[PATCH v1 18/23] x86: Optimize str{n}casecmp TOLOWER logic in strcmp-sse42.S
H.J. Lu
hjl.tools@gmail.com
Thu Mar 24 19:02:22 GMT 2022
On Wed, Mar 23, 2022 at 3:02 PM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
>
> Slightly faster method of doing TOLOWER that saves an
> instruction.
>
> Also replace the hard coded 5-byte no with .p2align 4. On builds with
> CET enabled this misaligned entry to strcasecmp.
>
> geometric_mean(N=40) of all benchmarks New / Original: .920
>
> All string/memory tests pass.
> ---
> Geomtric Mean N=40 runs; All functions page aligned
> length, align1, align2, max_char, New Time / Old Time
> 1, 1, 1, 127, 0.914
> 2, 2, 2, 127, 0.952
> 3, 3, 3, 127, 0.924
> 4, 4, 4, 127, 0.995
> 5, 5, 5, 127, 0.985
> 6, 6, 6, 127, 1.017
> 7, 7, 7, 127, 1.031
> 8, 0, 0, 127, 0.967
> 9, 1, 1, 127, 0.969
> 10, 2, 2, 127, 0.951
> 11, 3, 3, 127, 0.938
> 12, 4, 4, 127, 0.937
> 13, 5, 5, 127, 0.967
> 14, 6, 6, 127, 0.941
> 15, 7, 7, 127, 0.951
> 4, 0, 0, 127, 0.959
> 4, 0, 0, 254, 0.98
> 8, 0, 0, 254, 0.959
> 16, 0, 0, 127, 0.895
> 16, 0, 0, 254, 0.901
> 32, 0, 0, 127, 0.85
> 32, 0, 0, 254, 0.851
> 64, 0, 0, 127, 0.897
> 64, 0, 0, 254, 0.895
> 128, 0, 0, 127, 0.944
> 128, 0, 0, 254, 0.935
> 256, 0, 0, 127, 0.922
> 256, 0, 0, 254, 0.913
> 512, 0, 0, 127, 0.921
> 512, 0, 0, 254, 0.914
> 1024, 0, 0, 127, 0.845
> 1024, 0, 0, 254, 0.84
> 16, 1, 2, 127, 0.923
> 16, 2, 1, 254, 0.955
> 32, 2, 4, 127, 0.979
> 32, 4, 2, 254, 0.957
> 64, 3, 6, 127, 0.866
> 64, 6, 3, 254, 0.849
> 128, 4, 0, 127, 0.882
> 128, 0, 4, 254, 0.876
> 256, 5, 2, 127, 0.877
> 256, 2, 5, 254, 0.882
> 512, 6, 4, 127, 0.822
> 512, 4, 6, 254, 0.862
> 1024, 7, 6, 127, 0.903
> 1024, 6, 7, 254, 0.908
>
> sysdeps/x86_64/multiarch/strcmp-sse42.S | 83 +++++++++++--------------
> 1 file changed, 35 insertions(+), 48 deletions(-)
>
> diff --git a/sysdeps/x86_64/multiarch/strcmp-sse42.S b/sysdeps/x86_64/multiarch/strcmp-sse42.S
> index 580feb90e9..7805ae9d41 100644
> --- a/sysdeps/x86_64/multiarch/strcmp-sse42.S
> +++ b/sysdeps/x86_64/multiarch/strcmp-sse42.S
> @@ -88,9 +88,8 @@ ENTRY (GLABEL(__strcasecmp))
> movq __libc_tsd_LOCALE@gottpoff(%rip),%rax
> mov %fs:(%rax),%RDX_LP
>
> - // XXX 5 byte should be before the function
> - /* 5-byte NOP. */
> - .byte 0x0f,0x1f,0x44,0x00,0x00
> + /* Either 1 or 5 bytes (dependeing if CET is enabled). */
> + .p2align 4
> END (GLABEL(__strcasecmp))
> /* FALLTHROUGH to strcasecmp_l. */
> #endif
> @@ -99,9 +98,8 @@ ENTRY (GLABEL(__strncasecmp))
> movq __libc_tsd_LOCALE@gottpoff(%rip),%rax
> mov %fs:(%rax),%RCX_LP
>
> - // XXX 5 byte should be before the function
> - /* 5-byte NOP. */
> - .byte 0x0f,0x1f,0x44,0x00,0x00
> + /* Either 1 or 5 bytes (dependeing if CET is enabled). */
> + .p2align 4
> END (GLABEL(__strncasecmp))
> /* FALLTHROUGH to strncasecmp_l. */
> #endif
> @@ -169,27 +167,22 @@ STRCMP_SSE42:
> #if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
