On Fri, Mar 25, 2022 at 11:15 AM H.J. Lu via Libc-alpha wrote: > > On Thu, Mar 24, 2022 at 4:56 PM Noah Goldstein wrote: > > > > geometric_mean(N=40) of all benchmarks AVX2 / SSE42: .702 > > > > All string/memory tests pass. > > Reviewed-by: H.J. Lu > > --- > > sysdeps/x86_64/multiarch/Makefile | 4 + > > sysdeps/x86_64/multiarch/ifunc-impl-list.c | 28 +++ > > sysdeps/x86_64/multiarch/ifunc-strcasecmp.h | 12 + > > .../x86_64/multiarch/strcasecmp_l-avx2-rtm.S | 15 ++ > > sysdeps/x86_64/multiarch/strcasecmp_l-avx2.S | 23 ++ > > sysdeps/x86_64/multiarch/strcmp-avx2.S | 237 +++++++++++++++--- > > .../x86_64/multiarch/strncase_l-avx2-rtm.S | 16 ++ > > sysdeps/x86_64/multiarch/strncase_l-avx2.S | 27 ++ > > 8 files changed, 331 insertions(+), 31 deletions(-) > > create mode 100644 sysdeps/x86_64/multiarch/strcasecmp_l-avx2-rtm.S > > create mode 100644 sysdeps/x86_64/multiarch/strcasecmp_l-avx2.S > > create mode 100644 sysdeps/x86_64/multiarch/strncase_l-avx2-rtm.S > > create mode 100644 sysdeps/x86_64/multiarch/strncase_l-avx2.S > > > > diff --git a/sysdeps/x86_64/multiarch/Makefile b/sysdeps/x86_64/multiarch/Makefile > > index e7b413edad..06e1848823 100644 > > --- a/sysdeps/x86_64/multiarch/Makefile > > +++ b/sysdeps/x86_64/multiarch/Makefile > > @@ -55,6 +55,8 @@ sysdep_routines += \ > > stpncpy-sse2-unaligned \ > > stpncpy-ssse3 \ > > strcasecmp_l-avx \ > > + strcasecmp_l-avx2 \ > > + strcasecmp_l-avx2-rtm \ > > strcasecmp_l-sse2 \ > > strcasecmp_l-sse4_2 \ > > strcasecmp_l-ssse3 \ > > @@ -93,6 +95,8 @@ sysdep_routines += \ > > strlen-evex \ > > strlen-sse2 \ > > strncase_l-avx \ > > + strncase_l-avx2 \ > > + strncase_l-avx2-rtm \ > > strncase_l-sse2 \ > > strncase_l-sse4_2 \ > > strncase_l-ssse3 \ > > diff --git a/sysdeps/x86_64/multiarch/ifunc-impl-list.c b/sysdeps/x86_64/multiarch/ifunc-impl-list.c > > index a594f4176e..3c556d07ac 100644 > > --- a/sysdeps/x86_64/multiarch/ifunc-impl-list.c > > +++ b/sysdeps/x86_64/multiarch/ifunc-impl-list.c > > @@ -436,6 +436,13 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array, > > > > /* Support sysdeps/x86_64/multiarch/strcasecmp_l.c. */ > > IFUNC_IMPL (i, name, strcasecmp, > > + IFUNC_IMPL_ADD (array, i, strcasecmp, > > + CPU_FEATURE_USABLE (AVX2), > > + __strcasecmp_avx2) > > + IFUNC_IMPL_ADD (array, i, strcasecmp, > > + (CPU_FEATURE_USABLE (AVX2) > > + && CPU_FEATURE_USABLE (RTM)), > > + __strcasecmp_avx2_rtm) > > IFUNC_IMPL_ADD (array, i, strcasecmp, > > CPU_FEATURE_USABLE (AVX), > > __strcasecmp_avx) > > @@ -449,6 +456,13 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array, > > > > /* Support sysdeps/x86_64/multiarch/strcasecmp_l.c. */ > > IFUNC_IMPL (i, name, strcasecmp_l, > > + IFUNC_IMPL_ADD (array, i, strcasecmp, > > + CPU_FEATURE_USABLE (AVX2), > > + __strcasecmp_l_avx2) > > + IFUNC_IMPL_ADD (array, i, strcasecmp, > > + (CPU_FEATURE_USABLE (AVX2) > > + && CPU_FEATURE_USABLE (RTM)), > > + __strcasecmp_l_avx2_rtm) > > IFUNC_IMPL_ADD (array, i, strcasecmp_l, > > CPU_FEATURE_USABLE (AVX), > > __strcasecmp_l_avx) > > @@ -576,6 +590,13 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array, > > > > /* Support sysdeps/x86_64/multiarch/strncase_l.c. */ > > IFUNC_IMPL (i, name, strncasecmp, > > + IFUNC_IMPL_ADD (array, i, strncasecmp, > > + CPU_FEATURE_USABLE (AVX2), > > + __strncasecmp_avx2) > > + IFUNC_IMPL_ADD (array, i, strncasecmp, > > + (CPU_FEATURE_USABLE (AVX2) > > + && CPU_FEATURE_USABLE (RTM)), > > + __strncasecmp_avx2_rtm) > > IFUNC_IMPL_ADD (array, i, strncasecmp, > > CPU_FEATURE_USABLE (AVX), > > __strncasecmp_avx) > > @@ -590,6 +611,13 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array, > > > > /* Support sysdeps/x86_64/multiarch/strncase_l.c. */ > > IFUNC_IMPL (i, name, strncasecmp_l, > > + IFUNC_IMPL_ADD (array, i, strncasecmp, > > + CPU_FEATURE_USABLE (AVX2), > > + __strncasecmp_l_avx2) > > + IFUNC_IMPL_ADD (array, i, strncasecmp, > > + (CPU_FEATURE_USABLE (AVX2) > > + && CPU_FEATURE_USABLE (RTM)), > > + __strncasecmp_l_avx2_rtm) > > IFUNC_IMPL_ADD (array, i, strncasecmp_l, > > CPU_FEATURE_USABLE (AVX), > > __strncasecmp_l_avx) > > diff --git a/sysdeps/x86_64/multiarch/ifunc-strcasecmp.h b/sysdeps/x86_64/multiarch/ifunc-strcasecmp.h > > index 9e3cc61ac0..c4de111fd0 100644 > > --- a/sysdeps/x86_64/multiarch/ifunc-strcasecmp.h > > +++ b/sysdeps/x86_64/multiarch/ifunc-strcasecmp.h > > @@ -23,12 +23,24 @@ extern __typeof (REDIRECT_NAME) OPTIMIZE (sse2) attribute_hidden; > > extern __typeof (REDIRECT_NAME) OPTIMIZE (ssse3) attribute_hidden; > > extern __typeof (REDIRECT_NAME) OPTIMIZE (sse42) attribute_hidden; > > extern __typeof (REDIRECT_NAME) OPTIMIZE (avx) attribute_hidden; > > +extern __typeof (REDIRECT_NAME) OPTIMIZE (avx2) attribute_hidden; > > +extern __typeof (REDIRECT_NAME) OPTIMIZE (avx2_rtm) attribute_hidden; > > > > static inline void * > > IFUNC_SELECTOR (void) > > { > > const struct cpu_features* cpu_features = __get_cpu_features (); > > > > + if (CPU_FEATURE_USABLE_P (cpu_features, AVX2) > > + && CPU_FEATURES_ARCH_P (cpu_features, AVX_Fast_Unaligned_Load)) > > + { > > + if (CPU_FEATURE_USABLE_P (cpu_features, RTM)) > > + return OPTIMIZE (avx2_rtm); > > + > > + if (!CPU_FEATURES_ARCH_P (cpu_features, Prefer_No_VZEROUPPER)) > > + return OPTIMIZE (avx2); > > + } > > + > > if (CPU_FEATURE_USABLE_P (cpu_features, AVX)) > > return OPTIMIZE (avx); > > > > diff --git a/sysdeps/x86_64/multiarch/strcasecmp_l-avx2-rtm.S b/sysdeps/x86_64/multiarch/strcasecmp_l-avx2-rtm.S > > new file mode 100644 > > index 0000000000..09957fc3c5 > > --- /dev/null > > +++ b/sysdeps/x86_64/multiarch/strcasecmp_l-avx2-rtm.S > > @@ -0,0 +1,15 @@ > > +#ifndef STRCMP > > +# define STRCMP __strcasecmp_l_avx2_rtm > > +#endif > > + > > +#define _GLABEL(x) x ## _rtm > > +#define GLABEL(x) _GLABEL(x) > > + > > +#define ZERO_UPPER_VEC_REGISTERS_RETURN \ > > + ZERO_UPPER_VEC_REGISTERS_RETURN_XTEST > > + > > +#define VZEROUPPER_RETURN jmp L(return_vzeroupper) > > + > > +#define SECTION(p) p##.avx.rtm > > + > > +#include "strcasecmp_l-avx2.S" > > diff --git a/sysdeps/x86_64/multiarch/strcasecmp_l-avx2.S b/sysdeps/x86_64/multiarch/strcasecmp_l-avx2.S > > new file mode 100644 > > index 0000000000..e2762f2a22 > > --- /dev/null > > +++ b/sysdeps/x86_64/multiarch/strcasecmp_l-avx2.S > > @@ -0,0 +1,23 @@ > > +/* strcasecmp_l optimized with AVX2. > > + Copyright (C) 2017-2022 Free Software Foundation, Inc. > > + This file is part of the GNU C Library. > > + > > + The GNU C Library is free software; you can redistribute it and/or > > + modify it under the terms of the GNU Lesser General Public > > + License as published