From: Sunil Pandey <skpgkp2@gmail.com>
To: "H.J. Lu" <hjl.tools@gmail.com>,
Libc-stable Mailing List <libc-stable@sourceware.org>
Cc: Noah Goldstein <goldstein.w.n@gmail.com>,
GNU C Library <libc-alpha@sourceware.org>
Subject: Re: [PATCH v1 18/23] x86: Optimize str{n}casecmp TOLOWER logic in strcmp-sse42.S
Date: Thu, 12 May 2022 12:45:08 -0700 [thread overview]
Message-ID: <CAMAf5_dGf6F8H4OnMy1DZR4uBXxmL1cjSN_fe3=cqxOY0ViUMA@mail.gmail.com> (raw)
In-Reply-To: <CAMe9rOo-MzhNRiuyFhHpHKanbu50_OPr_Gaof9Yt16tJRwjYFA@mail.gmail.com>
On Thu, Mar 24, 2022 at 12:05 PM H.J. Lu via Libc-alpha
<libc-alpha@sourceware.org> wrote:
>
> 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.
I would like to backport this patch to release branches.
Any comments or objections?
--Sunil
next prev parent reply other threads:[~2022-05-12 19:45 UTC|newest]
Thread overview: 76+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-03-23 21:57 [PATCH v1 01/23] benchtests: Use json-lib in bench-strchr.c Noah Goldstein
2022-03-23 21:57 ` [PATCH v1 02/23] benchtests: Add random benchmark " Noah Goldstein
2022-03-24 18:44 ` H.J. Lu
2022-03-23 21:57 ` [PATCH v1 03/23] x86: Code cleanup in strchr-avx2 and comment justifying branch Noah Goldstein
2022-03-24 18:53 ` H.J. Lu
2022-03-24 19:20 ` Noah Goldstein
2022-03-24 19:36 ` H.J. Lu
2022-05-12 19:31 ` Sunil Pandey
2022-03-23 21:57 ` [PATCH v1 04/23] x86: Code cleanup in strchr-evex " Noah Goldstein
2022-03-24 18:54 ` H.J. Lu
2022-05-12 19:32 ` Sunil Pandey
2022-03-23 21:57 ` [PATCH v1 05/23] benchtests: Use json-lib in bench-strpbrk.c Noah Goldstein
2022-03-24 18:54 ` H.J. Lu
2022-03-23 21:57 ` [PATCH v1 06/23] benchtests: Use json-lib in bench-strspn.c Noah Goldstein
2022-03-24 18:54 ` H.J. Lu
2022-03-23 21:57 ` [PATCH v1 07/23] x86: Optimize strcspn and strpbrk in strcspn-c.c Noah Goldstein
2022-03-24 18:55 ` H.J. Lu
2022-05-12 19:34 ` Sunil Pandey
2022-03-23 21:57 ` [PATCH v1 08/23] x86: Optimize strspn in strspn-c.c Noah Goldstein
2022-03-24 18:56 ` H.J. Lu
2022-05-12 19:39 ` Sunil Pandey
2022-03-23 21:57 ` [PATCH v1 09/23] x86: Remove strcspn-sse2.S and use the generic implementation Noah Goldstein
2022-03-24 18:57 ` H.J. Lu
2022-05-12 19:40 ` Sunil Pandey
2022-03-23 21:57 ` [PATCH v1 10/23] x86: Remove strpbrk-sse2.S " Noah Goldstein
2022-03-24 18:57 ` H.J. Lu
2022-05-12 19:41 ` Sunil Pandey
2022-03-23 21:57 ` [PATCH v1 11/23] x86: Remove strspn-sse2.S " Noah Goldstein
2022-03-24 18:57 ` H.J. Lu
2022-05-12 19:42 ` Sunil Pandey
2022-03-23 21:57 ` [PATCH v1 12/23] x86: Fix fallback for wcsncmp_avx2 in strcmp-avx2.S [BZ #28896] Noah Goldstein
2022-03-24 18:59 ` H.J. Lu
