File: | out/../deps/v8/src/compiler/scheduler.cc |
Warning: | line 1229, column 14 Although the value stored to 'b2' is used in the enclosing expression, the value is never actually read from 'b2' |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | // Copyright 2013 the V8 project authors. All rights reserved. |
2 | // Use of this source code is governed by a BSD-style license that can be |
3 | // found in the LICENSE file. |
4 | |
5 | #include "src/compiler/scheduler.h" |
6 | |
7 | #include <iomanip> |
8 | |
9 | #include "src/base/iterator.h" |
10 | #include "src/builtins/profile-data-reader.h" |
11 | #include "src/codegen/tick-counter.h" |
12 | #include "src/compiler/common-operator.h" |
13 | #include "src/compiler/control-equivalence.h" |
14 | #include "src/compiler/graph.h" |
15 | #include "src/compiler/node-marker.h" |
16 | #include "src/compiler/node-properties.h" |
17 | #include "src/compiler/node.h" |
18 | #include "src/utils/bit-vector.h" |
19 | #include "src/zone/zone-containers.h" |
20 | |
21 | namespace v8 { |
22 | namespace internal { |
23 | namespace compiler { |
24 | |
25 | #define TRACE(...) \ |
26 | do { \ |
27 | if (FLAG_trace_turbo_scheduler) PrintF(__VA_ARGS__); \ |
28 | } while (false) |
29 | |
30 | Scheduler::Scheduler(Zone* zone, Graph* graph, Schedule* schedule, Flags flags, |
31 | size_t node_count_hint, TickCounter* tick_counter, |
32 | const ProfileDataFromFile* profile_data) |
33 | : zone_(zone), |
34 | graph_(graph), |
35 | schedule_(schedule), |
36 | flags_(flags), |
37 | scheduled_nodes_(zone), |
38 | schedule_root_nodes_(zone), |
39 | schedule_queue_(zone), |
40 | node_data_(zone), |
41 | tick_counter_(tick_counter), |
42 | profile_data_(profile_data), |
43 | common_dominator_cache_(zone) { |
44 | node_data_.reserve(node_count_hint); |
45 | node_data_.resize(graph->NodeCount(), DefaultSchedulerData()); |
46 | } |
47 | |
48 | Schedule* Scheduler::ComputeSchedule(Zone* zone, Graph* graph, Flags flags, |
49 | TickCounter* tick_counter, |
50 | const ProfileDataFromFile* profile_data) { |
51 | Zone* schedule_zone = |
52 | (flags & Scheduler::kTempSchedule) ? zone : graph->zone(); |
53 | |
54 | // Reserve 10% more space for nodes if node splitting is enabled to try to |
55 | // avoid resizing the vector since that would triple its zone memory usage. |
56 | float node_hint_multiplier = (flags & Scheduler::kSplitNodes) ? 1.1 : 1; |
57 | size_t node_count_hint = node_hint_multiplier * graph->NodeCount(); |
58 | |
59 | Schedule* schedule = |
60 | schedule_zone->New<Schedule>(schedule_zone, node_count_hint); |
61 | Scheduler scheduler(zone, graph, schedule, flags, node_count_hint, |
62 | tick_counter, profile_data); |
63 | |
64 | scheduler.BuildCFG(); |
65 | scheduler.ComputeSpecialRPONumbering(); |
66 | scheduler.GenerateDominatorTree(); |
67 | |
68 | scheduler.PrepareUses(); |
69 | scheduler.ScheduleEarly(); |
70 | scheduler.ScheduleLate(); |
71 | |
72 | scheduler.SealFinalSchedule(); |
73 | |
74 | return schedule; |
75 | } |
76 | |
77 | Scheduler::SchedulerData Scheduler::DefaultSchedulerData() { |
78 | SchedulerData def = {schedule_->start(), 0, kUnknown}; |
79 | return def; |
80 | } |
81 | |
82 | |
83 | Scheduler::SchedulerData* Scheduler::GetData(Node* node) { |
84 | return &node_data_[node->id()]; |
85 | } |
86 | |
87 | Scheduler::Placement Scheduler::InitializePlacement(Node* node) { |
88 | SchedulerData* data = GetData(node); |
89 | if (data->placement_ == kFixed) { |
90 | // Nothing to do for control nodes that have been already fixed in |
91 | // the schedule. |
92 | return data->placement_; |
93 | } |
94 | DCHECK_EQ(kUnknown, data->placement_)((void) 0); |
95 | switch (node->opcode()) { |
96 | case IrOpcode::kParameter: |
97 | case IrOpcode::kOsrValue: |
98 | // Parameters and OSR values are always fixed to the start block. |
99 | data->placement_ = kFixed; |
100 | break; |
101 | case IrOpcode::kPhi: |
102 | case IrOpcode::kEffectPhi: { |
103 | // Phis and effect phis are fixed if their control inputs are, whereas |
104 | // otherwise they are coupled to a floating control node. |
105 | Placement p = GetPlacement(NodeProperties::GetControlInput(node)); |
106 | data->placement_ = (p == kFixed ? kFixed : kCoupled); |
107 | break; |
108 | } |
109 | default: |
110 | // Control nodes that were not control-reachable from end may float. |
111 | data->placement_ = kSchedulable; |
112 | break; |
113 | } |
114 | return data->placement_; |
115 | } |
116 | |
117 | Scheduler::Placement Scheduler::GetPlacement(Node* node) { |
118 | return GetData(node)->placement_; |
119 | } |
120 | |
121 | bool Scheduler::IsLive(Node* node) { return GetPlacement(node) != kUnknown; } |
122 | |
123 | void Scheduler::UpdatePlacement(Node* node, Placement placement) { |
124 | SchedulerData* data = GetData(node); |
125 | if (data->placement_ == kUnknown) { |
126 | // We only update control nodes from {kUnknown} to {kFixed}. Ideally, we |
127 | // should check that {node} is a control node (including exceptional calls), |
128 | // but that is expensive. |
129 | DCHECK_EQ(Scheduler::kFixed, placement)((void) 0); |
130 | data->placement_ = placement; |
131 | return; |
132 | } |
133 | |
134 | switch (node->opcode()) { |
135 | case IrOpcode::kParameter: |
136 | // Parameters are fixed once and for all. |
137 | UNREACHABLE()V8_Fatal("unreachable code"); |
138 | case IrOpcode::kPhi: |
139 | case IrOpcode::kEffectPhi: { |
140 | // Phis and effect phis are coupled to their respective blocks. |
141 | DCHECK_EQ(Scheduler::kCoupled, data->placement_)((void) 0); |
142 | DCHECK_EQ(Scheduler::kFixed, placement)((void) 0); |
143 | Node* control = NodeProperties::GetControlInput(node); |
144 | BasicBlock* block = schedule_->block(control); |
145 | schedule_->AddNode(block, node); |
146 | break; |
147 | } |
148 | #define DEFINE_CONTROL_CASE(V) case IrOpcode::k##V: |
149 | CONTROL_OP_LIST(DEFINE_CONTROL_CASE)DEFINE_CONTROL_CASE(Start) DEFINE_CONTROL_CASE(Loop) DEFINE_CONTROL_CASE (Branch) DEFINE_CONTROL_CASE(Switch) DEFINE_CONTROL_CASE(IfTrue ) DEFINE_CONTROL_CASE(IfFalse) DEFINE_CONTROL_CASE(IfSuccess) DEFINE_CONTROL_CASE(IfException) DEFINE_CONTROL_CASE(IfValue ) DEFINE_CONTROL_CASE(IfDefault) DEFINE_CONTROL_CASE(Merge) DEFINE_CONTROL_CASE (Deoptimize) DEFINE_CONTROL_CASE(DeoptimizeIf) DEFINE_CONTROL_CASE (DeoptimizeUnless) DEFINE_CONTROL_CASE(TrapIf) DEFINE_CONTROL_CASE (TrapUnless) DEFINE_CONTROL_CASE(Return) DEFINE_CONTROL_CASE( TailCall) DEFINE_CONTROL_CASE(Terminate) DEFINE_CONTROL_CASE( Throw) DEFINE_CONTROL_CASE(End) |
150 | #undef DEFINE_CONTROL_CASE |
151 | { |
152 | // Control nodes force coupled uses to be placed. |
153 | for (auto use : node->uses()) { |
154 | if (GetPlacement(use) == Scheduler::kCoupled) { |
155 | DCHECK_EQ(node, NodeProperties::GetControlInput(use))((void) 0); |
156 | UpdatePlacement(use, placement); |
157 | } |
158 | } |
159 | break; |
160 | } |
161 | default: |
162 | DCHECK_EQ(Scheduler::kSchedulable, data->placement_)((void) 0); |
163 | DCHECK_EQ(Scheduler::kScheduled, placement)((void) 0); |
164 | break; |
165 | } |
166 | // Reduce the use count of the node's inputs to potentially make them |
167 | // schedulable. If all the uses of a node have been scheduled, then the node |
168 | // itself can be scheduled. |
169 | base::Optional<int> coupled_control_edge = GetCoupledControlEdge(node); |
170 | for (Edge const edge : node->input_edges()) { |
171 | DCHECK_EQ(node, edge.from())((void) 0); |
172 | if (edge.index() != coupled_control_edge) { |
173 | DecrementUnscheduledUseCount(edge.to(), node); |
174 | } |
175 | } |
176 | data->placement_ = placement; |
177 | } |
178 | |
179 | base::Optional<int> Scheduler::GetCoupledControlEdge(Node* node) { |
180 | if (GetPlacement(node) == kCoupled) { |
181 | return NodeProperties::FirstControlIndex(node); |
182 | } |
183 | return {}; |
184 | } |
185 | |
186 | void Scheduler::IncrementUnscheduledUseCount(Node* node, Node* from) { |
187 | // Tracking use counts for fixed nodes is useless. |
188 | if (GetPlacement(node) == kFixed) return; |
189 | |
190 | // Use count for coupled nodes is summed up on their control. |
191 | if (GetPlacement(node) == kCoupled) { |
192 | node = NodeProperties::GetControlInput(node); |
193 | DCHECK_NE(GetPlacement(node), Placement::kFixed)((void) 0); |
194 | DCHECK_NE(GetPlacement(node), Placement::kCoupled)((void) 0); |
195 | } |
196 | |
197 | ++(GetData(node)->unscheduled_count_); |
198 | if (FLAG_trace_turbo_scheduler) { |
199 | TRACE(" Use count of #%d:%s (used by #%d:%s)++ = %d\n", node->id(), |
200 | node->op()->mnemonic(), from->id(), from->op()->mnemonic(), |
201 | GetData(node)->unscheduled_count_); |
202 | } |
203 | } |
204 | |
205 | void Scheduler::DecrementUnscheduledUseCount(Node* node, Node* from) { |
206 | // Tracking use counts for fixed nodes is useless. |
207 | if (GetPlacement(node) == kFixed) return; |
208 | |
209 | // Use count for coupled nodes is summed up on their control. |
210 | if (GetPlacement(node) == kCoupled) { |
211 | node = NodeProperties::GetControlInput(node); |
212 | DCHECK_NE(GetPlacement(node), Placement::kFixed)((void) 0); |
213 | DCHECK_NE(GetPlacement(node), Placement::kCoupled)((void) 0); |
214 | } |
215 | |
216 | DCHECK_LT(0, GetData(node)->unscheduled_count_)((void) 0); |
217 | --(GetData(node)->unscheduled_count_); |
218 | if (FLAG_trace_turbo_scheduler) { |
219 | TRACE(" Use count of #%d:%s (used by #%d:%s)-- = %d\n", node->id(), |
220 | node->op()->mnemonic(), from->id(), from->op()->mnemonic(), |
221 | GetData(node)->unscheduled_count_); |
222 | } |
223 | if (GetData(node)->unscheduled_count_ == 0) { |
224 | TRACE(" newly eligible #%d:%s\n", node->id(), node->op()->mnemonic()); |
225 | schedule_queue_.push(node); |
226 | } |
227 | } |
228 | |
229 | // ----------------------------------------------------------------------------- |
230 | // Phase 1: Build control-flow graph. |
231 | |
232 | |
233 | // Internal class to build a control flow graph (i.e the basic blocks and edges |
234 | // between them within a Schedule) from the node graph. Visits control edges of |
235 | // the graph backwards from an end node in order to find the connected control |
236 | // subgraph, needed for scheduling. |
237 | class CFGBuilder : public ZoneObject { |
238 | public: |
239 | CFGBuilder(Zone* zone, Scheduler* scheduler) |
240 | : zone_(zone), |
241 | scheduler_(scheduler), |
242 | schedule_(scheduler->schedule_), |
243 | queued_(scheduler->graph_, 2), |
244 | queue_(zone), |
245 | control_(zone), |
246 | component_entry_(nullptr), |
247 | component_start_(nullptr), |
248 | component_end_(nullptr) {} |
249 | |
250 | // Run the control flow graph construction algorithm by walking the graph |
251 | // backwards from end through control edges, building and connecting the |
252 | // basic blocks for control nodes. |
253 | void Run() { |
254 | ResetDataStructures(); |
255 | Queue(scheduler_->graph_->end()); |
256 | |
257 | while (!queue_.empty()) { // Breadth-first backwards traversal. |
258 | scheduler_->tick_counter_->TickAndMaybeEnterSafepoint(); |
259 | Node* node = queue_.front(); |
260 | queue_.pop(); |
261 | int max = NodeProperties::PastControlIndex(node); |
262 | for (int i = NodeProperties::FirstControlIndex(node); i < max; i++) { |
263 | Queue(node->InputAt(i)); |
264 | } |
265 | } |
266 | |
267 | for (NodeVector::iterator i = control_.begin(); i != control_.end(); ++i) { |
268 | ConnectBlocks(*i); // Connect block to its predecessor/successors. |
269 | } |
270 | } |
271 | |
272 | // Run the control flow graph construction for a minimal control-connected |
273 | // component ending in {exit} and merge that component into an existing |
274 | // control flow graph at the bottom of {block}. |
275 | void Run(BasicBlock* block, Node* exit) { |
