[turbofan] Improve reduction of TruncateInt64ToInt32

When TruncateInt64ToInt32 is owned by Word32Op, it can be elided
because Word32Op automatically truncate int64 to int32.

Change-Id: Ia9dd4405f2b9b28710093dbc4c0471ea58df4e12
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/4100664
Reviewed-by: Tobias Tebbi <tebbi@chromium.org>
Commit-Queue: Hao A Xu <hao.a.xu@intel.com>
Cr-Commit-Position: refs/heads/main@{#85272}

src/compiler/backend/bitcast-elider.cc

@@ -20,6 +20,33 @@ bool IsBitcast(Node* node) {
          node->opcode() == IrOpcode::kBitcastWordToTaggedSigned;
 }
 
+bool OwnedByWord32Op(Node* node) {
+  for (Node* const use : node->uses()) {
+    switch (use->opcode()) {
+      case IrOpcode::kWord32Equal:
+      case IrOpcode::kInt32LessThan:
+      case IrOpcode::kInt32LessThanOrEqual:
+      case IrOpcode::kUint32LessThan:
+      case IrOpcode::kUint32LessThanOrEqual:
+      case IrOpcode::kChangeInt32ToInt64:
+#define Word32Op(Name) case IrOpcode::k##Name:
+        MACHINE_BINOP_32_LIST(Word32Op)
+#undef Word32Op
+        break;
+      default:
+        return false;
+    }
+  }
+  return true;
+}
+
+void Replace(Node* node, Node* replacement) {
+  for (Edge edge : node->use_edges()) {
+    edge.UpdateTo(replacement);
+  }
+  node->Kill();
+}
+
 }  // namespace
 
 void BitcastElider::Enqueue(Node* node) {
@@ -28,18 +55,21 @@ void BitcastElider::Enqueue(Node* node) {
   to_visit_.push(node);
 }
 
+void BitcastElider::Revisit(Node* node) { to_visit_.push(node); }
+
 void BitcastElider::VisitNode(Node* node) {
   for (int i = 0; i < node->InputCount(); i++) {
     Node* input = node->InputAt(i);
-    bool should_replace_input = false;
-    while (IsBitcast(input)) {
-      input = input->InputAt(0);
-      should_replace_input = true;
+    if (input->opcode() == IrOpcode::kTruncateInt64ToInt32 &&
+        OwnedByWord32Op(input)) {
+      Replace(input, input->InputAt(0));
+      Revisit(node);
+    } else if (is_builtin_ && IsBitcast(input)) {
+      Replace(input, input->InputAt(0));
+      Revisit(node);
+    } else {
+      Enqueue(input);
     }
-    if (should_replace_input) {
-      node->ReplaceInput(i, input);
-    }
-    Enqueue(input);
   }
 }
 
@@ -52,8 +82,11 @@ void BitcastElider::ProcessGraph() {
   }
 }
 
-BitcastElider::BitcastElider(Zone* zone, Graph* graph)
-    : graph_(graph), to_visit_(zone), seen_(graph, 2) {}
+BitcastElider::BitcastElider(Zone* zone, Graph* graph, bool is_builtin)
+    : graph_(graph),
+      to_visit_(zone),
+      seen_(graph, 2),
+      is_builtin_(is_builtin) {}
 
 void BitcastElider::Reduce() { ProcessGraph(); }
 

src/compiler/backend/bitcast-elider.h

@@ -15,16 +15,18 @@ namespace compiler {
 
 class Graph;
 
-// Elide all the Bitcast nodes which are required by MachineGraphVerifier. This
-// avoid generating redundant move instructions in instruction selection phase.
+// Elide all the Bitcast and TruncateInt64ToInt32 nodes which are required by
+// MachineGraphVerifier. This avoid generating redundant move instructions in
+// instruction selection phase.
 class BitcastElider {
  public:
-  BitcastElider(Zone* zone, Graph* graph);
+  BitcastElider(Zone* zone, Graph* graph, bool is_builtin);
   ~BitcastElider() = default;
 
   void Reduce();
 
   void Enqueue(Node* node);
+  void Revisit(Node* node);
   void VisitNode(Node* node);
   void ProcessGraph();
 
@@ -32,6 +34,7 @@ class BitcastElider {
   Graph* const graph_;
   ZoneQueue<Node*> to_visit_;
   NodeMarker<bool> seen_;
+  bool is_builtin_;
 };
 
 }  // namespace compiler

src/compiler/machine-operator-reducer.cc

@@ -1053,7 +1053,8 @@ Reduction MachineOperatorReducer::ReduceTruncateInt64ToInt32(Node* node) {
   Int64Matcher m(node->InputAt(0));
   if (m.HasResolvedValue())
     return ReplaceInt32(static_cast<int32_t>(m.ResolvedValue()));
-  if (m.IsChangeInt32ToInt64()) return Replace(m.node()->InputAt(0));
+  if (m.IsChangeInt32ToInt64() || m.IsChangeUint32ToUint64())
+    return Replace(m.node()->InputAt(0));
   // TruncateInt64ToInt32(BitcastTaggedToWordForTagAndSmiBits(Load(x))) =>
   // Load(x)
   // where the new Load uses Int32 rather than the tagged representation.

src/compiler/pipeline.cc

@@ -2487,8 +2487,8 @@ struct InstructionSelectionPhase {
 struct BitcastElisionPhase {
   DECL_PIPELINE_PHASE_CONSTANTS(BitcastElision)
 
-  void Run(PipelineData* data, Zone* temp_zone) {
-    BitcastElider bitcast_optimizer(temp_zone, data->graph());
+  void Run(PipelineData* data, Zone* temp_zone, bool is_builtin) {
+    BitcastElider bitcast_optimizer(temp_zone, data->graph(), is_builtin);
     bitcast_optimizer.Reduce();
   }
 };
@@ -3930,9 +3930,7 @@ bool PipelineImpl::SelectInstructions(Linkage* linkage) {
                               data->debug_name(), &temp_zone);
   }
 
-  if (Builtins::IsBuiltinId(data->info()->builtin())) {
-    Run<BitcastElisionPhase>();
-  }
+  Run<BitcastElisionPhase>(Builtins::IsBuiltinId(data->info()->builtin()));
 
   data->InitializeInstructionSequence(call_descriptor);