diff --git a/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp b/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
index 38331b614bceb..51c56ecea2c96 100644
--- a/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
+++ b/llvm/lib/Target/AMDGPU/SIFoldOperands.cpp
@@ -712,10 +712,15 @@ bool SIFoldOperandsImpl::updateOperand(FoldCandidate &Fold) const {
           TII->getRegClass(MI->getDesc(), Fold.UseOpNo, TRI)) {
     const TargetRegisterClass *NewRC =
         TRI->getRegClassForReg(*MRI, New->getReg());
-    const TargetRegisterClass *ConstrainRC =
-        TRI->findCommonRegClass(OpRC, Old.getSubReg(), NewRC, New->getSubReg());
-    if (!ConstrainRC)
-      return false;
+
+    const TargetRegisterClass *ConstrainRC = OpRC;
+    if (New->getSubReg()) {
+      ConstrainRC =
+          TRI->getMatchingSuperRegClass(NewRC, OpRC, New->getSubReg());
+
+      if (!ConstrainRC)
+        return false;
+    }
 
     if (!MRI->constrainRegClass(New->getReg(), ConstrainRC)) {
       LLVM_DEBUG(dbgs() << "Cannot constrain " << printReg(New->getReg(), TRI)