[llvm] [AMDGPU] In VectorLegalizer::Expand, if UnrollVectorOp returns Load, … (PR #88475)

[Craig Topper via llvm-commits]

================
@@ -1159,8 +1159,12 @@ void VectorLegalizer::Expand(SDNode *Node, SmallVectorImpl<SDValue> &Results) {
   }
 
   SDValue Unrolled = DAG.UnrollVectorOp(Node);
-  for (unsigned I = 0, E = Unrolled->getNumValues(); I != E; ++I)
-    Results.push_back(Unrolled.getValue(I));
+  const LoadSDNode *Ld = dyn_cast<LoadSDNode>(Unrolled.getNode());
+  if (Ld)
+    Results.push_back(Unrolled);
+  else
+    for (unsigned I = 0, E = Unrolled->getNumValues(); I != E; ++I)
+      Results.push_back(Unrolled.getValue(I));
----------------
topperc wrote:

Well even that's not technically right. If there is a single result expected, we shouldn't call getValue we should use the SDValue directly as the result. Folding might have allowed the SDValue to point to one result of a node with multiple results where the desired result isn't result 0.

If multiple results are expected, it should be a MERGE_VALUES node which isn't subject to folding.

Something like this is probably closer to correct.

```
if (Node->getValues() == 1) {
  // Only 1 result expect, use it directly. It isn't guaranteed to be result 0.
  Results.push_back(Unrolled);
} else {
  assert(Node->getValues() > 1);
  assert(Unrolled->getOpcode() == ISD::MERGE_VALUES);
  for (unsigned I = 0, E = Node->getNumValues(); I != E; ++I)
    Results.push_back(Unrolled.getValue(I));
}
```

https://github.com/llvm/llvm-project/pull/88475