[turbofan] Initial version of allocation folding and write barrier elimination.

src/compiler/memory-optimizer.cc

Index: src/compiler/memory-optimizer.cc
diff --git a/src/compiler/memory-optimizer.cc b/src/compiler/memory-optimizer.cc
new file mode 100644
index 0000000000000000000000000000000000000000..fb1d0ab6735fcbebfb06cf7110f9008a91e37f4c
--- /dev/null
+++ b/src/compiler/memory-optimizer.cc
@@ -0,0 +1,494 @@
+// Copyright 2016 the V8 project authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "src/compiler/memory-optimizer.h"
+
+#include "src/compiler/js-graph.h"
+#include "src/compiler/linkage.h"
+#include "src/compiler/node-matchers.h"
+#include "src/compiler/node-properties.h"
+#include "src/compiler/node.h"
+#include "src/compiler/simplified-operator.h"
+
+namespace v8 {
+namespace internal {
+namespace compiler {
+
+MemoryOptimizer::MemoryOptimizer(JSGraph* jsgraph, Zone* zone)
+    : jsgraph_(jsgraph),
+      empty_state_(new (zone) AllocationState()),
+      pending_(zone),
+      tokens_(zone),
+      zone_(zone) {}
+
+void MemoryOptimizer::Optimize() {
+  EnqueueUses(graph()->start(), empty_state());
+  while (!tokens_.empty()) {
+    Token const token = tokens_.front();
+    tokens_.pop();
+    VisitNode(token.node, token.state);
+  }
+  DCHECK(pending_.empty());
+  DCHECK(tokens_.empty());
+}
+
+MemoryOptimizer::AllocationGroup::AllocationGroup(Node* node,
+                                                  PretenureFlag pretenure,
+                                                  Zone* zone)
+    : node_ids_(zone), pretenure_(pretenure), size_(nullptr) {
+  node_ids_.insert(node->id());
+}
+
+MemoryOptimizer::AllocationGroup::AllocationGroup(Node* node,
+                                                  PretenureFlag pretenure,
+                                                  Node* size, Zone* zone)
+    : node_ids_(zone), pretenure_(pretenure), size_(size) {
+  node_ids_.insert(node->id());
+}
+
+void MemoryOptimizer::AllocationGroup::Add(Node* node) {
+  node_ids_.insert(node->id());
+}
+
+bool MemoryOptimizer::AllocationGroup::Contains(Node* node) const {
+  return node_ids_.find(node->id()) != node_ids_.end();
+}
+
+MemoryOptimizer::AllocationState::AllocationState()
+    : group_(nullptr), size_(std::numeric_limits<int>::max()), top_(nullptr) {}
+
+MemoryOptimizer::AllocationState::AllocationState(AllocationGroup* group)
+    : group_(group), size_(std::numeric_limits<int>::max()), top_(nullptr) {}
+
+MemoryOptimizer::AllocationState::AllocationState(AllocationGroup* group,
+                                                  int size, Node* top)
+    : group_(group), size_(size), top_(top) {}
+
+bool MemoryOptimizer::AllocationState::IsNewSpaceAllocation() const {
+  return group() && group()->IsNewSpaceAllocation();
+}
+
+void MemoryOptimizer::VisitNode(Node* node, AllocationState const* state) {
+  DCHECK(!node->IsDead());
+  DCHECK_LT(0, node->op()->EffectInputCount());
+  switch (node->opcode()) {
+    case IrOpcode::kAllocate:
+      return VisitAllocate(node, state);
+    case IrOpcode::kCall:
+      return VisitCall(node, state);
+    case IrOpcode::kLoadElement:
+      return VisitLoadElement(node, state);
+    case IrOpcode::kLoadField:
+      return VisitLoadField(node, state);
+    case IrOpcode::kStoreElement:
+      return VisitStoreElement(node, state);
+    case IrOpcode::kStoreField:
+      return VisitStoreField(node, state);
+    case IrOpcode::kCheckedLoad:
+    case IrOpcode::kCheckedStore:
+    case IrOpcode::kIfException:
+    case IrOpcode::kLoad:
+    case IrOpcode::kStore:
+      return VisitOtherEffect(node, state);
+    default:
+      break;
+  }
+  DCHECK_EQ(0, node->op()->EffectOutputCount());
+}
+
+void MemoryOptimizer::VisitAllocate(Node* node, AllocationState const* state) {
+  DCHECK_EQ(IrOpcode::kAllocate, node->opcode());
+  Node* value;
+  Node* size = node->InputAt(0);
+  Node* effect = node->InputAt(1);
+  Node* control = node->InputAt(2);
+  PretenureFlag pretenure = OpParameter<PretenureFlag>(node->op());
+
+  // Determine the top/limit addresses.
+  Node* top_address = jsgraph()->ExternalConstant(
+      pretenure == NOT_TENURED
+          ? ExternalReference::new_space_allocation_top_address(isolate())
+          : ExternalReference::old_space_allocation_top_address(isolate()));
+  Node* limit_address = jsgraph()->ExternalConstant(
+      pretenure == NOT_TENURED
+          ? ExternalReference::new_space_allocation_limit_address(isolate())
+          : ExternalReference::old_space_allocation_limit_address(isolate()));
+
+  // Check if we can fold this allocation into a previous allocation represented
+  // by the incoming {state}.
+  Int32Matcher m(size);
+  if (m.HasValue() && m.Value() < Page::kMaxRegularHeapObjectSize) {
+    int32_t const object_size = m.Value();
+    if (state->size() <= Page::kMaxRegularHeapObjectSize - object_size &&
+        state->group()->pretenure() == pretenure) {
+      // We can fold this Allocate {node} into the allocation {group}
+      // represented by the given {state}. Compute the upper bound for
+      // the new {state}.
+      int32_t const state_size = state->size() + object_size;

