summaryrefslogtreecommitdiff
path: root/Master/texmf-dist/source/latex/l3kernel/l3fp-basics.dtx
blob: 0b06d377232b24f5f3ab0175960d70f94611aa7d (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
% \iffalse meta-comment
%
%% File: l3fp-basics.dtx Copyright (C) 2011-2012 The LaTeX3 Project
%%
%% It may be distributed and/or modified under the conditions of the
%% LaTeX Project Public License (LPPL), either version 1.3c of this
%% license or (at your option) any later version.  The latest version
%% of this license is in the file
%%
%%    http://www.latex-project.org/lppl.txt
%%
%% This file is part of the "l3kernel bundle" (The Work in LPPL)
%% and all files in that bundle must be distributed together.
%%
%% The released version of this bundle is available from CTAN.
%%
%% -----------------------------------------------------------------------
%%
%% The development version of the bundle can be found at
%%
%%    http://www.latex-project.org/svnroot/experimental/trunk/
%%
%% for those people who are interested.
%%
%%%%%%%%%%%
%% NOTE: %%
%%%%%%%%%%%
%%
%%   Snapshots taken from the repository represent work in progress and may
%%   not work or may contain conflicting material!  We therefore ask
%%   people _not_ to put them into distributions, archives, etc. without
%%   prior consultation with the LaTeX Project Team.
%%
%% -----------------------------------------------------------------------
%%
%
%<*driver>
\RequirePackage{l3names}
\GetIdInfo$Id: l3fp-basics.dtx 3986 2012-07-15 19:23:51Z joseph $
  {L3 Floating-point arithmetic}
\documentclass[full]{l3doc}
\begin{document}
  \DocInput{\jobname.dtx}
\end{document}
%</driver>
% \fi
%
% \title{The \textsf{l3fp-basics} package\thanks{This file
%         has version number \ExplFileVersion, last
%         revised \ExplFileDate.}\\
% Floating point arithmetic}
% \author{^^A
%  The \LaTeX3 Project\thanks
%    {^^A
%      E-mail:
%        \href{mailto:latex-team@latex-project.org}
%          {latex-team@latex-project.org}^^A
%    }^^A
% }
% \date{Released \ExplFileDate}
%
% \maketitle
%
% \begin{documentation}
%
% \end{documentation}
%
% \begin{implementation}
%
% \section{\pkg{l3fp-basics} Implementation}
%
%    \begin{macrocode}
%<*initex|package>
%    \end{macrocode}
%
%    \begin{macrocode}
%<@@=fp>
%    \end{macrocode}
%
% All operations implemented in this module yield the outcome of
% rounding the infinitely precise result of the operation to the
% nearest representable number.
%
% ^^A begin[todo]: move
% \section{Internal storage of floating points numbers}
%
% A floating point number \meta{X} is stored as
% \begin{quote}
%   \cs{s_@@} \cs{@@_chk:w} \meta{case} \meta{sign} \meta{body} |;|
% \end{quote}
% Here, \meta{case} is 0 for $\pm 0$, 1 for normal numbers, 2 for $\pm
% \infty$, and 3 for \texttt{nan}, and \meta{sign} is $0$ for positive
% numbers, $1$ for \texttt{nan}s, and $2$ for negative numbers. The
% \meta{body} of normal numbers is \Arg{exponent} \Arg{X_1} \Arg{X_2}
% \Arg{X_3} \Arg{X_4}, with
% \[
% \meta{X} = (-1)^{\meta{sign}} 10^{-\meta{exponent}} \sum_i
% \meta{X_i} 10^{-4i}.
% \]
% Calculations are done in base $10000$, \emph{i.e.} one myriad.  The
% \meta{exponent} lies between $\pm\cs{c_@@_max_exponent_int} = \pm
% \the\csname\detokenize{c__fp_max_exponent_int}\endcsname$ inclusive.
%
% Additionally, positive and negative floating point numbers may only be
% stored with $1000\leq\meta{X_1}<10000$. This requirement is necessary
% in order to preserve accuracy and speed.
%
% ^^A end[todo]
%
% ^^A begin[todo]
%
% Some algorithms used below end up being quite similar to some
% described in \enquote{What Every Computer Scientist Should Know About
%   Floating Point Arithmetic}, by David Goldberg, which can be found at
% \texttt{http://cr.yp.to/2005-590/goldberg.pdf}. I need to compare them
% very carefully.
%
% ^^A end[todo]
%
%^^A todo sanitize, pack.
%
% \subsection{Common to several operations}
%
% \begin{macro}[EXP]
%   {
%     \@@_basics_pack_low:NNNNNw  ,
%     \@@_basics_pack_high:NNNNNw ,
%     \@@_basics_pack_high_carry:w
%   }
%   Addition and multiplication of mantissas are done in two steps:
%   first compute a (more or less) exact result,
%   then round and pack digits in the final (braced) form.
%   These functions take care of the packing, with special attention
%   given to the case where rounding has caused a carry.
%   In \cs{@@_basics_pack_high_carry:w}, |#1| should
%   always be $0000$.
%    \begin{macrocode}
\cs_new:Npn \@@_basics_pack_low:NNNNNw #1 #2#3#4#5 #6;
  {
    \if_meaning:w 2 #1
      + \c_one
    \fi:
    ; {#2#3#4#5} {#6} ;
  }
\cs_new:Npn \@@_basics_pack_high:NNNNNw #1 #2#3#4#5 #6;
  {
    \if_meaning:w 2 #1
      \@@_basics_pack_high_carry:w
    \fi:
    ; {#2#3#4#5} {#6}
  }
\cs_new:Npn \@@_basics_pack_high_carry:w \fi: ; #1
  { \fi: + \c_one ; {1000} }
%    \end{macrocode}
% \end{macro}
%
% \begin{macro}
%   {
%     \@@_basics_return_nan_nan:NNww ,
%     \@@_basics_return_zero:NNww    ,
%     \@@_basics_return_inf:NNww     ,
%     \@@_basics_return_i:NNNNww     ,
%     \@@_basics_return_ii:NNNNww    ,
%     \@@_basics_return_nan:NNNNww
%   }
%   \begin{syntax}
%     \cs{@@_basics_return_...:NNww}
%     ~~\meta{sign_1} \meta{sign_2} \meta{body_1} |;| \meta{body_2} |;|
%   \end{syntax}
%   Used for binary operations, to return a value for some special
%   cases (common to several operations). All functions expand once
%   after their arguments.
%
%   ^^A todo: redoc, changed.
%   The \texttt{nan_nan} function combines the \texttt{info} fields
%   of the two \texttt{nan}.\footnote{Bruno: check that messages are
%     kept.}
%   The \texttt{zero} and \texttt{inf} functions return $\pm 0$ or
%   $\pm\infty$ with a sign equal to the product of the two signs:
%   three \cs{exp_after:wN} are needed to escape out of the
%   conditional, and expand once after.
%   The \texttt{i} and \texttt{ii} functions return one of their
%   operands and expand after using \cs{@@_exp_after_o:w}.
%   In some cases, this could be optimized, since we know in advance
%   what case of number we have. However, it seems better to keep the
%   number of control sequences low: these functions are called only
%   in special cases anyways, so performance is not an issue.
%    \begin{macrocode}
\cs_new:Npn \@@_basics_return_nan_nan:NNww #1#2 #3; #4;
  { \@@_exp_after_o:w \s_@@ \@@_chk:w 3 1 #3 ; }
\cs_new:Npn \@@_basics_return_zero:NNww #1#2 #3; #4;
  {
    \if_meaning:w #1 #2
      \exp_after:wN \exp_after:wN \exp_after:wN \c_zero_fp
    \else:
      \exp_after:wN \exp_after:wN \exp_after:wN \c_minus_zero_fp
    \fi:
  }
\cs_new:Npn \@@_basics_return_inf:NNww #1#2 #3; #4;
  {
    \if_meaning:w #1 #2
      \exp_after:wN \exp_after:wN \exp_after:wN \c_inf_fp
    \else:
      \exp_after:wN \exp_after:wN \exp_after:wN \c_minus_inf_fp
    \fi:
  }
\cs_new:Npn \@@_basics_return_i:NNNNww #1#2 #3#4 #5; #6;
  { \@@_exp_after_o:w \s_@@ \@@_chk:w #1 #3 #5; }
\cs_new:Npn \@@_basics_return_ii:NNNNww #1#2 #3#4 #5; #6;
  { \@@_exp_after_o:w \s_@@ \@@_chk:w #2 #4 #6; }
\cs_new:Npn \@@_basics_return_nan:NNww #1#2
  {
    \if_meaning:w 1 #1
      \exp_after:wN \@@_basics_return_i:NNNNww
    \else:
      \exp_after:wN \@@_basics_return_ii:NNNNww
    \fi:
    3 3 #1 #2
  }
%    \end{macrocode}
% \end{macro}
%
% \subsection{Addition and subtraction}
%
% \begin{macro}[int, EXP]{\@@_+_o:ww}
%   For addition, everything is easy.  No need to grab the
%   \meta{body_2}.
%    \begin{macrocode}
\cs_new:cpn { @@_+_o:ww }
    \s_@@ \@@_chk:w #1 #2 #3 ; \s_@@ \@@_chk:w #4 #5
  { \@@_add_cases:NN #1 #4 #2 #5 #3 ; }
%    \end{macrocode}
% \end{macro}
%
% \begin{macro}[int, EXP]{\@@_-_o:ww}
%   Change the sign of the second argument.
