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/LoopInfo.h"
15 #include "llvm/Analysis/RegionInfo.h"
16 #include "llvm/Analysis/RegionIterator.h"
17 #include "llvm/Analysis/RegionPass.h"
18 #include "llvm/IR/Module.h"
19 #include "llvm/IR/PatternMatch.h"
20 #include "llvm/Support/Debug.h"
21 #include "llvm/Support/raw_ostream.h"
22 #include "llvm/Transforms/Utils/SSAUpdater.h"
23 
24 using namespace llvm;
25 using namespace llvm::PatternMatch;
26 
27 #define DEBUG_TYPE "structurizecfg"
28 
29 namespace {
30 
31 // Definition of the complex types used in this pass.
32 
33 typedef std::pair<BasicBlock *, Value *> BBValuePair;
34 
35 typedef SmallVector<RegionNode*, 8> RNVector;
36 typedef SmallVector<BasicBlock*, 8> BBVector;
37 typedef SmallVector<BranchInst*, 8> BranchVector;
38 typedef SmallVector<BBValuePair, 2> BBValueVector;
39 
40 typedef SmallPtrSet<BasicBlock *, 8> BBSet;
41 
42 typedef MapVector<PHINode *, BBValueVector> PhiMap;
43 typedef MapVector<BasicBlock *, BBVector> BB2BBVecMap;
44 
45 typedef DenseMap<DomTreeNode *, unsigned> DTN2UnsignedMap;
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 
53 static const char *const FlowBlockName = "Flow";
54 
55 /// @brief Find the nearest common dominator for multiple BasicBlocks
56 ///
57 /// Helper class for StructurizeCFG
58 /// TODO: Maybe move into common code
59 class NearestCommonDominator {
60   DominatorTree *DT;
61 
62   DTN2UnsignedMap IndexMap;
63 
64   BasicBlock *Result;
65   unsigned ResultIndex;
66   bool ExplicitMentioned;
67 
68 public:
69   /// \brief Start a new query
NearestCommonDominator(DominatorTree * DomTree)70   NearestCommonDominator(DominatorTree *DomTree) {
71     DT = DomTree;
72     Result = nullptr;
73   }
74 
75   /// \brief Add BB to the resulting dominator
addBlock(BasicBlock * BB,bool Remember=true)76   void addBlock(BasicBlock *BB, bool Remember = true) {
77     DomTreeNode *Node = DT->getNode(BB);
78 
79     if (!Result) {
80       unsigned Numbering = 0;
81       for (;Node;Node = Node->getIDom())
82         IndexMap[Node] = ++Numbering;
83       Result = BB;
84       ResultIndex = 1;
85       ExplicitMentioned = Remember;
86       return;
87     }
88 
89     for (;Node;Node = Node->getIDom())
90       if (IndexMap.count(Node))
91         break;
92       else
93         IndexMap[Node] = 0;
94 
95     assert(Node && "Dominator tree invalid!");
96 
97     unsigned Numbering = IndexMap[Node];
98     if (Numbering > ResultIndex) {
99       Result = Node->getBlock();
100       ResultIndex = Numbering;
101       ExplicitMentioned = Remember && (Result == BB);
102     } else if (Numbering == ResultIndex) {
103       ExplicitMentioned |= Remember;
104     }
105   }
106 
107   /// \brief Is "Result" one of the BBs added with "Remember" = True?
wasResultExplicitMentioned()108   bool wasResultExplicitMentioned() {
109     return ExplicitMentioned;
110   }
111 
112   /// \brief Get the query result
getResult()113   BasicBlock *getResult() {
114     return Result;
115   }
116 };
117 
118 /// @brief Transforms the control flow graph on one single entry/exit region
119 /// at a time.
120 ///
121 /// After the transform all "If"/"Then"/"Else" style control flow looks like
122 /// this:
123 ///
124 /// \verbatim
125 /// 1
126 /// ||
127 /// | |
128 /// 2 |
129 /// | /
130 /// |/
131 /// 3
132 /// ||   Where:
133 /// | |  1 = "If" block, calculates the condition
134 /// 4 |  2 = "Then" subregion, runs if the condition is true
135 /// | /  3 = "Flow" blocks, newly inserted flow blocks, rejoins the flow
136 /// |/   4 = "Else" optional subregion, runs if the condition is false
137 /// 5    5 = "End" block, also rejoins the control flow
138 /// \endverbatim
139 ///
140 /// Control flow is expressed as a branch where the true exit goes into the
141 /// "Then"/"Else" region, while the false exit skips the region
142 /// The condition for the optional "Else" region is expressed as a PHI node.
