|
 |
190885 |
From 98192464b47c056515b6ac5ff218c197bd75618d Mon Sep 17 00:00:00 2001
|
|
|
190885 |
From: "H.J. Lu" <hjl.tools@gmail.com>
|
|
|
190885 |
Date: Fri, 5 Mar 2021 06:36:50 -0800
|
|
|
190885 |
Subject: [PATCH] x86-64: Add strcpy family functions with 256-bit EVEX
|
|
|
190885 |
|
|
|
190885 |
Update ifunc-strcpy.h to select the function optimized with 256-bit EVEX
|
|
|
190885 |
instructions using YMM16-YMM31 registers to avoid RTM abort with usable
|
|
|
190885 |
AVX512VL and AVX512BW since VZEROUPPER isn't needed at function exit.
|
|
|
190885 |
|
|
|
190885 |
(cherry picked from commit 525bc2a32c9710df40371f951217c6ae7a923aee)
|
|
|
190885 |
---
|
|
|
190885 |
sysdeps/x86_64/multiarch/Makefile | 6 +
|
|
|
190885 |
sysdeps/x86_64/multiarch/ifunc-impl-list.c | 24 +
|
|
|
190885 |
sysdeps/x86_64/multiarch/ifunc-strcpy.h | 13 +-
|
|
|
190885 |
sysdeps/x86_64/multiarch/stpcpy-evex.S | 3 +
|
|
|
190885 |
sysdeps/x86_64/multiarch/stpncpy-evex.S | 4 +
|
|
|
190885 |
sysdeps/x86_64/multiarch/strcat-evex.S | 283 ++++++
|
|
|
190885 |
sysdeps/x86_64/multiarch/strcpy-evex.S | 1003 ++++++++++++++++++++
|
|
|
190885 |
sysdeps/x86_64/multiarch/strncat-evex.S | 3 +
|
|
|
190885 |
sysdeps/x86_64/multiarch/strncpy-evex.S | 3 +
|
|
|
190885 |
9 files changed, 1339 insertions(+), 3 deletions(-)
|
|
|
190885 |
create mode 100644 sysdeps/x86_64/multiarch/stpcpy-evex.S
|
|
|
190885 |
create mode 100644 sysdeps/x86_64/multiarch/stpncpy-evex.S
|
|
|
190885 |
create mode 100644 sysdeps/x86_64/multiarch/strcat-evex.S
|
|
|
190885 |
create mode 100644 sysdeps/x86_64/multiarch/strcpy-evex.S
|
|
|
190885 |
create mode 100644 sysdeps/x86_64/multiarch/strncat-evex.S
|
|
|
190885 |
create mode 100644 sysdeps/x86_64/multiarch/strncpy-evex.S
|
|
|
190885 |
|
|
|
190885 |
diff --git a/sysdeps/x86_64/multiarch/Makefile b/sysdeps/x86_64/multiarch/Makefile
|
|
|
190885 |
index 5ce85882..46783cd1 100644
|
|
|
190885 |
--- a/sysdeps/x86_64/multiarch/Makefile
|
|
|
190885 |
+++ b/sysdeps/x86_64/multiarch/Makefile
|
|
|
190885 |
@@ -43,11 +43,17 @@ sysdep_routines += strncat-c stpncpy-c strncpy-c \
|
|
|
190885 |
memchr-evex \
|
|
|
190885 |
memrchr-evex \
|
|
|
190885 |
rawmemchr-evex \
|
|
|
190885 |
+ stpcpy-evex \
|
|
|
190885 |
+ stpncpy-evex \
|
|
|
190885 |
+ strcat-evex \
|
|
|
190885 |
strchr-evex \
|
|
|
190885 |
strchrnul-evex \
|
|
|
190885 |
strcmp-evex \
|
|
|
190885 |
+ strcpy-evex \
|
|
|
190885 |
strlen-evex \
|
|
|
190885 |
+ strncat-evex \
|
|
|
190885 |
strncmp-evex \
|
|
|
190885 |
+ strncpy-evex \
|
|
|
190885 |
strnlen-evex \
|
|
|
190885 |
strrchr-evex
|
|
|
190885 |
CFLAGS-varshift.c += -msse4
|
|
|
190885 |
diff --git a/sysdeps/x86_64/multiarch/ifunc-impl-list.c b/sysdeps/x86_64/multiarch/ifunc-impl-list.c
|
|
|
190885 |
index bd7d9f19..082e4da3 100644
|
|
|
190885 |
--- a/sysdeps/x86_64/multiarch/ifunc-impl-list.c
|
|
|
190885 |
+++ b/sysdeps/x86_64/multiarch/ifunc-impl-list.c
|
|
|
190885 |
@@ -224,6 +224,10 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
|
|
|
190885 |
__stpncpy_ssse3)
|
|
|
190885 |
IFUNC_IMPL_ADD (array, i, stpncpy, CPU_FEATURE_USABLE (AVX2),
|
|
|
190885 |
__stpncpy_avx2)
|
|
|
190885 |
+ IFUNC_IMPL_ADD (array, i, stpncpy,
|
|
|
190885 |
+ (CPU_FEATURE_USABLE (AVX512VL)
|
|
|
190885 |
+ && CPU_FEATURE_USABLE (AVX512BW)),
|
|
|
190885 |
+ __stpncpy_evex)
|
|
|
190885 |
IFUNC_IMPL_ADD (array, i, stpncpy, 1,
|
|
|
190885 |
__stpncpy_sse2_unaligned)
|
|
|
190885 |
IFUNC_IMPL_ADD (array, i, stpncpy, 1, __stpncpy_sse2))
|
|
|
190885 |
@@ -234,6 +238,10 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
|
|
|
190885 |
__stpcpy_ssse3)
|
|
|
190885 |
IFUNC_IMPL_ADD (array, i, stpcpy, CPU_FEATURE_USABLE (AVX2),
|
|
|
190885 |
__stpcpy_avx2)
|
|
|
190885 |
+ IFUNC_IMPL_ADD (array, i, stpcpy,
|
|
|
190885 |
+ (CPU_FEATURE_USABLE (AVX512VL)
|
|
|
190885 |
+ && CPU_FEATURE_USABLE (AVX512BW)),
|
|
|
190885 |
+ __stpcpy_evex)
|
|
|
190885 |
IFUNC_IMPL_ADD (array, i, stpcpy, 1, __stpcpy_sse2_unaligned)
|
|
|
190885 |
IFUNC_IMPL_ADD (array, i, stpcpy, 1, __stpcpy_sse2))
|
|
|
190885 |
|
|
|
190885 |
@@ -268,6 +276,10 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
|
|
|
190885 |
IFUNC_IMPL (i, name, strcat,
|
|
|
190885 |
IFUNC_IMPL_ADD (array, i, strcat, CPU_FEATURE_USABLE (AVX2),
|
|
|
190885 |
__strcat_avx2)
|
|
|
190885 |
+ IFUNC_IMPL_ADD (array, i, strcat,
|
|
|
190885 |
+ (CPU_FEATURE_USABLE (AVX512VL)
|
|
|
190885 |
+ && CPU_FEATURE_USABLE (AVX512BW)),
|
|
|
190885 |
+ __strcat_evex)
|
|
|
190885 |
IFUNC_IMPL_ADD (array, i, strcat, CPU_FEATURE_USABLE (SSSE3),
|
|
|
190885 |
__strcat_ssse3)
|
|
|
190885 |
IFUNC_IMPL_ADD (array, i, strcat, 1, __strcat_sse2_unaligned)
|
|
|
190885 |
@@ -330,6 +342,10 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
|
|
|
190885 |
IFUNC_IMPL (i, name, strcpy,
|
|
|
190885 |
IFUNC_IMPL_ADD (array, i, strcpy, CPU_FEATURE_USABLE (AVX2),
|
|
|
190885 |
__strcpy_avx2)
|
|
|
190885 |
+ IFUNC_IMPL_ADD (array, i, strcpy,
|
|
|
190885 |
+ (CPU_FEATURE_USABLE (AVX512VL)
|
|
|
190885 |
+ && CPU_FEATURE_USABLE (AVX512BW)),
|
|
|
190885 |
+ __strcpy_evex)
|
|
|
190885 |
IFUNC_IMPL_ADD (array, i, strcpy, CPU_FEATURE_USABLE (SSSE3),
|
|
|
190885 |
__strcpy_ssse3)
|
|
|
190885 |
IFUNC_IMPL_ADD (array, i, strcpy, 1, __strcpy_sse2_unaligned)
|
|
|
190885 |
@@ -373,6 +389,10 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
|
|
|
190885 |
IFUNC_IMPL (i, name, strncat,
|
|
|
190885 |
IFUNC_IMPL_ADD (array, i, strncat, CPU_FEATURE_USABLE (AVX2),
|
|
|
190885 |
__strncat_avx2)
|
|
|
190885 |
+ IFUNC_IMPL_ADD (array, i, strncat,
|
|
|
190885 |
+ (CPU_FEATURE_USABLE (AVX512VL)
|
|
|
190885 |
+ && CPU_FEATURE_USABLE (AVX512BW)),
|
|
|
190885 |
+ __strncat_evex)
|
|
|
190885 |
IFUNC_IMPL_ADD (array, i, strncat, CPU_FEATURE_USABLE (SSSE3),
|
|
|
190885 |
__strncat_ssse3)
|
|
|
190885 |
IFUNC_IMPL_ADD (array, i, strncat, 1,
|
|
|
190885 |
@@ -383,6 +403,10 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
|
|
|
190885 |
IFUNC_IMPL (i, name, strncpy,
|
|
|
190885 |
IFUNC_IMPL_ADD (array, i, strncpy, CPU_FEATURE_USABLE (AVX2),
|
|
|
190885 |
__strncpy_avx2)
|
|
|
190885 |
+ IFUNC_IMPL_ADD (array, i, strncpy,
|
|
|
190885 |
+ (CPU_FEATURE_USABLE (AVX512VL)
|
|
|
190885 |
+ && CPU_FEATURE_USABLE (AVX512BW)),
|
|
|
190885 |
+ __strncpy_evex)
|
|
|
190885 |
IFUNC_IMPL_ADD (array, i, strncpy, CPU_FEATURE_USABLE (SSSE3),
|
|
|
190885 |
__strncpy_ssse3)
|
|
|
190885 |
IFUNC_IMPL_ADD (array, i, strncpy, 1,
|
|
|
190885 |
diff --git a/sysdeps/x86_64/multiarch/ifunc-strcpy.h b/sysdeps/x86_64/multiarch/ifunc-strcpy.h
|
|
|
190885 |
index 100dca5c..deae6348 100644
|
|
|
190885 |
--- a/sysdeps/x86_64/multiarch/ifunc-strcpy.h
|
|
|
190885 |
+++ b/sysdeps/x86_64/multiarch/ifunc-strcpy.h
|
|
|
190885 |
@@ -25,16 +25,23 @@ extern __typeof (REDIRECT_NAME) OPTIMIZE (sse2_unaligned)
|
|
|
190885 |
attribute_hidden;
|
|
|
190885 |
extern __typeof (REDIRECT_NAME) OPTIMIZE (ssse3) attribute_hidden;
|
|
|
190885 |
extern __typeof (REDIRECT_NAME) OPTIMIZE (avx2) attribute_hidden;
|
|
|
