From 8fedf4998449d5b6b909a1118fc2e152e4d2e6e7 Mon Sep 17 00:00:00 2001 From: Davide Quarta Date: Tue, 23 Jan 2024 19:36:49 +0100 Subject: replay mode support --- include/config.h | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) (limited to 'include/config.h') diff --git a/include/config.h b/include/config.h index 63340650..1649f110 100644 --- a/include/config.h +++ b/include/config.h @@ -83,7 +83,10 @@ will be kept and written to the crash/ directory as RECORD:... files. Note that every crash will be written, not only unique ones! */ -// #define AFL_PERSISTENT_RECORD +// #define AFL_PERSISTENT_RECORD + +/* Builds compiler-rt with support to replay persistent records */ +// #define AFL_PERSISTENT_REPLAY /* console output colors: There are three ways to configure its behavior * 1. default: colored outputs fixed on: defined USE_COLOR && defined -- cgit 1.4.1 From 023fc19ce04bffcbd623e27a1f2d1810c3ec0c3c Mon Sep 17 00:00:00 2001 From: Davide Quarta Date: Mon, 5 Feb 2024 18:26:46 +0100 Subject: better replay mode error handling, added replay mode documentation, code formatting --- include/afl-fuzz.h | 2 +- include/config.h | 10 +- include/persistent_replay.h | 152 +++++++++++++++++++--------- instrumentation/README.persistent_mode.md | 30 +++++- instrumentation/afl-compiler-rt.o.c | 58 ++++++----- src/afl-forkserver.c | 5 +- utils/persistent_mode/persistent_demo_new.c | 6 +- 7 files changed, 182 insertions(+), 81 deletions(-) (limited to 'include/config.h') diff --git a/include/afl-fuzz.h b/include/afl-fuzz.h index 864bc6b6..f95dcc20 100644 --- a/include/afl-fuzz.h +++ b/include/afl-fuzz.h @@ -125,7 +125,7 @@ #endif /* ^!SIMPLE_FILES */ #ifdef AFL_PERSISTENT_RECORD - #define RECORD_PREFIX "RECORD:" + #define RECORD_PREFIX "RECORD:" #endif #define STAGE_BUF_SIZE (64) /* usable size for stage name buf in afl_state */ diff --git a/include/config.h b/include/config.h index 1649f110..d44cda9c 100644 --- a/include/config.h +++ b/include/config.h @@ -83,10 +83,14 @@ will be kept and written to the crash/ directory as RECORD:... files. Note that every crash will be written, not only unique ones! */ -// #define AFL_PERSISTENT_RECORD +#define AFL_PERSISTENT_RECORD -/* Builds compiler-rt with support to replay persistent records */ -// #define AFL_PERSISTENT_REPLAY +/* Adds support in compiler-rt to replay persistent records */ +#define AFL_PERSISTENT_REPLAY + +/* Adds support in compiler-rt to replay persistent records in @@-style + * harnesses */ +// #define AFL_PERSISTENT_REPLAY_ARGPARSE /* console output colors: There are three ways to configure its behavior * 1. default: colored outputs fixed on: defined USE_COLOR && defined diff --git a/include/persistent_replay.h b/include/persistent_replay.h index b1a55e9f..58b22fb4 100644 --- a/include/persistent_replay.h +++ b/include/persistent_replay.h @@ -11,71 +11,116 @@ #include static unsigned short int is_replay_record; -static unsigned int replay_record; -static unsigned int replay_record_cnt; -static char replay_record_path[PATH_MAX]; -static char **record_arg; -static char *replay_record_dir; -static struct dirent **record_list; +static unsigned int replay_record; +static unsigned int replay_record_cnt; +static char replay_record_path[PATH_MAX]; +static char *replay_record_dir; +static struct dirent **record_list; + +#ifdef AFL_PERSISTENT_REPLAY_ARGPARSE +static char **record_arg = NULL; +#endif // AFL_PERSISTENT_REPLAY_ARGPARSE static int select_files(const struct dirent *dirbuf) { char fn[4096]; - if (dirbuf->d_name[0] == '.'){ + if (dirbuf->d_name[0] == '.') { + return 0; + } else { + snprintf(fn, sizeof(fn), "RECORD:%06u", replay_record); return !!strstr(dirbuf->d_name, fn); + } + } - + static int compare_files(const struct dirent **da, const struct dirent **db) { - - unsigned int c1=0, c2=0; + + unsigned int c1 = 0, c2 = 0; sscanf((*da)->d_name, "RECORD:%*u,cnt:%06u", &c1); sscanf((*db)->d_name, "RECORD:%*u,cnt:%06u", &c2); - return c1-c2; + return c1 - c2; + } -__attribute__((destructor)) static void __afl_record_replay_destroy(void){ - for (int i=0; i < replay_record_cnt; i++) { +__attribute__((destructor)) static void __afl_record_replay_destroy(void) { + + for (int i = 0; i < replay_record_cnt; i++) { + free(record_list[i]); + } + free(record_list); + } -__attribute__((constructor)) static void __afl_record_replay_init(int argc, char **argv) { - +__attribute__((constructor)) static void __afl_record_replay_init( +#ifdef AFL_PERSISTENT_REPLAY_ARGPARSE + int argc, char **argv +#endif // AFL_PERSISTENT_REPLAY_ARGPARSE +) { + +#ifdef AFL_PERSISTENT_REPLAY_ARGPARSE char **argp; +#endif // AFL_PERSISTENT_REPLAY_ARGPARSE + + struct stat sb; + + /* caveat: if harness uses @@ and we don't pass it, it will regardless loop + * the number of iterations defined for AFL_LOOP (on the same file)*/ + if (!(is_replay_record = !!getenv("AFL_PERSISTENT_REPLAY"))) { - /* caveat: if harness uses @@ and we don't pass it, it will regardless loop the number of iterations defined for AFL_LOOP (on the same file)*/ - if(!(is_replay_record = !!getenv("AFL_PERSISTENT_REPLAY"))){ // printf("[warning] AFL_PERSISTENT_REPLAY not set.\n"); return; + } replay_record = atoi(getenv("AFL_PERSISTENT_REPLAY")); replay_record_dir = getenv("AFL_PERSISTENT_DIR"); - replay_record_cnt = scandir(replay_record_dir ? replay_record_dir : "./", &record_list, select_files, compare_files); - if (!replay_record_cnt){ - printf("[error] Can't find the requested record!\n"); + if (!(stat(replay_record_dir, &sb) == 0 && S_ISDIR(sb.st_mode))) { + + fprintf(stderr, "[error] Can't find the requested record directory!\n"); is_replay_record = 0; + return; + } + replay_record_cnt = scandir(replay_record_dir ? replay_record_dir : "./", + &record_list, select_files, compare_files); + + if (!replay_record_cnt) { + + fprintf(stderr, "[error] Can't find the requested record!\n"); + is_replay_record = 0; + + } + +#ifdef AFL_PERSISTENT_REPLAY_ARGPARSE argp = argv; - while (*argp){ - if (!strcmp(*argp, "@@")){ + while (*argp) { + + if (!strcmp(*argp, "@@")) { + record_arg = argp; *record_arg = replay_record_path; break; + } + ++argp; + } +#endif // AFL_PERSISTENT_REPLAY_ARGPARSE + } /* only used if explictly included for compatibility @@ -83,67 +128,80 @@ __attribute__((constructor)) static void __afl_record_replay_init(int argc, char #ifdef AFL_COMPAT -#ifndef PATH_MAX - #define PATH_MAX 4096 -#endif + #ifndef PATH_MAX + #define PATH_MAX 4096 + #endif -#define FUZZ_BUF_SIZE 1024000 + #define FUZZ_BUF_SIZE 1024000 -// extern ssize_t read(int fildes, void *buf, size_t nbyte); + // extern ssize_t read(int fildes, void *buf, size_t nbyte); -//extern int __afl_persistent_loop(unsigned int max_cnt); -//extern unsigned char fuzz_buf[]; + // extern int __afl_persistent_loop(unsigned int max_cnt); + // extern unsigned char fuzz_buf[]; -#ifndef __AFL_HAVE_MANUAL_CONTROL - #define __AFL_HAVE_MANUAL_CONTROL -#endif + #ifndef __AFL_HAVE_MANUAL_CONTROL + #define __AFL_HAVE_MANUAL_CONTROL + #endif -#define __AFL_FUZZ_TESTCASE_LEN (read(0, fuzz_buf, FUZZ_BUF_SIZE)) -#define __AFL_FUZZ_TESTCASE_BUF fuzz_buf -#define __AFL_FUZZ_INIT() void sync(void); -#define __AFL_INIT() sync() -#define __AFL_LOOP(x) __afl_persistent_loop(x) + #define __AFL_FUZZ_TESTCASE_LEN (read(0, fuzz_buf, FUZZ_BUF_SIZE)) + #define __AFL_FUZZ_TESTCASE_BUF fuzz_buf + #define __AFL_FUZZ_INIT() void sync(void); + #define __AFL_INIT() sync() + #define __AFL_LOOP(x) __afl_persistent_loop(x) unsigned char fuzz_buf[FUZZ_BUF_SIZE]; int __afl_persistent_loop(unsigned int max_cnt) { - static unsigned int cycle_cnt = 1; + static unsigned int cycle_cnt = 1; static unsigned short int inited = 0; - char tcase[PATH_MAX]; + char tcase[PATH_MAX]; + + if (is_replay_record) { - if( is_replay_record ){ + if (!inited) { - if (!inited){ cycle_cnt = replay_record_cnt; inited = 1; + } snprintf(tcase, PATH_MAX, "%s/%s", - replay_record_dir ? replay_record_dir : "./", - record_list[replay_record_cnt-cycle_cnt]->d_name); - + replay_record_dir ? replay_record_dir : "./", + record_list[replay_record_cnt - cycle_cnt]->d_name); + #ifdef AFL_PERSISTENT_REPLAY_ARGPARSE if (record_arg) { + *record_arg = tcase; - } else { + + } else + + #endif // AFL_PERSISTENT_REPLAY_ARGPARSE + { + int fd = open(tcase, O_RDONLY); dup2(fd, 0); close(fd); + } } else { - if (!inited){ + if (!inited) { + cycle_cnt = max_cnt; inited = 1; + } } return cycle_cnt--; + } #endif // AFL_COMPAT -#endif // _HAVE_PERSISTENT_REPLAY_H \ No newline at end of file +#endif // _HAVE_PERSISTENT_REPLAY_H + diff --git a/instrumentation/README.persistent_mode.md b/instrumentation/README.persistent_mode.md index 14e59f4a..b5d982b0 100644 --- a/instrumentation/README.persistent_mode.md +++ b/instrumentation/README.persistent_mode.md @@ -195,4 +195,32 @@ Then as first line after the `__AFL_LOOP` while loop: int len = __AFL_FUZZ_TESTCASE_LEN; ``` -And that is all! \ No newline at end of file +And that is all! + +## 6) Persistent record, and replay + +If your software under test requires keeping a state between persistent loop iterations (i.e., a stateful network stack), you can use the `AFL_PERSISTENT_RECORD` variable as described in the [environment variables documentation](../docs/env_variables.md). + +To easily replay a crashing, or hanging record, you can use the persistent replay functionality by compiling AFL++ after uncommenting the `AFL_PERSISTENT_REPLAY` define in [config.h](../include/config.h). + +You can then run the test binary specifying the record number via the AFL_PERSISTENT_REPLAY environment variable (i.e., `RECORD:XXXXX`` -> `AFL_PERSISTENT_REPLAY=XXXXX`). +The directory where the record files live can be specified via the `AFL_PERSISTENT_DIR` environment varilable, otherwise by default it will be considered the current directory (`./`). + +If your harness reads the input files from arguments using the special `@@` argument you will need to define `AFL_PERSISTENT_ARGPARSE` in `config.h`, or before including the `persistent_replay.h` header file as show before. +In order to offer transparent support to harnesses using the `@@` command line argument, arguments are parsed by the `__afl_record_replay_init` init function. Since not all systems support passing arguments to initializers, this functionality is disabled by default, it's recommendable to use the `__AFL_FUZZ_TESTCASE_BUF/__AFL_FUZZ_TESTCASE_LEN` shared memory mechanism instead. + +### 7) Drop in replay functionality + +To use the replay functionality without having to use `afl-cc` you can just define `AFL_COMPAT` and include the [include/persistent_replay.h](../include/persistent_replay.h) self contained header file that provides a drop-in replacement for the persistent loop mechanism. + +```c +#ifndef __AFL_FUZZ_TESTCASE_LEN + #define AFL_COMPAT + // #define AFL_PERSISTENT_REPLAY_ARGPARSE + #include "persistent_replay.h" +#endif + +__AFL_FUZZ_INIT(); +``` + +A simple example is provided in [persistent_demo_new.c](../utils/persistent_mode/persistent_demo_new.c). \ No newline at end of file diff --git a/instrumentation/afl-compiler-rt.o.c b/instrumentation/afl-compiler-rt.o.c index 0fa22aee..037caaf0 100644 --- a/instrumentation/afl-compiler-rt.o.c +++ b/instrumentation/afl-compiler-rt.o.c @@ -84,7 +84,7 @@ #include #ifdef AFL_PERSISTENT_REPLAY -#include "persistent_replay.h" + #include "persistent_replay.h" #endif /* Globals needed by the injected instrumentation. The __afl_area_initial region @@ -1344,37 +1344,49 @@ int __afl_persistent_loop(unsigned int max_cnt) { #ifdef AFL_PERSISTENT_REPLAY -#ifndef PATH_MAX - #define PATH_MAX 4096 -#endif + #ifndef PATH_MAX + #define PATH_MAX 4096 + #endif - static u8 inited = 0; - char tcase[PATH_MAX]; + static u8 inited = 0; + char tcase[PATH_MAX]; - if( unlikely(is_replay_record) ){ + if (unlikely(is_replay_record)) { - if (!inited){ - cycle_cnt = replay_record_cnt; - inited = 1; - } + if (!inited) { - snprintf(tcase, PATH_MAX, "%s/%s", - replay_record_dir ? replay_record_dir : "./", - record_list[replay_record_cnt-cycle_cnt]->d_name); + cycle_cnt = replay_record_cnt; + inited = 1; + + } + + snprintf(tcase, PATH_MAX, "%s/%s", + replay_record_dir ? replay_record_dir : "./", + record_list[replay_record_cnt - cycle_cnt]->d_name); + + #ifdef AFL_PERSISTENT_REPLAY_ARGPARSE + if (record_arg) { + + *record_arg = tcase; + + } else + + #endif // AFL_PERSISTENT_REPLAY_ARGPARSE + { + + int fd = open(tcase, O_RDONLY); + dup2(fd, 0); + close(fd); + + } - if (record_arg) { - *record_arg = tcase; - } else { - int fd = open(tcase, O_RDONLY); - dup2(fd, 0); - close(fd); - } return cycle_cnt--; + } else -#endif +#endif - if (first_pass) { + if (first_pass) { /* Make sure that every iteration of __AFL_LOOP() starts with a clean slate. On subsequent calls, the parent will take care of that, but on the first diff --git a/src/afl-forkserver.c b/src/afl-forkserver.c index f8dd783f..36e46444 100644 --- a/src/afl-forkserver.c +++ b/src/afl-forkserver.c @@ -1593,7 +1593,7 @@ afl_fsrv_run_target(afl_forkserver_t *fsrv, u32 timeout, #ifdef AFL_PERSISTENT_RECORD fsrv_run_result_t retval = FSRV_RUN_OK; - char *persistent_out_fmt; + char *persistent_out_fmt; #endif #ifdef __linux__ @@ -1803,6 +1803,7 @@ afl_fsrv_run_target(afl_forkserver_t *fsrv, u32 timeout, persistent_out_fmt = "%s/hangs/RECORD:%06u,cnt:%06u"; goto store_persistent_record; #endif + } /* Did we crash? @@ -1841,7 +1842,7 @@ afl_fsrv_run_target(afl_forkserver_t *fsrv, u32 timeout, #ifdef AFL_PERSISTENT_RECORD store_persistent_record: - if (unlikely(retval == FSRV_RUN_CRASH || retval == FSRV_RUN_TMOUT) && + if (unlikely(retval == FSRV_RUN_CRASH || retval == FSRV_RUN_TMOUT) && unlikely(fsrv->persistent_record)) { char fn[PATH_MAX]; diff --git a/utils/persistent_mode/persistent_demo_new.c b/utils/persistent_mode/persistent_demo_new.c index 40ada9e1..3d9d90a6 100644 --- a/utils/persistent_mode/persistent_demo_new.c +++ b/utils/persistent_mode/persistent_demo_new.c @@ -31,8 +31,8 @@ /* this lets the source compile without afl-clang-fast/lto */ #ifndef __AFL_FUZZ_TESTCASE_LEN -#define AFL_COMPAT -#include "persistent_replay.h" + #define AFL_COMPAT + #include "persistent_replay.h" #endif __AFL_FUZZ_INIT(); @@ -86,8 +86,6 @@ int main(int argc, char **argv) { if (buf[5] == '!') { printf("six\n"); - char *nullo = NULL+1; - *nullo = 'p'; abort(); } -- cgit 1.4.1 From e405e721fad46e594b633147a6940cfdd602e4c2 Mon Sep 17 00:00:00 2001 From: Davide Quarta Date: Tue, 6 Feb 2024 18:19:52 +0100 Subject: reuse first_pass aux var in persistent loop for record replay mode, keep area ptr and loc logic intact in record replay mode, move replay record example to own dir in utils, update docs, move record compat layer to separate header file --- include/afl-persistent-replay.h | 131 +++++++++++++++++ include/afl-record-compat.h | 67 +++++++++ include/config.h | 6 +- include/persistent_replay.h | 207 --------------------------- instrumentation/README.persistent_mode.md | 20 +-- instrumentation/afl-compiler-rt.o.c | 87 +++++------ utils/persistent_mode/Makefile | 1 - utils/replay_record/Makefile | 8 ++ utils/replay_record/persistent_demo_replay.c | 148 +++++++++++++++++++ 9 files changed, 413 insertions(+), 262 deletions(-) create mode 100644 include/afl-persistent-replay.h create mode 100644 include/afl-record-compat.h delete mode 100644 include/persistent_replay.h create mode 100644 utils/replay_record/Makefile create mode 100644 utils/replay_record/persistent_demo_replay.c (limited to 'include/config.h') diff --git a/include/afl-persistent-replay.h b/include/afl-persistent-replay.h new file mode 100644 index 00000000..9e60ff9c --- /dev/null +++ b/include/afl-persistent-replay.h @@ -0,0 +1,131 @@ +#ifndef _HAVE_PERSISTENT_REPLAY_H +#define _HAVE_PERSISTENT_REPLAY_H + +#include +#include +#include +#include +#include +#include +#include +#include + +#ifndef PATH_MAX + #define PATH_MAX 4096 +#endif + +static unsigned short int is_replay_record; +static unsigned int replay_record; +static unsigned int replay_record_cnt; +static char replay_record_path[PATH_MAX]; +static char *replay_record_dir; +static struct dirent **record_list; + +#ifdef AFL_PERSISTENT_REPLAY_ARGPARSE +static char **record_arg = NULL; +#endif // AFL_PERSISTENT_REPLAY_ARGPARSE + +static int select_files(const struct dirent *dirbuf) { + + char fn[PATH_MAX]; + + if (dirbuf->d_name[0] == '.') { + + return 0; + + } else { + + snprintf(fn, sizeof(fn), "RECORD:%06u", replay_record); + return !!strstr(dirbuf->d_name, fn); + + } + +} + +static int compare_files(const struct dirent **da, const struct dirent **db) { + + unsigned int c1 = 0, c2 = 0; + + sscanf((*da)->d_name, "RECORD:%*u,cnt:%06u", &c1); + sscanf((*db)->d_name, "RECORD:%*u,cnt:%06u", &c2); + + return c1 - c2; + +} + +__attribute__((destructor)) static void __afl_record_replay_destroy(void) { + + for (int i = 0; i < replay_record_cnt; i++) { + + free(record_list[i]); + + } + + free(record_list); + +} + +__attribute__((constructor)) static void __afl_record_replay_init( +#ifdef AFL_PERSISTENT_REPLAY_ARGPARSE + int argc, char **argv +#endif // AFL_PERSISTENT_REPLAY_ARGPARSE +) { + +#ifdef AFL_PERSISTENT_REPLAY_ARGPARSE + char **argp; +#endif // AFL_PERSISTENT_REPLAY_ARGPARSE + + struct stat sb; + + /* caveat: if harness uses @@ and we don't pass it, it will regardless loop + * the number of iterations defined for AFL_LOOP (on the same file)*/ + if (!(is_replay_record = !!getenv("AFL_PERSISTENT_REPLAY"))) { + + // printf("[warning] AFL_PERSISTENT_REPLAY not set.\n"); + return; + + } + + replay_record = atoi(getenv("AFL_PERSISTENT_REPLAY")); + replay_record_dir = getenv("AFL_PERSISTENT_DIR"); + + if (!(stat(replay_record_dir, &sb) == 0 && S_ISDIR(sb.st_mode))) { + + fprintf(stderr, "[error] Can't find the requested record directory!\n"); + is_replay_record = 0; + return; + + } + + replay_record_cnt = scandir(replay_record_dir ? replay_record_dir : "./", + &record_list, select_files, compare_files); + + if (!replay_record_cnt) { + + fprintf(stderr, "[error] Can't find the requested record!\n"); + is_replay_record = 0; + + } + +#ifdef AFL_PERSISTENT_REPLAY_ARGPARSE + argp = argv; + while (*argp) { + + if (!strcmp(*argp, "@@")) { + + record_arg = argp; + *record_arg = replay_record_path; + break; + + } + + ++argp; + + } + +#endif // AFL_PERSISTENT_REPLAY_ARGPARSE + +} + +#endif // _HAVE_PERSISTENT_REPLAY_H + diff --git a/include/afl-record-compat.h b/include/afl-record-compat.h new file mode 100644 index 00000000..2c79595d --- /dev/null +++ b/include/afl-record-compat.h @@ -0,0 +1,67 @@ +#ifndef _HAVE_AFL_COMPAT_H +#define _HAVE_AFL_COMPAT_H + +#include + +#define FUZZ_BUF_SIZE 1024000 + +// extern ssize_t read(int fildes, void *buf, size_t nbyte); + +// extern int __afl_persistent_loop(unsigned int max_cnt); +// extern unsigned char fuzz_buf[]; + +#ifndef __AFL_HAVE_MANUAL_CONTROL + #define __AFL_HAVE_MANUAL_CONTROL +#endif + +#define __AFL_FUZZ_TESTCASE_LEN (read(0, fuzz_buf, FUZZ_BUF_SIZE)) +#define __AFL_FUZZ_TESTCASE_BUF fuzz_buf +#define __AFL_FUZZ_INIT() void sync(void); +#define __AFL_INIT() sync() +#define __AFL_LOOP(x) __afl_persistent_loop(x) + +unsigned char fuzz_buf[FUZZ_BUF_SIZE]; + +int __afl_persistent_loop(unsigned int max_cnt) { + + static unsigned int cycle_cnt = 1; + static unsigned short int inited = 0; + char tcase[PATH_MAX]; + + if (is_replay_record) { + + if (!inited) { + + cycle_cnt = replay_record_cnt; + inited = 1; + + } + + snprintf(tcase, PATH_MAX, "%s/%s", + replay_record_dir ? replay_record_dir : "./", + record_list[replay_record_cnt - cycle_cnt]->d_name); + +#ifdef AFL_PERSISTENT_REPLAY_ARGPARSE + if (record_arg) { + + *record_arg = tcase; + + } else + +#endif // AFL_PERSISTENT_REPLAY_ARGPARSE + { + + int fd = open(tcase, O_RDONLY); + dup2(fd, 0); + close(fd); + + } + + } + + return --cycle_cnt; + +} + +#endif // _HAVE_AFL_COMPAT_H + diff --git a/include/config.h b/include/config.h index d44cda9c..a5b6eba1 100644 --- a/include/config.h +++ b/include/config.h @@ -83,13 +83,11 @@ will be kept and written to the crash/ directory as RECORD:... files. Note that every crash will be written, not only unique ones! */ -#define AFL_PERSISTENT_RECORD - -/* Adds support in compiler-rt to replay persistent records */ -#define AFL_PERSISTENT_REPLAY +// #define AFL_PERSISTENT_RECORD /* Adds support in compiler-rt to replay persistent records in @@-style * harnesses */ + // #define AFL_PERSISTENT_REPLAY_ARGPARSE /* console output colors: There are three ways to configure its behavior diff --git a/include/persistent_replay.h b/include/persistent_replay.h deleted file mode 100644 index 58b22fb4..00000000 --- a/include/persistent_replay.h +++ /dev/null @@ -1,207 +0,0 @@ -#ifndef _HAVE_PERSISTENT_REPLAY_H -#define _HAVE_PERSISTENT_REPLAY_H - -#include -#include -#include -#include -#include -#include -#include -#include - -static unsigned short int is_replay_record; -static unsigned int replay_record; -static unsigned int replay_record_cnt; -static char replay_record_path[PATH_MAX]; -static char *replay_record_dir; -static struct dirent **record_list; - -#ifdef AFL_PERSISTENT_REPLAY_ARGPARSE -static char **record_arg = NULL; -#endif // AFL_PERSISTENT_REPLAY_ARGPARSE - -static int select_files(const struct dirent *dirbuf) { - - char fn[4096]; - - if (dirbuf->d_name[0] == '.') { - - return 0; - - } else { - - snprintf(fn, sizeof(fn), "RECORD:%06u", replay_record); - return !!strstr(dirbuf->d_name, fn); - - } - -} - -static int compare_files(const struct dirent **da, const struct dirent **db) { - - unsigned int c1 = 0, c2 = 0; - - sscanf((*da)->d_name, "RECORD:%*u,cnt:%06u", &c1); - sscanf((*db)->d_name, "RECORD:%*u,cnt:%06u", &c2); - - return c1 - c2; - -} - -__attribute__((destructor)) static void __afl_record_replay_destroy(void) { - - for (int i = 0; i < replay_record_cnt; i++) { - - free(record_list[i]); - - } - - free(record_list); - -} - -__attribute__((constructor)) static void __afl_record_replay_init( -#ifdef AFL_PERSISTENT_REPLAY_ARGPARSE - int argc, char **argv -#endif // AFL_PERSISTENT_REPLAY_ARGPARSE -) { - -#ifdef AFL_PERSISTENT_REPLAY_ARGPARSE - char **argp; -#endif // AFL_PERSISTENT_REPLAY_ARGPARSE - - struct stat sb; - - /* caveat: if harness uses @@ and we don't pass it, it will regardless loop - * the number of iterations defined for AFL_LOOP (on the same file)*/ - if (!(is_replay_record = !!getenv("AFL_PERSISTENT_REPLAY"))) { - - // printf("[warning] AFL_PERSISTENT_REPLAY not set.\n"); - return; - - } - - replay_record = atoi(getenv("AFL_PERSISTENT_REPLAY")); - replay_record_dir = getenv("AFL_PERSISTENT_DIR"); - - if (!(stat(replay_record_dir, &sb) == 0 && S_ISDIR(sb.st_mode))) { - - fprintf(stderr, "[error] Can't find the requested record directory!\n"); - is_replay_record = 0; - return; - - } - - replay_record_cnt = scandir(replay_record_dir ? replay_record_dir : "./", - &record_list, select_files, compare_files); - - if (!replay_record_cnt) { - - fprintf(stderr, "[error] Can't find the requested record!\n"); - is_replay_record = 0; - - } - -#ifdef AFL_PERSISTENT_REPLAY_ARGPARSE - argp = argv; - while (*argp) { - - if (!strcmp(*argp, "@@")) { - - record_arg = argp; - *record_arg = replay_record_path; - break; - - } - - ++argp; - - } - -#endif // AFL_PERSISTENT_REPLAY_ARGPARSE - -} - -/* only used if explictly included for compatibility - compiling without afl-cc */ - -#ifdef AFL_COMPAT - - #ifndef PATH_MAX - #define PATH_MAX 4096 - #endif - - #define FUZZ_BUF_SIZE 1024000 - - // extern ssize_t read(int fildes, void *buf, size_t nbyte); - - // extern int __afl_persistent_loop(unsigned int max_cnt); - // extern unsigned char fuzz_buf[]; - - #ifndef __AFL_HAVE_MANUAL_CONTROL - #define __AFL_HAVE_MANUAL_CONTROL - #endif - - #define __AFL_FUZZ_TESTCASE_LEN (read(0, fuzz_buf, FUZZ_BUF_SIZE)) - #define __AFL_FUZZ_TESTCASE_BUF fuzz_buf - #define __AFL_FUZZ_INIT() void sync(void); - #define __AFL_INIT() sync() - #define __AFL_LOOP(x) __afl_persistent_loop(x) - -unsigned char fuzz_buf[FUZZ_BUF_SIZE]; - -int __afl_persistent_loop(unsigned int max_cnt) { - - static unsigned int cycle_cnt = 1; - static unsigned short int inited = 0; - char tcase[PATH_MAX]; - - if (is_replay_record) { - - if (!inited) { - - cycle_cnt = replay_record_cnt; - inited = 1; - - } - - snprintf(tcase, PATH_MAX, "%s/%s", - replay_record_dir ? replay_record_dir : "./", - record_list[replay_record_cnt - cycle_cnt]->d_name); - - #ifdef AFL_PERSISTENT_REPLAY_ARGPARSE - if (record_arg) { - - *record_arg = tcase; - - } else - - #endif // AFL_PERSISTENT_REPLAY_ARGPARSE - { - - int fd = open(tcase, O_RDONLY); - dup2(fd, 0); - close(fd); - - } - - } else { - - if (!inited) { - - cycle_cnt = max_cnt; - inited = 1; - - } - - } - - return cycle_cnt--; - -} - -#endif // AFL_COMPAT - -#endif // _HAVE_PERSISTENT_REPLAY_H - diff --git a/instrumentation/README.persistent_mode.md b/instrumentation/README.persistent_mode.md index b5d982b0..8e4f6ae4 100644 --- a/instrumentation/README.persistent_mode.md +++ b/instrumentation/README.persistent_mode.md @@ -201,26 +201,28 @@ And that is all! If your software under test requires keeping a state between persistent loop iterations (i.e., a stateful network stack), you can use the `AFL_PERSISTENT_RECORD` variable as described in the [environment variables documentation](../docs/env_variables.md). -To easily replay a crashing, or hanging record, you can use the persistent replay functionality by compiling AFL++ after uncommenting the `AFL_PERSISTENT_REPLAY` define in [config.h](../include/config.h). - -You can then run the test binary specifying the record number via the AFL_PERSISTENT_REPLAY environment variable (i.e., `RECORD:XXXXX`` -> `AFL_PERSISTENT_REPLAY=XXXXX`). +When `AFL_PERSISTENT_RECORD` is enabled, replay functionality is also included in the compiler-rt library. To replay a specific record, assign the record number to the AFL_PERSISTENT_REPLAY environment variable (i.e., `RECORD:XXXXX`` -> `AFL_PERSISTENT_REPLAY=XXXXX`), and run the test binary as you would normally do. The directory where the record files live can be specified via the `AFL_PERSISTENT_DIR` environment varilable, otherwise by default it will be considered the current directory (`./`). -If your harness reads the input files from arguments using the special `@@` argument you will need to define `AFL_PERSISTENT_ARGPARSE` in `config.h`, or before including the `persistent_replay.h` header file as show before. +If your harness reads the input files from arguments using the special `@@` argument you will need to include support by enabling `AFL_PERSISTENT_ARGPARSE` in `config.h`. + In order to offer transparent support to harnesses using the `@@` command line argument, arguments are parsed by the `__afl_record_replay_init` init function. Since not all systems support passing arguments to initializers, this functionality is disabled by default, it's recommendable to use the `__AFL_FUZZ_TESTCASE_BUF/__AFL_FUZZ_TESTCASE_LEN` shared memory mechanism instead. -### 7) Drop in replay functionality +## 7) Drop-in persistent loop replay replacement -To use the replay functionality without having to use `afl-cc` you can just define `AFL_COMPAT` and include the [include/persistent_replay.h](../include/persistent_replay.h) self contained header file that provides a drop-in replacement for the persistent loop mechanism. +To use the replay functionality without having to use `afl-cc`, include the [include/record_compat.h](../include/afl-record_compat.h) header file. Together with the [include/afl-persistent-replay.h](../include/afl-persistent-replay.h) header included in it, `afl-record-compat.h` provides a drop-in replacement for the persistent loop mechanism. ```c #ifndef __AFL_FUZZ_TESTCASE_LEN - #define AFL_COMPAT // #define AFL_PERSISTENT_REPLAY_ARGPARSE - #include "persistent_replay.h" + #include "afl-record-compat.h" #endif __AFL_FUZZ_INIT(); ``` -A simple example is provided in [persistent_demo_new.c](../utils/persistent_mode/persistent_demo_new.c). \ No newline at end of file +A simple example is provided in [persistent_demo_replay.c](../utils/replay_record/persistent_demo_replay.c). + +Be aware that the [afl-record-compat.h](../include/afl-record-compat.h) header should only be included in a single compilation unit, or you will end up with clobbered functions and variables. + +If you need a cleaner solution, you'll have to move the functions and variables defined in [include/record_compat.h](../include/afl-record-compat.h) and [include/afl-persistent-replay.h](../include/afl-persistent-replay.h) in a C file, and add the relevant declarations to a header file. After including the new header file, the compilation unit resulting from compiling the C file can then be linked with your program. \ No newline at end of file diff --git a/instrumentation/afl-compiler-rt.o.c b/instrumentation/afl-compiler-rt.o.c index 037caaf0..4c5d4e79 100644 --- a/instrumentation/afl-compiler-rt.o.c +++ b/instrumentation/afl-compiler-rt.o.c @@ -83,8 +83,8 @@ #include #include -#ifdef AFL_PERSISTENT_REPLAY - #include "persistent_replay.h" +#ifdef AFL_PERSISTENT_RECORD + #include "afl-persistent-replay.h" #endif /* Globals needed by the injected instrumentation. The __afl_area_initial region @@ -1342,68 +1342,73 @@ int __afl_persistent_loop(unsigned int max_cnt) { static u8 first_pass = 1; static u32 cycle_cnt; -#ifdef AFL_PERSISTENT_REPLAY +#ifdef AFL_PERSISTENT_RECORD + char tcase[PATH_MAX]; +#endif - #ifndef PATH_MAX - #define PATH_MAX 4096 - #endif + if (first_pass) { - static u8 inited = 0; - char tcase[PATH_MAX]; + /* Make sure that every iteration of __AFL_LOOP() starts with a clean slate. + On subsequent calls, the parent will take care of that, but on the first + iteration, it's our job to erase any trace of whatever happened + before the loop. */ - if (unlikely(is_replay_record)) { + memset(__afl_area_ptr, 0, __afl_map_size); + __afl_area_ptr[0] = 1; + memset(__afl_prev_loc, 0, NGRAM_SIZE_MAX * sizeof(PREV_LOC_T)); - if (!inited) { + first_pass = 0; + __afl_selective_coverage_temp = 1; + +#ifdef AFL_PERSISTENT_RECORD + if (unlikely(is_replay_record)) { cycle_cnt = replay_record_cnt; - inited = 1; + goto persistent_record; - } + } else - snprintf(tcase, PATH_MAX, "%s/%s", - replay_record_dir ? replay_record_dir : "./", - record_list[replay_record_cnt - cycle_cnt]->d_name); +#endif + { - #ifdef AFL_PERSISTENT_REPLAY_ARGPARSE - if (record_arg) { + cycle_cnt = max_cnt; - *record_arg = tcase; + } - } else + return 1; - #endif // AFL_PERSISTENT_REPLAY_ARGPARSE - { + } else if (--cycle_cnt) { - int fd = open(tcase, O_RDONLY); - dup2(fd, 0); - close(fd); +#ifdef AFL_PERSISTENT_RECORD + if (unlikely(is_replay_record)) { - } + persistent_record: - return cycle_cnt--; + snprintf(tcase, PATH_MAX, "%s/%s", + replay_record_dir ? replay_record_dir : "./", + record_list[replay_record_cnt - cycle_cnt]->d_name); - } else + #ifdef AFL_PERSISTENT_REPLAY_ARGPARSE + if (unlikely(record_arg)) { -#endif + *record_arg = tcase; - if (first_pass) { + } else - /* Make sure that every iteration of __AFL_LOOP() starts with a clean slate. - On subsequent calls, the parent will take care of that, but on the first - iteration, it's our job to erase any trace of whatever happened - before the loop. */ + #endif // AFL_PERSISTENT_REPLAY_ARGPARSE + { - memset(__afl_area_ptr, 0, __afl_map_size); - __afl_area_ptr[0] = 1; - memset(__afl_prev_loc, 0, NGRAM_SIZE_MAX * sizeof(PREV_LOC_T)); + int fd = open(tcase, O_RDONLY); + dup2(fd, 0); + close(fd); - cycle_cnt = max_cnt; - first_pass = 0; - __afl_selective_coverage_temp = 1; + } - return 1; + return 1; - } else if (--cycle_cnt) { + } + +#endif raise(SIGSTOP); diff --git a/utils/persistent_mode/Makefile b/utils/persistent_mode/Makefile index 64de82a7..498aa3f8 100644 --- a/utils/persistent_mode/Makefile +++ b/utils/persistent_mode/Makefile @@ -1,7 +1,6 @@ all: ../../afl-clang-fast -o persistent_demo persistent_demo.c ../../afl-clang-fast -o persistent_demo_new persistent_demo_new.c - gcc -g -I ../../include -o persistent_demo_new_compat persistent_demo_new.c AFL_DONT_OPTIMIZE=1 ../../afl-clang-fast -o test-instr test-instr.c document: diff --git a/utils/replay_record/Makefile b/utils/replay_record/Makefile new file mode 100644 index 00000000..0d1cba92 --- /dev/null +++ b/utils/replay_record/Makefile @@ -0,0 +1,8 @@ +all: + test `grep '//[\s\t ]*#define[\s\t ]*AFL_PERSISTENT_RECORD' ../../include/config.h | wc -l` -eq 0 || (echo "AFL_PERSISTENT_RECORD must be enabled in config.h"; exit 1) + ../../afl-clang-fast -o persistent_demo_replay persistent_demo_replay.c + ${CC} -I ../../include -o persistent_demo_replay_compat persistent_demo_replay.c + ${CC} -g -I ../../include -DAFL_PERSISTENT_REPLAY_ARGPARSE -o persistent_demo_replay_argparse persistent_demo_replay.c + +clean: + rm -f persistent_demo_replay persistent_demo_replay_argparse persistent_demo_replay_compat diff --git a/utils/replay_record/persistent_demo_replay.c b/utils/replay_record/persistent_demo_replay.c new file mode 100644 index 00000000..6f6648f1 --- /dev/null +++ b/utils/replay_record/persistent_demo_replay.c @@ -0,0 +1,148 @@ +/* + american fuzzy lop++ - persistent mode example + -------------------------------------------- + + Originally written by Michal Zalewski + + Copyright 2015 Google Inc. All rights reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at: + + http://www.apache.org/licenses/LICENSE-2.0 + + This file demonstrates the high-performance "persistent mode" that may be + suitable for fuzzing certain fast and well-behaved libraries, provided that + they are stateless or that their internal state can be easily reset + across runs. + + To make this work, the library and this shim need to be compiled in LLVM + mode using afl-clang-fast (other compiler wrappers will *not* work). + + */ + +#include +#include +#include +#include +#include +#include + +#ifdef AFL_PERSISTENT_REPLAY_ARGPARSE + #include + #include +#endif + +/* this lets the source compile without afl-clang-fast/lto */ +#ifndef __AFL_FUZZ_TESTCASE_LEN + #include "afl-record-compat.h" +#endif + +__AFL_FUZZ_INIT(); + +/* Main entry point. */ + +/* To ensure checks are not optimized out it is recommended to disable + code optimization for the fuzzer harness main() */ +#pragma clang optimize off +#pragma GCC optimize("O0") + +int main(int argc, char **argv) { + + ssize_t len; /* how much input did we read? */ + unsigned char *buf; /* test case buffer pointer */ + +#ifdef AFL_PERSISTENT_REPLAY_ARGPARSE + int fd; + + if (argc < 2) { printf("Need an input file!"); } +#endif + + /* The number passed to __AFL_LOOP() controls the maximum number of + iterations before the loop exits and the program is allowed to + terminate normally. This limits the impact of accidental memory leaks + and similar hiccups. */ + + __AFL_INIT(); + +#ifdef AFL_PERSISTENT_REPLAY_ARGPARSE + buf = malloc(1000); +#else + buf = __AFL_FUZZ_TESTCASE_BUF; // this must be assigned before __AFL_LOOP! +#endif + + while (__AFL_LOOP(UINT_MAX)) { // increase if you have good stability + +#ifdef AFL_PERSISTENT_REPLAY_ARGPARSE + fd = open(argv[1], O_RDONLY); + len = read(fd, buf, 1000); + close(fd); +#else + len = __AFL_FUZZ_TESTCASE_LEN; // do not use the macro directly in a call! +#endif + + // fprintf(stderr, "input: %zd \"%s\"\n", len, buf); + + /* do we have enough data? */ + if (len < 8) continue; + + if (strcmp((char *)buf, "thisisateststring") == 0) printf("teststring\n"); + + if (buf[0] == 'f') { + + printf("one\n"); + if (buf[1] == 'o') { + + printf("two\n"); + if (buf[2] == 'o') { + + printf("three\n"); + if (buf[3] == '!') { + + printf("four\n"); + if (buf[4] == '!') { + + printf("five\n"); + if (buf[5] == '!') { + + printf("six\n"); + abort(); + + } else { + + if (buf[5] == 'O') { + + // hang + while (1) { + + continue; + + }; + + } + + } + + } + + } + + } + + } + + } + + /*** END PLACEHOLDER CODE ***/ + + } + + /* Once the loop is exited, terminate normally - AFL will restart the process + when this happens, with a clean slate when it comes to allocated memory, + leftover file descriptors, etc. */ + + return 0; + +} + -- cgit 1.4.1 From 038fef962c3d85fe7e37fcd8717270654f927881 Mon Sep 17 00:00:00 2001 From: vanhauser-thc Date: Thu, 8 Feb 2024 12:46:00 +0100 Subject: performance --- .gitignore | 1 + GNUmakefile | 59 +- docs/INSTALL.md | 5 +- include/config.h | 2 +- include/t1ha.h | 719 +++ include/t1ha0_ia32aes_b.h | 167 + include/t1ha_bits.h | 1254 +++++ include/t1ha_selfcheck.h | 76 + include/xxhash.h | 11013 ++++++++++++++++++++++++-------------------- src/afl-fuzz.c | 4 + src/afl-performance.c | 9 + utils/bench/Makefile | 8 + utils/bench/README.md | 2 + utils/bench/hash.c | 42 + 14 files changed, 8215 insertions(+), 5146 deletions(-) create mode 100644 include/t1ha.h create mode 100644 include/t1ha0_ia32aes_b.h create mode 100644 include/t1ha_bits.h create mode 100644 include/t1ha_selfcheck.h create mode 100644 utils/bench/Makefile create mode 100644 utils/bench/README.md create mode 100644 utils/bench/hash.c (limited to 'include/config.h') diff --git a/.gitignore b/.gitignore index 67feb240..8e191e29 100644 --- a/.gitignore +++ b/.gitignore @@ -99,6 +99,7 @@ unicorn_mode/samples/*/\.test-* utils/afl_network_proxy/afl-network-client utils/afl_network_proxy/afl-network-server utils/afl_proxy/afl-proxy +utils/bench/hash utils/optimin/build utils/optimin/optimin utils/persistent_mode/persistent_demo diff --git a/GNUmakefile b/GNUmakefile index 283c57c2..64dfc37f 100644 --- a/GNUmakefile +++ b/GNUmakefile @@ -84,21 +84,27 @@ else endif endif -#ifeq "$(shell echo 'int main() {return 0; }' | $(CC) -fno-move-loop-invariants -fdisable-tree-cunrolli -x c - -o .test 2>/dev/null && echo 1 || echo 0 ; rm -f .test )" "1" -# SPECIAL_PERFORMANCE += -fno-move-loop-invariants -fdisable-tree-cunrolli -#endif - -#ifeq "$(shell echo 'int main() {return 0; }' | $(CC) $(CFLAGS) -Werror -x c - -march=native -o .test 2>/dev/null && echo 1 || echo 0 ; rm -f .test )" "1" -# ifndef SOURCE_DATE_EPOCH -# HAVE_MARCHNATIVE = 1 -# CFLAGS_OPT += -march=native -# endif -#endif +ifdef PERFORMANCE + SPECIAL_PERFORMANCE := -D_AFL_SPECIAL_PERFORMANCE + ifeq "$(SYS)" "Linux" + ifeq "$(shell grep avx2 /proc/cpuinfo)" "" + else + SPECIAL_PERFORMANCE += -mavx2 -D_HAVE_AVX2 + endif + endif + ifeq "$(shell echo 'int main() {return 0; }' | $(CC) $(CFLAGS) -Werror -x c - -march=native -o .test 2>/dev/null && echo 1 || echo 0 ; rm -f .test )" "1" + HAVE_MARCHNATIVE = 1 + SPECIAL_PERFORMANCE += -march=native + endif + $(info SPECIAL_PERFORMANCE=$(SPECIAL_PERFORMANCE)) +else + SPECIAL_PERFORMANCE := +endif ifneq "$(SYS)" "Darwin" - #ifeq "$(HAVE_MARCHNATIVE)" "1" - # SPECIAL_PERFORMANCE += -march=native - #endif + #ifeq "$(HAVE_MARCHNATIVE)" "1" + # SPECIAL_PERFORMANCE += -march=native + #endif #ifndef DEBUG # CFLAGS_OPT += -D_FORTIFY_SOURCE=1 #endif @@ -389,6 +395,7 @@ help: @echo @echo Known build environment options: @echo "==========================================" + @echo "PERFORMANCE - compile with performance options that make the binary not transferable to other systems. Recommended!" @echo STATIC - compile AFL++ static @echo "CODE_COVERAGE - compile the target for code coverage (see docs/instrumentation/README.llvm.md)" @echo ASAN_BUILD - compiles AFL++ with memory sanitizer for debug purposes @@ -453,31 +460,31 @@ afl-as: src/afl-as.c include/afl-as.h $(COMM_HDR) | test_x86 @ln -sf afl-as as src/afl-performance.o : $(COMM_HDR) src/afl-performance.c include/hash.h - $(CC) $(CFLAGS) $(CFLAGS_OPT) -Iinclude -c src/afl-performance.c -o src/afl-performance.o + $(CC) $(CFLAGS) $(CFLAGS_OPT) $(SPECIAL_PERFORMANCE) -Iinclude -c src/afl-performance.c -o src/afl-performance.o src/afl-common.o : $(COMM_HDR) src/afl-common.c include/common.h - $(CC) $(CFLAGS) $(CFLAGS_FLTO) -c src/afl-common.c -o src/afl-common.o + $(CC) $(CFLAGS) $(CFLAGS_FLTO) $(SPECIAL_PERFORMANCE) -c src/afl-common.c -o src/afl-common.o src/afl-forkserver.o : $(COMM_HDR) src/afl-forkserver.c include/forkserver.h - $(CC) $(CFLAGS) $(CFLAGS_FLTO) -c src/afl-forkserver.c -o src/afl-forkserver.o + $(CC) $(CFLAGS) $(CFLAGS_FLTO) $(SPECIAL_PERFORMANCE) -c src/afl-forkserver.c -o src/afl-forkserver.o src/afl-sharedmem.o : $(COMM_HDR) src/afl-sharedmem.c include/sharedmem.h - $(CC) $(CFLAGS) $(CFLAGS_FLTO) -c src/afl-sharedmem.c -o src/afl-sharedmem.o + $(CC) $(CFLAGS) $(CFLAGS_FLTO) $(SPECIAL_PERFORMANCE) -c src/afl-sharedmem.c -o src/afl-sharedmem.o afl-fuzz: $(COMM_HDR) include/afl-fuzz.h $(AFL_FUZZ_FILES) src/afl-common.o src/afl-sharedmem.o src/afl-forkserver.o src/afl-performance.o | test_x86 - $(CC) $(CFLAGS) $(COMPILE_STATIC) $(CFLAGS_FLTO) $(AFL_FUZZ_FILES) src/afl-common.o src/afl-sharedmem.o src/afl-forkserver.o src/afl-performance.o -o $@ $(PYFLAGS) $(LDFLAGS) -lm + $(CC) $(CFLAGS) $(COMPILE_STATIC) $(CFLAGS_FLTO) $(SPECIAL_PERFORMANCE) $(AFL_FUZZ_FILES) src/afl-common.o src/afl-sharedmem.o src/afl-forkserver.o src/afl-performance.o -o $@ $(PYFLAGS) $(LDFLAGS) -lm afl-showmap: src/afl-showmap.c src/afl-common.o src/afl-sharedmem.o src/afl-forkserver.o src/afl-performance.o $(COMM_HDR) | test_x86 - $(CC) $(CFLAGS) $(COMPILE_STATIC) $(CFLAGS_FLTO) src/$@.c src/afl-fuzz-mutators.c src/afl-fuzz-python.c src/afl-common.o src/afl-sharedmem.o src/afl-forkserver.o src/afl-performance.o -o $@ $(PYFLAGS) $(LDFLAGS) + $(CC) $(CFLAGS) $(COMPILE_STATIC) $(CFLAGS_FLTO) $(SPECIAL_PERFORMANCE) src/$@.c src/afl-fuzz-mutators.c src/afl-fuzz-python.c src/afl-common.o src/afl-sharedmem.o src/afl-forkserver.o src/afl-performance.o -o $@ $(PYFLAGS) $(LDFLAGS) afl-tmin: src/afl-tmin.c src/afl-common.o src/afl-sharedmem.o src/afl-forkserver.o src/afl-performance.o $(COMM_HDR) | test_x86 - $(CC) $(CFLAGS) $(COMPILE_STATIC) $(CFLAGS_FLTO) src/$@.c src/afl-common.o src/afl-sharedmem.o src/afl-forkserver.o src/afl-performance.o -o $@ $(LDFLAGS) + $(CC) $(CFLAGS) $(COMPILE_STATIC) $(CFLAGS_FLTO) $(SPECIAL_PERFORMANCE) src/$@.c src/afl-common.o src/afl-sharedmem.o src/afl-forkserver.o src/afl-performance.o -o $@ $(LDFLAGS) afl-analyze: src/afl-analyze.c src/afl-common.o src/afl-sharedmem.o src/afl-performance.o src/afl-forkserver.o $(COMM_HDR) | test_x86 - $(CC) $(CFLAGS) $(COMPILE_STATIC) $(CFLAGS_FLTO) src/$@.c src/afl-common.o src/afl-sharedmem.o src/afl-performance.o src/afl-forkserver.o -o $@ $(LDFLAGS) + $(CC) $(CFLAGS) $(COMPILE_STATIC) $(CFLAGS_FLTO) $(SPECIAL_PERFORMANCE) src/$@.c src/afl-common.o src/afl-sharedmem.o src/afl-performance.o src/afl-forkserver.o -o $@ $(LDFLAGS) afl-gotcpu: src/afl-gotcpu.c src/afl-common.o $(COMM_HDR) | test_x86 - $(CC) $(CFLAGS) $(COMPILE_STATIC) $(CFLAGS_FLTO) src/$@.c src/afl-common.o -o $@ $(LDFLAGS) + $(CC) $(CFLAGS) $(COMPILE_STATIC) $(CFLAGS_FLTO) $(SPECIAL_PERFORMANCE) src/$@.c src/afl-common.o -o $@ $(LDFLAGS) .PHONY: document document: afl-fuzz-document @@ -494,17 +501,17 @@ unit_maybe_alloc: test/unittests/unit_maybe_alloc.o ./test/unittests/unit_maybe_alloc test/unittests/unit_hash.o : $(COMM_HDR) include/alloc-inl.h test/unittests/unit_hash.c $(AFL_FUZZ_FILES) src/afl-performance.o - @$(CC) $(CFLAGS) $(ASAN_CFLAGS) -c test/unittests/unit_hash.c -o test/unittests/unit_hash.o + @$(CC) $(CFLAGS) $(ASAN_CFLAGS) $(SPECIAL_PERFORMANCE) -c test/unittests/unit_hash.c -o test/unittests/unit_hash.o unit_hash: test/unittests/unit_hash.o src/afl-performance.o - @$(CC) $(CFLAGS) -Wl,--wrap=exit -Wl,--wrap=printf $^ -o test/unittests/unit_hash $(LDFLAGS) $(ASAN_LDFLAGS) -lcmocka + @$(CC) $(CFLAGS) $(SPECIAL_PERFORMANCE) -Wl,--wrap=exit -Wl,--wrap=printf $^ -o test/unittests/unit_hash $(LDFLAGS) $(ASAN_LDFLAGS) -lcmocka ./test/unittests/unit_hash test/unittests/unit_rand.o : $(COMM_HDR) include/alloc-inl.h test/unittests/unit_rand.c $(AFL_FUZZ_FILES) src/afl-performance.o - @$(CC) $(CFLAGS) $(ASAN_CFLAGS) -c test/unittests/unit_rand.c -o test/unittests/unit_rand.o + @$(CC) $(CFLAGS) $(ASAN_CFLAGS) $(SPECIAL_PERFORMANCE) -c test/unittests/unit_rand.c -o test/unittests/unit_rand.o unit_rand: test/unittests/unit_rand.o src/afl-common.o src/afl-performance.o - @$(CC) $(CFLAGS) $(ASAN_CFLAGS) -Wl,--wrap=exit -Wl,--wrap=printf $^ -o test/unittests/unit_rand $(LDFLAGS) $(ASAN_LDFLAGS) -lcmocka + @$(CC) $(CFLAGS) $(ASAN_CFLAGS) $(SPECIAL_PERFORMANCE) -Wl,--wrap=exit -Wl,--wrap=printf $^ -o test/unittests/unit_rand $(LDFLAGS) $(ASAN_LDFLAGS) -lcmocka ./test/unittests/unit_rand test/unittests/unit_list.o : $(COMM_HDR) include/list.h test/unittests/unit_list.c $(AFL_FUZZ_FILES) diff --git a/docs/INSTALL.md b/docs/INSTALL.md index 84bbe3ea..9f53afed 100644 --- a/docs/INSTALL.md +++ b/docs/INSTALL.md @@ -69,14 +69,15 @@ These build targets exist: [Unless you are on Mac OS X](https://developer.apple.com/library/archive/qa/qa1118/_index.html), you can also build statically linked versions of the AFL++ binaries by passing -the `STATIC=1` argument to make: +the `PERFORMANCE=1` argument to make: ```shell -make STATIC=1 +make PERFORMANCE=1 ``` These build options exist: +* PERFORMANCE - compile with performance options that make the binary not transferable to other systems. Recommended! * STATIC - compile AFL++ static * CODE_COVERAGE - compile the target for code coverage (see docs/instrumentation/README.llvm.md) * ASAN_BUILD - compiles AFL++ with memory sanitizer for debug purposes diff --git a/include/config.h b/include/config.h index 70ce2ae3..31d66b14 100644 --- a/include/config.h +++ b/include/config.h @@ -26,7 +26,7 @@ /* Version string: */ // c = release, a = volatile github dev, e = experimental branch -#define VERSION "++4.10c" +#define VERSION "++4.20a" /****************************************************** * * diff --git a/include/t1ha.h b/include/t1ha.h new file mode 100644 index 00000000..498f0dd6 --- /dev/null +++ b/include/t1ha.h @@ -0,0 +1,719 @@ +/* + * Copyright (c) 2016-2020 Positive Technologies, https://www.ptsecurity.com, + * Fast Positive Hash. + * + * Portions Copyright (c) 2010-2020 Leonid Yuriev , + * The 1Hippeus project (t1h). + * + * This software is provided 'as-is', without any express or implied + * warranty. In no event will the authors be held liable for any damages + * arising from the use of this software. + * + * Permission is granted to anyone to use this software for any purpose, + * including commercial applications, and to alter it and redistribute it + * freely, subject to the following restrictions: + * + * 1. The origin of this software must not be misrepresented; you must not + * claim that you wrote the original software. If you use this software + * in a product, an acknowledgement in the product documentation would be + * appreciated but is not required. + * 2. Altered source versions must be plainly marked as such, and must not be + * misrepresented as being the original software. + * 3. This notice may not be removed or altered from any source distribution. + */ + +/* + * t1ha = { Fast Positive Hash, aka "Позитивный Хэш" } + * by [Positive Technologies](https://www.ptsecurity.ru) + * + * Briefly, it is a 64-bit Hash Function: + * 1. Created for 64-bit little-endian platforms, in predominantly for x86_64, + * but portable and without