[llvm] [AMDGPU] Avoid quadratic erase in SIInstrWorklist (PR #211801)

[Arseniy Obolenskiy via llvm-commits]

================
@@ -93,7 +93,9 @@ struct SIInstrWorklist {
 
 private:
   /// InstrList contains the MachineInstrs.
-  SetVector<MachineInstr *> InstrList;
+  SmallVector<MachineInstr *> InstrList;
+  DenseSet<MachineInstr *> InSet;
----------------
aobolensk wrote:

Conducted an experiment: median max size 3, p90 14, p99 66, worst case observed 294 (according to all LIT CodeGen tests)

https://github.com/llvm/llvm-project/pull/211801