[NVPTX] don't erase CopyToRegs when folding movs into loads

We may still need to keep CopyToReg even after folding uses into vector
loads, since the original register may be used in other blocks.

Partially reverts 1fdbe69

Author: Prince781

llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp

@@ -5060,11 +5060,9 @@ combineUnpackingMovIntoLoad(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
           return !U.getUser()->use_empty();
         }
 
-        // Handle CopyToReg nodes that will become dead after our replacement
-        if (U.getUser()->getOpcode() == ISD::CopyToReg) {
-          DeadCopyToRegs.push_back(U.getUser());
+        // Peek through CopyToReg nodes
+        if (U.getUser()->getOpcode() == ISD::CopyToReg)
           return true;
-        }
 
         // Otherwise, this use prevents us from splitting a value.
         return false;
@@ -5132,10 +5130,6 @@ combineUnpackingMovIntoLoad(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
   for (unsigned I : seq(NewLoad->getNumValues() - NewNumOutputs))
     Results.push_back(NewLoad.getValue(NewNumOutputs + I));
 
-  // Remove dead CopyToReg nodes by folding them into the chain they reference
-  for (SDNode *CTR : DeadCopyToRegs)
-    DCI.CombineTo(CTR, CTR->getOperand(0));
-
   return DCI.DAG.getMergeValues(Results, DL);
 }
 
@@ -6544,4 +6538,4 @@ void NVPTXTargetLowering::computeKnownBitsForTargetNode(
   default:
     break;
   }
-}
+}