Re: [PATCH v9 6/6] x86: Update strlen-evex-base to use new reg/vec macros.

["H.J. Lu" <hjl.tools@gmail.com>]

On Fri, Oct 14, 2022 at 5:21 PM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
>
> To avoid duplicate the VMM / GPR / mask insn macros in all incoming
> evex512 files use the macros defined in 'reg-macros.h' and
> '{vec}-macros.h'
>
> This commit does not change libc.so
>
> Tested build on x86-64
> ---
>  sysdeps/x86_64/multiarch/strlen-evex-base.S | 116 +++++++-------------
>  sysdeps/x86_64/multiarch/strlen-evex512.S   |   4 +-
>  2 files changed, 44 insertions(+), 76 deletions(-)
>
> diff --git a/sysdeps/x86_64/multiarch/strlen-evex-base.S b/sysdeps/x86_64/multiarch/strlen-evex-base.S
> index 418e9f8411..c832b15a48 100644
> --- a/sysdeps/x86_64/multiarch/strlen-evex-base.S
> +++ b/sysdeps/x86_64/multiarch/strlen-evex-base.S
> @@ -36,42 +36,10 @@
>  #  define CHAR_SIZE    1
>  # endif
>
> -# define XMM0          xmm16
>  # define PAGE_SIZE     4096
>  # define CHAR_PER_VEC  (VEC_SIZE / CHAR_SIZE)
>
> -# if VEC_SIZE == 64
> -#  define KMOV         kmovq
> -#  define KORTEST      kortestq
> -#  define RAX          rax
> -#  define RCX          rcx
> -#  define RDX          rdx
> -#  define SHR          shrq
> -#  define TEXTSUFFIX   evex512
> -#  define VMM0         zmm16
> -#  define VMM1         zmm17
> -#  define VMM2         zmm18
> -#  define VMM3         zmm19
> -#  define VMM4         zmm20
> -#  define VMOVA                vmovdqa64
> -# elif VEC_SIZE == 32
> -/* Currently Unused.  */
> -#  define KMOV         kmovd
> -#  define KORTEST      kortestd
> -#  define RAX          eax
> -#  define RCX          ecx
> -#  define RDX          edx
> -#  define SHR          shrl
> -#  define TEXTSUFFIX   evex256
> -#  define VMM0         ymm16
> -#  define VMM1         ymm17
> -#  define VMM2         ymm18
> -#  define VMM3         ymm19
> -#  define VMM4         ymm20
> -#  define VMOVA                vmovdqa32
> -# endif
> -
> -       .section .text.TEXTSUFFIX, "ax", @progbits
> +       .section SECTION(.text),"ax",@progbits
>  /* Aligning entry point to 64 byte, provides better performance for
>     one vector length string.  */
>  ENTRY_P2ALIGN (STRLEN, 6)
> @@ -86,18 +54,18 @@ ENTRY_P2ALIGN (STRLEN, 6)
>  # endif
>
>         movl    %edi, %eax
> -       vpxorq  %XMM0, %XMM0, %XMM0
> +       vpxorq  %VMM_128(0), %VMM_128(0), %VMM_128(0)
>         andl    $(PAGE_SIZE - 1), %eax
>         cmpl    $(PAGE_SIZE - VEC_SIZE), %eax
>         ja      L(page_cross)
>
>         /* Compare [w]char for null, mask bit will be set for match.  */
> -       VPCMP   $0, (%rdi), %VMM0, %k0
> -       KMOV    %k0, %RAX
> -       test    %RAX, %RAX
> +       VPCMP   $0, (%rdi), %VMM(0), %k0
> +       KMOV    %k0, %VRAX
> +       test    %VRAX, %VRAX
>         jz      L(align_more)
>
> -       bsf     %RAX, %RAX
> +       bsf     %VRAX, %VRAX
>  # ifdef USE_AS_STRNLEN
