diff options
Diffstat (limited to 'include')
| -rw-r--r-- | include/afl-fuzz.h | 434 | ||||
| -rw-r--r-- | include/afl-prealloc.h | 2 | ||||
| -rw-r--r-- | include/alloc-inl.h | 355 | ||||
| -rw-r--r-- | include/android-ashmem.h | 87 | ||||
| -rw-r--r-- | include/cmplog.h | 27 | ||||
| -rw-r--r-- | include/common.h | 16 | ||||
| -rw-r--r-- | include/config.h | 157 | ||||
| -rw-r--r-- | include/coverage-32.h | 112 | ||||
| -rw-r--r-- | include/coverage-64.h | 189 | ||||
| -rw-r--r-- | include/debug.h | 169 | ||||
| -rw-r--r-- | include/envs.h | 50 | ||||
| -rw-r--r-- | include/forkserver.h | 35 | ||||
| -rw-r--r-- | include/list.h | 1 | ||||
| -rw-r--r-- | include/sharedmem.h | 1 | ||||
| -rw-r--r-- | include/snapshot-inl.h | 2 | ||||
| -rw-r--r-- | include/types.h | 38 | ||||
| -rw-r--r-- | include/xxh3.h | 3187 | ||||
| -rw-r--r-- | include/xxhash.h | 3688 |
18 files changed, 4594 insertions, 3956 deletions
diff --git a/include/afl-fuzz.h b/include/afl-fuzz.h index 1c1be711..5003b563 100644 --- a/include/afl-fuzz.h +++ b/include/afl-fuzz.h @@ -37,10 +37,6 @@ #define _FILE_OFFSET_BITS 64 #endif -#ifdef __ANDROID__ - #include "android-ashmem.h" -#endif - #include "config.h" #include "types.h" #include "debug.h" @@ -65,6 +61,9 @@ #include <dlfcn.h> #include <sched.h> +#include <netdb.h> +#include <netinet/in.h> + #include <sys/wait.h> #include <sys/time.h> #ifndef USEMMAP @@ -76,12 +75,18 @@ #include <sys/mman.h> #include <sys/ioctl.h> #include <sys/file.h> +#include <sys/types.h> #if defined(__APPLE__) || defined(__FreeBSD__) || defined(__OpenBSD__) || \ defined(__NetBSD__) || defined(__DragonFly__) #include <sys/sysctl.h> #endif /* __APPLE__ || __FreeBSD__ || __OpenBSD__ */ +#if defined(__HAIKU__) + #include <kernel/OS.h> + #include <kernel/scheduler.h> +#endif + /* For systems that have sched_setaffinity; right now just Linux, but one can hope... */ @@ -104,6 +109,7 @@ #include <kstat.h> #include <sys/sysinfo.h> #include <sys/pset.h> + #include <strings.h> #endif #endif /* __linux__ */ @@ -121,41 +127,72 @@ #define STAGE_BUF_SIZE (64) /* usable size for stage name buf in afl_state */ +// Little helper to access the ptr to afl->##name_buf - for use in afl_realloc. +#define AFL_BUF_PARAM(name) ((void **)&afl->name##_buf) + +#ifdef WORD_SIZE_64 + #define AFL_RAND_RETURN u64 +#else + #define AFL_RAND_RETURN u32 +#endif + extern s8 interesting_8[INTERESTING_8_LEN]; extern s16 interesting_16[INTERESTING_8_LEN + INTERESTING_16_LEN]; extern s32 interesting_32[INTERESTING_8_LEN + INTERESTING_16_LEN + INTERESTING_32_LEN]; +struct tainted { + + u32 pos; + u32 len; + struct tainted *next; + struct tainted *prev; + +}; + struct queue_entry { u8 *fname; /* File name for the test case */ u32 len; /* Input length */ + u32 id; /* entry number in queue_buf */ - u8 cal_failed, /* Calibration failed? */ - trim_done, /* Trimmed? */ + u8 colorized, /* Do not run redqueen stage again */ + cal_failed; /* Calibration failed? */ + bool trim_done, /* Trimmed? */ was_fuzzed, /* historical, but needed for MOpt */ passed_det, /* Deterministic stages passed? */ has_new_cov, /* Triggers new coverage? */ var_behavior, /* Variable behavior? */ favored, /* Currently favored? */ fs_redundant, /* Marked as redundant in the fs? */ - fully_colorized, /* Do not run redqueen stage again */ - is_ascii; /* Is the input just ascii text? */ + is_ascii, /* Is the input just ascii text? */ + disabled; /* Is disabled from fuzz selection */ u32 bitmap_size, /* Number of bits set in bitmap */ - fuzz_level; /* Number of fuzzing iterations */ + fuzz_level, /* Number of fuzzing iterations */ + n_fuzz_entry; /* offset in n_fuzz */ u64 exec_us, /* Execution time (us) */ handicap, /* Number of queue cycles behind */ - n_fuzz, /* Number of fuzz, does not overflow*/ depth, /* Path depth */ exec_cksum; /* Checksum of the execution trace */ u8 *trace_mini; /* Trace bytes, if kept */ u32 tc_ref; /* Trace bytes ref count */ - struct queue_entry *next, /* Next element, if any */ - *next_100; /* 100 elements ahead */ +#ifdef INTROSPECTION + u32 bitsmap_size; +#endif + + double perf_score, /* performance score */ + weight; + + u8 *testcase_buf; /* The testcase buffer, if loaded. */ + + u8 * cmplog_colorinput; /* the result buf of colorization */ + struct tainted *taint; /* Taint information from CmpLog */ + + struct queue_entry *mother; /* queue entry this based on */ }; @@ -167,6 +204,14 @@ struct extra_data { }; +struct auto_extra_data { + + u8 data[MAX_AUTO_EXTRA]; /* Dictionary token data */ + u32 len; /* Dictionary token length */ + u32 hit_cnt; /* Use count in the corpus */ + +}; + /* Fuzzing stages */ enum { @@ -207,7 +252,7 @@ enum { }; -#define operator_num 18 +#define operator_num 19 #define swarm_num 5 #define period_core 500000 @@ -223,18 +268,19 @@ enum { #define STAGE_OverWrite75 15 #define STAGE_OverWriteExtra 16 #define STAGE_InsertExtra 17 +#define STAGE_Splice 18 #define period_pilot 50000 enum { /* 00 */ EXPLORE, /* AFL default, Exploration-based constant schedule */ - /* 01 */ EXPLOIT, /* AFL's exploitation-based const. */ - /* 02 */ FAST, /* Exponential schedule */ - /* 03 */ COE, /* Cut-Off Exponential schedule */ - /* 04 */ LIN, /* Linear schedule */ - /* 05 */ QUAD, /* Quadratic schedule */ - /* 06 */ RARE, /* Rare edges */ - /* 07 */ MMOPT, /* Modified MOPT schedule */ + /* 01 */ MMOPT, /* Modified MOPT schedule */ + /* 02 */ EXPLOIT, /* AFL's exploitation-based const. */ + /* 03 */ FAST, /* Exponential schedule */ + /* 04 */ COE, /* Cut-Off Exponential schedule */ + /* 05 */ LIN, /* Linear schedule */ + /* 06 */ QUAD, /* Quadratic schedule */ + /* 07 */ RARE, /* Rare edges */ /* 08 */ SEEK, /* EXPLORE that ignores timings */ POWER_SCHEDULES_NUM @@ -272,16 +318,21 @@ enum { enum { /* 00 */ PY_FUNC_INIT, - /* 01 */ PY_FUNC_FUZZ, - /* 02 */ PY_FUNC_POST_PROCESS, - /* 03 */ PY_FUNC_INIT_TRIM, - /* 04 */ PY_FUNC_POST_TRIM, - /* 05 */ PY_FUNC_TRIM, - /* 06 */ PY_FUNC_HAVOC_MUTATION, - /* 07 */ PY_FUNC_HAVOC_MUTATION_PROBABILITY, - /* 08 */ PY_FUNC_QUEUE_GET, - /* 09 */ PY_FUNC_QUEUE_NEW_ENTRY, - /* 10 */ PY_FUNC_DEINIT, + /* 01 */ PY_FUNC_DEINIT, + /* FROM HERE ON BELOW ALL ARE OPTIONAL */ + /* 02 */ PY_OPTIONAL = 2, + /* 02 */ PY_FUNC_FUZZ = 2, + /* 03 */ PY_FUNC_FUZZ_COUNT, + /* 04 */ PY_FUNC_POST_PROCESS, + /* 05 */ PY_FUNC_INIT_TRIM, + /* 06 */ PY_FUNC_POST_TRIM, + /* 07 */ PY_FUNC_TRIM, + /* 08 */ PY_FUNC_HAVOC_MUTATION, + /* 09 */ PY_FUNC_HAVOC_MUTATION_PROBABILITY, + /* 10 */ PY_FUNC_QUEUE_GET, + /* 11 */ PY_FUNC_QUEUE_NEW_ENTRY, + /* 12 */ PY_FUNC_INTROSPECTION, + /* 13 */ PY_FUNC_DESCRIBE, PY_FUNC_COUNT }; @@ -296,8 +347,7 @@ typedef struct py_mutator { u8 * fuzz_buf; size_t fuzz_size; - u8 * post_process_buf; - size_t post_process_size; + Py_buffer post_process_buf; u8 * trim_buf; size_t trim_size; @@ -333,11 +383,14 @@ typedef struct afl_env_vars { u8 afl_skip_cpufreq, afl_exit_when_done, afl_no_affinity, afl_skip_bin_check, afl_dumb_forksrv, afl_import_first, afl_custom_mutator_only, afl_no_ui, afl_force_ui, afl_i_dont_care_about_missing_crashes, afl_bench_just_one, - afl_bench_until_crash, afl_debug_child_output, afl_autoresume, - afl_cal_fast, afl_cycle_schedules, afl_expand_havoc; + afl_bench_until_crash, afl_debug_child, afl_autoresume, afl_cal_fast, + afl_cycle_schedules, afl_expand_havoc, afl_statsd, afl_cmplog_only_new; u8 *afl_tmpdir, *afl_custom_mutator_library, *afl_python_module, *afl_path, - *afl_hang_tmout, *afl_skip_crashes, *afl_preload; + *afl_hang_tmout, *afl_forksrv_init_tmout, *afl_skip_crashes, *afl_preload, + *afl_max_det_extras, *afl_statsd_host, *afl_statsd_port, + *afl_crash_exitcode, *afl_statsd_tags_flavor, *afl_testcache_size, + *afl_testcache_entries, *afl_kill_signal; } afl_env_vars_t; @@ -351,7 +404,7 @@ struct afl_pass_stat { struct foreign_sync { u8 * dir; - time_t ctime; + time_t mtime; }; @@ -372,7 +425,8 @@ typedef struct afl_state { really makes no sense to haul them around as function parameters. */ u64 orig_hit_cnt_puppet, last_limit_time_start, tmp_pilot_time, total_pacemaker_time, total_puppet_find, temp_puppet_find, most_time_key, - most_time, most_execs_key, most_execs, old_hit_count, force_ui_update; + most_time, most_execs_key, most_execs, old_hit_count, force_ui_update, + prev_run_time; MOpt_globals_t mopt_globals_core, mopt_globals_pilot; @@ -424,6 +478,7 @@ typedef struct afl_state { u8 cal_cycles, /* Calibration cycles defaults */ cal_cycles_long, /* Calibration cycles defaults */ + havoc_stack_pow2, /* HAVOC_STACK_POW2 */ no_unlink, /* do not unlink cur_input */ debug, /* Debug mode */ custom_only, /* Custom mutator only mode */ @@ -464,14 +519,22 @@ typedef struct afl_state { disable_trim, /* Never trim in fuzz_one */ shmem_testcase_mode, /* If sharedmem testcases are used */ expand_havoc, /* perform expensive havoc after no find */ - cycle_schedules; /* cycle power schedules ? */ + cycle_schedules, /* cycle power schedules? */ + old_seed_selection; /* use vanilla afl seed selection */ u8 *virgin_bits, /* Regions yet untouched by fuzzing */ *virgin_tmout, /* Bits we haven't seen in tmouts */ *virgin_crash; /* Bits we haven't seen in crashes */ + double *alias_probability; /* alias weighted probabilities */ + u32 * alias_table; /* alias weighted random lookup table */ + u32 active_paths; /* enabled entries in the queue */ + u8 *var_bytes; /* Bytes that appear to be variable */ +#define N_FUZZ_SIZE (1 << 21) + u32 *n_fuzz; + volatile u8 stop_soon, /* Ctrl-C pressed? */ clear_screen; /* Window resized? */ @@ -490,7 +553,8 @@ typedef struct afl_state { useless_at_start, /* Number of useless starting paths */ var_byte_count, /* Bitmap bytes with var behavior */ current_entry, /* Current queue entry ID */ - havoc_div; /* Cycle count divisor for havoc */ + havoc_div, /* Cycle count divisor for havoc */ + max_det_extras; /* deterministic extra count (dicts)*/ u64 total_crashes, /* Total number of crashes */ unique_crashes, /* Crashes with unique signatures */ @@ -506,6 +570,7 @@ typedef struct afl_state { blocks_eff_total, /* Blocks subject to effector maps */ blocks_eff_select, /* Blocks selected as fuzzable */ start_time, /* Unix start time (ms) */ + last_sync_time, /* Time of last sync */ last_path_time, /* Time for most recent path (ms) */ last_crash_time, /* Time for most recent crash (ms) */ last_hang_time; /* Time for most recent hang (ms) */ @@ -519,7 +584,7 @@ typedef struct afl_state { u8 stage_name_buf[STAGE_BUF_SIZE]; /* reused stagename buf with len 64 */ - s32 stage_cur, stage_max; /* Stage progression */ + u32 stage_cur, stage_max; /* Stage progression */ s32 splicing_with; /* Splicing with which test case? */ u32 main_node_id, main_node_max; /* Main instance job splitting */ @@ -536,8 +601,9 @@ typedef struct afl_state { u32 rand_cnt; /* Random number counter */ - u64 rand_seed[4]; - s64 init_seed; + /* unsigned long rand_seed[3]; would also work */ + AFL_RAND_RETURN rand_seed[3]; + s64 init_seed; u64 total_cal_us, /* Total calibration time (us) */ total_cal_cycles; /* Total calibration cycles */ @@ -545,7 +611,8 @@ typedef struct afl_state { u64 total_bitmap_size, /* Total bit count for all bitmaps */ total_bitmap_entries; /* Number of bitmaps counted */ - s32 cpu_core_count; /* CPU core count */ + s32 cpu_core_count, /* CPU core count */ + cpu_to_bind; /* bind to specific CPU */ #ifdef HAVE_AFFINITY s32 cpu_aff; /* Selected CPU core */ @@ -553,20 +620,19 @@ typedef struct afl_state { struct queue_entry *queue, /* Fuzzing queue (linked list) */ *queue_cur, /* Current offset within the queue */ - *queue_top, /* Top of the list */ - *q_prev100; /* Previous 100 marker */ + *queue_top; /* Top of the list */ // growing buf struct queue_entry **queue_buf; - size_t queue_size; struct queue_entry **top_rated; /* Top entries for bitmap bytes */ struct extra_data *extras; /* Extra tokens to fuzz with */ u32 extras_cnt; /* Total number of tokens read */ - struct extra_data *a_extras; /* Automatically selected extras */ - u32 a_extras_cnt; /* Total number of tokens available */ + struct auto_extra_data + a_extras[MAX_AUTO_EXTRAS]; /* Automatically selected extras */ + u32 a_extras_cnt; /* Total number of tokens available */ /* afl_postprocess API - Now supported via custom mutators */ @@ -581,6 +647,10 @@ typedef struct afl_state { /* cmplog forkserver ids */ s32 cmplog_fsrv_ctl_fd, cmplog_fsrv_st_fd; u32 cmplog_prev_timed_out; + u32 cmplog_max_filesize; + u32 cmplog_lvl; + u32 colorize_success; + u8 cmplog_enable_arith, cmplog_enable_transform; struct afl_pass_stat *pass_stats; struct cmp_map * orig_cmp_map; @@ -590,10 +660,10 @@ typedef struct afl_state { unsigned long long int last_avg_exec_update; u32 last_avg_execs; - float last_avg_execs_saved; + double last_avg_execs_saved; /* foreign sync */ -#define FOREIGN_SYNCS_MAX 32 +#define FOREIGN_SYNCS_MAX 32U u8 foreign_sync_cnt; struct foreign_sync foreign_syncs[FOREIGN_SYNCS_MAX]; @@ -602,16 +672,24 @@ typedef struct afl_state { u32 document_counter; #endif - void *maybe_add_auto; - /* statistics file */ double last_bitmap_cvg, last_stability, last_eps; /* plot file saves from last run */ u32 plot_prev_qp, plot_prev_pf, plot_prev_pnf, plot_prev_ce, plot_prev_md; - u64 plot_prev_qc, plot_prev_uc, plot_prev_uh; + u64 plot_prev_qc, plot_prev_uc, plot_prev_uh, plot_prev_ed; u64 stats_last_stats_ms, stats_last_plot_ms, stats_last_ms, stats_last_execs; + + /* StatsD */ + u64 statsd_last_send_ms; + struct sockaddr_in statsd_server; + int statsd_sock; + char * statsd_tags_flavor; + char * statsd_tags_format; + char * statsd_metric_format; + int statsd_metric_format_type; + double stats_avg_exec; u8 *clean_trace; @@ -620,24 +698,23 @@ typedef struct afl_state { /*needed for afl_fuzz_one */ // TODO: see which we can reuse - u8 * out_buf; - size_t out_size; + u8 *out_buf; + + u8 *out_scratch_buf; - u8 * out_scratch_buf; - size_t out_scratch_size; + u8 *eff_buf; - u8 * eff_buf; - size_t eff_size; + u8 *in_buf; - u8 * in_buf; - size_t in_size; + u8 *in_scratch_buf; - u8 * in_scratch_buf; - size_t in_scratch_size; + u8 *ex_buf; - u8 * ex_buf; - size_t ex_size; - u32 custom_mutators_count; + u8 *testcase_buf, *splicecase_buf; + + u32 custom_mutators_count; + + struct custom_mutator *current_custom_fuzz; list_t custom_mutator_list; @@ -645,14 +722,49 @@ typedef struct afl_state { * they do not call another function */ u8 *map_tmp_buf; + /* queue entries ready for splicing count (len > 4) */ + u32 ready_for_splicing_count; + + /* This is the user specified maximum size to use for the testcase cache */ + u64 q_testcase_max_cache_size; + + /* This is the user specified maximum entries in the testcase cache */ + u32 q_testcase_max_cache_entries; + + /* How much of the testcase cache is used so far */ + u64 q_testcase_cache_size; + + /* highest cache count so far */ + u32 q_testcase_max_cache_count; + + /* How many queue entries currently have cached testcases */ + u32 q_testcase_cache_count; + + /* the smallest id currently known free entry */ + u32 q_testcase_smallest_free; + + /* How often did we evict from the cache (for statistics only) */ + u32 q_testcase_evictions; + + /* Refs to each queue entry with cached testcase (for eviction, if cache_count + * is too large) */ + struct queue_entry **q_testcase_cache; + +#ifdef INTROSPECTION + char mutation[8072]; + char m_tmp[4096]; + FILE *introspection_file; + u32 bitsmap_size; +#endif + } afl_state_t; struct custom_mutator { const char *name; + char * name_short; void * dh; u8 * post_process_buf; - size_t post_process_size; u8 stacked_custom_prob, stacked_custom; void *data; /* custom mutator data ptr */ @@ -669,11 +781,38 @@ struct custom_mutator { void *(*afl_custom_init)(afl_state_t *afl, unsigned int seed); /** + * When afl-fuzz was compiled with INTROSPECTION=1 then custom mutators can + * also give introspection information back with this function. + * + * @param data pointer returned in afl_custom_init by this custom mutator + * @return pointer to a text string (const char*) + */ + const char *(*afl_custom_introspection)(void *data); + + /** + * This method is called just before fuzzing a queue entry with the custom + * mutator, and receives the initial buffer. It should return the number of + * fuzzes to perform. + * + * A value of 0 means no fuzzing of this queue entry. + * + * The function is now allowed to change the data. + * + * (Optional) + * + * @param data pointer returned in afl_custom_init by this custom mutator + * @param buf Buffer containing the test case + * @param buf_size Size of the test case + * @return The amount of fuzzes to perform on this queue entry, 0 = skip + */ + u32 (*afl_custom_fuzz_count)(void *data, const u8 *buf, size_t buf_size); + + /** * Perform custom mutations on a given input * * (Optional for now. Required in the future) * - * @param data pointer returned in afl_custom_init for this fuzz case + * @param data pointer returned in afl_custom_init by this custom mutator * @param[in] buf Pointer to the input data to be mutated and the mutated * output * @param[in] buf_size Size of the input/output data @@ -689,13 +828,28 @@ struct custom_mutator { u8 *add_buf, size_t add_buf_size, size_t max_size); /** + * Describe the current testcase, generated by the last mutation. + * This will be called, for example, to give the written testcase a name + * after a crash ocurred. It can help to reproduce crashing mutations. + * + * (Optional) + * + * @param data pointer returned by afl_customm_init for this custom mutator + * @paramp[in] max_description_len maximum size avaliable for the description. + * A longer return string is legal, but will be truncated. + * @return A valid ptr to a 0-terminated string. + * An empty or NULL return will result in a default description + */ + const char *(*afl_custom_describe)(void *data, size_t max_description_len); + + /** * A post-processing function to use right before AFL writes the test case to * disk in order to execute the target. * * (Optional) If this functionality is not needed, simply don't define this * function. * - * @param[in] data pointer returned in afl_custom_init for this fuzz case + * @param[in] data pointer returned in afl_custom_init by this custom mutator * @param[in] buf Buffer containing the test case to be executed * @param[in] buf_size Size of the test case * @param[out] out_buf Pointer to the buffer storing the test case after @@ -722,7 +876,7 @@ struct custom_mutator { * * (Optional) * - * @param data pointer returned in afl_custom_init for this fuzz case + * @param data pointer returned in afl_custom_init by this custom mutator * @param buf Buffer containing the test case * @param buf_size Size of the test case * @return The amount of possible iteration steps to trim the input. @@ -741,7 +895,7 @@ struct custom_mutator { * * (Optional) * - * @param data pointer returned in afl_custom_init for this fuzz case + * @param data pointer returned in afl_custom_init by this custom mutator * @param[out] out_buf Pointer to the buffer containing the trimmed test case. * The library can reuse a buffer for each call * and will have to free the buf (for example in deinit) @@ -756,7 +910,7 @@ struct custom_mutator { * * (Optional) * - * @param data pointer returned in afl_custom_init for this fuzz case + * @param data pointer returned in afl_custom_init by this custom mutator * @param success Indicates if the last trim operation was successful. * @return The next trim iteration index (from 0 to the maximum amount of * steps returned in init_trim). Negative on error. @@ -769,7 +923,7 @@ struct custom_mutator { * * (Optional) * - * @param[in] data pointer returned in afl_custom_init for this fuzz case + * @param[in] data pointer returned in afl_custom_init by this custom mutator * @param[in] buf Pointer to the input data to be mutated and the mutated * output * @param[in] buf_size Size of input data @@ -788,7 +942,7 @@ struct custom_mutator { * * (Optional) * - * @param data pointer returned in afl_custom_init for this fuzz case + * @param data pointer returned in afl_custom_init by this custom mutator * @return The probability (0-100). */ u8 (*afl_custom_havoc_mutation_probability)(void *data); @@ -798,7 +952,7 @@ struct custom_mutator { * * (Optional) * - * @param data pointer returned in afl_custom_init for this fuzz case + * @param data pointer returned in afl_custom_init by this custom mutator * @param filename File name of the test case in the queue entry * @return Return True(1) if the fuzzer will fuzz the queue entry, and * False(0) otherwise. @@ -811,7 +965,7 @@ struct custom_mutator { * * (Optional) * - * @param data pointer returned in afl_custom_init for this fuzz case + * @param data pointer returned in afl_custom_init by this custom mutator * @param filename_new_queue File name of the new queue entry * @param filename_orig_queue File name of the original queue entry. This * argument can be NULL while initializing the fuzzer @@ -821,7 +975,7 @@ struct custom_mutator { /** * Deinitialize the custom mutator. * - * @param data pointer returned in afl_custom_init for this fuzz case + * @param data pointer returned in afl_custom_init by this custom mutator */ void (*afl_custom_deinit)(void *data); @@ -856,15 +1010,17 @@ u8 trim_case_custom(afl_state_t *, struct queue_entry *q, u8 *in_buf, struct custom_mutator *load_custom_mutator_py(afl_state_t *, char *); void finalize_py_module(void *); -size_t post_process_py(void *, u8 *, size_t, u8 **); -s32 init_trim_py(void *, u8 *, size_t); -s32 post_trim_py(void *, u8); -size_t trim_py(void *, u8 **); -size_t havoc_mutation_py(void *, u8 *, size_t, u8 **, size_t); -u8 havoc_mutation_probability_py(void *); -u8 queue_get_py(void *, const u8 *); -void queue_new_entry_py(void *, const u8 *, const u8 *); -void deinit_py(void *); +u32 fuzz_count_py(void *, const u8 *, size_t); +size_t post_process_py(void *, u8 *, size_t, u8 **); +s32 init_trim_py(void *, u8 *, size_t); +s32 post_trim_py(void *, u8); +size_t trim_py(void *, u8 **); +size_t havoc_mutation_py(void *, u8 *, size_t, u8 **, size_t); +u8 havoc_mutation_probability_py(void *); +u8 queue_get_py(void *, const u8 *); +const char *introspection_py(void *); +void queue_new_entry_py(void *, const u8 *, const u8 *); +void deinit_py(void *); #endif @@ -885,37 +1041,50 @@ void write_bitmap(afl_state_t *); u32 count_bits(afl_state_t *, u8 *); u32 count_bytes(afl_state_t *, u8 *); u32 count_non_255_bytes(afl_state_t *, u8 *); -#ifdef WORD_SIZE_64 -void simplify_trace(afl_state_t *, u64 *); +void simplify_trace(afl_state_t *, u8 *); void classify_counts(afl_forkserver_t *); +#ifdef WORD_SIZE_64 +void discover_word(u8 *ret, u64 *current, u64 *virgin); #else -void simplify_trace(afl_state_t *, u32 *); -void classify_counts(afl_forkserver_t *); +void discover_word(u8 *ret, u32 *current, u32 *virgin); #endif void init_count_class16(void); void minimize_bits(afl_state_t *, u8 *, u8 *); #ifndef SIMPLE_FILES -u8 *describe_op(afl_state_t *, u8); +u8 *describe_op(afl_state_t *, u8, size_t); #endif u8 save_if_interesting(afl_state_t *, void *, u32, u8); u8 has_new_bits(afl_state_t *, u8 *); +u8 has_new_bits_unclassified(afl_state_t *, u8 *); /* Extras */ void load_extras_file(afl_state_t *, u8 *, u32 *, u32 *, u32); void load_extras(afl_state_t *, u8 *); -void maybe_add_auto(void *, u8 *, u32); +void dedup_extras(afl_state_t *); +void deunicode_extras(afl_state_t *); +void add_extra(afl_state_t *afl, u8 *mem, u32 len); +void maybe_add_auto(afl_state_t *, u8 *, u32); void save_auto(afl_state_t *); void load_auto(afl_state_t *); void destroy_extras(afl_state_t *); /* Stats */ -void write_stats_file(afl_state_t *, double, double, double); -void maybe_update_plot_file(afl_state_t *, double, double); +void load_stats_file(afl_state_t *); +void write_setup_file(afl_state_t *, u32, char **); +void write_stats_file(afl_state_t *, u32, double, double, double); +void maybe_update_plot_file(afl_state_t *, u32, double, double); void show_stats(afl_state_t *); void show_init_stats(afl_state_t *); +/* StatsD */ + +void statsd_setup_format(afl_state_t *afl); +int statsd_socket_init(afl_state_t *afl); +int statsd_send_metric(afl_state_t *afl); +int statsd_format_metric(afl_state_t *afl, char *buff, size_t bufflen); + /* Run */ fsrv_run_result_t fuzz_run_target(afl_state_t *, afl_forkserver_t *fsrv, u32); @@ -939,7 +1108,7 @@ u8 fuzz_one(afl_state_t *); void bind_to_free_cpu(afl_state_t *); #endif void setup_post(afl_state_t *); -void read_testcases(afl_state_t *); +void read_testcases(afl_state_t *, u8 *); void perform_dry_run(afl_state_t *); void pivot_inputs(afl_state_t *); u32 find_start_position(afl_state_t *); @@ -947,6 +1116,8 @@ void find_timeout(afl_state_t *); double get_runnable_processes(void); void nuke_resume_dir(afl_state_t *); int check_main_node_exists(afl_state_t *); +u32 select_next_queue_entry(afl_state_t *afl); +void create_alias_table(afl_state_t *afl); void setup_dirs_fds(afl_state_t *); void setup_cmdline_file(afl_state_t *, char **); void setup_stdio_file(afl_state_t *); @@ -954,7 +1125,7 @@ void check_crash_handling(void); void check_cpu_governor(afl_state_t *); void get_core_count(afl_state_t *); void fix_up_sync(afl_state_t *); -void check_asan_opts(void); +void check_asan_opts(afl_state_t *); void check_binary(afl_state_t *, u8 *); void fix_up_banner(afl_state_t *, u8 *); void check_if_tty(afl_state_t *); @@ -967,11 +1138,13 @@ void read_foreign_testcases(afl_state_t *, int); u8 common_fuzz_cmplog_stuff(afl_state_t *afl, u8 *out_buf, u32 len); /* RedQueen */ -u8 input_to_state_stage(afl_state_t *afl, u8 *orig_buf, u8 *buf, u32 len, - u64 exec_cksum); +u8 input_to_state_stage(afl_state_t *afl, u8 *orig_buf, u8 *buf, u32 len); + +/* our RNG wrapper */ +AFL_RAND_RETURN rand_next(afl_state_t *afl); -/* xoshiro256** */ -uint64_t rand_next(afl_state_t *afl); +/* probability between 0.0 and 1.0 */ +double rand_next_percent(afl_state_t *afl); /**** Inline routines ****/ @@ -980,6 +1153,8 @@ uint64_t rand_next(afl_state_t *afl); static inline u32 rand_below(afl_state_t *afl, u32 limit) { + if (limit <= 1) return 0; + /* The boundary not being necessarily a power of 2, we need to ensure the result uniformity. */ if (unlikely(!afl->rand_cnt--) && likely(!afl->fixed_seed)) { @@ -991,7 +1166,44 @@ static inline u32 rand_below(afl_state_t *afl, u32 limit) { } - return rand_next(afl) % limit; + /* Modulo is biased - we don't want our fuzzing to be biased so let's do it + right. See: + https://stackoverflow.com/questions/10984974/why-do-people-say-there-is-modulo-bias-when-using-a-random-number-generator + */ + u64 unbiased_rnd; + do { + + unbiased_rnd = rand_next(afl); + + } while (unlikely(unbiased_rnd >= (UINT64_MAX - (UINT64_MAX % limit)))); + + return unbiased_rnd % limit; + +} + +/* we prefer lower range values here */ +/* this is only called with normal havoc, not MOpt, to have an equalizer for + expand havoc mode */ +static inline u32 rand_below_datalen(afl_state_t *afl, u32 limit) { + + if (limit <= 1) return 0; + + switch (rand_below(afl, 3)) { + + case 2: + return (rand_below(afl, limit) % (1 + rand_below(afl, limit - 1))) % + (1 + rand_below(afl, limit - 1)); + break; + case 1: + return rand_below(afl, limit) % (1 + rand_below(afl, limit - 1)); + break; + case 0: + return rand_below(afl, limit); + break; + + } + + return 1; // cannot be reached } @@ -1021,5 +1233,25 @@ static inline u64 next_p2(u64 val) { } +/* Returns the testcase buf from the file behind this queue entry. + Increases the refcount. */ +u8 *queue_testcase_get(afl_state_t *afl, struct queue_entry *q); + +/* If trimming changes the testcase size we have to reload it */ +void queue_testcase_retake(afl_state_t *afl, struct queue_entry *q, + u32 old_len); + +/* If trimming changes the testcase size we have to replace it */ +void queue_testcase_retake_mem(afl_state_t *afl, struct queue_entry *q, u8 *in, + u32 len, u32 old_len); + +/* Add a new queue entry directly to the cache */ + +void queue_testcase_store_mem(afl_state_t *afl, struct queue_entry *q, u8 *mem); + +#if TESTCASE_CACHE == 1 + #error define of TESTCASE_CACHE must be zero or larger than 1 +#endif + #endif diff --git a/include/afl-prealloc.h b/include/afl-prealloc.h index edf69a67..fa6c9b70 100644 --- a/include/afl-prealloc.h +++ b/include/afl-prealloc.h @@ -60,7 +60,7 @@ typedef enum prealloc_status { \ if ((prealloc_counter) >= (prealloc_size)) { \ \ - el_ptr = (void *)malloc(sizeof(*el_ptr)); \ + el_ptr = (element_t *)malloc(sizeof(*el_ptr)); \ if (!el_ptr) { FATAL("error in list.h -> out of memory for element!"); } \ el_ptr->pre_status = PRE_STATUS_MALLOC; \ \ diff --git a/include/alloc-inl.h b/include/alloc-inl.h index 832b2de4..c914da5f 100644 --- a/include/alloc-inl.h +++ b/include/alloc-inl.h @@ -30,12 +30,13 @@ #include <stdio.h> #include <stdlib.h> #include <string.h> +#include <stddef.h> #include "config.h" #include "types.h" #include "debug.h" -/* Initial size used for ck_maybe_grow */ +/* Initial size used for afl_realloc */ #define INITIAL_GROWTH_SIZE (64) // Be careful! _WANT_ORIGINAL_AFL_ALLOC is not compatible with custom mutators @@ -76,10 +77,6 @@ \ } while (0) - /* Allocator increments for ck_realloc_block(). */ - - #define ALLOC_BLK_INC 256 - /* Allocate a buffer, explicitly not zeroing it. Returns NULL for zero-sized requests. */ @@ -97,7 +94,8 @@ static inline void *DFL_ck_alloc_nozero(u32 size) { } -/* Allocate a buffer, returning zeroed memory. */ +/* Allocate a buffer, returning zeroed memory. + Returns null for 0 size */ static inline void *DFL_ck_alloc(u32 size) { @@ -149,15 +147,6 @@ static inline void *DFL_ck_realloc(void *orig, u32 size) { } -/* Re-allocate a buffer with ALLOC_BLK_INC increments (used to speed up - repeated small reallocs without complicating the user code). */ - -static inline void *DFL_ck_realloc_block(void *orig, u32 size) { - - return DFL_ck_realloc(orig, size); - -} - /* Create a buffer with a copy of a string. Returns NULL for NULL inputs. */ static inline u8 *DFL_ck_strdup(u8 *str) { @@ -177,53 +166,13 @@ static inline u8 *DFL_ck_strdup(u8 *str) { } -/* Create a buffer with a copy of a memory block. Returns NULL for zero-sized - or NULL inputs. */ - -static inline void *DFL_ck_memdup(void *mem, u32 size) { - - void *ret; - - if (!mem || !size) { return NULL; } - - ALLOC_CHECK_SIZE(size); - ret = malloc(size); - ALLOC_CHECK_RESULT(ret, size); - - return memcpy(ret, mem, size); - -} - -/* Create a buffer with a block of text, appending a NUL terminator at the end. - Returns NULL for zero-sized or NULL inputs. */ - -static inline u8 *DFL_ck_memdup_str(u8 *mem, u32 size) { - - u8 *ret; - - if (!mem || !size) { return NULL; } - - ALLOC_CHECK_SIZE(size); - ret = (u8 *)malloc(size + 1); - ALLOC_CHECK_RESULT(ret, size); - - memcpy(ret, mem, size); - ret[size] = 0; - - return ret; - -} - /* In non-debug mode, we just do straightforward aliasing of the above functions to user-visible names such as ck_alloc(). */ #define ck_alloc DFL_ck_alloc #define ck_alloc_nozero DFL_ck_alloc_nozero #define ck_realloc DFL_ck_realloc - #define ck_realloc_block DFL_ck_realloc_block #define ck_strdup DFL_ck_strdup - #define ck_memdup DFL_ck_memdup - #define ck_memdup_str DFL_ck_memdup_str #define ck_free DFL_ck_free #define alloc_report() @@ -278,10 +227,6 @@ static inline u8 *DFL_ck_memdup_str(u8 *mem, u32 size) { #define ALLOC_OFF_HEAD 8 #define ALLOC_OFF_TOTAL (ALLOC_OFF_HEAD + 1) - /* Allocator increments for ck_realloc_block(). */ - - #define ALLOC_BLK_INC 256 - /* Sanity-checking macros for pointers. */ #define CHECK_PTR(_p) \ @@ -326,7 +271,7 @@ static inline void *DFL_ck_alloc_nozero(u32 size) { ret = malloc(size + ALLOC_OFF_TOTAL); ALLOC_CHECK_RESULT(ret, size); - ret += ALLOC_OFF_HEAD; + ret = (char *)ret + ALLOC_OFF_HEAD; ALLOC_C1(ret) = ALLOC_MAGIC_C1; ALLOC_S(ret) = size; @@ -366,7 +311,7 @@ static inline void DFL_ck_free(void *mem) { ALLOC_C1(mem) = ALLOC_MAGIC_F; - free(mem - ALLOC_OFF_HEAD); + free((char *)mem - ALLOC_OFF_HEAD); } @@ -395,7 +340,7 @@ static inline void *DFL_ck_realloc(void *orig, u32 size) { #endif /* !DEBUG_BUILD */ old_size = ALLOC_S(orig); - orig -= ALLOC_OFF_HEAD; + orig = (char *)orig - ALLOC_OFF_HEAD; ALLOC_CHECK_SIZE(old_size); @@ -418,10 +363,11 @@ static inline void *DFL_ck_realloc(void *orig, u32 size) { if (orig) { - memcpy(ret + ALLOC_OFF_HEAD, orig + ALLOC_OFF_HEAD, MIN(size, old_size)); - memset(orig + ALLOC_OFF_HEAD, 0xFF, old_size); + memcpy((char *)ret + ALLOC_OFF_HEAD, (char *)orig + ALLOC_OFF_HEAD, + MIN(size, old_size)); + memset((char *)orig + ALLOC_OFF_HEAD, 0xFF, old_size); - ALLOC_C1(orig + ALLOC_OFF_HEAD) = ALLOC_MAGIC_F; + ALLOC_C1((char *)orig + ALLOC_OFF_HEAD) = ALLOC_MAGIC_F; free(orig); @@ -429,41 +375,18 @@ static inline void *DFL_ck_realloc(void *orig, u32 size) { #endif /* ^!DEBUG_BUILD */ - ret += ALLOC_OFF_HEAD; + ret = (char *)ret + ALLOC_OFF_HEAD; ALLOC_C1(ret) = ALLOC_MAGIC_C1; ALLOC_S(ret) = size; ALLOC_C2(ret) = ALLOC_MAGIC_C2; - if (size > old_size) memset(ret + old_size, 0, size - old_size); + if (size > old_size) memset((char *)ret + old_size, 0, size - old_size); return ret; } -/* Re-allocate a buffer with ALLOC_BLK_INC increments (used to speed up - repeated small reallocs without complicating the user code). */ - -static inline void *DFL_ck_realloc_block(void *orig, u32 size) { - - #ifndef DEBUG_BUILD - - if (orig) { - - CHECK_PTR(orig); - - if (ALLOC_S(orig) >= size) return orig; - - size += ALLOC_BLK_INC; - - } - - #endif /* !DEBUG_BUILD */ - - return DFL_ck_realloc(orig, size); - -} - /* Create a buffer with a copy of a string. Returns NULL for NULL inputs. */ static inline u8 *DFL_ck_strdup(u8 *str) { @@ -479,7 +402,7 @@ static inline u8 *DFL_ck_strdup(u8 *str) { ret = malloc(size + ALLOC_OFF_TOTAL); ALLOC_CHECK_RESULT(ret, size); - ret += ALLOC_OFF_HEAD; + ret = (char *)ret + ALLOC_OFF_HEAD; ALLOC_C1(ret) = ALLOC_MAGIC_C1; ALLOC_S(ret) = size; @@ -489,55 +412,6 @@ static inline u8 *DFL_ck_strdup(u8 *str) { } -/* Create a buffer with a copy of a memory block. Returns NULL for zero-sized - or NULL inputs. */ - -static inline void *DFL_ck_memdup(void *mem, u32 size) { - - void *ret; - - if (!mem || !size) return NULL; - - ALLOC_CHECK_SIZE(size); - ret = malloc(size + ALLOC_OFF_TOTAL); - ALLOC_CHECK_RESULT(ret, size); - - ret += ALLOC_OFF_HEAD; - - ALLOC_C1(ret) = ALLOC_MAGIC_C1; - ALLOC_S(ret) = size; - ALLOC_C2(ret) = ALLOC_MAGIC_C2; - - return memcpy(ret, mem, size); - -} - -/* Create a buffer with a block of text, appending a NUL terminator at the end. - Returns NULL for zero-sized or NULL inputs. */ - -static inline u8 *DFL_ck_memdup_str(u8 *mem, u32 size) { - - u8 *ret; - - if (!mem || !size) return NULL; - - ALLOC_CHECK_SIZE(size); - ret = malloc(size + ALLOC_OFF_TOTAL + 1); - ALLOC_CHECK_RESULT(ret, size); - - ret += ALLOC_OFF_HEAD; - - ALLOC_C1(ret) = ALLOC_MAGIC_C1; - ALLOC_S(ret) = size; - ALLOC_C2(ret) = ALLOC_MAGIC_C2; - - memcpy(ret, mem, size); - ret[size] = 0; - - return ret; - -} - #ifndef DEBUG_BUILD /* In non-debug mode, we just do straightforward aliasing of the above @@ -546,10 +420,7 @@ static inline u8 *DFL_ck_memdup_str(u8 *mem, u32 size) { #define ck_alloc DFL_ck_alloc #define ck_alloc_nozero DFL_ck_alloc_nozero #define ck_realloc DFL_ck_realloc - #define ck_realloc_block DFL_ck_realloc_block #define ck_strdup DFL_ck_strdup - #define ck_memdup DFL_ck_memdup - #define ck_memdup_str DFL_ck_memdup_str #define ck_free DFL_ck_free #define alloc_report() @@ -618,8 +489,8 @@ static inline void TRK_alloc_buf(void *ptr, const char *file, const char *func, /* No space available - allocate more. */ - TRK[bucket] = DFL_ck_realloc_block( - TRK[bucket], (TRK_cnt[bucket] + 1) * sizeof(struct TRK_obj)); + TRK[bucket] = DFL_ck_realloc(TRK[bucket], + (TRK_cnt[bucket] + 1) * sizeof(struct TRK_obj)); TRK[bucket][i].ptr = ptr; TRK[bucket][i].file = (char *)file; @@ -694,16 +565,6 @@ static inline void *TRK_ck_realloc(void *orig, u32 size, const char *file, } -static inline void *TRK_ck_realloc_block(void *orig, u32 size, const char *file, - const char *func, u32 line) { - - void *ret = DFL_ck_realloc_block(orig, size); - TRK_free_buf(orig, file, func, line); - TRK_alloc_buf(ret, file, func, line); - return ret; - -} - static inline void *TRK_ck_strdup(u8 *str, const char *file, const char *func, u32 line) { @@ -713,24 +574,6 @@ static inline void *TRK_ck_strdup(u8 *str, const char *file, const char *func, } -static inline void *TRK_ck_memdup(void *mem, u32 size, const char *file, - const char *func, u32 line) { - - void *ret = DFL_ck_memdup(mem, size); - TRK_alloc_buf(ret, file, func, line); - return ret; - -} - -static inline void *TRK_ck_memdup_str(void *mem, u32 size, const char *file, - const char *func, u32 line) { - - void *ret = DFL_ck_memdup_str(mem, size); - TRK_alloc_buf(ret, file, func, line); - return ret; - -} - static inline void TRK_ck_free(void *ptr, const char *file, const char *func, u32 line) { @@ -749,17 +592,8 @@ static inline void TRK_ck_free(void *ptr, const char *file, const char *func, #define ck_realloc(_p1, _p2) \ TRK_ck_realloc(_p1, _p2, __FILE__, __FUNCTION__, __LINE__) - #define ck_realloc_block(_p1, _p2) \ - TRK_ck_realloc_block(_p1, _p2, __FILE__, __FUNCTION__, __LINE__) - #define ck_strdup(_p1) TRK_ck_strdup(_p1, __FILE__, __FUNCTION__, __LINE__) - #define ck_memdup(_p1, _p2) \ - TRK_ck_memdup(_p1, _p2, __FILE__, __FUNCTION__, __LINE__) - - #define ck_memdup_str(_p1, _p2) \ - TRK_ck_memdup_str(_p1, _p2, __FILE__, __FUNCTION__, __LINE__) - #define ck_free(_p1) TRK_ck_free(_p1, __FILE__, __FUNCTION__, __LINE__) #endif /* ^!DEBUG_BUILD */ @@ -771,11 +605,14 @@ static inline void TRK_ck_free(void *ptr, const char *file, const char *func, */ static inline size_t next_pow2(size_t in) { - if (in == 0 || in > (size_t)-1) { - - return 0; /* avoid undefined behaviour under-/overflow */ + // Commented this out as this behavior doesn't change, according to unittests + // if (in == 0 || in > (size_t)-1) { - } + // + // return 0; /* avoid undefined behaviour under-/overflow + // */ + // + // } size_t out = in - 1; out |= out >> 1; @@ -787,6 +624,35 @@ static inline size_t next_pow2(size_t in) { } +/* AFL alloc buffer, the struct is here so we don't need to do fancy ptr + * arithmetics */ +struct afl_alloc_buf { + + /* The complete allocated size, including the header of len + * AFL_ALLOC_SIZE_OFFSET */ + size_t complete_size; + /* ptr to the first element of the actual buffer */ + u8 buf[0]; + +}; + +#define AFL_ALLOC_SIZE_OFFSET (offsetof(struct afl_alloc_buf, buf)) + +/* Returns the container element to this ptr */ +static inline struct afl_alloc_buf *afl_alloc_bufptr(void *buf) { + + return (struct afl_alloc_buf *)((u8 *)buf - AFL_ALLOC_SIZE_OFFSET); + +} + +/* Gets the maximum size of the buf contents (ptr->complete_size - + * AFL_ALLOC_SIZE_OFFSET) */ +static inline size_t afl_alloc_bufsize(void *buf) { + + return afl_alloc_bufptr(buf)->complete_size - AFL_ALLOC_SIZE_OFFSET; + +} + /* This function makes sure *size is > size_needed after call. It will realloc *buf otherwise. *size will grow exponentially as per: @@ -794,71 +660,116 @@ static inline size_t next_pow2(size_t in) { Will return NULL and free *buf if size_needed is <1 or realloc failed. @return For convenience, this function returns *buf. */ -static inline void *maybe_grow(void **buf, size_t *size, size_t size_needed) { +static inline void *afl_realloc(void **buf, size_t size_needed) { + + struct afl_alloc_buf *new_buf = NULL; + + size_t current_size = 0; + size_t next_size = 0; + + if (likely(*buf)) { + + /* the size is always stored at buf - 1*size_t */ + new_buf = (struct afl_alloc_buf *)afl_alloc_bufptr(*buf); + current_size = new_buf->complete_size; + + } + + size_needed += AFL_ALLOC_SIZE_OFFSET; /* No need to realloc */ - if (likely(size_needed && *size >= size_needed)) { return *buf; } + if (likely(current_size >= size_needed)) { return *buf; } /* No initial size was set */ - if (size_needed < INITIAL_GROWTH_SIZE) { size_needed = INITIAL_GROWTH_SIZE; } + if (size_needed < INITIAL_GROWTH_SIZE) { + + next_size = INITIAL_GROWTH_SIZE; - /* grow exponentially */ - size_t next_size = next_pow2(size_needed); + } else { - /* handle overflow and zero size_needed */ - if (!next_size) { next_size = size_needed; } + /* grow exponentially */ + next_size = next_pow2(size_needed); + + /* handle overflow: fall back to the original size_needed */ + if (unlikely(!next_size)) { next_size = size_needed; } + + } /* alloc */ - *buf = realloc(*buf, next_size); - *size = *buf ? next_size : 0; + struct afl_alloc_buf *newer_buf = + (struct afl_alloc_buf *)realloc(new_buf, next_size); + if (unlikely(!newer_buf)) { + + free(new_buf); // avoid a leak + *buf = NULL; + return NULL; + + } else { + + new_buf = newer_buf; + } + + new_buf->complete_size = next_size; + *buf = (void *)(new_buf->buf); return *buf; } -/* This function makes sure *size is > size_needed after call. - It will realloc *buf otherwise. - *size will grow exponentially as per: - https://blog.mozilla.org/nnethercote/2014/11/04/please-grow-your-buffers-exponentially/ - Will FATAL if size_needed is <1. - @return For convenience, this function returns *buf. - */ -static inline void *ck_maybe_grow(void **buf, size_t *size, - size_t size_needed) { +/* afl_realloc_exact uses afl alloc buffers but sets it to a specific size */ - /* Oops. found a bug? */ - if (unlikely(size_needed < 1)) { FATAL("cannot grow to non-positive size"); } +static inline void *afl_realloc_exact(void **buf, size_t size_needed) { - /* No need to realloc */ - if (likely(*size >= size_needed)) { return *buf; } + struct afl_alloc_buf *new_buf = NULL; - /* No initial size was set */ - if (size_needed < INITIAL_GROWTH_SIZE) { size_needed = INITIAL_GROWTH_SIZE; } + size_t current_size = 0; - /* grow exponentially */ - size_t next_size = next_pow2(size_needed); + if (likely(*buf)) { - /* handle overflow */ - if (!next_size) { next_size = size_needed; } + /* the size is always stored at buf - 1*size_t */ + new_buf = (struct afl_alloc_buf *)afl_alloc_bufptr(*buf); + current_size = new_buf->complete_size; + + } + + size_needed += AFL_ALLOC_SIZE_OFFSET; + + /* No need to realloc */ + if (unlikely(current_size == size_needed)) { return *buf; } /* alloc */ - *buf = ck_realloc(*buf, next_size); - *size = next_size; + struct afl_alloc_buf *newer_buf = + (struct afl_alloc_buf *)realloc(new_buf, size_needed); + if (unlikely(!newer_buf)) { + free(new_buf); // avoid a leak + *buf = NULL; + return NULL; + + } else { + + new_buf = newer_buf; + + } + + new_buf->complete_size = size_needed; + *buf = (void *)(new_buf->buf); return *buf; } +static inline void afl_free(void *buf) { + + if (buf) { free(afl_alloc_bufptr(buf)); } + +} + /* Swaps buf1 ptr and buf2 ptr, as well as their sizes */ -static inline void swap_bufs(void **buf1, size_t *size1, void **buf2, - size_t *size2) { +static inline void afl_swap_bufs(void **buf1, void **buf2) { - void * scratch_buf = *buf1; - size_t scratch_size = *size1; + void *scratch_buf = *buf1; *buf1 = *buf2; - *size1 = *size2; *buf2 = scratch_buf; - *size2 = scratch_size; } diff --git a/include/android-ashmem.h b/include/android-ashmem.h index 41d4d2da..91699b27 100644 --- a/include/android-ashmem.h +++ b/include/android-ashmem.h @@ -1,62 +1,34 @@ -/* - american fuzzy lop++ - android shared memory compatibility layer - ---------------------------------------------------------------- - - Originally written by Michal Zalewski - - Now maintained by Marc Heuse <mh@mh-sec.de>, - Heiko Eißfeldt <heiko.eissfeldt@hexco.de>, - Andrea Fioraldi <andreafioraldi@gmail.com>, - Dominik Maier <mail@dmnk.co> - - Copyright 2016, 2017 Google Inc. All rights reserved. - Copyright 2019-2020 AFLplusplus Project. 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 header re-defines the shared memory routines used by AFL++ - using the Andoid API. - - */ - -#ifndef _ANDROID_ASHMEM_H -#define _ANDROID_ASHMEM_H - #ifdef __ANDROID__ - - #include <fcntl.h> - #include <linux/shm.h> - #include <linux/ashmem.h> - #include <sys/ioctl.h> - #include <sys/mman.h> - - #if __ANDROID_API__ >= 26 - #define shmat bionic_shmat - #define shmctl bionic_shmctl - #define shmdt bionic_shmdt - #define shmget bionic_shmget - #endif - - #include <sys/shm.h> - #undef shmat - #undef shmctl - #undef shmdt - #undef shmget - #include <stdio.h> - - #define ASHMEM_DEVICE "/dev/ashmem" - -static inline int shmctl(int __shmid, int __cmd, struct shmid_ds *__buf) { + #ifndef _ANDROID_ASHMEM_H + #define _ANDROID_ASHMEM_H + + #include <fcntl.h> + #include <linux/ashmem.h> + #include <sys/ioctl.h> + #include <sys/mman.h> + + #if __ANDROID_API__ >= 26 + #define shmat bionic_shmat + #define shmctl bionic_shmctl + #define shmdt bionic_shmdt + #define shmget bionic_shmget + #endif + #include <sys/shm.h> + #undef shmat + #undef shmctl + #undef shmdt + #undef shmget + #include <stdio.h> + + #define ASHMEM_DEVICE "/dev/ashmem" + +int shmctl(int __shmid, int __cmd, struct shmid_ds *__buf) { int ret = 0; if (__cmd == IPC_RMID) { int length = ioctl(__shmid, ASHMEM_GET_SIZE, NULL); - struct ashmem_pin pin = {0, (unsigned int)length}; + struct ashmem_pin pin = {0, length}; ret = ioctl(__shmid, ASHMEM_UNPIN, &pin); close(__shmid); @@ -66,7 +38,7 @@ static inline int shmctl(int __shmid, int __cmd, struct shmid_ds *__buf) { } -static inline int shmget(key_t __key, size_t __size, int __shmflg) { +int shmget(key_t __key, size_t __size, int __shmflg) { (void)__shmflg; int fd, ret; @@ -90,7 +62,7 @@ error: } -static inline void *shmat(int __shmid, const void *__shmaddr, int __shmflg) { +void *shmat(int __shmid, const void *__shmaddr, int __shmflg) { (void)__shmflg; int size; @@ -106,7 +78,6 @@ static inline void *shmat(int __shmid, const void *__shmaddr, int __shmflg) { } -#endif /* __ANDROID__ */ - -#endif + #endif /* !_ANDROID_ASHMEM_H */ +#endif /* !__ANDROID__ */ diff --git a/include/cmplog.h b/include/cmplog.h index 74e6a3bb..878ed60c 100644 --- a/include/cmplog.h +++ b/include/cmplog.h @@ -29,26 +29,26 @@ #define _AFL_CMPLOG_H #include "config.h" -#include "forkserver.h" + +#define CMPLOG_LVL_MAX 3 #define CMP_MAP_W 65536 -#define CMP_MAP_H 256 +#define CMP_MAP_H 32 #define CMP_MAP_RTN_H (CMP_MAP_H / 4) #define SHAPE_BYTES(x) (x + 1) -#define CMP_TYPE_INS 0 -#define CMP_TYPE_RTN 1 +#define CMP_TYPE_INS 1 +#define CMP_TYPE_RTN 2 struct cmp_header { - unsigned hits : 20; - - unsigned cnt : 20; - unsigned id : 16; - - unsigned shape : 5; // from 0 to 31 - unsigned type : 1; + unsigned hits : 24; + unsigned id : 24; + unsigned shape : 5; + unsigned type : 2; + unsigned attribute : 4; + unsigned reserved : 5; } __attribute__((packed)); @@ -56,6 +56,8 @@ struct cmp_operands { u64 v0; u64 v1; + u64 v0_128; + u64 v1_128; }; @@ -77,7 +79,8 @@ struct cmp_map { /* Execs the child */ -void cmplog_exec_child(afl_forkserver_t *fsrv, char **argv); +struct afl_forkserver; +void cmplog_exec_child(struct afl_forkserver *fsrv, char **argv); #endif diff --git a/include/common.h b/include/common.h index 87a7425b..b7adbaec 100644 --- a/include/common.h +++ b/include/common.h @@ -31,14 +31,15 @@ #include <string.h> #include <unistd.h> #include <sys/time.h> +#include <stdbool.h> #include "types.h" -#include "stdbool.h" /* STRINGIFY_VAL_SIZE_MAX will fit all stringify_ strings. */ #define STRINGIFY_VAL_SIZE_MAX (16) -void detect_file_args(char **argv, u8 *prog_in, u8 *use_stdin); +void detect_file_args(char **argv, u8 *prog_in, bool *use_stdin); +void print_suggested_envs(char *mispelled_env); void check_environment_vars(char **env); char **argv_cpy_dup(int argc, char **argv); @@ -56,6 +57,11 @@ extern u8 *doc_path; /* path to documentation dir */ u8 *find_binary(u8 *fname); +/* Parses the kill signal environment variable, FATALs on error. + If the env is not set, sets the env to default_signal for the signal handlers + and returns the default_signal. */ +int parse_afl_kill_signal_env(u8 *afl_kill_signal_env, int default_signal); + /* Read a bitmap from file fname to memory This is for the -B option again. */ @@ -110,5 +116,11 @@ u8 *u_stringify_time_diff(u8 *buf, u64 cur_ms, u64 event_ms); /* Reads the map size from ENV */ u32 get_map_size(void); +/* create a stream file */ +FILE *create_ffile(u8 *fn); + +/* create a file */ +s32 create_file(u8 *fn); + #endif diff --git a/include/config.h b/include/config.h index 344a368f..29225f6b 100644 --- a/include/config.h +++ b/include/config.h @@ -10,7 +10,7 @@ Dominik Maier <mail@dmnk.co> Copyright 2016, 2017 Google Inc. All rights reserved. - Copyright 2019-2020 AFLplusplus Project. All rights reserved. + Copyright 2019-2021 AFLplusplus Project. 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. @@ -23,12 +23,10 @@ #ifndef _HAVE_CONFIG_H #define _HAVE_CONFIG_H -#include "types.h" - /* Version string: */ -// c = release, d = volatile github dev, e = experimental branch -#define VERSION "++2.66d" +// c = release, a = volatile github dev, e = experimental branch +#define VERSION "++3.12a" /****************************************************** * * @@ -36,15 +34,76 @@ * * ******************************************************/ +/* Default shared memory map size. Most targets just need a coverage map + between 20-250kb. Plus there is an auto-detection feature in afl-fuzz. + However if a target has problematic constructors and init arrays then + this can fail. Hence afl-fuzz deploys a larger default map. The largest + map seen so far is the xlsx fuzzer for libreoffice which is 5MB. + At runtime this value can be overriden via AFL_MAP_SIZE. + Default: 8MB (defined in bytes) */ +#define DEFAULT_SHMEM_SIZE (8 * 1024 * 1024) + +/* CMPLOG/REDQUEEN TUNING + * + * Here you can modify tuning and solving options for CMPLOG. + * Note that these are run-time options for afl-fuzz, no target + * recompilation required. + * + */ + +/* if TRANSFORM is enabled with '-l T', this additionally enables base64 + encoding/decoding */ +// #define CMPLOG_SOLVE_TRANSFORM_BASE64 + +/* If a redqueen pass finds more than one solution, try to combine them? */ +#define CMPLOG_COMBINE + +/* Minimum % of the corpus to perform cmplog on. Default: 10% */ +#define CMPLOG_CORPUS_PERCENT 5U + +/* Number of potential positions from which we decide if cmplog becomes + useless, default 8096 */ +#define CMPLOG_POSITIONS_MAX (12 * 1024) + +/* Maximum allowed fails per CMP value. Default: 128 */ +#define CMPLOG_FAIL_MAX 96 + +/* Now non-cmplog configuration options */ + +/* console output colors: There are three ways to configure its behavior + * 1. default: colored outputs fixed on: defined USE_COLOR && defined + * ALWAYS_COLORED The env var. AFL_NO_COLOR will have no effect + * 2. defined USE_COLOR && !defined ALWAYS_COLORED + * -> depending on env var AFL_NO_COLOR=1 colors can be switched off + * at run-time. Default is to use colors. + * 3. colored outputs fixed off: !defined USE_COLOR + * The env var. AFL_NO_COLOR will have no effect + */ + /* Comment out to disable terminal colors (note that this makes afl-analyze a lot less nice): */ #define USE_COLOR +#ifdef USE_COLOR + /* Comment in to always enable terminal colors */ + /* Comment out to enable runtime controlled terminal colors via AFL_NO_COLOR + */ + #define ALWAYS_COLORED 1 +#endif + +/* StatsD config + Config can be adjusted via AFL_STATSD_HOST and AFL_STATSD_PORT environment + variable. +*/ +#define STATSD_UPDATE_SEC 1 +#define STATSD_DEFAULT_PORT 8125 +#define STATSD_DEFAULT_HOST "127.0.0.1" + /* If you want to have the original afl internal memory corruption checks. Disabled by default for speed. it is better to use "make ASAN_BUILD=1". */ -//#define _WANT_ORIGINAL_AFL_ALLOC +// #define _WANT_ORIGINAL_AFL_ALLOC /* Comment out to disable fancy ANSI boxes and use poor man's 7-bit UI: */ @@ -55,69 +114,61 @@ /* Default timeout for fuzzed code (milliseconds). This is the upper bound, also used for detecting hangs; the actual value is auto-scaled: */ -#define EXEC_TIMEOUT 1000 +#define EXEC_TIMEOUT 1000U /* Timeout rounding factor when auto-scaling (milliseconds): */ -#define EXEC_TM_ROUND 20 +#define EXEC_TM_ROUND 20U /* 64bit arch MACRO */ #if (defined(__x86_64__) || defined(__arm64__) || defined(__aarch64__)) #define WORD_SIZE_64 1 #endif -/* Default memory limit for child process (MB): */ - -#ifndef __NetBSD__ - #ifndef WORD_SIZE_64 - #define MEM_LIMIT 25 - #else - #define MEM_LIMIT 50 - #endif /* ^!WORD_SIZE_64 */ -#else /* NetBSD's kernel needs more space for stack, see discussion for issue \ - #165 */ - #define MEM_LIMIT 200 -#endif -/* Default memory limit when running in QEMU mode (MB): */ +/* Default memory limit for child process (MB) 0 = disabled : */ + +#define MEM_LIMIT 0U -#define MEM_LIMIT_QEMU 200 +/* Default memory limit when running in QEMU mode (MB) 0 = disabled : */ -/* Default memory limit when running in Unicorn mode (MB): */ +#define MEM_LIMIT_QEMU 0U -#define MEM_LIMIT_UNICORN 200 +/* Default memory limit when running in Unicorn mode (MB) 0 = disabled : */ + +#define MEM_LIMIT_UNICORN 0U /* Number of calibration cycles per every new test case (and for test cases that show variable behavior): */ -#define CAL_CYCLES 8 -#define CAL_CYCLES_LONG 40 +#define CAL_CYCLES 8U +#define CAL_CYCLES_LONG 40U /* Number of subsequent timeouts before abandoning an input file: */ -#define TMOUT_LIMIT 250 +#define TMOUT_LIMIT 250U /* Maximum number of unique hangs or crashes to record: */ -#define KEEP_UNIQUE_HANG 500 -#define KEEP_UNIQUE_CRASH 5000 +#define KEEP_UNIQUE_HANG 500U +#define KEEP_UNIQUE_CRASH 5000U /* Baseline number of random tweaks during a single 'havoc' stage: */ -#define HAVOC_CYCLES 256 -#define HAVOC_CYCLES_INIT 1024 +#define HAVOC_CYCLES 256U +#define HAVOC_CYCLES_INIT 1024U /* Maximum multiplier for the above (should be a power of two, beware of 32-bit int overflows): */ -#define HAVOC_MAX_MULT 16 -#define HAVOC_MAX_MULT_MOPT 32 +#define HAVOC_MAX_MULT 64U +#define HAVOC_MAX_MULT_MOPT 64U /* Absolute minimum number of havoc cycles (after all adjustments): */ -#define HAVOC_MIN 16 +#define HAVOC_MIN 12U /* Power Schedule Divisor */ -#define POWER_BETA 1 +#define POWER_BETA 1U #define MAX_FACTOR (POWER_BETA * 32) /* Maximum stacking for havoc-stage tweaks. The actual value is calculated @@ -126,22 +177,22 @@ n = random between 1 and HAVOC_STACK_POW2 stacking = 2^n - In other words, the default (n = 7) produces 2, 4, 8, 16, 32, 64, or - 128 stacked tweaks: */ + In other words, the default (n = 4) produces 2, 4, 8, 16 + stacked tweaks: */ -#define HAVOC_STACK_POW2 7 +#define HAVOC_STACK_POW2 4U /* Caps on block sizes for cloning and deletion operations. Each of these ranges has a 33% probability of getting picked, except for the first two cycles where smaller blocks are favored: */ -#define HAVOC_BLK_SMALL 32 -#define HAVOC_BLK_MEDIUM 128 -#define HAVOC_BLK_LARGE 1500 +#define HAVOC_BLK_SMALL 32U +#define HAVOC_BLK_MEDIUM 128U +#define HAVOC_BLK_LARGE 1500U /* Extra-large blocks, selected very rarely (<5% of the time): */ -#define HAVOC_BLK_XL 32768 +#define HAVOC_BLK_XL 32768U /* Probabilities of skipping non-favored entries in the queue, expressed as percentages: */ @@ -169,9 +220,11 @@ #define TRIM_START_STEPS 16 #define TRIM_END_STEPS 1024 -/* Maximum size of input file, in bytes (keep under 100MB): */ +/* Maximum size of input file, in bytes (keep under 100MB, default 1MB): + (note that if this value is changed, several areas in afl-cc.c, afl-fuzz.c + and afl-fuzz-state.c have to be changed as well! */ -#define MAX_FILE (1 * 1024 * 1024) +#define MAX_FILE (1 * 1024 * 1024U) /* The same, for the test case minimizer: */ @@ -195,7 +248,7 @@ steps; past this point, the "extras/user" step will be still carried out, but with proportionally lower odds: */ -#define MAX_DET_EXTRAS 200 +#define MAX_DET_EXTRAS 256 /* Maximum number of auto-extracted dictionary tokens to actually use in fuzzing (first value), and to keep in memory as candidates. The latter should be much @@ -236,6 +289,11 @@ #define SYNC_INTERVAL 8 +/* Sync time (minimum time between syncing in ms, time is halfed for -M main + nodes) - default is 30 minutes: */ + +#define SYNC_TIME (30 * 60 * 1000) + /* Output directory reuse grace period (minutes): */ #define OUTPUT_GRACE 25 @@ -295,6 +353,13 @@ #define RESEED_RNG 100000 +/* The default maximum testcase cache size in MB, 0 = disable. + A value between 50 and 250 is a good default value. Note that the + number of entries will be auto assigned if not specified via the + AFL_TESTCACHE_ENTRIES env variable */ + +#define TESTCASE_CACHE_SIZE 50 + /* Maximum line length passed from GCC to 'as' and used for parsing configuration files: */ @@ -356,7 +421,7 @@ after changing this - otherwise, SEGVs may ensue. */ #define MAP_SIZE_POW2 16 -#define MAP_SIZE (1 << MAP_SIZE_POW2) +#define MAP_SIZE (1U << MAP_SIZE_POW2) /* Maximum allocator request size (keep well under INT_MAX): */ diff --git a/include/coverage-32.h b/include/coverage-32.h new file mode 100644 index 00000000..ca36c29f --- /dev/null +++ b/include/coverage-32.h @@ -0,0 +1,112 @@ +#include "config.h" +#include "types.h" + +u32 skim(const u32 *virgin, const u32 *current, const u32 *current_end); +u32 classify_word(u32 word); + +inline u32 classify_word(u32 word) { + + u16 mem16[2]; + memcpy(mem16, &word, sizeof(mem16)); + + mem16[0] = count_class_lookup16[mem16[0]]; + mem16[1] = count_class_lookup16[mem16[1]]; + + memcpy(&word, mem16, sizeof(mem16)); + return word; + +} + +void simplify_trace(afl_state_t *afl, u8 *bytes) { + + u32 *mem = (u32 *)bytes; + u32 i = (afl->fsrv.map_size >> 2); + + while (i--) { + + /* Optimize for sparse bitmaps. */ + + if (unlikely(*mem)) { + + u8 *mem8 = (u8 *)mem; + + mem8[0] = simplify_lookup[mem8[0]]; + mem8[1] = simplify_lookup[mem8[1]]; + mem8[2] = simplify_lookup[mem8[2]]; + mem8[3] = simplify_lookup[mem8[3]]; + + } else + + *mem = 0x01010101; + + mem++; + + } + +} + +inline void classify_counts(afl_forkserver_t *fsrv) { + + u32 *mem = (u32 *)fsrv->trace_bits; + u32 i = (fsrv->map_size >> 2); + + while (i--) { + + /* Optimize for sparse bitmaps. */ + + if (unlikely(*mem)) { *mem = classify_word(*mem); } + + mem++; + + } + +} + +/* Updates the virgin bits, then reflects whether a new count or a new tuple is + * seen in ret. */ +inline void discover_word(u8 *ret, u32 *current, u32 *virgin) { + + /* Optimize for (*current & *virgin) == 0 - i.e., no bits in current bitmap + that have not been already cleared from the virgin map - since this will + almost always be the case. */ + + if (*current & *virgin) { + + if (likely(*ret < 2)) { + + u8 *cur = (u8 *)current; + u8 *vir = (u8 *)virgin; + + /* Looks like we have not found any new bytes yet; see if any non-zero + bytes in current[] are pristine in virgin[]. */ + + if ((cur[0] && vir[0] == 0xff) || (cur[1] && vir[1] == 0xff) || + (cur[2] && vir[2] == 0xff) || (cur[3] && vir[3] == 0xff)) + *ret = 2; + else + *ret = 1; + + } + + *virgin &= ~*current; + + } + +} + +#define PACK_SIZE 16 +inline u32 skim(const u32 *virgin, const u32 *current, const u32 *current_end) { + + for (; current < current_end; virgin += 4, current += 4) { + + if (current[0] && classify_word(current[0]) & virgin[0]) return 1; + if (current[1] && classify_word(current[1]) & virgin[1]) return 1; + if (current[2] && classify_word(current[2]) & virgin[2]) return 1; + if (current[3] && classify_word(current[3]) & virgin[3]) return 1; + + } + + return 0; + +} + diff --git a/include/coverage-64.h b/include/coverage-64.h new file mode 100644 index 00000000..54fe9d33 --- /dev/null +++ b/include/coverage-64.h @@ -0,0 +1,189 @@ +#include "config.h" +#include "types.h" + +#if (defined(__AVX512F__) && defined(__AVX512DQ__)) || defined(__AVX2__) + #include <immintrin.h> +#endif + +u32 skim(const u64 *virgin, const u64 *current, const u64 *current_end); +u64 classify_word(u64 word); + +inline u64 classify_word(u64 word) { + + u16 mem16[4]; + memcpy(mem16, &word, sizeof(mem16)); + + mem16[0] = count_class_lookup16[mem16[0]]; + mem16[1] = count_class_lookup16[mem16[1]]; + mem16[2] = count_class_lookup16[mem16[2]]; + mem16[3] = count_class_lookup16[mem16[3]]; + + memcpy(&word, mem16, sizeof(mem16)); + return word; + +} + +void simplify_trace(afl_state_t *afl, u8 *bytes) { + + u64 *mem = (u64 *)bytes; + u32 i = (afl->fsrv.map_size >> 3); + + while (i--) { + + /* Optimize for sparse bitmaps. */ + + if (unlikely(*mem)) { + + u8 *mem8 = (u8 *)mem; + + mem8[0] = simplify_lookup[mem8[0]]; + mem8[1] = simplify_lookup[mem8[1]]; + mem8[2] = simplify_lookup[mem8[2]]; + mem8[3] = simplify_lookup[mem8[3]]; + mem8[4] = simplify_lookup[mem8[4]]; + mem8[5] = simplify_lookup[mem8[5]]; + mem8[6] = simplify_lookup[mem8[6]]; + mem8[7] = simplify_lookup[mem8[7]]; + + } else + + *mem = 0x0101010101010101ULL; + + mem++; + + } + +} + +inline void classify_counts(afl_forkserver_t *fsrv) { + + u64 *mem = (u64 *)fsrv->trace_bits; + u32 i = (fsrv->map_size >> 3); + + while (i--) { + + /* Optimize for sparse bitmaps. */ + + if (unlikely(*mem)) { *mem = classify_word(*mem); } + + mem++; + + } + +} + +/* Updates the virgin bits, then reflects whether a new count or a new tuple is + * seen in ret. */ +inline void discover_word(u8 *ret, u64 *current, u64 *virgin) { + + /* Optimize for (*current & *virgin) == 0 - i.e., no bits in current bitmap + that have not been already cleared from the virgin map - since this will + almost always be the case. */ + + if (*current & *virgin) { + + if (likely(*ret < 2)) { + + u8 *cur = (u8 *)current; + u8 *vir = (u8 *)virgin; + + /* Looks like we have not found any new bytes yet; see if any non-zero + bytes in current[] are pristine in virgin[]. */ + + if ((cur[0] && vir[0] == 0xff) || (cur[1] && vir[1] == 0xff) || + (cur[2] && vir[2] == 0xff) || (cur[3] && vir[3] == 0xff) || + (cur[4] && vir[4] == 0xff) || (cur[5] && vir[5] == 0xff) || + (cur[6] && vir[6] == 0xff) || (cur[7] && vir[7] == 0xff)) + *ret = 2; + else + *ret = 1; + + } + + *virgin &= ~*current; + + } + +} + +#if defined(__AVX512F__) && defined(__AVX512DQ__) + #define PACK_SIZE 64 +inline u32 skim(const u64 *virgin, const u64 *current, const u64 *current_end) { + + for (; current != current_end; virgin += 8, current += 8) { + + __m512i value = *(__m512i *)current; + __mmask8 mask = _mm512_testn_epi64_mask(value, value); + + /* All bytes are zero. */ + if (mask == 0xff) continue; + + /* Look for nonzero bytes and check for new bits. */ + #define UNROLL(x) \ + if (!(mask & (1 << x)) && classify_word(current[x]) & virgin[x]) return 1 + UNROLL(0); + UNROLL(1); + UNROLL(2); + UNROLL(3); + UNROLL(4); + UNROLL(5); + UNROLL(6); + UNROLL(7); + #undef UNROLL + + } + + return 0; + +} + +#endif + +#if !defined(PACK_SIZE) && defined(__AVX2__) + #define PACK_SIZE 32 +inline u32 skim(const u64 *virgin, const u64 *current, const u64 *current_end) { + + __m256i zeroes = _mm256_setzero_si256(); + + for (; current < current_end; virgin += 4, current += 4) { + + __m256i value = *(__m256i *)current; + __m256i cmp = _mm256_cmpeq_epi64(value, zeroes); + u32 mask = _mm256_movemask_epi8(cmp); + + /* All bytes are zero. */ + if (mask == (u32)-1) continue; + + /* Look for nonzero bytes and check for new bits. */ + if (!(mask & 0xff) && classify_word(current[0]) & virgin[0]) return 1; + if (!(mask & 0xff00) && classify_word(current[1]) & virgin[1]) return 1; + if (!(mask & 0xff0000) && classify_word(current[2]) & virgin[2]) return 1; + if (!(mask & 0xff000000) && classify_word(current[3]) & virgin[3]) return 1; + + } + + return 0; + +} + +#endif + +#if !defined(PACK_SIZE) + #define PACK_SIZE 32 +inline u32 skim(const u64 *virgin, const u64 *current, const u64 *current_end) { + + for (; current < current_end; virgin += 4, current += 4) { + + if (current[0] && classify_word(current[0]) & virgin[0]) return 1; + if (current[1] && classify_word(current[1]) & virgin[1]) return 1; + if (current[2] && classify_word(current[2]) & virgin[2]) return 1; + if (current[3] && classify_word(current[3]) & virgin[3]) return 1; + + } + + return 0; + +} + +#endif + diff --git a/include/debug.h b/include/debug.h index d1bd971b..fc1f39cb 100644 --- a/include/debug.h +++ b/include/debug.h @@ -28,11 +28,6 @@ #include "types.h" #include "config.h" -/* __FUNCTION__ is non-iso */ -#ifdef __func__ - #define __FUNCTION__ __func__ -#endif - /******************* * Terminal colors * *******************/ @@ -173,12 +168,84 @@ * Debug & error macros * ************************/ -/* Just print stuff to the appropriate stream. */ +#if defined USE_COLOR && !defined ALWAYS_COLORED + #include <unistd.h> + #pragma GCC diagnostic ignored "-Wformat-security" +static inline const char *colorfilter(const char *x) { + + static int once = 1; + static int disabled = 0; + + if (once) { + + /* when there is no tty -> we always want filtering + * when AFL_NO_UI is set filtering depends on AFL_NO_COLOR + * otherwise we want always colors + */ + disabled = + isatty(2) && (!getenv("AFL_NO_UI") || + (!getenv("AFL_NO_COLOR") && !getenv("AFL_NO_COLOUR"))); + once = 0; + + } + + if (likely(disabled)) return x; + + static char monochromestring[4096]; + char * d = monochromestring; + int in_seq = 0; + + while (*x) { + + if (in_seq && *x == 'm') { + + in_seq = 0; + + } else { + + if (!in_seq && *x == '\x1b') { in_seq = 1; } + if (!in_seq) { *d++ = *x; } + } + + ++x; + + } + + *d = '\0'; + return monochromestring; + +} + +#else + #define colorfilter(x) x /* no filtering necessary */ +#endif + +/* macro magic to transform the first parameter to SAYF + * through colorfilter which strips coloring */ +#define GET_MACRO(_1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, \ + _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, \ + _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, \ + _39, _40, NAME, ...) \ + NAME + +#define SAYF(...) \ + GET_MACRO(__VA_ARGS__, SAYF_N, SAYF_N, SAYF_N, SAYF_N, SAYF_N, SAYF_N, \ + SAYF_N, SAYF_N, SAYF_N, SAYF_N, SAYF_N, SAYF_N, SAYF_N, SAYF_N, \ + SAYF_N, SAYF_N, SAYF_N, SAYF_N, SAYF_N, SAYF_N, SAYF_N, SAYF_N, \ + SAYF_N, SAYF_N, SAYF_N, SAYF_N, SAYF_N, SAYF_N, SAYF_N, SAYF_N, \ + SAYF_N, SAYF_N, SAYF_N, SAYF_N, SAYF_N, SAYF_N, SAYF_N, SAYF_N, \ + SAYF_N, SAYF_1) \ + (__VA_ARGS__) + +#define SAYF_1(x) MY_SAYF(colorfilter(x)) +#define SAYF_N(x, ...) MY_SAYF(colorfilter(x), __VA_ARGS__) + +/* Just print stuff to the appropriate stream. */ #ifdef MESSAGES_TO_STDOUT - #define SAYF(x...) printf(x) + #define MY_SAYF(x...) printf(x) #else - #define SAYF(x...) fprintf(stderr, x) + #define MY_SAYF(x...) fprintf(stderr, x) #endif /* ^MESSAGES_TO_STDOUT */ /* Show a prefixed warning. */ @@ -223,43 +290,43 @@ /* Die with a verbose non-OS fatal error message. */ -#define FATAL(x...) \ - do { \ - \ - SAYF(bSTOP RESET_G1 CURSOR_SHOW cRST cLRD \ - "\n[-] PROGRAM ABORT : " cRST x); \ - SAYF(cLRD "\n Location : " cRST "%s(), %s:%u\n\n", __FUNCTION__, \ - __FILE__, __LINE__); \ - exit(1); \ - \ +#define FATAL(x...) \ + do { \ + \ + SAYF(bSTOP RESET_G1 CURSOR_SHOW cRST cLRD \ + "\n[-] PROGRAM ABORT : " cRST x); \ + SAYF(cLRD "\n Location : " cRST "%s(), %s:%u\n\n", __func__, \ + __FILE__, (u32)__LINE__); \ + exit(1); \ + \ } while (0) /* Die by calling abort() to provide a core dump. */ -#define ABORT(x...) \ - do { \ - \ - SAYF(bSTOP RESET_G1 CURSOR_SHOW cRST cLRD \ - "\n[-] PROGRAM ABORT : " cRST x); \ - SAYF(cLRD "\n Stop location : " cRST "%s(), %s:%u\n\n", __FUNCTION__, \ - __FILE__, __LINE__); \ - abort(); \ - \ +#define ABORT(x...) \ + do { \ + \ + SAYF(bSTOP RESET_G1 CURSOR_SHOW cRST cLRD \ + "\n[-] PROGRAM ABORT : " cRST x); \ + SAYF(cLRD "\n Stop location : " cRST "%s(), %s:%u\n\n", __func__, \ + __FILE__, (u32)__LINE__); \ + abort(); \ + \ } while (0) /* Die while also including the output of perror(). */ -#define PFATAL(x...) \ - do { \ - \ - fflush(stdout); \ - SAYF(bSTOP RESET_G1 CURSOR_SHOW cRST cLRD \ - "\n[-] SYSTEM ERROR : " cRST x); \ - SAYF(cLRD "\n Stop location : " cRST "%s(), %s:%u\n", __FUNCTION__, \ - __FILE__, __LINE__); \ - SAYF(cLRD " OS message : " cRST "%s\n", strerror(errno)); \ - exit(1); \ - \ +#define PFATAL(x...) \ + do { \ + \ + fflush(stdout); \ + SAYF(bSTOP RESET_G1 CURSOR_SHOW cRST cLRD \ + "\n[-] SYSTEM ERROR : " cRST x); \ + SAYF(cLRD "\n Stop location : " cRST "%s(), %s:%u\n", __func__, \ + __FILE__, (u32)__LINE__); \ + SAYF(cLRD " OS message : " cRST "%s\n", strerror(errno)); \ + exit(1); \ + \ } while (0) /* Die with FATAL() or PFATAL() depending on the value of res (used to @@ -275,22 +342,34 @@ \ } while (0) +/* Show a prefixed debug output. */ + +#define DEBUGF(x...) \ + do { \ + \ + fprintf(stderr, cMGN "[D] " cBRI "DEBUG: " cRST x); \ + fprintf(stderr, cRST ""); \ + \ + } while (0) + /* Error-checking versions of read() and write() that call RPFATAL() as appropriate. */ -#define ck_write(fd, buf, len, fn) \ - do { \ - \ - u32 _len = (len); \ - s32 _res = write(fd, buf, _len); \ - if (_res != _len) RPFATAL(_res, "Short write to %s", fn); \ - \ +#define ck_write(fd, buf, len, fn) \ + do { \ + \ + int _fd = (fd); \ + \ + s32 _len = (s32)(len); \ + s32 _res = write(_fd, (buf), _len); \ + if (_res != _len) RPFATAL(_res, "Short write to %s, fd %d", fn, _fd); \ + \ } while (0) #define ck_read(fd, buf, len, fn) \ do { \ \ - u32 _len = (len); \ + s32 _len = (s32)(len); \ s32 _res = read(fd, buf, _len); \ if (_res != _len) RPFATAL(_res, "Short read from %s", fn); \ \ diff --git a/include/envs.h b/include/envs.h index c1c7d387..cfd73b68 100644 --- a/include/envs.h +++ b/include/envs.h @@ -6,6 +6,7 @@ static char *afl_environment_deprecated[] = { "AFL_LLVM_WHITELIST", "AFL_GCC_WHITELIST", + "AFL_DEBUG_CHILD_OUTPUT", "AFL_DEFER_FORKSRV", "AFL_POST_LIBRARY", "AFL_PERSISTENT", @@ -27,32 +28,45 @@ static char *afl_environment_variables[] = { "AFL_CC", "AFL_CMIN_ALLOW_ANY", "AFL_CMIN_CRASHES_ONLY", + "AFL_CMPLOG_ONLY_NEW", "AFL_CODE_END", "AFL_CODE_START", "AFL_COMPCOV_BINNAME", "AFL_COMPCOV_LEVEL", + "AFL_CRASH_EXITCODE", "AFL_CUSTOM_MUTATOR_LIBRARY", "AFL_CUSTOM_MUTATOR_ONLY", "AFL_CXX", "AFL_CYCLE_SCHEDULES", "AFL_DEBUG", - "AFL_DEBUG_CHILD_OUTPUT", + "AFL_DEBUG_CHILD", "AFL_DEBUG_GDB", "AFL_DISABLE_TRIM", "AFL_DONT_OPTIMIZE", + "AFL_DRIVER_STDERR_DUPLICATE_FILENAME", "AFL_DUMB_FORKSRV", "AFL_ENTRYPOINT", "AFL_EXIT_WHEN_DONE", "AFL_FAST_CAL", "AFL_FORCE_UI", + "AFL_FUZZER_ARGS", // oss-fuzz + "AFL_GDB", + "AFL_GCC_ALLOWLIST", + "AFL_GCC_DENYLIST", + "AFL_GCC_BLOCKLIST", "AFL_GCC_INSTRUMENT_FILE", + "AFL_GCC_OUT_OF_LINE", + "AFL_GCC_SKIP_NEVERZERO", "AFL_GCJ", "AFL_HANG_TMOUT", + "AFL_FORKSRV_INIT_TMOUT", "AFL_HARDEN", "AFL_I_DONT_CARE_ABOUT_MISSING_CRASHES", + "AFL_IGNORE_UNKNOWN_ENVS", "AFL_IMPORT_FIRST", "AFL_INST_LIBS", "AFL_INST_RATIO", + "AFL_KILL_SIGNAL", "AFL_KEEP_TRACES", "AFL_KEEP_ASSEMBLY", "AFL_LD_HARD_FAIL", @@ -62,11 +76,18 @@ static char *afl_environment_variables[] = { "AFL_REAL_LD", "AFL_LD_PRELOAD", "AFL_LD_VERBOSE", + "AFL_LLVM_ALLOWLIST", + "AFL_LLVM_DENYLIST", + "AFL_LLVM_BLOCKLIST", "AFL_LLVM_CMPLOG", "AFL_LLVM_INSTRIM", + "AFL_LLVM_CALLER", "AFL_LLVM_CTX", - "AFL_LLVM_INSTRUMENT", + "AFL_LLVM_CTX_K", + "AFL_LLVM_DICT2FILE", + "AFL_LLVM_DOCUMENT_IDS", "AFL_LLVM_INSTRIM_LOOPHEAD", + "AFL_LLVM_INSTRUMENT", "AFL_LLVM_LTO_AUTODICTIONARY", "AFL_LLVM_AUTODICTIONARY", "AFL_LLVM_SKIPSINGLEBLOCK", @@ -88,30 +109,46 @@ static char *afl_environment_variables[] = { "AFL_LLVM_LTO_STARTID", "AFL_LLVM_LTO_DONTWRITEID", "AFL_NO_ARITH", + "AFL_NO_AUTODICT", "AFL_NO_BUILTIN", +#if defined USE_COLOR && !defined ALWAYS_COLORED + "AFL_NO_COLOR", + "AFL_NO_COLOUR", +#endif "AFL_NO_CPU_RED", "AFL_NO_FORKSRV", "AFL_NO_UI", "AFL_NO_PYTHON", "AFL_UNTRACER_FILE", "AFL_LLVM_USE_TRACE_PC", - "AFL_NO_X86", // not really an env but we dont want to warn on it "AFL_MAP_SIZE", "AFL_MAPSIZE", + "AFL_MAX_DET_EXTRAS", + "AFL_NO_X86", // not really an env but we dont want to warn on it + "AFL_NOOPT", + "AFL_PASSTHROUGH", "AFL_PATH", "AFL_PERFORMANCE_FILE", "AFL_PRELOAD", "AFL_PYTHON_MODULE", + "AFL_QEMU_CUSTOM_BIN", "AFL_QEMU_COMPCOV", "AFL_QEMU_COMPCOV_DEBUG", "AFL_QEMU_DEBUG_MAPS", "AFL_QEMU_DISABLE_CACHE", + "AFL_QEMU_DRIVER_NO_HOOK", + "AFL_QEMU_FORCE_DFL", "AFL_QEMU_PERSISTENT_ADDR", "AFL_QEMU_PERSISTENT_CNT", "AFL_QEMU_PERSISTENT_GPR", "AFL_QEMU_PERSISTENT_HOOK", + "AFL_QEMU_PERSISTENT_MEM", "AFL_QEMU_PERSISTENT_RET", "AFL_QEMU_PERSISTENT_RETADDR_OFFSET", + "AFL_QEMU_PERSISTENT_EXITS", + "AFL_QEMU_INST_RANGES", + "AFL_QEMU_EXCLUDE_RANGES", + "AFL_QEMU_SNAPSHOT", "AFL_QUIET", "AFL_RANDOM_ALLOC_CANARY", "AFL_REAL_PATH", @@ -119,6 +156,12 @@ static char *afl_environment_variables[] = { "AFL_SKIP_BIN_CHECK", "AFL_SKIP_CPUFREQ", "AFL_SKIP_CRASHES", + "AFL_STATSD", + "AFL_STATSD_HOST", + "AFL_STATSD_PORT", + "AFL_STATSD_TAGS_FLAVOR", + "AFL_TESTCACHE_SIZE", + "AFL_TESTCACHE_ENTRIES", "AFL_TMIN_EXACT", "AFL_TMPDIR", "AFL_TOKEN_FILE", @@ -131,6 +174,7 @@ static char *afl_environment_variables[] = { "AFL_WINE_PATH", "AFL_NO_SNAPSHOT", "AFL_EXPAND_HAVOC_NOW", + "AFL_USE_QASAN", NULL }; diff --git a/include/forkserver.h b/include/forkserver.h index 717493db..ac027f81 100644 --- a/include/forkserver.h +++ b/include/forkserver.h @@ -37,9 +37,7 @@ typedef struct afl_forkserver { /* a program that includes afl-forkserver needs to define these */ - u8 uses_asan; /* Target uses ASAN? */ u8 *trace_bits; /* SHM with instrumentation bitmap */ - u8 use_stdin; /* use stdin for sending data */ s32 fsrv_pid, /* PID of the fork server */ child_pid, /* PID of the fuzzed program */ @@ -53,9 +51,8 @@ typedef struct afl_forkserver { fsrv_ctl_fd, /* Fork server control pipe (write) */ fsrv_st_fd; /* Fork server status pipe (read) */ - u8 no_unlink; /* do not unlink cur_input */ - u32 exec_tmout; /* Configurable exec timeout (ms) */ + u32 init_tmout; /* Configurable init timeout (ms) */ u32 map_size; /* map size used by the target */ u32 snapshot; /* is snapshot feature used */ u64 mem_limit; /* Memory cap for child (MB) */ @@ -67,18 +64,29 @@ typedef struct afl_forkserver { FILE *plot_file; /* Gnuplot output file */ - /* Note: lat_run_timed_out is u32 to send it to the child as 4 byte array */ + /* Note: last_run_timed_out is u32 to send it to the child as 4 byte array */ u32 last_run_timed_out; /* Traced process timed out? */ u8 last_kill_signal; /* Signal that killed the child */ - u8 use_shmem_fuzz; /* use shared mem for test cases */ + bool use_shmem_fuzz; /* use shared mem for test cases */ + + bool support_shmem_fuzz; /* set by afl-fuzz */ + + bool use_fauxsrv; /* Fauxsrv for non-forking targets? */ + + bool qemu_mode; /* if running in qemu mode or not */ - u8 support_shmem_fuzz; /* set by afl-fuzz */ + bool use_stdin; /* use stdin for sending data */ - u8 use_fauxsrv; /* Fauxsrv for non-forking targets? */ + bool no_unlink; /* do not unlink cur_input */ - u8 qemu_mode; /* if running in qemu mode or not */ + bool uses_asan; /* Target uses ASAN? */ + + bool debug; /* debug mode? */ + + bool uses_crash_exitcode; /* Custom crash exitcode specified? */ + u8 crash_exitcode; /* The crash exitcode specified */ u32 *shmem_fuzz_len; /* length of the fuzzing test case */ @@ -89,9 +97,11 @@ typedef struct afl_forkserver { /* Function to kick off the forkserver child */ void (*init_child_func)(struct afl_forkserver *fsrv, char **argv); - u8 *function_opt; /* for autodictionary: afl ptr */ + u8 *afl_ptr; /* for autodictionary: afl ptr */ + + void (*add_extra_func)(void *afl_ptr, u8 *mem, u32 len); - void (*function_ptr)(void *afl_tmp, u8 *mem, u32 len); + u8 kill_signal; } afl_forkserver_t; @@ -110,11 +120,14 @@ void afl_fsrv_init(afl_forkserver_t *fsrv); void afl_fsrv_init_dup(afl_forkserver_t *fsrv_to, afl_forkserver_t *from); void afl_fsrv_start(afl_forkserver_t *fsrv, char **argv, volatile u8 *stop_soon_p, u8 debug_child_output); +u32 afl_fsrv_get_mapsize(afl_forkserver_t *fsrv, char **argv, + volatile u8 *stop_soon_p, u8 debug_child_output); void afl_fsrv_write_to_testcase(afl_forkserver_t *fsrv, u8 *buf, size_t len); fsrv_run_result_t afl_fsrv_run_target(afl_forkserver_t *fsrv, u32 timeout, volatile u8 *stop_soon_p); void afl_fsrv_killall(void); void afl_fsrv_deinit(afl_forkserver_t *fsrv); +void afl_fsrv_kill(afl_forkserver_t *fsrv); #ifdef __APPLE__ #define MSG_FORK_ON_APPLE \ diff --git a/include/list.h b/include/list.h index 88cbe062..7ec81cbe 100644 --- a/include/list.h +++ b/include/list.h @@ -81,6 +81,7 @@ static inline void list_append(list_t *list, void *el) { } element_t *el_box = NULL; + PRE_ALLOC(el_box, list->element_prealloc_buf, LIST_PREALLOC_SIZE, list->element_prealloc_count); if (!el_box) { FATAL("failed to allocate list element"); } diff --git a/include/sharedmem.h b/include/sharedmem.h index b15d0535..fdc947f9 100644 --- a/include/sharedmem.h +++ b/include/sharedmem.h @@ -51,6 +51,7 @@ typedef struct sharedmem { size_t map_size; /* actual allocated size */ int cmplog_mode; + int shmemfuzz_mode; struct cmp_map *cmp_map; } sharedmem_t; diff --git a/include/snapshot-inl.h b/include/snapshot-inl.h index a75d69c0..a18187ef 100644 --- a/include/snapshot-inl.h +++ b/include/snapshot-inl.h @@ -66,7 +66,7 @@ struct afl_snapshot_vmrange_args { static int afl_snapshot_dev_fd; -static int afl_snapshot_init() { +static int afl_snapshot_init(void) { afl_snapshot_dev_fd = open(AFL_SNAPSHOT_FILE_NAME, 0); return afl_snapshot_dev_fd; diff --git a/include/types.h b/include/types.h index 39f599a0..7b94fb83 100644 --- a/include/types.h +++ b/include/types.h @@ -25,10 +25,15 @@ #include <stdint.h> #include <stdlib.h> +#include "config.h" typedef uint8_t u8; typedef uint16_t u16; typedef uint32_t u32; +#ifdef WORD_SIZE_64 +typedef unsigned __int128 uint128_t; +typedef uint128_t u128; +#endif /* Extended forkserver option values */ @@ -50,7 +55,7 @@ typedef uint32_t u32; #define FS_OPT_SHDMEM_FUZZ 0x01000000 #define FS_OPT_OLD_AFLPP_WORKAROUND 0x0f000000 // FS_OPT_MAX_MAPSIZE is 8388608 = 0x800000 = 2^23 = 1 << 22 -#define FS_OPT_MAX_MAPSIZE ((0x00fffffe >> 1) + 1) +#define FS_OPT_MAX_MAPSIZE ((0x00fffffeU >> 1) + 1) #define FS_OPT_GET_MAPSIZE(x) (((x & 0x00fffffe) >> 1) + 1) #define FS_OPT_SET_MAPSIZE(x) \ (x <= 1 || x > FS_OPT_MAX_MAPSIZE ? 