xref: /llvm-project/llvm/lib/Transforms/Scalar/StructurizeCFG.cpp (revision 62c20d8b3b6f7fd006fc0c6eb1f14c2681a2bb31)
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<BasicBlock *, PhiMap> BBPhiMap;
47 typedef DenseMap<BasicBlock *, Value *> BBPredicates;
48 typedef DenseMap<BasicBlock *, BBPredicates> PredMap;
49 typedef DenseMap<BasicBlock *, BasicBlock*> BB2BBMap;
50 
51 // The name for newly created blocks.
52 static const char *const FlowBlockName = "Flow";
53 
54 /// Finds the nearest common dominator of a set of BasicBlocks.
55 ///
56 /// For every BB you add to the set, you can specify whether we "remember" the
57 /// block.  When you get the common dominator, you can also ask whether it's one
58 /// of the blocks we remembered.
59 class NearestCommonDominator {
60   DominatorTree *DT;
61   BasicBlock *Result = nullptr;
62   bool ResultIsRemembered = false;
63 
64   /// Add BB to the resulting dominator.
65   void addBlock(BasicBlock *BB, bool Remember) {
66     if (!Result) {
67       Result = BB;
68       ResultIsRemembered = Remember;
69       return;
70     }
71 
72     BasicBlock *NewResult = DT->findNearestCommonDominator(Result, BB);
73     if (NewResult != Result)
74       ResultIsRemembered = false;
75     if (NewResult == BB)
76       ResultIsRemembered |= Remember;
77     Result = NewResult;
78   }
79 
80 public:
81   explicit NearestCommonDominator(DominatorTree *DomTree) : DT(DomTree) {}
82 
83   void addBlock(BasicBlock *BB) {
84     addBlock(BB, /* Remember = */ false);
85   }
86 
87   void addAndRememberBlock(BasicBlock *BB) {
88     addBlock(BB, /* Remember = */ true);
89   }
90 
91   /// Get the nearest common dominator of all the BBs added via addBlock() and
92   /// addAndRememberBlock().
93   BasicBlock *result() { return Result; }
94 
95   /// Is the BB returned by getResult() one of the blocks we added to the set
96   /// with addAndRememberBlock()?
97   bool resultIsRememberedBlock() { return ResultIsRemembered; }
98 };
99 
100 /// @brief Transforms the control flow graph on one single entry/exit region
101 /// at a time.
102 ///
103 /// After the transform all "If"/"Then"/"Else" style control flow looks like
104 /// this:
105 ///
106 /// \verbatim
107 /// 1
108 /// ||
109 /// | |
110 /// 2 |
111 /// | /
112 /// |/
113 /// 3
114 /// ||   Where:
115 /// | |  1 = "If" block, calculates the condition
116 /// 4 |  2 = "Then" subregion, runs if the condition is true
117 /// | /  3 = "Flow" blocks, newly inserted flow blocks, rejoins the flow
118 /// |/   4 = "Else" optional subregion, runs if the condition is false
119 /// 5    5 = "End" block, also rejoins the control flow
120 /// \endverbatim
121 ///
122 /// Control flow is expressed as a branch where the true exit goes into the
123 /// "Then"/"Else" region, while the false exit skips the region
124 /// The condition for the optional "Else" region is expressed as a PHI node.
125 /// The incoming values of the PHI node are true for the "If" edge and false
126 /// for the "Then" edge.
127 ///
128 /// Additionally to that even complicated loops look like this:
129 ///
130 /// \verbatim
131 /// 1
132 /// ||
133 /// | |
134 /// 2 ^  Where:
135 /// | /  1 = "Entry" block
136 /// |/   2 = "Loop" optional subregion, with all exits at "Flow" block
137 /// 3    3 = "Flow" block, with back edge to entry block
138 /// |
139 /// \endverbatim
140 ///
141 /// The back edge of the "Flow" block is always on the false side of the branch
142 /// while the true side continues the general flow. So the loop condition
143 /// consist of a network of PHI nodes where the true incoming values expresses
144 /// breaks and the false values expresses continue states.
