Blame sysdeps/x86_64/memrchr.S
|
Packit |
6c4009 |
/* fast SSE2 memrchr with 64 byte loop and pmaxub instruction using
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
Copyright (C) 2011-2018 Free Software Foundation, Inc.
|
|
Packit |
6c4009 |
Contributed by Intel Corporation.
|
|
Packit |
6c4009 |
This file is part of the GNU C Library.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
The GNU C Library is free software; you can redistribute it and/or
|
|
Packit |
6c4009 |
modify it under the terms of the GNU Lesser General Public
|
|
Packit |
6c4009 |
License as published by the Free Software Foundation; either
|
|
Packit |
6c4009 |
version 2.1 of the License, or (at your option) any later version.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
The GNU C Library is distributed in the hope that it will be useful,
|
|
Packit |
6c4009 |
but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
Packit |
6c4009 |
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
Packit |
6c4009 |
Lesser General Public License for more details.
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
You should have received a copy of the GNU Lesser General Public
|
|
Packit |
6c4009 |
License along with the GNU C Library; if not, see
|
|
Packit |
6c4009 |
<http://www.gnu.org/licenses/>. */
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
#include <sysdep.h>
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.text
|
|
Packit |
6c4009 |
ENTRY (__memrchr)
|
|
Packit |
6c4009 |
movd %esi, %xmm1
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
sub $16, %rdx
|
|
Packit |
6c4009 |
jbe L(length_less16)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
punpcklbw %xmm1, %xmm1
|
|
Packit |
6c4009 |
punpcklbw %xmm1, %xmm1
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
add %rdx, %rdi
|
|
Packit |
6c4009 |
pshufd $0, %xmm1, %xmm1
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movdqu (%rdi), %xmm0
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm0
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
/* Check if there is a match. */
|
|
Packit |
6c4009 |
pmovmskb %xmm0, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jnz L(matches0)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
sub $64, %rdi
|
|
Packit |
6c4009 |
mov %edi, %ecx
|
|
Packit |
6c4009 |
and $15, %ecx
|
|
Packit |
6c4009 |
jz L(loop_prolog)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
add $16, %rdi
|
|
Packit |
6c4009 |
add $16, %rdx
|
|
Packit |
6c4009 |
and $-16, %rdi
|
|
Packit |
6c4009 |
sub %rcx, %rdx
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(loop_prolog):
|
|
Packit |
6c4009 |
sub $64, %rdx
|
|
Packit |
6c4009 |
jbe L(exit_loop)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movdqa 48(%rdi), %xmm0
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm0
|
|
Packit |
6c4009 |
pmovmskb %xmm0, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jnz L(matches48)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movdqa 32(%rdi), %xmm2
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm2
|
|
Packit |
6c4009 |
pmovmskb %xmm2, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jnz L(matches32)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movdqa 16(%rdi), %xmm3
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm3
|
|
Packit |
6c4009 |
pmovmskb %xmm3, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jnz L(matches16)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movdqa (%rdi), %xmm4
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm4
|
|
Packit |
6c4009 |
pmovmskb %xmm4, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jnz L(matches0)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
sub $64, %rdi
|
|
Packit |
6c4009 |
sub $64, %rdx
|
|
Packit |
6c4009 |
jbe L(exit_loop)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movdqa 48(%rdi), %xmm0
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm0
|
|
Packit |
6c4009 |
pmovmskb %xmm0, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jnz L(matches48)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movdqa 32(%rdi), %xmm2
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm2
|
|
Packit |
6c4009 |
pmovmskb %xmm2, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jnz L(matches32)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movdqa 16(%rdi), %xmm3
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm3
|
|
Packit |
6c4009 |
pmovmskb %xmm3, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jnz L(matches16)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movdqa (%rdi), %xmm3