190885 |
+extern __typeof (REDIRECT_NAME) OPTIMIZE (evex) attribute_hidden;
|
|
|
190885 |
|
|
|
190885 |
static inline void *
|
|
|
190885 |
IFUNC_SELECTOR (void)
|
|
|
190885 |
{
|
|
|
190885 |
const struct cpu_features* cpu_features = __get_cpu_features ();
|
|
|
190885 |
|
|
|
190885 |
- if (!CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_VZEROUPPER)
|
|
|
190885 |
- && CPU_FEATURE_USABLE_P (cpu_features, AVX2)
|
|
|
190885 |
+ if (CPU_FEATURE_USABLE_P (cpu_features, AVX2)
|
|
|
190885 |
&& CPU_FEATURES_ARCH_P (cpu_features, AVX_Fast_Unaligned_Load))
|
|
|
190885 |
- return OPTIMIZE (avx2);
|
|
|
190885 |
+ {
|
|
|
190885 |
+ if (CPU_FEATURE_USABLE_P (cpu_features, AVX512VL)
|
|
|
190885 |
+ && CPU_FEATURE_USABLE_P (cpu_features, AVX512BW))
|
|
|
190885 |
+ return OPTIMIZE (evex);
|
|
|
190885 |
+
|
|
|
190885 |
+ if (!CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_VZEROUPPER))
|
|
|
190885 |
+ return OPTIMIZE (avx2);
|
|
|
190885 |
+ }
|
|
|
190885 |
|
|
|
190885 |
if (CPU_FEATURES_ARCH_P (cpu_features, Fast_Unaligned_Load))
|
|
|
190885 |
return OPTIMIZE (sse2_unaligned);
|
|
|
190885 |
diff --git a/sysdeps/x86_64/multiarch/stpcpy-evex.S b/sysdeps/x86_64/multiarch/stpcpy-evex.S
|
|
|
190885 |
new file mode 100644
|
|
|
190885 |
index 00000000..7c6f26cd
|
|
|
190885 |
--- /dev/null
|
|
|
190885 |
+++ b/sysdeps/x86_64/multiarch/stpcpy-evex.S
|
|
|
190885 |
@@ -0,0 +1,3 @@
|
|
|
190885 |
+#define USE_AS_STPCPY
|
|
|
190885 |
+#define STRCPY __stpcpy_evex
|
|
|
190885 |
+#include "strcpy-evex.S"
|
|
|
190885 |
diff --git a/sysdeps/x86_64/multiarch/stpncpy-evex.S b/sysdeps/x86_64/multiarch/stpncpy-evex.S
|
|
|
190885 |
new file mode 100644
|
|
|
190885 |
index 00000000..1570014d
|
|
|
190885 |
--- /dev/null
|
|
|
190885 |
+++ b/sysdeps/x86_64/multiarch/stpncpy-evex.S
|
|
|
190885 |
@@ -0,0 +1,4 @@
|
|
|
190885 |
+#define USE_AS_STPCPY
|
|
|
190885 |
+#define USE_AS_STRNCPY
|
|
|
190885 |
+#define STRCPY __stpncpy_evex
|
|
|
190885 |
+#include "strcpy-evex.S"
|
|
|
190885 |
diff --git a/sysdeps/x86_64/multiarch/strcat-evex.S b/sysdeps/x86_64/multiarch/strcat-evex.S
|
|
|
190885 |
new file mode 100644
|
|
|
190885 |
index 00000000..97c3d85b
|
|
|
190885 |
--- /dev/null
|
|
|
190885 |
+++ b/sysdeps/x86_64/multiarch/strcat-evex.S
|
|
|
190885 |
@@ -0,0 +1,283 @@
|
|
|
190885 |
+/* strcat with 256-bit EVEX instructions.
|
|
|
190885 |
+ Copyright (C) 2021 Free Software Foundation, Inc.
|
|
|
190885 |
+ This file is part of the GNU C Library.
|
|
|
190885 |
+
|
|
|
190885 |
+ The GNU C Library is free software; you can redistribute it and/or
|
|
|
190885 |
+ modify it under the terms of the GNU Lesser General Public
|
|
|
190885 |
+ License as published by the Free Software Foundation; either
|
|
|
190885 |
+ version 2.1 of the License, or (at your option) any later version.
|
|
|
190885 |
+
|
|
|
190885 |
+ The GNU C Library is distributed in the hope that it will be useful,
|
|
|
190885 |
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
190885 |
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
190885 |
+ Lesser General Public License for more details.
|
|
|
190885 |
+
|
|
|
190885 |
+ You should have received a copy of the GNU Lesser General Public
|
|
|
190885 |
+ License along with the GNU C Library; if not, see
|
|
|
190885 |
+ <https://www.gnu.org/licenses/>. */
|
|
|
190885 |
+
|
|
|
190885 |
+#if IS_IN (libc)
|
|
|
190885 |
+
|
|
|
190885 |
+# include <sysdep.h>
|
|
|
190885 |
+
|
|
|
190885 |
+# ifndef STRCAT
|
|
|
190885 |
+# define STRCAT __strcat_evex
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+# define VMOVU vmovdqu64
|
|
|
190885 |
+# define VMOVA vmovdqa64
|
|
|
190885 |
+
|
|
|
190885 |
+/* zero register */
|
|
|
190885 |
+# define XMMZERO xmm16
|
|
|
190885 |
+# define YMMZERO ymm16
|
|
|
190885 |
+# define YMM0 ymm17
|
|
|
190885 |
+# define YMM1 ymm18
|
|
|
190885 |
+
|
|
|
190885 |
+# define USE_AS_STRCAT
|
|
|
190885 |
+
|
|
|
190885 |
+/* Number of bytes in a vector register */
|
|
|
190885 |
+# define VEC_SIZE 32
|
|
|
190885 |
+
|
|
|
190885 |
+ .section .text.evex,"ax",@progbits
|
|
|
190885 |
+ENTRY (STRCAT)
|
|
|
190885 |
+ mov %rdi, %r9
|
|
|
190885 |
+# ifdef USE_AS_STRNCAT
|
|
|
190885 |
+ mov %rdx, %r8
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+ xor %eax, %eax
|
|
|
190885 |
+ mov %edi, %ecx
|
|
|
190885 |
+ and $((VEC_SIZE * 4) - 1), %ecx
|
|
|
190885 |
+ vpxorq %XMMZERO, %XMMZERO, %XMMZERO
|
|
|
190885 |
+ cmp $(VEC_SIZE * 3), %ecx
|
|
|
190885 |
+ ja L(fourth_vector_boundary)
|
|
|
190885 |
+ vpcmpb $0, (%rdi), %YMMZERO, %k0
|
|
|
190885 |
+ kmovd %k0, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit_null_on_first_vector)
|
|
|
190885 |
+ mov %rdi, %rax
|
|
|
190885 |
+ and $-VEC_SIZE, %rax
|
|
|
190885 |
+ jmp L(align_vec_size_start)
|
|
|
190885 |
+L(fourth_vector_boundary):
|
|
|
190885 |
+ mov %rdi, %rax
|
|
|
190885 |
+ and $-VEC_SIZE, %rax
|
|
|
190885 |
+ vpcmpb $0, (%rax), %YMMZERO, %k0
|
|
|
190885 |
+ mov $-1, %r10d
|
|
|
190885 |
+ sub %rax, %rcx
|
|
|
190885 |
+ shl %cl, %r10d
|
|
|
190885 |
+ kmovd %k0, %edx
|
|
|
190885 |
+ and %r10d, %edx
|
|
|
190885 |
+ jnz L(exit)
|
|
|
190885 |
+
|
|
|
190885 |
+L(align_vec_size_start):
|
|
|
190885 |
+ vpcmpb $0, VEC_SIZE(%rax), %YMMZERO, %k0
|
|
|
190885 |
+ kmovd %k0, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit_null_on_second_vector)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, (VEC_SIZE * 2)(%rax), %YMMZERO, %k1
|
|
|
190885 |
+ kmovd %k1, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit_null_on_third_vector)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, (VEC_SIZE * 3)(%rax), %YMMZERO, %k2
|
|
|
190885 |
+ kmovd %k2, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit_null_on_fourth_vector)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, (VEC_SIZE * 4)(%rax), %YMMZERO, %k3
|
|
|
190885 |
+ kmovd %k3, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit_null_on_fifth_vector)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, (VEC_SIZE * 5)(%rax), %YMMZERO, %k4
|
|
|
190885 |
+ add $(VEC_SIZE * 4), %rax
|
|
|
190885 |
+ kmovd %k4, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit_null_on_second_vector)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, (VEC_SIZE * 2)(%rax), %YMMZERO, %k1
|
|
|
190885 |
+ kmovd %k1, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit_null_on_third_vector)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, (VEC_SIZE * 3)(%rax), %YMMZERO, %k2
|
|
|
190885 |
+ kmovd %k2, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit_null_on_fourth_vector)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, (VEC_SIZE * 4)(%rax), %YMMZERO, %k3
|
|
|
190885 |
+ kmovd %k3, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit_null_on_fifth_vector)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, (VEC_SIZE * 5)(%rax), %YMMZERO, %k4
|
|
|
190885 |
+ kmovd %k4, %edx
|
|
|
190885 |
+ add $(VEC_SIZE * 4), %rax
