Aart Bik | 30efb4e | 2015-07-30 12:14:31 -0700 | [diff] [blame^] | 1 | /* |
| 2 | * Copyright (C) 2015 The Android Open Source Project |
| 3 | * |
| 4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | * you may not use this file except in compliance with the License. |
| 6 | * You may obtain a copy of the License at |
| 7 | * |
| 8 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | * |
| 10 | * Unless required by applicable law or agreed to in writing, software |
| 11 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | * See the License for the specific language governing permissions and |
| 14 | * limitations under the License. |
| 15 | */ |
| 16 | |
| 17 | #include "induction_var_analysis.h" |
| 18 | |
| 19 | namespace art { |
| 20 | |
| 21 | /** |
| 22 | * Returns true if instruction is invariant within the given loop. |
| 23 | */ |
| 24 | static bool IsLoopInvariant(HLoopInformation* loop, HInstruction* instruction) { |
| 25 | HLoopInformation* other_loop = instruction->GetBlock()->GetLoopInformation(); |
| 26 | if (other_loop != loop) { |
| 27 | // If instruction does not occur in same loop, it is invariant |
| 28 | // if it appears in an outer loop (including no loop at all). |
| 29 | return other_loop == nullptr || loop->IsIn(*other_loop); |
| 30 | } |
| 31 | return false; |
| 32 | } |
| 33 | |
| 34 | /** |
| 35 | * Returns true if instruction is proper entry-phi-operation for given loop |
| 36 | * (referred to as mu-operation in Gerlek's paper). |
| 37 | */ |
| 38 | static bool IsEntryPhi(HLoopInformation* loop, HInstruction* instruction) { |
| 39 | return |
| 40 | instruction->IsPhi() && |
| 41 | instruction->InputCount() == 2 && |
| 42 | instruction->GetBlock() == loop->GetHeader(); |
| 43 | } |
| 44 | |
| 45 | // |
| 46 | // Class methods. |
| 47 | // |
| 48 | |
| 49 | HInductionVarAnalysis::HInductionVarAnalysis(HGraph* graph) |
| 50 | : HOptimization(graph, kInductionPassName), |
| 51 | global_depth_(0), |
| 52 | stack_(graph->GetArena()->Adapter()), |
| 53 | scc_(graph->GetArena()->Adapter()), |
| 54 | map_(std::less<int>(), graph->GetArena()->Adapter()), |
| 55 | cycle_(std::less<int>(), graph->GetArena()->Adapter()), |
| 56 | induction_(std::less<int>(), graph->GetArena()->Adapter()) { |
| 57 | } |
| 58 | |
| 59 | void HInductionVarAnalysis::Run() { |
| 60 | // Detects sequence variables (generalized induction variables) during an |
| 61 | // inner-loop-first traversal of all loops using Gerlek's algorithm. |
| 62 | for (HPostOrderIterator it_graph(*graph_); !it_graph.Done(); it_graph.Advance()) { |
| 63 | HBasicBlock* graph_block = it_graph.Current(); |
| 64 | if (graph_block->IsLoopHeader()) { |
| 65 | VisitLoop(graph_block->GetLoopInformation()); |
| 66 | } |
| 67 | } |
| 68 | } |
| 69 | |
| 70 | void HInductionVarAnalysis::VisitLoop(HLoopInformation* loop) { |
| 71 | // Find strongly connected components (SSCs) in the SSA graph of this loop using Tarjan's |
| 72 | // algorithm. Due to the descendant-first nature, classification happens "on-demand". |
| 73 | global_depth_ = 0; |
| 74 | CHECK(stack_.empty()); |
| 75 | map_.clear(); |
| 76 | |
| 77 | for (HBlocksInLoopIterator it_loop(*loop); !it_loop.Done(); it_loop.Advance()) { |
| 78 | HBasicBlock* loop_block = it_loop.Current(); |
