1 //===-- StructurizeCFG.cpp ------------------------------------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 10 #include "llvm/Transforms/Scalar.h" 11 #include "llvm/ADT/MapVector.h" 12 #include "llvm/ADT/PostOrderIterator.h" 13 #include "llvm/ADT/SCCIterator.h" 14 #include "llvm/Analysis/DivergenceAnalysis.h" 15 #include "llvm/Analysis/LoopInfo.h" 16 #include "llvm/Analysis/RegionInfo.h" 17 #include "llvm/Analysis/RegionIterator.h" 18 #include "llvm/Analysis/RegionPass.h" 19 #include "llvm/IR/Module.h" 20 #include "llvm/IR/PatternMatch.h" 21 #include "llvm/Support/Debug.h" 22 #include "llvm/Support/raw_ostream.h" 23 #include "llvm/Transforms/Utils/SSAUpdater.h" 24 25 using namespace llvm; 26 using namespace llvm::PatternMatch; 27 28 #define DEBUG_TYPE "structurizecfg" 29 30 namespace { 31 32 // Definition of the complex types used in this pass. 33 34 typedef std::pair<BasicBlock *, Value *> BBValuePair; 35 36 typedef SmallVector<RegionNode*, 8> RNVector; 37 typedef SmallVector<BasicBlock*, 8> BBVector; 38 typedef SmallVector<BranchInst*, 8> BranchVector; 39 typedef SmallVector<BBValuePair, 2> BBValueVector; 40 41 typedef SmallPtrSet<BasicBlock *, 8> BBSet; 42 43 typedef MapVector<PHINode *, BBValueVector> PhiMap; 44 typedef MapVector<BasicBlock *, BBVector> BB2BBVecMap; 45 46 typedef DenseMap<DomTreeNode *, unsigned> DTN2UnsignedMap; 47 typedef DenseMap<BasicBlock *, PhiMap> BBPhiMap; 48 typedef DenseMap<BasicBlock *, Value *> BBPredicates; 49 typedef DenseMap<BasicBlock *, BBPredicates> PredMap; 50 typedef DenseMap<BasicBlock *, BasicBlock*> BB2BBMap; 51 52 // The name for newly created blocks. 53 54 static const char *const FlowBlockName = "Flow"; 55 56 /// @brief Find the nearest common dominator for multiple BasicBlocks 57 /// 58 /// Helper class for StructurizeCFG 59 /// TODO: Maybe move into common code 60 class NearestCommonDominator { 61 DominatorTree *DT; 62 63 DTN2UnsignedMap IndexMap; 64 65 BasicBlock *Result; 66 unsigned ResultIndex; 67 bool ExplicitMentioned; 68 69 public: 70 /// \brief Start a new query 71 NearestCommonDominator(DominatorTree *DomTree) { 72 DT = DomTree; 73 Result = nullptr; 74 } 75 76 /// \brief Add BB to the resulting dominator 77 void addBlock(BasicBlock *BB, bool Remember = true) { 78 DomTreeNode *Node = DT->getNode(BB); 79 80 if (!Result) { 81 unsigned Numbering = 0; 82 for (;Node;Node = Node->getIDom()) 83 IndexMap[Node] = ++Numbering; 84 Result = BB; 85 ResultIndex = 1; 86 ExplicitMentioned = Remember; 87 return; 88 } 89 90 for (;Node;Node = Node->getIDom()) 91 if (IndexMap.count(Node)) 92 break; 93 else 94 IndexMap[Node] = 0; 95 96 assert(Node && "Dominator tree invalid!"); 97 98 unsigned Numbering = IndexMap[Node]; 99 if (Numbering > ResultIndex) { 100 Result = Node->getBlock(); 101 ResultIndex = Numbering; 102 ExplicitMentioned = Remember && (Result == BB); 103 } else if (Numbering == ResultIndex) { 104 ExplicitMentioned |= Remember; 105 } 106 } 107 108 /// \brief Is "Result" one of the BBs added with "Remember" = True? 109 bool wasResultExplicitMentioned() { 110 return ExplicitMentioned; 111 } 112 113 /// \brief Get the query result 114 BasicBlock *getResult() { 115 return Result; 116 } 117 }; 118 119 /// @brief Transforms the control flow graph on one single entry/exit region 120 /// at a time. 121 /// 122 /// After the transform all "If"/"Then"/"Else" style control flow looks like 123 /// this: 124 /// 125 /// \verbatim 126 /// 1 127 /// || 128 /// | | 129 /// 2 | 130 /// | / 131 /// |/ 132 /// 3 133 /// || Where: 134 /// | | 1 = "If" block, calculates the condition 135 /// 4 | 2 = "Then" subregion, runs if the condition is true 136 /// | / 3 = "Flow" blocks, newly inserted flow blocks, rejoins the flow 137 /// |/ 4 = "Else" optional subregion, runs if the condition is false 138 /// 5 5 = "End" block, also rejoins the control flow 139 /// \endverbatim 140 /// 141 /// Control flow is expressed as a branch where the true exit goes into the 142 /// "Then"/"Else" region, while the false exit skips the region 143 /// The condition for the optional "Else" region is expressed as a PHI node. 144 /// The incoming values of the PHI node are true for the "If" edge and false 145 /// for the "Then" edge. 146 /// 147 /// Additionally to that even complicated loops look like this: 148 /// 149 /// \verbatim 150 /// 1 151 /// || 152 /// | | 153 /// 2 ^ Where: 154 /// | / 1 = "Entry" block 155 /// |/ 2 = "Loop" optional subregion, with all exits at "Flow" block 156 /// 3 3 = "Flow" block, with back edge to entry block 157 /// | 158 /// \endverbatim 159 /// 160 /// The back edge of the "Flow" block is always on the false side of the branch 161 /// while the true side continues the general flow. So the loop condition 162 /// consist of a network of PHI nodes where the true incoming values expresses 163 /// breaks and the false values expresses continue states. 164 class StructurizeCFG : public RegionPass { 165 bool SkipUniformRegions; 166 167 Type *Boolean; 168 ConstantInt *BoolTrue; 169 ConstantInt *BoolFalse; 170 UndefValue *BoolUndef; 171 172 Function *Func; 173 Region *ParentRegion; 174 175 DominatorTree *DT; 176 LoopInfo *LI; 177 178 RNVector Order; 179 BBSet Visited; 180 181 BBPhiMap DeletedPhis; 182 BB2BBVecMap AddedPhis; 183 184 PredMap Predicates; 185 BranchVector Conditions; 186 187 BB2BBMap Loops; 188 PredMap LoopPreds; 189 BranchVector LoopConds; 190 191 RegionNode *PrevNode; 192 193 void orderNodes(); 194 195 void analyzeLoops(RegionNode *N); 196 197 Value *invert(Value *Condition); 198 199 Value *buildCondition(BranchInst *Term, unsigned Idx, bool Invert); 200 201 void gatherPredicates(RegionNode *N); 202 203 void collectInfos(); 204 205 void insertConditions(bool Loops); 206 207 void delPhiValues(BasicBlock *From, BasicBlock *To); 208 209 void addPhiValues(BasicBlock *From, BasicBlock *To); 210 211 void setPhiValues(); 212 213 void killTerminator(BasicBlock *BB); 214 215 void changeExit(RegionNode *Node, BasicBlock *NewExit, 216 bool IncludeDominator); 217 218 BasicBlock *getNextFlow(BasicBlock *Dominator); 219 220 BasicBlock *needPrefix(bool NeedEmpty); 221 222 BasicBlock *needPostfix(BasicBlock *Flow, bool ExitUseAllowed); 223 224 void setPrevNode(BasicBlock *BB); 225 226 bool dominatesPredicates(BasicBlock *BB, RegionNode *Node); 227 228 bool isPredictableTrue(RegionNode *Node); 229 230 void wireFlow(bool ExitUseAllowed, BasicBlock *LoopEnd); 231 232 void handleLoops(bool ExitUseAllowed, BasicBlock *LoopEnd); 233 234 void createFlow(); 235 236 void rebuildSSA(); 237 238 public: 239 static char ID; 240 241 StructurizeCFG() : 242 RegionPass(ID), SkipUniformRegions(false) { 243 initializeStructurizeCFGPass(*PassRegistry::getPassRegistry()); 244 } 245 246 StructurizeCFG(bool SkipUniformRegions) : 247 RegionPass(ID), SkipUniformRegions(SkipUniformRegions) { 248 initializeStructurizeCFGPass(*PassRegistry::getPassRegistry()); 249 } 250 251 using Pass::doInitialization; 252 bool doInitialization(Region *R, RGPassManager &RGM) override; 253 254 bool runOnRegion(Region *R, RGPassManager &RGM) override; 255 256 StringRef getPassName() const override { return "Structurize control flow"; } 257 258 void getAnalysisUsage(AnalysisUsage &AU) const override { 259 if (SkipUniformRegions) 260 AU.addRequired<DivergenceAnalysis>(); 261 