github_mirrors/vectorscan
3
0
Fork 1
mirror of https://github.com/VectorCamp/vectorscan.git synced 2025-11-18 18:20:35 +03:00
Code Issues Packages Projects Releases Wiki Activity

fix compilation with clang and some incomplete/wrong implementations for arm this time

Browse Source
This commit is contained in:
Konstantinos Margaritis
2021-12-06 18:22:58 +00:00
parent c68e46489b
commit ef2bc5cfbc
2 changed files with 264 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files

238
src/util/arch/arm/simd_utils.h
View File

@@ -122,24 +122,252 @@ m128 sub_2x64(m128 a, m128 b) {
return (m128) vsubq_u64((uint64x2_t)a, (uint64x2_t)b);
}
static really_really_inline
static really_inline
m128 lshift_m128(m128 a, unsigned b) {
return (m128) vshlq_n_u32((uint32x4_t)a, b);
#if defined(HAVE__BUILTIN_CONSTANT_P)
if (__builtin_constant_p(b)) {
return (m128) vshlq_n_u32((uint32x4_t)a, b);
}
#endif
#define CASE_LSHIFT_m128(a, offset) case offset: return (m128)vshlq_n_u32((int8x16_t)(a), (offset)); break;
switch (b) {
case 0: return a; break;
CASE_LSHIFT_m128(a, 1);
CASE_LSHIFT_m128(a, 2);
CASE_LSHIFT_m128(a, 3);
CASE_LSHIFT_m128(a, 4);
CASE_LSHIFT_m128(a, 5);
CASE_LSHIFT_m128(a, 6);
CASE_LSHIFT_m128(a, 7);
CASE_LSHIFT_m128(a, 8);
CASE_LSHIFT_m128(a, 9);
CASE_LSHIFT_m128(a, 10);
CASE_LSHIFT_m128(a, 11);
CASE_LSHIFT_m128(a, 12);
CASE_LSHIFT_m128(a, 13);
CASE_LSHIFT_m128(a, 14);
CASE_LSHIFT_m128(a, 15);
CASE_LSHIFT_m128(a, 16);
CASE_LSHIFT_m128(a, 17);
CASE_LSHIFT_m128(a, 18);
CASE_LSHIFT_m128(a, 19);
CASE_LSHIFT_m128(a, 20);
CASE_LSHIFT_m128(a, 21);
CASE_LSHIFT_m128(a, 22);
CASE_LSHIFT_m128(a, 23);
CASE_LSHIFT_m128(a, 24);
CASE_LSHIFT_m128(a, 25);
CASE_LSHIFT_m128(a, 26);
CASE_LSHIFT_m128(a, 27);
CASE_LSHIFT_m128(a, 28);
CASE_LSHIFT_m128(a, 29);
CASE_LSHIFT_m128(a, 30);
CASE_LSHIFT_m128(a, 31);
default: return zeroes128(); break;
}
#undef CASE_LSHIFT_m128
}
static really_really_inline
m128 rshift_m128(m128 a, unsigned b) {
return (m128) vshrq_n_u32((uint32x4_t)a, b);
#if defined(HAVE__BUILTIN_CONSTANT_P)
if (__builtin_constant_p(b)) {
return (m128) vshrq_n_u32((uint32x4_t)a, b);
}
#endif
#define CASE_RSHIFT_m128(a, offset) case offset: return (m128)vshrq_n_u32((int8x16_t)(a), (offset)); break;
switch (b) {
case 0: return a; break;
CASE_RSHIFT_m128(a, 1);
CASE_RSHIFT_m128(a, 2);
CASE_RSHIFT_m128(a, 3);
CASE_RSHIFT_m128(a, 4);
CASE_RSHIFT_m128(a, 5);
CASE_RSHIFT_m128(a, 6);
CASE_RSHIFT_m128(a, 7);
CASE_RSHIFT_m128(a, 8);
CASE_RSHIFT_m128(a, 9);
