice_pair:=false;
exit:end;

@ @<Declare unary action...@>=
procedure print_known_or_unknown_type(@!t:small_number;@!v:integer);
begin print_char("(");
if t<dependent then
  if t<>pair_type then print_type(t)
  else if nice_pair(v,pair_type) then print("pair")
  else print("unknown pair")
else print("unknown numeric");
print_char(")");
end;

@ @<Declare unary action...@>=
procedure bad_unary(@!c:quarterword);
begin exp_err("Not implemented: "); print_op(c);
@.Not implemented...@>
print_known_or_unknown_type(cur_type,cur_exp);
help3("I'm afraid I don't know how to apply that operation to that")@/
  ("particular type. Continue, and I'll simply return the")@/
  ("argument (shown above) as the result of the operation.");
put_get_error;
end;

@ @<Trace the current unary operation@>=
begin begin_diagnostic; print_nl("{"); print_op(c); print_char("(");@/
print_exp(null,0); {show the operand, but not verbosely}
print(")}"); end_diagnostic(false);
end

@ Negation is easy except when the current expression
is of type |independent|, or when it is a pair with one or more
|independent| components.

It is tempting to argue that the negative of an independent variable
is an independent variable, hence we don't have to do anything when
negating it. The fallacy is that other dependent variables pointing
to the current expression must change the sign of their
coefficients if we make no change to the current expression.

Instead, we work around the problem by copying the current expression
and recycling it afterwards (cf.~the |stash_in| routine).

@<Negate the current expression@>=
case cur_type of
pair_type,independent: begin q:=cur_exp; make_exp_copy(q);
  if cur_type=dependent then negate_dep_list(dep_list(cur_exp))
  else if cur_type=pair_type then
    begin p:=value(cur_exp);
    if type(x_part_loc(p))=known then negate(value(x_part_loc(p)))
    else negate_dep_list(dep_list(x_part_loc(p)));
    if type(y_part_loc(p))=known then negate(value(y_part_loc(p)))
    else negate_dep_list(dep_list(y_part_loc(p)));
    end; {if |cur_type=known| then |cur_exp=0|}
  recycle_value(q); free_node(q,value_node_size);
  end;
dependent,proto_dependent:negate_dep_list(dep_list(cur_exp));
known:negate(cur_exp);
picture_type:negate_edges(cur_exp);
othercases bad_unary(minus)
endcases

@ @<Declare unary action...@>=
procedure negate_dep_list(@!p:pointer);
label exit;
begin loop@+begin negate(value(p));
  if info(p)=null then return;
  p:=link(p);
  end;
exit:end;

@ @<Additional cases of unary operators@>=
not_op: if cur_type<>boolean_type then bad_unary(not_op)
  else cur_exp:=true_code+false_code-cur_exp;

@ @d three_sixty_units==23592960 {that's |360*unity|}
@d boolean_reset(#)==if # then cur_exp:=true_code@+else cur_exp:=false_code

@<Additional cases of unary operators@>=
sqrt_op,m_exp_op,m_log_op,sin_d_op,cos_d_op,floor_op,
 uniform_deviate,odd_op,char_exists_op:@t@>@;@/
  if cur_type<>known then bad_unary(c)
  else case c of
  sqrt_op:cur_exp:=square_rt(cur_exp);
  m_exp_op:cur_exp:=m_exp(cur_exp);
  m_log_op:cur_exp:=m_log(cur_exp);
  sin_d_op,cos_d_op:begin n_sin_cos((cur_exp mod three_sixty_units)*16);
    if c=sin_d_op then cur_exp:=round_fraction(n_sin)
    else cur_exp:=round_fraction(n_cos);
    end;
  floor_op:cur_exp:=floor_scaled(cur_exp);
  uniform_deviate:cur_exp:=unif_rand(cur_exp);
  odd_op: begin boolean_reset(odd(round_unscaled(cur_exp)));
    cur_type:=boolean_type;
    end;
  char_exists_op:@<Determine if a character has been shipped out@>;
  end; {there are no other cases}

@ @<Additional cases of unary operators@>=
angle_op:if nice_pair(cur_exp,cur_type) then
    begin p:=value(cur_exp);
    x:=n_arg(value(x_part_loc(p)),value(y_part_loc(p)));
    if x>=0 then flush_cur_exp((x+8)div 16)
    else flush_cur_exp(-((-x+8)div 16));
    end
  else bad_unary(angle_op);

@ If the current expression is a pair, but the context wants it to
be a path, we call |pair_to_path|.

@<Declare unary action...@>=
procedure pair_to_path;
begin cur_exp:=new_knot; cur_type:=path_type;
end;

@ @<Additional cases of unary operators@>=
x_part,y_part:if (cur_type<=pair_type)and(cur_type>=transform_type) then
    take_part(c)
  else bad_unary(c);
xx_part,xy_part,yx_part,yy_part: if cur_type=transform_type then take_part(c)
  else bad_unary(c);

@ In the following procedure, |cur_exp| points to a capsule, which points to
a big node. We want to delete all but one part of the big node.

@<Declare unary action...@>=
procedure take_part(@!c:quarterword);
var @!p:pointer; {the big node}
begin p:=value(cur_exp); value(temp_val):=p; type(temp_val):=cur_type;
link(p):=temp_val; free_node(cur_exp,value_node_size);
make_exp_copy(p+2*(c-x_part));
recycle_value(temp_val);
end;

@ @<Initialize table entries...@>=
name_type(temp_val):=capsule;

@ @<Additional cases of unary...@>=
char_op: if cur_type<>known then bad_unary(char_op)
  else  begin cur_exp:=round_unscaled(cur_exp) mod 256; cur_type:=string_type;
    if cur_exp<0 then cur_exp:=cur_exp+256;
    if length(cur_exp)<>1 then
      begin str_room(1); append_char(cur_exp); cur_exp:=make_string;
      end;
    end;
decimal: if cur_type<>known then bad_unary(decimal)
  else  begin old_setting:=selector; selector:=new_string;
    print_scaled(cur_exp); cur_exp:=make_string;
    selector:=old_setting; cur_type:=string_type;
    end;
oct_op,hex_op,ASCII_op: if cur_type<>string_type then bad_unary(c)
  else str_to_num(c);

@ @<Declare unary action...@>=
procedure str_to_num(@!c:quarterword); {converts a string to a number}
var @!n:integer; {accumulator}
@!m:ASCII_code; {current character}
@!k:pool_pointer; {index into |str_pool|}
@!b:8..16; {radix of conversion}
@!bad_char:boolean; {did the string contain an invalid digit?}
begin if c=ASCII_op then
  if length(cur_exp)=0 then n:=-1
  else n:=so(str_pool[str_start[cur_exp]])
else  begin if c=oct_op then b:=8@+else b:=16;
  n:=0; bad_char:=false;
  for k:=str_start[cur_exp] to str_start[cur_exp+1]-1 do
    begin m:=so(str_pool[k]);
    if (m>="0")and(m<="9") then m:=m-"0"
    else if (m>="A")and(m<="F") then m:=m-"A"+10
    else if (m>="a")and(m<="f") then m:=m-"a"+10
    else  begin bad_char:=true; m:=0;
      end;
    if m>=b then
      begin bad_char:=true; m:=0;
      end;
    if n<32768 div b then n:=n*b+m@+else n:=32767;
    end;
  @<Give error messages if |bad_char| or |n>=4096|@>;
  end;
flush_cur_exp(n*unity);
end;

@ @<Give error messages if |bad_char|...@>=
if bad_char then
  begin exp_err("String contains illegal digits");
@.String contains illegal digits@>
  if c=oct_op then
    help1("I zeroed out characters that weren't in the range 0..7.")
  else help1("I zeroed out characters that weren't hex digits.");
  put_get_error;
  end;
if n>4095 then
  begin print_err("Number too large ("); print_int(n); print_char(")");
@.Number too large@>
  help1("I have trouble with numbers greater than 4095; watch out.");
  put_get_error;
  end

@ The length operation is somewhat unusual in that it applies to a variety
of different types of operands.

@<Additional cases of unary...@>=
length_op: if cur_type=string_type then flush_cur_exp(length(cur_exp)*unity)
  else if cur_type=path_type then flush_cur_exp(path_length)
  else if cur_type=known then cur_exp:=abs(cur_exp)
  else if nice_pair(cur_exp,cur_type) then
    flush_cur_exp(pyth_add(value(x_part_loc(value(cur_exp))),@|
      value(y_part_loc(value(cur_exp)))))
  else bad_unary(c);

@ @<Declare unary action...@>=
function path_length:scaled; {computes the length of the current path}
var @!n:scaled; {the path length so far}
@!p:pointer; {traverser}
begin p:=cur_exp;
if left_type(p)=endpoint then n:=-unity@+else n:=0;
repeat p:=link(p); n:=n+unity;
until p=cur_exp;
path_length:=n;
end;

@ The turning number is computed only with respect to null pens. A different
pen might affect the turning number, in degenerate cases, because autorounding
will produce a slightly different path, or because excessively large coordinates
might be truncated.

@<Additional cases of unary...@>=
turning_op:if cur_type=pair_type then flush_cur_exp(0)
  else if cur_type<>path_type then bad_unary(turning_op)
  else if left_type(cur_exp)=endpoint then
     flush_cur_exp(0) {not a cyclic path}
  else  begin cur_pen:=null_pen; cur_path_type:=contour_code;
    cur_exp:=make_spec(cur_exp,
      fraction_one-half_unit-1-el_gordo,0);
    flush_cur_exp(turning_number*unity); {convert to |scaled|}
    end;

@ @d type_test_end== flush_cur_exp(true_code)
  else flush_cur_exp(false_code);
  cur_type:=boolean_type;
  end
@d type_range_end(#)==(cur_type<=#) then type_test_end
@d type_range(#)==begin if (cur_type>=#) and type_range_end
@d type_test(#)==begin if cur_type=# then type_test_end

@<Additional cases of unary operators@>=
boolean_type: type_range(boolean_type)(unknown_boolean);
string_type: type_range(string_type)(unknown_string);
pen_type: type_range(pen_type)(future_pen);
path_type: type_range(path_type)(unknown_path);
picture_type: type_range(picture_type)(unknown_picture);
transform_type,pair_type: type_test(c);
numeric_type: type_range(known)(independent);
known_op,unknown_op: test_known(c);

@ @<Declare unary action procedures@>=
procedure test_known(@!c:quarterword);
label done;
var @!b:true_code..false_code; {is the current expression known?}
@!p,@!q:pointer; {locations in a big node}
begin b:=false_code;
case cur_type of
vacuous,boolean_type,string_type,pen_type,future_pen,path_type,picture_type,
 known: b:=true_code;
transform_type,pair_type:begin p:=value(cur_exp); q:=p+big_node_size[cur_type];
  repeat q:=q-2;
  if type(q)<>known then goto done;
  until q=p;
  b:=true_code;
done:  end;
othercases do_nothing
endcases;
if c=known_op then flush_cur_exp(b)
else flush_cur_exp(true_code+false_code-b);
cur_type:=boolean_type;
end;

@ @<Additional cases of unary operators@>=
cycle_op: begin if cur_type<>path_type then flush_cur_exp(false_code)
  else if left_type(cur_exp)<>endpoint then flush_cur_exp(true_code)
  else flush_cur_exp(false_code);
  cur_type:=boolean_type;
  end;

@ @<Additional cases of unary operators@>=
make_pen_op: begin if cur_type=pair_type then pair_to_path;
  if cur_type=path_type then cur_type:=future_pen
  else bad_unary(make_pen_op);
  end;
make_path_op: begin if cur_type=future_pen then materialize_pen;
  if cur_type<>pen_type then bad_unary(make_path_op)
  else  begin flush_cur_exp(make_path(cur_exp)); cur_type:=path_type;
    end;
  end;
total_weight_op: if cur_type<>picture_type then bad_unary(total_weight_op)
  else flush_cur_exp(total_weight(cur_exp));
reverse: if cur_type=path_type then
    begin p:=htap_ypoc(cur_exp);
    if right_type(p)=endpoint then p:=link(p);
    toss_knot_list(cur_exp); cur_exp:=p;
    end
  else if cur_type=pair_type then pair_to_path
  else bad_unary(reverse);

@ Finally, we have the operations that combine a capsule~|p|
with the current expression.

@p @t\4@>@<Declare binary action procedures@>@;
procedure do_binary(@!p:pointer;@!c:quarterword);
label done,done1,exit;
var @!q,@!r,@!rr:pointer; {for list manipulation}
@!old_p,@!old_exp:pointer; {capsules to recycle}
@!v:integer; {for numeric manipulation}
begin check_arith;
if internal[tracing_commands]>two then
  @<Trace the current binary operation@>;
@<Sidestep |independent| cases in capsule |p|@>;
@<Sidestep |independent| cases in the current expression@>;
case c of
plus,minus:@<Add or subtract the current expression from |p|@>;
@t\4@>@<Additional cases of binary operators@>@;
end; {there are no other cases}
recycle_value(p); free_node(p,value_node_size); {|return| to avoid this}
exit:check_arith; @<Recycle any sidestepped |independent| capsules@>;
end;

@ @<Declare binary action...@>=
procedure bad_binary(@!p:pointer;@!c:quarterword);
begin disp_err(p,"");
exp_err("Not implemented: ");
@.Not implemented...@>
if c>=min_of then print_op(c);
print_known_or_unknown_type(type(p),p);
if c>=min_of then print("of")@+else print_op(c);
print_known_or_unknown_type(cur_type,cur_exp);@/
help3("I'm afraid I don't know how to apply that operation to that")@/
  ("combination of types. Continue, and I'll return the second")@/
  ("argument (see above) as the result of the operation.");
put_get_error;
end;

@ @<Trace the current binary operation@>=
begin begin_diagnostic; print_nl("{(");
print_exp(p,0); {show the operand, but not verbosely}
print_char(")"); print_op(c); print_char("(");@/
print_exp(null,0); print(")}"); end_diagnostic(false);
end

@ Several of the binary operations are potentially complicated by the
fact that |independent| values can sneak into capsules. For example,
we've seen an instance of this difficulty in the unary operation
of negation. In order to reduce the number of cases that need to be
handled, we first change the two operands (if necessary)
to rid them of |independent| components. The original operands are
put into capsules called |old_p| and |old_exp|, which will be
recycled after the binary operation has been safely carried out.

@<Recycle any sidestepped |independent| capsules@>=
if old_p<>null then
  begin recycle_value(old_p); free_node(old_p,value_node_size);
  end;
if old_exp<>null then
  begin recycle_value(old_exp); free_node(old_exp,value_node_size);
  end

@ A big node is considered to be ``tarnished'' if it contains at least one
independent component. We will define a simple function called `|tarnished|'
that returns |null| if and only if its argument is not tarnished.

@<Sidestep |independent| cases in capsule |p|@>=
case type(p) of
transform_type,pair_type: old_p:=tarnished(p);
independent: old_p:=void;
othercases old_p:=null
endcases;
if old_p<>null then
  begin q:=stash_cur_exp; old_p:=p; make_exp_copy(old_p);
  p:=stash_cur_exp; unstash_cur_exp(q);
  end;

@ @<Sidestep |independent| cases in the current expression@>=
case cur_type of
transform_type,pair_type:old_exp:=tarnished(cur_exp);
independent:old_exp:=void;
othercases old_exp:=null
endcases;
if old_exp<>null then
  begin old_exp:=cur_exp; make_exp_copy(old_exp);
  end

@ @<Declare binary action...@>=
function tarnished(@!p:pointer):pointer;
label exit;
var @!q:pointer; {beginning of the big node}
@!r:pointer; {current position in the big node}
begin q:=value(p); r:=q+big_node_size[type(p)];
repeat r:=r-2;
if type(r)=independent then
  begin tarnished:=void; return;
  end;
until r=q;
tarnished:=null;
exit:end;

@ @<Add or subtract the current expression from |p|@>=
if (cur_type<pair_type)or(type(p)<pair_type) then
  if (cur_type=picture_type)and(type(p)=picture_type) then
    begin if c=minus then negate_edges(cur_exp);
    cur_edges:=cur_exp; merge_edges(value(p));
    end
  else bad_binary(p,c)
else  if cur_type=pair_type then
    if type(p)<>pair_type then bad_binary(p,c)
    else  begin q:=value(p); r:=value(cur_exp);
      add_or_subtract(x_part_loc(q),x_part_loc(r),c);
      add_or_subtract(y_part_loc(q),y_part_loc(r),c);
      end
  else  if type(p)=pair_type then bad_binary(p,c)
    else add_or_subtract(p,null,c)

@ The first argument to |add_or_subtract| is the location of a value node
in a capsule or pair node that will soon be recycled. The second argument
is either a location within a pair or transform node of |cur_exp|,
or it is null (which means that |cur_exp| itself should be the second
argument).  The third argument is either |plus| or |minus|.

The sum or difference of the numeric quantities will replace the second
operand.  Arithmetic overflow may go undetected; users aren't supposed to
be monkeying around with really big values.
@^overflow in arithmetic@>

@<Declare binary action...@>=
@t\4@>@<Declare the procedure called |dep_finish|@>@;
procedure add_or_subtract(@!p,@!q:pointer;@!c:quarterword);
label done,exit;
var @!s,@!t:small_number; {operand types}
@!r:pointer; {list traverser}
@!v:integer; {second operand value}
begin if q=null then
  begin t:=cur_type;
  if t<dependent then v:=cur_exp@+else v:=dep_list(cur_exp);
  end
else  begin t:=type(q);
  if t<dependent then v:=value(q)@+else v:=dep_list(q);
  end;
if t=known then
  begin if c=minus then negate(v);
  if type(p)=known then
    begin v:=slow_add(value(p),v);
    if q=null then cur_exp:=v@+else value(q):=v;
    return;
    end;
  @<Add a known value to the constant term of |dep_list(p)|@>;
  end
else  begin if c=minus then negate_dep_list(v);
  @<Add operand |p| to the dependency list |v|@>;
  end;
exit:end;

@ @<Add a known value to the constant term of |dep_list(p)|@>=
r:=dep_list(p);
while info(r)<>null do r:=link(r);
value(r):=slow_add(value(r),v);
if q=null then
  begin q:=get_node(value_node_size); cur_exp:=q; cur_type:=type(p);
  name_type(q):=capsule;
  end;
dep_list(q):=dep_list(p); type(q):=type(p);
prev_dep(q):=prev_dep(p); link(prev_dep(p)):=q;
type(p):=known; {this will keep the recycler from collecting non-garbage}

@ We prefer |dependent| lists to |proto_dependent| ones, because it is
nice to retain the extra accuracy of |fraction| coefficients.
But we have to handle both kinds, and mixtures too.

@<Add operand |p| to the dependency list |v|@>=
if type(p)=known then
  @<Add the known |value(p)| to the constant term of |v|@>
else  begin s:=type(p); r:=dep_list(p);
  if t=dependent then
    begin if s=dependent then
     if max_coef(r)+max_coef(v)<coef_bound then
      begin v:=p_plus_q(v,r,dependent); goto done;
      end; {|fix_needed| will necessarily be false}
    t:=proto_dependent; v:=p_over_v(v,unity,dependent,proto_dependent);
    end;
  if s=proto_dependent then v:=p_plus_q(v,r,proto_dependent)
  else v:=p_plus_fq(v,unity,r,proto_dependent,dependent);
 done:  @<Output the answer, |v| (which might have become |known|)@>;
  end

@ @<Add the known |value(p)| to the constant term of |v|@>=
begin while info(v)<>null do v:=link(v);
value(v):=slow_add(value(p),value(v));
end

@ @<Output the answer, |v| (which might have become |known|)@>=
if q<>null then dep_finish(v,q,t)
else  begin cur_type:=t; dep_finish(v,null,t);
  end

@ Here's the current situation: The dependency list |v| of type |t|
should either be put into the current expression (if |q=null|) or
into location |q| within a pair node (otherwise). The destination (|cur_exp|
or |q|) formerly held a dependency list with the same
final pointer as the list |v|.

@<Declare the procedure called |dep_finish|@>=
procedure dep_finish(@!v,@!q:pointer;@!t:small_number);
var @!p:pointer; {the destination}
@!vv:scaled; {the value, if it is |known|}
begin if q=null then p:=cur_exp@+else p:=q;
dep_list(p):=v; type(p):=t;
if info(v)=null then
  begin vv:=value(v);
  if q=null then flush_cur_exp(vv)
  else  begin recycle_value(p); type(q):=known; value(q):=vv;
    end;
  end
else if q=null then cur_type:=t;
if fix_needed then fix_dependencies;
end;

@ Let's turn now to the six basic relations of comparison.

@<Additional cases of binary operators@>=
less_than,less_or_equal,greater_than,greater_or_equal,equal_to,unequal_to:
  begin@t@>@;
  if (cur_type>pair_type)and(type(p)>pair_type) then
    add_or_subtract(p,null,minus) {|cur_exp:=(p)-cur_exp|}
  else if cur_type<>type(p) then
    begin bad_binary(p,c); goto done;
    end
  else if cur_type=string_type then
    flush_cur_exp(str_vs_str(value(p),cur_exp))
  else if (cur_type=unknown_string)or(cur_type=unknown_boolean) then
    @<Check if unknowns have been equated@>
  else if (cur_type=pair_type)or(cur_type=transform_type) then
    @<Reduce comparison of big nodes to comparison of scalars@>
  else if cur_type=boolean_type then flush_cur_exp(cur_exp-value(p))
  else  begin bad_binary(p,c); goto done;
    end;
  @<Compare the current expression with zero@>;
done:  end;

@ @<Compare the current expression with zero@>=
if cur_type<>known then
  begin if cur_type<known then
    begin disp_err(p,"");
    help1("The quantities shown above have not been equated.")@/
    end
  else  help2("Oh dear. I can't decide if the expression above is positive,")@/
    ("negative, or zero. So this comparison test won't be `true'.");
  exp_err("Unknown relation will be considered false");
@.Unknown relation...@>
  put_get_flush_error(false_code);
  end
else case c of
  less_than: boolean_reset(cur_exp<0);
  less_or_equal: boolean_reset(cur_exp<=0);
  greater_than: boolean_reset(cur_exp>0);
  greater_or_equal: boolean_reset(cur_exp>=0);
  equal_to: boolean_reset(cur_exp=0);
  unequal_to: boolean_reset(cur_exp<>0);
  end; {there are no other cases}
 cur_type:=boolean_type

@ When two unknown strings are in the same ring, we know that they are
equal. Otherwise, we don't know whether they are equal or not, so we
make no change.

@<Check if unknowns have been equated@>=
begin q:=value(cur_exp);
while (q<>cur_exp)and(q<>p) do q:=value(q);
if q=p then flush_cur_exp(0);
end

@ @<Reduce comparison of big nodes to comparison of scalars@>=
begin q:=value(p); r:=value(cur_exp);
rr:=r+big_node_size[cur_type]-2;
loop@+  begin add_or_subtract(q,r,minus);
  if type(r)<>known then goto done1;
  if value(r)<>0 then goto done1;
  if r=rr then goto done1;
  q:=q+2; r:=r+2;
  end;
done1:take_part(x_part+half(r-value(cur_exp)));
end

@ Here we use the sneaky fact that |and_op-false_code=or_op-true_code|.

@<Additional cases of binary operators@>=
and_op,or_op: if (type(p)<>boolean_type)or(cur_type<>boolean_type) then
    bad_binary(p,c)
  else if value(p)=c+false_code-and_op then cur_exp:=value(p);

@ @<Additional cases of binary operators@>=
times: if (cur_type<pair_type)or(type(p)<pair_type) then bad_binary(p,times)
  else if (cur_type=known)or(type(p)=known) then
    @<Multiply when at least one operand is known@>
  else if (nice_pair(p,type(p))and(cur_type>pair_type))
      or(nice_pair(cur_exp,cur_type)and(type(p)>pair_type)) then
    begin hard_times(p); return;
    end
  else bad_binary(p,times);

@ @<Multiply when at least one operand is known@>=
begin if type(p)=known then
  begin v:=value(p); free_node(p,value_node_size);
  end
else  begin v:=cur_exp; unstash_cur_exp(p);
  end;
if cur_type=known then cur_exp:=take_scaled(cur_exp,v)
else if cur_type=pair_type then
  begin p:=value(cur_exp);
  dep_mult(x_part_loc(p),v,true);
  dep_mult(y_part_loc(p),v,true);
  end
else dep_mult(null,v,true);
return;
end

@ @<Declare binary action...@>=
procedure dep_mult(@!p:pointer;@!v:integer;@!v_is_scaled:boolean);
label exit;
var @!q:pointer; {the dependency list being multiplied by |v|}
@!s,@!t:small_number; {its type, before and after}
begin if p=null then q:=cur_exp
else if type(p)<>known then q:=p
else  begin if v_is_scaled then value(p):=take_scaled(value(p),v)
  else value(p):=take_fraction(value(p),v);
  return;
  end;
t:=type(q); q:=dep_list(q); s:=t;
if t=dependent then if v_is_scaled then
  if ab_vs_cd(max_coef(q),abs(v),coef_bound-1,unity)>=0 then t:=proto_dependent;
q:=p_times_v(q,v,s,t,v_is_scaled); dep_finish(q,p,t);
exit:end;

@ Here is a routine that is similar to |times|; but it is invoked only
internally, when |v| is a |fraction| whose magnitude is at most~1,
and when |cur_type>=pair_type|.

@p procedure frac_mult(@!n,@!d:scaled); {multiplies |cur_exp| by |n/d|}
var @!p:pointer; {a pair node}
@!old_exp:pointer; {a capsule to recycle}
@!v:fraction; {|n/d|}
begin if internal[tracing_commands]>two then
  @<Trace the fraction multiplication@>;
case cur_type of
transform_type,pair_type:old_exp:=tarnished(cur_exp);
independent:old_exp:=void;
othercases old_exp:=null
endcases;
if old_exp<>null then
  begin old_exp:=cur_exp; make_exp_copy(old_exp);
  end;
v:=make_fraction(n,d);
if cur_type=known then cur_exp:=take_fraction(cur_exp,v)
else if cur_type=pair_type then
  begin p:=value(cur_exp);
  dep_mult(x_part_loc(p),v,false);
  dep_mult(y_part_loc(p),v,false);
  end
else dep_mult(null,v,false);
if old_exp<>null then
  begin recycle_value(old_exp); free_node(old_exp,value_node_size);
  end
end;

@ @<Trace the fraction multiplication@>=
begin begin_diagnostic; print_nl("{("); print_scaled(n); print_char("/");
print_scaled(d); print(")*("); print_exp(null,0); print(")}");
end_diagnostic(false);
end

@ The |hard_times| routine multiplies a nice pair by a dependency list.

