AuroraMiddleware/glibc
1
0
Fork 0
mirror of https://sourceware.org/git/glibc.git synced 2024-11-22 04:50:07 +00:00
Code Issues Packages Projects Releases Wiki Activity
master
glibc/sysdeps/x86_64/multiarch/strcpy-evex.S

543 lines
12 KiB
ArmAsm
Raw Permalink Normal View History

x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
/* {wcs|wcp|str|stp}cpy with 256/512-bit EVEX instructions.
Update copyright dates with scripts/update-copyrights
2024-01-01 18:12:26 +00:00
Copyright (C) 2021-2024 Free Software Foundation, Inc.
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
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/>. */
x86: Add support to build st{p|r}{n}{cpy|cat} with explicit ISA level 1. Add default ISA level selection in non-multiarch/rtld implementations. 2. Add ISA level build guards to different implementations. - I.e strcpy-avx2.S which is ISA level 3 will only build if compiled ISA level <= 3. Otherwise there is no reason to include it as we will always use one of the ISA level 4 implementations (strcpy-evex.S). 3. Refactor the ifunc selector and ifunc implementation list to use the ISA level aware wrapper macros that allow functions below the compiled ISA level (with a guranteed replacement) to be skipped. Tested with and without multiarch on x86_64 for ISA levels: {generic, x86-64-v2, x86-64-v3, x86-64-v4} And m32 with and without multiarch.
2022-07-13 23:33:01 +00:00
#include <isa-level.h>
#if ISA_SHOULD_BUILD (4)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
/* Use evex-masked stores for small sizes. Turned off at the
moment. */
# define USE_EVEX_MASKED_STORE 0
/* Use movsb in page cross case to save code size. */
# define USE_MOVSB_IN_PAGE_CROSS 1
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# include <sysdep.h>
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# ifndef VEC_SIZE
# include "x86-evex256-vecs.h"
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# ifndef STRCPY
# define STRCPY __strcpy_evex
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# ifdef USE_AS_WCSCPY
# define VMOVU_MASK vmovdqu32
# define VPMIN vpminud
# define VPTESTN vptestnmd
# define VPTEST vptestmd
# define VPCMPEQ vpcmpeqd
# define CHAR_SIZE 4
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86/string: Use `movsl` instead of `movsd` [BZ #32344] `ld`, starting at 2.40, emits a warning when using `movsd`. There is no change to the actual code produced. Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
2024-11-08 19:18:17 +00:00
# define REP_MOVS rep movsl
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# define USE_WIDE_CHAR
# else
# define VMOVU_MASK vmovdqu8
# define VPMIN vpminub
# define VPTESTN vptestnmb
# define VPTEST vptestmb
# define VPCMPEQ vpcmpeqb
# define CHAR_SIZE 1
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# define REP_MOVS rep movsb
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# include "reg-macros.h"
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# ifdef USE_AS_STPCPY
# define END_REG rax
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# else
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# define END_REG rdi, %rdx, CHAR_SIZE
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# ifdef USE_AS_STRCAT
# define PAGE_ALIGN_REG edx
# define PAGE_ALIGN_REG_64 rdx
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# else
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# define PAGE_ALIGN_REG eax
# define PAGE_ALIGN_REG_64 rax
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# define VZERO VMM(7)
# define VZERO_128 VMM_128(7)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# define PAGE_SIZE 4096
# define CHAR_PER_VEC (VEC_SIZE / CHAR_SIZE)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
.section SECTION(.text), "ax", @progbits
ENTRY(STRCPY)
# ifdef USE_AS_STRCAT
movq %rdi, %rax
# include "strcat-strlen-evex.h.S"
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
movl %esi, %PAGE_ALIGN_REG
andl $(PAGE_SIZE - 1), %PAGE_ALIGN_REG
cmpl $(PAGE_SIZE - VEC_SIZE), %PAGE_ALIGN_REG
ja L(page_cross)
L(page_cross_continue):
VMOVU (%rsi), %VMM(0)
# if !defined USE_AS_STPCPY && !defined USE_AS_STRCAT
movq %rdi, %rax
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
/* Two short string implementations. One with traditional
branching approach and one with masked instructions (which
have potential for dramatically bad perf if dst splits a
page and is not in the TLB). */
# if USE_EVEX_MASKED_STORE
VPTEST %VMM(0), %VMM(0), %k0
KMOV %k0, %VRCX
# ifdef USE_AS_WCSCPY
subl $((1 << CHAR_PER_VEC)- 1), %VRCX
# else
inc %VRCX
# endif
jz L(more_1x_vec)
KMOV %VRCX, %k1
KXOR %k0, %k1, %k1
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
VMOVU_MASK %VMM(0), (%rdi){%k1}
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# ifdef USE_AS_STPCPY
bsf %VRCX, %VRCX
leaq (%rdi, %rcx, CHAR_SIZE), %rax
# endif
ret
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# else
VPTESTN %VMM(0), %VMM(0), %k0
KMOV %k0, %VRCX
test %VRCX, %VRCX
jz L(more_1x_vec)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
xorl %edx, %edx
bsf %VRCX, %VRDX
# ifdef USE_AS_STPCPY
leaq (%rdi, %rdx, CHAR_SIZE), %rax
# endif
/* Use mask bits in rcx to detect which copy we need. If the low
mask is zero then there must be a bit set in the upper half.
