Implement representation selection as part of SimplifiedLowering. Representation selection also req…

src/compiler/simplified-lowering.cc

 // Copyright 2014 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/simplified-lowering.h"
 
+#include <deque>
+#include <queue>
+
+#include "src/compiler/common-operator.h"
 #include "src/compiler/graph-inl.h"
 #include "src/compiler/node-properties-inl.h"
+#include "src/compiler/representation-change.h"
+#include "src/compiler/simplified-lowering.h"
+#include "src/compiler/simplified-operator.h"
 #include "src/objects.h"
 
 namespace v8 {
 namespace internal {
 namespace compiler {
 
+// Macro for outputting trace information from representation inference.
+#define TRACE(x) \
+  if (FLAG_trace_representation) PrintF x
+
+// Representation selection and lowering of {Simplified} operators to machine
+// operators are interwined. We use a fixpoint calculation to compute both the
+// output representation and the best possible lowering for {Simplified} nodes.
+// Representation change insertion ensures that all values are in the correct
+// machine representation after this phase, as dictated by the machine
+// operators themselves.
+enum Phase {

[rossberg, 2014/08/06 13:04:07]

I wish we could somehow factor this using some form of parameterisation, and not
a mode flag... :(

[titzer, 2014/08/08 09:16:13]

As discussed, the virtual method thing is going to result in pulling out major
portions into their own methods...

+  // 1.) PROPAGATE: Traverse the graph from the end, pushing usage information
+  //     backwards from uses to definitions, around cycles in phis, according
+  //     to local rules for each operator.
+  //     During this phase, the usage information for a node determines the best
+  //     possible lowering for each operator so far, and that in turn determines
+  //     the output representation.
+  //     Therefore, to be correct, this phase must iterate to a fixpoint before
+  //     the next phase can begin.
+  PROPAGATE,
+
+  // 2.) LOWER: perform lowering for all {Simplified} nodes by replacing some
+  //     operators for some nodes, expanding some nodes to multiple nodes, or
+  //     removing some (redundant) nodes.
+  //     During this phase, use the {RepresentationChanger} to insert
+  //     representation changes between uses that demand a particular
+  //     representation and nodes that produce a different representation.
+  LOWER
+};
+
+
+class RepresentationSelector {
+ public:
+  // Information for each node tracked during the fixpoint.
+  struct NodeInfo {
+    RepTypeUnion use : 14;     // Union of all usages for the node.
+    bool queued : 1;           // Bookkeeping for the traversal.
+    bool visited : 1;          // Bookkeeping for the traversal.
+    RepTypeUnion output : 16;  // Output type of the node.

[rossberg, 2014/08/06 13:04:07]

Why is this bitset larger than the other?

[titzer, 2014/08/08 09:16:13]

Done.

