From: Noah Goldstein <goldstein.w.n@gmail.com>
To: libc-alpha@sourceware.org
Subject: [PATCH v1 17/23] x86: Optimize str{n}casecmp TOLOWER logic in strcmp.S
Date: Wed, 23 Mar 2022 16:57:36 -0500 [thread overview]
Message-ID: <20220323215734.3927131-17-goldstein.w.n@gmail.com> (raw)
In-Reply-To: <20220323215734.3927131-1-goldstein.w.n@gmail.com>
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: .894
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, 0.903
2, 2, 2, 127, 0.905
3, 3, 3, 127, 0.877
4, 4, 4, 127, 0.888
5, 5, 5, 127, 0.901
6, 6, 6, 127, 0.954
7, 7, 7, 127, 0.932
8, 0, 0, 127, 0.918
9, 1, 1, 127, 0.914
10, 2, 2, 127, 0.877
11, 3, 3, 127, 0.909
12, 4, 4, 127, 0.876
13, 5, 5, 127, 0.886
14, 6, 6, 127, 0.914
15, 7, 7, 127, 0.939
4, 0, 0, 127, 0.963
4, 0, 0, 254, 0.943
8, 0, 0, 254, 0.927
16, 0, 0, 127, 0.876
16, 0, 0, 254, 0.865
32, 0, 0, 127, 0.865
32, 0, 0, 254, 0.862
64, 0, 0, 127, 0.863
64, 0, 0, 254, 0.896
128, 0, 0, 127, 0.885
128, 0, 0, 254, 0.882
256, 0, 0, 127, 0.87
256, 0, 0, 254, 0.869
512, 0, 0, 127, 0.832
512, 0, 0, 254, 0.848
1024, 0, 0, 127, 0.835
1024, 0, 0, 254, 0.843
16, 1, 2, 127, 0.914
16, 2, 1, 254, 0.949
32, 2, 4, 127, 0.955
32, 4, 2, 254, 1.004
64, 3, 6, 127, 0.844
64, 6, 3, 254, 0.905
128, 4, 0, 127, 0.889
128, 0, 4, 254, 0.845
256, 5, 2, 127, 0.929
256, 2, 5, 254, 0.907
512, 6, 4, 127, 0.837
512, 4, 6, 254, 0.862
1024, 7, 6, 127, 0.895
1024, 6, 7, 254, 0.89
sysdeps/x86_64/strcmp.S | 64 +++++++++++++++++++----------------------
1 file changed, 29 insertions(+), 35 deletions(-)
diff --git a/sysdeps/x86_64/strcmp.S b/sysdeps/x86_64/strcmp.S
index e2ab59c555..99d8b36f1d 100644
--- a/sysdeps/x86_64/strcmp.S
+++ b/sysdeps/x86_64/strcmp.S
@@ -75,9 +75,8 @@ ENTRY2 (__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
END2 (__strcasecmp)
# ifndef NO_NOLOCALE_ALIAS
weak_alias (__strcasecmp, strcasecmp)
@@ -94,9 +93,8 @@ ENTRY2 (__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
END2 (__strncasecmp)
# ifndef NO_NOLOCALE_ALIAS
weak_alias (__strncasecmp, strncasecmp)
@@ -146,22 +144,22 @@ ENTRY (STRCMP)
#if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
.section .rodata.cst16,"aM",@progbits,16
.align 16
-.Lbelowupper:
- .quad 0x4040404040404040
- .quad 0x4040404040404040
-.Ltopupper:
- .quad 0x5b5b5b5b5b5b5b5b
- .quad 0x5b5b5b5b5b5b5b5b
-.Ltouppermask:
+.Llcase_min:
+ .quad 0x3f3f3f3f3f3f3f3f
+ .quad 0x3f3f3f3f3f3f3f3f
+.Llcase_max:
+ .quad 0x9999999999999999
+ .quad 0x9999999999999999
+.Lcase_add:
.quad 0x2020202020202020
.quad 0x2020202020202020
.previous
- movdqa .Lbelowupper(%rip), %xmm5
-# define UCLOW_reg %xmm5
- movdqa .Ltopupper(%rip), %xmm6
-# define UCHIGH_reg %xmm6
- movdqa .Ltouppermask(%rip), %xmm7
-# define LCQWORD_reg %xmm7
+ movdqa .Llcase_min(%rip), %xmm5
+# define LCASE_MIN_reg %xmm5
+ movdqa .Llcase_max(%rip), %xmm6
+# define LCASE_MAX_reg %xmm6
+ movdqa .Lcase_add(%rip), %xmm7
+# define CASE_ADD_reg %xmm7
#endif
cmp $0x30, %ecx
ja LABEL(crosscache) /* rsi: 16-byte load will cross cache line */
@@ -172,22 +170,18 @@ ENTRY (STRCMP)
movhpd 8(%rdi), %xmm1
movhpd 8(%rsi), %xmm2
#if defined USE_AS_STRCASECMP_L || defined USE_AS_STRNCASECMP_L
-# define TOLOWER(reg1, reg2) \
- movdqa reg1, %xmm8; \
- movdqa UCHIGH_reg, %xmm9; \
- movdqa reg2, %xmm10; \
- movdqa UCHIGH_reg, %xmm11; \
- pcmpgtb UCLOW_reg, %xmm8; \
- pcmpgtb reg1, %xmm9; \
- pcmpgtb UCLOW_reg, %xmm10; \
- pcmpgtb reg2, %xmm11; \
- pand %xmm9, %xmm8; \
- pand %xmm11, %xmm10; \
- pand LCQWORD_reg, %xmm8; \
- pand LCQWORD_reg, %xmm10; \
- por %xmm8, reg1; \
- por %xmm10, reg2
- TOLOWER (%xmm1, %xmm2)
+# define TOLOWER(reg1, reg2) \
+ movdqa LCASE_MIN_reg, %xmm8; \
+ movdqa LCASE_MIN_reg, %xmm9; \
+ paddb reg1, %xmm8; \
+ paddb reg2, %xmm9; \
+ pcmpgtb LCASE_MAX_reg, %xmm8; \
+ pcmpgtb LCASE_MAX_reg, %xmm9; \
+ pandn CASE_ADD_reg, %xmm8; \
+ pandn CASE_ADD_reg, %xmm9; \
+ paddb %xmm8, reg1; \
+ paddb %xmm9, reg2
+ TOLOWER (%xmm1, %xmm2)
#else
# define TOLOWER(reg1, reg2)
#endif
--
2.25.1
next prev parent reply other threads:[~2022-03-23 22:01 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 ` Noah Goldstein [this message]
2022-03-24 19:02 ` [PATCH v1 17/23] x86: Optimize str{n}casecmp TOLOWER logic in strcmp.S 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
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=20220323215734.3927131-17-goldstein.w.n@gmail.com \
--to=goldstein.w.n@gmail.com \
--cc=libc-alpha@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).