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm3
|
|
Packit |
6c4009 |
pmovmskb %xmm3, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jnz L(matches0)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
mov %edi, %ecx
|
|
Packit |
6c4009 |
and $63, %ecx
|
|
Packit |
6c4009 |
jz L(align64_loop)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
add $64, %rdi
|
|
Packit |
6c4009 |
add $64, %rdx
|
|
Packit |
6c4009 |
and $-64, %rdi
|
|
Packit |
6c4009 |
sub %rcx, %rdx
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(align64_loop):
|
|
Packit |
6c4009 |
sub $64, %rdi
|
|
Packit |
6c4009 |
sub $64, %rdx
|
|
Packit |
6c4009 |
jbe L(exit_loop)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movdqa (%rdi), %xmm0
|
|
Packit |
6c4009 |
movdqa 16(%rdi), %xmm2
|
|
Packit |
6c4009 |
movdqa 32(%rdi), %xmm3
|
|
Packit |
6c4009 |
movdqa 48(%rdi), %xmm4
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm0
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm2
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm3
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm4
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
pmaxub %xmm3, %xmm0
|
|
Packit |
6c4009 |
pmaxub %xmm4, %xmm2
|
|
Packit |
6c4009 |
pmaxub %xmm0, %xmm2
|
|
Packit |
6c4009 |
pmovmskb %xmm2, %eax
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jz L(align64_loop)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
pmovmskb %xmm4, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jnz L(matches48)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
pmovmskb %xmm3, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jnz L(matches32)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movdqa 16(%rdi), %xmm2
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm2
|
|
Packit |
6c4009 |
pcmpeqb (%rdi), %xmm1
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
pmovmskb %xmm2, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jnz L(matches16)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
pmovmskb %xmm1, %eax
|
|
Packit |
6c4009 |
bsr %eax, %eax
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
add %rdi, %rax
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(exit_loop):
|
|
Packit |
6c4009 |
add $64, %edx
|
|
Packit |
6c4009 |
cmp $32, %edx
|
|
Packit |
6c4009 |
jbe L(exit_loop_32)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movdqa 48(%rdi), %xmm0
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm0
|
|
Packit |
6c4009 |
pmovmskb %xmm0, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jnz L(matches48)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movdqa 32(%rdi), %xmm2
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm2
|
|
Packit |
6c4009 |
pmovmskb %xmm2, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jnz L(matches32)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
movdqa 16(%rdi), %xmm3
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm3
|
|
Packit |
6c4009 |
pmovmskb %xmm3, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jnz L(matches16_1)
|
|
Packit |
6c4009 |
cmp $48, %edx
|
|
Packit |
6c4009 |
jbe L(return_null)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
pcmpeqb (%rdi), %xmm1
|
|
Packit |
6c4009 |
pmovmskb %xmm1, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jnz L(matches0_1)
|
|
Packit |
6c4009 |
xor %eax, %eax
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(exit_loop_32):
|
|
Packit |
6c4009 |
movdqa 48(%rdi), %xmm0
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm0
|
|
Packit |
6c4009 |
pmovmskb %xmm0, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jnz L(matches48_1)
|
|
Packit |
6c4009 |
cmp $16, %edx
|
|
Packit |
6c4009 |
jbe L(return_null)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
pcmpeqb 32(%rdi), %xmm1
|
|
Packit |
6c4009 |
pmovmskb %xmm1, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jnz L(matches32_1)
|
|
Packit |
6c4009 |
xor %eax, %eax
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(matches0):
|
|
Packit |
6c4009 |
bsr %eax, %eax
|
|
Packit |
6c4009 |
add %rdi, %rax
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(matches16):
|
|
Packit |
6c4009 |
bsr %eax, %eax
|
|
Packit |
6c4009 |
lea 16(%rax, %rdi), %rax
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(matches32):
|
|
Packit |
6c4009 |
bsr %eax, %eax
|
|
Packit |
6c4009 |
lea 32(%rax, %rdi), %rax
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(matches48):
|
|
Packit |
6c4009 |
bsr %eax, %eax
|
|
Packit |
6c4009 |
lea 48(%rax, %rdi), %rax
