ubuntu-buildroot/output/build/glibc-2.36-81-g4f4d7a13edfd.../sysdeps/x86_64/multiarch/memcmp-avx2-movbe.S

571 lines
13 KiB
ArmAsm

/* memcmp/wmemcmp optimized with AVX2.
Copyright (C) 2017-2022 Free Software Foundation, Inc.
This file is part of the GNU C Library.
The GNU C Library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public
License as published by the Free Software Foundation; either
version 2.1 of the License, or (at your option) any later version.
The GNU C Library is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public
License along with the GNU C Library; if not, see
<https://www.gnu.org/licenses/>. */
#include <isa-level.h>
#if ISA_SHOULD_BUILD (3)
/* memcmp/wmemcmp is implemented as:
1. Use ymm vector compares when possible. The only case where
vector compares is not possible for when size < VEC_SIZE
and loading from either s1 or s2 would cause a page cross.
2. For size from 2 to 7 bytes on page cross, load as big endian
with movbe and bswap to avoid branches.
3. Use xmm vector compare when size >= 4 bytes for memcmp or
size >= 8 bytes for wmemcmp.
4. Optimistically compare up to first 4 * VEC_SIZE one at a
to check for early mismatches. Only do this if its guranteed the
work is not wasted.
5. If size is 8 * VEC_SIZE or less, unroll the loop.
6. Compare 4 * VEC_SIZE at a time with the aligned first memory
area.
7. Use 2 vector compares when size is 2 * VEC_SIZE or less.
8. Use 4 vector compares when size is 4 * VEC_SIZE or less.
9. Use 8 vector compares when size is 8 * VEC_SIZE or less. */
# include <sysdep.h>
# ifndef MEMCMP
# define MEMCMP __memcmp_avx2_movbe
# endif
# ifdef USE_AS_WMEMCMP
# define CHAR_SIZE 4
# define VPCMPEQ vpcmpeqd
# else
# define CHAR_SIZE 1
# define VPCMPEQ vpcmpeqb
# endif
# ifndef VZEROUPPER
# define VZEROUPPER vzeroupper
# endif
# ifndef SECTION
# define SECTION(p) p##.avx
# endif
# define VEC_SIZE 32
# define PAGE_SIZE 4096
/* Warning!
wmemcmp has to use SIGNED comparison for elements.
memcmp has to use UNSIGNED comparison for elemnts.
*/
.section SECTION(.text),"ax",@progbits
ENTRY (MEMCMP)
# ifdef USE_AS_WMEMCMP
shl $2, %RDX_LP
# elif defined __ILP32__
/* Clear the upper 32 bits. */
movl %edx, %edx
# endif
cmp $VEC_SIZE, %RDX_LP
jb L(less_vec)
/* From VEC to 2 * VEC. No branch when size == VEC_SIZE. */
vmovdqu (%rsi), %ymm1
VPCMPEQ (%rdi), %ymm1, %ymm1
vpmovmskb %ymm1, %eax
/* NB: eax must be destination register if going to
L(return_vec_[0,2]). For L(return_vec_3 destination register
must be ecx. */
incl %eax
jnz L(return_vec_0)
cmpq $(VEC_SIZE * 2), %rdx
jbe L(last_1x_vec)
/* Check second VEC no matter what. */
vmovdqu VEC_SIZE(%rsi), %ymm2
VPCMPEQ VEC_SIZE(%rdi), %ymm2, %ymm2
vpmovmskb %ymm2, %eax
/* If all 4 VEC where equal eax will be all 1s so incl will
overflow and set zero flag. */
incl %eax
jnz L(return_vec_1)
/* Less than 4 * VEC. */
cmpq $(VEC_SIZE * 4), %rdx
jbe L(last_2x_vec)
/* Check third and fourth VEC no matter what. */
vmovdqu (VEC_SIZE * 2)(%rsi), %ymm3
VPCMPEQ (VEC_SIZE * 2)(%rdi), %ymm3, %ymm3
vpmovmskb %ymm3, %eax
incl %eax
jnz L(return_vec_2)
vmovdqu (VEC_SIZE * 3)(%rsi), %ymm4