+  };
+
+  RepresentationSelector(JSGraph* jsgraph, Zone* zone,
+                         RepresentationChanger* changer)
+      : jsgraph_(jsgraph),
+        zone_(zone),
+        count_(jsgraph->graph()->NodeCount()),
+        info_(zone->NewArray<NodeInfo>(count_)),
+        nodes_(NodeVector::allocator_type(zone)),
+        contains_js_nodes_(false),
+        phase_(PROPAGATE),
+        changer_(changer),
+        queue_(std::deque<Node*, NodePtrZoneAllocator>(
+            NodePtrZoneAllocator(zone))) {
+    memset(info_, 0, sizeof(NodeInfo) * count_);
+  }
+
+  void Run(SimplifiedLowering* lowering) {
+    // Run propagation phase to a fixpoint.
+    TRACE(("--{Propagation phase}--\n"));
+    phase_ = PROPAGATE;
+    Enqueue(jsgraph_->graph()->end());
+    // Process nodes from the queue until it is empty.
+    while (!queue_.empty()) {
+      Node* node = queue_.front();
+      NodeInfo* info = GetInfo(node);
+      queue_.pop();
+      info->queued = false;
+      TRACE((" visit #%d: %s\n", node->id(), node->op()->mnemonic()));
+      VisitNode(node, info->use, NULL);
+      TRACE(("  ==> output "));
+      PrintInfo(info->output);
+      TRACE(("\n"));
+    }
+
+    // Run lowering and change insertion phase.
+    TRACE(("--{Simplified lowering phase}--\n"));
+    phase_ = LOWER;
+    // Process nodes from the collected {nodes_} vector.
+    for (NodeVector::iterator i = nodes_.begin(); i != nodes_.end(); ++i) {
+      Node* node = *i;
+      TRACE((" visit #%d: %s\n", node->id(), node->op()->mnemonic()));
+      // Reuse {VisitNode()} so the representation rules are in one place.
+      VisitNode(node, GetUseInfo(node), lowering);
+    }
+  }
+
+  // Enqueue {node} if the {use} contains new information for that node.
+  // Add {node} to {nodes_} if this is the first time it's been visited.
+  void Enqueue(Node* node, RepTypeUnion use = 0) {
+    if (phase_ != PROPAGATE) return;
+    NodeInfo* info = GetInfo(node);
+    if (!info->visited) {
+      // First visit of this node.
+      info->visited = true;
+      info->queued = true;
+      nodes_.push_back(node);
+      queue_.push(node);
+      TRACE(("  initial: "));
+      info->use |= use;
+      PrintUseInfo(node);
+      return;
+    }
+    TRACE(("   queue?: "));
+    PrintUseInfo(node);
+    if ((info->use & use) != use) {
+      // New usage information for the node is available.
+      if (!info->queued) {
+        queue_.push(node);
+        info->queued = true;
+        TRACE(("   added: "));
+      } else {
+        TRACE((" inqueue: "));
+      }
+      info->use |= use;
+      PrintUseInfo(node);
+    }
+  }
+
+  bool lower() { return phase_ == LOWER; }
+
+  void Enqueue(Node* node, RepType use) {
+    Enqueue(node, static_cast<RepTypeUnion>(use));
+  }
+
+  void SetOutput(Node* node, RepTypeUnion output) {
+    // Every node should have at most one output representation. Note that
+    // phis can have 0, if they have not been used in a representation-inducing
+    // instruction.
+    DCHECK((output & rMask) == 0 || IsPowerOf2(output & rMask));
+    GetInfo(node)->output = output;
+  }
+
+  bool BothInputsAre(Node* node, Type* type) {
+    DCHECK_EQ(2, node->InputCount());
+    return NodeProperties::GetBounds(node->InputAt(0)).upper->Is(type) &&
+           NodeProperties::GetBounds(node->InputAt(1)).upper->Is(type);
+  }
+
+  void ProcessInput(Node* node, int index, RepTypeUnion use) {
+    Node* input = node->InputAt(index);
+    if (phase_ == PROPAGATE) {
+      // In the propagate phase, propagate the usage information backward.
+      Enqueue(input, use);
+    } else {
+      // In the change phase, insert a change before the use if necessary.
+      RepTypeUnion output = GetInfo(input)->output;
+      if ((output & rMask & use) == 0) {
+        // Output representation doesn't match usage.
+        TRACE(("  change: #%d:%s(@%d #%d:%s) ", node->id(),
+               node->op()->mnemonic(), index, input->id(),
+               input->op()->mnemonic()));
+        TRACE((" from "));
+        PrintInfo(output);
+        TRACE((" to "));
+        PrintInfo(use);
+        TRACE(("\n"));
+        Node* n = changer_->GetRepresentationFor(input, output, use);
+        node->ReplaceInput(index, n);
+      }
+    }
+  }
+
+  static const RepTypeUnion kFloat64 = rFloat64 | tNumber;
+  static const RepTypeUnion kInt32 = rWord32 | tInt32;
+  static const RepTypeUnion kUint32 = rWord32 | tUint32;
+  static const RepTypeUnion kInt64 = rWord64 | tInt64;
+  static const RepTypeUnion kUint64 = rWord64 | tUint64;
+  static const RepTypeUnion kAnyTagged = rTagged | tAny;
+
+  // The default, most general visitation case. For {node}, process all value,
+  // context, effect, and control inputs, assuming that value inputs should have
+  // {rTagged} representation and can observe all output values {tAny}.
+  void VisitInputs(Node* node) {
+    InputIter i = node->inputs().begin();
+    for (int j = NodeProperties::GetValueInputCount(node); j > 0; ++i, j--) {
+      ProcessInput(node, i.index(), kAnyTagged);  // Value inputs
+    }
+    for (int j = NodeProperties::GetContextInputCount(node); j > 0; ++i, j--) {
+      ProcessInput(node, i.index(), kAnyTagged);  // Context inputs
+    }
+    for (int j = NodeProperties::GetEffectInputCount(node); j > 0; ++i, j--) {
+      Enqueue(*i);  // Effect inputs: just visit
+    }
+    for (int j = NodeProperties::GetControlInputCount(node); j > 0; ++i, j--) {
+      Enqueue(*i);  // Control inputs: just visit
+    }
+    SetOutput(node, kAnyTagged);
+  }
+
+  // Helper for binops of the I x I -> O variety.
+  void VisitBinop(Node* node, RepTypeUnion input_use, RepTypeUnion output) {
+    DCHECK_EQ(2, node->InputCount());
+    ProcessInput(node, 0, input_use);
+    ProcessInput(node, 1, input_use);
+    SetOutput(node, output);
+  }
+
+  // Helper for unops of the I -> O variety.
+  void VisitUnop(Node* node, RepTypeUnion input_use, RepTypeUnion output) {
+    DCHECK_EQ(1, node->InputCount());
+    ProcessInput(node, 0, input_use);
+    SetOutput(node, output);
+  }
+
+  // Helper for leaf nodes.
+  void VisitLeaf(Node* node, RepType output) {
+    DCHECK_EQ(0, node->InputCount());
+    SetOutput(node, output);
+  }
+
+  // Helpers for specific types of binops.