%    \begin{macrocode}
\cs_new:cpn { @@_-_o:ww }
    \s_@@ \@@_chk:w #1 #2 #3 ; \s_@@ \@@_chk:w #4 #5
  {
    \exp_after:wN \@@_add_cases:NN
    \exp_after:wN #1
    \exp_after:wN #4
    \exp_after:wN #2
    \int_use:N \__int_eval:w \c_two - #5 \__int_eval_end:
    #3 ;
  }
%    \end{macrocode}
% \end{macro}
%
% \subsubsection{Sign, exponent, and special numbers}
%
% \begin{macro}[EXP,aux]{\@@_add_cases:NN}
%   \begin{syntax}
%     \cs{@@_add_cases:NN} \meta{case_1} \meta{case_2}
%     ~~\meta{sign_1} \meta{sign_2} \meta{body_1} |;| \meta{body_2} |;|
%   \end{syntax}
%   This performs the addition. it also expands the following tokens
%   on the input stream once.
%
%   Whenever \meta{case_1} is different from \meta{case_2}, the result
%   is simply the floating point number with the highest \meta{case}.
%   For instance, adding a normal number to a zero gives the normal
%   number, and adding a \texttt{nan} to any non-\texttt{nan} gives
%   that \texttt{nan}. Optimizing for addition of normal numbers,
%   we test for equality and then separate the \enquote{greater than}
%   and \enquote{less than} branches.
%    \begin{macrocode}
\cs_new:Npn \@@_add_cases:NN #1 #2
  {
    \if_int_compare:w #1 = #2 \exp_stop_f:
      \exp_after:wN \@@_add_cases_eq:N
    \else:
      \if_int_compare:w #1 < #2 \exp_stop_f:
        \exp_after:wN \exp_after:wN
        \exp_after:wN \@@_basics_return_ii:NNNNww
      \else:
        \exp_after:wN \exp_after:wN
        \exp_after:wN \@@_basics_return_i:NNNNww
      \fi:
      \exp_after:wN #1
    \fi:
    #2
  }
%    \end{macrocode}
%   If the first \meta{case} is larger, then the first number remains
%   untouched, while the second number is ignored. On the other hand,
%   if the second \meta{case} is larger, the opposite happens: we retain
%   the second number. In both cases, there needs to be one step of
%   expansion after.
%    \begin{macrocode}
%    \end{macrocode}
%   We are then ready for the equality case: we split according
%   to the \meta{case}.
%    \begin{macrocode}
\cs_new:Npn \@@_add_cases_eq:N #1
  {
    \if_case:w #1 \exp_stop_f:
         \exp_after:wN \@@_add_zeros:NNww
    \or: \exp_after:wN \@@_add_normal:NNww
    \or: \exp_after:wN \@@_add_inf:NNww
    \or: \exp_after:wN \@@_basics_return_nan_nan:NNww
    \fi:
  }
%    \end{macrocode}
%   Adding two zeros yields \cs{c_zero_fp}, except if both
%   zeros were $-0$.\footnote{Bruno: this should depend on the
%     rounding mode.}
%    \begin{macrocode}
\cs_new:Npn \@@_add_zeros:NNww #1#2 #3;
  {
    \if_int_compare:w #1 #2 = 02 \exp_stop_f:
      \@@_case_return_o:Nw \c_zero_fp
    \else:
      \@@_case_return_same_o:w
    \fi:
    \s_@@ \@@_chk:w 0 #2
  }
%    \end{macrocode}
%   If both infinities have the same sign, just return that infinity,
%   otherwise, it is an invalid operation.
%    \begin{macrocode}
\cs_new:Npn \@@_add_inf:NNww #1#2 #3;
  {
    \if_meaning:w #1 #2
      \@@_case_return_same_o:w
    \else:
      \@@_case_use:nw
        {
          \@@_invalid_operation:Nnww \c_nan_fp { + }
            \s_@@ \@@_chk:w 2 #1 #3 ;
        }
    \fi:
    \s_@@ \@@_chk:w 2 #2
  }
%    \end{macrocode}
% \end{macro}
%
% \begin{macro}{\@@_add_normal:NNww}
%   \begin{syntax}
%     \cs{@@_add_normal:NNww} \meta{sign_1} \meta{sign_2}
%     ~~\Arg{exp_1}  \meta{body_1} |;| \Arg{exp_2} \meta{body_2} |;|
%   \end{syntax}
%   We now have two normal numbers to add, and we have to check signs
%   and exponents more carefully before performing the addition.
%    \begin{macrocode}
\cs_new:Npn \@@_add_normal:NNww #1#2
  {
    \if_meaning:w #1#2
      \exp_after:wN \@@_add_npos:Nnwnw
    \else:
      \exp_after:wN \@@_sub_npos:Nnwnw
    \fi:
    #1
  }
%    \end{macrocode}
% \end{macro}
%
% \subsubsection{Absolute addition}
%
% In this subsection, we perform the addition
% of two positive normal numbers.
%
% \begin{macro}[EXP]{\@@_add_npos:Nnwnw}
%   \begin{syntax}
%     \cs{@@_add_npos:Nnwnw} \meta{sign}
%     ~~\Arg{exp_1}  \meta{body_1} |;| \Arg{exp_2} \meta{body_2} |;|
%   \end{syntax}
%   Since we are doing an addition, \meta{sign} will be the final sign.
%   The only special case which may arise is the case of an overflow.
%   This will be checked by \cs{@@_sanitize:Nw} at the end of
%   the calculation. We start an \cs{__int_eval:w}, responsible for
%   computing the exponent, which may receive a contribution of |+1|
%   in case of carry. The exponent should be stopped by |;| followed by
%   the overall \meta{sign} for the sanitizing to work properly.
%
%   Grab and compare the exponents. The smaller number is decimated until
%   its exponent reaches that of the bigger number. We need to bring the
%   final sign down in the midst of the calculation to do the rounding
%   correctly.
%    \begin{macrocode}
\cs_new:Npn \@@_add_npos:Nnwnw #1 #2#3; #4
  {
    \exp_after:wN \@@_sanitize:Nw
    \exp_after:wN #1
    \int_use:N \__int_eval:w
      \if_int_compare:w #2 > #4 \exp_stop_f:
        #2
        \exp_after:wN \@@_add_big_i:wNww \__int_value:w -
      \else:
        #4
        \exp_after:wN \@@_add_big_ii:wNww \__int_value:w
      \fi:
      \__int_eval:w #4 - #2 ; #1 #3;
  }
%    \end{macrocode}
% \end{macro}
%
% \begin{macro}{\@@_add_big_i:wNww,
%     \@@_add_big_ii:wNww}
%   \begin{syntax}
%     \cs{@@_add_big_i:wNww} \meta{shift} |;| \meta{sign}
%     ~~\meta{body_1} |;| \meta{body_2} |;|
%   \end{syntax}
%   Shift the mantissa of the small number, and then add with
%   \cs{@@_add_mantissa:NnnwnnnnN}.
%    \begin{macrocode}
\cs_new:Npn \@@_add_big_i:wNww #1; #2 #3; #4;
  {
    \@@_decimate:nNnnnn {#1}
      \@@_add_mantissa:NnnwnnnnN
      #4
    #3
    #2
  }
\cs_new:Npn \@@_add_big_ii:wNww #1; #2 #3; #4;
  {
    \@@_decimate:nNnnnn {#1}
      \@@_add_mantissa:NnnwnnnnN
      #3
    #4
    #2
  }
%    \end{macrocode}
% \end{macro}
%
% \begin{macro}{\@@_add_mantissa:NnnwnnnnN}
%   \begin{syntax}
%     \cs{@@_add_mantissa:NnnwnnnnN}
%     ~~\meta{rounding}
%     ~~\Arg{Y'_1} \Arg{Y'_2} \meta{extra-digits} |;|
%     ~~\Arg{X_1} \Arg{X_2} \Arg{X_3} \Arg{X_4}
%     ~~\meta{final sign}
%   \end{syntax}
%   To round properly, we must know at which digit the rounding
%   should occur. This requires to know whether the addition
%   produces an overall carry or not. Thus, we do the computation
%   now and check for a carry, then go back and do the rounding.
%   The rounding may cause a carry in very rare cases such as
%   $0.99\cdots 95 \to 1.00\cdots 0$, but this situation always
%   give an exact power of $10$, for which it is easy to correct
%   the result at the end.
%    \begin{macrocode}
\cs_new:Npn \@@_add_mantissa:NnnwnnnnN #1 #2#3 #4; #5#6#7#8
  {
    \exp_after:wN \@@_add_mantissa_test:N
    \int_use:N \__int_eval:w 1#5#6 + #2
      \exp_after:wN \@@_add_mantissa_pack:NNNNNNN
      \int_use:N \__int_eval:w 1#7#8 + #3 ; #1
  }
%    \end{macrocode}
%
%    \begin{macrocode}
\cs_new:Npn \@@_add_mantissa_pack:NNNNNNN #1 #2#3#4#5#6#7
  {
    \if:w 2 #1
      + \c_one
    \fi:
    ; #2 #3 #4 #5 #6 #7 ;
  }
\cs_new:Npn \@@_add_mantissa_test:N #1
  {
    \if:w 2 #1
      \exp_after:wN \@@_add_mantissa_carry:wwNNNN
    \else:
      \exp_after:wN \@@_add_mantissa_no_carry:wwNNNN
    \fi:
  }
%    \end{macrocode}
%
%   \begin{quote}
%     \cs{@@_add_mantissa_no_carry:wwNNNN}
%     ~~\meta{8d} |;| \meta{6d} |;| \meta{2d} |;|
%     ~~\meta{rounding} \meta{sign}
%   \end{quote}
%   If there's no carry, grab all the digits again, and just
%   set the rounding correctly.\footnote{Bruno: an optimization
%     would be to compute whether we need rounding or not,
%     and only grab digits if there is rounding.}
%
%    \begin{macrocode}
\cs_new:Npn \@@_add_mantissa_no_carry:wwNNNN
    #1; #2; #3#4 ; #5#6
  {
    \exp_after:wN \@@_basics_pack_high:NNNNNw
    \int_use:N \__int_eval:w 1 #1
      \exp_after:wN \@@_basics_pack_low:NNNNNw
      \int_use:N \__int_eval:w 1 #2 #3#4
        + \@@_round:NNN #6 #4 #5
        \exp_after:wN ;
  }
%    \end{macrocode}
%
%   The case where there is a carry is very similar: rounding can even
%   raise the first digit from $1$ to $2$ (but we don't need to check that).