145 class StructurizeCFG : public RegionPass {
146   bool SkipUniformRegions;
147 
148   Type *Boolean;
149   ConstantInt *BoolTrue;
150   ConstantInt *BoolFalse;
151   UndefValue *BoolUndef;
152 
153   Function *Func;
154   Region *ParentRegion;
155 
156   DominatorTree *DT;
157   LoopInfo *LI;
158 
159   SmallVector<RegionNode *, 8> Order;
160   BBSet Visited;
161 
162   BBPhiMap DeletedPhis;
163   BB2BBVecMap AddedPhis;
164 
165   PredMap Predicates;
166   BranchVector Conditions;
167 
168   BB2BBMap Loops;
169   PredMap LoopPreds;
170   BranchVector LoopConds;
171 
172   RegionNode *PrevNode;
173 
174   void orderNodes();
175 
176   void analyzeLoops(RegionNode *N);
177 
178   Value *invert(Value *Condition);
179 
180   Value *buildCondition(BranchInst *Term, unsigned Idx, bool Invert);
181 
182   void gatherPredicates(RegionNode *N);
183 
184   void collectInfos();
185 
186   void insertConditions(bool Loops);
187 
188   void delPhiValues(BasicBlock *From, BasicBlock *To);
189 
190   void addPhiValues(BasicBlock *From, BasicBlock *To);
191 
192   void setPhiValues();
193 
194   void killTerminator(BasicBlock *BB);
195 
196   void changeExit(RegionNode *Node, BasicBlock *NewExit,
197                   bool IncludeDominator);
198 
199   BasicBlock *getNextFlow(BasicBlock *Dominator);
200 
201   BasicBlock *needPrefix(bool NeedEmpty);
202 
203   BasicBlock *needPostfix(BasicBlock *Flow, bool ExitUseAllowed);
204 
205   void setPrevNode(BasicBlock *BB);
206 
207   bool dominatesPredicates(BasicBlock *BB, RegionNode *Node);
208 
209   bool isPredictableTrue(RegionNode *Node);
210 
211   void wireFlow(bool ExitUseAllowed, BasicBlock *LoopEnd);
212 
213   void handleLoops(bool ExitUseAllowed, BasicBlock *LoopEnd);
214 
215   void createFlow();
216 
217   void rebuildSSA();
218 
219 public:
220   static char ID;
221 
222   explicit StructurizeCFG(bool SkipUniformRegions = false)
223       : RegionPass(ID), SkipUniformRegions(SkipUniformRegions) {
224     initializeStructurizeCFGPass(*PassRegistry::getPassRegistry());
225   }
226 
227   bool doInitialization(Region *R, RGPassManager &RGM) override;
228 
229   bool runOnRegion(Region *R, RGPassManager &RGM) override;
230 
231   StringRef getPassName() const override { return "Structurize control flow"; }
232 
233   void getAnalysisUsage(AnalysisUsage &AU) const override {
234     if (SkipUniformRegions)
235       AU.addRequired<DivergenceAnalysis>();
236     AU.addRequiredID(LowerSwitchID);
237     AU.addRequired<DominatorTreeWrapperPass>();
238     AU.addRequired<LoopInfoWrapperPass>();
239 
240     AU.addPreserved<DominatorTreeWrapperPass>();
241     RegionPass::getAnalysisUsage(AU);
242   }
243 };
244 
245 } // end anonymous namespace
246 
247 char StructurizeCFG::ID = 0;
248 
249 INITIALIZE_PASS_BEGIN(StructurizeCFG, "structurizecfg", "Structurize the CFG",
250                       false, false)
251 INITIALIZE_PASS_DEPENDENCY(DivergenceAnalysis)
252 INITIALIZE_PASS_DEPENDENCY(LowerSwitch)
253 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
254 INITIALIZE_PASS_DEPENDENCY(RegionInfoPass)
255 INITIALIZE_PASS_END(StructurizeCFG, "structurizecfg", "Structurize the CFG",
256                     false, false)
257 
258 /// \brief Initialize the types and constants used in the pass
259 bool StructurizeCFG::doInitialization(Region *R, RGPassManager &RGM) {
260   LLVMContext &Context = R->getEntry()->getContext();
261 
262   Boolean = Type::getInt1Ty(Context);
263   BoolTrue = ConstantInt::getTrue(Context);
264   BoolFalse = ConstantInt::getFalse(Context);
265   BoolUndef = UndefValue::get(Boolean);
266 
267   return false;
268 }
269 
270 /// \brief Build up the general order of nodes
271 void StructurizeCFG::orderNodes() {
272   ReversePostOrderTraversal<Region*> RPOT(ParentRegion);
273   SmallDenseMap<Loop*, unsigned, 8> LoopBlocks;
274 
275   // The reverse post-order traversal of the list gives us an ordering close
276   // to what we want.  The only problem with it is that sometimes backedges
277   // for outer loops will be visited before backedges for inner loops.
