[llvm] [AMDGPU] Mark ASYNCMARK as meta instruction to fix hazard cycle miscounting (PR #189981)

[Sameer Sahasrabuddhe via llvm-commits]

================
@@ -3263,7 +3263,9 @@ bool SIInsertWaitcnts::insertWaitcntInBlock(MachineFunction &MF,
                                          E = Block.instr_end();
        Iter != E; ++Iter) {
     MachineInstr &Inst = *Iter;
-    if (Inst.isMetaInstruction())
+    // ASYNCMARK is meta instr but needs processing by
+    // generateWaitcntInstBefore and recordAsyncMark for vmcnt tracking.
+    if (Inst.isMetaInstruction() && Inst.getOpcode() != AMDGPU::ASYNCMARK)
----------------
ssahasra wrote:

This matches the assert above. I think it's worth capturing this as a static function with a useful name like `isIgnoredSomethingSomething()`. Then that function can be called in both places.

https://github.com/llvm/llvm-project/pull/189981