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

remove duplicate functions from previous merge

Browse Source
This commit is contained in:
Konstantinos Margaritis 2021-07-26 21:11:30 +03:00
parent e5050c9373
commit de30471edd
1 changed files with 1 additions and 93 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

94
src/nfa/shufti_simd.hpp
View File

@ -336,98 +336,6 @@ const u8 *shuftiExec(m128 mask_lo, m128 mask_hi, const u8 *buf,
return shuftiExecReal<VECTORSIZE>(mask_lo, mask_hi, buf, buf_end);
}
template <uint16_t S>
static really_inline const u8 *shuftiDoubleMini(SuperVector<S> mask1_lo, SuperVector<S> mask1_hi, SuperVector<S> mask2_lo, SuperVector<S> mask2_hi,
const u8 *buf, const u8 *buf_end){
uintptr_t len = buf_end - buf;
assert(len < S);
const SuperVector<S> low4bits = SuperVector<S>::dup_u8(0xf);
DEBUG_PRINTF("buf %p buf_end %p \n", buf, buf_end);
SuperVector<S> chars = SuperVector<S>::loadu_maskz(buf, len);
chars.print8("chars");
SuperVector<S> chars_lo = chars & low4bits;
chars_lo.print8("chars_lo");
SuperVector<S> chars_hi = chars.rshift64(4) & low4bits;
chars_hi.print8("chars_hi");
SuperVector<S> c1_lo = mask1_lo.pshufb_maskz(chars_lo, len);
c1_lo.print8("c1_lo");
SuperVector<S> c1_hi = mask1_hi.pshufb_maskz(chars_hi, len);
c1_hi.print8("c1_hi");
SuperVector<S> t1 = c1_lo | c1_hi;
t1.print8("t1");
SuperVector<S> c2_lo = mask2_lo.pshufb_maskz(chars_lo, len);
c2_lo.print8("c2_lo");
SuperVector<S> c2_hi = mask2_hi.pshufb_maskz(chars_hi, len);
c2_hi.print8("c2_hi");
SuperVector<S> t2 = c2_lo | c2_hi;
SuperVector<S> t = t1 | (t2.rshift128(1));
typename SuperVector<S>::movemask_type z = t.eqmask(SuperVector<S>::Ones());
DEBUG_PRINTF(" z: 0x%016llx\n", (u64a)z);
return firstMatch<S>(buf, z);
}
template <uint16_t S>
const u8 *shuftiDoubleExecReal(m128 mask1_lo, m128 mask1_hi, m128 mask2_lo, m128 mask2_hi,
const u8 *buf, const u8 *buf_end) {
assert(buf && buf_end);
assert(buf < buf_end);
DEBUG_PRINTF("shufti %p len %zu\n", buf, buf_end - buf);
DEBUG_PRINTF("b %s\n", buf);
const SuperVector<S> wide_mask1_lo(mask1_lo);
const SuperVector<S> wide_mask1_hi(mask1_hi);
const SuperVector<S> wide_mask2_lo(mask2_lo);
const SuperVector<S> wide_mask2_hi(mask2_hi);
const u8 *d = buf;
const u8 *rv;
DEBUG_PRINTF("start %p end %p \n", d, buf_end);
assert(d < buf_end);
if (d + S <= buf_end) {
// peel off first part to cacheline boundary
const u8 *d1 = ROUNDUP_PTR(d, S);
DEBUG_PRINTF("until aligned %p \n", d1);
if (d1 != d) {
SuperVector<S> chars = SuperVector<S>::loadu(d);
rv = fwdBlockDouble(wide_mask1_lo, wide_mask1_hi, wide_mask2_lo, wide_mask2_hi, chars, d);
DEBUG_PRINTF("rv %p \n", rv);
if (rv) return rv;
d = d1;
}
size_t loops = (buf_end - d) / S;
DEBUG_PRINTF("loops %ld \n", loops);
for (size_t i = 0; i < loops; i++, d+= S) {
DEBUG_PRINTF("it = %ld, d %p \n", i, d);
const u8 *base = ROUNDUP_PTR(d, S);
// On large packet buffers, this prefetch appears to get us about 2%.
__builtin_prefetch(base + 256);
SuperVector<S> chars = SuperVector<S>::load(d);
rv = fwdBlockDouble(wide_mask1_lo, wide_mask1_hi, wide_mask2_lo, wide_mask2_hi, chars, d);
if (rv) return rv;
}
}
DEBUG_PRINTF("tail d %p e %p \n", d, buf_end);
// finish off tail
if (d != buf_end) {
rv = shuftiDoubleMini(wide_mask1_lo, wide_mask1_hi, wide_mask2_lo, wide_mask2_hi, d, buf_end);
DEBUG_PRINTF("rv %p \n", rv);
if (rv >= buf && rv < buf_end) return rv;
}
return buf_end;
}
const u8 *rshuftiExec(m128 mask_lo, m128 mask_hi, const u8 *buf,
const u8 *buf_end) {
return rshuftiExecReal<VECTORSIZE>(mask_lo, mask_hi, buf, buf_end);
@ -437,4 +345,4 @@ const u8 *shuftiDoubleExec(m128 mask1_lo, m128 mask1_hi,
m128 mask2_lo, m128 mask2_hi,
const u8 *buf, const u8 *buf_end) {
return shuftiDoubleExecReal<VECTORSIZE>(mask1_lo, mask1_hi, mask2_lo, mask2_hi, buf, buf_end);
}
}