[PATCH] D125588: [MachineCombiner] Improve MachineCombiner's cost model

[Guozhi Wei via Phabricator via llvm-commits]

Carrot created this revision.
Carrot added reviewers: dmgreen, Gerolf, fhahn, snnw.
Herald added subscribers: pengfei, hiraditya.
Herald added a project: All.
Carrot requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

This patch contains following improvements to MachineCombiner's cost model.

  1 Ignore coalescable COPY instructions in computing latency because it will be deleted after RA.
  2 When computing CycleCount, use (FirstDepth + RootLatency) instead of (RootDepth + RootLatency) to avoid double counting of instructions when there are multiple instructions in either the old or new instruction sequence.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D125588

Files:
  llvm/lib/CodeGen/MachineCombiner.cpp
  llvm/test/CodeGen/AArch64/aarch64-combine-fmul-fsub.mir
  llvm/test/CodeGen/AArch64/arm64-fma-combines.ll
  llvm/test/CodeGen/AArch64/i128-math.ll
  llvm/test/CodeGen/AArch64/machine-combiner-madd.ll
  llvm/test/CodeGen/AArch64/madd-combiner.ll
  llvm/test/CodeGen/AArch64/madd-lohi.ll
  llvm/test/CodeGen/AArch64/mul-lohi.ll
  llvm/test/CodeGen/AArch64/srem-seteq-vec-nonsplat.ll
  llvm/test/CodeGen/X86/machine-combiner-int-vec.ll
  llvm/test/CodeGen/X86/machine-combiner-int.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D125588.429369.patch
Type: text/x-patch
Size: 58125 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220513/f8dccc82/attachment-0001.bin>