[llvm] [SystemZ] Don't lower float/double ATOMIC_[LOAD|STORE] to [LOAD|STORE] (PR #75879)

[Ulrich Weigand via llvm-commits]

================
@@ -1742,7 +1758,35 @@ void SystemZDAGToDAGISel::Select(SDNode *Node) {
     }
     break;
   }
+
+  case ISD::ATOMIC_STORE: {
+    auto *AtomOp = cast<AtomicSDNode>(Node);
+    // Replace the atomic_store with a regular store and select it. This is
+    // ok since we know all store instructions <= 8 bytes are atomic, and the
+    // 16 byte case is already handled during lowering.
+    StoreSDNode *St = cast<StoreSDNode>(CurDAG->getTruncStore(
+         AtomOp->getChain(), SDLoc(AtomOp), AtomOp->getVal(),
+         AtomOp->getBasePtr(), AtomOp->getMemoryVT(), AtomOp->getMemOperand()));
+    assert(St->getMemOperand()->isAtomic() && "Broken MMO.");
+    SDNode *Chain = St;
+    // We have to enforce sequential consistency by performing a
+    // serialization operation after the store.
+    if (AtomOp->getSuccessOrdering() == AtomicOrdering::SequentiallyConsistent)
+      Chain = CurDAG->getMachineNode(SystemZ::Serialize, SDLoc(AtomOp),
+                                     MVT::Other, SDValue(Chain, 0));
+    ReplaceNode(Node, Chain);
+    SelectCode(St);
+    return;
   }
+  }
+
+#ifndef NDEBUG
----------------
uweigand wrote:

I don't think we need that - nodes that are invalid that way already should just not be matched, right?

https://github.com/llvm/llvm-project/pull/75879