Re: [discuss] [PATCH] x86-64: memset optimization

[Stephen Hemminger <shemminger@linux-foundation.org>]

On Sun, 19 Aug 2007 20:24:24 +0200
Andi Kleen <ak@suse.de> wrote:

> 
> > I am looking at current source, built with current (non-experimental) GCC
> > from Fedora Core 7. If I dissassemble ether_setup, which is
> > 
> > void ether_setup(struct net_device *dev)
> > {
> > ...
> > 
> > 	memset(dev->broadcast, 0xFF, ETH_ALEN);
> > }
> > 
> > I see a tail recursion (jmp) to memset which is the code in arch/x86_64/lib/memset.S
> 
> That is likely gcc then deciding it can't use an inline memset for some reason.
> It does that for example if it can't figure out the alignment or similar.
> Honza (cc'ed) can probably give you more details why it happens, especially if you
> give him a preprocessed self contained test case.
> 
> A simple example like
> char x[6];
> 
> f()
> {
>         memset(x, 1, 6);
> }
> 
> gives with gcc 4.1:
> 
>         .text
>         .p2align 4,,15
> .globl f
>         .type   f, @function
> f:
> .LFB2:
>         movl    $16843009, x(%rip)
>         movw    $257, x+4(%rip)
>         ret
> .LFE2:
> 
> -Andi

The problem is with the optimization flags: passing -Os causes the compiler
to be stupid and not inline any memset/memcpy functions.


-- 
Stephen Hemminger <shemminger@linux-foundation.org>