VPCMPEQ (VEC_SIZE * 3)(%rdi), %ymm4, %ymm4
vpmovmskb %ymm4, %ecx
incl %ecx
jnz L(return_vec_3)
/* Go to 4x VEC loop. */
cmpq $(VEC_SIZE * 8), %rdx
ja L(more_8x_vec)
/* Handle remainder of size = 4 * VEC + 1 to 8 * VEC without any
branches. */
/* Load first two VEC from s2 before adjusting addresses. */
vmovdqu -(VEC_SIZE * 4)(%rsi, %rdx), %ymm1
vmovdqu -(VEC_SIZE * 3)(%rsi, %rdx), %ymm2
leaq -(4 * VEC_SIZE)(%rdi, %rdx), %rdi
leaq -(4 * VEC_SIZE)(%rsi, %rdx), %rsi
/* Wait to load from s1 until addressed adjust due to
unlamination of microfusion with complex address mode. */
VPCMPEQ (%rdi), %ymm1, %ymm1
VPCMPEQ (VEC_SIZE)(%rdi), %ymm2, %ymm2
vmovdqu (VEC_SIZE * 2)(%rsi), %ymm3
VPCMPEQ (VEC_SIZE * 2)(%rdi), %ymm3, %ymm3
vmovdqu (VEC_SIZE * 3)(%rsi), %ymm4
VPCMPEQ (VEC_SIZE * 3)(%rdi), %ymm4, %ymm4
/* Reduce VEC0 - VEC4. */
vpand %ymm1, %ymm2, %ymm5
vpand %ymm3, %ymm4, %ymm6
vpand %ymm5, %ymm6, %ymm7
vpmovmskb %ymm7, %ecx
incl %ecx
jnz L(return_vec_0_1_2_3)
/* NB: eax must be zero to reach here. */
VZEROUPPER_RETURN
.p2align 4
L(return_vec_0):
tzcntl %eax, %eax
# ifdef USE_AS_WMEMCMP
movl (%rdi, %rax), %ecx
xorl %edx, %edx
cmpl (%rsi, %rax), %ecx
/* NB: no partial register stall here because xorl zero idiom
above. */
setg %dl
leal -1(%rdx, %rdx), %eax
# else
movzbl (%rsi, %rax), %ecx
movzbl (%rdi, %rax), %eax
subl %ecx, %eax
# endif
L(return_vzeroupper):
ZERO_UPPER_VEC_REGISTERS_RETURN
.p2align 4
L(return_vec_1):
tzcntl %eax, %eax
# ifdef USE_AS_WMEMCMP
movl VEC_SIZE(%rdi, %rax), %ecx
xorl %edx, %edx
cmpl VEC_SIZE(%rsi, %rax), %ecx
setg %dl
leal -1(%rdx, %rdx), %eax
# else
movzbl VEC_SIZE(%rsi, %rax), %ecx
movzbl VEC_SIZE(%rdi, %rax), %eax
subl %ecx, %eax
# endif
VZEROUPPER_RETURN
.p2align 4
L(return_vec_2):
tzcntl %eax, %eax
# ifdef USE_AS_WMEMCMP
movl (VEC_SIZE * 2)(%rdi, %rax), %ecx
xorl %edx, %edx
cmpl (VEC_SIZE * 2)(%rsi, %rax), %ecx
setg %dl
leal -1(%rdx, %rdx), %eax
# else
movzbl (VEC_SIZE * 2)(%rsi, %rax), %ecx
movzbl (VEC_SIZE * 2)(%rdi, %rax), %eax
subl %ecx, %eax
# endif
VZEROUPPER_RETURN
/* NB: p2align 5 here to ensure 4x loop is 32 byte aligned. */
.p2align 5
L(8x_return_vec_0_1_2_3):
/* Returning from L(more_8x_vec) requires restoring rsi. */
addq %rdi, %rsi
L(return_vec_0_1_2_3):
vpmovmskb %ymm1, %eax
incl %eax
jnz L(return_vec_0)
vpmovmskb %ymm2, %eax
incl %eax
jnz L(return_vec_1)
vpmovmskb %ymm3, %eax
incl %eax
jnz L(return_vec_2)
L(return_vec_3):
tzcntl %ecx, %ecx
# ifdef USE_AS_WMEMCMP
movl (VEC_SIZE * 3)(%rdi, %rcx), %eax
xorl %edx, %edx
cmpl (VEC_SIZE * 3)(%rsi, %rcx), %eax
setg %dl
leal -1(%rdx, %rdx), %eax
# else
movzbl (VEC_SIZE * 3)(%rdi, %rcx), %eax
movzbl (VEC_SIZE * 3)(%rsi, %rcx), %ecx
subl %ecx, %eax
# endif
VZEROUPPER_RETURN
.p2align 4
L(more_8x_vec):
/* Set end of s1 in rdx. */
leaq -(VEC_SIZE * 4)(%rdi, %rdx), %rdx
/* rsi stores s2 - s1. This allows loop to only update one
pointer. */
subq %rdi, %rsi
/* Align s1 pointer. */
andq $-VEC_SIZE, %rdi
/* Adjust because first 4x vec where check already. */
subq $-(VEC_SIZE * 4), %rdi
.p2align 4
L(loop_4x_vec):
/* rsi has s2 - s1 so get correct address by adding s1 (in rdi).
