[clang] [llvm] [AMDGPU] Introduce asyncmark/wait intrinsics (PR #180467)

[Sameer Sahasrabuddhe via cfe-commits]

================
@@ -975,6 +999,14 @@ class WaitcntBrackets {
   // Store representative LDS DMA operations. The only useful info here is
   // alias info. One store is kept per unique AAInfo.
   SmallVector<const MachineInstr *> LDSDMAStores;
+
+  // State of all counters at each async mark encountered so far.
+  SmallVector<CounterValueArray> AsyncMarks;
+  static constexpr unsigned MaxAsyncMarks = 16;
----------------
ssahasra wrote:

AsyncMarks is a FIFO. Constantly enforcing MaxAsyncMarks as a data-structure limit would require removing elements from the front. Instead we treat MaxAsyncMarks as a soft algorithmic limit, which we enforce only when it matters.

The choice of value was explained in a different place. In #181095 I have now moved that explanation to this place.

https://github.com/llvm/llvm-project/pull/180467