276 | ResetDataStructures(); |
277 | Queue(exit); |
278 | |
279 | component_entry_ = nullptr; |
280 | component_start_ = block; |
281 | component_end_ = schedule_->block(exit); |
282 | scheduler_->equivalence_->Run(exit); |
283 | while (!queue_.empty()) { // Breadth-first backwards traversal. |
284 | scheduler_->tick_counter_->TickAndMaybeEnterSafepoint(); |
285 | Node* node = queue_.front(); |
286 | queue_.pop(); |
287 | |
288 | // Use control dependence equivalence to find a canonical single-entry |
289 | // single-exit region that makes up a minimal component to be scheduled. |
290 | if (IsSingleEntrySingleExitRegion(node, exit)) { |
291 | TRACE("Found SESE at #%d:%s\n", node->id(), node->op()->mnemonic()); |
292 | DCHECK(!component_entry_)((void) 0); |
293 | component_entry_ = node; |
294 | continue; |
295 | } |
296 | |
297 | int max = NodeProperties::PastControlIndex(node); |
298 | for (int i = NodeProperties::FirstControlIndex(node); i < max; i++) { |
299 | Queue(node->InputAt(i)); |
300 | } |
301 | } |
302 | DCHECK(component_entry_)((void) 0); |
303 | |
304 | for (NodeVector::iterator i = control_.begin(); i != control_.end(); ++i) { |
305 | ConnectBlocks(*i); // Connect block to its predecessor/successors. |
306 | } |
307 | } |
308 | |
309 | private: |
310 | friend class ScheduleLateNodeVisitor; |
311 | friend class Scheduler; |
312 | |
313 | void FixNode(BasicBlock* block, Node* node) { |
314 | schedule_->AddNode(block, node); |
315 | scheduler_->UpdatePlacement(node, Scheduler::kFixed); |
316 | } |
317 | |
318 | void Queue(Node* node) { |
319 | // Mark the connected control nodes as they are queued. |
320 | if (!queued_.Get(node)) { |
321 | BuildBlocks(node); |
322 | queue_.push(node); |
323 | queued_.Set(node, true); |
324 | control_.push_back(node); |
325 | } |
326 | } |
327 | |
328 | void BuildBlocks(Node* node) { |
329 | switch (node->opcode()) { |
330 | case IrOpcode::kEnd: |
331 | FixNode(schedule_->end(), node); |
332 | break; |
333 | case IrOpcode::kStart: |
334 | FixNode(schedule_->start(), node); |
335 | break; |
336 | case IrOpcode::kLoop: |
337 | case IrOpcode::kMerge: |
338 | BuildBlockForNode(node); |
339 | break; |
340 | case IrOpcode::kTerminate: { |
341 | // Put Terminate in the loop to which it refers. |
342 | Node* loop = NodeProperties::GetControlInput(node); |
343 | BasicBlock* block = BuildBlockForNode(loop); |
344 | FixNode(block, node); |
345 | break; |
346 | } |
347 | case IrOpcode::kBranch: |
348 | case IrOpcode::kSwitch: |
349 | BuildBlocksForSuccessors(node); |
350 | break; |
351 | #define BUILD_BLOCK_JS_CASE(Name, ...) case IrOpcode::k##Name: |
352 | JS_OP_LIST(BUILD_BLOCK_JS_CASE)BUILD_BLOCK_JS_CASE(JSEqual, Equal) BUILD_BLOCK_JS_CASE(JSStrictEqual , StrictEqual) BUILD_BLOCK_JS_CASE(JSLessThan, LessThan) BUILD_BLOCK_JS_CASE (JSGreaterThan, GreaterThan) BUILD_BLOCK_JS_CASE(JSLessThanOrEqual , LessThanOrEqual) BUILD_BLOCK_JS_CASE(JSGreaterThanOrEqual, GreaterThanOrEqual ) BUILD_BLOCK_JS_CASE(JSBitwiseOr, BitwiseOr) BUILD_BLOCK_JS_CASE (JSBitwiseXor, BitwiseXor) BUILD_BLOCK_JS_CASE(JSBitwiseAnd, BitwiseAnd ) BUILD_BLOCK_JS_CASE(JSShiftLeft, ShiftLeft) BUILD_BLOCK_JS_CASE (JSShiftRight, ShiftRight) BUILD_BLOCK_JS_CASE(JSShiftRightLogical , ShiftRightLogical) BUILD_BLOCK_JS_CASE(JSAdd, Add) BUILD_BLOCK_JS_CASE (JSSubtract, Subtract) BUILD_BLOCK_JS_CASE(JSMultiply, Multiply ) BUILD_BLOCK_JS_CASE(JSDivide, Divide) BUILD_BLOCK_JS_CASE(JSModulus , Modulus) BUILD_BLOCK_JS_CASE(JSExponentiate, Exponentiate) BUILD_BLOCK_JS_CASE (JSHasInPrototypeChain) BUILD_BLOCK_JS_CASE(JSInstanceOf) BUILD_BLOCK_JS_CASE (JSOrdinaryHasInstance) BUILD_BLOCK_JS_CASE(JSDecrement, Decrement ) BUILD_BLOCK_JS_CASE(JSIncrement, Increment) BUILD_BLOCK_JS_CASE (JSBitwiseNot, BitwiseNot) BUILD_BLOCK_JS_CASE(JSNegate, Negate ) BUILD_BLOCK_JS_CASE(JSToLength) BUILD_BLOCK_JS_CASE(JSToName ) BUILD_BLOCK_JS_CASE(JSToNumber) BUILD_BLOCK_JS_CASE(JSToNumberConvertBigInt ) BUILD_BLOCK_JS_CASE(JSToNumeric) BUILD_BLOCK_JS_CASE(JSToObject ) BUILD_BLOCK_JS_CASE(JSToString) BUILD_BLOCK_JS_CASE(JSParseInt ) BUILD_BLOCK_JS_CASE(JSCloneObject) BUILD_BLOCK_JS_CASE(JSCreate ) BUILD_BLOCK_JS_CASE(JSCreateArguments) BUILD_BLOCK_JS_CASE( JSCreateArray) BUILD_BLOCK_JS_CASE(JSCreateArrayFromIterable) BUILD_BLOCK_JS_CASE(JSCreateArrayIterator) BUILD_BLOCK_JS_CASE (JSCreateAsyncFunctionObject) BUILD_BLOCK_JS_CASE(JSCreateBoundFunction ) BUILD_BLOCK_JS_CASE(JSCreateClosure) BUILD_BLOCK_JS_CASE(JSCreateCollectionIterator ) BUILD_BLOCK_JS_CASE(JSCreateEmptyLiteralArray) BUILD_BLOCK_JS_CASE (JSCreateEmptyLiteralObject) BUILD_BLOCK_JS_CASE(JSCreateGeneratorObject ) BUILD_BLOCK_JS_CASE(JSCreateIterResultObject) BUILD_BLOCK_JS_CASE (JSCreateKeyValueArray) BUILD_BLOCK_JS_CASE(JSCreateLiteralArray ) BUILD_BLOCK_JS_CASE(JSCreateLiteralObject) BUILD_BLOCK_JS_CASE (JSCreateLiteralRegExp) BUILD_BLOCK_JS_CASE(JSCreateObject) BUILD_BLOCK_JS_CASE (JSCreatePromise) BUILD_BLOCK_JS_CASE(JSCreateStringIterator) BUILD_BLOCK_JS_CASE(JSCreateTypedArray) BUILD_BLOCK_JS_CASE( JSGetTemplateObject) BUILD_BLOCK_JS_CASE(JSLoadProperty) BUILD_BLOCK_JS_CASE (JSLoadNamed) BUILD_BLOCK_JS_CASE(JSLoadNamedFromSuper) BUILD_BLOCK_JS_CASE (JSLoadGlobal) BUILD_BLOCK_JS_CASE(JSSetKeyedProperty) BUILD_BLOCK_JS_CASE (JSDefineKeyedOwnProperty) BUILD_BLOCK_JS_CASE(JSSetNamedProperty ) BUILD_BLOCK_JS_CASE(JSDefineNamedOwnProperty) BUILD_BLOCK_JS_CASE (JSStoreGlobal) BUILD_BLOCK_JS_CASE(JSDefineKeyedOwnPropertyInLiteral ) BUILD_BLOCK_JS_CASE(JSStoreInArrayLiteral) BUILD_BLOCK_JS_CASE (JSDeleteProperty) BUILD_BLOCK_JS_CASE(JSHasProperty) BUILD_BLOCK_JS_CASE (JSGetSuperConstructor) BUILD_BLOCK_JS_CASE(JSHasContextExtension ) BUILD_BLOCK_JS_CASE(JSLoadContext) BUILD_BLOCK_JS_CASE(JSStoreContext ) BUILD_BLOCK_JS_CASE(JSCreateFunctionContext) BUILD_BLOCK_JS_CASE (JSCreateCatchContext) BUILD_BLOCK_JS_CASE(JSCreateWithContext ) BUILD_BLOCK_JS_CASE(JSCreateBlockContext) BUILD_BLOCK_JS_CASE (JSCall) BUILD_BLOCK_JS_CASE(JSCallForwardVarargs) BUILD_BLOCK_JS_CASE (JSCallWithArrayLike) BUILD_BLOCK_JS_CASE(JSCallWithSpread) BUILD_BLOCK_JS_CASE (JSWasmCall) BUILD_BLOCK_JS_CASE(JSConstructForwardVarargs) BUILD_BLOCK_JS_CASE (JSConstruct) BUILD_BLOCK_JS_CASE(JSConstructWithArrayLike) BUILD_BLOCK_JS_CASE (JSConstructWithSpread) BUILD_BLOCK_JS_CASE(JSAsyncFunctionEnter ) BUILD_BLOCK_JS_CASE(JSAsyncFunctionReject) BUILD_BLOCK_JS_CASE (JSAsyncFunctionResolve) BUILD_BLOCK_JS_CASE(JSCallRuntime) BUILD_BLOCK_JS_CASE (JSForInEnumerate) BUILD_BLOCK_JS_CASE(JSForInNext) BUILD_BLOCK_JS_CASE (JSForInPrepare) BUILD_BLOCK_JS_CASE(JSGetIterator) BUILD_BLOCK_JS_CASE (JSLoadMessage) BUILD_BLOCK_JS_CASE(JSStoreMessage) BUILD_BLOCK_JS_CASE (JSLoadModule) BUILD_BLOCK_JS_CASE(JSStoreModule) BUILD_BLOCK_JS_CASE (JSGetImportMeta) BUILD_BLOCK_JS_CASE(JSGeneratorStore) BUILD_BLOCK_JS_CASE (JSGeneratorRestoreContinuation) BUILD_BLOCK_JS_CASE(JSGeneratorRestoreContext ) BUILD_BLOCK_JS_CASE(JSGeneratorRestoreRegister) BUILD_BLOCK_JS_CASE (JSGeneratorRestoreInputOrDebugPos) BUILD_BLOCK_JS_CASE(JSFulfillPromise ) BUILD_BLOCK_JS_CASE(JSPerformPromiseThen) BUILD_BLOCK_JS_CASE (JSPromiseResolve) BUILD_BLOCK_JS_CASE(JSRejectPromise) BUILD_BLOCK_JS_CASE (JSResolvePromise) BUILD_BLOCK_JS_CASE(JSStackCheck) BUILD_BLOCK_JS_CASE (JSObjectIsArray) BUILD_BLOCK_JS_CASE(JSRegExpTest) BUILD_BLOCK_JS_CASE (JSDebugger) |
353 | // JS opcodes are just like calls => fall through. |
354 | #undef BUILD_BLOCK_JS_CASE |
355 | case IrOpcode::kCall: |
356 | if (NodeProperties::IsExceptionalCall(node)) { |
357 | BuildBlocksForSuccessors(node); |
358 | } |
359 | break; |
360 | default: |
361 | break; |
362 | } |
363 | } |
364 | |
365 | void ConnectBlocks(Node* node) { |
366 | switch (node->opcode()) { |
367 | case IrOpcode::kLoop: |
368 | case IrOpcode::kMerge: |
369 | ConnectMerge(node); |
370 | break; |
371 | case IrOpcode::kBranch: |
372 | scheduler_->UpdatePlacement(node, Scheduler::kFixed); |
373 | ConnectBranch(node); |
374 | break; |
375 | case IrOpcode::kSwitch: |
376 | scheduler_->UpdatePlacement(node, Scheduler::kFixed); |
377 | ConnectSwitch(node); |
378 | break; |
379 | case IrOpcode::kDeoptimize: |
380 | scheduler_->UpdatePlacement(node, Scheduler::kFixed); |
381 | ConnectDeoptimize(node); |
382 | break; |
383 | case IrOpcode::kTailCall: |
384 | scheduler_->UpdatePlacement(node, Scheduler::kFixed); |
385 | ConnectTailCall(node); |
386 | break; |
387 | case IrOpcode::kReturn: |
388 | scheduler_->UpdatePlacement(node, Scheduler::kFixed); |
389 | ConnectReturn(node); |
390 | break; |
391 | case IrOpcode::kThrow: |
392 | scheduler_->UpdatePlacement(node, Scheduler::kFixed); |
393 | ConnectThrow(node); |
394 | break; |
395 | #define CONNECT_BLOCK_JS_CASE(Name, ...) case IrOpcode::k##Name: |
396 | JS_OP_LIST(CONNECT_BLOCK_JS_CASE)CONNECT_BLOCK_JS_CASE(JSEqual, Equal) CONNECT_BLOCK_JS_CASE(JSStrictEqual , StrictEqual) CONNECT_BLOCK_JS_CASE(JSLessThan, LessThan) CONNECT_BLOCK_JS_CASE (JSGreaterThan, GreaterThan) CONNECT_BLOCK_JS_CASE(JSLessThanOrEqual , LessThanOrEqual) CONNECT_BLOCK_JS_CASE(JSGreaterThanOrEqual , GreaterThanOrEqual) CONNECT_BLOCK_JS_CASE(JSBitwiseOr, BitwiseOr ) CONNECT_BLOCK_JS_CASE(JSBitwiseXor, BitwiseXor) CONNECT_BLOCK_JS_CASE (JSBitwiseAnd, BitwiseAnd) CONNECT_BLOCK_JS_CASE(JSShiftLeft, ShiftLeft) CONNECT_BLOCK_JS_CASE(JSShiftRight, ShiftRight) CONNECT_BLOCK_JS_CASE (JSShiftRightLogical, ShiftRightLogical) CONNECT_BLOCK_JS_CASE (JSAdd, Add) CONNECT_BLOCK_JS_CASE(JSSubtract, Subtract) CONNECT_BLOCK_JS_CASE (JSMultiply, Multiply) CONNECT_BLOCK_JS_CASE(JSDivide, Divide ) CONNECT_BLOCK_JS_CASE(JSModulus, Modulus) CONNECT_BLOCK_JS_CASE (JSExponentiate, Exponentiate) CONNECT_BLOCK_JS_CASE(JSHasInPrototypeChain ) CONNECT_BLOCK_JS_CASE(JSInstanceOf) CONNECT_BLOCK_JS_CASE(JSOrdinaryHasInstance ) CONNECT_BLOCK_JS_CASE(JSDecrement, Decrement) CONNECT_BLOCK_JS_CASE (JSIncrement, Increment) CONNECT_BLOCK_JS_CASE(JSBitwiseNot, BitwiseNot ) CONNECT_BLOCK_JS_CASE(JSNegate, Negate) CONNECT_BLOCK_JS_CASE (JSToLength) CONNECT_BLOCK_JS_CASE(JSToName) CONNECT_BLOCK_JS_CASE (JSToNumber) CONNECT_BLOCK_JS_CASE(JSToNumberConvertBigInt) CONNECT_BLOCK_JS_CASE (JSToNumeric) CONNECT_BLOCK_JS_CASE(JSToObject) CONNECT_BLOCK_JS_CASE (JSToString) CONNECT_BLOCK_JS_CASE(JSParseInt) CONNECT_BLOCK_JS_CASE (JSCloneObject) CONNECT_BLOCK_JS_CASE(JSCreate) CONNECT_BLOCK_JS_CASE (JSCreateArguments) CONNECT_BLOCK_JS_CASE(JSCreateArray) CONNECT_BLOCK_JS_CASE (JSCreateArrayFromIterable) CONNECT_BLOCK_JS_CASE(JSCreateArrayIterator ) CONNECT_BLOCK_JS_CASE(JSCreateAsyncFunctionObject) CONNECT_BLOCK_JS_CASE (JSCreateBoundFunction) CONNECT_BLOCK_JS_CASE(JSCreateClosure ) CONNECT_BLOCK_JS_CASE(JSCreateCollectionIterator) CONNECT_BLOCK_JS_CASE (JSCreateEmptyLiteralArray) CONNECT_BLOCK_JS_CASE(JSCreateEmptyLiteralObject ) CONNECT_BLOCK_JS_CASE(JSCreateGeneratorObject) CONNECT_BLOCK_JS_CASE (JSCreateIterResultObject) CONNECT_BLOCK_JS_CASE(JSCreateKeyValueArray ) CONNECT_BLOCK_JS_CASE(JSCreateLiteralArray) CONNECT_BLOCK_JS_CASE (JSCreateLiteralObject) CONNECT_BLOCK_JS_CASE(JSCreateLiteralRegExp ) CONNECT_BLOCK_JS_CASE(JSCreateObject) CONNECT_BLOCK_JS_CASE (JSCreatePromise) CONNECT_BLOCK_JS_CASE(JSCreateStringIterator ) CONNECT_BLOCK_JS_CASE(JSCreateTypedArray) CONNECT_BLOCK_JS_CASE (JSGetTemplateObject) CONNECT_BLOCK_JS_CASE(JSLoadProperty) CONNECT_BLOCK_JS_CASE (JSLoadNamed) CONNECT_BLOCK_JS_CASE(JSLoadNamedFromSuper) CONNECT_BLOCK_JS_CASE (JSLoadGlobal) CONNECT_BLOCK_JS_CASE(JSSetKeyedProperty) CONNECT_BLOCK_JS_CASE (JSDefineKeyedOwnProperty) CONNECT_BLOCK_JS_CASE(JSSetNamedProperty ) CONNECT_BLOCK_JS_CASE(JSDefineNamedOwnProperty) CONNECT_BLOCK_JS_CASE (JSStoreGlobal) CONNECT_BLOCK_JS_CASE(JSDefineKeyedOwnPropertyInLiteral ) CONNECT_BLOCK_JS_CASE(JSStoreInArrayLiteral) CONNECT_BLOCK_JS_CASE (JSDeleteProperty) CONNECT_BLOCK_JS_CASE(JSHasProperty) CONNECT_BLOCK_JS_CASE (JSGetSuperConstructor) CONNECT_BLOCK_JS_CASE(JSHasContextExtension ) CONNECT_BLOCK_JS_CASE(JSLoadContext) CONNECT_BLOCK_JS_CASE( JSStoreContext) CONNECT_BLOCK_JS_CASE(JSCreateFunctionContext ) CONNECT_BLOCK_JS_CASE(JSCreateCatchContext) CONNECT_BLOCK_JS_CASE (JSCreateWithContext) CONNECT_BLOCK_JS_CASE(JSCreateBlockContext ) CONNECT_BLOCK_JS_CASE(JSCall) CONNECT_BLOCK_JS_CASE(JSCallForwardVarargs ) CONNECT_BLOCK_JS_CASE(JSCallWithArrayLike) CONNECT_BLOCK_JS_CASE (JSCallWithSpread) CONNECT_BLOCK_JS_CASE(JSWasmCall) CONNECT_BLOCK_JS_CASE (JSConstructForwardVarargs) CONNECT_BLOCK_JS_CASE(JSConstruct ) CONNECT_BLOCK_JS_CASE(JSConstructWithArrayLike) CONNECT_BLOCK_JS_CASE (JSConstructWithSpread) CONNECT_BLOCK_JS_CASE(JSAsyncFunctionEnter ) CONNECT_BLOCK_JS_CASE(JSAsyncFunctionReject) CONNECT_BLOCK_JS_CASE (JSAsyncFunctionResolve) CONNECT_BLOCK_JS_CASE(JSCallRuntime) CONNECT_BLOCK_JS_CASE(JSForInEnumerate) CONNECT_BLOCK_JS_CASE (JSForInNext) CONNECT_BLOCK_JS_CASE(JSForInPrepare) CONNECT_BLOCK_JS_CASE (JSGetIterator) CONNECT_BLOCK_JS_CASE(JSLoadMessage) CONNECT_BLOCK_JS_CASE (JSStoreMessage) CONNECT_BLOCK_JS_CASE(JSLoadModule) CONNECT_BLOCK_JS_CASE (JSStoreModule) CONNECT_BLOCK_JS_CASE(JSGetImportMeta) CONNECT_BLOCK_JS_CASE (JSGeneratorStore) CONNECT_BLOCK_JS_CASE(JSGeneratorRestoreContinuation ) CONNECT_BLOCK_JS_CASE(JSGeneratorRestoreContext) CONNECT_BLOCK_JS_CASE (JSGeneratorRestoreRegister) CONNECT_BLOCK_JS_CASE(JSGeneratorRestoreInputOrDebugPos ) CONNECT_BLOCK_JS_CASE(JSFulfillPromise) CONNECT_BLOCK_JS_CASE (JSPerformPromiseThen) CONNECT_BLOCK_JS_CASE(JSPromiseResolve ) CONNECT_BLOCK_JS_CASE(JSRejectPromise) CONNECT_BLOCK_JS_CASE (JSResolvePromise) CONNECT_BLOCK_JS_CASE(JSStackCheck) CONNECT_BLOCK_JS_CASE (JSObjectIsArray) CONNECT_BLOCK_JS_CASE(JSRegExpTest) CONNECT_BLOCK_JS_CASE (JSDebugger) |
397 | // JS opcodes are just like calls => fall through. |
398 | #undef CONNECT_BLOCK_JS_CASE |
399 | case IrOpcode::kCall: |
400 | if (NodeProperties::IsExceptionalCall(node)) { |
401 | scheduler_->UpdatePlacement(node, Scheduler::kFixed); |
402 | ConnectCall(node); |
403 | } |
404 | break; |
405 | default: |
406 | break; |
407 | } |
408 | } |
409 | |
410 | BasicBlock* BuildBlockForNode(Node* node) { |
411 | BasicBlock* block = schedule_->block(node); |
412 | if (block == nullptr) { |
413 | block = schedule_->NewBasicBlock(); |
414 | TRACE("Create block id:%d for #%d:%s\n", block->id().ToInt(), node->id(), |
415 | node->op()->mnemonic()); |
416 | FixNode(block, node); |
417 | } |
418 | return block; |
419 | } |
420 | |
421 | void BuildBlocksForSuccessors(Node* node) { |
422 | size_t const successor_cnt = node->op()->ControlOutputCount(); |
423 | Node** successors = zone_->NewArray<Node*>(successor_cnt); |
424 | NodeProperties::CollectControlProjections(node, successors, successor_cnt); |
425 | for (size_t index = 0; index < successor_cnt; ++index) { |
426 | BuildBlockForNode(successors[index]); |
427 | } |
428 | } |
429 | |
430 | void CollectSuccessorBlocks(Node* node, BasicBlock** successor_blocks, |
431 | size_t successor_cnt) { |
432 | Node** successors = reinterpret_cast<Node**>(successor_blocks); |
433 | NodeProperties::CollectControlProjections(node, successors, successor_cnt); |
434 | for (size_t index = 0; index < successor_cnt; ++index) { |
435 | successor_blocks[index] = schedule_->block(successors[index]); |
436 | } |
437 | } |
438 | |
439 | BasicBlock* FindPredecessorBlock(Node* node) { |
440 | BasicBlock* predecessor_block = nullptr; |
441 | while (true) { |
442 | predecessor_block = schedule_->block(node); |
443 | if (predecessor_block != nullptr) break; |
444 | node = NodeProperties::GetControlInput(node); |
445 | } |
446 | return predecessor_block; |
447 | } |
448 | |
449 | void ConnectCall(Node* call) { |
450 | BasicBlock* successor_blocks[2]; |
451 | CollectSuccessorBlocks(call, successor_blocks, arraysize(successor_blocks)(sizeof(ArraySizeHelper(successor_blocks)))); |
452 | |
453 | // Consider the exception continuation to be deferred. |
454 | successor_blocks[1]->set_deferred(true); |
455 | |
456 | Node* call_control = NodeProperties::GetControlInput(call); |
457 | BasicBlock* call_block = FindPredecessorBlock(call_control); |
458 | TraceConnect(call, call_block, successor_blocks[0]); |
459 | TraceConnect(call, call_block, successor_blocks[1]); |
460 | schedule_->AddCall(call_block, call, successor_blocks[0], |
461 | successor_blocks[1]); |
462 | } |
463 | |
464 | void ConnectBranch(Node* branch) { |
465 | BasicBlock* successor_blocks[2]; |
466 | CollectSuccessorBlocks(branch, successor_blocks, |
467 | arraysize(successor_blocks)(sizeof(ArraySizeHelper(successor_blocks)))); |
468 | |
469 | BranchHint hint_from_profile = BranchHint::kNone; |
470 | if (const ProfileDataFromFile* profile_data = scheduler_->profile_data()) { |
471 | double block_zero_count = |
472 | profile_data->GetCounter(successor_blocks[0]->id().ToSize()); |
473 | double block_one_count = |
474 | profile_data->GetCounter(successor_blocks[1]->id().ToSize()); |
475 | // If a branch is visited a non-trivial number of times and substantially |
476 | // more often than its alternative, then mark it as likely. |
477 | constexpr double kMinimumCount = 100000; |
478 | constexpr double kThresholdRatio = 4000; |
479 | if (block_zero_count > kMinimumCount && |
480 | block_zero_count / kThresholdRatio > block_one_count) { |
481 | hint_from_profile = BranchHint::kTrue; |
482 | } else if (block_one_count > kMinimumCount && |
483 | block_one_count / kThresholdRatio > block_zero_count) { |
484 | hint_from_profile = BranchHint::kFalse; |
485 | } |
486 | } |
487 | |
488 | // Consider branch hints. |
489 | switch (hint_from_profile) { |
490 | case BranchHint::kNone: |
491 | switch (BranchHintOf(branch->op())) { |
492 | case BranchHint::kNone: |
493 | break; |
494 | case BranchHint::kTrue: |
495 | successor_blocks[1]->set_deferred(true); |
496 | break; |
497 | case BranchHint::kFalse: |
498 | successor_blocks[0]->set_deferred(true); |
499 | break; |
500 | } |
501 | break; |
502 | case BranchHint::kTrue: |
503 | successor_blocks[1]->set_deferred(true); |
504 | break; |
505 | case BranchHint::kFalse: |
506 | successor_blocks[0]->set_deferred(true); |
507 | break; |
508 | } |
509 | |
510 | if (hint_from_profile != BranchHint::kNone && |
511 | BranchHintOf(branch->op()) != BranchHint::kNone && |
512 | hint_from_profile != BranchHintOf(branch->op())) { |
513 | PrintF("Warning: profiling data overrode manual branch hint.\n"); |
514 | } |
515 | |
516 | if (branch == component_entry_) { |
517 | TraceConnect(branch, component_start_, successor_blocks[0]); |
518 | TraceConnect(branch, component_start_, successor_blocks[1]); |
519 | schedule_->InsertBranch(component_start_, component_end_, branch, |
520 | successor_blocks[0], successor_blocks[1]); |
521 | } else { |
522 | Node* branch_control = NodeProperties::GetControlInput(branch); |
523 | BasicBlock* branch_block = FindPredecessorBlock(branch_control); |
524 | TraceConnect(branch, branch_block, successor_blocks[0]); |
525 | TraceConnect(branch, branch_block, successor_blocks[1]); |
526 | schedule_->AddBranch(branch_block, branch, successor_blocks[0], |
527 | successor_blocks[1]); |
528 | } |
529 | } |
530 | |
531 | void ConnectSwitch(Node* sw) { |
532 | size_t const successor_count = sw->op()->ControlOutputCount(); |
533 | BasicBlock** successor_blocks = |
534 | zone_->NewArray<BasicBlock*>(successor_count); |
535 | CollectSuccessorBlocks(sw, successor_blocks, successor_count); |
536 | |
537 | if (sw == component_entry_) { |
538 | for (size_t index = 0; index < successor_count; ++index) { |
539 | TraceConnect(sw, component_start_, successor_blocks[index]); |
540 | } |
541 | schedule_->InsertSwitch(component_start_, component_end_, sw, |
542 | successor_blocks, successor_count); |
543 | } else { |
544 | Node* switch_control = NodeProperties::GetControlInput(sw); |
545 | BasicBlock* switch_block = FindPredecessorBlock(switch_control); |
546 | for (size_t index = 0; index < successor_count; ++index) { |
547 | TraceConnect(sw, switch_block, successor_blocks[index]); |
548 | } |
549 | schedule_->AddSwitch(switch_block, sw, successor_blocks, successor_count); |
550 | } |
551 | for (size_t index = 0; index < successor_count; ++index) { |
552 | if (BranchHintOf(successor_blocks[index]->front()->op()) == |
553 | BranchHint::kFalse) { |
554 | successor_blocks[index]->set_deferred(true); |
555 | } |
556 | } |
557 | } |
558 | |
559 | void ConnectMerge(Node* merge) { |
560 | // Don't connect the special merge at the end to its predecessors. |
561 | if (IsFinalMerge(merge)) return; |
562 | |
563 | BasicBlock* block = schedule_->block(merge); |
564 | DCHECK_NOT_NULL(block)((void) 0); |
565 | // For all of the merge's control inputs, add a goto at the end to the |
566 | // merge's basic block. |
567 | for (Node* const input : merge->inputs()) { |
568 | BasicBlock* predecessor_block = FindPredecessorBlock(input); |
569 | TraceConnect(merge, predecessor_block, block); |
570 | schedule_->AddGoto(predecessor_block, block); |
571 | } |
572 | } |
573 | |
574 | void ConnectTailCall(Node* call) { |
575 | Node* call_control = NodeProperties::GetControlInput(call); |
576 | BasicBlock* call_block = FindPredecessorBlock(call_control); |
577 | TraceConnect(call, call_block, nullptr); |
578 | schedule_->AddTailCall(call_block, call); |
579 | } |
580 | |
581 | void ConnectReturn(Node* ret) { |
582 | Node* return_control = NodeProperties::GetControlInput(ret); |
583 | BasicBlock* return_block = FindPredecessorBlock(return_control); |
584 | TraceConnect(ret, return_block, nullptr); |
585 | schedule_->AddReturn(return_block, ret); |
586 | } |
587 | |
588 | void ConnectDeoptimize(Node* deopt) { |
589 | Node* deoptimize_control = NodeProperties::GetControlInput(deopt); |
590 | BasicBlock* deoptimize_block = FindPredecessorBlock(deoptimize_control); |
591 | TraceConnect(deopt, deoptimize_block, nullptr); |
592 | schedule_->AddDeoptimize(deoptimize_block, deopt); |
593 | } |
594 | |
595 | void ConnectThrow(Node* thr) { |
596 | Node* throw_control = NodeProperties::GetControlInput(thr); |
597 | BasicBlock* throw_block = FindPredecessorBlock(throw_control); |
598 | TraceConnect(thr, throw_block, nullptr); |
599 | schedule_->AddThrow(throw_block, thr); |
600 | } |
601 | |
602 | void TraceConnect(Node* node, BasicBlock* block, BasicBlock* succ) { |
603 | DCHECK_NOT_NULL(block)((void) 0); |
604 | if (succ == nullptr) { |
605 | TRACE("Connect #%d:%s, id:%d -> end\n", node->id(), |
606 | node->op()->mnemonic(), block->id().ToInt()); |
607 | } else { |
608 | TRACE("Connect #%d:%s, id:%d -> id:%d\n", node->id(), |
609 | node->op()->mnemonic(), block->id().ToInt(), succ->id().ToInt()); |
610 | } |
611 | } |
612 | |
613 | bool IsFinalMerge(Node* node) { |
614 | return (node->opcode() == IrOpcode::kMerge && |
615 | node == scheduler_->graph_->end()->InputAt(0)); |
616 | } |
617 | |
618 | bool IsSingleEntrySingleExitRegion(Node* entry, Node* exit) const { |
619 | size_t entry_class = scheduler_->equivalence_->ClassOf(entry); |
620 | size_t exit_class = scheduler_->equivalence_->ClassOf(exit); |
621 | return entry != exit && entry_class == exit_class; |
622 | } |
623 | |
624 | void ResetDataStructures() { |
625 | control_.clear(); |
626 | DCHECK(queue_.empty())((void) 0); |
627 | DCHECK(control_.empty())((void) 0); |
628 | } |
629 | |
630 | Zone* zone_; |
631 | Scheduler* scheduler_; |
632 | Schedule* schedule_; |
633 | NodeMarker<bool> queued_; // Mark indicating whether node is queued. |
634 | ZoneQueue<Node*> queue_; // Queue used for breadth-first traversal. |
635 | NodeVector control_; // List of encountered control nodes. |
636 | Node* component_entry_; // Component single-entry node. |
