[llvm] [AMDGPU][TTI] Add target hook for the custom instruction uniformity (PR #137639)

[Matt Arsenault via llvm-commits]

================
@@ -147,6 +147,17 @@ bool llvm::GenericUniformityAnalysisImpl<MachineSSAContext>::isDivergentUse(
   return isTemporalDivergent(*UseInstr->getParent(), *DefInstr);
 }
 
+template <>
+bool GenericUniformityAnalysisImpl<MachineSSAContext>::isOperandUniform(
+    const MachineInstr &I, InstructionUniformity IU) const {
+  switch (IU) {
+  case InstructionUniformity::EitherOfFirstTwoOp:
+    return !isDivergentUse(I.getOperand(0)) || !isDivergentUse(I.getOperand(1));
----------------
arsenm wrote:

This API will work poorly for machine instrs. This will typically be the def 

https://github.com/llvm/llvm-project/pull/137639