[llvm] [NVPTX] Use v2.u64 to load/store 128-bit values (PR #136638)

[Alex MacLean via llvm-commits]

================
@@ -5820,34 +5812,33 @@ static void ReplaceLoadVector(SDNode *N, SelectionDAG &DAG,
                                           LD->getMemOperand());
 
   SmallVector<SDValue> ScalarRes;
-  assert(NumElts <= ResVT.getVectorNumElements() &&
-         "NumElts should not increase, only decrease or stay the same.");
-  if (NumElts < ResVT.getVectorNumElements()) {
-    // If the number of elements has decreased, getVectorLoweringShape has
-    // upsized the element types
-    assert(EltVT.isVector() && EltVT.getSizeInBits() == 32 &&
-           EltVT.getVectorNumElements() <= 4 && "Unexpected upsized type.");
+  if (EltVT.isVector()) {
+    assert(EVT(EltVT.getVectorElementType()) == ResVT.getVectorElementType());
+    assert(NumElts * EltVT.getVectorNumElements() ==
+           ResVT.getVectorNumElements());
     // Generate EXTRACT_VECTOR_ELTs to split v2[i,f,bf]16/v4i8 subvectors back
     // into individual elements.
-    for (unsigned i = 0; i < NumElts; ++i) {
-      SDValue SubVector = NewLD.getValue(i);
+    for (const auto I : llvm::seq(NumElts)) {
----------------
AlexMaclean wrote:

Fixed

https://github.com/llvm/llvm-project/pull/136638