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

[Denys Vlasenko <vda.linux@googlemail.com>]

On Monday 20 August 2007 19:56, Jan Hubicka wrote:
> >  > > The problem is with the optimization flags: passing -Os causes the
> >  > > compiler to be stupid and not inline any memset/memcpy functions.
> >  >
> >  > you get what you ask for.. if you don't want that then don't ask for
> >  > it ;)
> >
> > Well, the compiler is really being dumb about -Os and in fact it's
> > giving bigger code, so I'm not really getting what I ask for.
> >
> > With my gcc at least (x86_64, gcc (GCC) 4.1.3 20070812 (prerelease)
> > (Ubuntu 4.1.2-15ubuntu2)) and Andi's example:
> >
> > #include <string.h>
> >
> > f(char x[6]) {
> >         memset(x, 1, 6);
> > }
> >
> > compiling with -O2 gives
> >
> > 0000000000000000 <f>:
> >    0:	c7 07 01 01 01 01    	movl   $0x1010101,(%rdi)
> >    6:	66 c7 47 04 01 01    	movw   $0x101,0x4(%rdi)
> >    c:	c3                   	retq
>
> GCC mainline (ie future GCC4.3.0) now give:
> 0000000000000000 <f>:
>    0:   b0 01                   mov    $0x1,%al
>    2:   b9 06 00 00 00          mov    $0x6,%ecx
>    7:   f3 aa                   rep stos %al,%es:(%rdi)
>    9:   c3                      retq
> That is smallest, definitly not fastest.
> GCC up to 4.3.0 won't be able to inline memset with non-0 operand...

No, it's not smallest. This one is smaller by 1 byte, maybe faster
(rep ... prefix is microcoded -> slower) and frees %ecx for other uses:

        mov    $0x01010101,%eax  # 5 bytes
        stosl                    # 1 byte
        stosw                    # 2 bytes
        retq
--
vda