143 /// The incomming values of the PHI node are true for the "If" edge and false
144 /// for the "Then" edge.
145 ///
146 /// Additionally to that even complicated loops look like this:
147 ///
148 /// \verbatim
149 /// 1
150 /// ||
151 /// | |
152 /// 2 ^  Where:
153 /// | /  1 = "Entry" block
154 /// |/   2 = "Loop" optional subregion, with all exits at "Flow" block
155 /// 3    3 = "Flow" block, with back edge to entry block
156 /// |
157 /// \endverbatim
158 ///
159 /// The back edge of the "Flow" block is always on the false side of the branch
160 /// while the true side continues the general flow. So the loop condition
161 /// consist of a network of PHI nodes where the true incoming values expresses
162 /// breaks and the false values expresses continue states.
163 class StructurizeCFG : public RegionPass {
164   Type *Boolean;
165   ConstantInt *BoolTrue;
166   ConstantInt *BoolFalse;
167   UndefValue *BoolUndef;
168 
169   Function *Func;
170   Region *ParentRegion;
171 
172   DominatorTree *DT;
173   LoopInfo *LI;
174 
175   RNVector Order;
176   BBSet Visited;
177 
178   BBPhiMap DeletedPhis;
179   BB2BBVecMap AddedPhis;
180 
181   PredMap Predicates;
182   BranchVector Conditions;
183 
184   BB2BBMap Loops;
185   PredMap LoopPreds;
186   BranchVector LoopConds;
187 
188   RegionNode *PrevNode;
189 
190   void orderNodes();
191 
192   void analyzeLoops(RegionNode *N);
193 
194   Value *invert(Value *Condition);
195 
196   Value *buildCondition(BranchInst *Term, unsigned Idx, bool Invert);
197 
198   void gatherPredicates(RegionNode *N);
199 
200   void collectInfos();
201 
202   void insertConditions(bool Loops);
203 
204   void delPhiValues(BasicBlock *From, BasicBlock *To);
205 
206   void addPhiValues(BasicBlock *From, BasicBlock *To);
207 
208   void setPhiValues();
209 
210   void killTerminator(BasicBlock *BB);
211 
212   void changeExit(RegionNode *Node, BasicBlock *NewExit,
213                   bool IncludeDominator);
214 
215   BasicBlock *getNextFlow(BasicBlock *Dominator);
216 
217   BasicBlock *needPrefix(bool NeedEmpty);
218 
219   BasicBlock *needPostfix(BasicBlock *Flow, bool ExitUseAllowed);
220 
221   void setPrevNode(BasicBlock *BB);
222 
223   bool dominatesPredicates(BasicBlock *BB, RegionNode *Node);
224 
225   bool isPredictableTrue(RegionNode *Node);
226 
227   void wireFlow(bool ExitUseAllowed, BasicBlock *LoopEnd);
228 
229   void handleLoops(bool ExitUseAllowed, BasicBlock *LoopEnd);
230 
231   void createFlow();
232 
233   void rebuildSSA();
234 
235 public:
236   static char ID;
237 
StructurizeCFG()238   StructurizeCFG() :
239     RegionPass(ID) {
240     initializeStructurizeCFGPass(*PassRegistry::getPassRegistry());
241   }
242 
243   using Pass::doInitialization;
244   bool doInitialization(Region *R, RGPassManager &RGM) override;
245 
246   bool runOnRegion(Region *R, RGPassManager &RGM) override;
247 
getPassName() const248   const char *getPassName() const override {
249     return "Structurize control flow";
250   }
251 
getAnalysisUsage(AnalysisUsage & AU) const252   void getAnalysisUsage(AnalysisUsage &AU) const override {
253     AU.addRequiredID(LowerSwitchID);
254     AU.addRequired<DominatorTreeWrapperPass>();
255     AU.addRequired<LoopInfoWrapperPass>();
256     AU.addPreserved<DominatorTreeWrapperPass>();
257     RegionPass::getAnalysisUsage(AU);
258   }
259 };
260 
261 } // end anonymous namespace
262 
263 char StructurizeCFG::ID = 0;
264 
265 INITIALIZE_PASS_BEGIN(StructurizeCFG, "structurizecfg", "Structurize the CFG",
266                       false, false)
INITIALIZE_PASS_DEPENDENCY(LowerSwitch)267 INITIALIZE_PASS_DEPENDENCY(LowerSwitch)
268 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
269 INITIALIZE_PASS_DEPENDENCY(RegionInfoPass)
270 INITIALIZE_PASS_END(StructurizeCFG, "structurizecfg", "Structurize the CFG",
271                     false, false)
272 
273 /// \brief Initialize the types and constants used in the pass
274 bool StructurizeCFG::doInitialization(Region *R, RGPassManager &RGM) {
275   LLVMContext &Context = R->getEntry()->getContext();
276 
277   Boolean = Type::getInt1Ty(Context);
278   BoolTrue = ConstantInt::getTrue(Context);
279   BoolFalse = ConstantInt::getFalse(Context);
280   BoolUndef = UndefValue::get(Boolean);
281 
282   return false;
283 }
284 
285 /// \brief Build up the general order of nodes
orderNodes()286 void StructurizeCFG::orderNodes() {
287   RNVector TempOrder;
288   ReversePostOrderTraversal<Region*> RPOT(ParentRegion);
289   TempOrder.append(RPOT.begin(), RPOT.end());
290 
291   std::map<Loop*, unsigned> LoopBlocks;
292 
293 
294   // The reverse post-order traversal of the list gives us an ordering close
295   // to what we want.  The only problem with it is that sometimes backedges
296   // for outer loops will be visited before backedges for inner loops.
