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: [committed] Add FMA support for MIPS


On Sun, Nov 7, 2010 at 1:32 PM, Richard Sandiford
<rdsandiford@googlemail.com> wrote:
> "Joseph S. Myers" <joseph@codesourcery.com> writes:
>> Are you sure these are fused operations? ?The MIPS32 manual I have to hand
>> right now (MD00086 Revision 2.62) says explicitly for MADD.fmt, for
>> example: "The intermediate product is rounded according to the current
>> rounding mode in FCSR. ... ?The results and flags are as if separate
>> floating-point multiply and add instructions were executed.".
>
> Ouch. ?Serves me right for not checking. ?They were fused in the
> original R8000 implementation, but now I look at the MIPS IV spec,
> it too explicitly says that only revisions of the R8000 used the
> fused model, and that all subsequent MIPS IV implementations used
> the unfused one.
>
> R8000 isn't important enough now to support a whole different model,
> so... patch reverted and brown paper bag duly adopted. ?Thanks a lot
> for catching my carelessness before it found its way into a release.

Well, the patch is still ok if you restrict the patterns to
flag_unsafe_math_optimizations.

For the -(a*b+c) variants we could introduce different named expanders
as well btw.

Richard.

> Richard
>


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