0 : ((x - 1) << 1)) @@ -61,6 +66,10 @@ typedef int8_t s8; typedef int16_t s16; typedef int32_t s32; typedef int64_t s64; +#ifdef WORD_SIZE_64 +typedef __int128 int128_t; +typedef int128_t s128; +#endif #ifndef MIN #define MIN(a, b) \ @@ -114,6 +123,33 @@ typedef int64_t s64; \ }) +// It is impossible to define 128 bit constants, so ... +#ifdef WORD_SIZE_64 + #define SWAPN(_x, _l) \ + ({ \ + \ + u128 _res = (_x), _ret; \ + char *d = (char *)&_ret, *s = (char *)&_res; \ + int i; \ + for (i = 0; i < 16; i++) \ + d[15 - i] = s[i]; \ + u32 sr = 128U - ((_l) << 3U); \ + (_ret >>= sr); \ + (u128) _ret; \ + \ + }) +#endif + +#define SWAPNN(_x, _y, _l) \ + ({ \ + \ + char *d = (char *)(_x), *s = (char *)(_y); \ + u32 i, l = (_l)-1; \ + for (i = 0; i <= l; i++) \ + d[l - i] = s[i]; \ + \ + }) + #ifdef AFL_LLVM_PASS #if defined(__linux__) || !defined(__ANDROID__) #define AFL_SR(s) (srandom(s)) diff --git a/include/xxh3.h b/include/xxh3.h deleted file mode 100644 index 2354bde9..00000000 --- a/include/xxh3.h +++ /dev/null @@ -1,3187 +0,0 @@ -/* - * xxHash - Extremely Fast Hash algorithm - * Development source file for `xxh3` - * Copyright (C) 2019-2020 Yann Collet - * - * BSD 2-Clause License (https://www.opensource.org/licenses/bsd-license.php) - * - * Redistribution and use in source and binary forms, with or without - * modification, are permitted provided that the following conditions are - * met: - * - * * Redistributions of source code must retain the above copyright - * notice, this list of conditions and the following disclaimer. - * * Redistributions in binary form must reproduce the above - * copyright notice, this list of conditions and the following disclaimer - * in the documentation and/or other materials provided with the - * distribution. - * - * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS - * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT - * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR - * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT - * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, - * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT - * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, - * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY - * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT - * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE - * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. - * - * You can contact the author at: - * - xxHash homepage: https://www.xxhash.com - * - xxHash source repository: https://github.com/Cyan4973/xxHash - */ - -/* - * Note: This file is separated for development purposes. - * It will be integrated into `xxhash.h` when development stage is completed. - * - * Credit: most of the work on vectorial and asm variants comes from - * @easyaspi314 - */ - -#ifndef XXH3_H_1397135465 -#define XXH3_H_1397135465 - -/* === Dependencies === */ -#ifndef XXHASH_H_5627135585666179 - /* special: when including `xxh3.h` directly, turn on XXH_INLINE_ALL */ - #undef XXH_INLINE_ALL /* avoid redefinition */ - #define XXH_INLINE_ALL -#endif -#include "xxhash.h" - -/* === Compiler specifics === */ - -#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L /* >= C99 */ - #define XXH_RESTRICT restrict -#else - /* Note: it might be useful to define __restrict or __restrict__ for some C++ - * compilers */ - #define XXH_RESTRICT /* disable */ -#endif - -#if (defined(__GNUC__) && (__GNUC__ >= 3)) || \ - (defined(__INTEL_COMPILER) && (__INTEL_COMPILER >= 800)) || \ - defined(__clang__) - #define XXH_likely(x) __builtin_expect(x, 1) - #define XXH_unlikely(x) __builtin_expect(x, 0) -#else - #define XXH_likely(x) (x) - #define XXH_unlikely(x) (x) -#endif - -#if defined(__GNUC__) - #if defined(__AVX2__) - #include <immintrin.h> - #elif defined(__SSE2__) - #include <emmintrin.h> - #elif defined(__ARM_NEON__) || defined(__ARM_NEON) - #define inline __inline__ /* clang bug */ - #include <arm_neon.h> - #undef inline - #endif -#elif defined(_MSC_VER) - #include <intrin.h> -#endif - -/* - * One goal of XXH3 is to make it fast on both 32-bit and 64-bit, while - * remaining a true 64-bit/128-bit hash function. - * - * This is done by prioritizing a subset of 64-bit operations that can be - * emulated without too many steps on the average 32-bit machine. - * - * For example, these two lines seem similar, and run equally fast on 64-bit: - * - * xxh_u64 x; - * x ^= (x >> 47); // good - * x ^= (x >> 13); // bad - * - * However, to a 32-bit machine, there is a major difference. - * - * x ^= (x >> 47) looks like this: - * - * x.lo ^= (x.hi >> (47 - 32)); - * - * while x ^= (x >> 13) looks like this: - * - * // note: funnel shifts are not usually cheap. - * x.lo ^= (x.lo >> 13) | (x.hi << (32 - 13)); - * x.hi ^= (x.hi >> 13); - * - * The first one is significantly faster than the second, simply because the - * shift is larger than 32. This means: - * - All the bits we need are in the upper 32 bits, so we can ignore the lower - * 32 bits in the shift. - * - The shift result will always fit in the lower 32 bits, and therefore, - * we can ignore the upper 32 bits in the xor. - * - * Thanks to this optimization, XXH3 only requires these features to be - * efficient: - * - * - Usable unaligned access - * - A 32-bit or 64-bit ALU - * - If 32-bit, a decent ADC instruction - * - A 32 or 64-bit multiply with a 64-bit result - * - For the 128-bit variant, a decent byteswap helps short inputs. - * - * The first two are already required by XXH32, and almost all 32-bit and 64-bit - * platforms which can run XXH32 can run XXH3 efficiently. - * - * Thumb-1, the classic 16-bit only subset of ARM's instruction set, is one - * notable exception. - * - * First of all, Thumb-1 lacks support for the UMULL instruction which - * performs the important long multiply. This means numerous __aeabi_lmul - * calls. - * - * Second of all, the 8 functional registers are just not enough. - * Setup for __aeabi_lmul, byteshift loads, pointers, and all arithmetic need - * Lo registers, and this shuffling results in thousands more MOVs than A32. - * - * A32 and T32 don't have this limitation. They can access all 14 registers, - * do a 32->64 multiply with UMULL, and the flexible operand allowing free - * shifts is helpful, too. - * - * Therefore, we do a quick sanity check. - * - * If compiling Thumb-1 for a target which supports ARM instructions, we will - * emit a warning, as it is not a "sane" platform to compile for. - * - * Usually, if this happens, it is because of an accident and you probably need - * to specify -march, as you likely meant to compile for a newer architecture. - */ -#if defined(__thumb__) && !defined(__thumb2__) && defined(__ARM_ARCH_ISA_ARM) - #warning "XXH3 is highly inefficient without ARM or Thumb-2." -#endif - -/* ========================================== - * Vectorization detection - * ========================================== */ -#define XXH_SCALAR 0 /* Portable scalar version */ -#define XXH_SSE2 1 /* SSE2 for Pentium 4 and all x86_64 */ -#define XXH_AVX2 2 /* AVX2 for Haswell and Bulldozer */ -#define XXH_AVX512 3 /* AVX512 for Skylake and Icelake */ -#define XXH_NEON 4 /* NEON for most ARMv7-A and all AArch64 */ -#define XXH_VSX 5 /* VSX and ZVector for POWER8/z13 */ - -#ifndef XXH_VECTOR /* can be defined on command line */ - #if defined(__AVX512F__) - #define XXH_VECTOR XXH_AVX512 - #elif defined(__AVX2__) - #define XXH_VECTOR XXH_AVX2 - #elif defined(__SSE2__) || defined(_M_AMD64) || defined(_M_X64) || \ - (defined(_M_IX86_FP) && (_M_IX86_FP == 2)) - #define XXH_VECTOR XXH_SSE2 - #elif defined(__GNUC__) /* msvc support maybe later */ \ - && (defined(__ARM_NEON__) || defined(__ARM_NEON)) && \ - (defined(__LITTLE_ENDIAN__) /* We only support little endian NEON */ \ - || \ - (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)) - #define XXH_VECTOR XXH_NEON - #elif (defined(__PPC64__) && defined(__POWER8_VECTOR__)) || \ - (defined(__s390x__) && defined(__VEC__)) && \ - defined(__GNUC__) /* TODO: IBM XL */ - #define XXH_VECTOR XXH_VSX - #else - #define XXH_VECTOR XXH_SCALAR - #endif -#endif - -/* - * Controls the alignment of the accumulator, - * for compatibility with aligned vector loads, which are usually faster. - */ -#ifndef XXH_ACC_ALIGN - #if defined(XXH_X86DISPATCH) - #define XXH_ACC_ALIGN 64 /* for compatibility with avx512 */ - #elif XXH_VECTOR == XXH_SCALAR /* scalar */ - #define XXH_ACC_ALIGN 8 - #elif XXH_VECTOR == XXH_SSE2 /* sse2 */ - #define XXH_ACC_ALIGN 16 - #elif XXH_VECTOR == XXH_AVX2 /* avx2 */ - #define XXH_ACC_ALIGN 32 - #elif XXH_VECTOR == XXH_NEON /* neon */ - #define XXH_ACC_ALIGN 16 - #elif XXH_VECTOR == XXH_VSX /* vsx */ - #define XXH_ACC_ALIGN 16 - #elif XXH_VECTOR == XXH_AVX512 /* avx512 */ - #define XXH_ACC_ALIGN 64 - #endif -#endif - -#if defined(XXH_X86DISPATCH) || XXH_VECTOR == XXH_SSE2 || \ - XXH_VECTOR == XXH_AVX2 || XXH_VECTOR == XXH_AVX512 - #define XXH_SEC_ALIGN XXH_ACC_ALIGN -#else - #define XXH_SEC_ALIGN 8 -#endif - -/* - * UGLY HACK: - * GCC usually generates the best code with -O3 for xxHash. - * - * However, when targeting AVX2, it is overzealous in its unrolling resulting - * in code roughly 3/4 the speed of Clang. - * - * There are other issues, such as GCC splitting _mm256_loadu_si256 into - * _mm_loadu_si128 + _mm256_inserti128_si256. This is an optimization which - * only applies to Sandy and Ivy Bridge... which don't even support AVX2. - * - * That is why when compiling the AVX2 version, it is recommended to use either - * -O2 -mavx2 -march=haswell - * or - * -O2 -mavx2 -mno-avx256-split-unaligned-load - * for decent performance, or to use Clang instead. - * - * Fortunately, we can control the first one with a pragma that forces GCC into - * -O2, but the other one we can't control without "failed to inline always - * inline function due to target mismatch" warnings. - */ -#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \ - && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \ - && defined(__OPTIMIZE__) && \ - !defined(__OPTIMIZE_SIZE__) /* respect -O0 and -Os */ - #pragma GCC push_options - #pragma GCC optimize("-O2") -#endif - -#if XXH_VECTOR == XXH_NEON - /* - * NEON's setup for vmlal_u32 is a little more complicated than it is on - * SSE2, AVX2, and VSX. - * - * While PMULUDQ and VMULEUW both perform a mask, VMLAL.U32 performs an - * upcast. - * - * To do the same operation, the 128-bit 'Q' register needs to be split into - * two 64-bit 'D' registers, performing this operation:: - * - * [ a | b ] - * | '---------. .--------' | - * | x | - * | .---------' '--------. | - * [ a & 0xFFFFFFFF | b & 0xFFFFFFFF ],[ a >> 32 | b >> 32 ] - * - * Due to significant changes in aarch64, the fastest method for aarch64 is - * completely different than the fastest method for ARMv7-A. - * - * ARMv7-A treats D registers as unions overlaying Q registers, so modifying - * D11 will modify the high half of Q5. This is similar to how modifying AH - * will only affect bits 8-15 of AX on x86. - * - * VZIP takes two registers, and puts even lanes in one register and odd lanes - * in the other. - * - * On ARMv7-A, this strangely modifies both parameters in place instead of - * taking the usual 3-operand form. - * - * Therefore, if we want to do this, we can simply use a D-form VZIP.32 on the - * lower and upper halves of the Q register to end up with the high and low - * halves where we want - all in one instruction. - * - * vzip.32 d10, d11 @ d10 = { d10[0], d11[0] }; d11 = { d10[1], - * d11[1] } - * - * Unfortunately we need inline assembly for this: Instructions modifying two - * registers at once is not possible in GCC or Clang's IR, and they have to - * create a copy. - * - * aarch64 requires a different approach. - * - * In order to make it easier to write a decent compiler for aarch64, many - * quirks were removed, such as conditional execution. - * - * NEON was also affected by this. - * - * aarch64 cannot access the high bits of a Q-form register, and writes to a - * D-form register zero the high bits, similar to how writes to W-form scalar - * registers (or DWORD registers on x86_64) work. - * - * The formerly free vget_high intrinsics now require a vext (with a few - * exceptions) - * - * Additionally, VZIP was replaced by ZIP1 and ZIP2, which are the equivalent - * of PUNPCKL* and PUNPCKH* in SSE, respectively, in order to only modify one - * operand. - * - * The equivalent of the VZIP.32 on the lower and upper halves would be this - * mess: - * - * ext v2.4s, v0.4s, v0.4s, #2 // v2 = { v0[2], v0[3], v0[0], v0[1] } - * zip1 v1.2s, v0.2s, v2.2s // v1 = { v0[0], v2[0] } - * zip2 v0.2s, v0.2s, v1.2s // v0 = { v0[1], v2[1] } - * - * Instead, we use a literal downcast, vmovn_u64 (XTN), and vshrn_n_u64 - * (SHRN): - * - * shrn v1.2s, v0.2d, #32 // v1 = (uint32x2_t)(v0 >> 32); - * xtn v0.2s, v0.2d // v0 = (uint32x2_t)(v0 & 0xFFFFFFFF); - * - * This is available on ARMv7-A, but is less efficient than a single VZIP.32. - */ - - /* - * Function-like macro: - * void XXH_SPLIT_IN_PLACE(uint64x2_t &in, uint32x2_t &outLo, uint32x2_t - * &outHi) - * { - - * outLo = (uint32x2_t)(in & 0xFFFFFFFF); - * outHi = (uint32x2_t)(in >> 32); - * in = UNDEFINED; - * } - */ - #if !defined(XXH_NO_VZIP_HACK) /* define to disable */ \ - && defined(__GNUC__) && !defined(__aarch64__) && !defined(__arm64__) - #define XXH_SPLIT_IN_PLACE(in, outLo, outHi) \ - do { \ - \ - /* Undocumented GCC/Clang operand modifier: %e0 = lower D half, %f0 = \ - * upper D half */ \ - /* https://github.com/gcc-mirror/gcc/blob/38cf91e5/gcc/config/arm/arm.c#L22486 \ - */ \ - /* https://github.com/llvm-mirror/llvm/blob/2c4ca683/lib/Target/ARM/ARMAsmPrinter.cpp#L399 \ - */ \ - __asm__("vzip.32 %e0, %f0" : "+w"(in)); \ - (outLo) = vget_low_u32(vreinterpretq_u32_u64(in)); \ - (outHi) = vget_high_u32(vreinterpretq_u32_u64(in)); \ - \ - } while (0) - - #else - #define XXH_SPLIT_IN_PLACE(in, outLo, outHi) \ - do { \ - \ - (outLo) = vmovn_u64(in); \ - (outHi) = vshrn_n_u64((in), 32); \ - \ - } while (0) - - #endif -#endif /* XXH_VECTOR == XXH_NEON */ - -/* - * VSX and Z Vector helpers. - * - * This is very messy, and any pull requests to clean this up are welcome. - * - * There are a lot of problems with supporting VSX and s390x, due to - * inconsistent intrinsics, spotty coverage, and multiple endiannesses. - */ -#if XXH_VECTOR == XXH_VSX - #if defined(__s390x__) - #include <s390intrin.h> - #else - #include <altivec.h> - #endif - - #undef vector /* Undo the pollution */ - -typedef __vector unsigned long long xxh_u64x2; -typedef __vector unsigned char xxh_u8x16; -typedef __vector unsigned xxh_u32x4; - - #ifndef XXH_VSX_BE - #if defined(__BIG_ENDIAN__) || \ - (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) - #define XXH_VSX_BE 1 - #elif defined(__VEC_ELEMENT_REG_ORDER__) && \ - __VEC_ELEMENT_REG_ORDER__ == __ORDER_BIG_ENDIAN__ - #warning "-maltivec=be is not recommended. Please use native endianness." - #define XXH_VSX_BE 1 - #else - #define XXH_VSX_BE 0 - #endif - #endif /* !defined(XXH_VSX_BE) */ - - #if XXH_VSX_BE - /* A wrapper for POWER9's vec_revb. */ - #if defined(__POWER9_VECTOR__) || (defined(__clang__) && defined(__s390x__)) - #define XXH_vec_revb vec_revb - #else -XXH_FORCE_INLINE xxh_u64x2 XXH_vec_revb(xxh_u64x2 val) { - - xxh_u8x16 const vByteSwap = {0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00, - 0x0F, 0x0E, 0x0D, 0x0C, 0x0B, 0x0A, 0x09, 0x08}; - return vec_perm(val, val, vByteSwap); - -} - - #endif - #endif /* XXH_VSX_BE */ - -/* - * Performs an unaligned load and byte swaps it on big endian. - */ -XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(const void *ptr) { - - xxh_u64x2 ret; - memcpy(&ret, ptr, sizeof(xxh_u64x2)); - #if XXH_VSX_BE - ret = XXH_vec_revb(ret); - #endif - return ret; - -} - - /* - * vec_mulo and vec_mule are very problematic intrinsics on PowerPC - * - * These intrinsics weren't added until GCC 8, despite existing for a while, - * and they are endian dependent. Also, their meaning swap depending on - * version. - * */ - #if defined(__s390x__) - /* s390x is always big endian, no issue on this platform */ - #define XXH_vec_mulo vec_mulo - #define XXH_vec_mule vec_mule - #elif defined(__clang__) && XXH_HAS_BUILTIN(__builtin_altivec_vmuleuw) - /* Clang has a better way to control this, we can just use the builtin which - * doesn't swap. */ - #define XXH_vec_mulo __builtin_altivec_vmulouw - #define XXH_vec_mule __builtin_altivec_vmuleuw - #else -/* gcc needs inline assembly */ -/* Adapted from - * https://github.com/google/highwayhash/blob/master/highwayhash/hh_vsx.h. */ -XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mulo(xxh_u32x4 a, xxh_u32x4 b) { - - xxh_u64x2 result; - __asm__("vmulouw %0, %1, %2" : "=v"(result) : "v"(a), "v"(b)); - return result; - -} - -XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mule(xxh_u32x4 a, xxh_u32x4 b) { - - xxh_u64x2 result; - __asm__("vmuleuw %0, %1, %2" : "=v"(result) : "v"(a), "v"(b)); - return result; - -} - - #endif /* XXH_vec_mulo, XXH_vec_mule */ -#endif /* XXH_VECTOR == XXH_VSX */ - -/* prefetch - * can be disabled, by declaring XXH_NO_PREFETCH build macro */ -#if defined(XXH_NO_PREFETCH) - #define XXH_PREFETCH(ptr) (void)(ptr) /* disabled */ -#else - #if defined(_MSC_VER) && \ - (defined(_M_X64) || \ - defined(_M_I86)) /* _mm_prefetch() is not defined outside of x86/x64 */ - #include <mmintrin.h> /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */ - #define XXH_PREFETCH(ptr) _mm_prefetch((const char *)(ptr), _MM_HINT_T0) - #elif defined(__GNUC__) && \ - ((__GNUC__ >= 4) || ((__GNUC__ == 3) && (__GNUC_MINOR__ >= 1))) - #define XXH_PREFETCH(ptr) \ - __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */) - #else - #define XXH_PREFETCH(ptr) (void)(ptr) /* disabled */ - #endif -#endif /* XXH_NO_PREFETCH */ - -/* ========================================== - * XXH3 default settings - * ========================================== */ - -#define XXH_SECRET_DEFAULT_SIZE 192 /* minimum XXH3_SECRET_SIZE_MIN */ - -#if (XXH_SECRET_DEFAULT_SIZE < XXH3_SECRET_SIZE_MIN) - #error "default keyset is not large enough" -#endif - -/* Pseudorandom secret taken directly from FARSH */ -XXH_ALIGN(64) -static const xxh_u8 XXH3_kSecret[XXH_SECRET_DEFAULT_SIZE] = { - - 0xb8, 0xfe, 0x6c, 0x39, 0x23, 0xa4, 0x4b, 0xbe, 0x7c, 0x01, 0x81, 0x2c, - 0xf7, 0x21, 0xad, 0x1c, 0xde, 0xd4, 0x6d, 0xe9, 0x83, 0x90, 0x97, 0xdb, - 0x72, 0x40, 0xa4, 0xa4, 0xb7, 0xb3, 0x67, 0x1f, 0xcb, 0x79, 0xe6, 0x4e, - 0xcc, 0xc0, 0xe5, 0x78, 0x82, 0x5a, 0xd0, 0x7d, 0xcc, 0xff, 0x72, 0x21, - 0xb8, 0x08, 0x46, 0x74, 0xf7, 0x43, 0x24, 0x8e, 0xe0, 0x35, 0x90, 0xe6, - 0x81, 0x3a, 0x26, 0x4c, 0x3c, 0x28, 0x52, 0xbb, 0x91, 0xc3, 0x00, 0xcb, - 0x88, 0xd0, 0x65, 0x8b, 0x1b, 0x53, 0x2e, 0xa3, 0x71, 0x64, 0x48, 0x97, - 0xa2, 0x0d, 0xf9, 0x4e, 0x38, 0x19, 0xef, 0x46, 0xa9, 0xde, 0xac, 0xd8, - 0xa8, 0xfa, 0x76, 0x3f, 0xe3, 0x9c, 0x34, 0x3f, 0xf9, 0xdc, 0xbb, 0xc7, - 0xc7, 0x0b, 0x4f, 0x1d, 0x8a, 0x51, 0xe0, 0x4b, 0xcd, 0xb4, 0x59, 0x31, - 0xc8, 0x9f, 0x7e, 0xc9, 0xd9, 0x78, 0x73, 0x64, - - 0xea, 0xc5, 0xac, 0x83, 0x34, 0xd3, 0xeb, 0xc3, 0xc5, 0x81, 0xa0, 0xff, - 0xfa, 0x13, 0x63, 0xeb, 0x17, 0x0d, 0xdd, 0x51, 0xb7, 0xf0, 0xda, 0x49, - 0xd3, 0x16, 0x55, 0x26, 0x29, 0xd4, 0x68, 0x9e, 0x2b, 0x16, 0xbe, 0x58, - 0x7d, 0x47, 0xa1, 0xfc, 0x8f, 0xf8, 0xb8, 0xd1, 0x7a, 0xd0, 0x31, 0xce, - 0x45, 0xcb, 0x3a, 0x8f, 0x95, 0x16, 0x04, 0x28, 0xaf, 0xd7, 0xfb, 0xca, - 0xbb, 0x4b, 0x40, 0x7e, - -}; - -#ifdef XXH_OLD_NAMES - #define kSecret XXH3_kSecret -#endif - -/* - * Calculates a 32-bit to 64-bit long multiply. - * - * Wraps __emulu on MSVC x86 because it tends to call __allmul when it doesn't - * need to (but it shouldn't need to anyways, it is about 7 instructions to do - * a 64x64 multiply...). Since we know that this will _always_ emit MULL, we - * use that instead of the normal method. - * - * If you are compiling for platforms like Thumb-1 and don't have a better - * option, you may also want to write your own long multiply routine here. - * - * XXH_FORCE_INLINE xxh_u64 XXH_mult32to64(xxh_u64 x, xxh_u64 y) - * { - - * return (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF); - * } - */ -#if defined(_MSC_VER) && defined(_M_IX86) - #include <intrin.h> - #define XXH_mult32to64(x, y) __emulu((unsigned)(x), (unsigned)(y)) -#else - /* - * Downcast + upcast is usually better than masking on older compilers like - * GCC 4.2 (especially 32-bit ones), all without affecting newer compilers. - * - * The other method, (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF), will AND both - * operands and perform a full 64x64 multiply -- entirely redundant on 32-bit. - */ - #define XXH_mult32to64(x, y) ((xxh_u64)(xxh_u32)(x) * (xxh_u64)(xxh_u32)(y)) -#endif - -/* - * Calculates a 64->128-bit long multiply. - * - * Uses __uint128_t and _umul128 if available, otherwise uses a scalar version. - */ -static XXH128_hash_t XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs) { - - /* - * GCC/Clang __uint128_t method. - * - * On most 64-bit targets, GCC and Clang define a __uint128_t type. - * This is usually the best way as it usually uses a native long 64-bit - * multiply, such as MULQ on x86_64 or MUL + UMULH on aarch64. - * - * Usually. - * - * Despite being a 32-bit platform, Clang (and emscripten) define this type - * despite not having the arithmetic for it. This results in a laggy - * compiler builtin call which calculates a full 128-bit multiply. - * In that case it is best to use the portable one. - * https://github.com/Cyan4973/xxHash/issues/211#issuecomment-515575677 - */ -#if defined(__GNUC__) && !defined(__wasm__) && defined(__SIZEOF_INT128__) || \ - (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128) - - __uint128_t const product = (__uint128_t)lhs * (__uint128_t)rhs; - XXH128_hash_t r128; - r128.low64 = (xxh_u64)(product); - r128.high64 = (xxh_u64)(product >> 64); - return r128; - - /* - * MSVC for x64's _umul128 method. - * - * xxh_u64 _umul128(xxh_u64 Multiplier, xxh_u64 Multiplicand, xxh_u64 - * *HighProduct); - * - * This compiles to single operand MUL on x64. - */ -#elif defined(_M_X64) || defined(_M_IA64) - - #ifndef _MSC_VER - #pragma intrinsic(_umul128) - #endif - xxh_u64 product_high; - xxh_u64 const product_low = _umul128(lhs, rhs, &product_high); - XXH128_hash_t r128; - r128.low64 = product_low; - r128.high64 = product_high; - return r128; - -#else - /* - * Portable scalar method. Optimized for 32-bit and 64-bit ALUs. - * - * This is a fast and simple grade school multiply, which is shown below - * with base 10 arithmetic instead of base 0x100000000. - * - * 9 3 // D2 lhs = 93 - * x 7 5 // D2 rhs = 75 - * ---------- - * 1 5 // D2 lo_lo = (93 % 10) * (75 % 10) = 15 - * 4 5 | // D2 hi_lo = (93 / 10) * (75 % 10) = 45 - * 2 1 | // D2 lo_hi = (93 % 10) * (75 / 10) = 21 - * + 6 3 | | // D2 hi_hi = (93 / 10) * (75 / 10) = 63 - * --------- - * 2 7 | // D2 cross = (15 / 10) + (45 % 10) + 21 = 27 - * + 6 7 | | // D2 upper = (27 / 10) + (45 / 10) + 63 = 67 - * --------- - * 6 9 7 5 // D4 res = (27 * 10) + (15 % 10) + (67 * 100) = 6975 - * - * The reasons for adding the products like this are: - * 1. It avoids manual carry tracking. Just like how - * (9 * 9) + 9 + 9 = 99, the same applies with this for UINT64_MAX. - * This avoids a lot of complexity. - * - * 2. It hints for, and on Clang, compiles to, the powerful UMAAL - * instruction available in ARM's Digital Signal Processing extension - * in 32-bit ARMv6 and later, which is shown below: - * - * void UMAAL(xxh_u32 *RdLo, xxh_u32 *RdHi, xxh_u32 Rn, xxh_u32 Rm) - * { - - * xxh_u64 product = (xxh_u64)*RdLo * (xxh_u64)*RdHi + Rn + Rm; - * *RdLo = (xxh_u32)(product & 0xFFFFFFFF); - * *RdHi = (xxh_u32)(product >> 32); - * } - * - * This instruction was designed for efficient long multiplication, and - * allows this to be calculated in only 4 instructions at speeds - * comparable to some 64-bit ALUs. - * - * 3. It isn't terrible on other platforms. Usually this will be a couple - * of 32-bit ADD/ADCs. - */ - - /* First calculate all of the cross products. */ - xxh_u64 const lo_lo = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs & 0xFFFFFFFF); - xxh_u64 const hi_lo = XXH_mult32to64(lhs >> 32, rhs & 0xFFFFFFFF); - xxh_u64 const lo_hi = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs >> 32); - xxh_u64 const hi_hi = XXH_mult32to64(lhs >> 32, rhs >> 32); - - /* Now add the products together. These will never overflow. */ - xxh_u64 const cross = (lo_lo >> 32) + (hi_lo & 0xFFFFFFFF) + lo_hi; - xxh_u64 const upper = (hi_lo >> 32) + (cross >> 32) + hi_hi; - xxh_u64 const lower = (cross << 32) | (lo_lo & 0xFFFFFFFF); - - XXH128_hash_t r128; - r128.low64 = lower; - r128.high64 = upper; - return r128; -#endif - -} - -/* - * Does a 64-bit to 128-bit multiply, then XOR folds it. - * - * The reason for the separate function is to prevent passing too many structs - * around by value. This will hopefully inline the multiply, but we don't force - * it. - */ -static xxh_u64 XXH3_mul128_fold64(xxh_u64 lhs, xxh_u64 rhs) { - - XXH128_hash_t product = XXH_mult64to128(lhs, rhs); - return product.low64 ^ product.high64; - -} - -/* Seems to produce slightly better code on GCC for some reason. */ -XXH_FORCE_INLINE xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift) { - - XXH_ASSERT(0 <= shift && shift < 64); - return v64 ^ (v64 >> shift); - -} - -/* - * We don't need to (or want to) mix as much as XXH64. - * - * Short hashes are more evenly distributed, so it isn't necessary. - */ -static XXH64_hash_t XXH3_avalanche(xxh_u64 h64) { - - h64 = XXH_xorshift64(h64, 37); - h64 *= 0x165667919E3779F9ULL; - h64 = XXH_xorshift64(h64, 32); - return h64; - -} - -/* ========================================== - * Short keys - * ========================================== - * One of the shortcomings of XXH32 and XXH64 was that their performance was - * sub-optimal on short lengths. It used an iterative algorithm which strongly - * favored lengths that were a multiple of 4 or 8. - * - * Instead of iterating over individual inputs, we use a set of single shot - * functions which piece together a range of lengths and operate in constant - * time. - * - * Additionally, the number of multiplies has been significantly reduced. This - * reduces latency, especially when emulating 64-bit multiplies on 32-bit. - * - * Depending on the platform, this may or may not be faster than XXH32, but it - * is almost guaranteed to be faster than XXH64. - */ - -/* - * At very short lengths, there isn't enough input to fully hide secrets, or use - * the entire secret. - * - * There is also only a limited amount of mixing we can do before significantly - * impacting performance. - * - * Therefore, we use different sections of the secret and always mix two secret - * samples with an XOR. This should have no effect on performance on the - * seedless or withSeed variants because everything _should_ be constant folded - * by modern compilers. - * - * The XOR mixing hides individual parts of the secret and increases entropy. - * - * This adds an extra layer of strength for custom secrets. - */ -XXH_FORCE_INLINE XXH64_hash_t XXH3_len_1to3_64b(const xxh_u8 *input, size_t len, - const xxh_u8 *secret, - XXH64_hash_t seed) { - - XXH_ASSERT(input != NULL); - XXH_ASSERT(1 <= len && len <= 3); - XXH_ASSERT(secret != NULL); - /* - * len = 1: combined = { input[0], 0x01, input[0], input[0] } - * len = 2: combined = { input[1], 0x02, input[0], input[1] } - * len = 3: combined = { input[2], 0x03, input[0], input[1] } - */ - { - - xxh_u8 const c1 = input[0]; - xxh_u8 const c2 = input[len >> 1]; - xxh_u8 const c3 = input[len - 1]; - xxh_u32 const combined = ((xxh_u32)c1 << 16) | ((xxh_u32)c2 << 24) | - ((xxh_u32)c3 << 0) | ((xxh_u32)len << 8); - xxh_u64 const bitflip = - (XXH_readLE32(secret) ^ XXH_readLE32(secret + 4)) + seed; - xxh_u64 const keyed = (xxh_u64)combined ^ bitflip; - xxh_u64 const mixed = keyed * XXH_PRIME64_1; - return XXH3_avalanche(mixed); - - } - -} - -XXH_FORCE_INLINE XXH64_hash_t XXH3_len_4to8_64b(const xxh_u8 *input, size_t len, - const xxh_u8 *secret, - XXH64_hash_t seed) { - - XXH_ASSERT(input != NULL); - XXH_ASSERT(secret != NULL); - XXH_ASSERT(4 <= len && len < 8); - seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32; - { - - xxh_u32 const input1 = XXH_readLE32(input); - xxh_u32 const input2 = XXH_readLE32(input + len - 4); - xxh_u64 const bitflip = - (XXH_readLE64(secret + 8) ^ XXH_readLE64(secret + 16)) - seed; - xxh_u64 const input64 = input2 + (((xxh_u64)input1) << 32); - xxh_u64 x = input64 ^ bitflip; - /* this mix is inspired by Pelle Evensen's rrmxmx */ - x ^= XXH_rotl64(x, 49) ^ XXH_rotl64(x, 24); - x *= 0x9FB21C651E98DF25ULL; - x ^= (x >> 35) + len; - x *= 0x9FB21C651E98DF25ULL; - return XXH_xorshift64(x, 28); - - } - -} - -XXH_FORCE_INLINE XXH64_hash_t XXH3_len_9to16_64b(const xxh_u8 *input, - size_t len, - const xxh_u8 *secret, - XXH64_hash_t seed) { - - XXH_ASSERT(input != NULL); - XXH_ASSERT(secret != NULL); - XXH_ASSERT(8 <= len && len <= 16); - { - - xxh_u64 const bitflip1 = - (XXH_readLE64(secret + 24) ^ XXH_readLE64(secret + 32)) + seed; - xxh_u64 const bitflip2 = - (XXH_readLE64(secret + 40) ^ XXH_readLE64(secret + 48)) - seed; - xxh_u64 const input_lo = XXH_readLE64(input) ^ bitflip1; - xxh_u64 const input_hi = XXH_readLE64(input + len - 8) ^ bitflip2; - xxh_u64 const acc = len + XXH_swap64(input_lo) + input_hi + - XXH3_mul128_fold64(input_lo, input_hi); - return XXH3_avalanche(acc); - - } - -} - -XXH_FORCE_INLINE XXH64_hash_t XXH3_len_0to16_64b(const xxh_u8 *input, - size_t len, - const xxh_u8 *secret, - XXH64_hash_t seed) { - - XXH_ASSERT(len <= 16); - { - - if (XXH_likely(len > 8)) - return XXH3_len_9to16_64b(input, len, secret, seed); - if (XXH_likely(len >= 4)) - return XXH3_len_4to8_64b(input, len, secret, seed); - if (len) return XXH3_len_1to3_64b(input, len, secret, seed); - return XXH3_avalanche((XXH_PRIME64_1 + seed) ^ (XXH_readLE64(secret + 56) ^ - XXH_readLE64(secret + 64))); - - } - -} - -/* - * DISCLAIMER: There are known *seed-dependent* multicollisions here due to - * multiplication by zero, affecting hashes of lengths 17 to 240. - * - * However, they are very unlikely. - * - * Keep this in mind when using the unseeded XXH3_64bits() variant: As with all - * unseeded non-cryptographic hashes, it does not attempt to defend itself - * against specially crafted inputs, only random inputs. - * - * Compared to classic UMAC where a 1 in 2^31 chance of 4 consecutive bytes - * cancelling out the secret is taken an arbitrary number of times (addressed - * in XXH3_accumulate_512), this collision is very unlikely with random inputs - * and/or proper seeding: - * - * This only has a 1 in 2^63 chance of 8 consecutive bytes cancelling out, in a - * function that is only called up to 16 times per hash with up to 240 bytes of - * input. - * - * This is not too bad for a non-cryptographic hash function, especially with - * only 64 bit outputs. - * - * The 128-bit variant (which trades some speed for strength) is NOT affected - * by this, although it is always a good idea to use a proper seed if you care - * about strength. - */ -XXH_FORCE_INLINE xxh_u64 XXH3_mix16B(const xxh_u8 *XXH_RESTRICT input, - const xxh_u8 *XXH_RESTRICT secret, - xxh_u64 seed64) { - -#if defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \ - && defined(__i386__) && defined(__SSE2__) /* x86 + SSE2 */ \ - && \ - !