diff options
Diffstat (limited to 'dviware')
39 files changed, 3806 insertions, 2745 deletions
diff --git a/dviware/dvisvgm/ChangeLog b/dviware/dvisvgm/ChangeLog index 725017aba8..4652cdb75d 100644 --- a/dviware/dvisvgm/ChangeLog +++ b/dviware/dvisvgm/ChangeLog @@ -1,3 +1,159 @@ +2020-08-12 Martin Gieseking <martin.gieseking@uos.de> + + * .appveyor.yml, .travis.yml, configure.ac, doc/dvisvgm.1, src/Doxyfile, + src/version.hpp: + set version to 2.10 + + * NEWS: + updated NEWS + +2020-07-27 Martin Gieseking <martin.gieseking@uos.de> + + * doc/dvisvgm.1, doc/dvisvgm.txt.in, libs/xxHash/xxh3.h, + libs/xxHash/xxhash.h: + updated xxHash to version 0.8 + +2020-07-24 Martin Gieseking <martin.gieseking@uos.de> + + * src/psdefs.cpp: + redefine internal GS operators .fillstroke and .eofillstroke +They are + used to implement PDF operators B and B*. +(closes #139) + +2020-07-16 Martin Gieseking <martin.gieseking@uos.de> + + * src/optimizer/TransformSimplifier.cpp: + don't unwind transform attributes of image elements + + * src/optimizer/GroupCollapser.cpp: + collapse groups w/o attributes first +in order to recurse into their + descendants. + + * src/optimizer/TransformSimplifier.cpp: + optimizer: unwind transform attributes of image elements +if it consists + only of scaling and translations. Merge the scaled translation +into x and + y, and simplify the remaining matrix(...) expression to scale(...) + +2020-07-15 Martin Gieseking <martin.gieseking@uos.de> + + * src/XMLNode.cpp, src/XMLNode.hpp, src/optimizer/GroupCollapser.cpp, + src/optimizer/RedundantElementRemover.cpp, + src/optimizer/WSNodeRemover.cpp: + renamed XMLNode::remove to XMLNode::detach + +2020-07-14 Martin Gieseking <martin.gieseking@uos.de> + + * src/optimizer/GroupCollapser.cpp: + don't move inner transforms to outer clipping paths + + * src/XMLNode.cpp: + fixed XMLElement::unwrap + + * src/optimizer/GroupCollapser.cpp: + improved optimizer module GroupCollapser + +2020-07-13 Martin Gieseking <martin.gieseking@uos.de> + + * src/optimizer/SVGOptimizer.cpp: + simplify transforms before optimizing groups + + * src/DvisvgmSpecialHandler.cpp: + fixed error message about invalid closing tags + + * src/DVIToSVGActions.cpp: + handle XML parser errors thrown in EOP + + * src/DvisvgmSpecialHandler.cpp, src/DvisvgmSpecialHandler.hpp: + reduced number of successive XML error messages + + * src/DvisvgmSpecialHandler.cpp, src/DvisvgmSpecialHandler.hpp: + renamed XMLParser::flush to XMLParser::finish + +2020-07-07 Martin Gieseking <martin.gieseking@uos.de> + + * src/FontCache.cpp, src/HashFunction.cpp, src/HashFunction.hpp, + src/MD5HashFunction.hpp, src/XXHashFunction.hpp, + tests/HashFunctionTest.cpp: + renamed HashFunction::digestValue() to HashFunction::digestBytes() + + * src/XXHashFunction.hpp: + added XXHashFunction::digestValue() to get the value as encoded by + xxhash + + * src/FileSystem.cpp, src/FileSystem.hpp: + create dedicated temp folders with each call of dvisvgm + +2020-06-25 Martin Gieseking <martin.gieseking@uos.de> + + * libs/xxHash/xxh3.h, libs/xxHash/xxhash.c, libs/xxHash/xxhash.h: + updated xxHash to version 0.7.4 + +2020-06-23 Martin Gieseking <martin.gieseking@uos.de> + + * src/DvisvgmSpecialHandler.cpp, tests/DvisvgmSpecialTest.cpp: + fixed computation of string boundaries after expanding {?(...)} +(closes + #136) + +2020-05-09 Martin Gieseking <martin.gieseking@uos.de> + + * .appveyor.yml: + updated FreeType to version 2.10.2 + +2020-05-03 Martin Gieseking <martin.gieseking@uos.de> + + * src/dvisvgm.cpp: + fixed indentation + +2020-04-24 John Bowman <bowman@ualberta.ca> + + * src/dvisvgm.cpp: + Fix indentation. + +2020-04-23 John Bowman <bowman@ualberta.ca> + + * src/dvisvgm.cpp: + Return error codes to shell. + +2020-04-13 Martin Gieseking <martin.gieseking@uos.de> + + * src/PathClipper.cpp, src/PathClipper.hpp, src/PsSpecialHandler.cpp: + changed signature of PathClipper::intersect() + + * src/PathClipper.cpp, src/PathClipper.hpp: + added unite operation to PathClipper + + * src/GraphicsPath.hpp, src/PsSpecialHandler.cpp: + fixed handling of clippath/eoclip combinations + +2020-04-11 Martin Gieseking <martin.gieseking@uos.de> + + * src/DVIToSVGActions.cpp: + fixed bbox debug rectangles + + * src/Font.cpp, src/Font.hpp: + added italic correction for fake slant + + * src/DVIReader.cpp, src/DVIReader.hpp, src/FontManager.cpp: + added evaluation of LuaTeX native font definitions + + * src/DVIReader.cpp: + minor code improvements + +2020-04-10 Martin Gieseking <martin.gieseking@uos.de> + + * src/psdefs.cpp: + fixed PS operator 'eoclip' + +2020-03-30 Martin Gieseking <martin.gieseking@uos.de> + + * src/optimizer/TransformSimplifier.cpp: + optimizer: fixed omission of second scale() parameter + 2020-03-21 Martin Gieseking <martin.gieseking@uos.de> * src/Matrix.cpp: diff --git a/dviware/dvisvgm/NEWS b/dviware/dvisvgm/NEWS index e898c38bd3..35d867ae4e 100644 --- a/dviware/dvisvgm/NEWS +++ b/dviware/dvisvgm/NEWS @@ -1,3 +1,22 @@ +dvisvgm-2.10 (2020-08-12) +- added evaluation of LuaTeX native font definitions +- added computation of of italic correction for fake slanted glyphs +- dvisvgm now creates dedicated temporary folders for each call of + the program to prevent file collisions between different processes +- added evaluation of internal GS operators '.fillstroke' and '.eofillstroke' + introduced with GS 9.52 to implement PDF operators 'B' and 'B*' + (GH issue #139) +- fixed and improved optimization of nested group elements +- fixed optimization of 'scale' transformation +- fixed optimization of 'transform' attributes applied to images clipping paths +- fixed processing of PS operator 'eoclip' +- fixed processing of clipping path intersections with different + fill rule properties +- fixed processing of {?(...)} expressions (GH issue #136) +- fixed error message about unmatched closing tags in XML parser +- updated bundled FreeType library to version 2.10.2 +- updated bundled xxHash library to version 0.8 + dvisvgm-2.9.1 (2020-03-21) - added macro {?(expr)} to enable the evaluation of math expressions in raw SVG fragments (GH #130) diff --git a/dviware/dvisvgm/configure b/dviware/dvisvgm/configure index 55663f51f2..a548bc57ea 100755 --- a/dviware/dvisvgm/configure +++ b/dviware/dvisvgm/configure @@ -1,6 +1,6 @@ #! /bin/sh # Guess values for system-dependent variables and create Makefiles. -# Generated by GNU Autoconf 2.69 for dvisvgm 2.9.1. +# Generated by GNU Autoconf 2.69 for dvisvgm 2.10. # # Report bugs to <martin.gieseking@uos.de>. # @@ -590,8 +590,8 @@ MAKEFLAGS= # Identity of this package. PACKAGE_NAME='dvisvgm' PACKAGE_TARNAME='dvisvgm' -PACKAGE_VERSION='2.9.1' -PACKAGE_STRING='dvisvgm 2.9.1' +PACKAGE_VERSION='2.10' +PACKAGE_STRING='dvisvgm 2.10' PACKAGE_BUGREPORT='martin.gieseking@uos.de' PACKAGE_URL='' @@ -1406,7 +1406,7 @@ if test "$ac_init_help" = "long"; then # Omit some internal or obsolete options to make the list less imposing. # This message is too long to be a string in the A/UX 3.1 sh. cat <<_ACEOF -\`configure' configures dvisvgm 2.9.1 to adapt to many kinds of systems. +\`configure' configures dvisvgm 2.10 to adapt to many kinds of systems. Usage: $0 [OPTION]... [VAR=VALUE]... @@ -1477,7 +1477,7 @@ fi if test -n "$ac_init_help"; then case $ac_init_help in - short | recursive ) echo "Configuration of dvisvgm 2.9.1:";; + short | recursive ) echo "Configuration of dvisvgm 2.10:";; esac cat <<\_ACEOF @@ -1625,7 +1625,7 @@ fi test -n "$ac_init_help" && exit $ac_status if $ac_init_version; then cat <<\_ACEOF -dvisvgm configure 2.9.1 +dvisvgm configure 2.10 generated by GNU Autoconf 2.69 Copyright (C) 2012 Free Software Foundation, Inc. @@ -2169,7 +2169,7 @@ cat >config.log <<_ACEOF This file contains any messages produced by compilers while running configure, to aid debugging if configure makes a mistake. -It was created by dvisvgm $as_me 2.9.1, which was +It was created by dvisvgm $as_me 2.10, which was generated by GNU Autoconf 2.69. Invocation command line was $ $0 $@ @@ -2527,7 +2527,7 @@ ac_link='$CC -o conftest$ac_exeext $CFLAGS $CPPFLAGS $LDFLAGS conftest.$ac_ext $ ac_compiler_gnu=$ac_cv_c_compiler_gnu -DATE="March 2020" +DATE="August 2020" ac_config_headers="$ac_config_headers config.h" @@ -3157,7 +3157,7 @@ fi # Define the identity of the package. PACKAGE='dvisvgm' - VERSION='2.9.1' + VERSION='2.10' cat >>confdefs.h <<_ACEOF @@ -20138,7 +20138,7 @@ cat >>$CONFIG_STATUS <<\_ACEOF || ac_write_fail=1 # report actual input values of CONFIG_FILES etc. instead of their # values after options handling. ac_log=" -This file was extended by dvisvgm $as_me 2.9.1, which was +This file was extended by dvisvgm $as_me 2.10, which was generated by GNU Autoconf 2.69. Invocation command line was CONFIG_FILES = $CONFIG_FILES @@ -20204,7 +20204,7 @@ _ACEOF cat >>$CONFIG_STATUS <<_ACEOF || ac_write_fail=1 ac_cs_config="`$as_echo "$ac_configure_args" | sed 's/^ //; s/[\\""\`\$]/\\\\&/g'`" ac_cs_version="\\ -dvisvgm config.status 2.9.1 +dvisvgm config.status 2.10 configured by $0, generated by GNU Autoconf 2.69, with options \\"\$ac_cs_config\\" diff --git a/dviware/dvisvgm/configure.ac b/dviware/dvisvgm/configure.ac index 783d2709a8..6b07ef13d2 100644 --- a/dviware/dvisvgm/configure.ac +++ b/dviware/dvisvgm/configure.ac @@ -4,8 +4,8 @@ # Process this file with autoconf to produce a configure script. AC_PREREQ(2.59) -AC_INIT([dvisvgm],[2.9.1],[martin.gieseking@uos.de]) -DATE="March 2020" +AC_INIT([dvisvgm],[2.10],[martin.gieseking@uos.de]) +DATE="August 2020" AC_CONFIG_SRCDIR(src) AC_CONFIG_HEADERS([config.h]) AC_CONFIG_MACRO_DIR([m4]) diff --git a/dviware/dvisvgm/doc/dvisvgm.1 b/dviware/dvisvgm/doc/dvisvgm.1 index 67cc037d26..085e67c03d 100644 --- a/dviware/dvisvgm/doc/dvisvgm.1 +++ b/dviware/dvisvgm/doc/dvisvgm.1 @@ -4,10 +4,10 @@ .\" Generator: DocBook XSL Stylesheets vsnapshot <http://docbook.sf.net/> .\" Date: 2020-03-19 .\" Manual: dvisvgm Manual -.\" Source: dvisvgm 2.9.1 +.\" Source: dvisvgm 2.10 .\" Language: English .\" -.TH "DVISVGM" "1" "2020\-03\-19" "dvisvgm 2\&.9\&.1" "dvisvgm Manual" +.TH "DVISVGM" "1" "2020\-03\-19" "dvisvgm 2\&.10" "dvisvgm Manual" .\" ----------------------------------------------------------------- .\" * Define some portability stuff .\" ----------------------------------------------------------------- @@ -775,7 +775,7 @@ of option \fBMD5\fR, \fBXXH32\fR, and \fBXXH64\fR, where the names can be given in lower case too, like -\fB\-\-page\-hashes=md5\fR\&. Since version 0\&.7\&.1, xxHash provides an experimental 128\-bit hash function, whose algorithm may still change with the next versions of the library\&. If the corresponding API is available, dvisvgm supports the new hash function and option +\fB\-\-page\-hashes=md5\fR\&. Since version 0\&.7\&.1, xxHash provides an experimental 128\-bit hash function whose algorithm has been stabilized as of version 0\&.8\&. When using a version prior to 0\&.8, the 128\-bit hash values can vary depending on the used xxHash version\&. If the corresponding API is available, dvisvgm supports the new hash function, and option \fB\-\-page\-hashes\fR additionally accepts the algorithm specifier \fBXXH128\fR\&. diff --git a/dviware/dvisvgm/doc/dvisvgm.txt.in b/dviware/dvisvgm/doc/dvisvgm.txt.in index 2b51b57c32..71fb3351b1 100644 --- a/dviware/dvisvgm/doc/dvisvgm.txt.in +++ b/dviware/dvisvgm/doc/dvisvgm.txt.in @@ -518,9 +518,10 @@ By default, dvisvgm uses the fast XXH64 hash algorithm to compute the values pro XXH32 and MD5 can be used as well. The desired algorithm is specified by argument 'params' of option *--page-hashes*. It takes one of the strings +MD5+, +XXH32+, and +XXH64+, where the names can be given in lower case too, like +--page-hashes=md5+. Since version 0.7.1, xxHash provides an experimental 128-bit -hash function, whose algorithm may still change with the next versions of the library. If the -corresponding API is available, dvisvgm supports the new hash function and option *--page-hashes* -additionally accepts the algorithm specifier +XXH128+. +hash function whose algorithm has been stabilized as of version 0.8. When using a version prior to 0.8, +the 128-bit hash values can vary depending on the used xxHash version. If the corresponding API is +available, dvisvgm supports the new hash function, and option *--page-hashes* additionally accepts the +algorithm specifier +XXH128+. + Finally, option *--page-hashes* can take a second argument that must be separated by a comma. Currently, only the two parameters 'list' and 'replace' are evaluated, e.g. +--page-hashes=md5,list+ diff --git a/dviware/dvisvgm/libs/xxHash/xxh3.h b/dviware/dvisvgm/libs/xxHash/xxh3.h index 9e3e88ad61..7e83e6418a 100644 --- a/dviware/dvisvgm/libs/xxHash/xxh3.h +++ b/dviware/dvisvgm/libs/xxHash/xxh3.h @@ -1,7 +1,7 @@ /* * xxHash - Extremely Fast Hash algorithm * Development source file for `xxh3` - * Copyright (C) 2019-present, Yann Collet + * Copyright (C) 2019-2020 Yann Collet * * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php) * @@ -34,2079 +34,22 @@ */ /* - * Note: This file is separated for development purposes. - * It will be integrated into `xxhash.h` when development stage is completed. + * Note: This file used to host the source code of XXH3_* variants. + * during the development period. + * The source code is now properly integrated within xxhash.h. * - * Credit: most of the work on vectorial and asm variants comes from @easyaspi314 - */ - -#ifndef XXH3_H_1397135465 -#define XXH3_H_1397135465 - -/* === Dependencies === */ -#ifndef XXHASH_H_5627135585666179 -/* special: when including `xxh3.h` directly, turn on XXH_INLINE_ALL */ -# undef XXH_INLINE_ALL /* avoid redefinition */ -# define XXH_INLINE_ALL -#endif -#include "xxhash.h" - - -/* === Compiler specifics === */ - -#if defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L /* >= C99 */ -# define XXH_RESTRICT restrict -#else -/* Note: it might be useful to define __restrict or __restrict__ for some C++ compilers */ -# define XXH_RESTRICT /* disable */ -#endif - -#if (defined(__GNUC__) && (__GNUC__ >= 3)) \ - || (defined(__INTEL_COMPILER) && (__INTEL_COMPILER >= 800)) \ - || defined(__clang__) -# define XXH_likely(x) __builtin_expect(x, 1) -# define XXH_unlikely(x) __builtin_expect(x, 0) -#else -# define XXH_likely(x) (x) -# define XXH_unlikely(x) (x) -#endif - -#if defined(__GNUC__) -# if defined(__AVX2__) -# include <immintrin.h> -# elif defined(__SSE2__) -# include <emmintrin.h> -# elif defined(__ARM_NEON__) || defined(__ARM_NEON) -# define inline __inline__ /* clang bug */ -# include <arm_neon.h> -# undef inline -# endif -#elif defined(_MSC_VER) -# include <intrin.h> -#endif - -/* - * One goal of XXH3 is to make it fast on both 32-bit and 64-bit, while - * remaining a true 64-bit/128-bit hash function. - * - * This is done by prioritizing a subset of 64-bit operations that can be - * emulated without too many steps on the average 32-bit machine. - * - * For example, these two lines seem similar, and run equally fast on 64-bit: - * - * xxh_u64 x; - * x ^= (x >> 47); // good - * x ^= (x >> 13); // bad - * - * However, to a 32-bit machine, there is a major difference. - * - * x ^= (x >> 47) looks like this: - * - * x.lo ^= (x.hi >> (47 - 32)); - * - * while x ^= (x >> 13) looks like this: - * - * // note: funnel shifts are not usually cheap. - * x.lo ^= (x.lo >> 13) | (x.hi << (32 - 13)); - * x.hi ^= (x.hi >> 13); - * - * The first one is significantly faster than the second, simply because the - * shift is larger than 32. This means: - * - All the bits we need are in the upper 32 bits, so we can ignore the lower - * 32 bits in the shift. - * - The shift result will always fit in the lower 32 bits, and therefore, - * we can ignore the upper 32 bits in the xor. - * - * Thanks to this optimization, XXH3 only requires these features to be efficient: - * - * - Usable unaligned access - * - A 32-bit or 64-bit ALU - * - If 32-bit, a decent ADC instruction - * - A 32 or 64-bit multiply with a 64-bit result - * - For the 128-bit variant, a decent byteswap helps short inputs. - * - * The first two are already required by XXH32, and almost all 32-bit and 64-bit - * platforms which can run XXH32 can run XXH3 efficiently. - * - * Thumb-1, the classic 16-bit only subset of ARM's instruction set, is one - * notable exception. - * - * First of all, Thumb-1 lacks support for the UMULL instruction which - * performs the important long multiply. This means numerous __aeabi_lmul - * calls. - * - * Second of all, the 8 functional registers are just not enough. - * Setup for __aeabi_lmul, byteshift loads, pointers, and all arithmetic need - * Lo registers, and this shuffling results in thousands more MOVs than A32. - * - * A32 and T32 don't have this limitation. They can access all 14 registers, - * do a 32->64 multiply with UMULL, and the flexible operand allowing free - * shifts is helpful, too. - * - * Therefore, we do a quick sanity check. - * - * If compiling Thumb-1 for a target which supports ARM instructions, we will - * emit a warning, as it is not a "sane" platform to compile for. - * - * Usually, if this happens, it is because of an accident and you probably need - * to specify -march, as you likely meant to compile for a newer architecture. - */ -#if defined(__thumb__) && !defined(__thumb2__) && defined(__ARM_ARCH_ISA_ARM) -# warning "XXH3 is highly inefficient without ARM or Thumb-2." -#endif - -/* ========================================== - * Vectorization detection - * ========================================== */ -#define XXH_SCALAR 0 /* Portable scalar version */ -#define XXH_SSE2 1 /* SSE2 for Pentium 4 and all x86_64 */ -#define XXH_AVX2 2 /* AVX2 for Haswell and Bulldozer */ -#define XXH_NEON 3 /* NEON for most ARMv7-A and all AArch64 */ -#define XXH_VSX 4 /* VSX and ZVector for POWER8/z13 */ - -#ifndef XXH_VECTOR /* can be defined on command line */ -# if defined(__AVX2__) -# define XXH_VECTOR XXH_AVX2 -# elif defined(__SSE2__) || defined(_M_AMD64) || defined(_M_X64) || (defined(_M_IX86_FP) && (_M_IX86_FP == 2)) -# define XXH_VECTOR XXH_SSE2 -# elif defined(__GNUC__) /* msvc support maybe later */ \ - && (defined(__ARM_NEON__) || defined(__ARM_NEON)) \ - && (defined(__LITTLE_ENDIAN__) /* We only support little endian NEON */ \ - || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)) -# define XXH_VECTOR XXH_NEON -# elif (defined(__PPC64__) && defined(__POWER8_VECTOR__)) \ - || (defined(__s390x__) && defined(__VEC__)) \ - && defined(__GNUC__) /* TODO: IBM XL */ -# define XXH_VECTOR XXH_VSX -# else -# define XXH_VECTOR XXH_SCALAR -# endif -#endif - -/* - * Controls the alignment of the accumulator. - * This is for compatibility with aligned vector loads, which are usually faster. - */ -#ifndef XXH_ACC_ALIGN -# if XXH_VECTOR == XXH_SCALAR /* scalar */ -# define XXH_ACC_ALIGN 8 -# elif XXH_VECTOR == XXH_SSE2 /* sse2 */ -# define XXH_ACC_ALIGN 16 -# elif XXH_VECTOR == XXH_AVX2 /* avx2 */ -# define XXH_ACC_ALIGN 32 -# elif XXH_VECTOR == XXH_NEON /* neon */ -# define XXH_ACC_ALIGN 16 -# elif XXH_VECTOR == XXH_VSX /* vsx */ -# define XXH_ACC_ALIGN 16 -# endif -#endif - -/* - * UGLY HACK: - * GCC usually generates the best code with -O3 for xxHash. - * - * However, when targeting AVX2, it is overzealous in its unrolling resulting - * in code roughly 3/4 the speed of Clang. - * - * There are other issues, such as GCC splitting _mm256_loadu_si256 into - * _mm_loadu_si128 + _mm256_inserti128_si256. This is an optimization which - * only applies to Sandy and Ivy Bridge... which don't even support AVX2. - * - * That is why when compiling the AVX2 version, it is recommended to use either - * -O2 -mavx2 -march=haswell - * or - * -O2 -mavx2 -mno-avx256-split-unaligned-load - * for decent performance, or to use Clang instead. - * - * Fortunately, we can control the first one with a pragma that forces GCC into - * -O2, but the other one we can't control without "failed to inline always - * inline function due to target mismatch" warnings. - */ -#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \ - && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \ - && defined(__OPTIMIZE__) && !defined(__OPTIMIZE_SIZE__) /* respect -O0 and -Os */ -# pragma GCC push_options -# pragma GCC optimize("-O2") -#endif - - -#if XXH_VECTOR == XXH_NEON -/* - * NEON's setup for vmlal_u32 is a little more complicated than it is on - * SSE2, AVX2, and VSX. - * - * While PMULUDQ and VMULEUW both perform a mask, VMLAL.U32 performs an upcast. - * - * To do the same operation, the 128-bit 'Q' register needs to be split into - * two 64-bit 'D' registers, performing this operation:: - * - * [ a | b ] - * | '---------. .--------' | - * | x | - * | .---------' '--------. | - * [ a & 0xFFFFFFFF | b & 0xFFFFFFFF ],[ a >> 32 | b >> 32 ] - * - * Due to significant changes in aarch64, the fastest method for aarch64 is - * completely different than the fastest method for ARMv7-A. - * - * ARMv7-A treats D registers as unions overlaying Q registers, so modifying - * D11 will modify the high half of Q5. This is similar to how modifying AH - * will only affect bits 8-15 of AX on x86. - * - * VZIP takes two registers, and puts even lanes in one register and odd lanes - * in the other. - * - * On ARMv7-A, this strangely modifies both parameters in place instead of - * taking the usual 3-operand form. - * - * Therefore, if we want to do this, we can simply use a D-form VZIP.32 on the - * lower and upper halves of the Q register to end up with the high and low - * halves where we want - all in one instruction. - * - * vzip.32 d10, d11 @ d10 = { d10[0], d11[0] }; d11 = { d10[1], d11[1] } - * - * Unfortunately we need inline assembly for this: Instructions modifying two - * registers at once is not possible in GCC or Clang's IR, and they have to - * create a copy. - * - * aarch64 requires a different approach. - * - * In order to make it easier to write a decent compiler for aarch64, many - * quirks were removed, such as conditional execution. - * - * NEON was also affected by this. - * - * aarch64 cannot access the high bits of a Q-form register, and writes to a - * D-form register zero the high bits, similar to how writes to W-form scalar - * registers (or DWORD registers on x86_64) work. - * - * The formerly free vget_high intrinsics now require a vext (with a few - * exceptions) - * - * Additionally, VZIP was replaced by ZIP1 and ZIP2, which are the equivalent - * of PUNPCKL* and PUNPCKH* in SSE, respectively, in order to only modify one - * operand. - * - * The equivalent of the VZIP.32 on the lower and upper halves would be this - * mess: - * - * ext v2.4s, v0.4s, v0.4s, #2 // v2 = { v0[2], v0[3], v0[0], v0[1] } - * zip1 v1.2s, v0.2s, v2.2s // v1 = { v0[0], v2[0] } - * zip2 v0.2s, v0.2s, v1.2s // v0 = { v0[1], v2[1] } - * - * Instead, we use a literal downcast, vmovn_u64 (XTN), and vshrn_n_u64 (SHRN): - * - * shrn v1.2s, v0.2d, #32 // v1 = (uint32x2_t)(v0 >> 32); - * xtn v0.2s, v0.2d // v0 = (uint32x2_t)(v0 & 0xFFFFFFFF); - * - * This is available on ARMv7-A, but is less efficient than a single VZIP.32. - */ - -/* - * Function-like macro: - * void XXH_SPLIT_IN_PLACE(uint64x2_t &in, uint32x2_t &outLo, uint32x2_t &outHi) - * { - * outLo = (uint32x2_t)(in & 0xFFFFFFFF); - * outHi = (uint32x2_t)(in >> 32); - * in = UNDEFINED; - * } - */ -# if !defined(XXH_NO_VZIP_HACK) /* define to disable */ \ - && defined(__GNUC__) \ - && !defined(__aarch64__) && !defined(__arm64__) -# define XXH_SPLIT_IN_PLACE(in, outLo, outHi) \ - do { \ - /* Undocumented GCC/Clang operand modifier: %e0 = lower D half, %f0 = upper D half */ \ - /* https://github.com/gcc-mirror/gcc/blob/38cf91e5/gcc/config/arm/arm.c#L22486 */ \ - /* https://github.com/llvm-mirror/llvm/blob/2c4ca683/lib/Target/ARM/ARMAsmPrinter.cpp#L399 */ \ - __asm__("vzip.32 %e0, %f0" : "+w" (in)); \ - (outLo) = vget_low_u32 (vreinterpretq_u32_u64(in)); \ - (outHi) = vget_high_u32(vreinterpretq_u32_u64(in)); \ - } while (0) -# else -# define XXH_SPLIT_IN_PLACE(in, outLo, outHi) \ - do { \ - (outLo) = vmovn_u64 (in); \ - (outHi) = vshrn_n_u64 ((in), 32); \ - } while (0) -# endif -#endif /* XXH_VECTOR == XXH_NEON */ - -/* - * VSX and Z Vector helpers. - * - * This is very messy, and any pull requests to clean this up are welcome. - * - * There are a lot of problems with supporting VSX and s390x, due to - * inconsistent intrinsics, spotty coverage, and multiple endiannesses. - */ -#if XXH_VECTOR == XXH_VSX -# if defined(__s390x__) -# include <s390intrin.h> -# else -# include <altivec.h> -# endif - -# undef vector /* Undo the pollution */ - -typedef __vector unsigned long long xxh_u64x2; -typedef __vector unsigned char xxh_u8x16; -typedef __vector unsigned xxh_u32x4; - -# ifndef XXH_VSX_BE -# if defined(__BIG_ENDIAN__) \ - || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) -# define XXH_VSX_BE 1 -# elif defined(__VEC_ELEMENT_REG_ORDER__) && __VEC_ELEMENT_REG_ORDER__ == __ORDER_BIG_ENDIAN__ -# warning "-maltivec=be is not recommended. Please use native endianness." -# define XXH_VSX_BE 1 -# else -# define XXH_VSX_BE 0 -# endif -# endif /* !defined(XXH_VSX_BE) */ - -# if XXH_VSX_BE -/* A wrapper for POWER9's vec_revb. */ -# if defined(__POWER9_VECTOR__) || (defined(__clang__) && defined(__s390x__)) -# define XXH_vec_revb vec_revb -# else -XXH_FORCE_INLINE xxh_u64x2 XXH_vec_revb(xxh_u64x2 val) -{ - xxh_u8x16 const vByteSwap = { 0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00, - 0x0F, 0x0E, 0x0D, 0x0C, 0x0B, 0x0A, 0x09, 0x08 }; - return vec_perm(val, val, vByteSwap); -} -# endif -# endif /* XXH_VSX_BE */ - -/* - * Performs an unaligned load and byte swaps it on big endian. - */ -XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(const void *ptr) -{ - xxh_u64x2 ret; - memcpy(&ret, ptr, sizeof(xxh_u64x2)); -# if XXH_VSX_BE - ret = XXH_vec_revb(ret); -# endif - return ret; -} - -/* - * vec_mulo and vec_mule are very problematic intrinsics on PowerPC - * - * These intrinsics weren't added until GCC 8, despite existing for a while, - * and they are endian dependent. Also, their meaning swap depending on version. - * */ -# if defined(__s390x__) - /* s390x is always big endian, no issue on this platform */ -# define XXH_vec_mulo vec_mulo -# define XXH_vec_mule vec_mule -# elif defined(__clang__) && __has_builtin(__builtin_altivec_vmuleuw) -/* Clang has a better way to control this, we can just use the builtin which doesn't swap. */ -# define XXH_vec_mulo __builtin_altivec_vmulouw -# define XXH_vec_mule __builtin_altivec_vmuleuw -# else -/* gcc needs inline assembly */ -/* Adapted from https://github.com/google/highwayhash/blob/master/highwayhash/hh_vsx.h. */ -XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mulo(xxh_u32x4 a, xxh_u32x4 b) -{ - xxh_u64x2 result; - __asm__("vmulouw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b)); - return result; -} -XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mule(xxh_u32x4 a, xxh_u32x4 b) -{ - xxh_u64x2 result; - __asm__("vmuleuw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b)); - return result; -} -# endif /* XXH_vec_mulo, XXH_vec_mule */ -#endif /* XXH_VECTOR == XXH_VSX */ - - -/* prefetch - * can be disabled, by declaring XXH_NO_PREFETCH build macro */ -#if defined(XXH_NO_PREFETCH) -# define XXH_PREFETCH(ptr) (void)(ptr) /* disabled */ -#else -# if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_I86)) /* _mm_prefetch() is not defined outside of x86/x64 */ -# include <mmintrin.h> /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */ -# define XXH_PREFETCH(ptr) _mm_prefetch((const char*)(ptr), _MM_HINT_T0) -# elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) ) -# define XXH_PREFETCH(ptr) __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */) -# else -# define XXH_PREFETCH(ptr) (void)(ptr) /* disabled */ -# endif -#endif /* XXH_NO_PREFETCH */ - - -/* ========================================== - * XXH3 default settings - * ========================================== */ - -#define XXH_SECRET_DEFAULT_SIZE 192 /* minimum XXH3_SECRET_SIZE_MIN */ - -#if (XXH_SECRET_DEFAULT_SIZE < XXH3_SECRET_SIZE_MIN) -# error "default keyset is not large enough" -#endif - -/* Pseudorandom secret taken directly from FARSH */ -XXH_ALIGN(64) static const xxh_u8 kSecret[XXH_SECRET_DEFAULT_SIZE] = { - 0xb8, 0xfe, 0x6c, 0x39, 0x23, 0xa4, 0x4b, 0xbe, 0x7c, 0x01, 0x81, 0x2c, 0xf7, 0x21, 0xad, 0x1c, - 0xde, 0xd4, 0x6d, 0xe9, 0x83, 0x90, 0x97, 0xdb, 0x72, 0x40, 0xa4, 0xa4, 0xb7, 0xb3, 0x67, 0x1f, - 0xcb, 0x79, 0xe6, 0x4e, 0xcc, 0xc0, 0xe5, 0x78, 0x82, 0x5a, 0xd0, 0x7d, 0xcc, 0xff, 0x72, 0x21, - 0xb8, 0x08, 0x46, 0x74, 0xf7, 0x43, 0x24, 0x8e, 0xe0, 0x35, 0x90, 0xe6, 0x81, 0x3a, 0x26, 0x4c, - 0x3c, 0x28, 0x52, 0xbb, 0x91, 0xc3, 0x00, 0xcb, 0x88, 0xd0, 0x65, 0x8b, 0x1b, 0x53, 0x2e, 0xa3, - 0x71, 0x64, 0x48, 0x97, 0xa2, 0x0d, 0xf9, 0x4e, 0x38, 0x19, 0xef, 0x46, 0xa9, 0xde, 0xac, 0xd8, - 0xa8, 0xfa, 0x76, 0x3f, 0xe3, 0x9c, 0x34, 0x3f, 0xf9, 0xdc, 0xbb, 0xc7, 0xc7, 0x0b, 0x4f, 0x1d, - 0x8a, 0x51, 0xe0, 0x4b, 0xcd, 0xb4, 0x59, 0x31, 0xc8, 0x9f, 0x7e, 0xc9, 0xd9, 0x78, 0x73, 0x64, - - 0xea, 0xc5, 0xac, 0x83, 0x34, 0xd3, 0xeb, 0xc3, 0xc5, 0x81, 0xa0, 0xff, 0xfa, 0x13, 0x63, 0xeb, - 0x17, 0x0d, 0xdd, 0x51, 0xb7, 0xf0, 0xda, 0x49, 0xd3, 0x16, 0x55, 0x26, 0x29, 0xd4, 0x68, 0x9e, - 0x2b, 0x16, 0xbe, 0x58, 0x7d, 0x47, 0xa1, 0xfc, 0x8f, 0xf8, 0xb8, 0xd1, 0x7a, 0xd0, 0x31, 0xce, - 0x45, 0xcb, 0x3a, 0x8f, 0x95, 0x16, 0x04, 0x28, 0xaf, 0xd7, 0xfb, 0xca, 0xbb, 0x4b, 0x40, 0x7e, -}; - -/* - * Does a 32-bit to 64-bit long multiply. - * - * Wraps __emulu on MSVC x86 because it tends to call __allmul when it doesn't - * need to (but it shouldn't need to anyways, it is about 7 instructions to do - * a 64x64 multiply...). Since we know that this will _always_ emit MULL, we - * use that instead of the normal method. - * - * If you are compiling for platforms like Thumb-1 and don't have a better option, - * you may also want to write your own long multiply routine here. - * - * XXH_FORCE_INLINE xxh_u64 XXH_mult32to64(xxh_u64 x, xxh_u64 y) - * { - * return (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF); - * } - */ -#if defined(_MSC_VER) && defined(_M_IX86) -# include <intrin.h> -# define XXH_mult32to64(x, y) __emulu((unsigned)(x), (unsigned)(y)) -#else -/* - * Downcast + upcast is usually better than masking on older compilers like - * GCC 4.2 (especially 32-bit ones), all without affecting newer compilers. - * - * The other method, (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF), will AND both operands - * and perform a full 64x64 multiply -- entirely redundant on 32-bit. - */ -# define XXH_mult32to64(x, y) ((xxh_u64)(xxh_u32)(x) * (xxh_u64)(xxh_u32)(y)) -#endif - -/* - * Calculates a 64->128-bit long multiply. - * - * Uses __uint128_t and _umul128 if available, otherwise uses a scalar version. - */ -static XXH128_hash_t -XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs) -{ - /* - * GCC/Clang __uint128_t method. - * - * On most 64-bit targets, GCC and Clang define a __uint128_t type. - * This is usually the best way as it usually uses a native long 64-bit - * multiply, such as MULQ on x86_64 or MUL + UMULH on aarch64. - * - * Usually. - * - * Despite being a 32-bit platform, Clang (and emscripten) define this type - * despite not having the arithmetic for it. This results in a laggy - * compiler builtin call which calculates a full 128-bit multiply. - * In that case it is best to use the portable one. - * https://github.com/Cyan4973/xxHash/issues/211#issuecomment-515575677 - */ -#if defined(__GNUC__) && !defined(__wasm__) \ - && defined(__SIZEOF_INT128__) \ - || (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128) - - __uint128_t product = (__uint128_t)lhs * (__uint128_t)rhs; - XXH128_hash_t const r128 = { (xxh_u64)(product), (xxh_u64)(product >> 64) }; - return r128; - - /* - * MSVC for x64's _umul128 method. - * - * xxh_u64 _umul128(xxh_u64 Multiplier, xxh_u64 Multiplicand, xxh_u64 *HighProduct); - * - * This compiles to single operand MUL on x64. - */ -#elif defined(_M_X64) || defined(_M_IA64) - -#ifndef _MSC_VER -# pragma intrinsic(_umul128) -#endif - xxh_u64 product_high; - xxh_u64 const product_low = _umul128(lhs, rhs, &product_high); - XXH128_hash_t const r128 = { product_low, product_high }; - return r128; - -#else - /* - * Portable scalar method. Optimized for 32-bit and 64-bit ALUs. - * - * This is a fast and simple grade school multiply, which is shown below - * with base 10 arithmetic instead of base 0x100000000. - * - * 9 3 // D2 lhs = 93 - * x 7 5 // D2 rhs = 75 - * ---------- - * 1 5 // D2 lo_lo = (93 % 10) * (75 % 10) = 15 - * 4 5 | // D2 hi_lo = (93 / 10) * (75 % 10) = 45 - * 2 1 | // D2 lo_hi = (93 % 10) * (75 / 10) = 21 - * + 6 3 | | // D2 hi_hi = (93 / 10) * (75 / 10) = 63 - * --------- - * 2 7 | // D2 cross = (15 / 10) + (45 % 10) + 21 = 27 - * + 6 7 | | // D2 upper = (27 / 10) + (45 / 10) + 63 = 67 - * --------- - * 6 9 7 5 // D4 res = (27 * 10) + (15 % 10) + (67 * 100) = 6975 - * - * The reasons for adding the products like this are: - * 1. It avoids manual carry tracking. Just like how - * (9 * 9) + 9 + 9 = 99, the same applies with this for UINT64_MAX. - * This avoids a lot of complexity. - * - * 2. It hints for, and on Clang, compiles to, the powerful UMAAL - * instruction available in ARM's Digital Signal Processing extension - * in 32-bit ARMv6 and later, which is shown below: - * - * void UMAAL(xxh_u32 *RdLo, xxh_u32 *RdHi, xxh_u32 Rn, xxh_u32 Rm) - * { - * xxh_u64 product = (xxh_u64)*RdLo * (xxh_u64)*RdHi + Rn + Rm; - * *RdLo = (xxh_u32)(product & 0xFFFFFFFF); - * *RdHi = (xxh_u32)(product >> 32); - * } - * - * This instruction was designed for efficient long multiplication, and - * allows this to be calculated in only 4 instructions at speeds - * comparable to some 64-bit ALUs. - * - * 3. It isn't terrible on other platforms. Usually this will be a couple - * of 32-bit ADD/ADCs. - */ - - /* First calculate all of the cross products. */ - xxh_u64 const lo_lo = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs & 0xFFFFFFFF); - xxh_u64 const hi_lo = XXH_mult32to64(lhs >> 32, rhs & 0xFFFFFFFF); - xxh_u64 const lo_hi = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs >> 32); - xxh_u64 const hi_hi = XXH_mult32to64(lhs >> 32, rhs >> 32); - - /* Now add the products together. These will never overflow. */ - xxh_u64 const cross = (lo_lo >> 32) + (hi_lo & 0xFFFFFFFF) + lo_hi; - xxh_u64 const upper = (hi_lo >> 32) + (cross >> 32) + hi_hi; - xxh_u64 const lower = (cross << 32) | (lo_lo & 0xFFFFFFFF); - - XXH128_hash_t r128 = { lower, upper }; - return r128; -#endif -} - -/* - * Does a 64-bit to 128-bit multiply, then XOR folds it. - * - * The reason for the separate function is to prevent passing too many structs - * around by value. This will hopefully inline the multiply, but we don't force it. - */ -static xxh_u64 -XXH3_mul128_fold64(xxh_u64 lhs, xxh_u64 rhs) -{ - XXH128_hash_t product = XXH_mult64to128(lhs, rhs); - return product.low64 ^ product.high64; -} - -/* Seems to produce slightly better code on GCC for some reason. */ -XXH_FORCE_INLINE xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift) -{ - XXH_ASSERT(0 <= shift && shift < 64); - return v64 ^ (v64 >> shift); -} - -/* - * We don't need to (or want to) mix as much as XXH64. - * - * Short hashes are more evenly distributed, so it isn't necessary. - */ -static XXH64_hash_t XXH3_avalanche(xxh_u64 h64) -{ - h64 = XXH_xorshift64(h64, 37); - h64 *= 0x165667919E3779F9ULL; - h64 = XXH_xorshift64(h64, 32); - return h64; -} - - -/* ========================================== - * Short keys - * ========================================== - * One of the shortcomings of XXH32 and XXH64 was that their performance was - * sub-optimal on short lengths. It used an iterative algorithm which strongly - * favored lengths that were a multiple of 4 or 8. - * - * Instead of iterating over individual inputs, we use a set of single shot - * functions which piece together a range of lengths and operate in constant time. + * xxh3.h is no longer useful, + * but it is still provided for compatibility with source code + * which used to include it directly. * - * Additionally, the number of multiplies has been significantly reduced. This - * reduces latency, especially when emulating 64-bit multiplies on 32-bit. + * Programs are now highly discourage to include xxh3.h. + * Include `xxhash.h` instead, which is the officially supported interface. * - * Depending on the platform, this may or may not be faster than XXH32, but it - * is almost guaranteed to be faster than XXH64. + * In the future, xxh3.h will start to generate warnings, then errors, + * then it will be removed from source package and from include directory. */ -/* - * At very short lengths, there isn't enough input to fully hide secrets, or use - * the entire secret. - * - * There is also only a limited amount of mixing we can do before significantly - * impacting performance. - * - * Therefore, we use different sections of the secret and always mix two secret - * samples with an XOR. This should have no effect on performance on the - * seedless or withSeed variants because everything _should_ be constant folded - * by modern compilers. - * - * The XOR mixing hides individual parts of the secret and increases entropy. - * - * This adds an extra layer of strength for custom secrets. - */ -XXH_FORCE_INLINE XXH64_hash_t -XXH3_len_1to3_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) -{ - XXH_ASSERT(input != NULL); - XXH_ASSERT(1 <= len && len <= 3); - XXH_ASSERT(secret != NULL); - /* - * len = 1: combined = { input[0], 0x01, input[0], input[0] } - * len = 2: combined = { input[1], 0x02, input[0], input[1] } - * len = 3: combined = { input[2], 0x03, input[0], input[1] } - */ - { xxh_u8 const c1 = input[0]; - xxh_u8 const c2 = input[len >> 1]; - xxh_u8 const c3 = input[len - 1]; - xxh_u32 const combined = ((xxh_u32)c1<<16) | (((xxh_u32)c2) << 24) | (((xxh_u32)c3) << 0) | (((xxh_u32)len) << 8); - xxh_u64 const bitflip = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed; - xxh_u64 const keyed = (xxh_u64)combined ^ bitflip; - xxh_u64 const mixed = keyed * PRIME64_1; - return XXH3_avalanche(mixed); - } -} - -XXH_FORCE_INLINE XXH64_hash_t -XXH3_len_4to8_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) -{ - XXH_ASSERT(input != NULL); - XXH_ASSERT(secret != NULL); - XXH_ASSERT(4 <= len && len < 8); - seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32; - { xxh_u32 const input1 = XXH_readLE32(input); - xxh_u32 const input2 = XXH_readLE32(input + len - 4); - xxh_u64 const bitflip = (XXH_readLE64(secret+8) ^ XXH_readLE64(secret+16)) - seed; - xxh_u64 const input64 = input2 + (((xxh_u64)input1) << 32); - xxh_u64 x = input64 ^ bitflip; - /* this mix is inspired by Pelle Evensen's rrmxmx */ - x ^= XXH_rotl64(x, 49) ^ XXH_rotl64(x, 24); - x *= 0x9FB21C651E98DF25ULL; - x ^= (x >> 35) + len ; - x *= 0x9FB21C651E98DF25ULL; - return XXH_xorshift64(x, 28); - } -} - -XXH_FORCE_INLINE XXH64_hash_t -XXH3_len_9to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) -{ - XXH_ASSERT(input != NULL); - XXH_ASSERT(secret != NULL); - XXH_ASSERT(8 <= len && len <= 16); - { xxh_u64 const bitflip1 = (XXH_readLE64(secret+24) ^ XXH_readLE64(secret+32)) + seed; - xxh_u64 const bitflip2 = (XXH_readLE64(secret+40) ^ XXH_readLE64(secret+48)) - seed; - xxh_u64 const input_lo = XXH_readLE64(input) ^ bitflip1; - xxh_u64 const input_hi = XXH_readLE64(input + len - 8) ^ bitflip2; - xxh_u64 const acc = len - + XXH_swap64(input_lo) + input_hi - + XXH3_mul128_fold64(input_lo, input_hi); - return XXH3_avalanche(acc); - } -} - -XXH_FORCE_INLINE XXH64_hash_t -XXH3_len_0to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) -{ - XXH_ASSERT(len <= 16); - { if (XXH_likely(len > 8)) return XXH3_len_9to16_64b(input, len, secret, seed); - if (XXH_likely(len >= 4)) return XXH3_len_4to8_64b(input, len, secret, seed); - if (len) return XXH3_len_1to3_64b(input, len, secret, seed); - return XXH3_avalanche((PRIME64_1 + seed) ^ (XXH_readLE64(secret+56) ^ XXH_readLE64(secret+64))); - } -} - -/* - * DISCLAIMER: There are known *seed-dependent* multicollisions here due to - * multiplication by zero, affecting hashes of lengths 17 to 240. - * - * However, they are very unlikely. - * - * Keep this in mind when using the unseeded XXH3_64bits() variant: As with all - * unseeded non-cryptographic hashes, it does not attempt to defend itself - * against specially crafted inputs, only random inputs. - * - * Compared to classic UMAC where a 1 in 2^31 chance of 4 consecutive bytes - * cancelling out the secret is taken an arbitrary number of times (addressed - * in XXH3_accumulate_512), this collision is very unlikely with random inputs - * and/or proper seeding: - * - * This only has a 1 in 2^63 chance of 8 consecutive bytes cancelling out, in a - * function that is only called up to 16 times per hash with up to 240 bytes of - * input. - * - * This is not too bad for a non-cryptographic hash function, especially with - * only 64 bit outputs. - * - * The 128-bit variant (which trades some speed for strength) is NOT affected - * by this, although it is always a good idea to use a proper seed if you care - * about strength. - */ -XXH_FORCE_INLINE xxh_u64 XXH3_mix16B(const xxh_u8* XXH_RESTRICT input, - const xxh_u8* XXH_RESTRICT secret, xxh_u64 seed64) -{ -#if defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \ - && defined(__i386__) && defined(__SSE2__) /* x86 + SSE2 */ \ - && !