[PATCH] D28219: [X86] Attempt to pre-truncate arithmetic operations if useful

[Elena Demikhovsky via Phabricator via llvm-commits]

delena added a comment.

One some targets pmulld is significantly slower than pmullw (FeatureSlowPMULLD). I assume that truncating even different (variable) inputs should be profitable in this case.

And your transformation may be profitable for this scenario:

trunc(binop (s/zext(x), s/zext(y)) when x and y are different variables.



================
Comment at: lib/Target/X86/X86ISelLowering.cpp:31795
+  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+  unsigned Opcode = Src.getOpcode();
+  EVT SrcVT = Src.getValueType();
----------------
Should you check hasOneUse() here ?


Repository:
  rL LLVM

https://reviews.llvm.org/D28219