[PATCH] D16836: [CodeGenPrepare] Don't transform select instructions into branches when both of operands are cheap

[Junmo Park via llvm-commits]

flyingforyou added a comment.

Hi Gerolf.

Below is what I answered about Sanjay's opinion in http://reviews.llvm.org/D17288. I think this could be answer about your opinion also.

We already consider that we apply "select to branch opt" for which one.

  // Do we have efficient codegen support for this kind of 'selects' ?
  if (TLI->isSelectSupported(SelectKind)) {
    // We have efficient codegen support for the select instruction.
    // Check if it is profitable to keep this 'select'.
    if (!TLI->isPredictableSelectExpensive() ||
        !isFormingBranchFromSelectProfitable(TTI, SI))
      return false;
  }

We already check through `TLI->isPredictableSelectExpensive()`. This can be set in X86ISelLowering.cpp.

  // A predictable cmov does not hurt on an in-order CPU.
  // FIXME: Use a CPU attribute to trigger this, not a CPU model.
  PredictableSelectIsExpensive = !Subtarget.isAtom();

If `PredictableSelectIsExpensive ` is not true, we don't transform "select to branch" opt.

And I think we can use `getSchedModel().isOutOfOrder()` for setting the flag `PredictableSelectIsExpensive`.

Junmo.


http://reviews.llvm.org/D16836