[llvm] SystemZ: Stop casting fp typed atomic loads in the IR (PR #90768)

[Ulrich Weigand via llvm-commits]

================
@@ -6263,8 +6283,25 @@ SystemZTargetLowering::LowerOperationWrapper(SDNode *N,
     MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
     SDValue Res = DAG.getMemIntrinsicNode(SystemZISD::ATOMIC_LOAD_128,
                                           DL, Tys, Ops, MVT::i128, MMO);
-    Results.push_back(lowerGR128ToI128(DAG, Res));
-    Results.push_back(Res.getValue(1));
+
+    EVT VT = N->getValueType(0);
+
+    if (VT == MVT::i128 || isTypeLegal(MVT::i128)) {
+      SDValue Lowered = lowerGR128ToI128(DAG, Res);
+      Results.push_back(DAG.getBitcast(VT, Lowered));
+      Results.push_back(Res.getValue(1));
+    } else {
+      // For the f128 case, after type legalization, we cannot produce a bitcast
+      // with an illegal type (i.e. i128), so introduce a stack store and reload
+      //
+      // FIXME: Really v2i64 should be legal, and should be used in place of
+      // unttyped. Then we could emit the bitcast which will potentially avoid
+      // the stack usage after combining.
----------------
uweigand wrote:

Comment should be updated now that we're no longer using the stack.

https://github.com/llvm/llvm-project/pull/90768