637 | BasicBlock* component_start_; // Component single-entry block. |
638 | BasicBlock* component_end_; // Component single-exit block. |
639 | }; |
640 | |
641 | |
642 | void Scheduler::BuildCFG() { |
643 | TRACE("--- CREATING CFG -------------------------------------------\n"); |
644 | |
645 | // Instantiate a new control equivalence algorithm for the graph. |
646 | equivalence_ = zone_->New<ControlEquivalence>(zone_, graph_); |
647 | |
648 | // Build a control-flow graph for the main control-connected component that |
649 | // is being spanned by the graph's start and end nodes. |
650 | control_flow_builder_ = zone_->New<CFGBuilder>(zone_, this); |
651 | control_flow_builder_->Run(); |
652 | |
653 | // Initialize per-block data. |
654 | // Reserve an extra 10% to avoid resizing vector when fusing floating control. |
655 | scheduled_nodes_.reserve(schedule_->BasicBlockCount() * 1.1); |
656 | scheduled_nodes_.resize(schedule_->BasicBlockCount()); |
657 | } |
658 | |
659 | |
660 | // ----------------------------------------------------------------------------- |
661 | // Phase 2: Compute special RPO and dominator tree. |
662 | |
663 | |
664 | // Compute the special reverse-post-order block ordering, which is essentially |
665 | // a RPO of the graph where loop bodies are contiguous. Properties: |
666 | // 1. If block A is a predecessor of B, then A appears before B in the order, |
667 | // unless B is a loop header and A is in the loop headed at B |
668 | // (i.e. A -> B is a backedge). |
669 | // => If block A dominates block B, then A appears before B in the order. |
670 | // => If block A is a loop header, A appears before all blocks in the loop |
671 | // headed at A. |
672 | // 2. All loops are contiguous in the order (i.e. no intervening blocks that |
673 | // do not belong to the loop.) |
674 | // Note a simple RPO traversal satisfies (1) but not (2). |
675 | class SpecialRPONumberer : public ZoneObject { |
676 | public: |
677 | SpecialRPONumberer(Zone* zone, Schedule* schedule) |
678 | : zone_(zone), |
679 | schedule_(schedule), |
680 | order_(nullptr), |
681 | beyond_end_(nullptr), |
682 | loops_(zone), |
683 | backedges_(zone), |
684 | stack_(zone), |
685 | previous_block_count_(0), |
686 | empty_(0, zone) {} |
687 | |
688 | // Computes the special reverse-post-order for the main control flow graph, |
689 | // that is for the graph spanned between the schedule's start and end blocks. |
690 | void ComputeSpecialRPO() { |
691 | DCHECK_EQ(0, schedule_->end()->SuccessorCount())((void) 0); |
692 | DCHECK(!order_)((void) 0); // Main order does not exist yet. |
693 | ComputeAndInsertSpecialRPO(schedule_->start(), schedule_->end()); |
694 | } |
695 | |
696 | // Computes the special reverse-post-order for a partial control flow graph, |
697 | // that is for the graph spanned between the given {entry} and {end} blocks, |
698 | // then updates the existing ordering with this new information. |
699 | void UpdateSpecialRPO(BasicBlock* entry, BasicBlock* end) { |
700 | DCHECK(order_)((void) 0); // Main order to be updated is present. |
701 | ComputeAndInsertSpecialRPO(entry, end); |
702 | } |
703 | |
704 | // Serialize the previously computed order as a special reverse-post-order |
705 | // numbering for basic blocks into the final schedule. |
706 | void SerializeRPOIntoSchedule() { |
707 | int32_t number = 0; |
708 | for (BasicBlock* b = order_; b != nullptr; b = b->rpo_next()) { |
709 | b->set_rpo_number(number++); |
710 | schedule_->rpo_order()->push_back(b); |
711 | } |
712 | BeyondEndSentinel()->set_rpo_number(number); |
713 | } |
714 | |
715 | // Print and verify the special reverse-post-order. |
716 | void PrintAndVerifySpecialRPO() { |
717 | #if DEBUG |
718 | if (FLAG_trace_turbo_scheduler) PrintRPO(); |
719 | VerifySpecialRPO(); |
720 | #endif |
721 | } |
722 | |
723 | const ZoneVector<BasicBlock*>& GetOutgoingBlocks(BasicBlock* block) { |
724 | if (HasLoopNumber(block)) { |
725 | LoopInfo const& loop = loops_[GetLoopNumber(block)]; |
726 | if (loop.outgoing) return *loop.outgoing; |
727 | } |
728 | return empty_; |
729 | } |
730 | |
731 | bool HasLoopBlocks() const { return loops_.size() != 0; } |
732 | |
733 | private: |
734 | using Backedge = std::pair<BasicBlock*, size_t>; |
735 | |
736 | // Numbering for BasicBlock::rpo_number for this block traversal: |
737 | static const int kBlockOnStack = -2; |
738 | static const int kBlockVisited1 = -3; |
739 | static const int kBlockVisited2 = -4; |
740 | static const int kBlockUnvisited1 = -1; |
741 | static const int kBlockUnvisited2 = kBlockVisited1; |
742 | |
743 | struct SpecialRPOStackFrame { |
744 | BasicBlock* block; |
745 | size_t index; |
746 | }; |
747 | |
748 | struct LoopInfo { |
749 | BasicBlock* header; |
750 | ZoneVector<BasicBlock*>* outgoing; |
751 | BitVector* members; |
752 | LoopInfo* prev; |
753 | BasicBlock* end; |
754 | BasicBlock* start; |
755 | |
756 | void AddOutgoing(Zone* zone, BasicBlock* block) { |
757 | if (outgoing == nullptr) { |
758 | outgoing = zone->New<ZoneVector<BasicBlock*>>(zone); |
759 | } |
760 | outgoing->push_back(block); |
761 | } |
762 | }; |
763 | |
764 | int Push(int depth, BasicBlock* child, int unvisited) { |
765 | if (child->rpo_number() == unvisited) { |
766 | stack_[depth].block = child; |
767 | stack_[depth].index = 0; |
768 | child->set_rpo_number(kBlockOnStack); |
769 | return depth + 1; |
770 | } |
771 | return depth; |
772 | } |
773 | |
774 | BasicBlock* PushFront(BasicBlock* head, BasicBlock* block) { |
775 | block->set_rpo_next(head); |
776 | return block; |
777 | } |
778 | |
779 | static int GetLoopNumber(BasicBlock* block) { return block->loop_number(); } |
780 | static void SetLoopNumber(BasicBlock* block, int loop_number) { |
781 | return block->set_loop_number(loop_number); |
782 | } |
783 | static bool HasLoopNumber(BasicBlock* block) { |
784 | return block->loop_number() >= 0; |
785 | } |
786 | |
787 | // We only need this special sentinel because some tests use the schedule's |
788 | // end block in actual control flow (e.g. with end having successors). |
789 | BasicBlock* BeyondEndSentinel() { |
790 | if (beyond_end_ == nullptr) { |
791 | BasicBlock::Id id = BasicBlock::Id::FromInt(-1); |
792 | beyond_end_ = schedule_->zone()->New<BasicBlock>(schedule_->zone(), id); |
793 | } |
794 | return beyond_end_; |
795 | } |
796 | |
797 | // Compute special RPO for the control flow graph between {entry} and {end}, |
798 | // mutating any existing order so that the result is still valid. |
799 | void ComputeAndInsertSpecialRPO(BasicBlock* entry, BasicBlock* end) { |
800 | // RPO should not have been serialized for this schedule yet. |
801 | CHECK_EQ(kBlockUnvisited1, schedule_->start()->loop_number())do { bool _cmp = ::v8::base::CmpEQImpl< typename ::v8::base ::pass_value_or_ref<decltype(kBlockUnvisited1)>::type, typename ::v8::base::pass_value_or_ref<decltype(schedule_->start ()->loop_number())>::type>((kBlockUnvisited1), (schedule_ ->start()->loop_number())); do { if ((__builtin_expect( !!(!(_cmp)), 0))) { V8_Fatal("Check failed: %s.", "kBlockUnvisited1" " " "==" " " "schedule_->start()->loop_number()"); } } while (false); } while (false); |
802 | CHECK_EQ(kBlockUnvisited1, schedule_->start()->rpo_number())do { bool _cmp = ::v8::base::CmpEQImpl< typename ::v8::base ::pass_value_or_ref<decltype(kBlockUnvisited1)>::type, typename ::v8::base::pass_value_or_ref<decltype(schedule_->start ()->rpo_number())>::type>((kBlockUnvisited1), (schedule_ ->start()->rpo_number())); do { if ((__builtin_expect(! !(!(_cmp)), 0))) { V8_Fatal("Check failed: %s.", "kBlockUnvisited1" " " "==" " " "schedule_->start()->rpo_number()"); } } while (false); } while (false); |
803 | CHECK_EQ(0, static_cast<int>(schedule_->rpo_order()->size()))do { bool _cmp = ::v8::base::CmpEQImpl< typename ::v8::base ::pass_value_or_ref<decltype(0)>::type, typename ::v8:: base::pass_value_or_ref<decltype(static_cast<int>(schedule_ ->rpo_order()->size()))>::type>((0), (static_cast <int>(schedule_->rpo_order()->size()))); do { if ( (__builtin_expect(!!(!(_cmp)), 0))) { V8_Fatal("Check failed: %s." , "0" " " "==" " " "static_cast<int>(schedule_->rpo_order()->size())" ); } } while (false); } while (false); |
804 | |
805 | // Find correct insertion point within existing order. |
806 | BasicBlock* insertion_point = entry->rpo_next(); |
807 | BasicBlock* order = insertion_point; |
808 | |
809 | // Perform an iterative RPO traversal using an explicit stack, |
810 | // recording backedges that form cycles. O(|B|). |
811 | DCHECK_LT(previous_block_count_, schedule_->BasicBlockCount())((void) 0); |
812 | stack_.resize(schedule_->BasicBlockCount() - previous_block_count_); |
813 | previous_block_count_ = schedule_->BasicBlockCount(); |
814 | int stack_depth = Push(0, entry, kBlockUnvisited1); |
815 | int num_loops = static_cast<int>(loops_.size()); |
816 | |
817 | while (stack_depth > 0) { |
818 | int current = stack_depth - 1; |
819 | SpecialRPOStackFrame* frame = &stack_[current]; |
820 | |
821 | if (frame->block != end && |
822 | frame->index < frame->block->SuccessorCount()) { |
823 | // Process the next successor. |
824 | BasicBlock* succ = frame->block->SuccessorAt(frame->index++); |
825 | if (succ->rpo_number() == kBlockVisited1) continue; |
826 | if (succ->rpo_number() == kBlockOnStack) { |
827 | // The successor is on the stack, so this is a backedge (cycle). |
828 | backedges_.push_back(Backedge(frame->block, frame->index - 1)); |
829 | if (!HasLoopNumber(succ)) { |
830 | // Assign a new loop number to the header if it doesn't have one. |
831 | SetLoopNumber(succ, num_loops++); |
832 | } |
833 | } else { |
834 | // Push the successor onto the stack. |
835 | DCHECK_EQ(kBlockUnvisited1, succ->rpo_number())((void) 0); |
836 | stack_depth = Push(stack_depth, succ, kBlockUnvisited1); |
837 | } |
838 | } else { |
839 | // Finished with all successors; pop the stack and add the block. |
840 | order = PushFront(order, frame->block); |
841 | frame->block->set_rpo_number(kBlockVisited1); |
842 | stack_depth--; |
843 | } |
844 | } |
845 | |
846 | // If no loops were encountered, then the order we computed was correct. |
847 | if (num_loops > static_cast<int>(loops_.size())) { |
848 | // Otherwise, compute the loop information from the backedges in order |
849 | // to perform a traversal that groups loop bodies together. |
850 | ComputeLoopInfo(&stack_, num_loops, &backedges_); |
851 | |
852 | // Initialize the "loop stack". Note the entry could be a loop header. |
853 | LoopInfo* loop = |
854 | HasLoopNumber(entry) ? &loops_[GetLoopNumber(entry)] : nullptr; |
855 | order = insertion_point; |
856 | |
857 | // Perform an iterative post-order traversal, visiting loop bodies before |
858 | // edges that lead out of loops. Visits each block once, but linking loop |
859 | // sections together is linear in the loop size, so overall is |
860 | // O(|B| + max(loop_depth) * max(|loop|)) |
861 | stack_depth = Push(0, entry, kBlockUnvisited2); |
862 | while (stack_depth > 0) { |
863 | SpecialRPOStackFrame* frame = &stack_[stack_depth - 1]; |
864 | BasicBlock* block = frame->block; |
865 | BasicBlock* succ = nullptr; |
866 | |
867 | if (block != end && frame->index < block->SuccessorCount()) { |
868 | // Process the next normal successor. |
869 | succ = block->SuccessorAt(frame->index++); |
870 | } else if (HasLoopNumber(block)) { |
871 | // Process additional outgoing edges from the loop header. |
872 | if (block->rpo_number() == kBlockOnStack) { |
873 | // Finish the loop body the first time the header is left on the |
874 | // stack. |
875 | DCHECK(loop != nullptr && loop->header == block)((void) 0); |
876 | loop->start = PushFront(order, block); |
877 | order = loop->end; |
878 | block->set_rpo_number(kBlockVisited2); |