| 79 | CHECK(loop_block->IsInLoop()); |
| 80 | if (loop_block->GetLoopInformation() != loop) { |
| 81 | continue; // Inner loops already visited. |
| 82 | } |
| 83 | // Visit phi-operations and instructions. |
| 84 | for (HInstructionIterator it(loop_block->GetPhis()); !it.Done(); it.Advance()) { |
| 85 | HInstruction* instruction = it.Current(); |
| 86 | if (!IsVisitedNode(instruction->GetId())) { |
| 87 | VisitNode(loop, instruction); |
| 88 | } |
| 89 | } |
| 90 | for (HInstructionIterator it(loop_block->GetInstructions()); !it.Done(); it.Advance()) { |
| 91 | HInstruction* instruction = it.Current(); |
| 92 | if (!IsVisitedNode(instruction->GetId())) { |
| 93 | VisitNode(loop, instruction); |
| 94 | } |
| 95 | } |
| 96 | } |
| 97 | |
| 98 | CHECK(stack_.empty()); |
| 99 | map_.clear(); |
| 100 | } |
| 101 | |
| 102 | void HInductionVarAnalysis::VisitNode(HLoopInformation* loop, HInstruction* instruction) { |
| 103 | const int id = instruction->GetId(); |
| 104 | const uint32_t d1 = ++global_depth_; |
| 105 | map_.Put(id, NodeInfo(d1)); |
| 106 | stack_.push_back(instruction); |
| 107 | |
| 108 | // Visit all descendants. |
| 109 | uint32_t low = d1; |
| 110 | for (size_t i = 0, count = instruction->InputCount(); i < count; ++i) { |
| 111 | low = std::min(low, VisitDescendant(loop, instruction->InputAt(i))); |
| 112 | } |
| 113 | |
| 114 | // Lower or found SCC? |
| 115 | if (low < d1) { |
| 116 | map_.find(id)->second.depth = low; |
| 117 | } else { |
| 118 | scc_.clear(); |
| 119 | cycle_.clear(); |
| 120 | |
| 121 | // Pop the stack to build the SCC for classification. |
| 122 | while (!stack_.empty()) { |
| 123 | HInstruction* x = stack_.back(); |
| 124 | scc_.push_back(x); |
| 125 | stack_.pop_back(); |
| 126 | map_.find(x->GetId())->second.done = true; |
| 127 | if (x == instruction) { |
| 128 | break; |
| 129 | } |
| 130 | } |
| 131 | |
| 132 | // Classify the SCC. |
| 133 | if (scc_.size() == 1 && !IsEntryPhi(loop, scc_[0])) { |
| 134 | ClassifyTrivial(loop, scc_[0]); |
| 135 | } else { |
| 136 | ClassifyNonTrivial(loop); |
| 137 | } |
| 138 | |
| 139 | scc_.clear(); |
| 140 | cycle_.clear(); |
| 141 | } |
| 142 | } |
| 143 | |
| 144 | uint32_t HInductionVarAnalysis::VisitDescendant(HLoopInformation* loop, HInstruction* instruction) { |
| 145 | // If the definition is either outside the loop (loop invariant entry value) |
| 146 | // or assigned in inner loop (inner exit value), the traversal stops. |
| 147 | HLoopInformation* otherLoop = instruction->GetBlock()->GetLoopInformation(); |
| 148 | if (otherLoop != loop) { |
| 149 | return global_depth_; |
| 150 | } |
| 151 | |
| 152 | // Inspect descendant node. |
| 153 | const int id = instruction->GetId(); |
| 154 | if (!IsVisitedNode(id)) { |
| 155 | VisitNode(loop, instruction); |
| 156 | return map_.find(id)->second.depth; |
| 157 | } else { |
| 158 | auto it = map_.find(id); |
| 159 | return it->second.done ? global_depth_ : it->second.depth; |
| 160 | } |
| 161 | } |
| 162 | |
| 163 | void HInductionVarAnalysis::ClassifyTrivial(HLoopInformation* loop, HInstruction* instruction) { |
| 164 | InductionInfo* info = nullptr; |
| 165 | if (instruction->IsPhi()) { |
| 166 | for (size_t i = 1, count = instruction->InputCount(); i < count; i++) { |