%   \begin{quote}
%     \cs{@@_add_mantissa_carry:wwNNNN}
%     ~~\meta{8d} |;| \meta{6d} |;| \meta{2d} |;|
%     ~~\meta{rounding} \meta{sign}
%   \end{quote}
%    \begin{macrocode}
\cs_new:Npn \@@_add_mantissa_carry:wwNNNN
    #1; #2; #3#4; #5#6
  {
    + \c_one
    \exp_after:wN \@@_add_mantissa_carry_pack:NNNNNNNNw
    \int_use:N \__int_eval:w 1 #1
      \exp_after:wN \@@_add_mantissa_carry_pack_ii:NNNNw
      \int_use:N \__int_eval:w 1 #2#3
        + \@@_round:NNNN #6 #3 #4 #5
        \exp_after:wN ;
  }
\cs_new:Npn \@@_add_mantissa_carry_pack_ii:NNNNw #1 #2#3#4 #5;
  {
    \if:w 2 #1
      + \c_one
    \fi:
    \__int_eval_end:
    #2#3#4; {#5} ;
  }
\cs_new:Npn \@@_add_mantissa_carry_pack:NNNNNNNNw
    #1#2#3#4 #5#6#7#8 #9; { ; {#1#2#3#4} {#5#6#7#8} {#9} }
%    \end{macrocode}
% \end{macro}
%
% \subsubsection{Absolute subtraction}
%
% \begin{macro}[EXP]{\@@_sub_npos:Nnwnw}
%   \begin{syntax}
%     \cs{@@_sub_npos:Nnwnw} \meta{sign}
%     ~~\Arg{exp_1}  \meta{body_1} |;| \Arg{exp_2} \meta{body_2} |;|
%   \end{syntax}
%   Rounding properly in some modes requires to know what the sign
%   of the result will be. For addition, this was easy. Here, besides
%   comparing the exponents to know how to decimate, we need to
%   check carefully which number is bigger when they have the same
%   exponent.
%    \begin{macrocode}
\cs_new:Npn \@@_sub_npos:Nnwnw #1 #2#3; #4 #5;
  {
    \exp_after:wN \@@_sanitize:wN
    \int_use:N \__int_eval:w
      \if_int_compare:w #2 > #4 \exp_stop_f:
        #2
        \exp_after:wN \@@_sub_big_i:wNww \__int_value:w -
      \else:
        #4
        \if_int_compare:w #2 = #4 \exp_stop_f:
          \@@_sub_exponent_eq:nnnnnnnn #3 #5
        \else:
          \exp_after:wN \@@_sub_big_ii:wNww \__int_value:w
        \fi:
      \fi:
      \__int_eval:w #4 - #2 ; #1 #3; #5;
  }
%    \end{macrocode}
% \end{macro}
%
% \begin{macro}{\@@_sub_exponent_eq:nnnnnnnn}
%    \begin{macrocode}
\cs_new:Npn \@@_sub_exponent_eq:nnnnnnnn #1#2#3#4 #5#6#7#8
  {
    \if_int_compare:w #1#2 > #5#6 \exp_stop_f:
      \exp_after:wN \@@_sub_big_i:wNww \__int_value:w
    \else:
      \if_int_compare:w #1#2 < #5#6 \exp_stop_f:
        \exp_after:wN \@@_sub_big_ii:wNww \__int_value:w
      \else:
        \if_int_compare:w #3#4 > #7#8 \exp_stop_f:
          \exp_after:wN \@@_sub_big_i:wNww \__int_value:w
        \else:
          \if_int_compare:w #3#4 < #7#8 \exp_stop_f:
            \exp_after:wN \@@_sub_big_ii:wNww \__int_value:w
          \else:
            \exp_after:wN \@@_sub_eq:wNww \__int_value:w
          \fi:
        \fi:
      \fi:
    \fi:
  }
\cs_new:Npn \@@_sub_eq:wNww #1; #2 #3; #4;
  { \exp_after:wN ; \exp_after:wN 1 \exp_after:wN ; }
%    \end{macrocode}
% \end{macro}
%
% \begin{macro}{\@@_sub_big_i:wNww,\@@_sub_big_ii:wNww}
%   \begin{syntax}
%     \cs{@@_sub_big_i:wNww} \meta{shift} |;| \meta{sign}
%     ~~\meta{body_1} |;| \meta{body_2} |;|
%   \end{syntax}
%   Shift the mantissa of the small number, and then subtract with
%   \cs{@@_sub_back_mantissa:NnnwNnnnn}.
%    \begin{macrocode}
\cs_new:Npn \@@_sub_big_i:wNww #1; #2 #3; #4;
  {
    \@@_decimate:nNnnnn {#1}
      \@@_sub_back_mantissa:NnnwNnnnn
      #4
    #2
    #3
  }
\cs_new:Npn \@@_sub_big_ii:wNww #1; #2 #3; #4;
  {
    \exp_after:wN \@@_sub_big_i:wNww
    \__int_value:w #1 \exp_after:wN ;
    \int_use:N \__int_eval:w 2 - #2 \__int_eval_end:
    #4; #3;
  }
%    \end{macrocode}
% \end{macro}
%
% \begin{macro}{\@@_sub_back_mantissa:NnnwNnnnn}
%   \begin{syntax}
%     \cs{@@_sub_back_mantissa:NnnwNnnnn}
%     ~~\meta{rounding} \Arg{Y'_1} \Arg{Y'_2} \meta{extra-digits}
%     ~~\meta{final sign}
%     ~~\Arg{X_1} \Arg{X_2} \Arg{X_3} \Arg{X_4}
%   \end{syntax}
%   At this stage, we know that \meta{Y} is less than \meta{X},
%   and we know the final sign.
%    \begin{macrocode}
\cs_new:Npn \@@_sub_back_mantissa:NnnwNnnnn #1 #2#3 #4; #5 #6#7#8#9
  {
    \exp_after:wN \@@_sub_back_mantissa_i:NNwNNNNwN
    \exp_after:wN #1
    \exp_after:wN #5
    \int_use:N \__int_eval:w 2#6#7 - #2 - \c_two +
      \exp_after:wN \@@_sub_back_mantissa_round:wNN
      \int_use:N \__int_eval:w 2#8#9 - #3 ; #1 #5
  }
%    \end{macrocode}
%   After the computation, we need to check whether the first digit of
%   the result is zero. This can only happen if the two numbers had the
%   same exponent, or exponents differing by $1$. In the latter case,
%   the \meta{rounding} digit is not quite enough to let us retrieve
%   the exact result (consider $\cdots25$ and $\cdots15$, both rounded
%   to $\cdots2$ in the usual mode), so we also move the result of
%   \cs{@@_round_neg:NNN} upstream as the digit $0$ or $1$.
%    \begin{macrocode}
\cs_new:Npn \@@_sub_back_mantissa_round:wNN #1; #2 #3
  {
    \exp_after:wN \@@_sub_back_mantissa_iii:N
    \__int_value:w
        \exp_after:wN \@@_round_neg:NNN
        \exp_after:wN #3
        \use_none:nnnnnnnn #1 #2
      + #1
    \exp_after:wN ;
  }
\cs_new:Npn \@@_sub_back_mantissa_iii:N #1
  {
    \exp_after:wN \@@_sub_back_mantissa_ii:NNNNNNw
    \exp_after:wN #1
    \int_use:N \__int_eval:w
      - #1
  }
\cs_new:Npn \@@_sub_back_mantissa_ii:NNNNNNw #1 #2 #3#4#5#6 #7;
  { #2 ; #1 {#3#4#5#6} {#7} ; }
%    \end{macrocode}
% \end{macro}
%
% \begin{macro}{\@@_sub_back_mantissa_i:NNwNNNNwN}
%   Here, |#3| should always be $2$, but we have to take is
%   as a normal undelimited argument since that would break
%   if |#2| is $2$.
%    \begin{macrocode}
\cs_new:Npn \@@_sub_back_mantissa_i:NNwNNNNwN #1#2 #3 #4#5#6#7 #8; #9
  {
    \if:w 0 #4
      \exp_after:wN \@@_sub_back_carry:NNwNnnnn
      \exp_after:wN #1
      \exp_after:wN #9
    \fi:
    ; #2
    {#4#5#6#7} {#8}
  }
%    \end{macrocode}
% \end{macro}
%
% \begin{macro}{\@@_sub_back_carry:NNwNnnnn}
%   \begin{syntax}
%     \cs{@@_sub_back_carry:NNwNnnnn}
%     ~~\meta{rounding} \meta{0 or 1} |;| \meta{final sign}
%     ~~\Arg{Z_1} \Arg{Z_2} \Arg{Z_3} \Arg{Z_4} |;|
%   \end{syntax}
%   This function is called when $\meta{Z_1}\leq 999$. We revert
%   the carry, which is given by \meta{0 or 1}, and subtract the
%   \meta{rounding} digit as appropriate, then feed the result,
%   of the form \meta{$\leq$ 7d} |;| \meta{9d} |;| to
%   \cs{@@_sub_back_carry_i:wwN}. The result is always exact.