AU.addRequiredID(LowerSwitchID); 262 AU.addRequired<DominatorTreeWrapperPass>(); 263 AU.addRequired<LoopInfoWrapperPass>(); 264 AU.addPreserved<DominatorTreeWrapperPass>(); 265 RegionPass::getAnalysisUsage(AU); 266 } 267 }; 268 269 } // end anonymous namespace 270 271 char StructurizeCFG::ID = 0; 272 273 INITIALIZE_PASS_BEGIN(StructurizeCFG, "structurizecfg", "Structurize the CFG", 274 false, false) 275 INITIALIZE_PASS_DEPENDENCY(DivergenceAnalysis) 276 INITIALIZE_PASS_DEPENDENCY(LowerSwitch) 277 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 278 INITIALIZE_PASS_DEPENDENCY(RegionInfoPass) 279 INITIALIZE_PASS_END(StructurizeCFG, "structurizecfg", "Structurize the CFG", 280 false, false) 281 282 /// \brief Initialize the types and constants used in the pass 283 bool StructurizeCFG::doInitialization(Region *R, RGPassManager &RGM) { 284 LLVMContext &Context = R->getEntry()->getContext(); 285 286 Boolean = Type::getInt1Ty(Context); 287 BoolTrue = ConstantInt::getTrue(Context); 288 BoolFalse = ConstantInt::getFalse(Context); 289 BoolUndef = UndefValue::get(Boolean); 290 291 return false; 292 } 293 294 /// \brief Build up the general order of nodes 295 void StructurizeCFG::orderNodes() { 296 RNVector TempOrder; 297 ReversePostOrderTraversal<Region*> RPOT(ParentRegion); 298 TempOrder.append(RPOT.begin(), RPOT.end()); 299 300 std::map<Loop*, unsigned> LoopBlocks; 301 302 303 // The reverse post-order traversal of the list gives us an ordering close 304 // to what we want. The only problem with it is that sometimes backedges 305 // for outer loops will be visited before backedges for inner loops. 306 for (RegionNode *RN : TempOrder) { 307 BasicBlock *BB = RN->getEntry(); 308 Loop *Loop = LI->getLoopFor(BB); 309 ++LoopBlocks[Loop]; 310 } 311 312 unsigned CurrentLoopDepth = 0; 313 Loop *CurrentLoop = nullptr; 314 BBSet TempVisited; 315 for (RNVector::iterator I = TempOrder.begin(), E = TempOrder.end(); I != E; ++I) { 316 BasicBlock *BB = (*I)->getEntry(); 317 unsigned LoopDepth = LI->getLoopDepth(BB); 318 319 if (is_contained(Order, *I)) 320 continue; 321 322 if (LoopDepth < CurrentLoopDepth) { 323 // Make sure we have visited all blocks in this loop before moving back to 324 // the outer loop. 325 326 RNVector::iterator LoopI = I; 327 while (unsigned &BlockCount = LoopBlocks[CurrentLoop]) { 328 LoopI++; 329 BasicBlock *LoopBB = (*LoopI)->getEntry(); 330 if (LI->getLoopFor(LoopBB) == CurrentLoop) { 331 --BlockCount; 332 Order.push_back(*LoopI); 333 } 334 } 335 } 336 337 CurrentLoop = LI->getLoopFor(BB); 338 if (CurrentLoop) { 339 LoopBlocks[CurrentLoop]--; 340 } 341 342 CurrentLoopDepth = LoopDepth; 343 Order.push_back(*I); 344 } 345 346 // This pass originally used a post-order traversal and then operated on 347 // the list in reverse. Now that we are using a reverse post-order traversal 348 // rather than re-working the whole pass to operate on the list in order, 349 // we just reverse the list and continue to operate on it in reverse. 350 std::reverse(Order.begin(), Order.end()); 351 } 352 353 /// \brief Determine the end of the loops 354 void StructurizeCFG::analyzeLoops(RegionNode *N) { 355 if (N->isSubRegion()) { 356 // Test for exit as back edge 357 BasicBlock *Exit = N->getNodeAs<Region>()->getExit(); 358 if (Visited.count(Exit)) 359 Loops[Exit] = N->getEntry(); 360 361 } else { 362 // Test for sucessors as back edge 363 BasicBlock *BB = N->getNodeAs<BasicBlock>(); 364 BranchInst *Term = cast<BranchInst>(BB->getTerminator()); 365 366 for (BasicBlock *Succ : Term->successors()) 367 if (Visited.count(Succ)) 368 Loops[Succ] = BB; 369 } 370 } 371 372 /// \brief Invert the given condition 373 Value *StructurizeCFG::invert(Value *Condition) { 374 // First: Check if it's a constant 375 if (Constant *C = dyn_cast<Constant>(Condition)) 376 return ConstantExpr::getNot(C); 377 378 // Second: If the condition is already inverted, return the original value 379 if (match(Condition, m_Not(m_Value(Condition)))) 380 return Condition; 381 382 if (Instruction *Inst = dyn_cast<Instruction>(Condition)) { 383 // Third: Check all the users for an invert 384 BasicBlock *Parent = Inst->getParent(); 385 for (User *U : Condition->users()) 386 if (Instruction *I = dyn_cast<Instruction>(U)) 387 if (I->getParent() == Parent && match(I, m_Not(m_Specific(Condition)))) 388 return I; 389 390 // Last option: Create a new instruction 391 return BinaryOperator::CreateNot(Condition, "", Parent->getTerminator()); 392 } 393 394 if (Argument *Arg = dyn_cast<Argument>(Condition)) { 395 BasicBlock &EntryBlock = Arg->getParent()->getEntryBlock(); 396 return BinaryOperator::CreateNot(Condition, 397 Arg->getName() + ".inv", 398 EntryBlock.getTerminator()); 399 } 400 401 llvm_unreachable("Unhandled condition to invert"); 402 } 403 404 /// \brief Build the condition for one edge 405 Value *StructurizeCFG::buildCondition(BranchInst *Term, unsigned Idx, 406 bool Invert) { 407 Value *Cond = Invert ? BoolFalse : BoolTrue; 408 if (Term->isConditional()) { 409 Cond = Term->getCondition(); 410 411 if (Idx != (unsigned)Invert) 412 Cond = invert(Cond); 413 } 414 return Cond; 415 } 416 417 /// \brief Analyze the predecessors of each block and build up predicates 418 void StructurizeCFG::gatherPredicates(RegionNode *N) { 419 RegionInfo *RI = ParentRegion->getRegionInfo(); 420 BasicBlock *BB = N->getEntry(); 421 BBPredicates &Pred = Predicates[BB]; 422 BBPredicates &LPred = LoopPreds[BB]; 423 424 for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 425 PI != PE; ++PI) { 426 427 // Ignore it if it's a branch from outside into our region entry 428 if (!ParentRegion->contains(*PI)) 429 continue; 430 431 Region *R = RI->getRegionFor(*PI); 432 if (R == ParentRegion) { 433 434 // It's a top level block in our region 435 BranchInst *Term = cast<BranchInst>((*PI)->getTerminator()); 436 for (unsigned i = 0, e = Term->getNumSuccessors(); i != e; ++i) { 437 BasicBlock *Succ = Term->getSuccessor(i); 438 if (Succ != BB) 439 continue; 440 441 if (Visited.count(*PI)) { 442 // Normal forward edge 443 if (Term->isConditional()) { 444 // Try to treat it like an ELSE block 445 BasicBlock *Other = Term->getSuccessor(!i); 446 if (Visited.count(Other) && !Loops.count(Other) && 447 !Pred.count(Other) && !Pred.count(*PI)) { 448 449 Pred[Other] = BoolFalse; 450 Pred[*PI] = BoolTrue; 451 continue; 452 } 453 } 454 Pred[*PI] = buildCondition(Term, i, false); 455 456 } else { 457 // Back edge 458 LPred[*PI] = buildCondition(Term, i, true); 459 } 460 } 461 462 } else { 463 464 // It's an exit from a sub region 465 while (R->getParent() != ParentRegion) 466 R = R->getParent(); 467 468 // Edge from inside a subregion to its entry, ignore it 469 if (*R == *N) 470 continue; 471 472 BasicBlock *Entry = R->getEntry(); 473 if (Visited.count(Entry)) 474 Pred[Entry] = BoolTrue; 475 else 476 LPred[Entry] = BoolFalse; 477 } 478 } 479 } 480 481 /// \brief Collect various loop and predicate infos 482 void StructurizeCFG::collectInfos() { 483 // Reset predicate 484 Predicates.clear(); 485 486 // and loop infos 487 Loops.clear(); 488 LoopPreds.clear(); 489 490 // Reset the visited nodes 491 Visited.clear(); 492 493 for (RegionNode *RN : reverse(Order)) { 494 495 DEBUG(dbgs() << "Visiting: " 496 << (RN->isSubRegion() ? "SubRegion with entry: " : "") 497 << RN->getEntry()->getName() << " Loop