This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] Expand round inline with SSE on x86_64 and i?86


On Sun, 29 Oct 2006, Roger Sayle wrote:

> 
> On Sun, 29 Oct 2006, Richard Guenther wrote:
> > 2006-08-23  Richard Guenther  <rguenther@suse.de>
> >
> > 	* config/i386/i386-protos.h (ix86_expand_round): Declare.
> > 	(ix86_expand_rounddf_32): Likewise.
> > 	* config/i386/i386.c (ix86_expand_round): New function expanding
> > 	round inline for SSE math and -fno-trapping-math.
> > 	(ix86_expand_rounddf_32): Same for DFmode on 32bit archs.
> > 	* config/i386/i386.md (rounddf2, roundsf2): New pattern expanding
> > 	round via ix86_expand_round.
> >
> > 	* gcc.target/i386/math-torture/round.c: New testcase.
> 
> This is OK for mainline.  Thanks.
> 
> I'm not sure how easy it would be to add some execution tests for these
> recent changes.  It's complicated by the fact that we need to pass flags
> to the compiler, and test for the presence of SSE at run-time, but all
> of your tests so far only confirm that we don't ICE the compiler, not
> that we get the correct result.  In theory, it might be nice to confirm
> that floor(2.6) == 2.0, and ceil(3.4) == 4.0 [but without compile-time
> rounding via real.c getting in the way].

Ok, I can try to produce something.  I have a benchmark/verification
that compares against glibc - I could take glibc results as valid and
put it in a table of expected results.

Does this sound like a plan?

Richard.

--
Richard Guenther <rguenther@suse.de>
Novell / SUSE Labs


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]