[PATCH] D156301: [WIP] Support FAdd/FSub global atomics in AMDGPUAtomicOptimizer.

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added inline comments.


================
Comment at: llvm/lib/Target/AMDGPU/AMDGPUAtomicOptimizer.cpp:219
+    // TODO: Support for double type
+    if (!isScanStrategyIterative() || I.getType()->isDoubleTy()) {
+      return;
----------------
I think this is a bad interpretation of the strategy option. Doing nothing just because you wanted something else is worse than just using an implemented path. Also you can just implement this with dpp?


================
Comment at: llvm/lib/Target/AMDGPU/AMDGPUAtomicOptimizer.cpp:329
+    // TODO: Support for double type
+    if (!isScanStrategyIterative() || I.getType()->isDoubleTy()) {
+      return;
----------------
Doesn't consider half

Should also handle <2 x half>, but atomicrmw doesn't support vectors now (you need the intrinsics for those)


================
Comment at: llvm/lib/Target/AMDGPU/AMDGPUAtomicOptimizer.cpp:610
+    if (isAtomicFloatingPointTy) {
+      AccumulatorBitCasted = B.CreateBitCast(Accumulator, B.getInt32Ty());
+      OldValuePhiBitCasted = B.CreateBitCast(OldValuePhi, B.getInt32Ty());
----------------
You shouldn't need a cast after D147732


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D156301/new/

https://reviews.llvm.org/D156301