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit_null_on_second_vector)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, (VEC_SIZE * 2)(%rax), %YMMZERO, %k1
|
|
|
190885 |
+ kmovd %k1, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit_null_on_third_vector)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, (VEC_SIZE * 3)(%rax), %YMMZERO, %k2
|
|
|
190885 |
+ kmovd %k2, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit_null_on_fourth_vector)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, (VEC_SIZE * 4)(%rax), %YMMZERO, %k3
|
|
|
190885 |
+ kmovd %k3, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit_null_on_fifth_vector)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, (VEC_SIZE * 5)(%rax), %YMMZERO, %k4
|
|
|
190885 |
+ add $(VEC_SIZE * 4), %rax
|
|
|
190885 |
+ kmovd %k4, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit_null_on_second_vector)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, (VEC_SIZE * 2)(%rax), %YMMZERO, %k1
|
|
|
190885 |
+ kmovd %k1, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit_null_on_third_vector)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, (VEC_SIZE * 3)(%rax), %YMMZERO, %k2
|
|
|
190885 |
+ kmovd %k2, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit_null_on_fourth_vector)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, (VEC_SIZE * 4)(%rax), %YMMZERO, %k3
|
|
|
190885 |
+ kmovd %k3, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit_null_on_fifth_vector)
|
|
|
190885 |
+
|
|
|
190885 |
+ test $((VEC_SIZE * 4) - 1), %rax
|
|
|
190885 |
+ jz L(align_four_vec_loop)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, (VEC_SIZE * 5)(%rax), %YMMZERO, %k4
|
|
|
190885 |
+ add $(VEC_SIZE * 5), %rax
|
|
|
190885 |
+ kmovd %k4, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit)
|
|
|
190885 |
+
|
|
|
190885 |
+ test $((VEC_SIZE * 4) - 1), %rax
|
|
|
190885 |
+ jz L(align_four_vec_loop)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, VEC_SIZE(%rax), %YMMZERO, %k0
|
|
|
190885 |
+ add $VEC_SIZE, %rax
|
|
|
190885 |
+ kmovd %k0, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit)
|
|
|
190885 |
+
|
|
|
190885 |
+ test $((VEC_SIZE * 4) - 1), %rax
|
|
|
190885 |
+ jz L(align_four_vec_loop)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, VEC_SIZE(%rax), %YMMZERO, %k0
|
|
|
190885 |
+ add $VEC_SIZE, %rax
|
|
|
190885 |
+ kmovd %k0, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit)
|
|
|
190885 |
+
|
|
|
190885 |
+ test $((VEC_SIZE * 4) - 1), %rax
|
|
|
190885 |
+ jz L(align_four_vec_loop)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, VEC_SIZE(%rax), %YMMZERO, %k1
|
|
|
190885 |
+ add $VEC_SIZE, %rax
|
|
|
190885 |
+ kmovd %k1, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit)
|
|
|
190885 |
+
|
|
|
190885 |
+ add $VEC_SIZE, %rax
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(align_four_vec_loop):
|
|
|
190885 |
+ VMOVA (%rax), %YMM0
|
|
|
190885 |
+ VMOVA (VEC_SIZE * 2)(%rax), %YMM1
|
|
|
190885 |
+ vpminub VEC_SIZE(%rax), %YMM0, %YMM0
|
|
|
190885 |
+ vpminub (VEC_SIZE * 3)(%rax), %YMM1, %YMM1
|
|
|
190885 |
+ vpminub %YMM0, %YMM1, %YMM0
|
|
|
190885 |
+ /* If K0 != 0, there is a null byte. */
|
|
|
190885 |
+ vpcmpb $0, %YMM0, %YMMZERO, %k0
|
|
|
190885 |
+ add $(VEC_SIZE * 4), %rax
|
|
|
190885 |
+ ktestd %k0, %k0
|
|
|
190885 |
+ jz L(align_four_vec_loop)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, -(VEC_SIZE * 4)(%rax), %YMMZERO, %k0
|
|
|
190885 |
+ sub $(VEC_SIZE * 5), %rax
|
|
|
190885 |
+ kmovd %k0, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit_null_on_second_vector)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, (VEC_SIZE * 2)(%rax), %YMMZERO, %k1
|
|
|
190885 |
+ kmovd %k1, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit_null_on_third_vector)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, (VEC_SIZE * 3)(%rax), %YMMZERO, %k2
|
|
|
190885 |
+ kmovd %k2, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(exit_null_on_fourth_vector)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, (VEC_SIZE * 4)(%rax), %YMMZERO, %k3
|
|
|
190885 |
+ kmovd %k3, %edx
|
|
|
190885 |
+ sub %rdi, %rax
|
|
|
190885 |
+ bsf %rdx, %rdx
|
|
|
190885 |
+ add %rdx, %rax
|
|
|
190885 |
+ add $(VEC_SIZE * 4), %rax
|
|
|
190885 |
+ jmp L(StartStrcpyPart)
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(exit):
|
|
|
190885 |
+ sub %rdi, %rax
|
|
|
190885 |
+L(exit_null_on_first_vector):
|
|
|
190885 |
+ bsf %rdx, %rdx
|
|
|
190885 |
+ add %rdx, %rax
|
|
|
190885 |
+ jmp L(StartStrcpyPart)
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(exit_null_on_second_vector):
|
|
|
190885 |
+ sub %rdi, %rax
|
|
|
190885 |
+ bsf %rdx, %rdx
|
|
|
190885 |
+ add %rdx, %rax
|
|
|
190885 |
+ add $VEC_SIZE, %rax
|
|
|
190885 |
+ jmp L(StartStrcpyPart)
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(exit_null_on_third_vector):
|
|
|
190885 |
+ sub %rdi, %rax
|
|
|
190885 |
+ bsf %rdx, %rdx
|
|
|
190885 |
+ add %rdx, %rax
|
|
|
190885 |
+ add $(VEC_SIZE * 2), %rax
|
|
|
190885 |
+ jmp L(StartStrcpyPart)
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(exit_null_on_fourth_vector):
|
|
|
190885 |
+ sub %rdi, %rax
|
|
|
190885 |
+ bsf %rdx, %rdx
|
|
|
190885 |
+ add %rdx, %rax
|
|
|
190885 |
+ add $(VEC_SIZE * 3), %rax
|
|
|
190885 |
+ jmp L(StartStrcpyPart)
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(exit_null_on_fifth_vector):
|
|
|
190885 |
+ sub %rdi, %rax
|
|
|
190885 |
+ bsf %rdx, %rdx
|
|
|
190885 |
+ add %rdx, %rax
|
|
|
190885 |
+ add $(VEC_SIZE * 4), %rax
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(StartStrcpyPart):
|
|
|
190885 |
+ lea (%r9, %rax), %rdi
|
|
|
190885 |
+ mov %rsi, %rcx
|
|
|
190885 |
+ mov %r9, %rax /* save result */
|
|
|
190885 |
+
|
|
|
190885 |
+# ifdef USE_AS_STRNCAT
|
|
|
190885 |
+ test %r8, %r8
|
|
|
190885 |
+ jz L(ExitZero)
|
|
|
190885 |
+# define USE_AS_STRNCPY
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+# include "strcpy-evex.S"
|
|
|
190885 |
+#endif
|
|
|
190885 |
diff --git a/sysdeps/x86_64/multiarch/strcpy-evex.S b/sysdeps/x86_64/multiarch/strcpy-evex.S
|
|
|
190885 |
new file mode 100644
|
|
|
190885 |
index 00000000..a343a1a6
|
|
|
190885 |
--- /dev/null
|
|
|
190885 |
+++ b/sysdeps/x86_64/multiarch/strcpy-evex.S
|
|
|
190885 |
@@ -0,0 +1,1003 @@
|
|
|
190885 |
+/* strcpy with 256-bit EVEX instructions.
|
|
|
190885 |
+ Copyright (C) 2021 Free Software Foundation, Inc.
|
|
|
190885 |
+ This file is part of the GNU C Library.
|
|
|
190885 |
+
|
|
|
190885 |
+ The GNU C Library is free software; you can redistribute it and/or
|
|
|
190885 |
+ modify it under the terms of the GNU Lesser General Public
|
|
|
190885 |
+ License as published by the Free Software Foundation; either
|
|
|
190885 |
+ version 2.1 of the License, or (at your option) any later version.
|
|
|
190885 |
+
|
|
|
190885 |
+ The GNU C Library is distributed in the hope that it will be useful,
|
|
|
190885 |
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
190885 |
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
190885 |
+ Lesser General Public License for more details.
