[turbofan] Optimize certain chains of Branch into a Switch.

This adds a new ControlFlowOptimizer that - for now - recognizes chains
of Branches generated by the SwitchBuilder for a subset of javascript
switches into Switch nodes. Those Switch nodes are then lowered to
either table or lookup switches.

Also rename Case to IfValue (and introduce IfDefault) for consistency.

BUG=v8:3872
LOG=n

Review URL: https://codereview.chromium.org/931623002

Cr-Commit-Position: refs/heads/master@{#26691}

BUILD.gn

@@ -527,6 +527,8 @@ source_set("v8_base") {
     "src/compiler/control-builders.cc",
     "src/compiler/control-builders.h",
     "src/compiler/control-equivalence.h",
+    "src/compiler/control-flow-optimizer.cc",
+    "src/compiler/control-flow-optimizer.h",
     "src/compiler/control-reducer.cc",
     "src/compiler/control-reducer.h",
     "src/compiler/diamond.h",

src/compiler/arm/code-generator-arm.cc

@@ -336,8 +336,12 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
       AssembleArchJump(i.InputRpo(0));
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
-    case kArchSwitch:
-      AssembleArchSwitch(instr);
+    case kArchLookupSwitch:
+      AssembleArchLookupSwitch(instr);
+      DCHECK_EQ(LeaveCC, i.OutputSBit());
+      break;
+    case kArchTableSwitch:
+      AssembleArchTableSwitch(instr);
       DCHECK_EQ(LeaveCC, i.OutputSBit());
       break;
     case kArchNop:
@@ -741,18 +745,6 @@ void CodeGenerator::AssembleArchJump(BasicBlock::RpoNumber target) {
 }
 
 
-void CodeGenerator::AssembleArchSwitch(Instruction* instr) {
-  ArmOperandConverter i(this, instr);
-  int const kNumLabels = static_cast<int>(instr->InputCount() - 1);
-  __ BlockConstPoolFor(kNumLabels + 2);
-  __ ldr(pc, MemOperand(pc, i.InputRegister(0), LSL, 2));
-  __ nop();
-  for (int index = 0; index < kNumLabels; ++index) {
-    __ dd(GetLabel(i.InputRpo(index + 1)));
-  }
-}
-
-
 // Assembles boolean materializations after an instruction.
 void CodeGenerator::AssembleArchBoolean(Instruction* instr,
                                         FlagsCondition condition) {
@@ -768,6 +760,31 @@ void CodeGenerator::AssembleArchBoolean(Instruction* instr,
 }
 
 
+void CodeGenerator::AssembleArchLookupSwitch(Instruction* instr) {
+  ArmOperandConverter i(this, instr);
+  Register input = i.InputRegister(0);
+  for (size_t index = 2; index < instr->InputCount(); index += 2) {
+    __ cmp(input, Operand(i.InputInt32(index + 0)));
+    __ b(eq, GetLabel(i.InputRpo(index + 1)));
+  }
+  AssembleArchJump(i.InputRpo(1));
+}
+
+
+void CodeGenerator::AssembleArchTableSwitch(Instruction* instr) {
+  ArmOperandConverter i(this, instr);
+  Register input = i.InputRegister(0);
+  size_t const case_count = instr->InputCount() - 2;
+  __ cmp(input, Operand(case_count));
+  __ BlockConstPoolFor(case_count + 2);
+  __ ldr(pc, MemOperand(pc, input, LSL, 2), lo);
+  __ b(GetLabel(i.InputRpo(1)));
+  for (size_t index = 0; index < case_count; ++index) {
+    __ dd(GetLabel(i.InputRpo(index + 2)));
+  }
+}
+
+
 void CodeGenerator::AssembleDeoptimizerCall(int deoptimization_id) {
   Address deopt_entry = Deoptimizer::GetDeoptimizationEntry(
       isolate(), deoptimization_id, Deoptimizer::LAZY);

src/compiler/arm/instruction-selector-arm.cc

@@ -1254,6 +1254,67 @@ void InstructionSelector::VisitBranch(Node* branch, BasicBlock* tbranch,
 }
 