Depth: " 498 << LI->getLoopDepth(RN->getEntry()) << "\n"); 499 500 // Analyze all the conditions leading to a node 501 gatherPredicates(RN); 502 503 // Remember that we've seen this node 504 Visited.insert(RN->getEntry()); 505 506 // Find the last back edges 507 analyzeLoops(RN); 508 } 509 } 510 511 /// \brief Insert the missing branch conditions 512 void StructurizeCFG::insertConditions(bool Loops) { 513 BranchVector &Conds = Loops ? LoopConds : Conditions; 514 Value *Default = Loops ? BoolTrue : BoolFalse; 515 SSAUpdater PhiInserter; 516 517 for (BranchInst *Term : Conds) { 518 assert(Term->isConditional()); 519 520 BasicBlock *Parent = Term->getParent(); 521 BasicBlock *SuccTrue = Term->getSuccessor(0); 522 BasicBlock *SuccFalse = Term->getSuccessor(1); 523 524 PhiInserter.Initialize(Boolean, ""); 525 PhiInserter.AddAvailableValue(&Func->getEntryBlock(), Default); 526 PhiInserter.AddAvailableValue(Loops ? SuccFalse : Parent, Default); 527 528 BBPredicates &Preds = Loops ? LoopPreds[SuccFalse] : Predicates[SuccTrue]; 529 530 NearestCommonDominator Dominator(DT); 531 Dominator.addBlock(Parent, false); 532 533 Value *ParentValue = nullptr; 534 for (BBPredicates::iterator PI = Preds.begin(), PE = Preds.end(); 535 PI != PE; ++PI) { 536 537 if (PI->first == Parent) { 538 ParentValue = PI->second; 539 break; 540 } 541 PhiInserter.AddAvailableValue(PI->first, PI->second); 542 Dominator.addBlock(PI->first); 543 } 544 545 if (ParentValue) { 546 Term->setCondition(ParentValue); 547 } else { 548 if (!Dominator.wasResultExplicitMentioned()) 549 PhiInserter.AddAvailableValue(Dominator.getResult(), Default); 550 551 Term->setCondition(PhiInserter.GetValueInMiddleOfBlock(Parent)); 552 } 553 } 554 } 555 556 /// \brief Remove all PHI values coming from "From" into "To" and remember 557 /// them in DeletedPhis 558 void StructurizeCFG::delPhiValues(BasicBlock *From, BasicBlock *To) { 559 PhiMap &Map = DeletedPhis[To]; 560 for (BasicBlock::iterator I = To->begin(), E = To->end(); 561 I != E && isa<PHINode>(*I);) { 562 563 PHINode &Phi = cast<PHINode>(*I++); 564 while (Phi.getBasicBlockIndex(From) != -1) { 565 Value *Deleted = Phi.removeIncomingValue(From, false); 566 Map[&Phi].push_back(std::make_pair(From, Deleted)); 567 } 568 } 569 } 570 571 /// \brief Add a dummy PHI value as soon as we knew the new predecessor 572 void StructurizeCFG::addPhiValues(BasicBlock *From, BasicBlock *To) { 573 for (BasicBlock::iterator I = To->begin(), E = To->end(); 574 I != E && isa<PHINode>(*I);) { 575 576 PHINode &Phi = cast<PHINode>(*I++); 577 Value *Undef = UndefValue::get(Phi.getType()); 578 Phi.addIncoming(Undef, From); 579 } 580 AddedPhis[To].push_back(From); 581 } 582 583 /// \brief Add the real PHI value as soon as everything is set up 584 void StructurizeCFG::setPhiValues() { 585 SSAUpdater Updater; 586 for (const auto &AddedPhi : AddedPhis) { 587 588 BasicBlock *To = AddedPhi.first; 589 const BBVector &From = AddedPhi.second; 590 591 if (!DeletedPhis.count(To)) 592 continue; 593 594 PhiMap &Map = DeletedPhis[To]; 595 for (const auto &PI : Map) { 596 597 PHINode *Phi = PI.first; 598 Value *Undef = UndefValue::get(Phi->getType()); 599 Updater.Initialize(Phi->getType(), ""); 600 Updater.AddAvailableValue(&Func->getEntryBlock(), Undef); 601 Updater.AddAvailableValue(To, Undef); 602 603 NearestCommonDominator Dominator(DT); 604 Dominator.addBlock(To, false); 605 for (const auto &VI : PI.second) { 606 607 Updater.AddAvailableValue(VI.first, VI.second); 608 Dominator.addBlock(VI.first); 609 } 610 611 if (!Dominator.wasResultExplicitMentioned()) 612 Updater.AddAvailableValue(Dominator.getResult(), Undef); 613 614 for (BasicBlock *FI : From) { 615 616 int Idx = Phi->getBasicBlockIndex(FI); 617 assert(Idx != -1); 618 Phi->setIncomingValue(Idx, Updater.GetValueAtEndOfBlock(FI)); 619 } 620 } 621 622 DeletedPhis.erase(To); 623 } 624 assert(DeletedPhis.empty()); 625 } 626 627 /// \brief Remove phi values from all successors and then remove the terminator. 