|
|
|
190885 |
+
|
|
|
190885 |
+ You should have received a copy of the GNU Lesser General Public
|
|
|
190885 |
+ License along with the GNU C Library; if not, see
|
|
|
190885 |
+ <https://www.gnu.org/licenses/>. */
|
|
|
190885 |
+
|
|
|
190885 |
+#if IS_IN (libc)
|
|
|
190885 |
+
|
|
|
190885 |
+# ifndef USE_AS_STRCAT
|
|
|
190885 |
+# include <sysdep.h>
|
|
|
190885 |
+
|
|
|
190885 |
+# ifndef STRCPY
|
|
|
190885 |
+# define STRCPY __strcpy_evex
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+# define VMOVU vmovdqu64
|
|
|
190885 |
+# define VMOVA vmovdqa64
|
|
|
190885 |
+
|
|
|
190885 |
+/* Number of bytes in a vector register */
|
|
|
190885 |
+# ifndef VEC_SIZE
|
|
|
190885 |
+# define VEC_SIZE 32
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+# define XMM2 xmm18
|
|
|
190885 |
+# define XMM3 xmm19
|
|
|
190885 |
+
|
|
|
190885 |
+# define YMM2 ymm18
|
|
|
190885 |
+# define YMM3 ymm19
|
|
|
190885 |
+# define YMM4 ymm20
|
|
|
190885 |
+# define YMM5 ymm21
|
|
|
190885 |
+# define YMM6 ymm22
|
|
|
190885 |
+# define YMM7 ymm23
|
|
|
190885 |
+
|
|
|
190885 |
+# ifndef USE_AS_STRCAT
|
|
|
190885 |
+
|
|
|
190885 |
+/* zero register */
|
|
|
190885 |
+# define XMMZERO xmm16
|
|
|
190885 |
+# define YMMZERO ymm16
|
|
|
190885 |
+# define YMM1 ymm17
|
|
|
190885 |
+
|
|
|
190885 |
+ .section .text.evex,"ax",@progbits
|
|
|
190885 |
+ENTRY (STRCPY)
|
|
|
190885 |
+# ifdef USE_AS_STRNCPY
|
|
|
190885 |
+ mov %RDX_LP, %R8_LP
|
|
|
190885 |
+ test %R8_LP, %R8_LP
|
|
|
190885 |
+ jz L(ExitZero)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ mov %rsi, %rcx
|
|
|
190885 |
+# ifndef USE_AS_STPCPY
|
|
|
190885 |
+ mov %rdi, %rax /* save result */
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+ vpxorq %XMMZERO, %XMMZERO, %XMMZERO
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+ and $((VEC_SIZE * 4) - 1), %ecx
|
|
|
190885 |
+ cmp $(VEC_SIZE * 2), %ecx
|
|
|
190885 |
+ jbe L(SourceStringAlignmentLessTwoVecSize)
|
|
|
190885 |
+
|
|
|
190885 |
+ and $-VEC_SIZE, %rsi
|
|
|
190885 |
+ and $(VEC_SIZE - 1), %ecx
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, (%rsi), %YMMZERO, %k0
|
|
|
190885 |
+ kmovd %k0, %edx
|
|
|
190885 |
+ shr %cl, %rdx
|
|
|
190885 |
+
|
|
|
190885 |
+# ifdef USE_AS_STRNCPY
|
|
|
190885 |
+# if defined USE_AS_STPCPY || defined USE_AS_STRCAT
|
|
|
190885 |
+ mov $VEC_SIZE, %r10
|
|
|
190885 |
+ sub %rcx, %r10
|
|
|
190885 |
+ cmp %r10, %r8
|
|
|
190885 |
+# else
|
|
|
190885 |
+ mov $(VEC_SIZE + 1), %r10
|
|
|
190885 |
+ sub %rcx, %r10
|
|
|
190885 |
+ cmp %r10, %r8
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ jbe L(CopyVecSizeTailCase2OrCase3)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(CopyVecSizeTail)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, VEC_SIZE(%rsi), %YMMZERO, %k1
|
|
|
190885 |
+ kmovd %k1, %edx
|
|
|
190885 |
+
|
|
|
190885 |
+# ifdef USE_AS_STRNCPY
|
|
|
190885 |
+ add $VEC_SIZE, %r10
|
|
|
190885 |
+ cmp %r10, %r8
|
|
|
190885 |
+ jbe L(CopyTwoVecSizeCase2OrCase3)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(CopyTwoVecSize)
|
|
|
190885 |
+
|
|
|
190885 |
+ VMOVU (%rsi, %rcx), %YMM2 /* copy VEC_SIZE bytes */
|
|
|
190885 |
+ VMOVU %YMM2, (%rdi)
|
|
|
190885 |
+
|
|
|
190885 |
+/* If source address alignment != destination address alignment */
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(UnalignVecSizeBoth):
|
|
|
190885 |
+ sub %rcx, %rdi
|
|
|
190885 |
+# ifdef USE_AS_STRNCPY
|
|
|
190885 |
+ add %rcx, %r8
|
|
|
190885 |
+ sbb %rcx, %rcx
|
|
|
190885 |
+ or %rcx, %r8
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ mov $VEC_SIZE, %rcx
|
|
|
190885 |
+ VMOVA (%rsi, %rcx), %YMM2
|
|
|
190885 |
+ VMOVU %YMM2, (%rdi, %rcx)
|
|
|
190885 |
+ VMOVA VEC_SIZE(%rsi, %rcx), %YMM2
|
|
|
190885 |
+ vpcmpb $0, %YMM2, %YMMZERO, %k0
|
|
|
190885 |
+ kmovd %k0, %edx
|
|
|
190885 |
+ add $VEC_SIZE, %rcx
|
|
|
190885 |
+# ifdef USE_AS_STRNCPY
|
|
|
190885 |
+ sub $(VEC_SIZE * 3), %r8
|
|
|
190885 |
+ jbe L(CopyVecSizeCase2OrCase3)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+# if defined USE_AS_STRNCPY && !defined USE_AS_STRCAT
|
|
|
190885 |
+ jnz L(CopyVecSizeUnalignedVec2)
|
|
|
190885 |
+# else
|
|
|
190885 |
+ jnz L(CopyVecSize)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+ VMOVU %YMM2, (%rdi, %rcx)
|
|
|
190885 |
+ VMOVA VEC_SIZE(%rsi, %rcx), %YMM3
|
|
|
190885 |
+ vpcmpb $0, %YMM3, %YMMZERO, %k0
|
|
|
190885 |
+ kmovd %k0, %edx
|
|
|
190885 |
+ add $VEC_SIZE, %rcx
|
|
|
190885 |
+# ifdef USE_AS_STRNCPY
|
|
|
190885 |
+ sub $VEC_SIZE, %r8
|
|
|
190885 |
+ jbe L(CopyVecSizeCase2OrCase3)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+# if defined USE_AS_STRNCPY && !defined USE_AS_STRCAT
|
|
|
190885 |
+ jnz L(CopyVecSizeUnalignedVec3)
|
|
|
190885 |
+# else
|
|
|
190885 |
+ jnz L(CopyVecSize)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+ VMOVU %YMM3, (%rdi, %rcx)
|
|
|
190885 |
+ VMOVA VEC_SIZE(%rsi, %rcx), %YMM4
|
|
|
190885 |
+ vpcmpb $0, %YMM4, %YMMZERO, %k0
|
|
|
190885 |
+ kmovd %k0, %edx
|
|
|
190885 |
+ add $VEC_SIZE, %rcx
|
|
|
190885 |
+# ifdef USE_AS_STRNCPY
|
|
|
190885 |
+ sub $VEC_SIZE, %r8
|
|
|
190885 |
+ jbe L(CopyVecSizeCase2OrCase3)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+# if defined USE_AS_STRNCPY && !defined USE_AS_STRCAT
|
|
|
190885 |
+ jnz L(CopyVecSizeUnalignedVec4)
|
|
|
190885 |
+# else
|
|
|
190885 |
+ jnz L(CopyVecSize)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+ VMOVU %YMM4, (%rdi, %rcx)
|
|
|
190885 |
+ VMOVA VEC_SIZE(%rsi, %rcx), %YMM2
|
|
|
190885 |
+ vpcmpb $0, %YMM2, %YMMZERO, %k0
|
|
|
190885 |
+ kmovd %k0, %edx
|
|
|
190885 |
+ add $VEC_SIZE, %rcx
|
|
|
190885 |
+# ifdef USE_AS_STRNCPY
|
|
|
190885 |
+ sub $VEC_SIZE, %r8
|
|
|
190885 |
+ jbe L(CopyVecSizeCase2OrCase3)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+# if defined USE_AS_STRNCPY && !defined USE_AS_STRCAT
|
|
|
190885 |
+ jnz L(CopyVecSizeUnalignedVec2)
|
|
|
190885 |
+# else
|
|
|
190885 |
+ jnz L(CopyVecSize)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+ VMOVU %YMM2, (%rdi, %rcx)
|
|
|
190885 |
+ VMOVA VEC_SIZE(%rsi, %rcx), %YMM2
|
|
|
190885 |
+ vpcmpb $0, %YMM2, %YMMZERO, %k0
|
|
|
190885 |
+ kmovd %k0, %edx
|
|
|
190885 |
+ add $VEC_SIZE, %rcx
|
|
|
190885 |
+# ifdef USE_AS_STRNCPY
|
|
|
190885 |
+ sub $VEC_SIZE, %r8
|
|
|
190885 |
+ jbe L(CopyVecSizeCase2OrCase3)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+# if defined USE_AS_STRNCPY && !defined USE_AS_STRCAT
|
|
|
190885 |
+ jnz L(CopyVecSizeUnalignedVec2)
|
|
|
190885 |
+# else
|
|
|
190885 |
+ jnz L(CopyVecSize)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+ VMOVA VEC_SIZE(%rsi, %rcx), %YMM3
|
|
|
190885 |
+ VMOVU %YMM2, (%rdi, %rcx)
|
|
|
190885 |
+ vpcmpb $0, %YMM3, %YMMZERO, %k0
|
|
|
190885 |
+ kmovd %k0, %edx
|
|
|
190885 |
+ add $VEC_SIZE, %rcx
|
|
|
190885 |
+# ifdef USE_AS_STRNCPY
|
|
|
190885 |