I.e if rcx != 0 and ecx == 0, then match must be upper 32
bits so we use L(copy_32_63). */
# if VEC_SIZE == 64
# ifdef USE_AS_WCSCPY
testb %cl, %cl
# else
testl %ecx, %ecx
# endif
jz L(copy_32_63)
# endif
# ifdef USE_AS_WCSCPY
testb $0xf, %cl
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# else
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
testw %cx, %cx
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
jz L(copy_16_31)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# ifdef USE_AS_WCSCPY
testb $0x3, %cl
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# else
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
testb %cl, %cl
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
jz L(copy_8_15)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# ifdef USE_AS_WCSCPY
vmovd %VMM_128(0), (%rdi)
/* No need to copy, we know its zero. */
movl $0, (%END_REG)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
ret
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# else
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
testb $0x7, %cl
jz L(copy_4_7)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
test %edx, %edx
jz L(set_null_term)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
/* NB: make this `vmovw` if support for AVX512-FP16 is added.
*/
vmovd %VMM_128(0), %esi
movw %si, (%rdi)
.p2align 4,, 1
L(set_null_term):
/* No need to copy, we know its zero. */
movb $0, (%END_REG)
ret
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# if VEC_SIZE == 64
.p2align 4,, 6
L(copy_32_63):
VMOVU -(32 - CHAR_SIZE)(%rsi, %rdx, CHAR_SIZE), %VMM_256(1)
VMOVU %VMM_256(0), (%rdi)
VMOVU %VMM_256(1), -(32 - CHAR_SIZE)(%END_REG)
ret
# endif
.p2align 4,, 6
L(copy_16_31):
/* Use xmm1 explicitly here as it won't require a `vzeroupper`
and will save code size. */
vmovdqu -(16 - CHAR_SIZE)(%rsi, %rdx, CHAR_SIZE), %xmm1
VMOVU %VMM_128(0), (%rdi)
vmovdqu %xmm1, -(16 - CHAR_SIZE)(%END_REG)
ret
.p2align 4,, 8
L(copy_8_15):
# ifdef USE_AS_WCSCPY
movl -(8 - CHAR_SIZE)(%rsi, %rdx, CHAR_SIZE), %ecx
# else
movq -(8 - CHAR_SIZE)(%rsi, %rdx, CHAR_SIZE), %rcx
# endif
vmovq %VMM_128(0), (%rdi)
movq %rcx, -(8 - CHAR_SIZE)(%END_REG)
ret
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# ifndef USE_AS_WCSCPY
.p2align 4,, 12
L(copy_4_7):
movl -(4 - CHAR_SIZE)(%rsi, %rdx, CHAR_SIZE), %ecx
vmovd %VMM_128(0), (%rdi)
movl %ecx, -(4 - CHAR_SIZE)(%END_REG)
ret
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
.p2align 4,, 8
L(more_1x_vec):
# if defined USE_AS_STPCPY || defined USE_AS_STRCAT
VMOVU %VMM(0), (%rdi)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
subq %rsi, %rdi
andq $-(VEC_SIZE), %rsi
addq %rsi, %rdi
VMOVA (VEC_SIZE * 1)(%rsi), %VMM(1)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
/* Ideally we store after moves to minimize impact of potential
false-dependencies. */
# if !defined USE_AS_STPCPY && !defined USE_AS_STRCAT
VMOVU %VMM(0), (%rax)
# endif
VPTESTN %VMM(1), %VMM(1), %k0
KMOV %k0, %VRCX
test %VRCX, %VRCX
jnz L(ret_vec_x1)
VMOVA (VEC_SIZE * 2)(%rsi), %VMM(2)
VMOVU %VMM(1), VEC_SIZE(%rdi)
VPTESTN %VMM(2), %VMM(2), %k0
KMOV %k0, %VRCX
test %VRCX, %VRCX
jnz L(ret_vec_x2)
VMOVA (VEC_SIZE * 3)(%rsi), %VMM(3)
VMOVU %VMM(2), (VEC_SIZE * 2)(%rdi)
VPTESTN %VMM(3), %VMM(3), %k0
KMOV %k0, %VRDX
test %VRDX, %VRDX
jnz L(ret_vec_x3)
VMOVA (VEC_SIZE * 4)(%rsi), %VMM(4)
VMOVU %VMM(3), (VEC_SIZE * 3)(%rdi)
VPTESTN %VMM(4), %VMM(4), %k0
KMOV %k0, %VRCX
test %VRCX, %VRCX
jnz L(ret_vec_x4)
VMOVU %VMM(4), (VEC_SIZE * 4)(%rdi)
/* Align for 4x loop. */
