Re: [PATCH 09/10] x86-32: use SSE for atomic64_read/set if available

[Andi Kleen <andi@firstfloor.org>]

On Thu, Feb 18, 2010 at 10:53:06AM +0100, Luca Barbieri wrote:
> > You seem to have forgotten to add benchmark results that show this is
> > actually worth while? And is there really any user on 32bit
> > that needs 64bit atomic_t?
> perf is currently the main user.
> On Core2, lock cmpxchg8b takes about 24 cycles and writes the
> cacheline, while movlps takes 1 cycle.
> clts/stts probably wipes out the savings if we need to use it, but we
> can keep TS off and restore it lazily on return to userspace.

s/probably/very likely/

CR changes are slow and synchronize the CPU. The later is always slow.

It sounds like you didn't time it?

> > I'm also suspicious of your use of global register variables.
> > This means they won't be saved on entry/exit of the functions.
> > Does that really work?
> I think it does.
> The functions never change the global register variables, and thus
> they are preserved.

Sounds fragile.

It'll generate worse code because gcc can't use these registers
at all in the C code. Some gcc versions also tend to give up when they run
out of registers too badly.

> Calls are done in inline assembly, which saves the variables if they
> are actually used as parameters (the global register variables are
> only visible in a portion of the C file, of course).

So why don't you simply use normal asm inputs/outputs?

-Andi

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