[Hannes Payer (out of office), 2016/05/10 07:49:33]

Hoist this computation and use state_size already in the if check above.

[Benedikt Meurer, 2016/05/10 08:01:45]

Cannot do because state->size() + object_size can overflow.

+
+      // Update the reservation check to the actual maximum upper bound.
+      AllocationGroup* const group = state->group();
+      if (OpParameter<int32_t>(group->size()) < state_size) {
+        NodeProperties::ChangeOp(group->size(),
+                                 common()->Int32Constant(state_size));
+      }
+
+      // Update the allocation top with the new object allocation.
+      // TODO(bmeurer): Defer writing back top as much as possible.
+      Node* top = graph()->NewNode(machine()->IntAdd(), state->top(),
+                                   jsgraph()->IntPtrConstant(object_size));
+      effect = graph()->NewNode(
+          machine()->Store(StoreRepresentation(
+              MachineType::PointerRepresentation(), kNoWriteBarrier)),
+          top_address, jsgraph()->IntPtrConstant(0), top, effect, control);
+
+      // Compute the effective inner allocated address.
+      value = graph()->NewNode(
+          machine()->BitcastWordToTagged(),
+          graph()->NewNode(machine()->IntAdd(), state->top(),
+                           jsgraph()->IntPtrConstant(kHeapObjectTag)));
+
+      // Extend the allocation {group}.
+      group->Add(value);
+      state = new (zone()) AllocationState(group, state_size, top);
+    } else {
+      // Setup a mutable reservation size node; will be patched as we fold
+      // additional allocations into this new group.
+      Node* size = graph()->NewNode(common()->Int32Constant(object_size));
+
+      // Load allocation top and limit.
+      Node* top = effect =
+          graph()->NewNode(machine()->Load(MachineType::Pointer()), top_address,
+                           jsgraph()->IntPtrConstant(0), effect, control);
+      Node* limit = effect = graph()->NewNode(
+          machine()->Load(MachineType::Pointer()), limit_address,
+          jsgraph()->IntPtrConstant(0), effect, control);
+
+      // Check if we need to collect garbage before we can start bump pointer
+      // allocation (always done for folded allocations).
+      Node* check = graph()->NewNode(
+          machine()->UintLessThan(),
+          graph()->NewNode(
+              machine()->IntAdd(), top,
+              machine()->Is64()
+                  ? graph()->NewNode(machine()->ChangeInt32ToInt64(), size)
+                  : size),
+          limit);
+      Node* branch =
+          graph()->NewNode(common()->Branch(BranchHint::kTrue), check, control);
+
+      Node* if_true = graph()->NewNode(common()->IfTrue(), branch);
+      Node* etrue = effect;
+      Node* vtrue = top;
+
+      Node* if_false = graph()->NewNode(common()->IfFalse(), branch);
+      Node* efalse = effect;
+      Node* vfalse;
+      {
+        Node* target = pretenure == NOT_TENURED
+                           ? jsgraph()->AllocateInNewSpaceStubConstant()
+                           : jsgraph()->AllocateInOldSpaceStubConstant();
+        if (!allocate_operator_.is_set()) {
+          CallDescriptor* descriptor =
+              Linkage::GetAllocateCallDescriptor(graph()->zone());
+          allocate_operator_.set(common()->Call(descriptor));
+        }
+        vfalse = efalse = graph()->NewNode(allocate_operator_.get(), target,
+                                           size, efalse, if_false);
+        vfalse = graph()->NewNode(machine()->IntSub(), vfalse,
+                                  jsgraph()->IntPtrConstant(kHeapObjectTag));
+      }
+
+      control = graph()->NewNode(common()->Merge(2), if_true, if_false);
+      effect = graph()->NewNode(common()->EffectPhi(2), etrue, efalse, control);
+      value = graph()->NewNode(
+          common()->Phi(MachineType::PointerRepresentation(), 2), vtrue, vfalse,
+          control);
+
+      // Compute the new top and write it back.
+      top = graph()->NewNode(machine()->IntAdd(), value,
+                             jsgraph()->IntPtrConstant(object_size));
+      effect = graph()->NewNode(
+          machine()->Store(StoreRepresentation(
+              MachineType::PointerRepresentation(), kNoWriteBarrier)),
+          top_address, jsgraph()->IntPtrConstant(0), top, effect, control);
+
+      // Compute the initial object address.
+      value = graph()->NewNode(
+          machine()->BitcastWordToTagged(),
+          graph()->NewNode(machine()->IntAdd(), value,
+                           jsgraph()->IntPtrConstant(kHeapObjectTag)));
+
+      // Start a new allocation group.
+      AllocationGroup* group =
+          new (zone()) AllocationGroup(value, pretenure, size, zone());
+      state = new (zone()) AllocationState(group, object_size, top);
+    }
+  } else {
+    // Load allocation top and limit.