@<Declare binary action procedures@>=
procedure hard_times(@!p:pointer);
var @!q:pointer; {a copy of the dependent variable |p|}
@!r:pointer; {the big node for the nice pair}
@!u,@!v:scaled; {the known values of the nice pair}
begin if type(p)=pair_type then
  begin q:=stash_cur_exp; unstash_cur_exp(p); p:=q;
  end; {now |cur_type=pair_type|}
r:=value(cur_exp); u:=value(x_part_loc(r)); v:=value(y_part_loc(r));
@<Move the dependent variable |p| into both parts of the pair node |r|@>;
dep_mult(x_part_loc(r),u,true); dep_mult(y_part_loc(r),v,true);
end;

@ @<Move the dependent variable |p|...@>=
type(y_part_loc(r)):=type(p);
new_dep(y_part_loc(r),copy_dep_list(dep_list(p)));@/
type(x_part_loc(r)):=type(p);
mem[value_loc(x_part_loc(r))]:=mem[value_loc(p)];
link(prev_dep(p)):=x_part_loc(r);
free_node(p,value_node_size)

@ @<Additional cases of binary operators@>=
over: if (cur_type<>known)or(type(p)<pair_type) then bad_binary(p,over)
  else  begin v:=cur_exp; unstash_cur_exp(p);
    if v=0 then @<Squeal about division by zero@>
    else  begin if cur_type=known then cur_exp:=make_scaled(cur_exp,v)
      else if cur_type=pair_type then
        begin p:=value(cur_exp);
        dep_div(x_part_loc(p),v);
        dep_div(y_part_loc(p),v);
        end
      else dep_div(null,v);
      end;
    return;
    end;

@ @<Declare binary action...@>=
procedure dep_div(@!p:pointer;@!v:scaled);
label exit;
var @!q:pointer; {the dependency list being divided by |v|}
@!s,@!t:small_number; {its type, before and after}
begin if p=null then q:=cur_exp
else if type(p)<>known then q:=p
else  begin value(p):=make_scaled(value(p),v); return;
  end;
t:=type(q); q:=dep_list(q); s:=t;
if t=dependent then
  if ab_vs_cd(max_coef(q),unity,coef_bound-1,abs(v))>=0 then t:=proto_dependent;
q:=p_over_v(q,v,s,t); dep_finish(q,p,t);
exit:end;

@ @<Squeal about division by zero@>=
begin exp_err("Division by zero");
@.Division by zero@>
help2("You're trying to divide the quantity shown above the error")@/
  ("message by zero. I'm going to divide it by one instead.");
put_get_error;
end

@ @<Additional cases of binary operators@>=
pythag_add,pythag_sub: if (cur_type=known)and(type(p)=known) then
    if c=pythag_add then cur_exp:=pyth_add(value(p),cur_exp)
    else cur_exp:=pyth_sub(value(p),cur_exp)
  else bad_binary(p,c);

@ The next few sections of the program deal with affine transformations
of coordinate data.

@<Additional cases of binary operators@>=
rotated_by,slanted_by,scaled_by,shifted_by,transformed_by,
 x_scaled,y_scaled,z_scaled: @t@>@;@/
  if (type(p)=path_type)or(type(p)=future_pen)or(type(p)=pen_type) then
    begin path_trans(p,c); return;
    end
  else if (type(p)=pair_type)or(type(p)=transform_type) then big_trans(p,c)
  else if type(p)=picture_type then
    begin edges_trans(p,c); return;
    end
  else bad_binary(p,c);

@ Let |c| be one of the eight transform operators. The procedure call
|set_up_trans(c)| first changes |cur_exp| to a transform that corresponds to
|c| and the original value of |cur_exp|. (In particular, |cur_exp| doesn't
change at all if |c=transformed_by|.)

Then, if all components of the resulting transform are |known|, they are
moved to the global variables |txx|, |txy|, |tyx|, |tyy|, |tx|, |ty|;
and |cur_exp| is changed to the known value zero.

@<Declare binary action...@>=
procedure set_up_trans(@!c:quarterword);
label done,exit;
var @!p,@!q,@!r:pointer; {list manipulation registers}
begin if (c<>transformed_by)or(cur_type<>transform_type) then
  @<Put the current transform into |cur_exp|@>;
@<If the current transform is entirely known, stash it in global variables;
  otherwise |return|@>;
exit:end;

@ @<Glob...@>=
@!txx,@!txy,@!tyx,@!tyy,@!tx,@!ty:scaled; {current transform coefficients}

@ @<Put the current transform...@>=
begin p:=stash_cur_exp; cur_exp:=id_transform; cur_type:=transform_type;
q:=value(cur_exp);
case c of
@<For each of the eight cases, change the relevant fields of |cur_exp|
  and |goto done|;
  but do nothing if capsule |p| doesn't have the appropriate type@>@;
end; {there are no other cases}
disp_err(p,"Improper transformation argument");
@.Improper transformation argument@>
help3("The expression shown above has the wrong type,")@/
  ("so I can't transform anything using it.")@/
  ("Proceed, and I'll omit the transformation.");
put_get_error;
done: recycle_value(p); free_node(p,value_node_size);
end

@ @<If the current transform is entirely known, ...@>=
q:=value(cur_exp); r:=q+transform_node_size;
repeat r:=r-2;
if type(r)<>known then return;
until r=q;
txx:=value(xx_part_loc(q));
txy:=value(xy_part_loc(q));
tyx:=value(yx_part_loc(q));
tyy:=value(yy_part_loc(q));
tx:=value(x_part_loc(q));
ty:=value(y_part_loc(q));
flush_cur_exp(0)

@ @<For each of the eight cases...@>=
rotated_by:if type(p)=known then
  @<Install sines and cosines, then |goto done|@>;
slanted_by:if type(p)>pair_type then
  begin install(xy_part_loc(q),p); goto done;
  end;
scaled_by:if type(p)>pair_type then
  begin install(xx_part_loc(q),p); install(yy_part_loc(q),p); goto done;
  end;
shifted_by:if type(p)=pair_type then
  begin r:=value(p); install(x_part_loc(q),x_part_loc(r));
  install(y_part_loc(q),y_part_loc(r)); goto done;
  end;
x_scaled:if type(p)>pair_type then
  begin install(xx_part_loc(q),p); goto done;
  end;
y_scaled:if type(p)>pair_type then
  begin install(yy_part_loc(q),p); goto done;
  end;
z_scaled:if type(p)=pair_type then
  @<Install a complex multiplier, then |goto done|@>;
transformed_by:do_nothing;

@ @<Install sines and cosines, then |goto done|@>=
begin n_sin_cos((value(p) mod three_sixty_units)*16);
value(xx_part_loc(q)):=round_fraction(n_cos);
value(yx_part_loc(q)):=round_fraction(n_sin);
value(xy_part_loc(q)):=-value(yx_part_loc(q));
value(yy_part_loc(q)):=value(xx_part_loc(q));
goto done;
end

@ @<Install a complex multiplier, then |goto done|@>=
begin r:=value(p);
install(xx_part_loc(q),x_part_loc(r));
install(yy_part_loc(q),x_part_loc(r));
install(yx_part_loc(q),y_part_loc(r));
if type(y_part_loc(r))=known then negate(value(y_part_loc(r)))
else negate_dep_list(dep_list(y_part_loc(r)));
install(xy_part_loc(q),y_part_loc(r));
goto done;
end

@ Procedure |set_up_known_trans| is like |set_up_trans|, but it
insists that the transformation be entirely known.

@<Declare binary action...@>=
procedure set_up_known_trans(@!c:quarterword);
begin set_up_trans(c);
if cur_type<>known then
  begin exp_err("Transform components aren't all known");
@.Transform components...@>
  help3("I'm unable to apply a partially specified transformation")@/
    ("except to a fully known pair or transform.")@/
    ("Proceed, and I'll omit the transformation.");
  put_get_flush_error(0);
  txx:=unity; txy:=0; tyx:=0; tyy:=unity; tx:=0; ty:=0;
  end;
end;

@ Here's a procedure that applies the transform |txx..ty| to a pair of
coordinates in locations |p| and~|q|.

@<Declare binary action...@>=
procedure trans(@!p,@!q:pointer);
var @!v:scaled; {the new |x| value}
begin v:=take_scaled(mem[p].sc,txx)+take_scaled(mem[q].sc,txy)+tx;
mem[q].sc:=take_scaled(mem[p].sc,tyx)+take_scaled(mem[q].sc,tyy)+ty;
mem[p].sc:=v;
end;

@ The simplest transformation procedure applies a transform to all
coordinates of a path. The |null_pen| remains unchanged if it isn't
being shifted.

@<Declare binary action...@>=
procedure path_trans(@!p:pointer;@!c:quarterword);
label exit;
var @!q:pointer; {list traverser}
begin set_up_known_trans(c); unstash_cur_exp(p);
if cur_type=pen_type then
  begin if max_offset(cur_exp)=0 then if tx=0 then if ty=0 then return;
  flush_cur_exp(make_path(cur_exp)); cur_type:=future_pen;
  end;
q:=cur_exp;
repeat if left_type(q)<>endpoint then
  trans(q+3,q+4); {that's |left_x| and |left_y|}
trans(q+1,q+2); {that's |x_coord| and |y_coord|}
if right_type(q)<>endpoint then
  trans(q+5,q+6); {that's |right_x| and |right_y|}
q:=link(q);
until q=cur_exp;
exit:end;

@ The next simplest transformation procedure applies to edges.
It is simple primarily because \MF\ doesn't allow very general
transformations to be made, and because the tricky subroutines
for edge transformation have already been written.

@<Declare binary action...@>=
procedure edges_trans(@!p:pointer;@!c:quarterword);
label exit;
begin set_up_known_trans(c); unstash_cur_exp(p); cur_edges:=cur_exp;
if empty_edges(cur_edges) then return; {the empty set is easy to transform}
if txx=0 then if tyy=0 then
 if txy mod unity=0 then if tyx mod unity=0 then
  begin xy_swap_edges; txx:=txy; tyy:=tyx; txy:=0; tyx:=0;
  if empty_edges(cur_edges) then return;
  end;
if txy=0 then if tyx=0 then
 if txx mod unity=0 then if tyy mod unity=0 then
  @<Scale the edges, shift them, and |return|@>;
print_err("That transformation is too hard");
@.That transformation...@>
help3("I can apply complicated transformations to paths,")@/
  ("but I can only do integer operations on pictures.")@/
  ("Proceed, and I'll omit the transformation.");
put_get_error;
exit:end;

@ @<Scale the edges, shift them, and |return|@>=
begin if (txx=0)or(tyy=0) then
  begin toss_edges(cur_edges);
  cur_exp:=get_node(edge_header_size); init_edges(cur_exp);
  end
else  begin if txx<0 then
    begin x_reflect_edges; txx:=-txx;
    end;
  if tyy<0 then
    begin y_reflect_edges; tyy:=-tyy;
    end;
  if txx<>unity then x_scale_edges(txx div unity);
  if tyy<>unity then y_scale_edges(tyy div unity);
  @<Shift the edges by |(tx,ty)|, rounded@>;
  end;
return;
end

@ @<Shift the edges...@>=
tx:=round_unscaled(tx); ty:=round_unscaled(ty);
if (m_min(cur_edges)+tx<=0)or(m_max(cur_edges)+tx>=8192)or@|
 (n_min(cur_edges)+ty<=0)or(n_max(cur_edges)+ty>=8191)or@|
 (abs(tx)>=4096)or(abs(ty)>=4096) then
  begin print_err("Too far to shift");
@.Too far to shift@>
  help3("I can't shift the picture as requested---it would")@/
    ("make some coordinates too large or too small.")@/
    ("Proceed, and I'll omit the transformation.");
  put_get_error;
  end
else  begin if tx<>0 then
    begin if not valid_range(m_offset(cur_edges)-tx) then fix_offset;
    m_min(cur_edges):=m_min(cur_edges)+tx;
    m_max(cur_edges):=m_max(cur_edges)+tx;
    m_offset(cur_edges):=m_offset(cur_edges)-tx;
    last_window_time(cur_edges):=0;
    end;
  if ty<>0 then
    begin n_min(cur_edges):=n_min(cur_edges)+ty;
    n_max(cur_edges):=n_max(cur_edges)+ty;
    n_pos(cur_edges):=n_pos(cur_edges)+ty;
    last_window_time(cur_edges):=0;
    end;
  end

@ The hard cases of transformation occur when big nodes are involved,
and when some of their components are unknown.

@<Declare binary action...@>=
@t\4@>@<Declare subroutines needed by |big_trans|@>@;
procedure big_trans(@!p:pointer;@!c:quarterword);
label exit;
var @!q,@!r,@!pp,@!qq:pointer; {list manipulation registers}
@!s:small_number; {size of a big node}
begin s:=big_node_size[type(p)]; q:=value(p); r:=q+s;
repeat r:=r-2;
if type(r)<>known then @<Transform an unknown big node and |return|@>;
until r=q;
@<Transform a known big node@>;
exit:end; {node |p| will now be recycled by |do_binary|}

@ @<Transform an unknown big node and |return|@>=
begin set_up_known_trans(c); make_exp_copy(p); r:=value(cur_exp);
if cur_type=transform_type then
  begin bilin1(yy_part_loc(r),tyy,xy_part_loc(q),tyx,0);
  bilin1(yx_part_loc(r),tyy,xx_part_loc(q),tyx,0);
  bilin1(xy_part_loc(r),txx,yy_part_loc(q),txy,0);
  bilin1(xx_part_loc(r),txx,yx_part_loc(q),txy,0);
  end;
bilin1(y_part_loc(r),tyy,x_part_loc(q),tyx,ty);
bilin1(x_part_loc(r),txx,y_part_loc(q),txy,tx);
return;
end

@ Let |p| point to a two-word value field inside a big node of |cur_exp|,
and let |q| point to a another value field. The |bilin1| procedure
replaces |p| by $p\cdot t+q\cdot u+\delta$.

@<Declare subroutines needed by |big_trans|@>=
procedure bilin1(@!p:pointer;@!t:scaled;@!q:pointer;@!u,@!delta:scaled);
var @!r:pointer; {list traverser}
begin if t<>unity then dep_mult(p,t,true);
if u<>0 then
  if type(q)=known then delta:=delta+take_scaled(value(q),u)
  else  begin @<Ensure that |type(p)=proto_dependent|@>;
    dep_list(p):=p_plus_fq(dep_list(p),u,dep_list(q),proto_dependent,type(q));
    end;
if type(p)=known then value(p):=value(p)+delta
else  begin r:=dep_list(p);
  while info(r)<>null do r:=link(r);
  delta:=value(r)+delta;
  if r<>dep_list(p) then value(r):=delta
  else  begin recycle_value(p); type(p):=known; value(p):=delta;
    end;
  end;
if fix_needed then fix_dependencies;
end;

@ @<Ensure that |type(p)=proto_dependent|@>=
if type(p)<>proto_dependent then
  begin if type(p)=known then new_dep(p,const_dependency(value(p)))
  else dep_list(p):=p_times_v(dep_list(p),unity,dependent,proto_dependent,true);
  type(p):=proto_dependent;
  end

@ @<Transform a known big node@>=
set_up_trans(c);
if cur_type=known then @<Transform known by known@>
else  begin pp:=stash_cur_exp; qq:=value(pp);
  make_exp_copy(p); r:=value(cur_exp);
  if cur_type=transform_type then
    begin bilin2(yy_part_loc(r),yy_part_loc(qq),
      value(xy_part_loc(q)),yx_part_loc(qq),null);
    bilin2(yx_part_loc(r),yy_part_loc(qq),
      value(xx_part_loc(q)),yx_part_loc(qq),null);
    bilin2(xy_part_loc(r),xx_part_loc(qq),
      value(yy_part_loc(q)),xy_part_loc(qq),null);
    bilin2(xx_part_loc(r),xx_part_loc(qq),
      value(yx_part_loc(q)),xy_part_loc(qq),null);
    end;
  bilin2(y_part_loc(r),yy_part_loc(qq),
    value(x_part_loc(q)),yx_part_loc(qq),y_part_loc(qq));
  bilin2(x_part_loc(r),xx_part_loc(qq),
    value(y_part_loc(q)),xy_part_loc(qq),x_part_loc(qq));
  recycle_value(pp); free_node(pp,value_node_size);
  end;

@ Let |p| be a |proto_dependent| value whose dependency list ends
at |dep_final|. The following procedure adds |v| times another
numeric quantity to~|p|.

@<Declare subroutines needed by |big_trans|@>=
procedure add_mult_dep(@!p:pointer;@!v:scaled;@!r:pointer);
begin if type(r)=known then
  value(dep_final):=value(dep_final)+take_scaled(value(r),v)
else  begin dep_list(p):=
   p_plus_fq(dep_list(p),v,dep_list(r),proto_dependent,type(r));
  if fix_needed then fix_dependencies;
  end;
end;

@ The |bilin2| procedure is something like |bilin1|, but with known
and unknown quantities reversed. Parameter |p| points to a value field
within the big node for |cur_exp|; and |type(p)=known|. Parameters
|t| and~|u| point to value fields elsewhere; so does parameter~|q|,
unless it is |null| (which stands for zero). Location~|p| will be
replaced by $p\cdot t+v\cdot u+q$.

@<Declare subroutines needed by |big_trans|@>=
procedure bilin2(@!p,@!t:pointer;@!v:scaled;@!u,@!q:pointer);
var @!vv:scaled; {temporary storage for |value(p)|}
begin vv:=value(p); type(p):=proto_dependent;
new_dep(p,const_dependency(0)); {this sets |dep_final|}
if vv<>0 then add_mult_dep(p,vv,t); {|dep_final| doesn't change}
if v<>0 then add_mult_dep(p,v,u);
if q<>null then add_mult_dep(p,unity,q);
if dep_list(p)=dep_final then
  begin vv:=value(dep_final); recycle_value(p);
  type(p):=known; value(p):=vv;
  end;
end;

@ @<Transform known by known@>=
begin make_exp_copy(p); r:=value(cur_exp);
if cur_type=transform_type then
  begin bilin3(yy_part_loc(r),tyy,value(xy_part_loc(q)),tyx,0);
  bilin3(yx_part_loc(r),tyy,value(xx_part_loc(q)),tyx,0);
  bilin3(xy_part_loc(r),txx,value(yy_part_loc(q)),txy,0);
  bilin3(xx_part_loc(r),txx,value(yx_part_loc(q)),txy,0);
  end;
bilin3(y_part_loc(r),tyy,value(x_part_loc(q)),tyx,ty);
bilin3(x_part_loc(r),txx,value(y_part_loc(q)),txy,tx);
end

@ Finally, in |bilin3| everything is |known|.

@<Declare subroutines needed by |big_trans|@>=
procedure bilin3(@!p:pointer;@!t,@!v,@!u,@!delta:scaled);
begin if t<>unity then delta:=delta+take_scaled(value(p),t)
else delta:=delta+value(p);
if u<>0 then value(p):=delta+take_scaled(v,u)
else value(p):=delta;
end;

@ @<Additional cases of binary operators@>=
concatenate: if (cur_type=string_type)and(type(p)=string_type) then cat(p)
  else bad_binary(p,concatenate);
substring_of: if nice_pair(p,type(p))and(cur_type=string_type) then
    chop_string(value(p))
  else bad_binary(p,substring_of);
subpath_of: begin if cur_type=pair_type then pair_to_path;
  if nice_pair(p,type(p))and(cur_type=path_type) then
    chop_path(value(p))
  else bad_binary(p,subpath_of);
  end;

@ @<Declare binary action...@>=
procedure cat(@!p:pointer);
var @!a,@!b:str_number; {the strings being concatenated}
@!k:pool_pointer; {index into |str_pool|}
begin a:=value(p); b:=cur_exp; str_room(length(a)+length(b));
for k:=str_start[a] to str_start[a+1]-1 do append_char(so(str_pool[k]));
for k:=str_start[b] to str_start[b+1]-1 do append_char(so(str_pool[k]));
cur_exp:=make_string; delete_str_ref(b);
end;

@ @<Declare binary action...@>=
procedure chop_string(@!p:pointer);
var @!a,@!b:integer; {start and stop points}
@!l:integer; {length of the original string}
@!k:integer; {runs from |a| to |b|}
@!s:str_number; {the original string}
@!reversed:boolean; {was |a>b|?}
begin a:=round_unscaled(value(x_part_loc(p)));
b:=round_unscaled(value(y_part_loc(p)));
if a<=b then reversed:=false
else  begin reversed:=true; k:=a; a:=b; b:=k;
  end;
s:=cur_exp; l:=length(s);
if a<0 then
  begin a:=0;
  if b<0 then b:=0;
  end;
if b>l then
  begin b:=l;
  if a>l then a:=l;
  end;
str_room(b-a);
if reversed then
  for k:=str_start[s]+b-1 downto str_start[s]+a do append_char(so(str_pool[k]))
else for k:=str_start[s]+a to str_start[s]+b-1 do append_char(so(str_pool[k]));
cur_exp:=make_string; delete_str_ref(s);
end;

@ @<Declare binary action...@>=
procedure chop_path(@!p:pointer);
var @!q:pointer; {a knot in the original path}
@!pp,@!qq,@!rr,@!ss:pointer; {link variables for copies of path nodes}
@!a,@!b,@!k,@!l:scaled; {indices for chopping}
@!reversed:boolean; {was |a>b|?}
begin l:=path_length; a:=value(x_part_loc(p)); b:=value(y_part_loc(p));
if a<=b then reversed:=false
else  begin reversed:=true; k:=a; a:=b; b:=k;
  end;
@<Dispense with the cases |a<0| and/or |b>l|@>;
q:=cur_exp;
while a>=unity do
  begin q:=link(q); a:=a-unity; b:=b-unity;
  end;
if b=a then @<Construct a path from |pp| to |qq| of length zero@>
else @<Construct a path from |pp| to |qq| of length $\lceil b\rceil$@>;
left_type(pp):=endpoint; right_type(qq):=endpoint; link(qq):=pp;
toss_knot_list(cur_exp);
if reversed then
  begin cur_exp:=link(htap_ypoc(pp)); toss_knot_list(pp);
  end
else cur_exp:=pp;
end;

@ @<Dispense with the cases |a<0| and/or |b>l|@>=
if a<0 then
  if left_type(cur_exp)=endpoint then
    begin a:=0; if b<0 then b:=0;
    end
  else  repeat a:=a+l; b:=b+l;
    until a>=0; {a cycle always has length |l>0|}
if b>l then if left_type(cur_exp)=endpoint then
    begin b:=l; if a>l then a:=l;
    end
  else while a>=l do
    begin a:=a-l; b:=b-l;
    end

@ @<Construct a path from |pp| to |qq| of length $\lceil b\rceil$@>=
begin pp:=copy_knot(q); qq:=pp;
repeat q:=link(q); rr:=qq; qq:=copy_knot(q); link(rr):=qq; b:=b-unity;
until b<=0;
if a>0 then
  begin ss:=pp; pp:=link(pp);
  split_cubic(ss,a*@'10000,x_coord(pp),y_coord(pp)); pp:=link(ss);
  free_node(ss,knot_node_size);
  if rr=ss then
    begin b:=make_scaled(b,unity-a); rr:=pp;
    end;
  end;
if b<0 then
  begin split_cubic(rr,(b+unity)*@'10000,x_coord(qq),y_coord(qq));
  free_node(qq,knot_node_size);
  qq:=link(rr);
  end;
end

@ @<Construct a path from |pp| to |qq| of length zero@>=
begin if a>0 then
  begin qq:=link(q);
  split_cubic(q,a*@'10000,x_coord(qq),y_coord(qq)); q:=link(q);
  end;
pp:=copy_knot(q); qq:=pp;
end

@ The |pair_value| routine changes the current expression to a
given ordered pair of values.

@<Declare binary action...@>=
procedure pair_value(@!x,@!y:scaled);
var @!p:pointer; {a pair node}
begin p:=get_node(value_node_size); flush_cur_exp(p); cur_type:=pair_type;
type(p):=pair_type; name_type(p):=capsule; init_big_node(p);
p:=value(p);@/
type(x_part_loc(p)):=known; value(x_part_loc(p)):=x;@/
type(y_part_loc(p)):=known; value(y_part_loc(p)):=y;@/
end;

@ @<Additional cases of binary operators@>=
point_of,precontrol_of,postcontrol_of: begin if cur_type=pair_type then
     pair_to_path;
  if (cur_type=path_type)and(type(p)=known) then
    find_point(value(p),c)
  else bad_binary(p,c);
  end;
pen_offset_of: begin if cur_type=future_pen then materialize_pen;
  if (cur_type=pen_type)and nice_pair(p,type(p)) then
    set_up_offset(value(p))
  else bad_binary(p,pen_offset_of);
  end;
direction_time_of: begin if cur_type=pair_type then pair_to_path;
  if (cur_type=path_type)and nice_pair(p,type(p)) then
    set_up_direction_time(value(p))
  else bad_binary(p,direction_time_of);
  end;

@ @<Declare binary action...@>=
procedure set_up_offset(@!p:pointer);
begin find_offset(value(x_part_loc(p)),value(y_part_loc(p)),cur_exp);
pair_value(cur_x,cur_y);
end;
@#
procedure set_up_direction_time(@!p:pointer);
begin flush_cur_exp(find_direction_time(value(x_part_loc(p)),
  value(y_part_loc(p)),cur_exp));
end;

@ @<Declare binary action...@>=
procedure find_point(@!v:scaled;@!c:quarterword);
var @!p:pointer; {the path}
@!n:scaled; {its length}
@!q:pointer; {successor of |p|}
begin p:=cur_exp;@/
if left_type(p)=endpoint then n:=-unity@+else n:=0;
repeat p:=link(p); n:=n+unity;
until p=cur_exp;
if n=0 then v:=0
else if v<0 then
  if left_type(p)=endpoint then v:=0
  else v:=n-1-((-v-1) mod n)
else if v>n then
  if left_type(p)=endpoint then v:=n
  else v:=v mod n;
p:=cur_exp;
while v>=unity do
  begin p:=link(p); v:=v-unity;
  end;
if v<>0 then @<Insert a fractional node by splitting the cubic@>;
@<Set the current expression to the desired path coordinates@>;
end;

@ @<Insert a fractional node...@>=
begin q:=link(p); split_cubic(p,v*@'10000,x_coord(q),y_coord(q)); p:=link(p);
end

@ @<Set the current expression to the desired path coordinates...@>=
case c of
point_of: pair_value(x_coord(p),y_coord(p));
precontrol_of: if left_type(p)=endpoint then pair_value(x_coord(p),y_coord(p))
  else pair_value(left_x(p),left_y(p));
postcontrol_of: if right_type(p)=endpoint then pair_value(x_coord(p),y_coord(p))
  else pair_value(right_x(p),right_y(p));
end {there are no other cases}

@ @<Additional cases of bin...@>=
intersect: begin if type(p)=pair_type then
    begin q:=stash_cur_exp; unstash_cur_exp(p);
    pair_to_path; p:=stash_cur_exp; unstash_cur_exp(q);
    end;
  if cur_type=pair_type then pair_to_path;
  if (cur_type=path_type)and(type(p)=path_type) then
    begin path_intersection(value(p),cur_exp);
    pair_value(cur_t,cur_tt);
    end
  else bad_binary(p,intersect);
  end;

@* \[43] Statements and commands.
The chief executive of \MF\ is the |do_statement| routine, which
contains the master switch that causes all the various pieces of \MF\
to do their things, in the right order.