%    \begin{macrocode}
\cs_new:Npn \@@_sub_back_carry:NNwNnnnn #1#2 ; #3 #4#5#6#7 ;
  {
    \exp_after:wN \@@_sub_back_carry_i:wwN
    \int_use:N \__int_eval:w #4 #5 - 1 + \exp_after:wN \@@_use_s:n
    \int_use:N \__int_eval:w 1 #6 #7 0 + #2 0 - #1 ; #3
  }
%    \end{macrocode}
%   Unless the first block is zero, check how many digits is has,
%   and shift the exponent down by the corresponding amount. Then
%   pack digits into blocks of $4$ (there are between $10$ and $16$
%   digits in front of \cs{@@_sub_back_carry_large:NNNNNNNNw}).
%    \begin{macrocode}
\cs_new:Npn \@@_sub_back_carry_i:wwN #1 ;
  {
    \if:w 0 #1
      - 8
      \exp_after:wN \@@_sub_back_carry_small:wN \__int_value:w
    \else:
      - \@@_sub_back_carry_ii:NNNNNNNNw #1 1234567;
      \exp_after:wN \@@_sub_back_carry_large:NNNNNNNNw
    \fi:
    #1
  }
%    \end{macrocode}
%   The case where the number is non-zero is slightly easier.
%    \begin{macrocode}
\cs_new:Npn \@@_sub_back_carry_ii:NNNNNNNNw #1#2#3#4#5#6#7#8#9; {#8}
\cs_new:Npn \@@_sub_back_carry_large:NNNNNNNNw #1#2#3#4 #5#6#7#8 #9;
  {
    \@@_sub_back_carry_large_ii:NNNNNNNNw
    #9 000000 ; {#1#2#3#4} {#5#6#7#8}
  }
\cs_new:Npn \@@_sub_back_carry_large_ii:NNNNNNNNw #1#2#3#4 #5#6#7#8 #9;
  { \@@_sub_back_carry_large_iii:nnnnN {#1#2#3#4} {#5#6#7#8} }
\cs_new:Npn \@@_sub_back_carry_large_iii:nnnnN #1#2 #3#4 #5
  { ; #5 {#3}{#4} {#1}{#2} ; }
%    \end{macrocode}
%   In the case of a \enquote{small} result, what comes after
%   \cs{@@_sub_back_carry_small:wN} has between $1$
%   and $9$ digits, and is not zero.
%    \begin{macrocode}
\cs_new:Npn \@@_sub_back_carry_small:wN #1;
  {
    - \exp_after:wN \@@_use_i_until_s:nw
      \use_none:nnnnnnnnn #1 012345678;
    \@@_sub_back_carry_small_ii:NNNNNNNN #1 00000000 ;
  }
\cs_new:Npn \@@_sub_back_carry_small_ii:NNNNNNNN #1#2#3#4 #5#6#7#8
  { \@@_sub_back_carry_small_iii:nnNwN {#1#2#3#4} {#5#6#7#8} }
\cs_new:Npn \@@_sub_back_carry_small_iii:nnNwN #1 #2 #3 #4; #5
  { ; #5 {#1} {#2} {#3000} {0000} ; }
%    \end{macrocode}
% \end{macro}
%
%
% \subsection{Multiplication}
%
% \begin{macro}[int, EXP]{\@@_*_o:ww}
%   For multiplication, everything is easy.  No need to grab the
%   \meta{body_2}.
%    \begin{macrocode}
\cs_new:cpn { @@_*_o:ww }
    \s_@@ \@@_chk:w #1 #2 #3 ; \s_@@ \@@_chk:w #4 #5
  { \@@_mul_cases:NN #1 #4 #2 #5 #3 ; }
%    \end{macrocode}
% \end{macro}
%
% \subsubsection{Signs, and special numbers}
%
% \begin{macro}[EXP,aux]{\@@_mul_cases:NN}
%   \begin{syntax}
%     \cs{@@_mul_cases:NN} \meta{case_1} \meta{case_2}
%     ~~\meta{sign_1} \meta{sign_2} \meta{body_1} |;| \meta{body_2} |;|
%   \end{syntax}
%   Expands the following tokens on the input stream once.
%   The special cases are coded at the start of this module,
%   and identical to the ones for division.\footnote{Bruno: \texttt{nan}
%     are not treated properly: $\infty\times 0$ should signal.}
%^^A todo: use the faster \if_meaning:w.
%    \begin{macrocode}
\cs_new:Npn \@@_mul_cases:NN #1 #2
  {
    \if_case:w \if_meaning:w 1 #1 #2 \else:
               \if_meaning:w 1 #2 #1 \else:
               \if_meaning:w #1#2 #1 \else:
               \if_int_compare:w \__int_eval:w #1 + #2 > \c_two
                 3 \else: 4 \fi: \fi: \fi: \fi:
               \exp_stop_f:
         \exp_after:wN \@@_basics_return_zero:NNww
    \or: \exp_after:wN \@@_mul_normal:NNww
    \or: \exp_after:wN \@@_basics_return_inf:NNww
    \or: \exp_after:wN \@@_basics_return_nan:NNww
    \or:
      \exp_after:wN \@@_mul_invalid:NNNNww
      \exp_after:wN #1
      \exp_after:wN #2
    \fi:
  }
\cs_new:Npn \@@_mul_invalid:NNNNww #1#2#3#4#5; #6;
  {
    \@@_invalid_operation:Nnww \c_nan_fp { * }
      \s_@@ \@@_chk:w #1 #3 #5 ;
      \s_@@ \@@_chk:w #2 #4 #6 ;
  }
%    \end{macrocode}
% \end{macro}
%
% \begin{macro}{\@@_mul_normal:NNww}
%   \begin{syntax}
%     \cs{@@_mul_normal:NNww} \meta{sign_1} \meta{sign_2}
%     ~~\Arg{exp_1}  \meta{body_1} |;| \Arg{exp_2} \meta{body_2} |;|
%   \end{syntax}
%   We now have two normal numbers to multiply. Combine the signs.
%    \begin{macrocode}
\cs_new:Npn \@@_mul_normal:NNww #1#2
  {
    \if:w #1#2
      \exp_after:wN \@@_mul_npos:Nnwnw
      \exp_after:wN 0
    \else:
      \exp_after:wN \@@_mul_npos:Nnwnw
      \exp_after:wN 2
    \fi:
  }
%    \end{macrocode}
% \end{macro}
%
% \subsubsection{Absolute multiplication}
%
% In this subsection, we perform the multiplication
% of two positive normal numbers.
%
% \begin{macro}[EXP]{\@@_mul_npos:Nnwnw}
%   \begin{syntax}
%     \cs{@@_mul_npos:Nnwnw} \meta{sign}
%     ~~\Arg{exp_1}  \meta{body_1} |;| \Arg{exp_2} \meta{body_2} |;|
%   \end{syntax}
%   As for addition, \meta{sign} is the final sign. After the computation,
%   \cs{@@_sanitize:Nw} checks for overflow or underflow.
%   As before, \cs{__int_eval:w} computes the exponent, catching any
%   shift coming from the computation in the mantissa. Again, the
%   \meta{sign} is needed for rounding to be done properly, so we move
%   it around with us. We setup the post-expansion here, triggered by
%   \cs{@@_mul_mantissa:nnnnNnnnn}.
%    \begin{macrocode}
\cs_new:Npn \@@_mul_npos:Nnwnw #1 #2#3; #4 #5;
  {
    \exp_after:wN \@@_sanitize:Nw
    \exp_after:wN #1
    \int_use:N \__int_eval:w
      #2 + #4
      \@@_mul_mantissa:nnnnNnnnn #3 #1 #5
    \exp_after:wN ;
  }
%    \end{macrocode}
% \end{macro}
%
% \begin{macro}{\@@_mul_mantissa:nnnnNnnnn}
%   \begin{syntax}
%     \cs{@@_mul_mantissa:nnnnNnnnn}
%     ~~\Arg{X_1} \Arg{X_2} \Arg{X_3} \Arg{X_4} \meta{sign}
%     ~~\Arg{Y_1} \Arg{Y_2} \Arg{Y_3} \Arg{Y_4} \meta{;}
%   \end{syntax}
%   After one expansion, the token following \meta{Y_4} must be a semicolon
%   (represented by \meta{;}).
%
%   The product of two $16$ digit integers has $31$ or $32$ digits,
%   but it is impossible to know which one before computing. The place
%   where we round depends on that number of digits, and may depend
%   on all digits until the last in some rare cases. The approach is
%   thus to compute the $5$ first blocks of $4$ digits (the first one
%   is between $100$ and $9999$ inclusive), and a compact version of
%   the remaining $3$ blocks. Afterwards, the number of digits is
%   known, and we can do the rounding within yet another set of
%   \cs{__int_eval:w}.