297   for (RegionNode *RN : TempOrder) {
298     BasicBlock *BB = RN->getEntry();
299     Loop *Loop = LI->getLoopFor(BB);
300     if (!LoopBlocks.count(Loop)) {
301       LoopBlocks[Loop] = 1;
302       continue;
303     }
304     LoopBlocks[Loop]++;
305   }
306 
307   unsigned CurrentLoopDepth = 0;
308   Loop *CurrentLoop = nullptr;
309   BBSet TempVisited;
310   for (RNVector::iterator I = TempOrder.begin(), E = TempOrder.end(); I != E; ++I) {
311     BasicBlock *BB = (*I)->getEntry();
312     unsigned LoopDepth = LI->getLoopDepth(BB);
313 
314     if (std::find(Order.begin(), Order.end(), *I) != Order.end())
315       continue;
316 
317     if (LoopDepth < CurrentLoopDepth) {
318       // Make sure we have visited all blocks in this loop before moving back to
319       // the outer loop.
320 
321       RNVector::iterator LoopI = I;
322       while(LoopBlocks[CurrentLoop]) {
323         LoopI++;
324         BasicBlock *LoopBB = (*LoopI)->getEntry();
325         if (LI->getLoopFor(LoopBB) == CurrentLoop) {
326           LoopBlocks[CurrentLoop]--;
327           Order.push_back(*LoopI);
328         }
329       }
330     }
331 
332     CurrentLoop = LI->getLoopFor(BB);
333     if (CurrentLoop) {
334       LoopBlocks[CurrentLoop]--;
335     }
336 
337     CurrentLoopDepth = LoopDepth;
338     Order.push_back(*I);
339   }
340 
341   // This pass originally used a post-order traversal and then operated on
342   // the list in reverse. Now that we are using a reverse post-order traversal
343   // rather than re-working the whole pass to operate on the list in order,
344   // we just reverse the list and continue to operate on it in reverse.
