1 //===- RegionInfo.h - SESE region analysis ----------------------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Calculate a program structure tree built out of single entry single exit
12 // The basic ideas are taken from "The Program Structure Tree - Richard Johnson,
13 // David Pearson, Keshav Pingali - 1994", however enriched with ideas from "The
14 // Refined Process Structure Tree - Jussi Vanhatalo, Hagen Voelyer, Jana
16 // The algorithm to calculate these data structures however is completely
17 // different, as it takes advantage of existing information already available
18 // in (Post)dominace tree and dominance frontier passes. This leads to a simpler
19 // and in practice hopefully better performing algorithm. The runtime of the
20 // algorithms described in the papers above are both linear in graph size,
21 // O(V+E), whereas this algorithm is not, as the dominance frontier information
22 // itself is not, but in practice runtime seems to be in the order of magnitude
23 // of dominance tree calculation.
25 //===----------------------------------------------------------------------===//
27 #ifndef LLVM_ANALYSIS_REGION_INFO_H
28 #define LLVM_ANALYSIS_REGION_INFO_H
30 #include "llvm/ADT/PointerIntPair.h"
31 #include "llvm/Analysis/DominanceFrontier.h"
32 #include "llvm/Analysis/PostDominators.h"
33 #include "llvm/Support/Allocator.h"
44 /// @brief Marker class to iterate over the elements of a Region in flat mode.
46 /// The class is used to either iterate in Flat mode or by not using it to not
47 /// iterate in Flat mode. During a Flat mode iteration all Regions are entered
48 /// and the iteration returns every BasicBlock. If the Flat mode is not
49 /// selected for SubRegions just one RegionNode containing the subregion is
51 template <class GraphType>
54 /// @brief A RegionNode represents a subregion or a BasicBlock that is part of a
57 RegionNode(const RegionNode &) LLVM_DELETED_FUNCTION;
58 const RegionNode &operator=(const RegionNode &) LLVM_DELETED_FUNCTION;
61 /// This is the entry basic block that starts this region node. If this is a
62 /// BasicBlock RegionNode, then entry is just the basic block, that this
63 /// RegionNode represents. Otherwise it is the entry of this (Sub)RegionNode.
65 /// In the BBtoRegionNode map of the parent of this node, BB will always map
66 /// to this node no matter which kind of node this one is.
68 /// The node can hold either a Region or a BasicBlock.
69 /// Use one bit to save, if this RegionNode is a subregion or BasicBlock
71 PointerIntPair<BasicBlock*, 1, bool> entry;
73 /// @brief The parent Region of this RegionNode.
78 /// @brief Create a RegionNode.
80 /// @param Parent The parent of this RegionNode.
81 /// @param Entry The entry BasicBlock of the RegionNode. If this
82 /// RegionNode represents a BasicBlock, this is the
83 /// BasicBlock itself. If it represents a subregion, this
84 /// is the entry BasicBlock of the subregion.
85 /// @param isSubRegion If this RegionNode represents a SubRegion.
86 inline RegionNode(Region* Parent, BasicBlock* Entry, bool isSubRegion = 0)
87 : entry(Entry, isSubRegion), parent(Parent) {}
89 /// @brief Get the parent Region of this RegionNode.
91 /// The parent Region is the Region this RegionNode belongs to. If for
92 /// example a BasicBlock is element of two Regions, there exist two
93 /// RegionNodes for this BasicBlock. Each with the getParent() function
94 /// pointing to the Region this RegionNode belongs to.
96 /// @return Get the parent Region of this RegionNode.
97 inline Region* getParent() const { return parent; }
99 /// @brief Get the entry BasicBlock of this RegionNode.
101 /// If this RegionNode represents a BasicBlock this is just the BasicBlock
102 /// itself, otherwise we return the entry BasicBlock of the Subregion
104 /// @return The entry BasicBlock of this RegionNode.
105 inline BasicBlock* getEntry() const { return entry.getPointer(); }
107 /// @brief Get the content of this RegionNode.
109 /// This can be either a BasicBlock or a subregion. Before calling getNodeAs()
110 /// check the type of the content with the isSubRegion() function call.