%    \begin{macrocode}
\cs_new:Npn \@@_mul_mantissa:nnnnNnnnn #1#2#3#4 #5 #6#7#8#9
  {
    \exp_after:wN \@@_mul_mantissa_after:NNN
    \exp_after:wN #5
    \int_use:N \__int_eval:w 99990000 + #1*#6 +
      \exp_after:wN \@@_mul_mantissa_keep:NNNNNw
      \int_use:N \__int_eval:w 99990000 + #1*#7 + #2*#6 +
        \exp_after:wN \@@_mul_mantissa_keep:NNNNNw
        \int_use:N \__int_eval:w 99990000 + #1*#8 + #2*#7 + #3*#6 +
          \exp_after:wN \@@_mul_mantissa_drop:NNNNNw
          \int_use:N \__int_eval:w  99990000 + #1*#9 + #2*#8 + #3*#7 + #4*#6 +
            \exp_after:wN \@@_mul_mantissa_drop:NNNNNw
            \int_use:N \__int_eval:w  99990000 + #2*#9 + #3*#8 + #4*#7 +
              \exp_after:wN \@@_mul_mantissa_drop:NNNNNw
              \int_use:N \__int_eval:w  99990000 + #3*#9 + #4*#8 +
                \exp_after:wN \@@_mul_mantissa_drop:NNNNNw
                \int_use:N \__int_eval:w 100000000 + #4*#9 \exp_after:wN ;
  }
\cs_new:Npn \@@_mul_mantissa_drop:NNNNNw #1#2#3#4#5 #6;
  { #1#2#3#4#5 ; + #6 }
\cs_new:Npn \@@_mul_mantissa_keep:NNNNNw #1#2#3#4#5 #6;
  { #1#2#3#4#5 ; #6 ; }
%    \end{macrocode}
%   Once the first \cs{int_use:N} \cs{__int_eval:w}, and all the
%   \cs{@@_mul_mantissa_...:NNNNNw} have been expanded,
%   we get
%   \begin{quote}
%     \cs{@@_mul_mantissa_after:NNN} \meta{sign} |1|
%     ~~\meta{digits 1--8} |;| \meta{digits 9--12} |;| \meta{digits 13--16} |;|
%     ~~|+| \meta{digits 17--20} |+| \meta{digits 21--24}
%     ~~|+| \meta{digits 25--28} |+| \meta{digits 29--32} |;|
%   \end{quote}
%   If the \meta{digit 1} is non-zero, then for rounding we only care
%   about the digits $16$ and $17$, and whether all other digits are zero
%   or not (check for exact ties). On the other hand, if \meta{digit 1}
%   is zero, we care about digits $17$ and $18$, and whether all others are
%   zero.
%    \begin{macrocode}
\cs_new:Npn \@@_mul_mantissa_after:NNN #1 #2 #3
  {
    \if:w 0 #3
      \exp_after:wN \@@_mul_mantissa_small:NNwwwN
    \else:
      \exp_after:wN \@@_mul_mantissa_large:NwwNNNN
    \fi:
    #1 #3
  }
%    \end{macrocode}
% \end{macro}
%
% \begin{macro}{\@@_mul_mantissa_large:NwwNNNN}
%   In this branch, \meta{digit 1} is non-zero. The result is thus
%   \meta{digits 1--16}, plus some rounding which depends on the digits
%   $16$, $17$, and whether all subsequent digits are zero or not.
%   Here, \cs{@@_round_s:NNNw} takes the \meta{sign}, followed by
%   digits $16$, $17$, and an integer expression which is zero if and
%   only if all further digits are zero.
%    \begin{macrocode}
\cs_new:Npn \@@_mul_mantissa_large:NwwNNNN #1 #2; #3; #4#5#6#7; +
  {
    \exp_after:wN \@@_basics_pack_high:NNNNNw
    \int_use:N \__int_eval:w 1#2
      \exp_after:wN \@@_basics_pack_low:NNNNNw
      \int_use:N \__int_eval:w 1#3#4#5#6#7 + \@@_round_s:NNNw #1 #7
  }
%    \end{macrocode}
% \end{macro}
%
% \begin{macro}{\@@_mul_mantissa_small:NNwwwN}
%   In this branch, \meta{digit 1} is zero. Our result will thus be
%   \meta{digits 2--17}, plus some rounding which depends on the digits
%   $17$, $18$, and whether all subsequent digits are zero or not.
%   The $8$ digits |1#3| are followed, after expansion of the
%   \texttt{small_pack} auxiliary, by the next digit, to form a $9$
%   digit number. Also, rounding may have caused a carry, which is
%   then converted to \cs{c_ten} rather than the usual \cs{c_one},
%   because of the shift.
%    \begin{macrocode}
\cs_new:Npn \@@_mul_mantissa_small:NNwwwN #1 #2#3; #4; #5; + #6
  {
    - \c_one
    \exp_after:wN \@@_basics_pack_high:NNNNNw
    \int_use:N \__int_eval:w 1#3
      \exp_after:wN \@@_mul_mantissa_small_pack:NNNNNNw
      \int_use:N \__int_eval:w 1#4#5#6 + \@@_round_s:NNNw #1 #6
  }
\cs_new:Npn \@@_mul_mantissa_small_pack:NNNNNNw #1#2 #3#4#5#6 #7;
  {
    #2
    \if:w 2 #1
      + \c_ten
    \fi:
    ; {#3#4#5#6} {#7} ;
  }
%    \end{macrocode}
% \end{macro}
%
% \subsection{Division}
%
% Time is now ripe to tackle the hardest of the four elementary
% operations: division.
%
% \begin{macro}[EXP]{\@@_/_o:ww}
%   For division we swap the two floating point numbers.
%    \begin{macrocode}
\cs_new:cpn { @@_/_o:ww }
    \s_@@ \@@_chk:w #1 #2 #3 ; \s_@@ \@@_chk:w #4 #5 #6 ;
  { \@@_div_back_cases:NN #4 #1 #5 #2 #6 ; #3 ; }
%    \end{macrocode}
% \end{macro}
%
% \subsubsection{Signs, and special numbers}
%
% In the case of division, the order of the operands matters,
% and it turns out to be slightly simpler if we internally
% compute the \enquote{backwards} division.
%
% \begin{macro}[EXP,aux]{\@@_div_back_cases:NN}
%   \begin{syntax}
%     \cs{@@_div_back_cases:NN} \meta{case_2} \meta{case_1}
%     ~~\meta{sign_2} \meta{sign_1} \meta{body_2} |;| \meta{body_1} |;|
%   \end{syntax}
%   Expands the following tokens on the input stream once.
%    \begin{macrocode}
\cs_new:Npn \@@_div_back_cases:NN #1 #2
  {
    \if_case:w \if_int_compare:w #1 = #2 \exp_stop_f:
                 #1 \exp_stop_f:
               \else:
                 \if_int_compare:w #1 < #2 \exp_stop_f:
                   \if:w 3 #2 \c_four \else: \c_five \fi:
                 \else:
                   \if:w 3 #1 \c_six \else: \c_seven \fi:
                 \fi:
               \fi:
         \exp_after:wN \@@_div_back_invalid:NNNww \exp_after:wN 0
    \or: \exp_after:wN \@@_div_back_normal:NNww
    \or: \exp_after:wN \@@_div_back_invalid:NNNww \exp_after:wN 2
    \or: \exp_after:wN \@@_basics_return_nan_nan:NNww
    \or:
      \exp_after:wN \@@_basics_return_ii:NNNNww
      \exp_after:wN #1
      \exp_after:wN #2
    \or: \exp_after:wN \@@_basics_return_inf:NNww
    \or:
      \exp_after:wN \@@_basics_return_i:NNNNww
      \exp_after:wN #1
      \exp_after:wN #2
    \or: \exp_after:wN \@@_basics_return_zero:NNww
    \fi:
  }
%    \end{macrocode}
%   Most of the special cases are common with some
%   previous operations. We only need to write the cases of
%   $0/0$ and $\infty/\infty$.
%    \begin{macrocode}
\cs_new:Npn \@@_div_back_invalid:NNNww #1#2#3 #4; #5;
  {
    \@@_invalid_operation:Nnww \c_nan_fp { / }
      \s_@@ \@@_chk:w #1 #3 #5 ;
      \s_@@ \@@_chk:w #1 #2 #4 ;
  }
%    \end{macrocode}
% \end{macro}
%
% \begin{macro}{\@@_div_back_normal:NNww}
%   \begin{syntax}
%     \cs{@@_div_back_normal:NNww} \meta{sign_1} \meta{sign_2}
%     ~~\Arg{exp_1}  \meta{body_1} |;| \Arg{exp_2} \meta{body_2} |;|
%   \end{syntax}
%   We now have two normal numbers to divide. Combine the signs.
%    \begin{macrocode}
\cs_new:Npn \@@_div_back_normal:NNww #1#2
  {
    \if:w #1#2
      \exp_after:wN \@@_div_back_npos:Nnwnw
      \exp_after:wN 0
    \else:
      \exp_after:wN \@@_div_back_npos:Nnwnw
      \exp_after:wN 2
    \fi:
  }
%    \end{macrocode}
% \end{macro}
%
% \subsubsection{Absolute (backwards) division}
%
% In this subsection, we perform the division
% of two positive normal numbers.
%
% \begin{macro}[EXP]{\@@_div_back_npos:Nnwnw}
%   \begin{syntax}
%     \cs{@@_div_back_npos:Nnwnw} \meta{sign}
%     ~~\Arg{exp Z} \Arg{Z_1} \Arg{Z_2} \Arg{Z_3} \Arg{Z_4} |;|
%     ~~\Arg{exp A} \Arg{A_1} \Arg{A_2} \Arg{A_3} \Arg{A_4} |;|
%   \end{syntax}
%   We want to compute $A/Z$. As for addition and multiplication,
%   \meta{sign} is the final sign. Checking for underflow and
%   overflow is done using the same auxiliary as for multiplication.
%   As explained just below, we first compute $y$, which is
%   the $5$ first digits of $Z$, plus $1$, and then compute pieces
%   of the quotient roughly $4$ digits at a time. Here, |#1| is
%   a single digit, |#2| and |#7| are the exponents (integers),
%   |#8| is three brace groups, and all others are each $4$ digits.
%    \begin{macrocode}
\cs_new:Npn \@@_div_back_npos:Nnwnw #1 #2 #3#4#5#6; #7 #8;
  {
    \exp_after:wN \@@_sanitize:Nw
    \exp_after:wN #1
    \int_use:N \__int_eval:w
      #7 - #2
      \@@_div_mantissa_i:wNwnn #3; #4;
      #8 {#3}{#4}{#5}{#6} #1
  }
%    \end{macrocode}
% \end{macro}
%
%   We are given two numbers, $A=0.A_{1}A_{2}A_{3}A_{4}$
%   and $Z=0.Z_{1}Z_{2}Z_{3}Z_{4}$, in blocks of $4$ digits,
%   and we know that the first digits of $A_{1}$ and of $Z_{1}$
%   are non-zero. To compute $A/Z$, we proceed as follows.