*/
vmovdqu (%rsi, %rdi), %ymm1
VPCMPEQ (%rdi), %ymm1, %ymm1
vmovdqu VEC_SIZE(%rsi, %rdi), %ymm2
VPCMPEQ VEC_SIZE(%rdi), %ymm2, %ymm2
vmovdqu (VEC_SIZE * 2)(%rsi, %rdi), %ymm3
VPCMPEQ (VEC_SIZE * 2)(%rdi), %ymm3, %ymm3
vmovdqu (VEC_SIZE * 3)(%rsi, %rdi), %ymm4
VPCMPEQ (VEC_SIZE * 3)(%rdi), %ymm4, %ymm4
vpand %ymm1, %ymm2, %ymm5
vpand %ymm3, %ymm4, %ymm6
vpand %ymm5, %ymm6, %ymm7
vpmovmskb %ymm7, %ecx
incl %ecx
jnz L(8x_return_vec_0_1_2_3)
subq $-(VEC_SIZE * 4), %rdi
/* Check if s1 pointer at end. */
cmpq %rdx, %rdi
jb L(loop_4x_vec)
subq %rdx, %rdi
/* rdi has 4 * VEC_SIZE - remaining length. */
cmpl $(VEC_SIZE * 3), %edi
jae L(8x_last_1x_vec)
/* Load regardless of branch. */
vmovdqu (VEC_SIZE * 2)(%rsi, %rdx), %ymm3
cmpl $(VEC_SIZE * 2), %edi
jae L(8x_last_2x_vec)
/* Check last 4 VEC. */
vmovdqu (%rsi, %rdx), %ymm1
VPCMPEQ (%rdx), %ymm1, %ymm1
vmovdqu VEC_SIZE(%rsi, %rdx), %ymm2
VPCMPEQ VEC_SIZE(%rdx), %ymm2, %ymm2
VPCMPEQ (VEC_SIZE * 2)(%rdx), %ymm3, %ymm3
vmovdqu (VEC_SIZE * 3)(%rsi, %rdx), %ymm4
VPCMPEQ (VEC_SIZE * 3)(%rdx), %ymm4, %ymm4
vpand %ymm1, %ymm2, %ymm5
vpand %ymm3, %ymm4, %ymm6
vpand %ymm5, %ymm6, %ymm7
vpmovmskb %ymm7, %ecx
/* Restore s1 pointer to rdi. */
movq %rdx, %rdi
incl %ecx
jnz L(8x_return_vec_0_1_2_3)
/* NB: eax must be zero to reach here. */
VZEROUPPER_RETURN
/* Only entry is from L(more_8x_vec). */
.p2align 4
L(8x_last_2x_vec):
/* Check second to last VEC. rdx store end pointer of s1 and
ymm3 has already been loaded with second to last VEC from s2.
