[llvm] [AMDGPU] Fix illegal AGPR reclassification in RewriteMFMAFormStage (PR #200972)

[via llvm-commits]

================
@@ -2263,6 +2265,25 @@ static bool isReachingDefAGPRForm(MachineInstr *RD,
   return false;
 }
 
+bool RewriteMFMAFormStage::hasUseRequiringVGPR(
+    ArrayRef<SlotIndex> Src2ReachingDefs,
+    const SmallPtrSetImpl<MachineInstr *> &RewriteSet) {
+  for (SlotIndex RDIdx : Src2ReachingDefs) {
+    const MachineInstr *RD = DAG.LIS->getInstructionFromIndex(RDIdx);
+    SmallVector<MachineOperand *, 8> ReachingUses;
+    findReachingUses(RD, DAG.LIS, ReachingUses);
----------------
xgxanq wrote:

> So this is gonna run per candidate and for each use look for reaching defs. Can the info be at least reused between heuristic init and the rewrite?

Actually, it’s not strictly necessary. Since this is a bug fix, let’s get it merged as soon as possible — this whole section will be refactored anyway later.

https://github.com/llvm/llvm-project/pull/200972