879 | // Pop the loop stack and continue visiting outgoing edges within |
880 | // the context of the outer loop, if any. |
881 | loop = loop->prev; |
882 | // We leave the loop header on the stack; the rest of this iteration |
883 | // and later iterations will go through its outgoing edges list. |
884 | } |
885 | |
886 | // Use the next outgoing edge if there are any. |
887 | size_t outgoing_index = frame->index - block->SuccessorCount(); |
888 | LoopInfo* info = &loops_[GetLoopNumber(block)]; |
889 | DCHECK(loop != info)((void) 0); |
890 | if (block != entry && info->outgoing != nullptr && |
891 | outgoing_index < info->outgoing->size()) { |
892 | succ = info->outgoing->at(outgoing_index); |
893 | frame->index++; |
894 | } |
895 | } |
896 | |
897 | if (succ != nullptr) { |
898 | // Process the next successor. |
899 | if (succ->rpo_number() == kBlockOnStack) continue; |
900 | if (succ->rpo_number() == kBlockVisited2) continue; |
901 | DCHECK_EQ(kBlockUnvisited2, succ->rpo_number())((void) 0); |
902 | if (loop != nullptr && !loop->members->Contains(succ->id().ToInt())) { |
903 | // The successor is not in the current loop or any nested loop. |
904 | // Add it to the outgoing edges of this loop and visit it later. |
905 | loop->AddOutgoing(zone_, succ); |
906 | } else { |
907 | // Push the successor onto the stack. |
908 | stack_depth = Push(stack_depth, succ, kBlockUnvisited2); |
909 | if (HasLoopNumber(succ)) { |
910 | // Push the inner loop onto the loop stack. |
911 | DCHECK(GetLoopNumber(succ) < num_loops)((void) 0); |
912 | LoopInfo* next = &loops_[GetLoopNumber(succ)]; |
913 | next->end = order; |
914 | next->prev = loop; |
915 | loop = next; |
916 | } |
917 | } |
918 | } else { |
919 | // Finished with all successors of the current block. |
920 | if (HasLoopNumber(block)) { |
921 | // If we are going to pop a loop header, then add its entire body. |
922 | LoopInfo* info = &loops_[GetLoopNumber(block)]; |
923 | for (BasicBlock* b = info->start; true; b = b->rpo_next()) { |
924 | if (b->rpo_next() == info->end) { |
925 | b->set_rpo_next(order); |
926 | info->end = order; |
927 | break; |
928 | } |
929 | } |
930 | order = info->start; |
931 | } else { |
932 | // Pop a single node off the stack and add it to the order. |
933 | order = PushFront(order, block); |
934 | block->set_rpo_number(kBlockVisited2); |
935 | } |
936 | stack_depth--; |
937 | } |
938 | } |
939 | } |
940 | |
941 | // Publish new order the first time. |
942 | if (order_ == nullptr) order_ = order; |
943 | |
944 | // Compute the correct loop headers and set the correct loop ends. |
945 | LoopInfo* current_loop = nullptr; |
946 | BasicBlock* current_header = entry->loop_header(); |
947 | int32_t loop_depth = entry->loop_depth(); |
948 | if (entry->IsLoopHeader()) --loop_depth; // Entry might be a loop header. |
949 | for (BasicBlock* b = order; b != insertion_point; b = b->rpo_next()) { |
950 | BasicBlock* current = b; |
951 | |
952 | // Reset BasicBlock::rpo_number again. |
953 | current->set_rpo_number(kBlockUnvisited1); |
954 | |
955 | // Finish the previous loop(s) if we just exited them. |
956 | while (current_header != nullptr && |
957 | current == current_header->loop_end()) { |
958 | DCHECK(current_header->IsLoopHeader())((void) 0); |
959 | DCHECK_NOT_NULL(current_loop)((void) 0); |
960 | current_loop = current_loop->prev; |
961 | current_header = |
962 | current_loop == nullptr ? nullptr : current_loop->header; |
963 | --loop_depth; |
964 | } |
965 | current->set_loop_header(current_header); |
966 | |
967 | // Push a new loop onto the stack if this loop is a loop header. |
968 | if (HasLoopNumber(current)) { |
969 | ++loop_depth; |
970 | current_loop = &loops_[GetLoopNumber(current)]; |
971 | BasicBlock* loop_end = current_loop->end; |
972 | current->set_loop_end(loop_end == nullptr ? BeyondEndSentinel() |
973 | : loop_end); |
974 | current_header = current_loop->header; |
975 | TRACE("id:%d is a loop header, increment loop depth to %d\n", |
976 | current->id().ToInt(), loop_depth); |
977 | } |
978 | |
979 | current->set_loop_depth(loop_depth); |
980 | |
981 | if (current->loop_header() == nullptr) { |
982 | TRACE("id:%d is not in a loop (depth == %d)\n", current->id().ToInt(), |
983 | current->loop_depth()); |
984 | } else { |
985 | TRACE("id:%d has loop header id:%d, (depth == %d)\n", |
986 | current->id().ToInt(), current->loop_header()->id().ToInt(), |
987 | current->loop_depth()); |
988 | } |
989 | } |
990 | } |
991 | |
992 | // Computes loop membership from the backedges of the control flow graph. |
993 | void ComputeLoopInfo(ZoneVector<SpecialRPOStackFrame>* queue, |
994 | size_t num_loops, ZoneVector<Backedge>* backedges) { |
995 | // Extend existing loop membership vectors. |
996 | for (LoopInfo& loop : loops_) { |
997 | loop.members->Resize(static_cast<int>(schedule_->BasicBlockCount()), |
998 | zone_); |
999 | } |
1000 | |
1001 | // Extend loop information vector. |
1002 | loops_.resize(num_loops, LoopInfo()); |
1003 | |
1004 | // Compute loop membership starting from backedges. |
1005 | // O(max(loop_depth) * max(|loop|) |
1006 | for (size_t i = 0; i < backedges->size(); i++) { |
1007 | BasicBlock* member = backedges->at(i).first; |
1008 | BasicBlock* header = member->SuccessorAt(backedges->at(i).second); |
1009 | size_t loop_num = GetLoopNumber(header); |
1010 | if (loops_[loop_num].header == nullptr) { |
1011 | loops_[loop_num].header = header; |
1012 | loops_[loop_num].members = zone_->New<BitVector>( |
1013 | static_cast<int>(schedule_->BasicBlockCount()), zone_); |
1014 | } |
1015 | |
1016 | int queue_length = 0; |
1017 | if (member != header) { |
1018 | // As long as the header doesn't have a backedge to itself, |
1019 | // Push the member onto the queue and process its predecessors. |
1020 | if (!loops_[loop_num].members->Contains(member->id().ToInt())) { |
1021 | loops_[loop_num].members->Add(member->id().ToInt()); |
1022 | } |
1023 | (*queue)[queue_length++].block = member; |
1024 | } |
1025 | |
1026 | // Propagate loop membership backwards. All predecessors of M up to the |
1027 | // loop header H are members of the loop too. O(|blocks between M and H|). |
1028 | while (queue_length > 0) { |
1029 | BasicBlock* block = (*queue)[--queue_length].block; |
1030 | for (size_t j = 0; j < block->PredecessorCount(); j++) { |
1031 | BasicBlock* pred = block->PredecessorAt(j); |
1032 | if (pred != header) { |
1033 | if (!loops_[loop_num].members->Contains(pred->id().ToInt())) { |
1034 | loops_[loop_num].members->Add(pred->id().ToInt()); |
1035 | (*queue)[queue_length++].block = pred; |
1036 | } |
1037 | } |
1038 | } |
1039 | } |
1040 | } |
1041 | } |
1042 | |
1043 | #if DEBUG |
1044 | void PrintRPO() { |
1045 | StdoutStream os; |
1046 | os << "RPO with " << loops_.size() << " loops"; |
1047 | if (loops_.size() > 0) { |
1048 | os << " ("; |
1049 | for (size_t i = 0; i < loops_.size(); i++) { |
1050 | if (i > 0) os << " "; |
1051 | os << "id:" << loops_[i].header->id(); |
1052 | } |
1053 | os << ")"; |
1054 | } |
1055 | os << ":\n"; |
1056 | |
1057 | for (BasicBlock* block = order_; block != nullptr; |
1058 | block = block->rpo_next()) { |
1059 | os << std::setw(5) << "B" << block->rpo_number() << ":"; |
1060 | for (size_t i = 0; i < loops_.size(); i++) { |
1061 | bool range = loops_[i].header->LoopContains(block); |
1062 | bool membership = loops_[i].header != block && range; |
1063 | os << (membership ? " |" : " "); |
1064 | os << (range ? "x" : " "); |
1065 | } |
1066 | os << " id:" << block->id() << ": "; |
1067 | if (block->loop_end() != nullptr) { |
1068 | os << " range: [B" << block->rpo_number() << ", B" |
1069 | << block->loop_end()->rpo_number() << ")"; |
1070 | } |
1071 | if (block->loop_header() != nullptr) { |
1072 | os << " header: id:" << block->loop_header()->id(); |
1073 | } |
1074 | if (block->loop_depth() > 0) { |
1075 | os << " depth: " << block->loop_depth(); |
1076 | } |
1077 | os << "\n"; |
1078 | } |
1079 | } |
1080 | |
1081 | void VerifySpecialRPO() { |
1082 | BasicBlockVector* order = schedule_->rpo_order(); |
1083 | DCHECK_LT(0, order->size())((void) 0); |
1084 | DCHECK_EQ(0, (*order)[0]->id().ToInt())((void) 0); // entry should be first. |
1085 | |
1086 | for (size_t i = 0; i < loops_.size(); i++) { |
1087 | LoopInfo* loop = &loops_[i]; |
1088 | BasicBlock* header = loop->header; |
1089 | BasicBlock* end = header->loop_end(); |
1090 | |
1091 | DCHECK_NOT_NULL(header)((void) 0); |
1092 | DCHECK_LE(0, header->rpo_number())((void) 0); |
1093 | DCHECK_LT(header->rpo_number(), order->size())((void) 0); |
1094 | DCHECK_NOT_NULL(end)((void) 0); |
1095 | DCHECK_LE(end->rpo_number(), order->size())((void) 0); |
1096 | DCHECK_GT(end->rpo_number(), header->rpo_number())((void) 0); |
1097 | DCHECK_NE(header->loop_header(), header)((void) 0); |
1098 | |
1099 | // Verify the start ... end list relationship. |
1100 | int links = 0; |
1101 | BasicBlock* block = loop->start; |
1102 | DCHECK_EQ(header, block)((void) 0); |
1103 | bool end_found; |
1104 | while (true) { |
1105 | if (block == nullptr || block == loop->end) { |
1106 | end_found = (loop->end == block); |
1107 | break; |
1108 | } |
1109 | // The list should be in same order as the final result. |
1110 | DCHECK(block->rpo_number() == links + header->rpo_number())((void) 0); |
1111 | links++; |
1112 | block = block->rpo_next(); |
1113 | DCHECK_LT(links, static_cast<int>(2 * order->size()))((void) 0); // cycle? |
1114 | } |
1115 | DCHECK_LT(0, links)((void) 0); |
1116 | DCHECK_EQ(links, end->rpo_number() - header->rpo_number())((void) 0); |
1117 | DCHECK(end_found)((void) 0); |
1118 | |
1119 | // Check loop depth of the header. |
1120 | int loop_depth = 0; |
1121 | for (LoopInfo* outer = loop; outer != nullptr; outer = outer->prev) { |
1122 | loop_depth++; |
1123 | } |
1124 | DCHECK_EQ(loop_depth, header->loop_depth())((void) 0); |
1125 | |
1126 | // Check the contiguousness of loops. |
1127 | int count = 0; |
1128 | for (int j = 0; j < static_cast<int>(order->size()); j++) { |
1129 | block = order->at(j); |
1130 | DCHECK_EQ(block->rpo_number(), j)((void) 0); |
1131 | if (j < header->rpo_number() || j >= end->rpo_number()) { |
1132 | DCHECK(!header->LoopContains(block))((void) 0); |
1133 | } else { |
1134 | DCHECK(header->LoopContains(block))((void) 0); |
1135 | DCHECK_GE(block->loop_depth(), loop_depth)((void) 0); |
1136 | count++; |
1137 | } |
1138 | } |
1139 | DCHECK_EQ(links, count)((void) 0); |
1140 | } |
1141 | } |
1142 | #endif // DEBUG |
1143 | |
1144 | Zone* zone_; |
1145 | Schedule* schedule_; |
1146 | BasicBlock* order_; |
1147 | BasicBlock* beyond_end_; |
1148 | ZoneVector<LoopInfo> loops_; |
1149 | ZoneVector<Backedge> backedges_; |
1150 | ZoneVector<SpecialRPOStackFrame> stack_; |
1151 | size_t previous_block_count_; |
1152 | ZoneVector<BasicBlock*> const empty_; |
1153 | }; |
1154 | |
1155 | |
1156 | BasicBlockVector* Scheduler::ComputeSpecialRPO(Zone* zone, Schedule* schedule) { |
1157 | SpecialRPONumberer numberer(zone, schedule); |
1158 | numberer.ComputeSpecialRPO(); |
1159 | numberer.SerializeRPOIntoSchedule(); |
1160 | numberer.PrintAndVerifySpecialRPO(); |
1161 | return schedule->rpo_order(); |
1162 | } |
1163 | |
1164 | |
1165 | void Scheduler::ComputeSpecialRPONumbering() { |
1166 | TRACE("--- COMPUTING SPECIAL RPO ----------------------------------\n"); |
1167 | |
1168 | // Compute the special reverse-post-order for basic blocks. |
1169 | special_rpo_ = zone_->New<SpecialRPONumberer>(zone_, schedule_); |
1170 | special_rpo_->ComputeSpecialRPO(); |
1171 | } |
1172 | |
1173 | BasicBlock* Scheduler::GetCommonDominatorIfCached(BasicBlock* b1, |
1174 | BasicBlock* b2) { |
1175 | auto entry1 = common_dominator_cache_.find(b1->id().ToInt()); |
1176 | if (entry1 == common_dominator_cache_.end()) return nullptr; |