| 167 | info = TransferPhi(LookupInfo(loop, instruction->InputAt(0)), |
| 168 | LookupInfo(loop, instruction->InputAt(i))); |
| 169 | } |
| 170 | } else if (instruction->IsAdd()) { |
| 171 | info = TransferAddSub(LookupInfo(loop, instruction->InputAt(0)), |
| 172 | LookupInfo(loop, instruction->InputAt(1)), kAdd); |
| 173 | } else if (instruction->IsSub()) { |
| 174 | info = TransferAddSub(LookupInfo(loop, instruction->InputAt(0)), |
| 175 | LookupInfo(loop, instruction->InputAt(1)), kSub); |
| 176 | } else if (instruction->IsMul()) { |
| 177 | info = TransferMul(LookupInfo(loop, instruction->InputAt(0)), |
| 178 | LookupInfo(loop, instruction->InputAt(1))); |
| 179 | } else if (instruction->IsNeg()) { |
| 180 | info = TransferNeg(LookupInfo(loop, instruction->InputAt(0))); |
| 181 | } |
| 182 | |
| 183 | // Successfully classified? |
| 184 | if (info != nullptr) { |
| 185 | AssignInfo(loop, instruction, info); |
| 186 | } |
| 187 | } |
| 188 | |
| 189 | void HInductionVarAnalysis::ClassifyNonTrivial(HLoopInformation* loop) { |
| 190 | const size_t size = scc_.size(); |
| 191 | CHECK_GE(size, 1u); |
| 192 | HInstruction* phi = scc_[size - 1]; |
| 193 | if (!IsEntryPhi(loop, phi)) { |
| 194 | return; |
| 195 | } |
| 196 | HInstruction* external = phi->InputAt(0); |
| 197 | HInstruction* internal = phi->InputAt(1); |
| 198 | InductionInfo* initial = LookupInfo(loop, external); |
| 199 | if (initial == nullptr || initial->induction_class != kInvariant) { |
| 200 | return; |
| 201 | } |
| 202 | |
| 203 | // Singleton entry-phi-operation may be a wrap-around induction. |
| 204 | if (size == 1) { |
| 205 | InductionInfo* update = LookupInfo(loop, internal); |
| 206 | if (update != nullptr) { |
| 207 | AssignInfo(loop, phi, NewInductionInfo(kWrapAround, kNop, initial, update, nullptr)); |
| 208 | } |
| 209 | return; |
| 210 | } |
| 211 | |
| 212 | // Inspect remainder of the cycle that resides in scc_. The cycle_ mapping assigns |
| 213 | // temporary meaning to its nodes. |
| 214 | cycle_.Overwrite(phi->GetId(), nullptr); |
| 215 | for (size_t i = 0; i < size - 1; i++) { |
| 216 | HInstruction* operation = scc_[i]; |
| 217 | InductionInfo* update = nullptr; |
| 218 | if (operation->IsPhi()) { |
| 219 | update = TransferCycleOverPhi(operation); |
| 220 | } else if (operation->IsAdd()) { |
| 221 | update = TransferCycleOverAddSub(loop, operation->InputAt(0), operation->InputAt(1), kAdd, true); |
| 222 | } else if (operation->IsSub()) { |
| 223 | update = TransferCycleOverAddSub(loop, operation->InputAt(0), operation->InputAt(1), kSub, true); |
| 224 | } |
| 225 | if (update == nullptr) { |
| 226 | return; |
| 227 | } |
| 228 | cycle_.Overwrite(operation->GetId(), update); |
| 229 | } |
| 230 | |
| 231 | // Success if the internal link received accumulated nonzero update. |
| 232 | auto it = cycle_.find(internal->GetId()); |
| 233 | if (it != cycle_.end() && it->second != nullptr) { |
| 234 | // Classify header phi and feed the cycle "on-demand". |
| 235 | AssignInfo(loop, phi, NewInductionInfo(kLinear, kNop, it->second, initial, nullptr)); |
| 236 | for (size_t i = 0; i < size - 1; i++) { |
| 237 | ClassifyTrivial(loop, scc_[i]); |
| 238 | } |
| 239 | } |
| 240 | } |
| 241 | |