 
+void InstructionSelector::VisitSwitch(Node* node, BasicBlock* default_branch,
+                                      BasicBlock** case_branches,
+                                      int32_t* case_values, size_t case_count,
+                                      int32_t min_value, int32_t max_value) {
+  ArmOperandGenerator g(this);
+  InstructionOperand value_operand = g.UseRegister(node->InputAt(0));
+  InstructionOperand default_operand = g.Label(default_branch);
+
+  // Note that {value_range} can be 0 if {min_value} is -2^31 and {max_value}
+  // is 2^31-1, so don't assume that it's non-zero below.
+  size_t value_range =
+      1u + bit_cast<uint32_t>(max_value) - bit_cast<uint32_t>(min_value);
+
+  // Determine whether to issue an ArchTableSwitch or an ArchLookupSwitch
+  // instruction.
+  size_t table_space_cost = 4 + value_range;
+  size_t table_time_cost = 3;
+  size_t lookup_space_cost = 3 + 2 * case_count;
+  size_t lookup_time_cost = case_count;
+  if (case_count > 0 &&
+      table_space_cost + 3 * table_time_cost <=
+          lookup_space_cost + 3 * lookup_time_cost &&
+      min_value > std::numeric_limits<int32_t>::min()) {
+    InstructionOperand index_operand = value_operand;
+    if (min_value) {
+      index_operand = g.TempRegister();
+      Emit(kArmSub | AddressingModeField::encode(kMode_Operand2_I),
+           index_operand, value_operand, g.TempImmediate(min_value));
+    }
+    size_t input_count = 2 + value_range;
+    auto* inputs = zone()->NewArray<InstructionOperand>(input_count);
+    inputs[0] = index_operand;
+    std::fill(&inputs[1], &inputs[input_count], default_operand);
+    for (size_t index = 0; index < case_count; ++index) {
+      size_t value = case_values[index] - min_value;
+      BasicBlock* branch = case_branches[index];
+      DCHECK_LE(0u, value);
+      DCHECK_LT(value + 2, input_count);
+      inputs[value + 2] = g.Label(branch);
+    }
+    Emit(kArchTableSwitch, 0, nullptr, input_count, inputs, 0, nullptr)
+        ->MarkAsControl();
+    return;
+  }
+
+  // Generate a sequence of conditional jumps.
+  size_t input_count = 2 + case_count * 2;
+  auto* inputs = zone()->NewArray<InstructionOperand>(input_count);
+  inputs[0] = value_operand;
+  inputs[1] = default_operand;
+  for (size_t index = 0; index < case_count; ++index) {
+    int32_t value = case_values[index];
+    BasicBlock* branch = case_branches[index];
+    inputs[index * 2 + 2 + 0] = g.TempImmediate(value);
+    inputs[index * 2 + 2 + 1] = g.Label(branch);
+  }
+  Emit(kArchLookupSwitch, 0, nullptr, input_count, inputs, 0, nullptr)
+      ->MarkAsControl();
+}
+
+
 void InstructionSelector::VisitWord32Equal(Node* const node) {
   FlagsContinuation cont(kEqual, node);
   Int32BinopMatcher m(node);

src/compiler/arm64/code-generator-arm64.cc

@@ -357,8 +357,11 @@ void CodeGenerator::AssembleArchInstruction(Instruction* instr) {
     case kArchJmp:
       AssembleArchJump(i.InputRpo(0));
       break;
-    case kArchSwitch:
-      AssembleArchSwitch(instr);
+    case kArchTableSwitch:
+      AssembleArchTableSwitch(instr);
+      break;
+    case kArchLookupSwitch:
+      AssembleArchLookupSwitch(instr);
       break;
     case kArchNop:
       // don't emit code for nops.
@@ -841,22 +844,6 @@ void CodeGenerator::AssembleArchJump(BasicBlock::RpoNumber target) {
 }
 
 
-void CodeGenerator::AssembleArchSwitch(Instruction* instr) {
-  Arm64OperandConverter i(this, instr);
-  UseScratchRegisterScope scope(masm());
-  Register reg = i.InputRegister(0);
-  Register tmp = scope.AcquireX();
-  Label table;
-  __ Adr(tmp, &table);
-  __ Add(tmp, tmp, Operand(reg, LSL, 2));
-  __ Br(tmp);
-  __ Bind(&table);
-  for (size_t index = 1; index < instr->InputCount(); ++index) {
-    __ B(GetLabel(i.InputRpo(index)));
-  }
-}
-
-
 // Assemble boolean materializations after this instruction.
 void CodeGenerator::AssembleArchBoolean(Instruction* instr,
                                         FlagsCondition condition) {
@@ -871,6 +858,38 @@ void CodeGenerator::AssembleArchBoolean(Instruction* instr,
 }
 