1177 | auto entry2 = entry1->second->find(b2->id().ToInt()); |
1178 | if (entry2 == entry1->second->end()) return nullptr; |
1179 | return entry2->second; |
1180 | } |
1181 | |
1182 | BasicBlock* Scheduler::GetCommonDominator(BasicBlock* b1, BasicBlock* b2) { |
1183 | // A very common fast case: |
1184 | if (b1 == b2) return b1; |
1185 | // Try to find the common dominator by walking, if there is a chance of |
1186 | // finding it quickly. |
1187 | constexpr int kCacheGranularity = 63; |
1188 | STATIC_ASSERT((kCacheGranularity & (kCacheGranularity + 1)) == 0)static_assert((kCacheGranularity & (kCacheGranularity + 1 )) == 0, "(kCacheGranularity & (kCacheGranularity + 1)) == 0" ); |
1189 | int depth_difference = b1->dominator_depth() - b2->dominator_depth(); |
1190 | if (depth_difference > -kCacheGranularity && |
1191 | depth_difference < kCacheGranularity) { |
1192 | for (int i = 0; i < kCacheGranularity; i++) { |
1193 | if (b1->dominator_depth() < b2->dominator_depth()) { |
1194 | b2 = b2->dominator(); |
1195 | } else { |
1196 | b1 = b1->dominator(); |
1197 | } |
1198 | if (b1 == b2) return b1; |
1199 | } |
1200 | // We might fall out of the loop here if the dominator tree has several |
1201 | // deep "parallel" subtrees. |
1202 | } |
1203 | // If it'd be a long walk, take the bus instead (i.e. use the cache). |
1204 | // To keep memory consumption low, there'll be a bus stop every 64 blocks. |
1205 | // First, walk to the nearest bus stop. |
1206 | if (b1->dominator_depth() < b2->dominator_depth()) std::swap(b1, b2); |
1207 | while ((b1->dominator_depth() & kCacheGranularity) != 0) { |
1208 | if (V8_LIKELY(b1->dominator_depth() > b2->dominator_depth())(__builtin_expect(!!(b1->dominator_depth() > b2->dominator_depth ()), 1))) { |
1209 | b1 = b1->dominator(); |
1210 | } else { |
1211 | b2 = b2->dominator(); |
1212 | } |
1213 | if (b1 == b2) return b1; |
1214 | } |
1215 | // Then, walk from bus stop to bus stop until we either find a bus (i.e. an |
1216 | // existing cache entry) or the result. Make a list of any empty bus stops |
1217 | // we'd like to populate for next time. |
1218 | constexpr int kMaxNewCacheEntries = 2 * 50; // Must be even. |
1219 | // This array stores a flattened list of pairs, e.g. if after finding the |
1220 | // {result}, we want to cache [(B11, B12) -> result, (B21, B22) -> result], |
1221 | // then we store [11, 12, 21, 22] here. |
1222 | int new_cache_entries[kMaxNewCacheEntries]; |
1223 | // Next free slot in {new_cache_entries}. |
1224 | int new_cache_entries_cursor = 0; |
1225 | while (b1 != b2) { |
1226 | if ((b1->dominator_depth() & kCacheGranularity) == 0) { |
1227 | BasicBlock* maybe_cache_hit = GetCommonDominatorIfCached(b1, b2); |
1228 | if (maybe_cache_hit != nullptr) { |
1229 | b1 = b2 = maybe_cache_hit; |
Although the value stored to 'b2' is used in the enclosing expression, the value is never actually read from 'b2' | |
1230 | break; |
1231 | } else if (new_cache_entries_cursor < kMaxNewCacheEntries) { |
1232 | new_cache_entries[new_cache_entries_cursor++] = b1->id().ToInt(); |
1233 | new_cache_entries[new_cache_entries_cursor++] = b2->id().ToInt(); |
1234 | } |
1235 | } |
1236 | if (V8_LIKELY(b1->dominator_depth() > b2->dominator_depth())(__builtin_expect(!!(b1->dominator_depth() > b2->dominator_depth ()), 1))) { |
1237 | b1 = b1->dominator(); |
1238 | } else { |
1239 | b2 = b2->dominator(); |
1240 | } |
1241 | } |
1242 | // Lastly, create new cache entries we noted down earlier. |
1243 | BasicBlock* result = b1; |
1244 | for (int i = 0; i < new_cache_entries_cursor;) { |
1245 | int id1 = new_cache_entries[i++]; |
1246 | int id2 = new_cache_entries[i++]; |
1247 | ZoneMap<int, BasicBlock*>* mapping; |
1248 | auto entry = common_dominator_cache_.find(id1); |
1249 | if (entry == common_dominator_cache_.end()) { |
1250 | mapping = zone_->New<ZoneMap<int, BasicBlock*>>(zone_); |
1251 | common_dominator_cache_[id1] = mapping; |
1252 | } else { |
1253 | mapping = entry->second; |
1254 | } |
1255 | // If there was an existing entry, we would have found it earlier. |
1256 | DCHECK_EQ(mapping->find(id2), mapping->end())((void) 0); |
1257 | mapping->insert({id2, result}); |
1258 | } |
1259 | return result; |
1260 | } |
1261 | |
1262 | void Scheduler::PropagateImmediateDominators(BasicBlock* block) { |
1263 | for (/*nop*/; block != nullptr; block = block->rpo_next()) { |
1264 | auto pred = block->predecessors().begin(); |
1265 | auto end = block->predecessors().end(); |
1266 | DCHECK(pred != end)((void) 0); // All blocks except start have predecessors. |
1267 | BasicBlock* dominator = *pred; |
1268 | bool deferred = dominator->deferred(); |
1269 | // For multiple predecessors, walk up the dominator tree until a common |
1270 | // dominator is found. Visitation order guarantees that all predecessors |
1271 | // except for backwards edges have been visited. |
1272 | // We use a one-element cache for previously-seen dominators. This gets |
1273 | // hit a lot for functions that have long chains of diamonds, and in |
1274 | // those cases turns quadratic into linear complexity. |
1275 | BasicBlock* cache = nullptr; |
1276 | for (++pred; pred != end; ++pred) { |
1277 | // Don't examine backwards edges. |
1278 | if ((*pred)->dominator_depth() < 0) continue; |
1279 | if ((*pred)->dominator_depth() > 3 && |
1280 | ((*pred)->dominator()->dominator() == cache || |
1281 | (*pred)->dominator()->dominator()->dominator() == cache)) { |
1282 | // Nothing to do, the last iteration covered this case. |
1283 | DCHECK_EQ(dominator, BasicBlock::GetCommonDominator(dominator, *pred))((void) 0); |
1284 | } else { |
1285 | dominator = BasicBlock::GetCommonDominator(dominator, *pred); |
1286 | } |
1287 | cache = (*pred)->dominator(); |
1288 | deferred = deferred & (*pred)->deferred(); |
1289 | } |
1290 | block->set_dominator(dominator); |
1291 | block->set_dominator_depth(dominator->dominator_depth() + 1); |
1292 | block->set_deferred(deferred | block->deferred()); |
1293 | TRACE("Block id:%d's idom is id:%d, depth = %d\n", block->id().ToInt(), |
1294 | dominator->id().ToInt(), block->dominator_depth()); |
1295 | } |
1296 | } |
1297 | |
1298 | void Scheduler::GenerateDominatorTree(Schedule* schedule) { |
1299 | // Seed start block to be the first dominator. |
1300 | schedule->start()->set_dominator_depth(0); |
1301 | |
1302 | // Build the block dominator tree resulting from the above seed. |
1303 | PropagateImmediateDominators(schedule->start()->rpo_next()); |
1304 | } |
1305 | |
1306 | void Scheduler::GenerateDominatorTree() { |
1307 | TRACE("--- IMMEDIATE BLOCK DOMINATORS -----------------------------\n"); |
1308 | GenerateDominatorTree(schedule_); |
1309 | } |
1310 | |
1311 | // ----------------------------------------------------------------------------- |
1312 | // Phase 3: Prepare use counts for nodes. |
1313 | |
1314 | |
1315 | class PrepareUsesVisitor { |
1316 | public: |
1317 | explicit PrepareUsesVisitor(Scheduler* scheduler, Graph* graph, Zone* zone) |
1318 | : scheduler_(scheduler), |
1319 | schedule_(scheduler->schedule_), |
1320 | graph_(graph), |
1321 | visited_(graph_->NodeCount(), false, zone), |
1322 | stack_(zone) {} |
1323 | |
1324 | void Run() { |
1325 | InitializePlacement(graph_->end()); |
1326 | while (!stack_.empty()) { |
1327 | Node* node = stack_.top(); |
1328 | stack_.pop(); |
1329 | VisitInputs(node); |
1330 | } |
1331 | } |
1332 | |
1333 | private: |
1334 | void InitializePlacement(Node* node) { |
1335 | TRACE("Pre #%d:%s\n", node->id(), node->op()->mnemonic()); |
1336 | DCHECK(!Visited(node))((void) 0); |
1337 | if (scheduler_->InitializePlacement(node) == Scheduler::kFixed) { |
1338 | // Fixed nodes are always roots for schedule late. |
1339 | scheduler_->schedule_root_nodes_.push_back(node); |
1340 | if (!schedule_->IsScheduled(node)) { |
1341 | // Make sure root nodes are scheduled in their respective blocks. |
1342 | TRACE("Scheduling fixed position node #%d:%s\n", node->id(), |
1343 | node->op()->mnemonic()); |
1344 | IrOpcode::Value opcode = node->opcode(); |
1345 | BasicBlock* block = |
1346 | opcode == IrOpcode::kParameter |
1347 | ? schedule_->start() |
1348 | : schedule_->block(NodeProperties::GetControlInput(node)); |
1349 | DCHECK_NOT_NULL(block)((void) 0); |
1350 | schedule_->AddNode(block, node); |
1351 | } |
1352 | } |
1353 | stack_.push(node); |
1354 | visited_[node->id()] = true; |
1355 | } |
1356 | |
1357 | void VisitInputs(Node* node) { |
1358 | DCHECK_NE(scheduler_->GetPlacement(node), Scheduler::kUnknown)((void) 0); |
1359 | bool is_scheduled = schedule_->IsScheduled(node); |
1360 | base::Optional<int> coupled_control_edge = |
1361 | scheduler_->GetCoupledControlEdge(node); |
1362 | for (auto edge : node->input_edges()) { |
1363 | Node* to = edge.to(); |
1364 | DCHECK_EQ(node, edge.from())((void) 0); |
1365 | if (!Visited(to)) { |
1366 | InitializePlacement(to); |
1367 | } |
1368 | TRACE("PostEdge #%d:%s->#%d:%s\n", node->id(), node->op()->mnemonic(), |
1369 | to->id(), to->op()->mnemonic()); |
1370 | DCHECK_NE(scheduler_->GetPlacement(to), Scheduler::kUnknown)((void) 0); |
1371 | if (!is_scheduled && edge.index() != coupled_control_edge) { |
1372 | scheduler_->IncrementUnscheduledUseCount(to, node); |
1373 | } |
1374 | } |
1375 | } |
1376 | |
1377 | bool Visited(Node* node) { return visited_[node->id()]; } |
1378 | |
1379 | Scheduler* scheduler_; |
1380 | Schedule* schedule_; |
1381 | Graph* graph_; |
1382 | BoolVector visited_; |
1383 | ZoneStack<Node*> stack_; |
1384 | }; |
1385 | |
1386 | |
1387 | void Scheduler::PrepareUses() { |
1388 | TRACE("--- PREPARE USES -------------------------------------------\n"); |
1389 | |
1390 | // Count the uses of every node, which is used to ensure that all of a |
1391 | // node's uses are scheduled before the node itself. |
1392 | PrepareUsesVisitor prepare_uses(this, graph_, zone_); |
1393 | prepare_uses.Run(); |
1394 | } |
1395 | |
1396 | |
1397 | // ----------------------------------------------------------------------------- |
1398 | // Phase 4: Schedule nodes early. |
1399 | |
1400 | |
1401 | class ScheduleEarlyNodeVisitor { |
1402 | public: |
1403 | ScheduleEarlyNodeVisitor(Zone* zone, Scheduler* scheduler) |
1404 | : scheduler_(scheduler), schedule_(scheduler->schedule_), queue_(zone) {} |
1405 | |
1406 | // Run the schedule early algorithm on a set of fixed root nodes. |
1407 | void Run(NodeVector* roots) { |
1408 | for (Node* const root : *roots) { |
1409 | queue_.push(root); |
1410 | } |
1411 | |
1412 | while (!queue_.empty()) { |
1413 | scheduler_->tick_counter_->TickAndMaybeEnterSafepoint(); |
1414 | VisitNode(queue_.front()); |
1415 | queue_.pop(); |
1416 | } |
1417 | } |
1418 | |
1419 | private: |
1420 | // Visits one node from the queue and propagates its current schedule early |
1421 | // position to all uses. This in turn might push more nodes onto the queue. |
1422 | void VisitNode(Node* node) { |
1423 | Scheduler::SchedulerData* data = scheduler_->GetData(node); |
1424 | |
1425 | // Fixed nodes already know their schedule early position. |
1426 | if (scheduler_->GetPlacement(node) == Scheduler::kFixed) { |
1427 | data->minimum_block_ = schedule_->block(node); |
1428 | TRACE("Fixing #%d:%s minimum_block = id:%d, dominator_depth = %d\n", |
1429 | node->id(), node->op()->mnemonic(), |
1430 | data->minimum_block_->id().ToInt(), |
1431 | data->minimum_block_->dominator_depth()); |
1432 | } |
1433 | |
1434 | // No need to propagate unconstrained schedule early positions. |
1435 | if (data->minimum_block_ == schedule_->start()) return; |
1436 | |
1437 | // Propagate schedule early position. |
1438 | DCHECK_NOT_NULL(data->minimum_block_)((void) 0); |
1439 | for (auto use : node->uses()) { |
1440 | if (scheduler_->IsLive(use)) { |