[rossberg, 2014/08/06 13:04:07]

Nit: not sure these are worth it, but I leave it up to you.

+  void VisitFloat64Binop(Node* node) { VisitBinop(node, kFloat64, kFloat64); }
+  void VisitInt32Binop(Node* node) { VisitBinop(node, kInt32, kInt32); }
+  void VisitUint32Binop(Node* node) { VisitBinop(node, kUint32, kUint32); }
+  void VisitInt64Binop(Node* node) { VisitBinop(node, kInt64, kInt64); }
+  void VisitUint64Binop(Node* node) { VisitBinop(node, kUint64, kUint64); }
+  void VisitFloat64Cmp(Node* node) { VisitBinop(node, kFloat64, rBit); }
+  void VisitInt32Cmp(Node* node) { VisitBinop(node, kInt32, rBit); }
+  void VisitUint32Cmp(Node* node) { VisitBinop(node, kUint32, rBit); }
+  void VisitInt64Cmp(Node* node) { VisitBinop(node, kInt64, rBit); }
+  void VisitUint64Cmp(Node* node) { VisitBinop(node, kUint64, rBit); }
+
+  // Helper for handling phis.
+  void VisitPhi(Node* node, RepTypeUnion use) {
+    // First, propagate the usage information to inputs of the phi.
+    int values = node->op()->InputCount();

[rossberg, 2014/08/06 13:04:07]

Doesn't this include non-values?

+    Node::Inputs inputs = node->inputs();
+    for (Node::Inputs::iterator iter(inputs.begin()); iter != inputs.end();
+         ++iter, --values) {
+      // Propagate {use} of the phi to value inputs, and 0 to control.
+      ProcessInput(node, iter.index(), values > 0 ? use : 0);

[rossberg, 2014/08/06 13:04:07]

Oh, this breaks the node properties abstraction quite nastily.

+    }
+    // Phis adapt to whatever output representation their uses demand.
+    RepTypeUnion use_rep = GetUseInfo(node) & rMask;
+    RepTypeUnion use_type = GetUseInfo(node) & tMask;
+    RepTypeUnion rep = 0;
+    if (use_rep & rTagged) {
+      rep = rTagged;  // Tagged overrides everything.
+    } else if (use_rep & rFloat64) {
+      rep = rFloat64;
+    } else if (use_rep & rWord64) {
+      rep = rWord64;
+    } else if (use_rep & rWord32) {
+      rep = rWord32;
+    } else if (use_rep & rBit) {
+      rep = rBit;
+    } else {
+      // There was no representation associated with any of the uses of this
+      // phi.
+      // Select the best one based on the usage type.
+      if (use_type & tAny) {
+        rep = rTagged;
+      } else if (use_rep & tNumber) {
+        rep = rFloat64;
+      } else if (use_rep & tInt64 || use_rep & tUint64) {
+        rep = rWord64;
+      } else if (use_rep & tInt32 || use_rep & tUint32) {
+        rep = rWord32;
+      } else if (use_rep & tBool) {
+        rep = rBit;
+      } else {
+        UNREACHABLE();  // should have at least a usage type!
+      }
+    }
+    // Preserve the usage type, but set the representation.
+    SetOutput(node, rep | use_type);
+  }
+
+  Operator* Int32Op(Node* node) {
+    return changer_->Int32OperatorFor(node->opcode());
+  }
+
+  Operator* Uint32Op(Node* node) {
+    return changer_->Uint32OperatorFor(node->opcode());
+  }
+
+  Operator* Float64Op(Node* node) {
+    return changer_->Float64OperatorFor(node->opcode());
+  }
+
+  // Dispatching routine for visiting the node {node} with the usage {use}.
+  // Depending on the operator, propagate new usage info to the inputs.
+  void VisitNode(Node* node, RepTypeUnion use, SimplifiedLowering* lowering) {
+    switch (node->opcode()) {
+      //------------------------------------------------------------------
+      // Common operators.
+      //------------------------------------------------------------------
+      case IrOpcode::kStart:
+        return VisitLeaf(node, rBit);

[rossberg, 2014/08/06 13:04:07]

Nit: could combine these two cases.

+      case IrOpcode::kDead:
+        return VisitLeaf(node, rBit);

[rossberg, 2014/08/06 13:04:07]

Why rBit?