penalties it can run on any 64-bit CPU. + * 2. In most cases up to 15% faster than City64, xxHash, mum-hash, metro-hash + * and all others portable hash-functions (which do not use specific + * hardware tricks). + * 3. Not suitable for cryptography. + * + * The Future will (be) Positive. Всё будет хорошо. + * + * ACKNOWLEDGEMENT: + * The t1ha was originally developed by Leonid Yuriev (Леонид Юрьев) + * for The 1Hippeus project - zerocopy messaging in the spirit of Sparta! + */ + +#pragma once + +/***************************************************************************** + * + * PLEASE PAY ATTENTION TO THE FOLLOWING NOTES + * about macros definitions which controls t1ha behaviour and/or performance. + * + * + * 1) T1HA_SYS_UNALIGNED_ACCESS = Defines the system/platform/CPU/architecture + * abilities for unaligned data access. + * + * By default, when the T1HA_SYS_UNALIGNED_ACCESS not defined, + * it will defined on the basis hardcoded knowledge about of capabilities + * of most common CPU architectures. But you could override this + * default behavior when build t1ha library itself: + * + * // To disable unaligned access at all. + * #define T1HA_SYS_UNALIGNED_ACCESS 0 + * + * // To enable unaligned access, but indicate that it significantly slow. + * #define T1HA_SYS_UNALIGNED_ACCESS 1 + * + * // To enable unaligned access, and indicate that it effecient. + * #define T1HA_SYS_UNALIGNED_ACCESS 2 + * + * + * 2) T1HA_USE_FAST_ONESHOT_READ = Controls the data reads at the end of buffer. + * + * When defined to non-zero, t1ha will use 'one shot' method for reading + * up to 8 bytes at the end of data. In this case just the one 64-bit read + * will be performed even when the available less than 8 bytes. + * + * This is little bit faster that switching by length of data tail. + * Unfortunately this will triggering a false-positive alarms from Valgrind, + * AddressSanitizer and other similar tool. + * + * By default, t1ha defines it to 1, but you could override this + * default behavior when build t1ha library itself: + * + * // For little bit faster and small code. + * #define T1HA_USE_FAST_ONESHOT_READ 1 + * + * // For calmness if doubt. + * #define T1HA_USE_FAST_ONESHOT_READ 0 + * + * + * 3) T1HA0_RUNTIME_SELECT = Controls choice fastest function in runtime. + * + * t1ha library offers the t1ha0() function as the fastest for current CPU. + * But actual CPU's features/capabilities and may be significantly different, + * especially on x86 platform. Therefore, internally, t1ha0() may require + * dynamic dispatching for choice best implementation. + * + * By default, t1ha enables such runtime choice and (may be) corresponding + * indirect calls if it reasonable, but you could override this default + * behavior when build t1ha library itself: + * + * // To enable runtime choice of fastest implementation. + * #define T1HA0_RUNTIME_SELECT 1 + * + * // To disable runtime choice of fastest implementation. + * #define T1HA0_RUNTIME_SELECT 0 + * + * When T1HA0_RUNTIME_SELECT is nonzero the t1ha0_resolve() function could + * be used to get actual t1ha0() implementation address at runtime. This is + * useful for two cases: + * - calling by local pointer-to-function usually is little + * bit faster (less overhead) than via a PLT thru the DSO boundary. + * - GNU Indirect functions (see below) don't supported by environment + * and calling by t1ha0_funcptr is not available and/or expensive. + * + * 4) T1HA_USE_INDIRECT_FUNCTIONS = Controls usage of GNU Indirect functions. + * + * In continue of T1HA0_RUNTIME_SELECT the T1HA_USE_INDIRECT_FUNCTIONS + * controls usage of ELF indirect functions feature. In general, when + * available, this reduces overhead of indirect function's calls though + * a DSO-bundary (https://sourceware.org/glibc/wiki/GNU_IFUNC). + * + * By default, t1ha engage GNU Indirect functions when it available + * and useful, but you could override this default behavior when build + * t1ha library itself: + * + * // To enable use of GNU ELF Indirect functions. + * #define T1HA_USE_INDIRECT_FUNCTIONS 1 + * + * // To disable use of GNU ELF Indirect functions. This may be useful + * // if the actual toolchain or the system's loader don't support ones. + * #define T1HA_USE_INDIRECT_FUNCTIONS 0 + * + * 5) T1HA0_AESNI_AVAILABLE = Controls AES-NI detection and dispatching on x86. + * + * In continue of T1HA0_RUNTIME_SELECT the T1HA0_AESNI_AVAILABLE controls + * detection and usage of AES-NI CPU's feature. On the other hand, this + * requires compiling parts of t1ha library with certain properly options, + * and could be difficult or inconvenient in some cases. + * + * By default, t1ha engade AES-NI for t1ha0() on the x86 platform, but + * you could override this default behavior when build t1ha library itself: + * + * // To disable detection and usage of AES-NI instructions for t1ha0(). + * // This may be useful when you unable to build t1ha library properly + * // or known that AES-NI will be unavailable at the deploy. + * #define T1HA0_AESNI_AVAILABLE 0 + * + * // To force detection and usage of AES-NI instructions for t1ha0(), + * // but I don't known reasons to anybody would need this. + * #define T1HA0_AESNI_AVAILABLE 1 + * + * 6) T1HA0_DISABLED, T1HA1_DISABLED, T1HA2_DISABLED = Controls availability of + * t1ha functions. + * + * In some cases could be useful to import/use only few of t1ha functions + * or just the one. So, this definitions allows disable corresponding parts + * of t1ha library. + * + * // To disable t1ha0(), t1ha0_32le(), t1ha0_32be() and all AES-NI. + * #define T1HA0_DISABLED + * + * // To disable t1ha1_le() and t1ha1_be(). + * #define T1HA1_DISABLED + * + * // To disable t1ha2_atonce(), t1ha2_atonce128() and so on. + * #define T1HA2_DISABLED + * + *****************************************************************************/ + +#define T1HA_VERSION_MAJOR 2 +#define T1HA_VERSION_MINOR 1 +#define T1HA_VERSION_RELEASE 1 + +#ifndef __has_attribute +#define __has_attribute(x) (0) +#endif + +#ifndef __has_include +#define __has_include(x) (0) +#endif + +#ifndef __GNUC_PREREQ +#if defined(__GNUC__) && defined(__GNUC_MINOR__) +#define __GNUC_PREREQ(maj, min) \ + ((__GNUC__ << 16) + __GNUC_MINOR__ >= ((maj) << 16) + (min)) +#else +#define __GNUC_PREREQ(maj, min) 0 +#endif +#endif /* __GNUC_PREREQ */ + +#ifndef __CLANG_PREREQ +#ifdef __clang__ +#define __CLANG_PREREQ(maj, min) \ + ((__clang_major__ << 16) + __clang_minor__ >= ((maj) << 16) + (min)) +#else +#define __CLANG_PREREQ(maj, min) (0) +#endif +#endif /* __CLANG_PREREQ */ + +#ifndef __LCC_PREREQ +#ifdef __LCC__ +#define __LCC_PREREQ(maj, min) \ + ((__LCC__ << 16) + __LCC_MINOR__ >= ((maj) << 16) + (min)) +#else +#define __LCC_PREREQ(maj, min) (0) +#endif +#endif /* __LCC_PREREQ */ + +/*****************************************************************************/ + +#ifdef _MSC_VER +/* Avoid '16' bytes padding added after data member 't1ha_context::total' + * and other warnings from std-headers if warning-level > 3. */ +#pragma warning(push, 3) +#endif + +#if defined(__cplusplus) && __cplusplus >= 201103L +#include +#include +#include +#else +#include +#include +#include +#endif + +/*****************************************************************************/ + +#if defined(i386) || defined(__386) || defined(__i386) || defined(__i386__) || \ + defined(i486) || defined(__i486) || defined(__i486__) || \ + defined(i586) | defined(__i586) || defined(__i586__) || defined(i686) || \ + defined(__i686) || defined(__i686__) || defined(_M_IX86) || \ + defined(_X86_) || defined(__THW_INTEL__) || defined(__I86__) || \ + defined(__INTEL__) || defined(__x86_64) || defined(__x86_64__) || \ + defined(__amd64__) || defined(__amd64) || defined(_M_X64) || \ + defined(_M_AMD64) || defined(__IA32__) || defined(__INTEL__) +#ifndef __ia32__ +/* LY: define neutral __ia32__ for x86 and x86-64 archs */ +#define __ia32__ 1 +#endif /* __ia32__ */ +#if !defined(__amd64__) && (defined(__x86_64) || defined(__x86_64__) || \ + defined(__amd64) || defined(_M_X64)) +/* LY: define trusty __amd64__ for all AMD64/x86-64 arch */ +#define __amd64__ 1 +#endif /* __amd64__ */ +#endif /* all x86 */ + +#if !defined(__BYTE_ORDER__) || !defined(__ORDER_LITTLE_ENDIAN__) || \ + !defined(__ORDER_BIG_ENDIAN__) + +/* *INDENT-OFF* */ +/* clang-format off */ + +#if defined(__GLIBC__) || defined(__GNU_LIBRARY__) || defined(__ANDROID__) || \ + defined(HAVE_ENDIAN_H) || __has_include() +#include +#elif defined(__APPLE__) || defined(__MACH__) || defined(__OpenBSD__) || \ + defined(HAVE_MACHINE_ENDIAN_H) || __has_include() +#include +#elif defined(HAVE_SYS_ISA_DEFS_H) || __has_include() +#include +#elif (defined(HAVE_SYS_TYPES_H) && defined(HAVE_SYS_ENDIAN_H)) || \ + (__has_include() && __has_include()) +#include +#include +#elif defined(__bsdi__) || defined(__DragonFly__) || defined(__FreeBSD__) || \ + defined(__NETBSD__) || defined(__NetBSD__) || \ + defined(HAVE_SYS_PARAM_H) || __has_include() +#include +#endif /* OS */ + +/* *INDENT-ON* */ +/* clang-format on */ + +#if defined(__BYTE_ORDER) && defined(__LITTLE_ENDIAN) && defined(__BIG_ENDIAN) +#define __ORDER_LITTLE_ENDIAN__ __LITTLE_ENDIAN +#define __ORDER_BIG_ENDIAN__ __BIG_ENDIAN +#define __BYTE_ORDER__ __BYTE_ORDER +#elif defined(_BYTE_ORDER) && defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN) +#define __ORDER_LITTLE_ENDIAN__ _LITTLE_ENDIAN +#define __ORDER_BIG_ENDIAN__ _BIG_ENDIAN +#define __BYTE_ORDER__ _BYTE_ORDER +#else +#define __ORDER_LITTLE_ENDIAN__ 1234 +#define __ORDER_BIG_ENDIAN__ 4321 + +#if defined(__LITTLE_ENDIAN__) || \ + (defined(_LITTLE_ENDIAN) && !defined(_BIG_ENDIAN)) || \ + defined(__ARMEL__) || defined(__THUMBEL__) || defined(__AARCH64EL__) || \ + defined(__MIPSEL__) || defined(_MIPSEL) || defined(__MIPSEL) || \ + defined(_M_ARM) || defined(_M_ARM64) || defined(__e2k__) || \ + defined(__elbrus_4c__) || defined(__elbrus_8c__) || defined(__bfin__) || \ + defined(__BFIN__) || defined(__ia64__) || defined(_IA64) || \ + defined(__IA64__) || defined(__ia64) || defined(_M_IA64) || \ + defined(__itanium__) || defined(__ia32__) || defined(__CYGWIN__) || \ + defined(_WIN64) || defined(_WIN32) || defined(__TOS_WIN__) || \ + defined(__WINDOWS__) +#define __BYTE_ORDER__ __ORDER_LITTLE_ENDIAN__ + +#elif defined(__BIG_ENDIAN__) || \ + (defined(_BIG_ENDIAN) && !defined(_LITTLE_ENDIAN)) || \ + defined(__ARMEB__) || defined(__THUMBEB__) || defined(__AARCH64EB__) || \ + defined(__MIPSEB__) || defined(_MIPSEB) || defined(__MIPSEB) || \ + defined(__m68k__) || defined(M68000) || defined(__hppa__) || \ + defined(__hppa) || defined(__HPPA__) || defined(__sparc__) || \ + defined(__sparc) || defined(__370__) || defined(__THW_370__) || \ + defined(__s390__) || defined(__s390x__) || defined(__SYSC_ZARCH__) +#define __BYTE_ORDER__ __ORDER_BIG_ENDIAN__ + +#else +#error __BYTE_ORDER__ should be defined. +#endif /* Arch */ + +#endif +#endif /* __BYTE_ORDER__ || __ORDER_LITTLE_ENDIAN__ || __ORDER_BIG_ENDIAN__ */ + +/*****************************************************************************/ + +#ifndef __dll_export +#if defined(_WIN32) || defined(_WIN64) || defined(__CYGWIN__) +#if defined(__GNUC__) || __has_attribute(dllexport) +#define __dll_export __attribute__((dllexport)) +#else +#define __dll_export __declspec(dllexport) +#endif +#elif defined(__GNUC__) || __has_attribute(__visibility__) +#define __dll_export __attribute__((__visibility__("default"))) +#else +#define __dll_export +#endif +#endif /* __dll_export */ + +#ifndef __dll_import +#if defined(_WIN32) || defined(_WIN64) || defined(__CYGWIN__) +#if defined(__GNUC__) || __has_attribute(dllimport) +#define __dll_import __attribute__((dllimport)) +#else +#define __dll_import __declspec(dllimport) +#endif +#elif defined(__GNUC__) || __has_attribute(__visibility__) +#define __dll_import __attribute__((__visibility__("default"))) +#else +#define __dll_import +#endif +#endif /* __dll_import */ + +#ifndef __force_inline +#ifdef _MSC_VER +#define __force_inline __forceinline +#elif __GNUC_PREREQ(3, 2) || __has_attribute(__always_inline__) +#define __force_inline __inline __attribute__((__always_inline__)) +#else +#define __force_inline __inline +#endif +#endif /* __force_inline */ + +#ifndef T1HA_API +#if defined(t1ha_EXPORTS) +#define T1HA_API __dll_export +#elif defined(t1ha_IMPORTS) +#define T1HA_API __dll_import +#else +#define T1HA_API +#endif +#endif /* T1HA_API */ + +#if defined(_MSC_VER) && defined(__ia32__) +#define T1HA_ALIGN_PREFIX __declspec(align(32)) /* required only for SIMD */ +#else +#define T1HA_ALIGN_PREFIX +#endif /* _MSC_VER */ + +#if defined(__GNUC__) && defined(__ia32__) +#define T1HA_ALIGN_SUFFIX \ + __attribute__((__aligned__(32))) /* required only for SIMD */ +#else +#define T1HA_ALIGN_SUFFIX +#endif /* GCC x86 */ + +#ifndef T1HA_USE_INDIRECT_FUNCTIONS +/* GNU ELF indirect functions usage control. For more info please see + * https://en.wikipedia.org/wiki/Executable_and_Linkable_Format + * and https://sourceware.org/glibc/wiki/GNU_IFUNC */ +#if defined(__ELF__) && defined(__amd64__) && \ + (__has_attribute(__ifunc__) || \ + (!defined(__clang__) && defined(__GNUC__) && __GNUC__ >= 4 && \ + !defined(__SANITIZE_ADDRESS__) && !defined(__SSP_ALL__))) +/* Enable gnu_indirect_function by default if : + * - ELF AND x86_64 + * - attribute(__ifunc__) is available OR + * GCC >= 4 WITHOUT -fsanitize=address NOR -fstack-protector-all */ +#define T1HA_USE_INDIRECT_FUNCTIONS 1 +#else +#define T1HA_USE_INDIRECT_FUNCTIONS 0 +#endif +#endif /* T1HA_USE_INDIRECT_FUNCTIONS */ + +#if __GNUC_PREREQ(4, 0) +#pragma GCC visibility push(hidden) +#endif /* __GNUC_PREREQ(4,0) */ + +#ifdef __cplusplus +extern "C" { +#endif + +typedef union T1HA_ALIGN_PREFIX t1ha_state256 { + uint8_t bytes[32]; + uint32_t u32[8]; + uint64_t u64[4]; + struct { + uint64_t a, b, c, d; + } n; +} t1ha_state256_t T1HA_ALIGN_SUFFIX; + +typedef struct t1ha_context { + t1ha_state256_t state; + t1ha_state256_t buffer; + size_t partial; + uint64_t total; +} t1ha_context_t; + +#ifdef _MSC_VER +#pragma warning(pop) +#endif + +/****************************************************************************** + * + * Self-testing API. + * + * Unfortunately, some compilers (exactly only Microsoft Visual C/C++) has + * a bugs which leads t1ha-functions to produce wrong results. This API allows + * check the correctness of the actual code in runtime. + * + * All check-functions returns 0 on success, or -1 in case the corresponding + * hash-function failed verification. PLEASE, always perform such checking at + * initialization of your code, if you using MSVC or other troubleful compilers. + */ + +T1HA_API int t1ha_selfcheck__all_enabled(void); + +#ifndef T1HA2_DISABLED +T1HA_API int t1ha_selfcheck__t1ha2_atonce(void); +T1HA_API int t1ha_selfcheck__t1ha2_atonce128(void); +T1HA_API int t1ha_selfcheck__t1ha2_stream(void); +T1HA_API int t1ha_selfcheck__t1ha2(void); +#endif /* T1HA2_DISABLED */ + +#ifndef T1HA1_DISABLED +T1HA_API int t1ha_selfcheck__t1ha1_le(void); +T1HA_API int t1ha_selfcheck__t1ha1_be(void); +T1HA_API int t1ha_selfcheck__t1ha1(void); +#endif /* T1HA1_DISABLED */ + +#ifndef T1HA0_DISABLED +T1HA_API int t1ha_selfcheck__t1ha0_32le(void); +T1HA_API int t1ha_selfcheck__t1ha0_32be(void); +T1HA_API int t1ha_selfcheck__t1ha0(void); + +/* Define T1HA0_AESNI_AVAILABLE to 0 for disable AES-NI support. */ +#ifndef T1HA0_AESNI_AVAILABLE +#if defined(__e2k__) || \ + (defined(__ia32__) && (!defined(_M_IX86) || _MSC_VER > 1800)) +#define T1HA0_AESNI_AVAILABLE 1 +#else +#define T1HA0_AESNI_AVAILABLE 0 +#endif +#endif /* ifndef T1HA0_AESNI_AVAILABLE */ + +#if T1HA0_AESNI_AVAILABLE +T1HA_API int t1ha_selfcheck__t1ha0_ia32aes_noavx(void); +T1HA_API int t1ha_selfcheck__t1ha0_ia32aes_avx(void); +#ifndef __e2k__ +T1HA_API int t1ha_selfcheck__t1ha0_ia32aes_avx2(void); +#endif +#endif /* if T1HA0_AESNI_AVAILABLE */ +#endif /* T1HA0_DISABLED */ + +/****************************************************************************** + * + * t1ha2 = 64 and 128-bit, SLIGHTLY MORE ATTENTION FOR QUALITY AND STRENGTH. + * + * - The recommended version of "Fast Positive Hash" with good quality + * for checksum, hash tables and fingerprinting. + * - Portable and extremely efficiency on modern 64-bit CPUs. + * Designed for 64-bit little-endian platforms, + * in other cases will runs slowly. + * - Great quality of hashing and still faster than other non-t1ha hashes. + * Provides streaming mode and 128-bit result. + * + * Note: Due performance reason 64- and 128-bit results are completely + * different each other, i.e. 64-bit result is NOT any part of 128-bit. + */ +#ifndef T1HA2_DISABLED + +/* The at-once variant with 64-bit result */ +T1HA_API uint64_t t1ha2_atonce(const void *data, size_t length, uint64_t seed); + +/* The at-once variant with 128-bit result. + * Argument `extra_result` is NOT optional and MUST be valid. + * The high 64-bit part of 128-bit hash will be always unconditionally + * stored to the address given by `extra_result` argument. */ +T1HA_API uint64_t t1ha2_atonce128(uint64_t *__restrict extra_result, + const void *__restrict data, size_t length, + uint64_t seed); + +/* The init/update/final trinity for streaming. + * Return 64 or 128-bit result depentently from `extra_result` argument. */ +T1HA_API void t1ha2_init(t1ha_context_t *ctx, uint64_t seed_x, uint64_t seed_y); +T1HA_API void t1ha2_update(t1ha_context_t *__restrict ctx, + const void *__restrict data, size_t length); + +/* Argument `extra_result` is optional and MAY be NULL. + * - If `extra_result` is NOT NULL then the 128-bit hash will be calculated, + * and high 64-bit part of it will be stored to the address given + * by `extra_result` argument. + * - Otherwise the 64-bit hash will be calculated + * and returned from function directly. + * + * Note: Due performance reason 64- and 128-bit results are completely + * different each other, i.e. 64-bit result is NOT any part of 128-bit. */ +T1HA_API uint64_t t1ha2_final(t1ha_context_t *__restrict ctx, + uint64_t *__restrict extra_result /* optional */); + +#endif /* T1HA2_DISABLED */ + +/****************************************************************************** + * + * t1ha1 = 64-bit, BASELINE FAST PORTABLE HASH: + * + * - Runs faster on 64-bit platforms in other cases may runs slowly. + * - Portable and stable, returns same 64-bit result + * on all architectures and CPUs. + * - Unfortunately it fails the "strict avalanche criteria", + * see test results at https://github.com/demerphq/smhasher. + * + * This flaw is insignificant for the t1ha1() purposes and imperceptible + * from a practical point of view. + * However, nowadays this issue has resolved in the next t1ha2(), + * that was initially planned to providing a bit more quality. + */ +#ifndef T1HA1_DISABLED + +/* The little-endian variant. */ +T1HA_API uint64_t t1ha1_le(const void *data, size_t length, uint64_t seed); + +/* The big-endian variant. */ +T1HA_API uint64_t t1ha1_be(const void *data, size_t length, uint64_t seed); + +#endif /* T1HA1_DISABLED */ + +/****************************************************************************** + * + * t1ha0 = 64-bit, JUST ONLY FASTER: + * + * - Provides fast-as-possible hashing for current CPU, including + * 32-bit systems and engaging the available hardware acceleration. + * - It is a facade that selects most quick-and-dirty hash + * for the current processor. For instance, on IA32 (x86) actual function + * will be selected in runtime, depending on current CPU capabilities + * + * BE CAREFUL!!! THIS IS MEANS: + * + * 1. The quality of hash is a subject for tradeoffs with performance. + * So, the quality and strength of t1ha0() may be lower than t1ha1(), + * especially on 32-bit targets, but then much faster. + * However, guaranteed that it passes all SMHasher tests. + * + * 2. No warranty that the hash result will be same for particular + * key on another machine or another version of libt1ha. + * + * Briefly, such hash-results and their derivatives, should be + * used only in runtime, but should not be persist or transferred + * over a network. + * + * + * When T1HA0_RUNTIME_SELECT is nonzero the t1ha0_resolve() function could + * be used to get actual t1ha0() implementation address at runtime. This is + * useful for two cases: + * - calling by local pointer-to-function usually is little + * bit faster (less overhead) than via a PLT thru the DSO boundary. + * - GNU Indirect functions (see below) don't supported by environment + * and calling by t1ha0_funcptr is not available and/or expensive. + */ + +#ifndef T1HA0_DISABLED + +/* The little-endian variant for 32-bit CPU. */ +uint64_t t1ha0_32le(const void *data, size_t length, uint64_t seed); +/* The big-endian variant for 32-bit CPU. */ +uint64_t t1ha0_32be(const void *data, size_t length, uint64_t seed); + +/* Define T1HA0_AESNI_AVAILABLE to 0 for disable AES-NI support. */ +#ifndef T1HA0_AESNI_AVAILABLE +#if defined(__e2k__) || \ + (defined(__ia32__) && (!defined(_M_IX86) || _MSC_VER > 1800)) +#define T1HA0_AESNI_AVAILABLE 1 +#else +#define T1HA0_AESNI_AVAILABLE 0 +#endif +#endif /* T1HA0_AESNI_AVAILABLE */ + +/* Define T1HA0_RUNTIME_SELECT to 0 for disable dispatching t1ha0 at runtime. */ +#ifndef T1HA0_RUNTIME_SELECT +#if T1HA0_AESNI_AVAILABLE && !defined(__e2k__) +#define T1HA0_RUNTIME_SELECT 1 +#else +#define T1HA0_RUNTIME_SELECT 0 +#endif +#endif /* T1HA0_RUNTIME_SELECT */ + +#if !T1HA0_RUNTIME_SELECT && !defined(T1HA0_USE_DEFINE) +#if defined(__LCC__) +#define T1HA0_USE_DEFINE 1 +#else +#define T1HA0_USE_DEFINE 0 +#endif +#endif /* T1HA0_USE_DEFINE */ + +#if T1HA0_AESNI_AVAILABLE +uint64_t t1ha0_ia32aes_noavx(const void *data, size_t length, uint64_t seed); +uint64_t t1ha0_ia32aes_avx(const void *data, size_t length, uint64_t seed); +#ifndef __e2k__ +uint64_t t1ha0_ia32aes_avx2(const void *data, size_t length, uint64_t seed); +#endif +#endif /* T1HA0_AESNI_AVAILABLE */ + +#if T1HA0_RUNTIME_SELECT +typedef uint64_t (*t1ha0_function_t)(const void *, size_t, uint64_t); +T1HA_API t1ha0_function_t t1ha0_resolve(void); +#if T1HA_USE_INDIRECT_FUNCTIONS +T1HA_API uint64_t t1ha0(const void *data, size_t length, uint64_t seed); +#else +/* Otherwise function pointer will be used. + * Unfortunately this may cause some overhead calling. */ +T1HA_API extern uint64_t (*t1ha0_funcptr)(const void *data, size_t length, + uint64_t seed); +static __force_inline uint64_t t1ha0(const void *data, size_t length, + uint64_t seed) { + return t1ha0_funcptr(data, length, seed); +} +#endif /* T1HA_USE_INDIRECT_FUNCTIONS */ + +#elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ + +#if T1HA0_USE_DEFINE + +#if (UINTPTR_MAX > 0xffffFFFFul || ULONG_MAX > 0xffffFFFFul) && \ + (!defined(T1HA1_DISABLED) || !defined(T1HA2_DISABLED)) +#if defined(T1HA1_DISABLED) +#define t1ha0 t1ha2_atonce +#else +#define t1ha0 t1ha1_be +#endif /* T1HA1_DISABLED */ +#else /* 32/64 */ +#define t1ha0 t1ha0_32be +#endif /* 32/64 */ + +#else /* T1HA0_USE_DEFINE */ + +static __force_inline uint64_t t1ha0(const void *data, size_t length, + uint64_t seed) { +#if (UINTPTR_MAX > 0xffffFFFFul || ULONG_MAX > 0xffffFFFFul) && \ + (!defined(T1HA1_DISABLED) || !defined(T1HA2_DISABLED)) +#if defined(T1HA1_DISABLED) + return t1ha2_atonce(data, length, seed); +#else + return t1ha1_be(data, length, seed); +#endif /* T1HA1_DISABLED */ +#else /* 32/64 */ + return t1ha0_32be(data, length, seed); +#endif /* 32/64 */ +} + +#endif /* !T1HA0_USE_DEFINE */ + +#else /* !T1HA0_RUNTIME_SELECT && __BYTE_ORDER__ != __ORDER_BIG_ENDIAN__ */ + +#if T1HA0_USE_DEFINE + +#if (UINTPTR_MAX > 0xffffFFFFul || ULONG_MAX > 0xffffFFFFul) && \ + (!defined(T1HA1_DISABLED) || !defined(T1HA2_DISABLED)) +#if defined(T1HA1_DISABLED) +#define t1ha0 t1ha2_atonce +#else +#define t1ha0 t1ha1_le +#endif /* T1HA1_DISABLED */ +#else /* 32/64 */ +#define t1ha0 t1ha0_32le +#endif /* 32/64 */ + +#else + +static __force_inline uint64_t t1ha0(const void *data, size_t length, + uint64_t seed) { +#if (UINTPTR_MAX > 0xffffFFFFul || ULONG_MAX > 0xffffFFFFul) && \ + (!defined(T1HA1_DISABLED) || !defined(T1HA2_DISABLED)) +#if defined(T1HA1_DISABLED) + return t1ha2_atonce(data, length, seed); +#else + return t1ha1_le(data, length, seed); +#endif /* T1HA1_DISABLED */ +#else /* 32/64 */ + return t1ha0_32le(data, length, seed); +#endif /* 32/64 */ +} + +#endif /* !T1HA0_USE_DEFINE */ + +#endif /* !T1HA0_RUNTIME_SELECT */ + +#endif /* T1HA0_DISABLED */ + +#ifdef __cplusplus +} +#endif + +#if __GNUC_PREREQ(4, 0) +#pragma GCC visibility pop +#endif /* __GNUC_PREREQ(4,0) */ diff --git a/include/t1ha0_ia32aes_b.h b/include/t1ha0_ia32aes_b.h new file mode 100644 index 00000000..e8e52638 --- /dev/null +++ b/include/t1ha0_ia32aes_b.h @@ -0,0 +1,167 @@ +/* + * Copyright (c) 2016-2020 Positive Technologies, https://www.ptsecurity.com, + * Fast Positive Hash. + * + * Portions Copyright (c) 2010-2020 Leonid Yuriev , + * The 1Hippeus project (t1h). + * + * This software is provided 'as-is', without any express or implied + * warranty. In no event will the authors be held liable for any damages + * arising from the use of this software. + * + * Permission is granted to anyone to use this software for any purpose, + * including commercial applications, and to alter it and redistribute it + * freely, subject to the following restrictions: + * + * 1. The origin of this software must not be misrepresented; you must not + * claim that you wrote the original software. If you use this software + * in a product, an acknowledgement in the product documentation would be + * appreciated but is not required. + * 2. Altered source versions must be plainly marked as such, and must not be + * misrepresented as being the original software. + * 3. This notice may not be removed or altered from any source distribution. + */ + +/* + * t1ha = { Fast Positive Hash, aka "Позитивный Хэш" } + * by [Positive Technologies](https://www.ptsecurity.ru) + * + * Briefly, it is a 64-bit Hash Function: + * 1. Created for 64-bit little-endian platforms, in predominantly for x86_64, + * but portable and without penalties it can run on any 64-bit CPU. + * 2. In most cases up to 15% faster than City64, xxHash, mum-hash, metro-hash + * and all others portable hash-functions (which do not use specific + * hardware tricks). + * 3. Not suitable for cryptography. + * + * The Future will (be) Positive. Всё будет хорошо. + * + * ACKNOWLEDGEMENT: + * The t1ha was originally developed by Leonid Yuriev (Леонид Юрьев) + * for The 1Hippeus project - zerocopy messaging in the spirit of Sparta! + */ + +#include "t1ha_bits.h" +#include "t1ha_selfcheck.h" + +#if T1HA0_AESNI_AVAILABLE + +uint64_t T1HA_IA32AES_NAME(const void *data, uint32_t len) { + uint64_t a = 0; + uint64_t b = len; + + if (likely(len > 32)) { + __m128i x = _mm_set_epi64x(a, b); + __m128i y = _mm_aesenc_si128(x, _mm_set_epi64x(prime_0, prime_1)); + + const __m128i *v = (const __m128i *)data; + const __m128i *const detent = + (const __m128i *)((const uint8_t *)data + (len & ~15ul)); + data = detent; + + if (len & 16) { + x = _mm_add_epi64(x, _mm_loadu_si128(v++)); + y = _mm_aesenc_si128(x, y); + } + len &= 15; + + if (v + 7 < detent) { + __m128i salt = y; + do { + __m128i t = _mm_aesenc_si128(_mm_loadu_si128(v++), salt); + t = _mm_aesdec_si128(t, _mm_loadu_si128(v++)); + t = _mm_aesdec_si128(t, _mm_loadu_si128(v++)); + t = _mm_aesdec_si128(t, _mm_loadu_si128(v++)); + + t = _mm_aesdec_si128(t, _mm_loadu_si128(v++)); + t = _mm_aesdec_si128(t, _mm_loadu_si128(v++)); + t = _mm_aesdec_si128(t, _mm_loadu_si128(v++)); + t = _mm_aesdec_si128(t, _mm_loadu_si128(v++)); + + salt = _mm_add_epi64(salt, _mm_set_epi64x(prime_5, prime_6)); + t = _mm_aesenc_si128(x, t); + x = _mm_add_epi64(y, x); + y = t; + } while (v + 7 < detent); + } + + while (v < detent) { + __m128i v0y = _mm_add_epi64(y, _mm_loadu_si128(v++)); + __m128i v1x = _mm_sub_epi64(x, _mm_loadu_si128(v++)); + x = _mm_aesdec_si128(x, v0y); + y = _mm_aesdec_si128(y, v1x); + } + + x = _mm_add_epi64(_mm_aesdec_si128(x, _mm_aesenc_si128(y, x)), y); +#if defined(__x86_64__) || defined(_M_X64) +#if defined(__SSE4_1__) || defined(__AVX__) + a = _mm_extract_epi64(x, 0); + b = _mm_extract_epi64(x, 1); +#else + a = _mm_cvtsi128_si64(x); + b = _mm_cvtsi128_si64(_mm_unpackhi_epi64(x, x)); +#endif +#else +#if defined(__SSE4_1__) || defined(__AVX__) + a = (uint32_t)_mm_extract_epi32(x, 0) | (uint64_t)_mm_extract_epi32(x, 1) + << 32; + b = (uint32_t)_mm_extract_epi32(x, 2) | (uint64_t)_mm_extract_epi32(x, 3) + << 32; +#else + a = (uint32_t)_mm_cvtsi128_si32(x); + a |= (uint64_t)_mm_cvtsi128_si32(_mm_shuffle_epi32(x, 1)) << 32; + x = _mm_unpackhi_epi64(x, x); + b = (uint32_t)_mm_cvtsi128_si32(x); + b |= (uint64_t)_mm_cvtsi128_si32(_mm_shuffle_epi32(x, 1)) << 32; +#endif +#endif +#ifdef __AVX__ + _mm256_zeroupper(); +#elif !(defined(_X86_64_) || defined(__x86_64__) || defined(_M_X64) || \ + defined(__e2k__)) + _mm_empty(); +#endif + } + + const uint64_t *v = (const uint64_t *)data; + switch (len) { + default: + mixup64(&a, &b, fetch64_le_unaligned(v++), prime_4); + /* fall through */ + case 24: + case 23: + case 22: + case 21: + case 20: + case 19: + case 18: + case 17: + mixup64(&b, &a, fetch64_le_unaligned(v++), prime_3); + /* fall through */ + case 16: + case 15: + case 14: + case 13: + case 12: + case 11: + case 10: + case 9: + mixup64(&a, &b, fetch64_le_unaligned(v++), prime_2); + /* fall through */ + case 8: + case 7: + case 6: + case 5: + case 4: + case 3: + case 2: + case 1: + mixup64(&b, &a, tail64_le_unaligned(v, len), prime_1); + /* fall through */ + case 0: + return final64(a, b); + } +} + +#endif /* T1HA0_AESNI_AVAILABLE */ +#undef T1HA_IA32AES_NAME diff --git a/include/t1ha_bits.h b/include/t1ha_bits.h new file mode 100644 index 00000000..539369aa --- /dev/null +++ b/include/t1ha_bits.h @@ -0,0 +1,1254 @@ +/* + * Copyright (c) 2016-2020 Positive Technologies, https://www.ptsecurity.com, + * Fast Positive Hash. + * + * Portions Copyright (c) 2010-2020 Leonid Yuriev , + * The 1Hippeus project (t1h). + * + * This software is provided 'as-is', without any express or implied + * warranty. In no event will the authors be held liable for any damages + * arising from the use of this software. + * + * Permission is granted to anyone to use this software for any purpose, + * including commercial applications, and to alter it and redistribute it + * freely, subject to the following restrictions: + * + * 1. The origin of this software must not be misrepresented; you must not + * claim that you wrote the original software. If you use this software + * in a product, an acknowledgement in the product documentation would be + * appreciated but is not required. + * 2. Altered source versions must be plainly marked as such, and must not be + * misrepresented as being the original software. + * 3. This notice may not be removed or altered from any source distribution. + */ + +/* + * t1ha = { Fast Positive Hash, aka "Позитивный Хэш" } + * by [Positive Technologies](https://www.ptsecurity.ru) + * + * Briefly, it is a 64-bit Hash Function: + * 1. Created for 64-bit little-endian platforms, in predominantly for x86_64, + * but portable and without penalties it can run on any 64-bit CPU. + * 2. In most cases up to 15% faster than City64, xxHash, mum-hash, metro-hash + * and all others portable hash-functions (which do not use specific + * hardware tricks). + * 3. Not suitable for cryptography. + * + * The Future will (be) Positive. Всё будет хорошо. + * + * ACKNOWLEDGEMENT: + * The t1ha was originally developed by Leonid Yuriev (Леонид Юрьев) + * for The 1Hippeus project - zerocopy messaging in the spirit of Sparta! + */ + +#pragma once + +#if defined(_MSC_VER) +#pragma warning(disable : 4201) /* nameless struct/union */ +#if _MSC_VER > 1800 +#pragma warning(disable : 4464) /* relative include path contains '..' */ +#endif /* 1800 */ +#endif /* MSVC */ +#include "t1ha.h" + +#ifndef T1HA_USE_FAST_ONESHOT_READ +/* Define it to 1 for little bit faster code. + * Unfortunately this may triggering a false-positive alarms from Valgrind, + * AddressSanitizer and other similar tool. + * So, define it to 0 for calmness if doubt. */ +#define T1HA_USE_FAST_ONESHOT_READ 1 +#endif /* T1HA_USE_FAST_ONESHOT_READ */ + +/*****************************************************************************/ + +#include /* for assert() */ +#include /* for bool */ +#include /* for memcpy() */ + +#if __BYTE_ORDER__ != __ORDER_LITTLE_ENDIAN__ && \ + __BYTE_ORDER__ != __ORDER_BIG_ENDIAN__ +#error Unsupported byte order. +#endif + +#define T1HA_UNALIGNED_ACCESS__UNABLE 0 +#define T1HA_UNALIGNED_ACCESS__SLOW 1 +#define T1HA_UNALIGNED_ACCESS__EFFICIENT 2 + +#ifndef T1HA_SYS_UNALIGNED_ACCESS +#if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) +#define T1HA_SYS_UNALIGNED_ACCESS T1HA_UNALIGNED_ACCESS__EFFICIENT +#elif defined(__ia32__) +#define T1HA_SYS_UNALIGNED_ACCESS T1HA_UNALIGNED_ACCESS__EFFICIENT +#elif defined(__e2k__) +#define T1HA_SYS_UNALIGNED_ACCESS T1HA_UNALIGNED_ACCESS__SLOW +#elif defined(__ARM_FEATURE_UNALIGNED) +#define T1HA_SYS_UNALIGNED_ACCESS T1HA_UNALIGNED_ACCESS__EFFICIENT +#else +#define T1HA_SYS_UNALIGNED_ACCESS T1HA_UNALIGNED_ACCESS__UNABLE +#endif +#endif /* T1HA_SYS_UNALIGNED_ACCESS */ + +#define ALIGNMENT_16 2 +#define ALIGNMENT_32 4 +#if UINTPTR_MAX > 0xffffFFFFul || ULONG_MAX > 0xffffFFFFul +#define ALIGNMENT_64 8 +#else +#define ALIGNMENT_64 4 +#endif + +#ifndef PAGESIZE +#define PAGESIZE 4096 +#endif /* PAGESIZE */ + +/***************************************************************************/ + +#ifndef __has_builtin +#define __has_builtin(x) (0) +#endif + +#ifndef __has_warning +#define __has_warning(x) (0) +#endif + +#ifndef __has_feature +#define __has_feature(x) (0) +#endif + +#ifndef __has_extension +#define __has_extension(x) (0) +#endif + +#if __has_feature(address_sanitizer) +#define __SANITIZE_ADDRESS__ 1 +#endif + +#ifndef __optimize +#if defined(__clang__) && !__has_attribute(__optimize__) +#define __optimize(ops) +#elif defined(__GNUC__) || __has_attribute(__optimize__) +#define __optimize(ops) __attribute__((__optimize__(ops))) +#else +#define __optimize(ops) +#endif +#endif /* __optimize */ + +#ifndef __cold +#if defined(__OPTIMIZE__) +#if defined(__e2k__) +#define __cold __optimize(1) __attribute__((__cold__)) +#elif defined(__clang__) && !__has_attribute(__cold__) && \ + __has_attribute(__section__) +/* just put infrequently used functions in separate section */ +#define __cold __attribute__((__section__("text.unlikely"))) __optimize("Os") +#elif defined(__GNUC__) || __has_attribute(__cold__) +#define __cold __attribute__((__cold__)) __optimize("Os") +#else +#define __cold __optimize("Os") +#endif +#else +#define __cold +#endif +#endif /* __cold */ + +#if __GNUC_PREREQ(4, 4) || defined(__clang__) + +#if defined(__ia32__) || defined(__e2k__) +#include +#endif + +#if defined(__ia32__) && !defined(__cpuid_count) +#include +#endif + +#if defined(__e2k__) +#include +#endif + +#ifndef likely +#define likely(cond) __builtin_expect(!!(cond), 1) +#endif + +#ifndef unlikely +#define unlikely(cond) __builtin_expect(!!(cond), 0) +#endif + +#if __GNUC_PREREQ(4, 5) || __has_builtin(__builtin_unreachable) +#define unreachable() __builtin_unreachable() +#endif + +#define bswap64(v) __builtin_bswap64(v) +#define bswap32(v) __builtin_bswap32(v) +#if __GNUC_PREREQ(4, 8) || __has_builtin(__builtin_bswap16) +#define bswap16(v) __builtin_bswap16(v) +#endif + +#if !defined(__maybe_unused) && \ + (__GNUC_PREREQ(4, 3) || __has_attribute(__unused__)) +#define __maybe_unused __attribute__((__unused__)) +#endif + +#if !defined(__always_inline) && \ + (__GNUC_PREREQ(3, 2) || __has_attribute(__always_inline__)) +#define __always_inline __inline __attribute__((__always_inline__)) +#endif + +#if defined(__e2k__) + +#if __iset__ >= 3 +#define mul_64x64_high(a, b) __builtin_e2k_umulhd(a, b) +#endif /* __iset__ >= 3 */ + +#if __iset__ >= 5 +static __maybe_unused __always_inline unsigned +e2k_add64carry_first(uint64_t base, uint64_t addend, uint64_t *sum) { + *sum = base + addend; + return (unsigned)__builtin_e2k_addcd_c(base, addend, 0); +} +#define add64carry_first(base, addend, sum) \ + e2k_add64carry_first(base, addend, sum) + +static __maybe_unused __always_inline unsigned +e2k_add64carry_next(unsigned carry, uint64_t base, uint64_t addend, + uint64_t *sum) { + *sum = __builtin_e2k_addcd(base, addend, carry); + return (unsigned)__builtin_e2k_addcd_c(base, addend, carry); +} +#define add64carry_next(carry, base, addend, sum) \ + e2k_add64carry_next(carry, base, addend, sum) + +static __maybe_unused __always_inline void e2k_add64carry_last(unsigned carry, + uint64_t base, + uint64_t addend, + uint64_t *sum) { + *sum = __builtin_e2k_addcd(base, addend, carry); +} +#define add64carry_last(carry, base, addend, sum) \ + e2k_add64carry_last(carry, base, addend, sum) +#endif /* __iset__ >= 5 */ + +#define fetch64_be_aligned(ptr) ((uint64_t)__builtin_e2k_ld_64s_be(ptr)) +#define fetch32_be_aligned(ptr) ((uint32_t)__builtin_e2k_ld_32u_be(ptr)) + +#endif /* __e2k__ Elbrus */ + +#elif defined(_MSC_VER) + +#if _MSC_FULL_VER < 190024234 && defined(_M_IX86) +#pragma message( \ + "For AES-NI at least \"Microsoft C/C++ Compiler\" version 19.00.24234 (Visual Studio 2015 Update 3) is required.") +#endif +#if _MSC_FULL_VER < 191526730 +#pragma message( \ + "It is recommended to use \"Microsoft C/C++ Compiler\" version 19.15.26730 (Visual Studio 2017 15.8) or newer.") +#endif +#if _MSC_FULL_VER < 180040629 +#error At least "Microsoft C/C++ Compiler" version 18.00.40629 (Visual Studio 2013 Update 5) is required. +#endif + +#pragma warning(push, 1) + +#include +#include +#define likely(cond) (cond) +#define unlikely(cond) (cond) +#define unreachable() __assume(0) +#define bswap64(v) _byteswap_uint64(v) +#define bswap32(v) _byteswap_ulong(v) +#define bswap16(v) _byteswap_ushort(v) +#define rot64(v, s) _rotr64(v, s) +#define rot32(v, s) _rotr(v, s) +#define __always_inline __forceinline + +#if defined(_M_X64) || defined(_M_IA64) +#pragma intrinsic(_umul128) +#define mul_64x64_128(a, b, ph) _umul128(a, b, ph) +#pragma intrinsic(_addcarry_u64) +#define add64carry_first(base, addend, sum) _addcarry_u64(0, base, addend, sum) +#define add64carry_next(carry, base, addend, sum) \ + _addcarry_u64(carry, base, addend, sum) +#define add64carry_last(carry, base, addend, sum) \ + (void)_addcarry_u64(carry, base, addend, sum) +#endif + +#if defined(_M_ARM64) || defined(_M_X64) || defined(_M_IA64) +#pragma intrinsic(__umulh) +#define mul_64x64_high(a, b) __umulh(a, b) +#endif + +#if defined(_M_IX86) +#pragma intrinsic(__emulu) +#define mul_32x32_64(a, b) __emulu(a, b) + +#if _MSC_VER >= 1915 /* LY: workaround for SSA-optimizer bug */ +#pragma intrinsic(_addcarry_u32) +#define add32carry_first(base, addend, sum) _addcarry_u32(0, base, addend, sum) +#define add32carry_next(carry, base, addend, sum) \ + _addcarry_u32(carry, base, addend, sum) +#define add32carry_last(carry, base, addend, sum) \ + (void)_addcarry_u32(carry, base, addend, sum) + +static __forceinline char +msvc32_add64carry_first(uint64_t base, uint64_t addend, uint64_t *sum) { + uint32_t *const sum32 = (uint32_t *)sum; + const uint32_t base_32l = (uint32_t)base; + const uint32_t base_32h = (uint32_t)(base >> 32); + const uint32_t addend_32l = (uint32_t)addend; + const uint32_t addend_32h = (uint32_t)(addend >> 32); + return add32carry_next(add32carry_first(base_32l, addend_32l, sum32), + base_32h, addend_32h, sum32 + 1); +} +#define add64carry_first(base, addend, sum) \ + msvc32_add64carry_first(base, addend, sum) + +static __forceinline char msvc32_add64carry_next(char carry, uint64_t base, + uint64_t addend, + uint64_t *sum) { + uint32_t *const sum32 = (uint32_t *)sum; + const uint32_t base_32l = (uint32_t)base; + const uint32_t base_32h = (uint32_t)(base >> 32); + const uint32_t addend_32l = (uint32_t)addend; + const uint32_t addend_32h = (uint32_t)(addend >> 32); + return add32carry_next(add32carry_next(carry, base_32l, addend_32l, sum32), + base_32h, addend_32h, sum32 + 1); +} +#define add64carry_next(carry, base, addend, sum) \ + msvc32_add64carry_next(carry, base, addend, sum) + +static __forceinline void msvc32_add64carry_last(char carry, uint64_t base, + uint64_t addend, + uint64_t *sum) { + uint32_t *const sum32 = (uint32_t *)sum; + const uint32_t base_32l = (uint32_t)base; + const uint32_t base_32h = (uint32_t)(base >> 32); + const uint32_t addend_32l = (uint32_t)addend; + const uint32_t addend_32h = (uint32_t)(addend >> 32); + add32carry_last(add32carry_next(carry, base_32l, addend_32l, sum32), base_32h, + addend_32h, sum32 + 1); +} +#define add64carry_last(carry, base, addend, sum) \ + msvc32_add64carry_last(carry, base, addend, sum) +#endif /* _MSC_FULL_VER >= 190024231 */ + +#elif defined(_M_ARM) +#define mul_32x32_64(a, b) _arm_umull(a, b) +#endif + +#pragma warning(pop) +#pragma warning(disable : 4514) /* 'xyz': unreferenced inline function \ + has been removed */ +#pragma warning(disable : 4710) /* 'xyz': function not inlined */ +#pragma warning(disable : 4711) /* function 'xyz' selected for \ + automatic inline expansion */ +#pragma warning(disable : 4127) /* conditional expression is constant */ +#pragma warning(disable : 4702) /* unreachable code */ +#endif /* Compiler */ + +#ifndef likely +#define likely(cond) (cond) +#endif +#ifndef unlikely +#define unlikely(cond) (cond) +#endif +#ifndef __maybe_unused +#define __maybe_unused +#endif +#ifndef __always_inline +#define __always_inline __inline +#endif +#ifndef unreachable +#define unreachable() \ + do { \ + } while (1) +#endif + +#ifndef bswap64 +#if defined(bswap_64) +#define bswap64 bswap_64 +#elif defined(__bswap_64) +#define bswap64 __bswap_64 +#else +static __always_inline uint64_t bswap64(uint64_t v) { + return v << 56 | v >> 56 | ((v << 40) & UINT64_C(0x00ff000000000000)) | + ((v << 24) & UINT64_C(0x0000ff0000000000)) | + ((v << 8) & UINT64_C(0x000000ff00000000)) | + ((v >> 8) & UINT64_C(0x00000000ff000000)) | + ((v >> 24) & UINT64_C(0x0000000000ff0000)) | + ((v >> 40) & UINT64_C(0x000000000000ff00)); +} +#endif +#endif /* bswap64 */ + +#ifndef bswap32 +#if defined(bswap_32) +#define bswap32 bswap_32 +#elif defined(__bswap_32) +#define bswap32 __bswap_32 +#else +static __always_inline uint32_t bswap32(uint32_t v) { + return v << 24 | v >> 24 | ((v << 8) & UINT32_C(0x00ff0000)) | + ((v >> 8) & UINT32_C(0x0000ff00)); +} +#endif +#endif /* bswap32 */ + +#ifndef bswap16 +#if defined(bswap_16) +#define bswap16 bswap_16 +#elif defined(__bswap_16) +#define bswap16 __bswap_16 +#else +static __always_inline uint16_t bswap16(uint16_t v) { return v << 8 | v >> 8; } +#endif +#endif /* bswap16 */ + +#if defined(__ia32__) || \ + T1HA_SYS_UNALIGNED_ACCESS == T1HA_UNALIGNED_ACCESS__EFFICIENT +/* The __builtin_assume_aligned() leads gcc/clang to load values into the + * registers, even when it is possible to directly use an operand from memory. + * This can lead to a shortage of registers and a significant slowdown. + * Therefore avoid unnecessary use of __builtin_assume_aligned() for x86. */ +#define read_unaligned(ptr, bits) (*(const uint##bits##_t *__restrict)(ptr)) +#define read_aligned(ptr, bits) (*(const uint##bits##_t *__restrict)(ptr)) +#endif /* __ia32__ */ + +#ifndef read_unaligned +#if defined(__GNUC__) || __has_attribute(__packed__) +typedef struct { + uint8_t unaligned_8; + uint16_t unaligned_16; + uint32_t unaligned_32; + uint64_t unaligned_64; +} __attribute__((__packed__)) t1ha_unaligned_proxy; +#define read_unaligned(ptr, bits) \ + (((const t1ha_unaligned_proxy *)((const uint8_t *)(ptr)-offsetof( \ + t1ha_unaligned_proxy, unaligned_##bits))) \ + ->unaligned_##bits) +#elif defined(_MSC_VER) +#pragma warning( \ + disable : 4235) /* nonstandard extension used: '__unaligned' \ + * keyword not supported on this architecture */ +#define read_unaligned(ptr, bits) (*(const __unaligned uint##bits##_t *)(ptr)) +#else +#pragma pack(push, 1) +typedef struct { + uint8_t unaligned_8; + uint16_t unaligned_16; + uint32_t unaligned_32; + uint64_t unaligned_64; +} t1ha_unaligned_proxy; +#pragma pack(pop) +#define read_unaligned(ptr, bits) \ + (((const t1ha_unaligned_proxy *)((const uint8_t *)(ptr)-offsetof( \ + t1ha_unaligned_proxy, unaligned_##bits))) \ + ->unaligned_##bits) +#endif +#endif /* read_unaligned */ + +#ifndef read_aligned +#if __GNUC_PREREQ(4, 8) || __has_builtin(__builtin_assume_aligned) +#define read_aligned(ptr, bits) \ + (*(const uint##bits##_t *)__builtin_assume_aligned(ptr, ALIGNMENT_##bits)) +#elif (__GNUC_PREREQ(3, 3) || __has_attribute(__aligned__)) && \ + !defined(__clang__) +#define read_aligned(ptr, bits) \ + (*(const uint##bits##_t \ + __attribute__((__aligned__(ALIGNMENT_##bits))) *)(ptr)) +#elif __has_attribute(__assume_aligned__) + +static __always_inline const + uint16_t *__attribute__((__assume_aligned__(ALIGNMENT_16))) + cast_aligned_16(const void *ptr) { + return (const uint16_t *)ptr; +} +static __always_inline const + uint32_t *__attribute__((__assume_aligned__(ALIGNMENT_32))) + cast_aligned_32(const void *ptr) { + return (const uint32_t *)ptr; +} +static __always_inline const + uint64_t *__attribute__((__assume_aligned__(ALIGNMENT_64))) + cast_aligned_64(const void *ptr) { + return (const uint64_t *)ptr; +} + +#define read_aligned(ptr, bits) (*cast_aligned_##bits(ptr)) + +#elif defined(_MSC_VER) +#define read_aligned(ptr, bits) \ + (*(const __declspec(align(ALIGNMENT_##bits)) uint##bits##_t *)(ptr)) +#else +#define read_aligned(ptr, bits) (*(const uint##bits##_t *)(ptr)) +#endif +#endif /* read_aligned */ + +#ifndef prefetch +#if (__GNUC_PREREQ(4, 0) || __has_builtin(__builtin_prefetch)) && \ + !defined(__ia32__) +#define prefetch(ptr) __builtin_prefetch(ptr) +#elif defined(_M_ARM64) || defined(_M_ARM) +#define prefetch(ptr) __prefetch(ptr) +#else +#define prefetch(ptr) \ + do { \ + (void)(ptr); \ + } while (0) +#endif +#endif /* prefetch */ + +#if __has_warning("-Wconstant-logical-operand") +#if defined(__clang__) +#pragma clang diagnostic ignored "-Wconstant-logical-operand" +#elif defined(__GNUC__) +#pragma GCC diagnostic ignored "-Wconstant-logical-operand" +#else +#pragma warning disable "constant-logical-operand" +#endif +#endif /* -Wconstant-logical-operand */ + +#if __has_warning("-Wtautological-pointer-compare") +#if defined(__clang__) +#pragma clang diagnostic ignored "-Wtautological-pointer-compare" +#elif defined(__GNUC__) +#pragma GCC diagnostic ignored "-Wtautological-pointer-compare" +#else +#pragma warning disable "tautological-pointer-compare" +#endif +#endif /* -Wtautological-pointer-compare */ + +/***************************************************************************/ + +#if __GNUC_PREREQ(4, 0) +#pragma GCC visibility push(hidden) +#endif /* __GNUC_PREREQ(4,0) */ + +/*---------------------------------------------------------- Little Endian */ + +#ifndef fetch16_le_aligned +static __maybe_unused __always_inline uint16_t +fetch16_le_aligned(const void *v) { + assert(((uintptr_t)v) % ALIGNMENT_16 == 