[llvm] [NVPTX] fold movs into loads and stores (PR #144581)

[Princeton Ferro via llvm-commits]

================
@@ -5047,26 +5043,229 @@ PerformFADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
   return SDValue();
 }
 
-static SDValue PerformStoreCombineHelper(SDNode *N, std::size_t Front,
-                                         std::size_t Back) {
+/// Combine extractelts into a load by increasing the number of return values.
+static SDValue
+combineUnpackingMovIntoLoad(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
+  // Don't run this optimization before the legalizer
+  if (DCI.isBeforeLegalize())
+    return SDValue();
+
+  EVT ElemVT = N->getValueType(0);
+  if (!Isv2x16VT(ElemVT))
+    return SDValue();
+
+  // Check whether all outputs are either used by an extractelt or are
+  // glue/chain nodes
+  if (!all_of(N->uses(), [&](SDUse &U) {
+        return U.getValueType() != ElemVT ||
+               (U.getUser()->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
+                // also check that the extractelt is used if this is an
+                // ISD::LOAD, otherwise it may be optimized by something else
+                (N->getOpcode() != ISD::LOAD || !U.getUser()->use_empty()));
----------------
Prince781 wrote:

Sure.

https://github.com/llvm/llvm-project/pull/144581