278   for (RegionNode *RN : RPOT) {
279     BasicBlock *BB = RN->getEntry();
280     Loop *Loop = LI->getLoopFor(BB);
281     ++LoopBlocks[Loop];
282   }
283 
284   unsigned CurrentLoopDepth = 0;
285   Loop *CurrentLoop = nullptr;
286   for (auto I = RPOT.begin(), E = RPOT.end(); I != E; ++I) {
287     BasicBlock *BB = (*I)->getEntry();
288     unsigned LoopDepth = LI->getLoopDepth(BB);
289 
290     if (is_contained(Order, *I))
291       continue;
292 
293     if (LoopDepth < CurrentLoopDepth) {
294       // Make sure we have visited all blocks in this loop before moving back to
295       // the outer loop.
296 
297       auto LoopI = I;
298       while (unsigned &BlockCount = LoopBlocks[CurrentLoop]) {
299         LoopI++;
300         BasicBlock *LoopBB = (*LoopI)->getEntry();
301         if (LI->getLoopFor(LoopBB) == CurrentLoop) {
302           --BlockCount;
303           Order.push_back(*LoopI);
304         }
305       }
306     }
307 
308     CurrentLoop = LI->getLoopFor(BB);
309     if (CurrentLoop)
310       LoopBlocks[CurrentLoop]--;
311 
312     CurrentLoopDepth = LoopDepth;
313     Order.push_back(*I);
314   }
315 
316   // This pass originally used a post-order traversal and then operated on
317   // the list in reverse. Now that we are using a reverse post-order traversal
318   // rather than re-working the whole pass to operate on the list in order,
319   // we just reverse the list and continue to operate on it in reverse.
320   std::reverse(Order.begin(), Order.end());
321 }
322 
323 /// \brief Determine the end of the loops
324 void StructurizeCFG::analyzeLoops(RegionNode *N) {
325   if (N->isSubRegion()) {
326     // Test for exit as back edge
327     BasicBlock *Exit = N->getNodeAs<Region>()->getExit();
328     if (Visited.count(Exit))
329       Loops[Exit] = N->getEntry();
330 
331   } else {
332     // Test for sucessors as back edge
333     BasicBlock *BB = N->getNodeAs<BasicBlock>();
334     BranchInst *Term = cast<BranchInst>(BB->getTerminator());
335 
336     for (BasicBlock *Succ : Term->successors())
337       if (Visited.count(Succ))
338         Loops[Succ] = BB;
339   }
340 }
341 
342 /// \brief Invert the given condition
343 Value *StructurizeCFG::invert(Value *Condition) {
344   // First: Check if it's a constant
345   if (Constant *C = dyn_cast<Constant>(Condition))
346     return ConstantExpr::getNot(C);
347 
348   // Second: If the condition is already inverted, return the original value
349   if (match(Condition, m_Not(m_Value(Condition))))
350     return Condition;
351 
352   if (Instruction *Inst = dyn_cast<Instruction>(Condition)) {
353     // Third: Check all the users for an invert
354     BasicBlock *Parent = Inst->getParent();
355     for (User *U : Condition->users())
356       if (Instruction *I = dyn_cast<Instruction>(U))
357         if (I->getParent() == Parent && match(I, m_Not(m_Specific(Condition))))
358           return I;
359 
360     // Last option: Create a new instruction
361     return BinaryOperator::CreateNot(Condition, "", Parent->getTerminator());
362   }
363 
364   if (Argument *Arg = dyn_cast<Argument>(Condition)) {
365     BasicBlock &EntryBlock = Arg->getParent()->getEntryBlock();
366     return BinaryOperator::CreateNot(Condition,
367                                      Arg->getName() + ".inv",
368                                      EntryBlock.getTerminator());
369   }
370 
371   llvm_unreachable("Unhandled condition to invert");
372 }
373 
374 /// \brief Build the condition for one edge
375 Value *StructurizeCFG::buildCondition(BranchInst *Term, unsigned Idx,
376                                       bool Invert) {