0); +#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ + return read_aligned(v, 16); +#else + return bswap16(read_aligned(v, 16)); +#endif +} +#endif /* fetch16_le_aligned */ + +#ifndef fetch16_le_unaligned +static __maybe_unused __always_inline uint16_t +fetch16_le_unaligned(const void *v) { +#if T1HA_SYS_UNALIGNED_ACCESS == T1HA_UNALIGNED_ACCESS__UNABLE + const uint8_t *p = (const uint8_t *)v; + return p[0] | (uint16_t)p[1] << 8; +#elif __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ + return read_unaligned(v, 16); +#else + return bswap16(read_unaligned(v, 16)); +#endif +} +#endif /* fetch16_le_unaligned */ + +#ifndef fetch32_le_aligned +static __maybe_unused __always_inline uint32_t +fetch32_le_aligned(const void *v) { + assert(((uintptr_t)v) % ALIGNMENT_32 == 0); +#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ + return read_aligned(v, 32); +#else + return bswap32(read_aligned(v, 32)); +#endif +} +#endif /* fetch32_le_aligned */ + +#ifndef fetch32_le_unaligned +static __maybe_unused __always_inline uint32_t +fetch32_le_unaligned(const void *v) { +#if T1HA_SYS_UNALIGNED_ACCESS == T1HA_UNALIGNED_ACCESS__UNABLE + return fetch16_le_unaligned(v) | + (uint32_t)fetch16_le_unaligned((const uint8_t *)v + 2) << 16; +#elif __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ + return read_unaligned(v, 32); +#else + return bswap32(read_unaligned(v, 32)); +#endif +} +#endif /* fetch32_le_unaligned */ + +#ifndef fetch64_le_aligned +static __maybe_unused __always_inline uint64_t +fetch64_le_aligned(const void *v) { + assert(((uintptr_t)v) % ALIGNMENT_64 == 0); +#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ + return read_aligned(v, 64); +#else + return bswap64(read_aligned(v, 64)); +#endif +} +#endif /* fetch64_le_aligned */ + +#ifndef fetch64_le_unaligned +static __maybe_unused __always_inline uint64_t +fetch64_le_unaligned(const void *v) { +#if T1HA_SYS_UNALIGNED_ACCESS == T1HA_UNALIGNED_ACCESS__UNABLE + return fetch32_le_unaligned(v) | + (uint64_t)fetch32_le_unaligned((const uint8_t *)v + 4) << 32; +#elif __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ + return read_unaligned(v, 64); +#else + return bswap64(read_unaligned(v, 64)); +#endif +} +#endif /* fetch64_le_unaligned */ + +static __maybe_unused __always_inline uint64_t tail64_le_aligned(const void *v, + size_t tail) { + const uint8_t *const p = (const uint8_t *)v; +#if T1HA_USE_FAST_ONESHOT_READ && !defined(__SANITIZE_ADDRESS__) + /* We can perform a 'oneshot' read, which is little bit faster. */ + const unsigned shift = ((8 - tail) & 7) << 3; + return fetch64_le_aligned(p) & ((~UINT64_C(0)) >> shift); +#else + uint64_t r = 0; + switch (tail & 7) { + default: + unreachable(); +/* fall through */ +#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ + /* For most CPUs this code is better when not needed byte reordering. */ + case 0: + return fetch64_le_aligned(p); + case 7: + r = (uint64_t)p[6] << 8; + /* fall through */ + case 6: + r += p[5]; + r <<= 8; + /* fall through */ + case 5: + r += p[4]; + r <<= 32; + /* fall through */ + case 4: + return r + fetch32_le_aligned(p); + case 3: + r = (uint64_t)p[2] << 16; + /* fall through */ + case 2: + return r + fetch16_le_aligned(p); + case 1: + return p[0]; +#else + case 0: + r = p[7] << 8; + /* fall through */ + case 7: + r += p[6]; + r <<= 8; + /* fall through */ + case 6: + r += p[5]; + r <<= 8; + /* fall through */ + case 5: + r += p[4]; + r <<= 8; + /* fall through */ + case 4: + r += p[3]; + r <<= 8; + /* fall through */ + case 3: + r += p[2]; + r <<= 8; + /* fall through */ + case 2: + r += p[1]; + r <<= 8; + /* fall through */ + case 1: + return r + p[0]; +#endif + } +#endif /* T1HA_USE_FAST_ONESHOT_READ */ +} + +#if T1HA_USE_FAST_ONESHOT_READ && \ + T1HA_SYS_UNALIGNED_ACCESS != T1HA_UNALIGNED_ACCESS__UNABLE && \ + defined(PAGESIZE) && PAGESIZE > 42 && !defined(__SANITIZE_ADDRESS__) +#define can_read_underside(ptr, size) \ + (((PAGESIZE - (size)) & (uintptr_t)(ptr)) != 0) +#endif /* T1HA_USE_FAST_ONESHOT_READ */ + +static __maybe_unused __always_inline uint64_t +tail64_le_unaligned(const void *v, size_t tail) { + const uint8_t *p = (const uint8_t *)v; +#if defined(can_read_underside) && \ + (UINTPTR_MAX > 0xffffFFFFul || ULONG_MAX > 0xffffFFFFul) + /* On some systems (e.g. x86_64) we can perform a 'oneshot' read, which + * is little bit faster. Thanks Marcin Żukowski + * for the reminder. */ + const unsigned offset = (8 - tail) & 7; + const unsigned shift = offset << 3; + if (likely(can_read_underside(p, 8))) { + p -= offset; + return fetch64_le_unaligned(p) >> shift; + } + return fetch64_le_unaligned(p) & ((~UINT64_C(0)) >> shift); +#else + uint64_t r = 0; + switch (tail & 7) { + default: + unreachable(); +/* fall through */ +#if T1HA_SYS_UNALIGNED_ACCESS == T1HA_UNALIGNED_ACCESS__EFFICIENT && \ + __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ + /* For most CPUs this code is better when not needed + * copying for alignment or byte reordering. */ + case 0: + return fetch64_le_unaligned(p); + case 7: + r = (uint64_t)p[6] << 8; + /* fall through */ + case 6: + r += p[5]; + r <<= 8; + /* fall through */ + case 5: + r += p[4]; + r <<= 32; + /* fall through */ + case 4: + return r + fetch32_le_unaligned(p); + case 3: + r = (uint64_t)p[2] << 16; + /* fall through */ + case 2: + return r + fetch16_le_unaligned(p); + case 1: + return p[0]; +#else + /* For most CPUs this code is better than a + * copying for alignment and/or byte reordering. */ + case 0: + r = p[7] << 8; + /* fall through */ + case 7: + r += p[6]; + r <<= 8; + /* fall through */ + case 6: + r += p[5]; + r <<= 8; + /* fall through */ + case 5: + r += p[4]; + r <<= 8; + /* fall through */ + case 4: + r += p[3]; + r <<= 8; + /* fall through */ + case 3: + r += p[2]; + r <<= 8; + /* fall through */ + case 2: + r += p[1]; + r <<= 8; + /* fall through */ + case 1: + return r + p[0]; +#endif + } +#endif /* can_read_underside */ +} + +/*------------------------------------------------------------- Big Endian */ + +#ifndef fetch16_be_aligned +static __maybe_unused __always_inline uint16_t +fetch16_be_aligned(const void *v) { + assert(((uintptr_t)v) % ALIGNMENT_16 == 0); +#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ + return read_aligned(v, 16); +#else + return bswap16(read_aligned(v, 16)); +#endif +} +#endif /* fetch16_be_aligned */ + +#ifndef fetch16_be_unaligned +static __maybe_unused __always_inline uint16_t +fetch16_be_unaligned(const void *v) { +#if T1HA_SYS_UNALIGNED_ACCESS == T1HA_UNALIGNED_ACCESS__UNABLE + const uint8_t *p = (const uint8_t *)v; + return (uint16_t)p[0] << 8 | p[1]; +#elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ + return read_unaligned(v, 16); +#else + return bswap16(read_unaligned(v, 16)); +#endif +} +#endif /* fetch16_be_unaligned */ + +#ifndef fetch32_be_aligned +static __maybe_unused __always_inline uint32_t +fetch32_be_aligned(const void *v) { + assert(((uintptr_t)v) % ALIGNMENT_32 == 0); +#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ + return read_aligned(v, 32); +#else + return bswap32(read_aligned(v, 32)); +#endif +} +#endif /* fetch32_be_aligned */ + +#ifndef fetch32_be_unaligned +static __maybe_unused __always_inline uint32_t +fetch32_be_unaligned(const void *v) { +#if T1HA_SYS_UNALIGNED_ACCESS == T1HA_UNALIGNED_ACCESS__UNABLE + return (uint32_t)fetch16_be_unaligned(v) << 16 | + fetch16_be_unaligned((const uint8_t *)v + 2); +#elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ + return read_unaligned(v, 32); +#else + return bswap32(read_unaligned(v, 32)); +#endif +} +#endif /* fetch32_be_unaligned */ + +#ifndef fetch64_be_aligned +static __maybe_unused __always_inline uint64_t +fetch64_be_aligned(const void *v) { + assert(((uintptr_t)v) % ALIGNMENT_64 == 0); +#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ + return read_aligned(v, 64); +#else + return bswap64(read_aligned(v, 64)); +#endif +} +#endif /* fetch64_be_aligned */ + +#ifndef fetch64_be_unaligned +static __maybe_unused __always_inline uint64_t +fetch64_be_unaligned(const void *v) { +#if T1HA_SYS_UNALIGNED_ACCESS == T1HA_UNALIGNED_ACCESS__UNABLE + return (uint64_t)fetch32_be_unaligned(v) << 32 | + fetch32_be_unaligned((const uint8_t *)v + 4); +#elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ + return read_unaligned(v, 64); +#else + return bswap64(read_unaligned(v, 64)); +#endif +} +#endif /* fetch64_be_unaligned */ + +static __maybe_unused __always_inline uint64_t tail64_be_aligned(const void *v, + size_t tail) { + const uint8_t *const p = (const uint8_t *)v; +#if T1HA_USE_FAST_ONESHOT_READ && !defined(__SANITIZE_ADDRESS__) + /* We can perform a 'oneshot' read, which is little bit faster. */ + const unsigned shift = ((8 - tail) & 7) << 3; + return fetch64_be_aligned(p) >> shift; +#else + switch (tail & 7) { + default: + unreachable(); +/* fall through */ +#if __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ + /* For most CPUs this code is better when not byte reordering. */ + case 1: + return p[0]; + case 2: + return fetch16_be_aligned(p); + case 3: + return (uint32_t)fetch16_be_aligned(p) << 8 | p[2]; + case 4: + return fetch32_be_aligned(p); + case 5: + return (uint64_t)fetch32_be_aligned(p) << 8 | p[4]; + case 6: + return (uint64_t)fetch32_be_aligned(p) << 16 | fetch16_be_aligned(p + 4); + case 7: + return (uint64_t)fetch32_be_aligned(p) << 24 | + (uint32_t)fetch16_be_aligned(p + 4) << 8 | p[6]; + case 0: + return fetch64_be_aligned(p); +#else + case 1: + return p[0]; + case 2: + return p[1] | (uint32_t)p[0] << 8; + case 3: + return p[2] | (uint32_t)p[1] << 8 | (uint32_t)p[0] << 16; + case 4: + return p[3] | (uint32_t)p[2] << 8 | (uint32_t)p[1] << 16 | + (uint32_t)p[0] << 24; + case 5: + return p[4] | (uint32_t)p[3] << 8 | (uint32_t)p[2] << 16 | + (uint32_t)p[1] << 24 | (uint64_t)p[0] << 32; + case 6: + return p[5] | (uint32_t)p[4] << 8 | (uint32_t)p[3] << 16 | + (uint32_t)p[2] << 24 | (uint64_t)p[1] << 32 | (uint64_t)p[0] << 40; + case 7: + return p[6] | (uint32_t)p[5] << 8 | (uint32_t)p[4] << 16 | + (uint32_t)p[3] << 24 | (uint64_t)p[2] << 32 | (uint64_t)p[1] << 40 | + (uint64_t)p[0] << 48; + case 0: + return p[7] | (uint32_t)p[6] << 8 | (uint32_t)p[5] << 16 | + (uint32_t)p[4] << 24 | (uint64_t)p[3] << 32 | (uint64_t)p[2] << 40 | + (uint64_t)p[1] << 48 | (uint64_t)p[0] << 56; +#endif + } +#endif /* T1HA_USE_FAST_ONESHOT_READ */ +} + +static __maybe_unused __always_inline uint64_t +tail64_be_unaligned(const void *v, size_t tail) { + const uint8_t *p = (const uint8_t *)v; +#if defined(can_read_underside) && \ + (UINTPTR_MAX > 0xffffFFFFul || ULONG_MAX > 0xffffFFFFul) + /* On some systems (e.g. x86_64) we can perform a 'oneshot' read, which + * is little bit faster. Thanks Marcin Żukowski + * for the reminder. */ + const unsigned offset = (8 - tail) & 7; + const unsigned shift = offset << 3; + if (likely(can_read_underside(p, 8))) { + p -= offset; + return fetch64_be_unaligned(p) & ((~UINT64_C(0)) >> shift); + } + return fetch64_be_unaligned(p) >> shift; +#else + switch (tail & 7) { + default: + unreachable(); +/* fall through */ +#if T1HA_SYS_UNALIGNED_ACCESS == T1HA_UNALIGNED_ACCESS__EFFICIENT && \ + __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ + /* For most CPUs this code is better when not needed + * copying for alignment or byte reordering. */ + case 1: + return p[0]; + case 2: + return fetch16_be_unaligned(p); + case 3: + return (uint32_t)fetch16_be_unaligned(p) << 8 | p[2]; + case 4: + return fetch32_be(p); + case 5: + return (uint64_t)fetch32_be_unaligned(p) << 8 | p[4]; + case 6: + return (uint64_t)fetch32_be_unaligned(p) << 16 | + fetch16_be_unaligned(p + 4); + case 7: + return (uint64_t)fetch32_be_unaligned(p) << 24 | + (uint32_t)fetch16_be_unaligned(p + 4) << 8 | p[6]; + case 0: + return fetch64_be_unaligned(p); +#else + /* For most CPUs this code is better than a + * copying for alignment and/or byte reordering. */ + case 1: + return p[0]; + case 2: + return p[1] | (uint32_t)p[0] << 8; + case 3: + return p[2] | (uint32_t)p[1] << 8 | (uint32_t)p[0] << 16; + case 4: + return p[3] | (uint32_t)p[2] << 8 | (uint32_t)p[1] << 16 | + (uint32_t)p[0] << 24; + case 5: + return p[4] | (uint32_t)p[3] << 8 | (uint32_t)p[2] << 16 | + (uint32_t)p[1] << 24 | (uint64_t)p[0] << 32; + case 6: + return p[5] | (uint32_t)p[4] << 8 | (uint32_t)p[3] << 16 | + (uint32_t)p[2] << 24 | (uint64_t)p[1] << 32 | (uint64_t)p[0] << 40; + case 7: + return p[6] | (uint32_t)p[5] << 8 | (uint32_t)p[4] << 16 | + (uint32_t)p[3] << 24 | (uint64_t)p[2] << 32 | (uint64_t)p[1] << 40 | + (uint64_t)p[0] << 48; + case 0: + return p[7] | (uint32_t)p[6] << 8 | (uint32_t)p[5] << 16 | + (uint32_t)p[4] << 24 | (uint64_t)p[3] << 32 | (uint64_t)p[2] << 40 | + (uint64_t)p[1] << 48 | (uint64_t)p[0] << 56; +#endif + } +#endif /* can_read_underside */ +} + +/***************************************************************************/ + +#ifndef rot64 +static __maybe_unused __always_inline uint64_t rot64(uint64_t v, unsigned s) { + return (v >> s) | (v << (64 - s)); +} +#endif /* rot64 */ + +#ifndef mul_32x32_64 +static __maybe_unused __always_inline uint64_t mul_32x32_64(uint32_t a, + uint32_t b) { + return a * (uint64_t)b; +} +#endif /* mul_32x32_64 */ + +#ifndef add64carry_first +static __maybe_unused __always_inline unsigned +add64carry_first(uint64_t base, uint64_t addend, uint64_t *sum) { +#if __has_builtin(__builtin_addcll) + unsigned long long carryout; + *sum = __builtin_addcll(base, addend, 0, &carryout); + return (unsigned)carryout; +#else + *sum = base + addend; + return *sum < addend; +#endif /* __has_builtin(__builtin_addcll) */ +} +#endif /* add64carry_fist */ + +#ifndef add64carry_next +static __maybe_unused __always_inline unsigned +add64carry_next(unsigned carry, uint64_t base, uint64_t addend, uint64_t *sum) { +#if __has_builtin(__builtin_addcll) + unsigned long long carryout; + *sum = __builtin_addcll(base, addend, carry, &carryout); + return (unsigned)carryout; +#else + *sum = base + addend + carry; + return *sum < addend || (carry && *sum == addend); +#endif /* __has_builtin(__builtin_addcll) */ +} +#endif /* add64carry_next */ + +#ifndef add64carry_last +static __maybe_unused __always_inline void +add64carry_last(unsigned carry, uint64_t base, uint64_t addend, uint64_t *sum) { +#if __has_builtin(__builtin_addcll) + unsigned long long carryout; + *sum = __builtin_addcll(base, addend, carry, &carryout); + (void)carryout; +#else + *sum = base + addend + carry; +#endif /* __has_builtin(__builtin_addcll) */ +} +#endif /* add64carry_last */ + +#ifndef mul_64x64_128 +static __maybe_unused __always_inline uint64_t mul_64x64_128(uint64_t a, + uint64_t b, + uint64_t *h) { +#if (defined(__SIZEOF_INT128__) || \ + (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128)) && \ + (!defined(__LCC__) || __LCC__ != 124) + __uint128_t r = (__uint128_t)a * (__uint128_t)b; + /* modern GCC could nicely optimize this */ + *h = (uint64_t)(r >> 64); + return (uint64_t)r; +#elif defined(mul_64x64_high) + *h = mul_64x64_high(a, b); + return a * b; +#else + /* performs 64x64 to 128 bit multiplication */ + const uint64_t ll = mul_32x32_64((uint32_t)a, (uint32_t)b); + const uint64_t lh = mul_32x32_64(a >> 32, (uint32_t)b); + const uint64_t hl = mul_32x32_64((uint32_t)a, b >> 32); + const uint64_t hh = mul_32x32_64(a >> 32, b >> 32); + + /* Few simplification are possible here for 32-bit architectures, + * but thus we would lost compatibility with the original 64-bit + * version. Think is very bad idea, because then 32-bit t1ha will + * still (relatively) very slowly and well yet not compatible. */ + uint64_t l; + add64carry_last(add64carry_first(ll, lh << 32, &l), hh, lh >> 32, h); + add64carry_last(add64carry_first(l, hl << 32, &l), *h, hl >> 32, h); + return l; +#endif +} +#endif /* mul_64x64_128() */ + +#ifndef mul_64x64_high +static __maybe_unused __always_inline uint64_t mul_64x64_high(uint64_t a, + uint64_t b) { + uint64_t h; + mul_64x64_128(a, b, &h); + return h; +} +#endif /* mul_64x64_high */ + +/***************************************************************************/ + +/* 'magic' primes */ +static const uint64_t prime_0 = UINT64_C(0xEC99BF0D8372CAAB); +static const uint64_t prime_1 = UINT64_C(0x82434FE90EDCEF39); +static const uint64_t prime_2 = UINT64_C(0xD4F06DB99D67BE4B); +static const uint64_t prime_3 = UINT64_C(0xBD9CACC22C6E9571); +static const uint64_t prime_4 = UINT64_C(0x9C06FAF4D023E3AB); +static const uint64_t prime_5 = UINT64_C(0xC060724A8424F345); +static const uint64_t prime_6 = UINT64_C(0xCB5AF53AE3AAAC31); + +/* xor high and low parts of full 128-bit product */ +static __maybe_unused __always_inline uint64_t mux64(uint64_t v, + uint64_t prime) { + uint64_t l, h; + l = mul_64x64_128(v, prime, &h); + return l ^ h; +} + +static __maybe_unused __always_inline uint64_t final64(uint64_t a, uint64_t b) { + uint64_t x = (a + rot64(b, 41)) * prime_0; + uint64_t y = (rot64(a, 23) + b) * prime_6; + return mux64(x ^ y, prime_5); +} + +static __maybe_unused __always_inline void mixup64(uint64_t *__restrict a, + uint64_t *__restrict b, + uint64_t v, uint64_t prime) { + uint64_t h; + *a ^= mul_64x64_128(*b + v, prime, &h); + *b += h; +} + +/***************************************************************************/ + +typedef union t1ha_uint128 { +#if defined(__SIZEOF_INT128__) || \ + (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128) + __uint128_t v; +#endif + struct { +#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ + uint64_t l, h; +#else + uint64_t h, l; +#endif + }; +} t1ha_uint128_t; + +static __maybe_unused __always_inline t1ha_uint128_t +not128(const t1ha_uint128_t v) { + t1ha_uint128_t r; +#if defined(__SIZEOF_INT128__) || \ + (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128) + r.v = ~v.v; +#else + r.l = ~v.l; + r.h = ~v.h; +#endif + return r; +} + +static __maybe_unused __always_inline t1ha_uint128_t +left128(const t1ha_uint128_t v, unsigned s) { + t1ha_uint128_t r; + assert(s < 128); +#if defined(__SIZEOF_INT128__) || \ + (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128) + r.v = v.v << s; +#else + r.l = (s < 64) ? v.l << s : 0; + r.h = (s < 64) ? (v.h << s) | (s ? v.l >> (64 - s) : 0) : v.l << (s - 64); +#endif + return r; +} + +static __maybe_unused __always_inline t1ha_uint128_t +right128(const t1ha_uint128_t v, unsigned s) { + t1ha_uint128_t r; + assert(s < 128); +#if defined(__SIZEOF_INT128__) || \ + (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128) + r.v = v.v >> s; +#else + r.l = (s < 64) ? (s ? v.h << (64 - s) : 0) | (v.l >> s) : v.h >> (s - 64); + r.h = (s < 64) ? v.h >> s : 0; +#endif + return r; +} + +static __maybe_unused __always_inline t1ha_uint128_t or128(t1ha_uint128_t x, + t1ha_uint128_t y) { + t1ha_uint128_t r; +#if defined(__SIZEOF_INT128__) || \ + (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128) + r.v = x.v | y.v; +#else + r.l = x.l | y.l; + r.h = x.h | y.h; +#endif + return r; +} + +static __maybe_unused __always_inline t1ha_uint128_t xor128(t1ha_uint128_t x, + t1ha_uint128_t y) { + t1ha_uint128_t r; +#if defined(__SIZEOF_INT128__) || \ + (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128) + r.v = x.v ^ y.v; +#else + r.l = x.l ^ y.l; + r.h = x.h ^ y.h; +#endif + return r; +} + +static __maybe_unused __always_inline t1ha_uint128_t rot128(t1ha_uint128_t v, + unsigned s) { + s &= 127; +#if defined(__SIZEOF_INT128__) || \ + (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128) + v.v = (v.v << (128 - s)) | (v.v >> s); + return v; +#else + return s ? or128(left128(v, 128 - s), right128(v, s)) : v; +#endif +} + +static __maybe_unused __always_inline t1ha_uint128_t add128(t1ha_uint128_t x, + t1ha_uint128_t y) { + t1ha_uint128_t r; +#if defined(__SIZEOF_INT128__) || \ + (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128) + r.v = x.v + y.v; +#else + add64carry_last(add64carry_first(x.l, y.l, &r.l), x.h, y.h, &r.h); +#endif + return r; +} + +static __maybe_unused __always_inline t1ha_uint128_t mul128(t1ha_uint128_t x, + t1ha_uint128_t y) { + t1ha_uint128_t r; +#if defined(__SIZEOF_INT128__) || \ + (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128) + r.v = x.v * y.v; +#else + r.l = mul_64x64_128(x.l, y.l, &r.h); + r.h += x.l * y.h + y.l * x.h; +#endif + return r; +} + +/***************************************************************************/ + +#if T1HA0_AESNI_AVAILABLE && defined(__ia32__) +uint64_t t1ha_ia32cpu_features(void); + +static __maybe_unused __always_inline bool +t1ha_ia32_AESNI_avail(uint64_t ia32cpu_features) { + /* check for AES-NI */ + return (ia32cpu_features & UINT32_C(0x02000000)) != 0; +} + +static __maybe_unused __always_inline bool +t1ha_ia32_AVX_avail(uint64_t ia32cpu_features) { + /* check for any AVX */ + return (ia32cpu_features & UINT32_C(0x1A000000)) == UINT32_C(0x1A000000); +} + +static __maybe_unused __always_inline bool +t1ha_ia32_AVX2_avail(uint64_t ia32cpu_features) { + /* check for 'Advanced Vector Extensions 2' */ + return ((ia32cpu_features >> 32) & 32) != 0; +} + +#endif /* T1HA0_AESNI_AVAILABLE && __ia32__ */ diff --git a/include/t1ha_selfcheck.h b/include/t1ha_selfcheck.h new file mode 100644 index 00000000..ff7c589c --- /dev/null +++ b/include/t1ha_selfcheck.h @@ -0,0 +1,76 @@ +/* + * Copyright (c) 2016-2020 Positive Technologies, https://www.ptsecurity.com, + * Fast Positive Hash. + * + * Portions Copyright (c) 2010-2020 Leonid Yuriev , + * The 1Hippeus project (t1h). + * + * This software is provided 'as-is', without any express or implied + * warranty. In no event will the authors be held liable for any damages + * arising from the use of this software. + * + * Permission is granted to anyone to use this software for any purpose, + * including commercial applications, and to alter it and redistribute it + * freely, subject to the following restrictions: + * + * 1. The origin of this software must not be misrepresented; you must not + * claim that you wrote the original software. If you use this software + * in a product, an acknowledgement in the product documentation would be + * appreciated but is not required. + * 2. Altered source versions must be plainly marked as such, and must not be + * misrepresented as being the original software. + * 3. This notice may not be removed or altered from any source distribution. + */ + +/* + * t1ha = { Fast Positive Hash, aka "Позитивный Хэш" } + * by [Positive Technologies](https://www.ptsecurity.ru) + * + * Briefly, it is a 64-bit Hash Function: + * 1. Created for 64-bit little-endian platforms, in predominantly for x86_64, + * but portable and without penalties it can run on any 64-bit CPU. + * 2. In most cases up to 15% faster than City64, xxHash, mum-hash, metro-hash + * and all others portable hash-functions (which do not use specific + * hardware tricks). + * 3. Not suitable for cryptography. + * + * The Future will (be) Positive. Всё будет хорошо. + * + * ACKNOWLEDGEMENT: + * The t1ha was originally developed by Leonid Yuriev (Леонид Юрьев) + * for The 1Hippeus project - zerocopy messaging in the spirit of Sparta! + */ + +#pragma once +#if defined(_MSC_VER) && _MSC_VER > 1800 +#pragma warning(disable : 4464) /* relative include path contains '..' */ +#endif /* MSVC */ +#include "t1ha.h" + +/***************************************************************************/ +/* Self-checking */ + +extern const uint8_t t1ha_test_pattern[64]; +int t1ha_selfcheck(uint64_t (*hash)(const void *, size_t, uint64_t), + const uint64_t *reference_values); + +#ifndef T1HA2_DISABLED +extern const uint64_t t1ha_refval_2atonce[81]; +extern const uint64_t t1ha_refval_2atonce128[81]; +extern const uint64_t t1ha_refval_2stream[81]; +extern const uint64_t t1ha_refval_2stream128[81]; +#endif /* T1HA2_DISABLED */ + +#ifndef T1HA1_DISABLED +extern const uint64_t t1ha_refval_64le[81]; +extern const uint64_t t1ha_refval_64be[81]; +#endif /* T1HA1_DISABLED */ + +#ifndef T1HA0_DISABLED +extern const uint64_t t1ha_refval_32le[81]; +extern const uint64_t t1ha_refval_32be[81]; +#if T1HA0_AESNI_AVAILABLE +extern const uint64_t t1ha_refval_ia32aes_a[81]; +extern const uint64_t t1ha_refval_ia32aes_b[81]; +#endif /* T1HA0_AESNI_AVAILABLE */ +#endif /* T1HA0_DISABLED */ diff --git a/include/xxhash.h b/include/xxhash.h index 9a880470..d11f0f63 100644 --- a/include/xxhash.h +++ b/include/xxhash.h @@ -1,7 +1,7 @@ /* * xxHash - Extremely Fast Hash algorithm * Header File - * Copyright (C) 2012-2024 Yann Collet + * Copyright (C) 2012-2023 Yann Collet * * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php) * @@ -32,328 +32,553 @@ * - xxHash homepage: https://www.xxhash.com * - xxHash source repository: https://github.com/Cyan4973/xxHash */ + /*! * @mainpage xxHash * + * xxHash is an extremely fast non-cryptographic hash algorithm, working at RAM speed + * limits. + * + * It is proposed in four flavors, in three families: + * 1. @ref XXH32_family + * - Classic 32-bit hash function. Simple, compact, and runs on almost all + * 32-bit and 64-bit systems. + * 2. @ref XXH64_family + * - Classic 64-bit adaptation of XXH32. Just as simple, and runs well on most + * 64-bit systems (but _not_ 32-bit systems). + * 3. @ref XXH3_family + * - Modern 64-bit and 128-bit hash function family which features improved + * strength and performance across the board, especially on smaller data. + * It benefits greatly from SIMD and 64-bit without requiring it. + * + * Benchmarks + * --- + * The reference system uses an Intel i7-9700K CPU, and runs Ubuntu x64 20.04. + * The open source benchmark program is compiled with clang v10.0 using -O3 flag. + * + * | Hash Name | ISA ext | Width | Large Data Speed | Small Data Velocity | + * | -------------------- | ------- | ----: | ---------------: | ------------------: | + * | XXH3_64bits() | @b AVX2 | 64 | 59.4 GB/s | 133.1 | + * | MeowHash | AES-NI | 128 | 58.2 GB/s | 52.5 | + * | XXH3_128bits() | @b AVX2 | 128 | 57.9 GB/s | 118.1 | + * | CLHash | PCLMUL | 64 | 37.1 GB/s | 58.1 | + * | XXH3_64bits() | @b SSE2 | 64 | 31.5 GB/s | 133.1 | + * | XXH3_128bits() | @b SSE2 | 128 | 29.6 GB/s | 118.1 | + * | RAM sequential read | | N/A | 28.0 GB/s | N/A | + * | ahash | AES-NI | 64 | 22.5 GB/s | 107.2 | + * | City64 | | 64 | 22.0 GB/s | 76.6 | + * | T1ha2 | | 64 | 22.0 GB/s | 99.0 | + * | City128 | | 128 | 21.7 GB/s | 57.7 | + * | FarmHash | AES-NI | 64 | 21.3 GB/s | 71.9 | + * | XXH64() | | 64 | 19.4 GB/s | 71.0 | + * | SpookyHash | | 64 | 19.3 GB/s | 53.2 | + * | Mum | | 64 | 18.0 GB/s | 67.0 | + * | CRC32C | SSE4.2 | 32 | 13.0 GB/s | 57.9 | + * | XXH32() | | 32 | 9.7 GB/s | 71.9 | + * | City32 | | 32 | 9.1 GB/s | 66.0 | + * | Blake3* | @b AVX2 | 256 | 4.4 GB/s | 8.1 | + * | Murmur3 | | 32 | 3.9 GB/s | 56.1 | + * | SipHash* | | 64 | 3.0 GB/s | 43.2 | + * | Blake3* | @b SSE2 | 256 | 2.4 GB/s | 8.1 | + * | HighwayHash | | 64 | 1.4 GB/s | 6.0 | + * | FNV64 | | 64 | 1.2 GB/s | 62.7 | + * | Blake2* | | 256 | 1.1 GB/s | 5.1 | + * | SHA1* | | 160 | 0.8 GB/s | 5.6 | + * | MD5* | | 128 | 0.6 GB/s | 7.8 | + * @note + * - Hashes which require a specific ISA extension are noted. SSE2 is also noted, + * even though it is mandatory on x64. + * - Hashes with an asterisk are cryptographic. Note that MD5 is non-cryptographic + * by modern standards. + * - Small data velocity is a rough average of algorithm's efficiency for small + * data. For more accurate information, see the wiki. + * - More benchmarks and strength tests are found on the wiki: + * https://github.com/Cyan4973/xxHash/wiki + * + * Usage + * ------ + * All xxHash variants use a similar API. Changing the algorithm is a trivial + * substitution. + * + * @pre + * For functions which take an input and length parameter, the following + * requirements are assumed: + * - The range from [`input`, `input + length`) is valid, readable memory. + * - The only exception is if the `length` is `0`, `input` may be `NULL`. + * - For C++, the objects must have the *TriviallyCopyable* property, as the + * functions access bytes directly as if it was an array of `unsigned char`. + * + * @anchor single_shot_example + * **Single Shot** + * + * These functions are stateless functions which hash a contiguous block of memory, + * immediately returning the result. They are the easiest and usually the fastest + * option. + * + * XXH32(), XXH64(), XXH3_64bits(), XXH3_128bits() + * + * @code{.c} + * #include + * #include "xxhash.h" + * + * // Example for a function which hashes a null terminated string with XXH32(). + * XXH32_hash_t hash_string(const char* string, XXH32_hash_t seed) + * { + * // NULL pointers are only valid if the length is zero + * size_t length = (string == NULL) ? 0 : strlen(string); + * return XXH32(string, length, seed); + * } + * @endcode + * + * + * @anchor streaming_example + * **Streaming** + * + * These groups of functions allow incremental hashing of unknown size, even + * more than what would fit in a size_t. + * + * XXH32_reset(), XXH64_reset(), XXH3_64bits_reset(), XXH3_128bits_reset() + * + * @code{.c} + * #include + * #include + * #include "xxhash.h" + * // Example for a function which hashes a FILE incrementally with XXH3_64bits(). + * XXH64_hash_t hashFile(FILE* f) + * { + * // Allocate a state struct. Do not just use malloc() or new. + * XXH3_state_t* state = XXH3_createState(); + * assert(state != NULL && "Out of memory!"); + * // Reset the state to start a new hashing session. + * XXH3_64bits_reset(state); + * char buffer[4096]; + * size_t count; + * // Read the file in chunks + * while ((count = fread(buffer, 1, sizeof(buffer), f)) != 0) { + * // Run update() as many times as necessary to process the data + * XXH3_64bits_update(state, buffer, count); + * } + * // Retrieve the finalized hash. This will not change the state. + * XXH64_hash_t result = XXH3_64bits_digest(state); + * // Free the state. Do not use free(). + * XXH3_freeState(state); + * return result; + * } + * @endcode + * + * Streaming functions generate the xxHash value from an incremental input. + * This method is slower than single-call functions, due to state management. + * For small inputs, prefer `XXH32()` and `XXH64()`, which are better optimized. + * + * An XXH state must first be allocated using `XXH*_createState()`. + * + * Start a new hash by initializing the state with a seed using `XXH*_reset()`. + * + * Then, feed the hash state by calling `XXH*_update()` as many times as necessary. + * + * The function returns an error code, with 0 meaning OK, and any other value + * meaning there is an error. + * + * Finally, a hash value can be produced anytime, by using `XXH*_digest()`. + * This function returns the nn-bits hash as an int or long long. + * + * It's still possible to continue inserting input into the hash state after a + * digest, and generate new hash values later on by invoking `XXH*_digest()`. + * + * When done, release the state using `XXH*_freeState()`. + * + * + * @anchor canonical_representation_example + * **Canonical Representation** + * + * The default return values from XXH functions are unsigned 32, 64 and 128 bit + * integers. + * This the simplest and fastest format for further post-processing. + * + * However, this leaves open the question of what is the order on the byte level, + * since little and big endian conventions will store the same number differently. + * + * The canonical representation settles this issue by mandating big-endian + * convention, the same convention as human-readable numbers (large digits first). + * + * When writing hash values to storage, sending them over a network, or printing + * them, it's highly recommended to use the canonical representation to ensure + * portability across a wider range of systems, present and future. + * + * The following functions allow transformation of hash values to and from + * canonical format. + * + * XXH32_canonicalFromHash(), XXH32_hashFromCanonical(), + * XXH64_canonicalFromHash(), XXH64_hashFromCanonical(), + * XXH128_canonicalFromHash(), XXH128_hashFromCanonical(), + * + * @code{.c} + * #include + * #include "xxhash.h" + * + * // Example for a function which prints XXH32_hash_t in human readable format + * void printXxh32(XXH32_hash_t hash) + * { + * XXH32_canonical_t cano; + * XXH32_canonicalFromHash(&cano, hash); + * size_t i; + * for(i = 0; i < sizeof(cano.digest); ++i) { + * printf("%02x", cano.digest[i]); + * } + * printf("\n"); + * } + * + * // Example for a function which converts XXH32_canonical_t to XXH32_hash_t + * XXH32_hash_t convertCanonicalToXxh32(XXH32_canonical_t cano) + * { + * XXH32_hash_t hash = XXH32_hashFromCanonical(&cano); + * return hash; + * } + * @endcode + * + * * @file xxhash.h * xxHash prototypes and implementation */ -/* TODO: update */ -/* Notice extracted from xxHash homepage: - -xxHash is an extremely fast hash algorithm, running at RAM speed limits. -It also successfully passes all tests from the SMHasher suite. - -Comparison (single thread, Windows Seven 32 bits, using SMHasher on a Core 2 Duo -@3GHz) - -Name Speed Q.Score Author -xxHash 5.4 GB/s 10 -CrapWow 3.2 GB/s 2 Andrew -MurmurHash 3a 2.7 GB/s 10 Austin Appleby -SpookyHash 2.0 GB/s 10 Bob Jenkins -SBox 1.4 GB/s 9 Bret Mulvey -Lookup3 1.2 GB/s 9 Bob Jenkins -SuperFastHash 1.2 GB/s 1 Paul Hsieh -CityHash64 1.05 GB/s 10 Pike & Alakuijala -FNV 0.55 GB/s 5 Fowler, Noll, Vo -CRC32 0.43 GB/s 9 -MD5-32 0.33 GB/s 10 Ronald L. Rivest -SHA1-32 0.28 GB/s 10 - -Q.Score is a measure of quality of the hash function. -It depends on successfully passing SMHasher test set. -10 is a perfect score. - -Note: SMHasher's CRC32 implementation is not the fastest one. -Other speed-oriented implementations can be faster, -especially in combination with PCLMUL instruction: -https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html?showComment=1552696407071#c3490092340461170735 - -A 64-bit version, named XXH64, is available since r35. -It offers much better speed, but for 64-bit applications only. -Name Speed on 64 bits Speed on 32 bits -XXH64 13.8 GB/s 1.9 GB/s -XXH32 6.8 GB/s 6.0 GB/s -*/ - -#if defined(__cplusplus) -extern "C" { +#if defined (__cplusplus) +extern "C" { #endif /* **************************** * INLINE mode ******************************/ /*! - * XXH_INLINE_ALL (and XXH_PRIVATE_API) + * @defgroup public Public API + * Contains details on the public xxHash functions. + * @{ + */ +#ifdef XXH_DOXYGEN +/*! + * @brief Gives access to internal state declaration, required for static allocation. + * + * Incompatible with dynamic linking, due to risks of ABI changes. + * + * Usage: + * @code{.c} + * #define XXH_STATIC_LINKING_ONLY + * #include "xxhash.h" + * @endcode + */ +# define XXH_STATIC_LINKING_ONLY +/* Do not undef XXH_STATIC_LINKING_ONLY for Doxygen */ + +/*! + * @brief Gives access to internal definitions. + * + * Usage: + * @code{.c} + * #define XXH_STATIC_LINKING_ONLY + * #define XXH_IMPLEMENTATION + * #include "xxhash.h" + * @endcode + */ +# define XXH_IMPLEMENTATION +/* Do not undef XXH_IMPLEMENTATION for Doxygen */ + +/*! + * @brief Exposes the implementation and marks all functions as `inline`. + * * Use these build macros to inline xxhash into the target unit. * Inlining improves performance on small inputs, especially when the length is * expressed as a compile-time constant: * - * https://fastcompression.blogspot.com/2018/03/xxhash-for-small-keys-impressive-power.html + * https://fastcompression.blogspot.com/2018/03/xxhash-for-small-keys-impressive-power.html * * It also keeps xxHash symbols private to the unit, so they are not exported. * * Usage: + * @code{.c} * #define XXH_INLINE_ALL * #include "xxhash.h" - * + * @endcode * Do not compile and link xxhash.o as a separate object, as it is not useful. */ -#if (defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)) && \ - !defined(XXH_INLINE_ALL_31684351384) -/* this section should be traversed only once */ - #define XXH_INLINE_ALL_31684351384 -/* give access to the advanced API, required to compile implementations */ - #undef XXH_STATIC_LINKING_ONLY /* avoid macro redef */ - #define XXH_STATIC_LINKING_ONLY -/* make all functions private */ - #undef XXH_PUBLIC_API - #if defined(__GNUC__) - #define XXH_PUBLIC_API static __inline __attribute__((unused)) - #elif defined(__cplusplus) || \ - (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) - #define XXH_PUBLIC_API static inline - #elif defined(_MSC_VER) - #define XXH_PUBLIC_API static __inline - #else - /* note: this version may generate warnings for unused static functions */ - #define XXH_PUBLIC_API static - #endif +# define XXH_INLINE_ALL +# undef XXH_INLINE_ALL +/*! + * @brief Exposes the implementation without marking functions as inline. + */ +# define XXH_PRIVATE_API +# undef XXH_PRIVATE_API +/*! + * @brief Emulate a namespace by transparently prefixing all symbols. + * + * If you want to include _and expose_ xxHash functions from within your own + * library, but also want to avoid symbol collisions with other libraries which + * may also include xxHash, you can use @ref XXH_NAMESPACE to automatically prefix + * any public symbol from xxhash library with the value of @ref XXH_NAMESPACE + * (therefore, avoid empty or numeric values). + * + * Note that no change is required within the calling program as long as it + * includes `xxhash.h`: Regular symbol names will be automatically translated + * by this header. + */ +# define XXH_NAMESPACE /* YOUR NAME HERE */ +# undef XXH_NAMESPACE +#endif -/* - * This part deals with the special case where a unit wants to inline xxHash, - * but "xxhash.h" has previously been included without XXH_INLINE_ALL, - * such as part of some previously included *.h header file. - * Without further action, the new include would just be ignored, - * and functions would effectively _not_ be inlined (silent failure). - * The following macros solve this situation by prefixing all inlined names, - * avoiding naming collision with previous inclusions. - */ -/* Before that, we unconditionally #undef all symbols, - * in case they were already defined with XXH_NAMESPACE. - * They will then be redefined for XXH_INLINE_ALL - */ - #undef XXH_versionNumber -/* XXH32 */ - #undef XXH32 - #undef XXH32_createState - #undef XXH32_freeState - #undef XXH32_reset - #undef XXH32_update - #undef XXH32_digest - #undef XXH32_copyState - #undef XXH32_canonicalFromHash - #undef XXH32_hashFromCanonical -/* XXH64 */ - #undef XXH64 - #undef XXH64_createState - #undef XXH64_freeState - #undef XXH64_reset - #undef XXH64_update - #undef XXH64_digest - #undef XXH64_copyState - #undef XXH64_canonicalFromHash - #undef XXH64_hashFromCanonical -/* XXH3_64bits */ - #undef XXH3_64bits - #undef XXH3_64bits_withSecret - #undef XXH3_64bits_withSeed - #undef XXH3_createState - #undef XXH3_freeState - #undef XXH3_copyState - #undef XXH3_64bits_reset - #undef XXH3_64bits_reset_withSeed - #undef XXH3_64bits_reset_withSecret - #undef XXH3_64bits_update - #undef XXH3_64bits_digest - #undef XXH3_generateSecret -/* XXH3_128bits */ - #undef XXH128 - #undef XXH3_128bits - #undef XXH3_128bits_withSeed - #undef XXH3_128bits_withSecret - #undef XXH3_128bits_reset - #undef XXH3_128bits_reset_withSeed - #undef XXH3_128bits_reset_withSecret - #undef XXH3_128bits_update - #undef XXH3_128bits_digest - #undef XXH128_isEqual - #undef XXH128_cmp - #undef XXH128_canonicalFromHash - #undef XXH128_hashFromCanonical -/* Finally, free the namespace itself */ - #undef XXH_NAMESPACE - -/* employ the namespace for XXH_INLINE_ALL */ - #define XXH_NAMESPACE XXH_INLINE_ -/* - * Some identifiers (enums, type names) are not symbols, - * but they must nonetheless be renamed to avoid redeclaration. - * Alternative solution: do not redeclare them. - * However, this requires some #ifdefs, and has a more dispersed impact. - * Meanwhile, renaming can be achieved in a single place. - */ - #define XXH_IPREF(Id) XXH_NAMESPACE##Id - #define XXH_OK XXH_IPREF(XXH_OK) - #define XXH_ERROR XXH_IPREF(XXH_ERROR) - #define XXH_errorcode XXH_IPREF(XXH_errorcode) - #define XXH32_canonical_t XXH_IPREF(XXH32_canonical_t) - #define XXH64_canonical_t XXH_IPREF(XXH64_canonical_t) - #define XXH128_canonical_t XXH_IPREF(XXH128_canonical_t) - #define XXH32_state_s XXH_IPREF(XXH32_state_s) - #define XXH32_state_t XXH_IPREF(XXH32_state_t) - #define XXH64_state_s XXH_IPREF(XXH64_state_s) - #define XXH64_state_t XXH_IPREF(XXH64_state_t) - #define XXH3_state_s XXH_IPREF(XXH3_state_s) - #define XXH3_state_t XXH_IPREF(XXH3_state_t) - #define XXH128_hash_t XXH_IPREF(XXH128_hash_t) -/* Ensure the header is parsed again, even if it was previously included */ - #undef XXHASH_H_5627135585666179 - #undef XXHASH_H_STATIC_13879238742 -#endif /* XXH_INLINE_ALL || XXH_PRIVATE_API */ +#if (defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)) \ + && !defined(XXH_INLINE_ALL_31684351384) + /* this section should be traversed only once */ +# define XXH_INLINE_ALL_31684351384 + /* give access to the advanced API, required to compile implementations */ +# undef XXH_STATIC_LINKING_ONLY /* avoid macro redef */ +# define XXH_STATIC_LINKING_ONLY + /* make all functions private */ +# undef XXH_PUBLIC_API +# if defined(__GNUC__) +# define XXH_PUBLIC_API static __inline __attribute__((unused)) +# elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) +# define XXH_PUBLIC_API static inline +# elif defined(_MSC_VER) +# define XXH_PUBLIC_API static __inline +# else + /* note: this version may generate warnings for unused static functions */ +# define XXH_PUBLIC_API static +# endif + + /* + * This part deals with the special case where a unit wants to inline xxHash, + * but "xxhash.h" has previously been included without XXH_INLINE_ALL, + * such as part of some previously included *.h header file. + * Without further action, the new include would just be ignored, + * and functions would effectively _not_ be inlined (silent failure). + * The following macros solve this situation by prefixing all inlined names, + * avoiding naming collision with previous inclusions. + */ + /* Before that, we unconditionally #undef all symbols, + * in case they were already defined with XXH_NAMESPACE. + * They will then be redefined for XXH_INLINE_ALL + */ +# undef XXH_versionNumber + /* XXH32 */ +# undef XXH32 +# undef XXH32_createState +# undef XXH32_freeState +# undef XXH32_reset +# undef XXH32_update +# undef XXH32_digest +# undef XXH32_copyState +# undef XXH32_canonicalFromHash +# undef XXH32_hashFromCanonical + /* XXH64 */ +# undef XXH64 +# undef XXH64_createState +# undef XXH64_freeState +# undef XXH64_reset +# undef XXH64_update +# undef XXH64_digest +# undef XXH64_copyState +# undef XXH64_canonicalFromHash +# undef XXH64_hashFromCanonical + /* XXH3_64bits */ +# undef XXH3_64bits +# undef XXH3_64bits_withSecret +# undef XXH3_64bits_withSeed +# undef XXH3_64bits_withSecretandSeed +# undef XXH3_createState +# undef XXH3_freeState +# undef XXH3_copyState +# undef XXH3_64bits_reset +# undef XXH3_64bits_reset_withSeed +# undef XXH3_64bits_reset_withSecret +# undef XXH3_64bits_update +# undef XXH3_64bits_digest +# undef XXH3_generateSecret + /* XXH3_128bits */ +# undef XXH128 +# undef XXH3_128bits +# undef XXH3_128bits_withSeed +# undef XXH3_128bits_withSecret +# undef XXH3_128bits_reset +# undef XXH3_128bits_reset_withSeed +# undef XXH3_128bits_reset_withSecret +# undef XXH3_128bits_reset_withSecretandSeed +# undef XXH3_128bits_update +# undef XXH3_128bits_digest +# undef XXH128_isEqual +# undef XXH128_cmp +# undef XXH128_canonicalFromHash +# undef XXH128_hashFromCanonical + /* Finally, free the namespace itself */ +# undef XXH_NAMESPACE + + /* employ the namespace for XXH_INLINE_ALL */ +# define XXH_NAMESPACE XXH_INLINE_ + /* + * Some identifiers (enums, type names) are not symbols, + * but they must nonetheless be renamed to avoid redeclaration. + * Alternative solution: do not redeclare them. + * However, this requires some #ifdefs, and has a more dispersed impact. + * Meanwhile, renaming can be achieved in a single place. + */ +# define XXH_IPREF(Id) XXH_NAMESPACE ## Id +# define XXH_OK XXH_IPREF(XXH_OK) +# define XXH_ERROR XXH_IPREF(XXH_ERROR) +# define XXH_errorcode XXH_IPREF(XXH_errorcode) +# define XXH32_canonical_t XXH_IPREF(XXH32_canonical_t) +# define XXH64_canonical_t XXH_IPREF(XXH64_canonical_t) +# define XXH128_canonical_t XXH_IPREF(XXH128_canonical_t) +# define XXH32_state_s XXH_IPREF(XXH32_state_s) +# define XXH32_state_t XXH_IPREF(XXH32_state_t) +# define XXH64_state_s XXH_IPREF(XXH64_state_s) +# define XXH64_state_t XXH_IPREF(XXH64_state_t) +# define XXH3_state_s XXH_IPREF(XXH3_state_s) +# define XXH3_state_t XXH_IPREF(XXH3_state_t) +# define XXH128_hash_t XXH_IPREF(XXH128_hash_t) + /* Ensure the header is parsed again, even if it was previously included */ +# undef XXHASH_H_5627135585666179 +# undef XXHASH_H_STATIC_13879238742 +#endif /* XXH_INLINE_ALL || XXH_PRIVATE_API */ /* **************************************************************** * Stable API *****************************************************************/ #ifndef XXHASH_H_5627135585666179 - #define XXHASH_H_5627135585666179 1 - - /*! - * @defgroup public Public API - * Contains details on the public xxHash functions. - * @{ - - */ - /* specific declaration modes for Windows */ - #if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API) - #if defined(WIN32) && defined(_MSC_VER) && \ - (defined(XXH_IMPORT) || defined(XXH_EXPORT)) - #ifdef XXH_EXPORT - #define XXH_PUBLIC_API __declspec(dllexport) - #elif XXH_IMPORT - #define XXH_PUBLIC_API __declspec(dllimport) - #endif - #else - #define XXH_PUBLIC_API /* do nothing */ - #endif - #endif - - #ifdef XXH_DOXYGEN - /*! - * @brief Emulate a namespace by transparently prefixing all symbols. - * - * If you want to include _and expose_ xxHash functions from within your own - * library, but also want to avoid symbol collisions with other libraries - * which may also include xxHash, you can use XXH_NAMESPACE to automatically - * prefix any public symbol from xxhash library with the value of - * XXH_NAMESPACE (therefore, avoid empty or numeric values). - * - * Note that no change is required within the calling program as long as it - * includes `xxhash.h`: Regular symbol names will be automatically - * translated by this header. - */ - #define XXH_NAMESPACE /* YOUR NAME HERE */ - #undef XXH_NAMESPACE - #endif - - #ifdef XXH_NAMESPACE - #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) - #define XXH32_reset XXH_NAME2(XXH_NAMESPACE, XXH32_reset) - #define XXH32_update XXH_NAME2(XXH_NAMESPACE, XXH32_update) - #define XXH32_digest XXH_NAME2(XXH_NAMESPACE, XXH32_digest) - #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) - #define XXH64_reset XXH_NAME2(XXH_NAMESPACE, XXH64_reset) - #define XXH64_update XXH_NAME2(XXH_NAMESPACE, XXH64_update) - #define XXH64_digest XXH_NAME2(XXH_NAMESPACE, XXH64_digest) - #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 - - /* ************************************* - * Version - ***************************************/ - #define XXH_VERSION_MAJOR 0 - #define XXH_VERSION_MINOR 8 - #define XXH_VERSION_RELEASE 1 - #define XXH_VERSION_NUMBER \ - (XXH_VERSION_MAJOR * 100 * 100 + XXH_VERSION_MINOR * 100 + \ - XXH_VERSION_RELEASE) +#define XXHASH_H_5627135585666179 1 + +/*! @brief Marks a global symbol. */ +#if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API) +# if defined(WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT)) +# ifdef XXH_EXPORT +# define XXH_PUBLIC_API __declspec(dllexport) +# elif XXH_IMPORT +# define XXH_PUBLIC_API __declspec(dllimport) +# endif +# else +# define XXH_PUBLIC_API /* do nothing */ +# endif +#endif + +#ifdef XXH_NAMESPACE +# 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) +# define XXH32_reset XXH_NAME2(XXH_NAMESPACE, XXH32_reset) +# define XXH32_update XXH_NAME2(XXH_NAMESPACE, XXH32_update) +# define XXH32_digest XXH_NAME2(XXH_NAMESPACE, XXH32_digest) +# 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) +# define XXH64_reset XXH_NAME2(XXH_NAMESPACE, XXH64_reset) +# define XXH64_update XXH_NAME2(XXH_NAMESPACE, XXH64_update) +# define XXH64_digest XXH_NAME2(XXH_NAMESPACE, XXH64_digest) +# 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_64bits_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecretandSeed) +# 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_reset_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecretandSeed) +# 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) +# define XXH3_generateSecret_fromSeed XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret_fromSeed) +/* 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_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecretandSeed) +# 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_reset_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecretandSeed) +# 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 + + +/* ************************************* +* Compiler specifics +***************************************/ + +/* specific declaration modes for Windows */ +#if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API) +# if defined(WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT)) +# ifdef XXH_EXPORT +# define XXH_PUBLIC_API __declspec(dllexport) +# elif XXH_IMPORT +# define XXH_PUBLIC_API __declspec(dllimport) +# endif +# else +# define XXH_PUBLIC_API /* do nothing */ +# endif +#endif + +#if defined (__GNUC__) +# define XXH_CONSTF __attribute__((const)) +# define XXH_PUREF __attribute__((pure)) +# define XXH_MALLOCF __attribute__((malloc)) +#else +# define XXH_CONSTF /* disable */ +# define XXH_PUREF +# define XXH_MALLOCF +#endif + +/* ************************************* +* Version +***************************************/ +#define XXH_VERSION_MAJOR 0 +#define XXH_VERSION_MINOR 8 +#define XXH_VERSION_RELEASE 2 +/*! @brief Version number, encoded as two digits each */ +#define XXH_VERSION_NUMBER (XXH_VERSION_MAJOR *100*100 + XXH_VERSION_MINOR *100 + XXH_VERSION_RELEASE) /*! * @brief Obtains the xxHash version. * - * This is only useful when xxHash is compiled as a shared library, as it is - * independent of the version defined in the header. + * This is mostly useful when xxHash is compiled as a shared library, + * since the returned value comes from the library, as opposed to header file. * - * @return `XXH_VERSION_NUMBER` as of when the libray was compiled. + * @return @ref XXH_VERSION_NUMBER of the invoked library. + */ +XXH_PUBLIC_API XXH_CONSTF unsigned XXH_versionNumber (void); + + +/* **************************** +* Common basic types +******************************/ +#include /* size_t */ +/*! + * @brief Exit code for the streaming API. */ -XXH_PUBLIC_API unsigned XXH_versionNumber(void); +typedef enum { + XXH_OK = 0, /*!