[PATCH] D55625: [Intrinsic] Unsigned Fixed Point Multiplication Intrinsic

[Leonard Chan via Phabricator via llvm-commits]

leonardchan added inline comments.


================
Comment at: llvm/test/CodeGen/X86/umul_fix.ll:262
+
+define <4 x i32> @vec2(<4 x i32> %x, <4 x i32> %y) nounwind {
+; X64-LABEL: vec2:
----------------
RKSimon wrote:
> leonardchan wrote:
> > RKSimon wrote:
> > > Can't you keep the vector codegen from scalarizing? See many of the other TargetLowering::expand* instructions which check for legal/custom vector operations before unrolling.
> > Could you clarify what you mean by keeping codegen from scalarizing? Vectorization isn't that high of a priority for us, but it would be nice to allow for accepting and returning vectors.
> Look at VectorLegalizer::ExpandAddSubSat + TargetLowering::expandAddSubSat - that does something very similar to what I'm suggesting.
Ah I see. Updated to expand both signed and unsigned fixed point multiplication to normal multiplication for scale of 0. Also updated the signed fixed point test.


Repository:
  rL LLVM

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D55625/new/

https://reviews.llvm.org/D55625