1 /*
2  * Copyright (C) 2016 The Android Open Source Project
3  *
4  * Licensed under the Apache License, Version 2.0 (the "License");
5  * you may not use this file except in compliance with the License.
6  * You may obtain a copy of the License at
7  *
8  *      http://www.apache.org/licenses/LICENSE-2.0
9  *
10  * Unless required by applicable law or agreed to in writing, software
11  * distributed under the License is distributed on an "AS IS" BASIS,
12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13  * See the License for the specific language governing permissions and
14  * limitations under the License.
15  */
16 
17 #ifndef ART_COMPILER_OPTIMIZING_SCHEDULER_H_
18 #define ART_COMPILER_OPTIMIZING_SCHEDULER_H_
19 
20 #include <fstream>
21 
22 #include "base/time_utils.h"
23 #include "driver/compiler_driver.h"
24 #include "nodes.h"
25 #include "optimization.h"
26 
27 namespace art {
28 
29 // General description of instruction scheduling.
30 //
31 // This pass tries to improve the quality of the generated code by reordering
32 // instructions in the graph to avoid execution delays caused by execution
33 // dependencies.
34 // Currently, scheduling is performed at the block level, so no `HInstruction`
35 // ever leaves its block in this pass.
36 //
37 // The scheduling process iterates through blocks in the graph. For blocks that
38 // we can and want to schedule:
39 // 1) Build a dependency graph for instructions.
40 //    It includes data dependencies (inputs/uses), but also environment
41 //    dependencies and side-effect dependencies.
42 // 2) Schedule the dependency graph.
43 //    This is a topological sort of the dependency graph, using heuristics to
44 //    decide what node to scheduler first when there are multiple candidates.
45 //
46 // A few factors impacting the quality of the scheduling are:
47 // - The heuristics used to decide what node to schedule in the topological sort
48 //   when there are multiple valid candidates. There is a wide range of
49 //   complexity possible here, going from a simple model only considering
50 //   latencies, to a super detailed CPU pipeline model.
51 // - Fewer dependencies in the dependency graph give more freedom for the
52 //   scheduling heuristics. For example de-aliasing can allow possibilities for
53 //   reordering of memory accesses.
54 // - The level of abstraction of the IR. It is easier to evaluate scheduling for
55 //   IRs that translate to a single assembly instruction than for IRs
56 //   that generate multiple assembly instructions or generate different code
57 //   depending on properties of the IR.
58 // - Scheduling is performed before register allocation, it is not aware of the
59 //   impact of moving instructions on register allocation.
60 //
61 //
62 // The scheduling code uses the terms predecessors, successors, and dependencies.
63 // This can be confusing at times, so here are clarifications.
64 // These terms are used from the point of view of the program dependency graph. So
65 // the inputs of an instruction are part of its dependencies, and hence part its
66 // predecessors. So the uses of an instruction are (part of) its successors.
67 // (Side-effect dependencies can yield predecessors or successors that are not
68 // inputs or uses.)
69 //
70 // Here is a trivial example. For the Java code:
71 //
72 //    int a = 1 + 2;
73 //
74 // we would have the instructions
75 //
76 //    i1 HIntConstant 1
77 //    i2 HIntConstant 2
78 //    i3 HAdd [i1,i2]
79 //
80 // `i1` and `i2` are predecessors of `i3`.
81 // `i3` is a successor of `i1` and a successor of `i2`.
82 // In a scheduling graph for this code we would have three nodes `n1`, `n2`,
83 // and `n3` (respectively for instructions `i1`, `i1`, and `i3`).
84 // Conceptually the program dependency graph for this would contain two edges
85 //
86 //    n1 -> n3
87 //    n2 -> n3
88 //
89 // Since we schedule backwards (starting from the last instruction in each basic
90 // block), the implementation of nodes keeps a list of pointers their
91 // predecessors. So `n3` would keep pointers to its predecessors `n1` and `n2`.
92 //
93 // Node dependencies are also referred to from the program dependency graph
94 // point of view: we say that node `B` immediately depends on `A` if there is an
95 // edge from `A` to `B` in the program dependency graph. `A` is a predecessor of
96 // `B`, `B` is a successor of `A`. In the example above `n3` depends on `n1` and
97 // `n2`.
98 // Since nodes in the scheduling graph keep a list of their predecessors, node
99 // `B` will have a pointer to its predecessor `A`.
100 // As we schedule backwards, `B` will be selected for scheduling before `A` is.
101 //
102 // So the scheduling for the example above could happen as follow
103 //
104 //    |---------------------------+------------------------|
105 //    | candidates for scheduling | instructions scheduled |
106 //    | --------------------------+------------------------|
107 //
108 // The only node without successors is `n3`, so it is the only initial
109 // candidate.
