[PATCH] D43142: Experimental pass to convert all floating point operations to the equivalent constrained intrinsics

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added inline comments.


================
Comment at: lib/Transforms/IPO/ForceConstrainedFP.cpp:59
+
+  for (auto &F : M.functions()) {
+    for (auto &BB : F) {
----------------
Why isn't this a function pass?


================
Comment at: lib/Transforms/IPO/ForceConstrainedFP.cpp:83-84
+    return false;
+  case Instruction::FAdd:
+    IID = Intrinsic::experimental_constrained_fadd;
+    break;
----------------
I would expect we would have opcode->intrinsic and intrinsic->opcode functions in utilities, and then this pass wouldn't need to be aware of the current set of constrained intrinsics


================
Comment at: lib/Transforms/IPO/PassManagerBuilder.cpp:407-410
+  // Experimental pass to exercise the constrained FP intrinsics.
+  if (ForceConstrainedFP)
+    MPM.add(createForceConstrainedFPPass());
+
----------------
I would drop this part


https://reviews.llvm.org/D43142