[PATCH] D41908: [X86][MMX] Add support for MMX zero vector creation

[Simon Pilgrim via Phabricator via llvm-commits]

RKSimon added a comment.

In https://reviews.llvm.org/D41908#972788, @efriedma wrote:

> > Wouldn't we already need/have handling for a zeroinitializer constant of x86_mmx type?
>
> From LangRef: "There are no arrays, vectors or constants of this type."


Yes, we can't attach constants to the x86_mmx type, hence having to alias it via MMX_MOVW2D



================
Comment at: lib/Target/X86/X86InstrInfo.cpp:8964
       Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()), 8);
+    else if (Opc == X86::MMX_SET0)
+      Ty = VectorType::get(Type::getInt32Ty(MF.getFunction().getContext()), 2);
----------------
RKSimon wrote:
> craig.topper wrote:
> > Do the test cases cover this?
> No I don't think they do. Any suggestions on suitable tests? I'm a little unclear as to when this kicks in given that the zero registers are all flagged as isReMaterializable.
I still haven't been able to cause this to fire, copying the approach from fold-pcmpeqd-2.ll doesn't seem to work - maybe just drop this part or add an assertion/unreachable?


Repository:
  rL LLVM

https://reviews.llvm.org/D41908