110 //
111 //    | n3                        | (none)                 |
112 //
113 // We schedule `n3` as the last (and only) instruction. All its predecessors
114 // that do not have any unscheduled successors become candidate. That is, `n1`
115 // and `n2` become candidates.
116 //
117 //    | n1, n2                    | n3                     |
118 //
119 // One of the candidates is selected. In practice this is where scheduling
120 // heuristics kick in, to decide which of the candidates should be selected.
121 // In this example, let it be `n1`. It is scheduled before previously scheduled
122 // nodes (in program order). There are no other nodes to add to the list of
123 // candidates.
124 //
125 //    | n2                        | n1                     |
126 //    |                           | n3                     |
127 //
128 // The only candidate available for scheduling is `n2`. Schedule it before
129 // (in program order) the previously scheduled nodes.
130 //
131 //    | (none)                    | n2                     |
132 //    |                           | n1                     |
133 //    |                           | n3                     |
134 //    |---------------------------+------------------------|
135 //
136 // So finally the instructions will be executed in the order `i2`, `i1`, and `i3`.
137 // In this trivial example, it does not matter which of `i1` and `i2` is
138 // scheduled first since they are constants. However the same process would
139 // apply if `i1` and `i2` were actual operations (for example `HMul` and `HDiv`).
140 
141 // Set to true to have instruction scheduling dump scheduling graphs to the file
142 // `scheduling_graphs.dot`. See `SchedulingGraph::DumpAsDotGraph()`.
143 static constexpr bool kDumpDotSchedulingGraphs = false;
144 
145 // Typically used as a default instruction latency.
146 static constexpr uint32_t kGenericInstructionLatency = 1;
147 
148 class HScheduler;
149 
150 /**
151  * A node representing an `HInstruction` in the `SchedulingGraph`.
152  */
153 class SchedulingNode : public ArenaObject<kArenaAllocScheduler> {
154  public:
SchedulingNode(HInstruction * instr,ArenaAllocator * arena,bool is_scheduling_barrier)155   SchedulingNode(HInstruction* instr, ArenaAllocator* arena, bool is_scheduling_barrier)
156       : latency_(0),
157         internal_latency_(0),
158         critical_path_(0),
159         instruction_(instr),
160         is_scheduling_barrier_(is_scheduling_barrier),
161         data_predecessors_(arena->Adapter(kArenaAllocScheduler)),
162         other_predecessors_(arena->Adapter(kArenaAllocScheduler)),
163         num_unscheduled_successors_(0) {
164     data_predecessors_.reserve(kPreallocatedPredecessors);
165   }
166 
AddDataPredecessor(SchedulingNode * predecessor)167   void AddDataPredecessor(SchedulingNode* predecessor) {
168     data_predecessors_.push_back(predecessor);
169     predecessor->num_unscheduled_successors_++;
170   }
171 
AddOtherPredecessor(SchedulingNode * predecessor)172   void AddOtherPredecessor(SchedulingNode* predecessor) {
173     other_predecessors_.push_back(predecessor);
174     predecessor->num_unscheduled_successors_++;
175   }
176 
DecrementNumberOfUnscheduledSuccessors()177   void DecrementNumberOfUnscheduledSuccessors() {
178     num_unscheduled_successors_--;
179   }
180 
MaybeUpdateCriticalPath(uint32_t other_critical_path)181   void MaybeUpdateCriticalPath(uint32_t other_critical_path) {
182     critical_path_ = std::max(critical_path_, other_critical_path);
183   }
184 
HasUnscheduledSuccessors()185   bool HasUnscheduledSuccessors() const {
186     return num_unscheduled_successors_ != 0;
187   }
188 
GetInstruction()189   HInstruction* GetInstruction() const { return instruction_; }
GetLatency()190   uint32_t GetLatency() const { return latency_; }
SetLatency(uint32_t latency)191   void SetLatency(uint32_t latency) { latency_ = latency; }
GetInternalLatency()192   uint32_t GetInternalLatency() const { return internal_latency_; }
SetInternalLatency(uint32_t internal_latency)193   void SetInternalLatency(uint32_t internal_latency) { internal_latency_ = internal_latency; }
GetCriticalPath()194   uint32_t GetCriticalPath() const { return critical_path_; }
IsSchedulingBarrier()195   bool IsSchedulingBarrier() const { return is_scheduling_barrier_; }
GetDataPredecessors()196   const ArenaVector<SchedulingNode*>& GetDataPredecessors() const { return data_predecessors_; }
GetOtherPredecessors()197   const ArenaVector<SchedulingNode*>& GetOtherPredecessors() const { return other_predecessors_; }
198 
199  private:
200   // The latency of this node. It represents the latency between the moment the
201   // last instruction for this node has executed to the moment the result
202   // produced by this node is available to users.