377   Value *Cond = Invert ? BoolFalse : BoolTrue;
378   if (Term->isConditional()) {
379     Cond = Term->getCondition();
380 
381     if (Idx != (unsigned)Invert)
382       Cond = invert(Cond);
383   }
384   return Cond;
385 }
386 
387 /// \brief Analyze the predecessors of each block and build up predicates
388 void StructurizeCFG::gatherPredicates(RegionNode *N) {
389   RegionInfo *RI = ParentRegion->getRegionInfo();
390   BasicBlock *BB = N->getEntry();
391   BBPredicates &Pred = Predicates[BB];
392   BBPredicates &LPred = LoopPreds[BB];
393 
394   for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
395        PI != PE; ++PI) {
396 
397     // Ignore it if it's a branch from outside into our region entry
398     if (!ParentRegion->contains(*PI))
399       continue;
400 
401     Region *R = RI->getRegionFor(*PI);
402     if (R == ParentRegion) {
403 
404       // It's a top level block in our region
405       BranchInst *Term = cast<BranchInst>((*PI)->getTerminator());
406       for (unsigned i = 0, e = Term->getNumSuccessors(); i != e; ++i) {
407         BasicBlock *Succ = Term->getSuccessor(i);
408         if (Succ != BB)
409           continue;
410 
411         if (Visited.count(*PI)) {
412           // Normal forward edge
413           if (Term->isConditional()) {
414             // Try to treat it like an ELSE block
415             BasicBlock *Other = Term->getSuccessor(!i);
416             if (Visited.count(Other) && !Loops.count(Other) &&
417                 !Pred.count(Other) && !Pred.count(*PI)) {
418 
419               Pred[Other] = BoolFalse;
420               Pred[*PI] = BoolTrue;
421               continue;
422             }
423           }
424           Pred[*PI] = buildCondition(Term, i, false);
425 
426         } else {
427           // Back edge
428           LPred[*PI] = buildCondition(Term, i, true);
429         }
430       }
431 
432     } else {
433 
434       // It's an exit from a sub region
435       while (R->getParent() != ParentRegion)
436         R = R->getParent();
437 
438       // Edge from inside a subregion to its entry, ignore it
439       if (*R == *N)
440         continue;
441 
442       BasicBlock *Entry = R->getEntry();
443       if (Visited.count(Entry))
444         Pred[Entry] = BoolTrue;
445       else
446         LPred[Entry] = BoolFalse;
447     }
448   }
449 }
450 
451 /// \brief Collect various loop and predicate infos
452 void StructurizeCFG::collectInfos() {
453   // Reset predicate
454   Predicates.clear();
455 
456   // and loop infos
457   Loops.clear();
458   LoopPreds.clear();
459 
460   // Reset the visited nodes
461   Visited.clear();
462 
463   for (RegionNode *RN : reverse(Order)) {
464 
465     DEBUG(dbgs() << "Visiting: "
466                  << (RN->isSubRegion() ? "SubRegion with entry: " : "")
467                  << RN->getEntry()->getName() << " Loop Depth: "
468                  << LI->getLoopDepth(RN->getEntry()) << "\n");
469 
470     // Analyze all the conditions leading to a node
471     gatherPredicates(RN);
472 
473     // Remember that we've seen this node
474     Visited.insert(RN->getEntry());
475 
476     // Find the last back edges
477     analyzeLoops(RN);
478   }
479 }
480 
481 /// \brief Insert the missing branch conditions
482 void StructurizeCFG::insertConditions(bool Loops) {
483   BranchVector &Conds = Loops ? LoopConds : Conditions;
484   Value *Default = Loops ? BoolTrue : BoolFalse;
485   SSAUpdater PhiInserter;
486 
487   for (BranchInst *Term : Conds) {
488     assert(Term->isConditional());
489 
490     BasicBlock *Parent = Term->getParent();
491     BasicBlock *SuccTrue = Term->getSuccessor(0);