| 242 | HInductionVarAnalysis::InductionInfo* HInductionVarAnalysis::TransferPhi(InductionInfo* a, |
| 243 | InductionInfo* b) { |
| 244 | // Transfer over a phi: if both inputs are identical, result is input. |
| 245 | if (InductionEqual(a, b)) { |
| 246 | return a; |
| 247 | } |
| 248 | return nullptr; |
| 249 | } |
| 250 | |
| 251 | HInductionVarAnalysis::InductionInfo* HInductionVarAnalysis::TransferAddSub(InductionInfo* a, |
| 252 | InductionInfo* b, |
| 253 | InductionOp op) { |
| 254 | // Transfer over an addition or subtraction: invariant or linear |
| 255 | // inputs combine into new invariant or linear result. |
| 256 | if (a != nullptr && b != nullptr) { |
| 257 | if (a->induction_class == kInvariant && b->induction_class == kInvariant) { |
| 258 | return NewInductionInfo(kInvariant, op, a, b, nullptr); |
| 259 | } else if (a->induction_class == kLinear && b->induction_class == kInvariant) { |
| 260 | return NewInductionInfo( |
| 261 | kLinear, |
| 262 | kNop, |
| 263 | a->op_a, |
| 264 | NewInductionInfo(kInvariant, op, a->op_b, b, nullptr), |
| 265 | nullptr); |
| 266 | } else if (a->induction_class == kInvariant && b->induction_class == kLinear) { |
| 267 | InductionInfo* ba = b->op_a; |
| 268 | if (op == kSub) { // negation required |
| 269 | ba = NewInductionInfo(kInvariant, kNeg, nullptr, ba, nullptr); |
| 270 | } |
| 271 | return NewInductionInfo( |
| 272 | kLinear, |
| 273 | kNop, |
| 274 | ba, |
| 275 | NewInductionInfo(kInvariant, op, a, b->op_b, nullptr), |
| 276 | nullptr); |
| 277 | } else if (a->induction_class == kLinear && b->induction_class == kLinear) { |
| 278 | return NewInductionInfo( |
| 279 | kLinear, |
| 280 | kNop, |
| 281 | NewInductionInfo(kInvariant, op, a->op_a, b->op_a, nullptr), |
| 282 | NewInductionInfo(kInvariant, op, a->op_b, b->op_b, nullptr), |
| 283 | nullptr); |
| 284 | } |
| 285 | } |
| 286 | return nullptr; |
| 287 | } |
| 288 | |
| 289 | HInductionVarAnalysis::InductionInfo* HInductionVarAnalysis::TransferMul(InductionInfo* a, |
| 290 | InductionInfo* b) { |
| 291 | // Transfer over a multiplication: invariant or linear |
| 292 | // inputs combine into new invariant or linear result. |
| 293 | // Two linear inputs would become quadratic. |
| 294 | if (a != nullptr && b != nullptr) { |
| 295 | if (a->induction_class == kInvariant && b->induction_class == kInvariant) { |
| 296 | return NewInductionInfo(kInvariant, kMul, a, b, nullptr); |
| 297 | } else if (a->induction_class == kLinear && b->induction_class == kInvariant) { |
| 298 | return NewInductionInfo( |
| 299 | kLinear, |
| 300 | kNop, |
| 301 | NewInductionInfo(kInvariant, kMul, a->op_a, b, nullptr), |
| 302 | NewInductionInfo(kInvariant, kMul, a->op_b, b, nullptr), |
| 303 | nullptr); |
| 304 | } else if (a->induction_class == kInvariant && b->induction_class == kLinear) { |
| 305 | return NewInductionInfo( |
| 306 | kLinear, |
| 307 | kNop, |
| 308 | NewInductionInfo(kInvariant, kMul, a, b->op_a, nullptr), |
| 309 | NewInductionInfo(kInvariant, kMul, a, b->op_b, nullptr), |
| 310 | nullptr); |
| 311 | } |
| 312 | } |
| 313 | return nullptr; |
| 314 | } |
| 315 | |
| 316 | HInductionVarAnalysis::InductionInfo* HInductionVarAnalysis::TransferNeg(InductionInfo* a) { |
