[llvm] [AMDGPU] Prefer vector i8s in PHI Nodes (PR #91016)

[Matt Arsenault via llvm-commits]

================
@@ -1123,6 +1137,19 @@ Value *GCNTTIImpl::rewriteIntrinsicWithAddressSpace(IntrinsicInst *II,
   }
 }
 
+InstructionCost GCNTTIImpl::getPHIScalarizationOverhead(Type *ScalarTy,
+                                                        VectorType *VTy) {
+  if (DL.getTypeSizeInBits(ScalarTy) != 8)
+    return 0;
----------------
arsenm wrote:

Specially treating i8 doesn't make sense, look for the multiples of legal type 

https://github.com/llvm/llvm-project/pull/91016