[Hannes Payer (out of office), 2016/05/10 07:49:33]

This code is similar to the one in the else case above. I guess it would make
sense to refactor the common path of bump pointer allocation. WDYT?

[Benedikt Meurer, 2016/05/10 08:01:45]

It's just almost the same, but not 100%. For now I'd like to keep it this way
and refactor it once the top is tagged.

+    Node* top = effect =
+        graph()->NewNode(machine()->Load(MachineType::Pointer()), top_address,
+                         jsgraph()->IntPtrConstant(0), effect, control);
+    Node* limit = effect =
+        graph()->NewNode(machine()->Load(MachineType::Pointer()), limit_address,
+                         jsgraph()->IntPtrConstant(0), effect, control);
+
+    // Compute the new top.
+    Node* new_top = graph()->NewNode(
+        machine()->IntAdd(), top,
+        machine()->Is64()
+            ? graph()->NewNode(machine()->ChangeInt32ToInt64(), size)
+            : size);
+
+    // Check if we can do bump pointer allocation here.
+    Node* check = graph()->NewNode(machine()->UintLessThan(), new_top, limit);
+    Node* branch =
+        graph()->NewNode(common()->Branch(BranchHint::kTrue), check, control);
+
+    Node* if_true = graph()->NewNode(common()->IfTrue(), branch);
+    Node* etrue = effect;
+    Node* vtrue;
+    {
+      etrue = graph()->NewNode(
+          machine()->Store(StoreRepresentation(
+              MachineType::PointerRepresentation(), kNoWriteBarrier)),
+          top_address, jsgraph()->IntPtrConstant(0), new_top, etrue, if_true);
+      vtrue = graph()->NewNode(
+          machine()->BitcastWordToTagged(),
+          graph()->NewNode(machine()->IntAdd(), top,
+                           jsgraph()->IntPtrConstant(kHeapObjectTag)));
+    }
+
+    Node* if_false = graph()->NewNode(common()->IfFalse(), branch);
+    Node* efalse = effect;
+    Node* vfalse;
+    {
+      Node* target = pretenure == NOT_TENURED
+                         ? jsgraph()->AllocateInNewSpaceStubConstant()
+                         : jsgraph()->AllocateInOldSpaceStubConstant();
+      if (!allocate_operator_.is_set()) {
+        CallDescriptor* descriptor =
+            Linkage::GetAllocateCallDescriptor(graph()->zone());
+        allocate_operator_.set(common()->Call(descriptor));
+      }
+      vfalse = efalse = graph()->NewNode(allocate_operator_.get(), target, size,
+                                         efalse, if_false);
+    }
+
+    control = graph()->NewNode(common()->Merge(2), if_true, if_false);
+    effect = graph()->NewNode(common()->EffectPhi(2), etrue, efalse, control);
+    value = graph()->NewNode(common()->Phi(MachineRepresentation::kTagged, 2),
+                             vtrue, vfalse, control);
+
+    // Create an unfoldable allocation group.
+    AllocationGroup* group =
+        new (zone()) AllocationGroup(value, pretenure, zone());
+    state = new (zone()) AllocationState(group);
+  }
+
+  // Replace all effect uses of {node} with the {effect}, enqueue the
+  // effect uses for further processing, and replace all value uses of
+  // {node} with the {value}.
+  for (Edge edge : node->use_edges()) {
+    if (NodeProperties::IsEffectEdge(edge)) {
+      EnqueueUse(edge.from(), edge.index(), state);
+      edge.UpdateTo(effect);
+    } else {
+      DCHECK(NodeProperties::IsValueEdge(edge));
+      edge.UpdateTo(value);
+    }
+  }
+
+  // Kill the {node} to make sure we don't leave dangling dead uses.
+  node->Kill();
+}
+
+void MemoryOptimizer::VisitCall(Node* node, AllocationState const* state) {
+  DCHECK_EQ(IrOpcode::kCall, node->opcode());
+  // If the call can allocate, we start with a fresh state.