< OK */ + XXH_ERROR /*!< Error */ +} XXH_errorcode; - /* **************************** - * Definitions - ******************************/ - #include /* size_t */ -typedef enum { XXH_OK = 0, XXH_ERROR } XXH_errorcode; - /*-********************************************************************** - * 32-bit hash - ************************************************************************/ - #if defined(XXH_DOXYGEN) /* Don't show include */ +/*-********************************************************************** +* 32-bit hash +************************************************************************/ +#if defined(XXH_DOXYGEN) /* Don't show include */ /*! * @brief An unsigned 32-bit integer. * @@ -361,51 +586,44 @@ typedef enum { XXH_OK = 0, XXH_ERROR } XXH_errorcode; */ typedef uint32_t XXH32_hash_t; - #elif !defined(__VMS) && \ - (defined(__cplusplus) || \ - (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)) - #include -typedef uint32_t XXH32_hash_t; - - #else - #include - #if UINT_MAX == 0xFFFFFFFFUL -typedef unsigned int XXH32_hash_t; - #else - #if ULONG_MAX == 0xFFFFFFFFUL -typedef unsigned long XXH32_hash_t; - #else - #error "unsupported platform: need a 32-bit type" - #endif - #endif - #endif +#elif !defined (__VMS) \ + && (defined (__cplusplus) \ + || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) ) +# include + typedef uint32_t XXH32_hash_t; + +#else +# include +# if UINT_MAX == 0xFFFFFFFFUL + typedef unsigned int XXH32_hash_t; +# elif ULONG_MAX == 0xFFFFFFFFUL + typedef unsigned long XXH32_hash_t; +# else +# error "unsupported platform: need a 32-bit type" +# endif +#endif /*! * @} * - * @defgroup xxh32_family XXH32 family + * @defgroup XXH32_family XXH32 family * @ingroup public * Contains functions used in the classic 32-bit xxHash algorithm. * * @note - * XXH32 is considered rather weak by today's standards. - * The @ref xxh3_family 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. + * XXH32 is useful for older platforms, with no or poor 64-bit performance. + * Note that the @ref XXH3_family provides competitive speed for both 32-bit + * and 64-bit systems, and offers true 64/128 bit hash results. * - * @see @ref xxh64_family, @ref xxh3_family : Other xxHash families - * @see @ref xxh32_impl for implementation details + * @see @ref XXH64_family, @ref XXH3_family : Other xxHash families + * @see @ref XXH32_impl for implementation details * @{ - */ /*! * @brief Calculates the 32-bit hash of @p input using xxHash32. * - * Speed on Core 2 Duo @ 3 GHz (single thread, SMHasher benchmark): 5.4 GB/s - * - * @param input The block of data to be hashed, at least @p length bytes in - * size. + * @param input The block of data to be hashed, at least @p length bytes in size. * @param length The length of @p input, in bytes. * @param seed The 32-bit seed to alter the hash's output predictably. * @@ -414,94 +632,46 @@ typedef unsigned long XXH32_hash_t; * readable, contiguous memory. However, if @p length is `0`, @p input may be * `NULL`. In C++, this also must be *TriviallyCopyable*. * - * @return The calculated 32-bit hash value. - * - * @see - * XXH64(), XXH3_64bits_withSeed(), XXH3_128bits_withSeed(), XXH128(): - * Direct equivalents for the other variants of xxHash. - * @see - * XXH32_createState(), XXH32_update(), XXH32_digest(): Streaming version. - */ -XXH_PUBLIC_API XXH32_hash_t XXH32(const void *input, size_t length, - XXH32_hash_t seed); - -/*! - * Streaming functions generate the xxHash value from an incremental input. - * This method is slower than single-call functions, due to state management. - * For small inputs, prefer `XXH32()` and `XXH64()`, which are better optimized. - * - * An XXH state must first be allocated using `XXH*_createState()`. - * - * Start a new hash by initializing the state with a seed using `XXH*_reset()`. - * - * Then, feed the hash state by calling `XXH*_update()` as many times as - * necessary. - * - * The function returns an error code, with 0 meaning OK, and any other value - * meaning there is an error. - * - * Finally, a hash value can be produced anytime, by using `XXH*_digest()`. - * This function returns the nn-bits hash as an int or long long. - * - * It's still possible to continue inserting input into the hash state after a - * digest, and generate new hash values later on by invoking `XXH*_digest()`. - * - * When done, release the state using `XXH*_freeState()`. - * - * Example code for incrementally hashing a file: - * @code{.c} - * #include - * #include - * #define BUFFER_SIZE 256 - * - * // Note: XXH64 and XXH3 use the same interface. - * XXH32_hash_t - * hashFile(FILE* stream) - * { - - * XXH32_state_t* state; - * unsigned char buf[BUFFER_SIZE]; - * size_t amt; - * XXH32_hash_t hash; + * @return The calculated 32-bit xxHash32 value. * - * state = XXH32_createState(); // Create a state - * assert(state != NULL); // Error check here - * XXH32_reset(state, 0xbaad5eed); // Reset state with our seed - * while ((amt = fread(buf, 1, sizeof(buf), stream)) != 0) { - - * XXH32_update(state, buf, amt); // Hash the file in chunks - * } - * hash = XXH32_digest(state); // Finalize the hash - * XXH32_freeState(state); // Clean up - * return hash; - * } - * @endcode + * @see @ref single_shot_example "Single Shot Example" for an example. */ +XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed); +#ifndef XXH_NO_STREAM /*! * @typedef struct XXH32_state_s XXH32_state_t * @brief The opaque state struct for the XXH32 streaming API. * * @see XXH32_state_s for details. + * @see @ref streaming_example "Streaming Example" */ typedef struct XXH32_state_s XXH32_state_t; /*! * @brief Allocates an @ref XXH32_state_t. * - * Must be freed with XXH32_freeState(). - * @return An allocated XXH32_state_t on success, `NULL` on failure. + * @return An allocated pointer of @ref XXH32_state_t on success. + * @return `NULL` on failure. + * + * @note Must be freed with XXH32_freeState(). + * + * @see @ref streaming_example "Streaming Example" */ -XXH_PUBLIC_API XXH32_state_t *XXH32_createState(void); +XXH_PUBLIC_API XXH_MALLOCF XXH32_state_t* XXH32_createState(void); /*! * @brief Frees an @ref XXH32_state_t. * - * Must be allocated with XXH32_createState(). - * @param statePtr A pointer to an @ref XXH32_state_t allocated with @ref - * XXH32_createState(). - * @return XXH_OK. + * @param statePtr A pointer to an @ref XXH32_state_t allocated with @ref XXH32_createState(). + * + * @return @ref XXH_OK. + * + * @note @p statePtr must be allocated with XXH32_createState(). + * + * @see @ref streaming_example "Streaming Example" + * */ -XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t *statePtr); +XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr); /*! * @brief Copies one @ref XXH32_state_t to another. * @@ -510,33 +680,31 @@ XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t *statePtr); * @pre * @p dst_state and @p src_state must not be `NULL` and must not overlap. */ -XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t *dst_state, - const XXH32_state_t *src_state); +XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dst_state, const XXH32_state_t* src_state); /*! * @brief Resets an @ref XXH32_state_t to begin a new hash. * - * This function resets and seeds a state. Call it before @ref XXH32_update(). - * * @param statePtr The state struct to reset. * @param seed The 32-bit seed to alter the hash result predictably. * * @pre * @p statePtr must not be `NULL`. * - * @return @ref XXH_OK on success, @ref XXH_ERROR on failure. + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note This function resets and seeds a state. Call it before @ref XXH32_update(). + * + * @see @ref streaming_example "Streaming Example" */ -XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t *statePtr, - XXH32_hash_t seed); +XXH_PUBLIC_API XXH_errorcode XXH32_reset (XXH32_state_t* statePtr, XXH32_hash_t seed); /*! * @brief Consumes a block of @p input to an @ref XXH32_state_t. * - * Call this to incrementally consume blocks of data. - * * @param statePtr The state struct to update. - * @param input The block of data to be hashed, at least @p length bytes in - * size. + * @param input The block of data to be hashed, at least @p length bytes in size. * @param length The length of @p input, in bytes. * * @pre @@ -546,70 +714,55 @@ XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t *statePtr, * readable, contiguous memory. However, if @p length is `0`, @p input may be * `NULL`. In C++, this also must be *TriviallyCopyable*. * - * @return @ref XXH_OK on success, @ref XXH_ERROR on failure. + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note Call this to incrementally consume blocks of data. + * + * @see @ref streaming_example "Streaming Example" */ -XXH_PUBLIC_API XXH_errorcode XXH32_update(XXH32_state_t *statePtr, - const void *input, size_t length); +XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* statePtr, const void* input, size_t length); /*! * @brief Returns the calculated hash value from an @ref XXH32_state_t. * - * @note - * Calling XXH32_digest() will not affect @p statePtr, so you can update, - * digest, and update again. - * * @param statePtr The state struct to calculate the hash from. * * @pre * @p statePtr must not be `NULL`. * - * @return The calculated xxHash32 value from that state. - */ -XXH_PUBLIC_API XXH32_hash_t XXH32_digest(const XXH32_state_t *statePtr); - -/******* Canonical representation *******/ - -/* - * The default return values from XXH functions are unsigned 32 and 64 bit - * integers. - * This the simplest and fastest format for further post-processing. - * - * However, this leaves open the question of what is the order on the byte - * level, since little and big endian conventions will store the same number - * differently. - * - * The canonical representation settles this issue by mandating big-endian - * convention, the same convention as human-readable numbers (large digits - * first). + * @return The calculated 32-bit xxHash32 value from that state. * - * When writing hash values to storage, sending them over a network, or printing - * them, it's highly recommended to use the canonical representation to ensure - * portability across a wider range of systems, present and future. + * @note + * Calling XXH32_digest() will not affect @p statePtr, so you can update, + * digest, and update again. * - * The following functions allow transformation of hash values to and from - * canonical format. + * @see @ref streaming_example "Streaming Example" */ +XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_digest (const XXH32_state_t* statePtr); +#endif /* !XXH_NO_STREAM */ + +/******* Canonical representation *******/ /*! * @brief Canonical (big endian) representation of @ref XXH32_hash_t. */ typedef struct { - - unsigned char digest[4]; /*!< Hash bytes, big endian */ - + unsigned char digest[4]; /*!< Hash bytes, big endian */ } XXH32_canonical_t; /*! * @brief Converts an @ref XXH32_hash_t to a big endian @ref XXH32_canonical_t. * - * @param dst The @ref XXH32_canonical_t pointer to be stored to. + * @param dst The @ref XXH32_canonical_t pointer to be stored to. * @param hash The @ref XXH32_hash_t to be converted. * * @pre * @p dst must not be `NULL`. + * + * @see @ref canonical_representation_example "Canonical Representation Example" */ -XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t *dst, - XXH32_hash_t hash); +XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash); /*! * @brief Converts an @ref XXH32_canonical_t to a native @ref XXH32_hash_t. @@ -620,103 +773,127 @@ XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t *dst, * @p src must not be `NULL`. * * @return The converted hash. + * + * @see @ref canonical_representation_example "Canonical Representation Example" + */ +XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src); + + +/*! @cond Doxygen ignores this part */ +#ifdef __has_attribute +# define XXH_HAS_ATTRIBUTE(x) __has_attribute(x) +#else +# define XXH_HAS_ATTRIBUTE(x) 0 +#endif +/*! @endcond */ + +/*! @cond Doxygen ignores this part */ +/* + * C23 __STDC_VERSION__ number hasn't been specified yet. For now + * leave as `201711L` (C17 + 1). + * TODO: Update to correct value when its been specified. + */ +#define XXH_C23_VN 201711L +/*! @endcond */ + +/*! @cond Doxygen ignores this part */ +/* C-language Attributes are added in C23. */ +#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN) && defined(__has_c_attribute) +# define XXH_HAS_C_ATTRIBUTE(x) __has_c_attribute(x) +#else +# define XXH_HAS_C_ATTRIBUTE(x) 0 +#endif +/*! @endcond */ + +/*! @cond Doxygen ignores this part */ +#if defined(__cplusplus) && defined(__has_cpp_attribute) +# define XXH_HAS_CPP_ATTRIBUTE(x) __has_cpp_attribute(x) +#else +# define XXH_HAS_CPP_ATTRIBUTE(x) 0 +#endif +/*! @endcond */ + +/*! @cond Doxygen ignores this part */ +/* + * Define XXH_FALLTHROUGH macro for annotating switch case with the 'fallthrough' attribute + * introduced in CPP17 and C23. + * CPP17 : https://en.cppreference.com/w/cpp/language/attributes/fallthrough + * C23 : https://en.cppreference.com/w/c/language/attributes/fallthrough + */ +#if XXH_HAS_C_ATTRIBUTE(fallthrough) || XXH_HAS_CPP_ATTRIBUTE(fallthrough) +# define XXH_FALLTHROUGH [[fallthrough]] +#elif XXH_HAS_ATTRIBUTE(__fallthrough__) +# define XXH_FALLTHROUGH __attribute__ ((__fallthrough__)) +#else +# define XXH_FALLTHROUGH /* fallthrough */ +#endif +/*! @endcond */ + +/*! @cond Doxygen ignores this part */ +/* + * Define XXH_NOESCAPE for annotated pointers in public API. + * https://clang.llvm.org/docs/AttributeReference.html#noescape + * As of writing this, only supported by clang. */ -XXH_PUBLIC_API XXH32_hash_t -XXH32_hashFromCanonical(const XXH32_canonical_t *src); - - #ifdef __has_attribute - #define XXH_HAS_ATTRIBUTE(x) __has_attribute(x) - #else - #define XXH_HAS_ATTRIBUTE(x) 0 - #endif - - /* C-language Attributes are added in C23. */ - #if defined(__STDC_VERSION__) && (__STDC_VERSION__ > 201710L) && \ - defined(__has_c_attribute) - #define XXH_HAS_C_ATTRIBUTE(x) __has_c_attribute(x) - #else - #define XXH_HAS_C_ATTRIBUTE(x) 0 - #endif - - #if defined(__cplusplus) && defined(__has_cpp_attribute) - #define XXH_HAS_CPP_ATTRIBUTE(x) __has_cpp_attribute(x) - #else - #define XXH_HAS_CPP_ATTRIBUTE(x) 0 - #endif - - /* - Define XXH_FALLTHROUGH macro for annotating switch case with the 'fallthrough' - attribute introduced in CPP17 and C23. CPP17 : - https://en.cppreference.com/w/cpp/language/attributes/fallthrough C23 : - https://en.cppreference.com/w/c/language/attributes/fallthrough - */ - #if XXH_HAS_C_ATTRIBUTE(x) - #define XXH_FALLTHROUGH [[fallthrough]] - #elif XXH_HAS_CPP_ATTRIBUTE(x) - #define XXH_FALLTHROUGH [[fallthrough]] - #elif XXH_HAS_ATTRIBUTE(__fallthrough__) - #define XXH_FALLTHROUGH __attribute__((fallthrough)) - #else - #define XXH_FALLTHROUGH - #endif +#if XXH_HAS_ATTRIBUTE(noescape) +# define XXH_NOESCAPE __attribute__((noescape)) +#else +# define XXH_NOESCAPE +#endif +/*! @endcond */ + /*! * @} * @ingroup public * @{ - */ - #ifndef XXH_NO_LONG_LONG - /*-********************************************************************** - * 64-bit hash - ************************************************************************/ - #if defined(XXH_DOXYGEN) /* don't include */ +#ifndef XXH_NO_LONG_LONG +/*-********************************************************************** +* 64-bit hash +************************************************************************/ +#if defined(XXH_DOXYGEN) /* don't include */ /*! * @brief An unsigned 64-bit integer. * * Not necessarily defined to `uint64_t` but functionally equivalent. */ typedef uint64_t XXH64_hash_t; - #elif !defined(__VMS) && \ - (defined(__cplusplus) || (defined(__STDC_VERSION__) && \ - (__STDC_VERSION__ >= 199901L) /* C99 */)) - #include -typedef uint64_t XXH64_hash_t; - #else - #include - #if defined(__LP64__) && ULONG_MAX == 0xFFFFFFFFFFFFFFFFULL -/* LP64 ABI says uint64_t is unsigned long */ -typedef unsigned long XXH64_hash_t; - #else -/* the following type must have a width of 64-bit */ -typedef unsigned long long XXH64_hash_t; - #endif - #endif +#elif !defined (__VMS) \ + && (defined (__cplusplus) \ + || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) ) +# include + typedef uint64_t XXH64_hash_t; +#else +# include +# if defined(__LP64__) && ULONG_MAX == 0xFFFFFFFFFFFFFFFFULL + /* LP64 ABI says uint64_t is unsigned long */ + typedef unsigned long XXH64_hash_t; +# else + /* the following type must have a width of 64-bit */ + typedef unsigned long long XXH64_hash_t; +# endif +#endif /*! * @} * - * @defgroup xxh64_family XXH64 family + * @defgroup XXH64_family XXH64 family * @ingroup public * @{ - * Contains functions used in the classic 64-bit xxHash algorithm. * * @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. + * and offers true 64/128 bit hash results. + * It provides better speed for systems with vector processing capabilities. */ /*! * @brief Calculates the 64-bit hash of @p input using xxHash64. * - * This function usually runs faster on 64-bit systems, but slower on 32-bit - * systems (see benchmark). - * - * @param input The block of data to be hashed, at least @p length bytes in - * size. + * @param input The block of data to be hashed, at least @p length bytes in size. * @param length The length of @p input, in bytes. * @param seed The 64-bit seed to alter the hash's output predictably. * @@ -725,62 +902,166 @@ typedef unsigned long long XXH64_hash_t; * readable, contiguous memory. However, if @p length is `0`, @p input may be * `NULL`. In C++, this also must be *TriviallyCopyable*. * - * @return The calculated 64-bit hash. + * @return The calculated 64-bit xxHash64 value. * - * @see - * XXH32(), XXH3_64bits_withSeed(), XXH3_128bits_withSeed(), XXH128(): - * Direct equivalents for the other variants of xxHash. - * @see - * XXH64_createState(), XXH64_update(), XXH64_digest(): Streaming version. + * @see @ref single_shot_example "Single Shot Example" for an example. */ -XXH_PUBLIC_API XXH64_hash_t XXH64(const void *input, size_t length, - XXH64_hash_t seed); +XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed); /******* Streaming *******/ +#ifndef XXH_NO_STREAM /*! * @brief The opaque state struct for the XXH64 streaming API. * * @see XXH64_state_s for details. + * @see @ref streaming_example "Streaming Example" */ -typedef struct XXH64_state_s XXH64_state_t; /* incomplete type */ -XXH_PUBLIC_API XXH64_state_t *XXH64_createState(void); -XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t *statePtr); -XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t *dst_state, - const XXH64_state_t *src_state); - -XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH64_state_t *statePtr, - XXH64_hash_t seed); -XXH_PUBLIC_API XXH_errorcode XXH64_update(XXH64_state_t *statePtr, - const void *input, size_t length); -XXH_PUBLIC_API XXH64_hash_t XXH64_digest(const XXH64_state_t *statePtr); - -/******* Canonical representation *******/ -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); +typedef struct XXH64_state_s XXH64_state_t; /* incomplete type */ /*! - * @} - * ************************************************************************ - * @defgroup xxh3_family XXH3 family - * @ingroup public - * @{ - + * @brief Allocates an @ref XXH64_state_t. * - * XXH3 is a more recent hash algorithm featuring: - * - Improved speed for both small and large inputs - * - True 64-bit and 128-bit outputs - * - SIMD acceleration - * - Improved 32-bit viability + * @return An allocated pointer of @ref XXH64_state_t on success. + * @return `NULL` on failure. * - * Speed analysis methodology is explained here: + * @note Must be freed with XXH64_freeState(). + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_MALLOCF XXH64_state_t* XXH64_createState(void); + +/*! + * @brief Frees an @ref XXH64_state_t. + * + * @param statePtr A pointer to an @ref XXH64_state_t allocated with @ref XXH64_createState(). + * + * @return @ref XXH_OK. + * + * @note @p statePtr must be allocated with XXH64_createState(). + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr); + +/*! + * @brief Copies one @ref XXH64_state_t to another. + * + * @param dst_state The state to copy to. + * @param src_state The state to copy from. + * @pre + * @p dst_state and @p src_state must not be `NULL` and must not overlap. + */ +XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dst_state, const XXH64_state_t* src_state); + +/*! + * @brief Resets an @ref XXH64_state_t to begin a new hash. + * + * @param statePtr The state struct to reset. + * @param seed The 64-bit seed to alter the hash result predictably. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note This function resets and seeds a state. Call it before @ref XXH64_update(). + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_errorcode XXH64_reset (XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed); + +/*! + * @brief Consumes a block of @p input to an @ref XXH64_state_t. + * + * @param statePtr The state struct to update. + * @param input The block of data to be hashed, at least @p length bytes in size. + * @param length The length of @p input, in bytes. + * + * @pre + * @p statePtr must not be `NULL`. + * @pre + * The memory between @p input and @p input + @p length must be valid, + * readable, contiguous memory. However, if @p length is `0`, @p input may be + * `NULL`. In C++, this also must be *TriviallyCopyable*. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note Call this to incrementally consume blocks of data. + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH_NOESCAPE XXH64_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length); + +/*! + * @brief Returns the calculated hash value from an @ref XXH64_state_t. + * + * @param statePtr The state struct to calculate the hash from. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return The calculated 64-bit xxHash64 value from that state. + * + * @note + * Calling XXH64_digest() will not affect @p statePtr, so you can update, + * digest, and update again. + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_digest (XXH_NOESCAPE const XXH64_state_t* statePtr); +#endif /* !XXH_NO_STREAM */ +/******* Canonical representation *******/ + +/*! + * @brief Canonical (big endian) representation of @ref XXH64_hash_t. + */ +typedef struct { unsigned char digest[sizeof(XXH64_hash_t)]; } XXH64_canonical_t; + +/*! + * @brief Converts an @ref XXH64_hash_t to a big endian @ref XXH64_canonical_t. + * + * @param dst The @ref XXH64_canonical_t pointer to be stored to. + * @param hash The @ref XXH64_hash_t to be converted. + * + * @pre + * @p dst must not be `NULL`. + * + * @see @ref canonical_representation_example "Canonical Representation Example" + */ +XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash); + +/*! + * @brief Converts an @ref XXH64_canonical_t to a native @ref XXH64_hash_t. + * + * @param src The @ref XXH64_canonical_t to convert. + * + * @pre + * @p src must not be `NULL`. + * + * @return The converted hash. + * + * @see @ref canonical_representation_example "Canonical Representation Example" + */ +XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src); + +#ifndef XXH_NO_XXH3 + +/*! + * @} + * ************************************************************************ + * @defgroup XXH3_family XXH3 family + * @ingroup public + * @{ + * + * XXH3 is a more recent hash algorithm featuring: + * - Improved speed for both small and large inputs + * - True 64-bit and 128-bit outputs + * - SIMD acceleration + * - Improved 32-bit viability + * + * Speed analysis methodology is explained here: * * https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html * @@ -790,16 +1071,26 @@ XXH64_hashFromCanonical(const XXH64_canonical_t *src); * * XXH3's speed benefits greatly from SIMD and 64-bit arithmetic, * but does not require it. - * Any 32-bit and 64-bit targets that can run XXH32 smoothly - * can run XXH3 at competitive speeds, even without vector support. - * Further details are explained in the implementation. - * - * Optimized implementations are provided for AVX512, AVX2, SSE2, NEON, POWER8, - * ZVector and scalar targets. This can be controlled via the XXH_VECTOR macro. + * Most 32-bit and 64-bit targets that can run XXH32 smoothly can run XXH3 + * at competitive speeds, even without vector support. Further details are + * explained in the implementation. + * + * XXH3 has a fast scalar implementation, but it also includes accelerated SIMD + * implementations for many common platforms: + * - AVX512 + * - AVX2 + * - SSE2 + * - ARM NEON + * - WebAssembly SIMD128 + * - POWER8 VSX + * - s390x ZVector + * This can be controlled via the @ref XXH_VECTOR macro, but it automatically + * selects the best version according to predefined macros. For the x86 family, an + * automatic runtime dispatcher is included separately in @ref xxh_x86dispatch.c. * * XXH3 implementation is portable: * it has a generic C90 formulation that can be compiled on any platform, - * all implementations generage exactly the same hash value on all platforms. + * all implementations generate exactly the same hash value on all platforms. * Starting from v0.8.0, it's also labelled "stable", meaning that * any future version will also generate the same hash value. * @@ -811,53 +1102,106 @@ XXH64_hashFromCanonical(const XXH64_canonical_t *src); * * The API supports one-shot hashing, streaming mode, and custom secrets. */ - /*-********************************************************************** - * XXH3 64-bit variant - ************************************************************************/ +* XXH3 64-bit variant +************************************************************************/ -/* 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); +/*! + * @brief Calculates 64-bit unseeded variant of XXH3 hash of @p input. + * + * @param input The block of data to be hashed, at least @p length bytes in size. + * @param length The length of @p input, in bytes. + * + * @pre + * The memory between @p input and @p input + @p length must be valid, + * readable, contiguous memory. However, if @p length is `0`, @p input may be + * `NULL`. In C++, this also must be *TriviallyCopyable*. + * + * @return The calculated 64-bit XXH3 hash value. + * + * @note + * This is equivalent to @ref XXH3_64bits_withSeed() with a seed of `0`, however + * it may have slightly better performance due to constant propagation of the + * defaults. + * + * @see + * XXH3_64bits_withSeed(), XXH3_64bits_withSecret(): other seeding variants + * @see @ref single_shot_example "Single Shot Example" for an example. + */ +XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length); -/* - * XXH3_64bits_withSeed(): - * This variant generates a custom secret on the fly - * based on default secret altered using the `seed` value. +/*! + * @brief Calculates 64-bit seeded variant of XXH3 hash of @p input. + * + * @param input The block of data to be hashed, at least @p length bytes in size. + * @param length The length of @p input, in bytes. + * @param seed The 64-bit seed to alter the hash result predictably. + * + * @pre + * The memory between @p input and @p input + @p length must be valid, + * readable, contiguous memory. However, if @p length is `0`, @p input may be + * `NULL`. In C++, this also must be *TriviallyCopyable*. + * + * @return The calculated 64-bit XXH3 hash value. + * + * @note + * seed == 0 produces the same results as @ref XXH3_64bits(). + * + * This variant generates a custom secret on the fly based on default secret + * altered using the @p 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(). + * + * @see @ref single_shot_example "Single Shot Example" for an example. */ -XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSeed(const void *data, size_t len, - XXH64_hash_t seed); +XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed); - /*! - * The bare minimum size for a custom secret. - * - * @see - * XXH3_64bits_withSecret(), XXH3_64bits_reset_withSecret(), - * XXH3_128bits_withSecret(), XXH3_128bits_reset_withSecret(). - */ - #define XXH3_SECRET_SIZE_MIN 136 +/*! + * The bare minimum size for a custom secret. + * + * @see + * XXH3_64bits_withSecret(), XXH3_64bits_reset_withSecret(), + * XXH3_128bits_withSecret(), XXH3_128bits_reset_withSecret(). + */ +#define XXH3_SECRET_SIZE_MIN 136 + +/*! + * @brief Calculates 64-bit variant of XXH3 with a custom "secret". + * + * @param data The block of data to be hashed, at least @p len bytes in size. + * @param len The length of @p data, in bytes. + * @param secret The secret data. + * @param secretSize The length of @p secret, in bytes. + * + * @return The calculated 64-bit XXH3 hash value. + * + * @pre + * The memory between @p data and @p data + @p len must be valid, + * readable, contiguous memory. However, if @p length is `0`, @p data may be + * `NULL`. In C++, this also must be *TriviallyCopyable*. + * + * 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 @p secretSize *must* be large enough (>= @ref XXH3_SECRET_SIZE_MIN). + * However, the quality of the secret impacts the dispersion of the hash algorithm. + * Therefore, the secret _must_ look like a bunch of random bytes. + * Avoid "trivial" or structured data such as repeated sequences or a text document. + * Whenever in doubt about the "randomness" of the blob of bytes, + * consider employing @ref XXH3_generateSecret() instead (see below). + * It will generate a proper high entropy secret derived from the blob of bytes. + * Another advantage of using XXH3_generateSecret() is that + * it guarantees that all bits within the initial blob of bytes + * will impact every bit of the output. + * This is not necessarily the case when using the blob of bytes directly + * because, when hashing _small_ inputs, only a portion of the secret is employed. + * + * @see @ref single_shot_example "Single Shot Example" for an example. + */ +XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize); -/* - * 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. - */ -XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSecret(const void *data, size_t len, - const void *secret, - size_t secretSize); /******* Streaming *******/ +#ifndef XXH_NO_STREAM /* * Streaming requires state maintenance. * This operation costs memory and CPU. @@ -866,52 +1210,143 @@ XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSecret(const void *data, size_t len, */ /*! - * @brief The state struct for the XXH3 streaming API. + * @brief The opaque state struct for the XXH3 streaming API. * * @see XXH3_state_s for details. + * @see @ref streaming_example "Streaming Example" */ -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); +typedef struct XXH3_state_s XXH3_state_t; +XXH_PUBLIC_API XXH_MALLOCF XXH3_state_t* XXH3_createState(void); +XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr); -/* - * XXH3_64bits_reset(): - * Initialize with default parameters. - * digest will be equivalent to `XXH3_64bits()`. +/*! + * @brief Copies one @ref XXH3_state_t to another. + * + * @param dst_state The state to copy to. + * @param src_state The state to copy from. + * @pre + * @p dst_state and @p src_state must not be `NULL` and must not overlap. */ -XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH3_state_t *statePtr); -/* - * XXH3_64bits_reset_withSeed(): - * Generate a custom secret from `seed`, and store it into `statePtr`. - * digest will be equivalent to `XXH3_64bits_withSeed()`. +XXH_PUBLIC_API void XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state); + +/*! + * @brief Resets an @ref XXH3_state_t to begin a new hash. + * + * @param statePtr The state struct to reset. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note + * - This function resets `statePtr` and generate a secret with default parameters. + * - Call this function before @ref XXH3_64bits_update(). + * - Digest will be equivalent to `XXH3_64bits()`. + * + * @see @ref streaming_example "Streaming Example" + * */ -XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH3_state_t *statePtr, - XXH64_hash_t seed); -/* - * XXH3_64bits_reset_withSecret(): - * `secret` is referenced, it _must outlive_ the hash streaming session. - * Similar to one-shot API, `secretSize` must be >= `XXH3_SECRET_SIZE_MIN`, +XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr); + +/*! + * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash. + * + * @param statePtr The state struct to reset. + * @param seed The 64-bit seed to alter the hash result predictably. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note + * - This function resets `statePtr` and generate a secret from `seed`. + * - Call this function before @ref XXH3_64bits_update(). + * - Digest will be equivalent to `XXH3_64bits_withSeed()`. + * + * @see @ref streaming_example "Streaming Example" + * + */ +XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed); + +/*! + * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash. + * + * @param statePtr The state struct to reset. + * @param secret The secret data. + * @param secretSize The length of @p secret, in bytes. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note + * `secret` is referenced, it _must outlive_ the hash streaming session. + * + * Similar to one-shot API, `secretSize` must be >= @ref 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). + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize); + +/*! + * @brief Consumes a block of @p input to an @ref XXH3_state_t. + * + * @param statePtr The state struct to update. + * @param input The block of data to be hashed, at least @p length bytes in size. + * @param length The length of @p input, in bytes. + * + * @pre + * @p statePtr must not be `NULL`. + * @pre + * The memory between @p input and @p input + @p length must be valid, + * readable, contiguous memory. However, if @p length is `0`, @p input may be + * `NULL`. In C++, this also must be *TriviallyCopyable*. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note Call this to incrementally consume blocks of data. + * + * @see @ref streaming_example "Streaming Example" */ -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 (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length); -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); +/*! + * @brief Returns the calculated XXH3 64-bit hash value from an @ref XXH3_state_t. + * + * @param statePtr The state struct to calculate the hash from. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return The calculated XXH3 64-bit hash value from that state. + * + * @note + * Calling XXH3_64bits_digest() will not affect @p statePtr, so you can update, + * digest, and update again. + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr); +#endif /* !XXH_NO_STREAM */ /* note : canonical representation of XXH3 is the same as XXH64 * since they both produce XXH64_hash_t values */ + /*-********************************************************************** - * XXH3 128-bit variant - ************************************************************************/ +* XXH3 128-bit variant +************************************************************************/ /*! * @brief The return value from 128-bit hashes. @@ -920,21 +1355,80 @@ XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest(const XXH3_state_t *statePtr); * endianness. */ typedef struct { - - XXH64_hash_t low64; /*!< `value & 0xFFFFFFFFFFFFFFFF` */ - XXH64_hash_t high64; /*!< `value >> 64` */ - + XXH64_hash_t low64; /*!< `value & 0xFFFFFFFFFFFFFFFF` */ + XXH64_hash_t high64; /*!< `value >> 64` */ } XXH128_hash_t; -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); -XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSecret(const void *data, - size_t len, - const void *secret, - size_t secretSize); +/*! + * @brief Calculates 128-bit unseeded variant of XXH3 of @p data. + * + * @param data The block of data to be hashed, at least @p length bytes in size. + * @param len The length of @p data, in bytes. + * + * @return The calculated 128-bit variant of XXH3 value. + * + * The 128-bit variant of XXH3 has more strength, but it has a bit of overhead + * for shorter inputs. + * + * This is equivalent to @ref XXH3_128bits_withSeed() with a seed of `0`, however + * it may have slightly better performance due to constant propagation of the + * defaults. + * + * @see XXH3_128bits_withSeed(), XXH3_128bits_withSecret(): other seeding variants + * @see @ref single_shot_example "Single Shot Example" for an example. + */ +XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits(XXH_NOESCAPE const void* data, size_t len); +/*! @brief Calculates 128-bit seeded variant of XXH3 hash of @p data. + * + * @param data The block of data to be hashed, at least @p length bytes in size. + * @param len The length of @p data, in bytes. + * @param seed The 64-bit seed to alter the hash result predictably. + * + * @return The calculated 128-bit variant of XXH3 value. + * + * @note + * seed == 0 produces the same results as @ref XXH3_64bits(). + * + * This variant generates a custom secret on the fly based on default secret + * altered using the @p seed value. + * + * While this operation is decently fast, note that it's not completely free. + * + * @see XXH3_128bits(), XXH3_128bits_withSecret(): other seeding variants + * @see @ref single_shot_example "Single Shot Example" for an example. + */ +XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSeed(XXH_NOESCAPE const void* data, size_t len, XXH64_hash_t seed); +/*! + * @brief Calculates 128-bit variant of XXH3 with a custom "secret". + * + * @param data The block of data to be hashed, at least @p len bytes in size. + * @param len The length of @p data, in bytes. + * @param secret The secret data. + * @param secretSize The length of @p secret, in bytes. + * + * @return The calculated 128-bit variant of XXH3 value. + * + * 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 @p secretSize *must* be large enough (>= @ref XXH3_SECRET_SIZE_MIN). + * However, the quality of the secret impacts the dispersion of the hash algorithm. + * Therefore, the secret _must_ look like a bunch of random bytes. + * Avoid "trivial" or structured data such as repeated sequences or a text document. + * Whenever in doubt about the "randomness" of the blob of bytes, + * consider employing @ref XXH3_generateSecret() instead (see below). + * It will generate a proper high entropy secret derived from the blob of bytes. + * Another advantage of using XXH3_generateSecret() is that + * it guarantees that all bits within the initial blob of bytes + * will impact every bit of the output. + * This is not necessarily the case when using the blob of bytes directly + * because, when hashing _small_ inputs, only a portion of the secret is employed. + * + * @see @ref single_shot_example "Single Shot Example" for an example. + */ +XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize); /******* Streaming *******/ +#ifndef XXH_NO_STREAM /* * Streaming requires state maintenance. * This operation costs memory and CPU. @@ -944,73 +1438,193 @@ XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSecret(const void *data, * 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. + * 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); - -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); - -/* 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(): - * Return: 1 if `h1` and `h2` are equal, 0 if they are not. + * @brief Resets an @ref XXH3_state_t to begin a new hash. + * + * @param statePtr The state struct to reset. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note + * - This function resets `statePtr` and generate a secret with default parameters. + * - Call it before @ref XXH3_128bits_update(). + * - Digest will be equivalent to `XXH3_128bits()`. + * + * @see @ref streaming_example "Streaming Example" */ -XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2); +XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr); /*! - * XXH128_cmp(): + * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash. * - * This comparator is compatible with stdlib's `qsort()`/`bsearch()`. + * @param statePtr The state struct to reset. + * @param seed The 64-bit seed to alter the hash result predictably. * - * 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); - + * @pre + * @p statePtr must not be `NULL`. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note + * - This function resets `statePtr` and generate a secret from `seed`. + * - Call it before @ref XXH3_128bits_update(). + * - Digest will be equivalent to `XXH3_128bits_withSeed()`. + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed); +/*! + * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash. + * + * @param statePtr The state struct to reset. + * @param secret The secret data. + * @param secretSize The length of @p secret, in bytes. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * `secret` is referenced, it _must outlive_ the hash streaming session. + * Similar to one-shot API, `secretSize` must be >= @ref 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). + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize); + +/*! + * @brief Consumes a block of @p input to an @ref XXH3_state_t. + * + * Call this to incrementally consume blocks of data. + * + * @param statePtr The state struct to update. + * @param input The block of data to be hashed, at least @p length bytes in size. + * @param length The length of @p input, in bytes. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @note + * The memory between @p input and @p input + @p length must be valid, + * readable, contiguous memory. However, if @p length is `0`, @p input may be + * `NULL`. In C++, this also must be *TriviallyCopyable*. + * + */ +XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length); + +/*! + * @brief Returns the calculated XXH3 128-bit hash value from an @ref XXH3_state_t. + * + * @param statePtr The state struct to calculate the hash from. + * + * @pre + * @p statePtr must not be `NULL`. + * + * @return The calculated XXH3 128-bit hash value from that state. + * + * @note + * Calling XXH3_128bits_digest() will not affect @p statePtr, so you can update, + * digest, and update again. + * + */ +XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr); +#endif /* !XXH_NO_STREAM */ + +/* 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 */ + +/*! + * @brief Check equality of two XXH128_hash_t values + * + * @param h1 The 128-bit hash value. + * @param h2 Another 128-bit hash value. + * + * @return `1` if `h1` and `h2` are equal. + * @return `0` if they are not. + */ +XXH_PUBLIC_API XXH_PUREF int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2); + +/*! + * @brief Compares two @ref XXH128_hash_t + * + * This comparator is compatible with stdlib's `qsort()`/`bsearch()`. + * + * @param h128_1 Left-hand side value + * @param h128_2 Right-hand side value + * + * @return >0 if @p h128_1 > @p h128_2 + * @return =0 if @p h128_1 == @p h128_2 + * @return <0 if @p h128_1 < @p h128_2 + */ +XXH_PUBLIC_API XXH_PUREF int XXH128_cmp(XXH_NOESCAPE const void* h128_1, XXH_NOESCAPE const void* h128_2); + + /******* Canonical representation *******/ -typedef struct { +typedef struct { unsigned char digest[sizeof(XXH128_hash_t)]; } XXH128_canonical_t; - unsigned char digest[sizeof(XXH128_hash_t)]; -} XXH128_canonical_t; +/*! + * @brief Converts an @ref XXH128_hash_t to a big endian @ref XXH128_canonical_t. + * + * @param dst The @ref XXH128_canonical_t pointer to be stored to. + * @param hash The @ref XXH128_hash_t to be converted. + * + * @pre + * @p dst must not be `NULL`. + * @see @ref canonical_representation_example "Canonical Representation Example" + */ +XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash); -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); +/*! + * @brief Converts an @ref XXH128_canonical_t to a native @ref XXH128_hash_t. + * + * @param src The @ref XXH128_canonical_t to convert. + * + * @pre + * @p src must not be `NULL`. + * + * @return The converted hash. + * @see @ref canonical_representation_example "Canonical Representation Example" + */ +XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src); - #endif /* XXH_NO_LONG_LONG */ + +#endif /* !XXH_NO_XXH3 */ +#endif /* XXH_NO_LONG_LONG */ /*! * @} */ -#endif /* XXHASH_H_5627135585666179 */ +#endif /* XXHASH_H_5627135585666179 */ + + #if defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742) - #define 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 @@ -1031,23 +1645,16 @@ XXH128_hashFromCanonical(const XXH128_canonical_t *src); * @see XXH64_state_s, XXH3_state_s */ struct XXH32_state_s { + XXH32_hash_t total_len_32; /*!