628 void StructurizeCFG::killTerminator(BasicBlock *BB) { 629 TerminatorInst *Term = BB->getTerminator(); 630 if (!Term) 631 return; 632 633 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); 634 SI != SE; ++SI) { 635 636 delPhiValues(BB, *SI); 637 } 638 639 Term->eraseFromParent(); 640 } 641 642 /// \brief Let node exit(s) point to NewExit 643 void StructurizeCFG::changeExit(RegionNode *Node, BasicBlock *NewExit, 644 bool IncludeDominator) { 645 if (Node->isSubRegion()) { 646 Region *SubRegion = Node->getNodeAs<Region>(); 647 BasicBlock *OldExit = SubRegion->getExit(); 648 BasicBlock *Dominator = nullptr; 649 650 // Find all the edges from the sub region to the exit 651 for (pred_iterator I = pred_begin(OldExit), E = pred_end(OldExit); 652 I != E;) { 653 654 BasicBlock *BB = *I++; 655 if (!SubRegion->contains(BB)) 656 continue; 657 658 // Modify the edges to point to the new exit 659 delPhiValues(BB, OldExit); 660 BB->getTerminator()->replaceUsesOfWith(OldExit, NewExit); 661 addPhiValues(BB, NewExit); 662 663 // Find the new dominator (if requested) 664 if (IncludeDominator) { 665 if (!Dominator) 666 Dominator = BB; 667 else 668 Dominator = DT->findNearestCommonDominator(Dominator, BB); 669 } 670 } 671 672 // Change the dominator (if requested) 673 if (Dominator) 674 DT->changeImmediateDominator(NewExit, Dominator); 675 676 // Update the region info 677 SubRegion->replaceExit(NewExit); 678 679 } else { 680 BasicBlock *BB = Node->getNodeAs<BasicBlock>(); 681 killTerminator(BB); 682 BranchInst::Create(NewExit, BB); 683 addPhiValues(BB, NewExit); 684 if (IncludeDominator) 685 DT->changeImmediateDominator(NewExit, BB); 686 } 687 } 688 689 /// \brief Create a new flow node and update dominator tree and region info 690 BasicBlock *StructurizeCFG::getNextFlow(BasicBlock *Dominator) { 691 LLVMContext &Context = Func->getContext(); 692 BasicBlock *Insert = Order.empty() ? ParentRegion->getExit() : 693 Order.back()->getEntry(); 694 BasicBlock *Flow = BasicBlock::Create(Context, FlowBlockName, 695 Func, Insert); 696 DT->addNewBlock(Flow, Dominator); 697 ParentRegion->getRegionInfo()->setRegionFor(Flow, ParentRegion); 698 return Flow; 699 } 700 701 /// \brief Create a new or reuse the previous node as flow node 702 BasicBlock *StructurizeCFG::needPrefix(bool NeedEmpty) { 703 BasicBlock *Entry = PrevNode->getEntry(); 704 705 if (!PrevNode->isSubRegion()) { 706 killTerminator(Entry); 707 if (!NeedEmpty || Entry->getFirstInsertionPt() == Entry->end()) 708 return Entry; 709 710 } 711 712 // create a new flow node 713 BasicBlock *Flow = getNextFlow(Entry); 714 715 // and wire it up 716 changeExit(PrevNode, Flow, true); 717 PrevNode = ParentRegion->getBBNode(Flow); 718 return Flow; 719 } 720 721 /// \brief Returns the region exit if possible, otherwise just a new flow node 722 BasicBlock *StructurizeCFG::needPostfix(BasicBlock *Flow, 723 bool ExitUseAllowed) { 724 if (Order.empty() && ExitUseAllowed) { 725 BasicBlock *Exit = ParentRegion->getExit(); 726 DT->changeImmediateDominator(Exit, Flow); 727 addPhiValues(Flow, Exit); 728 return Exit; 729 } 730 return getNextFlow(Flow); 731 } 732 733 /// \brief Set the previous node 734 void StructurizeCFG::setPrevNode(BasicBlock *BB) { 735 PrevNode = ParentRegion->contains(BB) ? ParentRegion->getBBNode(BB) 736 : nullptr; 737 } 738 739 /// \brief Does BB dominate all the predicates of Node ? 