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

From: Andi Kleen
Date: Thu Apr 23 2009 - 04:04:08 EST


On Thu, Apr 23, 2009 at 08:30:53AM +0200, Ingo Molnar wrote:
>
> * Andi Kleen <andi@xxxxxxxxxxxxxx> wrote:
>
> > > > Quick test here:
> > >
> > > How about you just compile the kernel with gcc-3.2 and compare
> > > the number of calls to memcpy before-and-after instead? That's
> > > the real test.
> >
> > I waited over 10 minutes for the full vmlinux objdumps to finish.
> > sorry lost patience. If someone has a fast disassembler we can try
> > it. I'll leave them running over night, maybe there are exact
> > numbers tomorrow.
>
> Uhm, the test Linus requested is very simple, it doesnt need 'full'
> objdumps, just a plain defconfig [*] - an objdump takes less than 10
> seconds here even on an old box i tried it on.
>
> I just did this - it all took less than 5 minutes to do the whole
> test with gcc34:
>
> vmlinux.gcc34.vanilla: 679 calls to memcpy
> vmlinux.gcc34.gcc-memcpy: 1393 calls to memcpy
>
> So your patch more than doubles the number of calls to out-of-line
> memcpy on older GCC. That's not really acceptable

How do you determine it's not acceptable?

It seems not nice to me, but not a fatal problem. I think
Linus was more interested in dramatic growth, but factor 2 over
a very large code base doesn't seem to be dramatic to me, especially
since it's very likely most of the are slow path code.

> Next time you send such patches please test with older GCCs straight
> away - it's a basic act of testing when doing a patch that 'lets GCC

I tested with 3.2.3, but not with 3.4.

3.2.3 doesn't do that. AFAIK 3.2 was a pretty common compiler; at least
several SUSE releases shipped with it.

-Andi

--
ak@xxxxxxxxxxxxxxx -- Speaking for myself only.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/