[turbofan] Compress HeapConstants in DecompressionOptimizer

We should be encountering this due to TaggedEquality.
DecompressionElimination used to take care of this, but it will not be
present in the new system.

Bug: v8:7703
Change-Id: I9fe00ee116ed1514cb4c465a8d19df6e785ef913
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/1868623Reviewed-by: Tobias Tebbi <tebbi@chromium.org>
Reviewed-by: Jakob Gruber <jgruber@chromium.org>
Commit-Queue: Santiago Aboy Solanes <solanes@chromium.org>
Cr-Commit-Position: refs/heads/master@{#64471}

src/compiler/decompression-optimizer.cc

@@ -20,15 +20,28 @@ bool IsMachineLoad(Node* const node) {
          opcode == IrOpcode::kUnalignedLoad;
 }
 
+bool IsHeapConstant(Node* const node) {
+  return node->opcode() == IrOpcode::kHeapConstant;
+}
+
+bool CanBeCompressed(Node* const node) {
+  return IsHeapConstant(node) ||
+         (IsMachineLoad(node) &&
+          CanBeTaggedPointer(
+              LoadRepresentationOf(node->op()).representation()));
+}
+
 }  // anonymous namespace
 
 DecompressionOptimizer::DecompressionOptimizer(Zone* zone, Graph* graph,
+                                               CommonOperatorBuilder* common,
                                                MachineOperatorBuilder* machine)
     : graph_(graph),
+      common_(common),
       machine_(machine),
       states_(graph, static_cast<uint32_t>(State::kNumberOfStates)),
       to_visit_(zone),
-      compressed_loads_(zone) {}
+      compressed_candidate_nodes_(zone) {}
 
 void DecompressionOptimizer::MarkNodes() {
   MaybeMarkAndQueueForRevisit(graph()->end(), State::kOnly32BitsObserved);
@@ -89,60 +102,72 @@ void DecompressionOptimizer::MaybeMarkAndQueueForRevisit(Node* const node,
     states_.Set(node, state);
     to_visit_.push_back(node);
 
-    // In the case of a TaggedPointer or TaggedAny Load that can be done in 32
-    // bits, we save it in compressed_loads_ to be changed later if necessary.
-    if (state == State::kOnly32BitsObserved && IsMachineLoad(node) &&
-        CanBeTaggedPointer(LoadRepresentationOf(node->op()).representation())) {
-      compressed_loads_.push_back(node);
+    if (state == State::kOnly32BitsObserved && CanBeCompressed(node)) {
+      compressed_candidate_nodes_.push_back(node);
     }
   }
 }
 
-void DecompressionOptimizer::ChangeLoads() {
-  for (Node* const node : compressed_loads_) {
-    // compressed_loads_ contains all the nodes that once had the
+void DecompressionOptimizer::ChangeHeapConstant(Node* const node) {
+  DCHECK(IsHeapConstant(node));
+  NodeProperties::ChangeOp(
+      node, common()->CompressedHeapConstant(HeapConstantOf(node->op())));
+}
+
+void DecompressionOptimizer::ChangeLoad(Node* const node) {
+  DCHECK(IsMachineLoad(node));
+  // Change to a Compressed MachRep to avoid the full decompression.
+  LoadRepresentation load_rep = LoadRepresentationOf(node->op());
+  LoadRepresentation compressed_load_rep;
+  if (load_rep == MachineType::AnyTagged()) {
+    compressed_load_rep = MachineType::AnyCompressed();
+  } else {
+    DCHECK_EQ(load_rep, MachineType::TaggedPointer());
+    compressed_load_rep = MachineType::CompressedPointer();
+  }
+
+  // Change to the Operator with the Compressed MachineRepresentation.
+  switch (node->opcode()) {
+    case IrOpcode::kLoad:
+      NodeProperties::ChangeOp(node, machine()->Load(compressed_load_rep));
+      break;
+    case IrOpcode::kPoisonedLoad:
+      NodeProperties::ChangeOp(node,
+                               machine()->PoisonedLoad(compressed_load_rep));
+      break;
+    case IrOpcode::kProtectedLoad:
+      NodeProperties::ChangeOp(node,
+                               machine()->ProtectedLoad(compressed_load_rep));
+      break;
+    case IrOpcode::kUnalignedLoad:
+      NodeProperties::ChangeOp(node,
+                               machine()->UnalignedLoad(compressed_load_rep));
+      break;
+    default:
+      UNREACHABLE();
+  }
+}
+
+void DecompressionOptimizer::ChangeNodes() {
+  for (Node* const node : compressed_candidate_nodes_) {
+    // compressed_candidate_nodes_ contains all the nodes that once had the
     // State::kOnly32BitsObserved. If we later updated the state to be
     // State::IsEverythingObserved, then we have to ignore them. This is less
-    // costly than removing them from the compressed_loads_ NodeVector when we
-    // update them to State::IsEverythingObserved.
+    // costly than removing them from the compressed_candidate_nodes_ NodeVector
+    // when we update them to State::IsEverythingObserved.
     if (IsEverythingObserved(node)) continue;
 