345   std::reverse(Order.begin(), Order.end());
346 }
347 
348 /// \brief Determine the end of the loops
analyzeLoops(RegionNode * N)349 void StructurizeCFG::analyzeLoops(RegionNode *N) {
350   if (N->isSubRegion()) {
351     // Test for exit as back edge
352     BasicBlock *Exit = N->getNodeAs<Region>()->getExit();
353     if (Visited.count(Exit))
354       Loops[Exit] = N->getEntry();
355 
356   } else {
357     // Test for sucessors as back edge
358     BasicBlock *BB = N->getNodeAs<BasicBlock>();
359     BranchInst *Term = cast<BranchInst>(BB->getTerminator());
360 
361     for (BasicBlock *Succ : Term->successors())
362       if (Visited.count(Succ))
363         Loops[Succ] = BB;
364   }
365 }
366 
367 /// \brief Invert the given condition
invert(Value * Condition)368 Value *StructurizeCFG::invert(Value *Condition) {
369   // First: Check if it's a constant
370   if (Condition == BoolTrue)
371     return BoolFalse;
372 
373   if (Condition == BoolFalse)
374     return BoolTrue;
375 
376   if (Condition == BoolUndef)
377     return BoolUndef;
378 
379   // Second: If the condition is already inverted, return the original value
380   if (match(Condition, m_Not(m_Value(Condition))))
381     return Condition;
382 
383   if (Instruction *Inst = dyn_cast<Instruction>(Condition)) {
384     // Third: Check all the users for an invert
385     BasicBlock *Parent = Inst->getParent();
386     for (User *U : Condition->users())
387       if (Instruction *I = dyn_cast<Instruction>(U))
388         if (I->getParent() == Parent && match(I, m_Not(m_Specific(Condition))))
389           return I;
390 
391     // Last option: Create a new instruction
392     return BinaryOperator::CreateNot(Condition, "", Parent->getTerminator());
393   }
394 
395   if (Argument *Arg = dyn_cast<Argument>(Condition)) {
396     BasicBlock &EntryBlock = Arg->getParent()->getEntryBlock();
397     return BinaryOperator::CreateNot(Condition,
398                                      Arg->getName() + ".inv",
399                                      EntryBlock.getTerminator());
400   }
401 
402   llvm_unreachable("Unhandled condition to invert");
403 }
404 
405 /// \brief Build the condition for one edge
buildCondition(BranchInst * Term,unsigned Idx,bool Invert)406 Value *StructurizeCFG::buildCondition(BranchInst *Term, unsigned Idx,
407                                       bool Invert) {
408   Value *Cond = Invert ? BoolFalse : BoolTrue;
409   if (Term->isConditional()) {
410     Cond = Term->getCondition();
411 
412     if (Idx != (unsigned)Invert)
413       Cond = invert(Cond);
414   }
415   return Cond;
416 }
417 
418 /// \brief Analyze the predecessors of each block and build up predicates
gatherPredicates(RegionNode * N)419 void StructurizeCFG::gatherPredicates(RegionNode *N) {
420   RegionInfo *RI = ParentRegion->getRegionInfo();
421   BasicBlock *BB = N->getEntry();
422   BBPredicates &Pred = Predicates[BB];
423   BBPredicates &LPred = LoopPreds[BB];
424 
425   for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
426        PI != PE; ++PI) {
427 
428     // Ignore it if it's a branch from outside into our region entry
429     if (!ParentRegion->contains(*PI))
430       continue;
431 
432     Region *R = RI->getRegionFor(*PI);
433     if (R == ParentRegion) {
434 
435       // It's a top level block in our region
436       BranchInst *Term = cast<BranchInst>((*PI)->getTerminator());
437       for (unsigned i = 0, e = Term->getNumSuccessors(); i != e; ++i) {
438         BasicBlock *Succ = Term->getSuccessor(i);
439         if (Succ != BB)
440           continue;
441 
442         if (Visited.count(*PI)) {
443           // Normal forward edge
444           if (Term->isConditional()) {
445             // Try to treat it like an ELSE block
446             BasicBlock *Other = Term->getSuccessor(!i);
447             if (Visited.count(Other) && !Loops.count(Other) &&
448                 !Pred.count(Other) && !Pred.count(*PI)) {
449 
450               Pred[Other] = BoolFalse;
451               Pred[*PI] = BoolTrue;
452               continue;
453             }
454           }
455           Pred[*PI] = buildCondition(Term, i, false);
456 
457         } else {
458           // Back edge
459           LPred[*PI] = buildCondition(Term, i, true);
460         }
461       }
462 
463     } else {
464 
465       // It's an exit from a sub region
466       while (R->getParent() != ParentRegion)
467         R = R->getParent();
468 
469       // Edge from inside a subregion to its entry, ignore it
470       if (*R == *N)
471         continue;
472 
473       BasicBlock *Entry = R->getEntry();
474       if (Visited.count(Entry))
475         Pred[Entry] = BoolTrue;
476       else
477         LPred[Entry] = BoolFalse;
478     }
479   }
480 }
481 
482 /// \brief Collect various loop and predicate infos
collectInfos()483 void StructurizeCFG::collectInfos() {
484   // Reset predicate
485   Predicates.clear();
486 
487   // and loop infos
488   Loops.clear();
489   LoopPreds.clear();
490 
491   // Reset the visited nodes
492   Visited.clear();
493 
