[clang] [llvm] [RegisterCoalescer] Improve register allocation for return values by limiting rematerialization (PR #163047)

[Alexis Engelke via llvm-commits]

================
@@ -1326,6 +1326,69 @@ bool RegisterCoalescer::reMaterializeDef(const CoalescerPair &CP,
   if (!TII->isAsCheapAsAMove(*DefMI))
     return false;
 
+  // Skip rematerialization for physical registers used as return values within
+  // the same basic block to enable better coalescing.
+  if (DstReg.isPhysical()) {
+    MachineBasicBlock *MBB = CopyMI->getParent();
+    if (DefMI->getParent() == MBB) {
+      // Check if there's already an identical instruction before CopyMI
+      // If so, allow rematerialization to avoid redundant instructions
+      bool FoundCopy = false;
+      for (MachineInstr &MI : *MBB) {
+        if (&MI == CopyMI) {
+          FoundCopy = true;
+          continue;
+        }
+
+        // Before CopyMI: check for duplicate instructions
+        if (!FoundCopy && &MI != DefMI &&
+            MI.isIdenticalTo(*DefMI, MachineInstr::IgnoreDefs)) {
----------------
aengelke wrote:

This seems like it could be a very expensive loop for large blocks.

https://github.com/llvm/llvm-project/pull/163047