*/
VPCMPEQ (VEC_SIZE * 2)(%rdx), %ymm3, %ymm3
vpmovmskb %ymm3, %eax
incl %eax
jnz L(8x_return_vec_2)
/* Check last VEC. */
.p2align 4
L(8x_last_1x_vec):
vmovdqu (VEC_SIZE * 3)(%rsi, %rdx), %ymm4
VPCMPEQ (VEC_SIZE * 3)(%rdx), %ymm4, %ymm4
vpmovmskb %ymm4, %eax
incl %eax
jnz L(8x_return_vec_3)
VZEROUPPER_RETURN
.p2align 4
L(last_2x_vec):
/* Check second to last VEC. */
vmovdqu -(VEC_SIZE * 2)(%rsi, %rdx), %ymm1
VPCMPEQ -(VEC_SIZE * 2)(%rdi, %rdx), %ymm1, %ymm1
vpmovmskb %ymm1, %eax
incl %eax
jnz L(return_vec_1_end)
/* Check last VEC. */
L(last_1x_vec):
vmovdqu -(VEC_SIZE * 1)(%rsi, %rdx), %ymm1
VPCMPEQ -(VEC_SIZE * 1)(%rdi, %rdx), %ymm1, %ymm1
vpmovmskb %ymm1, %eax
incl %eax
jnz L(return_vec_0_end)
VZEROUPPER_RETURN
.p2align 4
L(8x_return_vec_2):
subq $VEC_SIZE, %rdx
L(8x_return_vec_3):
tzcntl %eax, %eax
addq %rdx, %rax
# ifdef USE_AS_WMEMCMP
movl (VEC_SIZE * 3)(%rax), %ecx
xorl %edx, %edx
cmpl (VEC_SIZE * 3)(%rsi, %rax), %ecx
setg %dl
leal -1(%rdx, %rdx), %eax
# else
movzbl (VEC_SIZE * 3)(%rsi, %rax), %ecx
movzbl (VEC_SIZE * 3)(%rax), %eax
subl %ecx, %eax
# endif
VZEROUPPER_RETURN
.p2align 4
L(return_vec_1_end):
tzcntl %eax, %eax
addl %edx, %eax
# ifdef USE_AS_WMEMCMP
movl -(VEC_SIZE * 2)(%rdi, %rax), %ecx
xorl %edx, %edx
cmpl -(VEC_SIZE * 2)(%rsi, %rax), %ecx
setg %dl
leal -1(%rdx, %rdx), %eax
# else
movzbl -(VEC_SIZE * 2)(%rsi, %rax), %ecx
movzbl -(VEC_SIZE * 2)(%rdi, %rax), %eax
subl %ecx, %eax
# endif
VZEROUPPER_RETURN
.p2align 4
L(return_vec_0_end):
tzcntl %eax, %eax
addl %edx, %eax
# ifdef USE_AS_WMEMCMP
movl -VEC_SIZE(%rdi, %rax), %ecx
xorl %edx, %edx
cmpl -VEC_SIZE(%rsi, %rax), %ecx
setg %dl
leal -1(%rdx, %rdx), %eax
# else
movzbl -VEC_SIZE(%rsi, %rax), %ecx
movzbl -VEC_SIZE(%rdi, %rax), %eax
subl %ecx, %eax
# endif
VZEROUPPER_RETURN
.p2align 4
L(less_vec):
/* Check if one or less CHAR. This is necessary for size = 0 but
is also faster for size = CHAR_SIZE. */
cmpl $CHAR_SIZE, %edx
jbe L(one_or_less)
/* Check if loading one VEC from either s1 or s2 could cause a
page cross. This can have false positives but is by far the
fastest method. */
movl %edi, %eax
orl %esi, %eax
andl $(PAGE_SIZE - 1), %eax
cmpl $(PAGE_SIZE - VEC_SIZE), %eax
jg L(page_cross_less_vec)
/* No page cross possible. */
vmovdqu (%rsi), %ymm2
VPCMPEQ (%rdi), %ymm2, %ymm2
vpmovmskb %ymm2, %eax
incl %eax
/* Result will be zero if s1 and s2 match. Otherwise first set
bit will be first mismatch. */
bzhil %edx, %eax, %edx
jnz L(return_vec_0)
xorl %eax, %eax
VZEROUPPER_RETURN
.p2align 4
L(page_cross_less_vec):
/* if USE_AS_WMEMCMP it can only be 0, 4, 8, 12, 16, 20, 24, 28
bytes. */
cmpl $16, %edx