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable like XXH32 hack */ - /* - * UGLY HACK: - * GCC for x86 tends to autovectorize the 128-bit multiply, resulting in - * slower code. - * - * By forcing seed64 into a register, we disrupt the cost model and - * cause it to scalarize. See `XXH32_round()` - * - * FIXME: Clang's output is still _much_ faster -- On an AMD Ryzen 3600, - * XXH3_64bits @ len=240 runs at 4.6 GB/s with Clang 9, but 3.3 GB/s on - * GCC 9.2, despite both emitting scalar code. - * - * GCC generates much better scalar code than Clang for the rest of XXH3, - * which is why finding a more optimal codepath is an interest. - */ - __asm__("" : "+r"(seed64)); -#endif - { - - xxh_u64 const input_lo = XXH_readLE64(input); - xxh_u64 const input_hi = XXH_readLE64(input + 8); - return XXH3_mul128_fold64(input_lo ^ (XXH_readLE64(secret) + seed64), - input_hi ^ (XXH_readLE64(secret + 8) - seed64)); - - } - -} - -/* For mid range keys, XXH3 uses a Mum-hash variant. */ -XXH_FORCE_INLINE XXH64_hash_t XXH3_len_17to128_64b( - const xxh_u8 *XXH_RESTRICT input, size_t len, - const xxh_u8 *XXH_RESTRICT secret, size_t secretSize, XXH64_hash_t seed) { - - XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); - (void)secretSize; - XXH_ASSERT(16 < len && len <= 128); - - { - - xxh_u64 acc = len * XXH_PRIME64_1; - if (len > 32) { - - if (len > 64) { - - if (len > 96) { - - acc += XXH3_mix16B(input + 48, secret + 96, seed); - acc += XXH3_mix16B(input + len - 64, secret + 112, seed); - - } - - acc += XXH3_mix16B(input + 32, secret + 64, seed); - acc += XXH3_mix16B(input + len - 48, secret + 80, seed); - - } - - acc += XXH3_mix16B(input + 16, secret + 32, seed); - acc += XXH3_mix16B(input + len - 32, secret + 48, seed); - - } - - acc += XXH3_mix16B(input + 0, secret + 0, seed); - acc += XXH3_mix16B(input + len - 16, secret + 16, seed); - - return XXH3_avalanche(acc); - - } - -} - -#define XXH3_MIDSIZE_MAX 240 - -XXH_NO_INLINE XXH64_hash_t XXH3_len_129to240_64b( - const xxh_u8 *XXH_RESTRICT input, size_t len, - const xxh_u8 *XXH_RESTRICT secret, size_t secretSize, XXH64_hash_t seed) { - - XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); - (void)secretSize; - XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX); - -#define XXH3_MIDSIZE_STARTOFFSET 3 -#define XXH3_MIDSIZE_LASTOFFSET 17 - - { - - xxh_u64 acc = len * XXH_PRIME64_1; - int const nbRounds = (int)len / 16; - int i; - for (i = 0; i < 8; i++) { - - acc += XXH3_mix16B(input + (16 * i), secret + (16 * i), seed); - - } - - acc = XXH3_avalanche(acc); - XXH_ASSERT(nbRounds >= 8); -#if defined(__clang__) /* Clang */ \ - && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \ - && !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable */ - /* - * UGLY HACK: - * Clang for ARMv7-A tries to vectorize this loop, similar to GCC x86. - * In everywhere else, it uses scalar code. - * - * For 64->128-bit multiplies, even if the NEON was 100% optimal, it - * would still be slower than UMAAL (see XXH_mult64to128). - * - * Unfortunately, Clang doesn't handle the long multiplies properly and - * converts them to the nonexistent "vmulq_u64" intrinsic, which is then - * scalarized into an ugly mess of VMOV.32 instructions. - * - * This mess is difficult to avoid without turning autovectorization - * off completely, but they are usually relatively minor and/or not - * worth it to fix. - * - * This loop is the easiest to fix, as unlike XXH32, this pragma - * _actually works_ because it is a loop vectorization instead of an - * SLP vectorization. - */ - #pragma clang loop vectorize(disable) -#endif - for (i = 8; i < nbRounds; i++) { - - acc += - XXH3_mix16B(input + (16 * i), - secret + (16 * (i - 8)) + XXH3_MIDSIZE_STARTOFFSET, seed); - - } - - /* last bytes */ - acc += XXH3_mix16B(input + len - 16, - secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, - seed); - return XXH3_avalanche(acc); - - } - -} - -/* ======= Long Keys ======= */ - -#define XXH_STRIPE_LEN 64 -#define XXH_SECRET_CONSUME_RATE \ - 8 /* nb of secret bytes consumed at each accumulation */ -#define XXH_ACC_NB (XXH_STRIPE_LEN / sizeof(xxh_u64)) - -#ifdef XXH_OLD_NAMES - #define STRIPE_LEN XXH_STRIPE_LEN - #define ACC_NB XXH_ACC_NB -#endif - -typedef enum { XXH3_acc_64bits, XXH3_acc_128bits } XXH3_accWidth_e; - -XXH_FORCE_INLINE void XXH_writeLE64(void *dst, xxh_u64 v64) { - - if (!XXH_CPU_LITTLE_ENDIAN) v64 = XXH_swap64(v64); - memcpy(dst, &v64, sizeof(v64)); - -} - -/* Several intrinsic functions below are supposed to accept __int64 as argument, - * as documented in - * https://software.intel.com/sites/landingpage/IntrinsicsGuide/ . However, - * several environments do not define __int64 type, requiring a workaround. - */ -#if !defined(__VMS) && \ - (defined(__cplusplus) || \ - (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)) -typedef int64_t xxh_i64; -#else -/* the following type must have a width of 64-bit */ -typedef long long xxh_i64; -#endif - -/* - * XXH3_accumulate_512 is the tightest loop for long inputs, and it is the most - * optimized. - * - * It is a hardened version of UMAC, based off of FARSH's implementation. - * - * This was chosen because it adapts quite well to 32-bit, 64-bit, and SIMD - * implementations, and it is ridiculously fast. - * - * We harden it by mixing the original input to the accumulators as well as the - * product. - * - * This means that in the (relatively likely) case of a multiply by zero, the - * original input is preserved. - * - * On 128-bit inputs, we swap 64-bit pairs when we add the input to improve - * cross-pollination, as otherwise the upper and lower halves would be - * essentially independent. - * - * This doesn't matter on 64-bit hashes since they all get merged together in - * the end, so we skip the extra step. - * - * Both XXH3_64bits and XXH3_128bits use this subroutine. - */ - -#if (XXH_VECTOR == XXH_AVX512) || defined(XXH_X86DISPATCH) - - #ifndef XXH_TARGET_AVX512 - #define XXH_TARGET_AVX512 /* disable attribute target */ - #endif - -XXH_FORCE_INLINE XXH_TARGET_AVX512 void XXH3_accumulate_512_avx512( - void *XXH_RESTRICT acc, const void *XXH_RESTRICT input, - const void *XXH_RESTRICT secret, XXH3_accWidth_e accWidth) { - - XXH_ALIGN(64) __m512i *const xacc = (__m512i *)acc; - XXH_ASSERT((((size_t)acc) & 63) == 0); - XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i)); - - { - - /* data_vec = input[0]; */ - __m512i const data_vec = _mm512_loadu_si512(input); - /* key_vec = secret[0]; */ - __m512i const key_vec = _mm512_loadu_si512(secret); - /* data_key = data_vec ^ key_vec; */ - __m512i const data_key = _mm512_xor_si512(data_vec, key_vec); - /* data_key_lo = data_key >> 32; */ - __m512i const data_key_lo = - _mm512_shuffle_epi32(data_key, _MM_SHUFFLE(0, 3, 0, 1)); - /* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */ - __m512i const product = _mm512_mul_epu32(data_key, data_key_lo); - if (accWidth == XXH3_acc_128bits) { - - /* xacc[0] += swap(data_vec); */ - __m512i const data_swap = - _mm512_shuffle_epi32(data_vec, _MM_SHUFFLE(1, 0, 3, 2)); - __m512i const sum = _mm512_add_epi64(*xacc, data_swap); - /* xacc[0] += product; */ - *xacc = _mm512_add_epi64(product, sum); - - } else { /* XXH3_acc_64bits */ - - /* xacc[0] += data_vec; */ - __m512i const sum = _mm512_add_epi64(*xacc, data_vec); - /* xacc[0] += product; */ - *xacc = _mm512_add_epi64(product, sum); - - } - - } - -} - -/* - * XXH3_scrambleAcc: Scrambles the accumulators to improve mixing. - * - * Multiplication isn't perfect, as explained by Google in HighwayHash: - * - * // Multiplication mixes/scrambles bytes 0-7 of the 64-bit result to - * // varying degrees. In descending order of goodness, bytes - * // 3 4 2 5 1 6 0 7 have quality 228 224 164 160 100 96 36 32. - * // As expected, the upper and lower bytes are much worse. - * - * Source: - * https://github.com/google/highwayhash/blob/0aaf66b/highwayhash/hh_avx2.h#L291 - * - * Since our algorithm uses a pseudorandom secret to add some variance into the - * mix, we don't need to (or want to) mix as often or as much as HighwayHash - * does. - * - * This isn't as tight as XXH3_accumulate, but still written in SIMD to avoid - * extraction. - * - * Both XXH3_64bits and XXH3_128bits use this subroutine. - */ - -XXH_FORCE_INLINE XXH_TARGET_AVX512 void XXH3_scrambleAcc_avx512( - void *XXH_RESTRICT acc, const void *XXH_RESTRICT secret) { - - XXH_ASSERT((((size_t)acc) & 63) == 0); - XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i)); - { - - XXH_ALIGN(64) __m512i *const xacc = (__m512i *)acc; - const __m512i prime32 = _mm512_set1_epi32((int)XXH_PRIME32_1); - - /* xacc[0] ^= (xacc[0] >> 47) */ - __m512i const acc_vec = *xacc; - __m512i const shifted = _mm512_srli_epi64(acc_vec, 47); - __m512i const data_vec = _mm512_xor_si512(acc_vec, shifted); - /* xacc[0] ^= secret; */ - __m512i const key_vec = _mm512_loadu_si512(secret); - __m512i const data_key = _mm512_xor_si512(data_vec, key_vec); - - /* xacc[0] *= XXH_PRIME32_1; */ - __m512i const data_key_hi = - _mm512_shuffle_epi32(data_key, _MM_SHUFFLE(0, 3, 0, 1)); - __m512i const prod_lo = _mm512_mul_epu32(data_key, prime32); - __m512i const prod_hi = _mm512_mul_epu32(data_key_hi, prime32); - *xacc = _mm512_add_epi64(prod_lo, _mm512_slli_epi64(prod_hi, 32)); - - } - -} - -XXH_FORCE_INLINE XXH_TARGET_AVX512 void XXH3_initCustomSecret_avx512( - void *XXH_RESTRICT customSecret, xxh_u64 seed64) { - - XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 63) == 0); - XXH_STATIC_ASSERT(XXH_SEC_ALIGN == 64); - XXH_ASSERT(((size_t)customSecret & 63) == 0); - (void)(&XXH_writeLE64); - { - - int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m512i); - __m512i const seed = _mm512_mask_set1_epi64( - _mm512_set1_epi64((xxh_i64)seed64), 0xAA, -(xxh_i64)seed64); - - XXH_ALIGN(64) const __m512i *const src = (const __m512i *)XXH3_kSecret; - XXH_ALIGN(64) __m512i *const dest = (__m512i *)customSecret; - int i; - for (i = 0; i < nbRounds; ++i) { - - // GCC has a bug, _mm512_stream_load_si512 accepts 'void*', not 'void - // const*', this will warn "discards ‘const’ qualifier". - union { - - XXH_ALIGN(64) const __m512i *const cp; - XXH_ALIGN(64) void *const p; - - } const remote_const_void = {.cp = src + i}; - - dest[i] = - _mm512_add_epi64(_mm512_stream_load_si512(remote_const_void.p), seed); - - } - - } - -} - -#endif - -#if (XXH_VECTOR == XXH_AVX2) || defined(XXH_X86DISPATCH) - - #ifndef XXH_TARGET_AVX2 - #define XXH_TARGET_AVX2 /* disable attribute target */ - #endif - -XXH_FORCE_INLINE XXH_TARGET_AVX2 void XXH3_accumulate_512_avx2( - void *XXH_RESTRICT acc, const void *XXH_RESTRICT input, - const void *XXH_RESTRICT secret, XXH3_accWidth_e accWidth) { - - XXH_ASSERT((((size_t)acc) & 31) == 0); - { - - XXH_ALIGN(32) __m256i *const xacc = (__m256i *)acc; - /* Unaligned. This is mainly for pointer arithmetic, and because - * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. - */ - const __m256i *const xinput = (const __m256i *)input; - /* Unaligned. This is mainly for pointer arithmetic, and because - * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */ - const __m256i *const xsecret = (const __m256i *)secret; - - size_t i; - for (i = 0; i < XXH_STRIPE_LEN / sizeof(__m256i); i++) { - - /* data_vec = xinput[i]; */ - __m256i const data_vec = _mm256_loadu_si256(xinput + i); - /* key_vec = xsecret[i]; */ - __m256i const key_vec = _mm256_loadu_si256(xsecret + i); - /* data_key = data_vec ^ key_vec; */ - __m256i const data_key = _mm256_xor_si256(data_vec, key_vec); - /* data_key_lo = data_key >> 32; */ - __m256i const data_key_lo = - _mm256_shuffle_epi32(data_key, _MM_SHUFFLE(0, 3, 0, 1)); - /* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */ - __m256i const product = _mm256_mul_epu32(data_key, data_key_lo); - if (accWidth == XXH3_acc_128bits) { - - /* xacc[i] += swap(data_vec); */ - __m256i const data_swap = - _mm256_shuffle_epi32(data_vec, _MM_SHUFFLE(1, 0, 3, 2)); - __m256i const sum = _mm256_add_epi64(xacc[i], data_swap); - /* xacc[i] += product; */ - xacc[i] = _mm256_add_epi64(product, sum); - - } else { /* XXH3_acc_64bits */ - - /* xacc[i] += data_vec; */ - __m256i const sum = _mm256_add_epi64(xacc[i], data_vec); - /* xacc[i] += product; */ - xacc[i] = _mm256_add_epi64(product, sum); - - } - - } - - } - -} - -XXH_FORCE_INLINE XXH_TARGET_AVX2 void XXH3_scrambleAcc_avx2( - void *XXH_RESTRICT acc, const void *XXH_RESTRICT secret) { - - XXH_ASSERT((((size_t)acc) & 31) == 0); - { - - XXH_ALIGN(32) __m256i *const xacc = (__m256i *)acc; - /* Unaligned. This is mainly for pointer arithmetic, and because - * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */ - const __m256i *const xsecret = (const __m256i *)secret; - const __m256i prime32 = _mm256_set1_epi32((int)XXH_PRIME32_1); - - size_t i; - for (i = 0; i < XXH_STRIPE_LEN / sizeof(__m256i); i++) { - - /* xacc[i] ^= (xacc[i] >> 47) */ - __m256i const acc_vec = xacc[i]; - __m256i const shifted = _mm256_srli_epi64(acc_vec, 47); - __m256i const data_vec = _mm256_xor_si256(acc_vec, shifted); - /* xacc[i] ^= xsecret; */ - __m256i const key_vec = _mm256_loadu_si256(xsecret + i); - __m256i const data_key = _mm256_xor_si256(data_vec, key_vec); - - /* xacc[i] *= XXH_PRIME32_1; */ - __m256i const data_key_hi = - _mm256_shuffle_epi32(data_key, _MM_SHUFFLE(0, 3, 0, 1)); - __m256i const prod_lo = _mm256_mul_epu32(data_key, prime32); - __m256i const prod_hi = _mm256_mul_epu32(data_key_hi, prime32); - xacc[i] = _mm256_add_epi64(prod_lo, _mm256_slli_epi64(prod_hi, 32)); - - } - - } - -} - -XXH_FORCE_INLINE XXH_TARGET_AVX2 void XXH3_initCustomSecret_avx2( - void *XXH_RESTRICT customSecret, xxh_u64 seed64) { - - XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 31) == 0); - XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE / sizeof(__m256i)) == 6); - XXH_STATIC_ASSERT(XXH_SEC_ALIGN <= 64); - (void)(&XXH_writeLE64); - XXH_PREFETCH(customSecret); - { - - __m256i const seed = _mm256_set_epi64x(-(xxh_i64)seed64, (xxh_i64)seed64, - -(xxh_i64)seed64, (xxh_i64)seed64); - - XXH_ALIGN(64) const __m256i *const src = (const __m256i *)XXH3_kSecret; - XXH_ALIGN(64) __m256i * dest = (__m256i *)customSecret; - - #if defined(__GNUC__) || defined(__clang__) - /* - * On GCC & Clang, marking 'dest' as modified will cause the compiler: - * - do not extract the secret from sse registers in the internal loop - * - use less common registers, and avoid pushing these reg into stack - * The asm hack causes Clang to assume that XXH3_kSecretPtr aliases with - * customSecret, and on aarch64, this prevented LDP from merging two - * loads together for free. Putting the loads together before the stores - * properly generates LDP. - */ - __asm__("" : "+r"(dest)); - #endif - - /* GCC -O2 need unroll loop manually */ - dest[0] = _mm256_add_epi64(_mm256_stream_load_si256(src + 0), seed); - dest[1] = _mm256_add_epi64(_mm256_stream_load_si256(src + 1), seed); - dest[2] = _mm256_add_epi64(_mm256_stream_load_si256(src + 2), seed); - dest[3] = _mm256_add_epi64(_mm256_stream_load_si256(src + 3), seed); - dest[4] = _mm256_add_epi64(_mm256_stream_load_si256(src + 4), seed); - dest[5] = _mm256_add_epi64(_mm256_stream_load_si256(src + 5), seed); - - } - -} - -#endif - -#if (XXH_VECTOR == XXH_SSE2) || defined(XXH_X86DISPATCH) - - #ifndef XXH_TARGET_SSE2 - #define XXH_TARGET_SSE2 /* disable attribute target */ - #endif - -XXH_FORCE_INLINE XXH_TARGET_SSE2 void XXH3_accumulate_512_sse2( - void *XXH_RESTRICT acc, const void *XXH_RESTRICT input, - const void *XXH_RESTRICT secret, XXH3_accWidth_e accWidth) { - - /* SSE2 is just a half-scale version of the AVX2 version. */ - XXH_ASSERT((((size_t)acc) & 15) == 0); - { - - XXH_ALIGN(16) __m128i *const xacc = (__m128i *)acc; - /* Unaligned. This is mainly for pointer arithmetic, and because - * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */ - const __m128i *const xinput = (const __m128i *)input; - /* Unaligned. This is mainly for pointer arithmetic, and because - * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */ - const __m128i *const xsecret = (const __m128i *)secret; - - size_t i; - for (i = 0; i < XXH_STRIPE_LEN / sizeof(__m128i); i++) { - - /* data_vec = xinput[i]; */ - __m128i const data_vec = _mm_loadu_si128(xinput + i); - /* key_vec = xsecret[i]; */ - __m128i const key_vec = _mm_loadu_si128(xsecret + i); - /* data_key = data_vec ^ key_vec; */ - __m128i const data_key = _mm_xor_si128(data_vec, key_vec); - /* data_key_lo = data_key >> 32; */ - __m128i const data_key_lo = - _mm_shuffle_epi32(data_key, _MM_SHUFFLE(0, 3, 0, 1)); - /* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */ - __m128i const product = _mm_mul_epu32(data_key, data_key_lo); - if (accWidth == XXH3_acc_128bits) { - - /* xacc[i] += swap(data_vec); */ - __m128i const data_swap = - _mm_shuffle_epi32(data_vec, _MM_SHUFFLE(1, 0, 3, 2)); - __m128i const sum = _mm_add_epi64(xacc[i], data_swap); - /* xacc[i] += product; */ - xacc[i] = _mm_add_epi64(product, sum); - - } else { /* XXH3_acc_64bits */ - - /* xacc[i] += data_vec; */ - __m128i const sum = _mm_add_epi64(xacc[i], data_vec); - /* xacc[i] += product; */ - xacc[i] = _mm_add_epi64(product, sum); - - } - - } - - } - -} - -XXH_FORCE_INLINE XXH_TARGET_SSE2 void XXH3_scrambleAcc_sse2( - void *XXH_RESTRICT acc, const void *XXH_RESTRICT secret) { - - XXH_ASSERT((((size_t)acc) & 15) == 0); - { - - XXH_ALIGN(16) __m128i *const xacc = (__m128i *)acc; - /* Unaligned. This is mainly for pointer arithmetic, and because - * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */ - const __m128i *const xsecret = (const __m128i *)secret; - const __m128i prime32 = _mm_set1_epi32((int)XXH_PRIME32_1); - - size_t i; - for (i = 0; i < XXH_STRIPE_LEN / sizeof(__m128i); i++) { - - /* xacc[i] ^= (xacc[i] >> 47) */ - __m128i const acc_vec = xacc[i]; - __m128i const shifted = _mm_srli_epi64(acc_vec, 47); - __m128i const data_vec = _mm_xor_si128(acc_vec, shifted); - /* xacc[i] ^= xsecret[i]; */ - __m128i const key_vec = _mm_loadu_si128(xsecret + i); - __m128i const data_key = _mm_xor_si128(data_vec, key_vec); - - /* xacc[i] *= XXH_PRIME32_1; */ - __m128i const data_key_hi = - _mm_shuffle_epi32(data_key, _MM_SHUFFLE(0, 3, 0, 1)); - __m128i const prod_lo = _mm_mul_epu32(data_key, prime32); - __m128i const prod_hi = _mm_mul_epu32(data_key_hi, prime32); - xacc[i] = _mm_add_epi64(prod_lo, _mm_slli_epi64(prod_hi, 32)); - - } - - } - -} - -XXH_FORCE_INLINE XXH_TARGET_SSE2 void XXH3_initCustomSecret_sse2( - void *XXH_RESTRICT customSecret, xxh_u64 seed64) { - - XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0); - (void)(&XXH_writeLE64); - { - - int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m128i); - - #if defined(_MSC_VER) && defined(_M_IX86) && _MSC_VER < 1900 - // MSVC 32bit mode does not support _mm_set_epi64x before 2015 - XXH_ALIGN(16) - const xxh_i64 seed64x2[2] = {(xxh_i64)seed64, -(xxh_i64)seed64}; - __m128i const seed = _mm_load_si128((__m128i const *)seed64x2); - #else - __m128i const seed = _mm_set_epi64x(-(xxh_i64)seed64, (xxh_i64)seed64); - #endif - int i; - - XXH_ALIGN(64) const float *const src = (float const *)XXH3_kSecret; - XXH_ALIGN(XXH_SEC_ALIGN) __m128i *dest = (__m128i *)customSecret; - #if defined(__GNUC__) || defined(__clang__) - /* - * On GCC & Clang, marking 'dest' as modified will cause the compiler: - * - do not extract the secret from sse registers in the internal loop - * - use less common registers, and avoid pushing these reg into stack - */ - __asm__("" : "+r"(dest)); - #endif - - for (i = 0; i < nbRounds; ++i) { - - dest[i] = _mm_add_epi64(_mm_castps_si128(_mm_load_ps(src + i * 4)), seed); - - } - - } - -} - -#endif - -#if (XXH_VECTOR == XXH_NEON) - -XXH_FORCE_INLINE void XXH3_accumulate_512_neon(void *XXH_RESTRICT acc, - const void *XXH_RESTRICT input, - const void *XXH_RESTRICT secret, - XXH3_accWidth_e accWidth) { - - XXH_ASSERT((((size_t)acc) & 15) == 0); - { - - XXH_ALIGN(16) uint64x2_t *const xacc = (uint64x2_t *)acc; - /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. - */ - uint8_t const *const xinput = (const uint8_t *)input; - uint8_t const *const xsecret = (const uint8_t *)secret; - - size_t i; - for (i = 0; i < XXH_STRIPE_LEN / sizeof(uint64x2_t); i++) { - - /* data_vec = xinput[i]; */ - uint8x16_t data_vec = vld1q_u8(xinput + (i * 16)); - /* key_vec = xsecret[i]; */ - uint8x16_t key_vec = vld1q_u8(xsecret + (i * 16)); - uint64x2_t data_key; - uint32x2_t data_key_lo, data_key_hi; - if (accWidth == XXH3_acc_64bits) { - - /* xacc[i] += data_vec; */ - xacc[i] = vaddq_u64(xacc[i], vreinterpretq_u64_u8(data_vec)); - - } else { /* XXH3_acc_128bits */ - - /* xacc[i] += swap(data_vec); */ - uint64x2_t const data64 = vreinterpretq_u64_u8(data_vec); - uint64x2_t const swapped = vextq_u64(data64, data64, 1); - xacc[i] = vaddq_u64(xacc[i], swapped); - - } - - /* data_key = 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, - XXH3_accWidth_e accWidth) { - - 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; - - if (accWidth == XXH3_acc_64bits) { - - xacc[i] += data_vec; - - } else { /* XXH3_acc_128bits */ - - /* swap high and low halves */ - #ifdef __s390x__ - xxh_u64x2 const data_swapped = vec_permi(data_vec, data_vec, 2); - #else - xxh_u64x2 const data_swapped = vec_xxpermdi(data_vec, data_vec, 2); - #endif - xacc[i] += data_swapped; - - } - - } - -} - -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, XXH3_accWidth_e accWidth) { - - XXH_ALIGN(XXH_ACC_ALIGN) - xxh_u64 *const xacc = (xxh_u64 *)acc; /* presumed aligned */ - const xxh_u8 *const xinput = - (const xxh_u8 *)input; /* no alignment restriction */ - const xxh_u8 *const xsecret = - (const xxh_u8 *)secret; /* no alignment restriction */ - size_t i; - XXH_ASSERT(((size_t)acc & (XXH_ACC_ALIGN - 1)) == 0); - for (i = 0; i < XXH_ACC_NB; i++) { - - xxh_u64 const data_val = XXH_readLE64(xinput + 8 * i); - xxh_u64 const data_key = data_val ^ XXH_readLE64(xsecret + i * 8); - - if (accWidth == XXH3_acc_64bits) { - - xacc[i] += data_val; - - } else { - - xacc[i ^ 1] += data_val; /* swap adjacent lanes */ - - } - - xacc[i] += XXH_mult32to64(data_key & 0xFFFFFFFF, data_key >> 32); - - } - -} - -XXH_FORCE_INLINE void XXH3_scrambleAcc_scalar(void *XXH_RESTRICT acc, - const void *XXH_RESTRICT secret) { - - XXH_ALIGN(XXH_ACC_ALIGN) - xxh_u64 *const xacc = (xxh_u64 *)acc; /* presumed aligned */ - const xxh_u8 *const xsecret = - (const xxh_u8 *)secret; /* no alignment restriction */ - size_t i; - XXH_ASSERT((((size_t)acc) & (XXH_ACC_ALIGN - 1)) == 0); - for (i = 0; i < XXH_ACC_NB; i++) { - - xxh_u64 const key64 = XXH_readLE64(xsecret + 8 * i); - xxh_u64 acc64 = xacc[i]; - acc64 = XXH_xorshift64(acc64, 47); - acc64 ^= key64; - acc64 *= XXH_PRIME32_1; - xacc[i] = acc64; - - } - -} - -XXH_FORCE_INLINE void XXH3_initCustomSecret_scalar( - void *XXH_RESTRICT customSecret, xxh_u64 seed64) { - - /* - * We need a separate pointer for the hack below, - * which requires a non-const pointer. - * Any decent compiler will optimize this out otherwise. - */ - const xxh_u8 *kSecretPtr = XXH3_kSecret; - XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0); - -#if defined(__clang__) && defined(__aarch64__) - /* - * UGLY HACK: - * Clang generates a bunch of MOV/MOVK pairs for aarch64, and they are - * placed sequentially, in order, at the top of the unrolled loop. - * - * While MOVK is great for generating constants (2 cycles for a 64-bit - * constant compared to 4 cycles for LDR), long MOVK chains stall the - * integer pipelines: - * I L S - * MOVK - * MOVK - * MOVK - * MOVK - * ADD - * SUB STR - * STR - * By forcing loads from memory (as the asm line causes Clang to assume - * that XXH3_kSecretPtr has been changed), the pipelines are used more - * efficiently: - * I L S - * LDR - * ADD LDR - * SUB STR - * STR - * XXH3_64bits_withSeed, len == 256, Snapdragon 835 - * without hack: 2654.4 MB/s - * with hack: 3202.9 MB/s - */ - __asm__("" : "+r"(kSecretPtr)); -#endif - /* - * Note: in debug mode, this overrides the asm optimization - * and Clang will emit MOVK chains again. - */ - XXH_ASSERT(kSecretPtr == XXH3_kSecret); - - { - - int const nbRounds = XXH_SECRET_DEFAULT_SIZE / 16; - int i; - for (i = 0; i < nbRounds; i++) { - - /* - * The asm hack causes Clang to assume that kSecretPtr aliases with - * customSecret, and on aarch64, this prevented LDP from merging two - * loads together for free. Putting the loads together before the stores - * properly generates LDP. - */ - xxh_u64 lo = XXH_readLE64(kSecretPtr + 16 * i) + seed64; - xxh_u64 hi = XXH_readLE64(kSecretPtr + 16 * i + 8) - seed64; - XXH_writeLE64((xxh_u8 *)customSecret + 16 * i, lo); - XXH_writeLE64((xxh_u8 *)customSecret + 16 * i + 8, hi); - - } - - } - -} - -typedef void (*XXH3_f_accumulate_512)(void *XXH_RESTRICT, const void *, - const void *, XXH3_accWidth_e); -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_accWidth_e accWidth, - 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, accWidth); - - } - -} - -XXH_FORCE_INLINE void XXH3_hashLong_internal_loop( - xxh_u64 *XXH_RESTRICT acc, const xxh_u8 *XXH_RESTRICT input, size_t len, - const xxh_u8 *XXH_RESTRICT secret, size_t secretSize, - XXH3_accWidth_e accWidth, XXH3_f_accumulate_512 f_acc512, - XXH3_f_scrambleAcc f_scramble) { - - size_t const nb_rounds = - (secretSize - XXH_STRIPE_LEN) / XXH_SECRET_CONSUME_RATE; - size_t const block_len = XXH_STRIPE_LEN * nb_rounds; - size_t const nb_blocks = len / block_len; - - size_t n; - - XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); - - for (n = 0; n < nb_blocks; n++) { - - XXH3_accumulate(acc, input + n * block_len, secret, nb_rounds, accWidth, - f_acc512); - f_scramble(acc, secret + secretSize - XXH_STRIPE_LEN); - - } - - /* last partial block */ - XXH_ASSERT(len > XXH_STRIPE_LEN); - { - - size_t const nbStripes = (len - (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, - accWidth, f_acc512); - - /* last stripe */ - if (len & (XXH_STRIPE_LEN - 1)) { - - const xxh_u8 *const p = input + len - XXH_STRIPE_LEN; - /* Do not align on 8, so that the secret is different from the scrambler - */ -#define XXH_SECRET_LASTACC_START 7 - f_acc512(acc, p, - secret + secretSize - XXH_STRIPE_LEN - XXH_SECRET_LASTACC_START, - accWidth); - - } - - } - -} - -XXH_FORCE_INLINE xxh_u64 XXH3_mix2Accs(const xxh_u64 *XXH_RESTRICT acc, - const xxh_u8 *XXH_RESTRICT secret) { - - return XXH3_mul128_fold64(acc[0] ^ XXH_readLE64(secret), - acc[1] ^ XXH_readLE64(secret + 8)); - -} - -static XXH64_hash_t XXH3_mergeAccs(const xxh_u64 *XXH_RESTRICT acc, - const xxh_u8 *XXH_RESTRICT secret, - xxh_u64 start) { - - xxh_u64 result64 = start; - size_t i = 0; - - for (i = 0; i < 4; i++) { - - result64 += XXH3_mix2Accs(acc + 2 * i, secret + 16 * i); -#if defined(__clang__) /* Clang */ \ - && (defined(__arm__) || defined(__thumb__)) /* ARMv7 */ \ - && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \ - && !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable */ - /* - * UGLY HACK: - * Prevent autovectorization on Clang ARMv7-a. Exact same problem as - * the one in XXH3_len_129to240_64b. Speeds up shorter keys > 240b. - * XXH3_64bits, len == 256, Snapdragon 835: - * without hack: 2063.7 MB/s - * with hack: 2560.7 MB/s - */ - __asm__("" : "+r"(result64)); -#endif - - } - - return XXH3_avalanche(result64); - -} - -#define XXH3_INIT_ACC \ - { \ - \ - XXH_PRIME32_3, XXH_PRIME64_1, XXH_PRIME64_2, XXH_PRIME64_3, XXH_PRIME64_4, \ - XXH_PRIME32_2, XXH_PRIME64_5, XXH_PRIME32_1 \ - \ - } - -XXH_FORCE_INLINE XXH64_hash_t XXH3_hashLong_64b_internal( - const 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) { - - XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC; - - XXH3_hashLong_internal_loop(acc, input, len, secret, secretSize, - XXH3_acc_64bits, 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, 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 xxh_u8 *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); - -} - -/* - * 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 xxh_u8 *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 xxh_u8 *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 xxh_u8 *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((const xxh_u8 *)input, len, seed64, (const xxh_u8 *)secret, - secretLen); - -} - -/* === Public entry point === */ - -XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(const void *input, size_t len) { - - return XXH3_64bits_internal(input, len, 0, XXH3_kSecret, sizeof(XXH3_kSecret), - XXH3_hashLong_64b_withSecret); - -} - -XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSecret(const void *input, - size_t len, - const void *secret, - size_t secretSize) { - - return XXH3_64bits_internal(input, len, 0, secret, secretSize, - XXH3_hashLong_64b_withSecret); - -} - -XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSeed(const void *input, size_t len, - XXH64_hash_t seed) { - - return XXH3_64bits_internal(input, len, seed, XXH3_kSecret, - sizeof(XXH3_kSecret), XXH3_hashLong_64b_withSeed); - -} - -/* === XXH3 streaming === */ - -/* - * Malloc's a pointer that is always aligned to align. - * - * This must be freed with `XXH_alignedFree()`. - * - * malloc typically guarantees 16 byte alignment on 64-bit systems and 8 byte - * alignment on 32-bit. This isn't enough for the 32 byte aligned loads in AVX2 - * or on 32-bit, the 16 byte aligned loads in SSE2 and NEON. - * - * This underalignment previously caused a rather obvious crash which went - * completely unnoticed due to XXH3_createState() not actually being tested. - * Credit to RedSpah for noticing this bug. - * - * The alignment is done manually: Functions like posix_memalign or _mm_malloc - * are avoided: To maintain portability, we would have to write a fallback - * like this anyways, and besides, testing for the existence of library - * functions without relying on external build tools is impossible. - * - * The method is simple: Overallocate, manually align, and store the offset - * to the original behind the returned pointer. - * - * Align must be a power of 2 and 8 <= align <= 128. - */ -static void *XXH_alignedMalloc(size_t s, size_t align) { - - XXH_ASSERT(align <= 128 && align >= 8); /* range check */ - XXH_ASSERT((align & (align - 1)) == 0); /* power of 2 */ - XXH_ASSERT(s != 0 && s < (s + align)); /* empty/overflow */ - { /* Overallocate to make room for manual realignment and an offset byte */ - xxh_u8 *base = (xxh_u8 *)XXH_malloc(s + align); - if (base != NULL) { - - /* - * Get the offset needed to align this pointer. - * - * Even if the returned pointer is aligned, there will always be - * at least one byte to store the offset to the original pointer. - */ - size_t offset = align - ((size_t)base & (align - 1)); /* base % align */ - /* Add the offset for the now-aligned pointer */ - xxh_u8 *ptr = base + offset; - - XXH_ASSERT((size_t)ptr % align == 0); - - /* Store the offset immediately before the returned pointer. */ - ptr[-1] = (xxh_u8)offset; - return ptr; - - } - - return NULL; - - } - -} - -/* - * Frees an aligned pointer allocated by XXH_alignedMalloc(). Don't pass - * normal malloc'd pointers, XXH_alignedMalloc has a specific data layout. - */ -static void XXH_alignedFree(void *p) { - - if (p != NULL) { - - xxh_u8 *ptr = (xxh_u8 *)p; - /* Get the offset byte we added in XXH_malloc. */ - xxh_u8 offset = ptr[-1]; - /* Free the original malloc'd pointer */ - xxh_u8 *base = ptr - offset; - XXH_free(base); - - } - -} - -XXH_PUBLIC_API XXH3_state_t *XXH3_createState(void) { - - return (XXH3_state_t *)XXH_alignedMalloc(sizeof(XXH3_state_t), 64); - -} - -XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t *statePtr) { - - XXH_alignedFree(statePtr); - return XXH_OK; - -} - -XXH_PUBLIC_API void XXH3_copyState(XXH3_state_t * dst_state, - const XXH3_state_t *src_state) { - - memcpy(dst_state, src_state, sizeof(*dst_state)); - -} - -static void XXH3_64bits_reset_internal(XXH3_state_t *statePtr, - XXH64_hash_t seed, const xxh_u8 *secret, - size_t secretSize) { - - XXH_ASSERT(statePtr != NULL); - memset(statePtr, 0, sizeof(*statePtr)); - 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; - XXH_ASSERT(secret != NULL); - statePtr->extSecret = secret; - XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); - statePtr->secretLimit = secretSize - XXH_STRIPE_LEN; - statePtr->nbStripesPerBlock = statePtr->secretLimit / XXH_SECRET_CONSUME_RATE; - -} - -XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH3_state_t *statePtr) { - - if (statePtr == NULL) return XXH_ERROR; - XXH3_64bits_reset_internal(statePtr, 0, XXH3_kSecret, - XXH_SECRET_DEFAULT_SIZE); - return XXH_OK; - -} - -XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret( - XXH3_state_t *statePtr, const void *secret, size_t secretSize) { - - if (statePtr == NULL) return XXH_ERROR; - XXH3_64bits_reset_internal(statePtr, 0, (const xxh_u8 *)secret, secretSize); - if (secret == NULL) return XXH_ERROR; - if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR; - return XXH_OK; - -} - -XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH3_state_t *statePtr, - XXH64_hash_t seed) { - - if (statePtr == NULL) return XXH_ERROR; - XXH3_64bits_reset_internal(statePtr, seed, XXH3_kSecret, - XXH_SECRET_DEFAULT_SIZE); - XXH3_initCustomSecret(statePtr->customSecret, seed); - statePtr->extSecret = NULL; - return XXH_OK; - -} - -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 totalStripes, const xxh_u8 *XXH_RESTRICT secret, size_t secretLimit, - XXH3_accWidth_e accWidth, XXH3_f_accumulate_512 f_acc512, - XXH3_f_scrambleAcc f_scramble) { - - XXH_ASSERT(*nbStripesSoFarPtr < nbStripesPerBlock); - if (nbStripesPerBlock - *nbStripesSoFarPtr <= totalStripes) { - - /* need a scrambling operation */ - size_t const nbStripes = nbStripesPerBlock - *nbStripesSoFarPtr; - XXH3_accumulate(acc, input, - secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, - nbStripes, accWidth, f_acc512); - f_scramble(acc, secret + secretLimit); - XXH3_accumulate(acc, input + nbStripes * XXH_STRIPE_LEN, secret, - totalStripes - nbStripes, accWidth, f_acc512); - *nbStripesSoFarPtr = totalStripes - nbStripes; - - } else { - - XXH3_accumulate(acc, input, - secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, - totalStripes, accWidth, f_acc512); - *nbStripesSoFarPtr += totalStripes; - - } - -} - -/* - * Both XXH3_64bits_update and XXH3_128bits_update use this routine. - */ -XXH_FORCE_INLINE XXH_errorcode XXH3_update(XXH3_state_t *state, - const xxh_u8 *input, size_t len, - XXH3_accWidth_e accWidth, - XXH3_f_accumulate_512 f_acc512, - XXH3_f_scrambleAcc f_scramble) { - - if (input == NULL) -#if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && \ - (XXH_ACCEPT_NULL_INPUT_POINTER >= 1) - return XXH_OK; -#else - return XXH_ERROR; -#endif - - { - - const xxh_u8 *const bEnd = input + len; - const unsigned char *const secret = - (state->extSecret == NULL) ? state->customSecret : state->extSecret; - - state->totalLen += len; - - if (state->bufferedSize + len <= - XXH3_INTERNALBUFFER_SIZE) { /* fill in tmp buffer */ - XXH_memcpy(state->buffer + state->bufferedSize, input, len); - state->bufferedSize += (XXH32_hash_t)len; - return XXH_OK; - - } - - /* 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 */ - - /* - * There is some input left inside the internal buffer. - * Fill it, then consume it. - */ - if (state->bufferedSize) { - - size_t const loadSize = XXH3_INTERNALBUFFER_SIZE - state->bufferedSize; - XXH_memcpy(state->buffer + state->bufferedSize, input, loadSize); - input += loadSize; - XXH3_consumeStripes(state->acc, &state->nbStripesSoFar, - state->nbStripesPerBlock, state->buffer, - XXH3_INTERNALBUFFER_STRIPES, secret, - state->secretLimit, accWidth, f_acc512, f_scramble); - state->bufferedSize = 0; - - } - - /* Consume input by full buffer quantities */ - if (input + XXH3_INTERNALBUFFER_SIZE <= bEnd) { - - const xxh_u8 *const limit = bEnd - XXH3_INTERNALBUFFER_SIZE; - do { - - XXH3_consumeStripes(state->acc, &state->nbStripesSoFar, - state->nbStripesPerBlock, input, - XXH3_INTERNALBUFFER_STRIPES, secret, - state->secretLimit, accWidth, 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 (input < bEnd) { /* Some remaining input: buffer it */ - XXH_memcpy(state->buffer, input, (size_t)(bEnd - input)); - state->bufferedSize = (XXH32_hash_t)(bEnd - input); - - } - - } - - return XXH_OK; - -} - -XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update(XXH3_state_t *state, - const void *input, size_t len) { - - return XXH3_update(state, (const xxh_u8 *)input, len, XXH3_acc_64bits, - XXH3_accumulate_512, XXH3_scrambleAcc); - -} - -XXH_FORCE_INLINE void XXH3_digest_long(XXH64_hash_t * acc, - const XXH3_state_t * state, - const unsigned char *secret, - XXH3_accWidth_e accWidth) { - - /* - * Digest on a local copy. This way, the state remains unaltered, and it can - * continue ingesting more input afterwards. - */ - memcpy(acc, state->acc, sizeof(state->acc)); - if (state->bufferedSize >= XXH_STRIPE_LEN) { - - size_t const nbStripes = state->bufferedSize / XXH_STRIPE_LEN; - size_t nbStripesSoFar = state->nbStripesSoFar; - XXH3_consumeStripes(acc, &nbStripesSoFar, state->nbStripesPerBlock, - state->buffer, nbStripes, secret, state->secretLimit, - accWidth, XXH3_accumulate_512, XXH3_scrambleAcc); - if (state->bufferedSize % XXH_STRIPE_LEN) { /* one last partial stripe */ - XXH3_accumulate_512( - acc, state->buffer + state->bufferedSize - XXH_STRIPE_LEN, - secret + state->secretLimit - XXH_SECRET_LASTACC_START, accWidth); - - } - - } else { /* bufferedSize < XXH_STRIPE_LEN */ - - if (state->bufferedSize) { /* one last stripe */ - xxh_u8 lastStripe[XXH_STRIPE_LEN]; - size_t const catchupSize = XXH_STRIPE_LEN - state->bufferedSize; - memcpy(lastStripe, state->buffer + sizeof(state->buffer) - catchupSize, - catchupSize); - memcpy(lastStripe + catchupSize, state->buffer, state->bufferedSize); - XXH3_accumulate_512( - acc, lastStripe, - secret + state->secretLimit - XXH_SECRET_LASTACC_START, accWidth); - - } - - } - -} - -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, XXH3_acc_64bits); - return XXH3_mergeAccs(acc, secret + XXH_SECRET_MERGEACCS_START, - (xxh_u64)state->totalLen * XXH_PRIME64_1); - - } - - /* totalLen <= XXH3_MIDSIZE_MAX: digesting a short input */ - if (state->seed) - return XXH3_64bits_withSeed(state->buffer, (size_t)state->totalLen, - state->seed); - return XXH3_64bits_withSecret(state->buffer, (size_t)(state->totalLen), - secret, state->secretLimit + XXH_STRIPE_LEN); - -} - -#define XXH_MIN(x, y) (((x) > (y)) ? (y) : (x)) - -XXH_PUBLIC_API void XXH3_generateSecret(void * secretBuffer, - const void *customSeed, - size_t customSeedSize) { - - XXH_ASSERT(secretBuffer != NULL); - if (customSeedSize == 0) { - - memcpy(secretBuffer, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE); - return; - - } - - XXH_ASSERT(customSeed != NULL); - - { - - size_t const segmentSize = sizeof(XXH128_hash_t); - size_t const nbSegments = XXH_SECRET_DEFAULT_SIZE / segmentSize; - XXH128_canonical_t scrambler; - XXH64_hash_t seeds[12]; - size_t segnb; - XXH_ASSERT(nbSegments == 12); - XXH_ASSERT(segmentSize * nbSegments == - XXH_SECRET_DEFAULT_SIZE); /* exact multiple */ - XXH128_canonicalFromHash(&scrambler, XXH128(customSeed, customSeedSize, 0)); - - /* - * Copy customSeed to seeds[], truncating or repeating as necessary. - */ - { - - size_t toFill = XXH_MIN(customSeedSize, sizeof(seeds)); - size_t filled = toFill; - memcpy(seeds, customSeed, toFill); - while (filled < sizeof(seeds)) { - - toFill = XXH_MIN(filled, sizeof(seeds) - filled); - memcpy((char *)seeds + filled, seeds, toFill); - filled += toFill; - - } - - } - - /* generate secret */ - memcpy(secretBuffer, &scrambler, sizeof(scrambler)); - for (segnb = 1; segnb < nbSegments; segnb++) { - - size_t const segmentStart = segnb * segmentSize; - XXH128_canonical_t segment; - XXH128_canonicalFromHash(&segment, - XXH128(&scrambler, sizeof(scrambler), - XXH_readLE64(seeds + segnb) + segnb)); - memcpy((char *)secretBuffer + segmentStart, &segment, sizeof(segment)); - - } - - } - -} - -/* ========================================== - * XXH3 128 bits (a.k.a XXH128) - * ========================================== - * XXH3's 128-bit variant has better mixing and strength than the 64-bit - * variant, even without counting the significantly larger output size. - * - * For example, extra steps are taken to avoid the seed-dependent collisions - * in 17-240 byte inputs (See XXH3_mix16B and XXH128_mix32B). - * - * This strength naturally comes at the cost of some speed, especially on short - * lengths. Note that longer hashes are about as fast as the 64-bit version - * due to it using only a slight modification of the 64-bit loop. - * - * XXH128 is also more oriented towards 64-bit machines. It is still extremely - * fast for a _128-bit_ hash on 32-bit (it usually clears XXH64). - */ - -XXH_FORCE_INLINE XXH128_hash_t XXH3_len_1to3_128b(const xxh_u8 *input, - size_t len, - const xxh_u8 *secret, - XXH64_hash_t seed) { - - /* A doubled version of 1to3_64b with different constants. */ - XXH_ASSERT(input != NULL); - XXH_ASSERT(1 <= len && len <= 3); - XXH_ASSERT(secret != NULL); - /* - * len = 1: combinedl = { input[0], 0x01, input[0], input[0] } - * len = 2: combinedl = { input[1], 0x02, input[0], input[1] } - * len = 3: combinedl = { input[2], 0x03, input[0], input[1] } - */ - { - - xxh_u8 const c1 = input[0]; - xxh_u8 const c2 = input[len >> 1]; - xxh_u8 const c3 = input[len - 1]; - xxh_u32 const combinedl = ((xxh_u32)c1 << 16) | ((xxh_u32)c2 << 24) | - ((xxh_u32)c3 << 0) | ((xxh_u32)len << 8); - xxh_u32 const combinedh = XXH_rotl32(XXH_swap32(combinedl), 13); - xxh_u64 const bitflipl = - (XXH_readLE32(secret) ^ XXH_readLE32(secret + 4)) + seed; - xxh_u64 const bitfliph = - (XXH_readLE32(secret + 8) ^ XXH_readLE32(secret + 12)) - seed; - xxh_u64 const keyed_lo = (xxh_u64)combinedl ^ bitflipl; - xxh_u64 const keyed_hi = (xxh_u64)combinedh ^ bitfliph; - xxh_u64 const mixedl = keyed_lo * XXH_PRIME64_1; - xxh_u64 const mixedh = keyed_hi * XXH_PRIME64_5; - XXH128_hash_t h128; - h128.low64 = XXH3_avalanche(mixedl); - h128.high64 = XXH3_avalanche(mixedh); - return h128; - - } - -} - -XXH_FORCE_INLINE XXH128_hash_t XXH3_len_4to8_128b(const xxh_u8 *input, - size_t len, - const xxh_u8 *secret, - XXH64_hash_t seed) { - - XXH_ASSERT(input != NULL); - XXH_ASSERT(secret != NULL); - XXH_ASSERT(4 <= len && len <= 8); - seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32; - { - - xxh_u32 const input_lo = XXH_readLE32(input); - xxh_u32 const input_hi = XXH_readLE32(input + len - 4); - xxh_u64 const input_64 = input_lo + ((xxh_u64)input_hi << 32); - xxh_u64 const bitflip = - (XXH_readLE64(secret + 16) ^ XXH_readLE64(secret + 24)) + seed; - xxh_u64 const keyed = input_64 ^ bitflip; - - /* Shift len to the left to ensure it is even, this avoids even multiplies. - */ - XXH128_hash_t m128 = XXH_mult64to128(keyed, XXH_PRIME64_1 + (len << 2)); - - m128.high64 += (m128.low64 << 1); - m128.low64 ^= (m128.high64 >> 3); - - m128.low64 = XXH_xorshift64(m128.low64, 35); - m128.low64 *= 0x9FB21C651E98DF25ULL; - m128.low64 = XXH_xorshift64(m128.low64, 28); - m128.high64 = XXH3_avalanche(m128.high64); - return m128; - - } - -} - -XXH_FORCE_INLINE XXH128_hash_t XXH3_len_9to16_128b(const xxh_u8 *input, - size_t len, - const xxh_u8 *secret, - XXH64_hash_t seed) { - - XXH_ASSERT(input != NULL); - XXH_ASSERT(secret != NULL); - XXH_ASSERT(9 <= len && len <= 16); - { - - xxh_u64 const bitflipl = - (XXH_readLE64(secret + 32) ^ XXH_readLE64(secret + 40)) - seed; - xxh_u64 const bitfliph = - (XXH_readLE64(secret + 48) ^ XXH_readLE64(secret + 56)) + seed; - xxh_u64 const input_lo = XXH_readLE64(input); - xxh_u64 input_hi = XXH_readLE64(input + len - 8); - XXH128_hash_t m128 = - XXH_mult64to128(input_lo ^ input_hi ^ bitflipl, XXH_PRIME64_1); - /* - * Put len in the middle of m128 to ensure that the length gets mixed to - * both the low and high bits in the 128x64 multiply below. - */ - m128.low64 += (xxh_u64)(len - 1) << 54; - input_hi ^= bitfliph; - /* - * Add the high 32 bits of input_hi to the high 32 bits of m128, then - * add the long product of the low 32 bits of input_hi and XXH_PRIME32_2 to - * the high 64 bits of m128. - * - * The best approach to this operation is different on 32-bit and 64-bit. - */ - if (sizeof(void *) < sizeof(xxh_u64)) { /* 32-bit */ - /* - * 32-bit optimized version, which is more readable. - * - * On 32-bit, it removes an ADC and delays a dependency between the two - * halves of m128.high64, but it generates an extra mask on 64-bit. - */ - m128.high64 += (input_hi & 0xFFFFFFFF00000000) + - XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2); - - } else { - - /* - * 64-bit optimized (albeit more confusing) version. - * - * Uses some properties of addition and multiplication to remove the mask: - * - * Let: - * a = input_hi.lo = (input_hi & 0x00000000FFFFFFFF) - * b = input_hi.hi = (input_hi & 0xFFFFFFFF00000000) - * c = XXH_PRIME32_2 - * - * a + (b * c) - * Inverse Property: x + y - x == y - * a + (b * (1 + c - 1)) - * Distributive Property: x * (y + z) == (x * y) + (x * z) - * a + (b * 1) + (b * (c - 1)) - * Identity Property: x * 1 == x - * a + b + (b * (c - 1)) - * - * Substitute a, b, and c: - * input_hi.hi + input_hi.lo + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - - * 1)) - * - * Since input_hi.hi + input_hi.lo == input_hi, we get this: - * input_hi + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1)) - */ - m128.high64 += - input_hi + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2 - 1); - - } - - /* m128 ^= XXH_swap64(m128 >> 64); */ - m128.low64 ^= XXH_swap64(m128.high64); - - { /* 128x64 multiply: h128 = m128 * XXH_PRIME64_2; */ - XXH128_hash_t h128 = XXH_mult64to128(m128.low64, XXH_PRIME64_2); - h128.high64 += m128.high64 * XXH_PRIME64_2; - - h128.low64 = XXH3_avalanche(h128.low64); - h128.high64 = XXH3_avalanche(h128.high64); - return h128; - - } - - } - -} - -/* - * Assumption: `secret` size is >= XXH3_SECRET_SIZE_MIN - */ -XXH_FORCE_INLINE XXH128_hash_t XXH3_len_0to16_128b(const xxh_u8 *input, - size_t len, - const xxh_u8 *secret, - XXH64_hash_t seed) { - - XXH_ASSERT(len <= 16); - { - - if (len > 8) return XXH3_len_9to16_128b(input, len, secret, seed); - if (len >= 4) return XXH3_len_4to8_128b(input, len, secret, seed); - if (len) return XXH3_len_1to3_128b(input, len, secret, seed); - { - - XXH128_hash_t h128; - xxh_u64 const bitflipl = - XXH_readLE64(secret + 64) ^ XXH_readLE64(secret + 72); - xxh_u64 const bitfliph = - XXH_readLE64(secret + 80) ^ XXH_readLE64(secret + 88); - h128.low64 = XXH3_avalanche((XXH_PRIME64_1 + seed) ^ bitflipl); - h128.high64 = XXH3_avalanche((XXH_PRIME64_2 - seed) ^ bitfliph); - return h128; - - } - - } - -} - -/* - * A bit slower than XXH3_mix16B, but handles multiply by zero better. - */ -XXH_FORCE_INLINE XXH128_hash_t XXH128_mix32B(XXH128_hash_t acc, - const xxh_u8 *input_1, - const xxh_u8 *input_2, - const xxh_u8 *secret, - XXH64_hash_t seed) { - - acc.low64 += XXH3_mix16B(input_1, secret + 0, seed); - acc.low64 ^= XXH_readLE64(input_2) + XXH_readLE64(input_2 + 8); - acc.high64 += XXH3_mix16B(input_2, secret + 16, seed); - acc.high64 ^= XXH_readLE64(input_1) + XXH_readLE64(input_1 + 8); - return acc; - -} - -XXH_FORCE_INLINE XXH128_hash_t XXH3_len_17to128_128b( - const xxh_u8 *XXH_RESTRICT input, size_t len, - const xxh_u8 *XXH_RESTRICT secret, size_t secretSize, XXH64_hash_t seed) { - - XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); - (void)secretSize; - XXH_ASSERT(16 < len && len <= 128); - - { - - XXH128_hash_t acc; - acc.low64 = len * XXH_PRIME64_1; - acc.high64 = 0; - if (len > 32) { - - if (len > 64) { - - if (len > 96) { - - acc = XXH128_mix32B(acc, input + 48, input + len - 64, secret + 96, - seed); - - } - - acc = - XXH128_mix32B(acc, input + 32, input + len - 48, secret + 64, seed); - - } - - acc = XXH128_mix32B(acc, input + 16, input + len - 32, secret + 32, seed); - - } - - acc = XXH128_mix32B(acc, input, input + len - 16, secret, seed); - { - - XXH128_hash_t h128; - h128.low64 = acc.low64 + acc.high64; - h128.high64 = (acc.low64 * XXH_PRIME64_1) + (acc.high64 * XXH_PRIME64_4) + - ((len - seed) * XXH_PRIME64_2); - h128.low64 = XXH3_avalanche(h128.low64); - h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64); - return h128; - - } - - } - -} - -XXH_NO_INLINE XXH128_hash_t XXH3_len_129to240_128b( - const xxh_u8 *XXH_RESTRICT input, size_t len, - const xxh_u8 *XXH_RESTRICT secret, size_t secretSize, XXH64_hash_t seed) { - - XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); - (void)secretSize; - XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX); - - { - - XXH128_hash_t acc; - int const nbRounds = (int)len / 32; - int i; - acc.low64 = len * XXH_PRIME64_1; - acc.high64 = 0; - for (i = 0; i < 4; i++) { - - acc = XXH128_mix32B(acc, input + (32 * i), input + (32 * i) + 16, - secret + (32 * i), seed); - - } - - acc.low64 = XXH3_avalanche(acc.low64); - acc.high64 = XXH3_avalanche(acc.high64); - XXH_ASSERT(nbRounds >= 4); - for (i = 4; i < nbRounds; i++) { - - acc = XXH128_mix32B(acc, input + (32 * i), input + (32 * i) + 16, - secret + XXH3_MIDSIZE_STARTOFFSET + (32 * (i - 4)), - seed); - - } - - /* last bytes */ - acc = XXH128_mix32B( - acc, input + len - 16, input + len - 32, - secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET - 16, - 0ULL - seed); - - { - - XXH128_hash_t h128; - h128.low64 = acc.low64 + acc.high64; - h128.high64 = (acc.low64 * XXH_PRIME64_1) + (acc.high64 * XXH_PRIME64_4) + - ((len - seed) * XXH_PRIME64_2); - h128.low64 = XXH3_avalanche(h128.low64); - h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64); - return h128; - - } - - } - -} - -XXH_FORCE_INLINE XXH128_hash_t XXH3_hashLong_128b_internal( - const 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) { - - XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC; - - XXH3_hashLong_internal_loop(acc, input, len, secret, secretSize, - XXH3_acc_128bits, f_acc512, f_scramble); - - /* converge into final hash */ - XXH_STATIC_ASSERT(sizeof(acc) == 64); - XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START); - { - - XXH128_hash_t h128; - h128.low64 = XXH3_mergeAccs(acc, secret + XXH_SECRET_MERGEACCS_START, - (xxh_u64)len * XXH_PRIME64_1); - h128.high64 = XXH3_mergeAccs( - acc, secret + secretSize - sizeof(acc) - XXH_SECRET_MERGEACCS_START, - ~((xxh_u64)len * XXH_PRIME64_2)); - return h128; - - } - -} - -/* - * It's important for performance that XXH3_hashLong is not inlined. - */ -XXH_NO_INLINE XXH128_hash_t XXH3_hashLong_128b_defaultSecret( - const xxh_u8 *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_128b_internal(input, len, XXH3_kSecret, - sizeof(XXH3_kSecret), XXH3_accumulate_512, - XXH3_scrambleAcc); - -} - -/* - * It's important for performance that XXH3_hashLong is not inlined. - */ -XXH_NO_INLINE XXH128_hash_t XXH3_hashLong_128b_withSecret( - const xxh_u8 *XXH_RESTRICT input, size_t len, XXH64_hash_t seed64, - const xxh_u8 *XXH_RESTRICT secret, size_t secretLen) { - - (void)seed64; - return XXH3_hashLong_128b_internal(input, len, secret, secretLen, - XXH3_accumulate_512, XXH3_scrambleAcc); - -} - -XXH_FORCE_INLINE XXH128_hash_t XXH3_hashLong_128b_withSeed_internal( - const xxh_u8 *XXH_RESTRICT input, size_t len, XXH64_hash_t seed64, - XXH3_f_accumulate_512 f_acc512, XXH3_f_scrambleAcc f_scramble, - XXH3_f_initCustomSecret f_initSec) { - - if (seed64 == 0) - return XXH3_hashLong_128b_internal( - input, len, XXH3_kSecret, sizeof(XXH3_kSecret), f_acc512, f_scramble); - { - - XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE]; - f_initSec(secret, seed64); - return XXH3_hashLong_128b_internal(input, len, secret, sizeof(secret), - f_acc512, f_scramble); - - } - -} - -/* - * It's important for performance that XXH3_hashLong is not inlined. - */ -XXH_NO_INLINE XXH128_hash_t XXH3_hashLong_128b_withSeed( - const xxh_u8 *input, size_t len, XXH64_hash_t seed64, - const xxh_u8 *XXH_RESTRICT secret, size_t secretLen) { - - (void)secret; - (void)secretLen; - return XXH3_hashLong_128b_withSeed_internal( - input, len, seed64, XXH3_accumulate_512, XXH3_scrambleAcc, - XXH3_initCustomSecret); - -} - -typedef XXH128_hash_t (*XXH3_hashLong128_f)(const xxh_u8 *XXH_RESTRICT, size_t, - XXH64_hash_t, - const xxh_u8 *XXH_RESTRICT, size_t); - -XXH_FORCE_INLINE XXH128_hash_t -XXH3_128bits_internal(const void *input, size_t len, XXH64_hash_t seed64, - const xxh_u8 *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, secret, seed64); - if (len <= 128) - return XXH3_len_17to128_128b((const xxh_u8 *)input, len, secret, secretLen, - seed64); - if (len <= XXH3_MIDSIZE_MAX) - return XXH3_len_129to240_128b((const xxh_u8 *)input, len, secret, secretLen, - seed64); - return f_hl128((const xxh_u8 *)input, len, seed64, secret, secretLen); - -} - -/* === Public XXH128 API === */ - -XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void *input, size_t len) { - - return XXH3_128bits_internal(input, len, 0, XXH3_kSecret, - sizeof(XXH3_kSecret), - XXH3_hashLong_128b_withSecret); - -} - -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_defaultSecret); - -} - -XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSeed(const void * input, - size_t len, - XXH64_hash_t seed) { - - return XXH3_128bits_internal(input, len, seed, XXH3_kSecret, - sizeof(XXH3_kSecret), - XXH3_hashLong_128b_withSeed); - -} - -XXH_PUBLIC_API XXH128_hash_t XXH128(const void *input, size_t len, - XXH64_hash_t seed) { - - return XXH3_128bits_withSeed(input, len, seed); - -} - -/* === XXH3 128-bit streaming === */ - -/* - * All the functions are actually the same as for 64-bit streaming variant. - * The only difference is the finalizatiom routine. - */ - -static void XXH3_128bits_reset_internal(XXH3_state_t *statePtr, - XXH64_hash_t seed, const xxh_u8 *secret, - size_t secretSize) { - - XXH3_64bits_reset_internal(statePtr, seed, secret, secretSize); - -} - -XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH3_state_t *statePtr) { - - if (statePtr == NULL) return XXH_ERROR; - XXH3_128bits_reset_internal(statePtr, 0, XXH3_kSecret, - XXH_SECRET_DEFAULT_SIZE); - return XXH_OK; - -} - -XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret( - XXH3_state_t *statePtr, const void *secret, size_t secretSize) { - - if (statePtr == NULL) return XXH_ERROR; - XXH3_128bits_reset_internal(statePtr, 0, (const xxh_u8 *)secret, secretSize); - if (secret == NULL) return XXH_ERROR; - if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR; - return XXH_OK; - -} - -XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH3_state_t *statePtr, - XXH64_hash_t seed) { - - if (statePtr == NULL) return XXH_ERROR; - XXH3_128bits_reset_internal(statePtr, seed, XXH3_kSecret, - XXH_SECRET_DEFAULT_SIZE); - XXH3_initCustomSecret(statePtr->customSecret, seed); - statePtr->extSecret = NULL; - return XXH_OK; - -} - -XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update(XXH3_state_t *state, - const void * input, - size_t len) { - - return XXH3_update(state, (const xxh_u8 *)input, len, XXH3_acc_128bits, - XXH3_accumulate_512, XXH3_scrambleAcc); - -} - -XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest(const XXH3_state_t *state) { - - const unsigned char *const secret = - (state->extSecret == NULL) ? state->customSecret : state->extSecret; - if (state->totalLen > XXH3_MIDSIZE_MAX) { - - XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[XXH_ACC_NB]; - XXH3_digest_long(acc, state, secret, XXH3_acc_128bits); - XXH_ASSERT(state->secretLimit + XXH_STRIPE_LEN >= - sizeof(acc) + XXH_SECRET_MERGEACCS_START); - { - - XXH128_hash_t h128; - h128.low64 = XXH3_mergeAccs(acc, secret + XXH_SECRET_MERGEACCS_START, - (xxh_u64)state->totalLen * XXH_PRIME64_1); - h128.high64 = - XXH3_mergeAccs(acc, - secret + state->secretLimit + XXH_STRIPE_LEN - - sizeof(acc) - XXH_SECRET_MERGEACCS_START, - ~((xxh_u64)state->totalLen * XXH_PRIME64_2)); - return h128; - - } - - } - - /* len <= XXH3_MIDSIZE_MAX : short code */ - if (state->seed) - return XXH3_128bits_withSeed(state->buffer, (size_t)state->totalLen, - state->seed); - return XXH3_128bits_withSecret(state->buffer, (size_t)(state->totalLen), - secret, state->secretLimit + XXH_STRIPE_LEN); - -} - -/* 128-bit utility functions */ - -#include <string.h> /* memcmp, memcpy */ - -/* return : 1 is equal, 0 if different */ -XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2) { - - /* note : XXH128_hash_t is compact, it has no padding byte */ - return !(memcmp(&h1, &h2, sizeof(h1))); - -} - -/* This prototype is compatible with stdlib's qsort(). - * return : >0 if *h128_1 > *h128_2 - * <0 if *h128_1 < *h128_2 - * =0 if *h128_1 == *h128_2 */ -XXH_PUBLIC_API int XXH128_cmp(const void *h128_1, const void *h128_2) { - - XXH128_hash_t const h1 = *(const XXH128_hash_t *)h128_1; - XXH128_hash_t const h2 = *(const XXH128_hash_t *)h128_2; - int const hcmp = (h1.high64 > h2.high64) - (h2.high64 > h1.high64); - /* note : bets that, in most cases, hash values are different */ - if (hcmp) return hcmp; - return (h1.low64 > h2.low64) - (h2.low64 > h1.low64); - -} - -/*====== Canonical representation ======*/ -XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH128_canonical_t *dst, - XXH128_hash_t hash) { - - XXH_STATIC_ASSERT(sizeof(XXH128_canonical_t) == sizeof(XXH128_hash_t)); - if (XXH_CPU_LITTLE_ENDIAN) { - - hash.high64 = XXH_swap64(hash.high64); - hash.low64 = XXH_swap64(hash.low64); - - } - - memcpy(dst, &hash.high64, sizeof(hash.high64)); - memcpy((char *)dst + sizeof(hash.high64), &hash.low64, sizeof(hash.low64)); - -} - -XXH_PUBLIC_API XXH128_hash_t -XXH128_hashFromCanonical(const XXH128_canonical_t *src) { - - XXH128_hash_t h; - h.high64 = XXH_readBE64(src); - h.low64 = XXH_readBE64(src->digest + 8); - return h; - -} - -/* Pop our optimization override from above */ -#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \ - && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \ - && defined(__OPTIMIZE__) && \ - !defined(__OPTIMIZE_SIZE__) /* respect -O0 and -Os */ - #pragma GCC pop_options -#endif - -#endif /* XXH3_H_1397135465 */ - diff --git a/include/xxhash.h b/include/xxhash.h index 826f39bd..006d3f3d 100644 --- a/include/xxhash.h +++ b/include/xxhash.h @@ -197,6 +197,7 @@ extern "C" { #define XXH_CAT(A, B) A##B #define XXH_NAME2(A, B) XXH_CAT(A, B) #define XXH_versionNumber XXH_NAME2(XXH_NAMESPACE, XXH_versionNumber) + /* XXH32 */ #define XXH32 XXH_NAME2(XXH_NAMESPACE, XXH32) #define XXH32_createState XXH_NAME2(XXH_NAMESPACE, XXH32_createState) #define XXH32_freeState XXH_NAME2(XXH_NAMESPACE, XXH32_freeState) @@ -208,6 +209,7 @@ extern "C" { 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) @@ -219,14 +221,50 @@ extern "C" { 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 7 - #define XXH_VERSION_RELEASE 4 + #define XXH_VERSION_MINOR 8 + #define XXH_VERSION_RELEASE 0 #define XXH_VERSION_NUMBER \ (XXH_VERSION_MAJOR * 100 * 100 + XXH_VERSION_MINOR * 100 + \ XXH_VERSION_RELEASE) @@ -401,145 +439,56 @@ XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH64_canonical_t *dst, XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(const XXH64_canonical_t *src); - #endif /* XXH_NO_LONG_LONG */ - -#endif /* XXHASH_H_5627135585666179 */ - -#if defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742) - #define XXHASH_H_STATIC_13879238742 -/* **************************************************************************** - * This section contains declarations which are not guaranteed to remain stable. - * They may change in future versions, becoming incompatible with a different - * version of the library. - * These declarations should only be used with static linking. - * Never use them in association with dynamic linking! - ***************************************************************************** - */ +/*-********************************************************************** + * XXH3 64-bit variant + ************************************************************************/ -/* - * These definitions are only present to allow static allocation of an XXH - * state, for example, on the stack or in a struct. - * Never **ever** access members directly. +/* ************************************************************************ + * XXH3 is a new 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 + * + * In general, expect XXH3 to run about ~2x faster on large inputs and >3x + * faster on small ones compared to XXH64, though exact differences depend on + * the platform. + * + * The algorithm is portable: Like XXH32 and XXH64, it generates the same hash + * on all platforms. + * + * It benefits greatly from SIMD and 64-bit arithmetic, but does not require it. + * + * Almost all 32-bit and 64-bit targets that can run XXH32 smoothly can run + * XXH3 at competitive speeds, even if XXH64 runs slowly. 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 with the XXH_VECTOR macro. + * + * XXH3 offers 2 variants, _64bits and _128bits. + * When only 64 bits are needed, prefer calling the _64bits variant, as it + * reduces the amount of mixing, resulting in faster speed on small inputs. + * + * It's also generally simpler to manipulate a scalar return type than a struct. + * + * The 128-bit version adds additional strength, but it is slightly slower. + * + * Return values of XXH3 and XXH128 are officially finalized starting + * with v0.8.0 and will no longer change in future versions. + * Avoid storing values from before that release in long-term storage. + * + * Results produced by v0.7.x are not comparable with results from v0.7.y. + * However, the API is completely stable, and it can safely be used for + * ephemeral data (local sessions). + * + * The API supports one-shot hashing, streaming mode, and custom secrets. */ -struct XXH32_state_s { - - XXH32_hash_t total_len_32; - XXH32_hash_t large_len; - XXH32_hash_t v1; - XXH32_hash_t v2; - XXH32_hash_t v3; - XXH32_hash_t v4; - XXH32_hash_t mem32[4]; - XXH32_hash_t memsize; - XXH32_hash_t - reserved; /* never read nor write, might be removed in a future version */ - -}; /* typedef'd to XXH32_state_t */ - - #ifndef XXH_NO_LONG_LONG /* defined when there is no 64-bit support */ - -struct XXH64_state_s { - - XXH64_hash_t total_len; - XXH64_hash_t v1; - XXH64_hash_t v2; - XXH64_hash_t v3; - XXH64_hash_t v4; - XXH64_hash_t mem64[4]; - XXH32_hash_t memsize; - XXH32_hash_t reserved32; /* required for padding anyway */ - XXH64_hash_t reserved64; /* never read nor write, might be removed in a future - version */ - -}; /* typedef'd to XXH64_state_t */ - - /*-********************************************************************** - * XXH3 - * New experimental hash - ************************************************************************/ - - /* ************************************************************************ - * XXH3 is a new 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 - * - * In general, expect XXH3 to run about ~2x faster on large inputs and >3x - * faster on small ones compared to XXH64, though exact differences depend on - * the platform. - * - * The algorithm is portable: Like XXH32 and XXH64, it generates the same hash - * on all platforms. - * - * It benefits greatly from SIMD and 64-bit arithmetic, but does not require - * it. - * - * Almost all 32-bit and 64-bit targets that can run XXH32 smoothly can run - * XXH3 at competitive speeds, even if XXH64 runs slowly. 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 with the - * XXH_VECTOR macro. - * - * XXH3 offers 2 variants, _64bits and _128bits. - * When only 64 bits are needed, prefer calling the _64bits variant, as it - * reduces the amount of mixing, resulting in faster speed on small inputs. - * - * It's also generally simpler to manipulate a scalar return type than a - * struct. - * - * The 128-bit version adds additional strength, but it is slightly slower. - * - * The XXH3 algorithm is still in development. - * The results it produces may still change in future versions. - * - * Results produced by v0.7.x are not comparable with results from v0.7.y. - * However, the API is completely stable, and it can safely be used for - * ephemeral data (local sessions). - * - * Avoid storing values in long-term storage until the algorithm is finalized. - * - * Since v0.7.3, XXH3 has reached "release candidate" status, meaning that, if - * everything remains fine, its current format will be "frozen" and become the - * final one. - * - * After which, return values of XXH3 and XXH128 will no longer change in - * future versions. - * - * XXH3's return values will be officially finalized upon reaching v0.8.0. - * - * The API supports one-shot hashing, streaming mode, and custom secrets. - */ - - #ifdef XXH_NAMESPACE - #define XXH3_64bits XXH_NAME2(XXH_NAMESPACE, XXH3_64bits) - #define XXH3_64bits_withSecret \ - XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecret) - #define XXH3_64bits_withSeed \ - XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSeed) - - #define XXH3_createState XXH_NAME2(XXH_NAMESPACE, XXH3_createState) - #define XXH3_freeState XXH_NAME2(XXH_NAMESPACE, XXH3_freeState) - #define XXH3_copyState XXH_NAME2(XXH_NAMESPACE, XXH3_copyState) - - #define XXH3_64bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset) - #define XXH3_64bits_reset_withSeed \ - XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSeed) - #define XXH3_64bits_reset_withSecret \ - XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecret) - #define XXH3_64bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_update) - #define XXH3_64bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_digest) - - #define XXH3_generateSecret XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret) - #endif - /* XXH3_64bits(): * default 64-bit variant, using default secret and default seed of 0. * It's the fastest variant. */ @@ -547,8 +496,8 @@ XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(const void *data, size_t len); /* * XXH3_64bits_withSeed(): - * This variant generates a custom secret on the fly based on the default - * secret, altered using the `seed` value. + * This variant generates a custom secret on the fly + * based on default secret altered using the `seed` value. * While this operation is decently fast, note that it's not completely free. * Note: seed==0 produces the same results as XXH3_64bits(). */ @@ -559,74 +508,28 @@ XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSeed(const void *data, size_t len, * XXH3_64bits_withSecret(): * It's possible to provide any blob of bytes as a "secret" to generate the * hash. This makes it more difficult for an external actor to prepare an - * intentional collision. secretSize *must* be large enough (>= - * XXH3_SECRET_SIZE_MIN). The hash quality depends on the secret's high - * entropy, meaning that the secret should look like a bunch of random - * bytes. Avoid "trivial" sequences such as text or a bunch of repeated - * characters. If you are unsure of the "randonmess" of the blob of bytes, - * consider making it a "custom seed" instead, - * and use "XXH_generateSecret()" to generate a high quality secret. + * intentional collision. The main condition is that secretSize *must* be + * large enough (>= XXH3_SECRET_SIZE_MIN). However, the quality of produced + * hash values depends on secret's entropy. Technically, the secret must + * look like a bunch of random bytes. Avoid "trivial" or structured data + * such as repeated sequences or a text document. Whenever unsure about the + * "randomness" of the blob of bytes, consider relabelling it as a "custom + * seed" instead, and employ "XXH3_generateSecret()" (see below) to generate + * a high entropy secret derived from the custom seed. */ #define XXH3_SECRET_SIZE_MIN 136 XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSecret(const void *data, size_t len, const void *secret, size_t secretSize); - /* streaming 64-bit */ - - #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* C11+ */ - #include <stdalign.h> - #define XXH_ALIGN(n) alignas(n) - #elif defined(__GNUC__) - #define XXH_ALIGN(n) __attribute__((aligned(n))) - #elif defined(_MSC_VER) - #define XXH_ALIGN(n) __declspec(align(n)) - #else - #define XXH_ALIGN(n) /* disabled */ - #endif - - /* Old GCC versions only accept the attribute after the type in structures. - */ - #if !(defined(__STDC_VERSION__) && \ - (__STDC_VERSION__ >= 201112L)) /* C11+ */ \ - && defined(__GNUC__) - #define XXH_ALIGN_MEMBER(align, type) type XXH_ALIGN(align) - #else - #define XXH_ALIGN_MEMBER(align, type) XXH_ALIGN(align) type - #endif - -typedef struct XXH3_state_s XXH3_state_t; - - #define XXH3_INTERNALBUFFER_SIZE 256 - #define XXH3_SECRET_DEFAULT_SIZE 192 -struct XXH3_state_s { - - XXH_ALIGN_MEMBER(64, XXH64_hash_t acc[8]); - /* used to store a custom secret generated from a seed */ - XXH_ALIGN_MEMBER(64, unsigned char customSecret[XXH3_SECRET_DEFAULT_SIZE]); - XXH_ALIGN_MEMBER(64, unsigned char buffer[XXH3_INTERNALBUFFER_SIZE]); - XXH32_hash_t bufferedSize; - XXH32_hash_t reserved32; - size_t nbStripesPerBlock; - size_t nbStripesSoFar; - size_t secretLimit; - XXH64_hash_t totalLen; - XXH64_hash_t seed; - XXH64_hash_t reserved64; - const unsigned char *extSecret; /* reference to external secret; - * if == NULL, use .customSecret instead */ - /* note: there may be some padding at the end due to alignment on 64 bytes */ - -}; /* typedef'd to XXH3_state_t */ - - #undef XXH_ALIGN_MEMBER - +/******* Streaming *******/ /* * Streaming requires state maintenance. * This operation costs memory and CPU. * As a consequence, streaming is slower than one-shot hashing. - * For better performance, prefer one-shot functions whenever possible. + * For better performance, prefer one-shot functions whenever applicable. */ +typedef struct XXH3_state_s XXH3_state_t; XXH_PUBLIC_API XXH3_state_t *XXH3_createState(void); XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t *statePtr); XXH_PUBLIC_API void XXH3_copyState(XXH3_state_t * dst_state, @@ -634,8 +537,8 @@ XXH_PUBLIC_API void XXH3_copyState(XXH3_state_t * dst_state, /* * XXH3_64bits_reset(): - * Initialize with the default parameters. - * The result will be equivalent to `XXH3_64bits()`. + * Initialize with default parameters. + * digest will be equivalent to `XXH3_64bits()`. */ XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH3_state_t *statePtr); /* @@ -647,9 +550,12 @@ XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH3_state_t *statePtr, XXH64_hash_t seed); /* * XXH3_64bits_reset_withSecret(): - * `secret` is referenced, and must outlive the hash streaming session, so - * be careful when using stack arrays. - * `secretSize` must be >= `XXH3_SECRET_SIZE_MIN`. + * `secret` is referenced, it _must outlive_ the hash streaming session. + * Similar to one-shot API, `secretSize` must be >= `XXH3_SECRET_SIZE_MIN`, + * and the quality of produced hash values depends on secret's entropy + * (secret's content should look like a bunch of random bytes). + * When in doubt about the randomness of a candidate `secret`, + * consider employing `XXH3_generateSecret()` instead (see below). */ XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret( XXH3_state_t *statePtr, const void *secret, size_t secretSize); @@ -659,31 +565,12 @@ XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update(XXH3_state_t *statePtr, size_t length); XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest(const XXH3_state_t *statePtr); - /* 128-bit */ - - #ifdef XXH_NAMESPACE - #define XXH128 XXH_NAME2(XXH_NAMESPACE, XXH128) - #define XXH3_128bits XXH_NAME2(XXH_NAMESPACE, XXH3_128bits) - #define XXH3_128bits_withSeed \ - XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSeed) - #define XXH3_128bits_withSecret \ - XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecret) - - #define XXH3_128bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset) - #define XXH3_128bits_reset_withSeed \ - XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSeed) - #define XXH3_128bits_reset_withSecret \ - XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecret) - #define XXH3_128bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_update) - #define XXH3_128bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_digest) - - #define XXH128_isEqual XXH_NAME2(XXH_NAMESPACE, XXH128_isEqual) - #define XXH128_cmp XXH_NAME2(XXH_NAMESPACE, XXH128_cmp) - #define XXH128_canonicalFromHash \ - XXH_NAME2(XXH_NAMESPACE, XXH128_canonicalFromHash) - #define XXH128_hashFromCanonical \ - XXH_NAME2(XXH_NAMESPACE, XXH128_hashFromCanonical) - #endif +/* note : canonical representation of XXH3 is the same as XXH64 + * since they both produce XXH64_hash_t values */ + +/*-********************************************************************** + * XXH3 128-bit variant + ************************************************************************/ typedef struct { @@ -692,16 +579,28 @@ typedef struct { } XXH128_hash_t; -XXH_PUBLIC_API XXH128_hash_t XXH128(const void *data, size_t len, - XXH64_hash_t seed); XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void *data, size_t len); -XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSeed( - const void *data, size_t len, XXH64_hash_t seed); /* == XXH128() */ +XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSeed(const void *data, size_t len, + XXH64_hash_t seed); XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSecret(const void *data, size_t len, const void *secret, size_t secretSize); +/******* Streaming *******/ +/* + * Streaming requires state maintenance. + * This operation costs memory and CPU. + * As a consequence, streaming is slower than one-shot hashing. + * For better performance, prefer one-shot functions whenever applicable. + * + * XXH3_128bits uses the same XXH3_state_t as XXH3_64bits(). + * Use already declared XXH3_createState() and XXH3_freeState(). + * + * All reset and streaming functions have same meaning as their 64-bit + * counterpart. + */ + XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH3_state_t *statePtr); XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH3_state_t *statePtr, XXH64_hash_t seed); @@ -713,7 +612,10 @@ XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update(XXH3_state_t *statePtr, size_t length); XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest(const XXH3_state_t *statePtr); -/* Note: For better performance, these functions can be inlined using +/* 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 */ /*! @@ -745,6 +647,116 @@ XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH128_canonical_t *dst, XXH_PUBLIC_API XXH128_hash_t XXH128_hashFromCanonical(const XXH128_canonical_t *src); + #endif /* XXH_NO_LONG_LONG */ + +#endif /* XXHASH_H_5627135585666179 */ + +#if defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742) + #define XXHASH_H_STATIC_13879238742 +/* **************************************************************************** + * This section contains declarations which are not guaranteed to remain stable. + * They may change in future versions, becoming incompatible with a different + * version of the library. + * These declarations should only be used with static linking. + * Never use them in association with dynamic linking! + ***************************************************************************** + */ + +/* + * These definitions are only present to allow static allocation + * of XXH states, on stack or in a struct, for example. + * Never **ever** access their members directly. + */ + +struct XXH32_state_s { + + XXH32_hash_t total_len_32; + XXH32_hash_t large_len; + XXH32_hash_t v1; + XXH32_hash_t v2; + XXH32_hash_t v3; + XXH32_hash_t v4; + XXH32_hash_t mem32[4]; + XXH32_hash_t memsize; + XXH32_hash_t + reserved; /* never read nor write, might be removed in a future version */ + +}; /* typedef'd to XXH32_state_t */ + + #ifndef XXH_NO_LONG_LONG /* defined when there is no 64-bit support */ + +struct XXH64_state_s { + + XXH64_hash_t total_len; + XXH64_hash_t v1; + XXH64_hash_t v2; + XXH64_hash_t v3; + XXH64_hash_t v4; + XXH64_hash_t mem64[4]; + XXH32_hash_t memsize; + XXH32_hash_t reserved32; /* required for padding anyway */ + XXH64_hash_t reserved64; /* never read nor write, might be removed in a future + version */ + +}; /* typedef'd to XXH64_state_t */ + + #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* C11+ */ + #include <stdalign.h> + #define XXH_ALIGN(n) alignas(n) + #elif defined(__GNUC__) + #define XXH_ALIGN(n) __attribute__((aligned(n))) + #elif defined(_MSC_VER) + #define XXH_ALIGN(n) __declspec(align(n)) + #else + #define XXH_ALIGN(n) /* disabled */ + #endif + + /* Old GCC versions only accept the attribute after the type in structures. + */ + #if !