+ sub $VEC_SIZE, %r8
|
|
|
190885 |
+ jbe L(CopyVecSizeCase2OrCase3)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+# if defined USE_AS_STRNCPY && !defined USE_AS_STRCAT
|
|
|
190885 |
+ jnz L(CopyVecSizeUnalignedVec3)
|
|
|
190885 |
+# else
|
|
|
190885 |
+ jnz L(CopyVecSize)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+ VMOVU %YMM3, (%rdi, %rcx)
|
|
|
190885 |
+ mov %rsi, %rdx
|
|
|
190885 |
+ lea VEC_SIZE(%rsi, %rcx), %rsi
|
|
|
190885 |
+ and $-(VEC_SIZE * 4), %rsi
|
|
|
190885 |
+ sub %rsi, %rdx
|
|
|
190885 |
+ sub %rdx, %rdi
|
|
|
190885 |
+# ifdef USE_AS_STRNCPY
|
|
|
190885 |
+ lea (VEC_SIZE * 8)(%r8, %rdx), %r8
|
|
|
190885 |
+# endif
|
|
|
190885 |
+L(UnalignedFourVecSizeLoop):
|
|
|
190885 |
+ VMOVA (%rsi), %YMM4
|
|
|
190885 |
+ VMOVA VEC_SIZE(%rsi), %YMM5
|
|
|
190885 |
+ VMOVA (VEC_SIZE * 2)(%rsi), %YMM6
|
|
|
190885 |
+ VMOVA (VEC_SIZE * 3)(%rsi), %YMM7
|
|
|
190885 |
+ vpminub %YMM5, %YMM4, %YMM2
|
|
|
190885 |
+ vpminub %YMM7, %YMM6, %YMM3
|
|
|
190885 |
+ vpminub %YMM2, %YMM3, %YMM2
|
|
|
190885 |
+ /* If K7 != 0, there is a null byte. */
|
|
|
190885 |
+ vpcmpb $0, %YMM2, %YMMZERO, %k7
|
|
|
190885 |
+ kmovd %k7, %edx
|
|
|
190885 |
+# ifdef USE_AS_STRNCPY
|
|
|
190885 |
+ sub $(VEC_SIZE * 4), %r8
|
|
|
190885 |
+ jbe L(UnalignedLeaveCase2OrCase3)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(UnalignedFourVecSizeLeave)
|
|
|
190885 |
+
|
|
|
190885 |
+L(UnalignedFourVecSizeLoop_start):
|
|
|
190885 |
+ add $(VEC_SIZE * 4), %rdi
|
|
|
190885 |
+ add $(VEC_SIZE * 4), %rsi
|
|
|
190885 |
+ VMOVU %YMM4, -(VEC_SIZE * 4)(%rdi)
|
|
|
190885 |
+ VMOVA (%rsi), %YMM4
|
|
|
190885 |
+ VMOVU %YMM5, -(VEC_SIZE * 3)(%rdi)
|
|
|
190885 |
+ VMOVA VEC_SIZE(%rsi), %YMM5
|
|
|
190885 |
+ vpminub %YMM5, %YMM4, %YMM2
|
|
|
190885 |
+ VMOVU %YMM6, -(VEC_SIZE * 2)(%rdi)
|
|
|
190885 |
+ VMOVA (VEC_SIZE * 2)(%rsi), %YMM6
|
|
|
190885 |
+ VMOVU %YMM7, -VEC_SIZE(%rdi)
|
|
|
190885 |
+ VMOVA (VEC_SIZE * 3)(%rsi), %YMM7
|
|
|
190885 |
+ vpminub %YMM7, %YMM6, %YMM3
|
|
|
190885 |
+ vpminub %YMM2, %YMM3, %YMM2
|
|
|
190885 |
+ /* If K7 != 0, there is a null byte. */
|
|
|
190885 |
+ vpcmpb $0, %YMM2, %YMMZERO, %k7
|
|
|
190885 |
+ kmovd %k7, %edx
|
|
|
190885 |
+# ifdef USE_AS_STRNCPY
|
|
|
190885 |
+ sub $(VEC_SIZE * 4), %r8
|
|
|
190885 |
+ jbe L(UnalignedLeaveCase2OrCase3)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jz L(UnalignedFourVecSizeLoop_start)
|
|
|
190885 |
+
|
|
|
190885 |
+L(UnalignedFourVecSizeLeave):
|
|
|
190885 |
+ vpcmpb $0, %YMM4, %YMMZERO, %k1
|
|
|
190885 |
+ kmovd %k1, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(CopyVecSizeUnaligned_0)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, %YMM5, %YMMZERO, %k2
|
|
|
190885 |
+ kmovd %k2, %ecx
|
|
|
190885 |
+ test %ecx, %ecx
|
|
|
190885 |
+ jnz L(CopyVecSizeUnaligned_16)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, %YMM6, %YMMZERO, %k3
|
|
|
190885 |
+ kmovd %k3, %edx
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(CopyVecSizeUnaligned_32)
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, %YMM7, %YMMZERO, %k4
|
|
|
190885 |
+ kmovd %k4, %ecx
|
|
|
190885 |
+ bsf %ecx, %edx
|
|
|
190885 |
+ VMOVU %YMM4, (%rdi)
|
|
|
190885 |
+ VMOVU %YMM5, VEC_SIZE(%rdi)
|
|
|
190885 |
+ VMOVU %YMM6, (VEC_SIZE * 2)(%rdi)
|
|
|
190885 |
+# if defined USE_AS_STRNCPY && !defined USE_AS_STRCAT
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea (VEC_SIZE * 3)(%rdi, %rdx), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ VMOVU %YMM7, (VEC_SIZE * 3)(%rdi)
|
|
|
190885 |
+ add $(VEC_SIZE - 1), %r8
|
|
|
190885 |
+ sub %rdx, %r8
|
|
|
190885 |
+ lea ((VEC_SIZE * 3) + 1)(%rdi, %rdx), %rdi
|
|
|
190885 |
+ jmp L(StrncpyFillTailWithZero)
|
|
|
190885 |
+# else
|
|
|
190885 |
+ add $(VEC_SIZE * 3), %rsi
|
|
|
190885 |
+ add $(VEC_SIZE * 3), %rdi
|
|
|
190885 |
+ jmp L(CopyVecSizeExit)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+/* If source address alignment == destination address alignment */
|
|
|
190885 |
+
|
|
|
190885 |
+L(SourceStringAlignmentLessTwoVecSize):
|
|
|
190885 |
+ VMOVU (%rsi), %YMM3
|
|
|
190885 |
+ VMOVU VEC_SIZE(%rsi), %YMM2
|
|
|
190885 |
+ vpcmpb $0, %YMM3, %YMMZERO, %k0
|
|
|
190885 |
+ kmovd %k0, %edx
|
|
|
190885 |
+
|
|
|
190885 |
+# ifdef USE_AS_STRNCPY
|
|
|
190885 |
+# if defined USE_AS_STPCPY || defined USE_AS_STRCAT
|
|
|
190885 |
+ cmp $VEC_SIZE, %r8
|
|
|
190885 |
+# else
|
|
|
190885 |
+ cmp $(VEC_SIZE + 1), %r8
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ jbe L(CopyVecSizeTail1Case2OrCase3)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(CopyVecSizeTail1)
|
|
|
190885 |
+
|
|
|
190885 |
+ VMOVU %YMM3, (%rdi)
|
|
|
190885 |
+ vpcmpb $0, %YMM2, %YMMZERO, %k0
|
|
|
190885 |
+ kmovd %k0, %edx
|
|
|
190885 |
+
|
|
|
190885 |
+# ifdef USE_AS_STRNCPY
|
|
|
190885 |
+# if defined USE_AS_STPCPY || defined USE_AS_STRCAT
|
|
|
190885 |
+ cmp $(VEC_SIZE * 2), %r8
|
|
|
190885 |
+# else
|
|
|
190885 |
+ cmp $((VEC_SIZE * 2) + 1), %r8
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ jbe L(CopyTwoVecSize1Case2OrCase3)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+ jnz L(CopyTwoVecSize1)
|
|
|
190885 |
+
|
|
|
190885 |
+ and $-VEC_SIZE, %rsi
|
|
|
190885 |
+ and $(VEC_SIZE - 1), %ecx
|
|
|
190885 |
+ jmp L(UnalignVecSizeBoth)
|
|
|
190885 |
+
|
|
|
190885 |
+/*------End of main part with loops---------------------*/
|
|
|
190885 |
+
|
|
|
190885 |
+/* Case1 */
|
|
|
190885 |
+
|
|
|
190885 |
+# if (!defined USE_AS_STRNCPY) || (defined USE_AS_STRCAT)
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(CopyVecSize):
|
|
|
190885 |
+ add %rcx, %rdi
|
|
|
190885 |
+# endif
|
|
|
190885 |
+L(CopyVecSizeTail):
|
|
|
190885 |
+ add %rcx, %rsi
|
|
|
190885 |
+L(CopyVecSizeTail1):
|
|
|
190885 |
+ bsf %edx, %edx
|
|
|
190885 |
+L(CopyVecSizeExit):
|
|
|
190885 |
+ cmp $32, %edx
|
|
|
190885 |
+ jae L(Exit32_63)
|
|
|
190885 |
+ cmp $16, %edx
|
|
|
190885 |
+ jae L(Exit16_31)
|
|
|
190885 |
+ cmp $8, %edx
|
|
|
190885 |
+ jae L(Exit8_15)
|
|
|
190885 |
+ cmp $4, %edx
|
|
|
190885 |
+ jae L(Exit4_7)
|
|
|
190885 |
+ cmp $3, %edx
|
|
|
190885 |
+ je L(Exit3)
|
|
|
190885 |
+ cmp $1, %edx
|
|
|
190885 |
+ ja L(Exit2)
|
|
|
190885 |
+ je L(Exit1)
|
|
|
190885 |
+ movb $0, (%rdi)
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea (%rdi), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+# if defined USE_AS_STRNCPY && !defined USE_AS_STRCAT
|
|
|
190885 |
+ sub $1, %r8
|
|
|
190885 |
+ lea 1(%rdi), %rdi
|
|
|
190885 |
+ jnz L(StrncpyFillTailWithZero)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(CopyTwoVecSize1):
|
|
|
190885 |
+ add $VEC_SIZE, %rsi
|
|
|
190885 |
+ add $VEC_SIZE, %rdi
|
|
|
190885 |
+# if defined USE_AS_STRNCPY && !defined USE_AS_STRCAT
|
|
|
190885 |
+ sub $VEC_SIZE, %r8
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ jmp L(CopyVecSizeTail1)
