From: David Laight <David.Laight@ACULAB.COM>
To: 'Yury Norov' <yury.norov@gmail.com>,
Andy Shevchenko <andy.shevchenko@gmail.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>,
Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
Guenter Roeck <linux@roeck-us.net>,
"Dennis Zhou" <dennis@kernel.org>,
Russell King <linux@armlinux.org.uk>,
"Catalin Marinas" <catalin.marinas@arm.com>,
Andy Shevchenko <andriy.shevchenko@linux.intel.com>,
Rasmus Villemoes <linux@rasmusvillemoes.dk>,
Alexey Klimov <aklimov@redhat.com>,
Kees Cook <keescook@chromium.org>,
Andy Whitcroft <apw@canonical.com>
Subject: RE: [PATCH v2 1/3] lib/find_bit: introduce FIND_FIRST_BIT() macro
Date: Wed, 24 Aug 2022 14:18:40 +0000 [thread overview]
Message-ID: <eed65809f99d49acbbbf3ba3ce6568a0@AcuMS.aculab.com> (raw)
In-Reply-To: <YwYlWjlWO3fFrtQp@yury-laptop>
...
> And generated code looks almost the same, except that
> on x86_64 your version is bigger. Compare before:
> 0000000000000000 <_find_first_bit>:
> 0: mov %rsi,%rax
> 3: test %rsi,%rsi
> 6: je 35 <_find_first_bit+0x35>
> 8: xor %edx,%edx
> a: jmp 19 <_find_first_bit+0x19>
> c: add $0x40,%rdx // Track bits and
> 10: add $0x8,%rdi // index separately
That add is free - happens in parallel with other instrutcions
> 14: cmp %rax,%rdx
> 17: jae 35 <_find_first_bit+0x35>
The instructions below will (probably/hopefully) be
speculatively executed in parallel with the cmp/jae above
> 19: mov (%rdi),%rcx
> 1c: test %rcx,%rcx
> 1f: je c <_find_first_bit+0xc>
> 21: tzcnt %rcx,%rcx
> 26: add %rdx,%rcx
> 29: cmp %rcx,%rax
> 2c: cmova %rcx,%rax
> 30: jmp 35 <_find_first_bit+0x35>
> 35: jmp 3a <_find_first_bit+0x3a>
> 3a: nopw 0x0(%rax,%rax,1)
>
> And after:
> 0000000000000000 <_find_first_bit>:
> 0: mov %rsi,%rax
> 3: test %rsi,%rsi
> 6: je 39 <_find_first_bit+0x39>
> 8: xor %edx,%edx
> a: jmp 15 <_find_first_bit+0x15>
> c: add $0x40,%rdx // Track bits only
> 10: cmp %rdx,%rax
> 13: jbe 39 <_find_first_bit+0x39>
> 15: mov %rdx,%rcx
> 18: shr $0x6,%rcx // But divide here
> 1c: mov (%rdi,%rcx,8),%rcx
> 20: test %rcx,%rcx
That is a long register dependency chain involving %cx.
It will limit the execution speed to (at least 6) clocks/iteration.
The older version might be 3 clocks/iteration.
So this could easily run at half the speed.
David
> 23: je c <_find_first_bit+0xc>
> 25: tzcnt %rcx,%rcx
> 2a: add %rcx,%rdx
> 2d: cmp %rdx,%rax
> 30: cmova %rdx,%rax
> 34: jmp 39 <_find_first_bit+0x39>
> 39: jmp 3e <_find_first_bit+0x3e>
> 3e: xchg %ax,%ax // Which adds 4 bytes to .text
>
> Thanks,
> Yury
>
> > > + val = (EXPRESSION); \
> > > + if (val) { \
> > > + sz = min(idx * BITS_PER_LONG + __ffs(word_op(val)), sz);\
> >
> > sz = min(idx + __ffs(...));
> >
> > > + break; \
> > > + } \
> > > + } \
> > > + \
> > > + sz; \
> > > +})
> >
> >
> > --
> > With Best Regards,
> > Andy Shevchenko
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)
next prev parent reply other threads:[~2022-08-24 14:18 UTC|newest]
Thread overview: 23+ messages / expand[flat|nested] mbox.gz Atom feed top
2022-08-24 1:26 [PATCH v2 0/4] lib: optimize find_bit() functions Yury Norov
2022-08-24 1:26 ` [PATCH v2 1/3] lib/find_bit: introduce FIND_FIRST_BIT() macro Yury Norov
2022-08-24 9:10 ` Andy Shevchenko
2022-08-24 13:19 ` Yury Norov
2022-08-24 14:18 ` David Laight [this message]
2022-08-24 17:45 ` Andy Shevchenko
2022-08-24 17:59 ` Linus Torvalds
2022-08-24 1:26 ` [PATCH v2 2/3] lib/find_bit: create find_first_zero_bit_le() Yury Norov
2022-08-24 9:22 ` Andy Shevchenko
2022-08-24 9:24 ` Andy Shevchenko
2022-08-24 13:37 ` Yury Norov
2022-08-24 17:50 ` Andy Shevchenko
2022-08-24 17:58 ` Russell King (Oracle)
2022-08-24 20:03 ` Yury Norov
2022-08-24 18:11 ` Linus Torvalds
2022-08-24 22:09 ` Yury Norov
2022-08-24 1:26 ` [PATCH v2 3/3] lib/find_bit: optimize find_next_bit() functions Yury Norov
2022-08-24 9:19 ` Andy Shevchenko
2022-08-24 13:53 ` Yury Norov
2022-08-24 17:54 ` Andy Shevchenko
2022-08-24 17:56 ` Andy Shevchenko
2022-08-24 21:27 ` Yury Norov
2022-08-24 9:00 ` [PATCH v2 0/4] lib: optimize find_bit() functions Andy Shevchenko
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=eed65809f99d49acbbbf3ba3ce6568a0@AcuMS.aculab.com \
--to=david.laight@aculab.com \
--cc=aklimov@redhat.com \
--cc=andriy.shevchenko@linux.intel.com \
--cc=andy.shevchenko@gmail.com \
--cc=apw@canonical.com \
--cc=catalin.marinas@arm.com \
--cc=dennis@kernel.org \
--cc=keescook@chromium.org \
--cc=linux-kernel@vger.kernel.org \
--cc=linux@armlinux.org.uk \
--cc=linux@rasmusvillemoes.dk \
--cc=linux@roeck-us.net \
--cc=torvalds@linux-foundation.org \
--cc=yury.norov@gmail.com \
/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 NNTP newsgroup(s).