[PATCH] D63431: AMDGPU: Fold readlane/readfirstlane calls

[Nicolai Hähnle via Phabricator via llvm-commits]

nhaehnle added a comment.

Good idea, but needs a fix I think.



================
Comment at: lib/Transforms/InstCombine/InstCombineCalls.cpp:3787-3798
+    if (match(II->getArgOperand(0), m_Intrinsic<Intrinsic::amdgcn_readfirstlane>()))
+      return replaceInstUsesWith(*II, II->getArgOperand(0));
+
+    if (IID == Intrinsic::amdgcn_readfirstlane) {
+      // readfirstlane (readlane x, y) -> readlane x, y
+      if (match(II->getArgOperand(0), m_Intrinsic<Intrinsic::amdgcn_readlane>()))
+        return replaceInstUsesWith(*II, II->getArgOperand(0));
----------------
I believe all of these need to be guarded by being in the same basic block, since the user could be observing a read(first)lane inside a divergent loop.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D63431/new/

https://reviews.llvm.org/D63431