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(matches0_1):
|
|
Packit |
6c4009 |
bsr %eax, %eax
|
|
Packit |
6c4009 |
sub $64, %rdx
|
|
Packit |
6c4009 |
add %rax, %rdx
|
|
Packit |
6c4009 |
jl L(return_null)
|
|
Packit |
6c4009 |
add %rdi, %rax
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(matches16_1):
|
|
Packit |
6c4009 |
bsr %eax, %eax
|
|
Packit |
6c4009 |
sub $48, %rdx
|
|
Packit |
6c4009 |
add %rax, %rdx
|
|
Packit |
6c4009 |
jl L(return_null)
|
|
Packit |
6c4009 |
lea 16(%rdi, %rax), %rax
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(matches32_1):
|
|
Packit |
6c4009 |
bsr %eax, %eax
|
|
Packit |
6c4009 |
sub $32, %rdx
|
|
Packit |
6c4009 |
add %rax, %rdx
|
|
Packit |
6c4009 |
jl L(return_null)
|
|
Packit |
6c4009 |
lea 32(%rdi, %rax), %rax
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(matches48_1):
|
|
Packit |
6c4009 |
bsr %eax, %eax
|
|
Packit |
6c4009 |
sub $16, %rdx
|
|
Packit |
6c4009 |
add %rax, %rdx
|
|
Packit |
6c4009 |
jl L(return_null)
|
|
Packit |
6c4009 |
lea 48(%rdi, %rax), %rax
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(return_null):
|
|
Packit |
6c4009 |
xor %eax, %eax
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(length_less16_offset0):
|
|
Packit |
6c4009 |
test %edx, %edx
|
|
Packit |
6c4009 |
jz L(return_null)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
mov %dl, %cl
|
|
Packit |
6c4009 |
pcmpeqb (%rdi), %xmm1
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
mov $1, %edx
|
|
Packit |
6c4009 |
sal %cl, %edx
|
|
Packit |
6c4009 |
sub $1, %edx
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
pmovmskb %xmm1, %eax
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
and %edx, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jz L(return_null)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
bsr %eax, %eax
|
|
Packit |
6c4009 |
add %rdi, %rax
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(length_less16):
|
|
Packit |
6c4009 |
punpcklbw %xmm1, %xmm1
|
|
Packit |
6c4009 |
punpcklbw %xmm1, %xmm1
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
add $16, %edx
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
pshufd $0, %xmm1, %xmm1
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
mov %edi, %ecx
|
|
Packit |
6c4009 |
and $15, %ecx
|
|
Packit |
6c4009 |
jz L(length_less16_offset0)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
mov %cl, %dh
|
|
Packit |
6c4009 |
mov %ecx, %esi
|
|
Packit |
6c4009 |
add %dl, %dh
|
|
Packit |
6c4009 |
and $-16, %rdi
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
sub $16, %dh
|
|
Packit |
6c4009 |
ja L(length_less16_part2)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
pcmpeqb (%rdi), %xmm1
|
|
Packit |
6c4009 |
pmovmskb %xmm1, %eax
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
sar %cl, %eax
|
|
Packit |
6c4009 |
mov %dl, %cl
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
mov $1, %edx
|
|
Packit |
6c4009 |
sal %cl, %edx
|
|
Packit |
6c4009 |
sub $1, %edx
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
and %edx, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jz L(return_null)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
bsr %eax, %eax
|
|
Packit |
6c4009 |
add %rdi, %rax
|
|
Packit |
6c4009 |
add %rsi, %rax
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(length_less16_part2):
|
|
Packit |
6c4009 |
movdqa 16(%rdi), %xmm2
|
|
Packit |
6c4009 |
pcmpeqb %xmm1, %xmm2
|
|
Packit |
6c4009 |
pmovmskb %xmm2, %eax
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
mov %dh, %cl
|
|
Packit |
6c4009 |
mov $1, %edx
|
|
Packit |
6c4009 |
sal %cl, %edx
|
|
Packit |
6c4009 |
sub $1, %edx
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
and %edx, %eax
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jnz L(length_less16_part2_return)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
pcmpeqb (%rdi), %xmm1
|
|
Packit |
6c4009 |
pmovmskb %xmm1, %eax
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
mov %esi, %ecx
|
|
Packit |
6c4009 |
sar %cl, %eax
|
|
Packit |
6c4009 |
test %eax, %eax
|
|
Packit |
6c4009 |
jz L(return_null)
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
bsr %eax, %eax
|
|
Packit |
6c4009 |
add %rdi, %rax
|
|
Packit |
6c4009 |
add %rsi, %rax
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
.p2align 4
|
|
Packit |
6c4009 |
L(length_less16_part2_return):
|
|
Packit |
6c4009 |
bsr %eax, %eax
|
|
Packit |
6c4009 |
lea 16(%rax, %rdi), %rax
|
|
Packit |
6c4009 |
ret
|
|
Packit |
6c4009 |
|
|
Packit |
6c4009 |
END (__memrchr)
|
|
Packit |
6c4009 |
weak_alias (__memrchr, memrchr)
|