[clang] [llvm] [AMDGPU][SIInsertWaitcnt] Implement Waitcnt Expansion for Profiling (PR #169345)

[Jay Foad via cfe-commits]

================
@@ -1682,38 +1690,109 @@ bool WaitcntGeneratorPreGFX12::applyPreexistingWaitcnt(
 /// required counters in \p Wait
 bool WaitcntGeneratorPreGFX12::createNewWaitcnt(
     MachineBasicBlock &Block, MachineBasicBlock::instr_iterator It,
-    AMDGPU::Waitcnt Wait) {
+    AMDGPU::Waitcnt Wait, WaitcntBrackets *ScoreBrackets) {
   assert(ST);
   assert(isNormalMode(MaxCounter));
 
   bool Modified = false;
   const DebugLoc &DL = Block.findDebugLoc(It);
 
+  // Helper to emit expanded waitcnt sequence for profiling.
+  // Emits waitcnts from (Outstanding-1) down to Target, or just Target if
+  // nothing to expand. The EmitWaitcnt callback emits a single waitcnt.
+  auto EmitExpandedWaitcnt = [&](unsigned Outstanding, unsigned Target,
+                                 auto EmitWaitcnt) {
+    if (Outstanding > Target) {
----------------
jayfoad wrote:

You should not need this "if". The loop should work in all cases.

https://github.com/llvm/llvm-project/pull/169345