2022-03-24 19:18 ` Noah Goldstein
2022-03-24 19:34 ` H.J. Lu
2022-03-24 19:39 ` Noah Goldstein
2022-03-24 20:50 ` [PATCH v2 12/31] " Noah Goldstein
2022-03-24 21:26 ` H.J. Lu
2022-03-24 21:43 ` Noah Goldstein
2022-03-24 21:58 ` H.J. Lu
2022-05-04 6:05 ` Sunil Pandey
2022-03-23 21:57 ` [PATCH v1 13/23] benchtests: Use json-lib in bench-strcasecmp.c Noah Goldstein
2022-03-24 19:00 ` H.J. Lu
2022-03-23 21:57 ` [PATCH v1 14/23] benchtests: Use json-lib in bench-strncasecmp.c Noah Goldstein
2022-03-24 19:00 ` H.J. Lu
2022-03-23 21:57 ` [PATCH v1 15/23] string: Expand page cross tests in test-strcasecmp.c Noah Goldstein
2022-03-24 19:01 ` H.J. Lu
2022-03-23 21:57 ` [PATCH v1 16/23] string: Expand page cross tests in test-strncasecmp.c Noah Goldstein
2022-03-24 19:01 ` H.J. Lu
2022-03-23 21:57 ` [PATCH v1 17/23] x86: Optimize str{n}casecmp TOLOWER logic in strcmp.S Noah Goldstein
2022-03-24 19:02 ` H.J. Lu
2022-05-12 19:44 ` Sunil Pandey
2022-03-23 21:57 ` [PATCH v1 18/23] x86: Optimize str{n}casecmp TOLOWER logic in strcmp-sse42.S Noah Goldstein
2022-03-24 19:02 ` H.J. Lu
2022-05-12 19:45 ` Sunil Pandey [this message]
2022-03-23 21:57 ` [PATCH v1 19/23] string: Expand page cross test cases in test-strcmp.c Noah Goldstein
2022-03-24 19:02 ` H.J. Lu
2022-03-23 21:57 ` [PATCH v1 20/23] string: Expand page cross test cases in test-strncmp.c Noah Goldstein
2022-03-24 19:02 ` H.J. Lu
2022-03-23 21:57 ` [PATCH v1 21/23] x86: Add AVX2 optimized str{n}casecmp Noah Goldstein
2022-03-24 19:03 ` H.J. Lu
2022-03-24 22:41 ` [PATCH v3 " Noah Goldstein
2022-03-24 22:41 ` [PATCH v3 22/23] x86: Add EVEX " Noah Goldstein
2022-03-24 23:56 ` [PATCH v4 21/23] x86: Add AVX2 " Noah Goldstein
2022-03-24 23:56 ` [PATCH v4 22/23] x86: Add EVEX " Noah Goldstein
2022-03-25 18:15 ` H.J. Lu
2022-03-25 18:18 ` Noah Goldstein
2022-05-12 19:47 ` Sunil Pandey
2022-05-12 19:52 ` Sunil Pandey
2022-03-25 18:14 ` [PATCH v4 21/23] x86: Add AVX2 " H.J. Lu
2022-05-12 19:52 ` Sunil Pandey
2022-03-23 21:57 ` [PATCH v1 22/23] x86: Add EVEX " Noah Goldstein
2022-03-24 19:04 ` H.J. Lu
2022-03-23 21:57 ` [PATCH v1 23/23] x86: Remove AVX str{n}casecmp Noah Goldstein
2022-03-24 19:04 ` H.J. Lu
2022-05-12 19:54 ` Sunil Pandey
2022-03-24 18:43 ` [PATCH v1 01/23] benchtests: Use json-lib in bench-strchr.c H.J. Lu
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to='CAMAf5_dGf6F8H4OnMy1DZR4uBXxmL1cjSN_fe3=cqxOY0ViUMA@mail.gmail.com' \
--to=skpgkp2@gmail.com \
--cc=goldstein.w.n@gmail.com \
--cc=hjl.tools@gmail.com \
--cc=libc-alpha@sourceware.org \
--cc=libc-stable@sourceware.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).