203   uint32_t latency_;
204   // This represents the time spent *within* the generated code for this node.
205   // It should be zero for nodes that only generate a single instruction.
206   uint32_t internal_latency_;
207 
208   // The critical path from this instruction to the end of scheduling. It is
209   // used by the scheduling heuristics to measure the priority of this instruction.
210   // It is defined as
211   //     critical_path_ = latency_ + max((use.internal_latency_ + use.critical_path_) for all uses)
212   // (Note that here 'uses' is equivalent to 'data successors'. Also see comments in
213   // `HScheduler::Schedule(SchedulingNode* scheduling_node)`).
214   uint32_t critical_path_;
215 
216   // The instruction that this node represents.
217   HInstruction* const instruction_;
218 
219   // If a node is scheduling barrier, other nodes cannot be scheduled before it.
220   const bool is_scheduling_barrier_;
221 
222   // The lists of predecessors. They cannot be scheduled before this node. Once
223   // this node is scheduled, we check whether any of its predecessors has become a
224   // valid candidate for scheduling.
225   // Predecessors in `data_predecessors_` are data dependencies. Those in
226   // `other_predecessors_` contain side-effect dependencies, environment
227   // dependencies, and scheduling barrier dependencies.
228   ArenaVector<SchedulingNode*> data_predecessors_;
229   ArenaVector<SchedulingNode*> other_predecessors_;
230 
231   // The number of unscheduled successors for this node. This number is
232   // decremented as successors are scheduled. When it reaches zero this node
233   // becomes a valid candidate to schedule.
234   uint32_t num_unscheduled_successors_;
235 
236   static constexpr size_t kPreallocatedPredecessors = 4;
237 };
238 
239 /*
240  * Directed acyclic graph for scheduling.
241  */
242 class SchedulingGraph : public ValueObject {
243  public:
SchedulingGraph(const HScheduler * scheduler,ArenaAllocator * arena)244   SchedulingGraph(const HScheduler* scheduler, ArenaAllocator* arena)
245       : scheduler_(scheduler),
246         arena_(arena),
247         contains_scheduling_barrier_(false),
248         nodes_map_(arena_->Adapter(kArenaAllocScheduler)) {}
249 
250   SchedulingNode* AddNode(HInstruction* instr, bool is_scheduling_barrier = false) {
251     SchedulingNode* node = new (arena_) SchedulingNode(instr, arena_, is_scheduling_barrier);
252     nodes_map_.Insert(std::make_pair(instr, node));
253     contains_scheduling_barrier_ |= is_scheduling_barrier;
254     AddDependencies(instr, is_scheduling_barrier);
255     return node;
256   }
257 
Clear()258   void Clear() {
259     nodes_map_.Clear();
260     contains_scheduling_barrier_ = false;
261   }
262 
GetNode(const HInstruction * instr)263   SchedulingNode* GetNode(const HInstruction* instr) const {
264     auto it = nodes_map_.Find(instr);
265     if (it == nodes_map_.end()) {
266       return nullptr;
267     } else {
268       return it->second;
269     }
270   }
271 
272   bool IsSchedulingBarrier(const HInstruction* instruction) const;
273 
274   bool HasImmediateDataDependency(const SchedulingNode* node, const SchedulingNode* other) const;
275   bool HasImmediateDataDependency(const HInstruction* node, const HInstruction* other) const;
276   bool HasImmediateOtherDependency(const SchedulingNode* node, const SchedulingNode* other) const;
277   bool HasImmediateOtherDependency(const HInstruction* node, const HInstruction* other) const;
278 
Size()279   size_t Size() const {
280     return nodes_map_.Size();
281   }
282 
283   // Dump the scheduling graph, in dot file format, appending it to the file
284   // `scheduling_graphs.dot`.
285   void DumpAsDotGraph(const std::string& description,
286                       const ArenaVector<SchedulingNode*>& initial_candidates);
287 
288  protected:
289   void AddDependency(SchedulingNode* node, SchedulingNode* dependency, bool is_data_dependency);
AddDataDependency(SchedulingNode * node,SchedulingNode * dependency)290   void AddDataDependency(SchedulingNode* node, SchedulingNode* dependency) {
291     AddDependency(node, dependency, /*is_data_dependency*/true);
292   }
AddOtherDependency(SchedulingNode * node,SchedulingNode * dependency)293   void AddOtherDependency(SchedulingNode* node, SchedulingNode* dependency) {
294     AddDependency(node, dependency, /*is_data_dependency*/false);
295   }
296 
297   // Add dependencies nodes for the given `HInstruction`: inputs, environments, and side-effects.
