AuroraMiddleware/glibc
1
0
Fork 0
mirror of https://sourceware.org/git/glibc.git synced 2024-11-13 00:30:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

x86-64: Add strcpy family functions with 256-bit EVEX

Browse Source 
Update ifunc-strcpy.h to select the function optimized with 256-bit EVEX
instructions using YMM16-YMM31 registers to avoid RTM abort with usable
AVX512VL and AVX512BW since VZEROUPPER isn't needed at function exit.

(cherry picked from commit 525bc2a32c)
This commit is contained in:
H.J. Lu 2021-03-05 06:36:50 -08:00
parent db9071c0f6
commit 7257ba7bf2
9 changed files with 1339 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
sysdeps/x86_64/multiarch/Makefile
View File

@ -47,11 +47,17 @@ sysdep_routines += strncat-c stpncpy-c strncpy-c \
memchr-evex \
memrchr-evex \
rawmemchr-evex \
stpcpy-evex \
stpncpy-evex \
strcat-evex \
strchr-evex \
strchrnul-evex \
strcmp-evex \
strcpy-evex \
strlen-evex \
strncat-evex \
strncmp-evex \
strncpy-evex \
strnlen-evex \
strrchr-evex
CFLAGS-varshift.c += -msse4

24
sysdeps/x86_64/multiarch/ifunc-impl-list.c
View File

@ -224,6 +224,10 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
__stpncpy_ssse3)
IFUNC_IMPL_ADD (array, i, stpncpy, HAS_ARCH_FEATURE (AVX2_Usable),
__stpncpy_avx2)
IFUNC_IMPL_ADD (array, i, stpncpy,
(HAS_ARCH_FEATURE (AVX512VL_Usable)
&& HAS_ARCH_FEATURE (AVX512BW_Usable)),
__stpncpy_evex)
IFUNC_IMPL_ADD (array, i, stpncpy, 1,
__stpncpy_sse2_unaligned)
IFUNC_IMPL_ADD (array, i, stpncpy, 1, __stpncpy_sse2))
@ -234,6 +238,10 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
__stpcpy_ssse3)
IFUNC_IMPL_ADD (array, i, stpcpy, HAS_ARCH_FEATURE (AVX2_Usable),
__stpcpy_avx2)
IFUNC_IMPL_ADD (array, i, stpcpy,
(HAS_ARCH_FEATURE (AVX512VL_Usable)
&& HAS_ARCH_FEATURE (AVX512BW_Usable)),
__stpcpy_evex)
IFUNC_IMPL_ADD (array, i, stpcpy, 1, __stpcpy_sse2_unaligned)
IFUNC_IMPL_ADD (array, i, stpcpy, 1, __stpcpy_sse2))
@ -268,6 +276,10 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
IFUNC_IMPL (i, name, strcat,
IFUNC_IMPL_ADD (array, i, strcat, HAS_ARCH_FEATURE (AVX2_Usable),
__strcat_avx2)
IFUNC_IMPL_ADD (array, i, strcat,
(HAS_ARCH_FEATURE (AVX512VL_Usable)
&& HAS_ARCH_FEATURE (AVX512BW_Usable)),
__strcat_evex)
IFUNC_IMPL_ADD (array, i, strcat, HAS_CPU_FEATURE (SSSE3),
__strcat_ssse3)
IFUNC_IMPL_ADD (array, i, strcat, 1, __strcat_sse2_unaligned)
@ -330,6 +342,10 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
IFUNC_IMPL (i, name, strcpy,
IFUNC_IMPL_ADD (array, i, strcpy, HAS_ARCH_FEATURE (AVX2_Usable),
__strcpy_avx2)
IFUNC_IMPL_ADD (array, i, strcpy,
(HAS_ARCH_FEATURE (AVX512VL_Usable)
&& HAS_ARCH_FEATURE (AVX512BW_Usable)),
__strcpy_evex)
IFUNC_IMPL_ADD (array, i, strcpy, HAS_CPU_FEATURE (SSSE3),
__strcpy_ssse3)
IFUNC_IMPL_ADD (array, i, strcpy, 1, __strcpy_sse2_unaligned)
@ -373,6 +389,10 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
IFUNC_IMPL (i, name, strncat,
IFUNC_IMPL_ADD (array, i, strncat, HAS_ARCH_FEATURE (AVX2_Usable),
__strncat_avx2)
IFUNC_IMPL_ADD (array, i, strncat,
(HAS_ARCH_FEATURE (AVX512VL_Usable)
&& HAS_ARCH_FEATURE (AVX512BW_Usable)),
__strncat_evex)
IFUNC_IMPL_ADD (array, i, strncat, HAS_CPU_FEATURE (SSSE3),
__strncat_ssse3)
IFUNC_IMPL_ADD (array, i, strncat, 1,
@ -383,6 +403,10 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
IFUNC_IMPL (i, name, strncpy,
IFUNC_IMPL_ADD (array, i, strncpy, HAS_ARCH_FEATURE (AVX2_Usable),
__strncpy_avx2)
IFUNC_IMPL_ADD (array, i, strncpy,
(HAS_ARCH_FEATURE (AVX512VL_Usable)
&& HAS_ARCH_FEATURE (AVX512BW_Usable)),
__strncpy_evex)
IFUNC_IMPL_ADD (array, i, strncpy, HAS_CPU_FEATURE (SSSE3),
__strncpy_ssse3)
IFUNC_IMPL_ADD (array, i, strncpy, 1,