+  if (!(CallDescriptorOf(node->op())->flags() & CallDescriptor::kNoAllocate)) {
+    state = empty_state();
+  }
+  EnqueueUses(node, state);
+}
+
+void MemoryOptimizer::VisitLoadElement(Node* node,
+                                       AllocationState const* state) {
+  DCHECK_EQ(IrOpcode::kLoadElement, node->opcode());
+  ElementAccess const& access = ElementAccessOf(node->op());
+  Node* index = node->InputAt(1);
+  node->ReplaceInput(1, ComputeIndex(access, index));
+  NodeProperties::ChangeOp(node, machine()->Load(access.machine_type));
+  EnqueueUses(node, state);
+}
+
+void MemoryOptimizer::VisitLoadField(Node* node, AllocationState const* state) {
+  DCHECK_EQ(IrOpcode::kLoadField, node->opcode());
+  FieldAccess const& access = FieldAccessOf(node->op());
+  Node* offset = jsgraph()->IntPtrConstant(access.offset - access.tag());
+  node->InsertInput(graph()->zone(), 1, offset);
+  NodeProperties::ChangeOp(node, machine()->Load(access.machine_type));
+  EnqueueUses(node, state);
+}
+
+void MemoryOptimizer::VisitStoreElement(Node* node,
+                                        AllocationState const* state) {
+  DCHECK_EQ(IrOpcode::kStoreElement, node->opcode());
+  ElementAccess const& access = ElementAccessOf(node->op());
+  Node* object = node->InputAt(0);
+  Node* index = node->InputAt(1);
+  WriteBarrierKind write_barrier_kind =
+      ComputeWriteBarrierKind(object, state, access.write_barrier_kind);
+  node->ReplaceInput(1, ComputeIndex(access, index));
+  NodeProperties::ChangeOp(
+      node, machine()->Store(StoreRepresentation(
+                access.machine_type.representation(), write_barrier_kind)));
+  EnqueueUses(node, state);
+}
+
+void MemoryOptimizer::VisitStoreField(Node* node,
+                                      AllocationState const* state) {
+  DCHECK_EQ(IrOpcode::kStoreField, node->opcode());
+  FieldAccess const& access = FieldAccessOf(node->op());
+  Node* object = node->InputAt(0);
+  WriteBarrierKind write_barrier_kind =
+      ComputeWriteBarrierKind(object, state, access.write_barrier_kind);
+  Node* offset = jsgraph()->IntPtrConstant(access.offset - access.tag());
+  node->InsertInput(graph()->zone(), 1, offset);
+  NodeProperties::ChangeOp(
+      node, machine()->Store(StoreRepresentation(
+                access.machine_type.representation(), write_barrier_kind)));
+  EnqueueUses(node, state);
+}
+
+void MemoryOptimizer::VisitOtherEffect(Node* node,
+                                       AllocationState const* state) {
+  EnqueueUses(node, state);
+}
+
+Node* MemoryOptimizer::ComputeIndex(ElementAccess const& access, Node* key) {
+  Node* index = key;
+  int element_size_shift =
+      ElementSizeLog2Of(access.machine_type.representation());
+  if (element_size_shift) {
+    index = graph()->NewNode(machine()->Word32Shl(), index,
+                             jsgraph()->Int32Constant(element_size_shift));
+  }
+  const int fixed_offset = access.header_size - access.tag();
+  if (fixed_offset) {
+    index = graph()->NewNode(machine()->Int32Add(), index,
+                             jsgraph()->Int32Constant(fixed_offset));
+  }
+  if (machine()->Is64()) {
+    // TODO(turbofan): This is probably only correct for typed arrays, and only
+    // if the typed arrays are at most 2GiB in size, which happens to match
+    // exactly our current situation.
+    index = graph()->NewNode(machine()->ChangeUint32ToUint64(), index);
+  }
+  return index;
+}
+
+WriteBarrierKind MemoryOptimizer::ComputeWriteBarrierKind(
+    Node* object, AllocationState const* state,
+    WriteBarrierKind write_barrier_kind) {