> .section .rodata.cst16,"aM",@progbits,16
> .align 16
> -LABEL(belowupper):
> - .quad 0x4040404040404040
> - .quad 0x4040404040404040
> -LABEL(topupper):
> -# ifdef USE_AVX
> - .quad 0x5a5a5a5a5a5a5a5a
> - .quad 0x5a5a5a5a5a5a5a5a
> -# else
> - .quad 0x5b5b5b5b5b5b5b5b
> - .quad 0x5b5b5b5b5b5b5b5b
> -# endif
> -LABEL(touppermask):
> +LABEL(lcase_min):
> + .quad 0x3f3f3f3f3f3f3f3f
> + .quad 0x3f3f3f3f3f3f3f3f
> +LABEL(lcase_max):
> + .quad 0x9999999999999999
> + .quad 0x9999999999999999
> +LABEL(case_add):
> .quad 0x2020202020202020
> .quad 0x2020202020202020
> .previous
> - movdqa LABEL(belowupper)(%rip), %xmm4
> -# define UCLOW_reg %xmm4
> - movdqa LABEL(topupper)(%rip), %xmm5
> -# define UCHIGH_reg %xmm5
> - movdqa LABEL(touppermask)(%rip), %xmm6
> -# define LCQWORD_reg %xmm6
> + movdqa LABEL(lcase_min)(%rip), %xmm4
> +# define LCASE_MIN_reg %xmm4
> + movdqa LABEL(lcase_max)(%rip), %xmm5
> +# define LCASE_MAX_reg %xmm5
> + movdqa LABEL(case_add)(%rip), %xmm6
> +# define CASE_ADD_reg %xmm6
> #endif
> cmp $0x30, %ecx
> ja LABEL(crosscache)/* rsi: 16-byte load will cross cache line */
> @@ -200,32 +193,26 @@ LABEL(touppermask):
> #if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
> # ifdef USE_AVX
> # define TOLOWER(reg1, reg2) \
> - vpcmpgtb UCLOW_reg, reg1, %xmm7; \
> - vpcmpgtb UCHIGH_reg, reg1, %xmm8; \
> - vpcmpgtb UCLOW_reg, reg2, %xmm9; \
> - vpcmpgtb UCHIGH_reg, reg2, %xmm10; \
> - vpandn %xmm7, %xmm8, %xmm8; \
> - vpandn %xmm9, %xmm10, %xmm10; \
> - vpand LCQWORD_reg, %xmm8, %xmm8; \
> - vpand LCQWORD_reg, %xmm10, %xmm10; \
> - vpor reg1, %xmm8, reg1; \
> - vpor reg2, %xmm10, reg2
> + vpaddb LCASE_MIN_reg, reg1, %xmm7; \
> + vpaddb LCASE_MIN_reg, reg2, %xmm8; \
> + vpcmpgtb LCASE_MAX_reg, %xmm7, %xmm7; \
> + vpcmpgtb LCASE_MAX_reg, %xmm8, %xmm8; \
> + vpandn CASE_ADD_reg, %xmm7, %xmm7; \
> + vpandn CASE_ADD_reg, %xmm8, %xmm8; \
> + vpaddb %xmm7, reg1, reg1; \
> + vpaddb %xmm8, reg2, reg2
> # else
> # define TOLOWER(reg1, reg2) \
> - movdqa reg1, %xmm7; \
> - movdqa UCHIGH_reg, %xmm8; \
> - movdqa reg2, %xmm9; \
> - movdqa UCHIGH_reg, %xmm10; \
> - pcmpgtb UCLOW_reg, %xmm7; \
> - pcmpgtb reg1, %xmm8; \
> - pcmpgtb UCLOW_reg, %xmm9; \
> - pcmpgtb reg2, %xmm10; \
> - pand %xmm8, %xmm7; \
> - pand %xmm10, %xmm9; \
> - pand LCQWORD_reg, %xmm7; \
> - pand LCQWORD_reg, %xmm9; \
> - por %xmm7, reg1; \
> - por %xmm9, reg2
> + movdqa LCASE_MIN_reg, %xmm7; \
> + movdqa LCASE_MIN_reg, %xmm8; \
> + paddb reg1, %xmm7; \
> + paddb reg2, %xmm8; \
> + pcmpgtb LCASE_MAX_reg, %xmm7; \
> + pcmpgtb LCASE_MAX_reg, %xmm8; \
> + pandn CASE_ADD_reg, %xmm7; \
> + pandn CASE_ADD_reg, %xmm8; \
> + paddb %xmm7, reg1; \
> + paddb %xmm8, reg2
> # endif
> TOLOWER (%xmm1, %xmm2)
> #else
> --
> 2.25.1
>
LGTM.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
Thanks.
--
H.J.
More information about the Libc-alpha
mailing list