740 bool StructurizeCFG::dominatesPredicates(BasicBlock *BB, RegionNode *Node) { 741 BBPredicates &Preds = Predicates[Node->getEntry()]; 742 for (BBPredicates::iterator PI = Preds.begin(), PE = Preds.end(); 743 PI != PE; ++PI) { 744 745 if (!DT->dominates(BB, PI->first)) 746 return false; 747 } 748 return true; 749 } 750 751 /// \brief Can we predict that this node will always be called? 752 bool StructurizeCFG::isPredictableTrue(RegionNode *Node) { 753 BBPredicates &Preds = Predicates[Node->getEntry()]; 754 bool Dominated = false; 755 756 // Regionentry is always true 757 if (!PrevNode) 758 return true; 759 760 for (BBPredicates::iterator I = Preds.begin(), E = Preds.end(); 761 I != E; ++I) { 762 763 if (I->second != BoolTrue) 764 return false; 765 766 if (!Dominated && DT->dominates(I->first, PrevNode->getEntry())) 767 Dominated = true; 768 } 769 770 // TODO: The dominator check is too strict 771 return Dominated; 772 } 773 774 /// Take one node from the order vector and wire it up 775 void StructurizeCFG::wireFlow(bool ExitUseAllowed, 776 BasicBlock *LoopEnd) { 777 RegionNode *Node = Order.pop_back_val(); 778 Visited.insert(Node->getEntry()); 779 780 if (isPredictableTrue(Node)) { 781 // Just a linear flow 782 if (PrevNode) { 783 changeExit(PrevNode, Node->getEntry(), true); 784 } 785 PrevNode = Node; 786 787 } else { 788 // Insert extra prefix node (or reuse last one) 789 BasicBlock *Flow = needPrefix(false); 790 791 // Insert extra postfix node (or use exit instead) 792 BasicBlock *Entry = Node->getEntry(); 793 BasicBlock *Next = needPostfix(Flow, ExitUseAllowed); 794 795 // let it point to entry and next block 796 Conditions.push_back(BranchInst::Create(Entry, Next, BoolUndef, Flow)); 797 addPhiValues(Flow, Entry); 798 DT->changeImmediateDominator(Entry, Flow); 799 800 PrevNode = Node; 801 while (!Order.empty() && !Visited.count(LoopEnd) && 802 dominatesPredicates(Entry, Order.back())) { 803 handleLoops(false, LoopEnd); 804 } 805 806 changeExit(PrevNode, Next, false); 807 setPrevNode(Next); 808 } 809 } 810 811 void StructurizeCFG::handleLoops(bool ExitUseAllowed, 812 BasicBlock *LoopEnd) { 813 RegionNode *Node = Order.back(); 814 BasicBlock *LoopStart = Node->getEntry(); 815 816 if (!Loops.count(LoopStart)) { 817 wireFlow(ExitUseAllowed, LoopEnd); 818 return; 819 } 820 821 if (!isPredictableTrue(Node)) 822 LoopStart = needPrefix(true); 823 824 LoopEnd = Loops[Node->getEntry()]; 825 wireFlow(false, LoopEnd); 826 while (!Visited.count(LoopEnd)) { 827 handleLoops(false, LoopEnd); 828 } 829 830 // If the start of the loop is the entry block, we can't branch to it so 831 // insert a new dummy entry block. 832 Function *LoopFunc = LoopStart->getParent(); 833 if (LoopStart == &LoopFunc->getEntryBlock()) { 834 LoopStart->setName("entry.orig"); 835 836 BasicBlock *NewEntry = 837 BasicBlock::Create(LoopStart->getContext(), 838 "entry", 839 LoopFunc, 840 LoopStart); 841 BranchInst::Create(LoopStart, NewEntry); 842 } 843 844 // Create an extra loop end node 845 LoopEnd = needPrefix(false); 846 BasicBlock *Next = needPostfix(LoopEnd, ExitUseAllowed); 847 LoopConds.push_back(BranchInst::Create(Next, LoopStart, 848 BoolUndef, LoopEnd)); 849 addPhiValues(LoopEnd, LoopStart); 850 setPrevNode(Next); 851 } 852 853 /// After this function control flow looks like it should be, but 854 /// branches and PHI nodes only have undefined conditions. 