jae L(between_16_31)
# ifndef USE_AS_WMEMCMP
cmpl $8, %edx
jae L(between_8_15)
/* Fall through for [4, 7]. */
cmpl $4, %edx
jb L(between_2_3)
movbe (%rdi), %eax
movbe (%rsi), %ecx
shlq $32, %rax
shlq $32, %rcx
movbe -4(%rdi, %rdx), %edi
movbe -4(%rsi, %rdx), %esi
orq %rdi, %rax
orq %rsi, %rcx
subq %rcx, %rax
/* Fast path for return zero. */
jnz L(ret_nonzero)
/* No ymm register was touched. */
ret
.p2align 4
L(one_or_less):
jb L(zero)
movzbl (%rsi), %ecx
movzbl (%rdi), %eax
subl %ecx, %eax
/* No ymm register was touched. */
ret
.p2align 4,, 5
L(ret_nonzero):
sbbl %eax, %eax
orl $1, %eax
/* No ymm register was touched. */
ret
.p2align 4,, 2
L(zero):
xorl %eax, %eax
/* No ymm register was touched. */
ret
.p2align 4
L(between_8_15):
movbe (%rdi), %rax
movbe (%rsi), %rcx
subq %rcx, %rax
jnz L(ret_nonzero)
movbe -8(%rdi, %rdx), %rax
movbe -8(%rsi, %rdx), %rcx
subq %rcx, %rax
/* Fast path for return zero. */
jnz L(ret_nonzero)
/* No ymm register was touched. */
ret
# else
/* If USE_AS_WMEMCMP fall through into 8-15 byte case. */
vmovq (%rdi), %xmm1
vmovq (%rsi), %xmm2
VPCMPEQ %xmm1, %xmm2, %xmm2
vpmovmskb %xmm2, %eax
subl $0xffff, %eax
jnz L(return_vec_0)
/* Use overlapping loads to avoid branches. */
leaq -8(%rdi, %rdx), %rdi
leaq -8(%rsi, %rdx), %rsi
vmovq (%rdi), %xmm1
vmovq (%rsi), %xmm2
VPCMPEQ %xmm1, %xmm2, %xmm2
vpmovmskb %xmm2, %eax
subl $0xffff, %eax
/* Fast path for return zero. */
jnz L(return_vec_0)
/* No ymm register was touched. */
ret
# endif
.p2align 4,, 10
L(between_16_31):
/* From 16 to 31 bytes. No branch when size == 16. */
vmovdqu (%rsi), %xmm2
VPCMPEQ (%rdi), %xmm2, %xmm2
vpmovmskb %xmm2, %eax
subl $0xffff, %eax
jnz L(return_vec_0)
/* Use overlapping loads to avoid branches. */
vmovdqu -16(%rsi, %rdx), %xmm2
leaq -16(%rdi, %rdx), %rdi
leaq -16(%rsi, %rdx), %rsi
VPCMPEQ (%rdi), %xmm2, %xmm2
vpmovmskb %xmm2, %eax
subl $0xffff, %eax
/* Fast path for return zero. */
jnz L(return_vec_0)
/* No ymm register was touched. */
ret
# ifdef USE_AS_WMEMCMP
.p2align 4,, 2
L(zero):
xorl %eax, %eax
ret
.p2align 4
L(one_or_less):
jb L(zero)
movl (%rdi), %ecx
xorl %edx, %edx
cmpl (%rsi), %ecx
je L(zero)
setg %dl
leal -1(%rdx, %rdx), %eax
/* No ymm register was touched. */
ret
# else
.p2align 4
L(between_2_3):
/* Load as big endian to avoid branches. */
movzwl (%rdi), %eax
movzwl (%rsi), %ecx
bswap %eax
bswap %ecx
shrl %eax
shrl %ecx
movzbl -1(%rdi, %rdx), %edi
movzbl -1(%rsi, %rdx), %esi
orl %edi, %eax
orl %esi, %ecx
/* Subtraction is okay because the upper bit is zero. */
subl %ecx, %eax
/* No ymm register was touched. */
ret
# endif
END (MEMCMP)
#endif