13
sysdeps/x86_64/multiarch/ifunc-strcpy.h
View File

@ -25,16 +25,23 @@ extern __typeof (REDIRECT_NAME) OPTIMIZE (sse2_unaligned)
attribute_hidden;
extern __typeof (REDIRECT_NAME) OPTIMIZE (ssse3) attribute_hidden;
extern __typeof (REDIRECT_NAME) OPTIMIZE (avx2) attribute_hidden;
extern __typeof (REDIRECT_NAME) OPTIMIZE (evex) attribute_hidden;
static inline void *
IFUNC_SELECTOR (void)
{
const struct cpu_features* cpu_features = __get_cpu_features ();
if (!CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_VZEROUPPER)
&& CPU_FEATURES_ARCH_P (cpu_features, AVX2_Usable)
if (CPU_FEATURES_ARCH_P (cpu_features, AVX2_Usable)
&& CPU_FEATURES_ARCH_P (cpu_features, AVX_Fast_Unaligned_Load))
return OPTIMIZE (avx2);
{
if (CPU_FEATURES_ARCH_P (cpu_features, AVX512VL_Usable)
&& CPU_FEATURES_ARCH_P (cpu_features, AVX512BW_Usable))
return OPTIMIZE (evex);
if (!CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_VZEROUPPER))
return OPTIMIZE (avx2);
}
if (CPU_FEATURES_ARCH_P (cpu_features, Fast_Unaligned_Load))
return OPTIMIZE (sse2_unaligned);

3
sysdeps/x86_64/multiarch/stpcpy-evex.S Normal file
View File

@ -0,0 +1,3 @@
#define USE_AS_STPCPY
#define STRCPY __stpcpy_evex
#include "strcpy-evex.S"

4
sysdeps/x86_64/multiarch/stpncpy-evex.S Normal file
View File

@ -0,0 +1,4 @@
#define USE_AS_STPCPY
#define USE_AS_STRNCPY
#define STRCPY __stpncpy_evex
#include "strcpy-evex.S"