494   for (RNVector::reverse_iterator OI = Order.rbegin(), OE = Order.rend();
495        OI != OE; ++OI) {
496 
497     DEBUG(dbgs() << "Visiting: " <<
498                     ((*OI)->isSubRegion() ? "SubRegion with entry: " : "") <<
499                     (*OI)->getEntry()->getName() << " Loop Depth: " << LI->getLoopDepth((*OI)->getEntry()) << "\n");
500 
501     // Analyze all the conditions leading to a node
502     gatherPredicates(*OI);
503 
504     // Remember that we've seen this node
505     Visited.insert((*OI)->getEntry());
506 
507     // Find the last back edges
508     analyzeLoops(*OI);
509   }
510 }
511 
512 /// \brief Insert the missing branch conditions
insertConditions(bool Loops)513 void StructurizeCFG::insertConditions(bool Loops) {
514   BranchVector &Conds = Loops ? LoopConds : Conditions;
515   Value *Default = Loops ? BoolTrue : BoolFalse;
516   SSAUpdater PhiInserter;
517 
518   for (BranchInst *Term : Conds) {
519     assert(Term->isConditional());
520 
521     BasicBlock *Parent = Term->getParent();
522     BasicBlock *SuccTrue = Term->getSuccessor(0);
523     BasicBlock *SuccFalse = Term->getSuccessor(1);
524 
525     PhiInserter.Initialize(Boolean, "");
526     PhiInserter.AddAvailableValue(&Func->getEntryBlock(), Default);
527     PhiInserter.AddAvailableValue(Loops ? SuccFalse : Parent, Default);
528 
529     BBPredicates &Preds = Loops ? LoopPreds[SuccFalse] : Predicates[SuccTrue];
530 
531     NearestCommonDominator Dominator(DT);
532     Dominator.addBlock(Parent, false);
533 
534     Value *ParentValue = nullptr;
535     for (BBPredicates::iterator PI = Preds.begin(), PE = Preds.end();
536          PI != PE; ++PI) {
537 
538       if (PI->first == Parent) {
539         ParentValue = PI->second;
540         break;
541       }
542       PhiInserter.AddAvailableValue(PI->first, PI->second);
543       Dominator.addBlock(PI->first);
544     }
545 
546     if (ParentValue) {
547       Term->setCondition(ParentValue);
548     } else {
549       if (!Dominator.wasResultExplicitMentioned())
550         PhiInserter.AddAvailableValue(Dominator.getResult(), Default);
551 
552       Term->setCondition(PhiInserter.GetValueInMiddleOfBlock(Parent));
553     }
554   }
555 }
556 
557 /// \brief Remove all PHI values coming from "From" into "To" and remember
558 /// them in DeletedPhis
delPhiValues(BasicBlock * From,BasicBlock * To)559 void StructurizeCFG::delPhiValues(BasicBlock *From, BasicBlock *To) {
560   PhiMap &Map = DeletedPhis[To];
561   for (BasicBlock::iterator I = To->begin(), E = To->end();
562        I != E && isa<PHINode>(*I);) {
563 
564     PHINode &Phi = cast<PHINode>(*I++);
565     while (Phi.getBasicBlockIndex(From) != -1) {
566       Value *Deleted = Phi.removeIncomingValue(From, false);
567       Map[&Phi].push_back(std::make_pair(From, Deleted));
568     }
569   }
570 }
571 
572 /// \brief Add a dummy PHI value as soon as we knew the new predecessor
addPhiValues(BasicBlock * From,BasicBlock * To)573 void StructurizeCFG::addPhiValues(BasicBlock *From, BasicBlock *To) {
574   for (BasicBlock::iterator I = To->begin(), E = To->end();
575        I != E && isa<PHINode>(*I);) {
576 
577     PHINode &Phi = cast<PHINode>(*I++);
578     Value *Undef = UndefValue::get(Phi.getType());
579     Phi.addIncoming(Undef, From);
580   }
581   AddedPhis[To].push_back(From);
582 }
583 
584 /// \brief Add the real PHI value as soon as everything is set up
setPhiValues()585 void StructurizeCFG::setPhiValues() {
586   SSAUpdater Updater;
587   for (BB2BBVecMap::iterator AI = AddedPhis.begin(), AE = AddedPhis.end();
588        AI != AE; ++AI) {
589 
590     BasicBlock *To = AI->first;
591     BBVector &From = AI->second;
592 
593     if (!DeletedPhis.count(To))
594       continue;
595 
596     PhiMap &Map = DeletedPhis[To];
597     for (PhiMap::iterator PI = Map.begin(), PE = Map.end();
598          PI != PE; ++PI) {
599 
600       PHINode *Phi = PI->first;
601       Value *Undef = UndefValue::get(Phi->getType());
602       Updater.Initialize(Phi->getType(), "");
603       Updater.AddAvailableValue(&Func->getEntryBlock(), Undef);
604       Updater.AddAvailableValue(To, Undef);
605 
606       NearestCommonDominator Dominator(DT);
607       Dominator.addBlock(To, false);
608       for (BBValueVector::iterator VI = PI->second.begin(),
609            VE = PI->second.end(); VI != VE; ++VI) {
610 
611         Updater.AddAvailableValue(VI->first, VI->second);
612         Dominator.addBlock(VI->first);
613       }
614 
615       if (!Dominator.wasResultExplicitMentioned())
616         Updater.AddAvailableValue(Dominator.getResult(), Undef);
617 
618       for (BBVector::iterator FI = From.begin(), FE = From.end();
619            FI != FE; ++FI) {
620 
621         int Idx = Phi->getBasicBlockIndex(*FI);
622         assert(Idx != -1);
623         Phi->setIncomingValue(Idx, Updater.GetValueAtEndOfBlock(*FI));
624       }
625     }
626 
627     DeletedPhis.erase(To);
628   }
629   assert(DeletedPhis.empty());
630 }
631 
632 /// \brief Remove phi values from all successors and then remove the terminator.