+      case IrOpcode::kParameter:
+        return VisitLeaf(node, rTagged);
+      case IrOpcode::kInt32Constant:
+        return VisitLeaf(node, rWord32);
+      case IrOpcode::kInt64Constant:
+        return VisitLeaf(node, rWord64);
+      case IrOpcode::kFloat64Constant:
+        return VisitLeaf(node, rFloat64);
+      case IrOpcode::kExternalConstant:
+        return VisitLeaf(node, rPtr);
+      case IrOpcode::kNumberConstant:
+        return VisitLeaf(node, rTagged);
+      case IrOpcode::kHeapConstant:
+        return VisitLeaf(node, rTagged);
+
+      case IrOpcode::kEnd:
+      case IrOpcode::kIfTrue:
+      case IrOpcode::kIfFalse:
+      case IrOpcode::kReturn:
+      case IrOpcode::kMerge:
+      case IrOpcode::kThrow:
+        return VisitInputs(node);  // default visit for all node inputs.
+
+      case IrOpcode::kBranch:
+        ProcessInput(node, 0, rBit);
+        Enqueue(NodeProperties::GetControlInput(node, 0));
+        break;
+      case IrOpcode::kPhi:
+        return VisitPhi(node, use);
+
+//------------------------------------------------------------------

[rossberg, 2014/08/06 13:04:07]

Nit: weird indentation

[titzer, 2014/08/08 09:16:13]

clang-format keeps putting it back, so I'll just leave it.

+// JavaScript operators.
+//------------------------------------------------------------------
+// For now, we assume that all JS operators were too complex to lower
+// to Simplified and that they will always require tagged value inputs and
+// produce tagged value outputs.
+// TODO(turbofan): it might be possible to lower some JSOperators here,
+// but that responsibility really lies in the typed lowering phase.
+#define DEFINE_JS_CASE(x) case IrOpcode::k##x:
+        JS_OP_LIST(DEFINE_JS_CASE)
+#undef DEFINE_JS_CASE
+        contains_js_nodes_ = true;
+        VisitInputs(node);
+        return SetOutput(node, rTagged);
+
+      //------------------------------------------------------------------
+      // Simplified operators.
+      //------------------------------------------------------------------
+      case IrOpcode::kBooleanNot: {
+        if (lower()) {
+          RepTypeUnion input = GetInfo(node->InputAt(0))->output;
+          if (input & rBit) {
+            // BooleanNot(x: rBit) => WordEqual(x, #0)
+            node->set_op(lowering->machine()->WordEqual());
+            node->AppendInput(jsgraph_->zone(), jsgraph_->Int32Constant(0));
+          } else {
+            // BooleanNot(x: rTagged) => WordEqual(x, #false)

[rossberg, 2014/08/06 13:04:07]

Wait, why is BooleanNot overloaded on JS Booleans?

+            node->set_op(lowering->machine()->WordEqual());
+            node->AppendInput(jsgraph_->zone(), jsgraph_->FalseConstant());
+          }
+        } else {
+          // No representation requirement, since we handle both during
+          // lowering.
+          ProcessInput(node, 0, tBool);
+          SetOutput(node, rBit);
+        }
+        break;
+      }
+      case IrOpcode::kNumberEqual:
+      case IrOpcode::kNumberLessThan:
+      case IrOpcode::kNumberLessThanOrEqual: {
+        // Number comparisons reduce to integer comparisons for integer inputs.
+        if (BothInputsAre(node, Type::Signed32())) {
+          // => signed Int32Cmp
+          VisitInt32Cmp(node);
+          if (lower()) node->set_op(Int32Op(node));
+        } else if (BothInputsAre(node, Type::Unsigned32())) {
+          // => unsigned Int32Cmp
+          VisitUint32Cmp(node);
+          if (lower()) node->set_op(Uint32Op(node));
+        } else {
+          // => Float64Cmp
+          VisitFloat64Cmp(node);
+          if (lower()) node->set_op(Float64Op(node));
+        }
+        break;
+      }
+      case IrOpcode::kNumberAdd:
+      case IrOpcode::kNumberSubtract: {
+        // Add and subtract reduce to Int32Add/Sub if the inputs
+        // are already integers and all uses are truncating.
+        if (BothInputsAre(node, Type::Signed32()) &&
+            (use & (tUint32 | tNumber | tAny)) == 0) {
+          // => signed Int32Add/Sub
+          VisitInt32Binop(node);
+          if (lower()) node->set_op(Int32Op(node));
+        } else if (BothInputsAre(node, Type::Unsigned32()) &&
+                   (use & (tInt32 | tNumber | tAny)) == 0) {
+          // => unsigned Int32Add/Sub
+          VisitUint32Binop(node);
+          if (lower()) node->set_op(Uint32Op(node));
+        } else {
+          // => Float64Add/Sub
+          VisitFloat64Binop(node);
+          if (lower()) node->set_op(Float64Op(node));
+        }
+        break;
+      }
+      case IrOpcode::kNumberMultiply:
+      case IrOpcode::kNumberDivide:
+      case IrOpcode::kNumberModulus: {
+        // Float64Mul/Div/Mod
+        VisitFloat64Binop(node);
+        if (lower()) node->set_op(Float64Op(node));
+        break;
+      }
+      case IrOpcode::kNumberToInt32: {
+        if (lower()) {
+          RepTypeUnion in = GetInfo(node->InputAt(0))->output;
+          if ((in & rMask) == rWord32) {
+            // Input is already represented as a word32. This is a nop.
+            DeferReplacement(node, node->InputAt(0));
+          } else if ((in & tMask) == tInt32) {
+            // The input has type int32, just change representation.
+            VisitUnop(node, tInt32 | rWord32, tInt32 | rWord32);
+            DeferReplacement(node, node->InputAt(0));
+          } else {
+            // Require the input in float64 format and perform conversion.
+            // TODO(turbofan): could also avoid the conversion with a tag check.
+            VisitUnop(node, tInt32 | rFloat64, tInt32 | rWord32);
+            node->set_op(lowering->machine()->ChangeFloat64ToInt32());
+          }
+        } else {
+          // Propage a type to the input, but not a representation.