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(CopyTwoVecSize):
|
|
|
190885 |
+ bsf %edx, %edx
|
|
|
190885 |
+ add %rcx, %rsi
|
|
|
190885 |
+ add $VEC_SIZE, %edx
|
|
|
190885 |
+ sub %ecx, %edx
|
|
|
190885 |
+ jmp L(CopyVecSizeExit)
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(CopyVecSizeUnaligned_0):
|
|
|
190885 |
+ bsf %edx, %edx
|
|
|
190885 |
+# if defined USE_AS_STRNCPY && !defined USE_AS_STRCAT
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea (%rdi, %rdx), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ VMOVU %YMM4, (%rdi)
|
|
|
190885 |
+ add $((VEC_SIZE * 4) - 1), %r8
|
|
|
190885 |
+ sub %rdx, %r8
|
|
|
190885 |
+ lea 1(%rdi, %rdx), %rdi
|
|
|
190885 |
+ jmp L(StrncpyFillTailWithZero)
|
|
|
190885 |
+# else
|
|
|
190885 |
+ jmp L(CopyVecSizeExit)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(CopyVecSizeUnaligned_16):
|
|
|
190885 |
+ bsf %ecx, %edx
|
|
|
190885 |
+ VMOVU %YMM4, (%rdi)
|
|
|
190885 |
+# if defined USE_AS_STRNCPY && !defined USE_AS_STRCAT
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea VEC_SIZE(%rdi, %rdx), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ VMOVU %YMM5, VEC_SIZE(%rdi)
|
|
|
190885 |
+ add $((VEC_SIZE * 3) - 1), %r8
|
|
|
190885 |
+ sub %rdx, %r8
|
|
|
190885 |
+ lea (VEC_SIZE + 1)(%rdi, %rdx), %rdi
|
|
|
190885 |
+ jmp L(StrncpyFillTailWithZero)
|
|
|
190885 |
+# else
|
|
|
190885 |
+ add $VEC_SIZE, %rsi
|
|
|
190885 |
+ add $VEC_SIZE, %rdi
|
|
|
190885 |
+ jmp L(CopyVecSizeExit)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(CopyVecSizeUnaligned_32):
|
|
|
190885 |
+ bsf %edx, %edx
|
|
|
190885 |
+ VMOVU %YMM4, (%rdi)
|
|
|
190885 |
+ VMOVU %YMM5, VEC_SIZE(%rdi)
|
|
|
190885 |
+# if defined USE_AS_STRNCPY && !defined USE_AS_STRCAT
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea (VEC_SIZE * 2)(%rdi, %rdx), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ VMOVU %YMM6, (VEC_SIZE * 2)(%rdi)
|
|
|
190885 |
+ add $((VEC_SIZE * 2) - 1), %r8
|
|
|
190885 |
+ sub %rdx, %r8
|
|
|
190885 |
+ lea ((VEC_SIZE * 2) + 1)(%rdi, %rdx), %rdi
|
|
|
190885 |
+ jmp L(StrncpyFillTailWithZero)
|
|
|
190885 |
+# else
|
|
|
190885 |
+ add $(VEC_SIZE * 2), %rsi
|
|
|
190885 |
+ add $(VEC_SIZE * 2), %rdi
|
|
|
190885 |
+ jmp L(CopyVecSizeExit)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+# ifdef USE_AS_STRNCPY
|
|
|
190885 |
+# ifndef USE_AS_STRCAT
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(CopyVecSizeUnalignedVec6):
|
|
|
190885 |
+ VMOVU %YMM6, (%rdi, %rcx)
|
|
|
190885 |
+ jmp L(CopyVecSizeVecExit)
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(CopyVecSizeUnalignedVec5):
|
|
|
190885 |
+ VMOVU %YMM5, (%rdi, %rcx)
|
|
|
190885 |
+ jmp L(CopyVecSizeVecExit)
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(CopyVecSizeUnalignedVec4):
|
|
|
190885 |
+ VMOVU %YMM4, (%rdi, %rcx)
|
|
|
190885 |
+ jmp L(CopyVecSizeVecExit)
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(CopyVecSizeUnalignedVec3):
|
|
|
190885 |
+ VMOVU %YMM3, (%rdi, %rcx)
|
|
|
190885 |
+ jmp L(CopyVecSizeVecExit)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+/* Case2 */
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(CopyVecSizeCase2):
|
|
|
190885 |
+ add $VEC_SIZE, %r8
|
|
|
190885 |
+ add %rcx, %rdi
|
|
|
190885 |
+ add %rcx, %rsi
|
|
|
190885 |
+ bsf %edx, %edx
|
|
|
190885 |
+ cmp %r8d, %edx
|
|
|
190885 |
+ jb L(CopyVecSizeExit)
|
|
|
190885 |
+ jmp L(StrncpyExit)
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(CopyTwoVecSizeCase2):
|
|
|
190885 |
+ add %rcx, %rsi
|
|
|
190885 |
+ bsf %edx, %edx
|
|
|
190885 |
+ add $VEC_SIZE, %edx
|
|
|
190885 |
+ sub %ecx, %edx
|
|
|
190885 |
+ cmp %r8d, %edx
|
|
|
190885 |
+ jb L(CopyVecSizeExit)
|
|
|
190885 |
+ jmp L(StrncpyExit)
|
|
|
190885 |
+
|
|
|
190885 |
+L(CopyVecSizeTailCase2):
|
|
|
190885 |
+ add %rcx, %rsi
|
|
|
190885 |
+ bsf %edx, %edx
|
|
|
190885 |
+ cmp %r8d, %edx
|
|
|
190885 |
+ jb L(CopyVecSizeExit)
|
|
|
190885 |
+ jmp L(StrncpyExit)
|
|
|
190885 |
+
|
|
|
190885 |
+L(CopyVecSizeTail1Case2):
|
|
|
190885 |
+ bsf %edx, %edx
|
|
|
190885 |
+ cmp %r8d, %edx
|
|
|
190885 |
+ jb L(CopyVecSizeExit)
|
|
|
190885 |
+ jmp L(StrncpyExit)
|
|
|
190885 |
+
|
|
|
190885 |
+/* Case2 or Case3, Case3 */
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(CopyVecSizeCase2OrCase3):
|
|
|
190885 |
+ test %rdx, %rdx
|
|
|
190885 |
+ jnz L(CopyVecSizeCase2)
|
|
|
190885 |
+L(CopyVecSizeCase3):
|
|
|
190885 |
+ add $VEC_SIZE, %r8
|
|
|
190885 |
+ add %rcx, %rdi
|
|
|
190885 |
+ add %rcx, %rsi
|
|
|
190885 |
+ jmp L(StrncpyExit)
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(CopyTwoVecSizeCase2OrCase3):
|
|
|
190885 |
+ test %rdx, %rdx
|
|
|
190885 |
+ jnz L(CopyTwoVecSizeCase2)
|
|
|
190885 |
+ add %rcx, %rsi
|
|
|
190885 |
+ jmp L(StrncpyExit)
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(CopyVecSizeTailCase2OrCase3):
|
|
|
190885 |
+ test %rdx, %rdx
|
|
|
190885 |
+ jnz L(CopyVecSizeTailCase2)
|
|
|
190885 |
+ add %rcx, %rsi
|
|
|
190885 |
+ jmp L(StrncpyExit)
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(CopyTwoVecSize1Case2OrCase3):
|
|
|
190885 |
+ add $VEC_SIZE, %rdi
|
|
|
190885 |
+ add $VEC_SIZE, %rsi
|
|
|
190885 |
+ sub $VEC_SIZE, %r8
|
|
|
190885 |
+L(CopyVecSizeTail1Case2OrCase3):
|
|
|
190885 |
+ test %rdx, %rdx
|
|
|
190885 |
+ jnz L(CopyVecSizeTail1Case2)
|
|
|
190885 |
+ jmp L(StrncpyExit)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+/*------------End labels regarding with copying 1-VEC_SIZE bytes--and 1-(VEC_SIZE*2) bytes----*/
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(Exit1):
|
|
|
190885 |
+ movzwl (%rsi), %edx
|
|
|
190885 |
+ mov %dx, (%rdi)
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea 1(%rdi), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+# if defined USE_AS_STRNCPY && !defined USE_AS_STRCAT
|
|
|
190885 |
+ sub $2, %r8
|
|
|
190885 |
+ lea 2(%rdi), %rdi
|
|
|
190885 |
+ jnz L(StrncpyFillTailWithZero)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(Exit2):
|
|
|
190885 |
+ movzwl (%rsi), %ecx
|
|
|
190885 |
+ mov %cx, (%rdi)
|
|
|
190885 |
+ movb $0, 2(%rdi)
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea 2(%rdi), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+# if defined USE_AS_STRNCPY && !defined USE_AS_STRCAT
|
|
|
190885 |
+ sub $3, %r8
|
|
|
190885 |
+ lea 3(%rdi), %rdi
|
|
|
190885 |
+ jnz L(StrncpyFillTailWithZero)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(Exit3):
|
|
|
190885 |
+ mov (%rsi), %edx
|
|
|
190885 |
+ mov %edx, (%rdi)
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea 3(%rdi), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+# if defined USE_AS_STRNCPY && !defined USE_AS_STRCAT
|
|
|
190885 |
+ sub $4, %r8
|
|
|
190885 |
+ lea 4(%rdi), %rdi
|
|
|
190885 |
+ jnz L(StrncpyFillTailWithZero)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(Exit4_7):
|
|
|
190885 |
+ mov (%rsi), %ecx
|
|
|
190885 |
