glibc/sysdeps/x86_64/multiarch/memchr-evex-base.S
Sunil K Pandey 451c6e5854 x86_64: Implement evex512 version of memchr, rawmemchr and wmemchr
This patch implements following evex512 version of string functions.
evex512 version takes up to 30% less cycle as compared to evex,
depending on length and alignment.

- memchr function using 512 bit vectors.
- rawmemchr function using 512 bit vectors.
- wmemchr function using 512 bit vectors.

Code size data:

memchr-evex.o		762 byte
memchr-evex512.o	576 byte (-24%)

rawmemchr-evex.o	461 byte
rawmemchr-evex512.o	412 byte (-11%)

wmemchr-evex.o		794 byte
wmemchr-evex512.o	552 byte (-30%)

Placeholder function, not used by any processor at the moment.

Reviewed-by: Noah Goldstein <goldstein.w.n@gmail.com>
2022-10-18 13:26:33 -07:00

305 lines
6.0 KiB
ArmAsm

/* Placeholder function, not used by any processor at the moment.
Copyright (C) 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/>. */
/* UNUSED. Exists purely as reference implementation. */
#include <isa-level.h>
#if ISA_SHOULD_BUILD (4)
# include <sysdep.h>
# ifdef USE_AS_WMEMCHR
# define CHAR_SIZE 4
# define VPBROADCAST vpbroadcastd
# define VPCMPEQ vpcmpeqd
# define VPCMPNE vpcmpneqd
# define VPMINU vpminud
# define VPTESTNM vptestnmd
# else
# define CHAR_SIZE 1
# define VPBROADCAST vpbroadcastb
# define VPCMPEQ vpcmpeqb
# define VPCMPNE vpcmpneqb
# define VPMINU vpminub
# define VPTESTNM vptestnmb
# endif
# define PAGE_SIZE 4096
# define CHAR_PER_VEC (VEC_SIZE / CHAR_SIZE)
.section SECTION(.text), "ax", @progbits
/* Aligning entry point to 64 byte, provides better performance for
one vector length string. */
ENTRY_P2ALIGN (MEMCHR, 6)
# ifndef USE_AS_RAWMEMCHR
/* Check for zero length. */
test %RDX_LP, %RDX_LP
jz L(zero)
# ifdef __ILP32__
/* Clear the upper 32 bits. */
movl %edx, %edx
# endif
# endif
/* Broadcast CHAR to VMM(1). */
VPBROADCAST %esi, %VMM(1)
movl %edi, %eax
andl $(PAGE_SIZE - 1), %eax
cmpl $(PAGE_SIZE - VEC_SIZE), %eax
ja L(page_cross)
/* Compare [w]char for null, mask bit will be set for match. */
VPCMPEQ (%rdi), %VMM(1), %k0
KMOV %k0, %VRCX
# ifndef USE_AS_RAWMEMCHR
mov %rdx, %rsi
/* Need to use bsfq here as upper 32 bit of rsi may zero out
for 'bsf %ecx, %esi', if %ecx is 0. */
bsfq %rcx, %rsi
cmp $CHAR_PER_VEC, %rsi
ja L(align_more)
# ifdef USE_AS_WMEMCHR
leaq (%rdi, %rsi, CHAR_SIZE), %rdi
# else
addq %rsi, %rdi
# endif
xor %eax, %eax
cmp %rsi, %rdx
cmova %rdi, %rax
# else
bsf %VRCX, %VRAX
jz L(align_more)
add %rdi, %rax
# endif
ret
.p2align 5,,5
L(page_cross):
movl %eax, %ecx
andl $(VEC_SIZE - 1), %ecx
# ifdef USE_AS_WMEMCHR
shrl $2, %ecx
# endif
xorq %rdi, %rax
VPCMPEQ (PAGE_SIZE - VEC_SIZE)(%rax), %VMM(1), %k0
KMOV %k0, %VRSI
shr %cl, %VRSI
# ifndef USE_AS_RAWMEMCHR
jnz L(page_cross_end)
movl $CHAR_PER_VEC, %eax
sub %ecx, %eax
cmp %rax, %rdx
ja L(align_more)
# else
jz L(align_more)
# endif
L(page_cross_end):
# ifndef USE_AS_RAWMEMCHR
bsf %VRSI, %VRCX
jz L(zero)
leaq (%rdi, %rcx, CHAR_SIZE), %rdi
xor %eax, %eax