283
sysdeps/x86_64/multiarch/strcat-evex.S Normal file
View File

@ -0,0 +1,283 @@
/* strcat with 256-bit EVEX instructions.
Copyright (C) 2021 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/>. */
#if IS_IN (libc)
# include <sysdep.h>
# ifndef STRCAT
# define STRCAT __strcat_evex
# endif
# define VMOVU vmovdqu64
# define VMOVA vmovdqa64
/* zero register */
# define XMMZERO xmm16
# define YMMZERO ymm16
# define YMM0 ymm17
# define YMM1 ymm18
# define USE_AS_STRCAT
/* Number of bytes in a vector register */
# define VEC_SIZE 32
.section .text.evex,"ax",@progbits
ENTRY (STRCAT)
mov %rdi, %r9
# ifdef USE_AS_STRNCAT
mov %rdx, %r8
# endif
xor %eax, %eax
mov %edi, %ecx
and $((VEC_SIZE * 4) - 1), %ecx
vpxorq %XMMZERO, %XMMZERO, %XMMZERO
cmp $(VEC_SIZE * 3), %ecx
ja L(fourth_vector_boundary)
vpcmpb $0, (%rdi), %YMMZERO, %k0
kmovd %k0, %edx
test %edx, %edx
jnz L(exit_null_on_first_vector)
mov %rdi, %rax
and $-VEC_SIZE, %rax
jmp L(align_vec_size_start)
L(fourth_vector_boundary):
mov %rdi, %rax
and $-VEC_SIZE, %rax
vpcmpb $0, (%rax), %YMMZERO, %k0
mov $-1, %r10d
sub %rax, %rcx
shl %cl, %r10d
kmovd %k0, %edx
and %r10d, %edx
jnz L(exit)
L(align_vec_size_start):
vpcmpb $0, VEC_SIZE(%rax), %YMMZERO, %k0
kmovd %k0, %edx
test %edx, %edx
jnz L(exit_null_on_second_vector)
vpcmpb $0, (VEC_SIZE * 2)(%rax), %YMMZERO, %k1
kmovd %k1, %edx
test %edx, %edx
jnz L(exit_null_on_third_vector)
vpcmpb $0, (VEC_SIZE * 3)(%rax), %YMMZERO, %k2
kmovd %k2, %edx
test %edx, %edx
jnz L(exit_null_on_fourth_vector)
vpcmpb $0, (VEC_SIZE * 4)(%rax), %YMMZERO, %k3
kmovd %k3, %edx
test %edx, %edx
jnz L(exit_null_on_fifth_vector)
vpcmpb $0, (VEC_SIZE * 5)(%rax), %YMMZERO, %k4
add $(VEC_SIZE * 4), %rax
kmovd %k4, %edx
test %edx, %edx
jnz L(exit_null_on_second_vector)
vpcmpb $0, (VEC_SIZE * 2)(%rax), %YMMZERO, %k1
kmovd %k1, %edx
test %edx, %edx
jnz L(exit_null_on_third_vector)
vpcmpb $0, (VEC_SIZE * 3)(%rax), %YMMZERO, %k2
kmovd %k2, %edx
test %edx, %edx
jnz L(exit_null_on_fourth_vector)
vpcmpb $0, (VEC_SIZE * 4)(%rax), %YMMZERO, %k3
kmovd %k3, %edx
test %edx, %edx
jnz L(exit_null_on_fifth_vector)
vpcmpb $0, (VEC_SIZE * 5)(%rax), %YMMZERO, %k4
kmovd %k4, %edx
add $(VEC_SIZE * 4), %rax
test %edx, %edx
jnz L(exit_null_on_second_vector)
vpcmpb $0, (VEC_SIZE * 2)(%rax), %YMMZERO, %k1
kmovd %k1, %edx
test %edx, %edx
jnz L(exit_null_on_third_vector)
vpcmpb $0, (VEC_SIZE * 3)(%rax), %YMMZERO, %k2
kmovd %k2, %edx
test %edx, %edx
jnz L(exit_null_on_fourth_vector)
vpcmpb $0, (VEC_SIZE * 4)(%rax), %YMMZERO, %k3
kmovd %k3, %edx
test %edx, %edx
jnz L(exit_null_on_fifth_vector)
vpcmpb $0, (VEC_SIZE * 5)(%rax), %YMMZERO, %k4
add $(VEC_SIZE * 4), %rax
kmovd %k4, %edx
test %edx, %edx
jnz L(exit_null_on_second_vector)
vpcmpb $0, (VEC_SIZE * 2)(%rax), %YMMZERO, %k1
kmovd %k1, %edx
test %edx, %edx
jnz L(exit_null_on_third_vector)
vpcmpb $0, (VEC_SIZE * 3)(%rax), %YMMZERO, %k2
kmovd %k2, %edx
test %edx, %edx
jnz L(exit_null_on_fourth_vector)
