| 1 |
#include <string.h> |
| 2 |
|
| 3 |
#ifdef __wasm_simd128__ |
| 4 |
#include <wasm_simd128.h> |
| 5 |
#endif |
| 6 |
|
| 7 |
int memcmp(const void *vl, const void *vr, size_t n) |
| 8 |
{ |
| 9 |
#if defined(__wasm_simd128__) && defined(__wasilibc_simd_string) |
| 10 |
if (n >= sizeof(v128_t)) { |
| 11 |
// memcmp is allowed to read up to n bytes from each object. |
| 12 |
// Find the first different character in the objects. |
| 13 |
// Unaligned loads handle the case where the objects |
| 14 |
// have mismatching alignments. |
| 15 |
const v128_t *v1 = (v128_t *)vl; |
| 16 |
const v128_t *v2 = (v128_t *)vr; |
| 17 |
while (n) { |
| 18 |
const v128_t cmp = wasm_i8x16_eq(wasm_v128_load(v1), wasm_v128_load(v2)); |
| 19 |
// Bitmask is slow on AArch64, all_true is much faster. |
| 20 |
if (!wasm_i8x16_all_true(cmp)) { |
| 21 |
// Find the offset of the first zero bit (little-endian). |
| 22 |
size_t ctz = __builtin_ctz(~wasm_i8x16_bitmask(cmp)); |
| 23 |
const unsigned char *u1 = (unsigned char *)v1 + ctz; |
| 24 |
const unsigned char *u2 = (unsigned char *)v2 + ctz; |
| 25 |
// This may help the compiler if the function is inlined. |
| 26 |
__builtin_assume(*u1 - *u2 != 0); |
| 27 |
return *u1 - *u2; |
| 28 |
} |
| 29 |
// This makes n a multiple of sizeof(v128_t) |
| 30 |
// for every iteration except the first. |
| 31 |
size_t align = (n - 1) % sizeof(v128_t) + 1; |
| 32 |
v1 = (v128_t *)((char *)v1 + align); |
| 33 |
v2 = (v128_t *)((char *)v2 + align); |
| 34 |
n -= align; |
| 35 |
} |
| 36 |
return 0; |
| 37 |
} |
| 38 |
#endif |
| 39 |
|
| 40 |
const unsigned char *l=vl, *r=vr; |
| 41 |
for (; n && *l == *r; n--, l++, r++); |
| 42 |
return n ? *l-*r : 0; |
| 43 |
} |