[llvm] Calculate KnownBits from Metadata correctly for vector loads (PR #128908)

[Matt Arsenault via llvm-commits]

================
@@ -4084,29 +4084,32 @@ SDValue AMDGPUTargetLowering::performShlCombine(SDNode *N,
     }
   }
 
-  if (VT != MVT::i64)
+  if (VT.getScalarType() != MVT::i64)
     return SDValue();
 
   // i64 (shl x, C) -> (build_pair 0, (shl x, C -32))
 
   // On some subtargets, 64-bit shift is a quarter rate instruction. In the
   // common case, splitting this into a move and a 32-bit shift is faster and
   // the same code size.
-  EVT TargetType = VT.getHalfSizedIntegerVT(*DAG.getContext());
-  EVT TargetVecPairType = EVT::getVectorVT(*DAG.getContext(), TargetType, 2);
   KnownBits Known = DAG.computeKnownBits(RHS);
 
-  if (Known.getMinValue().getZExtValue() < TargetType.getSizeInBits())
+  EVT ElementType = VT.getScalarType();
+  EVT TargetScalarType = ElementType.getHalfSizedIntegerVT(*DAG.getContext());
+  EVT TargetType = (VT.isVector() ? VT.changeVectorElementType(TargetScalarType)
+                                  : TargetScalarType);
+
+  if (Known.getMinValue().getZExtValue() < TargetScalarType.getSizeInBits())
----------------
arsenm wrote:

The AMDGPU vector handling parts should be split into a separate PR 

https://github.com/llvm/llvm-project/pull/128908