< Total length hashed, modulo 2^32 */ + XXH32_hash_t large_len; /*!< Whether the hash is >= 16 (handles @ref total_len_32 overflow) */ + XXH32_hash_t v[4]; /*!< Accumulator lanes */ + XXH32_hash_t mem32[4]; /*!< Internal buffer for partial reads. Treated as unsigned char[16]. */ + XXH32_hash_t memsize; /*!< Amount of data in @ref mem32 */ + XXH32_hash_t reserved; /*!< Reserved field. Do not read nor write to it. */ +}; /* typedef'd to XXH32_state_t */ - XXH32_hash_t total_len_32; /*!< Total length hashed, modulo 2^32 */ - XXH32_hash_t large_len; /*!< Whether the hash is >= 16 (handles @ref - total_len_32 overflow) */ - XXH32_hash_t v1; /*!< First accumulator lane */ - XXH32_hash_t v2; /*!< Second accumulator lane */ - XXH32_hash_t v3; /*!< Third accumulator lane */ - XXH32_hash_t v4; /*!< Fourth accumulator lane */ - XXH32_hash_t mem32[4]; /*!< Internal buffer for partial reads. Treated as - unsigned char[16]. */ - XXH32_hash_t memsize; /*!< Amount of data in @ref mem32 */ - XXH32_hash_t reserved; /*!< Reserved field. Do not read or write to it, it may - be removed. */ -}; /* typedef'd to XXH32_state_t */ - - #ifndef XXH_NO_LONG_LONG /* defined when there is no 64-bit support */ +#ifndef XXH_NO_LONG_LONG /* defined when there is no 64-bit support */ /*! * @internal @@ -1062,64 +1669,57 @@ struct XXH32_state_s { * @see XXH32_state_s, XXH3_state_s */ struct XXH64_state_s { + XXH64_hash_t total_len; /*!< Total length hashed. This is always 64-bit. */ + XXH64_hash_t v[4]; /*!< Accumulator lanes */ + XXH64_hash_t mem64[4]; /*!< Internal buffer for partial reads. Treated as unsigned char[32]. */ + XXH32_hash_t memsize; /*!< Amount of data in @ref mem64 */ + XXH32_hash_t reserved32; /*!< Reserved field, needed for padding anyways*/ + XXH64_hash_t reserved64; /*!< Reserved field. Do not read or write to it. */ +}; /* typedef'd to XXH64_state_t */ + +#ifndef XXH_NO_XXH3 + +#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* >= C11 */ +# include +# define XXH_ALIGN(n) alignas(n) +#elif defined(__cplusplus) && (__cplusplus >= 201103L) /* >= C++11 */ +/* In C++ alignas() is a keyword */ +# 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 - XXH64_hash_t total_len; /*!< Total length hashed. This is always 64-bit. */ - XXH64_hash_t v1; /*!< First accumulator lane */ - XXH64_hash_t v2; /*!< Second accumulator lane */ - XXH64_hash_t v3; /*!< Third accumulator lane */ - XXH64_hash_t v4; /*!< Fourth accumulator lane */ - XXH64_hash_t mem64[4]; /*!< Internal buffer for partial reads. Treated as - unsigned char[32]. */ - XXH32_hash_t memsize; /*!< Amount of data in @ref mem64 */ - XXH32_hash_t reserved32; /*!< Reserved field, needed for padding anyways*/ - XXH64_hash_t reserved64; /*!< Reserved field. Do not read or write to it, it - may be removed. */ - -}; /* typedef'd to XXH64_state_t */ - - #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* >= C11 \ - */ - #include - #define XXH_ALIGN(n) alignas(n) - #elif defined(__cplusplus) && (__cplusplus >= 201103L) /* >= C++11 */ - /* In C++ alignas() is a keyword */ - #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(__cplusplus) && (__cplusplus >= 201103L)) /* >= C++11 */ \ - && defined(__GNUC__) - #define XXH_ALIGN_MEMBER(align, type) type XXH_ALIGN(align) - #else - #define XXH_ALIGN_MEMBER(align, type) XXH_ALIGN(align) type - #endif - - /*! - * @brief The size of the internal XXH3 buffer. - * - * This is the optimal update size for incremental hashing. - * - * @see XXH3_64b_update(), XXH3_128b_update(). - */ - #define XXH3_INTERNALBUFFER_SIZE 256 +/* Old GCC versions only accept the attribute after the type in structures. */ +#if !(defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)) /* C11+ */ \ + && ! (defined(__cplusplus) && (__cplusplus >= 201103L)) /* >= C++11 */ \ + && defined(__GNUC__) +# define XXH_ALIGN_MEMBER(align, type) type XXH_ALIGN(align) +#else +# define XXH_ALIGN_MEMBER(align, type) XXH_ALIGN(align) type +#endif - /*! - * @brief Default size of the secret buffer (and @ref XXH3_kSecret). - * - * This is the size used in @ref XXH3_kSecret and the seeded functions. - * - * Not to be confused with @ref XXH3_SECRET_SIZE_MIN. - */ - #define XXH3_SECRET_DEFAULT_SIZE 192 +/*! + * @brief The size of the internal XXH3 buffer. + * + * This is the optimal update size for incremental hashing. + * + * @see XXH3_64b_update(), XXH3_128b_update(). + */ +#define XXH3_INTERNALBUFFER_SIZE 256 + +/*! + * @internal + * @brief Default size of the secret buffer (and @ref XXH3_kSecret). + * + * This is the size used in @ref XXH3_kSecret and the seeded functions. + * + * Not to be confused with @ref XXH3_SECRET_SIZE_MIN. + */ +#define XXH3_SECRET_DEFAULT_SIZE 192 /*! * @internal @@ -1144,111 +1744,284 @@ struct XXH64_state_s { * @see XXH32_state_s, XXH64_state_s */ struct XXH3_state_s { + XXH_ALIGN_MEMBER(64, XXH64_hash_t acc[8]); + /*!< The 8 accumulators. See @ref XXH32_state_s::v and @ref XXH64_state_s::v */ + XXH_ALIGN_MEMBER(64, unsigned char customSecret[XXH3_SECRET_DEFAULT_SIZE]); + /*!< Used to store a custom secret generated from a seed. */ + XXH_ALIGN_MEMBER(64, unsigned char buffer[XXH3_INTERNALBUFFER_SIZE]); + /*!< The internal buffer. @see XXH32_state_s::mem32 */ + XXH32_hash_t bufferedSize; + /*!< The amount of memory in @ref buffer, @see XXH32_state_s::memsize */ + XXH32_hash_t useSeed; + /*!< Reserved field. Needed for padding on 64-bit. */ + size_t nbStripesSoFar; + /*!< Number or stripes processed. */ + XXH64_hash_t totalLen; + /*!< Total length hashed. 64-bit even on 32-bit targets. */ + size_t nbStripesPerBlock; + /*!< Number of stripes per block. */ + size_t secretLimit; + /*!< Size of @ref customSecret or @ref extSecret */ + XXH64_hash_t seed; + /*!< Seed for _withSeed variants. Must be zero otherwise, @see XXH3_INITSTATE() */ + XXH64_hash_t reserved64; + /*!< Reserved field. */ + const unsigned char* extSecret; + /*!< Reference to an external secret for the _withSecret variants, NULL + * for other variants. */ + /* 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 - XXH_ALIGN_MEMBER(64, XXH64_hash_t acc[8]); - /*!< The 8 accumulators. Similar to `vN` in @ref XXH32_state_s::v1 and @ref - * XXH64_state_s */ - XXH_ALIGN_MEMBER(64, unsigned char customSecret[XXH3_SECRET_DEFAULT_SIZE]); - /*!< Used to store a custom secret generated from a seed. */ - XXH_ALIGN_MEMBER(64, unsigned char buffer[XXH3_INTERNALBUFFER_SIZE]); - /*!< The internal buffer. @see XXH32_state_s::mem32 */ - XXH32_hash_t bufferedSize; - /*!< The amount of memory in @ref buffer, @see XXH32_state_s::memsize */ - XXH32_hash_t reserved32; - /*!< Reserved field. Needed for padding on 64-bit. */ - size_t nbStripesSoFar; - /*!< Number or stripes processed. */ - XXH64_hash_t totalLen; - /*!< Total length hashed. 64-bit even on 32-bit targets. */ - size_t nbStripesPerBlock; - /*!< Number of stripes per block. */ - size_t secretLimit; - /*!< Size of @ref customSecret or @ref extSecret */ - XXH64_hash_t seed; - /*!< Seed for _withSeed variants. Must be zero otherwise, @see - * XXH3_INITSTATE() */ - XXH64_hash_t reserved64; - /*!< Reserved field. */ - const unsigned char *extSecret; - /*!< Reference to an external secret for the _withSecret variants, NULL - * for other variants. */ - /* 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 - - /*! - * @brief Initializes a stack-allocated `XXH3_state_s`. - * - * When the @ref 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; } +/*! + * @brief Initializes a stack-allocated `XXH3_state_s`. + * + * When the @ref 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) \ + do { \ + XXH3_state_t* tmp_xxh3_state_ptr = (XXH3_state_ptr); \ + tmp_xxh3_state_ptr->seed = 0; \ + tmp_xxh3_state_ptr->extSecret = NULL; \ + } while(0) -/* === Experimental API === */ -/* Symbols defined below must be considered tied to a specific library version. + +/*! + * @brief Calculates the 128-bit hash of @p data using XXH3. + * + * @param data The block of data to be hashed, at least @p len bytes in size. + * @param len The length of @p data, in bytes. + * @param seed The 64-bit seed to alter the hash's output predictably. + * + * @pre + * The memory between @p data and @p data + @p len must be valid, + * readable, contiguous memory. However, if @p len is `0`, @p data may be + * `NULL`. In C++, this also must be *TriviallyCopyable*. + * + * @return The calculated 128-bit XXH3 value. + * + * @see @ref single_shot_example "Single Shot Example" for an example. */ +XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128(XXH_NOESCAPE const void* data, size_t len, XXH64_hash_t seed); -/* - * XXH3_generateSecret(): + +/* === Experimental API === */ +/* Symbols defined below must be considered tied to a specific library version. */ + +/*! + * @brief Derive a high-entropy secret from any user-defined content, named customSeed. + * + * @param secretBuffer A writable buffer for derived high-entropy secret data. + * @param secretSize Size of secretBuffer, in bytes. Must be >= XXH3_SECRET_DEFAULT_SIZE. + * @param customSeed A user-defined content. + * @param customSeedSize Size of customSeed, in bytes. * - * 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. + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * 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. + * and derives from it a high-entropy secret of length @p secretSize into an + * already allocated buffer @p secretBuffer. * * 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()` + * The functions @ref XXH3_128bits_withSecret(), @ref XXH3_64bits_withSecret(), + * @ref XXH3_128bits_reset_withSecret() and @ref 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) + * which must be large enough for implementation reasons (>= @ref 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. + * These conditions can be a high bar to meet, so @ref XXH3_generateSecret() can + * be employed to ensure proper quality. + * + * @p customSeed can be anything. It can have any size, even small ones, + * and its content can be anything, even "poor entropy" sources such as a bunch + * of zeroes. The resulting `secret` will nonetheless provide all required qualities. * - * 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. + * @pre + * - @p secretSize must be >= @ref XXH3_SECRET_SIZE_MIN + * - When @p customSeedSize > 0, supplying NULL as customSeed is undefined behavior. * - * Supplying NULL as the customSeed copies the default secret into - * `secretBuffer`. When customSeedSize > 0, supplying NULL as customSeed is - * undefined behavior. + * Example code: + * @code{.c} + * #include + * #include + * #include + * #define XXH_STATIC_LINKING_ONLY // expose unstable API + * #include "xxhash.h" + * // Hashes argv[2] using the entropy from argv[1]. + * int main(int argc, char* argv[]) + * { + * char secret[XXH3_SECRET_SIZE_MIN]; + * if (argv != 3) { return 1; } + * XXH3_generateSecret(secret, sizeof(secret), argv[1], strlen(argv[1])); + * XXH64_hash_t h = XXH3_64bits_withSecret( + * argv[2], strlen(argv[2]), + * secret, sizeof(secret) + * ); + * printf("%016llx\n", (unsigned long long) h); + * } + * @endcode */ -XXH_PUBLIC_API void XXH3_generateSecret(void *secretBuffer, - const void *customSeed, - size_t customSeedSize); +XXH_PUBLIC_API XXH_errorcode XXH3_generateSecret(XXH_NOESCAPE void* secretBuffer, size_t secretSize, XXH_NOESCAPE 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); +/*! + * @brief Generate the same secret as the _withSeed() variants. + * + * @param secretBuffer A writable buffer of @ref XXH3_SECRET_SIZE_MIN bytes + * @param seed The 64-bit seed to alter the hash result predictably. + * + * The generated secret can be used in combination with + *`*_withSecret()` and `_withSecretandSeed()` variants. + * + * Example C++ `std::string` hash class: + * @code{.cpp} + * #include + * #define XXH_STATIC_LINKING_ONLY // expose unstable API + * #include "xxhash.h" + * // Slow, seeds each time + * class HashSlow { + * XXH64_hash_t seed; + * public: + * HashSlow(XXH64_hash_t s) : seed{s} {} + * size_t operator()(const std::string& x) const { + * return size_t{XXH3_64bits_withSeed(x.c_str(), x.length(), seed)}; + * } + * }; + * // Fast, caches the seeded secret for future uses. + * class HashFast { + * unsigned char secret[XXH3_SECRET_SIZE_MIN]; + * public: + * HashFast(XXH64_hash_t s) { + * XXH3_generateSecret_fromSeed(secret, seed); + * } + * size_t operator()(const std::string& x) const { + * return size_t{ + * XXH3_64bits_withSecret(x.c_str(), x.length(), secret, sizeof(secret)) + * }; + * } + * }; + * @endcode + */ +XXH_PUBLIC_API void XXH3_generateSecret_fromSeed(XXH_NOESCAPE void* secretBuffer, XXH64_hash_t seed); + +/*! + * @brief Calculates 64/128-bit seeded variant of XXH3 hash of @p data. + * + * @param data The block of data to be hashed, at least @p len bytes in size. + * @param len The length of @p data, in bytes. + * @param secret The secret data. + * @param secretSize The length of @p secret, in bytes. + * @param seed The 64-bit seed to alter the hash result predictably. + * + * These variants generate hash values using either + * @p seed for "short" keys (< @ref XXH3_MIDSIZE_MAX = 240 bytes) + * or @p secret for "large" keys (>= @ref XXH3_MIDSIZE_MAX). + * + * This generally benefits speed, compared to `_withSeed()` or `_withSecret()`. + * `_withSeed()` has to generate the secret on the fly for "large" keys. + * It's fast, but can be perceptible for "not so large" keys (< 1 KB). + * `_withSecret()` has to generate the masks on the fly for "small" keys, + * which requires more instructions than _withSeed() variants. + * Therefore, _withSecretandSeed variant combines the best of both worlds. + * + * When @p secret has been generated by XXH3_generateSecret_fromSeed(), + * this variant produces *exactly* the same results as `_withSeed()` variant, + * hence offering only a pure speed benefit on "large" input, + * by skipping the need to regenerate the secret for every large input. + * + * Another usage scenario is to hash the secret to a 64-bit hash value, + * for example with XXH3_64bits(), which then becomes the seed, + * and then employ both the seed and the secret in _withSecretandSeed(). + * On top of speed, an added benefit is that each bit in the secret + * has a 50% chance to swap each bit in the output, via its impact to the seed. + * + * This is not guaranteed when using the secret directly in "small data" scenarios, + * because only portions of the secret are employed for small data. + */ +XXH_PUBLIC_API XXH_PUREF XXH64_hash_t +XXH3_64bits_withSecretandSeed(XXH_NOESCAPE const void* data, size_t len, + XXH_NOESCAPE const void* secret, size_t secretSize, + XXH64_hash_t seed); +/*! + * @brief Calculates 128-bit seeded variant of XXH3 hash of @p data. + * + * @param input The block of data to be hashed, at least @p len bytes in size. + * @param length The length of @p data, in bytes. + * @param secret The secret data. + * @param secretSize The length of @p secret, in bytes. + * @param seed64 The 64-bit seed to alter the hash result predictably. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @see XXH3_64bits_withSecretandSeed() + */ +XXH_PUBLIC_API XXH_PUREF XXH128_hash_t +XXH3_128bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t length, + XXH_NOESCAPE const void* secret, size_t secretSize, + XXH64_hash_t seed64); +#ifndef XXH_NO_STREAM +/*! + * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash. + * + * @param statePtr A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState(). + * @param secret The secret data. + * @param secretSize The length of @p secret, in bytes. + * @param seed64 The 64-bit seed to alter the hash result predictably. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @see XXH3_64bits_withSecretandSeed() + */ +XXH_PUBLIC_API XXH_errorcode +XXH3_64bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, + XXH_NOESCAPE const void* secret, size_t secretSize, + XXH64_hash_t seed64); +/*! + * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash. + * + * @param statePtr A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState(). + * @param secret The secret data. + * @param secretSize The length of @p secret, in bytes. + * @param seed64 The 64-bit seed to alter the hash result predictably. + * + * @return @ref XXH_OK on success. + * @return @ref XXH_ERROR on failure. + * + * @see XXH3_64bits_withSecretandSeed() + */ +XXH_PUBLIC_API XXH_errorcode +XXH3_128bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, + XXH_NOESCAPE const void* secret, size_t secretSize, + XXH64_hash_t seed64); +#endif /* !XXH_NO_STREAM */ + +#endif /* !XXH_NO_XXH3 */ +#endif /* XXH_NO_LONG_LONG */ +#if defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) +# define XXH_IMPLEMENTATION +#endif - #endif /* XXH_NO_LONG_LONG */ - #if defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) - #define XXH_IMPLEMENTATION - #endif +#endif /* defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742) */ -#endif /* defined(XXH_STATIC_LINKING_ONLY) && \ - !defined(XXHASH_H_STATIC_13879238742) */ /* ======================================================================== */ /* ======================================================================== */ /* ======================================================================== */ + /*-********************************************************************** * xxHash implementation *-********************************************************************** @@ -1271,424 +2044,477 @@ XXH_PUBLIC_API XXH128_hash_t XXH128(const void *data, size_t len, * which can then be linked into the final binary. ************************************************************************/ -#if (defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) || \ - defined(XXH_IMPLEMENTATION)) && \ - !defined(XXH_IMPLEM_13a8737387) - #define XXH_IMPLEM_13a8737387 - - /* ************************************* - * Tuning parameters - ***************************************/ - - /*! - * @defgroup tuning Tuning parameters - * @{ - - * - * Various macros to control xxHash's behavior. - */ - #ifdef XXH_DOXYGEN - /*! - * @brief Define this to disable 64-bit code. - * - * Useful if only using the @ref xxh32_family and you have a strict C90 - * compiler. - */ - #define XXH_NO_LONG_LONG - #undef XXH_NO_LONG_LONG /* don't actually */ - /*! - * @brief Controls how unaligned memory is accessed. - * - * By default, access to unaligned memory is controlled by `memcpy()`, which - * is safe and portable. - * - * Unfortunately, on some target/compiler combinations, the generated - * assembly is sub-optimal. - * - * The below switch allow selection of a different access method - * in the search for improved performance. - * - * @par Possible options: - * - * - `XXH_FORCE_MEMORY_ACCESS=0` (default): `memcpy` - * @par - * Use `memcpy()`. Safe and portable. Note that most modern compilers - * will eliminate the function call and treat it as an unaligned access. - * - * - `XXH_FORCE_MEMORY_ACCESS=1`: `__attribute__((packed))` - * @par - * Depends on compiler extensions and is therefore not portable. - * This method is safe _if_ your compiler supports it, - * and *generally* as fast or faster than `memcpy`. - * - * - `XXH_FORCE_MEMORY_ACCESS=2`: Direct cast - * @par - * Casts directly and dereferences. This method doesn't depend on the - * compiler, but it violates the C standard as it directly dereferences - * an unaligned pointer. 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. - * - * - `XXH_FORCE_MEMORY_ACCESS=3`: Byteshift - * @par - * Also portable. 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 systems which lack a native byteswap instruction. However, - * some compilers will emit literal byteshifts even if the target supports - * unaligned access. - * . - * - * @warning - * Methods 1 and 2 rely on implementation-defined behavior. Use these with - * care, as what works on one compiler/platform/optimization level may - * cause another to read garbage data or even crash. - * - * See https://stackoverflow.com/a/32095106/646947 for details. - * - * Prefer these methods in priority order (0 > 3 > 1 > 2) - */ - #define XXH_FORCE_MEMORY_ACCESS 0 - /*! - * @def XXH_ACCEPT_NULL_INPUT_POINTER - * @brief Whether to add explicit `NULL` checks. - * - * If the input pointer is `NULL` and the length is non-zero, 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. If it is, the result for null input pointers is the same as a - * zero-length input. - */ - #define XXH_ACCEPT_NULL_INPUT_POINTER 0 - /*! - * @def XXH_FORCE_ALIGN_CHECK - * @brief If defined to non-zero, adds a special path for aligned inputs - * (XXH32() and XXH64() only). - * - * 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 reads, resulting in _dramatically - * faster_ read speed. - * - * The check costs one initial branch per hash, which is generally - * negligible, but not zero. - * - * Moreover, it's not useful to generate an additional code path if memory - * access uses the same instruction for both aligned and unaligned - * addresses (e.g. x86 and aarch64). - * - * 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). - */ - #define XXH_FORCE_ALIGN_CHECK 0 - - /*! - * @def XXH_NO_INLINE_HINTS - * @brief When non-zero, sets all functions to `static`. - * - * By default, xxHash tries to force the compiler to inline almost all - * internal functions. - * - * This can usually improve performance due to reduced jumping and improved - * constant folding, but significantly increases the size of the binary - * which might not be favorable. - * - * Additionally, sometimes the forced inlining can be detrimental to - * performance, depending on the architecture. - * - * XXH_NO_INLINE_HINTS marks all internal functions as static, giving the - * compiler full control on whether to inline or not. - * - * When not optimizing (-O0), optimizing for size (-Os, -Oz), or using - * -fno-inline with GCC or Clang, this will automatically be defined. - */ - #define XXH_NO_INLINE_HINTS 0 - - /*! - * @def XXH_REROLL - * @brief Whether to reroll `XXH32_finalize`. - * - * For performance, `XXH32_finalize` uses an unrolled loop - * in the form of a switch statement. - * - * This is not always desirable, as it generates larger code, - * and depending on the architecture, may even be slower - * - * This is automatically defined with `-Os`/`-Oz` on GCC and Clang. - */ - #define XXH_REROLL 0 +#if ( defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) \ + || defined(XXH_IMPLEMENTATION) ) && !defined(XXH_IMPLEM_13a8737387) +# define XXH_IMPLEM_13a8737387 - /*! - * @internal - * @brief Redefines old internal names. - * - * For compatibility with code that uses xxHash's internals before the names - * were changed to improve namespacing. There is no other reason to use - * this. - */ - #define XXH_OLD_NAMES - #undef XXH_OLD_NAMES /* don't actually use, it is ugly. */ - #endif /* XXH_DOXYGEN */ - /*! - * @} - */ - - #ifndef XXH_FORCE_MEMORY_ACCESS /* can be defined externally, on command \ - line for example */ - /* prefer __packed__ structures (method 1) for gcc on armv7+ and mips */ - #if !defined(__clang__) && \ - ((defined(__INTEL_COMPILER) && !defined(_WIN32)) || \ - (defined(__GNUC__) && \ - ((defined(__ARM_ARCH) && __ARM_ARCH >= 7) || \ - (defined(__mips__) && (__mips <= 5 || __mips_isa_rev < 6) && \ - (!defined(__mips16) || defined(__mips_mips16e2)))))) - #define XXH_FORCE_MEMORY_ACCESS 1 - #endif - #endif - - #ifndef XXH_ACCEPT_NULL_INPUT_POINTER /* can be defined externally */ - #define XXH_ACCEPT_NULL_INPUT_POINTER 0 - #endif - - #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 - #endif - #endif - - #ifndef XXH_NO_INLINE_HINTS - #if defined(__OPTIMIZE_SIZE__) /* -Os, -Oz */ \ - || defined(__NO_INLINE__) /* -O0, -fno-inline */ - #define XXH_NO_INLINE_HINTS 1 - #else - #define XXH_NO_INLINE_HINTS 0 - #endif - #endif - - #ifndef XXH_REROLL - #if defined(__OPTIMIZE_SIZE__) /* -Os, -Oz */ || \ - (defined(__GNUC__) && !defined(__clang__)) - /* The if/then loop is preferable to switch/case on gcc (on x64) */ - #define XXH_REROLL 1 - #else - #define XXH_REROLL 0 - #endif - #endif - - /*! - * @defgroup impl Implementation - * @{ - - */ - - /* ************************************* - * Includes & Memory related functions - ***************************************/ - /* - * Modify the local functions below should you wish to use - * different memory routines for malloc() and free() - */ - #include +/* ************************************* +* Tuning parameters +***************************************/ /*! - * @internal - * @brief Modify this function to use a different routine than malloc(). + * @defgroup tuning Tuning parameters + * @{ + * + * Various macros to control xxHash's behavior. */ -static void *XXH_malloc(size_t s) { - - return malloc(s); - -} - +#ifdef XXH_DOXYGEN /*! - * @internal - * @brief Modify this function to use a different routine than free(). + * @brief Define this to disable 64-bit code. + * + * Useful if only using the @ref XXH32_family and you have a strict C90 compiler. */ -static void XXH_free(void *p) { - - free(p); - -} - - #include - +# define XXH_NO_LONG_LONG +# undef XXH_NO_LONG_LONG /* don't actually */ /*! - * @internal - * @brief Modify this function to use a different routine than memcpy(). + * @brief Controls how unaligned memory is accessed. + * + * By default, access to unaligned memory is controlled by `memcpy()`, which is + * safe and portable. + * + * Unfortunately, on some target/compiler combinations, the generated assembly + * is sub-optimal. + * + * The below switch allow selection of a different access method + * in the search for improved performance. + * + * @par Possible options: + * + * - `XXH_FORCE_MEMORY_ACCESS=0` (default): `memcpy` + * @par + * Use `memcpy()`. Safe and portable. Note that most modern compilers will + * eliminate the function call and treat it as an unaligned access. + * + * - `XXH_FORCE_MEMORY_ACCESS=1`: `__attribute__((aligned(1)))` + * @par + * Depends on compiler extensions and is therefore not portable. + * This method is safe _if_ your compiler supports it, + * and *generally* as fast or faster than `memcpy`. + * + * - `XXH_FORCE_MEMORY_ACCESS=2`: Direct cast + * @par + * Casts directly and dereferences. This method doesn't depend on the + * compiler, but it violates the C standard as it directly dereferences an + * unaligned pointer. 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. + * + * - `XXH_FORCE_MEMORY_ACCESS=3`: Byteshift + * @par + * Also portable. 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 + * systems which lack a native byteswap instruction. However, some compilers + * will emit literal byteshifts even if the target supports unaligned access. + * + * + * @warning + * Methods 1 and 2 rely on implementation-defined behavior. Use these with + * care, as what works on one compiler/platform/optimization level may cause + * another to read garbage data or even crash. + * + * See https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html for details. + * + * Prefer these methods in priority order (0 > 3 > 1 > 2) */ -static void *XXH_memcpy(void *dest, const void *src, size_t size) { - - return memcpy(dest, src, size); - -} - - #include /* ULLONG_MAX */ - - /* ************************************* - * Compiler Specific Options - ***************************************/ - #ifdef _MSC_VER /* Visual Studio warning fix */ - #pragma warning(disable : 4127) /* disable: C4127: conditional expression \ - is constant */ - #endif - - #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 - /* 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 - #define XXH_FORCE_INLINE static - #define XXH_NO_INLINE static - #endif - - /* ************************************* - * Debug - ***************************************/ - /*! - * @ingroup tuning - * @def XXH_DEBUGLEVEL - * @brief Sets the debugging level. - * - * XXH_DEBUGLEVEL is expected to be defined externally, typically via the - * compiler's command line options. The value must be a number. - */ - #ifndef XXH_DEBUGLEVEL - #ifdef DEBUGLEVEL /* backwards compat */ - #define XXH_DEBUGLEVEL DEBUGLEVEL - #else - #define XXH_DEBUGLEVEL 0 - #endif - #endif - - #if (XXH_DEBUGLEVEL >= 1) - #include /* note: can still be disabled with NDEBUG */ - #define XXH_ASSERT(c) assert(c) - #else - #define XXH_ASSERT(c) ((void)0) - #endif - - /* note: use after variable declarations */ - #ifndef XXH_STATIC_ASSERT - #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* C11 */ - #include - #define XXH_STATIC_ASSERT_WITH_MESSAGE(c, m) \ - do { \ - \ - static_assert((c), m); \ - \ - } while (0) - - #elif defined(__cplusplus) && (__cplusplus >= 201103L) /* C++11 */ - #define XXH_STATIC_ASSERT_WITH_MESSAGE(c, m) \ - do { \ - \ - static_assert((c), m); \ - \ - } while (0) - - #else - #define XXH_STATIC_ASSERT_WITH_MESSAGE(c, m) \ - do { \ - \ - struct xxh_sa { \ - \ - char x[(c) ? 1 : -1]; \ - \ - }; \ - \ - } while (0) - - #endif - #define XXH_STATIC_ASSERT(c) XXH_STATIC_ASSERT_WITH_MESSAGE((c), #c) - #endif - - /*! - * @internal - * @def XXH_COMPILER_GUARD(var) - * @brief Used to prevent unwanted optimizations for @p var. - * - * It uses an empty GCC inline assembly statement with a register constraint - * which forces @p var into a general purpose register (eg eax, ebx, ecx - * on x86) and marks it as modified. - * - * This is used in a few places to avoid unwanted autovectorization (e.g. - * XXH32_round()). All vectorization we want is explicit via intrinsics, - * and _usually_ isn't wanted elsewhere. - * - * We also use it to prevent unwanted constant folding for AArch64 in - * XXH3_initCustomSecret_scalar(). - */ - #ifdef __GNUC__ - #define XXH_COMPILER_GUARD(var) __asm__ __volatile__("" : "+r"(var)) - #else - #define XXH_COMPILER_GUARD(var) ((void)0) - #endif - - /* ************************************* - * Basic Types - ***************************************/ - #if !defined(__VMS) && \ - (defined(__cplusplus) || \ - (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)) - #include -typedef uint8_t xxh_u8; - #else -typedef unsigned char xxh_u8; - #endif -typedef XXH32_hash_t xxh_u32; +# define XXH_FORCE_MEMORY_ACCESS 0 - #ifdef XXH_OLD_NAMES - #define BYTE xxh_u8 - #define U8 xxh_u8 - #define U32 xxh_u32 - #endif - -/* *** Memory access *** */ +/*! + * @def XXH_SIZE_OPT + * @brief Controls how much xxHash optimizes for size. + * + * xxHash, when compiled, tends to result in a rather large binary size. This + * is mostly due to heavy usage to forced inlining and constant folding of the + * @ref XXH3_family to increase performance. + * + * However, some developers prefer size over speed. This option can + * significantly reduce the size of the generated code. When using the `-Os` + * or `-Oz` options on GCC or Clang, this is defined to 1 by default, + * otherwise it is defined to 0. + * + * Most of these size optimizations can be controlled manually. + * + * This is a number from 0-2. + * - `XXH_SIZE_OPT` == 0: Default. xxHash makes no size optimizations. Speed + * comes first. + * - `XXH_SIZE_OPT` == 1: Default for `-Os` and `-Oz`. xxHash is more + * conservative and disables hacks that increase code size. It implies the + * options @ref XXH_NO_INLINE_HINTS == 1, @ref XXH_FORCE_ALIGN_CHECK == 0, + * and @ref XXH3_NEON_LANES == 8 if they are not already defined. + * - `XXH_SIZE_OPT` == 2: xxHash tries to make itself as small as possible. + * Performance may cry. For example, the single shot functions just use the + * streaming API. + */ +# define XXH_SIZE_OPT 0 /*! - * @internal - * @fn xxh_u32 XXH_read32(const void* ptr) - * @brief Reads an unaligned 32-bit integer from @p ptr in native endianness. + * @def XXH_FORCE_ALIGN_CHECK + * @brief If defined to non-zero, adds a special path for aligned inputs (XXH32() + * and XXH64() only). * - * Affected by @ref XXH_FORCE_MEMORY_ACCESS. + * This is an important performance trick for architectures without decent + * unaligned memory access performance. * - * @param ptr The pointer to read from. - * @return The 32-bit native endian integer from the bytes at @p ptr. + * It checks for input alignment, and when conditions are met, uses a "fast + * path" employing direct 32-bit/64-bit reads, resulting in _dramatically + * faster_ read speed. + * + * The check costs one initial branch per hash, which is generally negligible, + * but not zero. + * + * Moreover, it's not useful to generate an additional code path if memory + * access uses the same instruction for both aligned and unaligned + * addresses (e.g. x86 and aarch64). + * + * 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, and some ARM chips + * which are platforms known to offer good unaligned memory accesses performance. + * + * It is also disabled by default when @ref XXH_SIZE_OPT >= 1. + * + * This option does not affect XXH3 (only XXH32 and XXH64). */ +# define XXH_FORCE_ALIGN_CHECK 0 /*! - * @internal - * @fn xxh_u32 XXH_readLE32(const void* ptr) - * @brief Reads an unaligned 32-bit little endian integer from @p ptr. + * @def XXH_NO_INLINE_HINTS + * @brief When non-zero, sets all functions to `static`. * - * Affected by @ref XXH_FORCE_MEMORY_ACCESS. + * By default, xxHash tries to force the compiler to inline almost all internal + * functions. * - * @param ptr The pointer to read from. - * @return The 32-bit little endian integer from the bytes at @p ptr. + * This can usually improve performance due to reduced jumping and improved + * constant folding, but significantly increases the size of the binary which + * might not be favorable. + * + * Additionally, sometimes the forced inlining can be detrimental to performance, + * depending on the architecture. + * + * XXH_NO_INLINE_HINTS marks all internal functions as static, giving the + * compiler full control on whether to inline or not. + * + * When not optimizing (-O0), using `-fno-inline` with GCC or Clang, or if + * @ref XXH_SIZE_OPT >= 1, this will automatically be defined. + */ +# define XXH_NO_INLINE_HINTS 0 + +/*! + * @def XXH3_INLINE_SECRET + * @brief Determines whether to inline the XXH3 withSecret code. + * + * When the secret size is known, the compiler can improve the performance + * of XXH3_64bits_withSecret() and XXH3_128bits_withSecret(). + * + * However, if the secret size is not known, it doesn't have any benefit. This + * happens when xxHash is compiled into a global symbol. Therefore, if + * @ref XXH_INLINE_ALL is *not* defined, this will be defined to 0. + * + * Additionally, this defaults to 0 on GCC 12+, which has an issue with function pointers + * that are *sometimes* force inline on -Og, and it is impossible to automatically + * detect this optimization level. + */ +# define XXH3_INLINE_SECRET 0 + +/*! + * @def XXH32_ENDJMP + * @brief Whether to use a jump for `XXH32_finalize`. + * + * For performance, `XXH32_finalize` uses multiple branches in the finalizer. + * This is generally preferable for performance, + * but depending on exact architecture, a jmp may be preferable. + * + * This setting is only possibly making a difference for very small inputs. + */ +# define XXH32_ENDJMP 0 + +/*! + * @internal + * @brief Redefines old internal names. + * + * For compatibility with code that uses xxHash's internals before the names + * were changed to improve namespacing. There is no other reason to use this. + */ +# define XXH_OLD_NAMES +# undef XXH_OLD_NAMES /* don't actually use, it is ugly. */ + +/*! + * @def XXH_NO_STREAM + * @brief Disables the streaming API. + * + * When xxHash is not inlined and the streaming functions are not used, disabling + * the streaming functions can improve code size significantly, especially with + * the @ref XXH3_family which tends to make constant folded copies of itself. + */ +# define XXH_NO_STREAM +# undef XXH_NO_STREAM /* don't actually */ +#endif /* XXH_DOXYGEN */ +/*! + * @} + */ + +#ifndef XXH_FORCE_MEMORY_ACCESS /* can be defined externally, on command line for example */ + /* prefer __packed__ structures (method 1) for GCC + * < ARMv7 with unaligned access (e.g. Raspbian armhf) still uses byte shifting, so we use memcpy + * which for some reason does unaligned loads. */ +# if defined(__GNUC__) && !(defined(__ARM_ARCH) && __ARM_ARCH < 7 && defined(__ARM_FEATURE_UNALIGNED)) +# define XXH_FORCE_MEMORY_ACCESS 1 +# endif +#endif + +#ifndef XXH_SIZE_OPT + /* default to 1 for -Os or -Oz */ +# if (defined(__GNUC__) || defined(__clang__)) && defined(__OPTIMIZE_SIZE__) +# define XXH_SIZE_OPT 1 +# else +# define XXH_SIZE_OPT 0 +# endif +#endif + +#ifndef XXH_FORCE_ALIGN_CHECK /* can be defined externally */ + /* don't check on sizeopt, x86, aarch64, or arm when unaligned access is available */ +# if XXH_SIZE_OPT >= 1 || \ + defined(__i386) || defined(__x86_64__) || defined(__aarch64__) || defined(__ARM_FEATURE_UNALIGNED) \ + || defined(_M_IX86) || defined(_M_X64) || defined(_M_ARM64) || defined(_M_ARM) /* visual */ +# define XXH_FORCE_ALIGN_CHECK 0 +# else +# define XXH_FORCE_ALIGN_CHECK 1 +# endif +#endif + +#ifndef XXH_NO_INLINE_HINTS +# if XXH_SIZE_OPT >= 1 || defined(__NO_INLINE__) /* -O0, -fno-inline */ +# define XXH_NO_INLINE_HINTS 1 +# else +# define XXH_NO_INLINE_HINTS 0 +# endif +#endif + +#ifndef XXH3_INLINE_SECRET +# if (defined(__GNUC__) && !defined(__clang__) && __GNUC__ >= 12) \ + || !defined(XXH_INLINE_ALL) +# define XXH3_INLINE_SECRET 0 +# else +# define XXH3_INLINE_SECRET 1 +# endif +#endif + +#ifndef XXH32_ENDJMP +/* generally preferable for performance */ +# define XXH32_ENDJMP 0 +#endif + +/*! + * @defgroup impl Implementation + * @{ + */ + + +/* ************************************* +* Includes & Memory related functions +***************************************/ +#if defined(XXH_NO_STREAM) +/* nothing */ +#elif defined(XXH_NO_STDLIB) + +/* When requesting to disable any mention of stdlib, + * the library loses the ability to invoked malloc / free. + * In practice, it means that functions like `XXH*_createState()` + * will always fail, and return NULL. + * This flag is useful in situations where + * xxhash.h is integrated into some kernel, embedded or limited environment + * without access to dynamic allocation. + */ + +static XXH_CONSTF void* XXH_malloc(size_t s) { (void)s; return NULL; } +static void XXH_free(void* p) { (void)p; } + +#else + +/* + * Modify the local functions below should you wish to use + * different memory routines for malloc() and free() + */ +#include + +/*! + * @internal + * @brief Modify this function to use a different routine than malloc(). + */ +static XXH_MALLOCF void* XXH_malloc(size_t s) { return malloc(s); } + +/*! + * @internal + * @brief Modify this function to use a different routine than free(). + */ +static void XXH_free(void* p) { free(p); } + +#endif /* XXH_NO_STDLIB */ + +#include + +/*! + * @internal + * @brief Modify this function to use a different routine than memcpy(). + */ +static void* XXH_memcpy(void* dest, const void* src, size_t size) +{ + return memcpy(dest,src,size); +} + +#include /* ULLONG_MAX */ + + +/* ************************************* +* Compiler Specific Options +***************************************/ +#ifdef _MSC_VER /* Visual Studio warning fix */ +# pragma warning(disable : 4127) /* disable: C4127: conditional expression is constant */ +#endif + +#if XXH_NO_INLINE_HINTS /* disable inlining hints */ +# if defined(__GNUC__) || defined(__clang__) +# define XXH_FORCE_INLINE static __attribute__((unused)) +# else +# define XXH_FORCE_INLINE static +# endif +# define XXH_NO_INLINE static +/* enable inlining hints */ +#elif defined(__GNUC__) || defined(__clang__) +# define XXH_FORCE_INLINE static __inline__ __attribute__((always_inline, unused)) +# define XXH_NO_INLINE static __attribute__((noinline)) +#elif defined(_MSC_VER) /* Visual Studio */ +# define XXH_FORCE_INLINE static __forceinline +# define XXH_NO_INLINE static __declspec(noinline) +#elif defined (__cplusplus) \ + || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L)) /* C99 */ +# define XXH_FORCE_INLINE static inline +# define XXH_NO_INLINE static +#else +# define XXH_FORCE_INLINE static +# define XXH_NO_INLINE static +#endif + +#if XXH3_INLINE_SECRET +# define XXH3_WITH_SECRET_INLINE XXH_FORCE_INLINE +#else +# define XXH3_WITH_SECRET_INLINE XXH_NO_INLINE +#endif + + +/* ************************************* +* Debug +***************************************/ +/*! + * @ingroup tuning + * @def XXH_DEBUGLEVEL + * @brief Sets the debugging level. + * + * XXH_DEBUGLEVEL is expected to be defined externally, typically via the + * compiler's command line options. The value must be a number. + */ +#ifndef XXH_DEBUGLEVEL +# ifdef DEBUGLEVEL /* backwards compat */ +# define XXH_DEBUGLEVEL DEBUGLEVEL +# else +# define XXH_DEBUGLEVEL 0 +# endif +#endif + +#if (XXH_DEBUGLEVEL>=1) +# include /* note: can still be disabled with NDEBUG */ +# define XXH_ASSERT(c) assert(c) +#else +# if defined(__INTEL_COMPILER) +# define XXH_ASSERT(c) XXH_ASSUME((unsigned char) (c)) +# else +# define XXH_ASSERT(c) XXH_ASSUME(c) +# endif +#endif + +/* note: use after variable declarations */ +#ifndef XXH_STATIC_ASSERT +# if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* C11 */ +# define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { _Static_assert((c),m); } while(0) +# elif defined(__cplusplus) && (__cplusplus >= 201103L) /* C++11 */ +# define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { static_assert((c),m); } while(0) +# else +# define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { struct xxh_sa { char x[(c) ? 1 : -1]; }; } while(0) +# endif +# define XXH_STATIC_ASSERT(c) XXH_STATIC_ASSERT_WITH_MESSAGE((c),#c) +#endif + +/*! + * @internal + * @def XXH_COMPILER_GUARD(var) + * @brief Used to prevent unwanted optimizations for @p var. + * + * It uses an empty GCC inline assembly statement with a register constraint + * which forces @p var into a general purpose register (eg eax, ebx, ecx + * on x86) and marks it as modified. + * + * This is used in a few places to avoid unwanted autovectorization (e.g. + * XXH32_round()). All vectorization we want is explicit via intrinsics, + * and _usually_ isn't wanted elsewhere. + * + * We also use it to prevent unwanted constant folding for AArch64 in + * XXH3_initCustomSecret_scalar(). + */ +#if defined(__GNUC__) || defined(__clang__) +# define XXH_COMPILER_GUARD(var) __asm__("" : "+r" (var)) +#else +# define XXH_COMPILER_GUARD(var) ((void)0) +#endif + +/* Specifically for NEON vectors which use the "w" constraint, on + * Clang. */ +#if defined(__clang__) && defined(__ARM_ARCH) && !defined(__wasm__) +# define XXH_COMPILER_GUARD_CLANG_NEON(var) __asm__("" : "+w" (var)) +#else +# define XXH_COMPILER_GUARD_CLANG_NEON(var) ((void)0) +#endif + +/* ************************************* +* Basic Types +***************************************/ +#if !defined (__VMS) \ + && (defined (__cplusplus) \ + || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) ) +# include + typedef uint8_t xxh_u8; +#else + typedef unsigned char xxh_u8; +#endif +typedef XXH32_hash_t xxh_u32; + +#ifdef XXH_OLD_NAMES +# warning "XXH_OLD_NAMES is planned to be removed starting v0.9. If the program depends on it, consider moving away from it by employing newer type names directly" +# define BYTE xxh_u8 +# define U8 xxh_u8 +# define U32 xxh_u32 +#endif + +/* *** Memory access *** */ + +/*! + * @internal + * @fn xxh_u32 XXH_read32(const void* ptr) + * @brief Reads an unaligned 32-bit integer from @p ptr in native endianness. + * + * Affected by @ref XXH_FORCE_MEMORY_ACCESS. + * + * @param ptr The pointer to read from. + * @return The 32-bit native endian integer from the bytes at @p ptr. + */ + +/*! + * @internal + * @fn xxh_u32 XXH_readLE32(const void* ptr) + * @brief Reads an unaligned 32-bit little endian integer from @p ptr. + * + * Affected by @ref XXH_FORCE_MEMORY_ACCESS. + * + * @param ptr The pointer to read from. + * @return The 32-bit little endian integer from the bytes at @p ptr. */ /*! @@ -1719,288 +2545,304 @@ typedef XXH32_hash_t xxh_u32; * @return The 32-bit little endian integer from the bytes at @p ptr. */ - #if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS == 3)) - /* - * Manual byteshift. Best for old compilers which don't inline memcpy. - * We actually directly use XXH_readLE32 and XXH_readBE32. - */ - #elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS == 2)) +#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3)) +/* + * Manual byteshift. Best for old compilers which don't inline memcpy. + * We actually directly use XXH_readLE32 and XXH_readBE32. + */ +#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2)) /* * Force direct memory access. Only works on CPU which support unaligned memory * access in hardware. */ -static xxh_u32 XXH_read32(const void *memPtr) { +static xxh_u32 XXH_read32(const void* memPtr) { return *(const xxh_u32*) memPtr; } - return *(const xxh_u32 *)memPtr; +#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1)) +/* + * __attribute__((aligned(1))) is supported by gcc and clang. Originally the + * documentation claimed that it only increased the alignment, but actually it + * can decrease it on gcc, clang, and icc: + * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502, + * https://gcc.godbolt.org/z/xYez1j67Y. + */ +#ifdef XXH_OLD_NAMES +typedef union { xxh_u32 u32; } __attribute__((packed)) unalign; +#endif +static xxh_u32 XXH_read32(const void* ptr) +{ + typedef __attribute__((aligned(1))) xxh_u32 xxh_unalign32; + return *((const xxh_unalign32*)ptr); } - #elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS == 1)) +#else - /* - * __pack instructions are safer but compiler specific, hence potentially - * problematic for some compilers. - * - * Currently only defined for GCC and ICC. - */ - #ifdef XXH_OLD_NAMES -typedef union { +/* + * Portable and safe solution. Generally efficient. + * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html + */ +static xxh_u32 XXH_read32(const void* memPtr) +{ + xxh_u32 val; + XXH_memcpy(&val, memPtr, sizeof(val)); + return val; +} - xxh_u32 u32; +#endif /* XXH_FORCE_DIRECT_MEMORY_ACCESS */ -} __attribute__((packed)) unalign; - #endif -static xxh_u32 XXH_read32(const void *ptr) { +/* *** Endianness *** */ - typedef union { +/*! + * @ingroup tuning + * @def XXH_CPU_LITTLE_ENDIAN + * @brief Whether the target is little endian. + * + * Defined to 1 if the target is little endian, or 0 if it is big endian. + * It can be defined externally, for example on the compiler command line. + * + * If it is not defined, + * a runtime check (which is usually constant folded) is used instead. + * + * @note + * This is not necessarily defined to an integer constant. + * + * @see XXH_isLittleEndian() for the runtime check. + */ +#ifndef XXH_CPU_LITTLE_ENDIAN +/* + * Try to detect endianness automatically, to avoid the nonstandard behavior + * in `XXH_isLittleEndian()` + */ +# if defined(_WIN32) /* Windows is always little endian */ \ + || defined(__LITTLE_ENDIAN__) \ + || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) +# define XXH_CPU_LITTLE_ENDIAN 1 +# elif defined(__BIG_ENDIAN__) \ + || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) +# define XXH_CPU_LITTLE_ENDIAN 0 +# else +/*! + * @internal + * @brief Runtime check for @ref XXH_CPU_LITTLE_ENDIAN. + * + * Most compilers will constant fold this. + */ +static int XXH_isLittleEndian(void) +{ + /* + * 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 }; + return one.c[0]; +} +# define XXH_CPU_LITTLE_ENDIAN XXH_isLittleEndian() +# endif +#endif - xxh_u32 u32; - } __attribute__((packed)) xxh_unalign; - return ((const xxh_unalign *)ptr)->u32; -} +/* **************************************** +* Compiler-specific Functions and Macros +******************************************/ +#define XXH_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__) + +#ifdef __has_builtin +# define XXH_HAS_BUILTIN(x) __has_builtin(x) +#else +# define XXH_HAS_BUILTIN(x) 0 +#endif + - #else /* - * Portable and safe solution. Generally efficient. - * see: https://stackoverflow.com/a/32095106/646947 + * C23 and future versions have standard "unreachable()". + * Once it has been implemented reliably we can add it as an + * additional case: + * + * ``` + * #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN) + * # include + * # ifdef unreachable + * # define XXH_UNREACHABLE() unreachable() + * # endif + * #endif + * ``` + * + * Note C++23 also has std::unreachable() which can be detected + * as follows: + * ``` + * #if defined(__cpp_lib_unreachable) && (__cpp_lib_unreachable >= 202202L) + * # include + * # define XXH_UNREACHABLE() std::unreachable() + * #endif + * ``` + * NB: `__cpp_lib_unreachable` is defined in the `` header. + * We don't use that as including `` in `extern "C"` blocks + * doesn't work on GCC12 */ -static xxh_u32 XXH_read32(const void *memPtr) { - xxh_u32 val; - memcpy(&val, memPtr, sizeof(val)); - return val; +#if XXH_HAS_BUILTIN(__builtin_unreachable) +# define XXH_UNREACHABLE() __builtin_unreachable() -} +#elif defined(_MSC_VER) +# define XXH_UNREACHABLE() __assume(0) - #endif /* XXH_FORCE_DIRECT_MEMORY_ACCESS */ +#else +# define XXH_UNREACHABLE() +#endif - /* *** Endianness *** */ +#if XXH_HAS_BUILTIN(__builtin_assume) +# define XXH_ASSUME(c) __builtin_assume(c) +#else +# define XXH_ASSUME(c) if (!