From: "H.J. Lu" <hjl.tools@gmail.com>
To: Noah Goldstein <goldstein.w.n@gmail.com>
Cc: GNU C Library <libc-alpha@sourceware.org>,
"Carlos O'Donell" <carlos@systemhalted.org>
Subject: Re: [PATCH v6 8/8] x86: Shrink code size of memchr-evex.S
Date: Tue, 7 Jun 2022 11:19:32 -0700 [thread overview]
Message-ID: <CAMe9rOq5wT=mOiTtd41izDYKBvQaK+L7qiD0NZfqEVyqppM11w@mail.gmail.com> (raw)
In-Reply-To: <20220607041134.2369903-8-goldstein.w.n@gmail.com>
On Mon, Jun 6, 2022 at 9:11 PM Noah Goldstein <goldstein.w.n@gmail.com> wrote:
>
> This is not meant as a performance optimization. The previous code was
> far to liberal in aligning targets and wasted code size unnecissarily.
>
> The total code size saving is: 64 bytes
>
> There are no non-negligible changes in the benchmarks.
> Geometric Mean of all benchmarks New / Old: 1.000
>
> Full xcheck passes on x86_64.
> ---
> sysdeps/x86_64/multiarch/memchr-evex.S | 46 ++++++++++++++------------
> 1 file changed, 25 insertions(+), 21 deletions(-)
>
> diff --git a/sysdeps/x86_64/multiarch/memchr-evex.S b/sysdeps/x86_64/multiarch/memchr-evex.S
> index cfaf02907d..0fd11b7632 100644
> --- a/sysdeps/x86_64/multiarch/memchr-evex.S
> +++ b/sysdeps/x86_64/multiarch/memchr-evex.S
> @@ -88,7 +88,7 @@
> # define PAGE_SIZE 4096
>
> .section SECTION(.text),"ax",@progbits
> -ENTRY (MEMCHR)
> +ENTRY_P2ALIGN (MEMCHR, 6)
> # ifndef USE_AS_RAWMEMCHR
> /* Check for zero length. */
> test %RDX_LP, %RDX_LP
> @@ -131,22 +131,24 @@ L(zero):
> xorl %eax, %eax
> ret
>
> - .p2align 5
> + .p2align 4
> L(first_vec_x0):
> - /* Check if first match was before length. */
> - tzcntl %eax, %eax
> - xorl %ecx, %ecx
> - cmpl %eax, %edx
> - leaq (%rdi, %rax, CHAR_SIZE), %rax
> - cmovle %rcx, %rax
> + /* Check if first match was before length. NB: tzcnt has false data-
> + dependency on destination. eax already had a data-dependency on esi
> + so this should have no affect here. */
> + tzcntl %eax, %esi
> +# ifdef USE_AS_WMEMCHR
> + leaq (%rdi, %rsi, CHAR_SIZE), %rdi
> +# else
> + addq %rsi, %rdi
> +# endif
> + xorl %eax, %eax
> + cmpl %esi, %edx
> + cmovg %rdi, %rax
> ret
> -# else
> - /* NB: first_vec_x0 is 17 bytes which will leave
> - cross_page_boundary (which is relatively cold) close enough
> - to ideal alignment. So only realign L(cross_page_boundary) if
> - rawmemchr. */
> - .p2align 4
> # endif
> +
> + .p2align 4
> L(cross_page_boundary):
> /* Save pointer before aligning as its original value is
> necessary for computer return address if byte is found or
> @@ -400,10 +402,14 @@ L(last_2x_vec):
> L(zero_end):
> ret
>
> +L(set_zero_end):
> + xorl %eax, %eax
> + ret
>
> .p2align 4
> L(first_vec_x1_check):
> - tzcntl %eax, %eax
> + /* eax must be non-zero. Use bsfl to save code size. */
> + bsfl %eax, %eax
> /* Adjust length. */
> subl $-(CHAR_PER_VEC * 4), %edx
> /* Check if match within remaining length. */
> @@ -412,9 +418,6 @@ L(first_vec_x1_check):
> /* NB: Multiply bytes by CHAR_SIZE to get the wchar_t count. */
> leaq VEC_SIZE(%rdi, %rax, CHAR_SIZE), %rax
> ret
> -L(set_zero_end):
> - xorl %eax, %eax
> - ret
>
> .p2align 4
> L(loop_4x_vec_end):
> @@ -464,7 +467,7 @@ L(loop_4x_vec_end):
> # endif
> ret
>
> - .p2align 4
> + .p2align 4,, 10
> L(last_vec_x1_return):
> tzcntl %eax, %eax
> # if defined USE_AS_WMEMCHR || RET_OFFSET != 0
> @@ -496,6 +499,7 @@ L(last_vec_x3_return):
> # endif
>
> # ifndef USE_AS_RAWMEMCHR
> + .p2align 4,, 5
> L(last_4x_vec_or_less_cmpeq):
> VPCMP $0, (VEC_SIZE * 5)(%rdi), %YMMMATCH, %k0
> kmovd %k0, %eax
> @@ -546,7 +550,7 @@ L(last_4x_vec):
> # endif
> andl %ecx, %eax
> jz L(zero_end2)
> - tzcntl %eax, %eax
> + bsfl %eax, %eax
> leaq (VEC_SIZE * 4)(%rdi, %rax, CHAR_SIZE), %rax
> L(zero_end2):
> ret
> @@ -562,6 +566,6 @@ L(last_vec_x3):
> leaq (VEC_SIZE * 3)(%rdi, %rax, CHAR_SIZE), %rax
> ret
> # endif
> -
> + /* 7 bytes from next cache line. */
> END (MEMCHR)
> #endif
> --
> 2.34.1
>
LGTM.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
Thanks.