492     BasicBlock *SuccFalse = Term->getSuccessor(1);
493 
494     PhiInserter.Initialize(Boolean, "");
495     PhiInserter.AddAvailableValue(&Func->getEntryBlock(), Default);
496     PhiInserter.AddAvailableValue(Loops ? SuccFalse : Parent, Default);
497 
498     BBPredicates &Preds = Loops ? LoopPreds[SuccFalse] : Predicates[SuccTrue];
499 
500     NearestCommonDominator Dominator(DT);
501     Dominator.addBlock(Parent);
502 
503     Value *ParentValue = nullptr;
504     for (BBPredicates::iterator PI = Preds.begin(), PE = Preds.end();
505          PI != PE; ++PI) {
506 
507       if (PI->first == Parent) {
508         ParentValue = PI->second;
509         break;
510       }
511       PhiInserter.AddAvailableValue(PI->first, PI->second);
512       Dominator.addAndRememberBlock(PI->first);
513     }
514 
515     if (ParentValue) {
516       Term->setCondition(ParentValue);
517     } else {
518       if (!Dominator.resultIsRememberedBlock())
519         PhiInserter.AddAvailableValue(Dominator.result(), Default);
520 
521       Term->setCondition(PhiInserter.GetValueInMiddleOfBlock(Parent));
522     }
523   }
524 }
525 
526 /// \brief Remove all PHI values coming from "From" into "To" and remember
527 /// them in DeletedPhis
528 void StructurizeCFG::delPhiValues(BasicBlock *From, BasicBlock *To) {
529   PhiMap &Map = DeletedPhis[To];
530   for (BasicBlock::iterator I = To->begin(), E = To->end();
531        I != E && isa<PHINode>(*I);) {
532 
533     PHINode &Phi = cast<PHINode>(*I++);
534     while (Phi.getBasicBlockIndex(From) != -1) {
535       Value *Deleted = Phi.removeIncomingValue(From, false);
536       Map[&Phi].push_back(std::make_pair(From, Deleted));
537     }
538   }
539 }
540 
541 /// \brief Add a dummy PHI value as soon as we knew the new predecessor
542 void StructurizeCFG::addPhiValues(BasicBlock *From, BasicBlock *To) {
543   for (BasicBlock::iterator I = To->begin(), E = To->end();
544        I != E && isa<PHINode>(*I);) {
545 
546     PHINode &Phi = cast<PHINode>(*I++);
547     Value *Undef = UndefValue::get(Phi.getType());
548     Phi.addIncoming(Undef, From);
549   }
550   AddedPhis[To].push_back(From);
551 }
552 
553 /// \brief Add the real PHI value as soon as everything is set up
554 void StructurizeCFG::setPhiValues() {
555   SSAUpdater Updater;
556   for (const auto &AddedPhi : AddedPhis) {
557 
558     BasicBlock *To = AddedPhi.first;
559     const BBVector &From = AddedPhi.second;
560 
561     if (!DeletedPhis.count(To))
562       continue;
563 
564     PhiMap &Map = DeletedPhis[To];
565     for (const auto &PI : Map) {
566 
567       PHINode *Phi = PI.first;
568       Value *Undef = UndefValue::get(Phi->getType());
569       Updater.Initialize(Phi->getType(), "");
570       Updater.AddAvailableValue(&Func->getEntryBlock(), Undef);
571       Updater.AddAvailableValue(To, Undef);
572 
573       NearestCommonDominator Dominator(DT);
574       Dominator.addBlock(To);
575       for (const auto &VI : PI.second) {
576 
577         Updater.AddAvailableValue(VI.first, VI.second);
578         Dominator.addAndRememberBlock(VI.first);
579       }
580 
581       if (!Dominator.resultIsRememberedBlock())
582         Updater.AddAvailableValue(Dominator.result(), Undef);
583 
584       for (BasicBlock *FI : From) {
585 
586         int Idx = Phi->getBasicBlockIndex(FI);
587         assert(Idx != -1);
588         Phi->setIncomingValue(Idx, Updater.GetValueAtEndOfBlock(FI));
589       }
590     }
591 
592     DeletedPhis.erase(To);
593   }
594   assert(DeletedPhis.empty());
595 }
596 
597 /// \brief Remove phi values from all successors and then remove the terminator.
