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

[Pankaj Dwivedi via llvm-commits]

================
@@ -1523,38 +1556,104 @@ 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) {
+      for (unsigned i = Outstanding - 1; i >= Target && i != ~0u; --i) {
+        EmitWaitcnt(i);
+        Modified = true;
+      }
+    } else {
+      EmitWaitcnt(Target);
+      Modified = true;
+    }
+  };
+
   // Waits for VMcnt, LKGMcnt and/or EXPcnt are encoded together into a
   // single instruction while VScnt has its own instruction.
   if (Wait.hasWaitExceptStoreCnt()) {
-    unsigned Enc = AMDGPU::encodeWaitcnt(IV, Wait);
-    [[maybe_unused]] auto SWaitInst =
-        BuildMI(Block, It, DL, TII->get(AMDGPU::S_WAITCNT)).addImm(Enc);
-    Modified = true;
+    // If profiling expansion is enabled and we have score brackets,
+    // emit an expanded sequence
+    if (ExpandWaitcntProfiling && ScoreBrackets) {
+      if (Wait.LoadCnt != ~0u) {
----------------
PankajDwivedi-25 wrote:

addressed.


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