Re: [RFC] [BZ15384] Enchance finite and isfinite.

[Marc Glisse <marc dot glisse at inria dot fr>]

On Sun, 21 Apr 2013, OndÅej BÃlka wrote:

> On Sun, Apr 21, 2013 at 03:35:19PM +0200, Marc Glisse wrote:
> > On Sun, 21 Apr 2013, OndÅej BÃlka wrote:
> > > However on x64 even gcc without optimizations expands finite to inline
> > > version which is slower than my version(see benchmark).
> >
> > This seems to depend on the CPU. Here:
> > model name	: Intel(R) Core(TM)2 Duo CPU     T9600  @ 2.80GHz
> Cannot duplicate
> on Intel(R) Core(TM)2 Quad  CPU   Q9300  @ 2.50GHz
> and  Intel(R) Core(TM)2 Duo CPU     E7200  @ 2.53GHz
> Could you try to run new version again?
>
> However on AMD Phenom(tm) II X6 1090T Processor my results are below.
>
> I fixed few mistakes in benchmark, now there should be correct version.
> One problem is that we are affected by gcc bugs, particulary
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54349

Funny, when I run your example from comment #2 in that PR, -march=native 
helps. On the other hand, -march=native hurts in 
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57024#c1

> Other explanation may be due bug in gcc that aligns loops only to 8
> bytes. One implementation can get faster just because it is 16 byte
> aligned so I changed that in assembly.

That sounds like a good reason.

> current
>
> real	0m0.816s
> user	0m0.813s
> sys	0m0.000s
>
> new # from mail
>
>
> real	0m0.738s
> user	0m0.737s
> sys	0m0.000s
>
> opt # with fixed http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54349
>
>
> real	0m0.703s
> user	0m0.700s
> sys	0m0.000s
>
> nonzero #from PR
>
>
> real	0m0.826s
> user	0m0.820s
> sys	0m0.003s

Different order here:
.81
.72
.76
.64

--
Marc Glisse