by the Free Software Foundation; either > > + version 2.1 of the License, or (at your option) any later version. > > + > > + The GNU C Library is distributed in the hope that it will be useful, > > + but WITHOUT ANY WARRANTY; without even the implied warranty of > > + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU > > + Lesser General Public License for more details. > > + > > + You should have received a copy of the GNU Lesser General Public > > + License along with the GNU C Library; if not, see > > + . */ > > + > > +#ifndef STRCMP > > +# define STRCMP __strcasecmp_l_avx2 > > +#endif > > +#define USE_AS_STRCASECMP_L > > +#include "strcmp-avx2.S" > > diff --git a/sysdeps/x86_64/multiarch/strcmp-avx2.S b/sysdeps/x86_64/multiarch/strcmp-avx2.S > > index 86a86b68e3..8da09bd86d 100644 > > --- a/sysdeps/x86_64/multiarch/strcmp-avx2.S > > +++ b/sysdeps/x86_64/multiarch/strcmp-avx2.S > > @@ -20,6 +20,10 @@ > > > > # include > > > > +# if defined USE_AS_STRCASECMP_L > > +# include "locale-defines.h" > > +# endif > > + > > # ifndef STRCMP > > # define STRCMP __strcmp_avx2 > > # endif > > @@ -74,13 +78,88 @@ > > # define VEC_OFFSET (-VEC_SIZE) > > # endif > > > > +# ifdef USE_AS_STRCASECMP_L > > +# define BYTE_LOOP_REG OFFSET_REG > > +# else > > +# define BYTE_LOOP_REG ecx > > +# endif > > + > > +# ifdef USE_AS_STRCASECMP_L > > +# ifdef USE_AS_STRNCMP > > +# define STRCASECMP __strncasecmp_avx2 > > +# define LOCALE_REG rcx > > +# define LOCALE_REG_LP RCX_LP > > +# define STRCASECMP_NONASCII __strncasecmp_l_nonascii > > +# else > > +# define STRCASECMP __strcasecmp_avx2 > > +# define LOCALE_REG rdx > > +# define LOCALE_REG_LP RDX_LP > > +# define STRCASECMP_NONASCII __strcasecmp_l_nonascii > > +# endif > > +# endif > > + > > # define xmmZERO xmm15 > > # define ymmZERO ymm15 > > > > +# define LCASE_MIN_ymm %ymm10 > > +# define LCASE_MAX_ymm %ymm11 > > +# define CASE_ADD_ymm %ymm12 > > + > > +# define LCASE_MIN_xmm %xmm10 > > +# define LCASE_MAX_xmm %xmm11 > > +# define CASE_ADD_xmm %xmm12 > > + > > + /* r11 is never use elsewhere so this is safe to maintain. */ > > +# define TOLOWER_BASE %r11 > > + > > # ifndef SECTION > > # define SECTION(p) p##.avx > > # endif > > > > +# ifdef USE_AS_STRCASECMP_L > > +# define REG(x, y) x ## y > > +# define TOLOWER(reg1_in, reg1_out, reg2_in, reg2_out, ext) \ > > + vpaddb REG(LCASE_MIN_, ext), reg1_in, REG(%ext, 8); \ > > + vpaddb REG(LCASE_MIN_, ext), reg2_in, REG(%ext, 9); \ > > + vpcmpgtb REG(LCASE_MAX_, ext), REG(%ext, 8), REG(%ext, 8); \ > > + vpcmpgtb REG(LCASE_MAX_, ext), REG(%ext, 9), REG(%ext, 9); \ > > + vpandn REG(CASE_ADD_, ext), REG(%ext, 8), REG(%ext, 8); \ > > + vpandn REG(CASE_ADD_, ext), REG(%ext, 9), REG(%ext, 9); \ > > + vpaddb REG(%ext, 8), reg1_in, reg1_out; \ > > + vpaddb REG(%ext, 9), reg2_in, reg2_out > > + > > +# define TOLOWER_gpr(src, dst) movl (TOLOWER_BASE, src, 4), dst > > +# define TOLOWER_ymm(...) TOLOWER(__VA_ARGS__, ymm) > > +# define TOLOWER_xmm(...) TOLOWER(__VA_ARGS__, xmm) > > + > > +# define CMP_R1_R2(s1_reg, s2_reg, scratch_reg, reg_out, ext) \ > > + TOLOWER (s1_reg, scratch_reg, s2_reg, s2_reg, ext); \ > > + VPCMPEQ scratch_reg, s2_reg, reg_out > > + > > +# define CMP_R1_S2(s1_reg, s2_mem, scratch_reg, reg_out, ext) \ > > + VMOVU s2_mem, reg_out; \ > > + CMP_R1_R2(s1_reg, reg_out, scratch_reg, reg_out, ext) > > + > > +# define CMP_R1_R2_ymm(...) CMP_R1_R2(__VA_ARGS__, ymm) > > +# define CMP_R1_R2_xmm(...) CMP_R1_R2(__VA_ARGS__, xmm) > > + > > +# define CMP_R1_S2_ymm(...) CMP_R1_S2(__VA_ARGS__, ymm) > > +# define CMP_R1_S2_xmm(...) CMP_R1_S2(__VA_ARGS__, xmm) > > + > > +# else > > +# define TOLOWER_gpr(...) > > +# define TOLOWER_ymm(...) > > +# define TOLOWER_xmm(...) > > + > > +# define CMP_R1_R2_ymm(s1_reg, s2_reg, scratch_reg, reg_out) \ > > + VPCMPEQ s2_reg, s1_reg, reg_out > > + > > +# define CMP_R1_R2_xmm(...) CMP_R1_R2_ymm(__VA_ARGS__) > > + > > +# define CMP_R1_S2_ymm(...) CMP_R1_R2_ymm(__VA_ARGS__) > > +# define CMP_R1_S2_xmm(...) CMP_R1_R2_xmm(__VA_ARGS__) > > +# endif > > + > > /* Warning! > > wcscmp/wcsncmp have to use SIGNED comparison for elements. > > strcmp/strncmp have to use UNSIGNED comparison for elements. > > @@ -102,8 +181,49 @@ > > returned. */ > > > > .section SECTION(.text), "ax", @progbits > > -ENTRY(STRCMP) > > + .align 16 > > + .type STRCMP, @function > > + .globl STRCMP > > + .hidden STRCMP > > + > > +# ifndef GLABEL > > +# define GLABEL(...) __VA_ARGS__ > > +# endif > > + > > +# ifdef USE_AS_STRCASECMP_L > > +ENTRY (GLABEL(STRCASECMP)) > > + movq __libc_tsd_LOCALE@gottpoff(%rip), %rax > > + mov %fs:(%rax), %LOCALE_REG_LP > > + > > + /* Either 1 or 5 bytes (dependeing if CET is enabled). */ > > + .p2align 4 > > +END (GLABEL(STRCASECMP)) > > + /* FALLTHROUGH to strcasecmp/strncasecmp_l. */ > > +# endif > > + > > + .p2align 4 > > +STRCMP: > > + cfi_startproc > > + _CET_ENDBR > > + CALL_MCOUNT > > + > > +# if defined USE_AS_STRCASECMP_L > > + /* We have to fall back on the C implementation for locales with > > + encodings not matching ASCII for single bytes. */ > > +# if LOCALE_T___LOCALES != 0 || LC_CTYPE != 0 > > + mov LOCALE_T___LOCALES + LC_CTYPE * LP_SIZE(%LOCALE_REG), %RAX_LP > > +# else > > + mov (%LOCALE_REG), %RAX_LP > > +# endif > > + testl $1, LOCALE_DATA_VALUES + _NL_CTYPE_NONASCII_CASE * SIZEOF_VALUES(%rax) > > + jne STRCASECMP_NONASCII > > + leaq _nl_C_LC_CTYPE_tolower + 128 * 4(%rip), TOLOWER_BASE > > +# endif > > + > > # ifdef USE_AS_STRNCMP > > + /* Don't overwrite LOCALE_REG (rcx) until we have pass > > + L(one_or_less). Otherwise we might use the wrong locale in > > + the OVERFLOW_STRCMP (strcasecmp_l). */ > > # ifdef __ILP32__ > > /* Clear the upper 32 bits. */ > > movl %edx, %edx > > @@ -128,6 +248,30 @@ ENTRY(STRCMP) > > # endif > > # endif > > vpxor %xmmZERO, %xmmZERO, %xmmZERO > > +# if defined USE_AS_STRCASECMP_L > > + .section .rodata.cst32, "aM", @progbits, 32 > > + .align 32 > > +L(lcase_min): > > + .quad 0x3f3f3f3f3f3f3f3f > > + .quad 0x3f3f3f3f3f3f3f3f > > + .quad 0x3f3f3f3f3f3f3f3f > > + .quad 0x3f3f3f3f3f3f3f3f > > +L(lcase_max): > > + .quad 0x9999999999999999 > > + .quad 0x9999999999999999 > > + .quad 0x9999999999999999 > > + .quad 0x9999999999999999 > > +L(case_add): > > + .quad 0x2020202020202020 > > + .quad 0x2020202020202020 > > + .quad 0x2020202020202020 > > + .quad 0x2020202020202020 > > + .previous > > + > > + vmovdqa L(lcase_min)(%rip), LCASE_MIN_ymm > > + vmovdqa L(lcase_max)(%rip), LCASE_MAX_ymm > > + vmovdqa L(case_add)(%rip), CASE_ADD_ymm > > +# endif > > movl %edi, %eax > > orl %esi, %eax > > sall $20, %eax > > @@ -138,8 +282,10 @@ ENTRY(STRCMP) > > L(no_page_cross): > > /* Safe to compare 4x vectors. */ > > VMOVU (%rdi), %ymm0 > > - /* 1s where s1 and s2 equal. */ > > - VPCMPEQ (%rsi), %ymm0, %ymm1 > > + /* 1s where s1 and s2 equal. Just VPCMPEQ if its not strcasecmp. > > + Otherwise converts ymm0 and load from rsi to lower. ymm2 is > > + scratch and ymm1 is the return. */ > > + CMP_R1_S2_ymm (%ymm0, (%rsi), %ymm2, %ymm1) > > /* 1s at null CHAR. */ > > VPCMPEQ %ymm0, %ymmZERO, %ymm2 > > /* 1s where s1 and s2 equal AND not null CHAR. */ > > @@ -172,6 +318,8 @@ L(return_vec_0): > > # else > > movzbl (%rdi, %rcx), %eax > > movzbl (%rsi, %rcx), %ecx > > + TOLOWER_gpr (%rax, %eax) > > + TOLOWER_gpr (%rcx, %ecx) > > subl %ecx, %eax > > # endif > > L(ret0): > > @@ -192,6 +340,10 @@ L(ret_zero): > > > > .p2align 4,, 5 > > L(one_or_less): > > +# ifdef USE_AS_STRCASECMP_L > > + /* Set locale argument for strcasecmp. */ > > + movq %LOCALE_REG, %rdx > > +# endif > > jb L(ret_zero) > > /* 'nbe' covers the case where length is negative (large > > unsigned). */ > > @@ -207,6 +359,8 @@ L(one_or_less): > > # else > > movzbl (%rdi), %eax > > movzbl (%rsi), %ecx > > + TOLOWER_gpr (%rax, %eax) > > + TOLOWER_gpr (%rcx, %ecx) > > subl %ecx, %eax > > # endif > > L(ret1): > > @@ -234,6 +388,8 @@ L(return_vec_1): > > # else > > movzbl VEC_SIZE(%rdi, %rcx), %eax > > movzbl VEC_SIZE(%rsi, %rcx), %ecx > > + TOLOWER_gpr (%rax, %eax) > > + TOLOWER_gpr (%rcx, %ecx) > > subl %ecx, %eax > > # endif > > L(ret2): > > @@ -265,6 +421,8 @@ L(return_vec_2): > > # else > > movzbl (VEC_SIZE * 2)(%rdi, %rcx), %eax > > movzbl (VEC_SIZE * 2)(%rsi, %rcx), %ecx > > + TOLOWER_gpr (%rax, %eax) > > + TOLOWER_gpr (%rcx, %ecx) > > subl %ecx, %eax > > # endif > > L(ret3): > > @@ -285,6 +443,8 @@ L(return_vec_3): > > # else > > movzbl (VEC_SIZE * 3)(%rdi, %rcx), %eax > > movzbl (VEC_SIZE * 3)(%rsi, %rcx), %ecx > > + TOLOWER_gpr (%rax, %eax) > > + TOLOWER_gpr (%rcx, %ecx) > > subl %ecx, %eax > > # endif > > L(ret4): > > @@ -295,7 +455,7 @@ L(ret4): > > L(more_3x_vec): > > /* Safe to compare 4x vectors. */ > > VMOVU VEC_SIZE(%rdi), %ymm0 > > - VPCMPEQ VEC_SIZE(%rsi), %ymm0, %ymm1 > > + CMP_R1_S2_ymm (%ymm0, VEC_SIZE(%rsi), %ymm2, %ymm1) > > VPCMPEQ %ymm0, %ymmZERO, %ymm2 > > vpandn %ymm1, %ymm2, %ymm1 > > vpmovmskb %ymm1, %ecx > > @@ -308,7 +468,7 @@ L(more_3x_vec): > > # endif > > > > VMOVU (VEC_SIZE * 2)(%rdi), %ymm0 > > - VPCMPEQ (VEC_SIZE * 2)(%rsi), %ymm0, %ymm1 > > + CMP_R1_S2_ymm (%ymm0, (VEC_SIZE * 2)(%rsi), %ymm2, %ymm1) > > VPCMPEQ %ymm0, %ymmZERO, %ymm2 > > vpandn %ymm1, %ymm2, %ymm1 > > vpmovmskb %ymm1, %ecx > > @@ -316,7 +476,7 @@ L(more_3x_vec): > > jnz L(return_vec_2) > > > > VMOVU (VEC_SIZE * 3)(%rdi), %ymm0 > > - VPCMPEQ (VEC_SIZE * 3)(%rsi), %ymm0, %ymm1 > > + CMP_R1_S2_ymm (%ymm0, (VEC_SIZE * 3)(%rsi), %ymm2, %ymm1) > > VPCMPEQ %ymm0, %ymmZERO, %ymm2 > > vpandn %ymm1, %ymm2, %ymm1 > > vpmovmskb %ymm1, %ecx > > @@ -391,12 +551,10 @@ L(loop_skip_page_cross_check): > > VMOVA (VEC_SIZE * 3)(%rdi), %ymm6 > > > > /* ymm1 all 1s where s1 and s2 equal. All 0s otherwise. */ > > - VPCMPEQ (VEC_SIZE * 0)(%rsi), %ymm0, %ymm1 > > - > > - VPCMPEQ (VEC_SIZE * 1)(%rsi), %ymm2, %ymm3 > > - VPCMPEQ (VEC_SIZE * 2)(%rsi), %ymm4, %ymm5 > > - VPCMPEQ (VEC_SIZE * 3)(%rsi), %ymm6, %ymm7 > > - > > + CMP_R1_S2_ymm (%ymm0, (VEC_SIZE * 0)(%rsi), %ymm3, %ymm1) > > + CMP_R1_S2_ymm (%ymm2, (VEC_SIZE * 1)(%rsi), %ymm5, %ymm3) > > + CMP_R1_S2_ymm (%ymm4, (VEC_SIZE * 2)(%rsi), %ymm7, %ymm5) > > + CMP_R1_S2_ymm (%ymm6, (VEC_SIZE * 3)(%rsi), %ymm13, %ymm7) > > > > /* If any mismatches or null CHAR then 0 CHAR, otherwise non- > > zero. */ > > @@ -465,6 +623,8 @@ L(return_vec_2_3_end): > > # else > > movzbl (VEC_SIZE * 2 - VEC_OFFSET)(%rdi, %LOOP_REG64), %eax > > movzbl (VEC_SIZE * 2 - VEC_OFFSET)(%rsi, %LOOP_REG64), %ecx > > + TOLOWER_gpr (%rax, %eax) > > + TOLOWER_gpr (%rcx, %ecx) > > subl %ecx, %eax > > xorl %r8d, %eax > > subl %r8d, %eax > > @@ -508,6 +668,8 @@ L(return_vec_0_end): > > # else > > movzbl (%rdi, %rcx), %eax > > movzbl (%rsi, %rcx), %ecx > > + TOLOWER_gpr (%rax, %eax) > > + TOLOWER_gpr (%rcx, %ecx) > > subl %ecx, %eax > > xorl %r8d, %eax > > subl %r8d, %eax > > @@ -530,6 +692,8 @@ L(return_vec_1_end): > > # else > > movzbl VEC_SIZE(%rdi, %rcx), %eax > > movzbl VEC_SIZE(%rsi, %rcx), %ecx > > + TOLOWER_gpr (%rax, %eax) > > + TOLOWER_gpr (%rcx, %ecx) > > subl %ecx, %eax > > xorl %r8d, %eax > > subl %r8d, %eax > > @@ -556,6 +720,8 @@ L(return_vec_2_end): > > # else > > movzbl (VEC_SIZE * 2)(%rdi, %rcx), %eax > > movzbl (VEC_SIZE * 2)(%rsi, %rcx), %ecx > > + TOLOWER_gpr (%rax, %eax) > > + TOLOWER_gpr (%rcx, %ecx) > > subl %ecx, %eax > > xorl %r8d, %eax > > subl %r8d, %eax > > @@ -583,7 +749,7 @@ L(page_cross_during_loop): > > jle L(less_1x_vec_till_page_cross) > > > > VMOVA (%rdi), %ymm0 > > - VPCMPEQ (%rsi), %ymm0, %ymm1 > > + CMP_R1_S2_ymm (%ymm0, (%rsi), %ymm2, %ymm1) > > VPCMPEQ %ymm0, %ymmZERO, %ymm2 > > vpandn %ymm1, %ymm2, %ymm1 > > vpmovmskb %ymm1, %ecx > > @@ -605,7 +771,7 @@ L(less_1x_vec_till_page_cross): > > here, it means the previous page (rdi - VEC_SIZE) has already > > been loaded earlier so must be valid. */ > > VMOVU -VEC_SIZE(%rdi, %rax), %ymm0 > > - VPCMPEQ -VEC_SIZE(%rsi, %rax), %ymm0, %ymm1 > > + CMP_R1_S2_ymm (%ymm0, -VEC_SIZE(%rsi, %rax), %ymm2, %ymm1) > > VPCMPEQ %ymm0, %ymmZERO, %ymm2 > > vpandn %ymm1, %ymm2, %ymm1 > > vpmovmskb %ymm1, %ecx > > @@ -647,6 +813,8 @@ L(return_page_cross_cmp_mem): > > # else > > movzbl VEC_OFFSET(%rdi, %rcx), %eax > > movzbl VEC_OFFSET(%rsi, %rcx), %ecx > > + TOLOWER_gpr (%rax, %eax) > > + TOLOWER_gpr (%rcx, %ecx) > > subl %ecx, %eax > > xorl %r8d, %eax > > subl %r8d, %eax > > @@ -673,7 +841,7 @@ L(more_2x_vec_till_page_cross): > > iteration here. */ > > > > VMOVU VEC_SIZE(%rdi), %ymm0 > > - VPCMPEQ VEC_SIZE(%rsi), %ymm0, %ymm1 > > + CMP_R1_S2_ymm (%ymm0, VEC_SIZE(%rsi), %ymm2, %ymm1) > > VPCMPEQ %ymm0, %ymmZERO, %ymm2 > > vpandn %ymm1, %ymm2, %ymm1 > > vpmovmskb %ymm1, %ecx > > @@ -689,7 +857,7 @@ L(more_2x_vec_till_page_cross): > > > > /* Safe to include comparisons from lower bytes. */ > > VMOVU -(VEC_SIZE * 2)(%rdi, %rax), %ymm0 > > - VPCMPEQ -(VEC_SIZE * 2)(%rsi, %rax), %ymm0, %ymm1 > > + CMP_R1_S2_ymm (%ymm0, -(VEC_SIZE * 2)(%rsi, %rax), %ymm2, %ymm1) > > VPCMPEQ %ymm0, %ymmZERO, %ymm2 > > vpandn %ymm1, %ymm2, %ymm1 > > vpmovmskb %ymm1, %ecx > > @@ -697,7 +865,7 @@ L(more_2x_vec_till_page_cross): > > jnz L(return_vec_page_cross_0) > > > > VMOVU -(VEC_SIZE * 1)(%rdi, %rax), %ymm0 > > - VPCMPEQ -(VEC_SIZE * 1)(%rsi, %rax), %ymm0, %ymm1 > > + CMP_R1_S2_ymm (%ymm0, -(VEC_SIZE * 1)(%rsi, %rax), %ymm2, %ymm1) > > VPCMPEQ %ymm0, %ymmZERO, %ymm2 > > vpandn %ymm1, %ymm2, %ymm1 > > vpmovmskb %ymm1, %ecx > > @@ -715,8 +883,8 @@ L(more_2x_vec_till_page_cross): > > VMOVA (VEC_SIZE * 2)(%rdi), %ymm4 > > VMOVA (VEC_SIZE * 3)(%rdi), %ymm6 > > > > - VPCMPEQ (VEC_SIZE * 2)(%rsi), %ymm4, %ymm5 > > - VPCMPEQ (VEC_SIZE * 3)(%rsi), %ymm6, %ymm7 > > + CMP_R1_S2_ymm (%ymm4, (VEC_SIZE * 2)(%rsi), %ymm7, %ymm5) > > + CMP_R1_S2_ymm (%ymm6, (VEC_SIZE * 3)(%rsi), %ymm13, %ymm7) > > vpand %ymm4, %ymm5, %ymm5 > > vpand %ymm6, %ymm7, %ymm7 > > VPMINU %ymm5, %ymm7, %ymm7 > > @@ -767,6 +935,8 @@ L(return_vec_page_cross_1): > > # else > > movzbl VEC_OFFSET(%rdi, %rcx), %eax > > movzbl VEC_OFFSET(%rsi, %rcx), %ecx > > + TOLOWER_gpr (%rax, %eax) > > + TOLOWER_gpr (%rcx, %ecx) > > subl %ecx, %eax > > xorl %r8d, %eax > > subl %r8d, %eax > > @@ -822,7 +992,7 @@ L(page_cross): > > L(page_cross_loop): > > > > VMOVU (%rdi, %OFFSET_REG64), %ymm0 > > - VPCMPEQ (%rsi, %OFFSET_REG64), %ymm0, %ymm1 > > + CMP_R1_S2_ymm (%ymm0, (%rsi, %OFFSET_REG64), %ymm2, %ymm1) > > VPCMPEQ %ymm0, %ymmZERO, %ymm2 > > vpandn %ymm1, %ymm2, %ymm1 > > vpmovmskb %ymm1, %ecx > > @@ -840,11 +1010,11 @@ L(page_cross_loop): > > subl %eax, %OFFSET_REG > > /* OFFSET_REG has distance to page cross - VEC_SIZE. Guranteed > > to not cross page so is safe to load. Since we have already > > - loaded at least 1 VEC from rsi it is also guranteed to be safe. > > - */ > > + loaded at least 1 VEC from rsi it is also guranteed to be > > + safe. */ > > > > VMOVU (%rdi, %OFFSET_REG64), %ymm0 > > - VPCMPEQ (%rsi, %OFFSET_REG64), %ymm0, %ymm1 > > + CMP_R1_S2_ymm (%ymm0, (%rsi, %OFFSET_REG64), %ymm2, %ymm1) > > VPCMPEQ %ymm0, %ymmZERO, %ymm2 > > vpandn %ymm1, %ymm2, %ymm1 > > vpmovmskb %ymm1, %ecx > > @@ -877,6 +1047,8 @@ L(ret_vec_page_cross_cont): > > # else > > movzbl (%rdi, %rcx), %eax > > movzbl (%rsi, %rcx), %ecx > > + TOLOWER_gpr (%rax, %eax) > > + TOLOWER_gpr (%rcx, %ecx) > > subl %ecx, %eax > > xorl %r8d, %eax > > subl %r8d, %eax > > @@ -930,7 +1102,7 @@ L(less_1x_vec_till_page): > > ja L(less_16_till_page) > > > > VMOVU (%rdi), %xmm0 > > - VPCMPEQ (%rsi), %xmm0, %xmm1 > > + CMP_R1_S2_xmm (%xmm0, (%rsi), %xmm2, %xmm1) > > VPCMPEQ %xmm0, %xmmZERO, %xmm2 > > vpandn %xmm1, %xmm2, %xmm1 > > vpmovmskb %ymm1, %ecx > > @@ -948,7 +1120,7 @@ L(less_1x_vec_till_page): > > # endif > > > > VMOVU (%rdi, %OFFSET_REG64), %xmm0 > > - VPCMPEQ (%rsi, %OFFSET_REG64), %xmm0, %xmm1 > > + CMP_R1_S2_xmm (%xmm0, (%rsi, %OFFSET_REG64), %xmm2, %xmm1) > > VPCMPEQ %xmm0, %xmmZERO, %xmm2 > > vpandn %xmm1, %xmm2, %xmm1 > > vpmovmskb %ymm1, %ecx > > @@ -986,7 +1158,7 @@ L(less_16_till_page): > > vmovq (%rdi), %xmm0 > > vmovq (%rsi), %xmm1 > > VPCMPEQ %xmm0, %xmmZERO, %xmm2 > > - VPCMPEQ %xmm1, %xmm0, %xmm1 > > + CMP_R1_R2_xmm (%xmm0, %xmm1, %xmm3, %xmm1) > > vpandn %xmm1, %xmm2, %xmm1 > > vpmovmskb %ymm1, %ecx > > incb %cl > > @@ -1006,7 +1178,7 @@ L(less_16_till_page): > > vmovq (%rdi, %OFFSET_REG64), %xmm0 > > vmovq (%rsi, %OFFSET_REG64), %xmm1 > > VPCMPEQ %xmm0, %xmmZERO, %xmm2 > > - VPCMPEQ %xmm1, %xmm0, %xmm1 > > + CMP_R1_R2_xmm (%xmm0, %xmm1, %xmm3, %xmm1) > > vpandn %xmm1, %xmm2, %xmm1 > > vpmovmskb %ymm1, %ecx > > incb %cl > > @@ -1062,7 +1234,7 @@ L(ret_less_8_wcs): > > vmovd (%rdi), %xmm0 > > vmovd (%rsi), %xmm1 > > VPCMPEQ %xmm0, %xmmZERO, %xmm2 > > - VPCMPEQ %xmm1, %xmm0, %xmm1 > > + CMP_R1_R2_xmm (%xmm0, %xmm1, %xmm3, %xmm1) > > vpandn %xmm1, %xmm2, %xmm1 > > vpmovmskb %ymm1, %ecx > > subl $0xf, %ecx > > @@ -1081,7 +1253,7 @@ L(ret_less_8_wcs): > > vmovd (%rdi, %OFFSET_REG64), %xmm0 > > vmovd (%rsi, %OFFSET_REG64), %xmm1 > > VPCMPEQ %xmm0, %xmmZERO, %xmm2 > > - VPCMPEQ %xmm1, %xmm0, %xmm1 > > + CMP_R1_R2_xmm (%xmm0, %xmm1, %xmm3, %xmm1) > > vpandn %xmm1, %xmm2, %xmm1 > > vpmovmskb %ymm1, %ecx > > subl $0xf, %ecx > > @@ -1115,7 +1287,9 @@ L(less_4_till_page): > > L(less_4_loop): > > movzbl (%rdi), %eax > > movzbl (%rsi, %rdi), %ecx > > - subl %ecx, %eax > > + TOLOWER_gpr (%rax, %eax) > > + TOLOWER_gpr (%rcx, %BYTE_LOOP_REG) > > + subl %BYTE_LOOP_REG, %eax > > jnz L(ret_less_4_loop) > > testl %ecx, %ecx > > jz L(ret_zero_4_loop) > > @@ -1142,5 +1316,6 @@ L(ret_less_4_loop): > > subl %r8d, %eax > > ret > > # endif > > -END(STRCMP) > > + cfi_endproc > > + .size STRCMP, .-STRCMP > > #endif > > diff --git a/sysdeps/x86_64/multiarch/strncase_l-avx2-rtm.S b/sysdeps/x86_64/multiarch/strncase_l-avx2-rtm.S > > new file mode 100644 > > index 0000000000..58c05dcfb8 > > --- /dev/null > > +++ b/sysdeps/x86_64/multiarch/strncase_l-avx2-rtm.S > > @@ -0,0 +1,16 @@ > > +#ifndef STRCMP > > +# define STRCMP __strncasecmp_l_avx2_rtm > > +#endif > > + > > +#define _GLABEL(x) x ## _rtm > > +#define GLABEL(x) _GLABEL(x) > > + > > +#define ZERO_UPPER_VEC_REGISTERS_RETURN \ > > + ZERO_UPPER_VEC_REGISTERS_RETURN_XTEST > > + > > +#define VZEROUPPER_RETURN jmp L(return_vzeroupper) > > + > > +#define SECTION(p) p##.avx.rtm > > +#define OVERFLOW_STRCMP __strcasecmp_l_avx2_rtm > > + > > +#include "strncase_l-avx2.S" > > diff --git a/sysdeps/x86_64/multiarch/strncase_l-avx2.S b/sysdeps/x86_64/multiarch/strncase_l-avx2.S > > new file mode 100644 > > index 0000000000..48c0aa21f8 > > --- /dev/null > > +++ b/sysdeps/x86_64/multiarch/strncase_l-avx2.S > > @@ -0,0 +1,27 @@ > > +/* strncasecmp_l optimized with AVX2. > > + Copyright (C) 2017-2022 Free Software Foundation, Inc. > > + This file is part of the GNU C Library. > > + > > + The GNU C Library is free software; you can redistribute it and/or > > + modify it under the terms of the GNU Lesser General Public > > + License as published by the Free Software Foundation; either > > + version 2.1 of the License, or (at your option) any later version. > > + > > + The GNU C Library is distributed in the hope that it will be useful, > > + but WITHOUT ANY WARRANTY; without even the implied warranty of > > + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU > > + Lesser General Public License for more details. > > + > > + You should have received a copy of the GNU Lesser General Public > > + License along with the GNU C Library; if not, see > > + . */ > > + > > +#ifndef STRCMP > > +# define STRCMP __strncasecmp_l_avx2 > > +#endif > > +#define USE_AS_STRCASECMP_L > > +#define USE_AS_STRNCMP > > +#ifndef OVERFLOW_STRCMP > > +# define OVERFLOW_STRCMP __strcasecmp_l_avx2 > > +#endif > > +#include "strcmp-avx2.S" > > -- > > 2.25.1 > > > > LGTM. > > Reviewed-by: H.J. Lu > > Thanks. > > -- > H.J. I would like to backport this patch to release branches. Any comments or objections? Conflict resolution patch attached. --Sunil