This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] Add FMA_EXPR, un-cse multiplications before expansion
On Thu, Oct 21, 2010 at 10:14:36PM +0000, Joseph S. Myers wrote:
> Presumably it is intended that after this patch target maintainers should
> describe fma instructions in the .md files *only* with fma RTL and not
> with RTL for a*b+c? Because FMA_EXPR will be generated at the GIMPLE
> level and, ultimately, FP_CONTRACT controls will operate at the GIMPLE
> level as well?
>
> (It might make sense to add -ffp-contract={on,off,fast} at an early stage,
> with "off" disabling the GIMPLE-level FMA formation. Implementation of
> the pragma, and distinguishing "on" (C99-conforming, contracting
> restricted to source expressions) from "fast" (no such restriction on
> contracting, the present state) could follow later.)
I would imagine there are some people in the world for whom the fma builtin was
created for in the C99 standard that don't want (a*b)+c automatically converted
into a FMA op, but do want to be able to use FMA explicitly when they need a
fused multiply/add. Right now on the powerpc, the fma builtin is NOT
predicated on whether the user turned off automatic conversion into a FMA
operation (-mno-fused-madd).
I had left the C99 standards committee before FMA came in, so I wasn't part of
the discussion for it.
--
Michael Meissner, IBM
5 Technology Place Drive, M/S 2757, Westford, MA 01886-3141, USA
meissner@linux.vnet.ibm.com