[rossberg, 2014/08/06 13:04:07]

Typo: Propagate

+          VisitUnop(node, tInt32, tInt32 | rWord32);
+        }
+        break;
+      }
+      case IrOpcode::kNumberToUint32: {
+        if (lower()) {
+          RepTypeUnion in = GetInfo(node->InputAt(0))->output;
+          if ((in & rMask) == rWord32) {
+            // Input is already represented as a word32. This is a nop.
+            DeferReplacement(node, node->InputAt(0));
+          } else if ((in & tMask) == tUint32) {
+            // The input has type int32, just change representation.
+            VisitUnop(node, tUint32 | rWord32, tUint32 | rWord32);
+            DeferReplacement(node, node->InputAt(0));
+          } else {
+            // Require the input in float64 format and perform conversion.
+            VisitUnop(node, tUint32 | rFloat64, tUint32 | rWord32);
+            node->set_op(lowering->machine()->ChangeFloat64ToInt32());
+          }
+        } else {
+          // Propage a type to the input, but not a representation.
+          VisitUnop(node, tUint32, tUint32 | rWord32);
+        }
+        break;
+      }
+      case IrOpcode::kReferenceEqual: {
+        VisitUnop(node, kAnyTagged, rBit);
+        if (lower()) node->set_op(lowering->machine()->WordEqual());
+      }
+      case IrOpcode::kStringEqual: {
+        VisitBinop(node, kAnyTagged, rBit);
+        // TODO(titzer): lower StringEqual to stub/runtime call.
+        break;
+      }
+      case IrOpcode::kStringLessThan: {
+        VisitBinop(node, kAnyTagged, rBit);
+        // TODO(titzer): lower StringLessThan to stub/runtime call.
+        break;
+      }
+      case IrOpcode::kStringLessThanOrEqual: {
+        VisitBinop(node, kAnyTagged, rBit);
+        // TODO(titzer): lower StringLessThanOrEqual to stub/runtime call.
+        break;
+      }
+      case IrOpcode::kStringAdd: {
+        VisitBinop(node, kAnyTagged, kAnyTagged);
+        // TODO(titzer): lower StringAdd to stub/runtime call.
+        break;
+      }
+      case IrOpcode::kLoadField: {
+        FieldAccess access = FieldAccessOf(node->op());
+        ProcessInput(node, 0, changer_->TypeForBasePointer(access));
+        SetOutput(node, changer_->TypeForField(access));
+        if (lower()) lowering->DoLoadField(node);
+        break;
+      }
+      case IrOpcode::kStoreField: {
+        FieldAccess access = FieldAccessOf(node->op());
+        ProcessInput(node, 0, changer_->TypeForBasePointer(access));
+        ProcessInput(node, 1, changer_->TypeForField(access));
+        SetOutput(node, 0);
+        if (lower()) lowering->DoStoreField(node);
+        break;
+      }
+      case IrOpcode::kLoadElement: {
+        ElementAccess access = ElementAccessOf(node->op());
+        ProcessInput(node, 0, changer_->TypeForBasePointer(access));
+        ProcessInput(node, 1, kInt32);  // element index
+        SetOutput(node, changer_->TypeForElement(access));
+        if (lower()) lowering->DoLoadElement(node);
+        break;
+      }
+      case IrOpcode::kStoreElement: {
+        ElementAccess access = ElementAccessOf(node->op());
+        ProcessInput(node, 0, changer_->TypeForBasePointer(access));
+        ProcessInput(node, 1, kInt32);  // element index
+        ProcessInput(node, 2, changer_->TypeForElement(access));
+        SetOutput(node, 0);
+        if (lower()) lowering->DoStoreElement(node);
+        break;
+      }
+
+      //------------------------------------------------------------------
+      // Machine-level operators.
+      //------------------------------------------------------------------
+      case IrOpcode::kLoad: {
+        // TODO(titzer): machine loads/stores need to know BaseTaggedness!?
+        RepType tBase = rTagged;
+        MachineRepresentation rep = OpParameter<MachineRepresentation>(node);
+        ProcessInput(node, 0, tBase);   // pointer or object
+        ProcessInput(node, 1, kInt32);  // index
+        SetOutput(node, changer_->TypeForMachineRepresentation(rep));
+        break;
+      }
+      case IrOpcode::kStore: {
+        // TODO(titzer): machine loads/stores need to know BaseTaggedness!?
+        RepType tBase = rTagged;
+        StoreRepresentation rep = OpParameter<StoreRepresentation>(node);
+        ProcessInput(node, 0, tBase);   // pointer or object
+        ProcessInput(node, 1, kInt32);  // index
+        ProcessInput(node, 2, changer_->TypeForMachineRepresentation(rep.rep));
+        SetOutput(node, 0);
+        break;
+      }
+      case IrOpcode::kWord32Shr:
+        // We use unsigned int32 as the output type for shift right, though

[rossberg, 2014/08/06 13:04:07]

This comment is confusing. Unsigned seems adequate here. Is this a bogus copy &
paste from the next case, perhaps?