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable like XXH32 hack */ - /* - * UGLY HACK: - * GCC for x86 tends to autovectorize the 128-bit multiply, resulting in - * slower code. - * - * By forcing seed64 into a register, we disrupt the cost model and - * cause it to scalarize. See `XXH32_round()` - * - * FIXME: Clang's output is still _much_ faster -- On an AMD Ryzen 3600, - * XXH3_64bits @ len=240 runs at 4.6 GB/s with Clang 9, but 3.3 GB/s on - * GCC 9.2, despite both emitting scalar code. - * - * GCC generates much better scalar code than Clang for the rest of XXH3, - * which is why finding a more optimal codepath is an interest. - */ - __asm__ ("" : "+r" (seed64)); -#endif - { xxh_u64 const input_lo = XXH_readLE64(input); - xxh_u64 const input_hi = XXH_readLE64(input+8); - return XXH3_mul128_fold64( - input_lo ^ (XXH_readLE64(secret) + seed64), - input_hi ^ (XXH_readLE64(secret+8) - seed64) - ); - } -} +/* Simulate the same impact as including the old xxh3.h source file */ -/* For mid range keys, XXH3 uses a Mum-hash variant. */ -XXH_FORCE_INLINE XXH64_hash_t -XXH3_len_17to128_64b(const xxh_u8* XXH_RESTRICT input, size_t len, - const xxh_u8* XXH_RESTRICT secret, size_t secretSize, - XXH64_hash_t seed) -{ - XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize; - XXH_ASSERT(16 < len && len <= 128); - - { xxh_u64 acc = len * PRIME64_1; - if (len > 32) { - if (len > 64) { - if (len > 96) { - acc += XXH3_mix16B(input+48, secret+96, seed); - acc += XXH3_mix16B(input+len-64, secret+112, seed); - } - acc += XXH3_mix16B(input+32, secret+64, seed); - acc += XXH3_mix16B(input+len-48, secret+80, seed); - } - acc += XXH3_mix16B(input+16, secret+32, seed); - acc += XXH3_mix16B(input+len-32, secret+48, seed); - } - acc += XXH3_mix16B(input+0, secret+0, seed); - acc += XXH3_mix16B(input+len-16, secret+16, seed); - - return XXH3_avalanche(acc); - } -} - -#define XXH3_MIDSIZE_MAX 240 - -XXH_NO_INLINE XXH64_hash_t -XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len, - const xxh_u8* XXH_RESTRICT secret, size_t secretSize, - XXH64_hash_t seed) -{ - XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize; - XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX); - - #define XXH3_MIDSIZE_STARTOFFSET 3 - #define XXH3_MIDSIZE_LASTOFFSET 17 - - { xxh_u64 acc = len * PRIME64_1; - int const nbRounds = (int)len / 16; - int i; - for (i=0; i<8; i++) { - acc += XXH3_mix16B(input+(16*i), secret+(16*i), seed); - } - acc = XXH3_avalanche(acc); - XXH_ASSERT(nbRounds >= 8); -#if defined(__clang__) /* Clang */ \ - && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \ - && !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable */ - /* - * UGLY HACK: - * Clang for ARMv7-A tries to vectorize this loop, similar to GCC x86. - * In everywhere else, it uses scalar code. - * - * For 64->128-bit multiplies, even if the NEON was 100% optimal, it - * would still be slower than UMAAL (see XXH_mult64to128). - * - * Unfortunately, Clang doesn't handle the long multiplies properly and - * converts them to the nonexistent "vmulq_u64" intrinsic, which is then - * scalarized into an ugly mess of VMOV.32 instructions. - * - * This mess is difficult to avoid without turning autovectorization - * off completely, but they are usually relatively minor and/or not - * worth it to fix. - * - * This loop is the easiest to fix, as unlike XXH32, this pragma - * _actually works_ because it is a loop vectorization instead of an - * SLP vectorization. - */ - #pragma clang loop vectorize(disable) -#endif - for (i=8 ; i < nbRounds; i++) { - acc += XXH3_mix16B(input+(16*i), secret+(16*(i-8)) + XXH3_MIDSIZE_STARTOFFSET, seed); - } - /* last bytes */ - acc += XXH3_mix16B(input + len - 16, secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, seed); - return XXH3_avalanche(acc); - } -} - - -/* === Long Keys === */ - -#define STRIPE_LEN 64 -#define XXH_SECRET_CONSUME_RATE 8 /* nb of secret bytes consumed at each accumulation */ -#define ACC_NB (STRIPE_LEN / sizeof(xxh_u64)) - -typedef enum { XXH3_acc_64bits, XXH3_acc_128bits } XXH3_accWidth_e; - -/* - * XXH3_accumulate_512 is the tightest loop for long inputs, and it is the most optimized. - * - * It is a hardened version of UMAC, based off of FARSH's implementation. - * - * This was chosen because it adapts quite well to 32-bit, 64-bit, and SIMD - * implementations, and it is ridiculously fast. - * - * We harden it by mixing the original input to the accumulators as well as the product. - * - * This means that in the (relatively likely) case of a multiply by zero, the - * original input is preserved. - * - * On 128-bit inputs, we swap 64-bit pairs when we add the input to improve - * cross-pollination, as otherwise the upper and lower halves would be - * essentially independent. - * - * This doesn't matter on 64-bit hashes since they all get merged together in - * the end, so we skip the extra step. - * - * Both XXH3_64bits and XXH3_128bits use this subroutine. - */ -XXH_FORCE_INLINE void -XXH3_accumulate_512( void* XXH_RESTRICT acc, - const void* XXH_RESTRICT input, - const void* XXH_RESTRICT secret, - XXH3_accWidth_e accWidth) -{ -#if (XXH_VECTOR == XXH_AVX2) - - XXH_ASSERT((((size_t)acc) & 31) == 0); - { XXH_ALIGN(32) __m256i* const xacc = (__m256i *) acc; - /* Unaligned. This is mainly for pointer arithmetic, and because - * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */ - const __m256i* const xinput = (const __m256i *) input; - /* Unaligned. This is mainly for pointer arithmetic, and because - * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */ - const __m256i* const xsecret = (const __m256i *) secret; - - size_t i; - for (i=0; i < STRIPE_LEN/sizeof(__m256i); i++) { - /* data_vec = xinput[i]; */ - __m256i const data_vec = _mm256_loadu_si256 (xinput+i); - /* key_vec = xsecret[i]; */ - __m256i const key_vec = _mm256_loadu_si256 (xsecret+i); - /* data_key = data_vec ^ key_vec; */ - __m256i const data_key = _mm256_xor_si256 (data_vec, key_vec); - /* data_key_lo = data_key >> 32; */ - __m256i const data_key_lo = _mm256_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1)); - /* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */ - __m256i const product = _mm256_mul_epu32 (data_key, data_key_lo); - if (accWidth == XXH3_acc_128bits) { - /* xacc[i] += swap(data_vec); */ - __m256i const data_swap = _mm256_shuffle_epi32(data_vec, _MM_SHUFFLE(1, 0, 3, 2)); - __m256i const sum = _mm256_add_epi64(xacc[i], data_swap); - /* xacc[i] += product; */ - xacc[i] = _mm256_add_epi64(product, sum); - } else { /* XXH3_acc_64bits */ - /* xacc[i] += data_vec; */ - __m256i const sum = _mm256_add_epi64(xacc[i], data_vec); - /* xacc[i] += product; */ - xacc[i] = _mm256_add_epi64(product, sum); - } - } } - -#elif (XXH_VECTOR == XXH_SSE2) - - /* SSE2 is just a half-scale version of the AVX2 version. */ - XXH_ASSERT((((size_t)acc) & 15) == 0); - { XXH_ALIGN(16) __m128i* const xacc = (__m128i *) acc; - /* Unaligned. This is mainly for pointer arithmetic, and because - * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */ - const __m128i* const xinput = (const __m128i *) input; - /* Unaligned. This is mainly for pointer arithmetic, and because - * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */ - const __m128i* const xsecret = (const __m128i *) secret; - - size_t i; - for (i=0; i < STRIPE_LEN/sizeof(__m128i); i++) { - /* data_vec = xinput[i]; */ - __m128i const data_vec = _mm_loadu_si128 (xinput+i); - /* key_vec = xsecret[i]; */ - __m128i const key_vec = _mm_loadu_si128 (xsecret+i); - /* data_key = data_vec ^ key_vec; */ - __m128i const data_key = _mm_xor_si128 (data_vec, key_vec); - /* data_key_lo = data_key >> 32; */ - __m128i const data_key_lo = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1)); - /* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */ - __m128i const product = _mm_mul_epu32 (data_key, data_key_lo); - if (accWidth == XXH3_acc_128bits) { - /* xacc[i] += swap(data_vec); */ - __m128i const data_swap = _mm_shuffle_epi32(data_vec, _MM_SHUFFLE(1,0,3,2)); - __m128i const sum = _mm_add_epi64(xacc[i], data_swap); - /* xacc[i] += product; */ - xacc[i] = _mm_add_epi64(product, sum); - } else { /* XXH3_acc_64bits */ - /* xacc[i] += data_vec; */ - __m128i const sum = _mm_add_epi64(xacc[i], data_vec); - /* xacc[i] += product; */ - xacc[i] = _mm_add_epi64(product, sum); - } - } } - -#elif (XXH_VECTOR == XXH_NEON) - - XXH_ASSERT((((size_t)acc) & 15) == 0); - { - XXH_ALIGN(16) uint64x2_t* const xacc = (uint64x2_t *) acc; - /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. */ - uint8_t const* const xinput = (const uint8_t *) input; - uint8_t const* const xsecret = (const uint8_t *) secret; - - size_t i; - for (i=0; i < STRIPE_LEN / sizeof(uint64x2_t); i++) { - /* data_vec = xinput[i]; */ - uint8x16_t data_vec = vld1q_u8(xinput + (i * 16)); - /* key_vec = xsecret[i]; */ - uint8x16_t key_vec = vld1q_u8(xsecret + (i * 16)); - /* data_key = data_vec ^ key_vec; */ - uint64x2_t data_key = vreinterpretq_u64_u8(veorq_u8(data_vec, key_vec)); - uint32x2_t data_key_lo, data_key_hi; - if (accWidth == XXH3_acc_64bits) { - /* xacc[i] += data_vec; */ - xacc[i] = vaddq_u64 (xacc[i], vreinterpretq_u64_u8(data_vec)); - } else { /* XXH3_acc_128bits */ - /* xacc[i] += swap(data_vec); */ - uint64x2_t const data64 = vreinterpretq_u64_u8(data_vec); - uint64x2_t const swapped = vextq_u64(data64, data64, 1); - xacc[i] = vaddq_u64 (xacc[i], swapped); - } - /* data_key_lo = (uint32x2_t) (data_key & 0xFFFFFFFF); - * data_key_hi = (uint32x2_t) (data_key >> 32); - * data_key = UNDEFINED; */ - XXH_SPLIT_IN_PLACE(data_key, data_key_lo, data_key_hi); - /* xacc[i] += (uint64x2_t) data_key_lo * (uint64x2_t) data_key_hi; */ - xacc[i] = vmlal_u32 (xacc[i], data_key_lo, data_key_hi); - - } - } - -#elif (XXH_VECTOR == XXH_VSX) - xxh_u64x2* const xacc = (xxh_u64x2*) acc; /* presumed aligned */ - xxh_u64x2 const* const xinput = (xxh_u64x2 const*) input; /* no alignment restriction */ - xxh_u64x2 const* const xsecret = (xxh_u64x2 const*) secret; /* no alignment restriction */ - xxh_u64x2 const v32 = { 32, 32 }; - size_t i; - for (i = 0; i < STRIPE_LEN / sizeof(xxh_u64x2); i++) { - /* data_vec = xinput[i]; */ - xxh_u64x2 const data_vec = XXH_vec_loadu(xinput + i); - /* key_vec = xsecret[i]; */ - xxh_u64x2 const key_vec = XXH_vec_loadu(xsecret + i); - xxh_u64x2 const data_key = data_vec ^ key_vec; - /* shuffled = (data_key << 32) | (data_key >> 32); */ - xxh_u32x4 const shuffled = (xxh_u32x4)vec_rl(data_key, v32); - /* product = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)shuffled & 0xFFFFFFFF); */ - xxh_u64x2 const product = XXH_vec_mulo((xxh_u32x4)data_key, shuffled); - xacc[i] += product; - - if (accWidth == XXH3_acc_64bits) { - xacc[i] += data_vec; - } else { /* XXH3_acc_128bits */ - /* swap high and low halves */ -#ifdef __s390x__ - xxh_u64x2 const data_swapped = vec_permi(data_vec, data_vec, 2); -#else - xxh_u64x2 const data_swapped = vec_xxpermdi(data_vec, data_vec, 2); -#endif - xacc[i] += data_swapped; - } - } - -#else /* scalar variant of Accumulator - universal */ - - XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64* const xacc = (xxh_u64*) acc; /* presumed aligned */ - const xxh_u8* const xinput = (const xxh_u8*) input; /* no alignment restriction */ - const xxh_u8* const xsecret = (const xxh_u8*) secret; /* no alignment restriction */ - size_t i; - XXH_ASSERT(((size_t)acc & (XXH_ACC_ALIGN-1)) == 0); - for (i=0; i < ACC_NB; i++) { - xxh_u64 const data_val = XXH_readLE64(xinput + 8*i); - xxh_u64 const data_key = data_val ^ XXH_readLE64(xsecret + i*8); - - if (accWidth == XXH3_acc_64bits) { - xacc[i] += data_val; - } else { - xacc[i ^ 1] += data_val; /* swap adjacent lanes */ - } - xacc[i] += XXH_mult32to64(data_key & 0xFFFFFFFF, data_key >> 32); - } -#endif -} - -/* - * XXH3_scrambleAcc: Scrambles the accumulators to improve mixing. - * - * Multiplication isn't perfect, as explained by Google in HighwayHash: - * - * // Multiplication mixes/scrambles bytes 0-7 of the 64-bit result to - * // varying degrees. In descending order of goodness, bytes - * // 3 4 2 5 1 6 0 7 have quality 228 224 164 160 100 96 36 32. - * // As expected, the upper and lower bytes are much worse. - * - * Source: https://github.com/google/highwayhash/blob/0aaf66b/highwayhash/hh_avx2.h#L291 - * - * Since our algorithm uses a pseudorandom secret to add some variance into the - * mix, we don't need to (or want to) mix as often or as much as HighwayHash does. - * - * This isn't as tight as XXH3_accumulate, but still written in SIMD to avoid - * extraction. - * - * Both XXH3_64bits and XXH3_128bits use this subroutine. - */ -XXH_FORCE_INLINE void -XXH3_scrambleAcc(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret) -{ -#if (XXH_VECTOR == XXH_AVX2) - - XXH_ASSERT((((size_t)acc) & 31) == 0); - { XXH_ALIGN(32) __m256i* const xacc = (__m256i*) acc; - /* Unaligned. This is mainly for pointer arithmetic, and because - * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */ - const __m256i* const xsecret = (const __m256i *) secret; - const __m256i prime32 = _mm256_set1_epi32((int)PRIME32_1); - - size_t i; - for (i=0; i < STRIPE_LEN/sizeof(__m256i); i++) { - /* xacc[i] ^= (xacc[i] >> 47) */ - __m256i const acc_vec = xacc[i]; - __m256i const shifted = _mm256_srli_epi64 (acc_vec, 47); - __m256i const data_vec = _mm256_xor_si256 (acc_vec, shifted); - /* xacc[i] ^= xsecret; */ - __m256i const key_vec = _mm256_loadu_si256 (xsecret+i); - __m256i const data_key = _mm256_xor_si256 (data_vec, key_vec); - - /* xacc[i] *= PRIME32_1; */ - __m256i const data_key_hi = _mm256_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1)); - __m256i const prod_lo = _mm256_mul_epu32 (data_key, prime32); - __m256i const prod_hi = _mm256_mul_epu32 (data_key_hi, prime32); - xacc[i] = _mm256_add_epi64(prod_lo, _mm256_slli_epi64(prod_hi, 32)); - } - } - -#elif (XXH_VECTOR == XXH_SSE2) - - XXH_ASSERT((((size_t)acc) & 15) == 0); - { XXH_ALIGN(16) __m128i* const xacc = (__m128i*) acc; - /* Unaligned. This is mainly for pointer arithmetic, and because - * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */ - const __m128i* const xsecret = (const __m128i *) secret; - const __m128i prime32 = _mm_set1_epi32((int)PRIME32_1); - - size_t i; - for (i=0; i < STRIPE_LEN/sizeof(__m128i); i++) { - /* xacc[i] ^= (xacc[i] >> 47) */ - __m128i const acc_vec = xacc[i]; - __m128i const shifted = _mm_srli_epi64 (acc_vec, 47); - __m128i const data_vec = _mm_xor_si128 (acc_vec, shifted); - /* xacc[i] ^= xsecret[i]; */ - __m128i const key_vec = _mm_loadu_si128 (xsecret+i); - __m128i const data_key = _mm_xor_si128 (data_vec, key_vec); - - /* xacc[i] *= PRIME32_1; */ - __m128i const data_key_hi = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1)); - __m128i const prod_lo = _mm_mul_epu32 (data_key, prime32); - __m128i const prod_hi = _mm_mul_epu32 (data_key_hi, prime32); - xacc[i] = _mm_add_epi64(prod_lo, _mm_slli_epi64(prod_hi, 32)); - } - } - -#elif (XXH_VECTOR == XXH_NEON) - - XXH_ASSERT((((size_t)acc) & 15) == 0); - - { uint64x2_t* xacc = (uint64x2_t*) acc; - uint8_t const* xsecret = (uint8_t const*) secret; - uint32x2_t prime = vdup_n_u32 (PRIME32_1); - - size_t i; - for (i=0; i < STRIPE_LEN/sizeof(uint64x2_t); i++) { - /* xacc[i] ^= (xacc[i] >> 47); */ - uint64x2_t acc_vec = xacc[i]; - uint64x2_t shifted = vshrq_n_u64 (acc_vec, 47); - uint64x2_t data_vec = veorq_u64 (acc_vec, shifted); - - /* xacc[i] ^= xsecret[i]; */ - uint8x16_t key_vec = vld1q_u8(xsecret + (i * 16)); - uint64x2_t data_key = veorq_u64(data_vec, vreinterpretq_u64_u8(key_vec)); - - /* xacc[i] *= PRIME32_1 */ - uint32x2_t data_key_lo, data_key_hi; - /* data_key_lo = (uint32x2_t) (xacc[i] & 0xFFFFFFFF); - * data_key_hi = (uint32x2_t) (xacc[i] >> 32); - * xacc[i] = UNDEFINED; */ - XXH_SPLIT_IN_PLACE(data_key, data_key_lo, data_key_hi); - { /* - * prod_hi = (data_key >> 32) * PRIME32_1; - * - * Avoid vmul_u32 + vshll_n_u32 since Clang 6 and 7 will - * incorrectly "optimize" this: - * tmp = vmul_u32(vmovn_u64(a), vmovn_u64(b)); - * shifted = vshll_n_u32(tmp, 32); - * to this: - * tmp = "vmulq_u64"(a, b); // no such thing! - * shifted = vshlq_n_u64(tmp, 32); - * - * However, unlike SSE, Clang lacks a 64-bit multiply routine - * for NEON, and it scalarizes two 64-bit multiplies instead. - * - * vmull_u32 has the same timing as vmul_u32, and it avoids - * this bug completely. - * See https://bugs.llvm.org/show_bug.cgi?id=39967 - */ - uint64x2_t prod_hi = vmull_u32 (data_key_hi, prime); - /* xacc[i] = prod_hi << 32; */ - xacc[i] = vshlq_n_u64(prod_hi, 32); - /* xacc[i] += (prod_hi & 0xFFFFFFFF) * PRIME32_1; */ - xacc[i] = vmlal_u32(xacc[i], data_key_lo, prime); - } - } } - -#elif (XXH_VECTOR == XXH_VSX) - - XXH_ASSERT((((size_t)acc) & 15) == 0); - - { xxh_u64x2* const xacc = (xxh_u64x2*) acc; - const xxh_u64x2* const xsecret = (const xxh_u64x2*) secret; - /* constants */ - xxh_u64x2 const v32 = { 32, 32 }; - xxh_u64x2 const v47 = { 47, 47 }; - xxh_u32x4 const prime = { PRIME32_1, PRIME32_1, PRIME32_1, PRIME32_1 }; - size_t i; - for (i = 0; i < STRIPE_LEN / sizeof(xxh_u64x2); i++) { - /* xacc[i] ^= (xacc[i] >> 47); */ - xxh_u64x2 const acc_vec = xacc[i]; - xxh_u64x2 const data_vec = acc_vec ^ (acc_vec >> v47); - - /* xacc[i] ^= xsecret[i]; */ - xxh_u64x2 const key_vec = XXH_vec_loadu(xsecret + i); - xxh_u64x2 const data_key = data_vec ^ key_vec; - - /* xacc[i] *= PRIME32_1 */ - /* prod_lo = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)prime & 0xFFFFFFFF); */ - xxh_u64x2 const prod_even = XXH_vec_mule((xxh_u32x4)data_key, prime); - /* prod_hi = ((xxh_u64x2)data_key >> 32) * ((xxh_u64x2)prime >> 32); */ - xxh_u64x2 const prod_odd = XXH_vec_mulo((xxh_u32x4)data_key, prime); - xacc[i] = prod_odd + (prod_even << v32); - } } - -#else /* scalar variant of Scrambler - universal */ - - XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64* const xacc = (xxh_u64*) acc; /* presumed aligned */ - const xxh_u8* const xsecret = (const xxh_u8*) secret; /* no alignment restriction */ - size_t i; - XXH_ASSERT((((size_t)acc) & (XXH_ACC_ALIGN-1)) == 0); - for (i=0; i < ACC_NB; i++) { - xxh_u64 const key64 = XXH_readLE64(xsecret + 8*i); - xxh_u64 acc64 = xacc[i]; - acc64 = XXH_xorshift64(acc64, 47); - acc64 ^= key64; - acc64 *= PRIME32_1; - xacc[i] = acc64; - } - -#endif -} - -#define XXH_PREFETCH_DIST 384 - -/* - * XXH3_accumulate() - * Loops over XXH3_accumulate_512(). - * Assumption: nbStripes will not overflow the secret size - */ -XXH_FORCE_INLINE void -XXH3_accumulate( xxh_u64* XXH_RESTRICT acc, - const xxh_u8* XXH_RESTRICT input, - const xxh_u8* XXH_RESTRICT secret, - size_t nbStripes, - XXH3_accWidth_e accWidth) -{ - size_t n; - for (n = 0; n < nbStripes; n++ ) { - const xxh_u8* const in = input + n*STRIPE_LEN; - XXH_PREFETCH(in + XXH_PREFETCH_DIST); - XXH3_accumulate_512(acc, - in, - secret + n*XXH_SECRET_CONSUME_RATE, - accWidth); - } -} - -XXH_FORCE_INLINE void -XXH3_hashLong_internal_loop( xxh_u64* XXH_RESTRICT acc, - const xxh_u8* XXH_RESTRICT input, size_t len, - const xxh_u8* XXH_RESTRICT secret, size_t secretSize, - XXH3_accWidth_e accWidth) -{ - size_t const nb_rounds = (secretSize - STRIPE_LEN) / XXH_SECRET_CONSUME_RATE; - size_t const block_len = STRIPE_LEN * nb_rounds; - size_t const nb_blocks = len / block_len; - - size_t n; - - XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); - - for (n = 0; n < nb_blocks; n++) { - XXH3_accumulate(acc, input + n*block_len, secret, nb_rounds, accWidth); - XXH3_scrambleAcc(acc, secret + secretSize - STRIPE_LEN); - } - - /* last partial block */ - XXH_ASSERT(len > STRIPE_LEN); - { size_t const nbStripes = (len - (block_len * nb_blocks)) / STRIPE_LEN; - XXH_ASSERT(nbStripes <= (secretSize / XXH_SECRET_CONSUME_RATE)); - XXH3_accumulate(acc, input + nb_blocks*block_len, secret, nbStripes, accWidth); - - /* last stripe */ - if (len & (STRIPE_LEN - 1)) { - const xxh_u8* const p = input + len - STRIPE_LEN; - /* Do not align on 8, so that the secret is different from the scrambler */ -#define XXH_SECRET_LASTACC_START 7 - XXH3_accumulate_512(acc, p, secret + secretSize - STRIPE_LEN - XXH_SECRET_LASTACC_START, accWidth); - } } -} - -XXH_FORCE_INLINE xxh_u64 -XXH3_mix2Accs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret) -{ - return XXH3_mul128_fold64( - acc[0] ^ XXH_readLE64(secret), - acc[1] ^ XXH_readLE64(secret+8) ); -} - -static XXH64_hash_t -XXH3_mergeAccs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret, xxh_u64 start) -{ - xxh_u64 result64 = start; - - result64 += XXH3_mix2Accs(acc+0, secret + 0); - result64 += XXH3_mix2Accs(acc+2, secret + 16); - result64 += XXH3_mix2Accs(acc+4, secret + 32); - result64 += XXH3_mix2Accs(acc+6, secret + 48); - - return XXH3_avalanche(result64); -} - -#define XXH3_INIT_ACC { PRIME32_3, PRIME64_1, PRIME64_2, PRIME64_3, \ - PRIME64_4, PRIME32_2, PRIME64_5, PRIME32_1 }; - -XXH_FORCE_INLINE XXH64_hash_t -XXH3_hashLong_internal(const xxh_u8* XXH_RESTRICT input, size_t len, - const xxh_u8* XXH_RESTRICT secret, size_t secretSize) -{ - XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[ACC_NB] = XXH3_INIT_ACC; - - XXH3_hashLong_internal_loop(acc, input, len, secret, secretSize, XXH3_acc_64bits); - - /* converge into final hash */ - XXH_STATIC_ASSERT(sizeof(acc) == 64); - /* do not align on 8, so that the secret is different from the accumulator */ -#define XXH_SECRET_MERGEACCS_START 11 - XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START); - return XXH3_mergeAccs(acc, secret + XXH_SECRET_MERGEACCS_START, (xxh_u64)len * PRIME64_1); -} - -/* - * It's important for performance that XXH3_hashLong is not inlined. Not sure - * why (uop cache maybe?), but the difference is large and easily measurable. - */ -XXH_NO_INLINE XXH64_hash_t -XXH3_hashLong_64b_defaultSecret(const xxh_u8* XXH_RESTRICT input, size_t len) -{ - return XXH3_hashLong_internal(input, len, kSecret, sizeof(kSecret)); -} - -/* - * It's important for performance that XXH3_hashLong is not inlined. Not sure - * why (uop cache maybe?), but the difference is large and easily measurable. - */ -XXH_NO_INLINE XXH64_hash_t -XXH3_hashLong_64b_withSecret(const xxh_u8* XXH_RESTRICT input, size_t len, - const xxh_u8* XXH_RESTRICT secret, size_t secretSize) -{ - return XXH3_hashLong_internal(input, len, secret, secretSize); -} - - -XXH_FORCE_INLINE void XXH_writeLE64(void* dst, xxh_u64 v64) -{ - if (!XXH_CPU_LITTLE_ENDIAN) v64 = XXH_swap64(v64); - memcpy(dst, &v64, sizeof(v64)); -} - -/* XXH3_initCustomSecret() : - * destination `customSecret` is presumed allocated and same size as `kSecret`. - */ -XXH_FORCE_INLINE void XXH3_initCustomSecret(xxh_u8* customSecret, xxh_u64 seed64) -{ - int const nbRounds = XXH_SECRET_DEFAULT_SIZE / 16; - int i; - - XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0); - - for (i=0; i < nbRounds; i++) { - XXH_writeLE64(customSecret + 16*i, XXH_readLE64(kSecret + 16*i) + seed64); - XXH_writeLE64(customSecret + 16*i + 8, XXH_readLE64(kSecret + 16*i + 8) - seed64); - } -} - - -/* - * XXH3_hashLong_64b_withSeed(): - * Generate a custom key based on alteration of default kSecret with the seed, - * and then use this key for long mode hashing. - * - * This operation is decently fast but nonetheless costs a little bit of time. - * Try to avoid it whenever possible (typically when seed==0). - * - * It's important for performance that XXH3_hashLong is not inlined. Not sure - * why (uop cache maybe?), but the difference is large and easily measurable. - */ -XXH_NO_INLINE XXH64_hash_t -XXH3_hashLong_64b_withSeed(const xxh_u8* input, size_t len, XXH64_hash_t seed) -{ - XXH_ALIGN(8) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE]; - if (seed==0) return XXH3_hashLong_64b_defaultSecret(input, len); - XXH3_initCustomSecret(secret, seed); - return XXH3_hashLong_internal(input, len, secret, sizeof(secret)); -} - -/* === Public entry point === */ - -XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(const void* input, size_t len) -{ - if (len <= 16) return XXH3_len_0to16_64b((const xxh_u8*)input, len, kSecret, 0); - if (len <= 128) return XXH3_len_17to128_64b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), 0); - if (len <= XXH3_MIDSIZE_MAX) return XXH3_len_129to240_64b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), 0); - return XXH3_hashLong_64b_defaultSecret((const xxh_u8*)input, len); -} - -XXH_PUBLIC_API XXH64_hash_t -XXH3_64bits_withSecret(const void* input, size_t len, const void* secret, size_t secretSize) -{ - XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); - /* - * If an action is to be taken if `secret` conditions are not respected, - * it should be done here. - * For now, it's a contract pre-condition. - * Adding a check and a branch here would cost performance at every hash. - */ - if (len <= 16) return XXH3_len_0to16_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, 0); - if (len <= 128) return XXH3_len_17to128_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, 0); - if (len <= XXH3_MIDSIZE_MAX) return XXH3_len_129to240_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, 0); - return XXH3_hashLong_64b_withSecret((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize); -} - -XXH_PUBLIC_API XXH64_hash_t -XXH3_64bits_withSeed(const void* input, size_t len, XXH64_hash_t seed) -{ - if (len <= 16) return XXH3_len_0to16_64b((const xxh_u8*)input, len, kSecret, seed); - if (len <= 128) return XXH3_len_17to128_64b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), seed); - if (len <= XXH3_MIDSIZE_MAX) return XXH3_len_129to240_64b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), seed); - return XXH3_hashLong_64b_withSeed((const xxh_u8*)input, len, seed); -} - -/* === XXH3 streaming === */ - -XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void) -{ - return (XXH3_state_t*)XXH_malloc(sizeof(XXH3_state_t)); -} - -XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr) -{ - XXH_free(statePtr); - return XXH_OK; -} - -XXH_PUBLIC_API void -XXH3_copyState(XXH3_state_t* dst_state, const XXH3_state_t* src_state) -{ - memcpy(dst_state, src_state, sizeof(*dst_state)); -} - -static void -XXH3_64bits_reset_internal(XXH3_state_t* statePtr, - XXH64_hash_t seed, - const xxh_u8* secret, size_t secretSize) -{ - XXH_ASSERT(statePtr != NULL); - memset(statePtr, 0, sizeof(*statePtr)); - statePtr->acc[0] = PRIME32_3; - statePtr->acc[1] = PRIME64_1; - statePtr->acc[2] = PRIME64_2; - statePtr->acc[3] = PRIME64_3; - statePtr->acc[4] = PRIME64_4; - statePtr->acc[5] = PRIME32_2; - statePtr->acc[6] = PRIME64_5; - statePtr->acc[7] = PRIME32_1; - statePtr->seed = seed; - XXH_ASSERT(secret != NULL); - statePtr->secret = secret; - XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); - statePtr->secretLimit = (XXH32_hash_t)(secretSize - STRIPE_LEN); - statePtr->nbStripesPerBlock = statePtr->secretLimit / XXH_SECRET_CONSUME_RATE; -} - -XXH_PUBLIC_API XXH_errorcode -XXH3_64bits_reset(XXH3_state_t* statePtr) -{ - if (statePtr == NULL) return XXH_ERROR; - XXH3_64bits_reset_internal(statePtr, 0, kSecret, XXH_SECRET_DEFAULT_SIZE); - return XXH_OK; -} - -XXH_PUBLIC_API XXH_errorcode -XXH3_64bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize) -{ - if (statePtr == NULL) return XXH_ERROR; - XXH3_64bits_reset_internal(statePtr, 0, (const xxh_u8*)secret, secretSize); - if (secret == NULL) return XXH_ERROR; - if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR; - return XXH_OK; -} - -XXH_PUBLIC_API XXH_errorcode -XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed) -{ - if (statePtr == NULL) return XXH_ERROR; - XXH3_64bits_reset_internal(statePtr, seed, kSecret, XXH_SECRET_DEFAULT_SIZE); - XXH3_initCustomSecret(statePtr->customSecret, seed); - statePtr->secret = statePtr->customSecret; - return XXH_OK; -} - -XXH_FORCE_INLINE void -XXH3_consumeStripes( xxh_u64* acc, - XXH32_hash_t* nbStripesSoFarPtr, XXH32_hash_t nbStripesPerBlock, - const xxh_u8* input, size_t totalStripes, - const xxh_u8* secret, size_t secretLimit, - XXH3_accWidth_e accWidth) -{ - XXH_ASSERT(*nbStripesSoFarPtr < nbStripesPerBlock); - if (nbStripesPerBlock - *nbStripesSoFarPtr <= totalStripes) { - /* need a scrambling operation */ - size_t const nbStripes = nbStripesPerBlock - *nbStripesSoFarPtr; - XXH3_accumulate(acc, input, secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, nbStripes, accWidth); - XXH3_scrambleAcc(acc, secret + secretLimit); - XXH3_accumulate(acc, input + nbStripes * STRIPE_LEN, secret, totalStripes - nbStripes, accWidth); - *nbStripesSoFarPtr = (XXH32_hash_t)(totalStripes - nbStripes); - } else { - XXH3_accumulate(acc, input, secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, totalStripes, accWidth); - *nbStripesSoFarPtr += (XXH32_hash_t)totalStripes; - } -} - -/* - * Both XXH3_64bits_update and XXH3_128bits_update use this routine. - */ -XXH_FORCE_INLINE XXH_errorcode -XXH3_update(XXH3_state_t* state, const xxh_u8* input, size_t len, XXH3_accWidth_e accWidth) -{ - if (input==NULL) -#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1) - return XXH_OK; -#else - return XXH_ERROR; -#endif - - { const xxh_u8* const bEnd = input + len; - - state->totalLen += len; - - if (state->bufferedSize + len <= XXH3_INTERNALBUFFER_SIZE) { /* fill in tmp buffer */ - XXH_memcpy(state->buffer + state->bufferedSize, input, len); - state->bufferedSize += (XXH32_hash_t)len; - return XXH_OK; - } - /* input is now > XXH3_INTERNALBUFFER_SIZE */ - - #define XXH3_INTERNALBUFFER_STRIPES (XXH3_INTERNALBUFFER_SIZE / STRIPE_LEN) - XXH_STATIC_ASSERT(XXH3_INTERNALBUFFER_SIZE % STRIPE_LEN == 0); /* clean multiple */ - - /* - * There is some input left inside the internal buffer. - * Fill it, then consume it. - */ - if (state->bufferedSize) { - size_t const loadSize = XXH3_INTERNALBUFFER_SIZE - state->bufferedSize; - XXH_memcpy(state->buffer + state->bufferedSize, input, loadSize); - input += loadSize; - XXH3_consumeStripes(state->acc, - &state->nbStripesSoFar, state->nbStripesPerBlock, - state->buffer, XXH3_INTERNALBUFFER_STRIPES, - state->secret, state->secretLimit, - accWidth); - state->bufferedSize = 0; - } - - /* Consume input by full buffer quantities */ - if (input+XXH3_INTERNALBUFFER_SIZE <= bEnd) { - const xxh_u8* const limit = bEnd - XXH3_INTERNALBUFFER_SIZE; - do { - XXH3_consumeStripes(state->acc, - &state->nbStripesSoFar, state->nbStripesPerBlock, - input, XXH3_INTERNALBUFFER_STRIPES, - state->secret, state->secretLimit, - accWidth); - input += XXH3_INTERNALBUFFER_SIZE; - } while (input<=limit); - } - - if (input < bEnd) { /* Some remaining input: buffer it */ - XXH_memcpy(state->buffer, input, (size_t)(bEnd-input)); - state->bufferedSize = (XXH32_hash_t)(bEnd-input); - } - } - - return XXH_OK; -} - -XXH_PUBLIC_API XXH_errorcode -XXH3_64bits_update(XXH3_state_t* state, const void* input, size_t len) -{ - return XXH3_update(state, (const xxh_u8*)input, len, XXH3_acc_64bits); -} - - -XXH_FORCE_INLINE void -XXH3_digest_long (XXH64_hash_t* acc, const XXH3_state_t* state, XXH3_accWidth_e accWidth) -{ - /* - * Digest on a local copy. This way, the state remains unaltered, and it can - * continue ingesting more input afterwards. - */ - memcpy(acc, state->acc, sizeof(state->acc)); - if (state->bufferedSize >= STRIPE_LEN) { - size_t const totalNbStripes = state->bufferedSize / STRIPE_LEN; - XXH32_hash_t nbStripesSoFar = state->nbStripesSoFar; - XXH3_consumeStripes(acc, - &nbStripesSoFar, state->nbStripesPerBlock, - state->buffer, totalNbStripes, - state->secret, state->secretLimit, - accWidth); - if (state->bufferedSize % STRIPE_LEN) { /* one last partial stripe */ - XXH3_accumulate_512(acc, - state->buffer + state->bufferedSize - STRIPE_LEN, - state->secret + state->secretLimit - XXH_SECRET_LASTACC_START, - accWidth); - } - } else { /* bufferedSize < STRIPE_LEN */ - if (state->bufferedSize) { /* one last stripe */ - xxh_u8 lastStripe[STRIPE_LEN]; - size_t const catchupSize = STRIPE_LEN - state->bufferedSize; - memcpy(lastStripe, state->buffer + sizeof(state->buffer) - catchupSize, catchupSize); - memcpy(lastStripe + catchupSize, state->buffer, state->bufferedSize); - XXH3_accumulate_512(acc, - lastStripe, - state->secret + state->secretLimit - XXH_SECRET_LASTACC_START, - accWidth); - } } -} - -XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (const XXH3_state_t* state) -{ - if (state->totalLen > XXH3_MIDSIZE_MAX) { - XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[ACC_NB]; - XXH3_digest_long(acc, state, XXH3_acc_64bits); - return XXH3_mergeAccs(acc, state->secret + XXH_SECRET_MERGEACCS_START, (xxh_u64)state->totalLen * PRIME64_1); - } - /* len <= XXH3_MIDSIZE_MAX : short code */ - if (state->seed) - return XXH3_64bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed); - return XXH3_64bits_withSecret(state->buffer, (size_t)(state->totalLen), state->secret, state->secretLimit + STRIPE_LEN); -} - -/* ========================================== - * XXH3 128 bits (=> XXH128) - * ========================================== */ - -XXH_FORCE_INLINE XXH128_hash_t -XXH3_len_1to3_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) -{ - XXH_ASSERT(input != NULL); - XXH_ASSERT(1 <= len && len <= 3); - XXH_ASSERT(secret != NULL); - /* - * len = 1: combinedl = { input[0], 0x01, input[0], input[0] } - * len = 2: combinedl = { input[1], 0x02, input[0], input[1] } - * len = 3: combinedl = { input[2], 0x03, input[0], input[1] } - */ - { xxh_u8 const c1 = input[0]; - xxh_u8 const c2 = input[len >> 1]; - xxh_u8 const c3 = input[len - 1]; - xxh_u32 const combinedl = ((xxh_u32)c1<<16) | (((xxh_u32)c2) << 24) | (((xxh_u32)c3) << 0) | (((xxh_u32)len) << 8); - xxh_u32 const combinedh = XXH_rotl32(XXH_swap32(combinedl), 13); - xxh_u64 const bitflipl = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed; - xxh_u64 const bitfliph = (XXH_readLE32(secret+8) ^ XXH_readLE32(secret+12)) - seed; - xxh_u64 const keyed_lo = (xxh_u64)combinedl ^ bitflipl; - xxh_u64 const keyed_hi = (xxh_u64)combinedh ^ bitfliph; - xxh_u64 const mixedl = keyed_lo * PRIME64_1; - xxh_u64 const mixedh = keyed_hi * PRIME64_5; - XXH128_hash_t const h128 = { XXH3_avalanche(mixedl) /*low64*/, XXH3_avalanche(mixedh) /*high64*/ }; - return h128; - } -} - -XXH_FORCE_INLINE XXH128_hash_t -XXH3_len_4to8_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) -{ - XXH_ASSERT(input != NULL); - XXH_ASSERT(secret != NULL); - XXH_ASSERT(4 <= len && len <= 8); - seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32; - { xxh_u32 const input_lo = XXH_readLE32(input); - xxh_u32 const input_hi = XXH_readLE32(input + len - 4); - xxh_u64 const input_64 = input_lo + ((xxh_u64)input_hi << 32); - xxh_u64 const bitflip = (XXH_readLE64(secret+16) ^ XXH_readLE64(secret+24)) + seed; - xxh_u64 const keyed = input_64 ^ bitflip; - - /* Shift len to the left to ensure it is even, this avoids even multiplies. */ - XXH128_hash_t m128 = XXH_mult64to128(keyed, PRIME64_1 + (len << 2)); - - m128.high64 += (m128.low64 << 1); - m128.low64 ^= (m128.high64 >> 3); - - m128.low64 = XXH_xorshift64(m128.low64, 35); - m128.low64 *= 0x9FB21C651E98DF25ULL; - m128.low64 = XXH_xorshift64(m128.low64, 28); - m128.high64 = XXH3_avalanche(m128.high64); - return m128; - } -} - -XXH_FORCE_INLINE XXH128_hash_t -XXH3_len_9to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) -{ - XXH_ASSERT(input != NULL); - XXH_ASSERT(secret != NULL); - XXH_ASSERT(9 <= len && len <= 16); - { xxh_u64 const bitflipl = (XXH_readLE64(secret+32) ^ XXH_readLE64(secret+40)) - seed; - xxh_u64 const bitfliph = (XXH_readLE64(secret+48) ^ XXH_readLE64(secret+56)) + seed; - xxh_u64 const input_lo = XXH_readLE64(input); - xxh_u64 input_hi = XXH_readLE64(input + len - 8); - XXH128_hash_t m128 = XXH_mult64to128(input_lo ^ input_hi ^ bitflipl, PRIME64_1); - /* - * Put len in the middle of m128 to ensure that the length gets mixed to - * both the low and high bits in the 128x64 multiply below. - */ - m128.low64 += (xxh_u64)(len - 1) << 54; - input_hi ^= bitfliph; - /* - * Add the high 32 bits of input_hi to the high 32 bits of m128, then - * add the long product of the low 32 bits of input_hi and PRIME32_2 to - * the high 64 bits of m128. - * - * The best approach to this operation is different on 32-bit and 64-bit. - */ - if (sizeof(void *) < sizeof(xxh_u64)) { /* 32-bit */ - /* - * 32-bit optimized version, which is more readable. - * - * On 32-bit, it removes an ADC and delays a dependency between the two - * halves of m128.high64, but it generates an extra mask on 64-bit. - */ - m128.high64 += (input_hi & 0xFFFFFFFF00000000) + XXH_mult32to64((xxh_u32)input_hi, PRIME32_2); - } else { - /* - * 64-bit optimized (albeit more confusing) version. - * - * Uses some properties of addition and multiplication to remove the mask: - * - * Let: - * a = input_hi.lo = (input_hi & 0x00000000FFFFFFFF) - * b = input_hi.hi = (input_hi & 0xFFFFFFFF00000000) - * c = PRIME32_2 - * - * a + (b * c) - * Inverse Property: x + y - x == y - * a + (b * (1 + c - 1)) - * Distributive Property: x * (y + z) == (x * y) + (x * z) - * a + (b * 1) + (b * (c - 1)) - * Identity Property: x * 1 == x - * a + b + (b * (c - 1)) - * - * Substitute a, b, and c: - * input_hi.hi + input_hi.lo + ((xxh_u64)input_hi.lo * (PRIME32_2 - 1)) - * - * Since input_hi.hi + input_hi.lo == input_hi, we get this: - * input_hi + ((xxh_u64)input_hi.lo * (PRIME32_2 - 1)) - */ - m128.high64 += input_hi + XXH_mult32to64((xxh_u32)input_hi, PRIME32_2 - 1); - } - /* m128 ^= XXH_swap64(m128 >> 64); */ - m128.low64 ^= XXH_swap64(m128.high64); - - { /* 128x64 multiply: h128 = m128 * PRIME64_2; */ - XXH128_hash_t h128 = XXH_mult64to128(m128.low64, PRIME64_2); - h128.high64 += m128.high64 * PRIME64_2; - - h128.low64 = XXH3_avalanche(h128.low64); - h128.high64 = XXH3_avalanche(h128.high64); - return h128; - } } -} - -/* Assumption : `secret` size is >= 16 - * Note : it should be >= XXH3_SECRET_SIZE_MIN anyway */ -XXH_FORCE_INLINE XXH128_hash_t -XXH3_len_0to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) -{ - XXH_ASSERT(len <= 16); - { if (len > 8) return XXH3_len_9to16_128b(input, len, secret, seed); - if (len >= 4) return XXH3_len_4to8_128b(input, len, secret, seed); - if (len) return XXH3_len_1to3_128b(input, len, secret, seed); - { XXH128_hash_t h128; - xxh_u64 const bitflipl = XXH_readLE64(secret+64) ^ XXH_readLE64(secret+72); - xxh_u64 const bitfliph = XXH_readLE64(secret+80) ^ XXH_readLE64(secret+88); - h128.low64 = XXH3_avalanche((PRIME64_1 + seed) ^ bitflipl); - h128.high64 = XXH3_avalanche((PRIME64_2 - seed) ^ bitfliph); - return h128; - } } -} - -/* - * A bit slower than XXH3_mix16B, but handles multiply by zero better. - */ -XXH_FORCE_INLINE XXH128_hash_t -XXH128_mix32B(XXH128_hash_t acc, const xxh_u8* input_1, const xxh_u8* input_2, const xxh_u8* secret, XXH64_hash_t seed) -{ - acc.low64 += XXH3_mix16B (input_1, secret+0, seed); - acc.low64 ^= XXH_readLE64(input_2) + XXH_readLE64(input_2 + 8); - acc.high64 += XXH3_mix16B (input_2, secret+16, seed); - acc.high64 ^= XXH_readLE64(input_1) + XXH_readLE64(input_1 + 8); - return acc; -} - - -XXH_FORCE_INLINE XXH128_hash_t -XXH3_len_17to128_128b(const xxh_u8* XXH_RESTRICT input, size_t len, - const xxh_u8* XXH_RESTRICT secret, size_t secretSize, - XXH64_hash_t seed) -{ - XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize; - XXH_ASSERT(16 < len && len <= 128); - - { XXH128_hash_t acc; - acc.low64 = len * PRIME64_1; - acc.high64 = 0; - if (len > 32) { - if (len > 64) { - if (len > 96) { - acc = XXH128_mix32B(acc, input+48, input+len-64, secret+96, seed); - } - acc = XXH128_mix32B(acc, input+32, input+len-48, secret+64, seed); - } - acc = XXH128_mix32B(acc, input+16, input+len-32, secret+32, seed); - } - acc = XXH128_mix32B(acc, input, input+len-16, secret, seed); - { xxh_u64 const low64 = acc.low64 + acc.high64; - xxh_u64 const high64 = (acc.low64 * PRIME64_1) + (acc.high64 * PRIME64_4) + ((len - seed) * PRIME64_2); - XXH128_hash_t const h128 = { XXH3_avalanche(low64), (XXH64_hash_t)0 - XXH3_avalanche(high64) }; - return h128; - } - } -} - -XXH_NO_INLINE XXH128_hash_t -XXH3_len_129to240_128b(const xxh_u8* XXH_RESTRICT input, size_t len, - const xxh_u8* XXH_RESTRICT secret, size_t secretSize, - XXH64_hash_t seed) -{ - XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize; - XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX); - - { XXH128_hash_t acc; - int const nbRounds = (int)len / 32; - int i; - acc.low64 = len * PRIME64_1; - acc.high64 = 0; - for (i=0; i<4; i++) { - acc = XXH128_mix32B(acc, input+(32*i), input+(32*i)+16, secret+(32*i), seed); - } - acc.low64 = XXH3_avalanche(acc.low64); - acc.high64 = XXH3_avalanche(acc.high64); - XXH_ASSERT(nbRounds >= 4); - for (i=4 ; i < nbRounds; i++) { - acc = XXH128_mix32B(acc, input+(32*i), input+(32*i)+16, secret+XXH3_MIDSIZE_STARTOFFSET+(32*(i-4)), seed); - } - /* last bytes */ - acc = XXH128_mix32B(acc, input + len - 16, input + len - 32, secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET - 16, 0ULL - seed); - - { xxh_u64 const low64 = acc.low64 + acc.high64; - xxh_u64 const high64 = (acc.low64 * PRIME64_1) + (acc.high64 * PRIME64_4) + ((len - seed) * PRIME64_2); - XXH128_hash_t const h128 = { XXH3_avalanche(low64), (XXH64_hash_t)0 - XXH3_avalanche(high64) }; - return h128; - } - } -} - -XXH_FORCE_INLINE XXH128_hash_t -XXH3_hashLong_128b_internal(const xxh_u8* XXH_RESTRICT input, size_t len, - const xxh_u8* XXH_RESTRICT secret, size_t secretSize) -{ - XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[ACC_NB] = XXH3_INIT_ACC; - - XXH3_hashLong_internal_loop(acc, input, len, secret, secretSize, XXH3_acc_128bits); - - /* converge into final hash */ - XXH_STATIC_ASSERT(sizeof(acc) == 64); - XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START); - { xxh_u64 const low64 = XXH3_mergeAccs(acc, secret + XXH_SECRET_MERGEACCS_START, (xxh_u64)len * PRIME64_1); - xxh_u64 const high64 = XXH3_mergeAccs(acc, secret + secretSize - sizeof(acc) - XXH_SECRET_MERGEACCS_START, ~((xxh_u64)len * PRIME64_2)); - XXH128_hash_t const h128 = { low64, high64 }; - return h128; - } -} - -/* - * It's important for performance that XXH3_hashLong is not inlined. Not sure - * why (uop cache maybe?), but the difference is large and easily measurable. - */ -XXH_NO_INLINE XXH128_hash_t -XXH3_hashLong_128b_defaultSecret(const xxh_u8* input, size_t len) -{ - return XXH3_hashLong_128b_internal(input, len, kSecret, sizeof(kSecret)); -} - -/* - * It's important for performance that XXH3_hashLong is not inlined. Not sure - * why (uop cache maybe?), but the difference is large and easily measurable. - */ -XXH_NO_INLINE XXH128_hash_t -XXH3_hashLong_128b_withSecret(const xxh_u8* input, size_t len, - const xxh_u8* secret, size_t secretSize) -{ - return XXH3_hashLong_128b_internal(input, len, secret, secretSize); -} - -/* - * It's important for performance that XXH3_hashLong is not inlined. Not sure - * why (uop cache maybe?), but the difference is large and easily measurable. - */ -XXH_NO_INLINE XXH128_hash_t -XXH3_hashLong_128b_withSeed(const xxh_u8* input, size_t len, XXH64_hash_t seed) -{ - XXH_ALIGN(8) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE]; - if (seed == 0) return XXH3_hashLong_128b_defaultSecret(input, len); - XXH3_initCustomSecret(secret, seed); - return XXH3_hashLong_128b_internal(input, len, secret, sizeof(secret)); -} - - -XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void* input, size_t len) -{ - if (len <= 16) return XXH3_len_0to16_128b((const xxh_u8*)input, len, kSecret, 0); - if (len <= 128) return XXH3_len_17to128_128b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), 0); - if (len <= XXH3_MIDSIZE_MAX) return XXH3_len_129to240_128b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), 0); - return XXH3_hashLong_128b_defaultSecret((const xxh_u8*)input, len); -} - -XXH_PUBLIC_API XXH128_hash_t -XXH3_128bits_withSecret(const void* input, size_t len, const void* secret, size_t secretSize) -{ - XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); - /* - * If an action is to be taken if `secret` conditions are not respected, - * it should be done here. - * For now, it's a contract pre-condition. - * Adding a check and a branch here would cost performance at every hash. - */ - if (len <= 16) return XXH3_len_0to16_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, 0); - if (len <= 128) return XXH3_len_17to128_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, 0); - if (len <= XXH3_MIDSIZE_MAX) return XXH3_len_129to240_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, 0); - return XXH3_hashLong_128b_withSecret((const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize); -} - -XXH_PUBLIC_API XXH128_hash_t -XXH3_128bits_withSeed(const void* input, size_t len, XXH64_hash_t seed) -{ - if (len <= 16) return XXH3_len_0to16_128b((const xxh_u8*)input, len, kSecret, seed); - if (len <= 128) return XXH3_len_17to128_128b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), seed); - if (len <= XXH3_MIDSIZE_MAX) return XXH3_len_129to240_128b((const xxh_u8*)input, len, kSecret, sizeof(kSecret), seed); - return XXH3_hashLong_128b_withSeed((const xxh_u8*)input, len, seed); -} - -XXH_PUBLIC_API XXH128_hash_t -XXH128(const void* input, size_t len, XXH64_hash_t seed) -{ - return XXH3_128bits_withSeed(input, len, seed); -} - - -/* === XXH3 128-bit streaming === */ - -/* all the functions are actually the same as for 64-bit streaming variant, - just the reset one is different (different initial acc values for 0,5,6,7), - and near the end of the digest function */ - -static void -XXH3_128bits_reset_internal(XXH3_state_t* statePtr, - XXH64_hash_t seed, - const xxh_u8* secret, size_t secretSize) -{ - XXH3_64bits_reset_internal(statePtr, seed, secret, secretSize); -} - -XXH_PUBLIC_API XXH_errorcode -XXH3_128bits_reset(XXH3_state_t* statePtr) -{ - if (statePtr == NULL) return XXH_ERROR; - XXH3_128bits_reset_internal(statePtr, 0, kSecret, XXH_SECRET_DEFAULT_SIZE); - return XXH_OK; -} - -XXH_PUBLIC_API XXH_errorcode -XXH3_128bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize) -{ - if (statePtr == NULL) return XXH_ERROR; - XXH3_128bits_reset_internal(statePtr, 0, (const xxh_u8*)secret, secretSize); - if (secret == NULL) return XXH_ERROR; - if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR; - return XXH_OK; -} - -XXH_PUBLIC_API XXH_errorcode -XXH3_128bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed) -{ - if (statePtr == NULL) return XXH_ERROR; - XXH3_128bits_reset_internal(statePtr, seed, kSecret, XXH_SECRET_DEFAULT_SIZE); - XXH3_initCustomSecret(statePtr->customSecret, seed); - statePtr->secret = statePtr->customSecret; - return XXH_OK; -} - -XXH_PUBLIC_API XXH_errorcode -XXH3_128bits_update(XXH3_state_t* state, const void* input, size_t len) -{ - return XXH3_update(state, (const xxh_u8*)input, len, XXH3_acc_128bits); -} - -XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (const XXH3_state_t* state) -{ - if (state->totalLen > XXH3_MIDSIZE_MAX) { - XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[ACC_NB]; - XXH3_digest_long(acc, state, XXH3_acc_128bits); - XXH_ASSERT(state->secretLimit + STRIPE_LEN >= sizeof(acc) + XXH_SECRET_MERGEACCS_START); - { xxh_u64 const low64 = XXH3_mergeAccs(acc, state->secret + XXH_SECRET_MERGEACCS_START, (xxh_u64)state->totalLen * PRIME64_1); - xxh_u64 const high64 = XXH3_mergeAccs(acc, state->secret + state->secretLimit + STRIPE_LEN - sizeof(acc) - XXH_SECRET_MERGEACCS_START, ~((xxh_u64)state->totalLen * PRIME64_2)); - XXH128_hash_t const h128 = { low64, high64 }; - return h128; - } - } - /* len <= XXH3_MIDSIZE_MAX : short code */ - if (state->seed) - return XXH3_128bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed); - return XXH3_128bits_withSecret(state->buffer, (size_t)(state->totalLen), state->secret, state->secretLimit + STRIPE_LEN); -} - -/* 128-bit utility functions */ - -#include <string.h> /* memcmp, memcpy */ - -/* return : 1 is equal, 0 if different */ -XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2) -{ - /* note : XXH128_hash_t is compact, it has no padding byte */ - return !(memcmp(&h1, &h2, sizeof(h1))); -} - -/* This prototype is compatible with stdlib's qsort(). - * return : >0 if *h128_1 > *h128_2 - * <0 if *h128_1 < *h128_2 - * =0 if *h128_1 == *h128_2 */ -XXH_PUBLIC_API int XXH128_cmp(const void* h128_1, const void* h128_2) -{ - XXH128_hash_t const h1 = *(const XXH128_hash_t*)h128_1; - XXH128_hash_t const h2 = *(const XXH128_hash_t*)h128_2; - int const hcmp = (h1.high64 > h2.high64) - (h2.high64 > h1.high64); - /* note : bets that, in most cases, hash values are different */ - if (hcmp) return hcmp; - return (h1.low64 > h2.low64) - (h2.low64 > h1.low64); -} - - -/*====== Canonical representation ======*/ -XXH_PUBLIC_API void -XXH128_canonicalFromHash(XXH128_canonical_t* dst, XXH128_hash_t hash) -{ - XXH_STATIC_ASSERT(sizeof(XXH128_canonical_t) == sizeof(XXH128_hash_t)); - if (XXH_CPU_LITTLE_ENDIAN) { - hash.high64 = XXH_swap64(hash.high64); - hash.low64 = XXH_swap64(hash.low64); - } - memcpy(dst, &hash.high64, sizeof(hash.high64)); - memcpy((char*)dst + sizeof(hash.high64), &hash.low64, sizeof(hash.low64)); -} - -XXH_PUBLIC_API XXH128_hash_t -XXH128_hashFromCanonical(const XXH128_canonical_t* src) -{ - XXH128_hash_t h; - h.high64 = XXH_readBE64(src); - h.low64 = XXH_readBE64(src->digest + 8); - return h; -} - -/* Pop our optimization override from above */ -#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \ - && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \ - && defined(__OPTIMIZE__) && !defined(__OPTIMIZE_SIZE__) /* respect -O0 and -Os */ -# pragma GCC pop_options -#endif - -#endif /* XXH3_H_1397135465 */ +#define XXH_INLINE_ALL +#include "xxhash.h" diff --git a/dviware/dvisvgm/libs/xxHash/xxhash.c b/dviware/dvisvgm/libs/xxHash/xxhash.c index 3e0da98f3e..0fae88c5d6 100644 --- a/dviware/dvisvgm/libs/xxHash/xxhash.c +++ b/dviware/dvisvgm/libs/xxHash/xxhash.c @@ -1,6 +1,6 @@ /* * xxHash - Extremely Fast Hash algorithm - * Copyright (C) 2012-present, Yann Collet + * Copyright (C) 2012-2020 Yann Collet * * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php) * diff --git a/dviware/dvisvgm/libs/xxHash/xxhash.h b/dviware/dvisvgm/libs/xxHash/xxhash.h index 3677f8287b..2d56d23c5d 100644 --- a/dviware/dvisvgm/libs/xxHash/xxhash.h +++ b/dviware/dvisvgm/libs/xxHash/xxhash.h @@ -1,7 +1,7 @@ /* * xxHash - Extremely Fast Hash algorithm * Header File - * Copyright (C) 2012-present, Yann Collet. + * Copyright (C) 2012-2020 Yann Collet * * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php) * @@ -195,6 +195,7 @@ extern "C" { # define XXH_CAT(A,B) A##B # define XXH_NAME2(A,B) XXH_CAT(A,B) # define XXH_versionNumber XXH_NAME2(XXH_NAMESPACE, XXH_versionNumber) +/* XXH32 */ # define XXH32 XXH_NAME2(XXH_NAMESPACE, XXH32) # define XXH32_createState XXH_NAME2(XXH_NAMESPACE, XXH32_createState) # define XXH32_freeState XXH_NAME2(XXH_NAMESPACE, XXH32_freeState) @@ -204,6 +205,7 @@ extern "C" { # define XXH32_copyState XXH_NAME2(XXH_NAMESPACE, XXH32_copyState) # define XXH32_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH32_canonicalFromHash) # define XXH32_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH32_hashFromCanonical) +/* XXH64 */ # define XXH64 XXH_NAME2(XXH_NAMESPACE, XXH64) # define XXH64_createState XXH_NAME2(XXH_NAMESPACE, XXH64_createState) # define XXH64_freeState XXH_NAME2(XXH_NAMESPACE, XXH64_freeState) @@ -213,6 +215,33 @@ extern "C" { # define XXH64_copyState XXH_NAME2(XXH_NAMESPACE, XXH64_copyState) # define XXH64_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH64_canonicalFromHash) # define XXH64_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH64_hashFromCanonical) +/* XXH3_64bits */ +# define XXH3_64bits XXH_NAME2(XXH_NAMESPACE, XXH3_64bits) +# define XXH3_64bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecret) +# define XXH3_64bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSeed) +# define XXH3_createState XXH_NAME2(XXH_NAMESPACE, XXH3_createState) +# define XXH3_freeState XXH_NAME2(XXH_NAMESPACE, XXH3_freeState) +# define XXH3_copyState XXH_NAME2(XXH_NAMESPACE, XXH3_copyState) +# define XXH3_64bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset) +# define XXH3_64bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSeed) +# define XXH3_64bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecret) +# define XXH3_64bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_update) +# define XXH3_64bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_digest) +# define XXH3_generateSecret XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret) +/* XXH3_128bits */ +# define XXH128 XXH_NAME2(XXH_NAMESPACE, XXH128) +# define XXH3_128bits XXH_NAME2(XXH_NAMESPACE, XXH3_128bits) +# define XXH3_128bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSeed) +# define XXH3_128bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecret) +# define XXH3_128bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset) +# define XXH3_128bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSeed) +# define XXH3_128bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecret) +# define XXH3_128bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_update) +# define XXH3_128bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_digest) +# define XXH128_isEqual XXH_NAME2(XXH_NAMESPACE, XXH128_isEqual) +# define XXH128_cmp XXH_NAME2(XXH_NAMESPACE, XXH128_cmp) +# define XXH128_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH128_canonicalFromHash) +# define XXH128_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH128_hashFromCanonical) #endif @@ -220,8 +249,8 @@ extern "C" { * Version ***************************************/ #define XXH_VERSION_MAJOR 0 -#define XXH_VERSION_MINOR 7 -#define XXH_VERSION_RELEASE 3 +#define XXH_VERSION_MINOR 8 +#define XXH_VERSION_RELEASE 0 #define XXH_VERSION_NUMBER (XXH_VERSION_MAJOR *100*100 + XXH_VERSION_MINOR *100 + XXH_VERSION_RELEASE) XXH_PUBLIC_API unsigned XXH_versionNumber (void); @@ -260,6 +289,10 @@ typedef enum { XXH_OK=0, XXH_ERROR } XXH_errorcode; * The memory between input & input+length must be valid (allocated and read-accessible). * "seed" can be used to alter the result predictably. * Speed on Core 2 Duo @ 3 GHz (single thread, SMHasher benchmark): 5.4 GB/s + * + * Note: XXH3 provides competitive speed for both 32-bit and 64-bit systems, + * and offers true 64/128 bit hash results. It provides a superior level of + * dispersion, and greatly reduces the risks of collisions. */ XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed); @@ -342,8 +375,13 @@ XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src * Returns the 64-bit hash of sequence of length @length stored at memory * address @input. * @seed can be used to alter the result predictably. + * * This function usually runs faster on 64-bit systems, but slower on 32-bit * systems (see benchmark). + * + * Note: XXH3 provides competitive speed for both 32-bit and 64-bit systems, + * and offers true 64/128 bit hash results. It provides a superior level of + * dispersion, and greatly reduces the risks of collisions. */ XXH_PUBLIC_API XXH64_hash_t XXH64 (const void* input, size_t length, XXH64_hash_t seed); @@ -358,64 +396,13 @@ XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH64_state_t* statePtr, const void* XXH_PUBLIC_API XXH64_hash_t XXH64_digest (const XXH64_state_t* statePtr); /******* Canonical representation *******/ -typedef struct { unsigned char digest[8]; } XXH64_canonical_t; +typedef struct { unsigned char digest[sizeof(XXH64_hash_t)]; } XXH64_canonical_t; XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t* dst, XXH64_hash_t hash); XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src); -#endif /* XXH_NO_LONG_LONG */ - -#endif /* XXHASH_H_5627135585666179 */ - -#define XXH_STATIC_LINKING_ONLY - -#if defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742) -#define XXHASH_H_STATIC_13879238742 -/* **************************************************************************** - * This section contains declarations which are not guaranteed to remain stable. - * They may change in future versions, becoming incompatible with a different - * version of the library. - * These declarations should only be used with static linking. - * Never use them in association with dynamic linking! - ***************************************************************************** */ - -/* - * These definitions are only present to allow static allocation of an XXH - * state, for example, on the stack or in a struct. - * Never **ever** access members directly. - */ - -struct XXH32_state_s { - XXH32_hash_t total_len_32; - XXH32_hash_t large_len; - XXH32_hash_t v1; - XXH32_hash_t v2; - XXH32_hash_t v3; - XXH32_hash_t v4; - XXH32_hash_t mem32[4]; - XXH32_hash_t memsize; - XXH32_hash_t reserved; /* never read nor write, might be removed in a future version */ -}; /* typedef'd to XXH32_state_t */ - - -#ifndef XXH_NO_LONG_LONG /* defined when there is no 64-bit support */ - -struct XXH64_state_s { - XXH64_hash_t total_len; - XXH64_hash_t v1; - XXH64_hash_t v2; - XXH64_hash_t v3; - XXH64_hash_t v4; - XXH64_hash_t mem64[4]; - XXH32_hash_t memsize; - XXH32_hash_t reserved32; /* required for padding anyway */ - XXH64_hash_t reserved64; /* never read nor write, might be removed in a future version */ -}; /* typedef'd to XXH64_state_t */ - - /*-********************************************************************** -* XXH3 -* New experimental hash +* XXH3 64-bit variant ************************************************************************/ /* ************************************************************************ @@ -439,11 +426,11 @@ struct XXH64_state_s { * It benefits greatly from SIMD and 64-bit arithmetic, but does not require it. * * Almost all 32-bit and 64-bit targets that can run XXH32 smoothly can run - * XXH3 at usable speeds, even if XXH64 runs slowly. Further details are + * XXH3 at competitive speeds, even if XXH64 runs slowly. Further details are * explained in the implementation. * - * Optimized implementations are provided for AVX2, SSE2, NEON, POWER8, ZVector, - * and scalar targets. This can be controlled with the XXH_VECTOR macro. + * Optimized implementations are provided for AVX512, AVX2, SSE2, NEON, POWER8, + * ZVector and scalar targets. This can be controlled with the XXH_VECTOR macro. * * XXH3 offers 2 variants, _64bits and _128bits. * When only 64 bits are needed, prefer calling the _64bits variant, as it @@ -461,116 +448,61 @@ struct XXH64_state_s { * ephemeral data (local sessions). * * Avoid storing values in long-term storage until the algorithm is finalized. + * XXH3's return values will be officially finalized upon reaching v0.8.0. + * + * After which, return values of XXH3 and XXH128 will no longer change in + * future versions. * * The API supports one-shot hashing, streaming mode, and custom secrets. */ -#ifdef XXH_NAMESPACE -# define XXH3_64bits XXH_NAME2(XXH_NAMESPACE, XXH3_64bits) -# define XXH3_64bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecret) -# define XXH3_64bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSeed) - -# define XXH3_createState XXH_NAME2(XXH_NAMESPACE, XXH3_createState) -# define XXH3_freeState XXH_NAME2(XXH_NAMESPACE, XXH3_freeState) -# define XXH3_copyState XXH_NAME2(XXH_NAMESPACE, XXH3_copyState) - -# define XXH3_64bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset) -# define XXH3_64bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSeed) -# define XXH3_64bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecret) -# define XXH3_64bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_update) -# define XXH3_64bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_digest) -#endif - /* XXH3_64bits(): * default 64-bit variant, using default secret and default seed of 0. * It's the fastest variant. */ XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(const void* data, size_t len); /* - * XXH3_64bits_withSecret(): - * It's possible to provide any blob of bytes as a "secret" to generate the hash. - * This makes it more difficult for an external actor to prepare an intentional - * collision. - * The secret *must* be large enough (>= XXH3_SECRET_SIZE_MIN). - * It should consist of random bytes. - * Avoid trivial sequences, such as repeating sequences and especially '\0', - * as this can cancel out itself. - * Failure to respect these conditions will result in a poor quality hash. - */ -#define XXH3_SECRET_SIZE_MIN 136 -XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSecret(const void* data, size_t len, const void* secret, size_t secretSize); - -/* * XXH3_64bits_withSeed(): - * This variant generates a custom secret on the fly based on the default - * secret, altered using the `seed` value. + * This variant generates a custom secret on the fly + * based on default secret altered using the `seed` value. * While this operation is decently fast, note that it's not completely free. * Note: seed==0 produces the same results as XXH3_64bits(). */ XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSeed(const void* data, size_t len, XXH64_hash_t seed); +/* + * XXH3_64bits_withSecret(): + * It's possible to provide any blob of bytes as a "secret" to generate the hash. + * This makes it more difficult for an external actor to prepare an intentional collision. + * The main condition is that secretSize *must* be large enough (>= XXH3_SECRET_SIZE_MIN). + * However, the quality of produced hash values depends on secret's entropy. + * Technically, the secret must look like a bunch of random bytes. + * Avoid "trivial" or structured data such as repeated sequences or a text document. + * Whenever unsure about the "randomness" of the blob of bytes, + * consider relabelling it as a "custom seed" instead, + * and employ "XXH3_generateSecret()" (see below) + * to generate a high entropy secret derived from the custom seed. + */ +#define XXH3_SECRET_SIZE_MIN 136 +XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSecret(const void* data, size_t len, const void* secret, size_t secretSize); -/* streaming 64-bit */ - -#if defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* C11+ */ -# include <stdalign.h> -# define XXH_ALIGN(n) alignas(n) -#elif defined(__GNUC__) -# define XXH_ALIGN(n) __attribute__ ((aligned(n))) -#elif defined(_MSC_VER) -# define XXH_ALIGN(n) __declspec(align(n)) -#else -# define XXH_ALIGN(n) /* disabled */ -#endif - -/* Old GCC versions only accept the attribute after the type in structures. */ -#if !(defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)) /* C11+ */ \ - && defined(__GNUC__) -# define XXH_ALIGN_MEMBER(align, type) type XXH_ALIGN(align) -#else -# define XXH_ALIGN_MEMBER(align, type) XXH_ALIGN(align) type -#endif - -typedef struct XXH3_state_s XXH3_state_t; - -#define XXH3_SECRET_DEFAULT_SIZE 192 /* minimum XXH3_SECRET_SIZE_MIN */ -#define XXH3_INTERNALBUFFER_SIZE 256 -struct XXH3_state_s { - XXH_ALIGN_MEMBER(64, XXH64_hash_t acc[8]); - /* used to store a custom secret generated from the seed. Makes state larger. - * Design might change */ - XXH_ALIGN_MEMBER(64, unsigned char customSecret[XXH3_SECRET_DEFAULT_SIZE]); - XXH_ALIGN_MEMBER(64, unsigned char buffer[XXH3_INTERNALBUFFER_SIZE]); - XXH32_hash_t bufferedSize; - XXH32_hash_t nbStripesPerBlock; - XXH32_hash_t nbStripesSoFar; - XXH32_hash_t secretLimit; - XXH32_hash_t reserved32; - XXH32_hash_t reserved32_2; - XXH64_hash_t totalLen; - XXH64_hash_t seed; - XXH64_hash_t reserved64; - /* note: there is some padding after due to alignment on 64 bytes */ - const unsigned char* secret; -}; /* typedef'd to XXH3_state_t */ - -#undef XXH_ALIGN_MEMBER +/******* Streaming *******/ /* * Streaming requires state maintenance. * This operation costs memory and CPU. * As a consequence, streaming is slower than one-shot hashing. - * For better performance, prefer one-shot functions whenever possible. + * For better performance, prefer one-shot functions whenever applicable. */ +typedef struct XXH3_state_s XXH3_state_t; XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void); XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr); XXH_PUBLIC_API void XXH3_copyState(XXH3_state_t* dst_state, const XXH3_state_t* src_state); - /* * XXH3_64bits_reset(): - * Initialize with the default parameters. - * The result will be equivalent to `XXH3_64bits()`. + * Initialize with default parameters. + * digest will be equivalent to `XXH3_64bits()`. */ XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH3_state_t* statePtr); /* @@ -581,46 +513,48 @@ XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH3_state_t* statePtr); XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed); /* * XXH3_64bits_reset_withSecret(): - * `secret` is referenced, and must outlive the hash streaming session, so - * be careful when using stack arrays. - * `secretSize` must be >= `XXH3_SECRET_SIZE_MIN`. + * `secret` is referenced, it _must outlive_ the hash streaming session. + * Similar to one-shot API, `secretSize` must be >= `XXH3_SECRET_SIZE_MIN`, + * and the quality of produced hash values depends on secret's entropy + * (secret's content should look like a bunch of random bytes). + * When in doubt about the randomness of a candidate `secret`, + * consider employing `XXH3_generateSecret()` instead (see below). */ XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize); XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update (XXH3_state_t* statePtr, const void* input, size_t length); XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (const XXH3_state_t* statePtr); +/* note : canonical representation of XXH3 is the same as XXH64 + * since they both produce XXH64_hash_t values */ -/* 128-bit */ - -#ifdef XXH_NAMESPACE -# define XXH128 XXH_NAME2(XXH_NAMESPACE, XXH128) -# define XXH3_128bits XXH_NAME2(XXH_NAMESPACE, XXH3_128bits) -# define XXH3_128bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSeed) -# define XXH3_128bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecret) -# define XXH3_128bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset) -# define XXH3_128bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSeed) -# define XXH3_128bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecret) -# define XXH3_128bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_update) -# define XXH3_128bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_digest) - -# define XXH128_isEqual XXH_NAME2(XXH_NAMESPACE, XXH128_isEqual) -# define XXH128_cmp XXH_NAME2(XXH_NAMESPACE, XXH128_cmp) -# define XXH128_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH128_canonicalFromHash) -# define XXH128_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH128_hashFromCanonical) -#endif +/*-********************************************************************** +* XXH3 128-bit variant +************************************************************************/ typedef struct { - XXH64_hash_t low64; - XXH64_hash_t high64; + XXH64_hash_t low64; + XXH64_hash_t high64; } XXH128_hash_t; -XXH_PUBLIC_API XXH128_hash_t XXH128(const void* data, size_t len, XXH64_hash_t seed); XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void* data, size_t len); -XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSeed(const void* data, size_t len, XXH64_hash_t seed); /* == XXH128() */ +XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSeed(const void* data, size_t len, XXH64_hash_t seed); XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSecret(const void* data, size_t len, const void* secret, size_t secretSize); +/******* Streaming *******/ +/* + * Streaming requires state maintenance. + * This operation costs memory and CPU. + * As a consequence, streaming is slower than one-shot hashing. + * For better performance, prefer one-shot functions whenever applicable. + * + * XXH3_128bits uses the same XXH3_state_t as XXH3_64bits(). + * Use already declared XXH3_createState() and XXH3_freeState(). + * + * All reset and streaming functions have same meaning as their 64-bit counterpart. + */ + XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH3_state_t* statePtr); XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed); XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize); @@ -628,8 +562,9 @@ XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH3_state_t* statePt XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update (XXH3_state_t* statePtr, const void* input, size_t length); XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (const XXH3_state_t* statePtr); - -/* Note: For better performance, these functions can be inlined using XXH_INLINE_ALL */ +/* Following helper functions make it possible to compare XXH128_hast_t values. + * Since XXH128_hash_t is a structure, this capability is not offered by the language. + * Note: For better performance, these functions can be inlined using XXH_INLINE_ALL */ /*! * XXH128_isEqual(): @@ -643,20 +578,161 @@ XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2); * This comparator is compatible with stdlib's `qsort()`/`bsearch()`. * * return: >0 if *h128_1 > *h128_2 - * <0 if *h128_1 < *h128_2 * =0 if *h128_1 == *h128_2 + * <0 if *h128_1 < *h128_2 */ XXH_PUBLIC_API int XXH128_cmp(const void* h128_1, const void* h128_2); /******* Canonical representation *******/ -typedef struct { unsigned char digest[16]; } XXH128_canonical_t; +typedef struct { unsigned char digest[sizeof(XXH128_hash_t)]; } XXH128_canonical_t; XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH128_canonical_t* dst, XXH128_hash_t hash); XXH_PUBLIC_API XXH128_hash_t XXH128_hashFromCanonical(const XXH128_canonical_t* src); #endif /* XXH_NO_LONG_LONG */ +#endif /* XXHASH_H_5627135585666179 */ + + + +#if defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742) +#define XXHASH_H_STATIC_13879238742 +/* **************************************************************************** + * This section contains declarations which are not guaranteed to remain stable. + * They may change in future versions, becoming incompatible with a different + * version of the library. + * These declarations should only be used with static linking. + * Never use them in association with dynamic linking! + ***************************************************************************** */ + +/* + * These definitions are only present to allow static allocation + * of XXH states, on stack or in a struct, for example. + * Never **ever** access their members directly. + */ + +struct XXH32_state_s { + XXH32_hash_t total_len_32; + XXH32_hash_t large_len; + XXH32_hash_t v1; + XXH32_hash_t v2; + XXH32_hash_t v3; + XXH32_hash_t v4; + XXH32_hash_t mem32[4]; + XXH32_hash_t memsize; + XXH32_hash_t reserved; /* never read nor write, might be removed in a future version */ +}; /* typedef'd to XXH32_state_t */ + + +#ifndef XXH_NO_LONG_LONG /* defined when there is no 64-bit support */ + +struct XXH64_state_s { + XXH64_hash_t total_len; + XXH64_hash_t v1; + XXH64_hash_t v2; + XXH64_hash_t v3; + XXH64_hash_t v4; + XXH64_hash_t mem64[4]; + XXH32_hash_t memsize; + XXH32_hash_t reserved32; /* required for padding anyway */ + XXH64_hash_t reserved64; /* never read nor write, might be removed in a future version */ +}; /* typedef'd to XXH64_state_t */ + +#if defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* C11+ */ +# include <stdalign.h> +# define XXH_ALIGN(n) alignas(n) +#elif defined(__GNUC__) +# define XXH_ALIGN(n) __attribute__ ((aligned(n))) +#elif defined(_MSC_VER) +# define XXH_ALIGN(n) __declspec(align(n)) +#else +# define XXH_ALIGN(n) /* disabled */ +#endif + +/* Old GCC versions only accept the attribute after the type in structures. */ +#if !(defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)) /* C11+ */ \ + && defined(__GNUC__) +# define XXH_ALIGN_MEMBER(align, type) type XXH_ALIGN(align) +#else +# define XXH_ALIGN_MEMBER(align, type) XXH_ALIGN(align) type +#endif + +#define XXH3_INTERNALBUFFER_SIZE 256 +#define XXH3_SECRET_DEFAULT_SIZE 192 +struct XXH3_state_s { + XXH_ALIGN_MEMBER(64, XXH64_hash_t acc[8]); + /* used to store a custom secret generated from a seed */ + XXH_ALIGN_MEMBER(64, unsigned char customSecret[XXH3_SECRET_DEFAULT_SIZE]); + XXH_ALIGN_MEMBER(64, unsigned char buffer[XXH3_INTERNALBUFFER_SIZE]); + XXH32_hash_t bufferedSize; + XXH32_hash_t reserved32; + size_t nbStripesSoFar; + XXH64_hash_t totalLen; + size_t nbStripesPerBlock; + size_t secretLimit; + XXH64_hash_t seed; + XXH64_hash_t reserved64; + const unsigned char* extSecret; /* reference to external secret; + * if == NULL, use .customSecret instead */ + /* note: there may be some padding at the end due to alignment on 64 bytes */ +}; /* typedef'd to XXH3_state_t */ + +#undef XXH_ALIGN_MEMBER + +/* When the XXH3_state_t structure is merely emplaced on stack, + * it should be initialized with XXH3_INITSTATE() or a memset() + * in case its first reset uses XXH3_NNbits_reset_withSeed(). + * This init can be omitted if the first reset uses default or _withSecret mode. + * This operation isn't necessary when the state is created with XXH3_createState(). + * Note that this doesn't prepare the state for a streaming operation, + * it's still necessary to use XXH3_NNbits_reset*() afterwards. + */ +#define XXH3_INITSTATE(XXH3_state_ptr) { (XXH3_state_ptr)->seed = 0; } + + +/* === Experimental API === */ +/* Symbols defined below must be considered tied to a specific library version. */ + +/* + * XXH3_generateSecret(): + * + * Derive a high-entropy secret from any user-defined content, named customSeed. + * The generated secret can be used in combination with `*_withSecret()` functions. + * The `_withSecret()` variants are useful to provide a higher level of protection than 64-bit seed, + * as it becomes much more difficult for an external actor to guess how to impact the calculation logic. + * + * The function accepts as input a custom seed of any length and any content, + * and derives from it a high-entropy secret of length XXH3_SECRET_DEFAULT_SIZE + * into an already allocated buffer secretBuffer. + * The generated secret is _always_ XXH_SECRET_DEFAULT_SIZE bytes long. + * + * The generated secret can then be used with any `*_withSecret()` variant. + * Functions `XXH3_128bits_withSecret()`, `XXH3_64bits_withSecret()`, + * `XXH3_128bits_reset_withSecret()` and `XXH3_64bits_reset_withSecret()` + * are part of this list. They all accept a `secret` parameter + * which must be very long for implementation reasons (>= XXH3_SECRET_SIZE_MIN) + * _and_ feature very high entropy (consist of random-looking bytes). + * These conditions can be a high bar to meet, so + * this function can be used to generate a secret of proper quality. + * + * customSeed can be anything. It can have any size, even small ones, + * and its content can be anything, even stupidly "low entropy" source such as a bunch of zeroes. + * The resulting `secret` will nonetheless provide all expected qualities. + * + * Supplying NULL as the customSeed copies the default secret into `secretBuffer`. + * When customSeedSize > 0, supplying NULL as customSeed is undefined behavior. + */ +XXH_PUBLIC_API void XXH3_generateSecret(void* secretBuffer, const void* customSeed, size_t customSeedSize); + + +/* simple short-cut to pre-selected XXH3_128bits variant */ +XXH_PUBLIC_API XXH128_hash_t XXH128(const void* data, size_t len, XXH64_hash_t seed); + + +#endif /* XXH_NO_LONG_LONG */ + + #if defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) # define XXH_IMPLEMENTATION #endif @@ -672,17 +748,23 @@ XXH_PUBLIC_API XXH128_hash_t XXH128_hashFromCanonical(const XXH128_canonical_t* /*-********************************************************************** * xxHash implementation *-********************************************************************** - * xxHash's implementation used to be found in xxhash.c. + * xxHash's implementation used to be hosted inside xxhash.c. * - * However, code inlining requires the implementation to be visible to the - * compiler, usually within the header. + * However, inlining requires implementation to be visible to the compiler, + * hence be included alongside the header. + * Previously, implementation was hosted inside xxhash.c, + * which was then #included when inlining was activated. + * This construction created issues with a few build and install systems, + * as it required xxhash.c to be stored in /include directory. * - * As a workaround, xxhash.c used to be included within xxhash.h. This caused - * some issues with some build systems, especially ones which treat .c files - * as source files. + * xxHash implementation is now directly integrated within xxhash.h. + * As a consequence, xxhash.c is no longer needed in /include. * - * Therefore, the implementation is now directly integrated within xxhash.h. - * Another small advantage is that xxhash.c is no longer needed in /include. + * xxhash.c is still available and is still useful. + * In a "normal" setup, when xxhash is not inlined, + * xxhash.h only exposes the prototypes and public symbols, + * while xxhash.c can be built into an object file xxhash.o + * which can then be linked into the final binary. ************************************************************************/ #if ( defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) \ @@ -700,10 +782,10 @@ XXH_PUBLIC_API XXH128_hash_t XXH128_hashFromCanonical(const XXH128_canonical_t* * Unfortunately, on some target/compiler combinations, the generated assembly * is sub-optimal. * - * The below switch allow to select a different access method for improved - * performance. + * The below switch allow selection of a different access method + * in the search for improved performance. * Method 0 (default): - * Use `memcpy()`. Safe and portable. + * Use `memcpy()`. Safe and portable. Default. * Method 1: * `__attribute__((packed))` statement. It depends on compiler extensions * and is therefore not portable. @@ -715,7 +797,7 @@ XXH_PUBLIC_API XXH128_hash_t XXH128_hashFromCanonical(const XXH128_canonical_t* * It can generate buggy code on targets which do not support unaligned * memory accesses. * But in some circumstances, it's the only known way to get the most - * performance (ie GCC + ARMv6) + * performance (example: GCC + ARMv6) * Method 3: * Byteshift. This can generate the best code on old compilers which don't * inline small `memcpy()` calls, and it might also be faster on big-endian @@ -733,7 +815,7 @@ XXH_PUBLIC_API XXH128_hash_t XXH128_hashFromCanonical(const XXH128_canonical_t* #endif /*! - *XXH_ACCEPT_NULL_INPUT_POINTER: + * XXH_ACCEPT_NULL_INPUT_POINTER: * If the input pointer is NULL, xxHash's default behavior is to dereference it, * triggering a segfault. * When this macro is enabled, xxHash actively checks the input for a null pointer. @@ -745,16 +827,26 @@ XXH_PUBLIC_API XXH128_hash_t XXH128_hashFromCanonical(const XXH128_canonical_t* /*! * XXH_FORCE_ALIGN_CHECK: - * This is a minor performance trick, only useful with lots of very small keys. - * It means: check for aligned/unaligned input. - * The check costs one initial branch per hash; - * Set it to 0 when the input is guaranteed to be aligned or when alignment - * doesn't matter for performance. + * This is an important performance trick + * for architectures without decent unaligned memory access performance. + * It checks for input alignment, and when conditions are met, + * uses a "fast path" employing direct 32-bit/64-bit read, + * resulting in _dramatically faster_ read speed. * - * This option does not affect XXH3. + * The check costs one initial branch per hash, which is generally negligible, but not zero. + * Moreover, it's not useful to generate binary for an additional code path + * if memory access uses same instruction for both aligned and unaligned adresses. + * + * In these cases, the alignment check can be removed by setting this macro to 0. + * Then the code will always use unaligned memory access. + * Align check is automatically disabled on x86, x64 & arm64, + * which are platforms known to offer good unaligned memory accesses performance. + * + * This option does not affect XXH3 (only XXH32 and XXH64). */ -#ifndef XXH_FORCE_ALIGN_CHECK /* can be defined externally */ -# if defined(__i386) || defined(_M_IX86) || defined(__x86_64__) || defined(_M_X64) +#ifndef XXH_FORCE_ALIGN_CHECK /* can be defined externally */ +# if defined(__i386) || defined(__x86_64__) || defined(__aarch64__) \ + || defined(_M_IX86) || defined(_M_X64) || defined(_M_ARM64) /* visual */ # define XXH_FORCE_ALIGN_CHECK 0 # else # define XXH_FORCE_ALIGN_CHECK 1 @@ -809,12 +901,14 @@ XXH_PUBLIC_API XXH128_hash_t XXH128_hashFromCanonical(const XXH128_canonical_t* * Includes & Memory related functions ***************************************/ /*! - * Modify the local functions below should you wish to use some other memory - * routines for malloc() and free() + * Modify the local functions below should you wish to use + * different memory routines for malloc() and free() */ #include <stdlib.h> + static void* XXH_malloc(size_t s) { return malloc(s); } -static void XXH_free (void* p) { free(p); } +static void XXH_free(void* p) { free(p); } + /*! and for memcpy() */ #include <string.h> static void* XXH_memcpy(void* dest, const void* src, size_t size) @@ -832,26 +926,27 @@ static void* XXH_memcpy(void* dest, const void* src, size_t size) # pragma warning(disable : 4127) /* disable: C4127: conditional expression is constant */ #endif -#if XXH_NO_INLINE_HINTS /* disable inlining hints */ -# define XXH_FORCE_INLINE static +#if XXH_NO_INLINE_HINTS /* disable inlining hints */ +# if defined(__GNUC__) +# define XXH_FORCE_INLINE static __attribute__((unused)) +# else +# define XXH_FORCE_INLINE static +# endif # define XXH_NO_INLINE static -#elif defined(_MSC_VER) /* Visual Studio */ +/* enable inlining hints */ +#elif defined(_MSC_VER) /* Visual Studio */ # define XXH_FORCE_INLINE static __forceinline # define XXH_NO_INLINE static __declspec(noinline) +#elif defined(__GNUC__) +# define XXH_FORCE_INLINE static __inline__ __attribute__((always_inline, unused)) +# define XXH_NO_INLINE static __attribute__((noinline)) +#elif defined (__cplusplus) \ + || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L)) /* C99 */ +# define XXH_FORCE_INLINE static inline +# define XXH_NO_INLINE static #else -# if defined (__cplusplus) \ - || defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L /* C99 */ -# ifdef __GNUC__ -# define XXH_FORCE_INLINE static inline __attribute__((always_inline)) -# define XXH_NO_INLINE static __attribute__((noinline)) -# else -# define XXH_FORCE_INLINE static inline -# define XXH_NO_INLINE static -# endif -# else -# define XXH_FORCE_INLINE static -# define XXH_NO_INLINE static -# endif /* __STDC_VERSION__ */ +# define XXH_FORCE_INLINE static +# define XXH_NO_INLINE static #endif @@ -860,14 +955,18 @@ static void* XXH_memcpy(void* dest, const void* src, size_t size) * Debug ***************************************/ /* - * DEBUGLEVEL is expected to be defined externally, typically via the compiler's - * command line options. The value must be a number. + * XXH_DEBUGLEVEL is expected to be defined externally, typically via the + * compiler's command line options. The value must be a number. */ -#ifndef DEBUGLEVEL -# define DEBUGLEVEL 0 +#ifndef XXH_DEBUGLEVEL +# ifdef DEBUGLEVEL /* backwards compat */ +# define XXH_DEBUGLEVEL DEBUGLEVEL +# else +# define XXH_DEBUGLEVEL 0 +# endif #endif -#if (DEBUGLEVEL>=1) +#if (XXH_DEBUGLEVEL>=1) # include <assert.h> /* note: can still be disabled with NDEBUG */ # define XXH_ASSERT(c) assert(c) #else @@ -875,7 +974,7 @@ static void* XXH_memcpy(void* dest, const void* src, size_t size) #endif /* note: use after variable declarations */ -#define XXH_STATIC_ASSERT(c) { enum { XXH_sa = 1/(int)(!!(c)) }; } +#define XXH_STATIC_ASSERT(c) do { enum { XXH_sa = 1/(int)(!!(c)) }; } while (0) /* ************************************* @@ -885,12 +984,17 @@ static void* XXH_memcpy(void* dest, const void* src, size_t size) && (defined (__cplusplus) \ || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) ) # include <stdint.h> - typedef uint8_t xxh_u8; + typedef uint8_t xxh_u8; #else - typedef unsigned char xxh_u8; + typedef unsigned char xxh_u8; #endif typedef XXH32_hash_t xxh_u32; +#ifdef XXH_OLD_NAMES +# define BYTE xxh_u8 +# define U8 xxh_u8 +# define U32 xxh_u32 +#endif /* *** Memory access *** */ @@ -915,8 +1019,14 @@ static xxh_u32 XXH_read32(const void* memPtr) { return *(const xxh_u32*) memPtr; * * Currently only defined for GCC and ICC. */ +#ifdef XXH_OLD_NAMES typedef union { xxh_u32 u32; } __attribute__((packed)) unalign; -static xxh_u32 XXH_read32(const void* ptr) { return ((const unalign*)ptr)->u32; } +#endif +static xxh_u32 XXH_read32(const void* ptr) +{ + typedef union { xxh_u32 u32; } __attribute__((packed)) xxh_unalign; + return ((const xxh_unalign*)ptr)->u32; +} #else @@ -958,10 +1068,13 @@ typedef enum { XXH_bigEndian=0, XXH_littleEndian=1 } XXH_endianess; || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) # define XXH_CPU_LITTLE_ENDIAN 0 # else +/* + * runtime test, presumed to simplify to a constant by compiler + */ static int XXH_isLittleEndian(void) { /* - * Nonstandard, but well-defined behavior in practice. + * Portable and well-defined behavior. * Don't use static: it is detrimental to performance. */ const union { xxh_u32 u; xxh_u8 c[4]; } one = { 1 }; @@ -979,12 +1092,14 @@ static int XXH_isLittleEndian(void) ******************************************/ #define XXH_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__) -#ifndef __has_builtin -# define __has_builtin(x) 0 +#ifdef __has_builtin +# define XXH_HAS_BUILTIN(x) __has_builtin(x) +#else +# define XXH_HAS_BUILTIN(x) 0 #endif -#if !defined(NO_CLANG_BUILTIN) && __has_builtin(__builtin_rotateleft32) \ - && __has_builtin(__builtin_rotateleft64) +#if !defined(NO_CLANG_BUILTIN) && XXH_HAS_BUILTIN(__builtin_rotateleft32) \ + && XXH_HAS_BUILTIN(__builtin_rotateleft64) # define XXH_rotl32 __builtin_rotateleft32 # define XXH_rotl64 __builtin_rotateleft64 /* Note: although _rotl exists for minGW (GCC under windows), performance seems poor */ @@ -1073,17 +1188,25 @@ XXH_PUBLIC_API unsigned XXH_versionNumber (void) { return XXH_VERSION_NUMBER; } /* ******************************************************************* * 32-bit hash functions *********************************************************************/ -static const xxh_u32 PRIME32_1 = 0x9E3779B1U; /* 0b10011110001101110111100110110001 */ -static const xxh_u32 PRIME32_2 = 0x85EBCA77U; /* 0b10000101111010111100101001110111 */ -static const xxh_u32 PRIME32_3 = 0xC2B2AE3DU; /* 0b11000010101100101010111000111101 */ -static const xxh_u32 PRIME32_4 = 0x27D4EB2FU; /* 0b00100111110101001110101100101111 */ -static const xxh_u32 PRIME32_5 = 0x165667B1U; /* 0b00010110010101100110011110110001 */ +static const xxh_u32 XXH_PRIME32_1 = 0x9E3779B1U; /* 0b10011110001101110111100110110001 */ +static const xxh_u32 XXH_PRIME32_2 = 0x85EBCA77U; /* 0b10000101111010111100101001110111 */ +static const xxh_u32 XXH_PRIME32_3 = 0xC2B2AE3DU; /* 0b11000010101100101010111000111101 */ +static const xxh_u32 XXH_PRIME32_4 = 0x27D4EB2FU; /* 0b00100111110101001110101100101111 */ +static const xxh_u32 XXH_PRIME32_5 = 0x165667B1U; /* 0b00010110010101100110011110110001 */ + +#ifdef XXH_OLD_NAMES +# define PRIME32_1 XXH_PRIME32_1 +# define PRIME32_2 XXH_PRIME32_2 +# define PRIME32_3 XXH_PRIME32_3 +# define PRIME32_4 XXH_PRIME32_4 +# define PRIME32_5 XXH_PRIME32_5 +#endif static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input) { - acc += input * PRIME32_2; + acc += input * XXH_PRIME32_2; acc = XXH_rotl32(acc, 13); - acc *= PRIME32_1; + acc *= XXH_PRIME32_1; #if defined(__GNUC__) && defined(__SSE4_1__) && !defined(XXH_ENABLE_AUTOVECTORIZE) /* * UGLY HACK: @@ -1139,9 +1262,9 @@ static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input) static xxh_u32 XXH32_avalanche(xxh_u32 h32) { h32 ^= h32 >> 15; - h32 *= PRIME32_2; + h32 *= XXH_PRIME32_2; h32 ^= h32 >> 13; - h32 *= PRIME32_3; + h32 *= XXH_PRIME32_3; h32 ^= h32 >> 16; return(h32); } @@ -1151,64 +1274,66 @@ static xxh_u32 XXH32_avalanche(xxh_u32 h32) static xxh_u32 XXH32_finalize(xxh_u32 h32, const xxh_u8* ptr, size_t len, XXH_alignment align) { -#define PROCESS1 \ - h32 += (*ptr++) * PRIME32_5; \ - h32 = XXH_rotl32(h32, 11) * PRIME32_1 ; +#define XXH_PROCESS1 do { \ + h32 += (*ptr++) * XXH_PRIME32_5; \ + h32 = XXH_rotl32(h32, 11) * XXH_PRIME32_1; \ +} while (0) -#define PROCESS4 \ - h32 += XXH_get32bits(ptr) * PRIME32_3; \ - ptr+=4; \ - h32 = XXH_rotl32(h32, 17) * PRIME32_4 ; +#define XXH_PROCESS4 do { \ + h32 += XXH_get32bits(ptr) * XXH_PRIME32_3; \ + ptr += 4; \ + h32 = XXH_rotl32(h32, 17) * XXH_PRIME32_4; \ +} while (0) /* Compact rerolled version */ if (XXH_REROLL) { len &= 15; while (len >= 4) { - PROCESS4; + XXH_PROCESS4; len -= 4; } while (len > 0) { - PROCESS1; + XXH_PROCESS1; --len; } return XXH32_avalanche(h32); } else { switch(len&15) /* or switch(bEnd - p) */ { - case 12: PROCESS4; + case 12: XXH_PROCESS4; /* fallthrough */ - case 8: PROCESS4; + case 8: XXH_PROCESS4; /* fallthrough */ - case 4: PROCESS4; + case 4: XXH_PROCESS4; return XXH32_avalanche(h32); - case 13: PROCESS4; + case 13: XXH_PROCESS4; /* fallthrough */ - case 9: PROCESS4; + case 9: XXH_PROCESS4; /* fallthrough */ - case 5: PROCESS4; - PROCESS1; + case 5: XXH_PROCESS4; + XXH_PROCESS1; return XXH32_avalanche(h32); - case 14: PROCESS4; + case 14: XXH_PROCESS4; /* fallthrough */ - case 10: PROCESS4; + case 10: XXH_PROCESS4; /* fallthrough */ - case 6: PROCESS4; - PROCESS1; - PROCESS1; + case 6: XXH_PROCESS4; + XXH_PROCESS1; + XXH_PROCESS1; return XXH32_avalanche(h32); - case 15: PROCESS4; + case 15: XXH_PROCESS4; /* fallthrough */ - case 11: PROCESS4; + case 11: XXH_PROCESS4; /* fallthrough */ - case 7: PROCESS4; + case 7: XXH_PROCESS4; /* fallthrough */ - case 3: PROCESS1; + case 3: XXH_PROCESS1; /* fallthrough */ - case 2: PROCESS1; + case 2: XXH_PROCESS1; /* fallthrough */ - case 1: PROCESS1; + case 1: XXH_PROCESS1; /* fallthrough */ case 0: return XXH32_avalanche(h32); } @@ -1217,6 +1342,14 @@ XXH32_finalize(xxh_u32 h32, const xxh_u8* ptr, size_t len, XXH_alignment align) } } +#ifdef XXH_OLD_NAMES +# define PROCESS1 XXH_PROCESS1 +# define PROCESS4 XXH_PROCESS4 +#else +# undef XXH_PROCESS1 +# undef XXH_PROCESS4 +#endif + XXH_FORCE_INLINE xxh_u32 XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment align) { @@ -1232,10 +1365,10 @@ XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment if (len>=16) { const xxh_u8* const limit = bEnd - 15; - xxh_u32 v1 = seed + PRIME32_1 + PRIME32_2; - xxh_u32 v2 = seed + PRIME32_2; + xxh_u32 v1 = seed + XXH_PRIME32_1 + XXH_PRIME32_2; + xxh_u32 v2 = seed + XXH_PRIME32_2; xxh_u32 v3 = seed + 0; - xxh_u32 v4 = seed - PRIME32_1; + xxh_u32 v4 = seed - XXH_PRIME32_1; do { v1 = XXH32_round(v1, XXH_get32bits(input)); input += 4; @@ -1247,7 +1380,7 @@ XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment h32 = XXH_rotl32(v1, 1) + XXH_rotl32(v2, 7) + XXH_rotl32(v3, 12) + XXH_rotl32(v4, 18); } else { - h32 = seed + PRIME32_5; + h32 = seed + XXH_PRIME32_5; } h32 += (xxh_u32)len; @@ -1299,10 +1432,10 @@ XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, XXH32_hash_t s { XXH32_state_t state; /* using a local state to memcpy() in order to avoid strict-aliasing warnings */ memset(&state, 0, sizeof(state)); - state.v1 = seed + PRIME32_1 + PRIME32_2; - state.v2 = seed + PRIME32_2; + state.v1 = seed + XXH_PRIME32_1 + XXH_PRIME32_2; + state.v2 = seed + XXH_PRIME32_2; state.v3 = seed + 0; - state.v4 = seed - PRIME32_1; + state.v4 = seed - XXH_PRIME32_1; /* do not write into reserved, planned to be removed in a future version */ memcpy(statePtr, &state, sizeof(state) - sizeof(state.reserved)); return XXH_OK; @@ -1383,7 +1516,7 @@ XXH_PUBLIC_API XXH32_hash_t XXH32_digest (const XXH32_state_t* state) + XXH_rotl32(state->v3, 12) + XXH_rotl32(state->v4, 18); } else { - h32 = state->v3 /* == seed */ + PRIME32_5; + h32 = state->v3 /* == seed */ + XXH_PRIME32_5; } h32 += state->total_len_32; @@ -1430,6 +1563,9 @@ XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src typedef XXH64_hash_t xxh_u64; +#ifdef XXH_OLD_NAMES +# define U64 xxh_u64 +#endif /*! * XXH_REROLL_XXH64: @@ -1478,8 +1614,14 @@ static xxh_u64 XXH_read64(const void* memPtr) { return *(const xxh_u64*) memPtr; * * Currently only defined for GCC and ICC. */ +#ifdef XXH_OLD_NAMES typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) unalign64; -static xxh_u64 XXH_read64(const void* ptr) { return ((const unalign64*)ptr)->u64; } +#endif +static xxh_u64 XXH_read64(const void* ptr) +{ + typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) xxh_unalign64; + return ((const xxh_unalign64*)ptr)->u64; +} #else @@ -1568,17 +1710,25 @@ XXH_readLE64_align(const void* ptr, XXH_alignment align) /******* xxh64 *******/ -static const xxh_u64 PRIME64_1 = 0x9E3779B185EBCA87ULL; /* 0b1001111000110111011110011011000110000101111010111100101010000111 */ -static const xxh_u64 PRIME64_2 = 0xC2B2AE3D27D4EB4FULL; /* 0b1100001010110010101011100011110100100111110101001110101101001111 */ -static const xxh_u64 PRIME64_3 = 0x165667B19E3779F9ULL; /* 0b0001011001010110011001111011000110011110001101110111100111111001 */ -static const xxh_u64 PRIME64_4 = 0x85EBCA77C2B2AE63ULL; /* 0b1000010111101011110010100111011111000010101100101010111001100011 */ -static const xxh_u64 PRIME64_5 = 0x27D4EB2F165667C5ULL; /* 0b0010011111010100111010110010111100010110010101100110011111000101 */ +static const xxh_u64 XXH_PRIME64_1 = 0x9E3779B185EBCA87ULL; /* 0b1001111000110111011110011011000110000101111010111100101010000111 */ +static const xxh_u64 XXH_PRIME64_2 = 0xC2B2AE3D27D4EB4FULL; /* 0b1100001010110010101011100011110100100111110101001110101101001111 */ +static const xxh_u64 XXH_PRIME64_3 = 0x165667B19E3779F9ULL; /* 0b0001011001010110011001111011000110011110001101110111100111111001 */ +static const xxh_u64 XXH_PRIME64_4 = 0x85EBCA77C2B2AE63ULL; /* 0b1000010111101011110010100111011111000010101100101010111001100011 */ +static const xxh_u64 XXH_PRIME64_5 = 0x27D4EB2F165667C5ULL; /* 0b0010011111010100111010110010111100010110010101100110011111000101 */ + +#ifdef XXH_OLD_NAMES +# define PRIME64_1 XXH_PRIME64_1 +# define PRIME64_2 XXH_PRIME64_2 +# define PRIME64_3 XXH_PRIME64_3 +# define PRIME64_4 XXH_PRIME64_4 +# define PRIME64_5 XXH_PRIME64_5 +#endif static xxh_u64 XXH64_round(xxh_u64 acc, xxh_u64 input) { - acc += input * PRIME64_2; + acc += input * XXH_PRIME64_2; acc = XXH_rotl64(acc, 31); - acc *= PRIME64_1; + acc *= XXH_PRIME64_1; return acc; } @@ -1586,16 +1736,16 @@ static xxh_u64 XXH64_mergeRound(xxh_u64 acc, xxh_u64 val) { val = XXH64_round(0, val); acc ^= val; - acc = acc * PRIME64_1 + PRIME64_4; + acc = acc * XXH_PRIME64_1 + XXH_PRIME64_4; return acc; } static xxh_u64 XXH64_avalanche(xxh_u64 h64) { h64 ^= h64 >> 33; - h64 *= PRIME64_2; + h64 *= XXH_PRIME64_2; h64 ^= h64 >> 29; - h64 *= PRIME64_3; + h64 *= XXH_PRIME64_3; h64 ^= h64 >> 32; return h64; } @@ -1606,117 +1756,119 @@ static xxh_u64 XXH64_avalanche(xxh_u64 h64) static xxh_u64 XXH64_finalize(xxh_u64 h64, const xxh_u8* ptr, size_t len, XXH_alignment align) { -#define PROCESS1_64 \ - h64 ^= (*ptr++) * PRIME64_5; \ - h64 = XXH_rotl64(h64, 11) * PRIME64_1; - -#define PROCESS4_64 \ - h64 ^= (xxh_u64)(XXH_get32bits(ptr)) * PRIME64_1; \ - ptr+=4; \ - h64 = XXH_rotl64(h64, 23) * PRIME64_2 + PRIME64_3; - -#define PROCESS8_64 { \ +#define XXH_PROCESS1_64 do { \ + h64 ^= (*ptr++) * XXH_PRIME64_5; \ + h64 = XXH_rotl64(h64, 11) * XXH_PRIME64_1; \ +} while (0) + +#define XXH_PROCESS4_64 do { \ + h64 ^= (xxh_u64)(XXH_get32bits(ptr)) * XXH_PRIME64_1; \ + ptr += 4; \ + h64 = XXH_rotl64(h64, 23) * XXH_PRIME64_2 + XXH_PRIME64_3; \ +} while (0) + +#define XXH_PROCESS8_64 do { \ xxh_u64 const k1 = XXH64_round(0, XXH_get64bits(ptr)); \ - ptr+=8; \ - h64 ^= k1; \ - h64 = XXH_rotl64(h64,27) * PRIME64_1 + PRIME64_4; \ -} + ptr += 8; \ + h64 ^= k1; \ + h64 = XXH_rotl64(h64,27) * XXH_PRIME64_1 + XXH_PRIME64_4; \ +} while (0) /* Rerolled version for 32-bit targets is faster and much smaller. */ if (XXH_REROLL || XXH_REROLL_XXH64) { len &= 31; while (len >= 8) { - PROCESS8_64; + XXH_PROCESS8_64; len -= 8; } if (len >= 4) { - PROCESS4_64; + XXH_PROCESS4_64; len -= 4; } while (len > 0) { - PROCESS1_64; + XXH_PROCESS1_64; --len; } return XXH64_avalanche(h64); } else { switch(len & 31) { - case 24: PROCESS8_64; + case 24: XXH_PROCESS8_64; /* fallthrough */ - case 16: PROCESS8_64; + case 16: XXH_PROCESS8_64; /* fallthrough */ - case 8: PROCESS8_64; + case 8: XXH_PROCESS8_64; return XXH64_avalanche(h64); - case 28: PROCESS8_64; + case 28: XXH_PROCESS8_64; /* fallthrough */ - case 20: PROCESS8_64; + case 20: XXH_PROCESS8_64; /* fallthrough */ - case 12: PROCESS8_64; + case 12: XXH_PROCESS8_64; /* fallthrough */ - case 4: PROCESS4_64; + case 4: XXH_PROCESS4_64; return XXH64_avalanche(h64); - case 25: PROCESS8_64; + case 25: XXH_PROCESS8_64; /* fallthrough */ - case 17: PROCESS8_64; + case 17: XXH_PROCESS8_64; /* fallthrough */ - case 9: PROCESS8_64; - PROCESS1_64; + case 9: XXH_PROCESS8_64; + XXH_PROCESS1_64; return XXH64_avalanche(h64); - case 29: PROCESS8_64; + case 29: XXH_PROCESS8_64; /* fallthrough */ - case 21: PROCESS8_64; + case 21: XXH_PROCESS8_64; /* fallthrough */ - case 13: PROCESS8_64; + case 13: XXH_PROCESS8_64; /* fallthrough */ - case 5: PROCESS4_64; - PROCESS1_64; + case 5: XXH_PROCESS4_64; + XXH_PROCESS1_64; return XXH64_avalanche(h64); - case 26: PROCESS8_64; + case 26: XXH_PROCESS8_64; /* fallthrough */ - case 18: PROCESS8_64; + case 18: XXH_PROCESS8_64; /* fallthrough */ - case 10: PROCESS8_64; - PROCESS1_64; - PROCESS1_64; + case 10: XXH_PROCESS8_64; + XXH_PROCESS1_64; + XXH_PROCESS1_64; return XXH64_avalanche(h64); - case 30: PROCESS8_64; + case 30: XXH_PROCESS8_64; /* fallthrough */ - case 22: PROCESS8_64; + case 22: XXH_PROCESS8_64; /* fallthrough */ - case 14: PROCESS8_64; + case 14: XXH_PROCESS8_64; /* fallthrough */ - case 6: PROCESS4_64; - PROCESS1_64; - PROCESS1_64; + case 6: XXH_PROCESS4_64; + XXH_PROCESS1_64; + XXH_PROCESS1_64; return XXH64_avalanche(h64); - case 27: PROCESS8_64; + case 27: XXH_PROCESS8_64; /* fallthrough */ - case 19: PROCESS8_64; + case 19: XXH_PROCESS8_64; /* fallthrough */ - case 11: PROCESS8_64; - PROCESS1_64; - PROCESS1_64; - PROCESS1_64; + case 11: XXH_PROCESS8_64; + XXH_PROCESS1_64; + XXH_PROCESS1_64; + XXH_PROCESS1_64; return XXH64_avalanche(h64); - case 31: PROCESS8_64; + case 31: XXH_PROCESS8_64; /* fallthrough */ - case 23: PROCESS8_64; + case 23: XXH_PROCESS8_64; /* fallthrough */ - case 15: PROCESS8_64; + case 15: XXH_PROCESS8_64; /* fallthrough */ - case 7: PROCESS4_64; + case 7: XXH_PROCESS4_64; /* fallthrough */ - case 3: PROCESS1_64; + case 3: XXH_PROCESS1_64; /* fallthrough */ - case 2: PROCESS1_64; + case 2: XXH_PROCESS1_64; /* fallthrough */ - case 1: PROCESS1_64; + case 1: XXH_PROCESS1_64; /* fallthrough */ case 0: return XXH64_avalanche(h64); } @@ -1726,6 +1878,16 @@ XXH64_finalize(xxh_u64 h64, const xxh_u8* ptr, size_t len, XXH_alignment align) return 0; /* unreachable, but some compilers complain without it */ } +#ifdef XXH_OLD_NAMES +# define PROCESS1_64 XXH_PROCESS1_64 +# define PROCESS4_64 XXH_PROCESS4_64 +# define PROCESS8_64 XXH_PROCESS8_64 +#else +# undef XXH_PROCESS1_64 +# undef XXH_PROCESS4_64 +# undef XXH_PROCESS8_64 +#endif + XXH_FORCE_INLINE xxh_u64 XXH64_endian_align(const xxh_u8* input, size_t len, xxh_u64 seed, XXH_alignment align) { @@ -1741,10 +1903,10 @@ XXH64_endian_align(const xxh_u8* input, size_t len, xxh_u64 seed, XXH_alignment if (len>=32) { const xxh_u8* const limit = bEnd - 32; - xxh_u64 v1 = seed + PRIME64_1 + PRIME64_2; - xxh_u64 v2 = seed + PRIME64_2; + xxh_u64 v1 = seed + XXH_PRIME64_1 + XXH_PRIME64_2; + xxh_u64 v2 = seed + XXH_PRIME64_2; xxh_u64 v3 = seed + 0; - xxh_u64 v4 = seed - PRIME64_1; + xxh_u64 v4 = seed - XXH_PRIME64_1; do { v1 = XXH64_round(v1, XXH_get64bits(input)); input+=8; @@ -1760,7 +1922,7 @@ XXH64_endian_align(const xxh_u8* input, size_t len, xxh_u64 seed, XXH_alignment h64 = XXH64_mergeRound(h64, v4); } else { - h64 = seed + PRIME64_5; + h64 = seed + XXH_PRIME64_5; } h64 += (xxh_u64) len; @@ -1811,10 +1973,10 @@ XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH64_state_t* statePtr, XXH64_hash_t s { XXH64_state_t state; /* use a local state to memcpy() in order to avoid strict-aliasing warnings */ memset(&state, 0, sizeof(state)); - state.v1 = seed + PRIME64_1 + PRIME64_2; - state.v2 = seed + PRIME64_2; + state.v1 = seed + XXH_PRIME64_1 + XXH_PRIME64_2; + state.v2 = seed + XXH_PRIME64_2; state.v3 = seed + 0; - state.v4 = seed - PRIME64_1; + state.v4 = seed - XXH_PRIME64_1; /* do not write into reserved64, might be removed in a future version */ memcpy(statePtr, &state, sizeof(state) - sizeof(state.reserved64)); return XXH_OK; @@ -1897,7 +2059,7 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_digest (const XXH64_state_t* state) h64 = XXH64_mergeRound(h64, v3); h64 = XXH64_mergeRound(h64, v4); } else { - h64 = state->v3 /*seed*/ + PRIME64_5; + h64 = state->v3 /*seed*/ + XXH_PRIME64_5; } h64 += (xxh_u64) state->total_len; @@ -1927,8 +2089,2671 @@ XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t* src * New generation hash designed for speed on small keys and vectorization ************************************************************************ */ -#include "xxh3.h" +/* === Compiler specifics === */ + +#if defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L /* >= C99 */ +# define XXH_RESTRICT restrict +#else +/* Note: it might be useful to define __restrict or __restrict__ for some C++ compilers */ +# define XXH_RESTRICT /* disable */ +#endif + +#if (defined(__GNUC__) && (__GNUC__ >= 3)) \ + || (defined(__INTEL_COMPILER) && (__INTEL_COMPILER >= 800)) \ + || defined(__clang__) +# define XXH_likely(x) __builtin_expect(x, 1) +# define XXH_unlikely(x) __builtin_expect(x, 0) +#else +# define XXH_likely(x) (x) +# define XXH_unlikely(x) (x) +#endif + +#if defined(__GNUC__) +# if defined(__AVX2__) +# include <immintrin.h> +# elif defined(__SSE2__) +# include <emmintrin.h> +# elif defined(__ARM_NEON__) || defined(__ARM_NEON) +# define inline __inline__ /* circumvent a clang bug */ +# include <arm_neon.h> +# undef inline +# endif +#elif defined(_MSC_VER) +# include <intrin.h> +#endif + +/* + * One goal of XXH3 is to make it fast on both 32-bit and 64-bit, while + * remaining a true 64-bit/128-bit hash function. + * + * This is done by prioritizing a subset of 64-bit operations that can be + * emulated without too many steps on the average 32-bit machine. + * + * For example, these two lines seem similar, and run equally fast on 64-bit: + * + * xxh_u64 x; + * x ^= (x >> 47); // good + * x ^= (x >> 13); // bad + * + * However, to a 32-bit machine, there is a major difference. + * + * x ^= (x >> 47) looks like this: + * + * x.lo ^= (x.hi >> (47 - 32)); + * + * while x ^= (x >> 13) looks like this: + * + * // note: funnel shifts are not usually cheap. + * x.lo ^= (x.lo >> 13) | (x.hi << (32 - 13)); + * x.hi ^= (x.hi >> 13); + * + * The first one is significantly faster than the second, simply because the + * shift is larger than 32. This means: + * - All the bits we need are in the upper 32 bits, so we can ignore the lower + * 32 bits in the shift. + * - The shift result will always fit in the lower 32 bits, and therefore, + * we can ignore the upper 32 bits in the xor. + * + * Thanks to this optimization, XXH3 only requires these features to be efficient: + * + * - Usable unaligned access + * - A 32-bit or 64-bit ALU + * - If 32-bit, a decent ADC instruction + * - A 32 or 64-bit multiply with a 64-bit result + * - For the 128-bit variant, a decent byteswap helps short inputs. + * + * The first two are already required by XXH32, and almost all 32-bit and 64-bit + * platforms which can run XXH32 can run XXH3 efficiently. + * + * Thumb-1, the classic 16-bit only subset of ARM's instruction set, is one + * notable exception. + * + * First of all, Thumb-1 lacks support for the UMULL instruction which + * performs the important long multiply. This means numerous __aeabi_lmul + * calls. + * + * Second of all, the 8 functional registers are just not enough. + * Setup for __aeabi_lmul, byteshift loads, pointers, and all arithmetic need + * Lo registers, and this shuffling results in thousands more MOVs than A32. + * + * A32 and T32 don't have this limitation. They can access all 14 registers, + * do a 32->64 multiply with UMULL, and the flexible operand allowing free + * shifts is helpful, too. + * + * Therefore, we do a quick sanity check. + * + * If compiling Thumb-1 for a target which supports ARM instructions, we will + * emit a warning, as it is not a "sane" platform to compile for. + * + * Usually, if this happens, it is because of an accident and you probably need + * to specify -march, as you likely meant to compile for a newer architecture. + * + * Credit: large sections of the vectorial and asm source code paths + * have been contributed by @easyaspi314 + */ +#if defined(__thumb__) && !defined(__thumb2__) && defined(__ARM_ARCH_ISA_ARM) +# warning "XXH3 is highly inefficient without ARM or Thumb-2." +#endif + +/* ========================================== + * Vectorization detection + * ========================================== */ +#define XXH_SCALAR 0 /* Portable scalar version */ +#define XXH_SSE2 1 /* SSE2 for Pentium 4 and all x86_64 */ +#define XXH_AVX2 2 /* AVX2 for Haswell and Bulldozer */ +#define XXH_AVX512 3 /* AVX512 for Skylake and Icelake */ +#define XXH_NEON 4 /* NEON for most ARMv7-A and all AArch64 */ +#define XXH_VSX 5 /* VSX and ZVector for POWER8/z13 */ + +#ifndef XXH_VECTOR /* can be defined on command line */ +# if defined(__AVX512F__) +# define XXH_VECTOR XXH_AVX512 +# elif defined(__AVX2__) +# define XXH_VECTOR XXH_AVX2 +# elif defined(__SSE2__) || defined(_M_AMD64) || defined(_M_X64) || (defined(_M_IX86_FP) && (_M_IX86_FP == 2)) +# define XXH_VECTOR XXH_SSE2 +# elif defined(__GNUC__) /* msvc support maybe later */ \ + && (defined(__ARM_NEON__) || defined(__ARM_NEON)) \ + && (defined(__LITTLE_ENDIAN__) /* We only support little endian NEON */ \ + || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)) +# define XXH_VECTOR XXH_NEON +# elif (defined(__PPC64__) && defined(__POWER8_VECTOR__)) \ + || (defined(__s390x__) && defined(__VEC__)) \ + && defined(__GNUC__) /* TODO: IBM XL */ +# define XXH_VECTOR XXH_VSX +# else +# define XXH_VECTOR XXH_SCALAR +# endif +#endif + +/* + * Controls the alignment of the accumulator, + * for compatibility with aligned vector loads, which are usually faster. + */ +#ifndef XXH_ACC_ALIGN +# if defined(XXH_X86DISPATCH) +# define XXH_ACC_ALIGN 64 /* for compatibility with avx512 */ +# elif XXH_VECTOR == XXH_SCALAR /* scalar */ +# define XXH_ACC_ALIGN 8 +# elif XXH_VECTOR == XXH_SSE2 /* sse2 */ +# define XXH_ACC_ALIGN 16 +# elif XXH_VECTOR == XXH_AVX2 /* avx2 */ +# define XXH_ACC_ALIGN 32 +# elif XXH_VECTOR == XXH_NEON /* neon */ +# define XXH_ACC_ALIGN 16 +# elif XXH_VECTOR == XXH_VSX /* vsx */ +# define XXH_ACC_ALIGN 16 +# elif XXH_VECTOR == XXH_AVX512 /* avx512 */ +# define XXH_ACC_ALIGN 64 +# endif +#endif + +#if defined(XXH_X86DISPATCH) || XXH_VECTOR == XXH_SSE2 \ + || XXH_VECTOR == XXH_AVX2 || XXH_VECTOR == XXH_AVX512 +# define XXH_SEC_ALIGN XXH_ACC_ALIGN +#else +# define XXH_SEC_ALIGN 8 +#endif + +/* + * UGLY HACK: + * GCC usually generates the best code with -O3 for xxHash. + * + * However, when targeting AVX2, it is overzealous in its unrolling resulting + * in code roughly 3/4 the speed of Clang. + * + * There are other issues, such as GCC splitting _mm256_loadu_si256 into + * _mm_loadu_si128 + _mm256_inserti128_si256. This is an optimization which + * only applies to Sandy and Ivy Bridge... which don't even support AVX2. + * + * That is why when compiling the AVX2 version, it is recommended to use either + * -O2 -mavx2 -march=haswell + * or + * -O2 -mavx2 -mno-avx256-split-unaligned-load + * for decent performance, or to use Clang instead. + * + * Fortunately, we can control the first one with a pragma that forces GCC into + * -O2, but the other one we can't control without "failed to inline always + * inline function due to target mismatch" warnings. + */ +#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \ + && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \ + && defined(__OPTIMIZE__) && !defined(__OPTIMIZE_SIZE__) /* respect -O0 and -Os */ +# pragma GCC push_options +# pragma GCC optimize("-O2") +#endif + + +#if XXH_VECTOR == XXH_NEON +/* + * NEON's setup for vmlal_u32 is a little more complicated than it is on + * SSE2, AVX2, and VSX. + * + * While PMULUDQ and VMULEUW both perform a mask, VMLAL.U32 performs an upcast. + * + * To do the same operation, the 128-bit 'Q' register needs to be split into + * two 64-bit 'D' registers, performing this operation:: + * + * [ a | b ] + * | '---------. .