%   \begin{itemize}
%   \item Find an integer $Q_{1} \simeq 10^4 A / Z$.
%   \item Replace $A$ by $B = 10^4 A - Q_{1} Z$.
%   \item Find an integer $Q_{2} \simeq 10^4 B / Z$.
%   \item Replace $B$ by $C = 10^4 B - Q_{2} Z$.
%   \item Find an integer $Q_{3} \simeq 10^4 C / Z$.
%   \item Replace $C$ by $D = 10^4 C - Q_{3} Z$.
%   \item Find an integer $Q_{4} \simeq 10^4 D / Z$.
%   \item Consider $E = 10^4 D - Q_{4} Z$, and ensure
%     correct rounding.
%   \end{itemize}
%   The calculations of $B$, $C$, $D$, and $E$ can be done
%   exactly with only $16$ (or $17$) digits.
%
%   Unfortunately, things are not as easy as they seem.
%   Firstly, we make sure that all intermediate steps are positive,
%   since negative results would require extra calculations at the end.
%   This requires that $Q_{1} \leq 10^4 A / Z$ etc. A reasonable
%   attempt would be to define $Q_{1}$ as
%   \[
%   \cs{int_eval:n} \left\{
%     \frac{ A_{1} A_{2} }{ Z_{1} + 1 } - 1 \right\}.
%   \]
%   Subtracting $1$ at the end takes care of the fact that e\TeX{}'s
%   \cs{__int_eval:w} rounds instead of truncating. We add $1$ to $Z_{1}$
%   because $ Z_{1} \leq 10^4 Z < Z_{1}+1$ and we need $Q_{1}$
%   to be an underestimate. However, we are now underestimating
%   $Q_{1}$ too much: it can be wrong by up to $100$, for instance
%   when $Z = 0.1$ and $A \simeq 1$. Then $B$ could take values up to
%   $10$ (maybe more), and a few steps down the line, we would run into
%   arithmetic overflow, since \TeX{} can only handle integers less than
%   roughly $2\cdot 10^9$.
%
%   A better formula is to take
%   \[
%   Q_{1} = \cs{int_eval:n} \left\{
%     \frac{ 10 \cdot A_{1} A_{2} }
%       { \left\lfloor 10^{-3} \cdot Z_{1} Z_{2} \right\rfloor + 1 }
%     - 1 \right\}.
%   \]
%   This is always less than $10^9 A / (10^5 Z)$, as we wanted.
%   In words, we take the $5$ first digits of $Z$ into account,
%   and the $8$ first digits of $A$, using $0$ as a $9$-th digit
%   rather than the true digit for efficiency reasons. We shall
%   prove that using this formula to define all the $Q_{i}$
%   leads to no overflow. For convenience, let us denote
%   \[
%   y = \left\lfloor 10^{-3} \cdot Z_{1} Z_{2} \right\rfloor + 1,
%   \]
%   so that, taking into account the fact that e\TeX{} rounds ties
%   away from zero,
%   \[
%   Q_{1} = \left\lfloor A_{1}A_{2}0/y - 1/2 \right\rfloor.
%   \]
%   Note that $10^4<y\leq 10^5$, and $999 \leq Q_{1} \leq 99989$.
%   Also note that this formula does not cause an overflow as long as
%   $A<2.147\cdots$, since the numerator involves an integer slightly
%   smaller than $10^9A$.
%
%   Let us bound $B$:
%   \begin{align*}
%   10^5 B
%   &=
%   A_{1}A_{2}0 + 10 \cdot 0.A_{3}A_{4}
%   - 10 \cdot Z_{1}.Z_{2}Z_{3}Z_{4}
%   \cdot \left\lfloor A_{1}A_{2}0/y - 1/2 \right\rfloor
%   \\
%   &<
%   A_{1}A_{2}0
%   \cdot \left( 1 - 10 \frac{Z_{1}.Z_{2}Z_{3}Z_{4}}{y} \right)
%   + \frac{3}{2} \cdot 10 \cdot Z_{1}.Z_{2}Z_{3}Z_{4} + 10
%   \\
%   &\leq
%   \frac{A_{1}A_{2}0 \cdot (y - 10 Z_{1}.Z_{2}Z_{3}Z_{4})}{y}
%   + \frac{3}{2} y + 10
%   \\
%   &\leq
%   \frac{A_{1}A_{2}0}{y} + \frac{3}{2} y + 10
%   \leq
%   \frac{10^9 A}{y} + 1.6 y
%   \end{align*}
%   At the last step, we hide $10$ into the second term
%   for later convenience. The same reasoning yields\footnote{Bruno:
%     I need to find much better notations. These are not great.}
%   \begin{align*}
%     10^5 B &< 10^9 A/y + 1.6 y, \\
%     10^5 C &< 10^9 B/y + 1.6 y, \\
%     10^5 D &< 10^9 C/y + 1.6 y, \\
%     10^5 E &< 10^9 D/y + 1.6 y. \\
%   \end{align*}
%   The goal is now to prove that none of $B$, $C$, $D$, and $E$
%   can go beyond $2.147\cdots$. Simply bounding each term on the
%   right-hand side separately will not be tight enough: for instance,
%   we would get $10^5 B < 10^5 + 1.6\cdot 10^5 = 2.6 \cdot 10^5$,
%   which is too large.
%
%   Combining the various inequalities together with $A<1$, we get
%   \begin{align*}
%     10^5 B &< 10^9/y + 1.6 y, \\
%     10^5 C &< 10^{13}/y^2 + 1.6 (y + 10^4), \\
%     10^5 D &< 10^{17}/y^3 + 1.6 (y + 10^4 + 10^8/y), \\
%     10^5 E &< 10^{21}/y^4 + 1.6 (y + 10^4 + 10^8/y + 10^{12}/y^2). \\
%   \end{align*}
%   All of those bounds are convex functions of $y$ (since every power
%   of $y$ involved is convex, and hte coefficients are positive), and
%   thus maximal at one of the end-points of the allowed range
%   $10^4<y\leq 10^5$. Thus,
%   \begin{align*}
%     10^5 B &< \mathrm{max} ( 1.16\cdot 10^5, 1.7 \cdot 10^5), \\
%     10^5 C &< \mathrm{max} ( 1.32\cdot 10^5, 1.77 \cdot 10^5), \\
%     10^5 D &< \mathrm{max} ( 1.48\cdot 10^5, 1.777 \cdot 10^5), \\
%     10^5 E &< \mathrm{max} ( 1.64\cdot 10^5, 1.7777 \cdot 10^5). \\
%   \end{align*}
%   All of those bounds are less than $2.147\cdot 10^5$, and
%   we are thus within \TeX{}'s bounds in all cases!\footnote{Bruno:
%     but I need to check this very carefully again.}
%
%   We will later need to have a bound on the $Q_{i}$. Their
%   definitions imply that $Q_{1} < 10^9 A/y - 1/2 < 10^5 A$ and
%   similarly for the other $Q_{i}$. Thus each of them is at most
%   $177770$.
%
%   The last step is to ensure correct rounding. We have
%   \[
%   A/Z = \sum_{i=1}^4 \left(10^{-4i} Q_{i}\right) + 10^{-16} E/Z
%   \]
%   exactly. Furthermore, we know that the result will be between
%   $0.1$ (inclusive) and $10$, so we only need to know the integer
%   part of $E/Z$, and a \enquote{rounding} digit encoding the rest
%   (see maybe addition for an explanation of why). Equivalently,
%   we need to find the integer part of $2E/Z$, and determine whether
%   it was an exact integer or not (this serves to detect ties). Since
%   \[
%   \frac{2E}{Z} = 2\frac{10^5 E}{10^5 Z}
%   \leq 2\frac{10^5 E}{10^4} < 36,
%   \]
%   this integer part is between $0$ and $35$ inclusive. We let
%   e\TeX{} round
%   \[
%   P = \cs{int_eval:n} \left\{
%     2 \frac{E_{1}E_{2}}{Z_{1}Z_{2}} \right\},
%   \]
%   which differs from $2E/Z$ by at most
%   \[
%   \frac{1}{2}
%   + 2 \left\lvert \frac{E}{Z} - \frac{E}{10^{-8} Z_{1}Z_{2}}\right\rvert
%   + 2 \left\lvert \frac{10^8 E - E_{1}E_{2}}{Z_{1}Z_{2}}\right\rvert
%   < 1,
%   \]
%   ($1/2$ comes from e\TeX{}'s rounding) because each absolute value
%   is less than $10^{-7}$. Thus $P$ is either the correct integer part,
%   or an overestimate by $1$ (impossible if $2E/Z$ is an integer). It
%   then suffices to compare $PZ$ with $2E$ to get the integer part of
%   $2E/Z$ and the information of whether it was an exact quotient or not.
%
% \begin{macro}[EXP]{\@@_div_mantissa_i:wNwnn}
%   \begin{syntax}
%     \cs{@@_div_mantissa_i:wNwnn}
%     ~~\meta{Z_1} |;| \meta{Z_2} |;|
%     ~~\Arg{A_1} \Arg{A_2} \Arg{A_3} \Arg{A_4}
%     ~~\Arg{Z_1} \Arg{Z_2} \Arg{Z_3} \Arg{Z_4} \meta{sign}
%   \end{syntax}
%   First compute $y$ from the first $5$ digits of $Z$, and
%   unbrace \meta{A_1} and \meta{A_2}.
%    \begin{macrocode}
\cs_new:Npn \@@_div_mantissa_i:wNwnn #1; #2 #3; #4 #5
  {
    \exp_after:wN \@@_div_mantissa_ii:ww
    \int_use:N \__int_eval:w #1#2 + \c_one ;
    #4 #5 ;
  }
%    \end{macrocode}
%   \begin{quote}
%     \cs{@@_div_mantissa_ii:ww}
%     ~~\meta{y} |;| \meta{A_1} \meta{A_2} |;| \Arg{A_3} \Arg{A_4}
%     ~~\Arg{Z_1} \Arg{Z_2} \Arg{Z_3} \Arg{Z_4} \meta{sign}
%   \end{quote}
%   Compute $Q_{1}$ by evaluating $\meta{A_1}\meta{A_2}0/y - 1$.
