[PATCH] x86: Add cmpmemsi for -minline-all-stringops

[H.J. Lu]

On Tue, May 19, 2020 at 1:48 AM Uros Bizjak <ubizjak@gmail.com> wrote:
>
> On Sun, May 17, 2020 at 7:06 PM H.J. Lu <hjl.tools@gmail.com> wrote:
> >
> > Duplicate the cmpstrn pattern for cmpmem.  The only difference is that
> > the length argument of cmpmem is guaranteed to be less than or equal to
> > lengths of 2 memory areas.  Since "repz cmpsb" can be much slower than
> > memcmp function implemented with vector instruction, see
> >
> > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=43052
> >
> > expand cmpmem to "repz cmpsb" only with -mgeneral-regs-only.
>
> If there is no benefit compared to the library implementation, then
> enable these patterns only when -minline-all-stringops is used.

Fixed.

> Eventually these should be reimplemented with SSE4 string instructions.
>
> Honza is the author of the block handling x86 system, I'll leave the
> review to him.

We used to expand memcmp to "repz cmpsb" via cmpstrnsi.  It was changed
by

commit 9b0f6f5e511ca512e4faeabc81d2fd3abad9b02f
Author: Nick Clifton <nickc@redhat.com>
Date:   Fri Aug 12 16:26:11 2011 +0000

    builtins.c (expand_builtin_memcmp): Do not use cmpstrnsi pattern.

            * builtins.c (expand_builtin_memcmp): Do not use cmpstrnsi
            pattern.
            * doc/md.texi (cmpstrn): Note that the comparison stops if both
            fetched bytes are zero.
            (cmpstr): Likewise.
            (cmpmem): Note that the comparison does not stop if both of the
            fetched bytes are zero.

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95151

is a regression.

Honza, can you take a look at this?

Thanks.

--
H.J.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-x86-Add-cmpmemsi-for-minline-all-stringops.patch
Type: text/x-patch
Size: 11702 bytes
Desc: not available
URL: <https://gcc.gnu.org/pipermail/gcc-patches/attachments/20200519/45946c26/attachment.bin>