(c)) { XXH_UNREACHABLE(); } +#endif - /*! - * @ingroup tuning - * @def XXH_CPU_LITTLE_ENDIAN - * @brief Whether the target is little endian. - * - * Defined to 1 if the target is little endian, or 0 if it is big endian. - * It can be defined externally, for example on the compiler command line. - * - * If it is not defined, - * a runtime check (which is usually constant folded) is used instead. - * - * @note - * This is not necessarily defined to an integer constant. - * - * @see XXH_isLittleEndian() for the runtime check. - */ - #ifndef XXH_CPU_LITTLE_ENDIAN - /* - * Try to detect endianness automatically, to avoid the nonstandard behavior - * in `XXH_isLittleEndian()` - */ - #if defined(_WIN32) /* Windows is always little endian */ \ - || defined(__LITTLE_ENDIAN__) || \ - (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) - #define XXH_CPU_LITTLE_ENDIAN 1 - #elif defined(__BIG_ENDIAN__) || \ - (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) - #define XXH_CPU_LITTLE_ENDIAN 0 - #else /*! * @internal - * @brief Runtime check for @ref XXH_CPU_LITTLE_ENDIAN. + * @def XXH_rotl32(x,r) + * @brief 32-bit rotate left. * - * Most compilers will constant fold this. + * @param x The 32-bit integer to be rotated. + * @param r The number of bits to rotate. + * @pre + * @p r > 0 && @p r < 32 + * @note + * @p x and @p r may be evaluated multiple times. + * @return The rotated result. + */ +#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 */ +#elif defined(_MSC_VER) +# define XXH_rotl32(x,r) _rotl(x,r) +# define XXH_rotl64(x,r) _rotl64(x,r) +#else +# define XXH_rotl32(x,r) (((x) << (r)) | ((x) >> (32 - (r)))) +# define XXH_rotl64(x,r) (((x) << (r)) | ((x) >> (64 - (r)))) +#endif + +/*! + * @internal + * @fn xxh_u32 XXH_swap32(xxh_u32 x) + * @brief A 32-bit byteswap. + * + * @param x The 32-bit integer to byteswap. + * @return @p x, byteswapped. */ -static int XXH_isLittleEndian(void) { - - /* - * 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}; - - return one.c[0]; - -} - - #define XXH_CPU_LITTLE_ENDIAN XXH_isLittleEndian() - #endif - #endif - - /* **************************************** - * Compiler-specific Functions and Macros - ******************************************/ - #define XXH_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__) - - #ifdef __has_builtin - #define XXH_HAS_BUILTIN(x) __has_builtin(x) - #else - #define XXH_HAS_BUILTIN(x) 0 - #endif - - /*! - * @internal - * @def XXH_rotl32(x,r) - * @brief 32-bit rotate left. - * - * @param x The 32-bit integer to be rotated. - * @param r The number of bits to rotate. - * @pre - * @p r > 0 && @p r < 32 - * @note - * @p x and @p r may be evaluated multiple times. - * @return The rotated result. - */ - #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 */ - #elif defined(_MSC_VER) - #define XXH_rotl32(x, r) _rotl(x, r) - #define XXH_rotl64(x, r) _rotl64(x, r) - #else - #define XXH_rotl32(x, r) (((x) << (r)) | ((x) >> (32 - (r)))) - #define XXH_rotl64(x, r) (((x) << (r)) | ((x) >> (64 - (r)))) - #endif - - /*! - * @internal - * @fn xxh_u32 XXH_swap32(xxh_u32 x) - * @brief A 32-bit byteswap. - * - * @param x The 32-bit integer to byteswap. - * @return @p x, byteswapped. - */ - #if defined(_MSC_VER) /* Visual Studio */ - #define XXH_swap32 _byteswap_ulong - #elif XXH_GCC_VERSION >= 403 - #define XXH_swap32 __builtin_bswap32 - #else -static xxh_u32 XXH_swap32(xxh_u32 x) { - - return ((x << 24) & 0xff000000) | ((x << 8) & 0x00ff0000) | - ((x >> 8) & 0x0000ff00) | ((x >> 24) & 0x000000ff); - -} - - #endif +#if defined(_MSC_VER) /* Visual Studio */ +# define XXH_swap32 _byteswap_ulong +#elif XXH_GCC_VERSION >= 403 +# define XXH_swap32 __builtin_bswap32 +#else +static xxh_u32 XXH_swap32 (xxh_u32 x) +{ + return ((x << 24) & 0xff000000 ) | + ((x << 8) & 0x00ff0000 ) | + ((x >> 8) & 0x0000ff00 ) | + ((x >> 24) & 0x000000ff ); +} +#endif + /* *************************** - * Memory reads - *****************************/ +* Memory reads +*****************************/ /*! * @internal * @brief Enum to indicate whether a pointer is aligned. */ typedef enum { - - XXH_aligned, /*!< Aligned */ - XXH_unaligned /*!< Possibly unaligned */ - + XXH_aligned, /*!< Aligned */ + XXH_unaligned /*!< Possibly unaligned */ } XXH_alignment; - /* - * XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load. - * - * This is ideal for older compilers which don't inline memcpy. - */ - #if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS == 3)) - -XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void *memPtr) { - - const xxh_u8 *bytePtr = (const xxh_u8 *)memPtr; - return bytePtr[0] | ((xxh_u32)bytePtr[1] << 8) | ((xxh_u32)bytePtr[2] << 16) | - ((xxh_u32)bytePtr[3] << 24); +/* + * XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load. + * + * This is ideal for older compilers which don't inline memcpy. + */ +#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3)) +XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* memPtr) +{ + const xxh_u8* bytePtr = (const xxh_u8 *)memPtr; + return bytePtr[0] + | ((xxh_u32)bytePtr[1] << 8) + | ((xxh_u32)bytePtr[2] << 16) + | ((xxh_u32)bytePtr[3] << 24); } -XXH_FORCE_INLINE xxh_u32 XXH_readBE32(const void *memPtr) { - - const xxh_u8 *bytePtr = (const xxh_u8 *)memPtr; - return bytePtr[3] | ((xxh_u32)bytePtr[2] << 8) | ((xxh_u32)bytePtr[1] << 16) | - ((xxh_u32)bytePtr[0] << 24); - +XXH_FORCE_INLINE xxh_u32 XXH_readBE32(const void* memPtr) +{ + const xxh_u8* bytePtr = (const xxh_u8 *)memPtr; + return bytePtr[3] + | ((xxh_u32)bytePtr[2] << 8) + | ((xxh_u32)bytePtr[1] << 16) + | ((xxh_u32)bytePtr[0] << 24); } - #else -XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void *ptr) { - - return XXH_CPU_LITTLE_ENDIAN ? XXH_read32(ptr) : XXH_swap32(XXH_read32(ptr)); - +#else +XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* ptr) +{ + return XXH_CPU_LITTLE_ENDIAN ? XXH_read32(ptr) : XXH_swap32(XXH_read32(ptr)); } -static xxh_u32 XXH_readBE32(const void *ptr) { - - return XXH_CPU_LITTLE_ENDIAN ? XXH_swap32(XXH_read32(ptr)) : XXH_read32(ptr); - +static xxh_u32 XXH_readBE32(const void* ptr) +{ + return XXH_CPU_LITTLE_ENDIAN ? XXH_swap32(XXH_read32(ptr)) : XXH_read32(ptr); } +#endif - #endif - -XXH_FORCE_INLINE xxh_u32 XXH_readLE32_align(const void *ptr, - XXH_alignment align) { - - if (align == XXH_unaligned) { - - return XXH_readLE32(ptr); - - } else { - - return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u32 *)ptr - : XXH_swap32(*(const xxh_u32 *)ptr); - - } - +XXH_FORCE_INLINE xxh_u32 +XXH_readLE32_align(const void* ptr, XXH_alignment align) +{ + if (align==XXH_unaligned) { + return XXH_readLE32(ptr); + } else { + return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u32*)ptr : XXH_swap32(*(const xxh_u32*)ptr); + } } + /* ************************************* - * Misc - ***************************************/ +* Misc +***************************************/ /*! @ingroup public */ -XXH_PUBLIC_API unsigned XXH_versionNumber(void) { +XXH_PUBLIC_API unsigned XXH_versionNumber (void) { return XXH_VERSION_NUMBER; } - return XXH_VERSION_NUMBER; - -} /* ******************************************************************* - * 32-bit hash functions - *********************************************************************/ +* 32-bit hash functions +*********************************************************************/ /*! * @} - * @defgroup xxh32_impl XXH32 implementation + * @defgroup XXH32_impl XXH32 implementation * @ingroup impl + * + * Details on the XXH32 implementation. * @{ - */ -/* #define instead of static const, to be used as initializers */ - #define XXH_PRIME32_1 0x9E3779B1U /*!< 0b10011110001101110111100110110001 */ - #define XXH_PRIME32_2 0x85EBCA77U /*!< 0b10000101111010111100101001110111 */ - #define XXH_PRIME32_3 0xC2B2AE3DU /*!< 0b11000010101100101010111000111101 */ - #define XXH_PRIME32_4 0x27D4EB2FU /*!< 0b00100111110101001110101100101111 */ - #define 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 + /* #define instead of static const, to be used as initializers */ +#define XXH_PRIME32_1 0x9E3779B1U /*!< 0b10011110001101110111100110110001 */ +#define XXH_PRIME32_2 0x85EBCA77U /*!< 0b10000101111010111100101001110111 */ +#define XXH_PRIME32_3 0xC2B2AE3DU /*!< 0b11000010101100101010111000111101 */ +#define XXH_PRIME32_4 0x27D4EB2FU /*!< 0b00100111110101001110101100101111 */ +#define 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 /*! * @internal @@ -2013,50 +2855,51 @@ XXH_PUBLIC_API unsigned XXH_versionNumber(void) { * @param input The stripe of input to mix. * @return The mixed accumulator lane. */ -static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input) { - - acc += input * XXH_PRIME32_2; - acc = XXH_rotl32(acc, 13); - acc *= XXH_PRIME32_1; - #if (defined(__SSE4_1__) || defined(__aarch64__)) && \ - !defined(XXH_ENABLE_AUTOVECTORIZE) - /* - * UGLY HACK: - * A compiler fence is the only thing that prevents GCC and Clang from - * autovectorizing the XXH32 loop (pragmas and attributes don't work for some - * reason) without globally disabling SSE4.1. - * - * The reason we want to avoid vectorization is because despite working on - * 4 integers at a time, there are multiple factors slowing XXH32 down on - * SSE4: - * - There's a ridiculous amount of lag from pmulld (10 cycles of latency on - * newer chips!) making it slightly slower to multiply four integers at - * once compared to four integers independently. Even when pmulld was - * fastest, Sandy/Ivy Bridge, it is still not worth it to go into SSE - * just to multiply unless doing a long operation. - * - * - Four instructions are required to rotate, - * movqda tmp, v // not required with VEX encoding - * pslld tmp, 13 // tmp <<= 13 - * psrld v, 19 // x >>= 19 - * por v, tmp // x |= tmp - * compared to one for scalar: - * roll v, 13 // reliably fast across the board - * shldl v, v, 13 // Sandy Bridge and later prefer this for some reason - * - * - Instruction level parallelism is actually more beneficial here because - * the SIMD actually serializes this operation: While v1 is rotating, v2 - * can load data, while v3 can multiply. SSE forces them to operate - * together. - * - * This is also enabled on AArch64, as Clang autovectorizes it incorrectly - * and it is pointless writing a NEON implementation that is basically the - * same speed as scalar for XXH32. - */ - XXH_COMPILER_GUARD(acc); - #endif - return acc; - +static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input) +{ + acc += input * XXH_PRIME32_2; + acc = XXH_rotl32(acc, 13); + acc *= XXH_PRIME32_1; +#if (defined(__SSE4_1__) || defined(__aarch64__) || defined(__wasm_simd128__)) && !defined(XXH_ENABLE_AUTOVECTORIZE) + /* + * UGLY HACK: + * A compiler fence is the only thing that prevents GCC and Clang from + * autovectorizing the XXH32 loop (pragmas and attributes don't work for some + * reason) without globally disabling SSE4.1. + * + * The reason we want to avoid vectorization is because despite working on + * 4 integers at a time, there are multiple factors slowing XXH32 down on + * SSE4: + * - There's a ridiculous amount of lag from pmulld (10 cycles of latency on + * newer chips!) making it slightly slower to multiply four integers at + * once compared to four integers independently. Even when pmulld was + * fastest, Sandy/Ivy Bridge, it is still not worth it to go into SSE + * just to multiply unless doing a long operation. + * + * - Four instructions are required to rotate, + * movqda tmp, v // not required with VEX encoding + * pslld tmp, 13 // tmp <<= 13 + * psrld v, 19 // x >>= 19 + * por v, tmp // x |= tmp + * compared to one for scalar: + * roll v, 13 // reliably fast across the board + * shldl v, v, 13 // Sandy Bridge and later prefer this for some reason + * + * - Instruction level parallelism is actually more beneficial here because + * the SIMD actually serializes this operation: While v1 is rotating, v2 + * can load data, while v3 can multiply. SSE forces them to operate + * together. + * + * This is also enabled on AArch64, as Clang is *very aggressive* in vectorizing + * the loop. NEON is only faster on the A53, and with the newer cores, it is less + * than half the speed. + * + * Additionally, this is used on WASM SIMD128 because it JITs to the same + * SIMD instructions and has the same issue. + */ + XXH_COMPILER_GUARD(acc); +#endif + return acc; } /*! @@ -2066,38 +2909,20 @@ static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input) { * The final mix ensures that all input bits have a chance to impact any bit in * the output digest, resulting in an unbiased distribution. * - * @param h32 The hash to avalanche. + * @param hash The hash to avalanche. * @return The avalanched hash. */ -static xxh_u32 XXH32_avalanche(xxh_u32 h32) { - - h32 ^= h32 >> 15; - h32 *= XXH_PRIME32_2; - h32 ^= h32 >> 13; - h32 *= XXH_PRIME32_3; - h32 ^= h32 >> 16; - return (h32); - +static xxh_u32 XXH32_avalanche(xxh_u32 hash) +{ + hash ^= hash >> 15; + hash *= XXH_PRIME32_2; + hash ^= hash >> 13; + hash *= XXH_PRIME32_3; + hash ^= hash >> 16; + return hash; } - #define XXH_get32bits(p) XXH_readLE32_align(p, align) - - #define XXH_PROCESS1 \ - do { \ - \ - h32 += (*ptr++) * XXH_PRIME32_5; \ - h32 = XXH_rotl32(h32, 11) * XXH_PRIME32_1; \ - \ - } while (0) - - #define XXH_PROCESS4 \ - do { \ - \ - h32 += XXH_get32bits(ptr) * XXH_PRIME32_3; \ - ptr += 4; \ - h32 = XXH_rotl32(h32, 17) * XXH_PRIME32_4; \ - \ - } while (0) +#define XXH_get32bits(p) XXH_readLE32_align(p, align) /*! * @internal @@ -2107,1515 +2932,1366 @@ static xxh_u32 XXH32_avalanche(xxh_u32 h32) { * This final stage will digest them to ensure that all input bytes are present * in the final mix. * - * @param h32 The hash to finalize. + * @param hash The hash to finalize. * @param ptr The pointer to the remaining input. * @param len The remaining length, modulo 16. * @param align Whether @p ptr is aligned. * @return The finalized hash. + * @see XXH64_finalize(). */ -static xxh_u32 XXH32_finalize(xxh_u32 h32, const xxh_u8 *ptr, size_t len, - XXH_alignment align) { - - /* Compact rerolled version */ - if (XXH_REROLL) { - - len &= 15; - while (len >= 4) { - - XXH_PROCESS4; - len -= 4; - - } - - while (len > 0) { - - XXH_PROCESS1; - --len; - - } - - return XXH32_avalanche(h32); - - } else { - - switch (len & 15) /* or switch(bEnd - p) */ { - - case 12: - XXH_PROCESS4; - XXH_FALLTHROUGH; - case 8: - XXH_PROCESS4; - XXH_FALLTHROUGH; - case 4: - XXH_PROCESS4; - return XXH32_avalanche(h32); - - case 13: - XXH_PROCESS4; - XXH_FALLTHROUGH; - case 9: - XXH_PROCESS4; - XXH_FALLTHROUGH; - case 5: - XXH_PROCESS4; - XXH_PROCESS1; - return XXH32_avalanche(h32); - - case 14: - XXH_PROCESS4; - XXH_FALLTHROUGH; - case 10: - XXH_PROCESS4; - XXH_FALLTHROUGH; - case 6: - XXH_PROCESS4; - XXH_PROCESS1; - XXH_PROCESS1; - return XXH32_avalanche(h32); - - case 15: - XXH_PROCESS4; - XXH_FALLTHROUGH; - case 11: - XXH_PROCESS4; - XXH_FALLTHROUGH; - case 7: - XXH_PROCESS4; - XXH_FALLTHROUGH; - case 3: - XXH_PROCESS1; - XXH_FALLTHROUGH; - case 2: - XXH_PROCESS1; - XXH_FALLTHROUGH; - case 1: - XXH_PROCESS1; - XXH_FALLTHROUGH; - case 0: - return XXH32_avalanche(h32); - +static XXH_PUREF xxh_u32 +XXH32_finalize(xxh_u32 hash, const xxh_u8* ptr, size_t len, XXH_alignment align) +{ +#define XXH_PROCESS1 do { \ + hash += (*ptr++) * XXH_PRIME32_5; \ + hash = XXH_rotl32(hash, 11) * XXH_PRIME32_1; \ +} while (0) + +#define XXH_PROCESS4 do { \ + hash += XXH_get32bits(ptr) * XXH_PRIME32_3; \ + ptr += 4; \ + hash = XXH_rotl32(hash, 17) * XXH_PRIME32_4; \ +} while (0) + + if (ptr==NULL) XXH_ASSERT(len == 0); + + /* Compact rerolled version; generally faster */ + if (!XXH32_ENDJMP) { + len &= 15; + while (len >= 4) { + XXH_PROCESS4; + len -= 4; + } + while (len > 0) { + XXH_PROCESS1; + --len; + } + return XXH32_avalanche(hash); + } else { + switch(len&15) /* or switch(bEnd - p) */ { + case 12: XXH_PROCESS4; + XXH_FALLTHROUGH; /* fallthrough */ + case 8: XXH_PROCESS4; + XXH_FALLTHROUGH; /* fallthrough */ + case 4: XXH_PROCESS4; + return XXH32_avalanche(hash); + + case 13: XXH_PROCESS4; + XXH_FALLTHROUGH; /* fallthrough */ + case 9: XXH_PROCESS4; + XXH_FALLTHROUGH; /* fallthrough */ + case 5: XXH_PROCESS4; + XXH_PROCESS1; + return XXH32_avalanche(hash); + + case 14: XXH_PROCESS4; + XXH_FALLTHROUGH; /* fallthrough */ + case 10: XXH_PROCESS4; + XXH_FALLTHROUGH; /* fallthrough */ + case 6: XXH_PROCESS4; + XXH_PROCESS1; + XXH_PROCESS1; + return XXH32_avalanche(hash); + + case 15: XXH_PROCESS4; + XXH_FALLTHROUGH; /* fallthrough */ + case 11: XXH_PROCESS4; + XXH_FALLTHROUGH; /* fallthrough */ + case 7: XXH_PROCESS4; + XXH_FALLTHROUGH; /* fallthrough */ + case 3: XXH_PROCESS1; + XXH_FALLTHROUGH; /* fallthrough */ + case 2: XXH_PROCESS1; + XXH_FALLTHROUGH; /* fallthrough */ + case 1: XXH_PROCESS1; + XXH_FALLTHROUGH; /* fallthrough */ + case 0: return XXH32_avalanche(hash); + } + XXH_ASSERT(0); + return hash; /* reaching this point is deemed impossible */ } - - XXH_ASSERT(0); - return h32; /* reaching this point is deemed impossible */ - - } - } - #ifdef XXH_OLD_NAMES - #define PROCESS1 XXH_PROCESS1 - #define PROCESS4 XXH_PROCESS4 - #else - #undef XXH_PROCESS1 - #undef XXH_PROCESS4 - #endif +#ifdef XXH_OLD_NAMES +# define PROCESS1 XXH_PROCESS1 +# define PROCESS4 XXH_PROCESS4 +#else +# undef XXH_PROCESS1 +# undef XXH_PROCESS4 +#endif /*! * @internal * @brief The implementation for @ref XXH32(). * - * @param input, len, seed Directly passed from @ref XXH32(). + * @param input , len , seed Directly passed from @ref XXH32(). * @param align Whether @p input is aligned. * @return The calculated hash. */ -XXH_FORCE_INLINE xxh_u32 XXH32_endian_align(const xxh_u8 *input, size_t len, - xxh_u32 seed, XXH_alignment align) { +XXH_FORCE_INLINE XXH_PUREF xxh_u32 +XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment align) +{ + xxh_u32 h32; + + if (input==NULL) XXH_ASSERT(len == 0); + + if (len>=16) { + const xxh_u8* const bEnd = input + len; + const xxh_u8* const limit = bEnd - 15; + 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 - XXH_PRIME32_1; + + do { + v1 = XXH32_round(v1, XXH_get32bits(input)); input += 4; + v2 = XXH32_round(v2, XXH_get32bits(input)); input += 4; + v3 = XXH32_round(v3, XXH_get32bits(input)); input += 4; + v4 = XXH32_round(v4, XXH_get32bits(input)); input += 4; + } while (input < limit); + + h32 = XXH_rotl32(v1, 1) + XXH_rotl32(v2, 7) + + XXH_rotl32(v3, 12) + XXH_rotl32(v4, 18); + } else { + h32 = seed + XXH_PRIME32_5; + } - const xxh_u8 *bEnd = input ? input + len : NULL; - xxh_u32 h32; + h32 += (xxh_u32)len; - #if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && \ - (XXH_ACCEPT_NULL_INPUT_POINTER >= 1) - if (input == NULL) { + return XXH32_finalize(h32, input, len&15, align); +} - len = 0; - bEnd = input = (const xxh_u8 *)(size_t)16; +/*! @ingroup XXH32_family */ +XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t len, XXH32_hash_t seed) +{ +#if !defined(XXH_NO_STREAM) && XXH_SIZE_OPT >= 2 + /* Simple version, good for code maintenance, but unfortunately slow for small inputs */ + XXH32_state_t state; + XXH32_reset(&state, seed); + XXH32_update(&state, (const xxh_u8*)input, len); + return XXH32_digest(&state); +#else + if (XXH_FORCE_ALIGN_CHECK) { + if ((((size_t)input) & 3) == 0) { /* Input is 4-bytes aligned, leverage the speed benefit */ + return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_aligned); + } } - } + return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned); +#endif +} - #endif - if (len >= 16) { - - const xxh_u8 *const limit = bEnd - 15; - 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 - XXH_PRIME32_1; - - do { - - v1 = XXH32_round(v1, XXH_get32bits(input)); - input += 4; - v2 = XXH32_round(v2, XXH_get32bits(input)); - input += 4; - v3 = XXH32_round(v3, XXH_get32bits(input)); - input += 4; - v4 = XXH32_round(v4, XXH_get32bits(input)); - input += 4; - - } while (input < limit); - - h32 = XXH_rotl32(v1, 1) + XXH_rotl32(v2, 7) + XXH_rotl32(v3, 12) + - XXH_rotl32(v4, 18); - - } else { - - h32 = seed + XXH_PRIME32_5; - - } - - h32 += (xxh_u32)len; - - return XXH32_finalize(h32, input, len & 15, align); - -} - -/*! @ingroup xxh32_family */ -XXH_PUBLIC_API XXH32_hash_t XXH32(const void *input, size_t len, - XXH32_hash_t seed) { - - #if 0 - /* Simple version, good for code maintenance, but unfortunately slow for small inputs */ - XXH32_state_t state; - XXH32_reset(&state, seed); - XXH32_update(&state, (const xxh_u8*)input, len); - return XXH32_digest(&state); - #else - if (XXH_FORCE_ALIGN_CHECK) { - - if ((((size_t)input) & 3) == - 0) { /* Input is 4-bytes aligned, leverage the speed benefit */ - return XXH32_endian_align((const xxh_u8 *)input, len, seed, XXH_aligned); - - } - - } - - return XXH32_endian_align((const xxh_u8 *)input, len, seed, XXH_unaligned); - #endif - -} /******* Hash streaming *******/ -/*! - * @ingroup xxh32_family - */ -XXH_PUBLIC_API XXH32_state_t *XXH32_createState(void) { - - return (XXH32_state_t *)XXH_malloc(sizeof(XXH32_state_t)); - -} - -/*! @ingroup xxh32_family */ -XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t *statePtr) { - - XXH_free(statePtr); - return XXH_OK; - -} - -/*! @ingroup xxh32_family */ -XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t *dstState, - const XXH32_state_t *srcState) { - - memcpy(dstState, srcState, sizeof(*dstState)); - +#ifndef XXH_NO_STREAM +/*! @ingroup XXH32_family */ +XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void) +{ + return (XXH32_state_t*)XXH_malloc(sizeof(XXH32_state_t)); +} +/*! @ingroup XXH32_family */ +XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr) +{ + XXH_free(statePtr); + return XXH_OK; } -/*! @ingroup xxh32_family */ -XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t *statePtr, - XXH32_hash_t seed) { - - 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 + XXH_PRIME32_1 + XXH_PRIME32_2; - state.v2 = seed + XXH_PRIME32_2; - state.v3 = seed + 0; - 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; - +/*! @ingroup XXH32_family */ +XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dstState, const XXH32_state_t* srcState) +{ + XXH_memcpy(dstState, srcState, sizeof(*dstState)); } -/*! @ingroup xxh32_family */ -XXH_PUBLIC_API XXH_errorcode XXH32_update(XXH32_state_t *state, - const void *input, size_t len) { - - if (input == NULL) - #if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && \ - (XXH_ACCEPT_NULL_INPUT_POINTER >= 1) +/*! @ingroup XXH32_family */ +XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, XXH32_hash_t seed) +{ + XXH_ASSERT(statePtr != NULL); + memset(statePtr, 0, sizeof(*statePtr)); + statePtr->v[0] = seed + XXH_PRIME32_1 + XXH_PRIME32_2; + statePtr->v[1] = seed + XXH_PRIME32_2; + statePtr->v[2] = seed + 0; + statePtr->v[3] = seed - XXH_PRIME32_1; return XXH_OK; - #else - return XXH_ERROR; - #endif - - { - - const xxh_u8 *p = (const xxh_u8 *)input; - const xxh_u8 *const bEnd = p + len; - - state->total_len_32 += (XXH32_hash_t)len; - state->large_len |= - (XXH32_hash_t)((len >= 16) | (state->total_len_32 >= 16)); - - if (state->memsize + len < 16) { /* fill in tmp buffer */ - XXH_memcpy((xxh_u8 *)(state->mem32) + state->memsize, input, len); - state->memsize += (XXH32_hash_t)len; - return XXH_OK; - - } - - if (state->memsize) { /* some data left from previous update */ - XXH_memcpy((xxh_u8 *)(state->mem32) + state->memsize, input, - 16 - state->memsize); - { - - const xxh_u32 *p32 = state->mem32; - state->v1 = XXH32_round(state->v1, XXH_readLE32(p32)); - p32++; - state->v2 = XXH32_round(state->v2, XXH_readLE32(p32)); - p32++; - state->v3 = XXH32_round(state->v3, XXH_readLE32(p32)); - p32++; - state->v4 = XXH32_round(state->v4, XXH_readLE32(p32)); - - } +} - p += 16 - state->memsize; - state->memsize = 0; +/*! @ingroup XXH32_family */ +XXH_PUBLIC_API XXH_errorcode +XXH32_update(XXH32_state_t* state, const void* input, size_t len) +{ + if (input==NULL) { + XXH_ASSERT(len == 0); + return XXH_OK; } - if (p <= bEnd - 16) { + { const xxh_u8* p = (const xxh_u8*)input; + const xxh_u8* const bEnd = p + len; - const xxh_u8 *const limit = bEnd - 16; - xxh_u32 v1 = state->v1; - xxh_u32 v2 = state->v2; - xxh_u32 v3 = state->v3; - xxh_u32 v4 = state->v4; + state->total_len_32 += (XXH32_hash_t)len; + state->large_len |= (XXH32_hash_t)((len>=16) | (state->total_len_32>=16)); - do { - - v1 = XXH32_round(v1, XXH_readLE32(p)); - p += 4; - v2 = XXH32_round(v2, XXH_readLE32(p)); - p += 4; - v3 = XXH32_round(v3, XXH_readLE32(p)); - p += 4; - v4 = XXH32_round(v4, XXH_readLE32(p)); - p += 4; - - } while (p <= limit); + if (state->memsize + len < 16) { /* fill in tmp buffer */ + XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, len); + state->memsize += (XXH32_hash_t)len; + return XXH_OK; + } - state->v1 = v1; - state->v2 = v2; - state->v3 = v3; - state->v4 = v4; + if (state->memsize) { /* some data left from previous update */ + XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, 16-state->memsize); + { const xxh_u32* p32 = state->mem32; + state->v[0] = XXH32_round(state->v[0], XXH_readLE32(p32)); p32++; + state->v[1] = XXH32_round(state->v[1], XXH_readLE32(p32)); p32++; + state->v[2] = XXH32_round(state->v[2], XXH_readLE32(p32)); p32++; + state->v[3] = XXH32_round(state->v[3], XXH_readLE32(p32)); + } + p += 16-state->memsize; + state->memsize = 0; + } - } + if (p <= bEnd-16) { + const xxh_u8* const limit = bEnd - 16; - if (p < bEnd) { + do { + state->v[0] = XXH32_round(state->v[0], XXH_readLE32(p)); p+=4; + state->v[1] = XXH32_round(state->v[1], XXH_readLE32(p)); p+=4; + state->v[2] = XXH32_round(state->v[2], XXH_readLE32(p)); p+=4; + state->v[3] = XXH32_round(state->v[3], XXH_readLE32(p)); p+=4; + } while (p<=limit); - XXH_memcpy(state->mem32, p, (size_t)(bEnd - p)); - state->memsize = (unsigned)(bEnd - p); + } + if (p < bEnd) { + XXH_memcpy(state->mem32, p, (size_t)(bEnd-p)); + state->memsize = (unsigned)(bEnd-p); + } } - } - - return XXH_OK; - + return XXH_OK; } -/*! @ingroup xxh32_family */ -XXH_PUBLIC_API XXH32_hash_t XXH32_digest(const XXH32_state_t *state) { - - xxh_u32 h32; - - if (state->large_len) { - h32 = XXH_rotl32(state->v1, 1) + XXH_rotl32(state->v2, 7) + - XXH_rotl32(state->v3, 12) + XXH_rotl32(state->v4, 18); +/*! @ingroup XXH32_family */ +XXH_PUBLIC_API XXH32_hash_t XXH32_digest(const XXH32_state_t* state) +{ + xxh_u32 h32; - } else { - - h32 = state->v3 /* == seed */ + XXH_PRIME32_5; - - } - - h32 += state->total_len_32; + if (state->large_len) { + h32 = XXH_rotl32(state->v[0], 1) + + XXH_rotl32(state->v[1], 7) + + XXH_rotl32(state->v[2], 12) + + XXH_rotl32(state->v[3], 18); + } else { + h32 = state->v[2] /* == seed */ + XXH_PRIME32_5; + } - return XXH32_finalize(h32, (const xxh_u8 *)state->mem32, state->memsize, - XXH_aligned); + h32 += state->total_len_32; + return XXH32_finalize(h32, (const xxh_u8*)state->mem32, state->memsize, XXH_aligned); } +#endif /* !XXH_NO_STREAM */ /******* Canonical representation *******/ -/*! - * @ingroup xxh32_family - * The default return values from XXH functions are unsigned 32 and 64 bit - * integers. - * - * The canonical representation uses big endian convention, the same convention - * as human-readable numbers (large digits first). - * - * This way, hash values can be written into a file or buffer, remaining - * comparable across different systems. - * - * The following functions allow transformation of hash values to and from their - * canonical format. - */ -XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t *dst, - XXH32_hash_t hash) { - - XXH_STATIC_ASSERT(sizeof(XXH32_canonical_t) == sizeof(XXH32_hash_t)); - if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap32(hash); - memcpy(dst, &hash, sizeof(*dst)); - +/*! @ingroup XXH32_family */ +XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash) +{ + XXH_STATIC_ASSERT(sizeof(XXH32_canonical_t) == sizeof(XXH32_hash_t)); + if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap32(hash); + XXH_memcpy(dst, &hash, sizeof(*dst)); } - -/*! @ingroup xxh32_family */ -XXH_PUBLIC_API XXH32_hash_t -XXH32_hashFromCanonical(const XXH32_canonical_t *src) { - - return XXH_readBE32(src); - +/*! @ingroup XXH32_family */ +XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src) +{ + return XXH_readBE32(src); } - #ifndef XXH_NO_LONG_LONG + +#ifndef XXH_NO_LONG_LONG /* ******************************************************************* - * 64-bit hash functions - *********************************************************************/ +* 64-bit hash functions +*********************************************************************/ /*! * @} * @ingroup impl * @{ - */ /******* Memory access *******/ typedef XXH64_hash_t xxh_u64; - #ifdef XXH_OLD_NAMES - #define U64 xxh_u64 - #endif - - #if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS == 3)) - /* - * Manual byteshift. Best for old compilers which don't inline memcpy. - * We actually directly use XXH_readLE64 and XXH_readBE64. - */ - #elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS == 2)) - -/* Force direct memory access. Only works on CPU which support unaligned memory - * access in hardware */ -static xxh_u64 XXH_read64(const void *memPtr) { +#ifdef XXH_OLD_NAMES +# define U64 xxh_u64 +#endif - return *(const xxh_u64 *)memPtr; +#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3)) +/* + * Manual byteshift. Best for old compilers which don't inline memcpy. + * We actually directly use XXH_readLE64 and XXH_readBE64. + */ +#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2)) +/* Force direct memory access. Only works on CPU which support unaligned memory access in hardware */ +static xxh_u64 XXH_read64(const void* memPtr) +{ + return *(const xxh_u64*) memPtr; } - #elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS == 1)) - - /* - * __pack instructions are safer, but compiler specific, hence potentially - * problematic for some compilers. - * - * Currently only defined for GCC and ICC. - */ - #ifdef XXH_OLD_NAMES -typedef union { - - xxh_u32 u32; - xxh_u64 u64; - -} __attribute__((packed)) unalign64; - - #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; +#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1)) +/* + * __attribute__((aligned(1))) is supported by gcc and clang. Originally the + * documentation claimed that it only increased the alignment, but actually it + * can decrease it on gcc, clang, and icc: + * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502, + * https://gcc.godbolt.org/z/xYez1j67Y. + */ +#ifdef XXH_OLD_NAMES +typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) unalign64; +#endif +static xxh_u64 XXH_read64(const void* ptr) +{ + typedef __attribute__((aligned(1))) xxh_u64 xxh_unalign64; + return *((const xxh_unalign64*)ptr); } - #else +#else /* * Portable and safe solution. Generally efficient. - * see: https://stackoverflow.com/a/32095106/646947 + * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html */ -static xxh_u64 XXH_read64(const void *memPtr) { - - xxh_u64 val; - memcpy(&val, memPtr, sizeof(val)); - return val; - -} - - #endif /* XXH_FORCE_DIRECT_MEMORY_ACCESS */ - - #if defined(_MSC_VER) /* Visual Studio */ - #define XXH_swap64 _byteswap_uint64 - #elif XXH_GCC_VERSION >= 403 - #define XXH_swap64 __builtin_bswap64 - #else -static xxh_u64 XXH_swap64(xxh_u64 x) { - - return ((x << 56) & 0xff00000000000000ULL) | - ((x << 40) & 0x00ff000000000000ULL) | - ((x << 24) & 0x0000ff0000000000ULL) | - ((x << 8) & 0x000000ff00000000ULL) | - ((x >> 8) & 0x00000000ff000000ULL) | - ((x >> 24) & 0x0000000000ff0000ULL) | - ((x >> 40) & 0x000000000000ff00ULL) | - ((x >> 56) & 0x00000000000000ffULL); - +static xxh_u64 XXH_read64(const void* memPtr) +{ + xxh_u64 val; + XXH_memcpy(&val, memPtr, sizeof(val)); + return val; +} + +#endif /* XXH_FORCE_DIRECT_MEMORY_ACCESS */ + +#if defined(_MSC_VER) /* Visual Studio */ +# define XXH_swap64 _byteswap_uint64 +#elif XXH_GCC_VERSION >= 403 +# define XXH_swap64 __builtin_bswap64 +#else +static xxh_u64 XXH_swap64(xxh_u64 x) +{ + return ((x << 56) & 0xff00000000000000ULL) | + ((x << 40) & 0x00ff000000000000ULL) | + ((x << 24) & 0x0000ff0000000000ULL) | + ((x << 8) & 0x000000ff00000000ULL) | + ((x >> 8) & 0x00000000ff000000ULL) | + ((x >> 24) & 0x0000000000ff0000ULL) | + ((x >> 40) & 0x000000000000ff00ULL) | + ((x >> 56) & 0x00000000000000ffULL); } +#endif - #endif - - /* XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load. */ - #if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS == 3)) - -XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void *memPtr) { - const xxh_u8 *bytePtr = (const xxh_u8 *)memPtr; - return bytePtr[0] | ((xxh_u64)bytePtr[1] << 8) | ((xxh_u64)bytePtr[2] << 16) | - ((xxh_u64)bytePtr[3] << 24) | ((xxh_u64)bytePtr[4] << 32) | - ((xxh_u64)bytePtr[5] << 40) | ((xxh_u64)bytePtr[6] << 48) | - ((xxh_u64)bytePtr[7] << 56); +/* XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load. */ +#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3)) +XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* memPtr) +{ + const xxh_u8* bytePtr = (const xxh_u8 *)memPtr; + return bytePtr[0] + | ((xxh_u64)bytePtr[1] << 8) + | ((xxh_u64)bytePtr[2] << 16) + | ((xxh_u64)bytePtr[3] << 24) + | ((xxh_u64)bytePtr[4] << 32) + | ((xxh_u64)bytePtr[5] << 40) + | ((xxh_u64)bytePtr[6] << 48) + | ((xxh_u64)bytePtr[7] << 56); } -XXH_FORCE_INLINE xxh_u64 XXH_readBE64(const void *memPtr) { - - const xxh_u8 *bytePtr = (const xxh_u8 *)memPtr; - return bytePtr[7] | ((xxh_u64)bytePtr[6] << 8) | ((xxh_u64)bytePtr[5] << 16) | - ((xxh_u64)bytePtr[4] << 24) | ((xxh_u64)bytePtr[3] << 32) | - ((xxh_u64)bytePtr[2] << 40) | ((xxh_u64)bytePtr[1] << 48) | - ((xxh_u64)bytePtr[0] << 56); - +XXH_FORCE_INLINE xxh_u64 XXH_readBE64(const void* memPtr) +{ + const xxh_u8* bytePtr = (const xxh_u8 *)memPtr; + return bytePtr[7] + | ((xxh_u64)bytePtr[6] << 8) + | ((xxh_u64)bytePtr[5] << 16) + | ((xxh_u64)bytePtr[4] << 24) + | ((xxh_u64)bytePtr[3] << 32) + | ((xxh_u64)bytePtr[2] << 40) + | ((xxh_u64)bytePtr[1] << 48) + | ((xxh_u64)bytePtr[0] << 56); } - #else -XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void *ptr) { - - return XXH_CPU_LITTLE_ENDIAN ? XXH_read64(ptr) : XXH_swap64(XXH_read64(ptr)); - +#else +XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* ptr) +{ + return XXH_CPU_LITTLE_ENDIAN ? XXH_read64(ptr) : XXH_swap64(XXH_read64(ptr)); } -static xxh_u64 XXH_readBE64(const void *ptr) { - - return XXH_CPU_LITTLE_ENDIAN ? XXH_swap64(XXH_read64(ptr)) : XXH_read64(ptr); - +static xxh_u64 XXH_readBE64(const void* ptr) +{ + return XXH_CPU_LITTLE_ENDIAN ? XXH_swap64(XXH_read64(ptr)) : XXH_read64(ptr); } +#endif - #endif - -XXH_FORCE_INLINE xxh_u64 XXH_readLE64_align(const void *ptr, - XXH_alignment align) { - - if (align == XXH_unaligned) - return XXH_readLE64(ptr); - else - return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u64 *)ptr - : XXH_swap64(*(const xxh_u64 *)ptr); - +XXH_FORCE_INLINE xxh_u64 +XXH_readLE64_align(const void* ptr, XXH_alignment align) +{ + if (align==XXH_unaligned) + return XXH_readLE64(ptr); + else + return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u64*)ptr : XXH_swap64(*(const xxh_u64*)ptr); } - /******* xxh64 *******/ - /*! - * @} - * @defgroup xxh64_impl XXH64 implementation - * @ingroup impl - * @{ - - */ - /* #define rather that static const, to be used as initializers */ - #define XXH_PRIME64_1 \ - 0x9E3779B185EBCA87ULL /*!< \ - 0b1001111000110111011110011011000110000101111010111100101010000111 \ - */ - #define XXH_PRIME64_2 \ - 0xC2B2AE3D27D4EB4FULL /*!< \ - 0b1100001010110010101011100011110100100111110101001110101101001111 \ - */ - #define XXH_PRIME64_3 \ - 0x165667B19E3779F9ULL /*!< \ - 0b0001011001010110011001111011000110011110001101110111100111111001 \ - */ - #define XXH_PRIME64_4 \ - 0x85EBCA77C2B2AE63ULL /*!< \ - 0b1000010111101011110010100111011111000010101100101010111001100011 \ - */ - #define 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 * XXH_PRIME64_2; - acc = XXH_rotl64(acc, 31); - acc *= XXH_PRIME64_1; - return acc; +/******* xxh64 *******/ +/*! + * @} + * @defgroup XXH64_impl XXH64 implementation + * @ingroup impl + * + * Details on the XXH64 implementation. + * @{ + */ +/* #define rather that static const, to be used as initializers */ +#define XXH_PRIME64_1 0x9E3779B185EBCA87ULL /*!< 0b1001111000110111011110011011000110000101111010111100101010000111 */ +#define XXH_PRIME64_2 0xC2B2AE3D27D4EB4FULL /*!< 0b1100001010110010101011100011110100100111110101001110101101001111 */ +#define XXH_PRIME64_3 0x165667B19E3779F9ULL /*!< 0b0001011001010110011001111011000110011110001101110111100111111001 */ +#define XXH_PRIME64_4 0x85EBCA77C2B2AE63ULL /*!