%   The result will be output to the left, in an \cs{__int_eval:w}
%   which we start now.
%    \begin{macrocode}
\cs_new:Npn \@@_div_mantissa_ii:ww #1; #2;
  {
    \exp_after:wN \@@_div_mantissa_iii:www
    \__int_value:w #1 \exp_after:wN ;
    \__int_value:w
      \exp_after:wN \@@_div_mantissa_calc:Nwwnnnnnn
      \int_use:N \__int_eval:w #20/#1 + 999999 ; #2 ;
  }
%    \end{macrocode}
%   \begin{quote}
%     \cs{@@_div_mantissa_calc:Nwwnnnnnn} \meta{$10^6+{}$Q_1} |;|
%     ~~\meta{A_1} \meta{A_2} |;| \Arg{A_3} \Arg{A_4}
%     ~~\Arg{Z_1} \Arg{Z_2} \Arg{Z_3} \Arg{Z_4} \meta{sign}
%   \end{quote}
%   The goal here is to expand to
%   \begin{quote}
%     \meta{$10^6+{}$Q_1} |;| \meta{B_1} \meta{B_2} |;| \Arg{B_3} \Arg{B_4}
%     ~~\Arg{Z_1} \Arg{Z_2} \Arg{Z_3} \Arg{Z_4} \meta{sign}
%   \end{quote}
%   where $B = 10^4 A - Q_{1} \cdot Z$. More generally, this function
%   is used with $A\to B$, $B\to C$ and $Q_{1} \to Q_{2}$, etc.
%
%   Computing the product $Q_{1} \cdot Z$ is almost simple, since
%   $Q_{1}$ is rather small, but not quite:
%   the product of $Q_{1}$ with each block of four digits $Z_{i}$
%   is within \TeX{}'s bounds, but we wouldn't be able to use the
%   usual trick of adding a large power of $10$ to ensure that the
%   number of digits is fixed (see other operations for many examples
%   of this). Instead, we split off the digit of $10^5$ in $Q_{1}$
%   (and more generally $Q_{i}$), and do something similar to the
%   case of the full multiplication.
%
%   We know that $0<Q_{i}<1.8\cdot 10^5$, so $10^6+Q_{i}$ starts
%   with the digit $1$, followed by $\#1 = 1$ or $0$, then |#2|, which
%   is $5$ more digits. It would be somewhat simpler if we got |#1|
%   to be two digits, and |#2| four, but we are constrained by the $9$
%   arguments limit.
%
%   The result we want is then (the overall power of $10$ is arbitrary):
%   \begin{align*}
%   &10^{-4} ( \#3 - \#2 \cdot \#6 - 10 \cdot \#1 \cdot \#6\#7 )
%   + 10^{-8} ( \#4 - \#2 \cdot \#7 - 10 \cdot \#1 \cdot \#8 ) \\
%   &+ 10^{-12}( \#5 - \#2 \cdot \#8 - 10 \cdot \#1 \cdot \#9 )
%   + 10^{-16}(     - \#2 \cdot \#9 ).
%   \end{align*}
%   The factors of $10$ come from the fact that
%   $Q_{i} = 10\cdot 10^4 \cdot \#1 + \#2$. As usual, to combine
%   all the terms, we need to choose some shifts which must ensure
%   that the number of digits of the second, third, and fourth terms
%   are each fixed. Here, a good choice is $2\cdot 10^9$.
%   We are flirting with \TeX{}'s limits once more.
%
%   If $\#1=0$, then each term in parentheses (omitting the first)
%   is in the open interval $(-10^9, 10^4)$. Thus, adding
%   $2\cdot 10^9$ to it gives a $10$ digits number.\footnote{Bruno:
%     check that the carry from below does not screw that up. This
%   requires slightly tighter bounds.}
%
%   If $\#1=1$, then $\#2 < 7.8 \cdot 10^4$, and each term
%   in parentheses (omitting the first) is in the interval
%   $(-8\cdot 10^8, 10^4)$, and we are even safer.
%
%   We add the terms containing $\#1$ in a slightly tricky way
%   for efficiency reasons: if $\#1=0$, no need to do any computation,
%   while if $\#1=1$ we want $10$ times some number, simply obtained
%   by appending a $0$ digit.
%    \begin{macrocode}
\cs_new:Npn \@@_div_mantissa_calc:Nwwnnnnnn 1#1#2; #3;#4#5 #6#7#8#9
  {
    1 #1 #2 \exp_after:wN ;
    \int_use:N \__int_eval:w
      - 200000 + #3 - #2 * #6
      \if_meaning:w 1 #1
        - #6#70
      \fi:
      +
      \exp_after:wN \@@_div_mantissa_calc_last:NNNNNN
      \int_use:N \__int_eval:w
        1999800000 + #4 - #2*#7
        \if_meaning:w 1 #1
          - #80
        \fi:
        +
        \exp_after:wN \@@_div_mantissa_calc_pack:NNNNNNw
        \int_use:N \__int_eval:w
          1999800000 + #5 - #2*#8
          \if_meaning:w 1 #1
            - #90
          \fi:
          +
          \exp_after:wN \@@_div_mantissa_calc_pack:NNNNNNw
          \int_use:N \__int_eval:w 2000000000 - #2*#9 ;
    {#6}{#7}{#8}{#9}
  }
\cs_new:Npn \@@_div_mantissa_calc_pack:NNNNNNw #1#2#3#4#5#6 #7;
  { #1#2#3#4#5#6 ; {#7} }
\cs_new:Npn \@@_div_mantissa_calc_last:NNNNNN #1#2#3#4#5#6
  { #1#2#3#4#5#6 \__int_eval_end: }
%    \end{macrocode}
%   \begin{quote}
%     \cs{@@_div_mantissa_iii:www} \meta{y} |;| \meta{$10^6+{}$Q_1} |;|
%     ~~\meta{B_1} \meta{B_2} |;| \Arg{B_3} \Arg{B_4}
%     ~~\Arg{Z_1} \Arg{Z_2} \Arg{Z_3} \Arg{Z_4} \meta{sign}
%   \end{quote}
%   \begin{macrocode}
\cs_new:Npn \@@_div_mantissa_iii:www #1; #2; #3;
  {
    \exp_after:wN \@@_div_mantissa_iii_after:w
    \int_use:N \__int_eval:w #2
      \exp_after:wN \@@_div_mantissa_iv:www
      \__int_value:w #1 \exp_after:wN ;
      \__int_value:w
        \exp_after:wN \@@_div_mantissa_calc:Nwwnnnnnn
        \int_use:N \__int_eval:w #30/#1 + 999999 ;
        #3 ;
  }
%    \end{macrocode}
%   \begin{quote}
%     \cs{@@_div_mantissa_iv:www} \meta{y} |;| \meta{$10^6+{}$Q_2} |;|
%     ~~\meta{C_1} \meta{C_2} |;| \Arg{C_3} \Arg{C_4}
%     ~~\Arg{Z_1} \Arg{Z_2} \Arg{Z_3} \Arg{Z_4} \meta{sign}
%   \end{quote}
%    \begin{macrocode}
\cs_new:Npn \@@_div_mantissa_iv:www #1; #2; #3;
  {
    \exp_after:wN \@@_div_mantissa_pack:NNN
    \int_use:N \__int_eval:w #2
      \exp_after:wN \@@_div_mantissa_v:www
      \__int_value:w #1 \exp_after:wN ;
      \__int_value:w
        \exp_after:wN \@@_div_mantissa_calc:Nwwnnnnnn
        \int_use:N \__int_eval:w #30/#1 + 999999 ;
        #3 ;
  }
%    \end{macrocode}
%   \begin{quote}
%     \cs{@@_div_mantissa_v:www} \meta{y} |;| \meta{$10^6+{}$Q_3} |;|
%     ~~\meta{D_1} \meta{D_2} |;| \Arg{D_3} \Arg{D_4}
%     ~~\Arg{Z_1} \Arg{Z_2} \Arg{Z_3} \Arg{Z_4} \meta{sign}
%   \end{quote}
%    \begin{macrocode}
\cs_new:Npn \@@_div_mantissa_v:www #1; #2; #3;
  {
    \exp_after:wN \@@_div_mantissa_pack:NNN
    \int_use:N \__int_eval:w #2
      \exp_after:wN \@@_div_mantissa_vi:wwnnnn
      \__int_value:w
        \exp_after:wN \@@_div_mantissa_calc:Nwwnnnnnn
        \int_use:N \__int_eval:w #30/#1 + 999999 ;
        #3 ;
  }
%    \end{macrocode}
%   \begin{quote}
%     \cs{@@_div_mantissa_vi:wwnnnn} \meta{$10^6+{}$Q_4} |;|
%     ~~\meta{E_1} \meta{E_2} |;| \Arg{E_3} \Arg{E_4}
%     ~~\Arg{Z_1} \Arg{Z_2} \Arg{Z_3} \Arg{Z_4} \meta{sign}
%   \end{quote}
%   We compute $P$ by rounding $2 E_{1} E_{2}/Z_{1}Z_{2}$.
%    \begin{macrocode}
\cs_new:Npn \@@_div_mantissa_vi:wwnnnn #1; #2;#3#4 #5#6
  {
    \exp_after:wN \@@_div_mantissa_pack:NNN
    \int_use:N \__int_eval:w #10
      \exp_after:wN \@@_div_mantissa_vii:wwnnnnnn
      \int_use:N \__int_eval:w (\c_two*#2)/#5#6 ; % <- P
      #2;{#3}{#4}
      {#5}{#6}
  }
%    \end{macrocode}
%   Note that we used |#10| instead of |#2| which we had previously.