In a sense, this is the grand climax of the program: It applies all the
tools that we have worked so hard to construct. In another sense, this is
the messiest part of the program: It necessarily refers to other pieces
of code all over the place, so that a person can't fully understand what is
going on without paging back and forth to be reminded of conventions that
are defined elsewhere. We are now at the hub of the web.

The structure of |do_statement| itself is quite simple.  The first token
of the statement is fetched using |get_x_next|.  If it can be the first
token of an expression, we look for an equation, an assignment, or a
title. Otherwise we use a \&{case} construction to branch at high speed to
the appropriate routine for various and sundry other types of commands,
each of which has an ``action procedure'' that does the necessary work.

The program uses the fact that
$$\hbox{|min_primary_command=max_statement_command=type_name|}$$
to interpret a statement that starts with, e.g., `\&{string}',
as a type declaration rather than a boolean expression.

@p @t\4@>@<Declare generic font output procedures@>@;
@t\4@>@<Declare action procedures for use by |do_statement|@>@;
procedure do_statement; {governs \MF's activities}
begin cur_type:=vacuous; get_x_next;
if cur_cmd>max_primary_command then @<Worry about bad statement@>
else if cur_cmd>max_statement_command then
  @<Do an equation, assignment, title, or
   `$\langle\,$expression$\,\rangle\,$\&{endgroup}'@>
else @<Do a statement that doesn't begin with an expression@>;
if cur_cmd<semicolon then
  @<Flush unparsable junk that was found after the statement@>;
error_count:=0;
end;

@ The only command codes |>max_primary_command| that can be present
at the beginning of a statement are |semicolon| and higher; these
occur when the statement is null.

@<Worry about bad statement@>=
begin if cur_cmd<semicolon then
  begin print_err("A statement can't begin with `");
@.A statement can't begin with x@>
  print_cmd_mod(cur_cmd,cur_mod); print_char("'");
  help5("I was looking for the beginning of a new statement.")@/
    ("If you just proceed without changing anything, I'll ignore")@/
    ("everything up to the next `;'. Please insert a semicolon")@/
    ("now in front of anything that you don't want me to delete.")@/
    ("(See Chapter 27 of The METAFONTbook for an example.)");@/
@:METAFONTbook}{\sl The {\logos METAFONT\/}book@>
  back_error; get_x_next;
  end;
end

@ The help message printed here says that everything is flushed up to
a semicolon, but actually the commands |end_group| and |stop| will
also terminate a statement.

@<Flush unparsable junk that was found after the statement@>=
begin print_err("Extra tokens will be flushed");
@.Extra tokens will be flushed@>
help6("I've just read as much of that statement as I could fathom,")@/
("so a semicolon should have been next. It's very puzzling...")@/
("but I'll try to get myself back together, by ignoring")@/
("everything up to the next `;'. Please insert a semicolon")@/
("now in front of anything that you don't want me to delete.")@/
("(See Chapter 27 of The METAFONTbook for an example.)");@/
@:METAFONTbook}{\sl The {\logos METAFONT\/}book@>
back_error; scanner_status:=flushing;
repeat get_next;
@<Decrease the string reference count...@>;
until end_of_statement; {|cur_cmd=semicolon|, |end_group|, or |stop|}
scanner_status:=normal;
end

@ If |do_statement| ends with |cur_cmd=end_group|, we should have
|cur_type=vacuous| unless the statement was simply an expression;
in the latter case, |cur_type| and |cur_exp| should represent that
expression.

@<Do a statement that doesn't...@>=
begin if internal[tracing_commands]>0 then show_cur_cmd_mod;
case cur_cmd of
type_name:do_type_declaration;
macro_def:if cur_mod>var_def then make_op_def
  else if cur_mod>end_def then scan_def;
@t\4@>@<Cases of |do_statement| that invoke particular commands@>@;
end; {there are no other cases}
cur_type:=vacuous;
end

@ The most important statements begin with expressions.

@<Do an equation, assignment, title, or...@>=
begin var_flag:=assignment; scan_expression;
if cur_cmd<end_group then
  begin if cur_cmd=equals then do_equation
  else if cur_cmd=assignment then do_assignment
  else if cur_type=string_type then @<Do a title@>
  else if cur_type<>vacuous then
    begin exp_err("Isolated expression");
@.Isolated expression@>
    help3("I couldn't find an `=' or `:=' after the")@/
      ("expression that is shown above this error message,")@/
      ("so I guess I'll just ignore it and carry on.");
    put_get_error;
    end;
  flush_cur_exp(0); cur_type:=vacuous;
  end;
end

@ @<Do a title@>=
begin if internal[tracing_titles]>0 then
  begin print_nl(""); slow_print(cur_exp); update_terminal;
  end;
if internal[proofing]>0 then
  @<Send the current expression as a title to the output file@>;
end

@ Equations and assignments are performed by the pair of mutually recursive
@^recursion@>
routines |do_equation| and |do_assignment|. These routines are called when
|cur_cmd=equals| and when |cur_cmd=assignment|, respectively; the left-hand
side is in |cur_type| and |cur_exp|, while the right-hand side is yet
to be scanned. After the routines are finished, |cur_type| and |cur_exp|
will be equal to the right-hand side (which will normally be equal
to the left-hand side).

@<Declare action procedures for use by |do_statement|@>=
@t\4@>@<Declare the procedure called |try_eq|@>@;
@t\4@>@<Declare the procedure called |make_eq|@>@;
procedure@?do_assignment; forward;@t\2@>@/
procedure do_equation;
var @!lhs:pointer; {capsule for the left-hand side}
@!p:pointer; {temporary register}
begin lhs:=stash_cur_exp; get_x_next; var_flag:=assignment; scan_expression;
if cur_cmd=equals then do_equation
else if cur_cmd=assignment then do_assignment;
if internal[tracing_commands]>two then @<Trace the current equation@>;
if cur_type=unknown_path then if type(lhs)=pair_type then
  begin p:=stash_cur_exp; unstash_cur_exp(lhs); lhs:=p;
  end; {in this case |make_eq| will change the pair to a path}
make_eq(lhs); {equate |lhs| to |(cur_type,cur_exp)|}
end;

@ And |do_assignment| is similar to |do_equation|:

@<Declare action procedures for use by |do_statement|@>=
procedure do_assignment;
var @!lhs:pointer; {token list for the left-hand side}
@!p:pointer; {where the left-hand value is stored}
@!q:pointer; {temporary capsule for the right-hand value}
begin if cur_type<>token_list then
  begin exp_err("Improper `:=' will be changed to `='");
@.Improper `:='@>
  help2("I didn't find a variable name at the left of the `:=',")@/
    ("so I'm going to pretend that you said `=' instead.");@/
  error; do_equation;
  end
else  begin lhs:=cur_exp; cur_type:=vacuous;@/
  get_x_next; var_flag:=assignment; scan_expression;
  if cur_cmd=equals then do_equation
  else if cur_cmd=assignment then do_assignment;
  if internal[tracing_commands]>two then @<Trace the current assignment@>;
  if info(lhs)>hash_end then
    @<Assign the current expression to an internal variable@>
  else @<Assign the current expression to the variable |lhs|@>;
  flush_node_list(lhs);
  end;
end;

@ @<Trace the current equation@>=
begin begin_diagnostic; print_nl("{("); print_exp(lhs,0);
print(")=("); print_exp(null,0); print(")}"); end_diagnostic(false);
end

@ @<Trace the current assignment@>=
begin begin_diagnostic; print_nl("{");
if info(lhs)>hash_end then slow_print(int_name[info(lhs)-(hash_end)])
else show_token_list(lhs,null,1000,0);
print(":="); print_exp(null,0); print_char("}"); end_diagnostic(false);
end

@ @<Assign the current expression to an internal variable@>=
if cur_type=known then internal[info(lhs)-(hash_end)]:=cur_exp
else  begin exp_err("Internal quantity `");
@.Internal quantity...@>
  slow_print(int_name[info(lhs)-(hash_end)]);
  print("' must receive a known value");
  help2("I can't set an internal quantity to anything but a known")@/
    ("numeric value, so I'll have to ignore this assignment.");
  put_get_error;
  end

@ @<Assign the current expression to the variable |lhs|@>=
begin p:=find_variable(lhs);
if p<>null then
  begin q:=stash_cur_exp; cur_type:=und_type(p); recycle_value(p);
  type(p):=cur_type; value(p):=null; make_exp_copy(p);
  p:=stash_cur_exp; unstash_cur_exp(q); make_eq(p);
  end
else  begin obliterated(lhs); put_get_error;
  end;
end


@ And now we get to the nitty-gritty. The |make_eq| procedure is given
a pointer to a capsule that is to be equated to the current expression.

@<Declare the procedure called |make_eq|@>=
procedure make_eq(@!lhs:pointer);
label restart,done, not_found;
var @!t:small_number; {type of the left-hand side}
@!v:integer; {value of the left-hand side}
@!p,@!q:pointer; {pointers inside of big nodes}
begin restart: t:=type(lhs);
if t<=pair_type then v:=value(lhs);
case t of
@t\4@>@<For each type |t|, make an equation and |goto done| unless |cur_type|
  is incompatible with~|t|@>@;
end; {all cases have been listed}
@<Announce that the equation cannot be performed@>;
done:check_arith; recycle_value(lhs); free_node(lhs,value_node_size);
end;

@ @<Announce that the equation cannot be performed@>=
disp_err(lhs,""); exp_err("Equation cannot be performed (");
@.Equation cannot be performed@>
if type(lhs)<=pair_type then print_type(type(lhs))@+else print("numeric");
print_char("=");
if cur_type<=pair_type then print_type(cur_type)@+else print("numeric");
print_char(")");@/
help2("I'm sorry, but I don't know how to make such things equal.")@/
  ("(See the two expressions just above the error message.)");
put_get_error

@ @<For each type |t|, make an equation and |goto done| unless...@>=
boolean_type,string_type,pen_type,path_type,picture_type:
  if cur_type=t+unknown_tag then
    begin nonlinear_eq(v,cur_exp,false); unstash_cur_exp(cur_exp); goto done;
    end
  else if cur_type=t then
    @<Report redundant or inconsistent equation and |goto done|@>;
unknown_types:if cur_type=t-unknown_tag then
    begin nonlinear_eq(cur_exp,lhs,true); goto done;
    end
  else if cur_type=t then
    begin ring_merge(lhs,cur_exp); goto done;
    end
  else if cur_type=pair_type then if t=unknown_path then
    begin pair_to_path; goto restart;
    end;
transform_type,pair_type:if cur_type=t then
    @<Do multiple equations and |goto done|@>;
known,dependent,proto_dependent,independent:if cur_type>=known then
    begin try_eq(lhs,null); goto done;
    end;
vacuous:do_nothing;

@ @<Report redundant or inconsistent equation and |goto done|@>=
begin if cur_type<=string_type then
  begin if cur_type=string_type then
    begin if str_vs_str(v,cur_exp)<>0 then goto not_found;
    end
  else if v<>cur_exp then goto not_found;
  @<Exclaim about a redundant equation@>; goto done;
  end;
print_err("Redundant or inconsistent equation");
@.Redundant or inconsistent equation@>
help2("An equation between already-known quantities can't help.")@/
  ("But don't worry; continue and I'll just ignore it.");
put_get_error; goto done;
not_found: print_err("Inconsistent equation");
@.Inconsistent equation@>
help2("The equation I just read contradicts what was said before.")@/
  ("But don't worry; continue and I'll just ignore it.");
put_get_error; goto done;
end

@ @<Do multiple equations and |goto done|@>=
begin p:=v+big_node_size[t]; q:=value(cur_exp)+big_node_size[t];
repeat p:=p-2; q:=q-2; try_eq(p,q);
until p=v;
goto done;
end

@ The first argument to |try_eq| is the location of a value node
in a capsule that will soon be recycled. The second argument is
either a location within a pair or transform node pointed to by
|cur_exp|, or it is |null| (which means that |cur_exp| itself
serves as the second argument). The idea is to leave |cur_exp| unchanged,
but to equate the two operands.

@<Declare the procedure called |try_eq|@>=
procedure try_eq(@!l,@!r:pointer);
label done,done1;
var @!p:pointer; {dependency list for right operand minus left operand}
@!t:known..independent; {the type of list |p|}
@!q:pointer; {the constant term of |p| is here}
@!pp:pointer; {dependency list for right operand}
@!tt:dependent..independent; {the type of list |pp|}
@!copied:boolean; {have we copied a list that ought to be recycled?}
begin @<Remove the left operand from its container, negate it, and
  put it into dependency list~|p| with constant term~|q|@>;
@<Add the right operand to list |p|@>;
if info(p)=null then @<Deal with redundant or inconsistent equation@>
else  begin linear_eq(p,t);
  if r=null then if cur_type<>known then if type(cur_exp)=known then
    begin pp:=cur_exp; cur_exp:=value(cur_exp); cur_type:=known;
    free_node(pp,value_node_size);
    end;
  end;
end;

@ @<Remove the left operand from its container, negate it, and...@>=
t:=type(l);
if t=known then
  begin t:=dependent; p:=const_dependency(-value(l)); q:=p;
  end
else if t=independent then
  begin t:=dependent; p:=single_dependency(l); negate(value(p));
  q:=dep_final;
  end
else  begin p:=dep_list(l); q:=p;
  loop@+  begin negate(value(q));
    if info(q)=null then goto done;
    q:=link(q);
    end;
 done:  link(prev_dep(l)):=link(q); prev_dep(link(q)):=prev_dep(l);
  type(l):=known;
  end

@ @<Deal with redundant or inconsistent equation@>=
begin if abs(value(p))>64 then {off by .001 or more}
  begin print_err("Inconsistent equation");@/
@.Inconsistent equation@>
  print(" (off by "); print_scaled(value(p)); print_char(")");
  help2("The equation I just read contradicts what was said before.")@/
    ("But don't worry; continue and I'll just ignore it.");
  put_get_error;
  end
else if r=null then @<Exclaim about a redundant equation@>;
free_node(p,dep_node_size);
end

@ @<Add the right operand to list |p|@>=
if r=null then
  if cur_type=known then
    begin value(q):=value(q)+cur_exp; goto done1;
    end
  else  begin tt:=cur_type;
    if tt=independent then pp:=single_dependency(cur_exp)
    else pp:=dep_list(cur_exp);
    end
else  if type(r)=known then
    begin value(q):=value(q)+value(r); goto done1;
    end
  else  begin tt:=type(r);
    if tt=independent then pp:=single_dependency(r)
    else pp:=dep_list(r);
    end;
if tt<>independent then copied:=false
else  begin copied:=true; tt:=dependent;
  end;
@<Add dependency list |pp| of type |tt| to dependency list~|p| of type~|t|@>;
if copied then flush_node_list(pp);
done1:

@ @<Add dependency list |pp| of type |tt| to dependency list~|p| of type~|t|@>=
watch_coefs:=false;
if t=tt then p:=p_plus_q(p,pp,t)
else if t=proto_dependent then
  p:=p_plus_fq(p,unity,pp,proto_dependent,dependent)
else  begin q:=p;
  while info(q)<>null do
    begin value(q):=round_fraction(value(q)); q:=link(q);
    end;
  t:=proto_dependent; p:=p_plus_q(p,pp,t);
  end;
watch_coefs:=true;

@ Our next goal is to process type declarations. For this purpose it's
convenient to have a procedure that scans a $\langle\,$declared
variable$\,\rangle$ and returns the corresponding token list. After the
following procedure has acted, the token after the declared variable
will have been scanned, so it will appear in |cur_cmd|, |cur_mod|,
and~|cur_sym|.

@<Declare the function called |scan_declared_variable|@>=
function scan_declared_variable:pointer;
label done;
var @!x:pointer; {hash address of the variable's root}
@!h,@!t:pointer; {head and tail of the token list to be returned}
@!l:pointer; {hash address of left bracket}
begin get_symbol; x:=cur_sym;
if cur_cmd<>tag_token then clear_symbol(x,false);
h:=get_avail; info(h):=x; t:=h;@/
loop@+  begin get_x_next;
  if cur_sym=0 then goto done;
  if cur_cmd<>tag_token then if cur_cmd<>internal_quantity then
    if cur_cmd=left_bracket then @<Descend past a collective subscript@>
    else goto done;
  link(t):=get_avail; t:=link(t); info(t):=cur_sym;
  end;
done: if eq_type(x) mod outer_tag<>tag_token then clear_symbol(x,false);
if equiv(x)=null then new_root(x);
scan_declared_variable:=h;
end;

@ If the subscript isn't collective, we don't accept it as part of the
declared variable.

@<Descend past a collective subscript@>=
begin l:=cur_sym; get_x_next;
if cur_cmd<>right_bracket then
  begin back_input; cur_sym:=l; cur_cmd:=left_bracket; goto done;
  end
else cur_sym:=collective_subscript;
end

@ Type declarations are introduced by the following primitive operations.

@<Put each...@>=
primitive("numeric",type_name,numeric_type);@/
@!@:numeric_}{\&{numeric} primitive@>
primitive("string",type_name,string_type);@/
@!@:string_}{\&{string} primitive@>
primitive("boolean",type_name,boolean_type);@/
@!@:boolean_}{\&{boolean} primitive@>
primitive("path",type_name,path_type);@/
@!@:path_}{\&{path} primitive@>
primitive("pen",type_name,pen_type);@/
@!@:pen_}{\&{pen} primitive@>
primitive("picture",type_name,picture_type);@/
@!@:picture_}{\&{picture} primitive@>
primitive("transform",type_name,transform_type);@/
@!@:transform_}{\&{transform} primitive@>
primitive("pair",type_name,pair_type);@/
@!@:pair_}{\&{pair} primitive@>

@ @<Cases of |print_cmd...@>=
type_name: print_type(m);

@ Now we are ready to handle type declarations, assuming that a
|type_name| has just been scanned.

@<Declare action procedures for use by |do_statement|@>=
procedure do_type_declaration;
var @!t:small_number; {the type being declared}
@!p:pointer; {token list for a declared variable}
@!q:pointer; {value node for the variable}
begin if cur_mod>=transform_type then t:=cur_mod@+else t:=cur_mod+unknown_tag;
repeat p:=scan_declared_variable;
flush_variable(equiv(info(p)),link(p),false);@/
q:=find_variable(p);
if q<>null then
  begin type(q):=t; value(q):=null;
  end
else  begin print_err("Declared variable conflicts with previous vardef");
@.Declared variable conflicts...@>
  help2("You can't use, e.g., `numeric foo[]' after `vardef foo'.")@/
    ("Proceed, and I'll ignore the illegal redeclaration.");
  put_get_error;
  end;
flush_list(p);
if cur_cmd<comma then @<Flush spurious symbols after the declared variable@>;
until end_of_statement;
end;

@ @<Flush spurious symbols after the declared variable@>=
begin print_err("Illegal suffix of declared variable will be flushed");
@.Illegal suffix...flushed@>
help5("Variables in declarations must consist entirely of")@/
  ("names and collective subscripts, e.g., `x[]a'.")@/
  ("Are you trying to use a reserved word in a variable name?")@/
  ("I'm going to discard the junk I found here,")@/
  ("up to the next comma or the end of the declaration.");
if cur_cmd=numeric_token then
  help_line[2]:="Explicit subscripts like `x15a' aren't permitted.";
put_get_error; scanner_status:=flushing;
repeat get_next;
@<Decrease the string reference count...@>;
until cur_cmd>=comma; {either |end_of_statement| or |cur_cmd=comma|}
scanner_status:=normal;
end

@ \MF's |main_control| procedure just calls |do_statement| repeatedly
until coming to the end of the user's program.
Each execution of |do_statement| concludes with
|cur_cmd=semicolon|, |end_group|, or |stop|.

@p procedure main_control;
begin repeat do_statement;
if cur_cmd=end_group then
  begin print_err("Extra `endgroup'");
@.Extra `endgroup'@>
  help2("I'm not currently working on a `begingroup',")@/
    ("so I had better not try to end anything.");
  flush_error(0);
  end;
until cur_cmd=stop;
end;

@ @<Put each...@>=
primitive("end",stop,0);@/
@!@:end_}{\&{end} primitive@>
primitive("dump",stop,1);@/
@!@:dump_}{\&{dump} primitive@>

@ @<Cases of |print_cmd...@>=
stop:if m=0 then print("end")@+else print("dump");

@* \[44] Commands.
Let's turn now to statements that are classified as ``commands'' because
of their imperative nature. We'll begin with simple ones, so that it
will be clear how to hook command processing into the |do_statement| routine;
then we'll tackle the tougher commands.

Here's one of the simplest:

@<Cases of |do_statement|...@>=
random_seed: do_random_seed;

@ @<Declare action procedures for use by |do_statement|@>=
procedure do_random_seed;
begin get_x_next;
if cur_cmd<>assignment then
  begin missing_err(":=");
@.Missing `:='@>
  help1("Always say `randomseed:=<numeric expression>'.");
  back_error;
  end;
get_x_next; scan_expression;
if cur_type<>known then
  begin exp_err("Unknown value will be ignored");
@.Unknown value...ignored@>
  help2("Your expression was too random for me to handle,")@/
    ("so I won't change the random seed just now.");@/
  put_get_flush_error(0);
  end
else @<Initialize the random seed to |cur_exp|@>;
end;

@ @<Initialize the random seed to |cur_exp|@>=
begin init_randoms(cur_exp);
if selector>=log_only then
  begin old_setting:=selector; selector:=log_only;
  print_nl("{randomseed:="); print_scaled(cur_exp); print_char("}");
  print_nl(""); selector:=old_setting;
  end;
end

@ And here's another simple one (somewhat different in flavor):

@<Cases of |do_statement|...@>=
mode_command: begin print_ln; interaction:=cur_mod;
  @<Initialize the print |selector| based on |interaction|@>;
  if log_opened then selector:=selector+2;
  get_x_next;
  end;

@ @<Put each...@>=
primitive("batchmode",mode_command,batch_mode);
@!@:batch_mode_}{\&{batchmode} primitive@>
primitive("nonstopmode",mode_command,nonstop_mode);
@!@:nonstop_mode_}{\&{nonstopmode} primitive@>
primitive("scrollmode",mode_command,scroll_mode);
@!@:scroll_mode_}{\&{scrollmode} primitive@>
primitive("errorstopmode",mode_command,error_stop_mode);
@!@:error_stop_mode_}{\&{errorstopmode} primitive@>

@ @<Cases of |print_cmd_mod|...@>=
mode_command: case m of
  batch_mode: print("batchmode");
  nonstop_mode: print("nonstopmode");
  scroll_mode: print("scrollmode");
  othercases print("errorstopmode")
  endcases;

@ The `\&{inner}' and `\&{outer}' commands are only slightly harder.

@<Cases of |do_statement|...@>=
protection_command: do_protection;

@ @<Put each...@>=
primitive("inner",protection_command,0);@/
@!@:inner_}{\&{inner} primitive@>
primitive("outer",protection_command,1);@/
@!@:outer_}{\&{outer} primitive@>

@ @<Cases of |print_cmd...@>=
protection_command: if m=0 then print("inner")@+else print("outer");

@ @<Declare action procedures for use by |do_statement|@>=
procedure do_protection;
var @!m:0..1; {0 to unprotect, 1 to protect}
@!t:halfword; {the |eq_type| before we change it}
begin m:=cur_mod;
repeat get_symbol; t:=eq_type(cur_sym);
  if m=0 then
    begin if t>=outer_tag then eq_type(cur_sym):=t-outer_tag;
    end
  else if t<outer_tag then eq_type(cur_sym):=t+outer_tag;
  get_x_next;
until cur_cmd<>comma;
end;

@ \MF\ never defines the tokens `\.(' and `\.)' to be primitives, but
plain \MF\ begins with the declaration `\&{delimiters} \.{()}'. Such a
declaration assigns the command code |left_delimiter| to `\.{(}' and
|right_delimiter| to `\.{)}'; the |equiv| of each delimiter is the
hash address of its mate.

@<Cases of |do_statement|...@>=
delimiters: def_delims;

@ @<Declare action procedures for use by |do_statement|@>=
procedure def_delims;
var l_delim,r_delim:pointer; {the new delimiter pair}
begin get_clear_symbol; l_delim:=cur_sym;@/
get_clear_symbol; r_delim:=cur_sym;@/
eq_type(l_delim):=left_delimiter; equiv(l_delim):=r_delim;@/
eq_type(r_delim):=right_delimiter; equiv(r_delim):=l_delim;@/
get_x_next;
end;

@ Here is a procedure that is called when \MF\ has reached a point
where some right delimiter is mandatory.

@<Declare the procedure called |check_delimiter|@>=
procedure check_delimiter(@!l_delim,@!r_delim:pointer);
label exit;
begin if cur_cmd=right_delimiter then if cur_mod=l_delim then return;
if cur_sym<>r_delim then
  begin  missing_err(text(r_delim));@/
@.Missing `)'@>
  help2("I found no right delimiter to match a left one. So I've")@/
    ("put one in, behind the scenes; this may fix the problem.");
  back_error;
  end
else  begin print_err("The token `"); slow_print(text(r_delim));
@.The token...delimiter@>
  print("' is no longer a right delimiter");
  help3("Strange: This token has lost its former meaning!")@/
    ("I'll read it as a right delimiter this time;")@/
    ("but watch out, I'll probably miss it later.");
  error;
  end;
exit:end;

@ The next four commands save or change the values associated with tokens.

