[PATCH] D156414: [AMDGPU] Break Large PHIs: Take whole PHI chains into account

[Pierre van Houtryve via Phabricator via llvm-commits]

Pierre-vh added a comment.

Note that I checked this fixes the rocRAND threefry regression, which was the last remaining one.
The threshold was found experimentally by checking rocRAND performance results. IIRC, 50% was good for MI100, but caused regressions on MI200 in a few tests. 66% seems like a good middleground that still benefits rocRAND on MI100 without hurting MI200.
I also think that threshold makes sense. We want to break all of the PHIs only when there is a very strong indication that it'll be beneficial. 50% is too much of a coinflip.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D156414/new/

https://reviews.llvm.org/D156414