subq %rsi, %rdi
/* + VEC_SIZE * 5 because we never added the original VEC_SIZE
we covered before aligning. */
subq $-(VEC_SIZE * 5), %rsi
andq $-(VEC_SIZE * 4), %rsi
/* Load first half of the loop before entry. */
VMOVA (VEC_SIZE * 0 + 0)(%rsi), %VMM(0)
VMOVA (VEC_SIZE * 1 + 0)(%rsi), %VMM(1)
VMOVA (VEC_SIZE * 2 + 0)(%rsi), %VMM(2)
VMOVA (VEC_SIZE * 3 + 0)(%rsi), %VMM(3)
VPMIN %VMM(0), %VMM(1), %VMM(4)
VPMIN %VMM(2), %VMM(3), %VMM(6)
VPTESTN %VMM(4), %VMM(4), %k2
VPTESTN %VMM(6), %VMM(6), %k4
KORTEST %k2, %k4
jnz L(loop_4x_done)
.p2align 4,, 11
L(loop_4x_vec):
VMOVU %VMM(0), (VEC_SIZE * 0 + 0)(%rdi, %rsi)
VMOVU %VMM(1), (VEC_SIZE * 1 + 0)(%rdi, %rsi)
VMOVU %VMM(2), (VEC_SIZE * 2 + 0)(%rdi, %rsi)
VMOVU %VMM(3), (VEC_SIZE * 3 + 0)(%rdi, %rsi)
subq $(VEC_SIZE * -4), %rsi
VMOVA (VEC_SIZE * 0 + 0)(%rsi), %VMM(0)
VMOVA (VEC_SIZE * 1 + 0)(%rsi), %VMM(1)
VMOVA (VEC_SIZE * 2 + 0)(%rsi), %VMM(2)
VMOVA (VEC_SIZE * 3 + 0)(%rsi), %VMM(3)
VPMIN %VMM(0), %VMM(1), %VMM(4)
VPMIN %VMM(2), %VMM(3), %VMM(6)
VPTESTN %VMM(4), %VMM(4), %k2
VPTESTN %VMM(6), %VMM(6), %k4
KORTEST %k2, %k4
jz L(loop_4x_vec)
L(loop_4x_done):
VPTESTN %VMM(0), %VMM(0), %k0
KMOV %k0, %VRCX
/* Restore rdi (%rdi). */
addq %rsi, %rdi
test %VRCX, %VRCX
jnz L(ret_vec_x0_end)
VMOVU %VMM(0), (VEC_SIZE * 0 + 0)(%rdi)
KMOV %k2, %VRCX
test %VRCX, %VRCX
jnz L(ret_vec_x1)
VMOVU %VMM(1), (VEC_SIZE * 1 + 0)(%rdi)
VPTESTN %VMM(2), %VMM(2), %k0
KMOV %k0, %VRCX
test %VRCX, %VRCX
jnz L(ret_vec_x2)
VMOVU %VMM(2), (VEC_SIZE * 2 + 0)(%rdi)
/* Place L(ret_vec_x4) here to save code size. We get a
meaningfuly benefit doing this for stpcpy. */
KMOV %k4, %VRDX
L(ret_vec_x3):
bsf %VRDX, %VRDX
VMOVU ((VEC_SIZE * 3)-(VEC_SIZE - CHAR_SIZE))(%rsi, %rdx, CHAR_SIZE), %VMM(0)
VMOVU %VMM(0), ((VEC_SIZE * 3 + 0)-(VEC_SIZE - CHAR_SIZE))(%rdi, %rdx, CHAR_SIZE)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# ifdef USE_AS_STPCPY
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
leaq (VEC_SIZE * 3 + 0)(%rdi, %rdx, CHAR_SIZE), %rax
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
L(return_end):
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
ret
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
.p2align 4,, 6
L(ret_vec_x0_end):
bsf %VRCX, %VRCX
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# ifdef USE_AS_STPCPY
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
leaq (%rdi, %rcx, CHAR_SIZE), %rax
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
inc %VRCX
VMOVU (-(VEC_SIZE))(%rsi, %rcx, CHAR_SIZE), %VMM(0)
VMOVU %VMM(0), (-(VEC_SIZE))(%rdi, %rcx, CHAR_SIZE)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
ret
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
.p2align 4,, 8
L(ret_vec_x1):
bsf %VRCX, %VRCX
VMOVU (VEC_SIZE -(VEC_SIZE - CHAR_SIZE))(%rsi, %rcx, CHAR_SIZE), %VMM(0)
VMOVU %VMM(0), (VEC_SIZE -(VEC_SIZE - CHAR_SIZE))(%rdi, %rcx, CHAR_SIZE)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# ifdef USE_AS_STPCPY
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
leaq VEC_SIZE(%rdi, %rcx, CHAR_SIZE), %rax
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
ret
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
.p2align 4,, 4
L(ret_vec_x2):
bsf %VRCX, %VRCX
VMOVU ((VEC_SIZE * 2)-(VEC_SIZE - CHAR_SIZE))(%rsi, %rcx, CHAR_SIZE), %VMM(0)
VMOVU %VMM(0), ((VEC_SIZE * 2)-(VEC_SIZE - CHAR_SIZE))(%rdi, %rcx, CHAR_SIZE)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# ifdef USE_AS_STPCPY
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
leaq (VEC_SIZE * 2)(%rdi, %rcx, CHAR_SIZE), %rax
