vectorscan/src/nfa/shufti_simd.hpp
2021-10-12 11:51:34 +03:00

299 lines
11 KiB
C++

/*
* Copyright (c) 2015-2017, Intel Corporation
* Copyright (c) 2020-2021, VectorCamp PC
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* * Redistributions of source code must retain the above copyright notice,
* this list of conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* * Neither the name of Intel Corporation nor the names of its contributors
* may be used to endorse or promote products derived from this software
* without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
* LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
* CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
* SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
* INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
* CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
* ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
* POSSIBILITY OF SUCH DAMAGE.
*/
/** \file
* \brief Shufti: character class acceleration.
*
* Utilises the SSSE3 pshufb shuffle instruction
*/
#include "shufti.h"
#include "ue2common.h"
#include "util/arch.h"
#include "util/bitutils.h"
#include "util/unaligned.h"
#include "util/supervector/supervector.hpp"
#include "util/match.hpp"
template <uint16_t S>
static really_inline
typename SuperVector<S>::movemask_type block(SuperVector<S> mask_lo, SuperVector<S> mask_hi,
SuperVector<S> chars, const SuperVector<S> low4bits) {
SuperVector<S> c_lo = chars & low4bits;
//printv_u8("c_lo", c_lo);
c_lo = mask_lo.pshufb(c_lo);
//printv_u8("c_lo", c_lo);
SuperVector<S> c_hi = mask_hi.pshufb(chars.rshift64(4) & low4bits);
SuperVector<S> t = c_lo & c_hi;
/*printv_u8("low4bits", low4bits);
printv_u8("mask_lo", mask_lo);
printv_u8("mask_hi", mask_hi);
printv_u8("chars", chars);
printv_u8("c_lo", c_lo);
printv_u8("c_hi", c_hi);
printv_u8("t", t);*/
return t.eqmask(SuperVector<S>::Zeroes());
}
template <uint16_t S>
static really_inline
const u8 *fwdBlock(SuperVector<S> mask_lo, SuperVector<S> mask_hi, SuperVector<S> chars,
const SuperVector<S> low4bits, const u8 *buf) {
typename SuperVector<S>::movemask_type z = block(mask_lo, mask_hi, chars, low4bits);
DEBUG_PRINTF("z %08x\n", z);
return firstMatch<S>(buf, z);
}
template <uint16_t S>
static really_inline
const u8 *shortShufti(SuperVector<S> mask_lo, SuperVector<S> mask_hi, const u8 *buf,
const u8 *buf_end, const SuperVector<S> low4bits) {
DEBUG_PRINTF("short shufti %p len %zu\n", buf, buf_end - buf);
uintptr_t len = buf_end - buf;
assert(len <= S);
SuperVector<S> chars = SuperVector<S>::loadu_maskz(buf, static_cast<uint8_t>(len));
//printv_u8("chars", chars);
uint8_t alignment = (uintptr_t)(buf) & 15;
typename SuperVector<S>::movemask_type maskb = 1 << alignment;
typename SuperVector<S>::movemask_type maske = SINGLE_LOAD_MASK(len - alignment);
typename SuperVector<S>::movemask_type z = block(mask_lo, mask_hi, chars, low4bits);
// reuse the load mask to indicate valid bytes
DEBUG_PRINTF("z %08x\n", z);
z &= maskb | maske;
DEBUG_PRINTF("z %08x\n", z);
return firstMatch<S>(buf, z);
}
template <uint16_t S>
static really_inline
const u8 *revBlock(SuperVector<S> mask_lo, SuperVector<S> mask_hi, SuperVector<S> chars,
const SuperVector<S> low4bits, const u8 *buf) {
typename SuperVector<S>::movemask_type z = block(mask_lo, mask_hi, chars, low4bits);
DEBUG_PRINTF("z %08x\n", z);
return lastMatch<S>(buf, z);
}
template <uint16_t S>