+        // technically it doesn't have a sign. Because JavaScript.
+        return VisitBinop(node, rWord32, rWord32 | tUint32);
+      case IrOpcode::kWord32And:
+      case IrOpcode::kWord32Or:
+      case IrOpcode::kWord32Xor:
+      case IrOpcode::kWord32Shl:
+      case IrOpcode::kWord32Sar:
+        // We use signed int32 as the output type for these word32 operations,
+        // though technically they don't have a sign. Because JavaScript.
+        return VisitBinop(node, rWord32, rWord32 | tInt32);
+      case IrOpcode::kWord32Equal:
+        return VisitBinop(node, rWord32, rBit);
+
+      case IrOpcode::kInt32Add:
+      case IrOpcode::kInt32Sub:
+      case IrOpcode::kInt32Mul:
+      case IrOpcode::kInt32Div:
+      case IrOpcode::kInt32Mod:
+        return VisitInt32Binop(node);
+      case IrOpcode::kInt32UDiv:
+      case IrOpcode::kInt32UMod:
+        return VisitUint32Binop(node);
+      case IrOpcode::kInt32LessThan:
+      case IrOpcode::kInt32LessThanOrEqual:
+        return VisitInt32Cmp(node);
+
+      case IrOpcode::kUint32LessThan:
+      case IrOpcode::kUint32LessThanOrEqual:
+        return VisitUint32Cmp(node);
+
+      case IrOpcode::kInt64Add:
+      case IrOpcode::kInt64Sub:
+      case IrOpcode::kInt64Mul:
+      case IrOpcode::kInt64Div:
+      case IrOpcode::kInt64Mod:
+        return VisitInt64Binop(node);
+      case IrOpcode::kInt64LessThan:
+      case IrOpcode::kInt64LessThanOrEqual:
+        return VisitInt64Cmp(node);
+
+      case IrOpcode::kInt64UDiv:
+      case IrOpcode::kInt64UMod:
+        return VisitUint64Binop(node);
+
+      case IrOpcode::kWord64And:
+      case IrOpcode::kWord64Or:
+      case IrOpcode::kWord64Xor:
+      case IrOpcode::kWord64Shl:
+      case IrOpcode::kWord64Shr:
+      case IrOpcode::kWord64Sar:
+        return VisitBinop(node, rWord64, rWord64);
+      case IrOpcode::kWord64Equal:
+        return VisitBinop(node, rWord64, rBit);
+
+      case IrOpcode::kConvertInt32ToInt64:
+        return VisitUnop(node, tInt32 | rWord32, tInt32 | rWord64);
+      case IrOpcode::kConvertInt64ToInt32:
+        return VisitUnop(node, tInt64 | rWord64, tInt32 | rWord32);
+
+      case IrOpcode::kChangeInt32ToFloat64:
+        return VisitUnop(node, tInt32 | rWord32, tInt32 | rFloat64);
+      case IrOpcode::kChangeUint32ToFloat64:
+        return VisitUnop(node, tUint32 | rWord32, tUint32 | rFloat64);
+      case IrOpcode::kChangeFloat64ToInt32:
+        return VisitUnop(node, tInt32 | rFloat64, tInt32 | rWord32);
+      case IrOpcode::kChangeFloat64ToUint32:
+        return VisitUnop(node, tUint32 | rFloat64, tUint32 | rWord32);
+
+      case IrOpcode::kFloat64Add:
+      case IrOpcode::kFloat64Sub:
+      case IrOpcode::kFloat64Mul:
+      case IrOpcode::kFloat64Div:
+      case IrOpcode::kFloat64Mod:
+        return VisitFloat64Binop(node);
+      case IrOpcode::kFloat64Equal:
+      case IrOpcode::kFloat64LessThan:
+      case IrOpcode::kFloat64LessThanOrEqual:
+        return VisitFloat64Cmp(node);
+      default:
+        VisitInputs(node);
+        break;
+    }
+  }
+
+  void DeferReplacement(Node* node, Node* replacement) {
+    if (replacement->id() < count_) {
+      // Replace with a previously existing node eagerly.
+      node->ReplaceUses(replacement);
+    } else {
+      // Otherwise, we are replacing a node with a representation change.
+      // Such a substitution must be done after all lowering is done, because
+      // new nodes do not have {NodeInfo} entries, and that would confuse
+      // the representation change insertion for uses of it.
+      // TODO(titzer): replace these after lowering.
+    }
+    node->RemoveAllInputs();  // Node is now dead.
+  }
+
+  void PrintUseInfo(Node* node) {
+    TRACE(("#%d:%-20s ", node->id(), node->op()->mnemonic()));
+    PrintInfo(GetUseInfo(node));
+    TRACE(("\n"));
+  }
+
+  void PrintInfo(RepTypeUnion info) {
+    if (FLAG_trace_representation) {
+      char buf[REP_TYPE_STRLEN];
+      RenderRepTypeUnion(buf, info);
+      TRACE(("%s", buf));
+    }
+  }
+
+ private:
+  JSGraph* jsgraph_;
+  Zone* zone_;
+  int count_;                       // number of nodes in the graph
+  NodeInfo* info_;                  // node id -> usage information
+  NodeVector nodes_;                // collected nodes
+  bool contains_js_nodes_;          // {true} if a JS operator was seen
+  Phase phase_;                     // current phase of algorithm
+  RepresentationChanger* changer_;  // for inserting representation changes
+
+  std::queue<Node*, std::deque<Node*, NodePtrZoneAllocator> > queue_;
+
+  NodeInfo* GetInfo(Node* node) {
+    DCHECK(node->id() >= 0);
+    DCHECK(node->id() < count_);
+    return &info_[node->id()];
+  }
+
+  RepTypeUnion GetUseInfo(Node* node) { return GetInfo(node)->use; }
+};
+
+
 Node* SimplifiedLowering::IsTagged(Node* node) {
   // TODO(titzer): factor this out to a TaggingScheme abstraction.
   STATIC_ASSERT(kSmiTagMask == 1);  // Only works if tag is the low bit.
   return graph()->NewNode(machine()->WordAnd(), node,
                           jsgraph()->Int32Constant(kSmiTagMask));
 }
 
