[PATCH] Refactor and enhance FMA combine

[Mehdi Amini]

Hi Olivier,

Tentatively CC: Steve who might have an input on the validity of the two transforms below without fast-math.

> On Apr 2, 2015, at 12:41 PM, Olivier Sallenave <ohsallen at us.ibm.com> wrote:
> 
> Hi Mehdi,
> 
>> I’d rather see the duplicated code (the one made obsolete by a correct canonicalization) removed from your patch (i.e. do not build technical debt), and a separate commits that implement the canonicalization part.
> 
> 
> Makes perfect sense, thanks for your feedback. I was able to do the canonicalization you suggest by adding the following transforms:
> 
> (fsub (fneg A), B) -> (fneg (fadd A, B))
> (fpext (fneg x)) -> (fneg (fpext x))
> 
> The problem is that I think those should be enabled only with -enable-unsafe-fp-math. So the two FMA combines that can be removed because of the canonicalization now only happen with -enable-unsafe-fp-math, whereas they used to work with -fp-contract=fast as well... Not sure what to do here.

I overlook the fact that we have these pesky fine grain flags :(

Now for these two particular transformations, they sound OK to me even without fast-math, but I rather have a numeric expert confirming because of all the possible edge cases.


Best,

Mehdi

.