 
+void CodeGenerator::AssembleArchLookupSwitch(Instruction* instr) {
+  Arm64OperandConverter i(this, instr);
+  Register input = i.InputRegister32(0);
+  for (size_t index = 2; index < instr->InputCount(); index += 2) {
+    __ Cmp(input, i.InputInt32(index + 0));
+    __ B(eq, GetLabel(i.InputRpo(index + 1)));
+  }
+  AssembleArchJump(i.InputRpo(1));
+}
+
+
+void CodeGenerator::AssembleArchTableSwitch(Instruction* instr) {
+  Arm64OperandConverter i(this, instr);
+  UseScratchRegisterScope scope(masm());
+  Register input = i.InputRegister32(0);
+  Register temp = scope.AcquireX();
+  size_t const case_count = instr->InputCount() - 2;
+  Label table;
+  __ Cmp(input, case_count);
+  __ B(hs, GetLabel(i.InputRpo(1)));
+  __ Adr(temp, &table);
+  __ Add(temp, temp, Operand(input, UXTW, 2));
+  __ Br(temp);
+  __ StartBlockPools();
+  __ Bind(&table);
+  for (size_t index = 0; index < case_count; ++index) {
+    __ B(GetLabel(i.InputRpo(index + 2)));
+  }
+  __ EndBlockPools();
+}
+
+
 void CodeGenerator::AssembleDeoptimizerCall(int deoptimization_id) {
   Address deopt_entry = Deoptimizer::GetDeoptimizationEntry(
       isolate(), deoptimization_id, Deoptimizer::LAZY);

src/compiler/arm64/instruction-selector-arm64.cc

@@ -1391,6 +1391,67 @@ void InstructionSelector::VisitBranch(Node* branch, BasicBlock* tbranch,
 }
 
 
+void InstructionSelector::VisitSwitch(Node* node, BasicBlock* default_branch,
+                                      BasicBlock** case_branches,
+                                      int32_t* case_values, size_t case_count,
+                                      int32_t min_value, int32_t max_value) {
+  Arm64OperandGenerator g(this);
+  InstructionOperand value_operand = g.UseRegister(node->InputAt(0));
+  InstructionOperand default_operand = g.Label(default_branch);
+
+  // Note that {value_range} can be 0 if {min_value} is -2^31 and {max_value}
+  // is 2^31-1, so don't assume that it's non-zero below.
+  size_t value_range =
+      1u + bit_cast<uint32_t>(max_value) - bit_cast<uint32_t>(min_value);
+
+  // Determine whether to issue an ArchTableSwitch or an ArchLookupSwitch
+  // instruction.
+  size_t table_space_cost = 4 + value_range;
+  size_t table_time_cost = 3;
+  size_t lookup_space_cost = 3 + 2 * case_count;
+  size_t lookup_time_cost = case_count;
+  if (case_count > 0 &&
+      table_space_cost + 3 * table_time_cost <=
+          lookup_space_cost + 3 * lookup_time_cost &&
+      min_value > std::numeric_limits<int32_t>::min()) {
+    InstructionOperand index_operand = value_operand;
+    if (min_value) {
+      index_operand = g.TempRegister();
+      Emit(kArm64Sub32, index_operand, value_operand,
+           g.TempImmediate(min_value));
+    }
+    size_t input_count = 2 + value_range;
+    auto* inputs = zone()->NewArray<InstructionOperand>(input_count);
+    inputs[0] = index_operand;
+    std::fill(&inputs[1], &inputs[input_count], default_operand);
+    for (size_t index = 0; index < case_count; ++index) {
+      size_t value = case_values[index] - min_value;
+      BasicBlock* branch = case_branches[index];
+      DCHECK_LE(0u, value);
+      DCHECK_LT(value + 2, input_count);
+      inputs[value + 2] = g.Label(branch);
+    }
+    Emit(kArchTableSwitch, 0, nullptr, input_count, inputs, 0, nullptr)
+        ->MarkAsControl();
+    return;
+  }
+
+  // Generate a sequence of conditional jumps.
+  size_t input_count = 2 + case_count * 2;
+  auto* inputs = zone()->NewArray<InstructionOperand>(input_count);
+  inputs[0] = value_operand;
+  inputs[1] = default_operand;
+  for (size_t index = 0; index < case_count; ++index) {
+    int32_t value = case_values[index];
+    BasicBlock* branch = case_branches[index];
+    inputs[index * 2 + 2 + 0] = g.TempImmediate(value);
+    inputs[index * 2 + 2 + 1] = g.Label(branch);
+  }
+  Emit(kArchLookupSwitch, 0, nullptr, input_count, inputs, 0, nullptr)
+      ->MarkAsControl();
+}
+
+
 void InstructionSelector::VisitWord32Equal(Node* const node) {
   Node* const user = node;
   FlagsContinuation cont(kEqual, node);