--------' | + * | x | + * | .---------' '--------. | + * [ a & 0xFFFFFFFF | b & 0xFFFFFFFF ],[ a >> 32 | b >> 32 ] + * + * Due to significant changes in aarch64, the fastest method for aarch64 is + * completely different than the fastest method for ARMv7-A. + * + * ARMv7-A treats D registers as unions overlaying Q registers, so modifying + * D11 will modify the high half of Q5. This is similar to how modifying AH + * will only affect bits 8-15 of AX on x86. + * + * VZIP takes two registers, and puts even lanes in one register and odd lanes + * in the other. + * + * On ARMv7-A, this strangely modifies both parameters in place instead of + * taking the usual 3-operand form. + * + * Therefore, if we want to do this, we can simply use a D-form VZIP.32 on the + * lower and upper halves of the Q register to end up with the high and low + * halves where we want - all in one instruction. + * + * vzip.32 d10, d11 @ d10 = { d10[0], d11[0] }; d11 = { d10[1], d11[1] } + * + * Unfortunately we need inline assembly for this: Instructions modifying two + * registers at once is not possible in GCC or Clang's IR, and they have to + * create a copy. + * + * aarch64 requires a different approach. + * + * In order to make it easier to write a decent compiler for aarch64, many + * quirks were removed, such as conditional execution. + * + * NEON was also affected by this. + * + * aarch64 cannot access the high bits of a Q-form register, and writes to a + * D-form register zero the high bits, similar to how writes to W-form scalar + * registers (or DWORD registers on x86_64) work. + * + * The formerly free vget_high intrinsics now require a vext (with a few + * exceptions) + * + * Additionally, VZIP was replaced by ZIP1 and ZIP2, which are the equivalent + * of PUNPCKL* and PUNPCKH* in SSE, respectively, in order to only modify one + * operand. + * + * The equivalent of the VZIP.32 on the lower and upper halves would be this + * mess: + * + * ext v2.4s, v0.4s, v0.4s, #2 // v2 = { v0[2], v0[3], v0[0], v0[1] } + * zip1 v1.2s, v0.2s, v2.2s // v1 = { v0[0], v2[0] } + * zip2 v0.2s, v0.2s, v1.2s // v0 = { v0[1], v2[1] } + * + * Instead, we use a literal downcast, vmovn_u64 (XTN), and vshrn_n_u64 (SHRN): + * + * shrn v1.2s, v0.2d, #32 // v1 = (uint32x2_t)(v0 >> 32); + * xtn v0.2s, v0.2d // v0 = (uint32x2_t)(v0 & 0xFFFFFFFF); + * + * This is available on ARMv7-A, but is less efficient than a single VZIP.32. + */ + +/* + * Function-like macro: + * void XXH_SPLIT_IN_PLACE(uint64x2_t &in, uint32x2_t &outLo, uint32x2_t &outHi) + * { + * outLo = (uint32x2_t)(in & 0xFFFFFFFF); + * outHi = (uint32x2_t)(in >> 32); + * in = UNDEFINED; + * } + */ +# if !defined(XXH_NO_VZIP_HACK) /* define to disable */ \ + && defined(__GNUC__) \ + && !defined(__aarch64__) && !defined(__arm64__) +# define XXH_SPLIT_IN_PLACE(in, outLo, outHi) \ + do { \ + /* Undocumented GCC/Clang operand modifier: %e0 = lower D half, %f0 = upper D half */ \ + /* https://github.com/gcc-mirror/gcc/blob/38cf91e5/gcc/config/arm/arm.c#L22486 */ \ + /* https://github.com/llvm-mirror/llvm/blob/2c4ca683/lib/Target/ARM/ARMAsmPrinter.cpp#L399 */ \ + __asm__("vzip.32 %e0, %f0" : "+w" (in)); \ + (outLo) = vget_low_u32 (vreinterpretq_u32_u64(in)); \ + (outHi) = vget_high_u32(vreinterpretq_u32_u64(in)); \ + } while (0) +# else +# define XXH_SPLIT_IN_PLACE(in, outLo, outHi) \ + do { \ + (outLo) = vmovn_u64 (in); \ + (outHi) = vshrn_n_u64 ((in), 32); \ + } while (0) +# endif +#endif /* XXH_VECTOR == XXH_NEON */ + +/* + * VSX and Z Vector helpers. + * + * This is very messy, and any pull requests to clean this up are welcome. + * + * There are a lot of problems with supporting VSX and s390x, due to + * inconsistent intrinsics, spotty coverage, and multiple endiannesses. + */ +#if XXH_VECTOR == XXH_VSX +# if defined(__s390x__) +# include <s390intrin.h> +# else +/* gcc's altivec.h can have the unwanted consequence to unconditionally + * #define bool, vector, and pixel keywords, + * with bad consequences for programs already using these keywords for other purposes. + * The paragraph defining these macros is skipped when __APPLE_ALTIVEC__ is defined. + * __APPLE_ALTIVEC__ is _generally_ defined automatically by the compiler, + * but it seems that, in some cases, it isn't. + * Force the build macro to be defined, so that keywords are not altered. + */ +# if defined(__GNUC__) && !defined(__APPLE_ALTIVEC__) +# define __APPLE_ALTIVEC__ +# endif +# include <altivec.h> +# endif + +typedef __vector unsigned long long xxh_u64x2; +typedef __vector unsigned char xxh_u8x16; +typedef __vector unsigned xxh_u32x4; + +# ifndef XXH_VSX_BE +# if defined(__BIG_ENDIAN__) \ + || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) +# define XXH_VSX_BE 1 +# elif defined(__VEC_ELEMENT_REG_ORDER__) && __VEC_ELEMENT_REG_ORDER__ == __ORDER_BIG_ENDIAN__ +# warning "-maltivec=be is not recommended. Please use native endianness." +# define XXH_VSX_BE 1 +# else +# define XXH_VSX_BE 0 +# endif +# endif /* !defined(XXH_VSX_BE) */ + +# if XXH_VSX_BE +/* A wrapper for POWER9's vec_revb. */ +# if defined(__POWER9_VECTOR__) || (defined(__clang__) && defined(__s390x__)) +# define XXH_vec_revb vec_revb +# else +XXH_FORCE_INLINE xxh_u64x2 XXH_vec_revb(xxh_u64x2 val) +{ + xxh_u8x16 const vByteSwap = { 0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00, + 0x0F, 0x0E, 0x0D, 0x0C, 0x0B, 0x0A, 0x09, 0x08 }; + return vec_perm(val, val, vByteSwap); +} +# endif +# endif /* XXH_VSX_BE */ + +/* + * Performs an unaligned load and byte swaps it on big endian. + */ +XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(const void *ptr) +{ + xxh_u64x2 ret; + memcpy(&ret, ptr, sizeof(xxh_u64x2)); +# if XXH_VSX_BE + ret = XXH_vec_revb(ret); +# endif + return ret; +} + +/* + * vec_mulo and vec_mule are very problematic intrinsics on PowerPC + * + * These intrinsics weren't added until GCC 8, despite existing for a while, + * and they are endian dependent. Also, their meaning swap depending on version. + * */ +# if defined(__s390x__) + /* s390x is always big endian, no issue on this platform */ +# define XXH_vec_mulo vec_mulo +# define XXH_vec_mule vec_mule +# elif defined(__clang__) && XXH_HAS_BUILTIN(__builtin_altivec_vmuleuw) +/* Clang has a better way to control this, we can just use the builtin which doesn't swap. */ +# define XXH_vec_mulo __builtin_altivec_vmulouw +# define XXH_vec_mule __builtin_altivec_vmuleuw +# else +/* gcc needs inline assembly */ +/* Adapted from https://github.com/google/highwayhash/blob/master/highwayhash/hh_vsx.h. */ +XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mulo(xxh_u32x4 a, xxh_u32x4 b) +{ + xxh_u64x2 result; + __asm__("vmulouw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b)); + return result; +} +XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mule(xxh_u32x4 a, xxh_u32x4 b) +{ + xxh_u64x2 result; + __asm__("vmuleuw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b)); + return result; +} +# endif /* XXH_vec_mulo, XXH_vec_mule */ +#endif /* XXH_VECTOR == XXH_VSX */ + + +/* prefetch + * can be disabled, by declaring XXH_NO_PREFETCH build macro */ +#if defined(XXH_NO_PREFETCH) +# define XXH_PREFETCH(ptr) (void)(ptr) /* disabled */ +#else +# if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_I86)) /* _mm_prefetch() is not defined outside of x86/x64 */ +# include <mmintrin.h> /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */ +# define XXH_PREFETCH(ptr) _mm_prefetch((const char*)(ptr), _MM_HINT_T0) +# elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) ) +# define XXH_PREFETCH(ptr) __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */) +# else +# define XXH_PREFETCH(ptr) (void)(ptr) /* disabled */ +# endif +#endif /* XXH_NO_PREFETCH */ + + +/* ========================================== + * XXH3 default settings + * ========================================== */ + +#define XXH_SECRET_DEFAULT_SIZE 192 /* minimum XXH3_SECRET_SIZE_MIN */ + +#if (XXH_SECRET_DEFAULT_SIZE < XXH3_SECRET_SIZE_MIN) +# error "default keyset is not large enough" +#endif + +/* Pseudorandom secret taken directly from FARSH */ +XXH_ALIGN(64) static const xxh_u8 XXH3_kSecret[XXH_SECRET_DEFAULT_SIZE] = { + 0xb8, 0xfe, 0x6c, 0x39, 0x23, 0xa4, 0x4b, 0xbe, 0x7c, 0x01, 0x81, 0x2c, 0xf7, 0x21, 0xad, 0x1c, + 0xde, 0xd4, 0x6d, 0xe9, 0x83, 0x90, 0x97, 0xdb, 0x72, 0x40, 0xa4, 0xa4, 0xb7, 0xb3, 0x67, 0x1f, + 0xcb, 0x79, 0xe6, 0x4e, 0xcc, 0xc0, 0xe5, 0x78, 0x82, 0x5a, 0xd0, 0x7d, 0xcc, 0xff, 0x72, 0x21, + 0xb8, 0x08, 0x46, 0x74, 0xf7, 0x43, 0x24, 0x8e, 0xe0, 0x35, 0x90, 0xe6, 0x81, 0x3a, 0x26, 0x4c, + 0x3c, 0x28, 0x52, 0xbb, 0x91, 0xc3, 0x00, 0xcb, 0x88, 0xd0, 0x65, 0x8b, 0x1b, 0x53, 0x2e, 0xa3, + 0x71, 0x64, 0x48, 0x97, 0xa2, 0x0d, 0xf9, 0x4e, 0x38, 0x19, 0xef, 0x46, 0xa9, 0xde, 0xac, 0xd8, + 0xa8, 0xfa, 0x76, 0x3f, 0xe3, 0x9c, 0x34, 0x3f, 0xf9, 0xdc, 0xbb, 0xc7, 0xc7, 0x0b, 0x4f, 0x1d, + 0x8a, 0x51, 0xe0, 0x4b, 0xcd, 0xb4, 0x59, 0x31, 0xc8, 0x9f, 0x7e, 0xc9, 0xd9, 0x78, 0x73, 0x64, + 0xea, 0xc5, 0xac, 0x83, 0x34, 0xd3, 0xeb, 0xc3, 0xc5, 0x81, 0xa0, 0xff, 0xfa, 0x13, 0x63, 0xeb, + 0x17, 0x0d, 0xdd, 0x51, 0xb7, 0xf0, 0xda, 0x49, 0xd3, 0x16, 0x55, 0x26, 0x29, 0xd4, 0x68, 0x9e, + 0x2b, 0x16, 0xbe, 0x58, 0x7d, 0x47, 0xa1, 0xfc, 0x8f, 0xf8, 0xb8, 0xd1, 0x7a, 0xd0, 0x31, 0xce, + 0x45, 0xcb, 0x3a, 0x8f, 0x95, 0x16, 0x04, 0x28, 0xaf, 0xd7, 0xfb, 0xca, 0xbb, 0x4b, 0x40, 0x7e, +}; + + +#ifdef XXH_OLD_NAMES +# define kSecret XXH3_kSecret +#endif + +/* + * Calculates a 32-bit to 64-bit long multiply. + * + * Wraps __emulu on MSVC x86 because it tends to call __allmul when it doesn't + * need to (but it shouldn't need to anyways, it is about 7 instructions to do + * a 64x64 multiply...). Since we know that this will _always_ emit MULL, we + * use that instead of the normal method. + * + * If you are compiling for platforms like Thumb-1 and don't have a better option, + * you may also want to write your own long multiply routine here. + * + * XXH_FORCE_INLINE xxh_u64 XXH_mult32to64(xxh_u64 x, xxh_u64 y) + * { + * return (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF); + * } + */ +#if defined(_MSC_VER) && defined(_M_IX86) +# include <intrin.h> +# define XXH_mult32to64(x, y) __emulu((unsigned)(x), (unsigned)(y)) +#else +/* + * Downcast + upcast is usually better than masking on older compilers like + * GCC 4.2 (especially 32-bit ones), all without affecting newer compilers. + * + * The other method, (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF), will AND both operands + * and perform a full 64x64 multiply -- entirely redundant on 32-bit. + */ +# define XXH_mult32to64(x, y) ((xxh_u64)(xxh_u32)(x) * (xxh_u64)(xxh_u32)(y)) +#endif + +/* + * Calculates a 64->128-bit long multiply. + * + * Uses __uint128_t and _umul128 if available, otherwise uses a scalar version. + */ +static XXH128_hash_t +XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs) +{ + /* + * GCC/Clang __uint128_t method. + * + * On most 64-bit targets, GCC and Clang define a __uint128_t type. + * This is usually the best way as it usually uses a native long 64-bit + * multiply, such as MULQ on x86_64 or MUL + UMULH on aarch64. + * + * Usually. + * + * Despite being a 32-bit platform, Clang (and emscripten) define this type + * despite not having the arithmetic for it. This results in a laggy + * compiler builtin call which calculates a full 128-bit multiply. + * In that case it is best to use the portable one. + * https://github.com/Cyan4973/xxHash/issues/211#issuecomment-515575677 + */ +#if defined(__GNUC__) && !defined(__wasm__) \ + && defined(__SIZEOF_INT128__) \ + || (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128) + + __uint128_t const product = (__uint128_t)lhs * (__uint128_t)rhs; + XXH128_hash_t r128; + r128.low64 = (xxh_u64)(product); + r128.high64 = (xxh_u64)(product >> 64); + return r128; + + /* + * MSVC for x64's _umul128 method. + * + * xxh_u64 _umul128(xxh_u64 Multiplier, xxh_u64 Multiplicand, xxh_u64 *HighProduct); + * + * This compiles to single operand MUL on x64. + */ +#elif defined(_M_X64) || defined(_M_IA64) + +#ifndef _MSC_VER +# pragma intrinsic(_umul128) +#endif + xxh_u64 product_high; + xxh_u64 const product_low = _umul128(lhs, rhs, &product_high); + XXH128_hash_t r128; + r128.low64 = product_low; + r128.high64 = product_high; + return r128; + +#else + /* + * Portable scalar method. Optimized for 32-bit and 64-bit ALUs. + * + * This is a fast and simple grade school multiply, which is shown below + * with base 10 arithmetic instead of base 0x100000000. + * + * 9 3 // D2 lhs = 93 + * x 7 5 // D2 rhs = 75 + * ---------- + * 1 5 // D2 lo_lo = (93 % 10) * (75 % 10) = 15 + * 4 5 | // D2 hi_lo = (93 / 10) * (75 % 10) = 45 + * 2 1 | // D2 lo_hi = (93 % 10) * (75 / 10) = 21 + * + 6 3 | | // D2 hi_hi = (93 / 10) * (75 / 10) = 63 + * --------- + * 2 7 | // D2 cross = (15 / 10) + (45 % 10) + 21 = 27 + * + 6 7 | | // D2 upper = (27 / 10) + (45 / 10) + 63 = 67 + * --------- + * 6 9 7 5 // D4 res = (27 * 10) + (15 % 10) + (67 * 100) = 6975 + * + * The reasons for adding the products like this are: + * 1. It avoids manual carry tracking. Just like how + * (9 * 9) + 9 + 9 = 99, the same applies with this for UINT64_MAX. + * This avoids a lot of complexity. + * + * 2. It hints for, and on Clang, compiles to, the powerful UMAAL + * instruction available in ARM's Digital Signal Processing extension + * in 32-bit ARMv6 and later, which is shown below: + * + * void UMAAL(xxh_u32 *RdLo, xxh_u32 *RdHi, xxh_u32 Rn, xxh_u32 Rm) + * { + * xxh_u64 product = (xxh_u64)*RdLo * (xxh_u64)*RdHi + Rn + Rm; + * *RdLo = (xxh_u32)(product & 0xFFFFFFFF); + * *RdHi = (xxh_u32)(product >> 32); + * } + * + * This instruction was designed for efficient long multiplication, and + * allows this to be calculated in only 4 instructions at speeds + * comparable to some 64-bit ALUs. + * + * 3. It isn't terrible on other platforms. Usually this will be a couple + * of 32-bit ADD/ADCs. + */ + + /* First calculate all of the cross products. */ + xxh_u64 const lo_lo = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs & 0xFFFFFFFF); + xxh_u64 const hi_lo = XXH_mult32to64(lhs >> 32, rhs & 0xFFFFFFFF); + xxh_u64 const lo_hi = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs >> 32); + xxh_u64 const hi_hi = XXH_mult32to64(lhs >> 32, rhs >> 32); + + /* Now add the products together. These will never overflow. */ + xxh_u64 const cross = (lo_lo >> 32) + (hi_lo & 0xFFFFFFFF) + lo_hi; + xxh_u64 const upper = (hi_lo >> 32) + (cross >> 32) + hi_hi; + xxh_u64 const lower = (cross << 32) | (lo_lo & 0xFFFFFFFF); + + XXH128_hash_t r128; + r128.low64 = lower; + r128.high64 = upper; + return r128; +#endif +} + +/* + * Does a 64-bit to 128-bit multiply, then XOR folds it. + * + * The reason for the separate function is to prevent passing too many structs + * around by value. This will hopefully inline the multiply, but we don't force it. + */ +static xxh_u64 +XXH3_mul128_fold64(xxh_u64 lhs, xxh_u64 rhs) +{ + XXH128_hash_t product = XXH_mult64to128(lhs, rhs); + return product.low64 ^ product.high64; +} + +/* Seems to produce slightly better code on GCC for some reason. */ +XXH_FORCE_INLINE xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift) +{ + XXH_ASSERT(0 <= shift && shift < 64); + return v64 ^ (v64 >> shift); +} + +/* + * This is a fast avalanche stage, + * suitable when input bits are already partially mixed + */ +static XXH64_hash_t XXH3_avalanche(xxh_u64 h64) +{ + h64 = XXH_xorshift64(h64, 37); + h64 *= 0x165667919E3779F9ULL; + h64 = XXH_xorshift64(h64, 32); + return h64; +} + +/* + * This is a stronger avalanche, + * inspired by Pelle Evensen's rrmxmx + * preferable when input has not been previously mixed + */ +static XXH64_hash_t XXH3_rrmxmx(xxh_u64 h64, xxh_u64 len) +{ + /* this mix is inspired by Pelle Evensen's rrmxmx */ + h64 ^= XXH_rotl64(h64, 49) ^ XXH_rotl64(h64, 24); + h64 *= 0x9FB21C651E98DF25ULL; + h64 ^= (h64 >> 35) + len ; + h64 *= 0x9FB21C651E98DF25ULL; + return XXH_xorshift64(h64, 28); +} + + +/* ========================================== + * Short keys + * ========================================== + * One of the shortcomings of XXH32 and XXH64 was that their performance was + * sub-optimal on short lengths. It used an iterative algorithm which strongly + * favored lengths that were a multiple of 4 or 8. + * + * Instead of iterating over individual inputs, we use a set of single shot + * functions which piece together a range of lengths and operate in constant time. + * + * Additionally, the number of multiplies has been significantly reduced. This + * reduces latency, especially when emulating 64-bit multiplies on 32-bit. + * + * Depending on the platform, this may or may not be faster than XXH32, but it + * is almost guaranteed to be faster than XXH64. + */ + +/* + * At very short lengths, there isn't enough input to fully hide secrets, or use + * the entire secret. + * + * There is also only a limited amount of mixing we can do before significantly + * impacting performance. + * + * Therefore, we use different sections of the secret and always mix two secret + * samples with an XOR. This should have no effect on performance on the + * seedless or withSeed variants because everything _should_ be constant folded + * by modern compilers. + * + * The XOR mixing hides individual parts of the secret and increases entropy. + * + * This adds an extra layer of strength for custom secrets. + */ +XXH_FORCE_INLINE XXH64_hash_t +XXH3_len_1to3_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(input != NULL); + XXH_ASSERT(1 <= len && len <= 3); + XXH_ASSERT(secret != NULL); + /* + * len = 1: combined = { input[0], 0x01, input[0], input[0] } + * len = 2: combined = { input[1], 0x02, input[0], input[1] } + * len = 3: combined = { input[2], 0x03, input[0], input[1] } + */ + { xxh_u8 const c1 = input[0]; + xxh_u8 const c2 = input[len >> 1]; + xxh_u8 const c3 = input[len - 1]; + xxh_u32 const combined = ((xxh_u32)c1 << 16) | ((xxh_u32)c2 << 24) + | ((xxh_u32)c3 << 0) | ((xxh_u32)len << 8); + xxh_u64 const bitflip = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed; + xxh_u64 const keyed = (xxh_u64)combined ^ bitflip; + return XXH64_avalanche(keyed); + } +} + +XXH_FORCE_INLINE XXH64_hash_t +XXH3_len_4to8_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(input != NULL); + XXH_ASSERT(secret != NULL); + XXH_ASSERT(4 <= len && len < 8); + seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32; + { xxh_u32 const input1 = XXH_readLE32(input); + xxh_u32 const input2 = XXH_readLE32(input + len - 4); + xxh_u64 const bitflip = (XXH_readLE64(secret+8) ^ XXH_readLE64(secret+16)) - seed; + xxh_u64 const input64 = input2 + (((xxh_u64)input1) << 32); + xxh_u64 const keyed = input64 ^ bitflip; + return XXH3_rrmxmx(keyed, len); + } +} + +XXH_FORCE_INLINE XXH64_hash_t +XXH3_len_9to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(input != NULL); + XXH_ASSERT(secret != NULL); + XXH_ASSERT(8 <= len && len <= 16); + { xxh_u64 const bitflip1 = (XXH_readLE64(secret+24) ^ XXH_readLE64(secret+32)) + seed; + xxh_u64 const bitflip2 = (XXH_readLE64(secret+40) ^ XXH_readLE64(secret+48)) - seed; + xxh_u64 const input_lo = XXH_readLE64(input) ^ bitflip1; + xxh_u64 const input_hi = XXH_readLE64(input + len - 8) ^ bitflip2; + xxh_u64 const acc = len + + XXH_swap64(input_lo) + input_hi + + XXH3_mul128_fold64(input_lo, input_hi); + return XXH3_avalanche(acc); + } +} + +XXH_FORCE_INLINE XXH64_hash_t +XXH3_len_0to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(len <= 16); + { if (XXH_likely(len > 8)) return XXH3_len_9to16_64b(input, len, secret, seed); + if (XXH_likely(len >= 4)) return XXH3_len_4to8_64b(input, len, secret, seed); + if (len) return XXH3_len_1to3_64b(input, len, secret, seed); + return XXH64_avalanche(seed ^ (XXH_readLE64(secret+56) ^ XXH_readLE64(secret+64))); + } +} + +/* + * DISCLAIMER: There are known *seed-dependent* multicollisions here due to + * multiplication by zero, affecting hashes of lengths 17 to 240. + * + * However, they are very unlikely. + * + * Keep this in mind when using the unseeded XXH3_64bits() variant: As with all + * unseeded non-cryptographic hashes, it does not attempt to defend itself + * against specially crafted inputs, only random inputs. + * + * Compared to classic UMAC where a 1 in 2^31 chance of 4 consecutive bytes + * cancelling out the secret is taken an arbitrary number of times (addressed + * in XXH3_accumulate_512), this collision is very unlikely with random inputs + * and/or proper seeding: + * + * This only has a 1 in 2^63 chance of 8 consecutive bytes cancelling out, in a + * function that is only called up to 16 times per hash with up to 240 bytes of + * input. + * + * This is not too bad for a non-cryptographic hash function, especially with + * only 64 bit outputs. + * + * The 128-bit variant (which trades some speed for strength) is NOT affected + * by this, although it is always a good idea to use a proper seed if you care + * about strength. + */ +XXH_FORCE_INLINE xxh_u64 XXH3_mix16B(const xxh_u8* XXH_RESTRICT input, + const xxh_u8* XXH_RESTRICT secret, xxh_u64 seed64) +{ +#if defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \ + && defined(__i386__) && defined(__SSE2__) /* x86 + SSE2 */ \ + && !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable like XXH32 hack */ + /* + * UGLY HACK: + * GCC for x86 tends to autovectorize the 128-bit multiply, resulting in + * slower code. + * + * By forcing seed64 into a register, we disrupt the cost model and + * cause it to scalarize. See `XXH32_round()` + * + * FIXME: Clang's output is still _much_ faster -- On an AMD Ryzen 3600, + * XXH3_64bits @ len=240 runs at 4.6 GB/s with Clang 9, but 3.3 GB/s on + * GCC 9.2, despite both emitting scalar code. + * + * GCC generates much better scalar code than Clang for the rest of XXH3, + * which is why finding a more optimal codepath is an interest. + */ + __asm__ ("" : "+r" (seed64)); +#endif + { xxh_u64 const input_lo = XXH_readLE64(input); + xxh_u64 const input_hi = XXH_readLE64(input+8); + return XXH3_mul128_fold64( + input_lo ^ (XXH_readLE64(secret) + seed64), + input_hi ^ (XXH_readLE64(secret+8) - seed64) + ); + } +} + +/* For mid range keys, XXH3 uses a Mum-hash variant. */ +XXH_FORCE_INLINE XXH64_hash_t +XXH3_len_17to128_64b(const xxh_u8* XXH_RESTRICT input, size_t len, + const xxh_u8* XXH_RESTRICT secret, size_t secretSize, + XXH64_hash_t seed) +{ + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize; + XXH_ASSERT(16 < len && len <= 128); + + { xxh_u64 acc = len * XXH_PRIME64_1; + if (len > 32) { + if (len > 64) { + if (len > 96) { + acc += XXH3_mix16B(input+48, secret+96, seed); + acc += XXH3_mix16B(input+len-64, secret+112, seed); + } + acc += XXH3_mix16B(input+32, secret+64, seed); + acc += XXH3_mix16B(input+len-48, secret+80, seed); + } + acc += XXH3_mix16B(input+16, secret+32, seed); + acc += XXH3_mix16B(input+len-32, secret+48, seed); + } + acc += XXH3_mix16B(input+0, secret+0, seed); + acc += XXH3_mix16B(input+len-16, secret+16, seed); + + return XXH3_avalanche(acc); + } +} + +#define XXH3_MIDSIZE_MAX 240 + +XXH_NO_INLINE XXH64_hash_t +XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len, + const xxh_u8* XXH_RESTRICT secret, size_t secretSize, + XXH64_hash_t seed) +{ + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize; + XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX); + + #define XXH3_MIDSIZE_STARTOFFSET 3 + #define XXH3_MIDSIZE_LASTOFFSET 17 + + { xxh_u64 acc = len * XXH_PRIME64_1; + int const nbRounds = (int)len / 16; + int i; + for (i=0; i<8; i++) { + acc += XXH3_mix16B(input+(16*i), secret+(16*i), seed); + } + acc = XXH3_avalanche(acc); + XXH_ASSERT(nbRounds >= 8); +#if defined(__clang__) /* Clang */ \ + && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \ + && !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable */ + /* + * UGLY HACK: + * Clang for ARMv7-A tries to vectorize this loop, similar to GCC x86. + * In everywhere else, it uses scalar code. + * + * For 64->128-bit multiplies, even if the NEON was 100% optimal, it + * would still be slower than UMAAL (see XXH_mult64to128). + * + * Unfortunately, Clang doesn't handle the long multiplies properly and + * converts them to the nonexistent "vmulq_u64" intrinsic, which is then + * scalarized into an ugly mess of VMOV.32 instructions. + * + * This mess is difficult to avoid without turning autovectorization + * off completely, but they are usually relatively minor and/or not + * worth it to fix. + * + * This loop is the easiest to fix, as unlike XXH32, this pragma + * _actually works_ because it is a loop vectorization instead of an + * SLP vectorization. + */ + #pragma clang loop vectorize(disable) +#endif + for (i=8 ; i < nbRounds; i++) { + acc += XXH3_mix16B(input+(16*i), secret+(16*(i-8)) + XXH3_MIDSIZE_STARTOFFSET, seed); + } + /* last bytes */ + acc += XXH3_mix16B(input + len - 16, secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, seed); + return XXH3_avalanche(acc); + } +} + + +/* ======= Long Keys ======= */ + +#define XXH_STRIPE_LEN 64 +#define XXH_SECRET_CONSUME_RATE 8 /* nb of secret bytes consumed at each accumulation */ +#define XXH_ACC_NB (XXH_STRIPE_LEN / sizeof(xxh_u64)) + +#ifdef XXH_OLD_NAMES +# define STRIPE_LEN XXH_STRIPE_LEN +# define ACC_NB XXH_ACC_NB +#endif + +XXH_FORCE_INLINE void XXH_writeLE64(void* dst, xxh_u64 v64) +{ + if (!XXH_CPU_LITTLE_ENDIAN) v64 = XXH_swap64(v64); + memcpy(dst, &v64, sizeof(v64)); +} + +/* Several intrinsic functions below are supposed to accept __int64 as argument, + * as documented in https://software.intel.com/sites/landingpage/IntrinsicsGuide/ . + * However, several environments do not define __int64 type, + * requiring a workaround. + */ +#if !defined (__VMS) \ + && (defined (__cplusplus) \ + || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) ) + typedef int64_t xxh_i64; +#else + /* the following type must have a width of 64-bit */ + typedef long long xxh_i64; +#endif + +/* + * XXH3_accumulate_512 is the tightest loop for long inputs, and it is the most optimized. + * + * It is a hardened version of UMAC, based off of FARSH's implementation. + * + * This was chosen because it adapts quite well to 32-bit, 64-bit, and SIMD + * implementations, and it is ridiculously fast. + * + * We harden it by mixing the original input to the accumulators as well as the product. + * + * This means that in the (relatively likely) case of a multiply by zero, the + * original input is preserved. + * + * On 128-bit inputs, we swap 64-bit pairs when we add the input to improve + * cross-pollination, as otherwise the upper and lower halves would be + * essentially independent. + * + * This doesn't matter on 64-bit hashes since they all get merged together in + * the end, so we skip the extra step. + * + * Both XXH3_64bits and XXH3_128bits use this subroutine. + */ + +#if (XXH_VECTOR == XXH_AVX512) || defined(XXH_X86DISPATCH) + +#ifndef XXH_TARGET_AVX512 +# define XXH_TARGET_AVX512 /* disable attribute target */ +#endif + +XXH_FORCE_INLINE XXH_TARGET_AVX512 void +XXH3_accumulate_512_avx512(void* XXH_RESTRICT acc, + const void* XXH_RESTRICT input, + const void* XXH_RESTRICT secret) +{ + XXH_ALIGN(64) __m512i* const xacc = (__m512i *) acc; + XXH_ASSERT((((size_t)acc) & 63) == 0); + XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i)); + + { + /* data_vec = input[0]; */ + __m512i const data_vec = _mm512_loadu_si512 (input); + /* key_vec = secret[0]; */ + __m512i const key_vec = _mm512_loadu_si512 (secret); + /* data_key = data_vec ^ key_vec; */ + __m512i const data_key = _mm512_xor_si512 (data_vec, key_vec); + /* data_key_lo = data_key >> 32; */ + __m512i const data_key_lo = _mm512_shuffle_epi32 (data_key, (_MM_PERM_ENUM)_MM_SHUFFLE(0, 3, 0, 1)); + /* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */ + __m512i const product = _mm512_mul_epu32 (data_key, data_key_lo); + /* xacc[0] += swap(data_vec); */ + __m512i const data_swap = _mm512_shuffle_epi32(data_vec, (_MM_PERM_ENUM)_MM_SHUFFLE(1, 0, 3, 2)); + __m512i const sum = _mm512_add_epi64(*xacc, data_swap); + /* xacc[0] += product; */ + *xacc = _mm512_add_epi64(product, sum); + } +} + +/* + * XXH3_scrambleAcc: Scrambles the accumulators to improve mixing. + * + * Multiplication isn't perfect, as explained by Google in HighwayHash: + * + * // Multiplication mixes/scrambles bytes 0-7 of the 64-bit result to + * // varying degrees. In descending order of goodness, bytes + * // 3 4 2 5 1 6 0 7 have quality 228 224 164 160 100 96 36 32. + * // As expected, the upper and lower bytes are much worse. + * + * Source: https://github.com/google/highwayhash/blob/0aaf66b/highwayhash/hh_avx2.h#L291 + * + * Since our algorithm uses a pseudorandom secret to add some variance into the + * mix, we don't need to (or want to) mix as often or as much as HighwayHash does. + * + * This isn't as tight as XXH3_accumulate, but still written in SIMD to avoid + * extraction. + * + * Both XXH3_64bits and XXH3_128bits use this subroutine. + */ + +XXH_FORCE_INLINE XXH_TARGET_AVX512 void +XXH3_scrambleAcc_avx512(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret) +{ + XXH_ASSERT((((size_t)acc) & 63) == 0); + XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i)); + { XXH_ALIGN(64) __m512i* const xacc = (__m512i*) acc; + const __m512i prime32 = _mm512_set1_epi32((int)XXH_PRIME32_1); + + /* xacc[0] ^= (xacc[0] >> 47) */ + __m512i const acc_vec = *xacc; + __m512i const shifted = _mm512_srli_epi64 (acc_vec, 47); + __m512i const data_vec = _mm512_xor_si512 (acc_vec, shifted); + /* xacc[0] ^= secret; */ + __m512i const key_vec = _mm512_loadu_si512 (secret); + __m512i const data_key = _mm512_xor_si512 (data_vec, key_vec); + + /* xacc[0] *= XXH_PRIME32_1; */ + __m512i const data_key_hi = _mm512_shuffle_epi32 (data_key, (_MM_PERM_ENUM)_MM_SHUFFLE(0, 3, 0, 1)); + __m512i const prod_lo = _mm512_mul_epu32 (data_key, prime32); + __m512i const prod_hi = _mm512_mul_epu32 (data_key_hi, prime32); + *xacc = _mm512_add_epi64(prod_lo, _mm512_slli_epi64(prod_hi, 32)); + } +} + +XXH_FORCE_INLINE XXH_TARGET_AVX512 void +XXH3_initCustomSecret_avx512(void* XXH_RESTRICT customSecret, xxh_u64 seed64) +{ + XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 63) == 0); + XXH_STATIC_ASSERT(XXH_SEC_ALIGN == 64); + XXH_ASSERT(((size_t)customSecret & 63) == 0); + (void)(&XXH_writeLE64); + { int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m512i); + __m512i const seed = _mm512_mask_set1_epi64(_mm512_set1_epi64((xxh_i64)seed64), 0xAA, -(xxh_i64)seed64); + + XXH_ALIGN(64) const __m512i* const src = (const __m512i*) XXH3_kSecret; + XXH_ALIGN(64) __m512i* const dest = ( __m512i*) customSecret; + int i; + for (i=0; i < nbRounds; ++i) { + /* GCC has a bug, _mm512_stream_load_si512 accepts 'void*', not 'void const*', + * this will warn "discards ‘const’ qualifier". */ + union { + XXH_ALIGN(64) const __m512i* cp; + XXH_ALIGN(64) void* p; + } remote_const_void; + remote_const_void.cp = src + i; + dest[i] = _mm512_add_epi64(_mm512_stream_load_si512(remote_const_void.p), seed); + } } +} + +#endif + +#if (XXH_VECTOR == XXH_AVX2) || defined(XXH_X86DISPATCH) + +#ifndef XXH_TARGET_AVX2 +# define XXH_TARGET_AVX2 /* disable attribute target */ +#endif + +XXH_FORCE_INLINE XXH_TARGET_AVX2 void +XXH3_accumulate_512_avx2( void* XXH_RESTRICT acc, + const void* XXH_RESTRICT input, + const void* XXH_RESTRICT secret) +{ + XXH_ASSERT((((size_t)acc) & 31) == 0); + { XXH_ALIGN(32) __m256i* const xacc = (__m256i *) acc; + /* Unaligned. This is mainly for pointer arithmetic, and because + * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */ + const __m256i* const xinput = (const __m256i *) input; + /* Unaligned. This is mainly for pointer arithmetic, and because + * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */ + const __m256i* const xsecret = (const __m256i *) secret; + + size_t i; + for (i=0; i < XXH_STRIPE_LEN/sizeof(__m256i); i++) { + /* data_vec = xinput[i]; */ + __m256i const data_vec = _mm256_loadu_si256 (xinput+i); + /* key_vec = xsecret[i]; */ + __m256i const key_vec = _mm256_loadu_si256 (xsecret+i); + /* data_key = data_vec ^ key_vec; */ + __m256i const data_key = _mm256_xor_si256 (data_vec, key_vec); + /* data_key_lo = data_key >> 32; */ + __m256i const data_key_lo = _mm256_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1)); + /* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */ + __m256i const product = _mm256_mul_epu32 (data_key, data_key_lo); + /* xacc[i] += swap(data_vec); */ + __m256i const data_swap = _mm256_shuffle_epi32(data_vec, _MM_SHUFFLE(1, 0, 3, 2)); + __m256i const sum = _mm256_add_epi64(xacc[i], data_swap); + /* xacc[i] += product; */ + xacc[i] = _mm256_add_epi64(product, sum); + } } +} + +XXH_FORCE_INLINE XXH_TARGET_AVX2 void +XXH3_scrambleAcc_avx2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret) +{ + XXH_ASSERT((((size_t)acc) & 31) == 0); + { XXH_ALIGN(32) __m256i* const xacc = (__m256i*) acc; + /* Unaligned. This is mainly for pointer arithmetic, and because + * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */ + const __m256i* const xsecret = (const __m256i *) secret; + const __m256i prime32 = _mm256_set1_epi32((int)XXH_PRIME32_1); + + size_t i; + for (i=0; i < XXH_STRIPE_LEN/sizeof(__m256i); i++) { + /* xacc[i] ^= (xacc[i] >> 47) */ + __m256i const acc_vec = xacc[i]; + __m256i const shifted = _mm256_srli_epi64 (acc_vec, 47); + __m256i const data_vec = _mm256_xor_si256 (acc_vec, shifted); + /* xacc[i] ^= xsecret; */ + __m256i const key_vec = _mm256_loadu_si256 (xsecret+i); + __m256i const data_key = _mm256_xor_si256 (data_vec, key_vec); + + /* xacc[i] *= XXH_PRIME32_1; */ + __m256i const data_key_hi = _mm256_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1)); + __m256i const prod_lo = _mm256_mul_epu32 (data_key, prime32); + __m256i const prod_hi = _mm256_mul_epu32 (data_key_hi, prime32); + xacc[i] = _mm256_add_epi64(prod_lo, _mm256_slli_epi64(prod_hi, 32)); + } + } +} + +XXH_FORCE_INLINE XXH_TARGET_AVX2 void XXH3_initCustomSecret_avx2(void* XXH_RESTRICT customSecret, xxh_u64 seed64) +{ + XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 31) == 0); + XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE / sizeof(__m256i)) == 6); + XXH_STATIC_ASSERT(XXH_SEC_ALIGN <= 64); + (void)(&XXH_writeLE64); + XXH_PREFETCH(customSecret); + { __m256i const seed = _mm256_set_epi64x(-(xxh_i64)seed64, (xxh_i64)seed64, -(xxh_i64)seed64, (xxh_i64)seed64); + + XXH_ALIGN(64) const __m256i* const src = (const __m256i*) XXH3_kSecret; + XXH_ALIGN(64) __m256i* dest = ( __m256i*) customSecret; + +# if defined(__GNUC__) || defined(__clang__) + /* + * On GCC & Clang, marking 'dest' as modified will cause the compiler: + * - do not extract the secret from sse registers in the internal loop + * - use less common registers, and avoid pushing these reg into stack + * The asm hack causes Clang to assume that XXH3_kSecretPtr aliases with + * customSecret, and on aarch64, this prevented LDP from merging two + * loads together for free. Putting the loads together before the stores + * properly generates LDP. + */ + __asm__("" : "+r" (dest)); +# endif + + /* GCC -O2 need unroll loop manually */ + dest[0] = _mm256_add_epi64(_mm256_stream_load_si256(src+0), seed); + dest[1] = _mm256_add_epi64(_mm256_stream_load_si256(src+1), seed); + dest[2] = _mm256_add_epi64(_mm256_stream_load_si256(src+2), seed); + dest[3] = _mm256_add_epi64(_mm256_stream_load_si256(src+3), seed); + dest[4] = _mm256_add_epi64(_mm256_stream_load_si256(src+4), seed); + dest[5] = _mm256_add_epi64(_mm256_stream_load_si256(src+5), seed); + } +} + +#endif + +#if (XXH_VECTOR == XXH_SSE2) || defined(XXH_X86DISPATCH) + +#ifndef XXH_TARGET_SSE2 +# define XXH_TARGET_SSE2 /* disable attribute target */ +#endif + +XXH_FORCE_INLINE XXH_TARGET_SSE2 void +XXH3_accumulate_512_sse2( void* XXH_RESTRICT acc, + const void* XXH_RESTRICT input, + const void* XXH_RESTRICT secret) +{ + /* SSE2 is just a half-scale version of the AVX2 version. */ + XXH_ASSERT((((size_t)acc) & 15) == 0); + { XXH_ALIGN(16) __m128i* const xacc = (__m128i *) acc; + /* Unaligned. This is mainly for pointer arithmetic, and because + * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */ + const __m128i* const xinput = (const __m128i *) input; + /* Unaligned. This is mainly for pointer arithmetic, and because + * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */ + const __m128i* const xsecret = (const __m128i *) secret; + + size_t i; + for (i=0; i < XXH_STRIPE_LEN/sizeof(__m128i); i++) { + /* data_vec = xinput[i]; */ + __m128i const data_vec = _mm_loadu_si128 (xinput+i); + /* key_vec = xsecret[i]; */ + __m128i const key_vec = _mm_loadu_si128 (xsecret+i); + /* data_key = data_vec ^ key_vec; */ + __m128i const data_key = _mm_xor_si128 (data_vec, key_vec); + /* data_key_lo = data_key >> 32; */ + __m128i const data_key_lo = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1)); + /* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */ + __m128i const product = _mm_mul_epu32 (data_key, data_key_lo); + /* xacc[i] += swap(data_vec); */ + __m128i const data_swap = _mm_shuffle_epi32(data_vec, _MM_SHUFFLE(1,0,3,2)); + __m128i const sum = _mm_add_epi64(xacc[i], data_swap); + /* xacc[i] += product; */ + xacc[i] = _mm_add_epi64(product, sum); + } } +} + +XXH_FORCE_INLINE XXH_TARGET_SSE2 void +XXH3_scrambleAcc_sse2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret) +{ + XXH_ASSERT((((size_t)acc) & 15) == 0); + { XXH_ALIGN(16) __m128i* const xacc = (__m128i*) acc; + /* Unaligned. This is mainly for pointer arithmetic, and because + * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */ + const __m128i* const xsecret = (const __m128i *) secret; + const __m128i prime32 = _mm_set1_epi32((int)XXH_PRIME32_1); + + size_t i; + for (i=0; i < XXH_STRIPE_LEN/sizeof(__m128i); i++) { + /* xacc[i] ^= (xacc[i] >> 47) */ + __m128i const acc_vec = xacc[i]; + __m128i const shifted = _mm_srli_epi64 (acc_vec, 47); + __m128i const data_vec = _mm_xor_si128 (acc_vec, shifted); + /* xacc[i] ^= xsecret[i]; */ + __m128i const key_vec = _mm_loadu_si128 (xsecret+i); + __m128i const data_key = _mm_xor_si128 (data_vec, key_vec); + + /* xacc[i] *= XXH_PRIME32_1; */ + __m128i const data_key_hi = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1)); + __m128i const prod_lo = _mm_mul_epu32 (data_key, prime32); + __m128i const prod_hi = _mm_mul_epu32 (data_key_hi, prime32); + xacc[i] = _mm_add_epi64(prod_lo, _mm_slli_epi64(prod_hi, 32)); + } + } +} + +XXH_FORCE_INLINE XXH_TARGET_SSE2 void XXH3_initCustomSecret_sse2(void* XXH_RESTRICT customSecret, xxh_u64 seed64) +{ + XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0); + (void)(&XXH_writeLE64); + { int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m128i); + +# if defined(_MSC_VER) && defined(_M_IX86) && _MSC_VER < 1900 + // MSVC 32bit mode does not support _mm_set_epi64x before 2015 + XXH_ALIGN(16) const xxh_i64 seed64x2[2] = { (xxh_i64)seed64, -(xxh_i64)seed64 }; + __m128i const seed = _mm_load_si128((__m128i const*)seed64x2); +# else + __m128i const seed = _mm_set_epi64x(-(xxh_i64)seed64, (xxh_i64)seed64); +# endif + int i; + + XXH_ALIGN(64) const float* const src = (float const*) XXH3_kSecret; + XXH_ALIGN(XXH_SEC_ALIGN) __m128i* dest = (__m128i*) customSecret; +# if defined(__GNUC__) || defined(__clang__) + /* + * On GCC & Clang, marking 'dest' as modified will cause the compiler: + * - do not extract the secret from sse registers in the internal loop + * - use less common registers, and avoid pushing these reg into stack + */ + __asm__("" : "+r" (dest)); +# endif + + for (i=0; i < nbRounds; ++i) { + dest[i] = _mm_add_epi64(_mm_castps_si128(_mm_load_ps(src+i*4)), seed); + } } +} + +#endif + +#if (XXH_VECTOR == XXH_NEON) + +XXH_FORCE_INLINE void +XXH3_accumulate_512_neon( void* XXH_RESTRICT acc, + const void* XXH_RESTRICT input, + const void* XXH_RESTRICT secret) +{ + XXH_ASSERT((((size_t)acc) & 15) == 0); + { + XXH_ALIGN(16) uint64x2_t* const xacc = (uint64x2_t *) acc; + /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. */ + uint8_t const* const xinput = (const uint8_t *) input; + uint8_t const* const xsecret = (const uint8_t *) secret; + + size_t i; + for (i=0; i < XXH_STRIPE_LEN / sizeof(uint64x2_t); i++) { + /* data_vec = xinput[i]; */ + uint8x16_t data_vec = vld1q_u8(xinput + (i * 16)); + /* key_vec = xsecret[i]; */ + uint8x16_t key_vec = vld1q_u8(xsecret + (i * 16)); + uint64x2_t data_key; + uint32x2_t data_key_lo, data_key_hi; + /* xacc[i] += swap(data_vec); */ + uint64x2_t const data64 = vreinterpretq_u64_u8(data_vec); + uint64x2_t const swapped = vextq_u64(data64, data64, 1); + xacc[i] = vaddq_u64 (xacc[i], swapped); + /* data_key = data_vec ^ key_vec; */ + data_key = vreinterpretq_u64_u8(veorq_u8(data_vec, key_vec)); + /* data_key_lo = (uint32x2_t) (data_key & 0xFFFFFFFF); + * data_key_hi = (uint32x2_t) (data_key >> 32); + * data_key = UNDEFINED; */ + XXH_SPLIT_IN_PLACE(data_key, data_key_lo, data_key_hi); + /* xacc[i] += (uint64x2_t) data_key_lo * (uint64x2_t) data_key_hi; */ + xacc[i] = vmlal_u32 (xacc[i], data_key_lo, data_key_hi); + + } + } +} + +XXH_FORCE_INLINE void +XXH3_scrambleAcc_neon(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret) +{ + XXH_ASSERT((((size_t)acc) & 15) == 0); + + { uint64x2_t* xacc = (uint64x2_t*) acc; + uint8_t const* xsecret = (uint8_t const*) secret; + uint32x2_t prime = vdup_n_u32 (XXH_PRIME32_1); + + size_t i; + for (i=0; i < XXH_STRIPE_LEN/sizeof(uint64x2_t); i++) { + /* xacc[i] ^= (xacc[i] >> 47); */ + uint64x2_t acc_vec = xacc[i]; + uint64x2_t shifted = vshrq_n_u64 (acc_vec, 47); + uint64x2_t data_vec = veorq_u64 (acc_vec, shifted); + + /* xacc[i] ^= xsecret[i]; */ + uint8x16_t key_vec = vld1q_u8(xsecret + (i * 16)); + uint64x2_t data_key = veorq_u64(data_vec, vreinterpretq_u64_u8(key_vec)); + + /* xacc[i] *= XXH_PRIME32_1 */ + uint32x2_t data_key_lo, data_key_hi; + /* data_key_lo = (uint32x2_t) (xacc[i] & 0xFFFFFFFF); + * data_key_hi = (uint32x2_t) (xacc[i] >> 32); + * xacc[i] = UNDEFINED; */ + XXH_SPLIT_IN_PLACE(data_key, data_key_lo, data_key_hi); + { /* + * prod_hi = (data_key >> 32) * XXH_PRIME32_1; + * + * Avoid vmul_u32 + vshll_n_u32 since Clang 6 and 7 will + * incorrectly "optimize" this: + * tmp = vmul_u32(vmovn_u64(a), vmovn_u64(b)); + * shifted = vshll_n_u32(tmp, 32); + * to this: + * tmp = "vmulq_u64"(a, b); // no such thing! + * shifted = vshlq_n_u64(tmp, 32); + * + * However, unlike SSE, Clang lacks a 64-bit multiply routine + * for NEON, and it scalarizes two 64-bit multiplies instead. + * + * vmull_u32 has the same timing as vmul_u32, and it avoids + * this bug completely. + * See https://bugs.llvm.org/show_bug.cgi?id=39967 + */ + uint64x2_t prod_hi = vmull_u32 (data_key_hi, prime); + /* xacc[i] = prod_hi << 32; */ + xacc[i] = vshlq_n_u64(prod_hi, 32); + /* xacc[i] += (prod_hi & 0xFFFFFFFF) * XXH_PRIME32_1; */ + xacc[i] = vmlal_u32(xacc[i], data_key_lo, prime); + } + } } +} + +#endif + +#if (XXH_VECTOR == XXH_VSX) + +XXH_FORCE_INLINE void +XXH3_accumulate_512_vsx( void* XXH_RESTRICT acc, + const void* XXH_RESTRICT input, + const void* XXH_RESTRICT secret) +{ + xxh_u64x2* const xacc = (xxh_u64x2*) acc; /* presumed aligned */ + xxh_u64x2 const* const xinput = (xxh_u64x2 const*) input; /* no alignment restriction */ + xxh_u64x2 const* const xsecret = (xxh_u64x2 const*) secret; /* no alignment restriction */ + xxh_u64x2 const v32 = { 32, 32 }; + size_t i; + for (i = 0; i < XXH_STRIPE_LEN / sizeof(xxh_u64x2); i++) { + /* data_vec = xinput[i]; */ + xxh_u64x2 const data_vec = XXH_vec_loadu(xinput + i); + /* key_vec = xsecret[i]; */ + xxh_u64x2 const key_vec = XXH_vec_loadu(xsecret + i); + xxh_u64x2 const data_key = data_vec ^ key_vec; + /* shuffled = (data_key << 32) | (data_key >> 32); */ + xxh_u32x4 const shuffled = (xxh_u32x4)vec_rl(data_key, v32); + /* product = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)shuffled & 0xFFFFFFFF); */ + xxh_u64x2 const product = XXH_vec_mulo((xxh_u32x4)data_key, shuffled); + xacc[i] += product; + + /* swap high and low halves */ +#ifdef __s390x__ + xacc[i] += vec_permi(data_vec, data_vec, 2); +#else + xacc[i] += vec_xxpermdi(data_vec, data_vec, 2); +#endif + } +} + +XXH_FORCE_INLINE void +XXH3_scrambleAcc_vsx(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret) +{ + XXH_ASSERT((((size_t)acc) & 15) == 0); + + { xxh_u64x2* const xacc = (xxh_u64x2*) acc; + const xxh_u64x2* const xsecret = (const xxh_u64x2*) secret; + /* constants */ + xxh_u64x2 const v32 = { 32, 32 }; + xxh_u64x2 const v47 = { 47, 47 }; + xxh_u32x4 const prime = { XXH_PRIME32_1, XXH_PRIME32_1, XXH_PRIME32_1, XXH_PRIME32_1 }; + size_t i; + for (i = 0; i < XXH_STRIPE_LEN / sizeof(xxh_u64x2); i++) { + /* xacc[i] ^= (xacc[i] >> 47); */ + xxh_u64x2 const acc_vec = xacc[i]; + xxh_u64x2 const data_vec = acc_vec ^ (acc_vec >> v47); + + /* xacc[i] ^= xsecret[i]; */ + xxh_u64x2 const key_vec = XXH_vec_loadu(xsecret + i); + xxh_u64x2 const data_key = data_vec ^ key_vec; + + /* xacc[i] *= XXH_PRIME32_1 */ + /* prod_lo = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)prime & 0xFFFFFFFF); */ + xxh_u64x2 const prod_even = XXH_vec_mule((xxh_u32x4)data_key, prime); + /* prod_hi = ((xxh_u64x2)data_key >> 32) * ((xxh_u64x2)prime >> 32); */ + xxh_u64x2 const prod_odd = XXH_vec_mulo((xxh_u32x4)data_key, prime); + xacc[i] = prod_odd + (prod_even << v32); + } } +} + +#endif + +/* scalar variants - universal */ + +XXH_FORCE_INLINE void +XXH3_accumulate_512_scalar(void* XXH_RESTRICT acc, + const void* XXH_RESTRICT input, + const void* XXH_RESTRICT secret) +{ + XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64* const xacc = (xxh_u64*) acc; /* presumed aligned */ + const xxh_u8* const xinput = (const xxh_u8*) input; /* no alignment restriction */ + const xxh_u8* const xsecret = (const xxh_u8*) secret; /* no alignment restriction */ + size_t i; + XXH_ASSERT(((size_t)acc & (XXH_ACC_ALIGN-1)) == 0); + for (i=0; i < XXH_ACC_NB; i++) { + xxh_u64 const data_val = XXH_readLE64(xinput + 8*i); + xxh_u64 const data_key = data_val ^ XXH_readLE64(xsecret + i*8); + xacc[i ^ 1] += data_val; /* swap adjacent lanes */ + xacc[i] += XXH_mult32to64(data_key & 0xFFFFFFFF, data_key >> 32); + } +} + +XXH_FORCE_INLINE void +XXH3_scrambleAcc_scalar(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret) +{ + XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64* const xacc = (xxh_u64*) acc; /* presumed aligned */ + const xxh_u8* const xsecret = (const xxh_u8*) secret; /* no alignment restriction */ + size_t i; + XXH_ASSERT((((size_t)acc) & (XXH_ACC_ALIGN-1)) == 0); + for (i=0; i < XXH_ACC_NB; i++) { + xxh_u64 const key64 = XXH_readLE64(xsecret + 8*i); + xxh_u64 acc64 = xacc[i]; + acc64 = XXH_xorshift64(acc64, 47); + acc64 ^= key64; + acc64 *= XXH_PRIME32_1; + xacc[i] = acc64; + } +} + +XXH_FORCE_INLINE void +XXH3_initCustomSecret_scalar(void* XXH_RESTRICT customSecret, xxh_u64 seed64) +{ + /* + * We need a separate pointer for the hack below, + * which requires a non-const pointer. + * Any decent compiler will optimize this out otherwise. + */ + const xxh_u8* kSecretPtr = XXH3_kSecret; + XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0); + +#if defined(__clang__) && defined(__aarch64__) + /* + * UGLY HACK: + * Clang generates a bunch of MOV/MOVK pairs for aarch64, and they are + * placed sequentially, in order, at the top of the unrolled loop. + * + * While MOVK is great for generating constants (2 cycles for a 64-bit + * constant compared to 4 cycles for LDR), long MOVK chains stall the + * integer pipelines: + * I L S + * MOVK + * MOVK + * MOVK + * MOVK + * ADD + * SUB STR + * STR + * By forcing loads from memory (as the asm line causes Clang to assume + * that XXH3_kSecretPtr has been changed), the pipelines are used more + * efficiently: + * I L S + * LDR + * ADD LDR + * SUB STR + * STR + * XXH3_64bits_withSeed, len == 256, Snapdragon 835 + * without hack: 2654.4 MB/s + * with hack: 3202.9 MB/s + */ + __asm__("" : "+r" (kSecretPtr)); +#endif + /* + * Note: in debug mode, this overrides the asm optimization + * and Clang will emit MOVK chains again. + */ + XXH_ASSERT(kSecretPtr == XXH3_kSecret); + + { int const nbRounds = XXH_SECRET_DEFAULT_SIZE / 16; + int i; + for (i=0; i < nbRounds; i++) { + /* + * The asm hack causes Clang to assume that kSecretPtr aliases with + * customSecret, and on aarch64, this prevented LDP from merging two + * loads together for free. Putting the loads together before the stores + * properly generates LDP. + */ + xxh_u64 lo = XXH_readLE64(kSecretPtr + 16*i) + seed64; + xxh_u64 hi = XXH_readLE64(kSecretPtr + 16*i + 8) - seed64; + XXH_writeLE64((xxh_u8*)customSecret + 16*i, lo); + XXH_writeLE64((xxh_u8*)customSecret + 16*i + 8, hi); + } } +} + + +typedef void (*XXH3_f_accumulate_512)(void* XXH_RESTRICT, const void*, const void*); +typedef void (*XXH3_f_scrambleAcc)(void* XXH_RESTRICT, const void*); +typedef void (*XXH3_f_initCustomSecret)(void* XXH_RESTRICT, xxh_u64); + + +#if (XXH_VECTOR == XXH_AVX512) + +#define XXH3_accumulate_512 XXH3_accumulate_512_avx512 +#define XXH3_scrambleAcc XXH3_scrambleAcc_avx512 +#define XXH3_initCustomSecret XXH3_initCustomSecret_avx512 + +#elif (XXH_VECTOR == XXH_AVX2) + +#define XXH3_accumulate_512 XXH3_accumulate_512_avx2 +#define XXH3_scrambleAcc XXH3_scrambleAcc_avx2 +#define XXH3_initCustomSecret XXH3_initCustomSecret_avx2 + +#elif (XXH_VECTOR == XXH_SSE2) + +#define XXH3_accumulate_512 XXH3_accumulate_512_sse2 +#define XXH3_scrambleAcc XXH3_scrambleAcc_sse2 +#define XXH3_initCustomSecret XXH3_initCustomSecret_sse2 + +#elif (XXH_VECTOR == XXH_NEON) + +#define XXH3_accumulate_512 XXH3_accumulate_512_neon +#define XXH3_scrambleAcc XXH3_scrambleAcc_neon +#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar + +#elif (XXH_VECTOR == XXH_VSX) + +#define XXH3_accumulate_512 XXH3_accumulate_512_vsx +#define XXH3_scrambleAcc XXH3_scrambleAcc_vsx +#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar + +#else /* scalar */ + +#define XXH3_accumulate_512 XXH3_accumulate_512_scalar +#define XXH3_scrambleAcc XXH3_scrambleAcc_scalar +#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar + +#endif + + + +#ifndef XXH_PREFETCH_DIST +# ifdef __clang__ +# define XXH_PREFETCH_DIST 320 +# else +# if (XXH_VECTOR == XXH_AVX512) +# define XXH_PREFETCH_DIST 512 +# else +# define XXH_PREFETCH_DIST 384 +# endif +# endif /* __clang__ */ +#endif /* XXH_PREFETCH_DIST */ + +/* + * XXH3_accumulate() + * Loops over XXH3_accumulate_512(). + * Assumption: nbStripes will not overflow the secret size + */ +XXH_FORCE_INLINE void +XXH3_accumulate( xxh_u64* XXH_RESTRICT acc, + const xxh_u8* XXH_RESTRICT input, + const xxh_u8* XXH_RESTRICT secret, + size_t nbStripes, + XXH3_f_accumulate_512 f_acc512) +{ + size_t n; + for (n = 0; n < nbStripes; n++ ) { + const xxh_u8* const in = input + n*XXH_STRIPE_LEN; + XXH_PREFETCH(in + XXH_PREFETCH_DIST); + f_acc512(acc, + in, + secret + n*XXH_SECRET_CONSUME_RATE); + } +} + +XXH_FORCE_INLINE void +XXH3_hashLong_internal_loop(xxh_u64* XXH_RESTRICT acc, + const xxh_u8* XXH_RESTRICT input, size_t len, + const xxh_u8* XXH_RESTRICT secret, size_t secretSize, + XXH3_f_accumulate_512 f_acc512, + XXH3_f_scrambleAcc f_scramble) +{ + size_t const nbStripesPerBlock = (secretSize - XXH_STRIPE_LEN) / XXH_SECRET_CONSUME_RATE; + size_t const block_len = XXH_STRIPE_LEN * nbStripesPerBlock; + size_t const nb_blocks = (len - 1) / block_len; + + size_t n; + + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); + + for (n = 0; n < nb_blocks; n++) { + XXH3_accumulate(acc, input + n*block_len, secret, nbStripesPerBlock, f_acc512); + f_scramble(acc, secret + secretSize - XXH_STRIPE_LEN); + } + + /* last partial block */ + XXH_ASSERT(len > XXH_STRIPE_LEN); + { size_t const nbStripes = ((len - 1) - (block_len * nb_blocks)) / XXH_STRIPE_LEN; + XXH_ASSERT(nbStripes <= (secretSize / XXH_SECRET_CONSUME_RATE)); + XXH3_accumulate(acc, input + nb_blocks*block_len, secret, nbStripes, f_acc512); + + /* last stripe */ + { const xxh_u8* const p = input + len - XXH_STRIPE_LEN; +#define XXH_SECRET_LASTACC_START 7 /* not aligned on 8, last secret is different from acc & scrambler */ + f_acc512(acc, p, secret + secretSize - XXH_STRIPE_LEN - XXH_SECRET_LASTACC_START); + } } +} + +XXH_FORCE_INLINE xxh_u64 +XXH3_mix2Accs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret) +{ + return XXH3_mul128_fold64( + acc[0] ^ XXH_readLE64(secret), + acc[1] ^ XXH_readLE64(secret+8) ); +} + +static XXH64_hash_t +XXH3_mergeAccs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret, xxh_u64 start) +{ + xxh_u64 result64 = start; + size_t i = 0; + + for (i = 0; i < 4; i++) { + result64 += XXH3_mix2Accs(acc+2*i, secret + 16*i); +#if defined(__clang__) /* Clang */ \ + && (defined(__arm__) || defined(__thumb__)) /* ARMv7 */ \ + && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \ + && !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable */ + /* + * UGLY HACK: + * Prevent autovectorization on Clang ARMv7-a. Exact same problem as + * the one in XXH3_len_129to240_64b. Speeds up shorter keys > 240b. + * XXH3_64bits, len == 256, Snapdragon 835: + * without hack: 2063.7 MB/s + * with hack: 2560.7 MB/s + */ + __asm__("" : "+r" (result64)); +#endif + } + + return XXH3_avalanche(result64); +} + +#define XXH3_INIT_ACC { XXH_PRIME32_3, XXH_PRIME64_1, XXH_PRIME64_2, XXH_PRIME64_3, \ + XXH_PRIME64_4, XXH_PRIME32_2, XXH_PRIME64_5, XXH_PRIME32_1 } + +XXH_FORCE_INLINE XXH64_hash_t +XXH3_hashLong_64b_internal(const void* XXH_RESTRICT input, size_t len, + const void* XXH_RESTRICT secret, size_t secretSize, + XXH3_f_accumulate_512 f_acc512, + XXH3_f_scrambleAcc f_scramble) +{ + XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC; + + XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, f_acc512, f_scramble); + + /* converge into final hash */ + XXH_STATIC_ASSERT(sizeof(acc) == 64); + /* do not align on 8, so that the secret is different from the accumulator */ +#define XXH_SECRET_MERGEACCS_START 11 + XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START); + return XXH3_mergeAccs(acc, (const xxh_u8*)secret + XXH_SECRET_MERGEACCS_START, (xxh_u64)len * XXH_PRIME64_1); +} + +/* + * It's important for performance that XXH3_hashLong is not inlined. + */ +XXH_NO_INLINE XXH64_hash_t +XXH3_hashLong_64b_withSecret(const void* XXH_RESTRICT input, size_t len, + XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen) +{ + (void)seed64; + return XXH3_hashLong_64b_internal(input, len, secret, secretLen, XXH3_accumulate_512, XXH3_scrambleAcc); +} + +/* + * It's important for performance that XXH3_hashLong is not inlined. + * Since the function is not inlined, the compiler may not be able to understand that, + * in some scenarios, its `secret` argument is actually a compile time constant. + * This variant enforces that the compiler can detect that, + * and uses this opportunity to streamline the generated code for better performance. + */ +XXH_NO_INLINE XXH64_hash_t +XXH3_hashLong_64b_default(const void* XXH_RESTRICT input, size_t len, + XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen) +{ + (void)seed64; (void)secret; (void)secretLen; + return XXH3_hashLong_64b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_accumulate_512, XXH3_scrambleAcc); +} + +/* + * XXH3_hashLong_64b_withSeed(): + * Generate a custom key based on alteration of default XXH3_kSecret with the seed, + * and then use this key for long mode hashing. + * + * This operation is decently fast but nonetheless costs a little bit of time. + * Try to avoid it whenever possible (typically when seed==0). + * + * It's important for performance that XXH3_hashLong is not inlined. Not sure + * why (uop cache maybe?), but the difference is large and easily measurable. + */ +XXH_FORCE_INLINE XXH64_hash_t +XXH3_hashLong_64b_withSeed_internal(const void* input, size_t len, + XXH64_hash_t seed, + XXH3_f_accumulate_512 f_acc512, + XXH3_f_scrambleAcc f_scramble, + XXH3_f_initCustomSecret f_initSec) +{ + if (seed == 0) + return XXH3_hashLong_64b_internal(input, len, + XXH3_kSecret, sizeof(XXH3_kSecret), + f_acc512, f_scramble); + { XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE]; + f_initSec(secret, seed); + return XXH3_hashLong_64b_internal(input, len, secret, sizeof(secret), + f_acc512, f_scramble); + } +} + +/* + * It's important for performance that XXH3_hashLong is not inlined. + */ +XXH_NO_INLINE XXH64_hash_t +XXH3_hashLong_64b_withSeed(const void* input, size_t len, + XXH64_hash_t seed, const xxh_u8* secret, size_t secretLen) +{ + (void)secret; (void)secretLen; + return XXH3_hashLong_64b_withSeed_internal(input, len, seed, + XXH3_accumulate_512, XXH3_scrambleAcc, XXH3_initCustomSecret); +} + + +typedef XXH64_hash_t (*XXH3_hashLong64_f)(const void* XXH_RESTRICT, size_t, + XXH64_hash_t, const xxh_u8* XXH_RESTRICT, size_t); + +XXH_FORCE_INLINE XXH64_hash_t +XXH3_64bits_internal(const void* XXH_RESTRICT input, size_t len, + XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen, + XXH3_hashLong64_f f_hashLong) +{ + XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN); + /* + * If an action is to be taken if `secretLen` condition is not respected, + * it should be done here. + * For now, it's a contract pre-condition. + * Adding a check and a branch here would cost performance at every hash. + * Also, note that function signature doesn't offer room to return an error. + */ + if (len <= 16) + return XXH3_len_0to16_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, seed64); + if (len <= 128) + return XXH3_len_17to128_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64); + if (len <= XXH3_MIDSIZE_MAX) + return XXH3_len_129to240_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64); + return f_hashLong(input, len, seed64, (const xxh_u8*)secret, secretLen); +} + + +/* === Public entry point === */ + +XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(const void* input, size_t len) +{ + return XXH3_64bits_internal(input, len, 0, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_default); +} + +XXH_PUBLIC_API XXH64_hash_t +XXH3_64bits_withSecret(const void* input, size_t len, const void* secret, size_t secretSize) +{ + return XXH3_64bits_internal(input, len, 0, secret, secretSize, XXH3_hashLong_64b_withSecret); +} + +XXH_PUBLIC_API XXH64_hash_t +XXH3_64bits_withSeed(const void* input, size_t len, XXH64_hash_t seed) +{ + return XXH3_64bits_internal(input, len, seed, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_withSeed); +} + + +/* === XXH3 streaming === */ + +/* + * Malloc's a pointer that is always aligned to align. + * + * This must be freed with `XXH_alignedFree()`. + * + * malloc typically guarantees 16 byte alignment on 64-bit systems and 8 byte + * alignment on 32-bit. This isn't enough for the 32 byte aligned loads in AVX2 + * or on 32-bit, the 16 byte aligned loads in SSE2 and NEON. + * + * This underalignment previously caused a rather obvious crash which went + * completely unnoticed due to XXH3_createState() not actually being tested. + * Credit to RedSpah for noticing this bug. + * + * The alignment is done manually: Functions like posix_memalign or _mm_malloc + * are avoided: To maintain portability, we would have to write a fallback + * like this anyways, and besides, testing for the existence of library + * functions without relying on external build tools is impossible. + * + * The method is simple: Overallocate, manually align, and store the offset + * to the original behind the returned pointer. + * + * Align must be a power of 2 and 8 <= align <= 128. + */ +static void* XXH_alignedMalloc(size_t s, size_t align) +{ + XXH_ASSERT(align <= 128 && align >= 8); /* range check */ + XXH_ASSERT((align & (align-1)) == 0); /* power of 2 */ + XXH_ASSERT(s != 0 && s < (s + align)); /* empty/overflow */ + { /* Overallocate to make room for manual realignment and an offset byte */ + xxh_u8* base = (xxh_u8*)XXH_malloc(s + align); + if (base != NULL) { + /* + * Get the offset needed to align this pointer. + * + * Even if the returned pointer is aligned, there will always be + * at least one byte to store the offset to the original pointer. + */ + size_t offset = align - ((size_t)base & (align - 1)); /* base % align */ + /* Add the offset for the now-aligned pointer */ + xxh_u8* ptr = base + offset; + + XXH_ASSERT((size_t)ptr % align == 0); + + /* Store the offset immediately before the returned pointer. */ + ptr[-1] = (xxh_u8)offset; + return ptr; + } + return NULL; + } +} +/* + * Frees an aligned pointer allocated by XXH_alignedMalloc(). Don't pass + * normal malloc'd pointers, XXH_alignedMalloc has a specific data layout. + */ +static void XXH_alignedFree(void* p) +{ + if (p != NULL) { + xxh_u8* ptr = (xxh_u8*)p; + /* Get the offset byte we added in XXH_malloc. */ + xxh_u8 offset = ptr[-1]; + /* Free the original malloc'd pointer */ + xxh_u8* base = ptr - offset; + XXH_free(base); + } +} +XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void) +{ + XXH3_state_t* const state = (XXH3_state_t*)XXH_alignedMalloc(sizeof(XXH3_state_t), 64); + if (state==NULL) return NULL; + XXH3_INITSTATE(state); + return state; +} + +XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr) +{ + XXH_alignedFree(statePtr); + return XXH_OK; +} + +XXH_PUBLIC_API void +XXH3_copyState(XXH3_state_t* dst_state, const XXH3_state_t* src_state) +{ + memcpy(dst_state, src_state, sizeof(*dst_state)); +} + +static void +XXH3_64bits_reset_internal(XXH3_state_t* statePtr, + XXH64_hash_t seed, + const void* secret, size_t secretSize) +{ + size_t const initStart = offsetof(XXH3_state_t, bufferedSize); + size_t const initLength = offsetof(XXH3_state_t, nbStripesPerBlock) - initStart; + XXH_ASSERT(offsetof(XXH3_state_t, nbStripesPerBlock) > initStart); + XXH_ASSERT(statePtr != NULL); + /* set members from bufferedSize to nbStripesPerBlock (excluded) to 0 */ + memset((char*)statePtr + initStart, 0, initLength); + statePtr->acc[0] = XXH_PRIME32_3; + statePtr->acc[1] = XXH_PRIME64_1; + statePtr->acc[2] = XXH_PRIME64_2; + statePtr->acc[3] = XXH_PRIME64_3; + statePtr->acc[4] = XXH_PRIME64_4; + statePtr->acc[5] = XXH_PRIME32_2; + statePtr->acc[6] = XXH_PRIME64_5; + statePtr->acc[7] = XXH_PRIME32_1; + statePtr->seed = seed; + statePtr->extSecret = (const unsigned char*)secret; + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); + statePtr->secretLimit = secretSize - XXH_STRIPE_LEN; + statePtr->nbStripesPerBlock = statePtr->secretLimit / XXH_SECRET_CONSUME_RATE; +} + +XXH_PUBLIC_API XXH_errorcode +XXH3_64bits_reset(XXH3_state_t* statePtr) +{ + if (statePtr == NULL) return XXH_ERROR; + XXH3_64bits_reset_internal(statePtr, 0, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE); + return XXH_OK; +} + +XXH_PUBLIC_API XXH_errorcode +XXH3_64bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize) +{ + if (statePtr == NULL) return XXH_ERROR; + XXH3_64bits_reset_internal(statePtr, 0, secret, secretSize); + if (secret == NULL) return XXH_ERROR; + if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR; + return XXH_OK; +} + +XXH_PUBLIC_API XXH_errorcode +XXH3_64bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed) +{ + if (statePtr == NULL) return XXH_ERROR; + if (seed==0) return XXH3_64bits_reset(statePtr); + if (seed != statePtr->seed) XXH3_initCustomSecret(statePtr->customSecret, seed); + XXH3_64bits_reset_internal(statePtr, seed, NULL, XXH_SECRET_DEFAULT_SIZE); + return XXH_OK; +} + +/* Note : when XXH3_consumeStripes() is invoked, + * there must be a guarantee that at least one more byte must be consumed from input + * so that the function can blindly consume all stripes using the "normal" secret segment */ +XXH_FORCE_INLINE void +XXH3_consumeStripes(xxh_u64* XXH_RESTRICT acc, + size_t* XXH_RESTRICT nbStripesSoFarPtr, size_t nbStripesPerBlock, + const xxh_u8* XXH_RESTRICT input, size_t nbStripes, + const xxh_u8* XXH_RESTRICT secret, size_t secretLimit, + XXH3_f_accumulate_512 f_acc512, + XXH3_f_scrambleAcc f_scramble) +{ + XXH_ASSERT(nbStripes <= nbStripesPerBlock); /* can handle max 1 scramble per invocation */ + XXH_ASSERT(*nbStripesSoFarPtr < nbStripesPerBlock); + if (nbStripesPerBlock - *nbStripesSoFarPtr <= nbStripes) { + /* need a scrambling operation */ + size_t const nbStripesToEndofBlock = nbStripesPerBlock - *nbStripesSoFarPtr; + size_t const nbStripesAfterBlock = nbStripes - nbStripesToEndofBlock; + XXH3_accumulate(acc, input, secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, nbStripesToEndofBlock, f_acc512); + f_scramble(acc, secret + secretLimit); + XXH3_accumulate(acc, input + nbStripesToEndofBlock * XXH_STRIPE_LEN, secret, nbStripesAfterBlock, f_acc512); + *nbStripesSoFarPtr = nbStripesAfterBlock; + } else { + XXH3_accumulate(acc, input, secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, nbStripes, f_acc512); + *nbStripesSoFarPtr += nbStripes; + } +} + +/* + * Both XXH3_64bits_update and XXH3_128bits_update use this routine. + */ +XXH_FORCE_INLINE XXH_errorcode +XXH3_update(XXH3_state_t* state, + const xxh_u8* input, size_t len, + XXH3_f_accumulate_512 f_acc512, + XXH3_f_scrambleAcc f_scramble) +{ + if (input==NULL) +#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && (XXH_ACCEPT_NULL_INPUT_POINTER>=1) + return XXH_OK; +#else + return XXH_ERROR; +#endif + + { const xxh_u8* const bEnd = input + len; + const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret; + state->totalLen += len; + + if (state->bufferedSize + len <= XXH3_INTERNALBUFFER_SIZE) { /* fill in tmp buffer */ + XXH_memcpy(state->buffer + state->bufferedSize, input, len); + state->bufferedSize += (XXH32_hash_t)len; + return XXH_OK; + } + /* total input is now > XXH3_INTERNALBUFFER_SIZE */ + + #define XXH3_INTERNALBUFFER_STRIPES (XXH3_INTERNALBUFFER_SIZE / XXH_STRIPE_LEN) + XXH_STATIC_ASSERT(XXH3_INTERNALBUFFER_SIZE % XXH_STRIPE_LEN == 0); /* clean multiple */ + + /* + * Internal buffer is partially filled (always, except at beginning) + * Complete it, then consume it. + */ + if (state->bufferedSize) { + size_t const loadSize = XXH3_INTERNALBUFFER_SIZE - state->bufferedSize; + XXH_memcpy(state->buffer + state->bufferedSize, input, loadSize); + input += loadSize; + XXH3_consumeStripes(state->acc, + &state->nbStripesSoFar, state->nbStripesPerBlock, + state->buffer, XXH3_INTERNALBUFFER_STRIPES, + secret, state->secretLimit, + f_acc512, f_scramble); + state->bufferedSize = 0; + } + XXH_ASSERT(input < bEnd); + + /* Consume input by a multiple of internal buffer size */ + if (input+XXH3_INTERNALBUFFER_SIZE < bEnd) { + const xxh_u8* const limit = bEnd - XXH3_INTERNALBUFFER_SIZE; + do { + XXH3_consumeStripes(state->acc, + &state->nbStripesSoFar, state->nbStripesPerBlock, + input, XXH3_INTERNALBUFFER_STRIPES, + secret, state->secretLimit, + f_acc512, f_scramble); + input += XXH3_INTERNALBUFFER_SIZE; + } while (input<limit); + /* for last partial stripe */ + memcpy(state->buffer + sizeof(state->buffer) - XXH_STRIPE_LEN, input - XXH_STRIPE_LEN, XXH_STRIPE_LEN); + } + XXH_ASSERT(input < bEnd); + + /* Some remaining input (always) : buffer it */ + XXH_memcpy(state->buffer, input, (size_t)(bEnd-input)); + state->bufferedSize = (XXH32_hash_t)(bEnd-input); + } + + return XXH_OK; +} + +XXH_PUBLIC_API XXH_errorcode +XXH3_64bits_update(XXH3_state_t* state, const void* input, size_t len) +{ + return XXH3_update(state, (const xxh_u8*)input, len, + XXH3_accumulate_512, XXH3_scrambleAcc); +} + + +XXH_FORCE_INLINE void +XXH3_digest_long (XXH64_hash_t* acc, + const XXH3_state_t* state, + const unsigned char* secret) +{ + /* + * Digest on a local copy. This way, the state remains unaltered, and it can + * continue ingesting more input afterwards. + */ + memcpy(acc, state->acc, sizeof(state->acc)); + if (state->bufferedSize >= XXH_STRIPE_LEN) { + size_t const nbStripes = (state->bufferedSize - 1) / XXH_STRIPE_LEN; + size_t nbStripesSoFar = state->nbStripesSoFar; + XXH3_consumeStripes(acc, + &nbStripesSoFar, state->nbStripesPerBlock, + state->buffer, nbStripes, + secret, state->secretLimit, + XXH3_accumulate_512, XXH3_scrambleAcc); + /* last stripe */ + XXH3_accumulate_512(acc, + state->buffer + state->bufferedSize - XXH_STRIPE_LEN, + secret + state->secretLimit - XXH_SECRET_LASTACC_START); + } else { /* bufferedSize < XXH_STRIPE_LEN */ + xxh_u8 lastStripe[XXH_STRIPE_LEN]; + size_t const catchupSize = XXH_STRIPE_LEN - state->bufferedSize; + XXH_ASSERT(state->bufferedSize > 0); /* there is always some input buffered */ + memcpy(lastStripe, state->buffer + sizeof(state->buffer) - catchupSize, catchupSize); + memcpy(lastStripe + catchupSize, state->buffer, state->bufferedSize); + XXH3_accumulate_512(acc, + lastStripe, + secret + state->secretLimit - XXH_SECRET_LASTACC_START); + } +} + +XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (const XXH3_state_t* state) +{ + const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret; + if (state->totalLen > XXH3_MIDSIZE_MAX) { + XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[XXH_ACC_NB]; + XXH3_digest_long(acc, state, secret); + return XXH3_mergeAccs(acc, + secret + XXH_SECRET_MERGEACCS_START, + (xxh_u64)state->totalLen * XXH_PRIME64_1); + } + /* totalLen <= XXH3_MIDSIZE_MAX: digesting a short input */ + if (state->seed) + return XXH3_64bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed); + return XXH3_64bits_withSecret(state->buffer, (size_t)(state->totalLen), + secret, state->secretLimit + XXH_STRIPE_LEN); +} + + +#define XXH_MIN(x, y) (((x) > (y)) ? (y) : (x)) + +XXH_PUBLIC_API void +XXH3_generateSecret(void* secretBuffer, const void* customSeed, size_t customSeedSize) +{ + XXH_ASSERT(secretBuffer != NULL); + if (customSeedSize == 0) { + memcpy(secretBuffer, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE); + return; + } + XXH_ASSERT(customSeed != NULL); + + { size_t const segmentSize = sizeof(XXH128_hash_t); + size_t const nbSegments = XXH_SECRET_DEFAULT_SIZE / segmentSize; + XXH128_canonical_t scrambler; + XXH64_hash_t seeds[12]; + size_t segnb; + XXH_ASSERT(nbSegments == 12); + XXH_ASSERT(segmentSize * nbSegments == XXH_SECRET_DEFAULT_SIZE); /* exact multiple */ + XXH128_canonicalFromHash(&scrambler, XXH128(customSeed, customSeedSize, 0)); + + /* + * Copy customSeed to seeds[], truncating or repeating as necessary. + */ + { size_t toFill = XXH_MIN(customSeedSize, sizeof(seeds)); + size_t filled = toFill; + memcpy(seeds, customSeed, toFill); + while (filled < sizeof(seeds)) { + toFill = XXH_MIN(filled, sizeof(seeds) - filled); + memcpy((char*)seeds + filled, seeds, toFill); + filled += toFill; + } } + + /* generate secret */ + memcpy(secretBuffer, &scrambler, sizeof(scrambler)); + for (segnb=1; segnb < nbSegments; segnb++) { + size_t const segmentStart = segnb * segmentSize; + XXH128_canonical_t segment; + XXH128_canonicalFromHash(&segment, + XXH128(&scrambler, sizeof(scrambler), XXH_readLE64(seeds + segnb) + segnb) ); + memcpy((char*)secretBuffer + segmentStart, &segment, sizeof(segment)); + } } +} + + +/* ========================================== + * XXH3 128 bits (a.k.a XXH128) + * ========================================== + * XXH3's 128-bit variant has better mixing and strength than the 64-bit variant, + * even without counting the significantly larger output size. + * + * For example, extra steps are taken to avoid the seed-dependent collisions + * in 17-240 byte inputs (See XXH3_mix16B and XXH128_mix32B). + * + * This strength naturally comes at the cost of some speed, especially on short + * lengths. Note that longer hashes are about as fast as the 64-bit version + * due to it using only a slight modification of the 64-bit loop. + * + * XXH128 is also more oriented towards 64-bit machines. It is still extremely + * fast for a _128-bit_ hash on 32-bit (it usually clears XXH64). + */ + +XXH_FORCE_INLINE XXH128_hash_t +XXH3_len_1to3_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + /* A doubled version of 1to3_64b with different constants. */ + XXH_ASSERT(input != NULL); + XXH_ASSERT(1 <= len && len <= 3); + XXH_ASSERT(secret != NULL); + /* + * len = 1: combinedl = { input[0], 0x01, input[0], input[0] } + * len = 2: combinedl = { input[1], 0x02, input[0], input[1] } + * len = 3: combinedl = { input[2], 0x03, input[0], input[1] } + */ + { xxh_u8 const c1 = input[0]; + xxh_u8 const c2 = input[len >> 1]; + xxh_u8 const c3 = input[len - 1]; + xxh_u32 const combinedl = ((xxh_u32)c1 <<16) | ((xxh_u32)c2 << 24) + | ((xxh_u32)c3 << 0) | ((xxh_u32)len << 8); + xxh_u32 const combinedh = XXH_rotl32(XXH_swap32(combinedl), 13); + xxh_u64 const bitflipl = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed; + xxh_u64 const bitfliph = (XXH_readLE32(secret+8) ^ XXH_readLE32(secret+12)) - seed; + xxh_u64 const keyed_lo = (xxh_u64)combinedl ^ bitflipl; + xxh_u64 const keyed_hi = (xxh_u64)combinedh ^ bitfliph; + XXH128_hash_t h128; + h128.low64 = XXH64_avalanche(keyed_lo); + h128.high64 = XXH64_avalanche(keyed_hi); + return h128; + } +} + +XXH_FORCE_INLINE XXH128_hash_t +XXH3_len_4to8_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(input != NULL); + XXH_ASSERT(secret != NULL); + XXH_ASSERT(4 <= len && len <= 8); + seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32; + { xxh_u32 const input_lo = XXH_readLE32(input); + xxh_u32 const input_hi = XXH_readLE32(input + len - 4); + xxh_u64 const input_64 = input_lo + ((xxh_u64)input_hi << 32); + xxh_u64 const bitflip = (XXH_readLE64(secret+16) ^ XXH_readLE64(secret+24)) + seed; + xxh_u64 const keyed = input_64 ^ bitflip; + + /* Shift len to the left to ensure it is even, this avoids even multiplies. */ + XXH128_hash_t m128 = XXH_mult64to128(keyed, XXH_PRIME64_1 + (len << 2)); + + m128.high64 += (m128.low64 << 1); + m128.low64 ^= (m128.high64 >> 3); + + m128.low64 = XXH_xorshift64(m128.low64, 35); + m128.low64 *= 0x9FB21C651E98DF25ULL; + m128.low64 = XXH_xorshift64(m128.low64, 