-    // Change to a Compressed MachRep to avoid the full decompression.
-    LoadRepresentation load_rep = LoadRepresentationOf(node->op());
-    LoadRepresentation compressed_load_rep;
-    if (load_rep == MachineType::AnyTagged()) {
-      compressed_load_rep = MachineType::AnyCompressed();
+    if (IsHeapConstant(node)) {
+      ChangeHeapConstant(node);
     } else {
-      DCHECK_EQ(load_rep, MachineType::TaggedPointer());
-      compressed_load_rep = MachineType::CompressedPointer();
-    }
-
-    // Change to the Operator with the Compressed MachineRepresentation.
-    switch (node->opcode()) {
-      case IrOpcode::kLoad:
-        NodeProperties::ChangeOp(node, machine()->Load(compressed_load_rep));
-        break;
-      case IrOpcode::kPoisonedLoad:
-        NodeProperties::ChangeOp(node,
-                                 machine()->PoisonedLoad(compressed_load_rep));
-        break;
-      case IrOpcode::kProtectedLoad:
-        NodeProperties::ChangeOp(node,
-                                 machine()->ProtectedLoad(compressed_load_rep));
-        break;
-      case IrOpcode::kUnalignedLoad:
-        NodeProperties::ChangeOp(node,
-                                 machine()->UnalignedLoad(compressed_load_rep));
-        break;
-      default:
-        UNREACHABLE();
+      ChangeLoad(node);
     }
   }
 }
 
 void DecompressionOptimizer::Reduce() {
   MarkNodes();
-  ChangeLoads();
+  ChangeNodes();
 }
 
 }  // namespace compiler

src/compiler/decompression-optimizer.h

@@ -5,6 +5,7 @@
 #ifndef V8_COMPILER_DECOMPRESSION_OPTIMIZER_H_
 #define V8_COMPILER_DECOMPRESSION_OPTIMIZER_H_
 
+#include "src/compiler/common-operator.h"
 #include "src/compiler/machine-operator.h"
 #include "src/compiler/node-marker.h"
 
@@ -15,9 +16,11 @@ namespace compiler {
 // Forward declare.
 class Graph;
 
-// DecompressionOptimizer purpose is to avoid the full decompression on Loads
-// whenever possible. Its scope is narrowed down to TaggedPointer and AnyTagged,
-// since TaggedSigned avoids full decompression always.
+// DecompressionOptimizer purpose is to hide the distinction between 32 bit and
+// 64 bit tagged values, while being able to use the compressed version of nodes
+// whenever possible. Its scope is narrowed down to loads of TaggedPointer and
+// AnyTagged (since TaggedSigned avoids full decompression always), and
+// HeapConstants.
 