855 void StructurizeCFG::createFlow() { 856 BasicBlock *Exit = ParentRegion->getExit(); 857 bool EntryDominatesExit = DT->dominates(ParentRegion->getEntry(), Exit); 858 859 DeletedPhis.clear(); 860 AddedPhis.clear(); 861 Conditions.clear(); 862 LoopConds.clear(); 863 864 PrevNode = nullptr; 865 Visited.clear(); 866 867 while (!Order.empty()) { 868 handleLoops(EntryDominatesExit, nullptr); 869 } 870 871 if (PrevNode) 872 changeExit(PrevNode, Exit, EntryDominatesExit); 873 else 874 assert(EntryDominatesExit); 875 } 876 877 /// Handle a rare case where the disintegrated nodes instructions 878 /// no longer dominate all their uses. Not sure if this is really nessasary 879 void StructurizeCFG::rebuildSSA() { 880 SSAUpdater Updater; 881 for (auto *BB : ParentRegion->blocks()) 882 for (BasicBlock::iterator II = BB->begin(), IE = BB->end(); 883 II != IE; ++II) { 884 885 bool Initialized = false; 886 for (auto I = II->use_begin(), E = II->use_end(); I != E;) { 887 Use &U = *I++; 888 Instruction *User = cast<Instruction>(U.getUser()); 889 if (User->getParent() == BB) { 890 continue; 891 892 } else if (PHINode *UserPN = dyn_cast<PHINode>(User)) { 893 if (UserPN->getIncomingBlock(U) == BB) 894 continue; 895 } 896 897 if (DT->dominates(&*II, User)) 898 continue; 899 900 if (!Initialized) { 901 Value *Undef = UndefValue::get(II->getType()); 902 Updater.Initialize(II->getType(), ""); 903 Updater.AddAvailableValue(&Func->getEntryBlock(), Undef); 904 Updater.AddAvailableValue(BB, &*II); 905 Initialized = true; 906 } 907 Updater.RewriteUseAfterInsertions(U); 908 } 909 } 910 } 911 912 static bool hasOnlyUniformBranches(const Region *R, 913 const DivergenceAnalysis &DA) { 914 for (const BasicBlock *BB : R->blocks()) { 915 const BranchInst *Br = dyn_cast<BranchInst>(BB->getTerminator()); 916 if (!Br || !Br->isConditional()) 917 continue; 918 919 if (!DA.isUniform(Br->getCondition())) 920 return false; 921 DEBUG(dbgs() << "BB: " << BB->getName() << " has uniform terminator\n"); 922 } 923 return true; 924 } 925 926 /// \brief Run the transformation for each region found 927 bool StructurizeCFG::runOnRegion(Region *R, RGPassManager &RGM) { 928 if (R->isTopLevelRegion()) 929 return false; 930 931 if (SkipUniformRegions) { 932 // TODO: We could probably be smarter here with how we handle sub-regions. 933 auto &DA = getAnalysis<DivergenceAnalysis>(); 934 if (hasOnlyUniformBranches(R, DA)) { 935 DEBUG(dbgs() << "Skipping region with uniform control flow: " << *R << '\n'); 936 937 // Mark all direct child block terminators as having been treated as 938 // uniform. To account for a possible future in which non-uniform 939 // sub-regions are treated more cleverly, indirect children are not 940 // marked as uniform. 941 MDNode *MD = MDNode::get(R->getEntry()->getParent()->getContext(), {}); 942 Region::element_iterator E = R->element_end(); 943 for (Region::element_iterator I = R->element_begin(); I != E; ++I) { 944 if (I->isSubRegion()) 945 continue; 946 947 if (Instruction *Term = I->getEntry()->getTerminator()) 948 Term->setMetadata("structurizecfg.uniform", MD); 949 } 950 951 return false; 952 } 953 } 954 955 Func = R->getEntry()->getParent(); 956 ParentRegion = R; 957 958 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 959 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 960 961 orderNodes(); 962 collectInfos(); 963 createFlow(); 964 insertConditions(false); 965 insertConditions(true); 966 setPhiValues(); 967 rebuildSSA(); 968 969 // Cleanup 970 Order.clear(); 971 Visited.clear(); 972 DeletedPhis.clear(); 973 AddedPhis.clear(); 974 Predicates.clear(); 975 Conditions.clear(); 976 Loops.clear(); 977 LoopPreds.clear(); 978 LoopConds.clear(); 979 980 return true; 981 } 982 983 Pass *llvm::createStructurizeCFGPass(bool SkipUniformRegions) { 984 return new StructurizeCFG(SkipUniformRegions); 985 } 986