28); + m128.high64 = XXH3_avalanche(m128.high64); + return m128; + } +} + +XXH_FORCE_INLINE XXH128_hash_t +XXH3_len_9to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(input != NULL); + XXH_ASSERT(secret != NULL); + XXH_ASSERT(9 <= len && len <= 16); + { xxh_u64 const bitflipl = (XXH_readLE64(secret+32) ^ XXH_readLE64(secret+40)) - seed; + xxh_u64 const bitfliph = (XXH_readLE64(secret+48) ^ XXH_readLE64(secret+56)) + seed; + xxh_u64 const input_lo = XXH_readLE64(input); + xxh_u64 input_hi = XXH_readLE64(input + len - 8); + XXH128_hash_t m128 = XXH_mult64to128(input_lo ^ input_hi ^ bitflipl, XXH_PRIME64_1); + /* + * Put len in the middle of m128 to ensure that the length gets mixed to + * both the low and high bits in the 128x64 multiply below. + */ + m128.low64 += (xxh_u64)(len - 1) << 54; + input_hi ^= bitfliph; + /* + * Add the high 32 bits of input_hi to the high 32 bits of m128, then + * add the long product of the low 32 bits of input_hi and XXH_PRIME32_2 to + * the high 64 bits of m128. + * + * The best approach to this operation is different on 32-bit and 64-bit. + */ + if (sizeof(void *) < sizeof(xxh_u64)) { /* 32-bit */ + /* + * 32-bit optimized version, which is more readable. + * + * On 32-bit, it removes an ADC and delays a dependency between the two + * halves of m128.high64, but it generates an extra mask on 64-bit. + */ + m128.high64 += (input_hi & 0xFFFFFFFF00000000ULL) + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2); + } else { + /* + * 64-bit optimized (albeit more confusing) version. + * + * Uses some properties of addition and multiplication to remove the mask: + * + * Let: + * a = input_hi.lo = (input_hi & 0x00000000FFFFFFFF) + * b = input_hi.hi = (input_hi & 0xFFFFFFFF00000000) + * c = XXH_PRIME32_2 + * + * a + (b * c) + * Inverse Property: x + y - x == y + * a + (b * (1 + c - 1)) + * Distributive Property: x * (y + z) == (x * y) + (x * z) + * a + (b * 1) + (b * (c - 1)) + * Identity Property: x * 1 == x + * a + b + (b * (c - 1)) + * + * Substitute a, b, and c: + * input_hi.hi + input_hi.lo + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1)) + * + * Since input_hi.hi + input_hi.lo == input_hi, we get this: + * input_hi + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1)) + */ + m128.high64 += input_hi + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2 - 1); + } + /* m128 ^= XXH_swap64(m128 >> 64); */ + m128.low64 ^= XXH_swap64(m128.high64); + + { /* 128x64 multiply: h128 = m128 * XXH_PRIME64_2; */ + XXH128_hash_t h128 = XXH_mult64to128(m128.low64, XXH_PRIME64_2); + h128.high64 += m128.high64 * XXH_PRIME64_2; + + h128.low64 = XXH3_avalanche(h128.low64); + h128.high64 = XXH3_avalanche(h128.high64); + return h128; + } } +} + +/* + * Assumption: `secret` size is >= XXH3_SECRET_SIZE_MIN + */ +XXH_FORCE_INLINE XXH128_hash_t +XXH3_len_0to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed) +{ + XXH_ASSERT(len <= 16); + { if (len > 8) return XXH3_len_9to16_128b(input, len, secret, seed); + if (len >= 4) return XXH3_len_4to8_128b(input, len, secret, seed); + if (len) return XXH3_len_1to3_128b(input, len, secret, seed); + { XXH128_hash_t h128; + xxh_u64 const bitflipl = XXH_readLE64(secret+64) ^ XXH_readLE64(secret+72); + xxh_u64 const bitfliph = XXH_readLE64(secret+80) ^ XXH_readLE64(secret+88); + h128.low64 = XXH64_avalanche(seed ^ bitflipl); + h128.high64 = XXH64_avalanche( seed ^ bitfliph); + return h128; + } } +} + +/* + * A bit slower than XXH3_mix16B, but handles multiply by zero better. + */ +XXH_FORCE_INLINE XXH128_hash_t +XXH128_mix32B(XXH128_hash_t acc, const xxh_u8* input_1, const xxh_u8* input_2, + const xxh_u8* secret, XXH64_hash_t seed) +{ + acc.low64 += XXH3_mix16B (input_1, secret+0, seed); + acc.low64 ^= XXH_readLE64(input_2) + XXH_readLE64(input_2 + 8); + acc.high64 += XXH3_mix16B (input_2, secret+16, seed); + acc.high64 ^= XXH_readLE64(input_1) + XXH_readLE64(input_1 + 8); + return acc; +} + + +XXH_FORCE_INLINE XXH128_hash_t +XXH3_len_17to128_128b(const xxh_u8* XXH_RESTRICT input, size_t len, + const xxh_u8* XXH_RESTRICT secret, size_t secretSize, + XXH64_hash_t seed) +{ + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize; + XXH_ASSERT(16 < len && len <= 128); + + { XXH128_hash_t acc; + acc.low64 = len * XXH_PRIME64_1; + acc.high64 = 0; + if (len > 32) { + if (len > 64) { + if (len > 96) { + acc = XXH128_mix32B(acc, input+48, input+len-64, secret+96, seed); + } + acc = XXH128_mix32B(acc, input+32, input+len-48, secret+64, seed); + } + acc = XXH128_mix32B(acc, input+16, input+len-32, secret+32, seed); + } + acc = XXH128_mix32B(acc, input, input+len-16, secret, seed); + { XXH128_hash_t h128; + h128.low64 = acc.low64 + acc.high64; + h128.high64 = (acc.low64 * XXH_PRIME64_1) + + (acc.high64 * XXH_PRIME64_4) + + ((len - seed) * XXH_PRIME64_2); + h128.low64 = XXH3_avalanche(h128.low64); + h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64); + return h128; + } + } +} + +XXH_NO_INLINE XXH128_hash_t +XXH3_len_129to240_128b(const xxh_u8* XXH_RESTRICT input, size_t len, + const xxh_u8* XXH_RESTRICT secret, size_t secretSize, + XXH64_hash_t seed) +{ + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize; + XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX); + + { XXH128_hash_t acc; + int const nbRounds = (int)len / 32; + int i; + acc.low64 = len * XXH_PRIME64_1; + acc.high64 = 0; + for (i=0; i<4; i++) { + acc = XXH128_mix32B(acc, + input + (32 * i), + input + (32 * i) + 16, + secret + (32 * i), + seed); + } + acc.low64 = XXH3_avalanche(acc.low64); + acc.high64 = XXH3_avalanche(acc.high64); + XXH_ASSERT(nbRounds >= 4); + for (i=4 ; i < nbRounds; i++) { + acc = XXH128_mix32B(acc, + input + (32 * i), + input + (32 * i) + 16, + secret + XXH3_MIDSIZE_STARTOFFSET + (32 * (i - 4)), + seed); + } + /* last bytes */ + acc = XXH128_mix32B(acc, + input + len - 16, + input + len - 32, + secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET - 16, + 0ULL - seed); + + { XXH128_hash_t h128; + h128.low64 = acc.low64 + acc.high64; + h128.high64 = (acc.low64 * XXH_PRIME64_1) + + (acc.high64 * XXH_PRIME64_4) + + ((len - seed) * XXH_PRIME64_2); + h128.low64 = XXH3_avalanche(h128.low64); + h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64); + return h128; + } + } +} + +XXH_FORCE_INLINE XXH128_hash_t +XXH3_hashLong_128b_internal(const void* XXH_RESTRICT input, size_t len, + const xxh_u8* XXH_RESTRICT secret, size_t secretSize, + XXH3_f_accumulate_512 f_acc512, + XXH3_f_scrambleAcc f_scramble) +{ + XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC; + + XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, secret, secretSize, f_acc512, f_scramble); + + /* converge into final hash */ + XXH_STATIC_ASSERT(sizeof(acc) == 64); + XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START); + { XXH128_hash_t h128; + h128.low64 = XXH3_mergeAccs(acc, + secret + XXH_SECRET_MERGEACCS_START, + (xxh_u64)len * XXH_PRIME64_1); + h128.high64 = XXH3_mergeAccs(acc, + secret + secretSize + - sizeof(acc) - XXH_SECRET_MERGEACCS_START, + ~((xxh_u64)len * XXH_PRIME64_2)); + return h128; + } +} + +/* + * It's important for performance that XXH3_hashLong is not inlined. + */ +XXH_NO_INLINE XXH128_hash_t +XXH3_hashLong_128b_default(const void* XXH_RESTRICT input, size_t len, + XXH64_hash_t seed64, + const void* XXH_RESTRICT secret, size_t secretLen) +{ + (void)seed64; (void)secret; (void)secretLen; + return XXH3_hashLong_128b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret), + XXH3_accumulate_512, XXH3_scrambleAcc); +} + +/* + * It's important for performance that XXH3_hashLong is not inlined. + */ +XXH_NO_INLINE XXH128_hash_t +XXH3_hashLong_128b_withSecret(const void* XXH_RESTRICT input, size_t len, + XXH64_hash_t seed64, + const void* XXH_RESTRICT secret, size_t secretLen) +{ + (void)seed64; + return XXH3_hashLong_128b_internal(input, len, (const xxh_u8*)secret, secretLen, + XXH3_accumulate_512, XXH3_scrambleAcc); +} + +XXH_FORCE_INLINE XXH128_hash_t +XXH3_hashLong_128b_withSeed_internal(const void* XXH_RESTRICT input, size_t len, + XXH64_hash_t seed64, + XXH3_f_accumulate_512 f_acc512, + XXH3_f_scrambleAcc f_scramble, + XXH3_f_initCustomSecret f_initSec) +{ + if (seed64 == 0) + return XXH3_hashLong_128b_internal(input, len, + XXH3_kSecret, sizeof(XXH3_kSecret), + f_acc512, f_scramble); + { XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE]; + f_initSec(secret, seed64); + return XXH3_hashLong_128b_internal(input, len, (const xxh_u8*)secret, sizeof(secret), + f_acc512, f_scramble); + } +} + +/* + * It's important for performance that XXH3_hashLong is not inlined. + */ +XXH_NO_INLINE XXH128_hash_t +XXH3_hashLong_128b_withSeed(const void* input, size_t len, + XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen) +{ + (void)secret; (void)secretLen; + return XXH3_hashLong_128b_withSeed_internal(input, len, seed64, + XXH3_accumulate_512, XXH3_scrambleAcc, XXH3_initCustomSecret); +} + +typedef XXH128_hash_t (*XXH3_hashLong128_f)(const void* XXH_RESTRICT, size_t, + XXH64_hash_t, const void* XXH_RESTRICT, size_t); + +XXH_FORCE_INLINE XXH128_hash_t +XXH3_128bits_internal(const void* input, size_t len, + XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen, + XXH3_hashLong128_f f_hl128) +{ + XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN); + /* + * If an action is to be taken if `secret` conditions are not respected, + * it should be done here. + * For now, it's a contract pre-condition. + * Adding a check and a branch here would cost performance at every hash. + */ + if (len <= 16) + return XXH3_len_0to16_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, seed64); + if (len <= 128) + return XXH3_len_17to128_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64); + if (len <= XXH3_MIDSIZE_MAX) + return XXH3_len_129to240_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64); + return f_hl128(input, len, seed64, secret, secretLen); +} + + +/* === Public XXH128 API === */ + +XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void* input, size_t len) +{ + return XXH3_128bits_internal(input, len, 0, + XXH3_kSecret, sizeof(XXH3_kSecret), + XXH3_hashLong_128b_default); +} + +XXH_PUBLIC_API XXH128_hash_t +XXH3_128bits_withSecret(const void* input, size_t len, const void* secret, size_t secretSize) +{ + return XXH3_128bits_internal(input, len, 0, + (const xxh_u8*)secret, secretSize, + XXH3_hashLong_128b_withSecret); +} + +XXH_PUBLIC_API XXH128_hash_t +XXH3_128bits_withSeed(const void* input, size_t len, XXH64_hash_t seed) +{ + return XXH3_128bits_internal(input, len, seed, + XXH3_kSecret, sizeof(XXH3_kSecret), + XXH3_hashLong_128b_withSeed); +} + +XXH_PUBLIC_API XXH128_hash_t +XXH128(const void* input, size_t len, XXH64_hash_t seed) +{ + return XXH3_128bits_withSeed(input, len, seed); +} + + +/* === XXH3 128-bit streaming === */ + +/* + * All the functions are actually the same as for 64-bit streaming variant. + * The only difference is the finalizatiom routine. + */ + +static void +XXH3_128bits_reset_internal(XXH3_state_t* statePtr, + XXH64_hash_t seed, + const void* secret, size_t secretSize) +{ + XXH3_64bits_reset_internal(statePtr, seed, secret, secretSize); +} + +XXH_PUBLIC_API XXH_errorcode +XXH3_128bits_reset(XXH3_state_t* statePtr) +{ + if (statePtr == NULL) return XXH_ERROR; + XXH3_128bits_reset_internal(statePtr, 0, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE); + return XXH_OK; +} + +XXH_PUBLIC_API XXH_errorcode +XXH3_128bits_reset_withSecret(XXH3_state_t* statePtr, const void* secret, size_t secretSize) +{ + if (statePtr == NULL) return XXH_ERROR; + XXH3_128bits_reset_internal(statePtr, 0, secret, secretSize); + if (secret == NULL) return XXH_ERROR; + if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR; + return XXH_OK; +} + +XXH_PUBLIC_API XXH_errorcode +XXH3_128bits_reset_withSeed(XXH3_state_t* statePtr, XXH64_hash_t seed) +{ + if (statePtr == NULL) return XXH_ERROR; + if (seed==0) return XXH3_128bits_reset(statePtr); + if (seed != statePtr->seed) XXH3_initCustomSecret(statePtr->customSecret, seed); + XXH3_128bits_reset_internal(statePtr, seed, NULL, XXH_SECRET_DEFAULT_SIZE); + return XXH_OK; +} + +XXH_PUBLIC_API XXH_errorcode +XXH3_128bits_update(XXH3_state_t* state, const void* input, size_t len) +{ + return XXH3_update(state, (const xxh_u8*)input, len, + XXH3_accumulate_512, XXH3_scrambleAcc); +} + +XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (const XXH3_state_t* state) +{ + const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret; + if (state->totalLen > XXH3_MIDSIZE_MAX) { + XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[XXH_ACC_NB]; + XXH3_digest_long(acc, state, secret); + XXH_ASSERT(state->secretLimit + XXH_STRIPE_LEN >= sizeof(acc) + XXH_SECRET_MERGEACCS_START); + { XXH128_hash_t h128; + h128.low64 = XXH3_mergeAccs(acc, + secret + XXH_SECRET_MERGEACCS_START, + (xxh_u64)state->totalLen * XXH_PRIME64_1); + h128.high64 = XXH3_mergeAccs(acc, + secret + state->secretLimit + XXH_STRIPE_LEN + - sizeof(acc) - XXH_SECRET_MERGEACCS_START, + ~((xxh_u64)state->totalLen * XXH_PRIME64_2)); + return h128; + } + } + /* len <= XXH3_MIDSIZE_MAX : short code */ + if (state->seed) + return XXH3_128bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed); + return XXH3_128bits_withSecret(state->buffer, (size_t)(state->totalLen), + secret, state->secretLimit + XXH_STRIPE_LEN); +} + +/* 128-bit utility functions */ + +#include <string.h> /* memcmp, memcpy */ + +/* return : 1 is equal, 0 if different */ +XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2) +{ + /* note : XXH128_hash_t is compact, it has no padding byte */ + return !(memcmp(&h1, &h2, sizeof(h1))); +} + +/* This prototype is compatible with stdlib's qsort(). + * return : >0 if *h128_1 > *h128_2 + * <0 if *h128_1 < *h128_2 + * =0 if *h128_1 == *h128_2 */ +XXH_PUBLIC_API int XXH128_cmp(const void* h128_1, const void* h128_2) +{ + XXH128_hash_t const h1 = *(const XXH128_hash_t*)h128_1; + XXH128_hash_t const h2 = *(const XXH128_hash_t*)h128_2; + int const hcmp = (h1.high64 > h2.high64) - (h2.high64 > h1.high64); + /* note : bets that, in most cases, hash values are different */ + if (hcmp) return hcmp; + return (h1.low64 > h2.low64) - (h2.low64 > h1.low64); +} + + +/*====== Canonical representation ======*/ +XXH_PUBLIC_API void +XXH128_canonicalFromHash(XXH128_canonical_t* dst, XXH128_hash_t hash) +{ + XXH_STATIC_ASSERT(sizeof(XXH128_canonical_t) == sizeof(XXH128_hash_t)); + if (XXH_CPU_LITTLE_ENDIAN) { + hash.high64 = XXH_swap64(hash.high64); + hash.low64 = XXH_swap64(hash.low64); + } + memcpy(dst, &hash.high64, sizeof(hash.high64)); + memcpy((char*)dst + sizeof(hash.high64), &hash.low64, sizeof(hash.low64)); +} + +XXH_PUBLIC_API XXH128_hash_t +XXH128_hashFromCanonical(const XXH128_canonical_t* src) +{ + XXH128_hash_t h; + h.high64 = XXH_readBE64(src); + h.low64 = XXH_readBE64(src->digest + 8); + return h; +} + +/* Pop our optimization override from above */ +#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \ + && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \ + && defined(__OPTIMIZE__) && !defined(__OPTIMIZE_SIZE__) /* respect -O0 and -Os */ +# pragma GCC pop_options +#endif #endif /* XXH_NO_LONG_LONG */ diff --git a/dviware/dvisvgm/src/DVIReader.cpp b/dviware/dvisvgm/src/DVIReader.cpp index 417efe9304..0581ad25c3 100644 --- a/dviware/dvisvgm/src/DVIReader.cpp +++ b/dviware/dvisvgm/src/DVIReader.cpp @@ -19,8 +19,6 @@ *************************************************************************/ #include <algorithm> -#include <cstdarg> -#include <fstream> #include <sstream> #include "Color.hpp" #include "DVIActions.hpp" @@ -28,6 +26,7 @@ #include "Font.hpp" #include "FontManager.hpp" #include "HashFunction.hpp" +#include "utility.hpp" #include "VectorStream.hpp" using namespace std; @@ -504,26 +503,64 @@ void DVIReader::cmdFontNum (int len) { } +/** Parses a sequence of font attributes given as key1=val1;key2=val2;...;keyn=valn */ +static map<string,string> parse_font_attribs (const string &str) { + map<string,string> attribs; + if (!str.empty()) { + for (const string &attr : util::split(str, ";")) { + vector<string> keyval = util::split(attr, "="); + if (keyval.size() == 2) + attribs[keyval[0]] = keyval[1]; + } + } + return attribs; +} + + /** Helper function to handle a font definition. * @param[in] fontnum local font number - * @param[in] name font name + * @param[in] name font name (or file path if enclosed in square brackets) * @param[in] cs checksum to be compared with TFM checksum - * @param[in] ds design size in PS point units - * @param[in] ss scaled size in PS point units */ -const Font* DVIReader::defineFont (uint32_t fontnum, const string &name, uint32_t cs, double ds, double ss) { + * @param[in] dsize design size of font in PS point units + * @param[in] ssize scaled size of font in PS point units */ +const Font* DVIReader::defineFont (uint32_t fontnum, const string &name, uint32_t cs, double dsize, double ssize) { FontManager &fm = FontManager::instance(); Font *font = fm.getFont(fontnum); - if (!font) { - int id = fm.registerFont(fontnum, name, cs, ds, ss); - font = fm.getFontById(id); - if (auto vf = dynamic_cast<VirtualFont*>(font)) { - // read vf file, register its font and character definitions - fm.enterVF(vf); - ifstream ifs(vf->path(), ios::binary); - VFReader vfReader(ifs); - vfReader.replaceActions(this); - vfReader.executeAll(); - fm.leaveVF(); + if (!font && !name.empty()) { // font not registered yet? + if (name[0] == '[') { // LuaTeX native font reference? + size_t last = name.rfind(']'); + if (last != string::npos) { + string path = name.substr(1, last-1); + FontStyle style; + int fontIndex=0; + if (name.size() > last && name[last+1] == ':') { // look for font attributes? + auto attribs = parse_font_attribs(name.substr(last+2)); + auto it = attribs.begin(); + if ((it = attribs.find("index")) != attribs.end()) + fontIndex = stoi(it->second); + if ((it = attribs.find("embolden")) != attribs.end()) + style.bold = stoi(it->second)/65536.0; + if ((it = attribs.find("extend")) != attribs.end()) + style.extend = stoi(it->second)/65536.0; + if ((it = attribs.find("slant")) != attribs.end()) + style.slant = stoi(it->second)/65536.0; + } + int id = fm.registerFont(fontnum, path, fontIndex, ssize, style, Color::BLACK); + font = fm.getFontById(id); + } + } + else { // TFM-based font specified by name + int id = fm.registerFont(fontnum, name, cs, dsize, ssize); + font = fm.getFontById(id); + if (auto vf = dynamic_cast<VirtualFont*>(font)) { + // read vf file, register its font and character definitions + fm.enterVF(vf); + ifstream ifs(vf->path(), ios::binary); + VFReader vfReader(ifs); + vfReader.replaceActions(this); + vfReader.executeAll(); + fm.leaveVF(); + } } } return font; diff --git a/dviware/dvisvgm/src/DVIReader.hpp b/dviware/dvisvgm/src/DVIReader.hpp index fc1424aca3..61c4866714 100644 --- a/dviware/dvisvgm/src/DVIReader.hpp +++ b/dviware/dvisvgm/src/DVIReader.hpp @@ -74,7 +74,7 @@ class DVIReader : public BasicDVIReader, public VFActions { virtual void moveDown (double dy, MoveMode mode); void putVFChar (Font *font, uint32_t c); double putGlyphArray (bool xonly, std::vector<double> &dx, std::vector<double> &dy, std::vector<uint16_t> &glyphs); - const Font* defineFont (uint32_t fontnum, const std::string &name, uint32_t cs, double ds, double ss); + const Font* defineFont (uint32_t fontnum, const std::string &name, uint32_t cs, double dsize, double ssize); void setFont (int num, SetFontMode mode); const DVIState& dviState() const {return _dviState;} double dvi2bp () const {return _dvi2bp;} diff --git a/dviware/dvisvgm/src/DVIToSVGActions.cpp b/dviware/dvisvgm/src/DVIToSVGActions.cpp index 6239be3b01..2464aec517 100644 --- a/dviware/dvisvgm/src/DVIToSVGActions.cpp +++ b/dviware/dvisvgm/src/DVIToSVGActions.cpp @@ -122,7 +122,7 @@ void DVIToSVGActions::setChar (double x, double y, unsigned c, bool vertical, co bbox.transform(getMatrix()); embed(bbox); #if 0 - XMLElement *rect = new XMLElement("rect"); + auto rect = util::make_unique<XMLElement>("rect"); rect->addAttribute("x", x-metrics.wl); rect->addAttribute("y", y-metrics.h); rect->addAttribute("width", metrics.wl+metrics.wr); @@ -130,26 +130,26 @@ void DVIToSVGActions::setChar (double x, double y, unsigned c, bool vertical, co rect->addAttribute("fill", "none"); rect->addAttribute("stroke", "red"); rect->addAttribute("stroke-width", "0.5"); - _svg.appendToPage(rect); + _svg.appendToPage(std::move(rect)); if (metrics.d > 0) { - XMLElement *line = new XMLElement("line"); + auto line = util::make_unique<XMLElement>("line"); line->addAttribute("x1", x-metrics.wl); line->addAttribute("y1", y); line->addAttribute("x2", x+metrics.wr); line->addAttribute("y2", y); line->addAttribute("stroke", "blue"); line->addAttribute("stroke-width", "0.5"); - _svg.appendToPage(line); + _svg.appendToPage(std::move(line)); } if (metrics.wl > 0) { - XMLElement *line = new XMLElement("line"); + auto line = util::make_unique<XMLElement>("line"); line->addAttribute("x1", x); line->addAttribute("y1", y-metrics.h); line->addAttribute("x2", x); line->addAttribute("y2", y+metrics.d); line->addAttribute("stroke", "blue"); line->addAttribute("stroke-width", "0.5"); - _svg.appendToPage(line); + _svg.appendToPage(std::move(line)); } #endif } @@ -227,7 +227,12 @@ void DVIToSVGActions::beginPage (unsigned pageno, const vector<int32_t>&) { /** This method is called when an "end of page (eop)" command was found in the DVI file. */ void DVIToSVGActions::endPage (unsigned pageno) { - SpecialManager::instance().notifyEndPage(pageno, *this); + try { + SpecialManager::instance().notifyEndPage(pageno, *this); + } + catch (const SpecialException &e) { + Message::estream(true) << "error in special: " << e.what() << '\n'; + } Matrix matrix = _dvireader->getPageTransformation(); _svg.transformPage(matrix); if (_bgcolor != Color::TRANSPARENT) { diff --git a/dviware/dvisvgm/src/DvisvgmSpecialHandler.cpp b/dviware/dvisvgm/src/DvisvgmSpecialHandler.cpp index 1f331659f0..97285db1a9 100644 --- a/dviware/dvisvgm/src/DvisvgmSpecialHandler.cpp +++ b/dviware/dvisvgm/src/DvisvgmSpecialHandler.cpp @@ -209,7 +209,9 @@ static void evaluate_expressions (string &str, SpecialActions &actions) { else { try { double val = calc.eval(expr); - str.replace(left, right-left+2, XMLString(val)); + XMLString valstr(val); + str.replace(left, right-left+2, valstr); + right = left+valstr.length()-1; } catch (CalculatorException &e) { throw SpecialException(string(e.what())+" in '{?("+expr+")}'"); @@ -413,8 +415,8 @@ void DvisvgmSpecialHandler::dviPreprocessingFinished () { void DvisvgmSpecialHandler::dviEndPage (unsigned, SpecialActions &actions) { - _defsParser.flush(actions); - _pageParser.flush(actions); + _defsParser.finish(actions); + _pageParser.finish(actions); actions.bbox().unlock(); for (auto &strvecpair : _macros) { StringVector &vec = strvecpair.second; @@ -437,8 +439,8 @@ vector<const char*> DvisvgmSpecialHandler::prefixes() const { /** Parses a fragment of XML code, creates corresponding XML nodes and adds them * to the SVG tree. The code may be split and processed by several calls of this * function. Incomplete chunks that can't be processed yet are stored and picked - * up again together with the next incoming XML fragment. If no further code should - * be appended, parameter 'finish' must be set. + * up again together with the next incoming XML fragment. If a call of this function + * is supposed to finish the parsing of an XML subtree, parameter 'finish' must be set. * @param[in] xml XML fragment to parse * @param[in] actions object providing the SVG tree functions * @param[in] finish if true, no more XML is expected and parsing is finished */ @@ -447,59 +449,65 @@ void DvisvgmSpecialHandler::XMLParser::parse (const string &xml, SpecialActions // incomplete tags are held back _xmlbuf += xml; size_t left=0, right; - while (left != string::npos) { - right = _xmlbuf.find('<', left); - if (left < right && left < _xmlbuf.length()) // plain text found? - (actions.svgTree().*_append)(util::make_unique<XMLText>(_xmlbuf.substr(left, right-left))); - if (right != string::npos) { - left = right; - if (_xmlbuf.compare(left, 9, "<![CDATA[") == 0) { - right = _xmlbuf.find("]]>", left+9); - if (right == string::npos) { - if (finish) throw SpecialException("expected ']]>' at end of CDATA block"); - break; + try { + while (left != string::npos) { + right = _xmlbuf.find('<', left); + if (left < right && left < _xmlbuf.length()) // plain text found? + (actions.svgTree().*_append)(util::make_unique<XMLText>(_xmlbuf.substr(left, right-left))); + if (right != string::npos) { + left = right; + if (_xmlbuf.compare(left, 9, "<![CDATA[") == 0) { + right = _xmlbuf.find("]]>", left+9); + if (right == string::npos) { + if (finish) throw SpecialException("expected ']]>' at end of CDATA block"); + break; + } + (actions.svgTree().*_append)(util::make_unique<XMLCData>(_xmlbuf.substr(left+9, right-left-9))); + right += 2; } - (actions.svgTree().*_append)(util::make_unique<XMLCData>(_xmlbuf.substr(left+9, right-left-9))); - right += 2; - } - else if (_xmlbuf.compare(left, 4, "<!--") == 0) { - right = _xmlbuf.find("-->", left+4); - if (right == string::npos) { - if (finish) throw SpecialException("expected '-->' at end of comment"); - break; + else if (_xmlbuf.compare(left, 4, "<!--") == 0) { + right = _xmlbuf.find("-->", left+4); + if (right == string::npos) { + if (finish) throw SpecialException("expected '-->' at end of comment"); + break; + } + (actions.svgTree().*_append)(util::make_unique<XMLComment>(_xmlbuf.substr(left+4, right-left-4))); + right += 2; } - (actions.svgTree().*_append)(util::make_unique<XMLComment>(_xmlbuf.substr(left+4, right-left-4))); - right += 2; - } - else if (_xmlbuf.compare(left, 2, "<?") == 0) { - right = _xmlbuf.find("?>", left+2); - if (right == string::npos) { - if (finish) throw SpecialException("expected '?>' at end of processing instruction"); - break; + else if (_xmlbuf.compare(left, 2, "<?") == 0) { + right = _xmlbuf.find("?>", left+2); + if (right == string::npos) { + if (finish) throw SpecialException("expected '?>' at end of processing instruction"); + break; + } + (actions.svgTree().*_append)(util::make_unique<XMLText>(_xmlbuf.substr(left, right-left+2))); + right++; } - (actions.svgTree().*_append)(util::make_unique<XMLText>(_xmlbuf.substr(left, right-left+2))); - right++; - } - else if (_xmlbuf.compare(left, 2, "</") == 0) { - right = _xmlbuf.find('>', left+2); - if (right == string::npos) { - if (finish) throw SpecialException("missing '>' at end of closing XML tag"); - break; + else if (_xmlbuf.compare(left, 2, "</") == 0) { + right = _xmlbuf.find('>', left+2); + if (right == string::npos) { + if (finish) throw SpecialException("missing '>' at end of closing XML tag"); + break; + } + closeElement(_xmlbuf.substr(left+2, right-left-2), actions); } - closeElement(_xmlbuf.substr(left+2, right-left-2), actions); - } - else { - right = _xmlbuf.find('>', left+1); - if (right == string::npos) { - if (finish) throw SpecialException("missing '>' or '/>' at end of opening XML tag"); - break; + else { + right = _xmlbuf.find('>', left+1); + if (right == string::npos) { + if (finish) throw SpecialException("missing '>' or '/>' at end of opening XML tag"); + break; + } + openElement(_xmlbuf.substr(left+1, right-left-1), actions); } - openElement(_xmlbuf.substr(left+1, right-left-1), actions); } + left = right; + if (right != string::npos) + left++; } - left = right; - if (right != string::npos) - left++; + } + catch (const SpecialException &e) { + _error = true; + throw; } if (left == string::npos) _xmlbuf.clear(); @@ -545,7 +553,7 @@ void DvisvgmSpecialHandler::XMLParser::closeElement (const string &tag, SpecialA if (_nameStack.empty()) throw SpecialException("spurious closing tag </" + name + ">"); if (_nameStack.back() != name) - throw SpecialException("expected </" + name + "> but found </" + _nameStack.back() + ">"); + throw SpecialException("expected </" + _nameStack.back() + "> but found </" + name + ">"); (actions.svgTree().*_popContext)(); _nameStack.pop_back(); } @@ -553,9 +561,10 @@ void DvisvgmSpecialHandler::XMLParser::closeElement (const string &tag, SpecialA /** Processes any remaining XML fragments, checks for missing closing tags, * and resets the parser state. */ -void DvisvgmSpecialHandler::XMLParser::flush (SpecialActions &actions) { +void DvisvgmSpecialHandler::XMLParser::finish (SpecialActions &actions) { if (!_xmlbuf.empty()) { - parse("", actions, true); + if (!_error) + parse("", actions, true); _xmlbuf.clear(); } string tags; @@ -563,8 +572,8 @@ void DvisvgmSpecialHandler::XMLParser::flush (SpecialActions &actions) { tags += "</"+_nameStack.back()+">, "; _nameStack.pop_back(); } - if (!tags.empty()) { + if (!tags.empty() && !_error) { tags.resize(tags.length()-2); - throw SpecialException("missing closing tags: "+tags); + throw SpecialException("missing closing tag(s): "+tags); } } diff --git a/dviware/dvisvgm/src/DvisvgmSpecialHandler.hpp b/dviware/dvisvgm/src/DvisvgmSpecialHandler.hpp index 0f73c73ecd..d6aa32c9ee 100644 --- a/dviware/dvisvgm/src/DvisvgmSpecialHandler.hpp +++ b/dviware/dvisvgm/src/DvisvgmSpecialHandler.hpp @@ -53,7 +53,7 @@ class DvisvgmSpecialHandler : public SpecialHandler { : _append(append), _pushContext(push), _popContext(pop) {} void parse (const std::string &xml, SpecialActions &actions, bool finish=false); - void flush (SpecialActions &actions); + void finish (SpecialActions &actions); protected: void openElement (const std::string &tag, SpecialActions &actions); @@ -65,6 +65,7 @@ class DvisvgmSpecialHandler : public SpecialHandler { PopFunc _popContext; std::string _xmlbuf; NameStack _nameStack; ///< names of nested elements still missing a closing tag + bool _error=false; }; using StringVector = std::vector<std::string>; diff --git a/dviware/dvisvgm/src/FileSystem.cpp b/dviware/dvisvgm/src/FileSystem.cpp index a1fafbdece..e1db59fa03 100644 --- a/dviware/dvisvgm/src/FileSystem.cpp +++ b/dviware/dvisvgm/src/FileSystem.cpp @@ -20,12 +20,14 @@ #include <config.h> #include <algorithm> +#include <chrono> #include <cstdlib> #include <cstring> #include <fstream> #include "FileSystem.hpp" #include "utility.hpp" #include "version.hpp" +#include "XXHashFunction.hpp" #ifdef HAVE_UNISTD_H #include <unistd.h> @@ -51,7 +53,7 @@ using namespace std; string FileSystem::TMPDIR; -const char *FileSystem::TMPSUBDIR = nullptr; +FileSystem::TemporaryDirectory FileSystem::_tmpdir; /** Private wrapper function for mkdir: creates a single folder. @@ -79,13 +81,6 @@ static bool inline s_rmdir (const string &dirname) { } -FileSystem::~FileSystem () { - // remove the subdirectory from the system's temp folder (if empty) - if (TMPSUBDIR) - s_rmdir(tmpdir()); -} - - bool FileSystem::remove (const string &fname) { return unlink(fname.c_str()) == 0; } @@ -192,36 +187,29 @@ const char* FileSystem::userdir () { /** Returns the path of the temporary folder. */ string FileSystem::tmpdir () { - string ret; - if (!TMPDIR.empty()) - ret = TMPDIR; - else { + if (_tmpdir.path().empty()) { + string basedir; + if (!TMPDIR.empty()) + basedir = TMPDIR; + else { #ifdef _WIN32 - char buf[MAX_PATH]; - if (GetTempPath(MAX_PATH, buf)) - ret = ensureForwardSlashes(buf); - else - ret = "."; + char buf[MAX_PATH]; + if (GetTempPath(MAX_PATH, buf)) + basedir = ensureForwardSlashes(buf); + else + basedir = "."; #else - if (const char *path = getenv("TMPDIR")) - ret = path; - else - ret = "/tmp"; + if (const char *path = getenv("TMPDIR")) + basedir = path; + else + basedir = "/tmp"; #endif - if (ret.back() == '/') - ret.pop_back(); - static bool initialized=false; - if (!initialized && ret != ".") { - TMPSUBDIR = PROGRAM_NAME; - s_mkdir(ret + "/" + TMPSUBDIR); - initialized = true; + if (basedir.back() == '/') + basedir.pop_back(); } - if (TMPSUBDIR) - ret += string("/") + TMPSUBDIR; + _tmpdir = TemporaryDirectory(basedir, PROGRAM_NAME); } - if (!ret.empty() && ret.back() != '/') - ret += '/'; - return ret; + return _tmpdir.path(); } @@ -368,3 +356,33 @@ int FileSystem::collect (const std::string &dirname, vector<string> &entries) { #endif return entries.size(); } + + +/** Creates a temporary directory in a given folder. + * @param[in] folder folder path in which the directory is to be created + * @param[in] prefix initial string of the directory name */ +FileSystem::TemporaryDirectory::TemporaryDirectory (const std::string &folder, string prefix) { + using namespace std::chrono; + auto now = system_clock::now().time_since_epoch(); + auto now_ms = duration_cast<milliseconds>(now).count(); + auto hash = XXH64HashFunction(to_string(now_ms)).digestValue(); + if (!prefix.empty() && prefix.back() != '-') + prefix += "-"; + for (int i=0; i < 10 && _path.empty(); i++) { + hash++; + stringstream oss; + oss << folder << '/' << prefix << hex << hash; + if (exists(oss.str())) + continue; + if (s_mkdir(oss.str())) + _path = oss.str() + "/"; + else + break; + } +} + + +FileSystem::TemporaryDirectory::~TemporaryDirectory () { + if (!