@<Cases of |do_statement|...@>=
save_command: repeat get_symbol; save_variable(cur_sym); get_x_next;
  until cur_cmd<>comma;
interim_command: do_interim;
let_command: do_let;
new_internal: do_new_internal;

@ @<Declare action procedures for use by |do_statement|@>=
procedure@?do_statement; forward;@t\2@>@/
procedure do_interim;
begin get_x_next;
if cur_cmd<>internal_quantity then
  begin print_err("The token `");
@.The token...quantity@>
  if cur_sym=0 then print("(%CAPSULE)")
  else slow_print(text(cur_sym));
  print("' isn't an internal quantity");
  help1("Something like `tracingonline' should follow `interim'.");
  back_error;
  end
else  begin save_internal(cur_mod); back_input;
  end;
do_statement;
end;

@ The following procedure is careful not to undefine the left-hand symbol
too soon, lest commands like `{\tt let x=x}' have a surprising effect.

@<Declare action procedures for use by |do_statement|@>=
procedure do_let;
var @!l:pointer; {hash location of the left-hand symbol}
begin get_symbol; l:=cur_sym; get_x_next;
if cur_cmd<>equals then if cur_cmd<>assignment then
  begin missing_err("=");
@.Missing `='@>
  help3("You should have said `let symbol = something'.")@/
    ("But don't worry; I'll pretend that an equals sign")@/
    ("was present. The next token I read will be `something'.");
  back_error;
  end;
get_symbol;
case cur_cmd of
defined_macro,secondary_primary_macro,tertiary_secondary_macro,
 expression_tertiary_macro: add_mac_ref(cur_mod);
othercases do_nothing
endcases;@/
clear_symbol(l,false); eq_type(l):=cur_cmd;
if cur_cmd=tag_token then equiv(l):=null
else equiv(l):=cur_mod;
get_x_next;
end;

@ @<Declare action procedures for use by |do_statement|@>=
procedure do_new_internal;
begin repeat if int_ptr=max_internal then
  overflow("number of internals",max_internal);
@:METAFONT capacity exceeded number of int}{\quad number of internals@>
get_clear_symbol; incr(int_ptr);
eq_type(cur_sym):=internal_quantity; equiv(cur_sym):=int_ptr;
int_name[int_ptr]:=text(cur_sym); internal[int_ptr]:=0;
get_x_next;
until cur_cmd<>comma;
end;

@ The various `\&{show}' commands are distinguished by modifier fields
in the usual way.

@d show_token_code=0 {show the meaning of a single token}
@d show_stats_code=1 {show current memory and string usage}
@d show_code=2 {show a list of expressions}
@d show_var_code=3 {show a variable and its descendents}
@d show_dependencies_code=4 {show dependent variables in terms of independents}

@<Put each...@>=
primitive("showtoken",show_command,show_token_code);@/
@!@:show_token_}{\&{showtoken} primitive@>
primitive("showstats",show_command,show_stats_code);@/
@!@:show_stats_}{\&{showstats} primitive@>
primitive("show",show_command,show_code);@/
@!@:show_}{\&{show} primitive@>
primitive("showvariable",show_command,show_var_code);@/
@!@:show_var_}{\&{showvariable} primitive@>
primitive("showdependencies",show_command,show_dependencies_code);@/
@!@:show_dependencies_}{\&{showdependencies} primitive@>

@ @<Cases of |print_cmd...@>=
show_command: case m of
  show_token_code:print("showtoken");
  show_stats_code:print("showstats");
  show_code:print("show");
  show_var_code:print("showvariable");
  othercases print("showdependencies")
  endcases;

@ @<Cases of |do_statement|...@>=
show_command:do_show_whatever;

@ The value of |cur_mod| controls the |verbosity| in the |print_exp| routine:
If it's |show_code|, complicated structures are abbreviated, otherwise
they aren't.

@<Declare action procedures for use by |do_statement|@>=
procedure do_show;
begin repeat get_x_next; scan_expression;
print_nl(">> ");
@.>>@>
print_exp(null,2); flush_cur_exp(0);
until cur_cmd<>comma;
end;

@ @<Declare action procedures for use by |do_statement|@>=
procedure disp_token;
begin print_nl("> ");
@.>\relax@>
if cur_sym=0 then @<Show a numeric or string or capsule token@>
else  begin slow_print(text(cur_sym)); print_char("=");
  if eq_type(cur_sym)>=outer_tag then print("(outer) ");
  print_cmd_mod(cur_cmd,cur_mod);
  if cur_cmd=defined_macro then
    begin print_ln; show_macro(cur_mod,null,100000);
    end; {this avoids recursion between |show_macro| and |print_cmd_mod|}
@^recursion@>
  end;
end;

@ @<Show a numeric or string or capsule token@>=
begin if cur_cmd=numeric_token then print_scaled(cur_mod)
else if cur_cmd=capsule_token then
  begin g_pointer:=cur_mod; print_capsule;
  end
else  begin print_char(""""); slow_print(cur_mod); print_char("""");
  delete_str_ref(cur_mod);
  end;
end

@ The following cases of |print_cmd_mod| might arise in connection
with |disp_token|, although they don't necessarily correspond to
primitive tokens.

@<Cases of |print_cmd_...@>=
left_delimiter,right_delimiter: begin if c=left_delimiter then print("lef")
  else print("righ");
  print("t delimiter that matches "); slow_print(text(m));
  end;
tag_token:if m=null then print("tag")@+else print("variable");
defined_macro: print("macro:");
secondary_primary_macro,tertiary_secondary_macro,expression_tertiary_macro:
  begin print_cmd_mod(macro_def,c); print("'d macro:");
  print_ln; show_token_list(link(link(m)),null,1000,0);
  end;
repeat_loop:print("[repeat the loop]");
internal_quantity:slow_print(int_name[m]);

@ @<Declare action procedures for use by |do_statement|@>=
procedure do_show_token;
begin repeat get_next; disp_token;
get_x_next;
until cur_cmd<>comma;
end;

@ @<Declare action procedures for use by |do_statement|@>=
procedure do_show_stats;
begin print_nl("Memory usage ");
@.Memory usage...@>
@!stat print_int(var_used); print_char("&"); print_int(dyn_used);
if false then@+tats@t@>@;@/
print("unknown");
print(" ("); print_int(hi_mem_min-lo_mem_max-1);
print(" still untouched)"); print_ln;
print_nl("String usage ");
print_int(str_ptr-init_str_ptr); print_char("&");
print_int(pool_ptr-init_pool_ptr);
print(" (");
print_int(max_strings-max_str_ptr); print_char("&");
print_int(pool_size-max_pool_ptr); print(" still untouched)"); print_ln;
get_x_next;
end;

@ Here's a recursive procedure that gives an abbreviated account
of a variable, for use by |do_show_var|.

@<Declare action procedures for use by |do_statement|@>=
procedure disp_var(@!p:pointer);
var @!q:pointer; {traverses attributes and subscripts}
@!n:0..max_print_line; {amount of macro text to show}
begin if type(p)=structured then @<Descend the structure@>
else if type(p)>=unsuffixed_macro then @<Display a variable macro@>
else if type(p)<>undefined then
  begin print_nl(""); print_variable_name(p); print_char("=");
  print_exp(p,0);
  end;
end;

@ @<Descend the structure@>=
begin q:=attr_head(p);
repeat disp_var(q); q:=link(q);
until q=end_attr;
q:=subscr_head(p);
while name_type(q)=subscr do
  begin disp_var(q); q:=link(q);
  end;
end

@ @<Display a variable macro@>=
begin print_nl(""); print_variable_name(p);
if type(p)>unsuffixed_macro then print("@@#"); {|suffixed_macro|}
print("=macro:");
if file_offset>=max_print_line-20 then n:=5
else n:=max_print_line-file_offset-15;
show_macro(value(p),null,n);
end

@ @<Declare action procedures for use by |do_statement|@>=
procedure do_show_var;
label done;
begin repeat get_next;
if cur_sym>0 then if cur_sym<=hash_end then
 if cur_cmd=tag_token then if cur_mod<>null then
  begin disp_var(cur_mod); goto done;
  end;
disp_token;
done:get_x_next;
until cur_cmd<>comma;
end;

@ @<Declare action procedures for use by |do_statement|@>=
procedure do_show_dependencies;
var @!p:pointer; {link that runs through all dependencies}
begin p:=link(dep_head);
while p<>dep_head do
  begin if interesting(p) then
    begin print_nl(""); print_variable_name(p);
    if type(p)=dependent then print_char("=")
    else print(" = "); {extra spaces imply proto-dependency}
    print_dependency(dep_list(p),type(p));
    end;
  p:=dep_list(p);
  while info(p)<>null do p:=link(p);
  p:=link(p);
  end;
get_x_next;
end;

@ Finally we are ready for the procedure that governs all of the
show commands.

@<Declare action procedures for use by |do_statement|@>=
procedure do_show_whatever;
begin if interaction=error_stop_mode then wake_up_terminal;
case cur_mod of
show_token_code:do_show_token;
show_stats_code:do_show_stats;
show_code:do_show;
show_var_code:do_show_var;
show_dependencies_code:do_show_dependencies;
end; {there are no other cases}
if internal[showstopping]>0 then
  begin print_err("OK");
@.OK@>
  if interaction<error_stop_mode then
    begin help0; decr(error_count);
    end
  else help1("This isn't an error message; I'm just showing something.");
  if cur_cmd=semicolon then error@+else put_get_error;
  end;
end;

@ The `\&{addto}' command needs the following additional primitives:

@d drop_code=0 {command modifier for `\&{dropping}'}
@d keep_code=1 {command modifier for `\&{keeping}'}

@<Put each...@>=
primitive("contour",thing_to_add,contour_code);@/
@!@:contour_}{\&{contour} primitive@>
primitive("doublepath",thing_to_add,double_path_code);@/
@!@:double_path_}{\&{doublepath} primitive@>
primitive("also",thing_to_add,also_code);@/
@!@:also_}{\&{also} primitive@>
primitive("withpen",with_option,pen_type);@/
@!@:with_pen_}{\&{withpen} primitive@>
primitive("withweight",with_option,known);@/
@!@:with_weight_}{\&{withweight} primitive@>
primitive("dropping",cull_op,drop_code);@/
@!@:dropping_}{\&{dropping} primitive@>
primitive("keeping",cull_op,keep_code);@/
@!@:keeping_}{\&{keeping} primitive@>

@ @<Cases of |print_cmd...@>=
thing_to_add:if m=contour_code then print("contour")
  else if m=double_path_code then print("doublepath")
  else print("also");
with_option:if m=pen_type then print("withpen")
  else print("withweight");
cull_op:if m=drop_code then print("dropping")
  else print("keeping");

@ @<Declare action procedures for use by |do_statement|@>=
function scan_with:boolean;
var @!t:small_number; {|known| or |pen_type|}
@!result:boolean; {the value to return}
begin t:=cur_mod; cur_type:=vacuous; get_x_next; scan_expression;
result:=false;
if cur_type<>t then @<Complain about improper type@>
else if cur_type=pen_type then result:=true
else @<Check the tentative weight@>;
scan_with:=result;
end;

@ @<Complain about improper type@>=
begin exp_err("Improper type");
@.Improper type@>
help2("Next time say `withweight <known numeric expression>';")@/
  ("I'll ignore the bad `with' clause and look for another.");
if t=pen_type then
  help_line[1]:="Next time say `withpen <known pen expression>';";
put_get_flush_error(0);
end

@ @<Check the tentative weight@>=
begin cur_exp:=round_unscaled(cur_exp);
if (abs(cur_exp)<4)and(cur_exp<>0) then result:=true
else  begin print_err("Weight must be -3, -2, -1, +1, +2, or +3");
@.Weight must be...@>
  help1("I'll ignore the bad `with' clause and look for another.");
  put_get_flush_error(0);
  end;
end

@ One of the things we need to do when we've parsed an \&{addto} or
similar command is set |cur_edges| to the header of a supposed \&{picture}
variable, given a token list for that variable.

@<Declare action procedures for use by |do_statement|@>=
procedure find_edges_var(@!t:pointer);
var @!p:pointer;
begin p:=find_variable(t); cur_edges:=null;
if p=null then
  begin obliterated(t); put_get_error;
  end
else if type(p)<>picture_type then
  begin print_err("Variable "); show_token_list(t,null,1000,0);
@.Variable x is the wrong type@>
  print(" is the wrong type ("); print_type(type(p)); print_char(")");
  help2("I was looking for a ""known"" picture variable.")@/
    ("So I'll not change anything just now."); put_get_error;
  end
else cur_edges:=value(p);
flush_node_list(t);
end;

@ @<Cases of |do_statement|...@>=
add_to_command: do_add_to;

@ @<Declare action procedures for use by |do_statement|@>=
procedure do_add_to;
label done, not_found;
var @!lhs,@!rhs:pointer; {variable on left, path on right}
@!w:integer; {tentative weight}
@!p:pointer; {list manipulation register}
@!q:pointer; {beginning of second half of doubled path}
@!add_to_type:double_path_code..also_code; {modifier of \&{addto}}
begin get_x_next; var_flag:=thing_to_add; scan_primary;
if cur_type<>token_list then
  @<Abandon edges command because there's no variable@>
else  begin lhs:=cur_exp; add_to_type:=cur_mod;@/
  cur_type:=vacuous; get_x_next; scan_expression;
  if add_to_type=also_code then @<Augment some edges by others@>
  else @<Get ready to fill a contour, and fill it@>;
  end;
end;

@ @<Abandon edges command because there's no variable@>=
begin exp_err("Not a suitable variable");
@.Not a suitable variable@>
help4("At this point I needed to see the name of a picture variable.")@/
  ("(Or perhaps you have indeed presented me with one; I might")@/
  ("have missed it, if it wasn't followed by the proper token.)")@/
  ("So I'll not change anything just now.");
put_get_flush_error(0);
end

@ @<Augment some edges by others@>=
begin find_edges_var(lhs);
if cur_edges=null then flush_cur_exp(0)
else if cur_type<>picture_type then
  begin exp_err("Improper `addto'");
@.Improper `addto'@>
  help2("This expression should have specified a known picture.")@/
    ("So I'll not change anything just now."); put_get_flush_error(0);
  end
else  begin merge_edges(cur_exp); flush_cur_exp(0);
  end;
end

@ @<Get ready to fill a contour...@>=
begin if cur_type=pair_type then pair_to_path;
if cur_type<>path_type then
  begin exp_err("Improper `addto'");
@.Improper `addto'@>
  help2("This expression should have been a known path.")@/
    ("So I'll not change anything just now.");
  put_get_flush_error(0); flush_token_list(lhs);
  end
else  begin rhs:=cur_exp; w:=1; cur_pen:=null_pen;
  while cur_cmd=with_option do
    if scan_with then
      if cur_type=known then w:=cur_exp
      else @<Change the tentative pen@>;
  @<Complete the contour filling operation@>;
  delete_pen_ref(cur_pen);
  end;
end

@ We could say `|add_pen_ref(cur_pen)|; |flush_cur_exp(0)|' after changing
|cur_pen| here.  But that would have no effect, because the current expression
will not be flushed. Thus we save a bit of code (at the risk of being too
tricky).

@<Change the tentative pen@>=
begin delete_pen_ref(cur_pen); cur_pen:=cur_exp;
end

@ @<Complete the contour filling...@>=
find_edges_var(lhs);
if cur_edges=null then toss_knot_list(rhs)
else  begin lhs:=null; cur_path_type:=add_to_type;
  if left_type(rhs)=endpoint then
    if cur_path_type=double_path_code then @<Double the path@>
    else @<Complain about non-cycle and |goto not_found|@>
  else if cur_path_type=double_path_code then lhs:=htap_ypoc(rhs);
  cur_wt:=w; rhs:=make_spec(rhs,max_offset(cur_pen),internal[tracing_specs]);
  @<Check the turning number@>;
  if max_offset(cur_pen)=0 then fill_spec(rhs)
  else fill_envelope(rhs);
  if lhs<>null then
    begin rev_turns:=true;
    lhs:=make_spec(lhs,max_offset(cur_pen),internal[tracing_specs]);
    rev_turns:=false;
    if max_offset(cur_pen)=0 then fill_spec(lhs)
    else fill_envelope(lhs);
    end;
not_found: end

@ @<Double the path@>=
if link(rhs)=rhs then @<Make a trivial one-point path cycle@>
else  begin p:=htap_ypoc(rhs); q:=link(p);@/
  right_x(path_tail):=right_x(q); right_y(path_tail):=right_y(q);
  right_type(path_tail):=right_type(q);
  link(path_tail):=link(q); free_node(q,knot_node_size);@/
  right_x(p):=right_x(rhs); right_y(p):=right_y(rhs);
  right_type(p):=right_type(rhs);
  link(p):=link(rhs); free_node(rhs,knot_node_size);@/
  rhs:=p;
  end

@ @<Make a trivial one-point path cycle@>=
begin right_x(rhs):=x_coord(rhs); right_y(rhs):=y_coord(rhs);
left_x(rhs):=x_coord(rhs); left_y(rhs):=y_coord(rhs);
left_type(rhs):=explicit; right_type(rhs):=explicit;
end

@ @<Complain about non-cycle...@>=
begin print_err("Not a cycle");
@.Not a cycle@>
help2("That contour should have ended with `..cycle' or `&cycle'.")@/
  ("So I'll not change anything just now."); put_get_error;
toss_knot_list(rhs); goto not_found;
end

@ @<Check the turning number@>=
if turning_number<=0 then
 if cur_path_type<>double_path_code then if internal[turning_check]>0 then
  if (turning_number<0)and(link(cur_pen)=null) then negate(cur_wt)
  else  begin if turning_number=0 then
      if (internal[turning_check]<=unity)and(link(cur_pen)=null) then goto done
      else print_strange("Strange path (turning number is zero)")
@.Strange path...@>
    else print_strange("Backwards path (turning number is negative)");
@.Backwards path...@>
    help3("The path doesn't have a counterclockwise orientation,")@/
      ("so I'll probably have trouble drawing it.")@/
      ("(See Chapter 27 of The METAFONTbook for more help.)");
@:METAFONTbook}{\sl The {\logos METAFONT\/}book@>
    put_get_error;
    end;
done:

@ @<Cases of |do_statement|...@>=
ship_out_command: do_ship_out;
display_command: do_display;
open_window: do_open_window;
cull_command: do_cull;

@ @<Declare action procedures for use by |do_statement|@>=
@t\4@>@<Declare the function called |tfm_check|@>@;
procedure do_ship_out;
label exit;
var @!c:integer; {the character code}
begin get_x_next; var_flag:=semicolon; scan_expression;
if cur_type<>token_list then
  if cur_type=picture_type then cur_edges:=cur_exp
  else  begin @<Abandon edges command because there's no variable@>;
    return;
    end
else  begin find_edges_var(cur_exp); cur_type:=vacuous;
  end;
if cur_edges<>null then
  begin c:=round_unscaled(internal[char_code]) mod 256;
  if c<0 then c:=c+256;
  @<Store the width information for character code~|c|@>;
  if internal[proofing]>=0 then ship_out(c);
  end;
flush_cur_exp(0);
exit:end;

@ @<Declare action procedures for use by |do_statement|@>=
procedure do_display;
label not_found,common_ending,exit;
var @!e:pointer; {token list for a picture variable}
begin get_x_next; var_flag:=in_window; scan_primary;
if cur_type<>token_list then
  @<Abandon edges command because there's no variable@>
else  begin e:=cur_exp; cur_type:=vacuous;
  get_x_next; scan_expression;
  if cur_type<>known then goto common_ending;
  cur_exp:=round_unscaled(cur_exp);
  if cur_exp<0 then goto not_found;
  if cur_exp>15 then goto not_found;
  if not window_open[cur_exp] then goto not_found;
  find_edges_var(e);
  if cur_edges<>null then disp_edges(cur_exp);
  return;
 not_found: cur_exp:=cur_exp*unity;
 common_ending: exp_err("Bad window number");
@.Bad window number@>
  help1("It should be the number of an open window.");
  put_get_flush_error(0); flush_token_list(e);
  end;
exit:end;

@ The only thing difficult about `\&{openwindow}' is that the syntax
allows the user to go astray in many ways. The following subroutine
helps keep the necessary program reasonably short and sweet.

@<Declare action procedures for use by |do_statement|@>=
function get_pair(@!c:command_code):boolean;
var @!p:pointer; {a pair of values that are known (we hope)}
@!b:boolean; {did we find such a pair?}
begin if cur_cmd<>c then get_pair:=false
else  begin get_x_next; scan_expression;
  if nice_pair(cur_exp,cur_type) then
    begin p:=value(cur_exp);
    cur_x:=value(x_part_loc(p)); cur_y:=value(y_part_loc(p));
    b:=true;
    end
  else b:=false;
  flush_cur_exp(0); get_pair:=b;
  end;
end;

@ @<Declare action procedures for use by |do_statement|@>=
procedure do_open_window;
label not_found,exit;
var @!k:integer; {the window number in question}
@!r0,@!c0,@!r1,@!c1:scaled; {window coordinates}
begin get_x_next; scan_expression;
if cur_type<>known then goto not_found;
k:=round_unscaled(cur_exp);
if k<0 then goto not_found;
if k>15 then goto not_found;
if not get_pair(from_token) then goto not_found;
r0:=cur_x; c0:=cur_y;
if not get_pair(to_token) then goto not_found;
r1:=cur_x; c1:=cur_y;
if not get_pair(at_token) then goto not_found;
open_a_window(k,r0,c0,r1,c1,cur_x,cur_y); return;
not_found:print_err("Improper `openwindow'");
@.Improper `openwindow'@>
help2("Say `openwindow k from (r0,c0) to (r1,c1) at (x,y)',")@/
  ("where all quantities are known and k is between 0 and 15.");
put_get_error;
exit:end;

@ @<Declare action procedures for use by |do_statement|@>=
procedure do_cull;
label not_found,exit;
var @!e:pointer; {token list for a picture variable}
@!keeping:drop_code..keep_code; {modifier of |cull_op|}
@!w,@!w_in,@!w_out:integer; {culling weights}
begin w:=1;
get_x_next; var_flag:=cull_op; scan_primary;
if cur_type<>token_list then
  @<Abandon edges command because there's no variable@>
else  begin e:=cur_exp; cur_type:=vacuous; keeping:=cur_mod;
  if not get_pair(cull_op) then goto not_found;
  while (cur_cmd=with_option)and(cur_mod=known) do
    if scan_with then w:=cur_exp;
  @<Set up the culling weights,
    or |goto not_found| if the thresholds are bad@>;
  find_edges_var(e);
  if cur_edges<>null then
    cull_edges(floor_unscaled(cur_x+unity-1),floor_unscaled(cur_y),w_out,w_in);
  return;
 not_found: print_err("Bad culling amounts");
@.Bad culling amounts@>
  help1("Always cull by known amounts that exclude 0.");
  put_get_error; flush_token_list(e);
  end;
exit:end;

@ @<Set up the culling weights, or |goto not_found| if the thresholds are bad@>=
if cur_x>cur_y then goto not_found;
if keeping=drop_code then
  begin if (cur_x>0)or(cur_y<0) then goto not_found;
  w_out:=w; w_in:=0;
  end
else  begin if (cur_x<=0)and(cur_y>=0) then goto not_found;
  w_out:=0; w_in:=w;
  end

@ The \&{everyjob} command simply assigns a nonzero value to the global variable
|start_sym|.

@<Cases of |do_statement|...@>=
every_job_command: begin get_symbol; start_sym:=cur_sym; get_x_next;
  end;

@ @<Glob...@>=
@!start_sym:halfword; {a symbolic token to insert at beginning of job}

@ @<Set init...@>=
start_sym:=0;

@ Finally, we have only the ``message'' commands remaining.

@d message_code=0
@d err_message_code=1
@d err_help_code=2

@<Put each...@>=
primitive("message",message_command,message_code);@/
@!@:message_}{\&{message} primitive@>
primitive("errmessage",message_command,err_message_code);@/
@!@:err_message_}{\&{errmessage} primitive@>
primitive("errhelp",message_command,err_help_code);@/
@!@:err_help_}{\&{errhelp} primitive@>

@ @<Cases of |print_cmd...@>=
message_command: if m<err_message_code then print("message")
  else if m=err_message_code then print("errmessage")
  else print("errhelp");

@ @<Cases of |do_statement|...@>=
message_command: do_message;

@ @<Declare action procedures for use by |do_statement|@>=
procedure do_message;
var @!m:message_code..err_help_code; {the type of message}
begin m:=cur_mod; get_x_next; scan_expression;
if cur_type<>string_type then
  begin exp_err("Not a string");
@.Not a string@>
  help1("A message should be a known string expression.");
  put_get_error;
  end
else  case m of
  message_code:begin print_nl(""); slow_print(cur_exp);
    end;
  err_message_code:@<Print string |cur_exp| as an error message@>;
  err_help_code:@<Save string |cur_exp| as the |err_help|@>;
  end; {there are no other cases}
flush_cur_exp(0);
end;

@ The global variable |err_help| is zero when the user has most recently
given an empty help string, or if none has ever been given.

@<Save string |cur_exp| as the |err_help|@>=
begin if err_help<>0 then delete_str_ref(err_help);
if length(cur_exp)=0 then err_help:=0
else  begin err_help:=cur_exp; add_str_ref(err_help);
  end;
end

@ If \&{errmessage} occurs often in |scroll_mode|, without user-defined
\&{errhelp}, we don't want to give a long help message each time. So we
give a verbose explanation only once.

@<Glob...@>=
@!long_help_seen:boolean; {has the long \&{errmessage} help been used?}

@ @<Set init...@>=long_help_seen:=false;

@ @<Print string |cur_exp| as an error message@>=
begin print_err(""); slow_print(cur_exp);
if err_help<>0 then use_err_help:=true
else if long_help_seen then help1("(That was another `errmessage'.)")
else  begin if interaction<error_stop_mode then long_help_seen:=true;
  help4("This error message was generated by an `errmessage'")@/
  ("command, so I can't give any explicit help.")@/
  ("Pretend that you're Miss Marple: Examine all clues,")@/
@^Marple, Jane@>
  ("and deduce the truth by inspired guesses.");
  end;
put_get_error; use_err_help:=false;
end

@* \[45] Font metric data.
\TeX\ gets its knowledge about fonts from font metric files, also called
\.{TFM} files; the `\.T' in `\.{TFM}' stands for \TeX,
but other programs know about them too. One of \MF's duties is to
write \.{TFM} files so that the user's fonts can readily be
applied to typesetting.
@:TFM files}{\.{TFM} files@>
@^font metric files@>

The information in a \.{TFM} file appears in a sequence of 8-bit bytes.
Since the number of bytes is always a multiple of~4, we could
also regard the file as a sequence of 32-bit words, but \MF\ uses the
byte interpretation. The format of \.{TFM} files was designed by
Lyle Ramshaw in 1980. The intent is to convey a lot of different kinds
@^Ramshaw, Lyle Harold@>
of information in a compact but useful form.