 
+void SimplifiedLowering::LowerAllNodes() {
+  SimplifiedOperatorBuilder simplified(graph()->zone());
+  RepresentationChanger changer(jsgraph(), &simplified, machine(),
+                                graph()->zone()->isolate());
+  RepresentationSelector selector(jsgraph(), zone(), &changer);
+  selector.Run(this);
+
+  LoweringBuilder::LowerAllNodes();
+}
+
+
 Node* SimplifiedLowering::Untag(Node* node) {
   // TODO(titzer): factor this out to a TaggingScheme abstraction.
   Node* shift_amount = jsgraph()->Int32Constant(kSmiTagSize + kSmiShiftSize);
   return graph()->NewNode(machine()->WordSar(), node, shift_amount);
 }
 
 
 Node* SimplifiedLowering::SmiTag(Node* node) {
   // TODO(titzer): factor this out to a TaggingScheme abstraction.
   Node* shift_amount = jsgraph()->Int32Constant(kSmiTagSize + kSmiShiftSize);
@@ skipping 125 matching lines @@
 
 
 void SimplifiedLowering::DoChangeFloat64ToTagged(Node* node, Node* effect,
                                                  Node* control) {
   return;  // TODO(titzer): need to call runtime to allocate in one branch
 }
 
 
 void SimplifiedLowering::DoChangeBoolToBit(Node* node, Node* effect,
                                            Node* control) {
-  Node* val = node->InputAt(0);
-  Operator* op =
-      kPointerSize == 8 ? machine()->Word64Equal() : machine()->Word32Equal();
-  Node* cmp = graph()->NewNode(op, val, jsgraph()->TrueConstant());
+  Node* cmp = graph()->NewNode(machine()->WordEqual(), node->InputAt(0),
+                               jsgraph()->TrueConstant());
   node->ReplaceUses(cmp);
 }
 
 
 void SimplifiedLowering::DoChangeBitToBool(Node* node, Node* effect,
                                            Node* control) {
   Node* val = node->InputAt(0);
   Node* branch = graph()->NewNode(common()->Branch(), val, control);
 
   // true branch.
@@ skipping 15 matching lines @@
     Type* type) {
   // TODO(turbofan): skip write barriers for Smis, etc.
   if (base_is_tagged == kTaggedBase && representation == kMachineTagged) {
     // Write barriers are only for writes into heap objects (i.e. tagged base).
     return kFullWriteBarrier;
   }
   return kNoWriteBarrier;
 }
 
 
-void SimplifiedLowering::DoLoadField(Node* node, Node* effect, Node* control) {
+void SimplifiedLowering::DoLoadField(Node* node) {
   const FieldAccess& access = FieldAccessOf(node->op());
   node->set_op(machine_.Load(access.representation));
   Node* offset = jsgraph()->Int32Constant(access.offset - access.tag());
   node->InsertInput(zone(), 1, offset);
 }
 
 
-void SimplifiedLowering::DoStoreField(Node* node, Node* effect, Node* control) {
+void SimplifiedLowering::DoStoreField(Node* node) {
   const FieldAccess& access = FieldAccessOf(node->op());
   WriteBarrierKind kind = ComputeWriteBarrierKind(
       access.base_is_tagged, access.representation, access.type);
   node->set_op(machine_.Store(access.representation, kind));
   Node* offset = jsgraph()->Int32Constant(access.offset - access.tag());
   node->InsertInput(zone(), 1, offset);
 }
 
