[turbofan][arm64] Replace...

[turbofan][arm64] Replace TruncateInt64ToInt32(BitcastTaggedToWordForTagAndSmiBits(Load(x))) with Load(x)

This allows arm64 to produce an extending load from ChangeInt32ToInt64(Load(x)) more frequently.
Reduces embedded code size by 0.66% for arm64.

This change gives 0.3% for Speedometer on an A55 machine.

Change-Id: Ie27a134cea3dfc8a26b87553f27ca01bf9f00f1a
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3803227Reviewed-by: Tobias Tebbi <tebbi@chromium.org>
Reviewed-by: Nico Hartmann <nicohartmann@chromium.org>
Commit-Queue: George Wort <george.wort@arm.com>
Cr-Commit-Position: refs/heads/main@{#82258}

src/compiler/machine-operator-reducer.cc

@@ -954,6 +954,36 @@ Reduction MachineOperatorReducer::ReduceTruncateInt64ToInt32(Node* node) {
   if (m.HasResolvedValue())
     return ReplaceInt32(static_cast<int32_t>(m.ResolvedValue()));
   if (m.IsChangeInt32ToInt64()) return Replace(m.node()->InputAt(0));
+  // TruncateInt64ToInt32(BitcastTaggedToWordForTagAndSmiBits(Load(x))) =>
+  // Load(x)
+  // where the new Load uses Int32 rather than the tagged representation.
+  if (m.IsBitcastTaggedToWordForTagAndSmiBits() && m.node()->UseCount() == 1) {
+    Node* input = m.node()->InputAt(0);
+    if (input->opcode() == IrOpcode::kLoad ||
+        input->opcode() == IrOpcode::kLoadImmutable) {
+      LoadRepresentation load_rep = LoadRepresentationOf(input->op());
+      if (ElementSizeLog2Of(load_rep.representation()) == 2) {
+        // Ensure that the value output of the load is only ever used by the
+        // BitcastTaggedToWordForTagAndSmiBits.
+        int value_edges = 0;
+        for (Edge edge : input->use_edges()) {
+          if (NodeProperties::IsValueEdge(edge)) ++value_edges;
+        }
+        if (value_edges == 1) {
+          // Removing the input is required as node is replaced by the Load, but
+          // is still used by the the BitcastTaggedToWordForTagAndSmiBits, so
+          // will prevent future CanCover calls being true.
+          m.node()->RemoveInput(0);
+          NodeProperties::ChangeOp(
+              input,
+              input->opcode() == IrOpcode::kLoad
+                  ? machine()->Load(LoadRepresentation::Int32())
+                  : machine()->LoadImmutable(LoadRepresentation::Int32()));
+          return Replace(input);
+        }
+      }
+    }
+  }
   return NoChange();
 }
 

test/unittests/compiler/machine-operator-reducer-unittest.cc

@@ -567,6 +567,37 @@ TEST_F(MachineOperatorReducerTest, TruncateInt64ToInt32WithConstant) {
   }
 }
 
+TEST_F(MachineOperatorReducerTest, TruncateInt64ToInt32AfterLoadAndBitcast) {
+  Node* value = Parameter(0);
+  Node* inputs[4] = {value, value, graph()->start(), graph()->start()};
+  LoadRepresentation load_reps[3] = {LoadRepresentation::AnyTagged(),
+                                     LoadRepresentation::TaggedPointer(),
+                                     LoadRepresentation::TaggedSigned()};
+  for (LoadRepresentation load_rep : load_reps) {
+    if (ElementSizeLog2Of(load_rep.representation()) != 2) continue;
+    {
+      Node* load = graph()->NewNode(machine()->Load(load_rep), 4, inputs);
+      Reduction reduction = Reduce(graph()->NewNode(
+          machine()->TruncateInt64ToInt32(),
+          graph()->NewNode(machine()->BitcastTaggedToWordForTagAndSmiBits(),
+                           load)));
+      ASSERT_TRUE(reduction.Changed());
+      EXPECT_EQ(load, reduction.replacement());
+      EXPECT_EQ(LoadRepresentationOf(load->op()), LoadRepresentation::Int32());
+    }
+    {
+      Node* load =
+          graph()->NewNode(machine()->LoadImmutable(load_rep), 2, inputs);
+      Reduction reduction = Reduce(graph()->NewNode(
+          machine()->TruncateInt64ToInt32(),
+          graph()->NewNode(machine()->BitcastTaggedToWordForTagAndSmiBits(),
+                           load)));
+      ASSERT_TRUE(reduction.Changed());
+      EXPECT_EQ(load, reduction.replacement());
+      EXPECT_EQ(LoadRepresentationOf(load->op()), LoadRepresentation::Int32());
+    }
+  }
+}
 
 // -----------------------------------------------------------------------------
 // RoundFloat64ToInt32