298   void AddDependencies(HInstruction* instruction, bool is_scheduling_barrier = false);
299 
300   const HScheduler* const scheduler_;
301 
302   ArenaAllocator* const arena_;
303 
304   bool contains_scheduling_barrier_;
305 
306   ArenaHashMap<const HInstruction*, SchedulingNode*> nodes_map_;
307 };
308 
309 /*
310  * The visitors derived from this base class are used by schedulers to evaluate
311  * the latencies of `HInstruction`s.
312  */
313 class SchedulingLatencyVisitor : public HGraphDelegateVisitor {
314  public:
315   // This class and its sub-classes will never be used to drive a visit of an
316   // `HGraph` but only to visit `HInstructions` one at a time, so we do not need
317   // to pass a valid graph to `HGraphDelegateVisitor()`.
SchedulingLatencyVisitor()318   SchedulingLatencyVisitor()
319       : HGraphDelegateVisitor(nullptr),
320         last_visited_latency_(0),
321         last_visited_internal_latency_(0) {}
322 
VisitInstruction(HInstruction * instruction)323   void VisitInstruction(HInstruction* instruction) OVERRIDE {
324     LOG(FATAL) << "Error visiting " << instruction->DebugName() << ". "
325         "Architecture-specific scheduling latency visitors must handle all instructions"
326         " (potentially by overriding the generic `VisitInstruction()`.";
327     UNREACHABLE();
328   }
329 
Visit(HInstruction * instruction)330   void Visit(HInstruction* instruction) {
331     instruction->Accept(this);
332   }
333 
CalculateLatency(SchedulingNode * node)334   void CalculateLatency(SchedulingNode* node) {
335     // By default nodes have no internal latency.
336     last_visited_internal_latency_ = 0;
337     Visit(node->GetInstruction());
338   }
339 
GetLastVisitedLatency()340   uint32_t GetLastVisitedLatency() const { return last_visited_latency_; }
GetLastVisitedInternalLatency()341   uint32_t GetLastVisitedInternalLatency() const { return last_visited_internal_latency_; }
342 
343  protected:
344   // The latency of the most recent visited SchedulingNode.
345   // This is for reporting the latency value to the user of this visitor.
346   uint32_t last_visited_latency_;
347   // This represents the time spent *within* the generated code for the most recent visited
348   // SchedulingNode. This is for reporting the internal latency value to the user of this visitor.
349   uint32_t last_visited_internal_latency_;
350 };
351 
352 class SchedulingNodeSelector : public ArenaObject<kArenaAllocScheduler> {
353  public:
354   virtual SchedulingNode* PopHighestPriorityNode(ArenaVector<SchedulingNode*>* nodes,
355                                                  const SchedulingGraph& graph) = 0;
~SchedulingNodeSelector()356   virtual ~SchedulingNodeSelector() {}
357  protected:
DeleteNodeAtIndex(ArenaVector<SchedulingNode * > * nodes,size_t index)358   static void DeleteNodeAtIndex(ArenaVector<SchedulingNode*>* nodes, size_t index) {
359     (*nodes)[index] = nodes->back();
360     nodes->pop_back();
361   }
362 };
363 
364 /*
365  * Select a `SchedulingNode` at random within the candidates.
366  */
367 class RandomSchedulingNodeSelector : public SchedulingNodeSelector {
368  public:
RandomSchedulingNodeSelector()369   explicit RandomSchedulingNodeSelector() : seed_(0) {
370     seed_  = static_cast<uint32_t>(NanoTime());
371     srand(seed_);
372   }
373 
PopHighestPriorityNode(ArenaVector<SchedulingNode * > * nodes,const SchedulingGraph & graph)374   SchedulingNode* PopHighestPriorityNode(ArenaVector<SchedulingNode*>* nodes,
375                                          const SchedulingGraph& graph) OVERRIDE {
376     UNUSED(graph);
377     DCHECK(!nodes->empty());
378     size_t select = rand_r(&seed_) % nodes->size();
379     SchedulingNode* select_node = (*nodes)[select];
380     DeleteNodeAtIndex(nodes, select);
381     return select_node;
382   }
383 
384   uint32_t seed_;
385 };
386 
387 /*
388  * Select a `SchedulingNode` according to critical path information,
389  * with heuristics to favor certain instruction patterns like materialized condition.