| 317 | // Transfer over a unary negation: invariant or linear input |
| 318 | // yields a similar, but negated result. |
| 319 | if (a != nullptr) { |
| 320 | if (a->induction_class == kInvariant) { |
| 321 | return NewInductionInfo(kInvariant, kNeg, nullptr, a, nullptr); |
| 322 | } else if (a->induction_class == kLinear) { |
| 323 | return NewInductionInfo( |
| 324 | kLinear, |
| 325 | kNop, |
| 326 | NewInductionInfo(kInvariant, kNeg, nullptr, a->op_a, nullptr), |
| 327 | NewInductionInfo(kInvariant, kNeg, nullptr, a->op_b, nullptr), |
| 328 | nullptr); |
| 329 | } |
| 330 | } |
| 331 | return nullptr; |
| 332 | } |
| 333 | |
| 334 | HInductionVarAnalysis::InductionInfo* HInductionVarAnalysis::TransferCycleOverPhi(HInstruction* phi) { |
| 335 | // Transfer within a cycle over a phi: only identical inputs |
| 336 | // can be combined into that input as result. |
| 337 | const size_t count = phi->InputCount(); |
| 338 | CHECK_GT(count, 0u); |
| 339 | auto ita = cycle_.find(phi->InputAt(0)->GetId()); |
| 340 | if (ita != cycle_.end()) { |
| 341 | InductionInfo* a = ita->second; |
| 342 | for (size_t i = 1; i < count; i++) { |
| 343 | auto itb = cycle_.find(phi->InputAt(i)->GetId()); |
| 344 | if (itb == cycle_.end() ||!HInductionVarAnalysis::InductionEqual(a, itb->second)) { |
| 345 | return nullptr; |
| 346 | } |
| 347 | } |
| 348 | return a; |
| 349 | } |
| 350 | return nullptr; |
| 351 | } |
| 352 | |
| 353 | HInductionVarAnalysis::InductionInfo* HInductionVarAnalysis::TransferCycleOverAddSub( |
| 354 | HLoopInformation* loop, |
| 355 | HInstruction* x, |
| 356 | HInstruction* y, |
| 357 | InductionOp op, |
| 358 | bool first) { |
| 359 | // Transfer within a cycle over an addition or subtraction: adding or |
| 360 | // subtracting an invariant value adds to the stride of the induction, |
| 361 | // starting with the phi value denoted by the unusual nullptr value. |
| 362 | auto it = cycle_.find(x->GetId()); |
| 363 | if (it != cycle_.end()) { |
| 364 | InductionInfo* a = it->second; |
| 365 | InductionInfo* b = LookupInfo(loop, y); |
| 366 | if (b != nullptr && b->induction_class == kInvariant) { |
| 367 | if (a == nullptr) { |
| 368 | if (op == kSub) { // negation required |
| 369 | return NewInductionInfo(kInvariant, kNeg, nullptr, b, nullptr); |
| 370 | } |
| 371 | return b; |
| 372 | } else if (a->induction_class == kInvariant) { |
| 373 | return NewInductionInfo(kInvariant, op, a, b, nullptr); |
| 374 | } |
| 375 | } |
| 376 | } |
| 377 | // On failure, try alternatives. |
| 378 | if (op == kAdd) { |
| 379 | // Try the other way around for an addition. |
| 380 | if (first) { |
| 381 | return TransferCycleOverAddSub(loop, y, x, op, false); |
| 382 | } |
| 383 | } |
| 384 | return nullptr; |
| 385 | } |
| 386 | |
| 387 | void HInductionVarAnalysis::PutInfo(int loop_id, int id, InductionInfo* info) { |
| 388 | auto it = induction_.find(loop_id); |
| 389 | if (it == induction_.end()) { |
| 390 | it = induction_.Put( |
| 391 | loop_id, ArenaSafeMap<int, InductionInfo*>(std::less<int>(), graph_->GetArena()->Adapter())); |
| 392 | } |
| 393 | it->second.Overwrite(id, info); |
| 394 | } |
| 395 | |
| 396 | HInductionVarAnalysis::InductionInfo* HInductionVarAnalysis::GetInfo(int loop_id, int id) { |