const u8 *shuftiExecReal(m128 mask_lo, m128 mask_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> low4bits = SuperVector<S>::dup_u8(0xf);
const SuperVector<S> wide_mask_lo(mask_lo);
const SuperVector<S> wide_mask_hi(mask_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) {
rv = shuftiFwdSlow((const u8 *)&mask_lo, (const u8 *)&mask_hi, d, d1);
// rv = shortShufti(wide_mask_lo, wide_mask_hi, d, d1, low4bits);
if (rv != d1) {
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("d %p \n", 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 = fwdBlock(wide_mask_lo, wide_mask_hi, chars, low4bits, d);
if (rv) return rv;
}
}
DEBUG_PRINTF("d %p e %p \n", d, buf_end);
// finish off tail
rv = buf_end;
if (d != buf_end) {
rv = shuftiFwdSlow((const u8 *)&mask_lo, (const u8 *)&mask_hi, d, buf_end);
// rv = shortShufti(wide_mask_lo, wide_mask_hi, buf_end - S, buf_end, low4bits);
DEBUG_PRINTF("rv %p \n", rv);
}
return rv;
}
template <uint16_t S>
const u8 *rshuftiExecReal(m128 mask_lo, m128 mask_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> low4bits = SuperVector<S>::dup_u8(0xf);
const SuperVector<S> wide_mask_lo(mask_lo);
const SuperVector<S> wide_mask_hi(mask_hi);
const u8 *d = buf_end;
const u8 *rv;
DEBUG_PRINTF("start %p end %p \n", buf, d);
assert(d > buf);
if (d - S >= buf) {
// peel off first part to cacheline boundary
const u8 *d1 = ROUNDDOWN_PTR(d, S);
DEBUG_PRINTF("until aligned %p \n", d1);
if (d1 != d) {
rv = shuftiRevSlow((const u8 *)&mask_lo, (const u8 *)&mask_hi, d1, d);
DEBUG_PRINTF("rv %p \n", rv);
// rv = shortShufti(wide_mask_lo, wide_mask_hi, d, d1, low4bits);
if (rv != d1 - 1) return rv;
d = d1;
}
while (d - S >= buf) {
d -= S;
DEBUG_PRINTF("d %p \n", d);
const u8 *base = ROUNDDOWN_PTR(buf, 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 = revBlock(wide_mask_lo, wide_mask_hi, chars, low4bits, d);
if (rv) return rv;
}
}
DEBUG_PRINTF("d %p e %p \n", buf, d);
// finish off tail
if (d != buf) {
rv = shuftiRevSlow((const u8 *)&mask_lo, (const u8 *)&mask_hi, buf, d);
// rv = shortShufti(wide_mask_lo, wide_mask_hi, buf_end - S, buf_end, low4bits);
DEBUG_PRINTF("rv %p \n", rv);
if (rv != d - 1) return rv;
}
return buf - 1;
}
template <uint16_t S>
static really_inline
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_lo = chars & low4bits;
SuperVector<S> chars_hi = chars.rshift64(4) & low4bits;
SuperVector<S> c1_lo = mask1_lo.pshufb(chars_lo);
SuperVector<S> c1_hi = mask1_hi.pshufb(chars_hi);
SuperVector<S> t1 = c1_lo | c1_hi;
SuperVector<S> c2_lo = mask2_lo.pshufb(chars_lo);
SuperVector<S> c2_hi = mask2_hi.pshufb(chars_hi);
SuperVector<S> t2 = c2_lo | c2_hi;
SuperVector<S> t = t1 | (t2 >> 1);
typename SuperVector<S>::movemask_type z = t.eqmask(SuperVector<S>::Ones());
DEBUG_PRINTF(" z: 0x%08x\n", 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> low4bits = SuperVector<S>::dup_u8(0xf);
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, low4bits, d);
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("d %p \n", 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, low4bits, d);
if (rv) return rv;
}
}
DEBUG_PRINTF("d %p e %p \n", d, buf_end);
// finish off tail
if (d != buf_end) {
SuperVector<S> chars = SuperVector<S>::loadu(buf_end - S);
rv = fwdBlockDouble(wide_mask1_lo, wide_mask1_hi, wide_mask2_lo, wide_mask2_hi, chars, low4bits, buf_end - S);
DEBUG_PRINTF("rv %p \n", rv);
if (rv) return rv;
}
return buf_end;
}