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
ret
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
/* ret_vec_x3 reuses return code after the loop. */
.p2align 4,, 6
L(ret_vec_x4):
bsf %VRCX, %VRCX
VMOVU ((VEC_SIZE * 4)-(VEC_SIZE - CHAR_SIZE))(%rsi, %rcx, CHAR_SIZE), %VMM(0)
VMOVU %VMM(0), ((VEC_SIZE * 4)-(VEC_SIZE - CHAR_SIZE))(%rdi, %rcx, CHAR_SIZE)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# ifdef USE_AS_STPCPY
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
leaq (VEC_SIZE * 4)(%rdi, %rcx, CHAR_SIZE), %rax
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
ret
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
.p2align 4,, 4
L(page_cross):
# ifndef USE_AS_STRCAT
vpxorq %VZERO_128, %VZERO_128, %VZERO_128
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
movq %rsi, %rcx
andq $(VEC_SIZE * -1), %rcx
VPCMPEQ (%rcx), %VZERO, %k0
KMOV %k0, %VRCX
# ifdef USE_AS_WCSCPY
andl $(VEC_SIZE - 1), %PAGE_ALIGN_REG
shrl $2, %PAGE_ALIGN_REG
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
shrx %VGPR(PAGE_ALIGN_REG_64), %VRCX, %VRCX
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# if USE_MOVSB_IN_PAGE_CROSS
/* Optimizing more aggressively for space as this is very cold
code. This saves 2x cache lines. */
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
/* This adds once to the later result which will get correct
copy bounds. NB: this can never zero-out a non-zero RCX as
to be in the page cross case rsi cannot be aligned and we
already right-shift rcx by the misalignment. */
shl %VRCX
jz L(page_cross_continue)
# if !defined USE_AS_STPCPY && !defined USE_AS_STRCAT
movq %rdi, %rax
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
bsf %VRCX, %VRCX
REP_MOVS
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# ifdef USE_AS_STPCPY
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
leaq -CHAR_SIZE(%rdi), %rax
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
ret
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# else
/* Check if we found zero-char before end of page. */
test %VRCX, %VRCX
jz L(page_cross_continue)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
/* Traditional copy case, essentially same as used in non-page-
cross case but since we can't reuse VMM(0) we need twice as
many loads from rsi. */
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# ifndef USE_AS_STRCAT
xorl %edx, %edx
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
/* Dependency on rdi must already have been satisfied. */
bsf %VRCX, %VRDX
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# ifdef USE_AS_STPCPY
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
leaq (%rdi, %rdx, CHAR_SIZE), %rax
# elif !defined USE_AS_STRCAT
movq %rdi, %rax
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# if VEC_SIZE == 64
# ifdef USE_AS_WCSCPY
testb %cl, %cl
# else
test %ecx, %ecx
# endif
jz L(page_cross_copy_32_63)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# ifdef USE_AS_WCSCPY
testb $0xf, %cl
# else
testw %cx, %cx
# endif
jz L(page_cross_copy_16_31)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# ifdef USE_AS_WCSCPY
testb $0x3, %cl
# else
testb %cl, %cl
# endif
jz L(page_cross_copy_8_15)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# ifdef USE_AS_WCSCPY
movl (%rsi), %esi
movl %esi, (%rdi)
movl $0, (%END_REG)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
ret
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# else
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
testb $0x7, %cl