(defined(__STDC_VERSION__) && \ + (__STDC_VERSION__ >= 201112L)) /* C11+ */ \ + && defined(__GNUC__) + #define XXH_ALIGN_MEMBER(align, type) type XXH_ALIGN(align) + #else + #define XXH_ALIGN_MEMBER(align, type) XXH_ALIGN(align) type + #endif + + #define XXH3_INTERNALBUFFER_SIZE 256 + #define XXH3_SECRET_DEFAULT_SIZE 192 +struct XXH3_state_s { + + XXH_ALIGN_MEMBER(64, XXH64_hash_t acc[8]); + /* used to store a custom secret generated from a seed */ + XXH_ALIGN_MEMBER(64, unsigned char customSecret[XXH3_SECRET_DEFAULT_SIZE]); + XXH_ALIGN_MEMBER(64, unsigned char buffer[XXH3_INTERNALBUFFER_SIZE]); + XXH32_hash_t bufferedSize; + XXH32_hash_t reserved32; + size_t nbStripesSoFar; + XXH64_hash_t totalLen; + size_t nbStripesPerBlock; + size_t secretLimit; + XXH64_hash_t seed; + XXH64_hash_t reserved64; + const unsigned char *extSecret; /* reference to external secret; + * if == NULL, use .customSecret instead */ + /* note: there may be some padding at the end due to alignment on 64 bytes */ + +}; /* typedef'd to XXH3_state_t */ + + #undef XXH_ALIGN_MEMBER + + /* When the XXH3_state_t structure is merely emplaced on stack, + * it should be initialized with XXH3_INITSTATE() or a memset() + * in case its first reset uses XXH3_NNbits_reset_withSeed(). + * This init can be omitted if the first reset uses default or _withSecret + * mode. This operation isn't necessary when the state is created with + * XXH3_createState(). Note that this doesn't prepare the state for a + * streaming operation, it's still necessary to use XXH3_NNbits_reset*() + * afterwards. + */ + #define XXH3_INITSTATE(XXH3_state_ptr) \ + { (XXH3_state_ptr)->seed = 0; } + /* === Experimental API === */ /* Symbols defined below must be considered tied to a specific library version. */ @@ -752,17 +764,19 @@ XXH128_hashFromCanonical(const XXH128_canonical_t *src); /* * XXH3_generateSecret(): * - * Derive a secret for use with `*_withSecret()` prototypes of XXH3. - * Use this if you need a higher level of security than the one provided by - * 64bit seed. + * 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. * - * Take as input a custom seed of any length and any content, - * generate from it a high-entropy secret of length XXH3_SECRET_DEFAULT_SIZE - * into already allocated buffer secretBuffer. - * The generated secret ALWAYS is XXH_SECRET_DEFAULT_SIZE bytes long. + * The function accepts as input a custom seed of any length and any content, + * and derives from it a high-entropy secret of length XXH3_SECRET_DEFAULT_SIZE + * into an already allocated buffer secretBuffer. + * The generated secret is _always_ XXH_SECRET_DEFAULT_SIZE bytes long. * * The generated secret can then be used with any `*_withSecret()` variant. - * The functions `XXH3_128bits_withSecret()`, `XXH3_64bits_withSecret()`, + * Functions `XXH3_128bits_withSecret()`, `XXH3_64bits_withSecret()`, * `XXH3_128bits_reset_withSecret()` and `XXH3_64bits_reset_withSecret()` * are part of this list. They all accept a `secret` parameter * which must be very long for implementation reasons (>= XXH3_SECRET_SIZE_MIN) @@ -771,8 +785,8 @@ XXH128_hashFromCanonical(const XXH128_canonical_t *src); * this function can be used to generate a secret of proper quality. * * customSeed can be anything. It can have any size, even small ones, - * and its content can be anything, even some "low entropy" source such as a - * bunch of zeroes. The resulting `secret` will nonetheless respect all expected + * and its content can be anything, even stupidly "low entropy" source such as a + * bunch of zeroes. The resulting `secret` will nonetheless provide all expected * qualities. * * Supplying NULL as the customSeed copies the default secret into @@ -783,6 +797,10 @@ XXH_PUBLIC_API void XXH3_generateSecret(void * secretBuffer, const void *customSeed, size_t customSeedSize); +/* simple short-cut to pre-selected XXH3_128bits variant */ +XXH_PUBLIC_API XXH128_hash_t XXH128(const void *data, size_t len, + XXH64_hash_t seed); + #endif /* XXH_NO_LONG_LONG */ #if defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) @@ -799,17 +817,23 @@ XXH_PUBLIC_API void XXH3_generateSecret(void * secretBuffer, /*-********************************************************************** * xxHash implementation *-********************************************************************** - * xxHash's implementation used to be found in xxhash.c. + * xxHash's implementation used to be hosted inside xxhash.c. * - * However, code inlining requires the implementation to be visible to the - * compiler, usually within the header. + * However, inlining requires implementation to be visible to the compiler, + * hence be included alongside the header. + * Previously, implementation was hosted inside xxhash.c, + * which was then #included when inlining was activated. + * This construction created issues with a few build and install systems, + * as it required xxhash.c to be stored in /include directory. * - * As a workaround, xxhash.c used to be included within xxhash.h. This caused - * some issues with some build systems, especially ones which treat .c files - * as source files. + * xxHash implementation is now directly integrated within xxhash.h. + * As a consequence, xxhash.c is no longer needed in /include. * - * Therefore, the implementation is now directly integrated within xxhash.h. - * Another small advantage is that xxhash.c is no longer needed in /include. + * xxhash.c is still available and is still useful. + * In a "normal" setup, when xxhash is not inlined, + * xxhash.h only exposes the prototypes and public symbols, + * while xxhash.c can be built into an object file xxhash.o + * which can then be linked into the final binary. ************************************************************************/ #if (defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) || \ @@ -828,10 +852,10 @@ XXH_PUBLIC_API void XXH3_generateSecret(void * secretBuffer, * Unfortunately, on some target/compiler combinations, the generated assembly * is sub-optimal. * - * The below switch allow to select a different access method for improved - * performance. + * The below switch allow selection of a different access method + * in the search for improved performance. * Method 0 (default): - * Use `memcpy()`. Safe and portable. + * Use `memcpy()`. Safe and portable. Default. * Method 1: * `__attribute__((packed))` statement. It depends on compiler extensions * and is therefore not portable. @@ -843,7 +867,7 @@ XXH_PUBLIC_API void XXH3_generateSecret(void * secretBuffer, * It can generate buggy code on targets which do not support unaligned * memory accesses. * But in some circumstances, it's the only known way to get the most - * performance (ie GCC + ARMv6) + * performance (example: GCC + ARMv6) * Method 3: * Byteshift. This can generate the best code on old compilers which don't * inline small `memcpy()` calls, and it might also be faster on @@ -924,7 +948,8 @@ XXH_PUBLIC_API void XXH3_generateSecret(void * secretBuffer, * -fno-inline with GCC or Clang, this will automatically be defined. */ #ifndef XXH_NO_INLINE_HINTS - #if defined(__OPTIMIZE_SIZE__) || defined(__NO_INLINE__) + #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 @@ -950,8 +975,8 @@ XXH_PUBLIC_API void XXH3_generateSecret(void * secretBuffer, * Includes & Memory related functions ***************************************/ /*! - * Modify the local functions below should you wish to use some other memory - * routines for malloc() and free() + * Modify the local functions below should you wish to use + * different memory routines for malloc() and free() */ #include <stdlib.h> @@ -1137,7 +1162,8 @@ typedef enum { XXH_bigEndian = 0, XXH_littleEndian = 1 } XXH_endianess; * Try to detect endianness automatically, to avoid the nonstandard behavior * in `XXH_isLittleEndian()` */ - #if defined(_WIN32) || defined(__LITTLE_ENDIAN__) || \ + #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__) || \ @@ -1778,13 +1804,16 @@ typedef XXH64_hash_t xxh_u64; * rerolled. */ #ifndef XXH_REROLL_XXH64 - #if (defined(__ILP32__) || defined(_ILP32)) || \ - !(defined(__x86_64__) || defined(_M_X64) || defined(_M_AMD64) || \ - defined(_M_ARM64) || defined(__aarch64__) || defined(__arm64__) || \ - defined(__PPC64__) || defined(__PPC64LE__) || \ - defined(__ppc64__) || defined(__powerpc64__) || \ - defined(__mips64__) || defined(__mips64)) || \ - (!defined(SIZE_MAX) || SIZE_MAX < ULLONG_MAX) + #if (defined(__ILP32__) || \ + defined(_ILP32)) /* ILP32 is often defined on 32-bit GCC family */ \ + || !(defined(__x86_64__) || defined(_M_X64) || \ + defined(_M_AMD64) /* x86-64 */ \ + || defined(_M_ARM64) || defined(__aarch64__) || \ + defined(__arm64__) /* aarch64 */ \ + || defined(__PPC64__) || defined(__PPC64LE__) || \ + defined(__ppc64__) || defined(__powerpc64__) /* ppc64 */ \ + || defined(__mips64__) || defined(__mips64)) /* mips64 */ \ + || (!defined(SIZE_MAX) || SIZE_MAX < ULLONG_MAX) /* check limits */ #define XXH_REROLL_XXH64 1 #else #define XXH_REROLL_XXH64 0 @@ -2428,7 +2457,3134 @@ XXH64_hashFromCanonical(const XXH64_canonical_t *src) { * New generation hash designed for speed on small keys and vectorization ************************************************************************ */ - #include "xxh3.h" + /* === Compiler specifics === */ + + #if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L /* >= C99 */ + #define XXH_RESTRICT restrict + #else + /* Note: it might be useful to define __restrict or __restrict__ for some + * C++ compilers */ + #define XXH_RESTRICT /* disable */ + #endif + + #if (defined(__GNUC__) && (__GNUC__ >= 3)) || \ + (defined(__INTEL_COMPILER) && (__INTEL_COMPILER >= 800)) || \ + defined(__clang__) + #define XXH_likely(x) __builtin_expect(x, 1) + #define XXH_unlikely(x) __builtin_expect(x, 0) + #else + #define XXH_likely(x) (x) + #define XXH_unlikely(x) (x) + #endif + + #if defined(__GNUC__) + #if defined(__AVX2__) + #include <immintrin.h> + #elif defined(__SSE2__) + #include <emmintrin.h> + #elif defined(__ARM_NEON__) || defined(__ARM_NEON) + #define inline __inline__ /* circumvent a clang bug */ + #include <arm_neon.h> + #undef inline + #endif + #elif defined(_MSC_VER) + #include <intrin.h> + #endif + + /* + * One goal of XXH3 is to make it fast on both 32-bit and 64-bit, while + * remaining a true 64-bit/128-bit hash function. + * + * This is done by prioritizing a subset of 64-bit operations that can be + * emulated without too many steps on the average 32-bit machine. + * + * For example, these two lines seem similar, and run equally fast on + * 64-bit: + * + * xxh_u64 x; + * x ^= (x >> 47); // good + * x ^= (x >> 13); // bad + * + * However, to a 32-bit machine, there is a major difference. + * + * x ^= (x >> 47) looks like this: + * + * x.lo ^= (x.hi >> (47 - 32)); + * + * while x ^= (x >> 13) looks like this: + * + * // note: funnel shifts are not usually cheap. + * x.lo ^= (x.lo >> 13) | (x.hi << (32 - 13)); + * x.hi ^= (x.hi >> 13); + * + * The first one is significantly faster than the second, simply because the + * shift is larger than 32. This means: + * - All the bits we need are in the upper 32 bits, so we can ignore the + * lower 32 bits in the shift. + * - The shift result will always fit in the lower 32 bits, and therefore, + * we can ignore the upper 32 bits in the xor. + * + * Thanks to this optimization, XXH3 only requires these features to be + * efficient: + * + * - Usable unaligned access + * - A 32-bit or 64-bit ALU + * - If 32-bit, a decent ADC instruction + * - A 32 or 64-bit multiply with a 64-bit result + * - For the 128-bit variant, a decent byteswap helps short inputs. + * + * The first two are already required by XXH32, and almost all 32-bit and + * 64-bit platforms which can run XXH32 can run XXH3 efficiently. + * + * Thumb-1, the classic 16-bit only subset of ARM's instruction set, is one + * notable exception. + * + * First of all, Thumb-1 lacks support for the UMULL instruction which + * performs the important long multiply. This means numerous __aeabi_lmul + * calls. + * + * Second of all, the 8 functional registers are just not enough. + * Setup for __aeabi_lmul, byteshift loads, pointers, and all arithmetic + * need Lo registers, and this shuffling results in thousands more MOVs than + * A32. + * + * A32 and T32 don't have this limitation. They can access all 14 registers, + * do a 32->64 multiply with UMULL, and the flexible operand allowing free + * shifts is helpful, too. + * + * Therefore, we do a quick sanity check. + * + * If compiling Thumb-1 for a target which supports ARM instructions, we + * will emit a warning, as it is not a "sane" platform to compile for. + * + * Usually, if this happens, it is because of an accident and you probably + * need to specify -march, as you likely meant to compile for a newer + * architecture. + * + * Credit: large sections of the vectorial and asm source code paths + * have been contributed by @easyaspi314 + */ + #if defined(__thumb__) && !defined(__thumb2__) && \ + defined(__ARM_ARCH_ISA_ARM) + #warning "XXH3 is highly inefficient without ARM or Thumb-2." + #endif + + /* ========================================== + * Vectorization detection + * ========================================== */ + #define XXH_SCALAR 0 /* Portable scalar version */ + #define XXH_SSE2 1 /* SSE2 for Pentium 4 and all x86_64 */ + #define XXH_AVX2 2 /* AVX2 for Haswell and Bulldozer */ + #define XXH_AVX512 3 /* AVX512 for Skylake and Icelake */ + #define XXH_NEON 4 /* NEON for most ARMv7-A and all AArch64 */ + #define XXH_VSX 5 /* VSX and ZVector for POWER8/z13 */ + + #ifndef XXH_VECTOR /* can be defined on command line */ + #if defined(__AVX512F__) + #define XXH_VECTOR XXH_AVX512 + #elif defined(__AVX2__) + #define XXH_VECTOR XXH_AVX2 + #elif defined(__SSE2__) || defined(_M_AMD64) || defined(_M_X64) || \ + (defined(_M_IX86_FP) && (_M_IX86_FP == 2)) + #define XXH_VECTOR XXH_SSE2 + #elif defined(__GNUC__) /* msvc support maybe later */ \ + && (defined(__ARM_NEON__) || defined(__ARM_NEON)) && \ + (defined(__LITTLE_ENDIAN__) /* We only support little endian NEON */ \ + || (defined(__BYTE_ORDER__) && \ + __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)) + #define XXH_VECTOR XXH_NEON + #elif (defined(__PPC64__) && defined(__POWER8_VECTOR__)) || \ + (defined(__s390x__) && defined(__VEC__)) && \ + defined(__GNUC__) /* TODO: IBM XL */ + #define XXH_VECTOR XXH_VSX + #else + #define XXH_VECTOR XXH_SCALAR + #endif + #endif + + /* + * Controls the alignment of the accumulator, + * for compatibility with aligned vector loads, which are usually faster. + */ + #ifndef XXH_ACC_ALIGN + #if defined(XXH_X86DISPATCH) + #define XXH_ACC_ALIGN 64 /* for compatibility with avx512 */ + #elif XXH_VECTOR == XXH_SCALAR /* scalar */ + #define XXH_ACC_ALIGN 8 + #elif XXH_VECTOR == XXH_SSE2 /* sse2 */ + #define XXH_ACC_ALIGN 16 + #elif XXH_VECTOR == XXH_AVX2 /* avx2 */ + #define XXH_ACC_ALIGN 32 + #elif XXH_VECTOR == XXH_NEON /* neon */ + #define XXH_ACC_ALIGN 16 + #elif XXH_VECTOR == XXH_VSX /* vsx */ + #define XXH_ACC_ALIGN 16 + #elif XXH_VECTOR == XXH_AVX512 /* avx512 */ + #define XXH_ACC_ALIGN 64 + #endif + #endif + + #if defined(XXH_X86DISPATCH) || XXH_VECTOR == XXH_SSE2 || \ + XXH_VECTOR == XXH_AVX2 || XXH_VECTOR == XXH_AVX512 + #define XXH_SEC_ALIGN XXH_ACC_ALIGN + #else + #define XXH_SEC_ALIGN 8 + #endif + + /* + * UGLY HACK: + * GCC usually generates the best code with -O3 for xxHash. + * + * However, when targeting AVX2, it is overzealous in its unrolling + * resulting in code roughly 3/4 the speed of Clang. + * + * There are other issues, such as GCC splitting _mm256_loadu_si256 into + * _mm_loadu_si128 + _mm256_inserti128_si256. This is an optimization which + * only applies to Sandy and Ivy Bridge... which don't even support AVX2. + * + * That is why when compiling the AVX2 version, it is recommended to use + * either -O2 -mavx2 -march=haswell or -O2 -mavx2 + * -mno-avx256-split-unaligned-load for decent performance, or to use Clang + * instead. + * + * Fortunately, we can control the first one with a pragma that forces GCC + * into -O2, but the other one we can't control without "failed to inline + * always inline function due to target mismatch" warnings. + */ + #if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \ + && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \ + && defined(__OPTIMIZE__) && \ + !defined(__OPTIMIZE_SIZE__) /* respect -O0 and -Os */ + #pragma GCC push_options + #pragma GCC optimize("-O2") + #endif + + #if XXH_VECTOR == XXH_NEON + /* + * NEON's setup for vmlal_u32 is a little more complicated than it is on + * SSE2, AVX2, and VSX. + * + * While PMULUDQ and VMULEUW both perform a mask, VMLAL.U32 performs an + * upcast. + * + * To do the same operation, the 128-bit 'Q' register needs to be split + * into two 64-bit 'D' registers, performing this operation:: + * + * [ a | b ] | + * '---------. .--------' | | x | + * | .---------' '--------. | + * [ a & 0xFFFFFFFF | b & 0xFFFFFFFF ],[ a >> 32 | b >> 32 ] + * + * Due to significant changes in aarch64, the fastest method for aarch64 + * is completely different than the fastest method for ARMv7-A. + * + * ARMv7-A treats D registers as unions overlaying Q registers, so + * modifying D11 will modify the high half of Q5. This is similar to how + * modifying AH will only affect bits 8-15 of AX on x86. + * + * VZIP takes two registers, and puts even lanes in one register and odd + * lanes in the other. + * + * On ARMv7-A, this strangely modifies both parameters in place instead of + * taking the usual 3-operand form. + * + * Therefore, if we want to do this, we can simply use a D-form VZIP.32 on + * the lower and upper halves of the Q register to end up with the high + * and low halves where we want - all in one instruction. + * + * vzip.32 d10, d11 @ d10 = { d10[0], d11[0] }; d11 = { d10[1], + * d11[1] } + * + * Unfortunately we need inline assembly for this: Instructions modifying + * two registers at once is not possible in GCC or Clang's IR, and they + * have to create a copy. + * + * aarch64 requires a different approach. + * + * In order to make it easier to write a decent compiler for aarch64, many + * quirks were removed, such as conditional execution. + * + * NEON was also affected by this. + * + * aarch64 cannot access the high bits of a Q-form register, and writes to + * a D-form register zero the high bits, similar to how writes to W-form + * scalar registers (or DWORD registers on x86_64) work. + * + * The formerly free vget_high intrinsics now require a vext (with a few + * exceptions) + * + * Additionally, VZIP was replaced by ZIP1 and ZIP2, which are the + * equivalent of PUNPCKL* and PUNPCKH* in SSE, respectively, in order to + * only modify one operand. + * + * The equivalent of the VZIP.32 on the lower and upper halves would be + * this mess: + * + * ext v2.4s, v0.4s, v0.4s, #2 // v2 = { v0[2], v0[3], v0[0], v0[1] + * } zip1 v1.2s, v0.2s, v2.2s // v1 = { v0[0], v2[0] } zip2 v0.2s, + * v0.2s, v1.2s // v0 = { v0[1], v2[1] } + * + * Instead, we use a literal downcast, vmovn_u64 (XTN), and vshrn_n_u64 + * (SHRN): + * + * shrn v1.2s, v0.2d, #32 // v1 = (uint32x2_t)(v0 >> 32); + * xtn v0.2s, v0.2d // v0 = (uint32x2_t)(v0 & 0xFFFFFFFF); + * + * This is available on ARMv7-A, but is less efficient than a single + * VZIP.32. + */ + + /* + * Function-like macro: + * void XXH_SPLIT_IN_PLACE(uint64x2_t &in, uint32x2_t &outLo, uint32x2_t + * &outHi) + * { + + * outLo = (uint32x2_t)(in & 0xFFFFFFFF); + * outHi = (uint32x2_t)(in >> 32); + * in = UNDEFINED; + * } + */ + #if !defined(XXH_NO_VZIP_HACK) /* define to disable */ \ + && defined(__GNUC__) && !defined(__aarch64__) && !defined(__arm64__) + #define XXH_SPLIT_IN_PLACE(in, outLo, outHi) \ + do { \ + \ + /* Undocumented GCC/Clang operand modifier: %e0 = lower D half, \ + * %f0 = upper D half */ \ + /* https://github.com/gcc-mirror/gcc/blob/38cf91e5/gcc/config/arm/arm.c#L22486 \ + */ \ + /* https://github.com/llvm-mirror/llvm/blob/2c4ca683/lib/Target/ARM/ARMAsmPrinter.cpp#L399 \ + */ \ + __asm__("vzip.32 %e0, %f0" : "+w"(in)); \ + (outLo) = vget_low_u32(vreinterpretq_u32_u64(in)); \ + (outHi) = vget_high_u32(vreinterpretq_u32_u64(in)); \ + \ + } while (0) + + #else + #define XXH_SPLIT_IN_PLACE(in, outLo, outHi) \ + do { \ + \ + (outLo) = vmovn_u64(in); \ + (outHi) = vshrn_n_u64((in), 32); \ + \ + } while (0) + + #endif + #endif /* XXH_VECTOR == XXH_NEON */ + + /* + * VSX and Z Vector helpers. + * + * This is very messy, and any pull requests to clean this up are welcome. + * + * There are a lot of problems with supporting VSX and s390x, due to + * inconsistent intrinsics, spotty coverage, and multiple endiannesses. + */ + #if XXH_VECTOR == XXH_VSX + #if defined(__s390x__) + #include <s390intrin.h> + #else + /* gcc's altivec.h can have the unwanted consequence to unconditionally + * #define bool, vector, and pixel keywords, + * with bad consequences for programs already using these keywords for + * other purposes. The paragraph defining these macros is skipped when + * __APPLE_ALTIVEC__ is defined. + * __APPLE_ALTIVEC__ is _generally_ defined automatically by the + * compiler, but it seems that, in some cases, it isn't. Force the build + * macro to be defined, so that keywords are not altered. + */ + #if defined(__GNUC__) && !defined(__APPLE_ALTIVEC__) + #define __APPLE_ALTIVEC__ + #endif + #include <altivec.h> + #endif + +typedef __vector unsigned long long xxh_u64x2; +typedef __vector unsigned char xxh_u8x16; +typedef __vector unsigned xxh_u32x4; + + #ifndef XXH_VSX_BE + #if defined(__BIG_ENDIAN__) || \ + (defined(__BYTE_ORDER__) && \ + __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) + #define XXH_VSX_BE 1 + #elif defined(__VEC_ELEMENT_REG_ORDER__) && \ + __VEC_ELEMENT_REG_ORDER__ == __ORDER_BIG_ENDIAN__ + #warning \ + "-maltivec=be is not recommended. Please use native endianness." + #define XXH_VSX_BE 1 + #else + #define XXH_VSX_BE 0 + #endif + #endif /* !defined(XXH_VSX_BE) */ + + #if XXH_VSX_BE + /* A wrapper for POWER9's vec_revb. */ + #if defined(__POWER9_VECTOR__) || \ + (defined(__clang__) && defined(__s390x__)) + #define XXH_vec_revb vec_revb + #else +XXH_FORCE_INLINE xxh_u64x2 XXH_vec_revb(xxh_u64x2 val) { + + xxh_u8x16 const vByteSwap = {0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00, + 0x0F, 0x0E, 0x0D, 0x0C, 0x0B, 0x0A, 0x09, 0x08}; + return vec_perm(val, val, vByteSwap); + +} + + #endif + #endif /* XXH_VSX_BE */ + +/* + * Performs an unaligned load and byte swaps it on big endian. + */ +XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(const void *ptr) { + + xxh_u64x2 ret; + memcpy(&ret, ptr, sizeof(xxh_u64x2)); + #if XXH_VSX_BE + ret = XXH_vec_revb(ret); + #endif + return ret; + +} + + /* + * vec_mulo and vec_mule are very problematic intrinsics on PowerPC + * + * These intrinsics weren't added until GCC 8, despite existing for a + * while, and they are endian dependent. Also, their meaning swap + * depending on version. + * */ + #if defined(__s390x__) + /* s390x is always big endian, no issue on this platform */ + #define XXH_vec_mulo vec_mulo + #define XXH_vec_mule vec_mule + #elif defined(__clang__) && XXH_HAS_BUILTIN(__builtin_altivec_vmuleuw) + /* Clang has a better way to control this, we can just use the builtin + * which doesn't swap. */ + #define XXH_vec_mulo __builtin_altivec_vmulouw + #define XXH_vec_mule __builtin_altivec_vmuleuw + #else +/* gcc needs inline assembly */ +/* Adapted from + * https://github.com/google/highwayhash/blob/master/highwayhash/hh_vsx.h. */ +XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mulo(xxh_u32x4 a, xxh_u32x4 b) { + + xxh_u64x2 result; + __asm__("vmulouw %0, %1, %2" : "=v"(result) : "v"(a), "v"(b)); + return result; + +} + +XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mule(xxh_u32x4 a, xxh_u32x4 b) { + + xxh_u64x2 result; + __asm__("vmuleuw %0, %1, %2" : "=v"(result) : "v"(a), "v"(b)); + return result; + +} + + #endif /* XXH_vec_mulo, XXH_vec_mule */ + #endif /* XXH_VECTOR == XXH_VSX */ + + /* prefetch + * can be disabled, by declaring XXH_NO_PREFETCH build macro */ + #if defined(XXH_NO_PREFETCH) + #define XXH_PREFETCH(ptr) (void)(ptr) /* disabled */ + #else + #if defined(_MSC_VER) && \ + (defined(_M_X64) || \ + defined( \ + _M_I86)) /* _mm_prefetch() is not defined outside of x86/x64 */ + #include <mmintrin.h> /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */ + #define XXH_PREFETCH(ptr) _mm_prefetch((const char *)(ptr), _MM_HINT_T0) + #elif defined(__GNUC__) && \ + ((__GNUC__ >= 4) || ((__GNUC__ == 3) && (__GNUC_MINOR__ >= 1))) + #define XXH_PREFETCH(ptr) \ + __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */) + #else + #define XXH_PREFETCH(ptr) (void)(ptr) /* disabled */ + #endif + #endif /* XXH_NO_PREFETCH */ + + /* ========================================== + * XXH3 default settings + * ========================================== */ + + #define XXH_SECRET_DEFAULT_SIZE 192 /* minimum XXH3_SECRET_SIZE_MIN */ + + #if (XXH_SECRET_DEFAULT_SIZE < XXH3_SECRET_SIZE_MIN) + #error "default keyset is not large enough" + #endif + +/* Pseudorandom secret taken directly from FARSH */ +XXH_ALIGN(64) +static const xxh_u8 XXH3_kSecret[XXH_SECRET_DEFAULT_SIZE] = { + + 0xb8, 0xfe, 0x6c, 0x39, 0x23, 0xa4, 0x4b, 0xbe, 0x7c, 0x01, 0x81, 0x2c, + 0xf7, 0x21, 0xad, 0x1c, 0xde, 0xd4, 0x6d, 0xe9, 0x83, 0x90, 0x97, 0xdb, + 0x72, 0x40, 0xa4, 0xa4, 0xb7, 0xb3, 0x67, 0x1f, 0xcb, 0x79, 0xe6, 0x4e, + 0xcc, 0xc0, 0xe5, 0x78, 0x82, 0x5a, 0xd0, 0x7d, 0xcc, 0xff, 0x72, 0x21, + 0xb8, 0x08, 0x46, 0x74, 0xf7, 0x43, 0x24, 0x8e, 0xe0, 0x35, 0x90, 0xe6, + 0x81, 0x3a, 0x26, 0x4c, 0x3c, 0x28, 0x52, 0xbb, 0x91, 0xc3, 0x00, 0xcb, + 0x88, 0xd0, 0x65, 0x8b, 0x1b, 0x53, 0x2e, 0xa3, 0x71, 0x64, 0x48, 0x97, + 0xa2, 0x0d, 0xf9, 0x4e, 0x38, 0x19, 0xef, 0x46, 0xa9, 0xde, 0xac, 0xd8, + 0xa8, 0xfa, 0x76, 0x3f, 0xe3, 0x9c, 0x34, 0x3f, 0xf9, 0xdc, 0xbb, 0xc7, + 0xc7, 0x0b, 0x4f, 0x1d, 0x8a, 0x51, 0xe0, 0x4b, 0xcd, 0xb4, 0x59, 0x31, + 0xc8, 0x9f, 0x7e, 0xc9, 0xd9, 0x78, 0x73, 0x64, 0xea, 0xc5, 0xac, 0x83, + 0x34, 0xd3, 0xeb, 0xc3, 0xc5, 0x81, 0xa0, 0xff, 0xfa, 0x13, 0x63, 0xeb, + 0x17, 0x0d, 0xdd, 0x51, 0xb7, 0xf0, 0xda, 0x49, 0xd3, 0x16, 0x55, 0x26, + 0x29, 0xd4, 0x68, 0x9e, 0x2b, 0x16, 0xbe, 0x58, 0x7d, 0x47, 0xa1, 0xfc, + 0x8f, 0xf8, 0xb8, 0xd1, 0x7a, 0xd0, 0x31, 0xce, 0x45, 0xcb, 0x3a, 0x8f, + 0x95, 0x16, 0x04, 0x28, 0xaf, 0xd7, 0xfb, 0xca, 0xbb, 0x4b, 0x40, 0x7e, + +}; + + #ifdef XXH_OLD_NAMES + #define kSecret XXH3_kSecret + #endif + + /* + * Calculates a 32-bit to 64-bit long multiply. + * + * Wraps __emulu on MSVC x86 because it tends to call __allmul when it + * doesn't need to (but it shouldn't need to anyways, it is about 7 + * instructions to do a 64x64 multiply...). Since we know that this will + * _always_ emit MULL, we use that instead of the normal method. + * + * If you are compiling for platforms like Thumb-1 and don't have a better + * option, you may also want to write your own long multiply routine here. + * + * XXH_FORCE_INLINE xxh_u64 XXH_mult32to64(xxh_u64 x, xxh_u64 y) + * { + + * return (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF); + * } + */ + #if defined(_MSC_VER) && defined(_M_IX86) + #include <intrin.h> + #define XXH_mult32to64(x, y) __emulu((unsigned)(x), (unsigned)(y)) + #else + /* + * Downcast + upcast is usually better than masking on older compilers + * like GCC 4.2 (especially 32-bit ones), all without affecting newer + * compilers. + * + * The other method, (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF), will AND both + * operands and perform a full 64x64 multiply -- entirely redundant on + * 32-bit. + */ + #define XXH_mult32to64(x, y) \ + ((xxh_u64)(xxh_u32)(x) * (xxh_u64)(xxh_u32)(y)) + #endif + +/* + * Calculates a 64->128-bit long multiply. + * + * Uses __uint128_t and _umul128 if available, otherwise uses a scalar version. + */ +static XXH128_hash_t XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs) { + + /* + * GCC/Clang __uint128_t method. + * + * On most 64-bit targets, GCC and Clang define a __uint128_t type. + * This is usually the best way as it usually uses a native long 64-bit + * multiply, such as MULQ on x86_64 or MUL + UMULH on aarch64. + * + * Usually. + * + * Despite being a 32-bit platform, Clang (and emscripten) define this type + * despite not having the arithmetic for it. This results in a laggy + * compiler builtin call which calculates a full 128-bit multiply. + * In that case it is best to use the portable one. + * https://github.com/Cyan4973/xxHash/issues/211#issuecomment-515575677 + */ + #if defined(__GNUC__) && !defined(__wasm__) && \ + defined(__SIZEOF_INT128__) || \ + (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128) + + __uint128_t const product = (__uint128_t)lhs * (__uint128_t)rhs; + XXH128_hash_t r128; + r128.low64 = (xxh_u64)(product); + r128.high64 = (xxh_u64)(product >> 64); + return r128; + + /* + * MSVC for x64's _umul128 method. + * + * xxh_u64 _umul128(xxh_u64 Multiplier, xxh_u64 Multiplicand, xxh_u64 + * *HighProduct); + * + * This compiles to single operand MUL on x64. + */ + #elif defined(_M_X64) || defined(_M_IA64) + + #ifndef _MSC_VER + #pragma intrinsic(_umul128) + #endif + xxh_u64 product_high; + xxh_u64 const product_low = _umul128(lhs, rhs, &product_high); + XXH128_hash_t r128; + r128.low64 = product_low; + r128.high64 = product_high; + return r128; + + #else + /* + * Portable scalar method. Optimized for 32-bit and 64-bit ALUs. + * + * This is a fast and simple grade school multiply, which is shown below + * with base 10 arithmetic instead of base 0x100000000. + * + * 9 3 // D2 lhs = 93 + * x 7 5 // D2 rhs = 75 + * ---------- + * 1 5 // D2 lo_lo = (93 % 10) * (75 % 10) = 15 + * 4 5 | // D2 hi_lo = (93 / 10) * (75 % 10) = 45 + * 2 1 | // D2 lo_hi = (93 % 10) * (75 / 10) = 21 + * + 6 3 | | // D2 hi_hi = (93 / 10) * (75 / 10) = 63 + * --------- + * 2 7 | // D2 cross = (15 / 10) + (45 % 10) + 21 = 27 + * + 6 7 | | // D2 upper = (27 / 10) + (45 / 10) + 63 = 67 + * --------- + * 6 9 7 5 // D4 res = (27 * 10) + (15 % 10) + (67 * 100) = 6975 + * + * The reasons for adding the products like this are: + * 1. It avoids manual carry tracking. Just like how + * (9 * 9) + 9 + 9 = 99, the same applies with this for UINT64_MAX. + * This avoids a lot of complexity. + * + * 2. It hints for, and on Clang, compiles to, the powerful UMAAL + * instruction available in ARM's Digital Signal Processing extension + * in 32-bit ARMv6 and later, which is shown below: + * + * void UMAAL(xxh_u32 *RdLo, xxh_u32 *RdHi, xxh_u32 Rn, xxh_u32 Rm) + * { + + * xxh_u64 product = (xxh_u64)*RdLo * (xxh_u64)*RdHi + Rn + Rm; + * *RdLo = (xxh_u32)(product & 0xFFFFFFFF); + * *RdHi = (xxh_u32)(product >> 32); + * } + * + * This instruction was designed for efficient long multiplication, and + * allows this to be calculated in only 4 instructions at speeds + * comparable to some 64-bit ALUs. + * + * 3. It isn't terrible on other platforms. Usually this will be a couple + * of 32-bit ADD/ADCs. + */ + + /* First calculate all of the cross products. */ + xxh_u64 const lo_lo = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs & 0xFFFFFFFF); + xxh_u64 const hi_lo = XXH_mult32to64(lhs >> 32, rhs & 0xFFFFFFFF); + xxh_u64 const lo_hi = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs >> 32); + xxh_u64 const hi_hi = XXH_mult32to64(lhs >> 32, rhs >> 32); + + /* Now add the products together. These will never overflow. */ + xxh_u64 const cross = (lo_lo >> 32) + (hi_lo & 0xFFFFFFFF) + lo_hi; + xxh_u64 const upper = (hi_lo >> 32) + (cross >> 32) + hi_hi; + xxh_u64 const lower = (cross << 32) | (lo_lo & 0xFFFFFFFF); + + XXH128_hash_t r128; + r128.low64 = lower; + r128.high64 = upper; + return r128; + #endif + +} + +/* + * Does a 64-bit to 128-bit multiply, then XOR folds it. + * + * The reason for the separate function is to prevent passing too many structs + * around by value. This will hopefully inline the multiply, but we don't force + * it. + */ +static xxh_u64 XXH3_mul128_fold64(xxh_u64 lhs, xxh_u64 rhs) { + + XXH128_hash_t product = XXH_mult64to128(lhs, rhs); + return product.low64 ^ product.high64; + +} + +/* Seems to produce slightly better code on GCC for some reason. */ +XXH_FORCE_INLINE xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift) { + + XXH_ASSERT(0 <= shift && shift < 64); + return v64 ^ (v64 >> shift); + +} + +/* + * This is a fast avalanche stage, + * suitable when input bits are already partially mixed + */ +static XXH64_hash_t XXH3_avalanche(xxh_u64 h64) { + + h64 = XXH_xorshift64(h64, 37); + h64 *= 0x165667919E3779F9ULL; + h64 = XXH_xorshift64(h64, 32); + return h64; + +} + +/* + * This is a stronger avalanche, + * inspired by Pelle Evensen's rrmxmx + * preferable when input has not been previously mixed + */ +static XXH64_hash_t XXH3_rrmxmx(xxh_u64 h64, xxh_u64 len) { + + /* this mix is inspired by Pelle Evensen's rrmxmx */ + h64 ^= XXH_rotl64(h64, 49) ^ XXH_rotl64(h64, 24); + h64 *= 0x9FB21C651E98DF25ULL; + h64 ^= (h64 >> 35) + len; + h64 *= 0x9FB21C651E98DF25ULL; + return XXH_xorshift64(h64, 28); + +} + +/* ========================================== + * Short keys + * ========================================== + * One of the shortcomings of XXH32 and XXH64 was that their performance was + * sub-optimal on short lengths. It used an iterative algorithm which strongly + * favored lengths that were a multiple of 4 or 8. + * + * Instead of iterating over individual inputs, we use a set of single shot + * functions which piece together a range of lengths and operate in constant + * time. + * + * Additionally, the number of multiplies has been significantly reduced. This + * reduces latency, especially when emulating 64-bit multiplies on 32-bit. + * + * Depending on the platform, this may or may not be faster than XXH32, but it + * is almost guaranteed to be faster than XXH64. + */ + +/* + * At very short lengths, there isn't enough input to fully hide secrets, or use + * the entire secret. + * + * There is also only a limited amount of mixing we can do before significantly + * impacting performance. + * + * Therefore, we use different sections of the secret and always mix two secret + * samples with an XOR. This should have no effect on performance on the + * seedless or withSeed variants because everything _should_ be constant folded + * by modern compilers. + * + * The XOR mixing hides individual parts of the secret and increases entropy. + * + * This adds an extra layer of strength for custom secrets. + */ +XXH_FORCE_INLINE XXH64_hash_t XXH3_len_1to3_64b(const xxh_u8 *input, size_t len, + const xxh_u8 *secret, + XXH64_hash_t seed) { + + XXH_ASSERT(input != NULL); + XXH_ASSERT(1 <= len && len <= 3); + XXH_ASSERT(secret != NULL); + /* + * len = 1: combined = { input[0], 0x01, input[0], input[0] } + * len = 2: combined = { input[1], 0x02, input[0], input[1] } + * len = 3: combined = { input[2], 0x03, input[0], input[1] } + */ + { + + xxh_u8 const c1 = input[0]; + xxh_u8 const c2 = input[len >> 1]; + xxh_u8 const c3 = input[len - 1]; + xxh_u32 const combined = ((xxh_u32)c1 << 16) | ((xxh_u32)c2 << 24) | + ((xxh_u32)c3 << 0) | ((xxh_u32)len << 8); + xxh_u64 const bitflip = + (XXH_readLE32(secret) ^ XXH_readLE32(secret + 4)) + seed; + xxh_u64 const keyed = (xxh_u64)combined ^ bitflip; + return XXH64_avalanche(keyed); + + } + +} + +XXH_FORCE_INLINE XXH64_hash_t XXH3_len_4to8_64b(const xxh_u8 *input, size_t len, + const xxh_u8 *secret, + XXH64_hash_t seed) { + + XXH_ASSERT(input != NULL); + XXH_ASSERT(secret != NULL); + XXH_ASSERT(4 <= len && len < 8); + seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32; + { + + xxh_u32 const input1 = XXH_readLE32(input); + xxh_u32 const input2 = XXH_readLE32(input + len - 4); + xxh_u64 const bitflip = + (XXH_readLE64(secret + 8) ^ XXH_readLE64(secret + 16)) - seed; + xxh_u64 const input64 = input2 + (((xxh_u64)input1) << 32); + xxh_u64 const keyed = input64 ^ bitflip; + return XXH3_rrmxmx(keyed, len); + + } + +} + +XXH_FORCE_INLINE XXH64_hash_t XXH3_len_9to16_64b(const xxh_u8 *input, + size_t len, + const xxh_u8 *secret, + XXH64_hash_t seed) { + + XXH_ASSERT(input != NULL); + XXH_ASSERT(secret != NULL); + XXH_ASSERT(8 <= len && len <= 16); + { + + xxh_u64 const bitflip1 = + (XXH_readLE64(secret + 24) ^ XXH_readLE64(secret + 32)) + seed; + xxh_u64 const bitflip2 = + (XXH_readLE64(secret + 40) ^ XXH_readLE64(secret + 48)) - seed; + xxh_u64 const input_lo = XXH_readLE64(input) ^ bitflip1; + xxh_u64 const input_hi = XXH_readLE64(input + len - 8) ^ bitflip2; + xxh_u64 const acc = len + XXH_swap64(input_lo) + input_hi + + XXH3_mul128_fold64(input_lo, input_hi); + return XXH3_avalanche(acc); + + } + +} + +XXH_FORCE_INLINE XXH64_hash_t XXH3_len_0to16_64b(const xxh_u8 *input, + size_t len, + const xxh_u8 *secret, + XXH64_hash_t seed) { + + XXH_ASSERT(len <= 16); + { + + if (XXH_likely(len > 8)) + return XXH3_len_9to16_64b(input, len, secret, seed); + if (XXH_likely(len >= 4)) + return XXH3_len_4to8_64b(input, len, secret, seed); + if (len) return XXH3_len_1to3_64b(input, len, secret, seed); + return XXH64_avalanche( + seed ^ (XXH_readLE64(secret + 56) ^ XXH_readLE64(secret + 64))); + + } + +} + +/* + * DISCLAIMER: There are known *seed-dependent* multicollisions here due to + * multiplication by zero, affecting hashes of lengths 17 to 240. + * + * However, they are very unlikely. + * + * Keep this in mind when using the unseeded XXH3_64bits() variant: As with all + * unseeded non-cryptographic hashes, it does not attempt to defend itself + * against specially crafted inputs, only random inputs. + * + * Compared to classic UMAC where a 1 in 2^31 chance of 4 consecutive bytes + * cancelling out the secret is taken an arbitrary number of times (addressed + * in XXH3_accumulate_512), this collision is very unlikely with random inputs + * and/or proper seeding: + * + * This only has a 1 in 2^63 chance of 8 consecutive bytes cancelling out, in a + * function that is only called up to 16 times per hash with up to 240 bytes of + * input. + * + * This is not too bad for a non-cryptographic hash function, especially with + * only 64 bit outputs. + * + * The 128-bit variant (which trades some speed for strength) is NOT affected + * by this, although it is always a good idea to use a proper seed if you care + * about strength. + */ +XXH_FORCE_INLINE xxh_u64 XXH3_mix16B(const xxh_u8 *XXH_RESTRICT input, + const xxh_u8 *XXH_RESTRICT secret, + xxh_u64 seed64) { + + #if defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \ + && defined(__i386__) && defined(__SSE2__) /* x86 + SSE2 */ \ + && \ + !defined( \ + XXH_ENABLE_AUTOVECTORIZE) /* Define to disable like XXH32 hack */ + /* + * UGLY HACK: + * GCC for x86 tends to autovectorize the 128-bit multiply, resulting in + * slower code. + * + * By forcing seed64 into a register, we disrupt the cost model and + * cause it to scalarize. See `XXH32_round()` + * + * FIXME: Clang's output is still _much_ faster -- On an AMD Ryzen 3600, + * XXH3_64bits @ len=240 runs at 4.6 GB/s with Clang 9, but 3.3 GB/s on + * GCC 9.2, despite both emitting scalar code. + * + * GCC generates much better scalar code than Clang for the rest of XXH3, + * which is why finding a more optimal codepath is an interest. + */ + __asm__("" : "+r"(seed64)); + #endif + { + + xxh_u64 const input_lo = XXH_readLE64(input); + xxh_u64 const input_hi = XXH_readLE64(input + 8); + return XXH3_mul128_fold64(input_lo ^ (XXH_readLE64(secret) + seed64), + input_hi ^ (XXH_readLE64(secret + 8) - seed64)); + + } + +} + +/* For mid range keys, XXH3 uses a Mum-hash variant. */ +XXH_FORCE_INLINE XXH64_hash_t XXH3_len_17to128_64b( + const xxh_u8 *XXH_RESTRICT input, size_t len, + const xxh_u8 *XXH_RESTRICT secret, size_t secretSize, XXH64_hash_t seed) { + + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); + (void)secretSize; + XXH_ASSERT(16 < len && len <= 128); + + { + + xxh_u64 acc = len * XXH_PRIME64_1; + if (len > 32) { + + if (len > 64) { + + if (len > 96) { + + acc += XXH3_mix16B(input + 48, secret + 96, seed); + acc += XXH3_mix16B(input + len - 64, secret + 112, seed); + + } + + acc += XXH3_mix16B(input + 32, secret + 64, seed); + acc += XXH3_mix16B(input + len - 48, secret + 80, seed); + + } + + acc += XXH3_mix16B(input + 16, secret + 32, seed); + acc += XXH3_mix16B(input + len - 32, secret + 48, seed); + + } + + acc += XXH3_mix16B(input + 0, secret + 0, seed); + acc += XXH3_mix16B(input + len - 16, secret + 16, seed); + + return XXH3_avalanche(acc); + + } + +} + + #define XXH3_MIDSIZE_MAX 240 + +XXH_NO_INLINE XXH64_hash_t XXH3_len_129to240_64b( + const xxh_u8 *XXH_RESTRICT input, size_t len, + const xxh_u8 *XXH_RESTRICT secret, size_t secretSize, XXH64_hash_t seed) { + + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); + (void)secretSize; + XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX); + + #define XXH3_MIDSIZE_STARTOFFSET 3 + #define XXH3_MIDSIZE_LASTOFFSET 17 + + { + + xxh_u64 acc = len * XXH_PRIME64_1; + int const nbRounds = (int)len / 16; + int i; + for (i = 0; i < 8; i++) { + + acc += XXH3_mix16B(input + (16 * i), secret + (16 * i), seed); + + } + + acc = XXH3_avalanche(acc); + XXH_ASSERT(nbRounds >= 8); + #if defined(__clang__) /* Clang */ \ + && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \ + && !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable */ + /* + * UGLY HACK: + * Clang for ARMv7-A tries to vectorize this loop, similar to GCC x86. + * In everywhere else, it uses scalar code. + * + * For 64->128-bit multiplies, even if the NEON was 100% optimal, it + * would still be slower than UMAAL (see XXH_mult64to128). + * + * Unfortunately, Clang doesn't handle the long multiplies properly and + * converts them to the nonexistent "vmulq_u64" intrinsic, which is then + * scalarized into an ugly mess of VMOV.32 instructions. + * + * This mess is difficult to avoid without turning autovectorization + * off completely, but they are usually relatively minor and/or not + * worth it to fix. + * + * This loop is the easiest to fix, as unlike XXH32, this pragma + * _actually works_ because it is a loop vectorization instead of an + * SLP vectorization. + */ + #pragma clang loop vectorize(disable) + #endif + for (i = 8; i < nbRounds; i++) { + + acc += + XXH3_mix16B(input + (16 * i), + secret + (16 * (i - 8)) + XXH3_MIDSIZE_STARTOFFSET, seed); + + } + + /* last bytes */ + acc += XXH3_mix16B(input + len - 16, + secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, + seed); + return XXH3_avalanche(acc); + + } + +} + + /* ======= Long Keys ======= */ + + #define XXH_STRIPE_LEN 64 + #define XXH_SECRET_CONSUME_RATE \ + 8 /* nb of secret bytes consumed at each accumulation */ + #define XXH_ACC_NB (XXH_STRIPE_LEN / sizeof(xxh_u64)) + + #ifdef XXH_OLD_NAMES + #define STRIPE_LEN XXH_STRIPE_LEN + #define ACC_NB XXH_ACC_NB + #endif + +XXH_FORCE_INLINE void XXH_writeLE64(void *dst, xxh_u64 v64) { + + if (!XXH_CPU_LITTLE_ENDIAN) v64 = XXH_swap64(v64); + memcpy(dst, &v64, sizeof(v64)); + +} + + /* Several intrinsic functions below are supposed to accept __int64 as + * argument, as documented in + * https://software.intel.com/sites/landingpage/IntrinsicsGuide/ . However, + * several environments do not define __int64 type, requiring a workaround. + */ + #if !defined(__VMS) && \ + (defined(__cplusplus) || (defined(__STDC_VERSION__) && \ + (__STDC_VERSION__ >= 199901L) /* C99 */)) +typedef int64_t xxh_i64; + #else +/* the following type must have a width of 64-bit */ +typedef long long xxh_i64; + #endif + + /* + * XXH3_accumulate_512 is the tightest loop for long inputs, and it is the + * most optimized. + * + * It is a hardened version of UMAC, based off of FARSH's implementation. + * + * This was chosen because it adapts quite well to 32-bit, 64-bit, and SIMD + * implementations, and it is ridiculously fast. + * + * We harden it by mixing the original input to the accumulators as well as + * the product. + * + * This means that in the (relatively likely) case of a multiply by zero, the + * original input is preserved. + * + * On 128-bit inputs, we swap 64-bit pairs when we add the input to improve + * cross-pollination, as otherwise the upper and lower halves would be + * essentially independent. + * + * This doesn't matter on 64-bit hashes since they all get merged together in + * the end, so we skip the extra step. + * + * Both XXH3_64bits and XXH3_128bits use this subroutine. + */ + + #if (XXH_VECTOR == XXH_AVX512) || defined(XXH_X86DISPATCH) + + #ifndef XXH_TARGET_AVX512 + #define XXH_TARGET_AVX512 /* disable attribute target */ + #endif + +XXH_FORCE_INLINE XXH_TARGET_AVX512 void XXH3_accumulate_512_avx512( + void *XXH_RESTRICT acc, const void *XXH_RESTRICT input, + const void *XXH_RESTRICT secret) { + + XXH_ALIGN(64) __m512i *const xacc = (__m512i *)acc; + XXH_ASSERT((((size_t)acc) & 63) == 0); + XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i)); + + { + + /* data_vec = input[0]; */ + __m512i const data_vec = _mm512_loadu_si512(input); + /* key_vec = secret[0]; */ + __m512i const key_vec = _mm512_loadu_si512(secret); + /* data_key = data_vec ^ key_vec; */ + __m512i const data_key = _mm512_xor_si512(data_vec, key_vec); + /* data_key_lo = data_key >> 32; */ + __m512i const data_key_lo = + _mm512_shuffle_epi32(data_key, (_MM_PERM_ENUM)_MM_SHUFFLE(0, 3, 0, 1)); + /* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */ + __m512i const product = _mm512_mul_epu32(data_key, data_key_lo); + /* xacc[0] += swap(data_vec); */ + __m512i const data_swap = + _mm512_shuffle_epi32(data_vec, (_MM_PERM_ENUM)_MM_SHUFFLE(1, 0, 3, 2)); + __m512i const sum = _mm512_add_epi64(*xacc, data_swap); + /* xacc[0] += product; */ + *xacc = _mm512_add_epi64(product, sum); + + } + +} + +/* + * XXH3_scrambleAcc: Scrambles the accumulators to improve mixing. + * + * Multiplication isn't perfect, as explained by Google in HighwayHash: + * + * // Multiplication mixes/scrambles bytes 0-7 of the 64-bit result to + * // varying degrees. In descending order of goodness, bytes + * // 3 4 2 5 1 6 0 7 have quality 228 224 164 160 100 96 36 32. + * // As expected, the upper and lower bytes are much worse. + * + * Source: + * https://github.com/google/highwayhash/blob/0aaf66b/highwayhash/hh_avx2.h#L291 + * + * Since our algorithm uses a pseudorandom secret to add some variance into the + * mix, we don't need to (or want to) mix as often or as much as HighwayHash + * does. + * + * This isn't as tight as XXH3_accumulate, but still written in SIMD to avoid + * extraction. + * + * Both XXH3_64bits and XXH3_128bits use this subroutine. + */ + +XXH_FORCE_INLINE XXH_TARGET_AVX512 void XXH3_scrambleAcc_avx512( + void *XXH_RESTRICT acc, const void *XXH_RESTRICT secret) { + + XXH_ASSERT((((size_t)acc) & 63) == 0); + XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i)); + { + + XXH_ALIGN(64) __m512i *const xacc = (__m512i *)acc; + const __m512i prime32 = _mm512_set1_epi32((int)XXH_PRIME32_1); + + /* xacc[0] ^= (xacc[0] >> 47) */ + __m512i const acc_vec = *xacc; + __m512i const shifted = _mm512_srli_epi64(acc_vec, 47); + __m512i const data_vec = _mm512_xor_si512(acc_vec, shifted); + /* xacc[0] ^= secret; */ + __m512i const key_vec = _mm512_loadu_si512(secret); + __m512i const data_key = _mm512_xor_si512(data_vec, key_vec); + + /* xacc[0] *= XXH_PRIME32_1; */ + __m512i const data_key_hi = + _mm512_shuffle_epi32(data_key, (_MM_PERM_ENUM)_MM_SHUFFLE(0, 3, 0, 1)); + __m512i const prod_lo = _mm512_mul_epu32(data_key, prime32); + __m512i const prod_hi = _mm512_mul_epu32(data_key_hi, prime32); + *xacc = _mm512_add_epi64(prod_lo, _mm512_slli_epi64(prod_hi, 32)); + + } + +} + +XXH_FORCE_INLINE XXH_TARGET_AVX512 void XXH3_initCustomSecret_avx512( + void *XXH_RESTRICT customSecret, xxh_u64 seed64) { + + XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 63) == 0); + XXH_STATIC_ASSERT(XXH_SEC_ALIGN == 64); + XXH_ASSERT(((size_t)customSecret & 63) == 0); + (void)(&XXH_writeLE64); + { + + int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m512i); + __m512i const seed = _mm512_mask_set1_epi64( + _mm512_set1_epi64((xxh_i64)seed64), 0xAA, -(xxh_i64)seed64); + + XXH_ALIGN(64) const __m512i *const src = (const __m512i *)XXH3_kSecret; + XXH_ALIGN(64) __m512i *const dest = (__m512i *)customSecret; + int i; + for (i = 0; i < nbRounds; ++i) { + + /* GCC has a bug, _mm512_stream_load_si512 accepts 'void*', not 'void + * const*', this will warn "discards ‘const’ qualifier". */ + union { + + XXH_ALIGN(64) const __m512i *cp; + XXH_ALIGN(64) void *p; + + } remote_const_void; + + remote_const_void.cp = src + i; + dest[i] = + _mm512_add_epi64(_mm512_stream_load_si512(remote_const_void.p), seed); + + } + + } + +} + + #endif + + #if (XXH_VECTOR == XXH_AVX2) || defined(XXH_X86DISPATCH) + + #ifndef XXH_TARGET_AVX2 + #define XXH_TARGET_AVX2 /* disable attribute target */ + #endif + +XXH_FORCE_INLINE XXH_TARGET_AVX2 void XXH3_accumulate_512_avx2( + void *XXH_RESTRICT acc, const void *XXH_RESTRICT input, + const void *XXH_RESTRICT secret) { + + XXH_ASSERT((((size_t)acc) & 31) == 0); + { + + XXH_ALIGN(32) __m256i *const xacc = (__m256i *)acc; + /* Unaligned. This is mainly for pointer arithmetic, and because + * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. + */ + const __m256i *const xinput = (const __m256i *)input; + /* Unaligned. This is mainly for pointer arithmetic, and because + * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */ + const __m256i *const xsecret = (const __m256i *)secret; + + size_t i; + for (i = 0; i < XXH_STRIPE_LEN / sizeof(__m256i); i++) { + + /* data_vec = xinput[i]; */ + __m256i const data_vec = _mm256_loadu_si256(xinput + i); + /* key_vec = xsecret[i]; */ + __m256i const key_vec = _mm256_loadu_si256(xsecret + i); + /* data_key = data_vec ^ key_vec; */ + __m256i const data_key = _mm256_xor_si256(data_vec, key_vec); + /* data_key_lo = data_key >> 32; */ + __m256i const data_key_lo = + _mm256_shuffle_epi32(data_key, _MM_SHUFFLE(0, 3, 0, 1)); + /* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */ + __m256i const product = _mm256_mul_epu32(data_key, data_key_lo); + /* xacc[i] += swap(data_vec); */ + __m256i const data_swap = + _mm256_shuffle_epi32(data_vec, _MM_SHUFFLE(1, 0, 3, 2)); + __m256i const sum = _mm256_add_epi64(xacc[i], data_swap); + /* xacc[i] += product; */ + xacc[i] = _mm256_add_epi64(product, sum); + + } + + } + +} + +XXH_FORCE_INLINE XXH_TARGET_AVX2 void XXH3_scrambleAcc_avx2( + void *XXH_RESTRICT acc, const void *XXH_RESTRICT secret) { + + XXH_ASSERT((((size_t)acc) & 31) == 0); + { + + XXH_ALIGN(32) __m256i *const xacc = (__m256i *)acc; + /* Unaligned. This is mainly for pointer arithmetic, and because + * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */ + const __m256i *const xsecret = (const __m256i *)secret; + const __m256i prime32 = _mm256_set1_epi32((int)XXH_PRIME32_1); + + size_t i; + for (i = 0; i < XXH_STRIPE_LEN / sizeof(__m256i); i++) { + + /* xacc[i] ^= (xacc[i] >> 47) */ + __m256i const acc_vec = xacc[i]; + __m256i const shifted = _mm256_srli_epi64(acc_vec, 47); + __m256i const data_vec = _mm256_xor_si256(acc_vec, shifted); + /* xacc[i] ^= xsecret; */ + __m256i const key_vec = _mm256_loadu_si256(xsecret + i); + __m256i const data_key = _mm256_xor_si256(data_vec, key_vec); + + /* xacc[i] *= XXH_PRIME32_1; */ + __m256i const data_key_hi = + _mm256_shuffle_epi32(data_key, _MM_SHUFFLE(0, 3, 0, 1)); + __m256i const prod_lo = _mm256_mul_epu32(data_key, prime32); + __m256i const prod_hi = _mm256_mul_epu32(data_key_hi, prime32); + xacc[i] = _mm256_add_epi64(prod_lo, _mm256_slli_epi64(prod_hi, 32)); + + } + + } + +} + +XXH_FORCE_INLINE XXH_TARGET_AVX2 void XXH3_initCustomSecret_avx2( + void *XXH_RESTRICT customSecret, xxh_u64 seed64) { + + XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 31) == 0); + XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE / sizeof(__m256i)) == 6); + XXH_STATIC_ASSERT(XXH_SEC_ALIGN <= 64); + (void)(&XXH_writeLE64); + XXH_PREFETCH(customSecret); + { + + __m256i const seed = _mm256_set_epi64x(-(xxh_i64)seed64, (xxh_i64)seed64, + -(xxh_i64)seed64, (xxh_i64)seed64); + + XXH_ALIGN(64) const __m256i *const src = (const __m256i *)XXH3_kSecret; + XXH_ALIGN(64) __m256i * dest = (__m256i *)customSecret; + + #if defined(__GNUC__) || defined(__clang__) + /* + * On GCC & Clang, marking 'dest' as modified will cause the compiler: + * - do not extract the secret from sse registers in the internal loop + * - use less common registers, and avoid pushing these reg into stack + * The asm hack causes Clang to assume that XXH3_kSecretPtr aliases with + * customSecret, and on aarch64, this prevented LDP from merging two + * loads together for free. Putting the loads together before the stores + * properly generates LDP. + */ + __asm__("" : "+r"(dest)); + #endif + + /* GCC -O2 need unroll loop manually */ + dest[0] = _mm256_add_epi64(_mm256_stream_load_si256(src + 0), seed); + dest[1] = _mm256_add_epi64(_mm256_stream_load_si256(src + 1), seed); + dest[2] = _mm256_add_epi64(_mm256_stream_load_si256(src + 2), seed); + dest[3] = _mm256_add_epi64(_mm256_stream_load_si256(src + 3), seed); + dest[4] = _mm256_add_epi64(_mm256_stream_load_si256(src + 4), seed); + dest[5] = _mm256_add_epi64(_mm256_stream_load_si256(src + 5), seed); + + } + +} + + #endif + + #if (XXH_VECTOR == XXH_SSE2) || defined(XXH_X86DISPATCH) + + #ifndef XXH_TARGET_SSE2 + #define XXH_TARGET_SSE2 /* disable attribute target */ + #endif + +XXH_FORCE_INLINE XXH_TARGET_SSE2 void XXH3_accumulate_512_sse2( + void *XXH_RESTRICT acc, const void *XXH_RESTRICT input, + const void *XXH_RESTRICT secret) { + + /* SSE2 is just a half-scale version of the AVX2 version. */ + XXH_ASSERT((((size_t)acc) & 15) == 0); + { + + XXH_ALIGN(16) __m128i *const xacc = (__m128i *)acc; + /* Unaligned. This is mainly for pointer arithmetic, and because + * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */ + const __m128i *const xinput = (const __m128i *)input; + /* Unaligned. This is mainly for pointer arithmetic, and because + * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */ + const __m128i *const xsecret = (const __m128i *)secret; + + size_t i; + for (i = 0; i < XXH_STRIPE_LEN / sizeof(__m128i); i++) { + + /* data_vec = xinput[i]; */ + __m128i const data_vec = _mm_loadu_si128(xinput + i); + /* key_vec = xsecret[i]; */ + __m128i const key_vec = _mm_loadu_si128(xsecret + i); + /* data_key = data_vec ^ key_vec; */ + __m128i const data_key = _mm_xor_si128(data_vec, key_vec); + /* data_key_lo = data_key >> 32; */ + __m128i const data_key_lo = + _mm_shuffle_epi32(data_key, _MM_SHUFFLE(0, 3, 0, 1)); + /* product = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */ + __m128i const product = _mm_mul_epu32(data_key, data_key_lo); + /* xacc[i] += swap(data_vec); */ + __m128i const data_swap = + _mm_shuffle_epi32(data_vec, _MM_SHUFFLE(1, 0, 3, 2)); + __m128i const sum = _mm_add_epi64(xacc[i], data_swap); + /* xacc[i] += product; */ + xacc[i] = _mm_add_epi64(product, sum); + + } + + } + +} + +XXH_FORCE_INLINE XXH_TARGET_SSE2 void XXH3_scrambleAcc_sse2( + void *XXH_RESTRICT acc, const void *XXH_RESTRICT secret) { + + XXH_ASSERT((((size_t)acc) & 15) == 0); + { + + XXH_ALIGN(16) __m128i *const xacc = (__m128i *)acc; + /* Unaligned. This is mainly for pointer arithmetic, and because + * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */ + const __m128i *const xsecret = (const __m128i *)secret; + const __m128i prime32 = _mm_set1_epi32((int)XXH_PRIME32_1); + + size_t i; + for (i = 0; i < XXH_STRIPE_LEN / sizeof(__m128i); i++) { + + /* xacc[i] ^= (xacc[i] >> 47) */ + __m128i const acc_vec = xacc[i]; + __m128i const shifted = _mm_srli_epi64(acc_vec, 47); + __m128i const data_vec = _mm_xor_si128(acc_vec, shifted); + /* xacc[i] ^= xsecret[i]; */ + __m128i const key_vec = _mm_loadu_si128(xsecret + i); + __m128i const data_key = _mm_xor_si128(data_vec, key_vec); + + /* xacc[i] *= XXH_PRIME32_1; */ + __m128i const data_key_hi = + _mm_shuffle_epi32(data_key, _MM_SHUFFLE(0, 3, 0, 1)); + __m128i const prod_lo = _mm_mul_epu32(data_key, prime32); + __m128i const prod_hi = _mm_mul_epu32(data_key_hi, prime32); + xacc[i] = _mm_add_epi64(prod_lo, _mm_slli_epi64(prod_hi, 32)); + + } + + } + +} + +XXH_FORCE_INLINE XXH_TARGET_SSE2 void XXH3_initCustomSecret_sse2( + void *XXH_RESTRICT customSecret, xxh_u64 seed64) { + + XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0); + (void)(&XXH_writeLE64); + { + + int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m128i); + + #if defined(_MSC_VER) && defined(_M_IX86) && _MSC_VER < 1900 + // MSVC 32bit mode does not support _mm_set_epi64x before 2015 + XXH_ALIGN(16) + const xxh_i64 seed64x2[2] = {(xxh_i64)seed64, -(xxh_i64)seed64}; + __m128i const seed = _mm_load_si128((__m128i const *)seed64x2); + #else + __m128i const seed = _mm_set_epi64x(-(xxh_i64)seed64, (xxh_i64)seed64); + #endif + int i; + + XXH_ALIGN(64) const float *const src = (float const *)XXH3_kSecret; + XXH_ALIGN(XXH_SEC_ALIGN) __m128i *dest = (__m128i *)customSecret; + #if defined(__GNUC__) || defined(__clang__) + /* + * On GCC & Clang, marking 'dest' as modified will cause the compiler: + * - do not extract the secret from sse registers in the internal loop + * - use less common registers, and avoid pushing these reg into stack + */ + __asm__("" : "+r"(dest)); + #endif + + for (i = 0; i < nbRounds; ++i) { + + dest[i] = _mm_add_epi64(_mm_castps_si128(_mm_load_ps(src + i * 4)), seed); + + } + + } + +} + + #endif + + #if (XXH_VECTOR == XXH_NEON) + +XXH_FORCE_INLINE void XXH3_accumulate_512_neon( + void *XXH_RESTRICT acc, const void *XXH_RESTRICT input, + const void *XXH_RESTRICT secret) { + + XXH_ASSERT((((size_t)acc) & 15) == 0); + { + + XXH_ALIGN(16) uint64x2_t *const xacc = (uint64x2_t *)acc; + /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. + */ + uint8_t const *const xinput = (const uint8_t *)input; + uint8_t const *const xsecret = (const uint8_t *)secret; + + size_t i; + for (i = 0; i < XXH_STRIPE_LEN / sizeof(uint64x2_t); i++) { + + /* data_vec = xinput[i]; */ + uint8x16_t data_vec = vld1q_u8(xinput + (i * 16)); + /* key_vec = xsecret[i]; */ + uint8x16_t key_vec = vld1q_u8(xsecret + (i * 16)); + uint64x2_t data_key; + uint32x2_t data_key_lo, data_key_hi; + /* xacc[i] += swap(data_vec); */ + uint64x2_t const data64 = vreinterpretq_u64_u8(data_vec); + uint64x2_t const swapped = vextq_u64(data64, data64, 1); + xacc[i] = vaddq_u64(xacc[i], swapped); + /* data_key = data_vec ^ key_vec; */ + data_key = vreinterpretq_u64_u8(veorq_u8(data_vec, key_vec)); + /* data_key_lo = (uint32x2_t) (data_key & 0xFFFFFFFF); + * data_key_hi = (uint32x2_t) (data_key >> 32); + * data_key = UNDEFINED; */ + XXH_SPLIT_IN_PLACE(data_key, data_key_lo, data_key_hi); + /* xacc[i] += (uint64x2_t) data_key_lo * (uint64x2_t) data_key_hi; */ + xacc[i] = vmlal_u32(xacc[i], data_key_lo, data_key_hi); + + } + + } + +} + +XXH_FORCE_INLINE void XXH3_scrambleAcc_neon(void *XXH_RESTRICT acc, + const void *XXH_RESTRICT secret) { + + XXH_ASSERT((((size_t)acc) & 15) == 0); + + { + + uint64x2_t * xacc = (uint64x2_t *)acc; + uint8_t const *xsecret = (uint8_t const *)secret; + uint32x2_t prime = vdup_n_u32(XXH_PRIME32_1); + + size_t i; + for (i = 0; i < XXH_STRIPE_LEN / sizeof(uint64x2_t); i++) { + + /* xacc[i] ^= (xacc[i] >> 47); */ + uint64x2_t acc_vec = xacc[i]; + uint64x2_t shifted = vshrq_n_u64(acc_vec, 47); + uint64x2_t data_vec = veorq_u64(acc_vec, shifted); + + /* xacc[i] ^= xsecret[i]; */ + uint8x16_t key_vec = vld1q_u8(xsecret + (i * 16)); + uint64x2_t data_key = veorq_u64(data_vec, vreinterpretq_u64_u8(key_vec)); + + /* xacc[i] *= XXH_PRIME32_1 */ + uint32x2_t data_key_lo, data_key_hi; + /* data_key_lo = (uint32x2_t) (xacc[i] & 0xFFFFFFFF); + * data_key_hi = (uint32x2_t) (xacc[i] >> 32); + * xacc[i] = UNDEFINED; */ + XXH_SPLIT_IN_PLACE(data_key, data_key_lo, data_key_hi); + { /* + * prod_hi = (data_key >> 32) * XXH_PRIME32_1; + * + * Avoid vmul_u32 + vshll_n_u32 since Clang 6 and 7 will + * incorrectly "optimize" this: + * tmp = vmul_u32(vmovn_u64(a), vmovn_u64(b)); + * shifted = vshll_n_u32(tmp, 32); + * to this: + * tmp = "vmulq_u64"(a, b); // no such thing! + * shifted = vshlq_n_u64(tmp, 32); + * + * However, unlike SSE, Clang lacks a 64-bit multiply routine + * for NEON, and it scalarizes two 64-bit multiplies instead. + * + * vmull_u32 has the same timing as vmul_u32, and it avoids + * this bug completely. + * See https://bugs.llvm.org/show_bug.cgi?id=39967 + */ + uint64x2_t prod_hi = vmull_u32(data_key_hi, prime); + /* xacc[i] = prod_hi << 32; */ + xacc[i] = vshlq_n_u64(prod_hi, 32); + /* xacc[i] += (prod_hi & 0xFFFFFFFF) * XXH_PRIME32_1; */ + xacc[i] = vmlal_u32(xacc[i], data_key_lo, prime); + + } + + } + + } + +} + + #endif + + #if (XXH_VECTOR == XXH_VSX) + +XXH_FORCE_INLINE void XXH3_accumulate_512_vsx(void *XXH_RESTRICT acc, + const void *XXH_RESTRICT input, + const void *XXH_RESTRICT secret) { + + xxh_u64x2 *const xacc = (xxh_u64x2 *)acc; /* presumed aligned */ + xxh_u64x2 const *const xinput = + (xxh_u64x2 const *)input; /* no alignment restriction */ + xxh_u64x2 const *const xsecret = + (xxh_u64x2 const *)secret; /* no alignment restriction */ + xxh_u64x2 const v32 = {32, 32}; + size_t i; + for (i = 0; i < XXH_STRIPE_LEN / sizeof(xxh_u64x2); i++) { + + /* data_vec = xinput[i]; */ + xxh_u64x2 const data_vec = XXH_vec_loadu(xinput + i); + /* key_vec = xsecret[i]; */ + xxh_u64x2 const key_vec = XXH_vec_loadu(xsecret + i); + xxh_u64x2 const data_key = data_vec ^ key_vec; + /* shuffled = (data_key << 32) | (data_key >> 32); */ + xxh_u32x4 const shuffled = (xxh_u32x4)vec_rl(data_key, v32); + /* product = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)shuffled & + * 0xFFFFFFFF); */ + xxh_u64x2 const product = XXH_vec_mulo((xxh_u32x4)data_key, shuffled); + xacc[i] += product; + + /* swap high and low halves */ + #ifdef __s390x__ + xacc[i] += vec_permi(data_vec, data_vec, 2); + #else + xacc[i] += vec_xxpermdi(data_vec, data_vec, 2); + #endif + + } + +} + +XXH_FORCE_INLINE void XXH3_scrambleAcc_vsx(void *XXH_RESTRICT acc, + const void *XXH_RESTRICT secret) { + + XXH_ASSERT((((size_t)acc) & 15) == 0); + + { + + xxh_u64x2 *const xacc = (xxh_u64x2 *)acc; + const xxh_u64x2 *const xsecret = (const xxh_u64x2 *)secret; + /* constants */ + xxh_u64x2 const v32 = {32, 32}; + xxh_u64x2 const v47 = {47, 47}; + xxh_u32x4 const prime = {XXH_PRIME32_1, XXH_PRIME32_1, XXH_PRIME32_1, + XXH_PRIME32_1}; + size_t i; + for (i = 0; i < XXH_STRIPE_LEN / sizeof(xxh_u64x2); i++) { + + /* xacc[i] ^= (xacc[i] >> 47); */ + xxh_u64x2 const acc_vec = xacc[i]; + xxh_u64x2 const data_vec = acc_vec ^ (acc_vec >> v47); + + /* xacc[i] ^= xsecret[i]; */ + xxh_u64x2 const key_vec = XXH_vec_loadu(xsecret + i); + xxh_u64x2 const data_key = data_vec ^ key_vec; + + /* xacc[i] *= XXH_PRIME32_1 */ + /* prod_lo = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)prime & + * 0xFFFFFFFF); */ + xxh_u64x2 const prod_even = XXH_vec_mule((xxh_u32x4)data_key, prime); + /* prod_hi = ((xxh_u64x2)data_key >> 32) * ((xxh_u64x2)prime >> 32); */ + xxh_u64x2 const prod_odd = XXH_vec_mulo((xxh_u32x4)data_key, prime); + xacc[i] = prod_odd + (prod_even << v32); + + } + + } + +} + + #endif + +/* scalar variants - universal */ + +XXH_FORCE_INLINE void XXH3_accumulate_512_scalar( + void *XXH_RESTRICT acc, const void *XXH_RESTRICT input, + const void *XXH_RESTRICT secret) { + + XXH_ALIGN(XXH_ACC_ALIGN) + xxh_u64 *const xacc = (xxh_u64 *)acc; /* presumed aligned */ + const xxh_u8 *const xinput = + (const xxh_u8 *)input; /* no alignment restriction */ + const xxh_u8 *const xsecret = + (const xxh_u8 *)secret; /* no alignment restriction */ + size_t i; + XXH_ASSERT(((size_t)acc & (XXH_ACC_ALIGN - 1)) == 0); + for (i = 0; i < XXH_ACC_NB; i++) { + + xxh_u64 const data_val = XXH_readLE64(xinput + 8 * i); + xxh_u64 const data_key = data_val ^ XXH_readLE64(xsecret + i * 8); + xacc[i ^ 1] += data_val; /* swap adjacent lanes */ + xacc[i] += XXH_mult32to64(data_key & 0xFFFFFFFF, data_key >> 32); + + } + +} + +XXH_FORCE_INLINE void XXH3_scrambleAcc_scalar(void *XXH_RESTRICT acc, + const void *XXH_RESTRICT secret) { + + XXH_ALIGN(XXH_ACC_ALIGN) + xxh_u64 *const xacc = (xxh_u64 *)acc; /* presumed aligned */ + const xxh_u8 *const xsecret = + (const xxh_u8 *)secret; /* no alignment restriction */ + size_t i; + XXH_ASSERT((((size_t)acc) & (XXH_ACC_ALIGN - 1)) == 0); + for (i = 0; i < XXH_ACC_NB; i++) { + + xxh_u64 const key64 = XXH_readLE64(xsecret + 8 * i); + xxh_u64 acc64 = xacc[i]; + acc64 = XXH_xorshift64(acc64, 47); + acc64 ^= key64; + acc64 *= XXH_PRIME32_1; + xacc[i] = acc64; + + } + +} + +XXH_FORCE_INLINE void XXH3_initCustomSecret_scalar( + void *XXH_RESTRICT customSecret, xxh_u64 seed64) { + + /* + * We need a separate pointer for the hack below, + * which requires a non-const pointer. + * Any decent compiler will optimize this out otherwise. + */ + const xxh_u8 *kSecretPtr = XXH3_kSecret; + XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0); + + #if defined(__clang__) && defined(__aarch64__) + /* + * UGLY HACK: + * Clang generates a bunch of MOV/MOVK pairs for aarch64, and they are + * placed sequentially, in order, at the top of the unrolled loop. + * + * While MOVK is great for generating constants (2 cycles for a 64-bit + * constant compared to 4 cycles for LDR), long MOVK chains stall the + * integer pipelines: + * I L S + * MOVK + * MOVK + * MOVK + * MOVK + * ADD + * SUB STR + * STR + * By forcing loads from memory (as the asm line causes Clang to assume + * that XXH3_kSecretPtr has been changed), the pipelines are used more + * efficiently: + * I L S + * LDR + * ADD LDR + * SUB STR + * STR + * XXH3_64bits_withSeed, len == 256, Snapdragon 835 + * without hack: 2654.4 MB/s + * with hack: 3202.9 MB/s + */ + __asm__("" : "+r"(kSecretPtr)); + #endif + /* + * Note: in debug mode, this overrides the asm optimization + * and Clang will emit MOVK chains again. + */ + XXH_ASSERT(kSecretPtr == XXH3_kSecret); + + { + + int const nbRounds = XXH_SECRET_DEFAULT_SIZE / 16; + int i; + for (i = 0; i < nbRounds; i++) { + + /* + * The asm hack causes Clang to assume that kSecretPtr aliases with + * customSecret, and on aarch64, this prevented LDP from merging two + * loads together for free. Putting the loads together before the stores + * properly generates LDP. + */ + xxh_u64 lo = XXH_readLE64(kSecretPtr + 16 * i) + seed64; + xxh_u64 hi = XXH_readLE64(kSecretPtr + 16 * i + 8) - seed64; + XXH_writeLE64((xxh_u8 *)customSecret + 16 * i, lo); + XXH_writeLE64((xxh_u8 *)customSecret + 16 * i + 8, hi); + + } + + } + +} + +typedef void (*XXH3_f_accumulate_512)(void *XXH_RESTRICT, const void *, + const void *); +typedef void (*XXH3_f_scrambleAcc)(void *XXH_RESTRICT, const void *); +typedef void (*XXH3_f_initCustomSecret)(void *XXH_RESTRICT, xxh_u64); + + #if (XXH_VECTOR == XXH_AVX512) + + #define XXH3_accumulate_512 XXH3_accumulate_512_avx512 + #define XXH3_scrambleAcc XXH3_scrambleAcc_avx512 + #define XXH3_initCustomSecret XXH3_initCustomSecret_avx512 + + #elif (XXH_VECTOR == XXH_AVX2) + + #define XXH3_accumulate_512 XXH3_accumulate_512_avx2 + #define XXH3_scrambleAcc XXH3_scrambleAcc_avx2 + #define XXH3_initCustomSecret XXH3_initCustomSecret_avx2 + + #elif (XXH_VECTOR == XXH_SSE2) + + #define XXH3_accumulate_512 XXH3_accumulate_512_sse2 + #define XXH3_scrambleAcc XXH3_scrambleAcc_sse2 + #define XXH3_initCustomSecret XXH3_initCustomSecret_sse2 + + #elif (XXH_VECTOR == XXH_NEON) + + #define XXH3_accumulate_512 XXH3_accumulate_512_neon + #define XXH3_scrambleAcc XXH3_scrambleAcc_neon + #define XXH3_initCustomSecret XXH3_initCustomSecret_scalar + + #elif (XXH_VECTOR == XXH_VSX) + + #define XXH3_accumulate_512 XXH3_accumulate_512_vsx + #define XXH3_scrambleAcc XXH3_scrambleAcc_vsx + #define XXH3_initCustomSecret XXH3_initCustomSecret_scalar + + #else /* scalar */ + + #define XXH3_accumulate_512 XXH3_accumulate_512_scalar + #define XXH3_scrambleAcc XXH3_scrambleAcc_scalar + #define XXH3_initCustomSecret XXH3_initCustomSecret_scalar + + #endif + + #ifndef XXH_PREFETCH_DIST + #ifdef __clang__ + #define XXH_PREFETCH_DIST 320 + #else + #if (XXH_VECTOR == XXH_AVX512) + #define XXH_PREFETCH_DIST 512 + #else + #define XXH_PREFETCH_DIST 384 + #endif + #endif /* __clang__ */ + #endif /* XXH_PREFETCH_DIST */ + +/* + * XXH3_accumulate() + * Loops over XXH3_accumulate_512(). + * Assumption: nbStripes will not overflow the secret size + */ +XXH_FORCE_INLINE void XXH3_accumulate(xxh_u64 *XXH_RESTRICT acc, + const xxh_u8 *XXH_RESTRICT input, + const xxh_u8 *XXH_RESTRICT secret, + size_t nbStripes, + XXH3_f_accumulate_512 f_acc512) { + + size_t n; + for (n = 0; n < nbStripes; n++) { + + const xxh_u8 *const in = input + n * XXH_STRIPE_LEN; + XXH_PREFETCH(in + XXH_PREFETCH_DIST); + f_acc512(acc, in, secret + n * XXH_SECRET_CONSUME_RATE); + + } + +} + +XXH_FORCE_INLINE void XXH3_hashLong_internal_loop( + xxh_u64 *XXH_RESTRICT acc, const xxh_u8 *XXH_RESTRICT input, size_t len, + const xxh_u8 *XXH_RESTRICT secret, size_t secretSize, + XXH3_f_accumulate_512 f_acc512, XXH3_f_scrambleAcc f_scramble) { + + size_t const nbStripesPerBlock = + (secretSize - XXH_STRIPE_LEN) / XXH_SECRET_CONSUME_RATE; + size_t const block_len = XXH_STRIPE_LEN * nbStripesPerBlock; + size_t const nb_blocks = (len - 1) / block_len; + + size_t n; + + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); + + for (n = 0; n < nb_blocks; n++) { + + XXH3_accumulate(acc, input + n * block_len, secret, nbStripesPerBlock, + f_acc512); + f_scramble(acc, secret + secretSize - XXH_STRIPE_LEN); + + } + + /* last partial block */ + XXH_ASSERT(len > XXH_STRIPE_LEN); + { + + size_t const nbStripes = + ((len - 1) - (block_len * nb_blocks)) / XXH_STRIPE_LEN; + XXH_ASSERT(nbStripes <= (secretSize / XXH_SECRET_CONSUME_RATE)); + XXH3_accumulate(acc, input + nb_blocks * block_len, secret, nbStripes, + f_acc512); + + /* last stripe */ + { + + const xxh_u8 *const p = input + len - XXH_STRIPE_LEN; + #define XXH_SECRET_LASTACC_START \ + 7 /* not aligned on 8, last secret is different from acc & scrambler */ + f_acc512(acc, p, + secret + secretSize - XXH_STRIPE_LEN - XXH_SECRET_LASTACC_START); + + } + + } + +} + +XXH_FORCE_INLINE xxh_u64 XXH3_mix2Accs(const xxh_u64 *XXH_RESTRICT acc, + const xxh_u8 *XXH_RESTRICT secret) { + + return XXH3_mul128_fold64(acc[0] ^ XXH_readLE64(secret), + acc[1] ^ XXH_readLE64(secret + 8)); + +} + +static XXH64_hash_t XXH3_mergeAccs(const xxh_u64 *XXH_RESTRICT acc, + const xxh_u8 *XXH_RESTRICT secret, + xxh_u64 start) { + + xxh_u64 result64 = start; + size_t i = 0; + + for (i = 0; i < 4; i++) { + + result64 += XXH3_mix2Accs(acc + 2 * i, secret + 16 * i); + #if defined(__clang__) /* Clang */ \ + && (defined(__arm__) || defined(__thumb__)) /* ARMv7 */ \ + && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \ + && !defined(XXH_ENABLE_AUTOVECTORIZE) /* Define to disable */ + /* + * UGLY HACK: + * Prevent autovectorization on Clang ARMv7-a. Exact same problem as + * the one in XXH3_len_129to240_64b. Speeds up shorter keys > 240b. + * XXH3_64bits, len == 256, Snapdragon 835: + * without hack: 2063.7 MB/s + * with hack: 2560.7 MB/s + */ + __asm__("" : "+r"(result64)); + #endif + + } + + return XXH3_avalanche(result64); + +} + + #define XXH3_INIT_ACC \ + { \ + \ + XXH_PRIME32_3, XXH_PRIME64_1, XXH_PRIME64_2, XXH_PRIME64_3, \ + XXH_PRIME64_4, XXH_PRIME32_2, XXH_PRIME64_5, XXH_PRIME32_1 \ + \ + } + +XXH_FORCE_INLINE XXH64_hash_t