1441 | PropagateMinimumPositionToNode(data->minimum_block_, use); |
1442 | } |
1443 | } |
1444 | } |
1445 | |
1446 | // Propagates {block} as another minimum position into the given {node}. This |
1447 | // has the net effect of computing the minimum dominator block of {node} that |
1448 | // still post-dominates all inputs to {node} when the queue is processed. |
1449 | void PropagateMinimumPositionToNode(BasicBlock* block, Node* node) { |
1450 | Scheduler::SchedulerData* data = scheduler_->GetData(node); |
1451 | |
1452 | // No need to propagate to fixed node, it's guaranteed to be a root. |
1453 | if (scheduler_->GetPlacement(node) == Scheduler::kFixed) return; |
1454 | |
1455 | // Coupled nodes influence schedule early position of their control. |
1456 | if (scheduler_->GetPlacement(node) == Scheduler::kCoupled) { |
1457 | Node* control = NodeProperties::GetControlInput(node); |
1458 | PropagateMinimumPositionToNode(block, control); |
1459 | } |
1460 | |
1461 | // Propagate new position if it is deeper down the dominator tree than the |
1462 | // current. Note that all inputs need to have minimum block position inside |
1463 | // the dominator chain of {node}'s minimum block position. |
1464 | DCHECK(InsideSameDominatorChain(block, data->minimum_block_))((void) 0); |
1465 | if (block->dominator_depth() > data->minimum_block_->dominator_depth()) { |
1466 | data->minimum_block_ = block; |
1467 | queue_.push(node); |
1468 | TRACE("Propagating #%d:%s minimum_block = id:%d, dominator_depth = %d\n", |
1469 | node->id(), node->op()->mnemonic(), |
1470 | data->minimum_block_->id().ToInt(), |
1471 | data->minimum_block_->dominator_depth()); |
1472 | } |
1473 | } |
1474 | |
1475 | #if DEBUG |
1476 | bool InsideSameDominatorChain(BasicBlock* b1, BasicBlock* b2) { |
1477 | BasicBlock* dominator = BasicBlock::GetCommonDominator(b1, b2); |
1478 | return dominator == b1 || dominator == b2; |
1479 | } |
1480 | #endif |
1481 | |
1482 | Scheduler* scheduler_; |
1483 | Schedule* schedule_; |
1484 | ZoneQueue<Node*> queue_; |
1485 | }; |
1486 | |
1487 | |
1488 | void Scheduler::ScheduleEarly() { |
1489 | if (!special_rpo_->HasLoopBlocks()) { |
1490 | TRACE("--- NO LOOPS SO SKIPPING SCHEDULE EARLY --------------------\n"); |
1491 | return; |
1492 | } |
1493 | |
1494 | TRACE("--- SCHEDULE EARLY -----------------------------------------\n"); |
1495 | if (FLAG_trace_turbo_scheduler) { |
1496 | TRACE("roots: "); |
1497 | for (Node* node : schedule_root_nodes_) { |
1498 | TRACE("#%d:%s ", node->id(), node->op()->mnemonic()); |
1499 | } |
1500 | TRACE("\n"); |
1501 | } |
1502 | |
1503 | // Compute the minimum block for each node thereby determining the earliest |
1504 | // position each node could be placed within a valid schedule. |
1505 | ScheduleEarlyNodeVisitor schedule_early_visitor(zone_, this); |
1506 | schedule_early_visitor.Run(&schedule_root_nodes_); |
1507 | } |
1508 | |
1509 | |
1510 | // ----------------------------------------------------------------------------- |
1511 | // Phase 5: Schedule nodes late. |
1512 | |
1513 | |
1514 | class ScheduleLateNodeVisitor { |
1515 | public: |
1516 | ScheduleLateNodeVisitor(Zone* zone, Scheduler* scheduler) |
1517 | : zone_(zone), |
1518 | scheduler_(scheduler), |
1519 | schedule_(scheduler_->schedule_), |
1520 | marked_(scheduler->zone_), |
1521 | marking_queue_(scheduler->zone_) {} |
1522 | |
1523 | // Run the schedule late algorithm on a set of fixed root nodes. |
1524 | void Run(NodeVector* roots) { |
1525 | for (Node* const root : *roots) { |
1526 | ProcessQueue(root); |
1527 | } |
1528 | } |
1529 | |
1530 | private: |
1531 | void ProcessQueue(Node* root) { |
1532 | ZoneQueue<Node*>* queue = &(scheduler_->schedule_queue_); |
1533 | for (Node* node : root->inputs()) { |
1534 | // Don't schedule coupled nodes on their own. |
1535 | if (scheduler_->GetPlacement(node) == Scheduler::kCoupled) { |
1536 | node = NodeProperties::GetControlInput(node); |
1537 | } |
1538 | |
1539 | // Test schedulability condition by looking at unscheduled use count. |
1540 | if (scheduler_->GetData(node)->unscheduled_count_ != 0) continue; |
1541 | |
1542 | queue->push(node); |
1543 | do { |
1544 | scheduler_->tick_counter_->TickAndMaybeEnterSafepoint(); |
1545 | Node* const n = queue->front(); |
1546 | queue->pop(); |
1547 | VisitNode(n); |
1548 | } while (!queue->empty()); |
1549 | } |
1550 | } |
1551 | |
1552 | // Visits one node from the queue of schedulable nodes and determines its |
1553 | // schedule late position. Also hoists nodes out of loops to find a more |
1554 | // optimal scheduling position. |
1555 | void VisitNode(Node* node) { |
1556 | DCHECK_EQ(0, scheduler_->GetData(node)->unscheduled_count_)((void) 0); |
1557 | |
1558 | // Don't schedule nodes that are already scheduled. |
1559 | if (schedule_->IsScheduled(node)) return; |
1560 | DCHECK_EQ(Scheduler::kSchedulable, scheduler_->GetPlacement(node))((void) 0); |
1561 | |
1562 | // Determine the dominating block for all of the uses of this node. It is |
1563 | // the latest block that this node can be scheduled in. |
1564 | TRACE("Scheduling #%d:%s\n", node->id(), node->op()->mnemonic()); |
1565 | BasicBlock* block = GetCommonDominatorOfUses(node); |
1566 | DCHECK_NOT_NULL(block)((void) 0); |
1567 | |
1568 | // The schedule early block dominates the schedule late block. |
1569 | BasicBlock* min_block = scheduler_->GetData(node)->minimum_block_; |
1570 | DCHECK_EQ(min_block, BasicBlock::GetCommonDominator(block, min_block))((void) 0); |
1571 | |
1572 | TRACE( |
1573 | "Schedule late of #%d:%s is id:%d at loop depth %d, minimum = id:%d\n", |
1574 | node->id(), node->op()->mnemonic(), block->id().ToInt(), |
1575 | block->loop_depth(), min_block->id().ToInt()); |
1576 | |
1577 | // Hoist nodes out of loops if possible. Nodes can be hoisted iteratively |
1578 | // into enclosing loop pre-headers until they would precede their schedule |
1579 | // early position. |
1580 | BasicBlock* hoist_block = GetHoistBlock(block); |
1581 | if (hoist_block && |
1582 | hoist_block->dominator_depth() >= min_block->dominator_depth()) { |
1583 | DCHECK(scheduler_->special_rpo_->HasLoopBlocks())((void) 0); |
1584 | do { |
1585 | TRACE(" hoisting #%d:%s to block id:%d\n", node->id(), |
1586 | node->op()->mnemonic(), hoist_block->id().ToInt()); |
1587 | DCHECK_LT(hoist_block->loop_depth(), block->loop_depth())((void) 0); |
1588 | block = hoist_block; |
1589 | hoist_block = GetHoistBlock(hoist_block); |
1590 | } while (hoist_block && |
1591 | hoist_block->dominator_depth() >= min_block->dominator_depth()); |
1592 | } else if (scheduler_->flags_ & Scheduler::kSplitNodes) { |
1593 | // Split the {node} if beneficial and return the new {block} for it. |
1594 | block = SplitNode(block, node); |
1595 | } |
1596 | |
1597 | // Schedule the node or a floating control structure. |
1598 | if (IrOpcode::IsMergeOpcode(node->opcode())) { |
1599 | ScheduleFloatingControl(block, node); |
1600 | } else if (node->opcode() == IrOpcode::kFinishRegion) { |
1601 | ScheduleRegion(block, node); |
1602 | } else { |
1603 | ScheduleNode(block, node); |
1604 | } |
1605 | } |
1606 | |
1607 | bool IsMarked(BasicBlock* block) const { |
1608 | DCHECK_LT(block->id().ToSize(), marked_.size())((void) 0); |
1609 | return marked_[block->id().ToSize()]; |
1610 | } |
1611 | |
1612 | void Mark(BasicBlock* block) { marked_[block->id().ToSize()] = true; } |
1613 | |
1614 | // Mark {block} and push its non-marked predecessor on the marking queue. |
1615 | void MarkBlock(BasicBlock* block) { |
1616 | DCHECK_LT(block->id().ToSize(), marked_.size())((void) 0); |
1617 | Mark(block); |
1618 | for (BasicBlock* pred_block : block->predecessors()) { |
1619 | if (IsMarked(pred_block)) continue; |
1620 | marking_queue_.push_back(pred_block); |
1621 | } |
1622 | } |
1623 | |
1624 | BasicBlock* SplitNode(BasicBlock* block, Node* node) { |
1625 | // For now, we limit splitting to pure nodes. |
1626 | if (!node->op()->HasProperty(Operator::kPure)) return block; |
1627 | // TODO(titzer): fix the special case of splitting of projections. |
1628 | if (node->opcode() == IrOpcode::kProjection) return block; |
1629 | |
1630 | // The {block} is common dominator of all uses of {node}, so we cannot |
1631 | // split anything unless the {block} has at least two successors. |
1632 | DCHECK_EQ(block, GetCommonDominatorOfUses(node))((void) 0); |
1633 | if (block->SuccessorCount() < 2) return block; |
1634 | |
1635 | // Clear marking bits. |
1636 | DCHECK(marking_queue_.empty())((void) 0); |
1637 | std::fill(marked_.begin(), marked_.end(), false); |
1638 | marked_.resize(schedule_->BasicBlockCount() + 1, false); |
1639 | |
1640 | // Check if the {node} has uses in {block}. |
1641 | for (Edge edge : node->use_edges()) { |
1642 | if (!scheduler_->IsLive(edge.from())) continue; |
1643 | BasicBlock* use_block = GetBlockForUse(edge); |
1644 | if (use_block == nullptr || IsMarked(use_block)) continue; |
1645 | if (use_block == block) { |
1646 | TRACE(" not splitting #%d:%s, it is used in id:%d\n", node->id(), |
1647 | node->op()->mnemonic(), block->id().ToInt()); |
1648 | marking_queue_.clear(); |
1649 | return block; |
1650 | } |
1651 | MarkBlock(use_block); |
1652 | } |
1653 | |
1654 | // Compute transitive marking closure; a block is marked if all its |
1655 | // successors are marked. |
1656 | do { |
1657 | BasicBlock* top_block = marking_queue_.front(); |
1658 | marking_queue_.pop_front(); |
1659 | if (IsMarked(top_block)) continue; |
1660 | bool marked = true; |
1661 | for (BasicBlock* successor : top_block->successors()) { |
1662 | if (!IsMarked(successor)) { |
1663 | marked = false; |
1664 | break; |
1665 | } |
1666 | } |
1667 | if (marked) MarkBlock(top_block); |
1668 | } while (!marking_queue_.empty()); |
1669 | |
1670 | // If the (common dominator) {block} is marked, we know that all paths from |
1671 | // {block} to the end contain at least one use of {node}, and hence there's |
1672 | // no point in splitting the {node} in this case. |
1673 | if (IsMarked(block)) { |
1674 | TRACE(" not splitting #%d:%s, its common dominator id:%d is perfect\n", |
1675 | node->id(), node->op()->mnemonic(), block->id().ToInt()); |
1676 | return block; |
1677 | } |
1678 | |
1679 | // Split {node} for uses according to the previously computed marking |
1680 | // closure. Every marking partition has a unique dominator, which get's a |
1681 | // copy of the {node} with the exception of the first partition, which get's |
1682 | // the {node} itself. |
1683 | ZoneMap<BasicBlock*, Node*> dominators(scheduler_->zone_); |
1684 | for (Edge edge : node->use_edges()) { |
1685 | if (!scheduler_->IsLive(edge.from())) continue; |
1686 | BasicBlock* use_block = GetBlockForUse(edge); |
1687 | if (use_block == nullptr) continue; |
1688 | while (IsMarked(use_block->dominator())) { |
1689 | use_block = use_block->dominator(); |
1690 | } |
1691 | auto& use_node = dominators[use_block]; |
1692 | if (use_node == nullptr) { |
1693 | if (dominators.size() == 1u) { |
1694 | // Place the {node} at {use_block}. |
1695 | block = use_block; |
1696 | use_node = node; |
1697 | TRACE(" pushing #%d:%s down to id:%d\n", node->id(), |
1698 | node->op()->mnemonic(), block->id().ToInt()); |
1699 | } else { |
1700 | // Place a copy of {node} at {use_block}. |
1701 | use_node = CloneNode(node); |
1702 | TRACE(" cloning #%d:%s for id:%d\n", use_node->id(), |
1703 | use_node->op()->mnemonic(), use_block->id().ToInt()); |
1704 | scheduler_->schedule_queue_.push(use_node); |
1705 | } |
1706 | } |
1707 | edge.UpdateTo(use_node); |
1708 | } |
1709 | return block; |
1710 | } |
1711 | |
1712 | BasicBlock* GetHoistBlock(BasicBlock* block) { |
1713 | if (!scheduler_->special_rpo_->HasLoopBlocks()) return nullptr; |
1714 | if (block->IsLoopHeader()) return block->dominator(); |