 // DecompressionOptimizer will run only when pointer compression is enabled. For
 // the moment, it's also requires FLAG_turbo_decompression_elimination to be
@@ -35,11 +38,12 @@ class Graph;
 class V8_EXPORT_PRIVATE DecompressionOptimizer final {
  public:
   DecompressionOptimizer(Zone* zone, Graph* graph,
+                         CommonOperatorBuilder* common,
                          MachineOperatorBuilder* machine);
   ~DecompressionOptimizer() = default;
 
-  // Assign States to the nodes, and then change the loads' Operator to avoid
-  // decompression if possible.
+  // Assign States to the nodes, and then change the node's Operator to use the
+  // compressed version if possible.
   void Reduce();
 
  private:
@@ -56,9 +60,15 @@ class V8_EXPORT_PRIVATE DecompressionOptimizer final {
     kNumberOfStates
   };
 
-  // Go through the already marked nodes and changed the operation for the loads
-  // that can avoid the full decompression.
-  void ChangeLoads();
+  // Change node's op from HeapConstant to CompressedHeapConstant.
+  void ChangeHeapConstant(Node* const node);
+
+  // Change node's load into a compressed one.
+  void ChangeLoad(Node* const node);
+
+  // Go through the already marked nodes and changed the operation for the nodes
+  // that can use compressed outputs.
+  void ChangeNodes();
 
   // Goes through the nodes to mark them all as appropriate. It will visit each
   // node at most twice: only when the node was unvisited, then marked as
@@ -74,9 +84,9 @@ class V8_EXPORT_PRIVATE DecompressionOptimizer final {
   // i.e either if:
   // * We are marking an unvisited node, or
   // * We are marking a node as needing 64 bits when we previously had the
-  // information that it could output 32 bits. Also, we store the TaggedPointer
-  // and AnyTagged loads that have their state set as kOnly32BitsObserved.
-  // If the node's state changes, we queue it for revisit.
+  // information that it could output 32 bits. Also, we store the HeapConstant
+  // and TaggedPointer and AnyTagged loads that have their state set as
+  // kOnly32BitsObserved. If the node's state changes, we queue it for revisit.
   void MaybeMarkAndQueueForRevisit(Node* const node, State state);
 
   bool IsEverythingObserved(Node* const node) {
@@ -84,19 +94,21 @@ class V8_EXPORT_PRIVATE DecompressionOptimizer final {
   }
 
   Graph* graph() const { return graph_; }
+  CommonOperatorBuilder* common() const { return common_; }
   MachineOperatorBuilder* machine() const { return machine_; }
 
   Graph* const graph_;
+  CommonOperatorBuilder* const common_;
   MachineOperatorBuilder* const machine_;
   NodeMarker<State> states_;
   // to_visit_ is a Deque but it's used as if it were a Queue. The reason why we
   // are using NodeDeque is because it attempts to reuse 'freed' zone memory
   // instead of always allocating a new region.
   NodeDeque to_visit_;
-  // Contains the AnyTagged and TaggedPointer loads that can avoid the full
-  // decompression. In a way, it functions as a NodeSet since each node will be
+  // Contains the nodes that can be changed into a compressed version of
+  // themselves. In a way, it functions as a NodeSet since each node will be
   // contained at most once. It's a Vector since we care about insertion speed.
-  NodeVector compressed_loads_;
+  NodeVector compressed_candidate_nodes_;
 
   DISALLOW_COPY_AND_ASSIGN(DecompressionOptimizer);
 };

src/compiler/pipeline.cc

@@ -1799,8 +1799,8 @@ struct DecompressionOptimizationPhase {
 
   void Run(PipelineData* data, Zone* temp_zone) {
     if (COMPRESS_POINTERS_BOOL && !FLAG_turbo_decompression_elimination) {
-      DecompressionOptimizer decompression_optimizer(temp_zone, data->graph(),
-                                                     data->machine());
+      DecompressionOptimizer decompression_optimizer(
+          temp_zone, data->graph(), data->common(), data->machine());
       decompression_optimizer.Reduce();
     }
   }

test/unittests/compiler/decompression-optimizer-unittest.cc

@@ -20,7 +20,8 @@ class DecompressionOptimizerTest : public GraphTest {
 
  protected:
   void Reduce() {
-    DecompressionOptimizer decompression_optimizer(zone(), graph(), machine());
+    DecompressionOptimizer decompression_optimizer(zone(), graph(), common(),
+                                                   machine());
     decompression_optimizer.Reduce();
   }