[llvm] [Codegen][LegalizeIntegerTypes] Improve shift through stack (PR #96151)

[Matt Arsenault via llvm-commits]

================
@@ -4603,23 +4616,23 @@ void DAGTypeLegalizer::ExpandIntRes_ShiftThroughStack(SDNode *N, SDValue &Lo,
   } else {
     AdjStackPtr = DAG.getMemBasePlusOffset(
         StackPtr, DAG.getConstant(VTByteWidth, dl, PtrTy), dl);
-    ByteOffset = DAG.getNegative(ByteOffset, dl, ShAmtVT);
+    Offset = DAG.getNegative(Offset, dl, ShAmtVT);
   }
 
   // Get the pointer somewhere into the stack slot from which we need to load.
-  ByteOffset = DAG.getSExtOrTrunc(ByteOffset, dl, PtrTy);
-  AdjStackPtr = DAG.getMemBasePlusOffset(AdjStackPtr, ByteOffset, dl);
+  Offset = DAG.getSExtOrTrunc(Offset, dl, PtrTy);
+  AdjStackPtr = DAG.getMemBasePlusOffset(AdjStackPtr, Offset, dl);
 
   // And load it! While the load is not legal, legalizing it is obvious.
   SDValue Res = DAG.getLoad(
       VT, dl, Ch, AdjStackPtr,
-      MachinePointerInfo::getUnknownStack(DAG.getMachineFunction()), Align(1));
-  // We've performed the shift by a CHAR_BIT * [_ShAmt / CHAR_BIT_]
+      MachinePointerInfo::getUnknownStack(DAG.getMachineFunction()));
----------------
arsenm wrote:

I don't think that's correct. It should be the common alignment of the base alignment and the element size 

https://github.com/llvm/llvm-project/pull/96151