| 397 | auto it = induction_.find(loop_id); |
| 398 | if (it != induction_.end()) { |
| 399 | auto loop_it = it->second.find(id); |
| 400 | if (loop_it != it->second.end()) { |
| 401 | return loop_it->second; |
| 402 | } |
| 403 | } |
| 404 | return nullptr; |
| 405 | } |
| 406 | |
| 407 | void HInductionVarAnalysis::AssignInfo(HLoopInformation* loop, |
| 408 | HInstruction* instruction, |
| 409 | InductionInfo* info) { |
| 410 | const int loopId = loop->GetHeader()->GetBlockId(); |
| 411 | const int id = instruction->GetId(); |
| 412 | PutInfo(loopId, id, info); |
| 413 | } |
| 414 | |
| 415 | HInductionVarAnalysis::InductionInfo* |
| 416 | HInductionVarAnalysis::LookupInfo(HLoopInformation* loop, |
| 417 | HInstruction* instruction) { |
| 418 | const int loop_id = loop->GetHeader()->GetBlockId(); |
| 419 | const int id = instruction->GetId(); |
| 420 | InductionInfo* info = GetInfo(loop_id, id); |
| 421 | if (info == nullptr && IsLoopInvariant(loop, instruction)) { |
| 422 | info = NewInductionInfo(kInvariant, kFetch, nullptr, nullptr, instruction); |
| 423 | PutInfo(loop_id, id, info); |
| 424 | } |
| 425 | return info; |
| 426 | } |
| 427 | |
| 428 | bool HInductionVarAnalysis::InductionEqual(InductionInfo* info1, |
| 429 | InductionInfo* info2) { |
| 430 | // Test structural equality only, without accounting for simplifications. |
| 431 | if (info1 != nullptr && info2 != nullptr) { |
| 432 | return |
| 433 | info1->induction_class == info2->induction_class && |
| 434 | info1->operation == info2->operation && |
| 435 | info1->fetch == info2->fetch && |
| 436 | InductionEqual(info1->op_a, info2->op_a) && |
| 437 | InductionEqual(info1->op_b, info2->op_b); |
| 438 | } |
| 439 | // Otherwise only two nullptrs are considered equal. |
| 440 | return info1 == info2; |
| 441 | } |
| 442 | |
| 443 | std::string HInductionVarAnalysis::InductionToString(InductionInfo* info) { |
| 444 | if (info != nullptr) { |
| 445 | if (info->induction_class == kInvariant) { |
| 446 | std::string inv = "("; |
| 447 | inv += InductionToString(info->op_a); |
| 448 | switch (info->operation) { |
| 449 | case kNop: inv += " ? "; break; |
| 450 | case kAdd: inv += " + "; break; |
| 451 | case kSub: |
| 452 | case kNeg: inv += " - "; break; |
| 453 | case kMul: inv += " * "; break; |
| 454 | case kDiv: inv += " / "; break; |
| 455 | case kFetch: |
| 456 | CHECK(info->fetch != nullptr); |
| 457 | inv += std::to_string(info->fetch->GetId()) + ":" + info->fetch->DebugName(); |
| 458 | break; |
| 459 | } |
| 460 | inv += InductionToString(info->op_b); |
| 461 | return inv + ")"; |
| 462 | } else { |
| 463 | CHECK(info->operation == kNop); |
| 464 | if (info->induction_class == kLinear) { |
| 465 | return "(" + InductionToString(info->op_a) + " * i + " + |
| 466 | InductionToString(info->op_b) + ")"; |
| 467 | } else if (info->induction_class == kWrapAround) { |
| 468 | return "wrap(" + InductionToString(info->op_a) + ", " + |
| 469 | InductionToString(info->op_b) + ")"; |
| 470 | } else if (info->induction_class == kPeriodic) { |
| 471 | return "periodic(" + InductionToString(info->op_a) + ", " + |
| 472 | InductionToString(info->op_b) + ")"; |
| 473 | } |
| 474 | } |
| 475 | } |
| 476 | return ""; |
| 477 | } |
| 478 | |
| 479 | } // namespace art |