+ mov %ecx, (%rdi)
|
|
|
190885 |
+ mov -3(%rsi, %rdx), %ecx
|
|
|
190885 |
+ mov %ecx, -3(%rdi, %rdx)
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea (%rdi, %rdx), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+# if defined USE_AS_STRNCPY && !defined USE_AS_STRCAT
|
|
|
190885 |
+ sub %rdx, %r8
|
|
|
190885 |
+ sub $1, %r8
|
|
|
190885 |
+ lea 1(%rdi, %rdx), %rdi
|
|
|
190885 |
+ jnz L(StrncpyFillTailWithZero)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(Exit8_15):
|
|
|
190885 |
+ mov (%rsi), %rcx
|
|
|
190885 |
+ mov -7(%rsi, %rdx), %r9
|
|
|
190885 |
+ mov %rcx, (%rdi)
|
|
|
190885 |
+ mov %r9, -7(%rdi, %rdx)
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea (%rdi, %rdx), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+# if defined USE_AS_STRNCPY && !defined USE_AS_STRCAT
|
|
|
190885 |
+ sub %rdx, %r8
|
|
|
190885 |
+ sub $1, %r8
|
|
|
190885 |
+ lea 1(%rdi, %rdx), %rdi
|
|
|
190885 |
+ jnz L(StrncpyFillTailWithZero)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(Exit16_31):
|
|
|
190885 |
+ VMOVU (%rsi), %XMM2
|
|
|
190885 |
+ VMOVU -15(%rsi, %rdx), %XMM3
|
|
|
190885 |
+ VMOVU %XMM2, (%rdi)
|
|
|
190885 |
+ VMOVU %XMM3, -15(%rdi, %rdx)
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea (%rdi, %rdx), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+# if defined USE_AS_STRNCPY && !defined USE_AS_STRCAT
|
|
|
190885 |
+ sub %rdx, %r8
|
|
|
190885 |
+ sub $1, %r8
|
|
|
190885 |
+ lea 1(%rdi, %rdx), %rdi
|
|
|
190885 |
+ jnz L(StrncpyFillTailWithZero)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(Exit32_63):
|
|
|
190885 |
+ VMOVU (%rsi), %YMM2
|
|
|
190885 |
+ VMOVU -31(%rsi, %rdx), %YMM3
|
|
|
190885 |
+ VMOVU %YMM2, (%rdi)
|
|
|
190885 |
+ VMOVU %YMM3, -31(%rdi, %rdx)
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea (%rdi, %rdx), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+# if defined USE_AS_STRNCPY && !defined USE_AS_STRCAT
|
|
|
190885 |
+ sub %rdx, %r8
|
|
|
190885 |
+ sub $1, %r8
|
|
|
190885 |
+ lea 1(%rdi, %rdx), %rdi
|
|
|
190885 |
+ jnz L(StrncpyFillTailWithZero)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+# ifdef USE_AS_STRNCPY
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(StrncpyExit1):
|
|
|
190885 |
+ movzbl (%rsi), %edx
|
|
|
190885 |
+ mov %dl, (%rdi)
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea 1(%rdi), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+# ifdef USE_AS_STRCAT
|
|
|
190885 |
+ movb $0, 1(%rdi)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(StrncpyExit2):
|
|
|
190885 |
+ movzwl (%rsi), %edx
|
|
|
190885 |
+ mov %dx, (%rdi)
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea 2(%rdi), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+# ifdef USE_AS_STRCAT
|
|
|
190885 |
+ movb $0, 2(%rdi)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(StrncpyExit3_4):
|
|
|
190885 |
+ movzwl (%rsi), %ecx
|
|
|
190885 |
+ movzwl -2(%rsi, %r8), %edx
|
|
|
190885 |
+ mov %cx, (%rdi)
|
|
|
190885 |
+ mov %dx, -2(%rdi, %r8)
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea (%rdi, %r8), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+# ifdef USE_AS_STRCAT
|
|
|
190885 |
+ movb $0, (%rdi, %r8)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(StrncpyExit5_8):
|
|
|
190885 |
+ mov (%rsi), %ecx
|
|
|
190885 |
+ mov -4(%rsi, %r8), %edx
|
|
|
190885 |
+ mov %ecx, (%rdi)
|
|
|
190885 |
+ mov %edx, -4(%rdi, %r8)
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea (%rdi, %r8), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+# ifdef USE_AS_STRCAT
|
|
|
190885 |
+ movb $0, (%rdi, %r8)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(StrncpyExit9_16):
|
|
|
190885 |
+ mov (%rsi), %rcx
|
|
|
190885 |
+ mov -8(%rsi, %r8), %rdx
|
|
|
190885 |
+ mov %rcx, (%rdi)
|
|
|
190885 |
+ mov %rdx, -8(%rdi, %r8)
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea (%rdi, %r8), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+# ifdef USE_AS_STRCAT
|
|
|
190885 |
+ movb $0, (%rdi, %r8)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(StrncpyExit17_32):
|
|
|
190885 |
+ VMOVU (%rsi), %XMM2
|
|
|
190885 |
+ VMOVU -16(%rsi, %r8), %XMM3
|
|
|
190885 |
+ VMOVU %XMM2, (%rdi)
|
|
|
190885 |
+ VMOVU %XMM3, -16(%rdi, %r8)
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea (%rdi, %r8), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+# ifdef USE_AS_STRCAT
|
|
|
190885 |
+ movb $0, (%rdi, %r8)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(StrncpyExit33_64):
|
|
|
190885 |
+ /* 0/32, 31/16 */
|
|
|
190885 |
+ VMOVU (%rsi), %YMM2
|
|
|
190885 |
+ VMOVU -VEC_SIZE(%rsi, %r8), %YMM3
|
|
|
190885 |
+ VMOVU %YMM2, (%rdi)
|
|
|
190885 |
+ VMOVU %YMM3, -VEC_SIZE(%rdi, %r8)
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea (%rdi, %r8), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+# ifdef USE_AS_STRCAT
|
|
|
190885 |
+ movb $0, (%rdi, %r8)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(StrncpyExit65):
|
|
|
190885 |
+ /* 0/32, 32/32, 64/1 */
|
|
|
190885 |
+ VMOVU (%rsi), %YMM2
|
|
|
190885 |
+ VMOVU 32(%rsi), %YMM3
|
|
|
190885 |
+ mov 64(%rsi), %cl
|
|
|
190885 |
+ VMOVU %YMM2, (%rdi)
|
|
|
190885 |
+ VMOVU %YMM3, 32(%rdi)
|
|
|
190885 |
+ mov %cl, 64(%rdi)
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea 65(%rdi), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+# ifdef USE_AS_STRCAT
|
|
|
190885 |
+ movb $0, 65(%rdi)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+# ifndef USE_AS_STRCAT
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(Fill1):
|
|
|
190885 |
+ mov %dl, (%rdi)
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(Fill2):
|
|
|
190885 |
+ mov %dx, (%rdi)
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(Fill3_4):
|
|
|
190885 |
+ mov %dx, (%rdi)
|
|
|
190885 |
+ mov %dx, -2(%rdi, %r8)
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(Fill5_8):
|
|
|
190885 |
+ mov %edx, (%rdi)
|
|
|
190885 |
+ mov %edx, -4(%rdi, %r8)
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(Fill9_16):
|
|
|
190885 |
+ mov %rdx, (%rdi)
|
|
|
190885 |
+ mov %rdx, -8(%rdi, %r8)
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(Fill17_32):
|
|
|
190885 |
+ VMOVU %XMMZERO, (%rdi)
|
|
|
190885 |
+ VMOVU %XMMZERO, -16(%rdi, %r8)
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(CopyVecSizeUnalignedVec2):
|
|
|
190885 |
+ VMOVU %YMM2, (%rdi, %rcx)
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(CopyVecSizeVecExit):
|
|
|
190885 |
+ bsf %edx, %edx
|
|
|
190885 |
+ add $(VEC_SIZE - 1), %r8
|
|
|
190885 |
+ add %rcx, %rdi
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea (%rdi, %rdx), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ sub %rdx, %r8
|
|
|
190885 |
+ lea 1(%rdi, %rdx), %rdi
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(StrncpyFillTailWithZero):
|
|
|
190885 |
+ xor %edx, %edx
|
|
|
190885 |
+ sub $VEC_SIZE, %r8
|
|
|
190885 |
+ jbe L(StrncpyFillExit)
