[PATCH] D147786: [AMDGPU] Less aggressively break large PHIs

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added inline comments.


================
Comment at: llvm/lib/Target/AMDGPU/AMDGPUCodeGenPrepare.cpp:1412-1431
+  // Check if this is a simple chain of insertelement that fills the vector. If
+  // that's the case, we can break up this PHI node profitably because the
+  // extractelement we will insert will get folded out.
+  BasicBlock *BB = IE->getParent();
+  BitVector EltsCovered(FVT->getNumElements());
+  InsertElementInst *Next = IE;
+  while (Next && !EltsCovered.all()) {
----------------
I'm worried this heuristic is too simple and doesn't really recognize canonical IR. If I run instcombine on the test cases, nearly all of them fold out the insertelement chains


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D147786/new/

https://reviews.llvm.org/D147786