[PATCH] D39352: [SimplifyCFG] Don't do if-conversion if there is a long dependence chain

[Guozhi Wei via Phabricator via llvm-commits]

Carrot added a comment.

In https://reviews.llvm.org/D39352#933324, @efriedma wrote:

> Rather than trying to prevent the flattening from happening in the first place, it might make more sense to reverse it later in the pass pipeline.  That would be more general, and avoid blocking other optimizations.  The x86 backend already does this in some limited cases; see X86CmovConversion.cpp.


Thank you for the alternative method. This optimization should be target independent. It impacts all OOO processors. I tested the same motivated test case on power8, and observed 10% performance difference.


https://reviews.llvm.org/D39352