On Wed, Sep 14, 2011 at 04:00:02AM -0400, David Miller wrote:
>
> Eric, this is a preliminary version of the FMA patch I've been
> working on. Just so you can see what I'm doing.
>
> First, ignore the fact that there are two configure tests for the
> presence of support for these instructions. I'm busy normalizing the
> -xarch options which binutils supports so that they are the same as
> Sun AS and therefore just one test is necessary.
>
> Second, like rs6000 the sparc negate fused multiply instructions
> negate the full result, not the multiply result. So we cannot use
> those instructions for the fnmadf4/fnmsdf4/fnmasf4/fnmssf4 patterns.
> Since rs6000 provides patterns for such negate operations (presumably
> just in case the combiner creates a match) I have done so for sparc
> as well.
>
> I was really surprised that cpu designers haven't settled on a
> consistent formula for negated fused multiply add/sub instructions.
> Ho hum...
On the powerpc, we have an issue with Spec 2006 and calculix when FMAs are
generated and -ffast-math is used, where line 307 of rubber.f is:
tt=datan2(dsqrt(1.d0-cn*cn),cn)/3.d0
The FNMSUB instruction generates a -0.0 while doing the multiply and subtract
generates +0.0. Dsqrt returns a -0.0 when given a -0.0, and datan2 (-0.0, 1.0)
returns -0.0. Note, calculix is nothing but nested FMAs, and if you disable
FMAs you get about a 10% drop in performance. I suspect that the issue may be
a powerpc backend issue where the wrong comparison is generated, but I haven't
tracked it down.
--
Michael Meissner, IBM
5 Technology Place Drive, M/S 2757, Westford, MA 01886-3141, USA
meiss...@linux.vnet.ibm.com fax +1 (978) 399-6899