390  */
391 class CriticalPathSchedulingNodeSelector : public SchedulingNodeSelector {
392  public:
CriticalPathSchedulingNodeSelector()393   CriticalPathSchedulingNodeSelector() : prev_select_(nullptr) {}
394 
395   SchedulingNode* PopHighestPriorityNode(ArenaVector<SchedulingNode*>* nodes,
396                                          const SchedulingGraph& graph) OVERRIDE;
397 
398  protected:
399   SchedulingNode* GetHigherPrioritySchedulingNode(SchedulingNode* candidate,
400                                                   SchedulingNode* check) const;
401 
402   SchedulingNode* SelectMaterializedCondition(ArenaVector<SchedulingNode*>* nodes,
403                                                const SchedulingGraph& graph) const;
404 
405  private:
406   const SchedulingNode* prev_select_;
407 };
408 
409 class HScheduler {
410  public:
HScheduler(ArenaAllocator * arena,SchedulingLatencyVisitor * latency_visitor,SchedulingNodeSelector * selector)411   HScheduler(ArenaAllocator* arena,
412              SchedulingLatencyVisitor* latency_visitor,
413              SchedulingNodeSelector* selector)
414       : arena_(arena),
415         latency_visitor_(latency_visitor),
416         selector_(selector),
417         only_optimize_loop_blocks_(true),
418         scheduling_graph_(this, arena),
419         cursor_(nullptr),
420         candidates_(arena_->Adapter(kArenaAllocScheduler)) {}
~HScheduler()421   virtual ~HScheduler() {}
422 
423   void Schedule(HGraph* graph);
424 
SetOnlyOptimizeLoopBlocks(bool loop_only)425   void SetOnlyOptimizeLoopBlocks(bool loop_only) { only_optimize_loop_blocks_ = loop_only; }
426 
427   // Instructions can not be rescheduled across a scheduling barrier.
428   virtual bool IsSchedulingBarrier(const HInstruction* instruction) const;
429 
430  protected:
431   void Schedule(HBasicBlock* block);
432   void Schedule(SchedulingNode* scheduling_node);
433   void Schedule(HInstruction* instruction);
434 
435   // Any instruction returning `false` via this method will prevent its
436   // containing basic block from being scheduled.
437   // This method is used to restrict scheduling to instructions that we know are
438   // safe to handle.
439   virtual bool IsSchedulable(const HInstruction* instruction) const;
440   bool IsSchedulable(const HBasicBlock* block) const;
441 
CalculateLatency(SchedulingNode * node)442   void CalculateLatency(SchedulingNode* node) {
443     latency_visitor_->CalculateLatency(node);
444     node->SetLatency(latency_visitor_->GetLastVisitedLatency());
445     node->SetInternalLatency(latency_visitor_->GetLastVisitedInternalLatency());
446   }
447 
448   ArenaAllocator* const arena_;
449   SchedulingLatencyVisitor* const latency_visitor_;
450   SchedulingNodeSelector* const selector_;
451   bool only_optimize_loop_blocks_;
452 
453   // We instantiate the members below as part of this class to avoid
454   // instantiating them locally for every chunk scheduled.
455   SchedulingGraph scheduling_graph_;
456   // A pointer indicating where the next instruction to be scheduled will be inserted.
457   HInstruction* cursor_;
458   // The list of candidates for scheduling. A node becomes a candidate when all
459   // its predecessors have been scheduled.
460   ArenaVector<SchedulingNode*> candidates_;
461 
462  private:
463   DISALLOW_COPY_AND_ASSIGN(HScheduler);
464 };
465 
IsSchedulingBarrier(const HInstruction * instruction)466 inline bool SchedulingGraph::IsSchedulingBarrier(const HInstruction* instruction) const {
467   return scheduler_->IsSchedulingBarrier(instruction);
468 }
469 
470 class HInstructionScheduling : public HOptimization {
471  public:
HInstructionScheduling(HGraph * graph,InstructionSet instruction_set)472   HInstructionScheduling(HGraph* graph, InstructionSet instruction_set)
473       : HOptimization(graph, kInstructionScheduling),
474         instruction_set_(instruction_set) {}
475 
Run()476   void Run() {
477     Run(/*only_optimize_loop_blocks*/ true, /*schedule_randomly*/ false);
478   }
479   void Run(bool only_optimize_loop_blocks, bool schedule_randomly);
480 
481   static constexpr const char* kInstructionScheduling = "scheduler";
482 
483   const InstructionSet instruction_set_;
484 
485  private:
486   DISALLOW_COPY_AND_ASSIGN(HInstructionScheduling);
487 };
488 
489 }  // namespace art
490 
491 #endif  // ART_COMPILER_OPTIMIZING_SCHEDULER_H_
492