1715 | // We have to check to make sure that the {block} dominates all |
1716 | // of the outgoing blocks. If it doesn't, then there is a path |
1717 | // out of the loop which does not execute this {block}, so we |
1718 | // can't hoist operations from this {block} out of the loop, as |
1719 | // that would introduce additional computations. |
1720 | if (BasicBlock* header_block = block->loop_header()) { |
1721 | for (BasicBlock* outgoing_block : |
1722 | scheduler_->special_rpo_->GetOutgoingBlocks(header_block)) { |
1723 | if (scheduler_->GetCommonDominator(block, outgoing_block) != block) { |
1724 | return nullptr; |
1725 | } |
1726 | } |
1727 | return header_block->dominator(); |
1728 | } |
1729 | return nullptr; |
1730 | } |
1731 | |
1732 | BasicBlock* GetCommonDominatorOfUses(Node* node) { |
1733 | BasicBlock* block = nullptr; |
1734 | for (Edge edge : node->use_edges()) { |
1735 | if (!scheduler_->IsLive(edge.from())) continue; |
1736 | BasicBlock* use_block = GetBlockForUse(edge); |
1737 | block = block == nullptr |
1738 | ? use_block |
1739 | : use_block == nullptr |
1740 | ? block |
1741 | : scheduler_->GetCommonDominator(block, use_block); |
1742 | } |
1743 | return block; |
1744 | } |
1745 | |
1746 | BasicBlock* FindPredecessorBlock(Node* node) { |
1747 | return scheduler_->control_flow_builder_->FindPredecessorBlock(node); |
1748 | } |
1749 | |
1750 | BasicBlock* GetBlockForUse(Edge edge) { |
1751 | // TODO(titzer): ignore uses from dead nodes (not visited in PrepareUses()). |
1752 | // Dead uses only occur if the graph is not trimmed before scheduling. |
1753 | Node* use = edge.from(); |
1754 | if (IrOpcode::IsPhiOpcode(use->opcode())) { |
1755 | // If the use is from a coupled (i.e. floating) phi, compute the common |
1756 | // dominator of its uses. This will not recurse more than one level. |
1757 | if (scheduler_->GetPlacement(use) == Scheduler::kCoupled) { |
1758 | TRACE(" inspecting uses of coupled #%d:%s\n", use->id(), |
1759 | use->op()->mnemonic()); |
1760 | // TODO(titzer): reenable once above TODO is addressed. |
1761 | // DCHECK_EQ(edge.to(), NodeProperties::GetControlInput(use)); |
1762 | return GetCommonDominatorOfUses(use); |
1763 | } |
1764 | // If the use is from a fixed (i.e. non-floating) phi, we use the |
1765 | // predecessor block of the corresponding control input to the merge. |
1766 | if (scheduler_->GetPlacement(use) == Scheduler::kFixed) { |
1767 | TRACE(" input@%d into a fixed phi #%d:%s\n", edge.index(), use->id(), |
1768 | use->op()->mnemonic()); |
1769 | Node* merge = NodeProperties::GetControlInput(use, 0); |
1770 | DCHECK(IrOpcode::IsMergeOpcode(merge->opcode()))((void) 0); |
1771 | Node* input = NodeProperties::GetControlInput(merge, edge.index()); |
1772 | return FindPredecessorBlock(input); |
1773 | } |
1774 | } else if (IrOpcode::IsMergeOpcode(use->opcode())) { |
1775 | // If the use is from a fixed (i.e. non-floating) merge, we use the |
1776 | // predecessor block of the current input to the merge. |
1777 | if (scheduler_->GetPlacement(use) == Scheduler::kFixed) { |
1778 | TRACE(" input@%d into a fixed merge #%d:%s\n", edge.index(), use->id(), |
1779 | use->op()->mnemonic()); |
1780 | return FindPredecessorBlock(edge.to()); |
1781 | } |
1782 | } |
1783 | BasicBlock* result = schedule_->block(use); |
1784 | if (result == nullptr) return nullptr; |
1785 | TRACE(" must dominate use #%d:%s in id:%d\n", use->id(), |
1786 | use->op()->mnemonic(), result->id().ToInt()); |
1787 | return result; |
1788 | } |
1789 | |
1790 | void ScheduleFloatingControl(BasicBlock* block, Node* node) { |
1791 | scheduler_->FuseFloatingControl(block, node); |
1792 | } |
1793 | |
1794 | void ScheduleRegion(BasicBlock* block, Node* region_end) { |
1795 | // We only allow regions of instructions connected into a linear |
1796 | // effect chain. The only value allowed to be produced by a node |
1797 | // in the chain must be the value consumed by the FinishRegion node. |
1798 | |
1799 | // We schedule back to front; we first schedule FinishRegion. |
1800 | CHECK_EQ(IrOpcode::kFinishRegion, region_end->opcode())do { bool _cmp = ::v8::base::CmpEQImpl< typename ::v8::base ::pass_value_or_ref<decltype(IrOpcode::kFinishRegion)>:: type, typename ::v8::base::pass_value_or_ref<decltype(region_end ->opcode())>::type>((IrOpcode::kFinishRegion), (region_end ->opcode())); do { if ((__builtin_expect(!!(!(_cmp)), 0))) { V8_Fatal("Check failed: %s.", "IrOpcode::kFinishRegion" " " "==" " " "region_end->opcode()"); } } while (false); } while (false); |
1801 | ScheduleNode(block, region_end); |
1802 | |
1803 | // Schedule the chain. |
1804 | Node* node = NodeProperties::GetEffectInput(region_end); |
1805 | while (node->opcode() != IrOpcode::kBeginRegion) { |
1806 | DCHECK_EQ(0, scheduler_->GetData(node)->unscheduled_count_)((void) 0); |
1807 | DCHECK_EQ(1, node->op()->EffectInputCount())((void) 0); |
1808 | DCHECK_EQ(1, node->op()->EffectOutputCount())((void) 0); |
1809 | DCHECK_EQ(0, node->op()->ControlOutputCount())((void) 0); |
1810 | // The value output (if there is any) must be consumed |
1811 | // by the EndRegion node. |
1812 | DCHECK(node->op()->ValueOutputCount() == 0 ||((void) 0) |
1813 | node == region_end->InputAt(0))((void) 0); |
1814 | ScheduleNode(block, node); |
1815 | node = NodeProperties::GetEffectInput(node); |
1816 | } |
1817 | // Schedule the BeginRegion node. |
1818 | DCHECK_EQ(0, scheduler_->GetData(node)->unscheduled_count_)((void) 0); |
1819 | ScheduleNode(block, node); |
1820 | } |
1821 | |
1822 | void ScheduleNode(BasicBlock* block, Node* node) { |
1823 | schedule_->PlanNode(block, node); |
1824 | size_t block_id = block->id().ToSize(); |
1825 | if (!scheduler_->scheduled_nodes_[block_id]) { |
1826 | scheduler_->scheduled_nodes_[block_id] = zone_->New<NodeVector>(zone_); |
1827 | } |
1828 | scheduler_->scheduled_nodes_[block_id]->push_back(node); |
1829 | scheduler_->UpdatePlacement(node, Scheduler::kScheduled); |
1830 | } |
1831 | |
1832 | Node* CloneNode(Node* node) { |
1833 | int const input_count = node->InputCount(); |
1834 | base::Optional<int> coupled_control_edge = |
1835 | scheduler_->GetCoupledControlEdge(node); |
1836 | for (int index = 0; index < input_count; ++index) { |
1837 | if (index != coupled_control_edge) { |
1838 | Node* const input = node->InputAt(index); |
1839 | scheduler_->IncrementUnscheduledUseCount(input, node); |
1840 | } |
1841 | } |
1842 | Node* const copy = scheduler_->graph_->CloneNode(node); |
1843 | TRACE(("clone #%d:%s -> #%d\n"), node->id(), node->op()->mnemonic(), |
1844 | copy->id()); |
1845 | scheduler_->node_data_.resize(copy->id() + 1, |
1846 | scheduler_->DefaultSchedulerData()); |
1847 | scheduler_->node_data_[copy->id()] = scheduler_->node_data_[node->id()]; |
1848 | return copy; |
1849 | } |
1850 | |
1851 | Zone* zone_; |
1852 | Scheduler* scheduler_; |
1853 | Schedule* schedule_; |
1854 | BoolVector marked_; |
1855 | ZoneDeque<BasicBlock*> marking_queue_; |
1856 | }; |
1857 | |
1858 | |
1859 | void Scheduler::ScheduleLate() { |
1860 | TRACE("--- SCHEDULE LATE ------------------------------------------\n"); |
1861 | if (FLAG_trace_turbo_scheduler) { |
1862 | TRACE("roots: "); |
1863 | for (Node* node : schedule_root_nodes_) { |
1864 | TRACE("#%d:%s ", node->id(), node->op()->mnemonic()); |
1865 | } |
1866 | TRACE("\n"); |
1867 | } |
1868 | |
1869 | // Schedule: Places nodes in dominator block of all their uses. |
1870 | ScheduleLateNodeVisitor schedule_late_visitor(zone_, this); |
1871 | schedule_late_visitor.Run(&schedule_root_nodes_); |
1872 | } |
1873 | |
1874 | |
1875 | // ----------------------------------------------------------------------------- |
1876 | // Phase 6: Seal the final schedule. |
1877 | |
1878 | |
1879 | void Scheduler::SealFinalSchedule() { |
1880 | TRACE("--- SEAL FINAL SCHEDULE ------------------------------------\n"); |
1881 | |
1882 | // Serialize the assembly order and reverse-post-order numbering. |
1883 | special_rpo_->SerializeRPOIntoSchedule(); |
1884 | special_rpo_->PrintAndVerifySpecialRPO(); |
1885 | |
1886 | // Add collected nodes for basic blocks to their blocks in the right order. |
1887 | int block_num = 0; |
1888 | for (NodeVector* nodes : scheduled_nodes_) { |
1889 | BasicBlock::Id id = BasicBlock::Id::FromInt(block_num++); |
1890 | BasicBlock* block = schedule_->GetBlockById(id); |
1891 | if (nodes) { |
1892 | for (Node* node : base::Reversed(*nodes)) { |
1893 | schedule_->AddNode(block, node); |
1894 | } |
1895 | } |
1896 | } |
1897 | } |
1898 | |
1899 | |
1900 | // ----------------------------------------------------------------------------- |
1901 | |
1902 | |
1903 | void Scheduler::FuseFloatingControl(BasicBlock* block, Node* node) { |
1904 | TRACE("--- FUSE FLOATING CONTROL ----------------------------------\n"); |
1905 | if (FLAG_trace_turbo_scheduler) { |
1906 | StdoutStream{} << "Schedule before control flow fusion:\n" << *schedule_; |
1907 | } |
1908 | |
1909 | // Iterate on phase 1: Build control-flow graph. |
1910 | control_flow_builder_->Run(block, node); |
1911 | |
1912 | // Iterate on phase 2: Compute special RPO and dominator tree. |
1913 | special_rpo_->UpdateSpecialRPO(block, schedule_->block(node)); |
1914 | // TODO(turbofan): Currently "iterate on" means "re-run". Fix that. |
1915 | for (BasicBlock* b = block->rpo_next(); b != nullptr; b = b->rpo_next()) { |
1916 | b->set_dominator_depth(-1); |
1917 | b->set_dominator(nullptr); |
1918 | } |
1919 | PropagateImmediateDominators(block->rpo_next()); |
1920 | |
1921 | // Iterate on phase 4: Schedule nodes early. |
1922 | // TODO(turbofan): The following loop gathering the propagation roots is a |
1923 | // temporary solution and should be merged into the rest of the scheduler as |
1924 | // soon as the approach settled for all floating loops. |
1925 | NodeVector propagation_roots(control_flow_builder_->control_); |
1926 | for (Node* control : control_flow_builder_->control_) { |
1927 | for (Node* use : control->uses()) { |
1928 | if (NodeProperties::IsPhi(use) && IsLive(use)) { |
1929 | propagation_roots.push_back(use); |
1930 | } |
1931 | } |
1932 | } |
1933 | if (FLAG_trace_turbo_scheduler) { |
1934 | TRACE("propagation roots: "); |
1935 | for (Node* r : propagation_roots) { |
1936 | TRACE("#%d:%s ", r->id(), r->op()->mnemonic()); |
1937 | } |
1938 | TRACE("\n"); |
1939 | } |
1940 | ScheduleEarlyNodeVisitor schedule_early_visitor(zone_, this); |
1941 | schedule_early_visitor.Run(&propagation_roots); |
1942 | |
1943 | // Move previously planned nodes. |
1944 | // TODO(turbofan): Improve that by supporting bulk moves. |
1945 | scheduled_nodes_.resize(schedule_->BasicBlockCount()); |
1946 | MovePlannedNodes(block, schedule_->block(node)); |
1947 | |
1948 | if (FLAG_trace_turbo_scheduler) { |
1949 | StdoutStream{} << "Schedule after control flow fusion:\n" << *schedule_; |
1950 | } |
1951 | } |
1952 | |
1953 | |
1954 | void Scheduler::MovePlannedNodes(BasicBlock* from, BasicBlock* to) { |
1955 | TRACE("Move planned nodes from id:%d to id:%d\n", from->id().ToInt(), |
1956 | to->id().ToInt()); |
1957 | NodeVector* from_nodes = scheduled_nodes_[from->id().ToSize()]; |
1958 | NodeVector* to_nodes = scheduled_nodes_[to->id().ToSize()]; |
1959 | if (!from_nodes) return; |
1960 | |
1961 | for (Node* const node : *from_nodes) { |
1962 | schedule_->SetBlockForNode(to, node); |
1963 | } |
1964 | if (to_nodes) { |
1965 | to_nodes->insert(to_nodes->end(), from_nodes->begin(), from_nodes->end()); |
1966 | from_nodes->clear(); |
1967 | } else { |
1968 | std::swap(scheduled_nodes_[from->id().ToSize()], |
1969 | scheduled_nodes_[to->id().ToSize()]); |
1970 | } |
1971 | } |
1972 | |
1973 | #undef TRACE |
1974 | |
1975 | } // namespace compiler |
1976 | } // namespace internal |
1977 | } // namespace v8 |