@<Glob...@>=
@!tfm_file:byte_file; {the font metric output goes here}
@!metric_file_name: str_number; {full name of the font metric file}

@ The first 24 bytes (6 words) of a \.{TFM} file contain twelve 16-bit
integers that give the lengths of the various subsequent portions
of the file. These twelve integers are, in order:
$$\vbox{\halign{\hfil#&$\null=\null$#\hfil\cr
|lf|&length of the entire file, in words;\cr
|lh|&length of the header data, in words;\cr
|bc|&smallest character code in the font;\cr
|ec|&largest character code in the font;\cr
|nw|&number of words in the width table;\cr
|nh|&number of words in the height table;\cr
|nd|&number of words in the depth table;\cr
|ni|&number of words in the italic correction table;\cr
|nl|&number of words in the lig/kern table;\cr
|nk|&number of words in the kern table;\cr
|ne|&number of words in the extensible character table;\cr
|np|&number of font parameter words.\cr}}$$
They are all nonnegative and less than $2^{15}$. We must have |bc-1<=ec<=255|,
|ne<=256|, and
$$\hbox{|lf=6+lh+(ec-bc+1)+nw+nh+nd+ni+nl+nk+ne+np|.}$$
Note that a font may contain as many as 256 characters (if |bc=0| and |ec=255|),
and as few as 0 characters (if |bc=ec+1|).

Incidentally, when two or more 8-bit bytes are combined to form an integer of
16 or more bits, the most significant bytes appear first in the file.
This is called BigEndian order.
@!@^BigEndian order@>

@ The rest of the \.{TFM} file may be regarded as a sequence of ten data
arrays having the informal specification
$$\def\arr$[#1]#2${\&{array} $[#1]$ \&{of} #2}
\tabskip\centering
\halign to\displaywidth{\hfil\\{#}\tabskip=0pt&$\,:\,$\arr#\hfil
 \tabskip\centering\cr
header&|[0..lh-1]@t\\{stuff}@>|\cr
char\_info&|[bc..ec]char_info_word|\cr
width&|[0..nw-1]fix_word|\cr
height&|[0..nh-1]fix_word|\cr
depth&|[0..nd-1]fix_word|\cr
italic&|[0..ni-1]fix_word|\cr
lig\_kern&|[0..nl-1]lig_kern_command|\cr
kern&|[0..nk-1]fix_word|\cr
exten&|[0..ne-1]extensible_recipe|\cr
param&|[1..np]fix_word|\cr}$$
The most important data type used here is a |@!fix_word|, which is
a 32-bit representation of a binary fraction. A |fix_word| is a signed
quantity, with the two's complement of the entire word used to represent
negation. Of the 32 bits in a |fix_word|, exactly 12 are to the left of the
binary point; thus, the largest |fix_word| value is $2048-2^{-20}$, and
the smallest is $-2048$. We will see below, however, that all but two of
the |fix_word| values must lie between $-16$ and $+16$.

@ The first data array is a block of header information, which contains
general facts about the font. The header must contain at least two words,
|header[0]| and |header[1]|, whose meaning is explained below.  Additional
header information of use to other software routines might also be
included, and \MF\ will generate it if the \.{headerbyte} command occurs.
For example, 16 more words of header information are in use at the Xerox
Palo Alto Research Center; the first ten specify the character coding
scheme used (e.g., `\.{XEROX TEXT}' or `\.{TEX MATHSY}'), the next five
give the font family name (e.g., `\.{HELVETICA}' or `\.{CMSY}'), and the
last gives the ``face byte.''

\yskip\hang|header[0]| is a 32-bit check sum that \MF\ will copy into
the \.{GF} output file. This helps ensure consistency between files,
since \TeX\ records the check sums from the \.{TFM}'s it reads, and these
should match the check sums on actual fonts that are used.  The actual
relation between this check sum and the rest of the \.{TFM} file is not
important; the check sum is simply an identification number with the
property that incompatible fonts almost always have distinct check sums.
@^check sum@>

\yskip\hang|header[1]| is a |fix_word| containing the design size of the
font, in units of \TeX\ points. This number must be at least 1.0; it is
fairly arbitrary, but usually the design size is 10.0 for a ``10 point''
font, i.e., a font that was designed to look best at a 10-point size,
whatever that really means. When a \TeX\ user asks for a font `\.{at}
$\delta$ \.{pt}', the effect is to override the design size and replace it
by $\delta$, and to multiply the $x$ and~$y$ coordinates of the points in
the font image by a factor of $\delta$ divided by the design size.  {\sl
All other dimensions in the\/ \.{TFM} file are |fix_word|\kern-1pt\
numbers in design-size units.} Thus, for example, the value of |param[6]|,
which defines the \.{em} unit, is often the |fix_word| value $2^{20}=1.0$,
since many fonts have a design size equal to one em.  The other dimensions
must be less than 16 design-size units in absolute value; thus,
|header[1]| and |param[1]| are the only |fix_word| entries in the whole
\.{TFM} file whose first byte might be something besides 0 or 255.
@^design size@>

@ Next comes the |char_info| array, which contains one |@!char_info_word|
per character. Each word in this part of the file contains six fields
packed into four bytes as follows.

\yskip\hang first byte: |@!width_index| (8 bits)\par
\hang second byte: |@!height_index| (4 bits) times 16, plus |@!depth_index|
  (4~bits)\par
\hang third byte: |@!italic_index| (6 bits) times 4, plus |@!tag|
  (2~bits)\par
\hang fourth byte: |@!remainder| (8 bits)\par
\yskip\noindent
The actual width of a character is \\{width}|[width_index]|, in design-size
units; this is a device for compressing information, since many characters
have the same width. Since it is quite common for many characters
to have the same height, depth, or italic correction, the \.{TFM} format
imposes a limit of 16 different heights, 16 different depths, and
64 different italic corrections.

Incidentally, the relation $\\{width}[0]=\\{height}[0]=\\{depth}[0]=
\\{italic}[0]=0$ should always hold, so that an index of zero implies a
value of zero.  The |width_index| should never be zero unless the
character does not exist in the font, since a character is valid if and
only if it lies between |bc| and |ec| and has a nonzero |width_index|.

@ The |tag| field in a |char_info_word| has four values that explain how to
interpret the |remainder| field.

\def\hangg#1 {\hang\hbox{#1 }}
\yskip\hangg|tag=0| (|no_tag|) means that |remainder| is unused.\par
\hangg|tag=1| (|lig_tag|) means that this character has a ligature/kerning
program starting at location |remainder| in the |lig_kern| array.\par
\hangg|tag=2| (|list_tag|) means that this character is part of a chain of
characters of ascending sizes, and not the largest in the chain.  The
|remainder| field gives the character code of the next larger character.\par
\hangg|tag=3| (|ext_tag|) means that this character code represents an
extensible character, i.e., a character that is built up of smaller pieces
so that it can be made arbitrarily large. The pieces are specified in
|@!exten[remainder]|.\par
\yskip\noindent
Characters with |tag=2| and |tag=3| are treated as characters with |tag=0|
unless they are used in special circumstances in math formulas. For example,
\TeX's \.{\\sum} operation looks for a |list_tag|, and the \.{\\left}
operation looks for both |list_tag| and |ext_tag|.

@d no_tag=0 {vanilla character}
@d lig_tag=1 {character has a ligature/kerning program}
@d list_tag=2 {character has a successor in a charlist}
@d ext_tag=3 {character is extensible}

@ The |lig_kern| array contains instructions in a simple programming language
that explains what to do for special letter pairs. Each word in this array is a
|@!lig_kern_command| of four bytes.

\yskip\hang first byte: |skip_byte|, indicates that this is the final program
  step if the byte is 128 or more, otherwise the next step is obtained by
  skipping this number of intervening steps.\par
\hang second byte: |next_char|, ``if |next_char| follows the current character,
  then perform the operation and stop, otherwise continue.''\par
\hang third byte: |op_byte|, indicates a ligature step if less than~128,
  a kern step otherwise.\par
\hang fourth byte: |remainder|.\par
\yskip\noindent
In a kern step, an
additional space equal to |kern[256*(op_byte-128)+remainder]| is inserted
between the current character and |next_char|. This amount is
often negative, so that the characters are brought closer together
by kerning; but it might be positive.

There are eight kinds of ligature steps, having |op_byte| codes $4a+2b+c$ where
$0\le a\le b+c$ and $0\le b,c\le1$. The character whose code is
|remainder| is inserted between the current character and |next_char|;
then the current character is deleted if $b=0$, and |next_char| is
deleted if $c=0$; then we pass over $a$~characters to reach the next
current character (which may have a ligature/kerning program of its own).

If the very first instruction of the |lig_kern| array has |skip_byte=255|,
the |next_char| byte is the so-called boundary character of this font;
the value of |next_char| need not lie between |bc| and~|ec|.
If the very last instruction of the |lig_kern| array has |skip_byte=255|,
there is a special ligature/kerning program for a boundary character at the
left, beginning at location |256*op_byte+remainder|.
The interpretation is that \TeX\ puts implicit boundary characters
before and after each consecutive string of characters from the same font.
These implicit characters do not appear in the output, but they can affect
ligatures and kerning.

If the very first instruction of a character's |lig_kern| program has
|skip_byte>128|, the program actually begins in location
|256*op_byte+remainder|. This feature allows access to large |lig_kern|
arrays, because the first instruction must otherwise
appear in a location |<=255|.

Any instruction with |skip_byte>128| in the |lig_kern| array must satisfy
the condition
$$\hbox{|256*op_byte+remainder<nl|.}$$
If such an instruction is encountered during
normal program execution, it denotes an unconditional halt; no ligature
or kerning command is performed.

@d stop_flag=128+min_quarterword
  {value indicating `\.{STOP}' in a lig/kern program}
@d kern_flag=128+min_quarterword {op code for a kern step}
@d skip_byte(#)==lig_kern[#].b0
@d next_char(#)==lig_kern[#].b1
@d op_byte(#)==lig_kern[#].b2
@d rem_byte(#)==lig_kern[#].b3

@ Extensible characters are specified by an |@!extensible_recipe|, which
consists of four bytes called |@!top|, |@!mid|, |@!bot|, and |@!rep| (in this
order). These bytes are the character codes of individual pieces used to
build up a large symbol.  If |top|, |mid|, or |bot| are zero, they are not
present in the built-up result. For example, an extensible vertical line is
like an extensible bracket, except that the top and bottom pieces are missing.

Let $T$, $M$, $B$, and $R$ denote the respective pieces, or an empty box
if the piece isn't present. Then the extensible characters have the form
$TR^kMR^kB$ from top to bottom, for some |k>=0|, unless $M$ is absent;
in the latter case we can have $TR^kB$ for both even and odd values of~|k|.
The width of the extensible character is the width of $R$; and the
height-plus-depth is the sum of the individual height-plus-depths of the
components used, since the pieces are butted together in a vertical list.

@d ext_top(#)==exten[#].b0 {|top| piece in a recipe}
@d ext_mid(#)==exten[#].b1 {|mid| piece in a recipe}
@d ext_bot(#)==exten[#].b2 {|bot| piece in a recipe}
@d ext_rep(#)==exten[#].b3 {|rep| piece in a recipe}

@ The final portion of a \.{TFM} file is the |param| array, which is another
sequence of |fix_word| values.

\yskip\hang|param[1]=slant| is the amount of italic slant, which is used
to help position accents. For example, |slant=.25| means that when you go
up one unit, you also go .25 units to the right. The |slant| is a pure
number; it is the only |fix_word| other than the design size itself that is
not scaled by the design size.
@^design size@>

\hang|param[2]=space| is the normal spacing between words in text.
Note that character @'40 in the font need not have anything to do with
blank spaces.

\hang|param[3]=space_stretch| is the amount of glue stretching between words.

\hang|param[4]=space_shrink| is the amount of glue shrinking between words.

\hang|param[5]=x_height| is the size of one ex in the font; it is also
the height of letters for which accents don't have to be raised or lowered.

\hang|param[6]=quad| is the size of one em in the font.

\hang|param[7]=extra_space| is the amount added to |param[2]| at the
ends of sentences.

\yskip\noindent
If fewer than seven parameters are present, \TeX\ sets the missing parameters
to zero.

@d slant_code=1
@d space_code=2
@d space_stretch_code=3
@d space_shrink_code=4
@d x_height_code=5
@d quad_code=6
@d extra_space_code=7

@ So that is what \.{TFM} files hold. One of \MF's duties is to output such
information, and it does this all at once at the end of a job.
In order to prepare for such frenetic activity, it squirrels away the
necessary facts in various arrays as information becomes available.

Character dimensions (\&{charwd}, \&{charht}, \&{chardp}, and \&{charic})
are stored respectively in |tfm_width|, |tfm_height|, |tfm_depth|, and
|tfm_ital_corr|. Other information about a character (e.g., about
its ligatures or successors) is accessible via the |char_tag| and
|char_remainder| arrays. Other information about the font as a whole
is kept in additional arrays called |header_byte|, |lig_kern|,
|kern|, |exten|, and |param|.

@d undefined_label==lig_table_size {an undefined local label}

@<Glob...@>=
@!bc,@!ec:eight_bits; {smallest and largest character codes shipped out}
@!tfm_width:array[eight_bits] of scaled; {\&{charwd} values}
@!tfm_height:array[eight_bits] of scaled; {\&{charht} values}
@!tfm_depth:array[eight_bits] of scaled; {\&{chardp} values}
@!tfm_ital_corr:array[eight_bits] of scaled; {\&{charic} values}
@!char_exists:array[eight_bits] of boolean; {has this code been shipped out?}
@!char_tag:array[eight_bits] of no_tag..ext_tag; {|remainder| category}
@!char_remainder:array[eight_bits] of 0..lig_table_size; {the |remainder| byte}
@!header_byte:array[1..header_size] of -1..255;
  {bytes of the \.{TFM} header, or $-1$ if unset}
@!lig_kern:array[0..lig_table_size] of four_quarters; {the ligature/kern table}
@!nl:0..32767-256; {the number of ligature/kern steps so far}
@!kern:array[0..max_kerns] of scaled; {distinct kerning amounts}
@!nk:0..max_kerns; {the number of distinct kerns so far}
@!exten:array[eight_bits] of four_quarters; {extensible character recipes}
@!ne:0..256; {the number of extensible characters so far}
@!param:array[1..max_font_dimen] of scaled; {\&{fontdimen} parameters}
@!np:0..max_font_dimen; {the largest \&{fontdimen} parameter specified so far}
@!nw,@!nh,@!nd,@!ni:0..256; {sizes of \.{TFM} subtables}
@!skip_table:array[eight_bits] of 0..lig_table_size; {local label status}
@!lk_started:boolean; {has there been a lig/kern step in this command yet?}
@!bchar:integer; {right boundary character}
@!bch_label:0..lig_table_size; {left boundary starting location}
@!ll,@!lll:0..lig_table_size; {registers used for lig/kern processing}
@!label_loc:array[0..256] of -1..lig_table_size; {lig/kern starting addresses}
@!label_char:array[1..256] of eight_bits; {characters for |label_loc|}
@!label_ptr:0..256; {highest position occupied in |label_loc|}

@ @<Set init...@>=
for k:=0 to 255 do
  begin tfm_width[k]:=0; tfm_height[k]:=0; tfm_depth[k]:=0; tfm_ital_corr[k]:=0;
  char_exists[k]:=false; char_tag[k]:=no_tag; char_remainder[k]:=0;
  skip_table[k]:=undefined_label;
  end;
for k:=1 to header_size do header_byte[k]:=-1;
bc:=255; ec:=0; nl:=0; nk:=0; ne:=0; np:=0;@/
internal[boundary_char]:=-unity;
bch_label:=undefined_label;@/
label_loc[0]:=-1; label_ptr:=0;

@ @<Declare the function called |tfm_check|@>=
function tfm_check(@!m:small_number):scaled;
begin if abs(internal[m])>=fraction_half then
  begin print_err("Enormous "); print(int_name[m]);
@.Enormous charwd...@>
@.Enormous chardp...@>
@.Enormous charht...@>
@.Enormous charic...@>
@.Enormous designsize...@>
  print(" has been reduced");
  help1("Font metric dimensions must be less than 2048pt.");
  put_get_error;
  if internal[m]>0 then tfm_check:=fraction_half-1
  else tfm_check:=1-fraction_half;
  end
else tfm_check:=internal[m];
end;

@ @<Store the width information for character code~|c|@>=
if c<bc then bc:=c;
if c>ec then ec:=c;
char_exists[c]:=true;
gf_dx[c]:=internal[char_dx]; gf_dy[c]:=internal[char_dy];
tfm_width[c]:=tfm_check(char_wd);
tfm_height[c]:=tfm_check(char_ht);
tfm_depth[c]:=tfm_check(char_dp);
tfm_ital_corr[c]:=tfm_check(char_ic)

@ Now let's consider \MF's special \.{TFM}-oriented commands.

@<Cases of |do_statement|...@>=
tfm_command: do_tfm_command;

@ @d char_list_code=0
@d lig_table_code=1
@d extensible_code=2
@d header_byte_code=3
@d font_dimen_code=4

@<Put each...@>=
primitive("charlist",tfm_command,char_list_code);@/
@!@:char_list_}{\&{charlist} primitive@>
primitive("ligtable",tfm_command,lig_table_code);@/
@!@:lig_table_}{\&{ligtable} primitive@>
primitive("extensible",tfm_command,extensible_code);@/
@!@:extensible_}{\&{extensible} primitive@>
primitive("headerbyte",tfm_command,header_byte_code);@/
@!@:header_byte_}{\&{headerbyte} primitive@>
primitive("fontdimen",tfm_command,font_dimen_code);@/
@!@:font_dimen_}{\&{fontdimen} primitive@>

@ @<Cases of |print_cmd...@>=
tfm_command: case m of
  char_list_code:print("charlist");
  lig_table_code:print("ligtable");
  extensible_code:print("extensible");
  header_byte_code:print("headerbyte");
  othercases print("fontdimen")
  endcases;

@ @<Declare action procedures for use by |do_statement|@>=
function get_code:eight_bits; {scans a character code value}
label found;
var @!c:integer; {the code value found}
begin get_x_next; scan_expression;
if cur_type=known then
  begin c:=round_unscaled(cur_exp);
  if c>=0 then if c<256 then goto found;
  end
else if cur_type=string_type then if length(cur_exp)=1 then
  begin c:=so(str_pool[str_start[cur_exp]]); goto found;
  end;
exp_err("Invalid code has been replaced by 0");
@.Invalid code...@>
help2("I was looking for a number between 0 and 255, or for a")@/
  ("string of length 1. Didn't find it; will use 0 instead.");
put_get_flush_error(0); c:=0;
found: get_code:=c;
end;

@ @<Declare action procedures for use by |do_statement|@>=
procedure set_tag(@!c:halfword;@!t:small_number;@!r:halfword);
begin if char_tag[c]=no_tag then
  begin char_tag[c]:=t; char_remainder[c]:=r;
  if t=lig_tag then
    begin incr(label_ptr); label_loc[label_ptr]:=r; label_char[label_ptr]:=c;
    end;
  end
else @<Complain about a character tag conflict@>;
end;

@ @<Complain about a character tag conflict@>=
begin print_err("Character ");
if (c>" ")and(c<127) then print(c)
else if c=256 then print("||")
else  begin print("code "); print_int(c);
  end;
print(" is already ");
@.Character c is already...@>
case char_tag[c] of
lig_tag: print("in a ligtable");
list_tag: print("in a charlist");
ext_tag: print("extensible");
end; {there are no other cases}
help2("It's not legal to label a character more than once.")@/
  ("So I'll not change anything just now.");
put_get_error; end

@ @<Declare action procedures for use by |do_statement|@>=
procedure do_tfm_command;
label continue,done;
var @!c,@!cc:0..256; {character codes}
@!k:0..max_kerns; {index into the |kern| array}
@!j:integer; {index into |header_byte| or |param|}
begin case cur_mod of
char_list_code: begin c:=get_code;
     {we will store a list of character successors}
  while cur_cmd=colon do
    begin cc:=get_code; set_tag(c,list_tag,cc); c:=cc;
    end;
  end;
lig_table_code: @<Store a list of ligature/kern steps@>;
extensible_code: @<Define an extensible recipe@>;
header_byte_code, font_dimen_code: begin c:=cur_mod; get_x_next;
  scan_expression;
  if (cur_type<>known)or(cur_exp<half_unit) then
    begin exp_err("Improper location");
@.Improper location@>
    help2("I was looking for a known, positive number.")@/
      ("For safety's sake I'll ignore the present command.");
    put_get_error;
    end
  else  begin j:=round_unscaled(cur_exp);
    if cur_cmd<>colon then
      begin missing_err(":");
@.Missing `:'@>
      help1("A colon should follow a headerbyte or fontdimen location.");
      back_error;
      end;
    if c=header_byte_code then @<Store a list of header bytes@>
    else @<Store a list of font dimensions@>;
    end;
  end;
end; {there are no other cases}
end;

@ @<Store a list of ligature/kern steps@>=
begin lk_started:=false;
continue: get_x_next;
if(cur_cmd=skip_to)and lk_started then
 @<Process a |skip_to| command and |goto done|@>;
if cur_cmd=bchar_label then
  begin c:=256; cur_cmd:=colon;@+end
else begin back_input; c:=get_code;@+end;
if(cur_cmd=colon)or(cur_cmd=double_colon)then
  @<Record a label in a lig/kern subprogram and |goto continue|@>;
if cur_cmd=lig_kern_token then @<Compile a ligature/kern command@>
else  begin print_err("Illegal ligtable step");
@.Illegal ligtable step@>
  help1("I was looking for `=:' or `kern' here.");
  back_error; next_char(nl):=qi(0); op_byte(nl):=qi(0); rem_byte(nl):=qi(0);@/
  skip_byte(nl):=stop_flag+1; {this specifies an unconditional stop}
  end;
if nl=lig_table_size then overflow("ligtable size",lig_table_size);
@:METAFONT capacity exceeded ligtable size}{\quad ligtable size@>
incr(nl);
if cur_cmd=comma then goto continue;
if skip_byte(nl-1)<stop_flag then skip_byte(nl-1):=stop_flag;
done:end

@ @<Put each...@>=
primitive("=:",lig_kern_token,0);
@!@:=:_}{\.{=:} primitive@>
primitive("=:|",lig_kern_token,1);
@!@:=:/_}{\.{=:\char'174} primitive@>
primitive("=:|>",lig_kern_token,5);
@!@:=:/>_}{\.{=:\char'174>} primitive@>
primitive("|=:",lig_kern_token,2);
@!@:=:/_}{\.{\char'174=:} primitive@>
primitive("|=:>",lig_kern_token,6);
@!@:=:/>_}{\.{\char'174=:>} primitive@>
primitive("|=:|",lig_kern_token,3);
@!@:=:/_}{\.{\char'174=:\char'174} primitive@>
primitive("|=:|>",lig_kern_token,7);
@!@:=:/>_}{\.{\char'174=:\char'174>} primitive@>
primitive("|=:|>>",lig_kern_token,11);
@!@:=:/>_}{\.{\char'174=:\char'174>>} primitive@>
primitive("kern",lig_kern_token,128);
@!@:kern_}{\&{kern} primitive@>

@ @<Cases of |print_cmd...@>=
lig_kern_token: case m of
0:print("=:");
1:print("=:|");
2:print("|=:");
3:print("|=:|");
5:print("=:|>");
6:print("|=:>");
7:print("|=:|>");
11:print("|=:|>>");
othercases print("kern")
endcases;

@ Local labels are implemented by maintaining the |skip_table| array,
where |skip_table[c]| is either |undefined_label| or the address of the
most recent lig/kern instruction that skips to local label~|c|. In the
latter case, the |skip_byte| in that instruction will (temporarily)
be zero if there were no prior skips to this label, or it will be the
distance to the prior skip.

We may need to cancel skips that span more than 127 lig/kern steps.

@d cancel_skips(#)==ll:=#;
  repeat lll:=qo(skip_byte(ll)); skip_byte(ll):=stop_flag; ll:=ll-lll;
  until lll=0
@d skip_error(#)==begin print_err("Too far to skip");
@.Too far to skip@>
  help1("At most 127 lig/kern steps can separate skipto1 from 1::.");
  error; cancel_skips(#);
  end

@<Process a |skip_to| command and |goto done|@>=
begin c:=get_code;
if nl-skip_table[c]>128 then
  begin skip_error(skip_table[c]); skip_table[c]:=undefined_label;
  end;
if skip_table[c]=undefined_label then skip_byte(nl-1):=qi(0)
else skip_byte(nl-1):=qi(nl-skip_table[c]-1);
skip_table[c]:=nl-1; goto done;
end

@ @<Record a label in a lig/kern subprogram and |goto continue|@>=
begin if cur_cmd=colon then
  if c=256 then bch_label:=nl
  else set_tag(c,lig_tag,nl)
else if skip_table[c]<undefined_label then
  begin ll:=skip_table[c]; skip_table[c]:=undefined_label;
  repeat lll:=qo(skip_byte(ll));
  if nl-ll>128 then
    begin skip_error(ll); goto continue;
    end;
  skip_byte(ll):=qi(nl-ll-1); ll:=ll-lll;
  until lll=0;
  end;
goto continue;
end

@ @<Compile a ligature/kern...@>=
begin next_char(nl):=qi(c); skip_byte(nl):=qi(0);
if cur_mod<128 then {ligature op}
  begin op_byte(nl):=qi(cur_mod); rem_byte(nl):=qi(get_code);
  end
else  begin get_x_next; scan_expression;
  if cur_type<>known then
    begin exp_err("Improper kern");
@.Improper kern@>
    help2("The amount of kern should be a known numeric value.")@/
      ("I'm zeroing this one. Proceed, with fingers crossed.");
    put_get_flush_error(0);
    end;
  kern[nk]:=cur_exp;
  k:=0;@+while kern[k]<>cur_exp do incr(k);
  if k=nk then
    begin if nk=max_kerns then overflow("kern",max_kerns);
@:METAFONT capacity exceeded kern}{\quad kern@>
    incr(nk);
    end;
  op_byte(nl):=kern_flag+(k div 256);
  rem_byte(nl):=qi((k mod 256));
  end;
lk_started:=true;
end

@ @d missing_extensible_punctuation(#)==
  begin missing_err(#);
@.Missing `\char`\#'@>
  help1("I'm processing `extensible c: t,m,b,r'."); back_error;
  end

@<Define an extensible recipe@>=
begin if ne=256 then overflow("extensible",256);
@:METAFONT capacity exceeded extensible}{\quad extensible@>
c:=get_code; set_tag(c,ext_tag,ne);
if cur_cmd<>colon then missing_extensible_punctuation(":");
ext_top(ne):=qi(get_code);
if cur_cmd<>comma then missing_extensible_punctuation(",");
ext_mid(ne):=qi(get_code);
if cur_cmd<>comma then missing_extensible_punctuation(",");
ext_bot(ne):=qi(get_code);
if cur_cmd<>comma then missing_extensible_punctuation(",");
ext_rep(ne):=qi(get_code);
incr(ne);
end

@ @<Store a list of header bytes@>=
repeat if j>header_size then overflow("headerbyte",header_size);
@:METAFONT capacity exceeded headerbyte}{\quad headerbyte@>
header_byte[j]:=get_code; incr(j);
until cur_cmd<>comma

@ @<Store a list of font dimensions@>=
repeat if j>max_font_dimen then overflow("fontdimen",max_font_dimen);
@:METAFONT capacity exceeded fontdimen}{\quad fontdimen@>
while j>np do
  begin incr(np); param[np]:=0;
  end;
get_x_next; scan_expression;
if cur_type<>known then
  begin exp_err("Improper font parameter");
@.Improper font parameter@>
  help1("I'm zeroing this one. Proceed, with fingers crossed.");
  put_get_flush_error(0);
  end;
param[j]:=cur_exp; incr(j);
until cur_cmd<>comma

@ OK: We've stored all the data that is needed for the \.{TFM} file.
All that remains is to output it in the correct format.

