Re: [PATCH v4] arch/x86: Improve 'rep movs{b|q}' usage in memmove_64.S

[Noah Goldstein <goldstein.w.n@gmail.com>]

On Fri, Nov 19, 2021 at 4:31 PM David Laight <David.Laight@aculab.com> wrote:
>
> From: Noah Goldstein
> > Sent: 17 November 2021 22:45
> >
> > On Wed, Nov 17, 2021 at 4:31 PM David Laight <David.Laight@aculab.com> wrote:
> > >
> > > From: Noah Goldstein
> > > > Sent: 17 November 2021 21:03
> > > >
> > > > Add check for "short distance movsb" for forwards FSRM usage and
> > > > entirely remove backwards 'rep movsq'. Both of these usages hit "slow
> > > > modes" that are an order of magnitude slower than usual.
> > > >
> > > > 'rep movsb' has some noticeable VERY slow modes that the current
> > > > implementation is either 1) not checking for or 2) intentionally
> > > > using.
> > >
> > > How does this relate to the decision that glibc made a few years
> > > ago to use backwards 'rep movs' for non-overlapping copies?
> >
> > GLIBC doesn't use backwards `rep movs`.  Since the regions are
> > non-overlapping it just uses forward copy. Backwards `rep movs` is
> > from setting the direction flag (`std`) and is a very slow byte
> > copy. For overlapping regions where backwards copy is necessary GLIBC
> > uses 4x vec copy loop.
>
> Try to find this commit 6fb8cbcb58a29fff73eb2101b34caa19a7f88eba
>
> Or follow links from https://www.win.tue.nl/~aeb/linux/misc/gcc-semibug.html
> But I can't find the actual patch.
>
> The claims were a massive performance increase for the reverse copy.
>

I don't think that's referring to optimizations around `rep movs`. It
appears to be referring to fallout from this patch:
https://sourceware.org/git/?p=glibc.git;a=commit;h=6fb8cbcb58a29fff73eb2101b34caa19a7f88eba

which broken programs misusing `memcpy` with overlapping regions
resulting in this fix:
https://sourceware.org/git/?p=glibc.git;a=commit;h=0354e355014b7bfda32622e0255399d859862fcd

AFAICT support for ERMS was only added around:
https://sourceware.org/git/?p=glibc.git;a=commit;h=13efa86ece61bf84daca50cab30db1b0902fe2db

Either way GLIBC memcpy/memmove moment most certainly does not
use backwards `rep movs`:
https://sourceware.org/git/?p=glibc.git;a=blob;f=sysdeps/x86_64/multiarch/memmove-vec-unaligned-erms.S;hb=HEAD#l655

as it is very slow.

> The pdf from www.agner.org/optimize may well indicate why some
> copies are unexpectedly slow due to cache access aliasing.

Even in the `4k` aliasing case `rep movsb` seems to stay within a
factor of 2 of optimal whereas the `std` backwards `rep movs` loses
by a factor of 10.

Either way, `4k` aliasing detection is mostly a concern of `memcpy` as
the direction of copy for `memmove` is a correctness question, not
an optimization.


>
> I'm pretty sure that Intel cpu (possibly from Ivy bridge onwards)
> can be persuaded to copy 8 bytes/clock for in-cache data with
> a fairly simple loop that contains 2 reads (maybe misaligned)
> and two writes (so 16 bytes per iteration).
> Extra unrolling just adds extra code top and bottom.
>
> You might want a loop like:
>         1:      mov     0(%rsi, %rcx),%rax
>                 mov     8(%rsi, %rcx),%rdx
>                 mov     %rax, 0(%rdi, %rcx)
>                 mov     %rdx, 8(%rdi, %rcx)
>                 add     $16, %rcx
>                 jnz     1b
>
>         David

The backwards loop already has 4x unrolled `movq` loop.

>
> -
> Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> Registration No: 1397386 (Wales)