Re: The speed of the compiler, was: Re: Combine four insns

[Nathan Froyd <froydnj at codesourcery dot com>]

On Tue, Aug 10, 2010 at 01:51:46PM -0600, Tom Tromey wrote:
> >>>>> "Andi" == Andi Kleen <andi@firstfloor.org> writes:
> 
> Richard> I'm sure there is a way to open-code this using integer math.
> 
> Andi> I don't think so. Take a look at what PCMPESTRI does.  There's no easy 
> Andi> replacement, even if you use all the Hacker's Delight tricks 
> Andi> (it's really a cool instruction, but also very complicated :-)
> 
> This is still pending:
> 
> http://gcc.gnu.org/ml/gcc-patches/2010-03/msg00526.html
> 
> I think any sort of hackery in this area is fine, if it speeds up the
> compiler.  All that is needed is someone with the time and motivation to
> do the testing.

FWIW, testing it on x86-64 on two source files lying about (admittedly a
rather small sample size) reduced the # of instructions executed cc1plus
-E by ~6% in one case and 0% (yes, exactly 0%) in the other.  Perhaps
there's a bug lying about (the patch wouldn't compile on a little-endian
machine, for one...).

I like Andi's PCMPESTRI idea, but that does nothing for the many people
without Intel's latest generation of chips (of which there are many),
let alone people on other platforms.

-Nathan