>         cmpq    %rsi, %rax
>         cmovnb  %rsi, %rax
> @@ -120,7 +88,7 @@ L(align_more):
>         movq    %rax, %rdx
>         subq    %rdi, %rdx
>  #  ifdef USE_AS_WCSLEN
> -       SHR     $2, %RDX
> +       shr     $2, %VRDX
>  #  endif
>         /* At this point rdx contains [w]chars already compared.  */
>         subq    %rsi, %rdx
> @@ -131,9 +99,9 @@ L(align_more):
>  # endif
>
>         /* Loop unroll 4 times for 4 vector loop.  */
> -       VPCMP   $0, (%rax), %VMM0, %k0
> -       KMOV    %k0, %RCX
> -       test    %RCX, %RCX
> +       VPCMP   $0, (%rax), %VMM(0), %k0
> +       KMOV    %k0, %VRCX
> +       test    %VRCX, %VRCX
>         jnz     L(ret_vec_x1)
>
>  # ifdef USE_AS_STRNLEN
> @@ -141,9 +109,9 @@ L(align_more):
>         jbe     L(ret_max)
>  # endif
>
> -       VPCMP   $0, VEC_SIZE(%rax), %VMM0, %k0
> -       KMOV    %k0, %RCX
> -       test    %RCX, %RCX
> +       VPCMP   $0, VEC_SIZE(%rax), %VMM(0), %k0
> +       KMOV    %k0, %VRCX
> +       test    %VRCX, %VRCX
>         jnz     L(ret_vec_x2)
>
>  # ifdef USE_AS_STRNLEN
> @@ -151,9 +119,9 @@ L(align_more):
>         jbe     L(ret_max)
>  # endif
>
> -       VPCMP   $0, (VEC_SIZE * 2)(%rax), %VMM0, %k0
> -       KMOV    %k0, %RCX
> -       test    %RCX, %RCX
> +       VPCMP   $0, (VEC_SIZE * 2)(%rax), %VMM(0), %k0
> +       KMOV    %k0, %VRCX
> +       test    %VRCX, %VRCX
>         jnz     L(ret_vec_x3)
>
>  # ifdef USE_AS_STRNLEN
> @@ -161,9 +129,9 @@ L(align_more):
>         jbe     L(ret_max)
>  # endif
>
> -       VPCMP   $0, (VEC_SIZE * 3)(%rax), %VMM0, %k0
> -       KMOV    %k0, %RCX
> -       test    %RCX, %RCX
> +       VPCMP   $0, (VEC_SIZE * 3)(%rax), %VMM(0), %k0
> +       KMOV    %k0, %VRCX
> +       test    %VRCX, %VRCX
>         jnz     L(ret_vec_x4)
>
>  # ifdef USE_AS_STRNLEN
> @@ -179,7 +147,7 @@ L(align_more):
>  # ifdef USE_AS_STRNLEN
>         subq    %rax, %rcx
>  #  ifdef USE_AS_WCSLEN
> -       SHR     $2, %RCX
> +       shr     $2, %VRCX
>  #  endif
>         /* rcx contains number of [w]char will be recompared due to
>            alignment fixes.  rdx must be incremented by rcx to offset
> @@ -199,42 +167,42 @@ L(loop_entry):
>  # endif
>         /* VPMINU and VPCMP combination provide better performance as
>            compared to alternative combinations.  */
> -       VMOVA   (VEC_SIZE * 4)(%rax), %VMM1
> -       VPMINU  (VEC_SIZE * 5)(%rax), %VMM1, %VMM2
> -       VMOVA   (VEC_SIZE * 6)(%rax), %VMM3
> -       VPMINU  (VEC_SIZE * 7)(%rax), %VMM3, %VMM4
> +       VMOVA   (VEC_SIZE * 4)(%rax), %VMM(1)
> +       VPMINU  (VEC_SIZE * 5)(%rax), %VMM(1), %VMM(2)
> +       VMOVA   (VEC_SIZE * 6)(%rax), %VMM(3)
> +       VPMINU  (VEC_SIZE * 7)(%rax), %VMM(3), %VMM(4)
>
> -       VPTESTN %VMM2, %VMM2, %k0
> -       VPTESTN %VMM4, %VMM4, %k1
> +       VPTESTN %VMM(2), %VMM(2), %k0
> +       VPTESTN %VMM(4), %VMM(4), %k1
>
>         subq    $-(VEC_SIZE * 4), %rax