vpcmpb $0, (VEC_SIZE * 4)(%rax), %YMMZERO, %k3
kmovd %k3, %edx
test %edx, %edx
jnz L(exit_null_on_fifth_vector)
test $((VEC_SIZE * 4) - 1), %rax
jz L(align_four_vec_loop)
vpcmpb $0, (VEC_SIZE * 5)(%rax), %YMMZERO, %k4
add $(VEC_SIZE * 5), %rax
kmovd %k4, %edx
test %edx, %edx
jnz L(exit)
test $((VEC_SIZE * 4) - 1), %rax
jz L(align_four_vec_loop)
vpcmpb $0, VEC_SIZE(%rax), %YMMZERO, %k0
add $VEC_SIZE, %rax
kmovd %k0, %edx
test %edx, %edx
jnz L(exit)
test $((VEC_SIZE * 4) - 1), %rax
jz L(align_four_vec_loop)
vpcmpb $0, VEC_SIZE(%rax), %YMMZERO, %k0
add $VEC_SIZE, %rax
kmovd %k0, %edx
test %edx, %edx
jnz L(exit)
test $((VEC_SIZE * 4) - 1), %rax
jz L(align_four_vec_loop)
vpcmpb $0, VEC_SIZE(%rax), %YMMZERO, %k1
add $VEC_SIZE, %rax
kmovd %k1, %edx
test %edx, %edx
jnz L(exit)
add $VEC_SIZE, %rax
.p2align 4
L(align_four_vec_loop):
VMOVA (%rax), %YMM0
VMOVA (VEC_SIZE * 2)(%rax), %YMM1
vpminub VEC_SIZE(%rax), %YMM0, %YMM0
vpminub (VEC_SIZE * 3)(%rax), %YMM1, %YMM1
vpminub %YMM0, %YMM1, %YMM0
/* If K0 != 0, there is a null byte. */
vpcmpb $0, %YMM0, %YMMZERO, %k0
add $(VEC_SIZE * 4), %rax
ktestd %k0, %k0
jz L(align_four_vec_loop)
vpcmpb $0, -(VEC_SIZE * 4)(%rax), %YMMZERO, %k0
sub $(VEC_SIZE * 5), %rax
kmovd %k0, %edx
test %edx, %edx
jnz L(exit_null_on_second_vector)
vpcmpb $0, (VEC_SIZE * 2)(%rax), %YMMZERO, %k1
kmovd %k1, %edx
test %edx, %edx
jnz L(exit_null_on_third_vector)
vpcmpb $0, (VEC_SIZE * 3)(%rax), %YMMZERO, %k2
kmovd %k2, %edx
test %edx, %edx
jnz L(exit_null_on_fourth_vector)
vpcmpb $0, (VEC_SIZE * 4)(%rax), %YMMZERO, %k3
kmovd %k3, %edx
sub %rdi, %rax
bsf %rdx, %rdx
add %rdx, %rax
add $(VEC_SIZE * 4), %rax
jmp L(StartStrcpyPart)
.p2align 4
L(exit):
sub %rdi, %rax
L(exit_null_on_first_vector):
bsf %rdx, %rdx
add %rdx, %rax
jmp L(StartStrcpyPart)
.p2align 4
L(exit_null_on_second_vector):
sub %rdi, %rax
bsf %rdx, %rdx
add %rdx, %rax
add $VEC_SIZE, %rax
jmp L(StartStrcpyPart)
.p2align 4
L(exit_null_on_third_vector):
sub %rdi, %rax
bsf %rdx, %rdx
add %rdx, %rax
add $(VEC_SIZE * 2), %rax
jmp L(StartStrcpyPart)
.p2align 4
L(exit_null_on_fourth_vector):
sub %rdi, %rax
bsf %rdx, %rdx
add %rdx, %rax
add $(VEC_SIZE * 3), %rax
jmp L(StartStrcpyPart)
.p2align 4
L(exit_null_on_fifth_vector):
sub %rdi, %rax
bsf %rdx, %rdx
add %rdx, %rax
add $(VEC_SIZE * 4), %rax
.p2align 4
L(StartStrcpyPart):
lea (%r9, %rax), %rdi
mov %rsi, %rcx
mov %r9, %rax /* save result */
# ifdef USE_AS_STRNCAT
test %r8, %r8
jz L(ExitZero)
# define USE_AS_STRNCPY
# endif
# include "strcpy-evex.S"
#endif

1003
sysdeps/x86_64/multiarch/strcpy-evex.S Normal file
View File

File diff suppressed because it is too large Load Diff

3
sysdeps/x86_64/multiarch/strncat-evex.S Normal file
View File

@ -0,0 +1,3 @@
#define USE_AS_STRNCAT
#define STRCAT __strncat_evex
#include "strcat-evex.S"

3
sysdeps/x86_64/multiarch/strncpy-evex.S Normal file
View File

@ -0,0 +1,3 @@
#define USE_AS_STRNCPY
#define STRCPY __strncpy_evex
#include "strcpy-evex.S"