CASE_RSHIFT_m128(a, 10);
CASE_RSHIFT_m128(a, 11);
CASE_RSHIFT_m128(a, 12);
CASE_RSHIFT_m128(a, 13);
CASE_RSHIFT_m128(a, 14);
CASE_RSHIFT_m128(a, 15);
CASE_RSHIFT_m128(a, 16);
CASE_RSHIFT_m128(a, 17);
CASE_RSHIFT_m128(a, 18);
CASE_RSHIFT_m128(a, 19);
CASE_RSHIFT_m128(a, 20);
CASE_RSHIFT_m128(a, 21);
CASE_RSHIFT_m128(a, 22);
CASE_RSHIFT_m128(a, 23);
CASE_RSHIFT_m128(a, 24);
CASE_RSHIFT_m128(a, 25);
CASE_RSHIFT_m128(a, 26);
CASE_RSHIFT_m128(a, 27);
CASE_RSHIFT_m128(a, 28);
CASE_RSHIFT_m128(a, 29);
CASE_RSHIFT_m128(a, 30);
CASE_RSHIFT_m128(a, 31);
default: return zeroes128(); break;
}
#undef CASE_RSHIFT_m128
}
static really_really_inline
m128 lshift64_m128(m128 a, unsigned b) {
return (m128) vshlq_n_u64((uint64x2_t)a, b);
#if defined(HAVE__BUILTIN_CONSTANT_P)
if (__builtin_constant_p(b)) {
return (m128) vshlq_n_u64((uint64x2_t)a, b);
}
#endif
#define CASE_LSHIFT64_m128(a, offset) case offset: return (m128)vshlq_n_u64((int8x16_t)(a), (offset)); break;
switch (b) {
case 0: return a; break;
CASE_LSHIFT64_m128(a, 1);
CASE_LSHIFT64_m128(a, 2);
CASE_LSHIFT64_m128(a, 3);
CASE_LSHIFT64_m128(a, 4);
CASE_LSHIFT64_m128(a, 5);
CASE_LSHIFT64_m128(a, 6);
CASE_LSHIFT64_m128(a, 7);
CASE_LSHIFT64_m128(a, 8);
CASE_LSHIFT64_m128(a, 9);
CASE_LSHIFT64_m128(a, 10);
CASE_LSHIFT64_m128(a, 11);
CASE_LSHIFT64_m128(a, 12);
CASE_LSHIFT64_m128(a, 13);
CASE_LSHIFT64_m128(a, 14);
CASE_LSHIFT64_m128(a, 15);
CASE_LSHIFT64_m128(a, 16);
CASE_LSHIFT64_m128(a, 17);
CASE_LSHIFT64_m128(a, 18);
CASE_LSHIFT64_m128(a, 19);
CASE_LSHIFT64_m128(a, 20);
CASE_LSHIFT64_m128(a, 21);
CASE_LSHIFT64_m128(a, 22);
CASE_LSHIFT64_m128(a, 23);
CASE_LSHIFT64_m128(a, 24);
CASE_LSHIFT64_m128(a, 25);
CASE_LSHIFT64_m128(a, 26);
CASE_LSHIFT64_m128(a, 27);
CASE_LSHIFT64_m128(a, 28);
CASE_LSHIFT64_m128(a, 29);
CASE_LSHIFT64_m128(a, 30);
CASE_LSHIFT64_m128(a, 31);
CASE_LSHIFT64_m128(a, 32);
CASE_LSHIFT64_m128(a, 33);
CASE_LSHIFT64_m128(a, 34);
CASE_LSHIFT64_m128(a, 35);
CASE_LSHIFT64_m128(a, 36);
CASE_LSHIFT64_m128(a, 37);
CASE_LSHIFT64_m128(a, 38);
CASE_LSHIFT64_m128(a, 39);
CASE_LSHIFT64_m128(a, 40);
CASE_LSHIFT64_m128(a, 41);
CASE_LSHIFT64_m128(a, 42);
CASE_LSHIFT64_m128(a, 43);
CASE_LSHIFT64_m128(a, 44);
CASE_LSHIFT64_m128(a, 45);
CASE_LSHIFT64_m128(a, 46);
CASE_LSHIFT64_m128(a, 47);
CASE_LSHIFT64_m128(a, 48);
CASE_LSHIFT64_m128(a, 49);
CASE_LSHIFT64_m128(a, 50);
CASE_LSHIFT64_m128(a, 51);
CASE_LSHIFT64_m128(a, 52);
CASE_LSHIFT64_m128(a, 53);
CASE_LSHIFT64_m128(a, 54);
CASE_LSHIFT64_m128(a, 55);
