[PATCH] D45710: Fast Math Flag mapping into SDNode

[Sanjay Patel via Phabricator via llvm-commits]

spatel added inline comments.


================
Comment at: lib/Target/AMDGPU/SIISelLowering.cpp:6700
   if ((Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath ||
-       (N0->getFlags().hasUnsafeAlgebra() &&
-        N1->getFlags().hasUnsafeAlgebra())) &&
+       (N0->getFlags().hasAllowReassociation() &&
+        N1->getFlags().hasAllowReassociation())) &&
----------------
mcberg2017 wrote:
> rampitec wrote:
> > mcberg2017 wrote:
> > > rampitec wrote:
> > > > Production of a fused fma opcode is not tied to reassociation. The question here is legality of intermediate result rounding.
> > > Stanislav, what change if any do you request then for this code?
> > I would say AllowContract can be added here, but reassociation is not a sufficient or required condition.
> Ok, will do.
So this is an existing bug in the code? It shouldn't be checking hasUnsafeAlgebra (or Options.UnsafeFPMath). Can we add a test for that?


https://reviews.llvm.org/D45710