598 void StructurizeCFG::killTerminator(BasicBlock *BB) {
599   TerminatorInst *Term = BB->getTerminator();
600   if (!Term)
601     return;
602 
603   for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB);
604        SI != SE; ++SI) {
605 
606     delPhiValues(BB, *SI);
607   }
608 
609   Term->eraseFromParent();
610 }
611 
612 /// \brief Let node exit(s) point to NewExit
613 void StructurizeCFG::changeExit(RegionNode *Node, BasicBlock *NewExit,
614                                 bool IncludeDominator) {
615   if (Node->isSubRegion()) {
616     Region *SubRegion = Node->getNodeAs<Region>();
617     BasicBlock *OldExit = SubRegion->getExit();
618     BasicBlock *Dominator = nullptr;
619 
620     // Find all the edges from the sub region to the exit
621     for (pred_iterator I = pred_begin(OldExit), E = pred_end(OldExit);
622          I != E;) {
623 
624       BasicBlock *BB = *I++;
625       if (!SubRegion->contains(BB))
626         continue;
627 
628       // Modify the edges to point to the new exit
629       delPhiValues(BB, OldExit);
630       BB->getTerminator()->replaceUsesOfWith(OldExit, NewExit);
631       addPhiValues(BB, NewExit);
632 
633       // Find the new dominator (if requested)
634       if (IncludeDominator) {
635         if (!Dominator)
636           Dominator = BB;
637         else
638           Dominator = DT->findNearestCommonDominator(Dominator, BB);
639       }
640     }
641 
642     // Change the dominator (if requested)
643     if (Dominator)
644       DT->changeImmediateDominator(NewExit, Dominator);
645 
646     // Update the region info
647     SubRegion->replaceExit(NewExit);
648 
649   } else {
650     BasicBlock *BB = Node->getNodeAs<BasicBlock>();
651     killTerminator(BB);
652     BranchInst::Create(NewExit, BB);
653     addPhiValues(BB, NewExit);
654     if (IncludeDominator)
655       DT->changeImmediateDominator(NewExit, BB);
656   }
657 }
658 
659 /// \brief Create a new flow node and update dominator tree and region info
660 BasicBlock *StructurizeCFG::getNextFlow(BasicBlock *Dominator) {
661   LLVMContext &Context = Func->getContext();
662   BasicBlock *Insert = Order.empty() ? ParentRegion->getExit() :
663                        Order.back()->getEntry();
664   BasicBlock *Flow = BasicBlock::Create(Context, FlowBlockName,
665                                         Func, Insert);
666   DT->addNewBlock(Flow, Dominator);
667   ParentRegion->getRegionInfo()->setRegionFor(Flow, ParentRegion);
668   return Flow;
669 }
670 
671 /// \brief Create a new or reuse the previous node as flow node
672 BasicBlock *StructurizeCFG::needPrefix(bool NeedEmpty) {
673   BasicBlock *Entry = PrevNode->getEntry();
674 
675   if (!PrevNode->isSubRegion()) {
676     killTerminator(Entry);
677     if (!NeedEmpty || Entry->getFirstInsertionPt() == Entry->end())
678       return Entry;
679 
680   }
681 
682   // create a new flow node
683   BasicBlock *Flow = getNextFlow(Entry);
684 
685   // and wire it up
686   changeExit(PrevNode, Flow, true);
687   PrevNode = ParentRegion->getBBNode(Flow);
688   return Flow;
689 }
690 
691 /// \brief Returns the region exit if possible, otherwise just a new flow node
692 BasicBlock *StructurizeCFG::needPostfix(BasicBlock *Flow,
693                                         bool ExitUseAllowed) {
694   if (Order.empty() && ExitUseAllowed) {
695     BasicBlock *Exit = ParentRegion->getExit();
696     DT->changeImmediateDominator(Exit, Flow);
697     addPhiValues(Flow, Exit);
698     return Exit;
699   }
700   return getNextFlow(Flow);
701 }
702 
703 /// \brief Set the previous node
704 void StructurizeCFG::setPrevNode(BasicBlock *BB) {
705   PrevNode = ParentRegion->contains(BB) ? ParentRegion->getBBNode(BB)
706                                         : nullptr;
707 }
708 
709 /// \brief Does BB dominate all the predicates of Node ?