killTerminator(BasicBlock * BB)633 void StructurizeCFG::killTerminator(BasicBlock *BB) {
634   TerminatorInst *Term = BB->getTerminator();
635   if (!Term)
636     return;
637 
638   for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB);
639        SI != SE; ++SI) {
640 
641     delPhiValues(BB, *SI);
642   }
643 
644   Term->eraseFromParent();
645 }
646 
647 /// \brief Let node exit(s) point to NewExit
changeExit(RegionNode * Node,BasicBlock * NewExit,bool IncludeDominator)648 void StructurizeCFG::changeExit(RegionNode *Node, BasicBlock *NewExit,
649                                 bool IncludeDominator) {
650   if (Node->isSubRegion()) {
651     Region *SubRegion = Node->getNodeAs<Region>();
652     BasicBlock *OldExit = SubRegion->getExit();
653     BasicBlock *Dominator = nullptr;
654 
655     // Find all the edges from the sub region to the exit
656     for (pred_iterator I = pred_begin(OldExit), E = pred_end(OldExit);
657          I != E;) {
658 
659       BasicBlock *BB = *I++;
660       if (!SubRegion->contains(BB))
661         continue;
662 
663       // Modify the edges to point to the new exit
664       delPhiValues(BB, OldExit);
665       BB->getTerminator()->replaceUsesOfWith(OldExit, NewExit);
666       addPhiValues(BB, NewExit);
667 
668       // Find the new dominator (if requested)
669       if (IncludeDominator) {
670         if (!Dominator)
671           Dominator = BB;
672         else
673           Dominator = DT->findNearestCommonDominator(Dominator, BB);
674       }
675     }
676 
677     // Change the dominator (if requested)
678     if (Dominator)
679       DT->changeImmediateDominator(NewExit, Dominator);
680 
681     // Update the region info
682     SubRegion->replaceExit(NewExit);
683 
684   } else {
685     BasicBlock *BB = Node->getNodeAs<BasicBlock>();
686     killTerminator(BB);
687     BranchInst::Create(NewExit, BB);
688     addPhiValues(BB, NewExit);
689     if (IncludeDominator)
690       DT->changeImmediateDominator(NewExit, BB);
691   }
692 }
693 
694 /// \brief Create a new flow node and update dominator tree and region info
getNextFlow(BasicBlock * Dominator)695 BasicBlock *StructurizeCFG::getNextFlow(BasicBlock *Dominator) {
696   LLVMContext &Context = Func->getContext();
697   BasicBlock *Insert = Order.empty() ? ParentRegion->getExit() :
698                        Order.back()->getEntry();
699   BasicBlock *Flow = BasicBlock::Create(Context, FlowBlockName,
700                                         Func, Insert);
701   DT->addNewBlock(Flow, Dominator);
702   ParentRegion->getRegionInfo()->setRegionFor(Flow, ParentRegion);
703   return Flow;
704 }
705 
706 /// \brief Create a new or reuse the previous node as flow node
needPrefix(bool NeedEmpty)707 BasicBlock *StructurizeCFG::needPrefix(bool NeedEmpty) {
708   BasicBlock *Entry = PrevNode->getEntry();
709 
710   if (!PrevNode->isSubRegion()) {
711     killTerminator(Entry);
712     if (!NeedEmpty || Entry->getFirstInsertionPt() == Entry->end())
713       return Entry;
714 
715   }
716 
717   // create a new flow node
718   BasicBlock *Flow = getNextFlow(Entry);
719 
720   // and wire it up
721   changeExit(PrevNode, Flow, true);
722   PrevNode = ParentRegion->getBBNode(Flow);
723   return Flow;
724 }
725 
726 /// \brief Returns the region exit if possible, otherwise just a new flow node
needPostfix(BasicBlock * Flow,bool ExitUseAllowed)727 BasicBlock *StructurizeCFG::needPostfix(BasicBlock *Flow,
728                                         bool ExitUseAllowed) {
729   if (Order.empty() && ExitUseAllowed) {
730     BasicBlock *Exit = ParentRegion->getExit();
731     DT->changeImmediateDominator(Exit, Flow);
732     addPhiValues(Flow, Exit);
733     return Exit;
734   }
735   return getNextFlow(Flow);
736 }
737 
738 /// \brief Set the previous node
setPrevNode(BasicBlock * BB)739 void StructurizeCFG::setPrevNode(BasicBlock *BB) {
740   PrevNode = ParentRegion->contains(BB) ? ParentRegion->getBBNode(BB)
741                                         : nullptr;
742 }
743 
744 /// \brief Does BB dominate all the predicates of Node ?
