[PATCH] D37467: Add a new pass to speculate around PHI nodes with constant (integer) operands when profitable.

[Roman Lebedev via Phabricator via llvm-commits]

lebedev.ri added a comment.

I guess i still didn't post a reproducer, but i've seen that the IR after this transform
undergoes LSR, and ends up having pretty negative performance effect in the end.
I'm rather unconvinced that this optimization holds in the IR.
We just can't be sure that in the end we will actually have the instructions we costmodel here.
I would much rather see this being a (late) codegen pass.


Repository:
  rL LLVM

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D37467/new/

https://reviews.llvm.org/D37467