An interesting problem needs to be solved in this connection, because
the \.{TFM} format allows at most 256~widths, 16~heights, 16~depths,
and 64~italic corrections. If the data has more distinct values than
this, we want to meet the necessary restrictions by perturbing the
given values as little as possible.

\MF\ solves this problem in two steps. First the values of a given
kind (widths, heights, depths, or italic corrections) are sorted;
then the list of sorted values is perturbed, if necessary.

The sorting operation is facilitated by having a special node of
essentially infinite |value| at the end of the current list.

@<Initialize table entries...@>=
value(inf_val):=fraction_four;

@ Straight linear insertion is good enough for sorting, since the lists
are usually not terribly long. As we work on the data, the current list
will start at |link(temp_head)| and end at |inf_val|; the nodes in this
list will be in increasing order of their |value| fields.

Given such a list, the |sort_in| function takes a value and returns a pointer
to where that value can be found in the list. The value is inserted in
the proper place, if necessary.

At the time we need to do these operations, most of \MF's work has been
completed, so we will have plenty of memory to play with. The value nodes
that are allocated for sorting will never be returned to free storage.

@d clear_the_list==link(temp_head):=inf_val

@p function sort_in(@!v:scaled):pointer;
label found;
var @!p,@!q,@!r:pointer; {list manipulation registers}
begin p:=temp_head;
loop@+  begin q:=link(p);
  if v<=value(q) then goto found;
  p:=q;
  end;
found: if v<value(q) then
  begin r:=get_node(value_node_size); value(r):=v; link(r):=q; link(p):=r;
  end;
sort_in:=link(p);
end;

@ Now we come to the interesting part, where we reduce the list if necessary
until it has the required size. The |min_cover| routine is basic to this
process; it computes the minimum number~|m| such that the values of the
current sorted list can be covered by |m|~intervals of width~|d|. It
also sets the global value |perturbation| to the smallest value $d'>d$
such that the covering found by this algorithm would be different.

In particular, |min_cover(0)| returns the number of distinct values in the
current list and sets |perturbation| to the minimum distance between
adjacent values.

@p function min_cover(@!d:scaled):integer;
var @!p:pointer; {runs through the current list}
@!l:scaled; {the least element covered by the current interval}
@!m:integer; {lower bound on the size of the minimum cover}
begin m:=0; p:=link(temp_head); perturbation:=el_gordo;
while p<>inf_val do
  begin incr(m); l:=value(p);
  repeat p:=link(p);
  until value(p)>l+d;
  if value(p)-l<perturbation then perturbation:=value(p)-l;
  end;
min_cover:=m;
end;

@ @<Glob...@>=
@!perturbation:scaled; {quantity related to \.{TFM} rounding}
@!excess:integer; {the list is this much too long}

@ The smallest |d| such that a given list can be covered with |m| intervals
is determined by the |threshold| routine, which is sort of an inverse
to |min_cover|. The idea is to increase the interval size rapidly until
finding the range, then to go sequentially until the exact borderline has
been discovered.

@p function threshold(@!m:integer):scaled;
var @!d:scaled; {lower bound on the smallest interval size}
begin excess:=min_cover(0)-m;
if excess<=0 then threshold:=0
else  begin repeat d:=perturbation;
  until min_cover(d+d)<=m;
  while min_cover(d)>m do d:=perturbation;
  threshold:=d;
  end;
end;

@ The |skimp| procedure reduces the current list to at most |m| entries,
by changing values if necessary. It also sets |info(p):=k| if |value(p)|
is the |k|th distinct value on the resulting list, and it sets
|perturbation| to the maximum amount by which a |value| field has
been changed. The size of the resulting list is returned as the
value of |skimp|.

@p function skimp(@!m:integer):integer;
var @!d:scaled; {the size of intervals being coalesced}
@!p,@!q,@!r:pointer; {list manipulation registers}
@!l:scaled; {the least value in the current interval}
@!v:scaled; {a compromise value}
begin d:=threshold(m); perturbation:=0;
q:=temp_head; m:=0; p:=link(temp_head);
while p<>inf_val do
  begin incr(m); l:=value(p); info(p):=m;
  if value(link(p))<=l+d then
    @<Replace an interval of values by its midpoint@>;
  q:=p; p:=link(p);
  end;
skimp:=m;
end;

@ @<Replace an interval...@>=
begin repeat p:=link(p); info(p):=m;
decr(excess);@+if excess=0 then d:=0;
until value(link(p))>l+d;
v:=l+half(value(p)-l);
if value(p)-v>perturbation then perturbation:=value(p)-v;
r:=q;
repeat r:=link(r); value(r):=v;
until r=p;
link(q):=p; {remove duplicate values from the current list}
end

@ A warning message is issued whenever something is perturbed by
more than 1/16\thinspace pt.

@p procedure tfm_warning(@!m:small_number);
begin print_nl("(some "); print(int_name[m]);
@.some charwds...@>
@.some chardps...@>
@.some charhts...@>
@.some charics...@>
print(" values had to be adjusted by as much as ");
print_scaled(perturbation); print("pt)");
end;

@ Here's an example of how we use these routines.
The width data needs to be perturbed only if there are 256 distinct
widths, but \MF\ must check for this case even though it is
highly unusual.

An integer variable |k| will be defined when we use this code.
The |dimen_head| array will contain pointers to the sorted
lists of dimensions.

@<Massage the \.{TFM} widths@>=
clear_the_list;
for k:=bc to ec do if char_exists[k] then
  tfm_width[k]:=sort_in(tfm_width[k]);
nw:=skimp(255)+1; dimen_head[1]:=link(temp_head);
if perturbation>=@'10000 then tfm_warning(char_wd)

@ @<Glob...@>=
@!dimen_head:array[1..4] of pointer; {lists of \.{TFM} dimensions}

@ Heights, depths, and italic corrections are different from widths
not only because their list length is more severely restricted, but
also because zero values do not need to be put into the lists.

@<Massage the \.{TFM} heights, depths, and italic corrections@>=
clear_the_list;
for k:=bc to ec do if char_exists[k] then
  if tfm_height[k]=0 then tfm_height[k]:=zero_val
  else tfm_height[k]:=sort_in(tfm_height[k]);
nh:=skimp(15)+1; dimen_head[2]:=link(temp_head);
if perturbation>=@'10000 then tfm_warning(char_ht);
clear_the_list;
for k:=bc to ec do if char_exists[k] then
  if tfm_depth[k]=0 then tfm_depth[k]:=zero_val
  else tfm_depth[k]:=sort_in(tfm_depth[k]);
nd:=skimp(15)+1; dimen_head[3]:=link(temp_head);
if perturbation>=@'10000 then tfm_warning(char_dp);
clear_the_list;
for k:=bc to ec do if char_exists[k] then
  if tfm_ital_corr[k]=0 then tfm_ital_corr[k]:=zero_val
  else tfm_ital_corr[k]:=sort_in(tfm_ital_corr[k]);
ni:=skimp(63)+1; dimen_head[4]:=link(temp_head);
if perturbation>=@'10000 then tfm_warning(char_ic)

@ @<Initialize table entries...@>=
value(zero_val):=0; info(zero_val):=0;

@ Bytes 5--8 of the header are set to the design size, unless the user has
some crazy reason for specifying them differently.
@^design size@>

Error messages are not allowed at the time this procedure is called,
so a warning is printed instead.

The value of |max_tfm_dimen| is calculated so that
$$\hbox{|make_scaled(16*max_tfm_dimen,internal[design_size])|}
 < \\{three\_bytes}.$$

@d three_bytes==@'100000000 {$2^{24}$}

@p procedure fix_design_size;
var @!d:scaled; {the design size}
begin d:=internal[design_size];
if (d<unity)or(d>=fraction_half) then
  begin if d<>0 then
    print_nl("(illegal design size has been changed to 128pt)");
@.illegal design size...@>
  d:=@'40000000; internal[design_size]:=d;
  end;
if header_byte[5]<0 then if header_byte[6]<0 then
  if header_byte[7]<0 then if header_byte[8]<0 then
  begin header_byte[5]:=d div @'4000000;
  header_byte[6]:=(d div 4096) mod 256;
  header_byte[7]:=(d div 16) mod 256;
  header_byte[8]:=(d mod 16)*16;
  end;
max_tfm_dimen:=16*internal[design_size]-1-internal[design_size] div @'10000000;
if max_tfm_dimen>=fraction_half then max_tfm_dimen:=fraction_half-1;
end;

@ The |dimen_out| procedure computes a |fix_word| relative to the
design size. If the data was out of range, it is corrected and the
global variable |tfm_changed| is increased by~one.

@p function dimen_out(@!x:scaled):integer;
begin if abs(x)>max_tfm_dimen then
  begin incr(tfm_changed);
  if x>0 then x:=max_tfm_dimen@+else x:=-max_tfm_dimen;
  end;
x:=make_scaled(x*16,internal[design_size]);
dimen_out:=x;
end;

@ @<Glob...@>=
@!max_tfm_dimen:scaled; {bound on widths, heights, kerns, etc.}
@!tfm_changed:integer; {the number of data entries that were out of bounds}

@ If the user has not specified any of the first four header bytes,
the |fix_check_sum| procedure replaces them by a ``check sum'' computed
from the |tfm_width| data relative to the design size.
@^check sum@>

@p procedure fix_check_sum;
label exit;
var @!k:eight_bits; {runs through character codes}
@!b1,@!b2,@!b3,@!b4:eight_bits; {bytes of the check sum}
@!x:integer; {hash value used in check sum computation}
begin if header_byte[1]<0 then if header_byte[2]<0 then
  if header_byte[3]<0 then if header_byte[4]<0 then
  begin @<Compute a check sum in |(b1,b2,b3,b4)|@>;
  header_byte[1]:=b1; header_byte[2]:=b2;
  header_byte[3]:=b3; header_byte[4]:=b4; return;
  end;
for k:=1 to 4 do if header_byte[k]<0 then header_byte[k]:=0;
exit:end;

@ @<Compute a check sum in |(b1,b2,b3,b4)|@>=
b1:=bc; b2:=ec; b3:=bc; b4:=ec; tfm_changed:=0;
for k:=bc to ec do if char_exists[k] then
  begin x:=dimen_out(value(tfm_width[k]))+(k+4)*@'20000000; {this is positive}
  b1:=(b1+b1+x) mod 255;
  b2:=(b2+b2+x) mod 253;
  b3:=(b3+b3+x) mod 251;
  b4:=(b4+b4+x) mod 247;
  end

@ Finally we're ready to actually write the \.{TFM} information.
Here are some utility routines for this purpose.

@d tfm_out(#)==write(tfm_file,#) {output one byte to |tfm_file|}

@p procedure tfm_two(@!x:integer); {output two bytes to |tfm_file|}
begin tfm_out(x div 256); tfm_out(x mod 256);
end;
@#
procedure tfm_four(@!x:integer); {output four bytes to |tfm_file|}
begin if x>=0 then tfm_out(x div three_bytes)
else  begin x:=x+@'10000000000; {use two's complement for negative values}
  x:=x+@'10000000000;
  tfm_out((x div three_bytes) + 128);
  end;
x:=x mod three_bytes; tfm_out(x div unity);
x:=x mod unity; tfm_out(x div @'400);
tfm_out(x mod @'400);
end;
@#
procedure tfm_qqqq(@!x:four_quarters); {output four quarterwords to |tfm_file|}
begin tfm_out(qo(x.b0)); tfm_out(qo(x.b1)); tfm_out(qo(x.b2));
tfm_out(qo(x.b3));
end;

@ @<Finish the \.{TFM} file@>=
if job_name=0 then open_log_file;
pack_job_name(".tfm");
while not b_open_out(tfm_file) do
  prompt_file_name("file name for font metrics",".tfm");
metric_file_name:=b_make_name_string(tfm_file);
@<Output the subfile sizes and header bytes@>;
@<Output the character information bytes, then
  output the dimensions themselves@>;
@<Output the ligature/kern program@>;
@<Output the extensible character recipes and the font metric parameters@>;
@!stat if internal[tracing_stats]>0 then
  @<Log the subfile sizes of the \.{TFM} file@>;@;@+tats@/
print_nl("Font metrics written on "); slow_print(metric_file_name);
print_char(".");
@.Font metrics written...@>
b_close(tfm_file)

@ Integer variables |lh|, |k|, and |lk_offset| will be defined when we use
this code.

@<Output the subfile sizes and header bytes@>=
k:=header_size;
while header_byte[k]<0 do decr(k);
lh:=(k+3) div 4; {this is the number of header words}
if bc>ec then bc:=1; {if there are no characters, |ec=0| and |bc=1|}
@<Compute the ligature/kern program offset and implant the
  left boundary label@>;
tfm_two(6+lh+(ec-bc+1)+nw+nh+nd+ni+nl+lk_offset+nk+ne+np);
  {this is the total number of file words that will be output}
tfm_two(lh); tfm_two(bc); tfm_two(ec); tfm_two(nw); tfm_two(nh);
tfm_two(nd); tfm_two(ni); tfm_two(nl+lk_offset); tfm_two(nk); tfm_two(ne);
tfm_two(np);
for k:=1 to 4*lh do
  begin if header_byte[k]<0 then header_byte[k]:=0;
  tfm_out(header_byte[k]);
  end

@ @<Output the character information bytes...@>=
for k:=bc to ec do
  if not char_exists[k] then tfm_four(0)
  else  begin tfm_out(info(tfm_width[k])); {the width index}
    tfm_out((info(tfm_height[k]))*16+info(tfm_depth[k]));
    tfm_out((info(tfm_ital_corr[k]))*4+char_tag[k]);
    tfm_out(char_remainder[k]);
    end;
tfm_changed:=0;
for k:=1 to 4 do
  begin tfm_four(0); p:=dimen_head[k];
  while p<>inf_val do
    begin tfm_four(dimen_out(value(p))); p:=link(p);
    end;
  end

@ We need to output special instructions at the beginning of the
|lig_kern| array in order to specify the right boundary character
and/or to handle starting addresses that exceed 255. The |label_loc|
and |label_char| arrays have been set up to record all the
starting addresses; we have $-1=|label_loc|[0]<|label_loc|[1]\le\cdots
\le|label_loc|[|label_ptr]|$.

@<Compute the ligature/kern program offset...@>=
bchar:=round_unscaled(internal[boundary_char]);
if(bchar<0)or(bchar>255)then
  begin bchar:=-1; lk_started:=false; lk_offset:=0;@+end
else begin lk_started:=true; lk_offset:=1;@+end;
@<Find the minimum |lk_offset| and adjust all remainders@>;
if bch_label<undefined_label then
  begin skip_byte(nl):=qi(255); next_char(nl):=qi(0);
  op_byte(nl):=qi(((bch_label+lk_offset)div 256));
  rem_byte(nl):=qi(((bch_label+lk_offset)mod 256));
  incr(nl); {possibly |nl=lig_table_size+1|}
  end

@ @<Find the minimum |lk_offset|...@>=
k:=label_ptr; {pointer to the largest unallocated label}
if label_loc[k]+lk_offset>255 then
  begin lk_offset:=0; lk_started:=false; {location 0 can do double duty}
  repeat char_remainder[label_char[k]]:=lk_offset;
  while label_loc[k-1]=label_loc[k] do
    begin decr(k); char_remainder[label_char[k]]:=lk_offset;
    end;
  incr(lk_offset); decr(k);
  until lk_offset+label_loc[k]<256;
    {N.B.: |lk_offset=256| satisfies this when |k=0|}
  end;
if lk_offset>0 then
  while k>0 do
    begin char_remainder[label_char[k]]
     :=char_remainder[label_char[k]]+lk_offset;
    decr(k);
    end

@ @<Output the ligature/kern program@>=
for k:=0 to 255 do if skip_table[k]<undefined_label then
  begin print_nl("(local label "); print_int(k); print(":: was missing)");
@.local label l:: was missing@>
  cancel_skips(skip_table[k]);
  end;
if lk_started then {|lk_offset=1| for the special |bchar|}
  begin tfm_out(255); tfm_out(bchar); tfm_two(0);
  end
else for k:=1 to lk_offset do {output the redirection specs}
  begin ll:=label_loc[label_ptr];
  if bchar<0 then
    begin tfm_out(254); tfm_out(0);
    end
  else begin tfm_out(255); tfm_out(bchar);
    end;
  tfm_two(ll+lk_offset);
  repeat decr(label_ptr);
  until label_loc[label_ptr]<ll;
  end;
for k:=0 to nl-1 do tfm_qqqq(lig_kern[k]);
for k:=0 to nk-1 do tfm_four(dimen_out(kern[k]))

@ @<Output the extensible character recipes...@>=
for k:=0 to ne-1 do tfm_qqqq(exten[k]);
for k:=1 to np do
  if k=1 then
    if abs(param[1])<fraction_half then tfm_four(param[1]*16)
    else  begin incr(tfm_changed);
      if param[1]>0 then tfm_four(el_gordo)
      else tfm_four(-el_gordo);
      end
  else tfm_four(dimen_out(param[k]));
if tfm_changed>0 then
  begin if tfm_changed=1 then print_nl("(a font metric dimension")
@.a font metric dimension...@>
  else  begin print_nl("("); print_int(tfm_changed);
@.font metric dimensions...@>
    print(" font metric dimensions");
    end;
  print(" had to be decreased)");
  end

@ @<Log the subfile sizes of the \.{TFM} file@>=
begin wlog_ln(' ');
if bch_label<undefined_label then decr(nl);
wlog_ln('(You used ',nw:1,'w,',@| nh:1,'h,',@| nd:1,'d,',@| ni:1,'i,',@|
 nl:1,'l,',@| nk:1,'k,',@| ne:1,'e,',@|
 np:1,'p metric file positions');
wlog_ln('  out of ',@| '256w,16h,16d,64i,',@|
 lig_table_size:1,'l,',max_kerns:1,'k,256e,',@|
 max_font_dimen:1,'p)');
end

@* \[46] Generic font file format.
The most important output produced by a typical run of \MF\ is the
``generic font'' (\.{GF}) file that specifies the bit patterns of the
characters that have been drawn. The term {\sl generic\/} indicates that
this file format doesn't match the conventions of any name-brand manufacturer;
but it is easy to convert \.{GF} files to the special format required by
almost all digital phototypesetting equipment. There's a strong analogy
between the \.{DVI} files written by \TeX\ and the \.{GF} files written
by \MF; and, in fact, the file formats have a lot in common.

A \.{GF} file is a stream of 8-bit bytes that may be
regarded as a series of commands in a machine-like language. The first
byte of each command is the operation code, and this code is followed by
zero or more bytes that provide parameters to the command. The parameters
themselves may consist of several consecutive bytes; for example, the
`|boc|' (beginning of character) command has six parameters, each of
which is four bytes long. Parameters are usually regarded as nonnegative
integers; but four-byte-long parameters can be either positive or
negative, hence they range in value from $-2^{31}$ to $2^{31}-1$.
As in \.{TFM} files, numbers that occupy
more than one byte position appear in BigEndian order,
and negative numbers appear in two's complement notation.

A \.{GF} file consists of a ``preamble,'' followed by a sequence of one or
more ``characters,'' followed by a ``postamble.'' The preamble is simply a
|pre| command, with its parameters that introduce the file; this must come
first.  Each ``character'' consists of a |boc| command, followed by any
number of other commands that specify ``black'' pixels,
followed by an |eoc| command. The characters appear in the order that \MF\
generated them. If we ignore no-op commands (which are allowed between any
two commands in the file), each |eoc| command is immediately followed by a
|boc| command, or by a |post| command; in the latter case, there are no
more characters in the file, and the remaining bytes form the postamble.
Further details about the postamble will be explained later.

Some parameters in \.{GF} commands are ``pointers.'' These are four-byte
quantities that give the location number of some other byte in the file;
the first file byte is number~0, then comes number~1, and so on.

@ The \.{GF} format is intended to be both compact and easily interpreted
by a machine. Compactness is achieved by making most of the information
relative instead of absolute. When a \.{GF}-reading program reads the
commands for a character, it keeps track of two quantities: (a)~the current
column number,~|m|; and (b)~the current row number,~|n|.  These are 32-bit
signed integers, although most actual font formats produced from \.{GF}
files will need to curtail this vast range because of practical
limitations. (\MF\ output will never allow $\vert m\vert$ or $\vert
n\vert$ to get extremely large, but the \.{GF} format tries to be more general.)

How do \.{GF}'s row and column numbers correspond to the conventions
of \TeX\ and \MF? Well, the ``reference point'' of a character, in \TeX's
view, is considered to be at the lower left corner of the pixel in row~0
and column~0. This point is the intersection of the baseline with the left
edge of the type; it corresponds to location $(0,0)$ in \MF\ programs.
Thus the pixel in \.{GF} row~0 and column~0 is \MF's unit square, comprising the
region of the plane whose coordinates both lie between 0 and~1. The
pixel in \.{GF} row~|n| and column~|m| consists of the points whose \MF\
coordinates |(x,y)| satisfy |m<=x<=m+1| and |n<=y<=n+1|.  Negative values of
|m| and~|x| correspond to columns of pixels {\sl left\/} of the reference
point; negative values of |n| and~|y| correspond to rows of pixels {\sl
below\/} the baseline.

Besides |m| and |n|, there's also a third aspect of the current
state, namely the @!|paint_switch|, which is always either |black| or
|white|. Each \\{paint} command advances |m| by a specified amount~|d|,
and blackens the intervening pixels if |paint_switch=black|; then
the |paint_switch| changes to the opposite state. \.{GF}'s commands are
designed so that |m| will never decrease within a row, and |n| will never
increase within a character; hence there is no way to whiten a pixel that
has been blackened.

@ Here is a list of all the commands that may appear in a \.{GF} file. Each
command is specified by its symbolic name (e.g., |boc|), its opcode byte
(e.g., 67), and its parameters (if any). The parameters are followed
by a bracketed number telling how many bytes they occupy; for example,
`|d[2]|' means that parameter |d| is two bytes long.

\yskip\hang|paint_0| 0. This is a \\{paint} command with |d=0|; it does
nothing but change the |paint_switch| from \\{black} to \\{white} or vice~versa.

\yskip\hang\\{paint\_1} through \\{paint\_63} (opcodes 1 to 63).
These are \\{paint} commands with |d=1| to~63, defined as follows: If
|paint_switch=black|, blacken |d|~pixels of the current row~|n|,
in columns |m| through |m+d-1| inclusive. Then, in any case,
complement the |paint_switch| and advance |m| by~|d|.

\yskip\hang|paint1| 64 |d[1]|. This is a \\{paint} command with a specified
value of~|d|; \MF\ uses it to paint when |64<=d<256|.

\yskip\hang|@!paint2| 65 |d[2]|. Same as |paint1|, but |d|~can be as high
as~65535.

\yskip\hang|@!paint3| 66 |d[3]|. Same as |paint1|, but |d|~can be as high
as $2^{24}-1$. \MF\ never needs this command, and it is hard to imagine
anybody making practical use of it; surely a more compact encoding will be
desirable when characters can be this large. But the command is there,
anyway, just in case.

\yskip\hang|boc| 67 |c[4]| |p[4]| |min_m[4]| |max_m[4]| |min_n[4]|
|max_n[4]|. Beginning of a character:  Here |c| is the character code, and
|p| points to the previous character beginning (if any) for characters having
this code number modulo 256.  (The pointer |p| is |-1| if there was no
prior character with an equivalent code.) The values of registers |m| and |n|
defined by the instructions that follow for this character must
satisfy |min_m<=m<=max_m| and |min_n<=n<=max_n|.  (The values of |max_m| and
|min_n| need not be the tightest bounds possible.)  When a \.{GF}-reading
program sees a |boc|, it can use |min_m|, |max_m|, |min_n|, and |max_n| to
initialize the bounds of an array. Then it sets |m:=min_m|, |n:=max_n|, and
|paint_switch:=white|.

\yskip\hang|boc1| 68 |c[1]| |@!del_m[1]| |max_m[1]| |@!del_n[1]| |max_n[1]|.
Same as |boc|, but |p| is assumed to be~$-1$; also |del_m=max_m-min_m|
and |del_n=max_n-min_n| are given instead of |min_m| and |min_n|.
The one-byte parameters must be between 0 and 255, inclusive.
\ (This abbreviated |boc| saves 19~bytes per character, in common cases.)

\yskip\hang|eoc| 69. End of character: All pixels blackened so far
constitute the pattern for this character. In particular, a completely
blank character might have |eoc| immediately following |boc|.

\yskip\hang|skip0| 70. Decrease |n| by 1 and set |m:=min_m|,
|paint_switch:=white|. \ (This finishes one row and begins another,
ready to whiten the leftmost pixel in the new row.)

\yskip\hang|skip1| 71 |d[1]|. Decrease |n| by |d+1|, set |m:=min_m|, and set
|paint_switch:=white|. This is a way to produce |d| all-white rows.

\yskip\hang|@!skip2| 72 |d[2]|. Same as |skip1|, but |d| can be as large
as 65535.

\yskip\hang|@!skip3| 73 |d[3]|. Same as |skip1|, but |d| can be as large
as $2^{24}-1$. \MF\ obviously never needs this command.

