[llvm] [AMDGPU] Fold saddsat(x, -C) to ssubsat(x, C) when C is an inline constant (PR #216592)

[Madhur Kumar via llvm-commits]

================
@@ -1029,14 +1033,35 @@ class OpSelBinOpClampPat<SDPatternOperator node,
 let SubtargetPredicate = isGFX9Plus, True16Predicate = NotHasTrue16BitInsts in {
   def : OpSelBinOpClampPat<saddsat, V_ADD_I16_e64>;
   def : OpSelBinOpClampPat<ssubsat, V_SUB_I16_e64>;
+
+  let AddedComplexity = 15 in {
----------------
MadhurKumar004 wrote:

  `def : OpSelBinOpClampPat<saddsat, V_ADD_I16_e64>; ` has complexity of 21, and now defined pattern had 7, so is it better to keep it JUST above 21 by adding complexity of 15 or it should have some margin? 

https://github.com/llvm/llvm-project/pull/216592