_path.empty()) + s_rmdir(_path); +} diff --git a/dviware/dvisvgm/src/FileSystem.hpp b/dviware/dvisvgm/src/FileSystem.hpp index f82ff1e870..9a45c4ac19 100644 --- a/dviware/dvisvgm/src/FileSystem.hpp +++ b/dviware/dvisvgm/src/FileSystem.hpp @@ -25,8 +25,23 @@ #include <vector> class FileSystem { + class TemporaryDirectory { + friend class FileSystem; + public: + TemporaryDirectory (const std::string &folder, std::string prefix); + TemporaryDirectory (TemporaryDirectory &&tmpdir) =default; + ~TemporaryDirectory (); + TemporaryDirectory& operator = (TemporaryDirectory &&tmpdir) =default; + const std::string& path () const {return _path;} + + protected: + TemporaryDirectory () =default; + + private: + std::string _path; + }; + public: - ~FileSystem (); static bool remove (const std::string &fname); static bool rename (const std::string &oldname, const std::string &newname); static bool copy (const std::string &src, const std::string &dest, bool remove_src=false); @@ -48,8 +63,9 @@ class FileSystem { protected: FileSystem () =default; - bool system_tmpdir_available (); - static const char* TMPSUBDIR; ///< subdirectory of the system's temporary folder + + private: + static TemporaryDirectory _tmpdir; }; #endif diff --git a/dviware/dvisvgm/src/Font.cpp b/dviware/dvisvgm/src/Font.cpp index e40ad8a392..8cbcb1db47 100644 --- a/dviware/dvisvgm/src/Font.cpp +++ b/dviware/dvisvgm/src/Font.cpp @@ -607,8 +607,13 @@ PhysicalFont::Type NativeFont::type () const { double NativeFont::charWidth (int c) const { FontEngine::instance().setFont(*this); int upem = FontEngine::instance().getUnitsPerEM(); - double w = upem ? (scaledSize()*FontEngine::instance().getAdvance(c)/upem*_style.extend) : 0; - w += abs(_style.slant*charHeight(c)); + return upem ? (scaledSize()*FontEngine::instance().getAdvance(c)/upem*_style.extend) : 0; +} + + +double NativeFont::italicCorr(int c) const { + double w = abs(_style.slant*charHeight(c)); // slant := tan(phi) = dx/height + w *= _style.extend; return w; } diff --git a/dviware/dvisvgm/src/Font.hpp b/dviware/dvisvgm/src/Font.hpp index 59556e7a81..0a92cbeff5 100644 --- a/dviware/dvisvgm/src/Font.hpp +++ b/dviware/dvisvgm/src/Font.hpp @@ -266,7 +266,7 @@ class NativeFont : public PhysicalFont { double charWidth (int c) const override; double charDepth (int c) const override; double charHeight (int c) const override; - double italicCorr (int c) const override {return 0;} + double italicCorr (int c) const override; const FontMetrics* getMetrics () const override {return nullptr;} const FontStyle* style () const override {return &_style;} Color color () const override {return _color;} diff --git a/dviware/dvisvgm/src/FontCache.cpp b/dviware/dvisvgm/src/FontCache.cpp index e024558913..1fd13c9c11 100644 --- a/dviware/dvisvgm/src/FontCache.cpp +++ b/dviware/dvisvgm/src/FontCache.cpp @@ -160,7 +160,7 @@ bool FontCache::write (const string &fontname, ostream &os) const { XXH32HashFunction hashfunc; sw.writeUnsigned(FORMAT_VERSION, 1, hashfunc); - sw.writeBytes(hashfunc.digestValue()); // space for checksum + sw.writeBytes(hashfunc.digestBytes()); // space for checksum sw.writeString(fontname, hashfunc, true); sw.writeUnsigned(_glyphs.size(), 4, hashfunc); WriteActions actions(sw, hashfunc); @@ -171,7 +171,7 @@ bool FontCache::write (const string &fontname, ostream &os) const { glyph.iterate(actions, false); } os.seekp(1); - auto digest = hashfunc.digestValue(); + auto digest = hashfunc.digestBytes(); sw.writeBytes(digest); // insert checksum os.seekp(0, ios::end); return true; @@ -215,7 +215,7 @@ bool FontCache::read (const string &fontname, istream &is) { auto hashcmp = sr.readBytes(hashfunc.digestSize()); hashfunc.update(is); - if (hashfunc.digestValue() != hashcmp) + if (hashfunc.digestBytes() != hashcmp) return false; is.clear(); @@ -309,7 +309,7 @@ bool FontCache::fontinfo (std::istream &is, FontInfo &info) { info.checksum = sr.readBytes(hashfunc.digestSize()); hashfunc.update(is); - if (hashfunc.digestValue() != info.checksum) + if (hashfunc.digestBytes() != info.checksum) return false; is.clear(); diff --git a/dviware/dvisvgm/src/FontManager.cpp b/dviware/dvisvgm/src/FontManager.cpp index 58d2dfc261..eabb4a3d30 100644 --- a/dviware/dvisvgm/src/FontManager.cpp +++ b/dviware/dvisvgm/src/FontManager.cpp @@ -255,8 +255,6 @@ int FontManager::registerFont (uint32_t fontnum, string filename, int fontIndex, if (id >= 0) return id; - if (!filename.empty() && filename[0] == '[' && filename[filename.size()-1] == ']') - filename = filename.substr(1, filename.size()-2); string fontname = NativeFont::uniqueName(filename, style); const char *path = filename.c_str(); unique_ptr<Font> newfont; diff --git a/dviware/dvisvgm/src/GraphicsPath.hpp b/dviware/dvisvgm/src/GraphicsPath.hpp index 565251c0d0..64c112ad0b 100644 --- a/dviware/dvisvgm/src/GraphicsPath.hpp +++ b/dviware/dvisvgm/src/GraphicsPath.hpp @@ -732,7 +732,7 @@ class GraphicsPath { private: std::deque<CommandVariant> _commands; ///< sequence of path commands - WindingRule _windingRule; + WindingRule _windingRule = WindingRule::NON_ZERO; Point _startPoint; ///< start point of final sub-path Point _finalPoint; ///< final point reached by last command in path }; diff --git a/dviware/dvisvgm/src/HashFunction.cpp b/dviware/dvisvgm/src/HashFunction.cpp index 7f99dacaf0..16661fb3d6 100644 --- a/dviware/dvisvgm/src/HashFunction.cpp +++ b/dviware/dvisvgm/src/HashFunction.cpp @@ -96,7 +96,7 @@ void HashFunction::update (istream &is) { string HashFunction::digestString () const { ostringstream oss; oss << hex << setfill('0'); - for (int byte : digestValue()) + for (int byte : digestBytes()) oss << setw(2) << byte; return oss.str(); } diff --git a/dviware/dvisvgm/src/HashFunction.hpp b/dviware/dvisvgm/src/HashFunction.hpp index b13ee39a50..e31040075e 100644 --- a/dviware/dvisvgm/src/HashFunction.hpp +++ b/dviware/dvisvgm/src/HashFunction.hpp @@ -35,7 +35,7 @@ class HashFunction { virtual void update (const char *data, size_t length) =0; virtual void update (const std::string &data) =0; virtual void update (const std::vector<uint8_t> &data) =0; - virtual std::vector<uint8_t> digestValue () const =0; + virtual std::vector<uint8_t> digestBytes () const =0; void update (std::istream &is); std::string digestString () const; static std::vector<std::string> supportedAlgorithms (); diff --git a/dviware/dvisvgm/src/MD5HashFunction.hpp b/dviware/dvisvgm/src/MD5HashFunction.hpp index f0eda874f2..27539576ed 100644 --- a/dviware/dvisvgm/src/MD5HashFunction.hpp +++ b/dviware/dvisvgm/src/MD5HashFunction.hpp @@ -42,7 +42,7 @@ class MD5HashFunction : public HashFunction { void update (const std::string &data) override {update(data.data(), data.length());} void update (const std::vector<uint8_t> &data) override {update(reinterpret_cast<const char*>(data.data()), data.size());} - std::vector<uint8_t> digestValue () const override { + std::vector<uint8_t> digestBytes () const override { std::vector<uint8_t> hash(16); MD5_CTX savedContext = _context; MD5_Final(hash.data(), &_context); // also erases the context structure diff --git a/dviware/dvisvgm/src/PathClipper.cpp b/dviware/dvisvgm/src/PathClipper.cpp index 4aa4373f46..f1cca09fd6 100644 --- a/dviware/dvisvgm/src/PathClipper.cpp +++ b/dviware/dvisvgm/src/PathClipper.cpp @@ -310,22 +310,37 @@ inline PolyFillType polyFillType (CurvedPath::WindingRule wr) { } -/** Computes the intersection of to curved paths. +/** Combines two curved paths by applying a boolean operation on them. + * @param[in] op operation to perform * @param[in] p1 first curved path * @param[in] p2 second curved path - * @param[out] result intersection of p1 and p2 */ -void PathClipper::intersect (const CurvedPath &p1, const CurvedPath &p2, CurvedPath &result) { - if (p1.size() < 2 || p2.size() < 2) - return; - Clipper clipper; - Polygons polygons; - flatten(p1, polygons); - clipper.AddPaths(polygons, ptSubject, true); - polygons.clear(); - flatten(p2, polygons); - clipper.AddPaths(polygons, ptClip, true); - clipper.ZFillFunction(callback); - Polygons flattenedPath; - clipper.Execute(ctIntersection, flattenedPath, polyFillType(p1.windingRule()), polyFillType(p2.windingRule())); - reconstruct(flattenedPath, result); + * @return intersection of p1 and p2 */ +CurvedPath PathClipper::combine (ClipType op, const CurvedPath &p1, const CurvedPath &p2) { + CurvedPath result; + if (p1.size() > 1 && p2.size() > 1) { + Clipper clipper; + Polygons polygons; + flatten(p1, polygons); + clipper.AddPaths(polygons, ptSubject, true); + polygons.clear(); + flatten(p2, polygons); + clipper.AddPaths(polygons, ptClip, true); + clipper.ZFillFunction(callback); + Polygons flattenedPath; + clipper.Execute(op, flattenedPath, polyFillType(p1.windingRule()), polyFillType(p2.windingRule())); + reconstruct(flattenedPath, result); + } + return result; +} + + +/** Returns the intersection of two curved paths. */ +CurvedPath PathClipper::intersect (const CurvedPath &p1, const CurvedPath &p2) { + return combine(ctIntersection, p1, p2); +} + + +/** Returns the union of two curved paths. */ +CurvedPath PathClipper::unite (const CurvedPath &p1, const CurvedPath &p2) { + return combine(ctUnion, p1, p2); } diff --git a/dviware/dvisvgm/src/PathClipper.hpp b/dviware/dvisvgm/src/PathClipper.hpp index e3e6bd1681..7c42f5de6d 100644 --- a/dviware/dvisvgm/src/PathClipper.hpp +++ b/dviware/dvisvgm/src/PathClipper.hpp @@ -36,9 +36,11 @@ class PathClipper { using CurvedPath = GraphicsPath<double>; public: - void intersect (const CurvedPath &p1, const CurvedPath &p2, CurvedPath &result); + CurvedPath intersect (const CurvedPath &p1, const CurvedPath &p2); + CurvedPath unite (const CurvedPath &p1, const CurvedPath &p2); protected: + CurvedPath combine (ClipperLib::ClipType op, const CurvedPath &p1, const CurvedPath &p2); void flatten (const CurvedPath &gp, ClipperLib::Paths &polygons); void reconstruct (const ClipperLib::Path &polygon, CurvedPath &path); void reconstruct (const ClipperLib::Paths &polygons, CurvedPath &path); diff --git a/dviware/dvisvgm/src/PsSpecialHandler.cpp b/dviware/dvisvgm/src/PsSpecialHandler.cpp index 2301eeb571..9641556dd5 100644 --- a/dviware/dvisvgm/src/PsSpecialHandler.cpp +++ b/dviware/dvisvgm/src/PsSpecialHandler.cpp @@ -941,7 +941,7 @@ void PsSpecialHandler::clip (Path path, bool evenodd) { if (!_actions->getMatrix().isIdentity()) path.transform(_actions->getMatrix()); if (_clipStack.prependedPath()) - path.prepend(*_clipStack.prependedPath()); + path = PathClipper().unite(*_clipStack.prependedPath(), path); int oldID = _clipStack.topID(); @@ -954,9 +954,7 @@ void PsSpecialHandler::clip (Path path, bool evenodd) { else { // compute the intersection of the current clipping path with the current graphics path const Path *oldPath = _clipStack.path(); - Path intersectedPath(windingRule); - PathClipper clipper; - clipper.intersect(*oldPath, path, intersectedPath); + Path intersectedPath = PathClipper().intersect(*oldPath, path); pathReplaced = _clipStack.replace(intersectedPath); intersectedPath.writeSVG(oss, SVGTree::RELATIVE_PATH_CMDS); } diff --git a/dviware/dvisvgm/src/XMLNode.cpp b/dviware/dvisvgm/src/XMLNode.cpp index db07d95dd2..363115fbca 100644 --- a/dviware/dvisvgm/src/XMLNode.cpp +++ b/dviware/dvisvgm/src/XMLNode.cpp @@ -259,7 +259,7 @@ XMLElement* XMLElement::wrap (XMLNode *first, XMLNode *last, const string &name) XMLNode *child = first; while (child && child != last) { XMLNode *next = child->next(); - wrapper->insertLast(remove(child)); + wrapper->insertLast(detach(child)); child = next; } XMLElement *ret = wrapper.get(); @@ -277,44 +277,41 @@ XMLElement* XMLElement::wrap (XMLNode *first, XMLNode *last, const string &name) * Example: unwrap a child element b of a: * <a>text1<b><c/>text2<d/></b></a> => <a>text1<c/>text2<d/></a> * @param[in] child child element to unwrap - * @return raw pointer to the first node C1 of the unwrapped sequence */ -XMLNode* XMLElement::unwrap (XMLElement *child) { - if (!child || !child->parent()) + * @return raw pointer to the first node C1 of the unwrapped sequence or nullptr if element was empty */ +XMLNode* XMLElement::unwrap (XMLElement *element) { + if (!element || !element->parent()) return nullptr; - XMLElement *parent = child->parent()->toElement(); - auto removedChild = remove(child); - if (child->empty()) - return child->next(); - XMLNode *firstGrandchild = child->firstChild(); - XMLNode *prev = child->prev(); - unique_ptr<XMLNode> grandchild = std::move(child->_firstChild); - while (grandchild) { - prev = parent->insertAfter(std::move(grandchild), prev); - grandchild = std::move(prev->_next); - } - return firstGrandchild; -} - - -/** Removes a child node from the element. - * @param[in] child pointer to child to remove - * @return pointer to removed child or nullptr if given child doesn't belong to this element */ -unique_ptr<XMLNode> XMLElement::remove (XMLNode *child) { - unique_ptr<XMLNode> node; - if (child && child->parent()) { - XMLElement *parent = child->parent()->toElement(); - if (child == parent->_lastChild) - parent->_lastChild = child->prev(); - if (child != parent->firstChild()) - node = child->prev()->removeNext(); + XMLElement *parent = element->parent()->toElement(); + XMLNode *prev = element->prev(); + auto unlinkedElement = util::static_unique_ptr_cast<XMLElement>(detach(element)); + if (unlinkedElement->empty()) + return nullptr; + XMLNode *firstChild = unlinkedElement->firstChild(); + while (auto child = detach(unlinkedElement->firstChild())) + prev = parent->insertAfter(std::move(child), prev); + return firstChild; +} + + +/** Isolates a node and its descendants from a subtree. + * @param[in] node raw pointer to node to be detached + * @return unique pointer to the detached node. */ +unique_ptr<XMLNode> XMLElement::detach (XMLNode *node) { + unique_ptr<XMLNode> uniqueNode; + if (node && node->parent()) { + XMLElement *parent = node->parent()->toElement(); + if (node == parent->_lastChild) + parent->_lastChild = node->prev(); + if (node != parent->firstChild()) + uniqueNode = node->prev()->removeNext(); else { - node = std::move(parent->_firstChild); - if ((parent->_firstChild = std::move(node->_next))) + uniqueNode = std::move(parent->_firstChild); + if ((parent->_firstChild = std::move(uniqueNode->_next))) parent->_firstChild->prev(nullptr); } - child->parent(nullptr); + node->parent(nullptr); } - return node; + return uniqueNode; } diff --git a/dviware/dvisvgm/src/XMLNode.hpp b/dviware/dvisvgm/src/XMLNode.hpp index 534ae87bd7..65bd69fab8 100644 --- a/dviware/dvisvgm/src/XMLNode.hpp +++ b/dviware/dvisvgm/src/XMLNode.hpp @@ -157,7 +157,7 @@ class XMLElement : public XMLNode { const XMLElement* toElement () const override {return this;} const Attribute* getAttribute (const std::string &name) const; - static std::unique_ptr<XMLNode> remove (XMLNode *child); + static std::unique_ptr<XMLNode> detach (XMLNode *node); static XMLElement* wrap (XMLNode *first, XMLNode *last, const std::string &name); static XMLNode* unwrap (XMLElement *child); diff --git a/dviware/dvisvgm/src/XXHashFunction.hpp b/dviware/dvisvgm/src/XXHashFunction.hpp index e44083f182..d26a4e4984 100644 --- a/dviware/dvisvgm/src/XXHashFunction.hpp +++ b/dviware/dvisvgm/src/XXHashFunction.hpp @@ -36,6 +36,7 @@ struct XXHInterface { template<> struct XXHInterface<4> { using State = XXH32_state_t; + using Digest = XXH32_hash_t; static constexpr auto createState = &XXH32_createState; static constexpr auto freeState = &XXH32_freeState; static constexpr auto reset = &XXH32_reset; @@ -46,6 +47,7 @@ struct XXHInterface<4> { template<> struct XXHInterface<8> { using State = XXH64_state_t; + using Digest = XXH64_hash_t; static constexpr auto createState = &XXH64_createState; static constexpr auto freeState = &XXH64_freeState; static constexpr auto reset = &XXH64_reset; @@ -57,6 +59,7 @@ struct XXHInterface<8> { template<> struct XXHInterface<16> { using State = XXH3_state_t; + using Digest = XXH128_hash_t; static constexpr auto createState = &XXH3_createState; static constexpr auto freeState = &XXH3_freeState; static constexpr auto reset = &XXH3_128bits_reset_withSeed; @@ -71,7 +74,7 @@ class XXHashFunction : public HashFunction { using Interface = XXHInterface<HASH_BYTES>; public: XXHashFunction () : _state(Interface::createState()) {Interface::reset(_state, 0);} - XXHashFunction(const char *data, size_t length) : XXHashFunction() {update(data, length);} + XXHashFunction (const char *data, size_t length) : XXHashFunction() {update(data, length);} explicit XXHashFunction(const std::string &data) : XXHashFunction() {update(data);} explicit XXHashFunction(const std::vector<uint8_t> &data) : XXHashFunction() {update(data);} ~XXHashFunction () override {Interface::freeState(_state);} @@ -83,10 +86,11 @@ class XXHashFunction : public HashFunction { using HashFunction::update; // unhide update(istream &is) defined in base class - std::vector<uint8_t> digestValue () const override { + std::vector<uint8_t> digestBytes () const override { return util::bytes(Interface::digest(_state), HASH_BYTES); } + typename Interface::Digest digestValue () const {return Interface::digest(_state);} static unsigned version () {return XXH_versionNumber();} private: @@ -100,7 +104,7 @@ using XXH64HashFunction = XXHashFunction<8>; using XXH128HashFunction = XXHashFunction<16>; template<> -inline std::vector<uint8_t> XXHashFunction<16>::digestValue () const { +inline std::vector<uint8_t> XXHashFunction<16>::digestBytes () const { std::vector<uint8_t> hash; auto digest = Interface::digest(_state); for (auto chunk : {digest.high64, digest.low64}) { diff --git a/dviware/dvisvgm/src/dvisvgm.cpp b/dviware/dvisvgm/src/dvisvgm.cpp index 74a8e83abc..803e34e717 100644 --- a/dviware/dvisvgm/src/dvisvgm.cpp +++ b/dviware/dvisvgm/src/dvisvgm.cpp @@ -480,16 +480,20 @@ int main (int argc, char *argv[]) { } catch (DVIException &e) { Message::estream() << "\nDVI error: " << e.what() << '\n'; + return -1; } catch (PSException &e) { Message::estream() << "\nPostScript error: " << e.what() << '\n'; + return -2; } catch (SignalException &e) { Message::wstream().clearline(); Message::wstream(true) << "execution interrupted by user\n"; + return -3; } catch (exception &e) { Message::estream(true) << e.what() << '\n'; + return -4; } return 0; } diff --git a/dviware/dvisvgm/src/optimizer/GroupCollapser.cpp b/dviware/dvisvgm/src/optimizer/GroupCollapser.cpp index b8a32786c1..c3060be788 100644 --- a/dviware/dvisvgm/src/optimizer/GroupCollapser.cpp +++ b/dviware/dvisvgm/src/optimizer/GroupCollapser.cpp @@ -61,7 +61,7 @@ static void remove_ws_nodes (XMLElement *elem) { node = node->next(); else { XMLNode *next = node->next(); - XMLElement::remove(node); + XMLElement::detach(node); node = next; } } @@ -69,29 +69,31 @@ static void remove_ws_nodes (XMLElement *elem) { /** Recursively removes all redundant group elements from the given context element - * and moves their attributes to the corresponding parent element. + * and moves their attributes to the corresponding parent elements. * @param[in] context root of the subtree to process */ void GroupCollapser::execute (XMLElement *context) { if (!context) return; - XMLNode *node=context->firstChild(); - while (node) { - XMLNode *next = node->next(); // keep safe pointer to next node - if (XMLElement *elem = node->toElement()) - execute(elem); - node = next; - } - if (context->name() == "g" && context->attributes().empty()) { - // unwrap group without attributes - remove_ws_nodes(context); - XMLElement::unwrap(context); + + XMLNode *child=context->firstChild(); + while (child) { + XMLNode *next=child->next(); + if (XMLElement *childElement = child->toElement()) { + execute(childElement); + // check for groups without attributes and remove them + if (childElement->name() == "g" && childElement->attributes().empty()) { + remove_ws_nodes(childElement); + if (XMLNode *firstUnwrappedNode = XMLElement::unwrap(childElement)) + next = firstUnwrappedNode; + } + } + child = next; } - else { - XMLElement *child = only_child_element(context); - if (child && collapsible(*context)) { - if (child->name() == "g" && unwrappable(*child, *context) && moveAttributes(*child, *context)) { + if (XMLElement *childElement = only_child_element(context)) { + if (collapsible(*context)) { + if (childElement->name() == "g" && unwrappable(*childElement, *context) && moveAttributes(*childElement, *context)) { remove_ws_nodes(context); - XMLElement::unwrap(child); + XMLElement::unwrap(childElement); } } } @@ -147,12 +149,15 @@ bool GroupCollapser::collapsible (const XMLElement &element) { * @param[in] source element whose children and attributes should be moved * @param[in] dest element that should receive the children and attributes */ bool GroupCollapser::unwrappable (const XMLElement &source, const XMLElement &dest) { - // check for colliding clip-path attributes - if (const char *cp1 = source.getAttributeValue("clip-path")) { - if (const char *cp2 = dest.getAttributeValue("clip-path")) { - if (string(cp1) != cp2) - return false; - } + const char *cp1 = source.getAttributeValue("clip-path"); + const char *cp2 = dest.getAttributeValue("clip-path"); + if (cp2) { + // check for colliding clip-path attributes + if (cp1 && string(cp1) != string(cp2)) + return false; + // don't apply inner transformations to outer clipping paths + if (source.hasAttribute("transform")) + return false; } // these attributes prevent a group from being unwrapped static const char *attribs[] = { diff --git a/dviware/dvisvgm/src/optimizer/RedundantElementRemover.cpp b/dviware/dvisvgm/src/optimizer/RedundantElementRemover.cpp index 906ccc2d14..a8aca1a993 100644 --- a/dviware/dvisvgm/src/optimizer/RedundantElementRemover.cpp +++ b/dviware/dvisvgm/src/optimizer/RedundantElementRemover.cpp @@ -64,6 +64,6 @@ void RedundantElementRemover::execute (XMLElement *defs, XMLElement *context) { descendants.clear(); for (const string &str : idTree.getKeys()) { XMLElement *node = defs->getFirstDescendant("clipPath", "id", str.c_str()); - XMLElement::remove(node); + XMLElement::detach(node); } } diff --git a/dviware/dvisvgm/src/optimizer/SVGOptimizer.cpp b/dviware/dvisvgm/src/optimizer/SVGOptimizer.cpp index b2d45176e7..ac4a78fe20 100644 --- a/dviware/dvisvgm/src/optimizer/SVGOptimizer.cpp +++ b/dviware/dvisvgm/src/optimizer/SVGOptimizer.cpp @@ -39,9 +39,9 @@ SVGOptimizer::SVGOptimizer (SVGTree *svg) : _svg(svg) { // optimizer modules available to the user; must be listed in default order // _moduleEntries.emplace_back(ModuleEntry("remove-ws", util::make_unique<WSNodeRemover>())); _moduleEntries.emplace_back(ModuleEntry("simplify-text", util::make_unique<TextSimplifier>())); + _moduleEntries.emplace_back(ModuleEntry("simplify-transform", util::make_unique<TransformSimplifier>())); _moduleEntries.emplace_back(ModuleEntry("group-attributes", util::make_unique<AttributeExtractor>())); _moduleEntries.emplace_back(ModuleEntry("collapse-groups", util::make_unique<GroupCollapser>())); - _moduleEntries.emplace_back(ModuleEntry("simplify-transform", util::make_unique<TransformSimplifier>())); _moduleEntries.emplace_back(ModuleEntry("remove-clippath", util::make_unique<RedundantElementRemover>())); } diff --git a/dviware/dvisvgm/src/optimizer/TransformSimplifier.cpp b/dviware/dvisvgm/src/optimizer/TransformSimplifier.cpp index 461698b62a..2541631b0e 100644 --- a/dviware/dvisvgm/src/optimizer/TransformSimplifier.cpp +++ b/dviware/dvisvgm/src/optimizer/TransformSimplifier.cpp @@ -79,8 +79,9 @@ bool TransformSimplifier::incorporateTransform (XMLElement *elem, const Matrix & if (const char *ystr = elem->getAttributeValue("y")) y = strtod(ystr, nullptr); // width and height attributes must not become negative. Hence, only apply the scaling - // values if they are non-negative. Otherwise, keep a scaling matrix - if (sx < 0 || sy < 0) { + // values if they are non-negative. Otherwise, keep a scaling matrix. Also retain scaling + // transformations in image elements to avoid the need of attribute 'preseveAspectRatio'. + if (sx < 0 || sy < 0 || elem->name() == "image") { x += (sx == 0 ? 0 : tx/sx); y += (sy == 0 ? 0 : ty/sy); elem->addAttribute("transform", "scale("+XMLString(sx)+","+XMLString(sy)+")"); @@ -120,7 +121,7 @@ static string scale_cmd (double sx, double sy) { XMLString sxstr(sx), systr(sy); if (sxstr != "1" || systr != "1") { ret = "scale("+sxstr; - if (systr != "1") + if (systr != sxstr) ret += " "+systr; ret += ')'; } diff --git a/dviware/dvisvgm/src/optimizer/WSNodeRemover.cpp b/dviware/dvisvgm/src/optimizer/WSNodeRemover.cpp index be58bf70fd..3f83011f5a 100644 --- a/dviware/dvisvgm/src/optimizer/WSNodeRemover.cpp +++ b/dviware/dvisvgm/src/optimizer/WSNodeRemover.cpp @@ -34,7 +34,7 @@ void WSNodeRemover::execute (XMLElement *context) { while (child) { if (removeWS && child->toWSNode()) { XMLNode *next = child->next(); - XMLElement::remove(child); + XMLElement::detach(child); child = next; continue; } diff --git a/dviware/dvisvgm/src/psdefs.cpp b/dviware/dvisvgm/src/psdefs.cpp index 9c5991217d..9c8380816d 100644 --- a/dviware/dvisvgm/src/psdefs.cpp +++ b/dviware/dvisvgm/src/psdefs.cpp @@ -66,88 +66,90 @@ const char *PSInterpreter::PSDEFS = "felse}put @SD/fill{@dodraw @GD/@nulldev get not and{prcolor 0 1(newpath)prcmd " "prpath 0(fill)prcmd :newpath}{:fill}ifelse}put @SD/eofill{@dodraw @GD/@nulldev" " get not and{prcolor 0 1(newpath)prcmd prpath 0(eofill)prcmd :newpath}{:eofill" -"}ifelse}put @SD/clip{:clip @GD/@nulldev get not{0 1(newpath)prcmd prpath 0(cli" -"p)prcmd}if}put @SD/eoclip{:eoclip @GD/@nulldev get not{0 1(newpath)prcmd prpat" -"h 0(eoclip)prcmd}}put @SD/shfill{begin currentdict/ShadingType known currentdi" -"ct/ColorSpace known and currentdict/DataSource known and currentdict/Function " -"known not and ShadingType 4 ge{DataSource type/arraytype eq{<</DeviceGray 1/De" -"viceRGB 3/DeviceCMYK 4/bgknown currentdict/Background known/bbknown currentdic" -"t/BBox known>>begin currentdict ColorSpace known{ShadingType ColorSpace load b" -"gknown{1 Background aload pop}{0}ifelse bbknown{1 BBox aload pop}{0}ifelse Sha" -"dingType 5 eq{VerticesPerRow}if DataSource aload length 4 add bgknown{ColorSpa" -"ce load add}if bbknown{4 add}if ShadingType 5 eq{1 add}if(shfill)prcmd}if end}" -"if}if end}put @SD/image{dup type/dicttype eq{dup}{<</Width 6 index/Height 7 in" -"dex/colorimg false>>}ifelse @execimg}put @SD/colorimage{<<2 index{/Width 2 ind" -"ex 8 add index/Height 4 index 9 add index}{/Width 8 index/Height 9 index}ifels" -"e/colorimg true>>@execimg}put/@imgbase(./)def/@imgdevice(jpeg)def/@execimg{@GD" -"/@imgcnt 2 copy .knownget{1 add}{1}ifelse put begin<</imgdev null/imgid @GD/@i" -"mgcnt get/ispng @imgdevice 0 3 getinterval(png)eq dup/suffix exch{(.png)}{(.jp" -"g)}ifelse/colorimg currentdict/colorimg .knownget dup{pop}if/colordev 1 index " -"currentcolorspace dup length 1 ne exch 0 get/DeviceGray ne or or>>begin @imgde" -"vice(png)ne @imgdevice(jpeg)ne and{@imgdevice cvn}{colordev{ispng{/png16m}{/jp" -"eg}ifelse}{ispng{/pnggray}{/jpeggray}ifelse}ifelse}ifelse dup devicedict exch " -"known{:gsave/imgdev exch finddevice def mark/OutputFile @imgbase imgid 20 stri" -"ng cvs strconcat suffix strconcat/PageSize[Width Height]/UseFastColor true isp" -"ng{@imgdevice(pngmonod)eq{/MinFeatureSize where{pop/MinFeatureSize MinFeatureS" -"ize}if}if}{/JPEGQ where{pop/JPEGQ JPEGQ}if}ifelse imgdev putdeviceprops setdev" -"ice[Width 0 0 Height neg 0 Height]/setmatrix sysexec colorimg{:colorimage}{:im" -"age}ifelse/copypage sysexec mark/OutputFile()imgdev putdeviceprops pop :gresto" -"re imgid Width Height 3(image)prcmd}{pop colorimg{:colorimage}{:image}ifelse}i" -"felse end end}def/@rect{4 -2 roll moveto exch dup 0 rlineto exch 0 exch rlinet" -"o neg 0 rlineto closepath}bind def/@rectcc{4 -2 roll moveto 2 copy 0 lt exch 0" -" lt xor{dup 0 exch rlineto exch 0 rlineto neg 0 exch rlineto}{exch dup 0 rline" -"to exch 0 exch rlineto neg 0 rlineto}ifelse closepath}bind def @SD/rectclip{:n" -"ewpath dup type/arraytype eq{aload length 4 idiv{@rectcc}repeat}{@rectcc}ifels" -"e clip :newpath}put @SD/rectfill{:gsave :newpath dup type/arraytype eq{aload l" -"ength 4 idiv{@rectcc}repeat}{@rectcc}ifelse fill :grestore}put @SD/rectstroke{" -"gsave :newpath dup type/arraytype eq{aload length 4 idiv{@rect}repeat}{@rect}i" -"felse stroke grestore}put false setglobal @SD readonly pop/initclip 0 defpr/cl" -"ippath 0 defpr/sysexec{@SD exch get exec}def/adddot{dup length 1 add string du" -"p 0 46 put dup 3 -1 roll 1 exch putinterval}def/setlinewidth{dup/setlinewidth " -"sysexec 1(setlinewidth)prcmd}def/setlinecap 1 defpr/setlinejoin 1 defpr/setmit" -"erlimit 1 defpr/setdash{mark 3 1 roll 2 copy/setdash sysexec exch aload length" -" 1 add -1 roll counttomark(setdash)prcmd pop}def/@setpagedevice{pop<<>>/setpag" -"edevice sysexec matrix setmatrix newpath 0(setpagedevice)prcmd}def/@checknulld" -"ev{@GD/@nulldev get{currentpagedevice maxlength 0 ne{@GD/@nulldev false put 0 " -"1(setnulldevice)prcmd}if}if}def/prcolor{currentcolorspace @setcolorspace curre" -"ntrgbcolor 3(setrgbcolor)prcmd}def/printgstate{@dodraw @GD/@nulldev get not an" -"d{matrix currentmatrix aload pop 6(setmatrix)prcmd applyscalevals currentlinew" -"idth 1(setlinewidth)prcmd currentlinecap 1(setlinecap)prcmd currentlinejoin 1(" -"setlinejoin)prcmd currentmiterlimit 1(setmiterlimit)prcmd prcolor currentdash " -"mark 3 1 roll exch aload length 1 add -1 roll counttomark(setdash)prcmd pop}if" -"}def/strconcat{exch dup length 2 index length add string dup dup 4 2 roll copy" -" length 4 -1 roll putinterval}def/setgstate{/setgstate sysexec printgstate}def" -"/save{@UD begin/@saveID vmstatus pop pop def end :save @saveID 1(save)prcmd}de" -"f/restore{:restore @checknulldev printgstate @UD/@saveID known{@UD begin @save" -"ID end}{0}ifelse 1(restore)prcmd}def/gsave 0 defpr/grestore{:grestore @checknu" -"lldev printgstate 0(grestore)prcmd}def/grestoreall{:grestoreall @checknulldev " -"setstate 0(grestoreall)prcmd}def/rotate{dup type/arraytype ne @dodraw and{dup " -"1(rotate)prcmd}if/rotate sysexec applyscalevals}def/scale{dup type/arraytype n" -"e @dodraw and{2 copy 2(scale)prcmd}if/scale sysexec applyscalevals}def/transla" -"te{dup type/arraytype ne @dodraw and{2 copy 2(translate)prcmd}if/translate sys" -"exec}def/setmatrix{dup/setmatrix sysexec @dodraw{aload pop 6(setmatrix)prcmd a" -"pplyscalevals}{pop}ifelse}def/initmatrix{matrix setmatrix}def/concat{matrix cu" -"rrentmatrix matrix concatmatrix setmatrix}def/makepattern{gsave<</mx 3 -1 roll" -">>begin<</XUID[1000000 @patcnt]>>copy mx/makepattern sysexec dup begin Pattern" -"Type 2 lt{PatternType @patcnt BBox aload pop XStep YStep PaintType mx aload po" -"p 15(makepattern)prcmd :newpath matrix setmatrix dup PaintProc 0 1(makepattern" -")prcmd @GD/@patcnt @patcnt 1 add put}if end end grestore}def/setpattern{begin " -"PatternType 1 eq{PaintType 1 eq{XUID aload pop exch pop 1}{:gsave[currentcolor" -"space aload length -1 roll pop]/setcolorspace sysexec/setcolor sysexec XUID al" -"oad pop exch pop currentrgbcolor :grestore 4}ifelse(setpattern)prcmd currentco" -"lorspace 0 get/Pattern ne{[/Pattern currentcolorspace]/setcolorspace sysexec}i" -"f currentcolorspace @setcolorspace}{/setpattern sysexec}ifelse end}def/setcolo" -"r{dup type/dicttype eq{setpattern}{/setcolor sysexec/currentrgbcolor sysexec s" -"etrgbcolor}ifelse}def/setcolorspace{dup/setcolorspace sysexec @setcolorspace}d" -"ef/@setcolorspace{dup type/arraytype eq{0 get}if/Pattern eq{1}{0}ifelse 1(setc" -"olorspace)prcmd}def/setgray 1 defpr/setcmykcolor 4 defpr/sethsbcolor 3 defpr/s" -"etrgbcolor 3 defpr/.setopacityalpha{dup/.setopacityalpha sysexec 1(setopacitya" -"lpha)prcmd}def/.setshapealpha{dup/.setshapealpha sysexec 1(setshapealpha)prcmd" -"}def/.setblendmode{dup/.setblendmode sysexec<</Normal 0/Compatible 0/Multiply " -"1/Screen 2/Overlay 3/SoftLight 4/HardLight 5/ColorDodge 6/ColorBurn 7/Darken 8" -"/Lighten 9/Difference 10/Exclusion 11/Hue 12/Saturation 13/Color 14/Luminosity" -" 15/CompatibleOverprint 16>>exch get 1(setblendmode)prcmd}def/@pdfpagecount{(r" -")file runpdfbegin pdfpagecount runpdfend}def/@pdfpagebox{(r)file runpdfbegin d" -"up dup 1 lt exch pdfpagecount gt or{pop}{pdfgetpage/MediaBox pget pop aload po" -"p}ifelse runpdfend}def DELAYBIND{.bindnow}if "; +"}ifelse}put/.fillstroke{:gsave fill :grestore .swapcolors stroke .swapcolors}b" +"ind def/.eofillstroke{:gsave eofill :grestore .swapcolors stroke .swapcolors}b" +"ind def @SD/clip{:clip @GD/@nulldev get not{0 1(newpath)prcmd prpath 0(clip)pr" +"cmd}if}put @SD/eoclip{:eoclip @GD/@nulldev get not{0 1(newpath)prcmd prpath 0(" +"eoclip)prcmd}if}put @SD/shfill{begin currentdict/ShadingType known currentdict" +"/ColorSpace known and currentdict/DataSource known and currentdict/Function kn" +"own not and ShadingType 4 ge{DataSource type/arraytype eq{<</DeviceGray 1/Devi" +"ceRGB 3/DeviceCMYK 4/bgknown currentdict/Background known/bbknown currentdict/" +"BBox known>>begin currentdict ColorSpace known{ShadingType ColorSpace load bgk" +"nown{1 Background aload pop}{0}ifelse bbknown{1 BBox aload pop}{0}ifelse Shadi" +"ngType 5 eq{VerticesPerRow}if DataSource aload length 4 add bgknown{ColorSpace" +" load add}if bbknown{4 add}if ShadingType 5 eq{1 add}if(shfill)prcmd}if end}if" +"}if end}put @SD/image{dup type/dicttype eq{dup}{<</Width 6 index/Height 7 inde" +"x/colorimg false>>}ifelse @execimg}put @SD/colorimage{<<2 index{/Width 2 index" +" 8 add index/Height 4 index 9 add index}{/Width 8 index/Height 9 index}ifelse/" +"colorimg true>>@execimg}put/@imgbase(./)def/@imgdevice(jpeg)def/@execimg{@GD/@" +"imgcnt 2 copy .knownget{1 add}{1}ifelse put begin<</imgdev null/imgid @GD/@img" +"cnt get/ispng @imgdevice 0 3 getinterval(png)eq dup/suffix exch{(.png)}{(.jpg)" +"}ifelse/colorimg currentdict/colorimg .knownget dup{pop}if/colordev 1 index cu" +"rrentcolorspace dup length 1 ne exch 0 get/DeviceGray ne or or>>begin @imgdevi" +"ce(png)ne @imgdevice(jpeg)ne and{@imgdevice cvn}{colordev{ispng{/png16m}{/jpeg" +"}ifelse}{ispng{/pnggray}{/jpeggray}ifelse}ifelse}ifelse dup devicedict exch kn" +"own{:gsave/imgdev exch finddevice def mark/OutputFile @imgbase imgid 20 string" +" cvs strconcat suffix strconcat/PageSize[Width Height]/UseFastColor true ispng" +"{@imgdevice(pngmonod)eq{/MinFeatureSize where{pop/MinFeatureSize MinFeatureSiz" +"e}if}if}{/JPEGQ where{pop/JPEGQ JPEGQ}if}ifelse imgdev putdeviceprops setdevic" +"e[Width 0 0 Height neg 0 Height]/setmatrix sysexec colorimg{:colorimage}{:imag" +"e}ifelse/copypage sysexec mark/OutputFile()imgdev putdeviceprops pop :grestore" +" imgid Width Height 3(image)prcmd}{pop colorimg{:colorimage}{:image}ifelse}ife" +"lse end end}def/@rect{4 -2 roll moveto exch dup 0 rlineto exch 0 exch rlineto " +"neg 0 rlineto closepath}bind def/@rectcc{4 -2 roll moveto 2 copy 0 lt exch 0 l" +"t xor{dup 0 exch rlineto exch 0 rlineto neg 0 exch rlineto}{exch dup 0 rlineto" +" exch 0 exch rlineto neg 0 rlineto}ifelse closepath}bind def @SD/rectclip{:new" +"path dup type/arraytype eq{aload length 4 idiv{@rectcc}repeat}{@rectcc}ifelse " +"clip :newpath}put @SD/rectfill{:gsave :newpath dup type/arraytype eq{aload len" +"gth 4 idiv{@rectcc}repeat}{@rectcc}ifelse fill :grestore}put @SD/rectstroke{gs" +"ave :newpath dup type/arraytype eq{aload length 4 idiv{@rect}repeat}{@rect}ife" +"lse stroke grestore}put false setglobal @SD readonly pop/initclip 0 defpr/clip" +"path 0 defpr/sysexec{@SD exch get exec}def/adddot{dup length 1 add string dup " +"0 46 put dup 3 -1 roll 1 exch putinterval}def/setlinewidth{dup/setlinewidth sy" +"sexec 1(setlinewidth)prcmd}def/setlinecap 1 defpr/setlinejoin 1 defpr/setmiter" +"limit 1 defpr/setdash{mark 3 1 roll 2 copy/setdash sysexec exch aload length 1" +" add -1 roll counttomark(setdash)prcmd pop}def/@setpagedevice{pop<<>>/setpaged" +"evice sysexec matrix setmatrix newpath 0(setpagedevice)prcmd}def/@checknulldev" +"{@GD/@nulldev get{currentpagedevice maxlength 0 ne{@GD/@nulldev false put 0 1(" +"setnulldevice)prcmd}if}if}def/prcolor{currentcolorspace @setcolorspace current" +"rgbcolor 3(setrgbcolor)prcmd}def/printgstate{@dodraw @GD/@nulldev get not and{" +"matrix currentmatrix aload pop 6(setmatrix)prcmd applyscalevals currentlinewid" +"th 1(setlinewidth)prcmd currentlinecap 1(setlinecap)prcmd currentlinejoin 1(se" +"tlinejoin)prcmd currentmiterlimit 1(setmiterlimit)prcmd prcolor currentdash ma" +"rk 3 1 roll exch aload length 1 add -1 roll counttomark(setdash)prcmd pop}if}d" +"ef/strconcat{exch dup length 2 index length add string dup dup 4 2 roll copy l" +"ength 4 -1 roll putinterval}def/setgstate{/setgstate sysexec printgstate}def/s" +"ave{@UD begin/@saveID vmstatus pop pop def end :save @saveID 1(save)prcmd}def/" +"restore{:restore @checknulldev printgstate @UD/@saveID known{@UD begin @saveID" +" end}{0}ifelse 1(restore)prcmd}def/gsave 0 defpr/grestore{:grestore @checknull" +"dev printgstate 0(grestore)prcmd}def/grestoreall{:grestoreall @checknulldev se" +"tstate 0(grestoreall)prcmd}def/rotate{dup type/arraytype ne @dodraw and{dup 1(" +"rotate)prcmd}if/rotate sysexec applyscalevals}def/scale{dup type/arraytype ne " +"@dodraw and{2 copy 2(scale)prcmd}if/scale sysexec applyscalevals}def/translate" +"{dup type/arraytype ne @dodraw and{2 copy 2(translate)prcmd}if/translate sysex" +"ec}def/setmatrix{dup/setmatrix sysexec @dodraw{aload pop 6(setmatrix)prcmd app" +"lyscalevals}{pop}ifelse}def/initmatrix{matrix setmatrix}def/concat{matrix curr" +"entmatrix matrix concatmatrix setmatrix}def/makepattern{gsave<</mx 3 -1 roll>>" +"begin<</XUID[1000000 @patcnt]>>copy mx/makepattern sysexec dup begin PatternTy" +"pe 2 lt{PatternType @patcnt BBox aload pop XStep YStep PaintType mx aload pop " +"15(makepattern)prcmd :newpath matrix setmatrix dup PaintProc 0 1(makepattern)p" +"rcmd @GD/@patcnt @patcnt 1 add put}if end end grestore}def/setpattern{begin Pa" +"tternType 1 eq{PaintType 1 eq{XUID aload pop exch pop 1}{:gsave[currentcolorsp" +"ace aload length -1 roll pop]/setcolorspace sysexec/setcolor sysexec XUID aloa" +"d pop exch pop currentrgbcolor :grestore 4}ifelse(setpattern)prcmd currentcolo" +"rspace 0 get/Pattern ne{[/Pattern currentcolorspace]/setcolorspace sysexec}if " +"currentcolorspace @setcolorspace}{/setpattern sysexec}ifelse end}def/setcolor{" +"dup type/dicttype eq{setpattern}{/setcolor sysexec/currentrgbcolor sysexec set" +"rgbcolor}ifelse}def/setcolorspace{dup/setcolorspace sysexec @setcolorspace}def" +"/@setcolorspace{dup type/arraytype eq{0 get}if/Pattern eq{1}{0}ifelse 1(setcol" +"orspace)prcmd}def/setgray 1 defpr/setcmykcolor 4 defpr/sethsbcolor 3 defpr/set" +"rgbcolor 3 defpr/.setopacityalpha{dup/.setopacityalpha sysexec 1(setopacityalp" +"ha)prcmd}def/.setshapealpha{dup/.setshapealpha sysexec 1(setshapealpha)prcmd}d" +"ef/.setblendmode{dup/.setblendmode sysexec<</Normal 0/Compatible 0/Multiply 1/" +"Screen 2/Overlay 3/SoftLight 4/HardLight 5/ColorDodge 6/ColorBurn 7/Darken 8/L" +"ighten 9/Difference 10/Exclusion 11/Hue 12/Saturation 13/Color 14/Luminosity 1" +"5/CompatibleOverprint 16>>exch get 1(setblendmode)prcmd}def/@pdfpagecount{(r)f" +"ile runpdfbegin pdfpagecount runpdfend}def/@pdfpagebox{(r)file runpdfbegin dup" +" dup 1 lt exch pdfpagecount gt or{pop}{pdfgetpage/MediaBox pget pop aload pop}" +"ifelse runpdfend}def DELAYBIND{.bindnow}if "; diff --git a/dviware/dvisvgm/tests/DvisvgmSpecialTest.cpp b/dviware/dvisvgm/tests/DvisvgmSpecialTest.cpp index b8e67403ae..89dba64201 100644 --- a/dviware/dvisvgm/tests/DvisvgmSpecialTest.cpp +++ b/dviware/dvisvgm/tests/DvisvgmSpecialTest.cpp @@ -92,10 +92,10 @@ TEST_F(DvisvgmSpecialTest, rawText) { EXPECT_TRUE(recorder.defsEquals("")); EXPECT_TRUE(recorder.pageEquals("<g id='page1'>first\n-42,14</g>")) << recorder.pageString(); - iss.clear(); iss.str("raw \t ;{?(-x+2*y-5)}second {?bbox dummy} \t"); + iss.clear(); iss.str("raw \t ;{?(-x+2*y-5)}{?(-y+2*x-5)}second {?bbox dummy} \t"); handler.process("", iss, recorder); EXPECT_TRUE(recorder.defsEquals("")); - EXPECT_TRUE(recorder.pageEquals("<g id='page1'>first\n-42,14;65second 0 0 0 0</g>")) << recorder.pageString(); + EXPECT_TRUE(recorder.pageEquals("<g id='page1'>first\n-42,14;65-103second 0 0 0 0</g>")) << recorder.pageString(); } diff --git a/dviware/dvisvgm/tests/HashFunctionTest.cpp b/dviware/dvisvgm/tests/HashFunctionTest.cpp index 6b287014df..0d5b7bc1b4 100644 --- a/dviware/dvisvgm/tests/HashFunctionTest.cpp +++ b/dviware/dvisvgm/tests/HashFunctionTest.cpp @@ -36,7 +36,7 @@ TEST(HashFunctionTest, md5) { md5.update("0123456789"); EXPECT_EQ(md5.digestString(), "781e5e245d69b566979b86e28d23f2c7"); vector<uint8_t> bytes = {0x78, 0x1e, 0x5e, 0x24, 0x5d, 0x69, 0xb5, 0x66, 0x97, 0x9b, 0x86, 0xe2, 0x8d, 0x23, 0xf2, 0xc7}; - EXPECT_EQ(md5.digestValue(), bytes); + EXPECT_EQ(md5.digestBytes(), bytes); } @@ -51,7 +51,7 @@ TEST(HashFunctionTest, xxh32) { xxh32.update("0123456789"); EXPECT_EQ(xxh32.digestString(), "950c9c0a"); vector<uint8_t> bytes = {0x95, 0x0c, 0x9c, 0x0a}; - EXPECT_EQ(xxh32.digestValue(), bytes); + EXPECT_EQ(xxh32.digestBytes(), bytes); } @@ -66,7 +66,7 @@ TEST(HashFunctionTest, xxh64) { xxh64.update("0123456789"); EXPECT_EQ(xxh64.digestString(), "3f5fc178a81867e7"); vector<uint8_t> bytes = {0x3f, 0x5f, 0xc1, 0x78, 0xa8, 0x18, 0x67, 0xe7}; - EXPECT_EQ(xxh64.digestValue(), bytes); + EXPECT_EQ(xxh64.digestBytes(), bytes); } @@ -82,7 +82,7 @@ TEST(HashFunctionTest, xxh128) { xxh128.update("0123456789"); EXPECT_EQ(xxh128.digestString(), "e353667619ec664b49655fc9692165fb"); vector<uint8_t> bytes = {0xe3, 0x53, 0x66, 0x76, 0x19, 0xec, 0x66, 0x4b, 0x49, 0x65, 0x5F, 0xc9, 0x69, 0x21, 0x65, 0xfb}; - EXPECT_EQ(xxh128.digestValue(), bytes); + EXPECT_EQ(xxh128.digestBytes(), bytes); } #endif |