710 bool StructurizeCFG::dominatesPredicates(BasicBlock *BB, RegionNode *Node) {
711   BBPredicates &Preds = Predicates[Node->getEntry()];
712   for (BBPredicates::iterator PI = Preds.begin(), PE = Preds.end();
713        PI != PE; ++PI) {
714 
715     if (!DT->dominates(BB, PI->first))
716       return false;
717   }
718   return true;
719 }
720 
721 /// \brief Can we predict that this node will always be called?
722 bool StructurizeCFG::isPredictableTrue(RegionNode *Node) {
723   BBPredicates &Preds = Predicates[Node->getEntry()];
724   bool Dominated = false;
725 
726   // Regionentry is always true
727   if (!PrevNode)
728     return true;
729 
730   for (BBPredicates::iterator I = Preds.begin(), E = Preds.end();
731        I != E; ++I) {
732 
733     if (I->second != BoolTrue)
734       return false;
735 
736     if (!Dominated && DT->dominates(I->first, PrevNode->getEntry()))
737       Dominated = true;
738   }
739 
740   // TODO: The dominator check is too strict
741   return Dominated;
742 }
743 
744 /// Take one node from the order vector and wire it up
745 void StructurizeCFG::wireFlow(bool ExitUseAllowed,
746                               BasicBlock *LoopEnd) {
747   RegionNode *Node = Order.pop_back_val();
748   Visited.insert(Node->getEntry());
749 
750   if (isPredictableTrue(Node)) {
751     // Just a linear flow
752     if (PrevNode) {
753       changeExit(PrevNode, Node->getEntry(), true);
754     }
755     PrevNode = Node;
756 
757   } else {
758     // Insert extra prefix node (or reuse last one)
759     BasicBlock *Flow = needPrefix(false);
760 
761     // Insert extra postfix node (or use exit instead)
762     BasicBlock *Entry = Node->getEntry();
763     BasicBlock *Next = needPostfix(Flow, ExitUseAllowed);
764 
765     // let it point to entry and next block
766     Conditions.push_back(BranchInst::Create(Entry, Next, BoolUndef, Flow));
767     addPhiValues(Flow, Entry);
768     DT->changeImmediateDominator(Entry, Flow);
769 
770     PrevNode = Node;
771     while (!Order.empty() && !Visited.count(LoopEnd) &&
772            dominatesPredicates(Entry, Order.back())) {
773       handleLoops(false, LoopEnd);
774     }
775 
776     changeExit(PrevNode, Next, false);
777     setPrevNode(Next);
778   }
779 }
780 
781 void StructurizeCFG::handleLoops(bool ExitUseAllowed,
782                                  BasicBlock *LoopEnd) {
783   RegionNode *Node = Order.back();
784   BasicBlock *LoopStart = Node->getEntry();
785 
786   if (!Loops.count(LoopStart)) {
787     wireFlow(ExitUseAllowed, LoopEnd);
788     return;
789   }
790 
791   if (!isPredictableTrue(Node))
792     LoopStart = needPrefix(true);
793 
794   LoopEnd = Loops[Node->getEntry()];
795   wireFlow(false, LoopEnd);
796   while (!Visited.count(LoopEnd)) {
797     handleLoops(false, LoopEnd);
798   }
799 
800   // If the start of the loop is the entry block, we can't branch to it so
801   // insert a new dummy entry block.
802   Function *LoopFunc = LoopStart->getParent();
803   if (LoopStart == &LoopFunc->getEntryBlock()) {
804     LoopStart->setName("entry.orig");
805 
806     BasicBlock *NewEntry =
807       BasicBlock::Create(LoopStart->getContext(),
808                          "entry",
809                          LoopFunc,
810                          LoopStart);
811     BranchInst::Create(LoopStart, NewEntry);
812   }
813 
814   // Create an extra loop end node
815   LoopEnd = needPrefix(false);
816   BasicBlock *Next = needPostfix(LoopEnd, ExitUseAllowed);
817   LoopConds.push_back(BranchInst::Create(Next, LoopStart,
818                                          BoolUndef, LoopEnd));
819   addPhiValues(LoopEnd, LoopStart);
820   setPrevNode(Next);
821 }
822 
823 /// After this function control flow looks like it should be, but
824 /// branches and PHI nodes only have undefined conditions.