XXH3_hashLong_64b_internal( + const void *XXH_RESTRICT input, size_t len, const void *XXH_RESTRICT secret, + size_t secretSize, XXH3_f_accumulate_512 f_acc512, + XXH3_f_scrambleAcc f_scramble) { + + XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC; + + XXH3_hashLong_internal_loop(acc, (const xxh_u8 *)input, len, + (const xxh_u8 *)secret, secretSize, f_acc512, + f_scramble); + + /* converge into final hash */ + XXH_STATIC_ASSERT(sizeof(acc) == 64); + /* do not align on 8, so that the secret is different from the accumulator + */ + #define XXH_SECRET_MERGEACCS_START 11 + XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START); + return XXH3_mergeAccs(acc, + (const xxh_u8 *)secret + XXH_SECRET_MERGEACCS_START, + (xxh_u64)len * XXH_PRIME64_1); + +} + +/* + * It's important for performance that XXH3_hashLong is not inlined. + */ +XXH_NO_INLINE XXH64_hash_t XXH3_hashLong_64b_withSecret( + const void *XXH_RESTRICT input, size_t len, XXH64_hash_t seed64, + const xxh_u8 *XXH_RESTRICT secret, size_t secretLen) { + + (void)seed64; + return XXH3_hashLong_64b_internal(input, len, secret, secretLen, + XXH3_accumulate_512, XXH3_scrambleAcc); + +} + +/* + * It's important for performance that XXH3_hashLong is not inlined. + * Since the function is not inlined, the compiler may not be able to understand + * that, in some scenarios, its `secret` argument is actually a compile time + * constant. This variant enforces that the compiler can detect that, and uses + * this opportunity to streamline the generated code for better performance. + */ +XXH_NO_INLINE XXH64_hash_t XXH3_hashLong_64b_default( + const void *XXH_RESTRICT input, size_t len, XXH64_hash_t seed64, + const xxh_u8 *XXH_RESTRICT secret, size_t secretLen) { + + (void)seed64; + (void)secret; + (void)secretLen; + return XXH3_hashLong_64b_internal(input, len, XXH3_kSecret, + sizeof(XXH3_kSecret), XXH3_accumulate_512, + XXH3_scrambleAcc); + +} + +/* + * XXH3_hashLong_64b_withSeed(): + * Generate a custom key based on alteration of default XXH3_kSecret with the + * seed, and then use this key for long mode hashing. + * + * This operation is decently fast but nonetheless costs a little bit of time. + * Try to avoid it whenever possible (typically when seed==0). + * + * It's important for performance that XXH3_hashLong is not inlined. Not sure + * why (uop cache maybe?), but the difference is large and easily measurable. + */ +XXH_FORCE_INLINE XXH64_hash_t XXH3_hashLong_64b_withSeed_internal( + const void *input, size_t len, XXH64_hash_t seed, + XXH3_f_accumulate_512 f_acc512, XXH3_f_scrambleAcc f_scramble, + XXH3_f_initCustomSecret f_initSec) { + + if (seed == 0) + return XXH3_hashLong_64b_internal( + input, len, XXH3_kSecret, sizeof(XXH3_kSecret), f_acc512, f_scramble); + { + + XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE]; + f_initSec(secret, seed); + return XXH3_hashLong_64b_internal(input, len, secret, sizeof(secret), + f_acc512, f_scramble); + + } + +} + +/* + * It's important for performance that XXH3_hashLong is not inlined. + */ +XXH_NO_INLINE XXH64_hash_t XXH3_hashLong_64b_withSeed(const void * input, + size_t len, + XXH64_hash_t seed, + const xxh_u8 *secret, + size_t secretLen) { + + (void)secret; + (void)secretLen; + return XXH3_hashLong_64b_withSeed_internal( + input, len, seed, XXH3_accumulate_512, XXH3_scrambleAcc, + XXH3_initCustomSecret); + +} + +typedef XXH64_hash_t (*XXH3_hashLong64_f)(const void *XXH_RESTRICT, size_t, + XXH64_hash_t, + const xxh_u8 *XXH_RESTRICT, size_t); + +XXH_FORCE_INLINE XXH64_hash_t +XXH3_64bits_internal(const void *XXH_RESTRICT input, size_t len, + XXH64_hash_t seed64, const void *XXH_RESTRICT secret, + size_t secretLen, XXH3_hashLong64_f f_hashLong) { + + XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN); + /* + * If an action is to be taken if `secretLen` condition is not respected, + * it should be done here. + * For now, it's a contract pre-condition. + * Adding a check and a branch here would cost performance at every hash. + * Also, note that function signature doesn't offer room to return an error. + */ + if (len <= 16) + return XXH3_len_0to16_64b((const xxh_u8 *)input, len, + (const xxh_u8 *)secret, seed64); + if (len <= 128) + return XXH3_len_17to128_64b((const xxh_u8 *)input, len, + (const xxh_u8 *)secret, secretLen, seed64); + if (len <= XXH3_MIDSIZE_MAX) + return XXH3_len_129to240_64b((const xxh_u8 *)input, len, + (const xxh_u8 *)secret, secretLen, seed64); + return f_hashLong(input, len, seed64, (const xxh_u8 *)secret, secretLen); + +} + +/* === Public entry point === */ + +XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(const void *input, size_t len) { + + return XXH3_64bits_internal(input, len, 0, XXH3_kSecret, sizeof(XXH3_kSecret), + XXH3_hashLong_64b_default); + +} + +XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSecret(const void *input, + size_t len, + const void *secret, + size_t secretSize) { + + return XXH3_64bits_internal(input, len, 0, secret, secretSize, + XXH3_hashLong_64b_withSecret); + +} + +XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_withSeed(const void *input, size_t len, + XXH64_hash_t seed) { + + return XXH3_64bits_internal(input, len, seed, XXH3_kSecret, + sizeof(XXH3_kSecret), XXH3_hashLong_64b_withSeed); + +} + +/* === XXH3 streaming === */ + +/* + * Malloc's a pointer that is always aligned to align. + * + * This must be freed with `XXH_alignedFree()`. + * + * malloc typically guarantees 16 byte alignment on 64-bit systems and 8 byte + * alignment on 32-bit. This isn't enough for the 32 byte aligned loads in AVX2 + * or on 32-bit, the 16 byte aligned loads in SSE2 and NEON. + * + * This underalignment previously caused a rather obvious crash which went + * completely unnoticed due to XXH3_createState() not actually being tested. + * Credit to RedSpah for noticing this bug. + * + * The alignment is done manually: Functions like posix_memalign or _mm_malloc + * are avoided: To maintain portability, we would have to write a fallback + * like this anyways, and besides, testing for the existence of library + * functions without relying on external build tools is impossible. + * + * The method is simple: Overallocate, manually align, and store the offset + * to the original behind the returned pointer. + * + * Align must be a power of 2 and 8 <= align <= 128. + */ +static void *XXH_alignedMalloc(size_t s, size_t align) { + + XXH_ASSERT(align <= 128 && align >= 8); /* range check */ + XXH_ASSERT((align & (align - 1)) == 0); /* power of 2 */ + XXH_ASSERT(s != 0 && s < (s + align)); /* empty/overflow */ + { /* Overallocate to make room for manual realignment and an offset byte */ + xxh_u8 *base = (xxh_u8 *)XXH_malloc(s + align); + if (base != NULL) { + + /* + * Get the offset needed to align this pointer. + * + * Even if the returned pointer is aligned, there will always be + * at least one byte to store the offset to the original pointer. + */ + size_t offset = align - ((size_t)base & (align - 1)); /* base % align */ + /* Add the offset for the now-aligned pointer */ + xxh_u8 *ptr = base + offset; + + XXH_ASSERT((size_t)ptr % align == 0); + + /* Store the offset immediately before the returned pointer. */ + ptr[-1] = (xxh_u8)offset; + return ptr; + + } + + return NULL; + + } + +} + +/* + * Frees an aligned pointer allocated by XXH_alignedMalloc(). Don't pass + * normal malloc'd pointers, XXH_alignedMalloc has a specific data layout. + */ +static void XXH_alignedFree(void *p) { + + if (p != NULL) { + + xxh_u8 *ptr = (xxh_u8 *)p; + /* Get the offset byte we added in XXH_malloc. */ + xxh_u8 offset = ptr[-1]; + /* Free the original malloc'd pointer */ + xxh_u8 *base = ptr - offset; + XXH_free(base); + + } + +} + +XXH_PUBLIC_API XXH3_state_t *XXH3_createState(void) { + + XXH3_state_t *const state = + (XXH3_state_t *)XXH_alignedMalloc(sizeof(XXH3_state_t), 64); + if (state == NULL) return NULL; + XXH3_INITSTATE(state); + return state; + +} + +XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t *statePtr) { + + XXH_alignedFree(statePtr); + return XXH_OK; + +} + +XXH_PUBLIC_API void XXH3_copyState(XXH3_state_t * dst_state, + const XXH3_state_t *src_state) { + + memcpy(dst_state, src_state, sizeof(*dst_state)); + +} + +static void XXH3_64bits_reset_internal(XXH3_state_t *statePtr, + XXH64_hash_t seed, const void *secret, + size_t secretSize) { + + size_t const initStart = offsetof(XXH3_state_t, bufferedSize); + size_t const initLength = + offsetof(XXH3_state_t, nbStripesPerBlock) - initStart; + XXH_ASSERT(offsetof(XXH3_state_t, nbStripesPerBlock) > initStart); + XXH_ASSERT(statePtr != NULL); + /* set members from bufferedSize to nbStripesPerBlock (excluded) to 0 */ + memset((char *)statePtr + initStart, 0, initLength); + statePtr->acc[0] = XXH_PRIME32_3; + statePtr->acc[1] = XXH_PRIME64_1; + statePtr->acc[2] = XXH_PRIME64_2; + statePtr->acc[3] = XXH_PRIME64_3; + statePtr->acc[4] = XXH_PRIME64_4; + statePtr->acc[5] = XXH_PRIME32_2; + statePtr->acc[6] = XXH_PRIME64_5; + statePtr->acc[7] = XXH_PRIME32_1; + statePtr->seed = seed; + statePtr->extSecret = (const unsigned char *)secret; + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); + statePtr->secretLimit = secretSize - XXH_STRIPE_LEN; + statePtr->nbStripesPerBlock = statePtr->secretLimit / XXH_SECRET_CONSUME_RATE; + +} + +XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH3_state_t *statePtr) { + + if (statePtr == NULL) return XXH_ERROR; + XXH3_64bits_reset_internal(statePtr, 0, XXH3_kSecret, + XXH_SECRET_DEFAULT_SIZE); + return XXH_OK; + +} + +XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret( + XXH3_state_t *statePtr, const void *secret, size_t secretSize) { + + if (statePtr == NULL) return XXH_ERROR; + XXH3_64bits_reset_internal(statePtr, 0, secret, secretSize); + if (secret == NULL) return XXH_ERROR; + if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR; + return XXH_OK; + +} + +XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH3_state_t *statePtr, + XXH64_hash_t seed) { + + if (statePtr == NULL) return XXH_ERROR; + if (seed == 0) return XXH3_64bits_reset(statePtr); + if (seed != statePtr->seed) + XXH3_initCustomSecret(statePtr->customSecret, seed); + XXH3_64bits_reset_internal(statePtr, seed, NULL, XXH_SECRET_DEFAULT_SIZE); + return XXH_OK; + +} + +/* Note : when XXH3_consumeStripes() is invoked, + * there must be a guarantee that at least one more byte must be consumed from + * input + * so that the function can blindly consume all stripes using the "normal" + * secret segment */ +XXH_FORCE_INLINE void XXH3_consumeStripes( + xxh_u64 *XXH_RESTRICT acc, size_t *XXH_RESTRICT nbStripesSoFarPtr, + size_t nbStripesPerBlock, const xxh_u8 *XXH_RESTRICT input, + size_t nbStripes, const xxh_u8 *XXH_RESTRICT secret, size_t secretLimit, + XXH3_f_accumulate_512 f_acc512, XXH3_f_scrambleAcc f_scramble) { + + XXH_ASSERT(nbStripes <= + nbStripesPerBlock); /* can handle max 1 scramble per invocation */ + XXH_ASSERT(*nbStripesSoFarPtr < nbStripesPerBlock); + if (nbStripesPerBlock - *nbStripesSoFarPtr <= nbStripes) { + + /* need a scrambling operation */ + size_t const nbStripesToEndofBlock = nbStripesPerBlock - *nbStripesSoFarPtr; + size_t const nbStripesAfterBlock = nbStripes - nbStripesToEndofBlock; + XXH3_accumulate(acc, input, + secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, + nbStripesToEndofBlock, f_acc512); + f_scramble(acc, secret + secretLimit); + XXH3_accumulate(acc, input + nbStripesToEndofBlock * XXH_STRIPE_LEN, secret, + nbStripesAfterBlock, f_acc512); + *nbStripesSoFarPtr = nbStripesAfterBlock; + + } else { + + XXH3_accumulate(acc, input, + secret + nbStripesSoFarPtr[0] * XXH_SECRET_CONSUME_RATE, + nbStripes, f_acc512); + *nbStripesSoFarPtr += nbStripes; + + } + +} + +/* + * Both XXH3_64bits_update and XXH3_128bits_update use this routine. + */ +XXH_FORCE_INLINE XXH_errorcode XXH3_update(XXH3_state_t *state, + const xxh_u8 *input, size_t len, + XXH3_f_accumulate_512 f_acc512, + XXH3_f_scrambleAcc f_scramble) { + + if (input == NULL) + #if defined(XXH_ACCEPT_NULL_INPUT_POINTER) && \ + (XXH_ACCEPT_NULL_INPUT_POINTER >= 1) + return XXH_OK; + #else + return XXH_ERROR; + #endif + + { + + const xxh_u8 *const bEnd = input + len; + const unsigned char *const secret = + (state->extSecret == NULL) ? state->customSecret : state->extSecret; + + state->totalLen += len; + + if (state->bufferedSize + len <= + XXH3_INTERNALBUFFER_SIZE) { /* fill in tmp buffer */ + XXH_memcpy(state->buffer + state->bufferedSize, input, len); + state->bufferedSize += (XXH32_hash_t)len; + return XXH_OK; + + } + + /* total input is now > XXH3_INTERNALBUFFER_SIZE */ + + #define XXH3_INTERNALBUFFER_STRIPES \ + (XXH3_INTERNALBUFFER_SIZE / XXH_STRIPE_LEN) + XXH_STATIC_ASSERT(XXH3_INTERNALBUFFER_SIZE % XXH_STRIPE_LEN == + 0); /* clean multiple */ + + /* + * Internal buffer is partially filled (always, except at beginning) + * Complete it, then consume it. + */ + if (state->bufferedSize) { + + size_t const loadSize = XXH3_INTERNALBUFFER_SIZE - state->bufferedSize; + XXH_memcpy(state->buffer + state->bufferedSize, input, loadSize); + input += loadSize; + XXH3_consumeStripes(state->acc, &state->nbStripesSoFar, + state->nbStripesPerBlock, state->buffer, + XXH3_INTERNALBUFFER_STRIPES, secret, + state->secretLimit, f_acc512, f_scramble); + state->bufferedSize = 0; + + } + + XXH_ASSERT(input < bEnd); + + /* Consume input by a multiple of internal buffer size */ + if (input + XXH3_INTERNALBUFFER_SIZE < bEnd) { + + const xxh_u8 *const limit = bEnd - XXH3_INTERNALBUFFER_SIZE; + do { + + XXH3_consumeStripes(state->acc, &state->nbStripesSoFar, + state->nbStripesPerBlock, input, + XXH3_INTERNALBUFFER_STRIPES, secret, + state->secretLimit, f_acc512, f_scramble); + input += XXH3_INTERNALBUFFER_SIZE; + + } while (input < limit); + + /* for last partial stripe */ + memcpy(state->buffer + sizeof(state->buffer) - XXH_STRIPE_LEN, + input - XXH_STRIPE_LEN, XXH_STRIPE_LEN); + + } + + XXH_ASSERT(input < bEnd); + + /* Some remaining input (always) : buffer it */ + XXH_memcpy(state->buffer, input, (size_t)(bEnd - input)); + state->bufferedSize = (XXH32_hash_t)(bEnd - input); + + } + + return XXH_OK; + +} + +XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update(XXH3_state_t *state, + const void *input, size_t len) { + + return XXH3_update(state, (const xxh_u8 *)input, len, XXH3_accumulate_512, + XXH3_scrambleAcc); + +} + +XXH_FORCE_INLINE void XXH3_digest_long(XXH64_hash_t * acc, + const XXH3_state_t * state, + const unsigned char *secret) { + + /* + * Digest on a local copy. This way, the state remains unaltered, and it can + * continue ingesting more input afterwards. + */ + memcpy(acc, state->acc, sizeof(state->acc)); + if (state->bufferedSize >= XXH_STRIPE_LEN) { + + size_t const nbStripes = (state->bufferedSize - 1) / XXH_STRIPE_LEN; + size_t nbStripesSoFar = state->nbStripesSoFar; + XXH3_consumeStripes(acc, &nbStripesSoFar, state->nbStripesPerBlock, + state->buffer, nbStripes, secret, state->secretLimit, + XXH3_accumulate_512, XXH3_scrambleAcc); + /* last stripe */ + XXH3_accumulate_512(acc, + state->buffer + state->bufferedSize - XXH_STRIPE_LEN, + secret + state->secretLimit - XXH_SECRET_LASTACC_START); + + } else { /* bufferedSize < XXH_STRIPE_LEN */ + + xxh_u8 lastStripe[XXH_STRIPE_LEN]; + size_t const catchupSize = XXH_STRIPE_LEN - state->bufferedSize; + XXH_ASSERT(state->bufferedSize > + 0); /* there is always some input buffered */ + memcpy(lastStripe, state->buffer + sizeof(state->buffer) - catchupSize, + catchupSize); + memcpy(lastStripe + catchupSize, state->buffer, state->bufferedSize); + XXH3_accumulate_512(acc, lastStripe, + secret + state->secretLimit - XXH_SECRET_LASTACC_START); + + } + +} + +XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest(const XXH3_state_t *state) { + + const unsigned char *const secret = + (state->extSecret == NULL) ? state->customSecret : state->extSecret; + if (state->totalLen > XXH3_MIDSIZE_MAX) { + + XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[XXH_ACC_NB]; + XXH3_digest_long(acc, state, secret); + return XXH3_mergeAccs(acc, secret + XXH_SECRET_MERGEACCS_START, + (xxh_u64)state->totalLen * XXH_PRIME64_1); + + } + + /* totalLen <= XXH3_MIDSIZE_MAX: digesting a short input */ + if (state->seed) + return XXH3_64bits_withSeed(state->buffer, (size_t)state->totalLen, + state->seed); + return XXH3_64bits_withSecret(state->buffer, (size_t)(state->totalLen), + secret, state->secretLimit + XXH_STRIPE_LEN); + +} + + #define XXH_MIN(x, y) (((x) > (y)) ? (y) : (x)) + +XXH_PUBLIC_API void XXH3_generateSecret(void * secretBuffer, + const void *customSeed, + size_t customSeedSize) { + + XXH_ASSERT(secretBuffer != NULL); + if (customSeedSize == 0) { + + memcpy(secretBuffer, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE); + return; + + } + + XXH_ASSERT(customSeed != NULL); + + { + + size_t const segmentSize = sizeof(XXH128_hash_t); + size_t const nbSegments = XXH_SECRET_DEFAULT_SIZE / segmentSize; + XXH128_canonical_t scrambler; + XXH64_hash_t seeds[12]; + size_t segnb; + XXH_ASSERT(nbSegments == 12); + XXH_ASSERT(segmentSize * nbSegments == + XXH_SECRET_DEFAULT_SIZE); /* exact multiple */ + XXH128_canonicalFromHash(&scrambler, XXH128(customSeed, customSeedSize, 0)); + + /* + * Copy customSeed to seeds[], truncating or repeating as necessary. + */ + { + + size_t toFill = XXH_MIN(customSeedSize, sizeof(seeds)); + size_t filled = toFill; + memcpy(seeds, customSeed, toFill); + while (filled < sizeof(seeds)) { + + toFill = XXH_MIN(filled, sizeof(seeds) - filled); + memcpy((char *)seeds + filled, seeds, toFill); + filled += toFill; + + } + + } + + /* generate secret */ + memcpy(secretBuffer, &scrambler, sizeof(scrambler)); + for (segnb = 1; segnb < nbSegments; segnb++) { + + size_t const segmentStart = segnb * segmentSize; + XXH128_canonical_t segment; + XXH128_canonicalFromHash(&segment, + XXH128(&scrambler, sizeof(scrambler), + XXH_readLE64(seeds + segnb) + segnb)); + memcpy((char *)secretBuffer + segmentStart, &segment, sizeof(segment)); + + } + + } + +} + +/* ========================================== + * XXH3 128 bits (a.k.a XXH128) + * ========================================== + * XXH3's 128-bit variant has better mixing and strength than the 64-bit + * variant, even without counting the significantly larger output size. + * + * For example, extra steps are taken to avoid the seed-dependent collisions + * in 17-240 byte inputs (See XXH3_mix16B and XXH128_mix32B). + * + * This strength naturally comes at the cost of some speed, especially on short + * lengths. Note that longer hashes are about as fast as the 64-bit version + * due to it using only a slight modification of the 64-bit loop. + * + * XXH128 is also more oriented towards 64-bit machines. It is still extremely + * fast for a _128-bit_ hash on 32-bit (it usually clears XXH64). + */ + +XXH_FORCE_INLINE XXH128_hash_t XXH3_len_1to3_128b(const xxh_u8 *input, + size_t len, + const xxh_u8 *secret, + XXH64_hash_t seed) { + + /* A doubled version of 1to3_64b with different constants. */ + XXH_ASSERT(input != NULL); + XXH_ASSERT(1 <= len && len <= 3); + XXH_ASSERT(secret != NULL); + /* + * len = 1: combinedl = { input[0], 0x01, input[0], input[0] } + * len = 2: combinedl = { input[1], 0x02, input[0], input[1] } + * len = 3: combinedl = { input[2], 0x03, input[0], input[1] } + */ + { + + xxh_u8 const c1 = input[0]; + xxh_u8 const c2 = input[len >> 1]; + xxh_u8 const c3 = input[len - 1]; + xxh_u32 const combinedl = ((xxh_u32)c1 << 16) | ((xxh_u32)c2 << 24) | + ((xxh_u32)c3 << 0) | ((xxh_u32)len << 8); + xxh_u32 const combinedh = XXH_rotl32(XXH_swap32(combinedl), 13); + xxh_u64 const bitflipl = + (XXH_readLE32(secret) ^ XXH_readLE32(secret + 4)) + seed; + xxh_u64 const bitfliph = + (XXH_readLE32(secret + 8) ^ XXH_readLE32(secret + 12)) - seed; + xxh_u64 const keyed_lo = (xxh_u64)combinedl ^ bitflipl; + xxh_u64 const keyed_hi = (xxh_u64)combinedh ^ bitfliph; + XXH128_hash_t h128; + h128.low64 = XXH64_avalanche(keyed_lo); + h128.high64 = XXH64_avalanche(keyed_hi); + return h128; + + } + +} + +XXH_FORCE_INLINE XXH128_hash_t XXH3_len_4to8_128b(const xxh_u8 *input, + size_t len, + const xxh_u8 *secret, + XXH64_hash_t seed) { + + XXH_ASSERT(input != NULL); + XXH_ASSERT(secret != NULL); + XXH_ASSERT(4 <= len && len <= 8); + seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32; + { + + xxh_u32 const input_lo = XXH_readLE32(input); + xxh_u32 const input_hi = XXH_readLE32(input + len - 4); + xxh_u64 const input_64 = input_lo + ((xxh_u64)input_hi << 32); + xxh_u64 const bitflip = + (XXH_readLE64(secret + 16) ^ XXH_readLE64(secret + 24)) + seed; + xxh_u64 const keyed = input_64 ^ bitflip; + + /* Shift len to the left to ensure it is even, this avoids even multiplies. + */ + XXH128_hash_t m128 = XXH_mult64to128(keyed, XXH_PRIME64_1 + (len << 2)); + + m128.high64 += (m128.low64 << 1); + m128.low64 ^= (m128.high64 >> 3); + + m128.low64 = XXH_xorshift64(m128.low64, 35); + m128.low64 *= 0x9FB21C651E98DF25ULL; + m128.low64 = XXH_xorshift64(m128.low64, 28); + m128.high64 = XXH3_avalanche(m128.high64); + return m128; + + } + +} + +XXH_FORCE_INLINE XXH128_hash_t XXH3_len_9to16_128b(const xxh_u8 *input, + size_t len, + const xxh_u8 *secret, + XXH64_hash_t seed) { + + XXH_ASSERT(input != NULL); + XXH_ASSERT(secret != NULL); + XXH_ASSERT(9 <= len && len <= 16); + { + + xxh_u64 const bitflipl = + (XXH_readLE64(secret + 32) ^ XXH_readLE64(secret + 40)) - seed; + xxh_u64 const bitfliph = + (XXH_readLE64(secret + 48) ^ XXH_readLE64(secret + 56)) + seed; + xxh_u64 const input_lo = XXH_readLE64(input); + xxh_u64 input_hi = XXH_readLE64(input + len - 8); + XXH128_hash_t m128 = + XXH_mult64to128(input_lo ^ input_hi ^ bitflipl, XXH_PRIME64_1); + /* + * Put len in the middle of m128 to ensure that the length gets mixed to + * both the low and high bits in the 128x64 multiply below. + */ + m128.low64 += (xxh_u64)(len - 1) << 54; + input_hi ^= bitfliph; + /* + * Add the high 32 bits of input_hi to the high 32 bits of m128, then + * add the long product of the low 32 bits of input_hi and XXH_PRIME32_2 to + * the high 64 bits of m128. + * + * The best approach to this operation is different on 32-bit and 64-bit. + */ + if (sizeof(void *) < sizeof(xxh_u64)) { /* 32-bit */ + /* + * 32-bit optimized version, which is more readable. + * + * On 32-bit, it removes an ADC and delays a dependency between the two + * halves of m128.high64, but it generates an extra mask on 64-bit. + */ + m128.high64 += (input_hi & 0xFFFFFFFF00000000ULL) + + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2); + + } else { + + /* + * 64-bit optimized (albeit more confusing) version. + * + * Uses some properties of addition and multiplication to remove the mask: + * + * Let: + * a = input_hi.lo = (input_hi & 0x00000000FFFFFFFF) + * b = input_hi.hi = (input_hi & 0xFFFFFFFF00000000) + * c = XXH_PRIME32_2 + * + * a + (b * c) + * Inverse Property: x + y - x == y + * a + (b * (1 + c - 1)) + * Distributive Property: x * (y + z) == (x * y) + (x * z) + * a + (b * 1) + (b * (c - 1)) + * Identity Property: x * 1 == x + * a + b + (b * (c - 1)) + * + * Substitute a, b, and c: + * input_hi.hi + input_hi.lo + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - + * 1)) + * + * Since input_hi.hi + input_hi.lo == input_hi, we get this: + * input_hi + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1)) + */ + m128.high64 += + input_hi + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2 - 1); + + } + + /* m128 ^= XXH_swap64(m128 >> 64); */ + m128.low64 ^= XXH_swap64(m128.high64); + + { /* 128x64 multiply: h128 = m128 * XXH_PRIME64_2; */ + XXH128_hash_t h128 = XXH_mult64to128(m128.low64, XXH_PRIME64_2); + h128.high64 += m128.high64 * XXH_PRIME64_2; + + h128.low64 = XXH3_avalanche(h128.low64); + h128.high64 = XXH3_avalanche(h128.high64); + return h128; + + } + + } + +} + +/* + * Assumption: `secret` size is >= XXH3_SECRET_SIZE_MIN + */ +XXH_FORCE_INLINE XXH128_hash_t XXH3_len_0to16_128b(const xxh_u8 *input, + size_t len, + const xxh_u8 *secret, + XXH64_hash_t seed) { + + XXH_ASSERT(len <= 16); + { + + if (len > 8) return XXH3_len_9to16_128b(input, len, secret, seed); + if (len >= 4) return XXH3_len_4to8_128b(input, len, secret, seed); + if (len) return XXH3_len_1to3_128b(input, len, secret, seed); + { + + XXH128_hash_t h128; + xxh_u64 const bitflipl = + XXH_readLE64(secret + 64) ^ XXH_readLE64(secret + 72); + xxh_u64 const bitfliph = + XXH_readLE64(secret + 80) ^ XXH_readLE64(secret + 88); + h128.low64 = XXH64_avalanche(seed ^ bitflipl); + h128.high64 = XXH64_avalanche(seed ^ bitfliph); + return h128; + + } + + } + +} + +/* + * A bit slower than XXH3_mix16B, but handles multiply by zero better. + */ +XXH_FORCE_INLINE XXH128_hash_t XXH128_mix32B(XXH128_hash_t acc, + const xxh_u8 *input_1, + const xxh_u8 *input_2, + const xxh_u8 *secret, + XXH64_hash_t seed) { + + acc.low64 += XXH3_mix16B(input_1, secret + 0, seed); + acc.low64 ^= XXH_readLE64(input_2) + XXH_readLE64(input_2 + 8); + acc.high64 += XXH3_mix16B(input_2, secret + 16, seed); + acc.high64 ^= XXH_readLE64(input_1) + XXH_readLE64(input_1 + 8); + return acc; + +} + +XXH_FORCE_INLINE XXH128_hash_t XXH3_len_17to128_128b( + const xxh_u8 *XXH_RESTRICT input, size_t len, + const xxh_u8 *XXH_RESTRICT secret, size_t secretSize, XXH64_hash_t seed) { + + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); + (void)secretSize; + XXH_ASSERT(16 < len && len <= 128); + + { + + XXH128_hash_t acc; + acc.low64 = len * XXH_PRIME64_1; + acc.high64 = 0; + if (len > 32) { + + if (len > 64) { + + if (len > 96) { + + acc = XXH128_mix32B(acc, input + 48, input + len - 64, secret + 96, + seed); + + } + + acc = + XXH128_mix32B(acc, input + 32, input + len - 48, secret + 64, seed); + + } + + acc = XXH128_mix32B(acc, input + 16, input + len - 32, secret + 32, seed); + + } + + acc = XXH128_mix32B(acc, input, input + len - 16, secret, seed); + { + + XXH128_hash_t h128; + h128.low64 = acc.low64 + acc.high64; + h128.high64 = (acc.low64 * XXH_PRIME64_1) + (acc.high64 * XXH_PRIME64_4) + + ((len - seed) * XXH_PRIME64_2); + h128.low64 = XXH3_avalanche(h128.low64); + h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64); + return h128; + + } + + } + +} + +XXH_NO_INLINE XXH128_hash_t XXH3_len_129to240_128b( + const xxh_u8 *XXH_RESTRICT input, size_t len, + const xxh_u8 *XXH_RESTRICT secret, size_t secretSize, XXH64_hash_t seed) { + + XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); + (void)secretSize; + XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX); + + { + + XXH128_hash_t acc; + int const nbRounds = (int)len / 32; + int i; + acc.low64 = len * XXH_PRIME64_1; + acc.high64 = 0; + for (i = 0; i < 4; i++) { + + acc = XXH128_mix32B(acc, input + (32 * i), input + (32 * i) + 16, + secret + (32 * i), seed); + + } + + acc.low64 = XXH3_avalanche(acc.low64); + acc.high64 = XXH3_avalanche(acc.high64); + XXH_ASSERT(nbRounds >= 4); + for (i = 4; i < nbRounds; i++) { + + acc = XXH128_mix32B(acc, input + (32 * i), input + (32 * i) + 16, + secret + XXH3_MIDSIZE_STARTOFFSET + (32 * (i - 4)), + seed); + + } + + /* last bytes */ + acc = XXH128_mix32B( + acc, input + len - 16, input + len - 32, + secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET - 16, + 0ULL - seed); + + { + + XXH128_hash_t h128; + h128.low64 = acc.low64 + acc.high64; + h128.high64 = (acc.low64 * XXH_PRIME64_1) + (acc.high64 * XXH_PRIME64_4) + + ((len - seed) * XXH_PRIME64_2); + h128.low64 = XXH3_avalanche(h128.low64); + h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64); + return h128; + + } + + } + +} + +XXH_FORCE_INLINE XXH128_hash_t XXH3_hashLong_128b_internal( + const void *XXH_RESTRICT input, size_t len, + const xxh_u8 *XXH_RESTRICT secret, size_t secretSize, + XXH3_f_accumulate_512 f_acc512, XXH3_f_scrambleAcc f_scramble) { + + XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC; + + XXH3_hashLong_internal_loop(acc, (const xxh_u8 *)input, len, secret, + secretSize, f_acc512, f_scramble); + + /* converge into final hash */ + XXH_STATIC_ASSERT(sizeof(acc) == 64); + XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START); + { + + XXH128_hash_t h128; + h128.low64 = XXH3_mergeAccs(acc, secret + XXH_SECRET_MERGEACCS_START, + (xxh_u64)len * XXH_PRIME64_1); + h128.high64 = XXH3_mergeAccs( + acc, secret + secretSize - sizeof(acc) - XXH_SECRET_MERGEACCS_START, + ~((xxh_u64)len * XXH_PRIME64_2)); + return h128; + + } + +} + +/* + * It's important for performance that XXH3_hashLong is not inlined. + */ +XXH_NO_INLINE XXH128_hash_t XXH3_hashLong_128b_default( + const void *XXH_RESTRICT input, size_t len, XXH64_hash_t seed64, + const void *XXH_RESTRICT secret, size_t secretLen) { + + (void)seed64; + (void)secret; + (void)secretLen; + return XXH3_hashLong_128b_internal(input, len, XXH3_kSecret, + sizeof(XXH3_kSecret), XXH3_accumulate_512, + XXH3_scrambleAcc); + +} + +/* + * It's important for performance that XXH3_hashLong is not inlined. + */ +XXH_NO_INLINE XXH128_hash_t XXH3_hashLong_128b_withSecret( + const void *XXH_RESTRICT input, size_t len, XXH64_hash_t seed64, + const void *XXH_RESTRICT secret, size_t secretLen) { + + (void)seed64; + return XXH3_hashLong_128b_internal(input, len, (const xxh_u8 *)secret, + secretLen, XXH3_accumulate_512, + XXH3_scrambleAcc); + +} + +XXH_FORCE_INLINE XXH128_hash_t XXH3_hashLong_128b_withSeed_internal( + const void *XXH_RESTRICT input, size_t len, XXH64_hash_t seed64, + XXH3_f_accumulate_512 f_acc512, XXH3_f_scrambleAcc f_scramble, + XXH3_f_initCustomSecret f_initSec) { + + if (seed64 == 0) + return XXH3_hashLong_128b_internal( + input, len, XXH3_kSecret, sizeof(XXH3_kSecret), f_acc512, f_scramble); + { + + XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE]; + f_initSec(secret, seed64); + return XXH3_hashLong_128b_internal(input, len, (const xxh_u8 *)secret, + sizeof(secret), f_acc512, f_scramble); + + } + +} + +/* + * It's important for performance that XXH3_hashLong is not inlined. + */ +XXH_NO_INLINE XXH128_hash_t +XXH3_hashLong_128b_withSeed(const void *input, size_t len, XXH64_hash_t seed64, + const void *XXH_RESTRICT secret, size_t secretLen) { + + (void)secret; + (void)secretLen; + return XXH3_hashLong_128b_withSeed_internal( + input, len, seed64, XXH3_accumulate_512, XXH3_scrambleAcc, + XXH3_initCustomSecret); + +} + +typedef XXH128_hash_t (*XXH3_hashLong128_f)(const void *XXH_RESTRICT, size_t, + XXH64_hash_t, + const void *XXH_RESTRICT, size_t); + +XXH_FORCE_INLINE XXH128_hash_t +XXH3_128bits_internal(const void *input, size_t len, XXH64_hash_t seed64, + const void *XXH_RESTRICT secret, size_t secretLen, + XXH3_hashLong128_f f_hl128) { + + XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN); + /* + * If an action is to be taken if `secret` conditions are not respected, + * it should be done here. + * For now, it's a contract pre-condition. + * Adding a check and a branch here would cost performance at every hash. + */ + if (len <= 16) + return XXH3_len_0to16_128b((const xxh_u8 *)input, len, + (const xxh_u8 *)secret, seed64); + if (len <= 128) + return XXH3_len_17to128_128b((const xxh_u8 *)input, len, + (const xxh_u8 *)secret, secretLen, seed64); + if (len <= XXH3_MIDSIZE_MAX) + return XXH3_len_129to240_128b((const xxh_u8 *)input, len, + (const xxh_u8 *)secret, secretLen, seed64); + return f_hl128(input, len, seed64, secret, secretLen); + +} + +/* === Public XXH128 API === */ + +XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(const void *input, size_t len) { + + return XXH3_128bits_internal(input, len, 0, XXH3_kSecret, + sizeof(XXH3_kSecret), + XXH3_hashLong_128b_default); + +} + +XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSecret(const void *input, + size_t len, + const void *secret, + size_t secretSize) { + + return XXH3_128bits_internal(input, len, 0, (const xxh_u8 *)secret, + secretSize, XXH3_hashLong_128b_withSecret); + +} + +XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_withSeed(const void * input, + size_t len, + XXH64_hash_t seed) { + + return XXH3_128bits_internal(input, len, seed, XXH3_kSecret, + sizeof(XXH3_kSecret), + XXH3_hashLong_128b_withSeed); + +} + +XXH_PUBLIC_API XXH128_hash_t XXH128(const void *input, size_t len, + XXH64_hash_t seed) { + + return XXH3_128bits_withSeed(input, len, seed); + +} + +/* === XXH3 128-bit streaming === */ + +/* + * All the functions are actually the same as for 64-bit streaming variant. + * The only difference is the finalizatiom routine. + */ + +static void XXH3_128bits_reset_internal(XXH3_state_t *statePtr, + XXH64_hash_t seed, const void *secret, + size_t secretSize) { + + XXH3_64bits_reset_internal(statePtr, seed, secret, secretSize); + +} + +XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH3_state_t *statePtr) { + + if (statePtr == NULL) return XXH_ERROR; + XXH3_128bits_reset_internal(statePtr, 0, XXH3_kSecret, + XXH_SECRET_DEFAULT_SIZE); + return XXH_OK; + +} + +XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret( + XXH3_state_t *statePtr, const void *secret, size_t secretSize) { + + if (statePtr == NULL) return XXH_ERROR; + XXH3_128bits_reset_internal(statePtr, 0, secret, secretSize); + if (secret == NULL) return XXH_ERROR; + if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR; + return XXH_OK; + +} + +XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH3_state_t *statePtr, + XXH64_hash_t seed) { + + if (statePtr == NULL) return XXH_ERROR; + if (seed == 0) return XXH3_128bits_reset(statePtr); + if (seed != statePtr->seed) + XXH3_initCustomSecret(statePtr->customSecret, seed); + XXH3_128bits_reset_internal(statePtr, seed, NULL, XXH_SECRET_DEFAULT_SIZE); + return XXH_OK; + +} + +XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update(XXH3_state_t *state, + const void * input, + size_t len) { + + return XXH3_update(state, (const xxh_u8 *)input, len, XXH3_accumulate_512, + XXH3_scrambleAcc); + +} + +XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest(const XXH3_state_t *state) { + + const unsigned char *const secret = + (state->extSecret == NULL) ? state->customSecret : state->extSecret; + if (state->totalLen > XXH3_MIDSIZE_MAX) { + + XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[XXH_ACC_NB]; + XXH3_digest_long(acc, state, secret); + XXH_ASSERT(state->secretLimit + XXH_STRIPE_LEN >= + sizeof(acc) + XXH_SECRET_MERGEACCS_START); + { + + XXH128_hash_t h128; + h128.low64 = XXH3_mergeAccs(acc, secret + XXH_SECRET_MERGEACCS_START, + (xxh_u64)state->totalLen * XXH_PRIME64_1); + h128.high64 = + XXH3_mergeAccs(acc, + secret + state->secretLimit + XXH_STRIPE_LEN - + sizeof(acc) - XXH_SECRET_MERGEACCS_START, + ~((xxh_u64)state->totalLen * XXH_PRIME64_2)); + return h128; + + } + + } + + /* len <= XXH3_MIDSIZE_MAX : short code */ + if (state->seed) + return XXH3_128bits_withSeed(state->buffer, (size_t)state->totalLen, + state->seed); + return XXH3_128bits_withSecret(state->buffer, (size_t)(state->totalLen), + secret, state->secretLimit + XXH_STRIPE_LEN); + +} + + /* 128-bit utility functions */ + + #include <string.h> /* memcmp, memcpy */ + +/* return : 1 is equal, 0 if different */ +XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2) { + + /* note : XXH128_hash_t is compact, it has no padding byte */ + return !(memcmp(&h1, &h2, sizeof(h1))); + +} + +/* This prototype is compatible with stdlib's qsort(). + * return : >0 if *h128_1 > *h128_2 + * <0 if *h128_1 < *h128_2 + * =0 if *h128_1 == *h128_2 */ +XXH_PUBLIC_API int XXH128_cmp(const void *h128_1, const void *h128_2) { + + XXH128_hash_t const h1 = *(const XXH128_hash_t *)h128_1; + XXH128_hash_t const h2 = *(const XXH128_hash_t *)h128_2; + int const hcmp = (h1.high64 > h2.high64) - (h2.high64 > h1.high64); + /* note : bets that, in most cases, hash values are different */ + if (hcmp) return hcmp; + return (h1.low64 > h2.low64) - (h2.low64 > h1.low64); + +} + +/*====== Canonical representation ======*/ +XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH128_canonical_t *dst, + XXH128_hash_t hash) { + + XXH_STATIC_ASSERT(sizeof(XXH128_canonical_t) == sizeof(XXH128_hash_t)); + if (XXH_CPU_LITTLE_ENDIAN) { + + hash.high64 = XXH_swap64(hash.high64); + hash.low64 = XXH_swap64(hash.low64); + + } + + memcpy(dst, &hash.high64, sizeof(hash.high64)); + memcpy((char *)dst + sizeof(hash.high64), &hash.low64, sizeof(hash.low64)); + +} + +XXH_PUBLIC_API XXH128_hash_t +XXH128_hashFromCanonical(const XXH128_canonical_t *src) { + + XXH128_hash_t h; + h.high64 = XXH_readBE64(src); + h.low64 = XXH_readBE64(src->digest + 8); + return h; + +} + + /* Pop our optimization override from above */ + #if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \ + && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \ + && defined(__OPTIMIZE__) && \ + !defined(__OPTIMIZE_SIZE__) /* respect -O0 and -Os */ + #pragma GCC pop_options + #endif #endif /* XXH_NO_LONG_LONG */ |