This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] Add i386 torture harness
On Wed, Oct 25, 2006 at 03:55:11PM +0200, Richard Guenther wrote:
> On Wed, 25 Oct 2006, Uros Bizjak wrote:
>
> > Richard Guenther wrote:
> >
> > + { -O0 -msse -mno-sse2 -mfpmath=sse,387 } \
> > + { -O0 -msse -msse2 -mfpmath=sse,387 } \
> >
> > Is there a reason to use -mfpmath=sse,387 instead of plain
> > -mfpmath=387? -mfpmath=sse,387 is considered deprecated, as the
> > register allocator can't make a clear decision when FP pseudo should
> > go into SSE and when into x87 register set. This usually results in
> > register moves between both register sets.
>
> Well, as long as it is a valid combination (like if you have a
> processor without SSE2 but only SSE and still want to use SSE math
> for floats) we should make sure we don't ICE during expansion. Which
> is the point of this torture - detect ICEs for all those weird
> combinations. But you are right that for at least x86_64 an extra
> combination with -mfpmath=387 could be useful ;)
>
> So I added
>
> { -O0 -mfpmath=387 } \
> { -O0 -mfpmath=387 -ffast-math } \
> { -O2 -mfpmath=387 } \
> { -O2 -mfpmath=387 -ffast-math } \
This (on top of the submitted patch) is ok.
> (is there a way to generate the combinations automatically with tcl?)
I'm sure there is, but as with anything with tcl it would take awhile to
get right.
Janis