lkml.org 
[lkml]   [2020]   [Mar]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
Subject[PATCH] x86: Alias memset to __builtin_memset.
From
> Not sure we should modify this for someone's downstream tree to work
> around an issue they introduced. If you want to file a bug
> internally, I'd be happy to comment on it.

I don't have a strong opinion on that. I don't know the policy of the
linux kernel w.r.t. this. There is an internal bug for this, where
kernel maintainers suggested I upstream this patch.

> Does __builtin_memset detect support for `rep stosb`, then patch the
> kernel to always use it or not?

__builtin_memset just allows the compiler to recognize that this has the
semantics of a memset, exactly like it happens when -freestanding is not
specified.

In terms of what compilers do when expanding the memset, it depends on
the size.
gcc or clang obviously do not generate vector code when -no-sse is
specified, as this would break promises.

clang inlines stores for small sizes and switches to memset as the size
increases: https://godbolt.org/z/_X16xt

gcc inlines stores for tiny sizes, then switches to repstos for larger
sizes: https://godbolt.org/z/m-G233 This behaviour is additionally
configurable through command line flags: https://godbolt.org/z/wsoJpQ

> Did you verify what this change does for other compilers?

Are there other compilers are used to build the kernel on x86 ?

icc does the same as gcc and clang: https://godbolt.org/z/yCju_D

> yet it doesn't use vector operations for the general case

I'm not sure how vector operations relate to freestanding, or this change.

> Adding -ffreestanding to optimize one hot memset in
> one function is using a really big hammer to solve the wrong
> problem.

-ffreestanding was not added to optimize anything. It was already there.
If anything -ffreestanding actually pessimizes a lot of the code
generation, as it prevents the compiler from recognizing the semantics
of common primitives. This is what this change is trying to fix.
Removing -ffreestanding from the options is another (easier?) way to fix
the problem. This change is a stab at accomodating both those who want
freestanding and those who want performance.

The hot memset I mentionned was just the hottest one. But as you can imagine
there are many constant-size memsets spread across many functions, some of
which are also very hot, the others constituting a long tail which is not
negligible.


\
 
 \ /
  Last update: 2020-03-25 09:00    [W:0.144 / U:28.688 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site