%   Two reasons: firstly, since we dropped $y$, the argument which
%   holds $Q_{i}$ has changed, and secondly, we will want the
%   fourth piece of the result to have $5$ digits, including the
%   \meta{rounding} digit, which we shall compute now from $P$.
%   \begin{quote}
%     \cs{@@_div_mantissa_vii:wwnnnnnn} \meta{P} |;|
%     ~~\meta{E_1} \meta{E_2} |;| \Arg{E_3} \Arg{E_4}
%     ~~\Arg{Z_1} \Arg{Z_2} \Arg{Z_3} \Arg{Z_4} \meta{sign}
%   \end{quote}
%   Then compute $2E-PZ$. Once more, we need to be careful and show
%   that the calculation $\#1\cdot\#5\#6$ below does not cause an
%   overflow: naively, $P$ can be up to $35$, and $\#5\#6$ up to
%   $10^8$, but both cannot happen simultaneously. To show that things
%   are fine, we split in two (non-disjoint) cases.
%   \begin{itemize}
%   \item For small $P$, say, $P< 10$, the product obeys
%     $P\cdot\#5\#6 < 10^8 \cdot P < 10^9 $.
%   \item For large $P$, say, $P\geq 3$, the rounding error on $P$,
%     which is at most $1$, is less than a factor of $2$, hence
%     $P\leq 4E/Z$, and $P\cdot \#5\#6 \leq 4E\cdot 10^8 < 10^9$.
%   \end{itemize}
%   Both inequalities could be made tighter if needed.
%
%   Note however that $P\cdot \#7\#8$ may overflow,
%   since the two factors are now independent, and the result may reach
%   $3.5\cdot 10^9$.
%
%   Also, we add $10\cdot P/2$ to the \enquote{fourth piece} of the result
%   as a first estimate of $10$ times $E/Z$. The goal is that the last digit
%   (for now $0$ or $5$) should be the \meta{rounding} digit. More precisely,
%   it will be corrected later by adding or subtracting $1$ depending on
%   whether $F$ was the correct integer part, or an overestimate (and nothing
%   is added when the quotient was exact). This does not give the
%   \enquote{correct} \meta{rounding} digit, but it always gives a digit
%   in the right \enquote{class} ($0$, $[1,4]$, $5$, or $[6-9]$), enough
%   for rounding purposes.
%    \begin{macrocode}
\cs_new:Npn \@@_div_mantissa_vii:wwnnnnnn #1; #2;#3#4 #5#6#7#8
  {
    + \c_five * #1
    \exp_after:wN \@@_div_mantissa_ix:Nww
    \int_use:N \__int_eval:w -20 + 2*#2 - #1*#5#6 +
      \exp_after:wN \@@_div_mantissa_viii:NNw
      \int_use:N \__int_eval:w 199980 + 2*#3 - #1*#7 +
        \exp_after:wN \@@_div_mantissa_viii:NNw
        \int_use:N \__int_eval:w 200000 + 2*#4 - #1*#8 ; ;
  }
\cs_new:Npn \@@_div_mantissa_viii:NNw #1#2#3; { #1#2 ; + #3 }
%    \end{macrocode}
%   \begin{quote}
%     \cs{@@_div_mantissa_ix:Nww}
%     ~~\meta{F_1} \meta{F_2} |;| |+| \meta{F_3} |+| \meta{F_4} |;| \meta{sign}
%   \end{quote}
%   where $F=2E-PZ$. We only need to know whether it is positive,
%   negative, or exactly zero.
%    \begin{macrocode}
\cs_new:Npn \@@_div_mantissa_ix:Nww #1#2;#3;
  {
    \if_meaning:w 0 #1
      \exp_after:wN \@@_div_mantissa_x:w
      \int_use:N \__int_eval:w #3
    \else:
      \if_meaning:w - #1
        -
      \else:
        +
      \fi:
      \c_one
    \fi:
    ;
  }
\cs_new:Npn \@@_div_mantissa_x:w #1;
  {
    \if_int_compare:w #1 > \c_zero
      + \c_one
    \fi:
    ;
  }
%    \end{macrocode}
%   We now obtain the following code, where \TeX{} is in the process
%   of expanding each of the integer expressions, and thus expands
%   the function at the bottom before the ones above it.
%   \begin{quote}
%     \cs{@@_div_mantissa_iii_after:w}  $10^6 + Q_{1}$
%     \cs{@@_div_mantissa_pack:NNN} $10^6 + Q_{2}$
%     \cs{@@_div_mantissa_pack:NNN} $10^6 + Q_{3}$
%     \cs{@@_div_mantissa_pack:NNN}
%     $10^7 + 10\cdot Q_{4} + 5 \cdot P + \varepsilon$ |;| \meta{sign}
%   \end{quote}
%   Here, $\varepsilon$ is $0$ in case $2E=PZ$ (\emph{i.e.}, $F=0$),
%   $1$ in case $2E>PZ$, which means that $P$ was the correct value,
%   but not with an exact quotient, and $-1$ if $2E<PZ$, \emph{i.e.},
%   $P$ was an overestimate.
%    \begin{macrocode}
\cs_new:Npn \@@_div_mantissa_pack:NNN 1 #1 #2 { + #1 #2 ; }
%    \end{macrocode}
%   Once those have been expanded, we get
%   \begin{quote}
%     \cs{@@_div_mantissa_iii_after:w} |1| |0| \meta{5d} |;|
%     ~~\meta{4d} |;| \meta{4d} |;| \meta{5d} |;| \meta{sign}
%   \end{quote}
%   The reason we know that the first two digits are |1| and |0|
%   is that the final result is known to be between $0.1$ (inclusive)
%   and $10$, hence $\widetilde{Q_{1}}$ (the tilde denoting the
%   contribution from the other $Q_{i}$) is at most $99999$,
%   and $10^6+\widetilde{Q_{1}} = 10\cdots$.
%
%   It is now time to round. This depends on how many digits
%   the final result will have.
%    \begin{macrocode}
\cs_new:Npn \@@_div_mantissa_iii_after:w 10 #1
  {
    \if_meaning:w 0 #1
      \exp_after:wN \@@_div_mantissa_small:wwwNNNNwN
    \else:
      \exp_after:wN \@@_div_mantissa_large:wwwNNNNwN
    \fi:
    #1
  }
%    \end{macrocode}
%   \begin{quote}
%     \cs{@@_div_mantissa_small:wwwNNNNwN} |0| \meta{4d} |;|
%     ~~\meta{4d} |;| \meta{4d} |;| \meta{5d} |;| \meta{sign}
%   \end{quote}
%    \begin{macrocode}
\cs_new:Npn \@@_div_mantissa_small:wwwNNNNwN
    0 #1; #2; #3; #4#5#6#7#8; #9
  {
    \exp_after:wN \@@_basics_pack_high:NNNNNw
    \int_use:N \__int_eval:w 1 #1#2
      \exp_after:wN \@@_basics_pack_low:NNNNNw
      \int_use:N \__int_eval:w 1 #3#4#5#6#7
        + \@@_round:NNN #9 #7 #8
        \exp_after:wN ;
  }
%    \end{macrocode}
%   \begin{quote}
%     \cs{@@_div_mantissa_large:wwwNNNNwN} \meta{5d} |;|
%     ~~\meta{4d} |;| \meta{4d} |;| \meta{5d} |;| \meta{sign}
%   \end{quote}
%   \footnote{Bruno: rename the \enquote{add mantissa carry pack} function.}
%    \begin{macrocode}
\cs_new:Npn \@@_div_mantissa_large:wwwNNNNwN
    #1; #2; #3; #4#5#6#7#8; #9
  {
    + \c_one
    \exp_after:wN \@@_div_mantissa_large_pack:NNNNNNNNw
    \int_use:N \__int_eval:w 1 #1 #2 %<- 1+9d
      \exp_after:wN \@@_add_mantissa_carry_pack_ii:NNNNw
      \int_use:N \__int_eval:w 1 #3 #4 #5 #6
        + \@@_round:NNNN #9 #6 #7 #8
        \exp_after:wN ;
  }
\cs_new:Npn \@@_div_mantissa_large_pack:NNNNNNNNw
    1 #1#2#3#4 #5#6#7#8 #9; { ; {#1#2#3#4} {#5#6#7#8} {#9} }
%    \end{macrocode}
% \end{macro}
%
% \subsection{Unary operations}
%
% \begin{macro}{\@@_neg:w}
%   This function flips the sign of the \meta{floating point}
%   and expands after it in the input stream, just like
%   \cs{@@_+_o:ww} etc.
%    \begin{macrocode}
\cs_new:Npn \@@_neg:w \s_@@ \@@_chk:w #1 #2
  {
    \exp_after:wN \@@_exp_after_o:w
    \exp_after:wN \s_@@
    \exp_after:wN \@@_chk:w
    \exp_after:wN #1
    \int_use:N \__int_eval:w \c_two - #2 \__int_eval_end:
  }
%    \end{macrocode}
% \end{macro}
%
% \begin{macro}{\@@_abs:w}
%   This function sets the sign of the \meta{floating point} to be
%   positive, and expands after itself in the input stream, just like
%   \cs{@@_neg:w}.
%    \begin{macrocode}
\cs_new:Npn \@@_abs:w \s_@@ \@@_chk:w #1 #2
  { \@@_exp_after_o:w \s_@@ \@@_chk:w #1 0 }
%    \end{macrocode}
% \end{macro}
%
%    \begin{macrocode}
%</initex|package>
%    \end{macrocode}
%
% \end{implementation}
%
% \PrintChanges
%
% \PrintIndex