 
 Node* SimplifiedLowering::ComputeIndex(const ElementAccess& access,
@@ skipping 24 matching lines @@
     index = graph()->NewNode(machine()->Int32Mul(),
                              jsgraph()->Int32Constant(element_size), index);
   }
   int fixed_offset = access.header_size - access.tag();
   if (fixed_offset == 0) return index;
   return graph()->NewNode(machine()->Int32Add(),
                           jsgraph()->Int32Constant(fixed_offset), index);
 }
 
 
-void SimplifiedLowering::DoLoadElement(Node* node, Node* effect,
-                                       Node* control) {
+void SimplifiedLowering::DoLoadElement(Node* node) {
   const ElementAccess& access = ElementAccessOf(node->op());
   node->set_op(machine_.Load(access.representation));
   node->ReplaceInput(1, ComputeIndex(access, node->InputAt(1)));
 }
 
 
-void SimplifiedLowering::DoStoreElement(Node* node, Node* effect,
-                                        Node* control) {
+void SimplifiedLowering::DoStoreElement(Node* node) {
   const ElementAccess& access = ElementAccessOf(node->op());
   WriteBarrierKind kind = ComputeWriteBarrierKind(
       access.base_is_tagged, access.representation, access.type);
   node->set_op(machine_.Store(access.representation, kind));
   node->ReplaceInput(1, ComputeIndex(access, node->InputAt(1)));
 }
 
 
-void SimplifiedLowering::Lower(Node* node) {
-  Node* start = graph()->start();
+void SimplifiedLowering::Lower(Node* node) {}

[rossberg, 2014/08/06 13:04:07]

What's this still good for?

+
+
+void SimplifiedLowering::LowerChange(Node* node, Node* effect, Node* control) {
   switch (node->opcode()) {
-    case IrOpcode::kBooleanNot:
-    case IrOpcode::kNumberEqual:
-    case IrOpcode::kNumberLessThan:
-    case IrOpcode::kNumberLessThanOrEqual:
-    case IrOpcode::kNumberAdd:
-    case IrOpcode::kNumberSubtract:
-    case IrOpcode::kNumberMultiply:
-    case IrOpcode::kNumberDivide:
-    case IrOpcode::kNumberModulus:
-    case IrOpcode::kNumberToInt32:
-    case IrOpcode::kNumberToUint32:
-    case IrOpcode::kReferenceEqual:
-    case IrOpcode::kStringEqual:
-    case IrOpcode::kStringLessThan:
-    case IrOpcode::kStringLessThanOrEqual:
-    case IrOpcode::kStringAdd:
-      break;
     case IrOpcode::kChangeTaggedToInt32:
-      DoChangeTaggedToUI32(node, start, start, true);
+      DoChangeTaggedToUI32(node, effect, control, true);
       break;
     case IrOpcode::kChangeTaggedToUint32:
-      DoChangeTaggedToUI32(node, start, start, false);
+      DoChangeTaggedToUI32(node, effect, control, false);
       break;
     case IrOpcode::kChangeTaggedToFloat64:
-      DoChangeTaggedToFloat64(node, start, start);
+      DoChangeTaggedToFloat64(node, effect, control);
       break;
     case IrOpcode::kChangeInt32ToTagged:
-      DoChangeUI32ToTagged(node, start, start, true);
+      DoChangeUI32ToTagged(node, effect, control, true);
       break;
     case IrOpcode::kChangeUint32ToTagged:
-      DoChangeUI32ToTagged(node, start, start, false);
+      DoChangeUI32ToTagged(node, effect, control, false);
       break;
     case IrOpcode::kChangeFloat64ToTagged:
-      DoChangeFloat64ToTagged(node, start, start);
+      DoChangeFloat64ToTagged(node, effect, control);
       break;
     case IrOpcode::kChangeBoolToBit:
-      DoChangeBoolToBit(node, start, start);
+      DoChangeBoolToBit(node, effect, control);
       break;
     case IrOpcode::kChangeBitToBool:
-      DoChangeBitToBool(node, start, start);
-      break;
-    case IrOpcode::kLoadField:

[rossberg, 2014/08/06 13:04:07]

Why did these cases disappear?

-      DoLoadField(node, start, start);
-      break;
-    case IrOpcode::kStoreField:
-      DoStoreField(node, start, start);
-      break;
-    case IrOpcode::kLoadElement:
-      DoLoadElement(node, start, start);
-      break;
-    case IrOpcode::kStoreElement:
-      DoStoreElement(node, start, start);
+      DoChangeBitToBool(node, effect, control);
       break;
     default:
+      UNREACHABLE();
       break;
   }
 }
 
 }  // namespace compiler
 }  // namespace internal
 }  // namespace v8