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

[Matt Arsenault via llvm-commits]

================
@@ -363,6 +363,10 @@ class TargetTransformInfoImplBase {
     return 0;
   }
 
+  InstructionCost getPHIScalarizationOverhead(Type *ScalarTy, VectorType *VTy) {
+    return 0;
----------------
arsenm wrote:

For i32/pointer/multiples of 2 x 16 should be 0 

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