1
/* chunkset_ssse3.c -- SSSE3 inline functions to copy small data chunks.
2
* For conditions of distribution and use, see copyright notice in zlib.h
7
/* This requires SSE2 support. While it's implicit with SSSE3, we can minimize
8
* code size by sharing the chunkcopy functions, which will certainly compile
9
* to identical machine code */
10
#if defined(X86_SSSE3) && defined(X86_SSE2)
12
#include "../generic/chunk_permute_table.h"
14
typedef __m128i chunk_t;
18
#define HAVE_CHUNKMEMSET_2
19
#define HAVE_CHUNKMEMSET_4
20
#define HAVE_CHUNKMEMSET_8
23
#define HAVE_CHUNKUNROLL
25
static const lut_rem_pair perm_idx_lut[13] = {
27
{0, 0}, /* don't care */
31
{0 * 32, 0}, /* don't care */
42
static inline void chunkmemset_2(uint8_t *from, chunk_t *chunk) {
44
memcpy(&tmp, from, sizeof(tmp));
45
*chunk = _mm_set1_epi16(tmp);
48
static inline void chunkmemset_4(uint8_t *from, chunk_t *chunk) {
50
memcpy(&tmp, from, sizeof(tmp));
51
*chunk = _mm_set1_epi32(tmp);
54
static inline void chunkmemset_8(uint8_t *from, chunk_t *chunk) {
56
memcpy(&tmp, from, sizeof(tmp));
57
*chunk = _mm_set1_epi64x(tmp);
60
static inline void loadchunk(uint8_t const *s, chunk_t *chunk) {
61
*chunk = _mm_loadu_si128((__m128i *)s);
64
static inline void storechunk(uint8_t *out, chunk_t *chunk) {
65
_mm_storeu_si128((__m128i *)out, *chunk);
68
static inline chunk_t GET_CHUNK_MAG(uint8_t *buf, uint32_t *chunk_rem, uint32_t dist) {
69
lut_rem_pair lut_rem = perm_idx_lut[dist - 3];
70
__m128i perm_vec, ret_vec;
72
* This is _not_ to subvert the memory sanitizer but to instead unpoison some
73
* bytes we willingly and purposefully load uninitialized that we swizzle over
74
* in a vector register, anyway. If what we assume is wrong about what is used,
75
* the memory sanitizer will still usefully flag it */
76
__msan_unpoison(buf + dist, 16 - dist);
77
ret_vec = _mm_loadu_si128((__m128i*)buf);
78
*chunk_rem = lut_rem.remval;
80
perm_vec = _mm_load_si128((__m128i*)(permute_table + lut_rem.idx));
81
ret_vec = _mm_shuffle_epi8(ret_vec, perm_vec);
86
extern uint8_t* chunkcopy_sse2(uint8_t *out, uint8_t const *from, unsigned len);
87
extern uint8_t* chunkunroll_sse2(uint8_t *out, unsigned *dist, unsigned *len);
89
#define CHUNKSIZE chunksize_ssse3
90
#define CHUNKMEMSET chunkmemset_ssse3
91
#define CHUNKMEMSET_SAFE chunkmemset_safe_ssse3
92
#define CHUNKCOPY chunkcopy_sse2
93
#define CHUNKUNROLL chunkunroll_sse2
95
#include "chunkset_tpl.h"
97
#define INFLATE_FAST inflate_fast_ssse3
99
#include "inffast_tpl.h"