\yskip\hang|new_row_0| 74. Decrease |n| by 1 and set |m:=min_m|,
|paint_switch:=black|. \ (This finishes one row and begins another,
ready to {\sl blacken\/} the leftmost pixel in the new row.)

\yskip\hang|@!new_row_1| through |@!new_row_164| (opcodes 75 to 238). Same as
|new_row_0|, but with |m:=min_m+1| through |min_m+164|, respectively.

\yskip\hang|xxx1| 239 |k[1]| |x[k]|. This command is undefined in
general; it functions as a $(k+2)$-byte |no_op| unless special \.{GF}-reading
programs are being used. \MF\ generates \\{xxx} commands when encountering
a \&{special} string; this occurs in the \.{GF} file only between
characters, after the preamble, and before the postamble. However,
\\{xxx} commands might appear within characters,
in \.{GF} files generated by other
processors. It is recommended that |x| be a string having the form of a
keyword followed by possible parameters relevant to that keyword.

\yskip\hang|@!xxx2| 240 |k[2]| |x[k]|. Like |xxx1|, but |0<=k<65536|.

\yskip\hang|xxx3| 241 |k[3]| |x[k]|. Like |xxx1|, but |0<=k<@t$2^{24}$@>|.
\MF\ uses this when sending a \&{special} string whose length exceeds~255.

\yskip\hang|@!xxx4| 242 |k[4]| |x[k]|. Like |xxx1|, but |k| can be
ridiculously large; |k| mustn't be negative.

\yskip\hang|yyy| 243 |y[4]|. This command is undefined in general;
it functions as a 5-byte |no_op| unless special \.{GF}-reading programs
are being used. \MF\ puts |scaled| numbers into |yyy|'s, as a
result of \&{numspecial} commands; the intent is to provide numeric
parameters to \\{xxx} commands that immediately precede.

\yskip\hang|@!no_op| 244. No operation, do nothing. Any number of |no_op|'s
may occur between \.{GF} commands, but a |no_op| cannot be inserted between
a command and its parameters or between two parameters.

\yskip\hang|char_loc| 245 |c[1]| |dx[4]| |dy[4]| |w[4]| |p[4]|.
This command will appear only in the postamble, which will be explained shortly.

\yskip\hang|@!char_loc0| 246 |c[1]| |@!dm[1]| |w[4]| |p[4]|.
Same as |char_loc|, except that |dy| is assumed to be zero, and the value
of~|dx| is taken to be |65536*dm|, where |0<=dm<256|.

\yskip\hang|pre| 247 |i[1]| |k[1]| |x[k]|.
Beginning of the preamble; this must come at the very beginning of the
file. Parameter |i| is an identifying number for \.{GF} format, currently
131. The other information is merely commentary; it is not given
special interpretation like \\{xxx} commands are. (Note that \\{xxx}
commands may immediately follow the preamble, before the first |boc|.)

\yskip\hang|post| 248. Beginning of the postamble, see below.

\yskip\hang|post_post| 249. Ending of the postamble, see below.

\yskip\noindent Commands 250--255 are undefined at the present time.

@d gf_id_byte=131 {identifies the kind of \.{GF} files described here}

@ \MF\ refers to the following opcodes explicitly.

@d paint_0=0 {beginning of the \\{paint} commands}
@d paint1=64 {move right a given number of columns, then
  black${}\swap{}$white}
@d boc=67 {beginning of a character}
@d boc1=68 {short form of |boc|}
@d eoc=69 {end of a character}
@d skip0=70 {skip no blank rows}
@d skip1=71 {skip over blank rows}
@d new_row_0=74 {move down one row and then right}
@d max_new_row=164 {the largest \\{new\_row} command is |new_row_164|}
@d xxx1=239 {for \&{special} strings}
@d xxx3=241 {for long \&{special} strings}
@d yyy=243 {for \&{numspecial} numbers}
@d char_loc=245 {character locators in the postamble}
@d pre=247 {preamble}
@d post=248 {postamble beginning}
@d post_post=249 {postamble ending}

@ The last character in a \.{GF} file is followed by `|post|'; this command
introduces the postamble, which summarizes important facts that \MF\ has
accumulated. The postamble has the form
$$\vbox{\halign{\hbox{#\hfil}\cr
  |post| |p[4]| |@!ds[4]| |@!cs[4]| |@!hppp[4]| |@!vppp[4]|
   |@!min_m[4]| |@!max_m[4]| |@!min_n[4]| |@!max_n[4]|\cr
  $\langle\,$character locators$\,\rangle$\cr
  |post_post| |q[4]| |i[1]| 223's$[{\G}4]$\cr}}$$
Here |p| is a pointer to the byte following the final |eoc| in the file
(or to the byte following the preamble, if there are no characters);
it can be used to locate the beginning of \\{xxx} commands
that might have preceded the postamble. The |ds| and |cs| parameters
@^design size@> @^check sum@>
give the design size and check sum, respectively, which are exactly the
values put into the header of the \.{TFM} file that \MF\ produces (or
would produce) on this run. Parameters |hppp| and |vppp| are the ratios of
pixels per point, horizontally and vertically, expressed as |scaled| integers
(i.e., multiplied by $2^{16}$); they can be used to correlate the font
with specific device resolutions, magnifications, and ``at sizes.''  Then
come |min_m|, |max_m|, |min_n|, and |max_n|, which bound the values that
registers |m| and~|n| assume in all characters in this \.{GF} file.
(These bounds need not be the best possible; |max_m| and |min_n| may, on the
other hand, be tighter than the similar bounds in |boc| commands. For
example, some character may have |min_n=-100| in its |boc|, but it might
turn out that |n| never gets lower than |-50| in any character; then
|min_n| can have any value |<=-50|. If there are no characters in the file,
it's possible to have |min_m>max_m| and/or |min_n>max_n|.)

@ Character locators are introduced by |char_loc| commands,
which specify a character residue~|c|, character escapements (|dx,dy|),
a character width~|w|, and a pointer~|p|
to the beginning of that character. (If two or more characters have the
same code~|c| modulo 256, only the last will be indicated; the others can be
located by following backpointers. Characters whose codes differ by a
multiple of 256 are assumed to share the same font metric information,
hence the \.{TFM} file contains only residues of character codes modulo~256.
This convention is intended for oriental languages, when there are many
character shapes but few distinct widths.)
@^oriental characters@>@^Chinese characters@>@^Japanese characters@>

The character escapements (|dx,dy|) are the values of \MF's \&{chardx}
and \&{chardy} parameters; they are in units of |scaled| pixels;
i.e., |dx| is in horizontal pixel units times $2^{16}$, and |dy| is in
vertical pixel units times $2^{16}$.  This is the intended amount of
displacement after typesetting the character; for \.{DVI} files, |dy|
should be zero, but other document file formats allow nonzero vertical
escapement.

The character width~|w| duplicates the information in the \.{TFM} file; it
is a |fix_word| value relative to the design size, and it should be
independent of magnification.

The backpointer |p| points to the character's |boc|, or to the first of
a sequence of consecutive \\{xxx} or |yyy| or |no_op| commands that
immediately precede the |boc|, if such commands exist; such ``special''
commands essentially belong to the characters, while the special commands
after the final character belong to the postamble (i.e., to the font
as a whole). This convention about |p| applies also to the backpointers
in |boc| commands, even though it wasn't explained in the description
of~|boc|. @^backpointers@>

Pointer |p| might be |-1| if the character exists in the \.{TFM} file
but not in the \.{GF} file. This unusual situation can arise in \MF\ output
if the user had |proofing<0| when the character was being shipped out,
but then made |proofing>=0| in order to get a \.{GF} file.

@ The last part of the postamble, following the |post_post| byte that
signifies the end of the character locators, contains |q|, a pointer to the
|post| command that started the postamble.  An identification byte, |i|,
comes next; this currently equals~131, as in the preamble.

The |i| byte is followed by four or more bytes that are all equal to
the decimal number 223 (i.e., @'337 in octal). \MF\ puts out four to seven of
these trailing bytes, until the total length of the file is a multiple of
four bytes, since this works out best on machines that pack four bytes per
word; but any number of 223's is allowed, as long as there are at least four
of them. In effect, 223 is a sort of signature that is added at the very end.
@^Fuchs, David Raymond@>

This curious way to finish off a \.{GF} file makes it feasible for
\.{GF}-reading programs to find the postamble first, on most computers,
even though \MF\ wants to write the postamble last. Most operating
systems permit random access to individual words or bytes of a file, so
the \.{GF} reader can start at the end and skip backwards over the 223's
until finding the identification byte. Then it can back up four bytes, read
|q|, and move to byte |q| of the file. This byte should, of course,
contain the value 248 (|post|); now the postamble can be read, so the
\.{GF} reader can discover all the information needed for individual characters.

Unfortunately, however, standard \PASCAL\ does not include the ability to
@^system dependencies@>
access a random position in a file, or even to determine the length of a file.
Almost all systems nowadays provide the necessary capabilities, so \.{GF}
format has been designed to work most efficiently with modern operating systems.
But if \.{GF} files have to be processed under the restrictions of standard
\PASCAL, one can simply read them from front to back. This will
be adequate for most applications. However, the postamble-first approach
would facilitate a program that merges two \.{GF} files, replacing data
from one that is overridden by corresponding data in the other.

@* \[47] Shipping characters out.
The |ship_out| procedure, to be described below, is given a pointer to
an edge structure. Its mission is to describe the positive pixels
in \.{GF} form, outputting a ``character'' to |gf_file|.

Several global variables hold information about the font file as a whole:\
|gf_min_m|, |gf_max_m|, |gf_min_n|, and |gf_max_n| are the minimum and
maximum \.{GF} coordinates output so far; |gf_prev_ptr| is the byte number
following the preamble or the last |eoc| command in the output;
|total_chars| is the total number of characters (i.e., |boc..eoc| segments)
shipped out.  There's also an array, |char_ptr|, containing the starting
positions of each character in the file, as required for the postamble. If
character code~|c| has not yet been output, |char_ptr[c]=-1|.

@<Glob...@>=
@!gf_min_m,@!gf_max_m,@!gf_min_n,@!gf_max_n:integer; {bounding rectangle}
@!gf_prev_ptr:integer; {where the present/next character started/starts}
@!total_chars:integer; {the number of characters output so far}
@!char_ptr:array[eight_bits] of integer; {where individual characters started}
@!gf_dx,@!gf_dy:array[eight_bits] of integer; {device escapements}

@ @<Set init...@>=
gf_prev_ptr:=0; total_chars:=0;

@ The \.{GF} bytes are output to a buffer instead of being sent
byte-by-byte to |gf_file|, because this tends to save a lot of
subroutine-call overhead. \MF\ uses the same conventions for |gf_file|
as \TeX\ uses for its \\{dvi\_file}; hence if system-dependent
changes are needed, they should probably be the same for both programs.

The output buffer is divided into two parts of equal size; the bytes found
in |gf_buf[0..half_buf-1]| constitute the first half, and those in
|gf_buf[half_buf..gf_buf_size-1]| constitute the second. The global
variable |gf_ptr| points to the position that will receive the next
output byte. When |gf_ptr| reaches |gf_limit|, which is always equal
to one of the two values |half_buf| or |gf_buf_size|, the half buffer that
is about to be invaded next is sent to the output and |gf_limit| is
changed to its other value. Thus, there is always at least a half buffer's
worth of information present, except at the very beginning of the job.

Bytes of the \.{GF} file are numbered sequentially starting with 0;
the next byte to be generated will be number |gf_offset+gf_ptr|.

@<Types...@>=
@!gf_index=0..gf_buf_size; {an index into the output buffer}

@ Some systems may find it more efficient to make |gf_buf| a |packed|
array, since output of four bytes at once may be facilitated.
@^system dependencies@>

@<Glob...@>=
@!gf_buf:array[gf_index] of eight_bits; {buffer for \.{GF} output}
@!half_buf:gf_index; {half of |gf_buf_size|}
@!gf_limit:gf_index; {end of the current half buffer}
@!gf_ptr:gf_index; {the next available buffer address}
@!gf_offset:integer; {|gf_buf_size| times the number of times the
  output buffer has been fully emptied}

@ Initially the buffer is all in one piece; we will output half of it only
after it first fills up.

@<Set init...@>=
half_buf:=gf_buf_size div 2; gf_limit:=gf_buf_size; gf_ptr:=0;
gf_offset:=0;

@ The actual output of |gf_buf[a..b]| to |gf_file| is performed by calling
|write_gf(a,b)|. It is safe to assume that |a| and |b+1| will both be
multiples of 4 when |write_gf(a,b)| is called; therefore it is possible on
many machines to use efficient methods to pack four bytes per word and to
output an array of words with one system call.
@^system dependencies@>

@<Declare generic font output procedures@>=
procedure write_gf(@!a,@!b:gf_index);
var k:gf_index;
begin for k:=a to b do write(gf_file,gf_buf[k]);
end;

@ To put a byte in the buffer without paying the cost of invoking a procedure
each time, we use the macro |gf_out|.

@d gf_out(#)==@+begin gf_buf[gf_ptr]:=#; incr(gf_ptr);
  if gf_ptr=gf_limit then gf_swap;
  end

@<Declare generic font output procedures@>=
procedure gf_swap; {outputs half of the buffer}
begin if gf_limit=gf_buf_size then
  begin write_gf(0,half_buf-1); gf_limit:=half_buf;
  gf_offset:=gf_offset+gf_buf_size; gf_ptr:=0;
  end
else  begin write_gf(half_buf,gf_buf_size-1); gf_limit:=gf_buf_size;
  end;
end;

@ Here is how we clean out the buffer when \MF\ is all through; |gf_ptr|
will be a multiple of~4.

@<Empty the last bytes out of |gf_buf|@>=
if gf_limit=half_buf then write_gf(half_buf,gf_buf_size-1);
if gf_ptr>0 then write_gf(0,gf_ptr-1)

@ The |gf_four| procedure outputs four bytes in two's complement notation,
without risking arithmetic overflow.

@<Declare generic font output procedures@>=
procedure gf_four(@!x:integer);
begin if x>=0 then gf_out(x div three_bytes)
else  begin x:=x+@'10000000000;
  x:=x+@'10000000000;
  gf_out((x div three_bytes) + 128);
  end;
x:=x mod three_bytes; gf_out(x div unity);
x:=x mod unity; gf_out(x div @'400);
gf_out(x mod @'400);
end;

@ Of course, it's even easier to output just two or three bytes.

@<Declare generic font output procedures@>=
procedure gf_two(@!x:integer);
begin gf_out(x div @'400); gf_out(x mod @'400);
end;
@#
procedure gf_three(@!x:integer);
begin gf_out(x div unity); gf_out((x mod unity) div @'400);
gf_out(x mod @'400);
end;

@ We need a simple routine to generate a \\{paint}
command of the appropriate type.

@<Declare generic font output procedures@>=
procedure gf_paint(@!d:integer); {here |0<=d<65536|}
begin if d<64 then gf_out(paint_0+d)
else if d<256 then
  begin gf_out(paint1); gf_out(d);
  end
else  begin gf_out(paint1+1); gf_two(d);
  end;
end;

@ And |gf_string| outputs one or two strings. If the first string number
is nonzero, an \\{xxx} command is generated.

@<Declare generic font output procedures@>=
procedure gf_string(@!s,@!t:str_number);
var @!k:pool_pointer;
@!l:integer; {length of the strings to output}
begin if s<>0 then
  begin l:=length(s);
  if t<>0 then l:=l+length(t);
  if l<=255 then
    begin gf_out(xxx1); gf_out(l);
    end
  else  begin gf_out(xxx3); gf_three(l);
    end;
  for k:=str_start[s] to str_start[s+1]-1 do gf_out(so(str_pool[k]));
  end;
if t<>0 then for k:=str_start[t] to str_start[t+1]-1 do gf_out(so(str_pool[k]));
end;

@ The choice between |boc| commands is handled by |gf_boc|.

@d one_byte(#)== #>=0 then if #<256

@<Declare generic font output procedures@>=
procedure gf_boc(@!min_m,@!max_m,@!min_n,@!max_n:integer);
label exit;
begin if min_m<gf_min_m then gf_min_m:=min_m;
if max_n>gf_max_n then gf_max_n:=max_n;
if boc_p=-1 then if one_byte(boc_c) then
 if one_byte(max_m-min_m) then if one_byte(max_m) then
  if one_byte(max_n-min_n) then if one_byte(max_n) then
  begin gf_out(boc1); gf_out(boc_c);@/
  gf_out(max_m-min_m); gf_out(max_m);
  gf_out(max_n-min_n); gf_out(max_n); return;
  end;
gf_out(boc); gf_four(boc_c); gf_four(boc_p);@/
gf_four(min_m); gf_four(max_m); gf_four(min_n); gf_four(max_n);
exit: end;

@ Two of the parameters to |gf_boc| are global.

@<Glob...@>=
@!boc_c,@!boc_p:integer; {parameters of the next |boc| command}

@ Here is a routine that gets a \.{GF} file off to a good start.

@d check_gf==@t@>@+if output_file_name=0 then init_gf

@<Declare generic font output procedures@>=
procedure init_gf;
var @!k:eight_bits; {runs through all possible character codes}
@!t:integer; {the time of this run}
begin gf_min_m:=4096; gf_max_m:=-4096; gf_min_n:=4096; gf_max_n:=-4096;
for k:=0 to 255 do char_ptr[k]:=-1;
@<Determine the file extension, |gf_ext|@>;
set_output_file_name;
gf_out(pre); gf_out(gf_id_byte); {begin to output the preamble}
old_setting:=selector; selector:=new_string; print(" METAFONT output ");
print_int(round_unscaled(internal[year])); print_char(".");
print_dd(round_unscaled(internal[month])); print_char(".");
print_dd(round_unscaled(internal[day])); print_char(":");@/
t:=round_unscaled(internal[time]);
print_dd(t div 60); print_dd(t mod 60);@/
selector:=old_setting; gf_out(cur_length);
gf_string(0,make_string); decr(str_ptr);
pool_ptr:=str_start[str_ptr]; {flush that string from memory}
gf_prev_ptr:=gf_offset+gf_ptr;
end;

@ @<Determine the file extension...@>=
if internal[hppp]<=0 then gf_ext:=".gf"
else  begin old_setting:=selector; selector:=new_string; print_char(".");
  print_int(make_scaled(internal[hppp],59429463));
    {$2^{32}/72.27\approx59429463.07$}
  print("gf"); gf_ext:=make_string; selector:=old_setting;
  end

@ With those preliminaries out of the way, |ship_out| is not especially
difficult.

@<Declare generic font output procedures@>=
procedure ship_out(@!c:eight_bits);
label done;
var @!f:integer; {current character extension}
@!prev_m,@!m,@!mm:integer; {previous and current pixel column numbers}
@!prev_n,@!n:integer; {previous and current pixel row numbers}
@!p,@!q:pointer; {for list traversal}
@!prev_w,@!w,@!ww:integer; {old and new weights}
@!d:integer; {data from edge-weight node}
@!delta:integer; {number of rows to skip}
@!cur_min_m:integer; {starting column, relative to the current offset}
@!x_off,@!y_off:integer; {offsets, rounded to integers}
begin check_gf; f:=round_unscaled(internal[char_ext]);@/
x_off:=round_unscaled(internal[x_offset]);
y_off:=round_unscaled(internal[y_offset]);
if term_offset>max_print_line-9 then print_ln
else if (term_offset>0)or(file_offset>0) then print_char(" ");
print_char("["); print_int(c);
if f<>0 then
  begin print_char("."); print_int(f);
  end;
update_terminal;
boc_c:=256*f+c; boc_p:=char_ptr[c]; char_ptr[c]:=gf_prev_ptr;@/
if internal[proofing]>0 then @<Send nonzero offsets to the output file@>;
@<Output the character represented in |cur_edges|@>;
gf_out(eoc); gf_prev_ptr:=gf_offset+gf_ptr; incr(total_chars);
print_char("]"); update_terminal; {progress report}
if internal[tracing_output]>0 then
  print_edges(" (just shipped out)",true,x_off,y_off);
end;

@ @<Send nonzero offsets to the output file@>=
begin if x_off<>0 then
  begin gf_string("xoffset",0); gf_out(yyy); gf_four(x_off*unity);
  end;
if y_off<>0 then
  begin gf_string("yoffset",0); gf_out(yyy); gf_four(y_off*unity);
  end;
end

@ @<Output the character represented in |cur_edges|@>=
prev_n:=4096; p:=knil(cur_edges); n:=n_max(cur_edges)-zero_field;
while p<>cur_edges do
  begin @<Output the pixels of edge row |p| to font row |n|@>;
  p:=knil(p); decr(n);
  end;
if prev_n=4096 then @<Finish off an entirely blank character@>
else if prev_n+y_off<gf_min_n then
  gf_min_n:=prev_n+y_off

@ @<Finish off an entirely blank...@>=
begin gf_boc(0,0,0,0);
if gf_max_m<0 then gf_max_m:=0;
if gf_min_n>0 then gf_min_n:=0;
end

@ In this loop, |prev_w| represents the weight at column |prev_m|, which is
the most recent column reflected in the output so far; |w| represents the
weight at column~|m|, which is the most recent column in the edge data.
Several edges might cancel at the same column position, so we need to
look ahead to column~|mm| before actually outputting anything.

@<Output the pixels of edge row |p| to font row |n|@>=
if unsorted(p)>void then sort_edges(p);
q:=sorted(p); w:=0; prev_m:=-fraction_one; {$|fraction_one|\approx\infty$}
ww:=0; prev_w:=0; m:=prev_m;
repeat if q=sentinel then mm:=fraction_one
else  begin d:=ho(info(q)); mm:=d div 8; ww:=ww+(d mod 8)-zero_w;
  end;
if mm<>m then
  begin if prev_w<=0 then
    begin if w>0 then @<Start black at $(m,n)$@>;
    end
  else if w<=0 then @<Stop black at $(m,n)$@>;
  m:=mm;
  end;
w:=ww; q:=link(q);
until mm=fraction_one;
if w<>0 then {this should be impossible}
  print_nl("(There's unbounded black in character shipped out!)");
@.There's unbounded black...@>
if prev_m-m_offset(cur_edges)+x_off>gf_max_m then
  gf_max_m:=prev_m-m_offset(cur_edges)+x_off


@ @<Start black at $(m,n)$@>=
begin if prev_m=-fraction_one then @<Start a new row at $(m,n)$@>
else gf_paint(m-prev_m);
prev_m:=m; prev_w:=w;
end

@ @<Stop black at $(m,n)$@>=
begin gf_paint(m-prev_m); prev_m:=m; prev_w:=w;
end

@ @<Start a new row at $(m,n)$@>=
begin if prev_n=4096 then
  begin gf_boc(m_min(cur_edges)+x_off-zero_field,
    m_max(cur_edges)+x_off-zero_field,@|
    n_min(cur_edges)+y_off-zero_field,n+y_off);
  cur_min_m:=m_min(cur_edges)-zero_field+m_offset(cur_edges);
  end
else if prev_n>n+1 then @<Skip down |prev_n-n| rows@>
else @<Skip to column $m$ in the next row and |goto done|, or skip zero rows@>;
gf_paint(m-cur_min_m); {skip to column $m$, painting white}
done:prev_n:=n;
end

@ @<Skip to column $m$ in the next row...@>=
begin delta:=m-cur_min_m;
if delta>max_new_row then gf_out(skip0)
else  begin gf_out(new_row_0+delta); goto done;
  end;
end

@ @<Skip down...@>=
begin delta:=prev_n-n-1;
if delta<@'400 then
  begin gf_out(skip1); gf_out(delta);
  end
else  begin gf_out(skip1+1); gf_two(delta);
  end;
end

@ Now that we've finished |ship_out|, let's look at the other commands
by which a user can send things to the \.{GF} file.

@<Cases of |do_statement|...@>=
special_command: do_special;

@ @<Put each...@>=
primitive("special",special_command,string_type);@/
@!@:special_}{\&{special} primitive@>
primitive("numspecial",special_command,known);@/
@!@:num_special_}{\&{numspecial} primitive@>

@ @<Declare action procedures for use by |do_statement|@>=
procedure do_special;
var @!m:small_number; {either |string_type| or |known|}
begin m:=cur_mod; get_x_next; scan_expression;
if internal[proofing]>=0 then
  if cur_type<>m then @<Complain about improper special operation@>
  else  begin check_gf;
    if m=string_type then gf_string(cur_exp,0)
    else  begin gf_out(yyy); gf_four(cur_exp);
      end;
    end;
flush_cur_exp(0);
end;

@ @<Complain about improper special operation@>=
begin exp_err("Unsuitable expression");
@.Unsuitable expression@>
help1("The expression shown above has the wrong type to be output.");
put_get_error;
end

@ @<Send the current expression as a title to the output file@>=
begin check_gf; gf_string("title ",cur_exp);
@.title@>
end

@ @<Cases of |print_cmd...@>=
special_command:if m=known then print("numspecial")
  else print("special");

@ @<Determine if a character has been shipped out@>=
begin cur_exp:=round_unscaled(cur_exp) mod 256;
if cur_exp<0 then cur_exp:=cur_exp+256;
boolean_reset(char_exists[cur_exp]); cur_type:=boolean_type;
end

@ At the end of the program we must finish things off by writing the postamble.
The \.{TFM} information should have been computed first.

An integer variable |k| and a |scaled| variable |x| will be declared for
use by this routine.