dominatesPredicates(BasicBlock * BB,RegionNode * Node)745 bool StructurizeCFG::dominatesPredicates(BasicBlock *BB, RegionNode *Node) {
746   BBPredicates &Preds = Predicates[Node->getEntry()];
747   for (BBPredicates::iterator PI = Preds.begin(), PE = Preds.end();
748        PI != PE; ++PI) {
749 
750     if (!DT->dominates(BB, PI->first))
751       return false;
752   }
753   return true;
754 }
755 
756 /// \brief Can we predict that this node will always be called?
isPredictableTrue(RegionNode * Node)757 bool StructurizeCFG::isPredictableTrue(RegionNode *Node) {
758   BBPredicates &Preds = Predicates[Node->getEntry()];
759   bool Dominated = false;
760 
761   // Regionentry is always true
762   if (!PrevNode)
763     return true;
764 
765   for (BBPredicates::iterator I = Preds.begin(), E = Preds.end();
766        I != E; ++I) {
767 
768     if (I->second != BoolTrue)
769       return false;
770 
771     if (!Dominated && DT->dominates(I->first, PrevNode->getEntry()))
772       Dominated = true;
773   }
774 
775   // TODO: The dominator check is too strict
776   return Dominated;
777 }
778 
779 /// Take one node from the order vector and wire it up
wireFlow(bool ExitUseAllowed,BasicBlock * LoopEnd)780 void StructurizeCFG::wireFlow(bool ExitUseAllowed,
781                               BasicBlock *LoopEnd) {
782   RegionNode *Node = Order.pop_back_val();
783   Visited.insert(Node->getEntry());
784 
785   if (isPredictableTrue(Node)) {
786     // Just a linear flow
787     if (PrevNode) {
788       changeExit(PrevNode, Node->getEntry(), true);
789     }
790     PrevNode = Node;
791 
792   } else {
793     // Insert extra prefix node (or reuse last one)
794     BasicBlock *Flow = needPrefix(false);
795 
796     // Insert extra postfix node (or use exit instead)
797     BasicBlock *Entry = Node->getEntry();
798     BasicBlock *Next = needPostfix(Flow, ExitUseAllowed);
799 
800     // let it point to entry and next block
801     Conditions.push_back(BranchInst::Create(Entry, Next, BoolUndef, Flow));
802     addPhiValues(Flow, Entry);
803     DT->changeImmediateDominator(Entry, Flow);
804 
805     PrevNode = Node;
806     while (!Order.empty() && !Visited.count(LoopEnd) &&
807            dominatesPredicates(Entry, Order.back())) {
808       handleLoops(false, LoopEnd);
809     }
810 
811     changeExit(PrevNode, Next, false);
812     setPrevNode(Next);
813   }
814 }
815 
handleLoops(bool ExitUseAllowed,BasicBlock * LoopEnd)816 void StructurizeCFG::handleLoops(bool ExitUseAllowed,
817                                  BasicBlock *LoopEnd) {
818   RegionNode *Node = Order.back();
819   BasicBlock *LoopStart = Node->getEntry();
820 
821   if (!Loops.count(LoopStart)) {
822     wireFlow(ExitUseAllowed, LoopEnd);
823     return;
824   }
825 
826   if (!isPredictableTrue(Node))
827     LoopStart = needPrefix(true);
828 
829   LoopEnd = Loops[Node->getEntry()];
830   wireFlow(false, LoopEnd);
831   while (!Visited.count(LoopEnd)) {
832     handleLoops(false, LoopEnd);
833   }
834 
835   // If the start of the loop is the entry block, we can't branch to it so
836   // insert a new dummy entry block.