src/compiler/code-generator.h

@@ -69,9 +69,10 @@ class CodeGenerator FINAL : public GapResolver::Assembler {
 
   void AssembleArchInstruction(Instruction* instr);
   void AssembleArchJump(BasicBlock::RpoNumber target);
-  void AssembleArchSwitch(Instruction* instr);
   void AssembleArchBranch(Instruction* instr, BranchInfo* branch);
   void AssembleArchBoolean(Instruction* instr, FlagsCondition condition);
+  void AssembleArchLookupSwitch(Instruction* instr);
+  void AssembleArchTableSwitch(Instruction* instr);
 
   void AssembleDeoptimizerCall(int deoptimization_id);
 

src/compiler/common-operator.cc

@@ -36,12 +36,6 @@ BranchHint BranchHintOf(const Operator* const op) {
 }
 
 
-size_t CaseIndexOf(const Operator* const op) {
-  DCHECK_EQ(IrOpcode::kCase, op->opcode());
-  return OpParameter<size_t>(op);
-}
-
-
 bool operator==(SelectParameters const& lhs, SelectParameters const& rhs) {
   return lhs.type() == rhs.type() && lhs.hint() == rhs.hint();
 }
@@ -121,6 +115,7 @@ size_t ProjectionIndexOf(const Operator* const op) {
   V(End, Operator::kKontrol, 0, 0, 1, 0, 0, 0)             \
   V(IfTrue, Operator::kKontrol, 0, 0, 1, 0, 0, 1)          \
   V(IfFalse, Operator::kKontrol, 0, 0, 1, 0, 0, 1)         \
+  V(IfDefault, Operator::kKontrol, 0, 0, 1, 0, 0, 1)       \
   V(Throw, Operator::kFoldable, 1, 1, 1, 0, 0, 1)          \
   V(Return, Operator::kNoThrow, 1, 1, 1, 0, 0, 1)          \
   V(OsrNormalEntry, Operator::kFoldable, 0, 1, 1, 0, 1, 1) \
@@ -256,20 +251,20 @@ const Operator* CommonOperatorBuilder::Branch(BranchHint hint) {
 
 
 const Operator* CommonOperatorBuilder::Switch(size_t control_output_count) {
-  DCHECK_GE(control_output_count, 2u);        // Disallow trivial switches.
+  DCHECK_GE(control_output_count, 3u);        // Disallow trivial switches.
   return new (zone()) Operator(               // --
       IrOpcode::kSwitch, Operator::kKontrol,  // opcode
       "Switch",                               // name
       1, 0, 1, 0, 0, control_output_count);   // counts
 }
 
 
-const Operator* CommonOperatorBuilder::Case(size_t index) {
-  return new (zone()) Operator1<size_t>(    // --
-      IrOpcode::kCase, Operator::kKontrol,  // opcode
-      "Case",                               // name
-      0, 0, 1, 0, 0, 1,                     // counts
-      index);                               // parameter
+const Operator* CommonOperatorBuilder::IfValue(int32_t index) {
+  return new (zone()) Operator1<int32_t>(      // --
+      IrOpcode::kIfValue, Operator::kKontrol,  // opcode
+      "IfValue",                               // name
+      0, 0, 1, 0, 0, 1,                        // counts
+      index);                                  // parameter
 }
 
 

src/compiler/common-operator.h

@@ -33,9 +33,6 @@ std::ostream& operator<<(std::ostream&, BranchHint);
 BranchHint BranchHintOf(const Operator* const);
 
 
-size_t CaseIndexOf(const Operator* const);
-
-
 class SelectParameters FINAL {
  public:
   explicit SelectParameters(MachineType type,
@@ -176,7 +173,8 @@ class CommonOperatorBuilder FINAL : public ZoneObject {
   const Operator* IfTrue();
   const Operator* IfFalse();
   const Operator* Switch(size_t control_output_count);
-  const Operator* Case(size_t index);
+  const Operator* IfValue(int32_t value);
+  const Operator* IfDefault();
   const Operator* Throw();
   const Operator* Return();