[PATCH] D133902: X86: Stop assigning register costs for longer encodings

[Mircea Trofin via Phabricator via llvm-commits]

mtrofin added a comment.

LGTM

In D133902#3791035 <https://reviews.llvm.org/D133902#3791035>, @MatzeB wrote:

> I guess this somewhat of an RFC, if we would accept better regalloc at slightly bigger programs: I found that dropping CostPerUse produced a small but measurable improvement on X86 regalloc.

FWIW, tried it on a server app (was curious if the "slightly larger" bit meant more iCache pressure), doesn't seem to be an issue.

> I would also love to hear if others can explain why the increased register costs for the larger encodings result in slightly smaller programs, even though I don't us prefering registers with more or less uses when ordering eviction and assignment choices (and as far as the AllocationOrder goes the register with longer encodings are still on the back anyway).




Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D133902/new/

https://reviews.llvm.org/D133902