@<Finish the \.{GF} file@>=
begin gf_out(post); {beginning of the postamble}
gf_four(gf_prev_ptr); gf_prev_ptr:=gf_offset+gf_ptr-5; {|post| location}
gf_four(internal[design_size]*16);
for k:=1 to 4 do gf_out(header_byte[k]); {the check sum}
gf_four(internal[hppp]);
gf_four(internal[vppp]);@/
gf_four(gf_min_m); gf_four(gf_max_m);
gf_four(gf_min_n); gf_four(gf_max_n);
for k:=0 to 255 do if char_exists[k] then
  begin x:=gf_dx[k] div unity;
  if (gf_dy[k]=0)and(x>=0)and(x<256)and(gf_dx[k]=x*unity) then
    begin gf_out(char_loc+1); gf_out(k); gf_out(x);
    end
  else  begin gf_out(char_loc); gf_out(k);
    gf_four(gf_dx[k]); gf_four(gf_dy[k]);
    end;
  x:=value(tfm_width[k]);
  if abs(x)>max_tfm_dimen then
    if x>0 then x:=three_bytes-1@+else x:=1-three_bytes
  else x:=make_scaled(x*16,internal[design_size]);
  gf_four(x); gf_four(char_ptr[k]);
  end;
gf_out(post_post); gf_four(gf_prev_ptr); gf_out(gf_id_byte);@/
k:=4+((gf_buf_size-gf_ptr) mod 4); {the number of 223's}
while k>0 do
  begin gf_out(223); decr(k);
  end;
@<Empty the last bytes out of |gf_buf|@>;
print_nl("Output written on "); slow_print(output_file_name);
@.Output written...@>
print(" ("); print_int(total_chars); print(" character");
if total_chars<>1 then print_char("s");
print(", "); print_int(gf_offset+gf_ptr); print(" bytes).");
b_close(gf_file);
end

@* \[48] Dumping and undumping the tables.
After \.{INIMF} has seen a collection of macros, it
can write all the necessary information on an auxiliary file so
that production versions of \MF\ are able to initialize their
memory at high speed. The present section of the program takes
care of such output and input. We shall consider simultaneously
the processes of storing and restoring,
so that the inverse relation between them is clear.
@.INIMF@>

The global variable |base_ident| is a string that is printed right
after the |banner| line when \MF\ is ready to start. For \.{INIMF} this
string says simply `\.{(INIMF)}'; for other versions of \MF\ it says,
for example, `\.{(preloaded base=plain 1984.2.29)}', showing the year,
month, and day that the base file was created. We have |base_ident=0|
before \MF's tables are loaded.

@<Glob...@>=
@!base_ident:str_number;

@ @<Set init...@>=
base_ident:=0;

@ @<Initialize table entries...@>=
base_ident:=" (INIMF)";

@ @<Declare act...@>=
@!init procedure store_base_file;
var @!k:integer; {all-purpose index}
@!p,@!q: pointer; {all-purpose pointers}
@!x: integer; {something to dump}
@!w: four_quarters; {four ASCII codes}
begin @<Create the |base_ident|, open the base file,
  and inform the user that dumping has begun@>;
@<Dump constants for consistency check@>;
@<Dump the string pool@>;
@<Dump the dynamic memory@>;
@<Dump the table of equivalents and the hash table@>;
@<Dump a few more things and the closing check word@>;
@<Close the base file@>;
end;
tini

@ Corresponding to the procedure that dumps a base file, we also have a function
that reads~one~in. The function returns |false| if the dumped base is
incompatible with the present \MF\ table sizes, etc.

@d off_base=6666 {go here if the base file is unacceptable}
@d too_small(#)==begin wake_up_terminal;
  wterm_ln('---! Must increase the ',#);
@.Must increase the x@>
  goto off_base;
  end

@p @t\4@>@<Declare the function called |open_base_file|@>@;
function load_base_file:boolean;
label off_base,exit;
var @!k:integer; {all-purpose index}
@!p,@!q: pointer; {all-purpose pointers}
@!x: integer; {something undumped}
@!w: four_quarters; {four ASCII codes}
begin @<Undump constants for consistency check@>;
@<Undump the string pool@>;
@<Undump the dynamic memory@>;
@<Undump the table of equivalents and the hash table@>;
@<Undump a few more things and the closing check word@>;
load_base_file:=true; return; {it worked!}
off_base: wake_up_terminal;
  wterm_ln('(Fatal base file error; I''m stymied)');
@.Fatal base file error@>
load_base_file:=false;
exit:end;

@ Base files consist of |memory_word| items, and we use the following
macros to dump words of different types:

@d dump_wd(#)==begin base_file^:=#; put(base_file);@+end
@d dump_int(#)==begin base_file^.int:=#; put(base_file);@+end
@d dump_hh(#)==begin base_file^.hh:=#; put(base_file);@+end
@d dump_qqqq(#)==begin base_file^.qqqq:=#; put(base_file);@+end

@<Glob...@>=
@!base_file:word_file; {for input or output of base information}

@ The inverse macros are slightly more complicated, since we need to check
the range of the values we are reading in. We say `|undump(a)(b)(x)|' to
read an integer value |x| that is supposed to be in the range |a<=x<=b|.
System error messages should be suppressed when undumping.
@^system dependencies@>

@d undump_wd(#)==begin get(base_file); #:=base_file^;@+end
@d undump_int(#)==begin get(base_file); #:=base_file^.int;@+end
@d undump_hh(#)==begin get(base_file); #:=base_file^.hh;@+end
@d undump_qqqq(#)==begin get(base_file); #:=base_file^.qqqq;@+end
@d undump_end_end(#)==#:=x;@+end
@d undump_end(#)==(x>#) then goto off_base@+else undump_end_end
@d undump(#)==begin undump_int(x); if (x<#) or undump_end
@d undump_size_end_end(#)==too_small(#)@+else undump_end_end
@d undump_size_end(#)==if x># then undump_size_end_end
@d undump_size(#)==begin undump_int(x);
  if x<# then goto off_base; undump_size_end

@ The next few sections of the program should make it clear how we use the
dump/undump macros.

@<Dump constants for consistency check@>=
dump_int(@$);@/
dump_int(mem_min);@/
dump_int(mem_top);@/
dump_int(hash_size);@/
dump_int(hash_prime);@/
dump_int(max_in_open)

@ Sections of a \.{WEB} program that are ``commented out'' still contribute
strings to the string pool; therefore \.{INIMF} and \MF\ will have
the same strings. (And it is, of course, a good thing that they do.)
@.WEB@>
@^string pool@>

@<Undump constants for consistency check@>=
x:=base_file^.int;
if x<>@$ then goto off_base; {check that strings are the same}
undump_int(x);
if x<>mem_min then goto off_base;
undump_int(x);
if x<>mem_top then goto off_base;
undump_int(x);
if x<>hash_size then goto off_base;
undump_int(x);
if x<>hash_prime then goto off_base;
undump_int(x);
if x<>max_in_open then goto off_base

@ @d dump_four_ASCII==
  w.b0:=qi(so(str_pool[k])); w.b1:=qi(so(str_pool[k+1]));
  w.b2:=qi(so(str_pool[k+2])); w.b3:=qi(so(str_pool[k+3]));
  dump_qqqq(w)

@<Dump the string pool@>=
dump_int(pool_ptr);
dump_int(str_ptr);
for k:=0 to str_ptr do dump_int(str_start[k]);
k:=0;
while k+4<pool_ptr do
  begin dump_four_ASCII; k:=k+4;
  end;
k:=pool_ptr-4; dump_four_ASCII;
print_ln; print_int(str_ptr); print(" strings of total length ");
print_int(pool_ptr)

@ @d undump_four_ASCII==
  undump_qqqq(w);
  str_pool[k]:=si(qo(w.b0)); str_pool[k+1]:=si(qo(w.b1));
  str_pool[k+2]:=si(qo(w.b2)); str_pool[k+3]:=si(qo(w.b3))

@<Undump the string pool@>=
undump_size(0)(pool_size)('string pool size')(pool_ptr);
undump_size(0)(max_strings)('max strings')(str_ptr);
for k:=0 to str_ptr do
  begin undump(0)(pool_ptr)(str_start[k]); str_ref[k]:=max_str_ref;
  end;
k:=0;
while k+4<pool_ptr do
  begin undump_four_ASCII; k:=k+4;
  end;
k:=pool_ptr-4; undump_four_ASCII;
init_str_ptr:=str_ptr; init_pool_ptr:=pool_ptr;
max_str_ptr:=str_ptr; max_pool_ptr:=pool_ptr

@ By sorting the list of available spaces in the variable-size portion of
|mem|, we are usually able to get by without having to dump very much
of the dynamic memory.

We recompute |var_used| and |dyn_used|, so that \.{INIMF} dumps valid
information even when it has not been gathering statistics.

@<Dump the dynamic memory@>=
sort_avail; var_used:=0;
dump_int(lo_mem_max); dump_int(rover);
p:=mem_min; q:=rover; x:=0;
repeat for k:=p to q+1 do dump_wd(mem[k]);
x:=x+q+2-p; var_used:=var_used+q-p;
p:=q+node_size(q); q:=rlink(q);
until q=rover;
var_used:=var_used+lo_mem_max-p; dyn_used:=mem_end+1-hi_mem_min;@/
for k:=p to lo_mem_max do dump_wd(mem[k]);
x:=x+lo_mem_max+1-p;
dump_int(hi_mem_min); dump_int(avail);
for k:=hi_mem_min to mem_end do dump_wd(mem[k]);
x:=x+mem_end+1-hi_mem_min;
p:=avail;
while p<>null do
  begin decr(dyn_used); p:=link(p);
  end;
dump_int(var_used); dump_int(dyn_used);
print_ln; print_int(x);
print(" memory locations dumped; current usage is ");
print_int(var_used); print_char("&"); print_int(dyn_used)

@ @<Undump the dynamic memory@>=
undump(lo_mem_stat_max+1000)(hi_mem_stat_min-1)(lo_mem_max);
undump(lo_mem_stat_max+1)(lo_mem_max)(rover);
p:=mem_min; q:=rover;
repeat for k:=p to q+1 do undump_wd(mem[k]);
p:=q+node_size(q);
if (p>lo_mem_max)or((q>=rlink(q))and(rlink(q)<>rover)) then goto off_base;
q:=rlink(q);
until q=rover;
for k:=p to lo_mem_max do undump_wd(mem[k]);
undump(lo_mem_max+1)(hi_mem_stat_min)(hi_mem_min);
undump(null)(mem_top)(avail); mem_end:=mem_top;
for k:=hi_mem_min to mem_end do undump_wd(mem[k]);
undump_int(var_used); undump_int(dyn_used)

@ A different scheme is used to compress the hash table, since its lower region
is usually sparse. When |text(p)<>0| for |p<=hash_used|, we output three
words: |p|, |hash[p]|, and |eqtb[p]|. The hash table is, of course, densely
packed for |p>=hash_used|, so the remaining entries are output in~a~block.

@<Dump the table of equivalents and the hash table@>=
dump_int(hash_used); st_count:=frozen_inaccessible-1-hash_used;
for p:=1 to hash_used do if text(p)<>0 then
  begin dump_int(p); dump_hh(hash[p]); dump_hh(eqtb[p]); incr(st_count);
  end;
for p:=hash_used+1 to hash_end do
  begin dump_hh(hash[p]); dump_hh(eqtb[p]);
  end;
dump_int(st_count);@/
print_ln; print_int(st_count); print(" symbolic tokens")

@ @<Undump the table of equivalents and the hash table@>=
undump(1)(frozen_inaccessible)(hash_used); p:=0;
repeat undump(p+1)(hash_used)(p); undump_hh(hash[p]); undump_hh(eqtb[p]);
until p=hash_used;
for p:=hash_used+1 to hash_end do
  begin undump_hh(hash[p]); undump_hh(eqtb[p]);
  end;
undump_int(st_count)

@ We have already printed a lot of statistics, so we set |tracing_stats:=0|
to prevent them from appearing again.

@<Dump a few more things and the closing check word@>=
dump_int(int_ptr);
for k:=1 to int_ptr do
  begin dump_int(internal[k]); dump_int(int_name[k]);
  end;
dump_int(start_sym); dump_int(interaction); dump_int(base_ident);
dump_int(bg_loc); dump_int(eg_loc); dump_int(serial_no); dump_int(69069);
internal[tracing_stats]:=0

@ @<Undump a few more things and the closing check word@>=
undump(max_given_internal)(max_internal)(int_ptr);
for k:=1 to int_ptr do
  begin undump_int(internal[k]);
  undump(0)(str_ptr)(int_name[k]);
  end;
undump(0)(frozen_inaccessible)(start_sym);
undump(batch_mode)(error_stop_mode)(interaction);
undump(0)(str_ptr)(base_ident);
undump(1)(hash_end)(bg_loc);
undump(1)(hash_end)(eg_loc);
undump_int(serial_no);@/
undump_int(x);@+if (x<>69069)or eof(base_file) then goto off_base

@ @<Create the |base_ident|...@>=
selector:=new_string;
print(" (preloaded base="); print(job_name); print_char(" ");
print_int(round_unscaled(internal[year])); print_char(".");
print_int(round_unscaled(internal[month])); print_char(".");
print_int(round_unscaled(internal[day])); print_char(")");
if interaction=batch_mode then selector:=log_only
else selector:=term_and_log;
str_room(1); base_ident:=make_string; str_ref[base_ident]:=max_str_ref;@/
pack_job_name(base_extension);
while not w_open_out(base_file) do
 prompt_file_name("base file name",base_extension);
print_nl("Beginning to dump on file ");
@.Beginning to dump...@>
slow_print(w_make_name_string(base_file)); flush_string(str_ptr-1);
print_nl(""); slow_print(base_ident)

@ @<Close the base file@>=
w_close(base_file)

@* \[49] The main program.
This is it: the part of \MF\ that executes all those procedures we have
written.

Well---almost. We haven't put the parsing subroutines into the
program yet; and we'd better leave space for a few more routines that may
have been forgotten.

@p @<Declare the basic parsing subroutines@>@;
@<Declare miscellaneous procedures that were declared |forward|@>@;
@<Last-minute procedures@>

@ We've noted that there are two versions of \MF84. One, called \.{INIMF},
@.INIMF@>
has to be run first; it initializes everything from scratch, without
reading a base file, and it has the capability of dumping a base file.
The other one is called `\.{VIRMF}'; it is a ``virgin'' program that needs
@.VIRMF@>
to input a base file in order to get started. \.{VIRMF} typically has
a bit more memory capacity than \.{INIMF}, because it does not need the
space consumed by the dumping/undumping routines and the numerous calls on
|primitive|, etc.

The \.{VIRMF} program cannot read a base file instantaneously, of course;
the best implementations therefore allow for production versions of \MF\ that
not only avoid the loading routine for \PASCAL\ object code, they also have
a base file pre-loaded. This is impossible to do if we stick to standard
\PASCAL; but there is a simple way to fool many systems into avoiding the
initialization, as follows:\quad(1)~We declare a global integer variable
called |ready_already|. The probability is negligible that this
variable holds any particular value like 314159 when \.{VIRMF} is first
loaded.\quad(2)~After we have read in a base file and initialized
everything, we set |ready_already:=314159|.\quad(3)~Soon \.{VIRMF}
will print `\.*', waiting for more input; and at this point we
interrupt the program and save its core image in some form that the
operating system can reload speedily.\quad(4)~When that core image is
activated, the program starts again at the beginning; but now
|ready_already=314159| and all the other global variables have
their initial values too. The former chastity has vanished!

In other words, if we allow ourselves to test the condition
|ready_already=314159|, before |ready_already| has been
assigned a value, we can avoid the lengthy initialization. Dirty tricks
rarely pay off so handsomely.
@^dirty \PASCAL@>
@^system dependencies@>

On systems that allow such preloading, the standard program called \.{MF}
should be the one that has \.{plain} base preloaded, since that agrees
with {\sl The {\logos METAFONT\/}book}.  Other versions, e.g., \.{CMMF},
should also be provided for commonly used bases such as \.{cmbase}.
@:METAFONTbook}{\sl The {\logos METAFONT\/}book@>
@.cmbase@>
@.plain@>

@<Glob...@>=
@!ready_already:integer; {a sacrifice of purity for economy}

@ Now this is really it: \MF\ starts and ends here.

The initial test involving |ready_already| should be deleted if the
\PASCAL\ runtime system is smart enough to detect such a ``mistake.''
@^system dependencies@>

@p begin @!{|start_here|}
history:=fatal_error_stop; {in case we quit during initialization}
t_open_out; {open the terminal for output}
if ready_already=314159 then goto start_of_MF;
@<Check the ``constant'' values...@>@;
if bad>0 then
  begin wterm_ln('Ouch---my internal constants have been clobbered!',
    '---case ',bad:1);
@.Ouch...clobbered@>
  goto final_end;
  end;
initialize; {set global variables to their starting values}
@!init if not get_strings_started then goto final_end;
init_tab; {initialize the tables}
init_prim; {call |primitive| for each primitive}
init_str_ptr:=str_ptr; init_pool_ptr:=pool_ptr;@/
max_str_ptr:=str_ptr; max_pool_ptr:=pool_ptr; fix_date_and_time;
tini@/
ready_already:=314159;
start_of_MF: @<Initialize the output routines@>;
@<Get the first line of input and prepare to start@>;
history:=spotless; {ready to go!}
if start_sym>0 then {insert the `\&{everyjob}' symbol}
  begin cur_sym:=start_sym; back_input;
  end;
main_control; {come to life}
final_cleanup; {prepare for death}
end_of_MF: close_files_and_terminate;
final_end: ready_already:=0;
end.

@ Here we do whatever is needed to complete \MF's job gracefully on the
local operating system. The code here might come into play after a fatal
error; it must therefore consist entirely of ``safe'' operations that
cannot produce error messages. For example, it would be a mistake to call
|str_room| or |make_string| at this time, because a call on |overflow|
might lead to an infinite loop.
@^system dependencies@>

If |final_cleanup| is bypassed, this program doesn't bother to close
the input files that may still be open.

@<Last-minute...@>=
procedure close_files_and_terminate;
var @!k:integer; {all-purpose index}
@!lh:integer; {the length of the \.{TFM} header, in words}
@!lk_offset:0..256; {extra words inserted at beginning of |lig_kern| array}
@!p:pointer; {runs through a list of \.{TFM} dimensions}
@!x:scaled; {a |tfm_width| value being output to the \.{GF} file}
begin
@!stat if internal[tracing_stats]>0 then
  @<Output statistics about this job@>;@;@+tats@/
wake_up_terminal; @<Finish the \.{TFM} and \.{GF} files@>;
if log_opened then
  begin wlog_cr;
  a_close(log_file); selector:=selector-2;
  if selector=term_only then
    begin print_nl("Transcript written on ");
@.Transcript written...@>
    slow_print(log_name); print_char(".");
    end;
  end;
end;

@ We want to finish the \.{GF} file if and only if it has already been started;
this will be true if and only if |gf_prev_ptr| is positive.
We want to produce a \.{TFM} file if and only if |fontmaking| is positive.
The \.{TFM} widths must be computed if there's a \.{GF} file, even if
there's going to be no \.{TFM}~file.

We reclaim all of the variable-size memory at this point, so that
there is no chance of another memory overflow after the memory capacity
has already been exceeded.

@<Finish the \.{TFM} and \.{GF} files@>=
if (gf_prev_ptr>0)or(internal[fontmaking]>0) then
  begin @<Make the dynamic memory into one big available node@>;
  @<Massage the \.{TFM} widths@>;
  fix_design_size; fix_check_sum;
  if internal[fontmaking]>0 then
    begin @<Massage the \.{TFM} heights, depths, and italic corrections@>;
    internal[fontmaking]:=0; {avoid loop in case of fatal error}
    @<Finish the \.{TFM} file@>;
    end;
  if gf_prev_ptr>0 then @<Finish the \.{GF} file@>;
  end

@ @<Make the dynamic memory into one big available node@>=
rover:=lo_mem_stat_max+1; link(rover):=empty_flag; lo_mem_max:=hi_mem_min-1;
if lo_mem_max-rover>max_halfword then lo_mem_max:=max_halfword+rover;
node_size(rover):=lo_mem_max-rover; llink(rover):=rover; rlink(rover):=rover;
link(lo_mem_max):=null; info(lo_mem_max):=null

@ The present section goes directly to the log file instead of using
|print| commands, because there's no need for these strings to take
up |str_pool| memory when a non-{\bf stat} version of \MF\ is being used.

@<Output statistics...@>=
if log_opened then
  begin wlog_ln(' ');
  wlog_ln('Here is how much of METAFONT''s memory',' you used:');
@.Here is how much...@>
  wlog(' ',max_str_ptr-init_str_ptr:1,' string');
  if max_str_ptr<>init_str_ptr+1 then wlog('s');
  wlog_ln(' out of ', max_strings-init_str_ptr:1);@/
  wlog_ln(' ',max_pool_ptr-init_pool_ptr:1,' string characters out of ',
    pool_size-init_pool_ptr:1);@/
  wlog_ln(' ',lo_mem_max-mem_min+mem_end-hi_mem_min+2:1,@|
    ' words of memory out of ',mem_end+1-mem_min:1);@/
  wlog_ln(' ',st_count:1,' symbolic tokens out of ',
    hash_size:1);@/
  wlog_ln(' ',max_in_stack:1,'i,',@|
    int_ptr:1,'n,',@|
    max_rounding_ptr:1,'r,',@|
    max_param_stack:1,'p,',@|
    max_buf_stack+1:1,'b stack positions out of ',@|
    stack_size:1,'i,',
    max_internal:1,'n,',
    max_wiggle:1,'r,',
    param_size:1,'p,',
    buf_size:1,'b');
  end

@ We get to the |final_cleanup| routine when \&{end} or \&{dump} has
been scanned.

@<Last-minute...@>=
procedure final_cleanup;
label exit;
var c:small_number; {0 for \&{end}, 1 for \&{dump}}
begin c:=cur_mod;
if job_name=0 then open_log_file;
while input_ptr>0 do
  if token_state then end_token_list@+else end_file_reading;
while loop_ptr<>null do stop_iteration;
while open_parens>0 do
  begin print(" )"); decr(open_parens);
  end;
while cond_ptr<>null do
  begin print_nl("(end occurred when ");@/
@.end occurred...@>
  print_cmd_mod(fi_or_else,cur_if);
    {`\.{if}' or `\.{elseif}' or `\.{else}'}
  if if_line<>0 then
    begin print(" on line "); print_int(if_line);
    end;
  print(" was incomplete)");
  if_line:=if_line_field(cond_ptr);
  cur_if:=name_type(cond_ptr); loop_ptr:=cond_ptr;
  cond_ptr:=link(cond_ptr); free_node(loop_ptr,if_node_size);
  end;
if history<>spotless then
 if ((history=warning_issued)or(interaction<error_stop_mode)) then
  if selector=term_and_log then
  begin selector:=term_only;
  print_nl("(see the transcript file for additional information)");
@.see the transcript file...@>
  selector:=term_and_log;
  end;
if c=1 then
  begin @!init store_base_file; return;@+tini@/
  print_nl("(dump is performed only by INIMF)"); return;
@.dump...only by INIMF@>
  end;
exit:end;

@ @<Last-minute...@>=
@!init procedure init_prim; {initialize all the primitives}
begin
@<Put each...@>;
end;
@#
procedure init_tab; {initialize other tables}
var @!k:integer; {all-purpose index}
begin @<Initialize table entries (done by \.{INIMF} only)@>@;
end;
tini

@ When we begin the following code, \MF's tables may still contain garbage;
the strings might not even be present. Thus we must proceed cautiously to get
bootstrapped in.

But when we finish this part of the program, \MF\ is ready to call on the
|main_control| routine to do its work.

@<Get the first line...@>=
begin @<Initialize the input routines@>;
if (base_ident=0)or(buffer[loc]="&") then
  begin if base_ident<>0 then initialize; {erase preloaded base}
  if not open_base_file then goto final_end;
  if not load_base_file then
    begin w_close(base_file); goto final_end;
    end;
  w_close(base_file);
  while (loc<limit)and(buffer[loc]=" ") do incr(loc);
  end;
buffer[limit]:="%";@/
fix_date_and_time; init_randoms(sys_time+sys_day*unity);@/
@<Initialize the print |selector|...@>;
if loc<limit then if buffer[loc]<>"\" then start_input; {\&{input} assumed}
end

@* \[50] Debugging.
Once \MF\ is working, you should be able to diagnose most errors with
the \.{show} commands and other diagnostic features. But for the initial
stages of debugging, and for the revelation of really deep mysteries, you
can compile \MF\ with a few more aids, including the \PASCAL\ runtime
checks and its debugger. An additional routine called |debug_help|
will also come into play when you type `\.D' after an error message;
|debug_help| also occurs just before a fatal error causes \MF\ to succumb.
@^debugging@>
@^system dependencies@>

The interface to |debug_help| is primitive, but it is good enough when used
with a \PASCAL\ debugger that allows you to set breakpoints and to read
variables and change their values. After getting the prompt `\.{debug \#}', you
type either a negative number (this exits |debug_help|), or zero (this
goes to a location where you can set a breakpoint, thereby entering into
dialog with the \PASCAL\ debugger), or a positive number |m| followed by
an argument |n|. The meaning of |m| and |n| will be clear from the
program below. (If |m=13|, there is an additional argument, |l|.)
@.debug \#@>

@d breakpoint=888 {place where a breakpoint is desirable}

@<Last-minute...@>=
@!debug procedure debug_help; {routine to display various things}
label breakpoint,exit;
var @!k,@!l,@!m,@!n:integer;
begin clear_terminal;
  loop begin wake_up_terminal;
  print_nl("debug # (-1 to exit):"); update_terminal;
@.debug \#@>
  read(term_in,m);
  if m<0 then return
  else if m=0 then
    begin goto breakpoint;@/ {go to every declared label at least once}
    breakpoint: m:=0; @{'BREAKPOINT'@}@/
    end
  else  begin read(term_in,n);
    case m of
    @t\4@>@<Numbered cases for |debug_help|@>@;
    othercases print("?")
    endcases;
    end;
  end;
exit:end;
gubed

@ @<Numbered cases...@>=
1: print_word(mem[n]); {display |mem[n]| in all forms}
2: print_int(info(n));
3: print_int(link(n));
4: begin print_int(eq_type(n)); print_char(":"); print_int(equiv(n));
  end;
5: print_variable_name(n);
6: print_int(internal[n]);
7: do_show_dependencies;
9: show_token_list(n,null,100000,0);
10: slow_print(n);
11: check_mem(n>0); {check wellformedness; print new busy locations if |n>0|}
12: search_mem(n); {look for pointers to |n|}
13: begin read(term_in,l); print_cmd_mod(n,l);
  end;
14: for k:=0 to n do print(buffer[k]);
15: panicking:=not panicking;

@* \[51] System-dependent changes.
This section should be replaced, if necessary, by any special
modifications of the program
that are necessary to make \MF\ work at a particular installation.
It is usually best to design your change file so that all changes to
previous sections preserve the section numbering; then everybody's version
will be consistent with the published program. More extensive changes,
which introduce new sections, can be inserted here; then only the index
itself will get a new section number.
@^system dependencies@>

@* \[52] Index.
Here is where you can find all uses of each identifier in the program,
with underlined entries pointing to where the identifier was defined.
If the identifier is only one letter long, however, you get to see only
the underlined entries. {\sl All references are to section numbers instead of
page numbers.}

This index also lists error messages and other aspects of the program
that you might want to look up some day. For example, the entry
for ``system dependencies'' lists all sections that should receive
special attention from people who are installing \MF\ in a new
operating environment. A list of various things that can't happen appears
under ``this can't happen''.
Approximately 25 sections are listed under ``inner loop''; these account
for more than 60\pct! of \MF's running time, exclusive of input and output.