[llvm] [AMDGPU] GCNHazardRecognizer: refactor getWaitStatesSince (NFCI) (PR #108347)

[Carl Ritson via llvm-commits]

================
@@ -495,51 +496,104 @@ hasHazard(StateT State,
   return false;
 }
 
-// Returns a minimum wait states since \p I walking all predecessors.
-// Only scans until \p IsExpired does not return true.
-// Can only be run in a hazard recognizer mode.
-static int getWaitStatesSince(
+// Update \p WaitStates while iterating from \p I to hazard in \p MBB.
+static HazardFnResult countWaitStatesSince(
     GCNHazardRecognizer::IsHazardFn IsHazard, const MachineBasicBlock *MBB,
-    MachineBasicBlock::const_reverse_instr_iterator I, int WaitStates,
-    IsExpiredFn IsExpired, DenseSet<const MachineBasicBlock *> &Visited,
-    GetNumWaitStatesFn GetNumWaitStates = SIInstrInfo::getNumWaitStates) {
+    MachineBasicBlock::const_reverse_instr_iterator I, int &WaitStates,
+    IsExpiredFn IsExpired, GetNumWaitStatesFn GetNumWaitStates) {
   for (auto E = MBB->instr_rend(); I != E; ++I) {
     // Don't add WaitStates for parent BUNDLE instructions.
     if (I->isBundle())
       continue;
 
     if (IsHazard(*I))
-      return WaitStates;
+      return HazardFound;
 
     if (I->isInlineAsm())
       continue;
 
     WaitStates += GetNumWaitStates(*I);
 
     if (IsExpired(*I, WaitStates))
-      return std::numeric_limits<int>::max();
+      return HazardExpired;
+  }
+
+  return NoHazardFound;
+}
+
+// Implements predecessor search for getWaitStatesSince.
+static int getWaitStatesSinceImpl(
+    GCNHazardRecognizer::IsHazardFn IsHazard,
+    const MachineBasicBlock *InitialMBB, int InitialWaitStates,
+    IsExpiredFn IsExpired,
+    GetNumWaitStatesFn GetNumWaitStates = SIInstrInfo::getNumWaitStates) {
+  DenseMap<const MachineBasicBlock *, int> Visited;
+
+  // Build worklist of predecessors.
+  // Note: use queue so search is breadth first, which reduces search space
+  // when a hazard is found.
+  std::queue<const MachineBasicBlock *> Worklist;
----------------
perlfu wrote:

Yes, testing agrees using SmallVector is a bit faster (probably more so in common cases).
However, this method does mean the worklist will essentially grow to compass the entire search.
I have added some code to reclaim space when the list is (very) large.

https://github.com/llvm/llvm-project/pull/108347