|
|
|
190885 |
+
|
|
|
190885 |
+ VMOVU %YMMZERO, (%rdi)
|
|
|
190885 |
+ add $VEC_SIZE, %rdi
|
|
|
190885 |
+
|
|
|
190885 |
+ mov %rdi, %rsi
|
|
|
190885 |
+ and $(VEC_SIZE - 1), %esi
|
|
|
190885 |
+ sub %rsi, %rdi
|
|
|
190885 |
+ add %rsi, %r8
|
|
|
190885 |
+ sub $(VEC_SIZE * 4), %r8
|
|
|
190885 |
+ jb L(StrncpyFillLessFourVecSize)
|
|
|
190885 |
+
|
|
|
190885 |
+L(StrncpyFillLoopVmovdqa):
|
|
|
190885 |
+ VMOVA %YMMZERO, (%rdi)
|
|
|
190885 |
+ VMOVA %YMMZERO, VEC_SIZE(%rdi)
|
|
|
190885 |
+ VMOVA %YMMZERO, (VEC_SIZE * 2)(%rdi)
|
|
|
190885 |
+ VMOVA %YMMZERO, (VEC_SIZE * 3)(%rdi)
|
|
|
190885 |
+ add $(VEC_SIZE * 4), %rdi
|
|
|
190885 |
+ sub $(VEC_SIZE * 4), %r8
|
|
|
190885 |
+ jae L(StrncpyFillLoopVmovdqa)
|
|
|
190885 |
+
|
|
|
190885 |
+L(StrncpyFillLessFourVecSize):
|
|
|
190885 |
+ add $(VEC_SIZE * 2), %r8
|
|
|
190885 |
+ jl L(StrncpyFillLessTwoVecSize)
|
|
|
190885 |
+ VMOVA %YMMZERO, (%rdi)
|
|
|
190885 |
+ VMOVA %YMMZERO, VEC_SIZE(%rdi)
|
|
|
190885 |
+ add $(VEC_SIZE * 2), %rdi
|
|
|
190885 |
+ sub $VEC_SIZE, %r8
|
|
|
190885 |
+ jl L(StrncpyFillExit)
|
|
|
190885 |
+ VMOVA %YMMZERO, (%rdi)
|
|
|
190885 |
+ add $VEC_SIZE, %rdi
|
|
|
190885 |
+ jmp L(Fill)
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(StrncpyFillLessTwoVecSize):
|
|
|
190885 |
+ add $VEC_SIZE, %r8
|
|
|
190885 |
+ jl L(StrncpyFillExit)
|
|
|
190885 |
+ VMOVA %YMMZERO, (%rdi)
|
|
|
190885 |
+ add $VEC_SIZE, %rdi
|
|
|
190885 |
+ jmp L(Fill)
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(StrncpyFillExit):
|
|
|
190885 |
+ add $VEC_SIZE, %r8
|
|
|
190885 |
+L(Fill):
|
|
|
190885 |
+ cmp $17, %r8d
|
|
|
190885 |
+ jae L(Fill17_32)
|
|
|
190885 |
+ cmp $9, %r8d
|
|
|
190885 |
+ jae L(Fill9_16)
|
|
|
190885 |
+ cmp $5, %r8d
|
|
|
190885 |
+ jae L(Fill5_8)
|
|
|
190885 |
+ cmp $3, %r8d
|
|
|
190885 |
+ jae L(Fill3_4)
|
|
|
190885 |
+ cmp $1, %r8d
|
|
|
190885 |
+ ja L(Fill2)
|
|
|
190885 |
+ je L(Fill1)
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+/* end of ifndef USE_AS_STRCAT */
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(UnalignedLeaveCase2OrCase3):
|
|
|
190885 |
+ test %rdx, %rdx
|
|
|
190885 |
+ jnz L(UnalignedFourVecSizeLeaveCase2)
|
|
|
190885 |
+L(UnalignedFourVecSizeLeaveCase3):
|
|
|
190885 |
+ lea (VEC_SIZE * 4)(%r8), %rcx
|
|
|
190885 |
+ and $-VEC_SIZE, %rcx
|
|
|
190885 |
+ add $(VEC_SIZE * 3), %r8
|
|
|
190885 |
+ jl L(CopyVecSizeCase3)
|
|
|
190885 |
+ VMOVU %YMM4, (%rdi)
|
|
|
190885 |
+ sub $VEC_SIZE, %r8
|
|
|
190885 |
+ jb L(CopyVecSizeCase3)
|
|
|
190885 |
+ VMOVU %YMM5, VEC_SIZE(%rdi)
|
|
|
190885 |
+ sub $VEC_SIZE, %r8
|
|
|
190885 |
+ jb L(CopyVecSizeCase3)
|
|
|
190885 |
+ VMOVU %YMM6, (VEC_SIZE * 2)(%rdi)
|
|
|
190885 |
+ sub $VEC_SIZE, %r8
|
|
|
190885 |
+ jb L(CopyVecSizeCase3)
|
|
|
190885 |
+ VMOVU %YMM7, (VEC_SIZE * 3)(%rdi)
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ lea (VEC_SIZE * 4)(%rdi), %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+# ifdef USE_AS_STRCAT
|
|
|
190885 |
+ movb $0, (VEC_SIZE * 4)(%rdi)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(UnalignedFourVecSizeLeaveCase2):
|
|
|
190885 |
+ xor %ecx, %ecx
|
|
|
190885 |
+ vpcmpb $0, %YMM4, %YMMZERO, %k1
|
|
|
190885 |
+ kmovd %k1, %edx
|
|
|
190885 |
+ add $(VEC_SIZE * 3), %r8
|
|
|
190885 |
+ jle L(CopyVecSizeCase2OrCase3)
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+# ifndef USE_AS_STRCAT
|
|
|
190885 |
+ jnz L(CopyVecSizeUnalignedVec4)
|
|
|
190885 |
+# else
|
|
|
190885 |
+ jnz L(CopyVecSize)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ vpcmpb $0, %YMM5, %YMMZERO, %k2
|
|
|
190885 |
+ kmovd %k2, %edx
|
|
|
190885 |
+ VMOVU %YMM4, (%rdi)
|
|
|
190885 |
+ add $VEC_SIZE, %rcx
|
|
|
190885 |
+ sub $VEC_SIZE, %r8
|
|
|
190885 |
+ jbe L(CopyVecSizeCase2OrCase3)
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+# ifndef USE_AS_STRCAT
|
|
|
190885 |
+ jnz L(CopyVecSizeUnalignedVec5)
|
|
|
190885 |
+# else
|
|
|
190885 |
+ jnz L(CopyVecSize)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, %YMM6, %YMMZERO, %k3
|
|
|
190885 |
+ kmovd %k3, %edx
|
|
|
190885 |
+ VMOVU %YMM5, VEC_SIZE(%rdi)
|
|
|
190885 |
+ add $VEC_SIZE, %rcx
|
|
|
190885 |
+ sub $VEC_SIZE, %r8
|
|
|
190885 |
+ jbe L(CopyVecSizeCase2OrCase3)
|
|
|
190885 |
+ test %edx, %edx
|
|
|
190885 |
+# ifndef USE_AS_STRCAT
|
|
|
190885 |
+ jnz L(CopyVecSizeUnalignedVec6)
|
|
|
190885 |
+# else
|
|
|
190885 |
+ jnz L(CopyVecSize)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+ vpcmpb $0, %YMM7, %YMMZERO, %k4
|
|
|
190885 |
+ kmovd %k4, %edx
|
|
|
190885 |
+ VMOVU %YMM6, (VEC_SIZE * 2)(%rdi)
|
|
|
190885 |
+ lea VEC_SIZE(%rdi, %rcx), %rdi
|
|
|
190885 |
+ lea VEC_SIZE(%rsi, %rcx), %rsi
|
|
|
190885 |
+ bsf %edx, %edx
|
|
|
190885 |
+ cmp %r8d, %edx
|
|
|
190885 |
+ jb L(CopyVecSizeExit)
|
|
|
190885 |
+L(StrncpyExit):
|
|
|
190885 |
+ cmp $65, %r8d
|
|
|
190885 |
+ je L(StrncpyExit65)
|
|
|
190885 |
+ cmp $33, %r8d
|
|
|
190885 |
+ jae L(StrncpyExit33_64)
|
|
|
190885 |
+ cmp $17, %r8d
|
|
|
190885 |
+ jae L(StrncpyExit17_32)
|
|
|
190885 |
+ cmp $9, %r8d
|
|
|
190885 |
+ jae L(StrncpyExit9_16)
|
|
|
190885 |
+ cmp $5, %r8d
|
|
|
190885 |
+ jae L(StrncpyExit5_8)
|
|
|
190885 |
+ cmp $3, %r8d
|
|
|
190885 |
+ jae L(StrncpyExit3_4)
|
|
|
190885 |
+ cmp $1, %r8d
|
|
|
190885 |
+ ja L(StrncpyExit2)
|
|
|
190885 |
+ je L(StrncpyExit1)
|
|
|
190885 |
+# ifdef USE_AS_STPCPY
|
|
|
190885 |
+ mov %rdi, %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+# ifdef USE_AS_STRCAT
|
|
|
190885 |
+ movb $0, (%rdi)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+ .p2align 4
|
|
|
190885 |
+L(ExitZero):
|
|
|
190885 |
+# ifndef USE_AS_STRCAT
|
|
|
190885 |
+ mov %rdi, %rax
|
|
|
190885 |
+# endif
|
|
|
190885 |
+ ret
|
|
|
190885 |
+
|
|
|
190885 |
+# endif
|
|
|
190885 |
+
|
|
|
190885 |
+# ifndef USE_AS_STRCAT
|
|
|
190885 |
+END (STRCPY)
|
|
|
190885 |
+# else
|
|
|
190885 |
+END (STRCAT)
|
|
|
190885 |
+# endif
|
|
|
190885 |
+#endif
|
|
|
190885 |
diff --git a/sysdeps/x86_64/multiarch/strncat-evex.S b/sysdeps/x86_64/multiarch/strncat-evex.S
|
|
|
190885 |
new file mode 100644
|
|
|
190885 |
index 00000000..8884f023
|
|
|
190885 |
--- /dev/null
|
|
|
190885 |
+++ b/sysdeps/x86_64/multiarch/strncat-evex.S
|
|
|
190885 |
@@ -0,0 +1,3 @@
|
|
|
190885 |
+#define USE_AS_STRNCAT
|
|
|
190885 |
+#define STRCAT __strncat_evex
|
|
|
190885 |
+#include "strcat-evex.S"
|
|
|
190885 |
diff --git a/sysdeps/x86_64/multiarch/strncpy-evex.S b/sysdeps/x86_64/multiarch/strncpy-evex.S
|
|
|
190885 |
new file mode 100644
|
|
|
190885 |
index 00000000..40e391f0
|
|
|
190885 |
--- /dev/null
|
|
|
190885 |
+++ b/sysdeps/x86_64/multiarch/strncpy-evex.S
|
|
|
190885 |
@@ -0,0 +1,3 @@
|
|
|
190885 |
+#define USE_AS_STRNCPY
|
|
|
190885 |
+#define STRCPY __strncpy_evex
|
|
|
190885 |
+#include "strcpy-evex.S"
|
|
|
190885 |
--
|
|
|
190885 |
GitLab
|
|
|
190885 |
|