--
H.J.
next prev parent reply other threads:[~2022-06-07 18:20 UTC|newest]
Thread overview: 82+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-06-03 4:42 [PATCH v1 1/8] x86: Create header for VEC classes in x86 strings library Noah Goldstein
2022-06-03 4:42 ` [PATCH v1 2/8] x86: Add COND_VZEROUPPER that can replace vzeroupper if no `ret` Noah Goldstein
2022-06-03 20:04 ` [PATCH v2 1/8] x86: Create header for VEC classes in x86 strings library Noah Goldstein
2022-06-03 20:04 ` [PATCH v2 2/8] x86: Add COND_VZEROUPPER that can replace vzeroupper if no `ret` Noah Goldstein
2022-06-03 23:12 ` H.J. Lu
2022-06-03 23:33 ` Noah Goldstein
2022-06-03 20:04 ` [PATCH v2 3/8] Benchtests: Improve memrchr benchmarks Noah Goldstein
2022-06-03 20:04 ` [PATCH v2 4/8] x86: Optimize memrchr-sse2.S Noah Goldstein
2022-06-03 20:04 ` [PATCH v2 5/8] x86: Optimize memrchr-evex.S Noah Goldstein
2022-06-03 20:04 ` [PATCH v2 6/8] x86: Optimize memrchr-avx2.S Noah Goldstein
2022-06-03 20:04 ` [PATCH v2 7/8] x86: Shrink code size of memchr-avx2.S Noah Goldstein
2022-06-03 20:04 ` [PATCH v2 8/8] x86: Shrink code size of memchr-evex.S Noah Goldstein
2022-06-03 23:09 ` [PATCH v2 1/8] x86: Create header for VEC classes in x86 strings library H.J. Lu
2022-06-03 23:49 ` Noah Goldstein
2022-06-03 23:49 ` [PATCH v3 " Noah Goldstein
2022-06-03 23:49 ` [PATCH v3 2/8] x86: Add COND_VZEROUPPER that can replace vzeroupper if no `ret` Noah Goldstein
2022-06-06 21:30 ` H.J. Lu
2022-06-06 22:38 ` Noah Goldstein
2022-06-03 23:49 ` [PATCH v3 3/8] Benchtests: Improve memrchr benchmarks Noah Goldstein
2022-06-03 23:49 ` [PATCH v3 4/8] x86: Optimize memrchr-sse2.S Noah Goldstein
2022-06-03 23:49 ` [PATCH v3 5/8] x86: Optimize memrchr-evex.S Noah Goldstein
2022-06-03 23:49 ` [PATCH v3 6/8] x86: Optimize memrchr-avx2.S Noah Goldstein
2022-06-03 23:50 ` [PATCH v3 7/8] x86: Shrink code size of memchr-avx2.S Noah Goldstein
2022-06-03 23:50 ` [PATCH v3 8/8] x86: Shrink code size of memchr-evex.S Noah Goldstein
2022-06-06 22:37 ` [PATCH v4 1/8] x86: Create header for VEC classes in x86 strings library Noah Goldstein
2022-06-06 22:37 ` [PATCH v4 2/8] x86: Add COND_VZEROUPPER that can replace vzeroupper if no `ret` Noah Goldstein
2022-06-07 2:45 ` H.J. Lu
2022-07-14 2:12 ` Sunil Pandey
2022-06-06 22:37 ` [PATCH v4 3/8] Benchtests: Improve memrchr benchmarks Noah Goldstein
2022-06-07 2:44 ` H.J. Lu
2022-06-07 4:10 ` Noah Goldstein
2022-06-06 22:37 ` [PATCH v4 4/8] x86: Optimize memrchr-sse2.S Noah Goldstein
2022-06-06 22:37 ` [PATCH v4 5/8] x86: Optimize memrchr-evex.S Noah Goldstein
2022-06-07 2:41 ` H.J. Lu
2022-06-07 4:09 ` Noah Goldstein
2022-06-07 4:12 ` Noah Goldstein
2022-06-06 22:37 ` [PATCH v4 6/8] x86: Optimize memrchr-avx2.S Noah Goldstein
2022-06-07 2:35 ` H.J. Lu
2022-06-07 4:06 ` Noah Goldstein
2022-06-06 22:37 ` [PATCH v4 7/8] x86: Shrink code size of memchr-avx2.S Noah Goldstein
