[WIP][turbofan] Instruction scheduler for Turbofan.

src/compiler/instruction-scheduler.h

+// Copyright 2015 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.
+
+#ifndef V8_COMPILER_INSTRUCTION_SCHEDULER_H_
+#define V8_COMPILER_INSTRUCTION_SCHEDULER_H_
+
+#include "src/compiler/instruction.h"
+#include "src/zone-containers.h"
+
+namespace v8 {
+namespace internal {
+namespace compiler {
+
+// A set of flags describing properties of the instructions so that the
+// scheduler is aware of dependencies between instructions.
+enum ArchOpcodeFlags {
+  kNoOpcodeFlags = 0,
+  kIsBlockTerminator = 1,  // The instruction marks the end of a basic block
+                           // e.g.: jump and return instructions.
+  kHasSideEffect = 2,      // The instruction has some side effects (memory
+                           // store, function call...)
+  kIsLoadOperation = 4,    // The instruction is a memory load.
+};
+
+
+class InstructionScheduler final : public ZoneObject {
+ public:
+  InstructionScheduler(Zone* zone, InstructionSequence* sequence);
+
+  void StartBlock(RpoNumber rpo);
+  void EndBlock(RpoNumber rpo);
+
+  void AddInstruction(Instruction* instr);
+
+ private:
+  // A scheduling graph node.
+  // Represent an instruction and their dependencies.
+  class ScheduleGraphNode: public ZoneObject {
+   public:
+    ScheduleGraphNode(Zone* zone, Instruction* instr);
+
+    // Mark the instruction represented by 'node' as a dependecy of this one.
+    // The current instruction will be registered as an unscheduled predecessor
+    // of 'node' (i.e. it must be scheduled before 'node').
+    void AddSuccessor(ScheduleGraphNode* node);
+
+    // Check if all the predecessors of this instruction have been scheduled.
+    bool HasUnscheduledPredecessor() {
+      return unscheduled_predecessors_count_ != 0;
+    }
+
+    // Record that we have scheduled one of the predecessor of this node.

[Jarin, 2015/11/24 11:43:03]

predecessor -> predecessors.

[baptiste.afsa1, 2015/12/01 09:08:42]

Done.

+    void DropUnscheduledPredecessor() {
+      DCHECK(unscheduled_predecessors_count_ > 0);
+      unscheduled_predecessors_count_--;
+    }
+
+    Instruction* instruction() { return instr_; }
+    ZoneDeque<ScheduleGraphNode*>& successors() { return successors_; }
+    int latency() const { return latency_; }
+
+    int total_latency() const { return total_latency_; }
+    void set_total_latency(int latency) { total_latency_ = latency; }
+
+    int start_cycle() const { return start_cycle_; }
+    void set_start_cycle(int start_cycle) { start_cycle_ = start_cycle; }
+
+   private:
+    Instruction* instr_;
+    ZoneDeque<ScheduleGraphNode*> successors_;
+
+    // Number of unscheduled predecessors for this node.
+    int unscheduled_predecessors_count_;
+
+    // Estimate of the instruction latency (the number of cycles it takes for
+    // instruction to complete).
+    int latency_;
+
+    // The sum of all the latencies on the path from this node to the end of
+    // the graph (i.e. a node with no successor).
+    int total_latency_;
+
+    // The scheduler keeps a nominal cycle count to keep track of when the
+    // result of an instruction is available. This field is updated by the
+    // scheduler to indicate when the value of all the operands of this
+    // instruction will be available.
+    int start_cycle_;
+  };
+
+  // Compare the two nodes and return true if node1 is a better candidate than
+  // node2 (i.e. node1 should be scheduled before node2).
+  bool CompareNodes(ScheduleGraphNode *node1, ScheduleGraphNode *node2) const;
+
+  // Perform scheduling for the current block.
+  void ScheduleBlock();
+
+  // Return the scheduling properties of the given instruction.
+  int GetInstructionFlags(const Instruction* instr) const;
+  int GetTargetInstructionFlags(const Instruction* instr) const;
+
+  // Return true if instr2 uses any value defined by instr1.
+  bool HasOperandDependency(const Instruction* instr1,
+                            const Instruction* instr2) const;
+
+  // Return true if the instruction is a basic block terminator.
+  bool IsBlockTerminator(const Instruction* instr) const;
+
+  // Check whether the given instruction has side effects (e.g. function call,
+  // memory store).
+  bool HasSideEffect(const Instruction* instr) const {
+    return GetInstructionFlags(instr) & kHasSideEffect;
+  }
+
+  // Return true if the instruction is a memory load.
+  bool IsLoadOperation(const Instruction* instr) const {
+    return GetInstructionFlags(instr) & kIsLoadOperation;
+  }
+
+  // Identify nops used as a definition point for live-in registers at
+  // function entry.
+  bool IsLiveInNop(const Instruction* instr) const {

[Jarin, 2015/11/24 11:43:03]

Funny name. Could we invent a better name? For example IsFixedRegisterParameter.


(Yes, it is unfortunate that we misused kArchNop for this.)

[baptiste.afsa1, 2015/12/01 09:08:42]

Done.

+    return (instr->arch_opcode() == kArchNop) &&
+      (instr->OutputCount() == 1) &&
+      (instr->OutputAt(0)->IsUnallocated()) &&
+      UnallocatedOperand::cast(instr->OutputAt(0))->HasFixedRegisterPolicy();
+  }
+
+  void ComputeTotalLatencies();
+
+  static int GetInstructionLatency(const Instruction* instr);
+
+  Zone* zone() { return zone_; }
+  InstructionSequence* sequence() { return sequence_; }
+
+  Zone* zone_;
+  InstructionSequence* sequence_;
+  ZoneVector<ScheduleGraphNode*> graph_;
+
+  // Last side effect instruction encountered while building the graph.
+  ScheduleGraphNode* last_side_effect_instr_;
+
+  // Set of load instructions encountered since the last side effect instruction
+  // which will be added as predecessors of the next instruction with side
+  // effects.
+  ZoneVector<ScheduleGraphNode*> pending_loads_;
+
+  // Live-in register markers are nop instructions which are emitted at the
+  // beginning of a basic block so that the register allocator will find a
+  // defining instruction for live-in values. They must not be moved.
+  // All these nops are chained together and added as a predecessor of every
+  // other instructions in the basic block.
+  ScheduleGraphNode* last_live_in_reg_marker_;
+};
+
+}  // namespace compiler
+}  // namespace internal
+}  // namespace v8
+
+#endif  // V8_COMPILER_INSTRUCTION_SCHEDULER_H_