jz L(page_cross_copy_4_7)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
test %edx, %edx
jz L(page_cross_set_null_term)
movzwl (%rsi), %ecx
movw %cx, (%rdi)
L(page_cross_set_null_term):
movb $0, (%END_REG)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
ret
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
.p2align 4,, 4
L(page_cross_copy_4_7):
movl (%rsi), %ecx
movl -(4 - CHAR_SIZE)(%rsi, %rdx, CHAR_SIZE), %esi
movl %ecx, (%rdi)
movl %esi, -(4 - CHAR_SIZE)(%END_REG)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
ret
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
# if VEC_SIZE == 64
.p2align 4,, 4
L(page_cross_copy_32_63):
VMOVU (%rsi), %VMM_256(0)
VMOVU -(32 - CHAR_SIZE)(%rsi, %rdx, CHAR_SIZE), %VMM_256(1)
VMOVU %VMM_256(0), (%rdi)
VMOVU %VMM_256(1), -(32 - CHAR_SIZE)(%END_REG)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
ret
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
.p2align 4,, 4
L(page_cross_copy_16_31):
vmovdqu (%rsi), %xmm0
vmovdqu -(16 - CHAR_SIZE)(%rsi, %rdx, CHAR_SIZE), %xmm1
vmovdqu %xmm0, (%rdi)
vmovdqu %xmm1, -(16 - CHAR_SIZE)(%END_REG)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
ret
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
.p2align 4,, 4
L(page_cross_copy_8_15):
movq (%rsi), %rcx
movq -(8 - CHAR_SIZE)(%rsi, %rdx, CHAR_SIZE), %rsi
movq %rcx, (%rdi)
movq %rsi, -(8 - CHAR_SIZE)(%END_REG)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
ret
# endif
x86: Optimize and shrink st{r|p}{n}{cat|cpy}-evex functions Optimizations are: 1. Use more overlapping stores to avoid branches. 2. Reduce how unrolled the aligning copies are (this is more of a code-size save, its a negative for some sizes in terms of perf). 3. Improve the loop a bit (similiar to what we do in strlen with 2x vpminu + kortest instead of 3x vpminu + kmov + test). 4. For st{r|p}n{cat|cpy} re-order the branches to minimize the number that are taken. Performance Changes: Times are from N = 10 runs of the benchmark suite and are reported as geometric mean of all ratios of New Implementation / Old Implementation. stpcpy-evex -> 0.922 strcat-evex -> 0.985 strcpy-evex -> 0.880 strncpy-evex -> 0.831 stpncpy-evex -> 0.780 strncat-evex -> 0.958 Code Size Changes: function -> Bytes New / Bytes Old -> Ratio strcat-evex -> 819 / 1874 -> 0.437 strcpy-evex -> 700 / 1074 -> 0.652 stpcpy-evex -> 735 / 1094 -> 0.672 strncpy-evex -> 1397 / 2611 -> 0.535 stpncpy-evex -> 1489 / 2691 -> 0.553 strncat-evex -> 1184 / 2832 -> 0.418 Notes: 1. Because of the significant difference between the implementations they are split into three files. strcpy-evex.S -> strcpy, stpcpy, strcat strncpy-evex.S -> strncpy strncat-evex.S > strncat I couldn't find a way to merge them without making the ifdefs incredibly difficult to follow. 2. All implementations can be made evex512 by including "x86-evex512-vecs.h" at the top. 3. All implementations have an optional define: `USE_EVEX_MASKED_STORE` Setting to one uses evex-masked stores for handling short strings. This saves code size and branches. It's disabled for all implementations are the moment as there are some serious drawbacks to masked stores in certain cases, but that may be fixed on future architectures. Full check passes on x86-64 and build succeeds for all ISA levels w/ and w/o multiarch.
2022-11-09 01:38:38 +00:00
END(STRCPY)
x86-64: Add strcpy family functions with 256-bit EVEX 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.
2021-03-05 14:36:50 +00:00
#endif
Reference in New Issue Copy Permalink