< 0b1000010111101011110010100111011111000010101100101010111001100011 */ +#define 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 +/*! @copydoc XXH32_round */ +static xxh_u64 XXH64_round(xxh_u64 acc, xxh_u64 input) +{ + acc += input * XXH_PRIME64_2; + acc = XXH_rotl64(acc, 31); + acc *= XXH_PRIME64_1; + return acc; } -static xxh_u64 XXH64_mergeRound(xxh_u64 acc, xxh_u64 val) { - - val = XXH64_round(0, val); - acc ^= val; - acc = acc * XXH_PRIME64_1 + XXH_PRIME64_4; - return acc; - +static xxh_u64 XXH64_mergeRound(xxh_u64 acc, xxh_u64 val) +{ + val = XXH64_round(0, val); + acc ^= val; + acc = acc * XXH_PRIME64_1 + XXH_PRIME64_4; + return acc; } -static xxh_u64 XXH64_avalanche(xxh_u64 h64) { - - h64 ^= h64 >> 33; - h64 *= XXH_PRIME64_2; - h64 ^= h64 >> 29; - h64 *= XXH_PRIME64_3; - h64 ^= h64 >> 32; - return h64; - +/*! @copydoc XXH32_avalanche */ +static xxh_u64 XXH64_avalanche(xxh_u64 hash) +{ + hash ^= hash >> 33; + hash *= XXH_PRIME64_2; + hash ^= hash >> 29; + hash *= XXH_PRIME64_3; + hash ^= hash >> 32; + return hash; } - #define XXH_get64bits(p) XXH_readLE64_align(p, align) - -static xxh_u64 XXH64_finalize(xxh_u64 h64, const xxh_u8 *ptr, size_t len, - XXH_alignment align) { - - len &= 31; - while (len >= 8) { - - xxh_u64 const k1 = XXH64_round(0, XXH_get64bits(ptr)); - ptr += 8; - h64 ^= k1; - h64 = XXH_rotl64(h64, 27) * XXH_PRIME64_1 + XXH_PRIME64_4; - len -= 8; - - } - - if (len >= 4) { - - h64 ^= (xxh_u64)(XXH_get32bits(ptr)) * XXH_PRIME64_1; - ptr += 4; - h64 = XXH_rotl64(h64, 23) * XXH_PRIME64_2 + XXH_PRIME64_3; - len -= 4; - - } - - while (len > 0) { - h64 ^= (*ptr++) * XXH_PRIME64_5; - h64 = XXH_rotl64(h64, 11) * XXH_PRIME64_1; - --len; - - } - - return XXH64_avalanche(h64); +#define XXH_get64bits(p) XXH_readLE64_align(p, align) +/*! + * @internal + * @brief Processes the last 0-31 bytes of @p ptr. + * + * There may be up to 31 bytes remaining to consume from the input. + * This final stage will digest them to ensure that all input bytes are present + * in the final mix. + * + * @param hash The hash to finalize. + * @param ptr The pointer to the remaining input. + * @param len The remaining length, modulo 32. + * @param align Whether @p ptr is aligned. + * @return The finalized hash + * @see XXH32_finalize(). + */ +static XXH_PUREF xxh_u64 +XXH64_finalize(xxh_u64 hash, const xxh_u8* ptr, size_t len, XXH_alignment align) +{ + if (ptr==NULL) XXH_ASSERT(len == 0); + len &= 31; + while (len >= 8) { + xxh_u64 const k1 = XXH64_round(0, XXH_get64bits(ptr)); + ptr += 8; + hash ^= k1; + hash = XXH_rotl64(hash,27) * XXH_PRIME64_1 + XXH_PRIME64_4; + len -= 8; + } + if (len >= 4) { + hash ^= (xxh_u64)(XXH_get32bits(ptr)) * XXH_PRIME64_1; + ptr += 4; + hash = XXH_rotl64(hash, 23) * XXH_PRIME64_2 + XXH_PRIME64_3; + len -= 4; + } + while (len > 0) { + hash ^= (*ptr++) * XXH_PRIME64_5; + hash = XXH_rotl64(hash, 11) * XXH_PRIME64_1; + --len; + } + return XXH64_avalanche(hash); } - #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) { - - const xxh_u8 *bEnd = input ? input + len : NULL; - xxh_u64 h64; - - #if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && \ - (XXH_ACCEPT_NULL_INPUT_POINTER >= 1) - if (input == NULL) { - - len = 0; - bEnd = input = (const xxh_u8 *)(size_t)32; - - } - - #endif - - if (len >= 32) { - - const xxh_u8 *const limit = bEnd - 32; - 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 - XXH_PRIME64_1; - - do { - - v1 = XXH64_round(v1, XXH_get64bits(input)); - input += 8; - v2 = XXH64_round(v2, XXH_get64bits(input)); - input += 8; - v3 = XXH64_round(v3, XXH_get64bits(input)); - input += 8; - v4 = XXH64_round(v4, XXH_get64bits(input)); - input += 8; - - } while (input <= limit); - - h64 = XXH_rotl64(v1, 1) + XXH_rotl64(v2, 7) + XXH_rotl64(v3, 12) + - XXH_rotl64(v4, 18); - h64 = XXH64_mergeRound(h64, v1); - h64 = XXH64_mergeRound(h64, v2); - h64 = XXH64_mergeRound(h64, v3); - h64 = XXH64_mergeRound(h64, v4); - - } else { - - h64 = seed + XXH_PRIME64_5; +#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 - } +/*! + * @internal + * @brief The implementation for @ref XXH64(). + * + * @param input , len , seed Directly passed from @ref XXH64(). + * @param align Whether @p input is aligned. + * @return The calculated hash. + */ +XXH_FORCE_INLINE XXH_PUREF xxh_u64 +XXH64_endian_align(const xxh_u8* input, size_t len, xxh_u64 seed, XXH_alignment align) +{ + xxh_u64 h64; + if (input==NULL) XXH_ASSERT(len == 0); + + if (len>=32) { + const xxh_u8* const bEnd = input + len; + const xxh_u8* const limit = bEnd - 31; + 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 - XXH_PRIME64_1; + + do { + v1 = XXH64_round(v1, XXH_get64bits(input)); input+=8; + v2 = XXH64_round(v2, XXH_get64bits(input)); input+=8; + v3 = XXH64_round(v3, XXH_get64bits(input)); input+=8; + v4 = XXH64_round(v4, XXH_get64bits(input)); input+=8; + } while (input= 2 /* Simple version, good for code maintenance, but unfortunately slow for small inputs */ XXH64_state_t state; XXH64_reset(&state, seed); XXH64_update(&state, (const xxh_u8*)input, len); return XXH64_digest(&state); - #else - if (XXH_FORCE_ALIGN_CHECK) { - - if ((((size_t)input) & 7) == - 0) { /* Input is aligned, let's leverage the speed advantage */ - return XXH64_endian_align((const xxh_u8 *)input, len, seed, XXH_aligned); - - } - - } +#else + if (XXH_FORCE_ALIGN_CHECK) { + if ((((size_t)input) & 7)==0) { /* Input is aligned, let's leverage the speed advantage */ + return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_aligned); + } } - return XXH64_endian_align((const xxh_u8 *)input, len, seed, XXH_unaligned); - - #endif + return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned); +#endif } /******* Hash Streaming *******/ - -/*! @ingroup xxh64_family*/ -XXH_PUBLIC_API XXH64_state_t *XXH64_createState(void) { - - return (XXH64_state_t *)XXH_malloc(sizeof(XXH64_state_t)); - -} - -/*! @ingroup xxh64_family */ -XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t *statePtr) { - - XXH_free(statePtr); - return XXH_OK; - -} - -/*! @ingroup xxh64_family */ -XXH_PUBLIC_API void XXH64_copyState(XXH64_state_t *dstState, - const XXH64_state_t *srcState) { - - memcpy(dstState, srcState, sizeof(*dstState)); - +#ifndef XXH_NO_STREAM +/*! @ingroup XXH64_family*/ +XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void) +{ + return (XXH64_state_t*)XXH_malloc(sizeof(XXH64_state_t)); +} +/*! @ingroup XXH64_family */ +XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr) +{ + XXH_free(statePtr); + return XXH_OK; } -/*! @ingroup xxh64_family */ -XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH64_state_t *statePtr, - XXH64_hash_t seed) { - - 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 + XXH_PRIME64_1 + XXH_PRIME64_2; - state.v2 = seed + XXH_PRIME64_2; - state.v3 = seed + 0; - 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; - +/*! @ingroup XXH64_family */ +XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dstState, const XXH64_state_t* srcState) +{ + XXH_memcpy(dstState, srcState, sizeof(*dstState)); } -/*! @ingroup xxh64_family */ -XXH_PUBLIC_API XXH_errorcode XXH64_update(XXH64_state_t *state, - const void *input, size_t len) { - - if (input == NULL) - #if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && \ - (XXH_ACCEPT_NULL_INPUT_POINTER >= 1) +/*! @ingroup XXH64_family */ +XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed) +{ + XXH_ASSERT(statePtr != NULL); + memset(statePtr, 0, sizeof(*statePtr)); + statePtr->v[0] = seed + XXH_PRIME64_1 + XXH_PRIME64_2; + statePtr->v[1] = seed + XXH_PRIME64_2; + statePtr->v[2] = seed + 0; + statePtr->v[3] = seed - XXH_PRIME64_1; return XXH_OK; - #else - return XXH_ERROR; - #endif - - { - - const xxh_u8 *p = (const xxh_u8 *)input; - const xxh_u8 *const bEnd = p + len; - - state->total_len += len; - - if (state->memsize + len < 32) { /* fill in tmp buffer */ - XXH_memcpy(((xxh_u8 *)state->mem64) + state->memsize, input, len); - state->memsize += (xxh_u32)len; - return XXH_OK; - - } - - if (state->memsize) { /* tmp buffer is full */ - XXH_memcpy(((xxh_u8 *)state->mem64) + state->memsize, input, - 32 - state->memsize); - state->v1 = XXH64_round(state->v1, XXH_readLE64(state->mem64 + 0)); - state->v2 = XXH64_round(state->v2, XXH_readLE64(state->mem64 + 1)); - state->v3 = XXH64_round(state->v3, XXH_readLE64(state->mem64 + 2)); - state->v4 = XXH64_round(state->v4, XXH_readLE64(state->mem64 + 3)); - p += 32 - state->memsize; - state->memsize = 0; +} +/*! @ingroup XXH64_family */ +XXH_PUBLIC_API XXH_errorcode +XXH64_update (XXH_NOESCAPE XXH64_state_t* state, XXH_NOESCAPE const void* input, size_t len) +{ + if (input==NULL) { + XXH_ASSERT(len == 0); + return XXH_OK; } - if (p + 32 <= bEnd) { - - const xxh_u8 *const limit = bEnd - 32; - xxh_u64 v1 = state->v1; - xxh_u64 v2 = state->v2; - xxh_u64 v3 = state->v3; - xxh_u64 v4 = state->v4; + { const xxh_u8* p = (const xxh_u8*)input; + const xxh_u8* const bEnd = p + len; - do { + state->total_len += len; - v1 = XXH64_round(v1, XXH_readLE64(p)); - p += 8; - v2 = XXH64_round(v2, XXH_readLE64(p)); - p += 8; - v3 = XXH64_round(v3, XXH_readLE64(p)); - p += 8; - v4 = XXH64_round(v4, XXH_readLE64(p)); - p += 8; - - } while (p <= limit); + if (state->memsize + len < 32) { /* fill in tmp buffer */ + XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, len); + state->memsize += (xxh_u32)len; + return XXH_OK; + } - state->v1 = v1; - state->v2 = v2; - state->v3 = v3; - state->v4 = v4; + if (state->memsize) { /* tmp buffer is full */ + XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, 32-state->memsize); + state->v[0] = XXH64_round(state->v[0], XXH_readLE64(state->mem64+0)); + state->v[1] = XXH64_round(state->v[1], XXH_readLE64(state->mem64+1)); + state->v[2] = XXH64_round(state->v[2], XXH_readLE64(state->mem64+2)); + state->v[3] = XXH64_round(state->v[3], XXH_readLE64(state->mem64+3)); + p += 32 - state->memsize; + state->memsize = 0; + } - } + if (p+32 <= bEnd) { + const xxh_u8* const limit = bEnd - 32; - if (p < bEnd) { + do { + state->v[0] = XXH64_round(state->v[0], XXH_readLE64(p)); p+=8; + state->v[1] = XXH64_round(state->v[1], XXH_readLE64(p)); p+=8; + state->v[2] = XXH64_round(state->v[2], XXH_readLE64(p)); p+=8; + state->v[3] = XXH64_round(state->v[3], XXH_readLE64(p)); p+=8; + } while (p<=limit); - XXH_memcpy(state->mem64, p, (size_t)(bEnd - p)); - state->memsize = (unsigned)(bEnd - p); + } + if (p < bEnd) { + XXH_memcpy(state->mem64, p, (size_t)(bEnd-p)); + state->memsize = (unsigned)(bEnd-p); + } } - } - - return XXH_OK; - + return XXH_OK; } -/*! @ingroup xxh64_family */ -XXH_PUBLIC_API XXH64_hash_t XXH64_digest(const XXH64_state_t *state) { - - xxh_u64 h64; - - if (state->total_len >= 32) { - - xxh_u64 const v1 = state->v1; - xxh_u64 const v2 = state->v2; - xxh_u64 const v3 = state->v3; - xxh_u64 const v4 = state->v4; - h64 = XXH_rotl64(v1, 1) + XXH_rotl64(v2, 7) + XXH_rotl64(v3, 12) + - XXH_rotl64(v4, 18); - h64 = XXH64_mergeRound(h64, v1); - h64 = XXH64_mergeRound(h64, v2); - h64 = XXH64_mergeRound(h64, v3); - h64 = XXH64_mergeRound(h64, v4); +/*! @ingroup XXH64_family */ +XXH_PUBLIC_API XXH64_hash_t XXH64_digest(XXH_NOESCAPE const XXH64_state_t* state) +{ + xxh_u64 h64; - } else { - - h64 = state->v3 /*seed*/ + XXH_PRIME64_5; - - } - - h64 += (xxh_u64)state->total_len; + if (state->total_len >= 32) { + h64 = XXH_rotl64(state->v[0], 1) + XXH_rotl64(state->v[1], 7) + XXH_rotl64(state->v[2], 12) + XXH_rotl64(state->v[3], 18); + h64 = XXH64_mergeRound(h64, state->v[0]); + h64 = XXH64_mergeRound(h64, state->v[1]); + h64 = XXH64_mergeRound(h64, state->v[2]); + h64 = XXH64_mergeRound(h64, state->v[3]); + } else { + h64 = state->v[2] /*seed*/ + XXH_PRIME64_5; + } - return XXH64_finalize(h64, (const xxh_u8 *)state->mem64, - (size_t)state->total_len, XXH_aligned); + h64 += (xxh_u64) state->total_len; + return XXH64_finalize(h64, (const xxh_u8*)state->mem64, (size_t)state->total_len, XXH_aligned); } +#endif /* !XXH_NO_STREAM */ /******* Canonical representation *******/ -/*! @ingroup xxh64_family */ -XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t *dst, - XXH64_hash_t hash) { - - XXH_STATIC_ASSERT(sizeof(XXH64_canonical_t) == sizeof(XXH64_hash_t)); - if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap64(hash); - memcpy(dst, &hash, sizeof(*dst)); - +/*! @ingroup XXH64_family */ +XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash) +{ + XXH_STATIC_ASSERT(sizeof(XXH64_canonical_t) == sizeof(XXH64_hash_t)); + if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap64(hash); + XXH_memcpy(dst, &hash, sizeof(*dst)); } -/*! @ingroup xxh64_family */ -XXH_PUBLIC_API XXH64_hash_t -XXH64_hashFromCanonical(const XXH64_canonical_t *src) { - - return XXH_readBE64(src); - +/*! @ingroup XXH64_family */ +XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src) +{ + return XXH_readBE64(src); } - #ifndef XXH_NO_XXH3 - - /* ********************************************************************* - * XXH3 - * New generation hash designed for speed on small keys and vectorization - ************************************************************************ */ - /*! - * @} - * @defgroup xxh3_impl XXH3 implementation - * @ingroup impl - * @{ - - */ +#ifndef XXH_NO_XXH3 - /* === Compiler specifics === */ - - #if ((defined(sun) || defined(__sun)) && \ - __cplusplus) /* Solaris includes __STDC_VERSION__ with C++. Tested \ - with GCC 5.5 */ - #define XXH_RESTRICT /* disable */ - #elif 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 - #elif defined(__SSE2__) - #include - #elif defined(__ARM_NEON__) || defined(__ARM_NEON) - #define inline __inline__ /* circumvent a clang bug */ - #include - #undef inline - #endif - #elif defined(_MSC_VER) - #include - #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 - * ========================================== */ - - #ifdef XXH_DOXYGEN - /*! - * @ingroup tuning - * @brief Overrides the vectorization implementation chosen for XXH3. - * - * Can be defined to 0 to disable SIMD or any of the values mentioned in - * @ref XXH_VECTOR_TYPE. - * - * If this is not defined, it uses predefined macros to determine the - * best implementation. - */ - #define XXH_VECTOR XXH_SCALAR +/* ********************************************************************* +* XXH3 +* New generation hash designed for speed on small keys and vectorization +************************************************************************ */ /*! - * @ingroup tuning - * @brief Possible values for @ref XXH_VECTOR. - * - * Note that these are actually implemented as macros. - * - * If this is not defined, it is detected automatically. - * @ref XXH_X86DISPATCH overrides this. + * @} + * @defgroup XXH3_impl XXH3 implementation + * @ingroup impl + * @{ */ -enum XXH_VECTOR_TYPE /* fake enum */ { - XXH_SCALAR = 0, /*!< Portable scalar version */ - XXH_SSE2 = 1, /*!< - * SSE2 for Pentium 4, Opteron, all x86_64. - * - * @note SSE2 is also guaranteed on Windows 10, macOS, and - * Android x86. - */ - XXH_AVX2 = 2, /*!< AVX2 for Haswell and Bulldozer */ - XXH_AVX512 = 3, /*!< AVX512 for Skylake and Icelake */ - XXH_NEON = 4, /*!< NEON for most ARMv7-A and all AArch64 */ - XXH_VSX = 5, /*!< VSX and ZVector for POWER8/z13 (64-bit) */ +/* === Compiler specifics === */ -}; +#if ((defined(sun) || defined(__sun)) && __cplusplus) /* Solaris includes __STDC_VERSION__ with C++. Tested with GCC 5.5 */ +# define XXH_RESTRICT /* disable */ +#elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L /* >= C99 */ +# define XXH_RESTRICT restrict +#elif (defined (__GNUC__) && ((__GNUC__ > 3) || (__GNUC__ == 3 && __GNUC_MINOR__ >= 1))) \ + || (defined (__clang__)) \ + || (defined (_MSC_VER) && (_MSC_VER >= 1400)) \ + || (defined (__INTEL_COMPILER) && (__INTEL_COMPILER >= 1300)) +/* + * There are a LOT more compilers that recognize __restrict but this + * covers the major ones. + */ +# define XXH_RESTRICT __restrict +#else +# define XXH_RESTRICT /* disable */ +#endif - /*! - * @ingroup tuning - * @brief Selects the minimum alignment for XXH3's accumulators. - * - * When using SIMD, this should match the alignment reqired for said - * vector type, so, for example, 32 for AVX2. - * - * Default: Auto detected. - */ - #define XXH_ACC_ALIGN 8 - #endif - - /* Actual definition */ - #ifndef XXH_DOXYGEN - #define XXH_SCALAR 0 - #define XXH_SSE2 1 - #define XXH_AVX2 2 - #define XXH_AVX512 3 - #define XXH_NEON 4 - #define XXH_VSX 5 - #endif - - #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 = { +#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 - * 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. - */ +#ifndef XXH_HAS_INCLUDE +# ifdef __has_include +/* + * Not defined as XXH_HAS_INCLUDE(x) (function-like) because + * this causes segfaults in Apple Clang 4.2 (on Mac OS X 10.7 Lion) + */ +# define XXH_HAS_INCLUDE __has_include +# else +# define XXH_HAS_INCLUDE(x) 0 +# endif +#endif - /*! - * Function-like macro: - * void XXH_SPLIT_IN_PLACE(uint64x2_t &in, uint32x2_t &outLo, uint32x2_t - * &outHi) - * { +#if defined(__GNUC__) || defined(__clang__) +# if defined(__ARM_FEATURE_SVE) +# include +# endif +# if defined(__ARM_NEON__) || defined(__ARM_NEON) \ + || (defined(_M_ARM) && _M_ARM >= 7) \ + || defined(_M_ARM64) || defined(_M_ARM64EC) \ + || (defined(__wasm_simd128__) && XXH_HAS_INCLUDE()) /* WASM SIMD128 via SIMDe */ +# define inline __inline__ /* circumvent a clang bug */ +# include +# undef inline +# elif defined(__AVX2__) +# include +# elif defined(__SSE2__) +# include +# endif +#endif - * 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 - #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 - #endif +#if defined(_MSC_VER) +# include +#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 - #if defined(__POWER9_VECTOR__) || \ - (defined(__clang__) && defined(__s390x__)) - #define XXH_vec_revb vec_revb - #else +/* + * 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 + * ========================================== */ + +#ifdef XXH_DOXYGEN /*! - * A polyfill for POWER9's vec_revb(). + * @ingroup tuning + * @brief Overrides the vectorization implementation chosen for XXH3. + * + * Can be defined to 0 to disable SIMD or any of the values mentioned in + * @ref XXH_VECTOR_TYPE. + * + * If this is not defined, it uses predefined macros to determine the best + * implementation. + */ +# define XXH_VECTOR XXH_SCALAR +/*! + * @ingroup tuning + * @brief Possible values for @ref XXH_VECTOR. + * + * Note that these are actually implemented as macros. + * + * If this is not defined, it is detected automatically. + * internal macro XXH_X86DISPATCH overrides this. + */ +enum XXH_VECTOR_TYPE /* fake enum */ { + XXH_SCALAR = 0, /*!< Portable scalar version */ + XXH_SSE2 = 1, /*!< + * SSE2 for Pentium 4, Opteron, all x86_64. + * + * @note SSE2 is also guaranteed on Windows 10, macOS, and + * Android x86. + */ + XXH_AVX2 = 2, /*!< AVX2 for Haswell and Bulldozer */ + XXH_AVX512 = 3, /*!< AVX512 for Skylake and Icelake */ + XXH_NEON = 4, /*!< + * NEON for most ARMv7-A, all AArch64, and WASM SIMD128 + * via the SIMDeverywhere polyfill provided with the + * Emscripten SDK. + */ + XXH_VSX = 5, /*!< VSX and ZVector for POWER8/z13 (64-bit) */ + XXH_SVE = 6, /*!< SVE for some ARMv8-A and ARMv9-A */ +}; +/*! + * @ingroup tuning + * @brief Selects the minimum alignment for XXH3's accumulators. + * + * When using SIMD, this should match the alignment required for said vector + * type, so, for example, 32 for AVX2. + * + * Default: Auto detected. */ -XXH_FORCE_INLINE xxh_u64x2 XXH_vec_revb(xxh_u64x2 val) { +# define XXH_ACC_ALIGN 8 +#endif - 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); +/* Actual definition */ +#ifndef XXH_DOXYGEN +# define XXH_SCALAR 0 +# define XXH_SSE2 1 +# define XXH_AVX2 2 +# define XXH_AVX512 3 +# define XXH_NEON 4 +# define XXH_VSX 5 +# define XXH_SVE 6 +#endif -} +#ifndef XXH_VECTOR /* can be defined on command line */ +# if defined(__ARM_FEATURE_SVE) +# define XXH_VECTOR XXH_SVE +# elif ( \ + defined(__ARM_NEON__) || defined(__ARM_NEON) /* gcc */ \ + || defined(_M_ARM) || defined(_M_ARM64) || defined(_M_ARM64EC) /* msvc */ \ + || (defined(__wasm_simd128__) && XXH_HAS_INCLUDE()) /* wasm simd128 via SIMDe */ \ + ) && ( \ + defined(_WIN32) || defined(__LITTLE_ENDIAN__) /* little endian only */ \ + || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \ + ) +# define XXH_VECTOR XXH_NEON +# elif 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(__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 - #endif - #endif /* XXH_VSX_BE */ +/* __ARM_FEATURE_SVE is only supported by GCC & Clang. */ +#if (XXH_VECTOR == XXH_SVE) && !defined(__ARM_FEATURE_SVE) +# ifdef _MSC_VER +# pragma warning(once : 4606) +# else +# warning "__ARM_FEATURE_SVE isn't supported. Use SCALAR instead." +# endif +# undef XXH_VECTOR +# define XXH_VECTOR XXH_SCALAR +#endif -/*! - * Performs an unaligned vector load and byte swaps it on big endian. +/* + * Controls the alignment of the accumulator, + * for compatibility with aligned vector loads, which are usually faster. */ -XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(const void *ptr) { +#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 +# elif XXH_VECTOR == XXH_SVE /* sve */ +# define XXH_ACC_ALIGN 64 +# endif +#endif - xxh_u64x2 ret; - memcpy(&ret, ptr, sizeof(xxh_u64x2)); - #if XXH_VSX_BE - ret = XXH_vec_revb(ret); - #endif - return ret; +#if defined(XXH_X86DISPATCH) || XXH_VECTOR == XXH_SSE2 \ + || XXH_VECTOR == XXH_AVX2 || XXH_VECTOR == XXH_AVX512 +# define XXH_SEC_ALIGN XXH_ACC_ALIGN +#elif XXH_VECTOR == XXH_SVE +# define XXH_SEC_ALIGN XXH_ACC_ALIGN +#else +# define XXH_SEC_ALIGN 8 +#endif -} +#if defined(__GNUC__) || defined(__clang__) +# define XXH_ALIASING __attribute__((may_alias)) +#else +# define XXH_ALIASING /* nothing */ +#endif - /* - * 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) { +/* + * 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__) && XXH_SIZE_OPT <= 0 /* respect -O0 and -Os */ +# pragma GCC push_options +# pragma GCC optimize("-O2") +#endif + +#if XXH_VECTOR == XXH_NEON + +/* + * UGLY HACK: While AArch64 GCC on Linux does not seem to care, on macOS, GCC -O3 + * optimizes out the entire hashLong loop because of the aliasing violation. + * + * However, GCC is also inefficient at load-store optimization with vld1q/vst1q, + * so the only option is to mark it as aliasing. + */ +typedef uint64x2_t xxh_aliasing_uint64x2_t XXH_ALIASING; - xxh_u64x2 result; - __asm__("vmulouw %0, %1, %2" : "=v"(result) : "v"(a), "v"(b)); - return result; +/*! + * @internal + * @brief `vld1q_u64` but faster and alignment-safe. + * + * On AArch64, unaligned access is always safe, but on ARMv7-a, it is only + * *conditionally* safe (`vld1` has an alignment bit like `movdq[ua]` in x86). + * + * GCC for AArch64 sees `vld1q_u8` as an intrinsic instead of a load, so it + * prohibits load-store optimizations. Therefore, a direct dereference is used. + * + * Otherwise, `vld1q_u8` is used with `vreinterpretq_u8_u64` to do a safe + * unaligned load. + */ +#if defined(__aarch64__) && defined(__GNUC__) && !defined(__clang__) +XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr) /* silence -Wcast-align */ +{ + return *(xxh_aliasing_uint64x2_t const *)ptr; +} +#else +XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr) +{ + return vreinterpretq_u64_u8(vld1q_u8((uint8_t const*)ptr)); +} +#endif +/*! + * @internal + * @brief `vmlal_u32` on low and high halves of a vector. + * + * This is a workaround for AArch64 GCC < 11 which implemented arm_neon.h with + * inline assembly and were therefore incapable of merging the `vget_{low, high}_u32` + * with `vmlal_u32`. + */ +#if defined(__aarch64__) && defined(__GNUC__) && !defined(__clang__) && __GNUC__ < 11 +XXH_FORCE_INLINE uint64x2_t +XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs) +{ + /* Inline assembly is the only way */ + __asm__("umlal %0.2d, %1.2s, %2.2s" : "+w" (acc) : "w" (lhs), "w" (rhs)); + return acc; +} +XXH_FORCE_INLINE uint64x2_t +XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs) +{ + /* This intrinsic works as expected */ + return vmlal_high_u32(acc, lhs, rhs); +} +#else +/* Portable intrinsic versions */ +XXH_FORCE_INLINE uint64x2_t +XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs) +{ + return vmlal_u32(acc, vget_low_u32(lhs), vget_low_u32(rhs)); +} +/*! @copydoc XXH_vmlal_low_u32 + * Assume the compiler converts this to vmlal_high_u32 on aarch64 */ +XXH_FORCE_INLINE uint64x2_t +XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs) +{ + return vmlal_u32(acc, vget_high_u32(lhs), vget_high_u32(rhs)); } +#endif + +/*! + * @ingroup tuning + * @brief Controls the NEON to scalar ratio for XXH3 + * + * This can be set to 2, 4, 6, or 8. + * + * ARM Cortex CPUs are _very_ sensitive to how their pipelines are used. + * + * For example, the Cortex-A73 can dispatch 3 micro-ops per cycle, but only 2 of those + * can be NEON. If you are only using NEON instructions, you are only using 2/3 of the CPU + * bandwidth. + * + * This is even more noticeable on the more advanced cores like the Cortex-A76 which + * can dispatch 8 micro-ops per cycle, but still only 2 NEON micro-ops at once. + * + * Therefore, to make the most out of the pipeline, it is beneficial to run 6 NEON lanes + * and 2 scalar lanes, which is chosen by default. + * + * This does not apply to Apple processors or 32-bit processors, which run better with + * full NEON. These will default to 8. Additionally, size-optimized builds run 8 lanes. + * + * This change benefits CPUs with large micro-op buffers without negatively affecting + * most other CPUs: + * + * | Chipset | Dispatch type | NEON only | 6:2 hybrid | Diff. | + * |:----------------------|:--------------------|----------:|-----------:|------:| + * | Snapdragon 730 (A76) | 2 NEON/8 micro-ops | 8.8 GB/s | 10.1 GB/s | ~16% | + * | Snapdragon 835 (A73) | 2 NEON/3 micro-ops | 5.1 GB/s | 5.3 GB/s | ~5% | + * | Marvell PXA1928 (A53) | In-order dual-issue | 1.9 GB/s | 1.9 GB/s | 0% | + * | Apple M1 | 4 NEON/8 micro-ops | 37.3 GB/s | 36.1 GB/s | ~-3% | + * + * It also seems to fix some bad codegen on GCC, making it almost as fast as clang. + * + * When using WASM SIMD128, if this is 2 or 6, SIMDe will scalarize 2 of the lanes meaning + * it effectively becomes worse 4. + * + * @see XXH3_accumulate_512_neon() + */ +# ifndef XXH3_NEON_LANES +# if (defined(__aarch64__) || defined(__arm64__) || defined(_M_ARM64) || defined(_M_ARM64EC)) \ + && !defined(__APPLE__) && XXH_SIZE_OPT <= 0 +# define XXH3_NEON_LANES 6 +# else +# define XXH3_NEON_LANES XXH_ACC_NB +# endif +# 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 +/* Annoyingly, these headers _may_ define three macros: `bool`, `vector`, + * and `pixel`. This is a problem for obvious reasons. + * + * These keywords are unnecessary; the spec literally says they are + * equivalent to `__bool`, `__vector`, and `__pixel` and may be undef'd + * after including the header. + * + * We use pragma push_macro/pop_macro to keep the namespace clean. */ +# pragma push_macro("bool") +# pragma push_macro("vector") +# pragma push_macro("pixel") +/* silence potential macro redefined warnings */ +# undef bool +# undef vector +# undef pixel + +# if defined(__s390x__) +# include +# else +# include +# endif + +/* Restore the original macro values, if applicable. */ +# pragma pop_macro("pixel") +# pragma pop_macro("vector") +# pragma pop_macro("bool") -XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mule(xxh_u32x4 a, xxh_u32x4 b) { +typedef __vector unsigned long long xxh_u64x2; +typedef __vector unsigned char xxh_u8x16; +typedef __vector unsigned xxh_u32x4; - xxh_u64x2 result; - __asm__("vmuleuw %0, %1, %2" : "=v"(result) : "v"(a), "v"(b)); - return result; +/* + * UGLY HACK: Similar to aarch64 macOS GCC, s390x GCC has the same aliasing issue. + */ +typedef xxh_u64x2 xxh_aliasing_u64x2 XXH_ALIASING; + +# 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 +# if defined(__POWER9_VECTOR__) || (defined(__clang__) && defined(__s390x__)) +# define XXH_vec_revb vec_revb +# else +/*! + * A polyfill for POWER9's vec_revb(). + */ +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 vector load and byte swaps it on big endian. + */ +XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(const void *ptr) +{ + xxh_u64x2 ret; + XXH_memcpy(&ret, ptr, sizeof(xxh_u64x2)); +# if XXH_VSX_BE + ret = XXH_vec_revb(ret); +# endif + return ret; } - #endif /* XXH_vec_mulo, XXH_vec_mule */ - #endif /* XXH_VECTOR == XXH_VSX */ +/* + * 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) && !defined(__ibmxl__) +/* Clang has a better way to control this, we can just use the builtin which doesn't swap. */ + /* The IBM XL Compiler (which defined __clang__) only implements the vec_* operations */ +# 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 */ + +#if XXH_VECTOR == XXH_SVE +#define ACCRND(acc, offset) \ +do { \ + svuint64_t input_vec = svld1_u64(mask, xinput + offset); \ + svuint64_t secret_vec = svld1_u64(mask, xsecret + offset); \ + svuint64_t mixed = sveor_u64_x(mask, secret_vec, input_vec); \ + svuint64_t swapped = svtbl_u64(input_vec, kSwap); \ + svuint64_t mixed_lo = svextw_u64_x(mask, mixed); \ + svuint64_t mixed_hi = svlsr_n_u64_x(mask, mixed, 32); \ + svuint64_t mul = svmad_u64_x(mask, mixed_lo, mixed_hi, swapped); \ + acc = svadd_u64_x(mask, acc, mul); \ +} while (0) +#endif /* XXH_VECTOR == XXH_SVE */ + +/* 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 XXH_SIZE_OPT >= 1 +# define XXH_PREFETCH(ptr) (void)(ptr) +# elif defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86)) /* _mm_prefetch() not defined outside of x86/x64 */ +# include /* 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 */ - /* 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_IX86)) /* _mm_prefetch() not defined outside of x86/x64 */ - #include /* 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 - * ========================================== */ +/* ========================================== + * XXH3 default settings + * ========================================== */ - #define XXH_SECRET_DEFAULT_SIZE 192 /* minimum XXH3_SECRET_SIZE_MIN */ +#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 +#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, - +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 +static const xxh_u64 PRIME_MX1 = 0x165667919E3779F9ULL; /*!< 0b0001011001010110011001111001000110011110001101110111100111111001 */ +static const xxh_u64 PRIME_MX2 = 0x9FB21C651E98DF25ULL; /*!< 0b1001111110110010000111000110010100011110100110001101111100100101 */ + +#ifdef XXH_OLD_NAMES +# define kSecret XXH3_kSecret +#endif - #ifdef XXH_DOXYGEN +#ifdef XXH_DOXYGEN /*! * @brief Calculates a 32-bit to 64-bit long multiply. * * Implemented as a macro. * - * 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. + * 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. + * 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. * * @param x, y Numbers to be multiplied * @return 64-bit product of the low 32 bits of @p x and @p y. */ -XXH_FORCE_INLINE xxh_u64 XXH_mult32to64(xxh_u64 x, xxh_u64 y) { - - return (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF); - -} - - #elif defined(_MSC_VER) && defined(_M_IX86) - #include - #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 +XXH_FORCE_INLINE xxh_u64 +XXH_mult32to64(xxh_u64 x, xxh_u64 y) +{ + return (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF); +} +#elif defined(_MSC_VER) && defined(_M_IX86) +# 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 /*! * @brief Calculates a 64->128-bit long multiply. @@ -3623,157 +4299,167 @@ XXH_FORCE_INLINE xxh_u64 XXH_mult32to64(xxh_u64 x, xxh_u64 y) { * Uses `__uint128_t` and `_umul128` if available, otherwise uses a scalar * version. * - * @param lhs, rhs The 64-bit integers to be multiplied + * @param lhs , rhs The 64-bit integers to be multiplied * @return The 128-bit result represented in an @ref XXH128_hash_t. */ -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; +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(__clang__)) && !defined(__wasm__) \ + && defined(__SIZEOF_INT128__) \ + || (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128) - /* - * 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 + __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)) && !defined(_M_ARM64EC) + +#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; + + /* + * MSVC for ARM64's __umulh method. + * + * This compiles to the same MUL + UMULH as GCC/Clang's __uint128_t method. + */ +#elif defined(_M_ARM64) || defined(_M_ARM64EC) + +#ifndef _MSC_VER +# pragma intrinsic(__umulh) +#endif + XXH128_hash_t r128; + r128.low64 = lhs * rhs; + r128.high64 = __umulh(lhs, rhs); + 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 } /*! * @brief Calculates 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. + * around by value. This will hopefully inline the multiply, but we don't force it. * - * @param lhs, rhs The 64-bit integers to multiply + * @param lhs , rhs The 64-bit integers to multiply * @return The low 64 bits of the product XOR'd by the high 64 bits. * @see XXH_mult64to128() */ -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; - +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); - +XXH_FORCE_INLINE XXH_CONSTF 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; - +static XXH64_hash_t XXH3_avalanche(xxh_u64 h64) +{ + h64 = XXH_xorshift64(h64, 37); + h64 *= PRIME_MX1; + h64 = XXH_xorshift64(h64, 32); + return h64; } /* @@ -3781,17 +4467,17 @@ static XXH64_hash_t XXH3_avalanche(xxh_u64 h64) { * 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); - +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 *= PRIME_MX2; + h64 ^= (h64 >> 35) + len ; + h64 *= PRIME_MX2; + return XXH_xorshift64(h64, 28); } + /* ========================================== * Short keys * ========================================== @@ -3800,8 +4486,7 @@ static XXH64_hash_t XXH3_rrmxmx(xxh_u64 h64, xxh_u64 len) { * 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. + * 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. @@ -3826,98 +4511,70 @@ static XXH64_hash_t XXH3_rrmxmx(xxh_u64 h64, xxh_u64 len) { * * 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 XXH_PUREF 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 XXH_PUREF 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(9 <= 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 XXH_PUREF 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(9 <= 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))); - - } - +XXH_FORCE_INLINE XXH_PUREF 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))); + } } /* @@ -3946,113 +4603,106 @@ XXH_FORCE_INLINE XXH64_hash_t XXH3_len_0to16_64b(const xxh_u8 *input, * 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. - */ - XXH_COMPILER_GUARD(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)); - - } - +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. + */ + XXH_COMPILER_GUARD(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); - +XXH_FORCE_INLINE XXH_PUREF 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 XXH_SIZE_OPT >= 1 + /* Smaller and cleaner, but slightly slower. */ + unsigned int i = (unsigned int)(len - 1) / 32; + do { + acc += XXH3_mix16B(input+16 * i, secret+32*i, seed); + acc += XXH3_mix16B(input+len-16*(i+1), secret+32*i+16, seed); + } while (i-- != 0); +#else + 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 + 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); +#endif + return XXH3_avalanche(acc); } - - 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 */ +/*! + * @brief Maximum size of "short" key in bytes. + */ +#define XXH3_MIDSIZE_MAX 240 + +XXH_NO_INLINE XXH_PUREF 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; + xxh_u64 acc_end; + unsigned int const nbRounds = (unsigned int)len / 16; + unsigned int i; + XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX); + for (i=0; i<8; i++) { + acc += XXH3_mix16B(input+(16*i), secret+(16*i), seed); + } + /* last bytes */ + acc_end = XXH3_mix16B(input + len - 16, secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, seed); + XXH_ASSERT(nbRounds >= 8); + acc = XXH3_avalanche(acc); +#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. @@ -4065,2205 +4715,2334 @@ XXH_NO_INLINE XXH64_hash_t XXH3_len_129to240_64b( * 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_DISPATCH_AVX512) && XXH_DISPATCH_AVX512 != 0) - - #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) { - - __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)); - { - - __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)(0U - seed64)); - - const __m512i *const src = (const __m512i *)((const void *)XXH3_kSecret); - __m512i *const dest = (__m512i *)customSecret; - int i; - XXH_ASSERT(((size_t)src & 63) == 0); /* control alignment */ - XXH_ASSERT(((size_t)dest & 63) == 0); - 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 { - - const __m512i *cp; - 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_DISPATCH_AVX2) && XXH_DISPATCH_AVX2 != 0) - - #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); - { - - __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); - { - - __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)(0U - seed64), (xxh_i64)seed64, - (xxh_i64)(0U - seed64), (xxh_i64)seed64); - - const __m256i *const src = (const __m256i *)((const void *)XXH3_kSecret); - __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 - */ - XXH_COMPILER_GUARD(dest); - #endif - XXH_ASSERT(((size_t)src & 31) == 0); /* control alignment */ - XXH_ASSERT(((size_t)dest & 31) == 0); - - /* 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 - - /* x86dispatch always generates SSE2 */ - #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); - { - - __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); - { - - __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)(0U - seed64)}; - __m128i const seed = _mm_load_si128((__m128i const *)seed64x2); - #else - __m128i const seed = - _mm_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64); - #endif - int i; - - const void *const src16 = XXH3_kSecret; - __m128i *dst16 = (__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 - */ - XXH_COMPILER_GUARD(dst16); - #endif - XXH_ASSERT(((size_t)src16 & 15) == 0); /* control alignment */ - XXH_ASSERT(((size_t)dst16 & 15) == 0); - - for (i = 0; i < nbRounds; ++i) { - - dst16[i] = - _mm_add_epi64(_mm_load_si128((const __m128i *)src16 + i), 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); - { - - 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_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_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 - */ - XXH_COMPILER_GUARD(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 - */ - XXH_COMPILER_GUARD(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 === */ - -/*! @ingroup xxh3_family */ -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); - + * 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++) { + /* + * Prevents clang for unrolling the acc loop and interleaving with this one. + */ + XXH_COMPILER_GUARD(acc); + acc_end += XXH3_mix16B(input+(16*i), secret+(16*(i-8)) + XXH3_MIDSIZE_STARTOFFSET, seed); + } + return XXH3_avalanche(acc + acc_end); + } } -/*! @ingroup xxh3_family */ -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); +/* ======= 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)) -/*! @ingroup xxh3_family */ -XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSeed(const void *input, size_t len, - XXH64_hash_t seed) { +#ifdef XXH_OLD_NAMES +# define STRIPE_LEN XXH_STRIPE_LEN +# define ACC_NB XXH_ACC_NB +#endif - return XXH3_64bits_internal(input, len, seed, XXH3_kSecret, - sizeof(XXH3_kSecret), XXH3_hashLong_64b_withSeed); +#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 */ -} +/* + * These macros are to generate an XXH3_accumulate() function. + * The two arguments select the name suffix and target attribute. + * + * The name of this symbol is XXH3_accumulate_() and it calls + * XXH3_accumulate_512_(). + * + * It may be useful to hand implement this function if the compiler fails to + * optimize the inline function. + */ +#define XXH3_ACCUMULATE_TEMPLATE(name) \ +void \ +XXH3_accumulate_##name(xxh_u64* XXH_RESTRICT acc, \ + const xxh_u8* XXH_RESTRICT input, \ + const xxh_u8* XXH_RESTRICT secret, \ + size_t nbStripes) \ +{ \ + 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); \ + XXH3_accumulate_512_##name( \ + acc, \ + in, \ + secret + n*XXH_SECRET_CONSUME_RATE); \ + } \ +} + + +XXH_FORCE_INLINE void XXH_writeLE64(void* dst, xxh_u64 v64) +{ + if (!XXH_CPU_LITTLE_ENDIAN) v64 = XXH_swap64(v64); + XXH_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 streaming === */ /* - * Malloc's a pointer that is always aligned to align. + * XXH3_accumulate_512 is the tightest loop for long inputs, and it is the most optimized. * - * This must be freed with `XXH_alignedFree()`. + * It is a hardened version of UMAC, based off of FARSH's implementation. * - * 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 was chosen because it adapts quite well to 32-bit, 64-bit, and SIMD + * implementations, and it is ridiculously fast. * - * 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. + * We harden it by mixing the original input to the accumulators as well as the product. * - * 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. + * This means that in the (relatively likely) case of a multiply by zero, the + * original input is preserved. * - * The method is simple: Overallocate, manually align, and store the offset - * to the original behind the returned pointer. + * 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. * - * Align must be a power of 2 and 8 <= align <= 128. + * 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. */ -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; +#if (XXH_VECTOR == XXH_AVX512) \ + || (defined(XXH_DISPATCH_AVX512) && XXH_DISPATCH_AVX512 != 0) - XXH_ASSERT((size_t)ptr % align == 0); +#ifndef XXH_TARGET_AVX512 +# define XXH_TARGET_AVX512 /* disable attribute target */ +#endif - /* Store the offset immediately before the returned pointer. */ - ptr[-1] = (xxh_u8)offset; - return ptr; +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) +{ + __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_srli_epi64 (data_key, 32); + /* 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); } - - return NULL; - - } - } +XXH_FORCE_INLINE XXH_TARGET_AVX512 XXH3_ACCUMULATE_TEMPLATE(avx512) /* - * Frees an aligned pointer allocated by XXH_alignedMalloc(). Don't pass - * normal malloc'd pointers, XXH_alignedMalloc has a specific data layout. + * 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. */ -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_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)); + { __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); + /* xacc[0] ^= secret; */ + __m512i const key_vec = _mm512_loadu_si512 (secret); + __m512i const data_key = _mm512_ternarylogic_epi32(key_vec, acc_vec, shifted, 0x96 /* key_vec ^ acc_vec ^ shifted */); + + /* xacc[0] *= XXH_PRIME32_1; */ + __m512i const data_key_hi = _mm512_srli_epi64 (data_key, 32); + __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)); + } } -/*! @ingroup xxh3_family */ -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_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_pos = _mm512_set1_epi64((xxh_i64)seed64); + __m512i const seed = _mm512_mask_sub_epi64(seed_pos, 0xAA, _mm512_set1_epi8(0), seed_pos); + const __m512i* const src = (const __m512i*) ((const void*) XXH3_kSecret); + __m512i* const dest = ( __m512i*) customSecret; + int i; + XXH_ASSERT(((size_t)src & 63) == 0); /* control alignment */ + XXH_ASSERT(((size_t)dest & 63) == 0); + for (i=0; i < nbRounds; ++i) { + dest[i] = _mm512_add_epi64(_mm512_load_si512(src + i), seed); + } } } -/*! @ingroup xxh3_family */ -XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t *statePtr) { - - XXH_alignedFree(statePtr); - return XXH_OK; - -} +#endif -/*! @ingroup xxh3_family */ -XXH_PUBLIC_API void XXH3_copyState(XXH3_state_t *dst_state, - const XXH3_state_t *src_state) { +#if (XXH_VECTOR == XXH_AVX2) \ + || (defined(XXH_DISPATCH_AVX2) && XXH_DISPATCH_AVX2 != 0) - memcpy(dst_state, src_state, sizeof(*dst_state)); +#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); + { __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_srli_epi64 (data_key, 32); + /* 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 XXH3_ACCUMULATE_TEMPLATE(avx2) + +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); + { __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_srli_epi64 (data_key, 32); + __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)); + } + } } -static void XXH3_reset_internal(XXH3_state_t *statePtr, XXH64_hash_t seed, - const void *secret, size_t secretSize) { +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)(0U - seed64), (xxh_i64)seed64, (xxh_i64)(0U - seed64), (xxh_i64)seed64); - 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; + const __m256i* const src = (const __m256i*) ((const void*) XXH3_kSecret); + __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 + */ + XXH_COMPILER_GUARD(dest); +# endif + XXH_ASSERT(((size_t)src & 31) == 0); /* control alignment */ + XXH_ASSERT(((size_t)dest & 31) == 0); + + /* GCC -O2 need unroll loop manually */ + dest[0] = _mm256_add_epi64(_mm256_load_si256(src+0), seed); + dest[1] = _mm256_add_epi64(_mm256_load_si256(src+1), seed); + dest[2] = _mm256_add_epi64(_mm256_load_si256(src+2), seed); + dest[3] = _mm256_add_epi64(_mm256_load_si256(src+3), seed); + dest[4] = _mm256_add_epi64(_mm256_load_si256(src+4), seed); + dest[5] = _mm256_add_epi64(_mm256_load_si256(src+5), seed); + } } -/*! @ingroup xxh3_family */ -XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH3_state_t *statePtr) { - - if (statePtr == NULL) return XXH_ERROR; - XXH3_reset_internal(statePtr, 0, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE); - return XXH_OK; - -} +#endif -/*! @ingroup xxh3_family */ -XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret( - XXH3_state_t *statePtr, const void *secret, size_t secretSize) { +/* x86dispatch always generates SSE2 */ +#if (XXH_VECTOR == XXH_SSE2) || defined(XXH_X86DISPATCH) - if (statePtr == NULL) return XXH_ERROR; - XXH3_reset_internal(statePtr, 0, secret, secretSize); - if (secret == NULL) return XXH_ERROR; - if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR; - return XXH_OK; +#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); + { __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 XXH3_ACCUMULATE_TEMPLATE(sse2) + +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); + { __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)); + } + } } -/*! @ingroup xxh3_family */ -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_reset_internal(statePtr, seed, NULL, XXH_SECRET_DEFAULT_SIZE); - return XXH_OK; +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)(0U - seed64) }; + __m128i const seed = _mm_load_si128((__m128i const*)seed64x2); +# else + __m128i const seed = _mm_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64); +# endif + int i; + + const void* const src16 = XXH3_kSecret; + __m128i* dst16 = (__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 + */ + XXH_COMPILER_GUARD(dst16); +# endif + XXH_ASSERT(((size_t)src16 & 15) == 0); /* control alignment */ + XXH_ASSERT(((size_t)dst16 & 15) == 0); + for (i=0; i < nbRounds; ++i) { + dst16[i] = _mm_add_epi64(_mm_load_si128((const __m128i *)src16+i), seed); + } } } -/* 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 { +#endif - XXH3_accumulate(acc, input, - secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, - nbStripes, f_acc512); - *nbStripesSoFarPtr += nbStripes; +#if (XXH_VECTOR == XXH_NEON) - } +/* forward declarations for the scalar routines */ +XXH_FORCE_INLINE void +XXH3_scalarRound(void* XXH_RESTRICT acc, void const* XXH_RESTRICT input, + void const* XXH_RESTRICT secret, size_t lane); -} +XXH_FORCE_INLINE void +XXH3_scalarScrambleRound(void* XXH_RESTRICT acc, + void const* XXH_RESTRICT secret, size_t lane); -/* - * Both XXH3_64bits_update and XXH3_128bits_update use this routine. +/*! + * @internal + * @brief The bulk processing loop for NEON and WASM SIMD128. + * + * The NEON code path is actually partially scalar when running on AArch64. This + * is to optimize the pipelining and can have up to 15% speedup depending on the + * CPU, and it also mitigates some GCC codegen issues. + * + * @see XXH3_NEON_LANES for configuring this and details about this optimization. + * + * NEON's 32-bit to 64-bit long multiply takes a half vector of 32-bit + * integers instead of the other platforms which mask full 64-bit vectors, + * so the setup is more complicated than just shifting right. + * + * Additionally, there is an optimization for 4 lanes at once noted below. + * + * Since, as stated, the most optimal amount of lanes for Cortexes is 6, + * there needs to be *three* versions of the accumulate operation used + * for the remaining 2 lanes. + * + * WASM's SIMD128 uses SIMDe's arm_neon.h polyfill because the intrinsics overlap + * nearly perfectly. */ -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; - XXH_ASSERT(state->bufferedSize <= XXH3_INTERNALBUFFER_SIZE); - - 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; +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_STATIC_ASSERT(XXH3_NEON_LANES > 0 && XXH3_NEON_LANES <= XXH_ACC_NB && XXH3_NEON_LANES % 2 == 0); + { /* GCC for darwin arm64 does not like aliasing here */ + xxh_aliasing_uint64x2_t* const xacc = (xxh_aliasing_uint64x2_t*) acc; + /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. */ + uint8_t const* xinput = (const uint8_t *) input; + uint8_t const* xsecret = (const uint8_t *) secret; + + size_t i; +#ifdef __wasm_simd128__ + /* + * On WASM SIMD128, Clang emits direct address loads when XXH3_kSecret + * is constant propagated, which results in it converting it to this + * inside the loop: + * + * a = v128.load(XXH3_kSecret + 0 + $secret_offset, offset = 0) + * b = v128.load(XXH3_kSecret + 16 + $secret_offset, offset = 0) + * ... + * + * This requires a full 32-bit address immediate (and therefore a 6 byte + * instruction) as well as an add for each offset. + * + * Putting an asm guard prevents it from folding (at the cost of losing + * the alignment hint), and uses the free offset in `v128.load` instead + * of adding secret_offset each time which overall reduces code size by + * about a kilobyte and improves performance. + */ + XXH_COMPILER_GUARD(xsecret); +#endif + /* Scalar lanes use the normal scalarRound routine */ + for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) { + XXH3_scalarRound(acc, input, secret, i); + } + i = 0; + /* 4 NEON lanes at a time. */ + for (; i+1 < XXH3_NEON_LANES / 2; i+=2) { + /* data_vec = xinput[i]; */ + uint64x2_t data_vec_1 = XXH_vld1q_u64(xinput + (i * 16)); + uint64x2_t data_vec_2 = XXH_vld1q_u64(xinput + ((i+1) * 16)); + /* key_vec = xsecret[i]; */ + uint64x2_t key_vec_1 = XXH_vld1q_u64(xsecret + (i * 16)); + uint64x2_t key_vec_2 = XXH_vld1q_u64(xsecret + ((i+1) * 16)); + /* data_swap = swap(data_vec) */ + uint64x2_t data_swap_1 = vextq_u64(data_vec_1, data_vec_1, 1); + uint64x2_t data_swap_2 = vextq_u64(data_vec_2, data_vec_2, 1); + /* data_key = data_vec ^ key_vec; */ + uint64x2_t data_key_1 = veorq_u64(data_vec_1, key_vec_1); + uint64x2_t data_key_2 = veorq_u64(data_vec_2, key_vec_2); + + /* + * If we reinterpret the 64x2 vectors as 32x4 vectors, we can use a + * de-interleave operation for 4 lanes in 1 step with `vuzpq_u32` to + * get one vector with the low 32 bits of each lane, and one vector + * with the high 32 bits of each lane. + * + * The intrinsic returns a double vector because the original ARMv7-a + * instruction modified both arguments in place. AArch64 and SIMD128 emit + * two instructions from this intrinsic. + * + * [ dk11L | dk11H | dk12L | dk12H ] -> [ dk11L | dk12L | dk21L | dk22L ] + * [ dk21L | dk21H | dk22L | dk22H ] -> [ dk11H | dk12H | dk21H | dk22H ] + */ + uint32x4x2_t unzipped = vuzpq_u32( + vreinterpretq_u32_u64(data_key_1), + vreinterpretq_u32_u64(data_key_2) + ); + /* data_key_lo = data_key & 0xFFFFFFFF */ + uint32x4_t data_key_lo = unzipped.val[0]; + /* data_key_hi = data_key >> 32 */ + uint32x4_t data_key_hi = unzipped.val[1]; + /* + * Then, we can split the vectors horizontally and multiply which, as for most + * widening intrinsics, have a variant that works on both high half vectors + * for free on AArch64. A similar instruction is available on SIMD128. + * + * sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi + */ + uint64x2_t sum_1 = XXH_vmlal_low_u32(data_swap_1, data_key_lo, data_key_hi); + uint64x2_t sum_2 = XXH_vmlal_high_u32(data_swap_2, data_key_lo, data_key_hi); + /* + * Clang reorders + * a += b * c; // umlal swap.2d, dkl.2s, dkh.2s + * c += a; // add acc.2d, acc.2d, swap.2d + * to + * c += a; // add acc.2d, acc.2d, swap.2d + * c += b * c; // umlal acc.2d, dkl.2s, dkh.2s + * + * While it would make sense in theory since the addition is faster, + * for reasons likely related to umlal being limited to certain NEON + * pipelines, this is worse. A compiler guard fixes this. + */ + XXH_COMPILER_GUARD_CLANG_NEON(sum_1); + XXH_COMPILER_GUARD_CLANG_NEON(sum_2); + /* xacc[i] = acc_vec + sum; */ + xacc[i] = vaddq_u64(xacc[i], sum_1); + xacc[i+1] = vaddq_u64(xacc[i+1], sum_2); + } + /* Operate on the remaining NEON lanes 2 at a time. */ + for (; i < XXH3_NEON_LANES / 2; i++) { + /* data_vec = xinput[i]; */ + uint64x2_t data_vec = XXH_vld1q_u64(xinput + (i * 16)); + /* key_vec = xsecret[i]; */ + uint64x2_t key_vec = XXH_vld1q_u64(xsecret + (i * 16)); + /* acc_vec_2 = swap(data_vec) */ + uint64x2_t data_swap = vextq_u64(data_vec, data_vec, 1); + /* data_key = data_vec ^ key_vec; */ + uint64x2_t data_key = veorq_u64(data_vec, key_vec); + /* For two lanes, just use VMOVN and VSHRN. */ + /* data_key_lo = data_key & 0xFFFFFFFF; */ + uint32x2_t data_key_lo = vmovn_u64(data_key); + /* data_key_hi = data_key >> 32; */ + uint32x2_t data_key_hi = vshrn_n_u64(data_key, 32); + /* sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi; */ + uint64x2_t sum = vmlal_u32(data_swap, data_key_lo, data_key_hi); + /* Same Clang workaround as before */ + XXH_COMPILER_GUARD_CLANG_NEON(sum); + /* xacc[i] = acc_vec + sum; */ + xacc[i] = vaddq_u64 (xacc[i], sum); + } } +} +XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(neon) - /* 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 */ +XXH_FORCE_INLINE void +XXH3_scrambleAcc_neon(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret) +{ + XXH_ASSERT((((size_t)acc) & 15) == 0); - /* - * Internal buffer is partially filled (always, except at beginning) - * Complete it, then consume it. - */ - if (state->bufferedSize) { + { xxh_aliasing_uint64x2_t* xacc = (xxh_aliasing_uint64x2_t*) acc; + uint8_t const* xsecret = (uint8_t const*) secret; - 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; + size_t i; + /* WASM uses operator overloads and doesn't need these. */ +#ifndef __wasm_simd128__ + /* { prime32_1, prime32_1 } */ + uint32x2_t const kPrimeLo = vdup_n_u32(XXH_PRIME32_1); + /* { 0, prime32_1, 0, prime32_1 } */ + uint32x4_t const kPrimeHi = vreinterpretq_u32_u64(vdupq_n_u64((xxh_u64)XXH_PRIME32_1 << 32)); +#endif + /* AArch64 uses both scalar and neon at the same time */ + for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) { + XXH3_scalarScrambleRound(acc, secret, i); + } + for (i=0; i < XXH3_NEON_LANES / 2; 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]; */ + uint64x2_t key_vec = XXH_vld1q_u64(xsecret + (i * 16)); + uint64x2_t data_key = veorq_u64(data_vec, key_vec); + /* xacc[i] *= XXH_PRIME32_1 */ +#ifdef __wasm_simd128__ + /* SIMD128 has multiply by u64x2, use it instead of expanding and scalarizing */ + xacc[i] = data_key * XXH_PRIME32_1; +#else + /* + * Expanded version with portable NEON intrinsics + * + * lo(x) * lo(y) + (hi(x) * lo(y) << 32) + * + * prod_hi = hi(data_key) * lo(prime) << 32 + * + * Since we only need 32 bits of this multiply a trick can be used, reinterpreting the vector + * as a uint32x4_t and multiplying by { 0, prime, 0, prime } to cancel out the unwanted bits + * and avoid the shift. + */ + uint32x4_t prod_hi = vmulq_u32 (vreinterpretq_u32_u64(data_key), kPrimeHi); + /* Extract low bits for vmlal_u32 */ + uint32x2_t data_key_lo = vmovn_u64(data_key); + /* xacc[i] = prod_hi + lo(data_key) * XXH_PRIME32_1; */ + xacc[i] = vmlal_u32(vreinterpretq_u64_u32(prod_hi), data_key_lo, kPrimeLo); +#endif + } } +} +#endif - XXH_ASSERT(input < bEnd); - - /* Consume input by a multiple of internal buffer size */ - if (bEnd - input > XXH3_INTERNALBUFFER_SIZE) { - - 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); - +#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) +{ + /* presumed aligned */ + xxh_aliasing_u64x2* const xacc = (xxh_aliasing_u64x2*) acc; + xxh_u8 const* const xinput = (xxh_u8 const*) input; /* no alignment restriction */ + xxh_u8 const* const xsecret = (xxh_u8 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 + 16*i); + /* key_vec = xsecret[i]; */ + xxh_u64x2 const key_vec = XXH_vec_loadu(xsecret + 16*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); + /* acc_vec = xacc[i]; */ + xxh_u64x2 acc_vec = xacc[i]; + acc_vec += product; + + /* swap high and low halves */ +#ifdef __s390x__ + acc_vec += vec_permi(data_vec, data_vec, 2); +#else + acc_vec += vec_xxpermdi(data_vec, data_vec, 2); +#endif + xacc[i] = acc_vec; } - - 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; - } - -/*! @ingroup xxh3_family */ -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 XXH3_ACCUMULATE_TEMPLATE(vsx) + +XXH_FORCE_INLINE void +XXH3_scrambleAcc_vsx(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret) +{ + XXH_ASSERT((((size_t)acc) & 15) == 0); + + { xxh_aliasing_u64x2* const xacc = (xxh_aliasing_u64x2*) acc; + const xxh_u8* const xsecret = (const xxh_u8*) 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 + 16*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); + } } } -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); - - } +#endif +#if (XXH_VECTOR == XXH_SVE) + +XXH_FORCE_INLINE void +XXH3_accumulate_512_sve( void* XXH_RESTRICT acc, + const void* XXH_RESTRICT input, + const void* XXH_RESTRICT secret) +{ + uint64_t *xacc = (uint64_t *)acc; + const uint64_t *xinput = (const uint64_t *)(const void *)input; + const uint64_t *xsecret = (const uint64_t *)(const void *)secret; + svuint64_t kSwap = sveor_n_u64_z(svptrue_b64(), svindex_u64(0, 1), 1); + uint64_t element_count = svcntd(); + if (element_count >= 8) { + svbool_t mask = svptrue_pat_b64(SV_VL8); + svuint64_t vacc = svld1_u64(mask, xacc); + ACCRND(vacc, 0); + svst1_u64(mask, xacc, vacc); + } else if (element_count == 2) { /* sve128 */ + svbool_t mask = svptrue_pat_b64(SV_VL2); + svuint64_t acc0 = svld1_u64(mask, xacc + 0); + svuint64_t acc1 = svld1_u64(mask, xacc + 2); + svuint64_t acc2 = svld1_u64(mask, xacc + 4); + svuint64_t acc3 = svld1_u64(mask, xacc + 6); + ACCRND(acc0, 0); + ACCRND(acc1, 2); + ACCRND(acc2, 4); + ACCRND(acc3, 6); + svst1_u64(mask, xacc + 0, acc0); + svst1_u64(mask, xacc + 2, acc1); + svst1_u64(mask, xacc + 4, acc2); + svst1_u64(mask, xacc + 6, acc3); + } else { + svbool_t mask = svptrue_pat_b64(SV_VL4); + svuint64_t acc0 = svld1_u64(mask, xacc + 0); + svuint64_t acc1 = svld1_u64(mask, xacc + 4); + ACCRND(acc0, 0); + ACCRND(acc1, 4); + svst1_u64(mask, xacc + 0, acc0); + svst1_u64(mask, xacc + 4, acc1); + } } -/*! @ingroup xxh3_family */ -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); - +XXH_FORCE_INLINE void +XXH3_accumulate_sve(xxh_u64* XXH_RESTRICT acc, + const xxh_u8* XXH_RESTRICT input, + const xxh_u8* XXH_RESTRICT secret, + size_t nbStripes) +{ + if (nbStripes != 0) { + uint64_t *xacc = (uint64_t *)acc; + const uint64_t *xinput = (const uint64_t *)(const void *)input; + const uint64_t *xsecret = (const uint64_t *)(const void *)secret; + svuint64_t kSwap = sveor_n_u64_z(svptrue_b64(), svindex_u64(0, 1), 1); + uint64_t element_count = svcntd(); + if (element_count >= 8) { + svbool_t mask = svptrue_pat_b64(SV_VL8); + svuint64_t vacc = svld1_u64(mask, xacc + 0); + do { + /* svprfd(svbool_t, void *, enum svfprop); */ + svprfd(mask, xinput + 128, SV_PLDL1STRM); + ACCRND(vacc, 0); + xinput += 8; + xsecret += 1; + nbStripes--; + } while (nbStripes != 0); + + svst1_u64(mask, xacc + 0, vacc); + } else if (element_count == 2) { /* sve128 */ + svbool_t mask = svptrue_pat_b64(SV_VL2); + svuint64_t acc0 = svld1_u64(mask, xacc + 0); + svuint64_t acc1 = svld1_u64(mask, xacc + 2); + svuint64_t acc2 = svld1_u64(mask, xacc + 4); + svuint64_t acc3 = svld1_u64(mask, xacc + 6); + do { + svprfd(mask, xinput + 128, SV_PLDL1STRM); + ACCRND(acc0, 0); + ACCRND(acc1, 2); + ACCRND(acc2, 4); + ACCRND(acc3, 6); + xinput += 8; + xsecret += 1; + nbStripes--; + } while (nbStripes != 0); + + svst1_u64(mask, xacc + 0, acc0); + svst1_u64(mask, xacc + 2, acc1); + svst1_u64(mask, xacc + 4, acc2); + svst1_u64(mask, xacc + 6, acc3); + } else { + svbool_t mask = svptrue_pat_b64(SV_VL4); + svuint64_t acc0 = svld1_u64(mask, xacc + 0); + svuint64_t acc1 = svld1_u64(mask, xacc + 4); + do { + svprfd(mask, xinput + 128, SV_PLDL1STRM); + ACCRND(acc0, 0); + ACCRND(acc1, 4); + xinput += 8; + xsecret += 1; + nbStripes--; + } while (nbStripes != 0); + + svst1_u64(mask, xacc + 0, acc0); + svst1_u64(mask, xacc + 4, acc1); + } + } } - #define XXH_MIN(x, y) (((x) > (y)) ? (y) : (x)) - -/*! @ingroup xxh3_family */ -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); +#endif - { +/* scalar variants - universal */ - 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)); +#if defined(__aarch64__) && (defined(__GNUC__) || defined(__clang__)) +/* + * In XXH3_scalarRound(), GCC and Clang have a similar codegen issue, where they + * emit an excess mask and a full 64-bit multiply-add (MADD X-form). + * + * While this might not seem like much, as AArch64 is a 64-bit architecture, only + * big Cortex designs have a full 64-bit multiplier. + * + * On the little cores, the smaller 32-bit multiplier is used, and full 64-bit + * multiplies expand to 2-3 multiplies in microcode. This has a major penalty + * of up to 4 latency cycles and 2 stall cycles in the multiply pipeline. + * + * Thankfully, AArch64 still provides the 32-bit long multiply-add (UMADDL) which does + * not have this penalty and does the mask automatically. + */ +XXH_FORCE_INLINE xxh_u64 +XXH_mult32to64_add64(xxh_u64 lhs, xxh_u64 rhs, xxh_u64 acc) +{ + xxh_u64 ret; + /* note: %x = 64-bit register, %w = 32-bit register */ + __asm__("umaddl %x0, %w1, %w2, %x3" : "=r" (ret) : "r" (lhs), "r" (rhs), "r" (acc)); + return ret; +} +#else +XXH_FORCE_INLINE xxh_u64 +XXH_mult32to64_add64(xxh_u64 lhs, xxh_u64 rhs, xxh_u64 acc) +{ + return XXH_mult32to64((xxh_u32)lhs, (xxh_u32)rhs) + acc; +} +#endif - /* - * Copy customSeed to seeds[], truncating or repeating as necessary. - */ +/*! + * @internal + * @brief Scalar round for @ref XXH3_accumulate_512_scalar(). + * + * This is extracted to its own function because the NEON path uses a combination + * of NEON and scalar. + */ +XXH_FORCE_INLINE void +XXH3_scalarRound(void* XXH_RESTRICT acc, + void const* XXH_RESTRICT input, + void const* XXH_RESTRICT secret, + size_t lane) +{ + xxh_u64* xacc = (xxh_u64*) acc; + xxh_u8 const* xinput = (xxh_u8 const*) input; + xxh_u8 const* xsecret = (xxh_u8 const*) secret; + XXH_ASSERT(lane < XXH_ACC_NB); + XXH_ASSERT(((size_t)acc & (XXH_ACC_ALIGN-1)) == 0); { - - 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; - - } - + xxh_u64 const data_val = XXH_readLE64(xinput + lane * 8); + xxh_u64 const data_key = data_val ^ XXH_readLE64(xsecret + lane * 8); + xacc[lane ^ 1] += data_val; /* swap adjacent lanes */ + xacc[lane] = XXH_mult32to64_add64(data_key /* & 0xFFFFFFFF */, data_key >> 32, xacc[lane]); } +} - /* 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)); - +/*! + * @internal + * @brief Processes a 64 byte block of data using the scalar path. + */ +XXH_FORCE_INLINE void +XXH3_accumulate_512_scalar(void* XXH_RESTRICT acc, + const void* XXH_RESTRICT input, + const void* XXH_RESTRICT secret) +{ + size_t i; + /* ARM GCC refuses to unroll this loop, resulting in a 24% slowdown on ARMv6. */ +#if defined(__GNUC__) && !defined(__clang__) \ + && (defined(__arm__) || defined(__thumb2__)) \ + && defined(__ARM_FEATURE_UNALIGNED) /* no unaligned access just wastes bytes */ \ + && XXH_SIZE_OPT <= 0 +# pragma GCC unroll 8 +#endif + for (i=0; i < XXH_ACC_NB; i++) { + XXH3_scalarRound(acc, input, secret, i); } - - } - } +XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(scalar) -/* ========================================== - * 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. +/*! + * @internal + * @brief Scalar scramble step for @ref XXH3_scrambleAcc_scalar(). * - * 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). + * This is extracted to its own function because the NEON path uses a combination + * of NEON and scalar. */ - -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 void +XXH3_scalarScrambleRound(void* XXH_RESTRICT acc, + void const* XXH_RESTRICT secret, + size_t lane) +{ + xxh_u64* const xacc = (xxh_u64*) acc; /* presumed aligned */ + const xxh_u8* const xsecret = (const xxh_u8*) secret; /* no alignment restriction */ + XXH_ASSERT((((size_t)acc) & (XXH_ACC_ALIGN-1)) == 0); + XXH_ASSERT(lane < XXH_ACC_NB); + { + xxh_u64 const key64 = XXH_readLE64(xsecret + lane * 8); + xxh_u64 acc64 = xacc[lane]; + acc64 = XXH_xorshift64(acc64, 47); + acc64 ^= key64; + acc64 *= XXH_PRIME32_1; + xacc[lane] = acc64; + } } -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); - { +/*! + * @internal + * @brief Scrambles the accumulators after a large chunk has been read + */ +XXH_FORCE_INLINE void +XXH3_scrambleAcc_scalar(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret) +{ + size_t i; + for (i=0; i < XXH_ACC_NB; i++) { + XXH3_scalarScrambleRound(acc, secret, i); + } +} - 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); +XXH_FORCE_INLINE void +XXH3_initCustomSecret_scalar(void* XXH_RESTRICT customSecret, xxh_u64 seed64) +{ /* - * 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. + * We need a separate pointer for the hack below, + * which requires a non-const pointer. + * Any decent compiler will optimize this out otherwise. */ - m128.low64 += (xxh_u64)(len - 1) << 54; - input_hi ^= bitfliph; + const xxh_u8* kSecretPtr = XXH3_kSecret; + XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0); + +#if defined(__GNUC__) && defined(__aarch64__) /* - * 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. + * UGLY HACK: + * GCC and Clang generate a bunch of MOV/MOVK pairs for aarch64, and they are + * placed sequentially, in order, at the top of the unrolled loop. * - * The best approach to this operation is different on 32-bit and 64-bit. + * While MOVK is great for generating constants (2 cycles for a 64-bit + * constant compared to 4 cycles for LDR), it fights for bandwidth with + * the arithmetic instructions. + * + * I L S + * MOVK + * MOVK + * MOVK + * MOVK + * ADD + * SUB STR + * STR + * By forcing loads from memory (as the asm line causes the compiler to assume + * that XXH3_kSecretPtr has been changed), the pipelines are used more + * efficiently: + * I L S + * LDR + * ADD LDR + * SUB STR + * STR + * + * See XXH3_NEON_LANES for details on the pipsline. + * + * XXH3_64bits_withSeed, len == 256, Snapdragon 835 + * without hack: 2654.4 MB/s + * with hack: 3202.9 MB/s */ - 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; - - } - - } - + XXH_COMPILER_GUARD(kSecretPtr); +#endif + { int const nbRounds = XXH_SECRET_DEFAULT_SIZE / 16; + int i; + for (i=0; i < nbRounds; i++) { + /* + * The asm hack causes the compiler 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); + } } } -/* - * 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; - - } +typedef void (*XXH3_f_accumulate)(xxh_u64* XXH_RESTRICT, const xxh_u8* XXH_RESTRICT, const xxh_u8* XXH_RESTRICT, size_t); +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) -/* - * 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) { +#define XXH3_accumulate_512 XXH3_accumulate_512_avx512 +#define XXH3_accumulate XXH3_accumulate_avx512 +#define XXH3_scrambleAcc XXH3_scrambleAcc_avx512 +#define XXH3_initCustomSecret XXH3_initCustomSecret_avx512 - 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; +#elif (XXH_VECTOR == XXH_AVX2) -} +#define XXH3_accumulate_512 XXH3_accumulate_512_avx2 +#define XXH3_accumulate XXH3_accumulate_avx2 +#define XXH3_scrambleAcc XXH3_scrambleAcc_avx2 +#define XXH3_initCustomSecret XXH3_initCustomSecret_avx2 -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) { +#elif (XXH_VECTOR == XXH_SSE2) - XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); - (void)secretSize; - XXH_ASSERT(16 < len && len <= 128); +#define XXH3_accumulate_512 XXH3_accumulate_512_sse2 +#define XXH3_accumulate XXH3_accumulate_sse2 +#define XXH3_scrambleAcc XXH3_scrambleAcc_sse2 +#define XXH3_initCustomSecret XXH3_initCustomSecret_sse2 - { +#elif (XXH_VECTOR == XXH_NEON) - XXH128_hash_t acc; - acc.low64 = len * XXH_PRIME64_1; - acc.high64 = 0; - if (len > 32) { +#define XXH3_accumulate_512 XXH3_accumulate_512_neon +#define XXH3_accumulate XXH3_accumulate_neon +#define XXH3_scrambleAcc XXH3_scrambleAcc_neon +#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar - if (len > 64) { +#elif (XXH_VECTOR == XXH_VSX) - if (len > 96) { +#define XXH3_accumulate_512 XXH3_accumulate_512_vsx +#define XXH3_accumulate XXH3_accumulate_vsx +#define XXH3_scrambleAcc XXH3_scrambleAcc_vsx +#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar - acc = XXH128_mix32B(acc, input + 48, input + len - 64, secret + 96, - seed); +#elif (XXH_VECTOR == XXH_SVE) +#define XXH3_accumulate_512 XXH3_accumulate_512_sve +#define XXH3_accumulate XXH3_accumulate_sve +#define XXH3_scrambleAcc XXH3_scrambleAcc_scalar +#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar - } +#else /* scalar */ - acc = - XXH128_mix32B(acc, input + 32, input + len - 48, secret + 64, seed); +#define XXH3_accumulate_512 XXH3_accumulate_512_scalar +#define XXH3_accumulate XXH3_accumulate_scalar +#define XXH3_scrambleAcc XXH3_scrambleAcc_scalar +#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar - } +#endif - acc = XXH128_mix32B(acc, input + 16, input + len - 32, secret + 32, seed); +#if XXH_SIZE_OPT >= 1 /* don't do SIMD for initialization */ +# undef XXH3_initCustomSecret +# define XXH3_initCustomSecret XXH3_initCustomSecret_scalar +#endif - } +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 f_acc, + 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; - acc = XXH128_mix32B(acc, input, input + len - 16, secret, seed); - { + size_t n; - 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_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); + for (n = 0; n < nb_blocks; n++) { + f_acc(acc, input + n*block_len, secret, nbStripesPerBlock); + 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)); + f_acc(acc, input + nb_blocks*block_len, secret, nbStripes); + /* 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 */ + XXH3_accumulate_512(acc, p, secret + secretSize - XXH_STRIPE_LEN - XXH_SECRET_LASTACC_START); + } } } -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); +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; - 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); - + 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 + */ + XXH_COMPILER_GUARD(result64); +#endif } - 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); + 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 } - /* last bytes */ - acc = XXH128_mix32B( - acc, input + len - 16, input + len - 32, - secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET - 16, - 0ULL - seed); +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 f_acc, + 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_acc, f_scramble); - 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; + /* 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 to transmit secret's size (when it's static) + * so that the compiler can properly optimize the vectorized loop. + * This makes a big performance difference for "medium" keys (<1 KB) when using AVX instruction set. + * When the secret size is unknown, or on GCC 12 where the mix of NO_INLINE and FORCE_INLINE + * breaks -Og, this is XXH_NO_INLINE. + */ +XXH3_WITH_SECRET_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, XXH3_scrambleAcc); +} - } +/* + * It's preferable for performance that XXH3_hashLong is not inlined, + * as it results in a smaller function for small data, easier to the instruction cache. + * Note that inside this no_inline function, we do inline the internal loop, + * and provide a statically defined secret size to allow optimization of vector loop. + */ +XXH_NO_INLINE XXH_PUREF 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, 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 f_acc, + XXH3_f_scrambleAcc f_scramble, + XXH3_f_initCustomSecret f_initSec) +{ +#if XXH_SIZE_OPT <= 0 + if (seed == 0) + return XXH3_hashLong_64b_internal(input, len, + XXH3_kSecret, sizeof(XXH3_kSecret), + f_acc, f_scramble); +#endif + { 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_acc, f_scramble); + } } -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) { +/* + * It's important for performance that XXH3_hashLong is not inlined. + */ +XXH_NO_INLINE XXH64_hash_t +XXH3_hashLong_64b_withSeed(const void* XXH_RESTRICT input, size_t len, + XXH64_hash_t seed, const xxh_u8* XXH_RESTRICT secret, size_t secretLen) +{ + (void)secret; (void)secretLen; + return XXH3_hashLong_64b_withSeed_internal(input, len, seed, + XXH3_accumulate, XXH3_scrambleAcc, XXH3_initCustomSecret); +} - 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); +typedef XXH64_hash_t (*XXH3_hashLong64_f)(const void* XXH_RESTRICT, size_t, + XXH64_hash_t, const xxh_u8* XXH_RESTRICT, size_t); - /* converge into final hash */ - XXH_STATIC_ASSERT(sizeof(acc) == 64); - XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START); - { +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); +} - 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; - } +/* === Public entry point === */ +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length) +{ + return XXH3_64bits_internal(input, length, 0, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_default); } -/* - * 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) { +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH64_hash_t +XXH3_64bits_withSecret(XXH_NOESCAPE const void* input, size_t length, XXH_NOESCAPE const void* secret, size_t secretSize) +{ + return XXH3_64bits_internal(input, length, 0, secret, secretSize, XXH3_hashLong_64b_withSecret); +} - (void)seed64; - (void)secret; - (void)secretLen; - return XXH3_hashLong_128b_internal(input, len, XXH3_kSecret, - sizeof(XXH3_kSecret), XXH3_accumulate_512, - XXH3_scrambleAcc); +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH64_hash_t +XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed) +{ + return XXH3_64bits_internal(input, length, seed, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_withSeed); +} +XXH_PUBLIC_API XXH64_hash_t +XXH3_64bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t length, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed) +{ + if (length <= XXH3_MIDSIZE_MAX) + return XXH3_64bits_internal(input, length, seed, XXH3_kSecret, sizeof(XXH3_kSecret), NULL); + return XXH3_hashLong_64b_withSecret(input, length, seed, (const xxh_u8*)secret, secretSize); } + +/* === XXH3 streaming === */ +#ifndef XXH_NO_STREAM /* - * It's important for performance that XXH3_hashLong is not inlined. + * 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. */ -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); +static XXH_MALLOCF 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); + } +} +/*! @ingroup XXH3_family */ +/*! + * @brief Allocate an @ref XXH3_state_t. + * + * @return An allocated pointer of @ref XXH3_state_t on success. + * @return `NULL` on failure. + * + * @note Must be freed with XXH3_freeState(). + * + * @see @ref streaming_example "Streaming Example" + */ +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; +} +/*! @ingroup XXH3_family */ +/*! + * @brief Frees an @ref XXH3_state_t. + * + * @param statePtr A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState(). + * + * @return @ref XXH_OK. + * + * @note Must be allocated with XXH3_createState(). + * + * @see @ref streaming_example "Streaming Example" + */ +XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr) +{ + XXH_alignedFree(statePtr); + return XXH_OK; } -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) { +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API void +XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state) +{ + XXH_memcpy(dst_state, src_state, sizeof(*dst_state)); +} + +static void +XXH3_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->useSeed = (seed != 0); + 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; +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr) +{ + if (statePtr == NULL) return XXH_ERROR; + XXH3_reset_internal(statePtr, 0, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE); + return XXH_OK; +} - if (seed64 == 0) - return XXH3_hashLong_128b_internal( - input, len, XXH3_kSecret, sizeof(XXH3_kSecret), f_acc512, f_scramble); - { +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize) +{ + if (statePtr == NULL) return XXH_ERROR; + XXH3_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_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); +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_64bits_reset_withSeed(XXH_NOESCAPE 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) || (statePtr->extSecret != NULL)) + XXH3_initCustomSecret(statePtr->customSecret, seed); + XXH3_reset_internal(statePtr, seed, NULL, XXH_SECRET_DEFAULT_SIZE); + return XXH_OK; +} - } +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_64bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed64) +{ + if (statePtr == NULL) return XXH_ERROR; + if (secret == NULL) return XXH_ERROR; + if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR; + XXH3_reset_internal(statePtr, seed64, secret, secretSize); + statePtr->useSeed = 1; /* always, even if seed64==0 */ + return XXH_OK; +} +/*! + * @internal + * @brief Processes a large input for XXH3_update() and XXH3_digest_long(). + * + * Unlike XXH3_hashLong_internal_loop(), this can process data that overlaps a block. + * + * @param acc Pointer to the 8 accumulator lanes + * @param nbStripesSoFarPtr In/out pointer to the number of leftover stripes in the block* + * @param nbStripesPerBlock Number of stripes in a block + * @param input Input pointer + * @param nbStripes Number of stripes to process + * @param secret Secret pointer + * @param secretLimit Offset of the last block in @p secret + * @param f_acc Pointer to an XXH3_accumulate implementation + * @param f_scramble Pointer to an XXH3_scrambleAcc implementation + * @return Pointer past the end of @p input after processing + */ +XXH_FORCE_INLINE const xxh_u8 * +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 f_acc, + XXH3_f_scrambleAcc f_scramble) +{ + const xxh_u8* initialSecret = secret + *nbStripesSoFarPtr * XXH_SECRET_CONSUME_RATE; + /* Process full blocks */ + if (nbStripes >= (nbStripesPerBlock - *nbStripesSoFarPtr)) { + /* Process the initial partial block... */ + size_t nbStripesThisIter = nbStripesPerBlock - *nbStripesSoFarPtr; + + do { + /* Accumulate and scramble */ + f_acc(acc, input, initialSecret, nbStripesThisIter); + f_scramble(acc, secret + secretLimit); + input += nbStripesThisIter * XXH_STRIPE_LEN; + nbStripes -= nbStripesThisIter; + /* Then continue the loop with the full block size */ + nbStripesThisIter = nbStripesPerBlock; + initialSecret = secret; + } while (nbStripes >= nbStripesPerBlock); + *nbStripesSoFarPtr = 0; + } + /* Process a partial block */ + if (nbStripes > 0) { + f_acc(acc, input, initialSecret, nbStripes); + input += nbStripes * XXH_STRIPE_LEN; + *nbStripesSoFarPtr += nbStripes; + } + /* Return end pointer */ + return input; } +#ifndef XXH3_STREAM_USE_STACK +# if XXH_SIZE_OPT <= 0 && !defined(__clang__) /* clang doesn't need additional stack space */ +# define XXH3_STREAM_USE_STACK 1 +# endif +#endif /* - * It's important for performance that XXH3_hashLong is not inlined. + * Both XXH3_64bits_update and XXH3_128bits_update use this routine. */ -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) { +XXH_FORCE_INLINE XXH_errorcode +XXH3_update(XXH3_state_t* XXH_RESTRICT const state, + const xxh_u8* XXH_RESTRICT input, size_t len, + XXH3_f_accumulate f_acc, + XXH3_f_scrambleAcc f_scramble) +{ + if (input==NULL) { + XXH_ASSERT(len == 0); + return XXH_OK; + } + + XXH_ASSERT(state != NULL); + { const xxh_u8* const bEnd = input + len; + const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret; +#if defined(XXH3_STREAM_USE_STACK) && XXH3_STREAM_USE_STACK >= 1 + /* For some reason, gcc and MSVC seem to suffer greatly + * when operating accumulators directly into state. + * Operating into stack space seems to enable proper optimization. + * clang, on the other hand, doesn't seem to need this trick */ + XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[8]; + XXH_memcpy(acc, state->acc, sizeof(acc)); +#else + xxh_u64* XXH_RESTRICT const acc = state->acc; +#endif + state->totalLen += len; + XXH_ASSERT(state->bufferedSize <= XXH3_INTERNALBUFFER_SIZE); + + /* small input : just fill in tmp buffer */ + if (len <= XXH3_INTERNALBUFFER_SIZE - state->bufferedSize) { + XXH_memcpy(state->buffer + state->bufferedSize, input, len); + state->bufferedSize += (XXH32_hash_t)len; + return XXH_OK; + } - (void)secret; - (void)secretLen; - return XXH3_hashLong_128b_withSeed_internal( - input, len, seed64, XXH3_accumulate_512, XXH3_scrambleAcc, - XXH3_initCustomSecret); + /* 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(acc, + &state->nbStripesSoFar, state->nbStripesPerBlock, + state->buffer, XXH3_INTERNALBUFFER_STRIPES, + secret, state->secretLimit, + f_acc, f_scramble); + state->bufferedSize = 0; + } + XXH_ASSERT(input < bEnd); + if (bEnd - input > XXH3_INTERNALBUFFER_SIZE) { + size_t nbStripes = (size_t)(bEnd - 1 - input) / XXH_STRIPE_LEN; + input = XXH3_consumeStripes(acc, + &state->nbStripesSoFar, state->nbStripesPerBlock, + input, nbStripes, + secret, state->secretLimit, + f_acc, f_scramble); + XXH_memcpy(state->buffer + sizeof(state->buffer) - XXH_STRIPE_LEN, input - XXH_STRIPE_LEN, XXH_STRIPE_LEN); -typedef XXH128_hash_t (*XXH3_hashLong128_f)(const void *XXH_RESTRICT, size_t, - XXH64_hash_t, - const void *XXH_RESTRICT, size_t); + } + /* Some remaining input (always) : buffer it */ + XXH_ASSERT(input < bEnd); + XXH_ASSERT(bEnd - input <= XXH3_INTERNALBUFFER_SIZE); + XXH_ASSERT(state->bufferedSize == 0); + XXH_memcpy(state->buffer, input, (size_t)(bEnd-input)); + state->bufferedSize = (XXH32_hash_t)(bEnd-input); +#if defined(XXH3_STREAM_USE_STACK) && XXH3_STREAM_USE_STACK >= 1 + /* save stack accumulators into state */ + XXH_memcpy(state->acc, acc, sizeof(acc)); +#endif + } -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); + return XXH_OK; +} +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_64bits_update(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len) +{ + return XXH3_update(state, (const xxh_u8*)input, len, + XXH3_accumulate, XXH3_scrambleAcc); } -/* === Public XXH128 API === */ -/*! @ingroup xxh3_family */ -XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void *input, size_t len) { +XXH_FORCE_INLINE void +XXH3_digest_long (XXH64_hash_t* acc, + const XXH3_state_t* state, + const unsigned char* secret) +{ + xxh_u8 lastStripe[XXH_STRIPE_LEN]; + const xxh_u8* lastStripePtr; - return XXH3_128bits_internal(input, len, 0, XXH3_kSecret, - sizeof(XXH3_kSecret), - XXH3_hashLong_128b_default); + /* + * Digest on a local copy. This way, the state remains unaltered, and it can + * continue ingesting more input afterwards. + */ + XXH_memcpy(acc, state->acc, sizeof(state->acc)); + if (state->bufferedSize >= XXH_STRIPE_LEN) { + /* Consume remaining stripes then point to remaining data in buffer */ + 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, XXH3_scrambleAcc); + lastStripePtr = state->buffer + state->bufferedSize - XXH_STRIPE_LEN; + } else { /* bufferedSize < XXH_STRIPE_LEN */ + /* Copy to temp buffer */ + size_t const catchupSize = XXH_STRIPE_LEN - state->bufferedSize; + XXH_ASSERT(state->bufferedSize > 0); /* there is always some input buffered */ + XXH_memcpy(lastStripe, state->buffer + sizeof(state->buffer) - catchupSize, catchupSize); + XXH_memcpy(lastStripe + catchupSize, state->buffer, state->bufferedSize); + lastStripePtr = lastStripe; + } + /* Last stripe */ + XXH3_accumulate_512(acc, + lastStripePtr, + secret + state->secretLimit - XXH_SECRET_LASTACC_START); +} +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (XXH_NOESCAPE 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->useSeed) + 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); } +#endif /* !XXH_NO_STREAM */ -/*! @ingroup xxh3_family */ -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); +/* ========================================== + * 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 XXH_PUREF 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; + } } -/*! @ingroup xxh3_family */ -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_FORCE_INLINE XXH_PUREF 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 *= PRIME_MX2; + m128.low64 = XXH_xorshift64(m128.low64, 28); + m128.high64 = XXH3_avalanche(m128.high64); + return m128; + } } -/*! @ingroup xxh3_family */ -XXH_PUBLIC_API XXH128_hash_t XXH128(const void *input, size_t len, - XXH64_hash_t seed) { +XXH_FORCE_INLINE XXH_PUREF 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); - return XXH3_128bits_withSeed(input, len, seed); + { /* 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; + } } } -/* === XXH3 128-bit streaming === */ +/* + * Assumption: `secret` size is >= XXH3_SECRET_SIZE_MIN + */ +XXH_FORCE_INLINE XXH_PUREF 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; + } } +} /* - * All the functions are actually the same as for 64-bit streaming variant. - * The only difference is the finalization routine. + * 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 XXH_PUREF 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 XXH_SIZE_OPT >= 1 + { + /* Smaller, but slightly slower. */ + unsigned int i = (unsigned int)(len - 1) / 32; + do { + acc = XXH128_mix32B(acc, input+16*i, input+len-16*(i+1), secret+32*i, seed); + } while (i-- != 0); + } +#else + 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); +#endif + { 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; + } + } +} -/*! @ingroup xxh3_family */ -XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH3_state_t *statePtr) { +XXH_NO_INLINE XXH_PUREF 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); - if (statePtr == NULL) return XXH_ERROR; - XXH3_reset_internal(statePtr, 0, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE); - return XXH_OK; + { XXH128_hash_t acc; + unsigned i; + acc.low64 = len * XXH_PRIME64_1; + acc.high64 = 0; + /* + * We set as `i` as offset + 32. We do this so that unchanged + * `len` can be used as upper bound. This reaches a sweet spot + * where both x86 and aarch64 get simple agen and good codegen + * for the loop. + */ + for (i = 32; i < 160; i += 32) { + acc = XXH128_mix32B(acc, + input + i - 32, + input + i - 16, + secret + i - 32, + seed); + } + acc.low64 = XXH3_avalanche(acc.low64); + acc.high64 = XXH3_avalanche(acc.high64); + /* + * NB: `i <= len` will duplicate the last 32-bytes if + * len % 32 was zero. This is an unfortunate necessity to keep + * the hash result stable. + */ + for (i=160; i <= len; i += 32) { + acc = XXH128_mix32B(acc, + input + i - 32, + input + i - 16, + secret + XXH3_MIDSIZE_STARTOFFSET + i - 160, + seed); + } + /* last bytes */ + acc = XXH128_mix32B(acc, + input + len - 16, + input + len - 32, + secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET - 16, + (XXH64_hash_t)0 - 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 f_acc, + 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_acc, 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; + } } -/*! @ingroup xxh3_family */ -XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret( - XXH3_state_t *statePtr, const void *secret, size_t secretSize) { +/* + * It's important for performance that XXH3_hashLong() is not inlined. + */ +XXH_NO_INLINE XXH_PUREF 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, XXH3_scrambleAcc); +} - if (statePtr == NULL) return XXH_ERROR; - XXH3_reset_internal(statePtr, 0, secret, secretSize); - if (secret == NULL) return XXH_ERROR; - if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR; - return XXH_OK; +/* + * It's important for performance to pass @p secretLen (when it's static) + * to the compiler, so that it can properly optimize the vectorized loop. + * + * When the secret size is unknown, or on GCC 12 where the mix of NO_INLINE and FORCE_INLINE + * breaks -Og, this is XXH_NO_INLINE. + */ +XXH3_WITH_SECRET_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, 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 f_acc, + 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_acc, 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_acc, f_scramble); + } } -/*! @ingroup xxh3_family */ -XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH3_state_t *statePtr, - XXH64_hash_t seed) { +/* + * 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, XXH3_scrambleAcc, XXH3_initCustomSecret); +} - if (statePtr == NULL) return XXH_ERROR; - if (seed == 0) return XXH3_128bits_reset(statePtr); - if (seed != statePtr->seed) - XXH3_initCustomSecret(statePtr->customSecret, seed); - XXH3_reset_internal(statePtr, seed, NULL, XXH_SECRET_DEFAULT_SIZE); - return XXH_OK; +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); } -/*! @ingroup xxh3_family */ -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); +/* === Public XXH128 API === */ +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(XXH_NOESCAPE const void* input, size_t len) +{ + return XXH3_128bits_internal(input, len, 0, + XXH3_kSecret, sizeof(XXH3_kSecret), + XXH3_hashLong_128b_default); } -/*! @ingroup xxh3_family */ -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) { +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH128_hash_t +XXH3_128bits_withSecret(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize) +{ + return XXH3_128bits_internal(input, len, 0, + (const xxh_u8*)secret, secretSize, + XXH3_hashLong_128b_withSecret); +} - 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); - { +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH128_hash_t +XXH3_128bits_withSeed(XXH_NOESCAPE 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); +} - 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; +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH128_hash_t +XXH3_128bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed) +{ + if (len <= XXH3_MIDSIZE_MAX) + return XXH3_128bits_internal(input, len, seed, XXH3_kSecret, sizeof(XXH3_kSecret), NULL); + return XXH3_hashLong_128b_withSecret(input, len, seed, secret, secretSize); +} - } +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH128_hash_t +XXH128(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed) +{ + return XXH3_128bits_withSeed(input, len, seed); +} - } - /* 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); +/* === XXH3 128-bit streaming === */ +#ifndef XXH_NO_STREAM +/* + * All initialization and update functions are identical to 64-bit streaming variant. + * The only difference is the finalization routine. + */ +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr) +{ + return XXH3_64bits_reset(statePtr); +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize) +{ + return XXH3_64bits_reset_withSecret(statePtr, secret, secretSize); +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed) +{ + return XXH3_64bits_reset_withSeed(statePtr, seed); +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_128bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed) +{ + return XXH3_64bits_reset_withSecretandSeed(statePtr, secret, secretSize, seed); +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_128bits_update(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len) +{ + return XXH3_64bits_update(state, input, len); +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (XXH_NOESCAPE 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); } +#endif /* !XXH_NO_STREAM */ +/* 128-bit utility functions */ - /* 128-bit utility functions */ - - #include /* memcmp, memcpy */ +#include /* memcmp, memcpy */ /* return : 1 is equal, 0 if different */ -/*! @ingroup xxh3_family */ -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))); - +/*! @ingroup XXH3_family */ +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 */ -/*! @ingroup xxh3_family */ -XXH_PUBLIC_API int XXH128_cmp(const void *h128_1, const void *h128_2) { + * @return : >0 if *h128_1 > *h128_2 + * <0 if *h128_1 < *h128_2 + * =0 if *h128_1 == *h128_2 */ +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API int XXH128_cmp(XXH_NOESCAPE const void* h128_1, XXH_NOESCAPE 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); +} - 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 ======*/ +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API void +XXH128_canonicalFromHash(XXH_NOESCAPE 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); + } + XXH_memcpy(dst, &hash.high64, sizeof(hash.high64)); + XXH_memcpy((char*)dst + sizeof(hash.high64), &hash.low64, sizeof(hash.low64)); } -/*====== Canonical representation ======*/ -/*! @ingroup xxh3_family */ -XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH128_canonical_t *dst, - XXH128_hash_t hash) { +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH128_hash_t +XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src) +{ + XXH128_hash_t h; + h.high64 = XXH_readBE64(src); + h.low64 = XXH_readBE64(src->digest + 8); + return h; +} - 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); - } +/* ========================================== + * Secret generators + * ========================================== + */ +#define XXH_MIN(x, y) (((x) > (y)) ? (y) : (x)) + +XXH_FORCE_INLINE void XXH3_combine16(void* dst, XXH128_hash_t h128) +{ + XXH_writeLE64( dst, XXH_readLE64(dst) ^ h128.low64 ); + XXH_writeLE64( (char*)dst+8, XXH_readLE64((char*)dst+8) ^ h128.high64 ); +} + +/*! @ingroup XXH3_family */ +XXH_PUBLIC_API XXH_errorcode +XXH3_generateSecret(XXH_NOESCAPE void* secretBuffer, size_t secretSize, XXH_NOESCAPE const void* customSeed, size_t customSeedSize) +{ +#if (XXH_DEBUGLEVEL >= 1) + XXH_ASSERT(secretBuffer != NULL); + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); +#else + /* production mode, assert() are disabled */ + if (secretBuffer == NULL) return XXH_ERROR; + if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR; +#endif - memcpy(dst, &hash.high64, sizeof(hash.high64)); - memcpy((char *)dst + sizeof(hash.high64), &hash.low64, sizeof(hash.low64)); + if (customSeedSize == 0) { + customSeed = XXH3_kSecret; + customSeedSize = XXH_SECRET_DEFAULT_SIZE; + } +#if (XXH_DEBUGLEVEL >= 1) + XXH_ASSERT(customSeed != NULL); +#else + if (customSeed == NULL) return XXH_ERROR; +#endif + /* Fill secretBuffer with a copy of customSeed - repeat as needed */ + { size_t pos = 0; + while (pos < secretSize) { + size_t const toCopy = XXH_MIN((secretSize - pos), customSeedSize); + memcpy((char*)secretBuffer + pos, customSeed, toCopy); + pos += toCopy; + } } + + { size_t const nbSeg16 = secretSize / 16; + size_t n; + XXH128_canonical_t scrambler; + XXH128_canonicalFromHash(&scrambler, XXH128(customSeed, customSeedSize, 0)); + for (n=0; ndigest + 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 +/* Pop our optimization override from above */ +#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \ + && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \ + && defined(__OPTIMIZE__) && XXH_SIZE_OPT <= 0 /* respect -O0 and -Os */ +# pragma GCC pop_options +#endif - #endif /* XXH_NO_LONG_LONG */ +#endif /* XXH_NO_LONG_LONG */ - #endif /* XXH_NO_XXH3 */ +#endif /* XXH_NO_XXH3 */ /*! * @} */ -#endif /* XXH_IMPLEMENTATION */ - -#if defined(__cplusplus) +#endif /* XXH_IMPLEMENTATION */ -} +#if defined (__cplusplus) +} /* extern "C" */ #endif - diff --git a/src/afl-fuzz.c b/src/afl-fuzz.c index ea8f1423..48e32996 100644 --- a/src/afl-fuzz.c +++ b/src/afl-fuzz.c @@ -401,6 +401,10 @@ static void usage(u8 *argv0, int more_help) { SAYF("Compiled with _AFL_DOCUMENT_MUTATIONS.\n"); #endif +#ifdef _AFL_SPECIAL_PERFORMANCE + SAYF("Compiled with special performance options for this specific system, it might not work on other platforms!\n"); +#endif + SAYF("For additional help please consult %s/README.md :)\n\n", doc_path); exit(1); diff --git a/src/afl-performance.c b/src/afl-performance.c index 07c1b527..22cf4dec 100644 --- a/src/afl-performance.c +++ b/src/afl-performance.c @@ -2,9 +2,18 @@ #include "afl-fuzz.h" #include "types.h" +#ifdef _HAVE_AVX2 +#define T1HA0_AESNI_AVAILABLE 1 +#define T1HA_USE_FAST_ONESHOT_READ 1 +#define T1HA_USE_INDIRECT_FUNCTIONS 1 +#define T1HA_IA32AES_NAME XXH3_64bits +#include "t1ha0_ia32aes_b.h" +#else #define XXH_INLINE_ALL #include "xxhash.h" #undef XXH_INLINE_ALL +#endif + void rand_set_seed(afl_state_t *afl, s64 init_seed) { diff --git a/utils/bench/Makefile b/utils/bench/Makefile new file mode 100644 index 00000000..e7d2f3a1 --- /dev/null +++ b/utils/bench/Makefile @@ -0,0 +1,8 @@ +all: hash + +hash: hash.c + gcc -O3 -mavx2 -march=native -I../../include -o hash hash.c + +clean: + rm -f hash + diff --git a/utils/bench/README.md b/utils/bench/README.md new file mode 100644 index 00000000..772c117b --- /dev/null +++ b/utils/bench/README.md @@ -0,0 +1,2 @@ +# Internal AFL++ benchmarking + diff --git a/utils/bench/hash.c b/utils/bench/hash.c new file mode 100644 index 00000000..013a5321 --- /dev/null +++ b/utils/bench/hash.c @@ -0,0 +1,42 @@ +#include +#include +#include + +#define T1HA0_AESNI_AVAILABLE 1 +#define T1HA_USE_FAST_ONESHOT_READ 1 +#define T1HA_USE_INDIRECT_FUNCTIONS 1 +#define T1HA_IA32AES_NAME t1ha0_ia32aes +#include "t1ha0_ia32aes_b.h" + +#define XXH_INLINE_ALL +#include "xxhash.h" +#undef XXH_INLINE_ALL + +int main() { + char *data = malloc(4097); + struct timespec start, end; + long long duration; + int i; + uint64_t res; + + clock_gettime(CLOCK_MONOTONIC, &start); + for (i = 0; i < 100000000; ++i) { + res = XXH3_64bits(data, 4097); + memcpy(data + 16, (char*)&res, 8); + } + clock_gettime(CLOCK_MONOTONIC, &end); + duration = (end.tv_sec - start.tv_sec) * 1000000000LL + (end.tv_nsec - start.tv_nsec); + printf("xxh3 duration: %lld ns\n", duration); + + memset(data, 0, 4097); + clock_gettime(CLOCK_MONOTONIC, &start); + for (i = 0; i < 100000000; ++i) { + res = t1ha0_ia32aes(data, 4097); + memcpy(data + 16, (char*)&res, 8); + } + clock_gettime(CLOCK_MONOTONIC, &end); + duration = (end.tv_sec - start.tv_sec) * 1000000000LL + (end.tv_nsec - start.tv_nsec); + printf("t1ha0_ia32aes duration: %lld ns\n", duration); + + return 0; +} -- cgit 1.4.1 From e01307a993387bfe842df1deb23ec7facffd4859 Mon Sep 17 00:00:00 2001 From: vanhauser-thc Date: Sat, 13 Apr 2024 11:39:19 +0200 Subject: v4.20c --- README.md | 4 ++-- docs/Changelog.md | 2 +- include/config.h | 2 +- 3 files changed, 4 insertions(+), 4 deletions(-) (limited to 'include/config.h') diff --git a/README.md b/README.md index f15089c2..2583407e 100644 --- a/README.md +++ b/README.md @@ -2,9 +2,9 @@ AFL++ logo -Release version: [4.10c](https://github.com/AFLplusplus/AFLplusplus/releases) +Release version: [4.20c](https://github.com/AFLplusplus/AFLplusplus/releases) -GitHub version: 4.20a +GitHub version: 4.20c Repository: [https://github.com/AFLplusplus/AFLplusplus](https://github.com/AFLplusplus/AFLplusplus) diff --git a/docs/Changelog.md b/docs/Changelog.md index 116134ff..2428d63f 100644 --- a/docs/Changelog.md +++ b/docs/Changelog.md @@ -3,7 +3,7 @@ This is the list of all noteworthy changes made in every public release of the tool. See README.md for the general instruction manual. -### Version ++4.20a (dev) +### Version ++4.20c (release) ! A new forkserver communication model is now introduced. afl-fuzz is backward compatible to old compiled targets if they are not built for CMPLOG/Redqueen, but new compiled targets will not work with diff --git a/include/config.h b/include/config.h index 31d66b14..3ea059ff 100644 --- a/include/config.h +++ b/include/config.h @@ -26,7 +26,7 @@ /* Version string: */ // c = release, a = volatile github dev, e = experimental branch -#define VERSION "++4.20a" +#define VERSION "++4.20c" /****************************************************** * * -- cgit 1.4.1