837   Function *LoopFunc = LoopStart->getParent();
838   if (LoopStart == &LoopFunc->getEntryBlock()) {
839     LoopStart->setName("entry.orig");
840 
841     BasicBlock *NewEntry =
842       BasicBlock::Create(LoopStart->getContext(),
843                          "entry",
844                          LoopFunc,
845                          LoopStart);
846     BranchInst::Create(LoopStart, NewEntry);
847   }
848 
849   // Create an extra loop end node
850   LoopEnd = needPrefix(false);
851   BasicBlock *Next = needPostfix(LoopEnd, ExitUseAllowed);
852   LoopConds.push_back(BranchInst::Create(Next, LoopStart,
853                                          BoolUndef, LoopEnd));
854   addPhiValues(LoopEnd, LoopStart);
855   setPrevNode(Next);
856 }
857 
858 /// After this function control flow looks like it should be, but
859 /// branches and PHI nodes only have undefined conditions.
createFlow()860 void StructurizeCFG::createFlow() {
861   BasicBlock *Exit = ParentRegion->getExit();
862   bool EntryDominatesExit = DT->dominates(ParentRegion->getEntry(), Exit);
863 
864   DeletedPhis.clear();
865   AddedPhis.clear();
866   Conditions.clear();
867   LoopConds.clear();
868 
869   PrevNode = nullptr;
870   Visited.clear();
871 
872   while (!Order.empty()) {
873     handleLoops(EntryDominatesExit, nullptr);
874   }
875 
876   if (PrevNode)
877     changeExit(PrevNode, Exit, EntryDominatesExit);
878   else
879     assert(EntryDominatesExit);
880 }
881 
882 /// Handle a rare case where the disintegrated nodes instructions
883 /// no longer dominate all their uses. Not sure if this is really nessasary
rebuildSSA()884 void StructurizeCFG::rebuildSSA() {
885   SSAUpdater Updater;
886   for (auto *BB : ParentRegion->blocks())
887     for (BasicBlock::iterator II = BB->begin(), IE = BB->end();
888          II != IE; ++II) {
889 
890       bool Initialized = false;
891       for (auto I = II->use_begin(), E = II->use_end(); I != E;) {
892         Use &U = *I++;
893         Instruction *User = cast<Instruction>(U.getUser());
894         if (User->getParent() == BB) {
895           continue;
896 
897         } else if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
898           if (UserPN->getIncomingBlock(U) == BB)
899             continue;
900         }
901 
902         if (DT->dominates(&*II, User))
903           continue;
904 
905         if (!Initialized) {
906           Value *Undef = UndefValue::get(II->getType());
907           Updater.Initialize(II->getType(), "");
908           Updater.AddAvailableValue(&Func->getEntryBlock(), Undef);
909           Updater.AddAvailableValue(BB, &*II);
910           Initialized = true;
911         }
912         Updater.RewriteUseAfterInsertions(U);
913       }
914     }
915 }
916 
917 /// \brief Run the transformation for each region found
runOnRegion(Region * R,RGPassManager & RGM)918 bool StructurizeCFG::runOnRegion(Region *R, RGPassManager &RGM) {
919   if (R->isTopLevelRegion())
920     return false;
921 
922   Func = R->getEntry()->getParent();
923   ParentRegion = R;
924 
925   DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
926   LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
927 
928   orderNodes();
929   collectInfos();
930   createFlow();
931   insertConditions(false);
932   insertConditions(true);
933   setPhiValues();
934   rebuildSSA();
935 
936   // Cleanup
937   Order.clear();
938   Visited.clear();
939   DeletedPhis.clear();
940   AddedPhis.clear();
941   Predicates.clear();
942   Conditions.clear();
943   Loops.clear();
944   LoopPreds.clear();
945   LoopConds.clear();
946 
947   return true;
948 }
949 
950 /// \brief Create the pass
createStructurizeCFGPass()951 Pass *llvm::createStructurizeCFGPass() {
952   return new StructurizeCFG();
953 }
954