[llvm] [AMDGPU] Examine instructions in pending queues during scheduling (PR #147653)

[Jeffrey Byrnes via llvm-commits]

================
@@ -319,17 +327,45 @@ void GCNSchedStrategy::initCandidate(SchedCandidate &Cand, SUnit *SU,
   }
 }
 
+static bool shouldCheckPending(SchedBoundary &Zone,
+                               const TargetSchedModel *SchedModel) {
+  const unsigned ReadyListLimit = 256;
+  bool HasBufferedModel =
+      SchedModel->hasInstrSchedModel() && SchedModel->getMicroOpBufferSize();
+  return ExaminePendingQueue &&
+         Zone.Available.size() + Zone.Pending.size() <= ReadyListLimit &&
+         HasBufferedModel;
+}
+
+static SUnit *pickOnlyChoice(SchedBoundary &Zone,
+                             const TargetSchedModel *SchedModel) {
+  if (!shouldCheckPending(Zone, SchedModel) || Zone.Pending.empty())
+    return Zone.pickOnlyChoice();
+  return nullptr;
----------------
jrbyrnes wrote:

We can probably pick the only pending instruction so long as we handle the cycle.

https://github.com/llvm/llvm-project/pull/147653