cmp %rcx, %rdx
cmova %rdi, %rax
# else
bsf %VRSI, %VRAX
add %rdi, %rax
# endif
ret
# ifndef USE_AS_RAWMEMCHR
L(zero):
xorl %eax, %eax
ret
# endif
L(ret_vec_x2):
subq $-VEC_SIZE, %rdi
L(ret_vec_x1):
bsf %VRAX, %VRAX
# ifndef USE_AS_RAWMEMCHR
cmp %rax, %rdx
jbe L(zero)
# endif
# ifdef USE_AS_WMEMCHR
leaq (%rdi, %rax, CHAR_SIZE), %rax
# else
add %rdi, %rax
# endif
ret
.p2align 5,,5
L(align_more):
# ifndef USE_AS_RAWMEMCHR
mov %rdi, %rax
# endif
subq $-VEC_SIZE, %rdi
/* Align rdi to VEC_SIZE. */
andq $-VEC_SIZE, %rdi
# ifndef USE_AS_RAWMEMCHR
subq %rdi, %rax
# ifdef USE_AS_WMEMCHR
sar $2, %rax
# endif
addq %rax, %rdx
# endif
/* Loop unroll 4 times for 4 vector loop. */
VPCMPEQ (%rdi), %VMM(1), %k0
KMOV %k0, %VRAX
test %VRAX, %VRAX
jnz L(ret_vec_x1)
# ifndef USE_AS_RAWMEMCHR
subq $CHAR_PER_VEC, %rdx
jbe L(zero)
# endif
VPCMPEQ VEC_SIZE(%rdi), %VMM(1), %k0
KMOV %k0, %VRAX
test %VRAX, %VRAX
jnz L(ret_vec_x2)
# ifndef USE_AS_RAWMEMCHR
subq $CHAR_PER_VEC, %rdx
jbe L(zero)
# endif
VPCMPEQ (VEC_SIZE * 2)(%rdi), %VMM(1), %k0
KMOV %k0, %VRAX
test %VRAX, %VRAX
jnz L(ret_vec_x3)
# ifndef USE_AS_RAWMEMCHR
subq $CHAR_PER_VEC, %rdx
jbe L(zero)
# endif
VPCMPEQ (VEC_SIZE * 3)(%rdi), %VMM(1), %k0
KMOV %k0, %VRAX
test %VRAX, %VRAX
jnz L(ret_vec_x4)
# ifndef USE_AS_RAWMEMCHR
subq $CHAR_PER_VEC, %rdx
jbe L(zero)
/* Save pointer to find alignment adjustment. */
movq %rdi, %rax
# endif
/* Align address to VEC_SIZE * 4 for loop. */
andq $-(VEC_SIZE * 4), %rdi
/* Add alignment difference to rdx. */
# ifndef USE_AS_RAWMEMCHR
subq %rdi, %rax
# ifdef USE_AS_WMEMCHR
shr $2, %VRAX
# endif
addq %rax, %rdx
# endif
/* 4 vector loop. */
.p2align 5,,11
L(loop):
VPCMPNE (VEC_SIZE * 4)(%rdi), %VMM(1), %k1
vpxorq (VEC_SIZE * 5)(%rdi), %VMM(1), %VMM(2)
vpxorq (VEC_SIZE * 6)(%rdi), %VMM(1), %VMM(3)
VPCMPEQ (VEC_SIZE * 7)(%rdi), %VMM(1), %k3
VPMINU %VMM(2), %VMM(3), %VMM(3){%k1}{z}
VPTESTNM %VMM(3), %VMM(3), %k2
subq $-(VEC_SIZE * 4), %rdi
KORTEST %k2, %k3
# ifdef USE_AS_RAWMEMCHR
jz L(loop)
# else
jnz L(loopend)
subq $(CHAR_PER_VEC * 4), %rdx
ja L(loop)
L(zero_2):
xor %eax, %eax
ret
# endif
L(loopend):
VPCMPEQ (%rdi), %VMM(1), %k1
KMOV %k1, %VRAX
test %VRAX, %VRAX
jnz L(ret_vec_x1)
# ifndef USE_AS_RAWMEMCHR
subq $CHAR_PER_VEC, %rdx
jbe L(zero_2)
# endif
VPCMPEQ VEC_SIZE(%rdi), %VMM(1), %k1
KMOV %k1, %VRAX
test %VRAX, %VRAX
jnz L(ret_vec_x2)
# ifndef USE_AS_RAWMEMCHR
subq $CHAR_PER_VEC, %rdx
jbe L(zero_2)
# endif
VPCMPEQ (VEC_SIZE * 2)(%rdi), %VMM(1), %k1
KMOV %k1, %VRAX
test %VRAX, %VRAX
jnz L(ret_vec_x3)
# ifndef USE_AS_RAWMEMCHR
subq $CHAR_PER_VEC, %rdx
jbe L(zero_2)
# endif
/* At this point null [w]char must be in the fourth vector so no
need to check. */
KMOV %k3, %VRAX
L(ret_vec_x4):
bsf %VRAX, %VRAX
# ifndef USE_AS_RAWMEMCHR
cmp %rax, %rdx
jbe L(zero)
# endif
leaq (VEC_SIZE * 3)(%rdi, %rax, CHAR_SIZE), %rax
ret
.p2align 5,,5
L(ret_vec_x3):
bsf %VRAX, %VRAX
# ifndef USE_AS_RAWMEMCHR
cmp %rax, %rdx
jbe L(zero)
# endif
leaq (VEC_SIZE * 2)(%rdi, %rax, CHAR_SIZE), %rax
ret
END (MEMCHR)
#endif