RedundantFlagCalculationElimination: optimize parity globally

Signed-off-by: Alyssa Rosenzweig <[email protected]>

FEXCore/Source/Interface/IR/Passes/RedundantFlagCalculationElimination.cpp

@@ -96,7 +96,7 @@ class DeadFlagCalculationEliminination final : public FEXCore::IR::Pass {
   void FoldBranch(IREmitter* IREmit, IRListView& CurrentIR, IROp_CondJump* Op, Ref CodeNode);
   CondClassType X86ToArmFloatCond(CondClassType X86);
   bool ProcessBlock(IREmitter* IREmit, IRListView& CurrentIR, Ref Block, ControlFlowGraph& CFG, bool& ReadsParity);
-  void OptimizeParity(IREmitter* IREmit, IRListView& CurrentIR);
+  void OptimizeParity(IREmitter* IREmit, IRListView& CurrentIR, ControlFlowGraph& CFG);
 };
 
 unsigned DeadFlagCalculationEliminination::FlagForReg(unsigned Reg) {
@@ -571,11 +571,31 @@ bool DeadFlagCalculationEliminination::ProcessBlock(IREmitter* IREmit, IRListVie
   return Progress;
 }
 
-void DeadFlagCalculationEliminination::OptimizeParity(IREmitter* IREmit, IRListView& CurrentIR) {
-  // Pass to determine if 8-bit parity is required.
+void DeadFlagCalculationEliminination::OptimizeParity(IREmitter* IREmit, IRListView& CurrentIR, ControlFlowGraph& CFG) {
+  // Mapping for flags inside this pass.
+  const uint8_t PARTIAL = 0;
+  const uint8_t FULL = 1;
+
+  // Initialize conservatively: all blocks need full parity. This initialization
+  // matters for proper handling of backedges.
+  for (auto [Block, BlockHeader] : CurrentIR.GetBlocks()) {
+    CFG.Get(Block)->Flags = FULL;
+  }
+
   for (auto [Block, BlockHeader] : CurrentIR.GetBlocks()) {
-    // TODO: This is local for now, but we should propagate this globally.
-    bool Full = true;
+    bool Full = false;
+    auto Predecessors = CFG.Get(Block)->Predecessors;
+
+    if (Predecessors.empty()) {
+      // Conservatively assume there was full parity before the start block
+      Full = true;
+    } else {
+      // If any predecessor needs full parity at the end, we need full parity.
+      for (auto Pred : Predecessors) {
+        Full |= (CFG.Get(Pred)->Flags == FULL);
+      }
+    }
+
     for (auto [CodeNode, IROp] : CurrentIR.GetCode(Block)) {
       if (IROp->Op == OP_STOREREGISTER) {
         auto Op = IROp->CW<IR::IROp_StoreRegister>();
@@ -611,6 +631,9 @@ void DeadFlagCalculationEliminination::OptimizeParity(IREmitter* IREmit, IRListV
         IREmit->Remove(CodeNode);
       }
     }
+
+    // Record our final state for our successors to read.
+    CFG.Get(Block)->Flags = Full ? FULL : PARTIAL;
   }
 }
 
@@ -675,7 +698,7 @@ void DeadFlagCalculationEliminination::Run(IREmitter* IREmit) {
   }
 
   if (ReadsParity) {
-    OptimizeParity(IREmit, CurrentIR);
+    OptimizeParity(IREmit, CurrentIR, CFG);
   }
 }