github_mirrors/vectorscan
3
0
Fork 1
mirror of https://github.com/VectorCamp/vectorscan.git synced 2025-06-28 16:41:01 +03:00
Code Issues Packages Projects Releases Wiki Activity

use rshift128() instead of vector-wide right shift

Browse Source
This commit is contained in:
Konstantinos Margaritis 2021-07-20 14:33:03 +03:00 committed by Konstantinos Margaritis
parent 6c51f7f591
commit b67cd7dfd0
1 changed files with 3 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
src/nfa/shufti_simd.hpp
View File

@ -221,6 +221,7 @@ template <uint16_t S>
static really_inline static really_inline
const u8 *fwdBlockDouble(SuperVector<S> mask1_lo, SuperVector<S> mask1_hi, SuperVector<S> mask2_lo, SuperVector<S> mask2_hi, const u8 *fwdBlockDouble(SuperVector<S> mask1_lo, SuperVector<S> mask1_hi, SuperVector<S> mask2_lo, SuperVector<S> mask2_hi,
SuperVector<S> chars, const SuperVector<S> low4bits, const u8 *buf) { SuperVector<S> chars, const SuperVector<S> low4bits, const u8 *buf) {
SuperVector<S> chars_lo = chars & low4bits; SuperVector<S> chars_lo = chars & low4bits;
SuperVector<S> chars_hi = chars.rshift64(4) & low4bits; SuperVector<S> chars_hi = chars.rshift64(4) & low4bits;
SuperVector<S> c1_lo = mask1_lo.pshufb(chars_lo); SuperVector<S> c1_lo = mask1_lo.pshufb(chars_lo);
@ -230,7 +231,7 @@ const u8 *fwdBlockDouble(SuperVector<S> mask1_lo, SuperVector<S> mask1_hi, Super
SuperVector<S> c2_lo = mask2_lo.pshufb(chars_lo); SuperVector<S> c2_lo = mask2_lo.pshufb(chars_lo);
SuperVector<S> c2_hi = mask2_hi.pshufb(chars_hi); SuperVector<S> c2_hi = mask2_hi.pshufb(chars_hi);
SuperVector<S> t2 = c2_lo | c2_hi; SuperVector<S> t2 = c2_lo | c2_hi;
SuperVector<S> t = t1 | (t2 >> 1); SuperVector<S> t = t1 | (t2.rshift128(1));
typename SuperVector<S>::movemask_type z = t.eqmask(SuperVector<S>::Ones()); typename SuperVector<S>::movemask_type z = t.eqmask(SuperVector<S>::Ones());
DEBUG_PRINTF(" z: 0x%016llx\n", (u64a)z); DEBUG_PRINTF(" z: 0x%016llx\n", (u64a)z);
@ -264,6 +265,7 @@ const u8 *shuftiDoubleExecReal(m128 mask1_lo, m128 mask1_hi,
if (d1 != d) { if (d1 != d) {
SuperVector<S> chars = SuperVector<S>::loadu(d); SuperVector<S> chars = SuperVector<S>::loadu(d);
rv = fwdBlockDouble(wide_mask1_lo, wide_mask1_hi, wide_mask2_lo, wide_mask2_hi, chars, low4bits, d); rv = fwdBlockDouble(wide_mask1_lo, wide_mask1_hi, wide_mask2_lo, wide_mask2_hi, chars, low4bits, d);
DEBUG_PRINTF("rv %p \n", rv);
if (rv) return rv; if (rv) return rv;
d = d1; d = d1;
} }