2022-06-06 22:37 ` [PATCH v4 8/8] x86: Shrink code size of memchr-evex.S Noah Goldstein
2022-06-07 4:05 ` [PATCH v5 1/8] x86: Create header for VEC classes in x86 strings library Noah Goldstein
2022-06-07 4:05 ` [PATCH v5 2/8] x86: Add COND_VZEROUPPER that can replace vzeroupper if no `ret` Noah Goldstein
2022-06-07 4:05 ` [PATCH v5 3/8] Benchtests: Improve memrchr benchmarks Noah Goldstein
2022-06-07 4:05 ` [PATCH v5 4/8] x86: Optimize memrchr-sse2.S Noah Goldstein
2022-06-07 4:05 ` [PATCH v5 5/8] x86: Optimize memrchr-evex.S Noah Goldstein
2022-06-07 4:05 ` [PATCH v5 6/8] x86: Optimize memrchr-avx2.S Noah Goldstein
2022-06-07 4:05 ` [PATCH v5 7/8] x86: Shrink code size of memchr-avx2.S Noah Goldstein
2022-06-07 4:05 ` [PATCH v5 8/8] x86: Shrink code size of memchr-evex.S Noah Goldstein
2022-06-07 4:11 ` [PATCH v6 1/8] x86: Create header for VEC classes in x86 strings library Noah Goldstein
2022-06-07 4:11 ` [PATCH v6 2/8] x86: Add COND_VZEROUPPER that can replace vzeroupper if no `ret` Noah Goldstein
2022-06-07 4:11 ` [PATCH v6 3/8] Benchtests: Improve memrchr benchmarks Noah Goldstein
2022-06-07 18:03 ` H.J. Lu
2022-06-07 4:11 ` [PATCH v6 4/8] x86: Optimize memrchr-sse2.S Noah Goldstein
2022-06-07 18:04 ` H.J. Lu
2022-07-14 2:19 ` Sunil Pandey
2022-06-07 4:11 ` [PATCH v6 5/8] x86: Optimize memrchr-evex.S Noah Goldstein
2022-06-07 18:21 ` H.J. Lu
2022-07-14 2:21 ` Sunil Pandey
2022-06-07 4:11 ` [PATCH v6 6/8] x86: Optimize memrchr-avx2.S Noah Goldstein
2022-06-07 18:17 ` H.J. Lu
2022-07-14 2:26 ` Sunil Pandey
2022-07-14 2:43 ` Noah Goldstein
2022-06-07 4:11 ` [PATCH v6 7/8] x86: Shrink code size of memchr-avx2.S Noah Goldstein
2022-06-07 18:18 ` H.J. Lu
2022-07-14 2:31 ` Sunil Pandey
2022-07-14 2:41 ` Noah Goldstein
2022-06-07 4:11 ` [PATCH v6 8/8] x86: Shrink code size of memchr-evex.S Noah Goldstein
2022-06-07 18:19 ` H.J. Lu [this message]
2022-07-14 2:32 ` Sunil Pandey
2022-06-07 18:04 ` [PATCH v6 1/8] x86: Create header for VEC classes in x86 strings library H.J. Lu
2022-07-14 2:07 ` Sunil Pandey
2022-06-03 4:42 ` [PATCH v1 3/8] Benchtests: Improve memrchr benchmarks Noah Goldstein
2022-06-03 4:42 ` [PATCH v1 4/8] x86: Optimize memrchr-sse2.S Noah Goldstein
2022-06-03 4:47 ` Noah Goldstein
2022-06-03 4:42 ` [PATCH v1 5/8] x86: Optimize memrchr-evex.S Noah Goldstein
2022-06-03 4:49 ` Noah Goldstein
2022-06-03 4:42 ` [PATCH v1 6/8] x86: Optimize memrchr-avx2.S Noah Goldstein
2022-06-03 4:50 ` Noah Goldstein
2022-06-03 4:42 ` [PATCH v1 7/8] x86: Shrink code size of memchr-avx2.S Noah Goldstein
2022-06-03 4:42 ` [PATCH v1 8/8] x86: Shrink code size of memchr-evex.S Noah Goldstein
2022-06-03 4:51 ` [PATCH v1 1/8] x86: Create header for VEC classes in x86 strings library Noah Goldstein
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='CAMe9rOq5wT=mOiTtd41izDYKBvQaK+L7qiD0NZfqEVyqppM11w@mail.gmail.com' \
--to=hjl.tools@gmail.com \
--cc=carlos@systemhalted.org \
--cc=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).