825 void StructurizeCFG::createFlow() {
826   BasicBlock *Exit = ParentRegion->getExit();
827   bool EntryDominatesExit = DT->dominates(ParentRegion->getEntry(), Exit);
828 
829   DeletedPhis.clear();
830   AddedPhis.clear();
831   Conditions.clear();
832   LoopConds.clear();
833 
834   PrevNode = nullptr;
835   Visited.clear();
836 
837   while (!Order.empty()) {
838     handleLoops(EntryDominatesExit, nullptr);
839   }
840 
841   if (PrevNode)
842     changeExit(PrevNode, Exit, EntryDominatesExit);
843   else
844     assert(EntryDominatesExit);
845 }
846 
847 /// Handle a rare case where the disintegrated nodes instructions
848 /// no longer dominate all their uses. Not sure if this is really nessasary
849 void StructurizeCFG::rebuildSSA() {
850   SSAUpdater Updater;
851   for (auto *BB : ParentRegion->blocks())
852     for (BasicBlock::iterator II = BB->begin(), IE = BB->end();
853          II != IE; ++II) {
854 
855       bool Initialized = false;
856       for (auto I = II->use_begin(), E = II->use_end(); I != E;) {
857         Use &U = *I++;
858         Instruction *User = cast<Instruction>(U.getUser());
859         if (User->getParent() == BB) {
860           continue;
861 
862         } else if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
863           if (UserPN->getIncomingBlock(U) == BB)
864             continue;
865         }
866 
867         if (DT->dominates(&*II, User))
868           continue;
869 
870         if (!Initialized) {
871           Value *Undef = UndefValue::get(II->getType());
872           Updater.Initialize(II->getType(), "");
873           Updater.AddAvailableValue(&Func->getEntryBlock(), Undef);
874           Updater.AddAvailableValue(BB, &*II);
875           Initialized = true;
876         }
877         Updater.RewriteUseAfterInsertions(U);
878       }
879     }
880 }
881 
882 static bool hasOnlyUniformBranches(const Region *R,
883                                    const DivergenceAnalysis &DA) {
884   for (const BasicBlock *BB : R->blocks()) {
885     const BranchInst *Br = dyn_cast<BranchInst>(BB->getTerminator());
886     if (!Br || !Br->isConditional())
887       continue;
888 
889     if (!DA.isUniform(Br->getCondition()))
890       return false;
891     DEBUG(dbgs() << "BB: " << BB->getName() << " has uniform terminator\n");
892   }
893   return true;
894 }
895 
896 /// \brief Run the transformation for each region found
897 bool StructurizeCFG::runOnRegion(Region *R, RGPassManager &RGM) {
898   if (R->isTopLevelRegion())
899     return false;
900 
901   if (SkipUniformRegions) {
902     // TODO: We could probably be smarter here with how we handle sub-regions.
903     auto &DA = getAnalysis<DivergenceAnalysis>();
904     if (hasOnlyUniformBranches(R, DA)) {
905       DEBUG(dbgs() << "Skipping region with uniform control flow: " << *R << '\n');
906 
907       // Mark all direct child block terminators as having been treated as
908       // uniform. To account for a possible future in which non-uniform
909       // sub-regions are treated more cleverly, indirect children are not
910       // marked as uniform.
911       MDNode *MD = MDNode::get(R->getEntry()->getParent()->getContext(), {});
912       for (RegionNode *E : R->elements()) {
913         if (E->isSubRegion())
914           continue;
915 
916         if (Instruction *Term = E->getEntry()->getTerminator())
917           Term->setMetadata("structurizecfg.uniform", MD);
918       }
919 
920       return false;
921     }
922   }
923 
924   Func = R->getEntry()->getParent();
925   ParentRegion = R;
926 
927   DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
928   LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
929 
930   orderNodes();
931   collectInfos();
932   createFlow();
933   insertConditions(false);
934   insertConditions(true);
935   setPhiValues();
936   rebuildSSA();
937 
938   // Cleanup
939   Order.clear();
940   Visited.clear();
941   DeletedPhis.clear();
942   AddedPhis.clear();
943   Predicates.clear();
944   Conditions.clear();
945   Loops.clear();
946   LoopPreds.clear();
947   LoopConds.clear();
948 
949   return true;
950 }
951 
952 Pass *llvm::createStructurizeCFGPass(bool SkipUniformRegions) {
953   return new StructurizeCFG(SkipUniformRegions);
954 }
955