CASE_LSHIFT64_m128(a, 56);
CASE_LSHIFT64_m128(a, 57);
CASE_LSHIFT64_m128(a, 58);
CASE_LSHIFT64_m128(a, 59);
CASE_LSHIFT64_m128(a, 60);
CASE_LSHIFT64_m128(a, 61);
CASE_LSHIFT64_m128(a, 62);
CASE_LSHIFT64_m128(a, 63);
default: return zeroes128(); break;
}
#undef CASE_LSHIFT64_m128
}
static really_really_inline
m128 rshift64_m128(m128 a, unsigned b) {
return (m128) vshrq_n_u64((uint64x2_t)a, b);
#if defined(HAVE__BUILTIN_CONSTANT_P)
if (__builtin_constant_p(b)) {
return (m128) vshrq_n_u64((uint64x2_t)a, b);
}
#endif
#define CASE_RSHIFT64_m128(a, offset) case offset: return (m128)vshrq_n_u64((int8x16_t)(a), (offset)); break;
switch (b) {
case 0: return a; break;
CASE_RSHIFT64_m128(a, 1);
CASE_RSHIFT64_m128(a, 2);
CASE_RSHIFT64_m128(a, 3);
CASE_RSHIFT64_m128(a, 4);
CASE_RSHIFT64_m128(a, 5);
CASE_RSHIFT64_m128(a, 6);
CASE_RSHIFT64_m128(a, 7);
CASE_RSHIFT64_m128(a, 8);
CASE_RSHIFT64_m128(a, 9);
CASE_RSHIFT64_m128(a, 10);
CASE_RSHIFT64_m128(a, 11);
CASE_RSHIFT64_m128(a, 12);
CASE_RSHIFT64_m128(a, 13);
CASE_RSHIFT64_m128(a, 14);
CASE_RSHIFT64_m128(a, 15);
CASE_RSHIFT64_m128(a, 16);
CASE_RSHIFT64_m128(a, 17);
CASE_RSHIFT64_m128(a, 18);
CASE_RSHIFT64_m128(a, 19);
CASE_RSHIFT64_m128(a, 20);
CASE_RSHIFT64_m128(a, 21);
CASE_RSHIFT64_m128(a, 22);
CASE_RSHIFT64_m128(a, 23);
CASE_RSHIFT64_m128(a, 24);
CASE_RSHIFT64_m128(a, 25);
CASE_RSHIFT64_m128(a, 26);
CASE_RSHIFT64_m128(a, 27);
CASE_RSHIFT64_m128(a, 28);
CASE_RSHIFT64_m128(a, 29);
CASE_RSHIFT64_m128(a, 30);
CASE_RSHIFT64_m128(a, 31);
CASE_RSHIFT64_m128(a, 32);
CASE_RSHIFT64_m128(a, 33);
CASE_RSHIFT64_m128(a, 34);
CASE_RSHIFT64_m128(a, 35);
CASE_RSHIFT64_m128(a, 36);
CASE_RSHIFT64_m128(a, 37);
CASE_RSHIFT64_m128(a, 38);
CASE_RSHIFT64_m128(a, 39);
CASE_RSHIFT64_m128(a, 40);
CASE_RSHIFT64_m128(a, 41);
CASE_RSHIFT64_m128(a, 42);
CASE_RSHIFT64_m128(a, 43);
CASE_RSHIFT64_m128(a, 44);
CASE_RSHIFT64_m128(a, 45);
CASE_RSHIFT64_m128(a, 46);
CASE_RSHIFT64_m128(a, 47);
CASE_RSHIFT64_m128(a, 48);
CASE_RSHIFT64_m128(a, 49);
CASE_RSHIFT64_m128(a, 50);
CASE_RSHIFT64_m128(a, 51);
CASE_RSHIFT64_m128(a, 52);
CASE_RSHIFT64_m128(a, 53);
CASE_RSHIFT64_m128(a, 54);
CASE_RSHIFT64_m128(a, 55);
CASE_RSHIFT64_m128(a, 56);
CASE_RSHIFT64_m128(a, 57);
CASE_RSHIFT64_m128(a, 58);
CASE_RSHIFT64_m128(a, 59);
CASE_RSHIFT64_m128(a, 60);
CASE_RSHIFT64_m128(a, 61);
CASE_RSHIFT64_m128(a, 62);
CASE_RSHIFT64_m128(a, 63);
default: return zeroes128(); break;
}
#undef CASE_RSHIFT64_m128
}
static really_inline m128 eq128(m128 a, m128 b) {