Re: [PATCH] X86-32: Let gcc decide whether to inline memcpy was Re: New x86 warning

[Andi Kleen <andi@firstfloor.org>]

On Wed, Apr 22, 2009 at 01:56:54PM -0700, Linus Torvalds wrote:
> 
> 
> On Wed, 22 Apr 2009, Andi Kleen wrote:
> > 
> > Modern gcc (and that is all that is supported now) should be able to
> > generate this code on its own already.  So if you call __builtin_* it
> > will  just work (that is what 64bit does) without that explicit code.
> 
> Last time we tried that, it wasn't true. Gcc wouldn't inline even trivial 
> cases of constant sizes.

AFAIK it's all true on 3.2+ when it can figure out the alignment
(but some gcc versions had problems passing the alignment around e.g.
through inlining), under the assumption that out of line can do
a better job with unaligned data. That's not true with my patch,
but could be true in theory.

Quick test here:

char a[10];
char b[2];
char c[4];
char d[8];

short x;
long y;

char xyz[100];


f()
{
#define C(x) memcpy(&x, xyz, sizeof(x));
        C(x)
        C(y)
        C(a)
        C(b)
        C(c)
        C(d)
}

and everything gets inlined with gcc 3.2 which is the oldest
we still care about:

gcc version 3.2.3

        movzwl  xyz+8(%rip), %eax
        movzwl  xyz(%rip), %ecx
        movq    xyz(%rip), %rdx
        movw    %ax, a+8(%rip)
        movw    %cx, x(%rip)
        movw    %cx, b(%rip)
        movl    xyz(%rip), %eax
        movq    %rdx, y(%rip)
        movq    %rdx, a(%rip)
        movq    %rdx, d(%rip)
        movl    %eax, c(%rip)
        ret

-Andi
-- 
ak@linux.intel.com -- Speaking for myself only.