[llvm] [AMDGPU] Enable vectorization of i8 values. (PR #134934)

[Shilei Tian via llvm-commits]

================
@@ -537,6 +538,12 @@ InstructionCost GCNTTIImpl::getArithmeticInstrCost(
 
   MVT::SimpleValueType SLT = LT.second.getScalarType().SimpleTy;
 
+  VectorType *VecTy = dyn_cast<VectorType>(Ty);
+  InstructionCost LTTypeCost = LT.first;
+  if (VecTy &&
+      VecTy->getElementType() == IntegerType::getInt8Ty(VecTy->getContext()))
+    LTTypeCost = (((LT.first - 1) / 4) + 1);
----------------
shiltian wrote:

Can you put a comment on the logic behind this?

https://github.com/llvm/llvm-project/pull/134934