>         KORTEST %k0, %k1
>         jz      L(loop)
>
> -       VPTESTN %VMM1, %VMM1, %k2
> -       KMOV    %k2, %RCX
> -       test    %RCX, %RCX
> +       VPTESTN %VMM(1), %VMM(1), %k2
> +       KMOV    %k2, %VRCX
> +       test    %VRCX, %VRCX
>         jnz     L(ret_vec_x1)
>
> -       KMOV    %k0, %RCX
> +       KMOV    %k0, %VRCX
>         /* At this point, if k0 is non zero, null char must be in the
>            second vector.  */
> -       test    %RCX, %RCX
> +       test    %VRCX, %VRCX
>         jnz     L(ret_vec_x2)
>
> -       VPTESTN %VMM3, %VMM3, %k3
> -       KMOV    %k3, %RCX
> -       test    %RCX, %RCX
> +       VPTESTN %VMM(3), %VMM(3), %k3
> +       KMOV    %k3, %VRCX
> +       test    %VRCX, %VRCX
>         jnz     L(ret_vec_x3)
>         /* At this point null [w]char must be in the fourth vector so no
>            need to check.  */
> -       KMOV    %k1, %RCX
> +       KMOV    %k1, %VRCX
>
>         /* Fourth, third, second vector terminating are pretty much
>            same, implemented this way to avoid branching and reuse code
>            from pre loop exit condition.  */
>  L(ret_vec_x4):
> -       bsf     %RCX, %RCX
> +       bsf     %VRCX, %VRCX
>         subq    %rdi, %rax
>  # ifdef USE_AS_WCSLEN
>         subq    $-(VEC_SIZE * 3), %rax
> @@ -250,7 +218,7 @@ L(ret_vec_x4):
>         ret
>
>  L(ret_vec_x3):
> -       bsf     %RCX, %RCX
> +       bsf     %VRCX, %VRCX
>         subq    %rdi, %rax
>  # ifdef USE_AS_WCSLEN
>         subq    $-(VEC_SIZE * 2), %rax
> @@ -268,7 +236,7 @@ L(ret_vec_x3):
>  L(ret_vec_x2):
>         subq    $-VEC_SIZE, %rax
>  L(ret_vec_x1):
> -       bsf     %RCX, %RCX
> +       bsf     %VRCX, %VRCX
>         subq    %rdi, %rax
>  # ifdef USE_AS_WCSLEN
>         shrq    $2, %rax
> @@ -289,13 +257,13 @@ L(page_cross):
>         /* ecx contains number of w[char] to be skipped as a result
>            of address alignment.  */
>         xorq    %rdi, %rax
> -       VPCMP   $0, (PAGE_SIZE - VEC_SIZE)(%rax), %VMM0, %k0
> -       KMOV    %k0, %RAX
> +       VPCMP   $0, (PAGE_SIZE - VEC_SIZE)(%rax), %VMM(0), %k0
> +       KMOV    %k0, %VRAX
>         /* Ignore number of character for alignment adjustment.  */
> -       SHR     %cl, %RAX
> +       shr     %cl, %VRAX
>         jz      L(align_more)
>
> -       bsf     %RAX, %RAX
> +       bsf     %VRAX, %VRAX
>  # ifdef USE_AS_STRNLEN
>         cmpq    %rsi, %rax
>         cmovnb  %rsi, %rax
> diff --git a/sysdeps/x86_64/multiarch/strlen-evex512.S b/sysdeps/x86_64/multiarch/strlen-evex512.S
> index 116f8981c8..10c3415c8a 100644
> --- a/sysdeps/x86_64/multiarch/strlen-evex512.S
> +++ b/sysdeps/x86_64/multiarch/strlen-evex512.S
> @@ -2,6 +2,6 @@
>  # define STRLEN                __strlen_evex512
>  #endif
>
> -#define VEC_SIZE       64
> -
> +#include "x86-evex512-vecs.h"
> +#include "reg-macros.h"
>  #include "strlen-evex-base.S"
> --
> 2.34.1
>

LGTM.

Thanks.

-- 
H.J.