+  if (state->IsNewSpaceAllocation() && state->group()->Contains(object)) {
+    write_barrier_kind = kNoWriteBarrier;
+  }
+  return write_barrier_kind;
+}
+
+MemoryOptimizer::AllocationState const* MemoryOptimizer::MergeStates(
+    AllocationStates const& states) {
+  // Check if all states are the same; or at least if all allocation
+  // states belong to the same allocation group.
+  AllocationState const* state = states.front();
+  AllocationGroup* group = state->group();
+  for (size_t i = 1; i < states.size(); ++i) {
+    if (states[i] != state) state = nullptr;
+    if (states[i]->group() != group) group = nullptr;
+  }
+  if (state == nullptr) {
+    if (group != nullptr) {
+      // We cannot fold any more allocations into this group, but we can still
+      // eliminate write barriers on stores to this group.
+      // TODO(bmeurer): We could potentially just create a Phi here to merge
+      // the various tops; but we need to pay special attention not to create
+      // an unschedulable graph.
+      state = new (zone()) AllocationState(group);
+    } else {
+      // The states are from different allocation groups.
+      state = empty_state();
+    }
+  }
+  return state;
+}
+
+void MemoryOptimizer::EnqueueMerge(Node* node, int index,
+                                   AllocationState const* state) {
+  DCHECK_EQ(IrOpcode::kEffectPhi, node->opcode());
+  int const input_count = node->InputCount() - 1;
+  DCHECK_LT(0, input_count);
+  Node* const control = node->InputAt(input_count);
+  if (control->opcode() == IrOpcode::kLoop) {
+    // For loops we always start with an empty state at the beginning.
+    if (index == 0) EnqueueUses(node, empty_state());
+  } else {
+    DCHECK_EQ(IrOpcode::kMerge, control->opcode());
+    // Check if we already know about this pending merge.
+    NodeId const id = node->id();
+    auto it = pending_.find(id);
+    if (it == pending_.end()) {
+      // Insert a new pending merge.
+      it = pending_.insert(std::make_pair(id, AllocationStates(zone()))).first;
+    }
+    // Add the next input state.
+    it->second.push_back(state);
+    // Check if states for all inputs are available by now.
+    if (it->second.size() == static_cast<size_t>(input_count)) {
+      // All inputs to this effect merge are done, merge the states given all
+      // input constraints, drop the pending merge and enqueue uses of the
+      // EffectPhi {node}.
+      state = MergeStates(it->second);
+      EnqueueUses(node, state);
+      pending_.erase(it);
+    }
+  }
+}
+
+void MemoryOptimizer::EnqueueUses(Node* node, AllocationState const* state) {
+  for (Edge const edge : node->use_edges()) {
+    if (NodeProperties::IsEffectEdge(edge)) {
+      EnqueueUse(edge.from(), edge.index(), state);
+    }
+  }
+}
+
+void MemoryOptimizer::EnqueueUse(Node* node, int index,
+                                 AllocationState const* state) {
+  if (node->opcode() == IrOpcode::kEffectPhi) {
+    // An EffectPhi represents a merge of different effect chains, which
+    // needs special handling depending on whether the merge is part of a
+    // loop or just a normal control join.
+    EnqueueMerge(node, index, state);
+  } else {
+    Token token = {node, state};
+    tokens_.push(token);
+  }
+}
+
+Graph* MemoryOptimizer::graph() const { return jsgraph()->graph(); }
+
+Isolate* MemoryOptimizer::isolate() const { return jsgraph()->isolate(); }
+
+CommonOperatorBuilder* MemoryOptimizer::common() const {
+  return jsgraph()->common();
+}
+
+MachineOperatorBuilder* MemoryOptimizer::machine() const {
+  return jsgraph()->machine();
+}
+
+}  // namespace compiler
+}  // namespace internal
+}  // namespace v8