[llvm] [AMDGPU] Downgrade cluster loads in strict mode (PR #218999)

[Stanislav Mekhanoshin via llvm-commits]

================
@@ -11792,12 +11792,31 @@ SITargetLowering::lowerStructBufferAtomicIntrin(SDValue Op, SelectionDAG &DAG,
                                  M->getMemOperand());
 }
 
+static void initializeM0ToZeroForClusterLoad(SDValue Op, SelectionDAG &DAG,
+                                             SDLoc DL) {
+  SDNode *N = Op.getNode();
+  SDValue Zero = DAG.getConstant(0, DL, MVT::i32);
+  unsigned NumOperands = N->getNumOperands();
+  if (N->getOperand(NumOperands - 1) == Zero)
+    return;
+  SmallVector<SDValue, 7> Ops(N->ops());
+  Ops[NumOperands - 1] = Zero; // M0 = 0
+  DAG.UpdateNodeOperands(N, Ops);
+}
+
 SDValue SITargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op,
                                                  SelectionDAG &DAG) const {
   unsigned IntrID = Op.getConstantOperandVal(1);
   SDLoc DL(Op);
 
   switch (IntrID) {
+  case Intrinsic::amdgcn_cluster_load_b32:
----------------
rampitec wrote:

Apparently yes. It is not implemented in downstream too, so that is a todo.

https://github.com/llvm/llvm-project/pull/218999