112 /// @return The content of this RegionNode.
114 inline T* getNodeAs() const;
116 /// @brief Is this RegionNode a subregion?
118 /// @return True if it contains a subregion. False if it contains a
120 inline bool isSubRegion() const {
121 return entry.getInt();
125 /// Print a RegionNode.
126 inline raw_ostream &operator<<(raw_ostream &OS, const RegionNode &Node);
129 inline BasicBlock* RegionNode::getNodeAs<BasicBlock>() const {
130 assert(!isSubRegion() && "This is not a BasicBlock RegionNode!");
135 inline Region* RegionNode::getNodeAs<Region>() const {
136 assert(isSubRegion() && "This is not a subregion RegionNode!");
137 return reinterpret_cast<Region*>(const_cast<RegionNode*>(this));
140 //===----------------------------------------------------------------------===//
141 /// @brief A single entry single exit Region.
143 /// A Region is a connected subgraph of a control flow graph that has exactly
144 /// two connections to the remaining graph. It can be used to analyze or
145 /// optimize parts of the control flow graph.
147 /// A <em> simple Region </em> is connected to the remaining graph by just two
148 /// edges. One edge entering the Region and another one leaving the Region.
150 /// An <em> extended Region </em> (or just Region) is a subgraph that can be
151 /// transform into a simple Region. The transformation is done by adding
152 /// BasicBlocks that merge several entry or exit edges so that after the merge
153 /// just one entry and one exit edge exists.
155 /// The \e Entry of a Region is the first BasicBlock that is passed after
156 /// entering the Region. It is an element of the Region. The entry BasicBlock
157 /// dominates all BasicBlocks in the Region.
159 /// The \e Exit of a Region is the first BasicBlock that is passed after
160 /// leaving the Region. It is not an element of the Region. The exit BasicBlock,
161 /// postdominates all BasicBlocks in the Region.
163 /// A <em> canonical Region </em> cannot be constructed by combining smaller
166 /// Region A is the \e parent of Region B, if B is completely contained in A.
168 /// Two canonical Regions either do not intersect at all or one is
169 /// the parent of the other.
171 /// The <em> Program Structure Tree</em> is a graph (V, E) where V is the set of
172 /// Regions in the control flow graph and E is the \e parent relation of these
178 /// A simple control flow graph, that contains two regions.
188 /// \ |/ Region A: 1 -> 9 {1,2,3,4,5,6,7,8}
189 /// 9 Region B: 2 -> 9 {2,4,5,6,7}
192 /// You can obtain more examples by either calling
194 /// <tt> "opt -regions -analyze anyprogram.ll" </tt>
196 /// <tt> "opt -view-regions-only anyprogram.ll" </tt>
198 /// on any LLVM file you are interested in.
200 /// The first call returns a textual representation of the program structure
201 /// tree, the second one creates a graphical representation using graphviz.
202 class Region : public RegionNode {
203 friend class RegionInfo;
204 Region(const Region &) LLVM_DELETED_FUNCTION;
205 const Region &operator=(const Region &) LLVM_DELETED_FUNCTION;
207 // Information necessary to manage this Region.
211 // The exit BasicBlock of this region.
212 // (The entry BasicBlock is part of RegionNode)
215 typedef std::vector<Region*> RegionSet;
217 // The subregions of this region.
220 typedef std::map<BasicBlock*, RegionNode*> BBNodeMapT;
222 // Save the BasicBlock RegionNodes that are element of this Region.
223 mutable BBNodeMapT BBNodeMap;
225 /// verifyBBInRegion - Check if a BB is in this Region. This check also works
226 /// if the region is incorrectly built. (EXPENSIVE!)
227 void verifyBBInRegion(BasicBlock* BB) const;
229 /// verifyWalk - Walk over all the BBs of the region starting from BB and
230 /// verify that all reachable basic blocks are elements of the region.
232 void verifyWalk(BasicBlock* BB, std::set<BasicBlock*>* visitedBB) const;
234 /// verifyRegionNest - Verify if the region and its children are valid
235 /// regions (EXPENSIVE!)
236 void verifyRegionNest() const;
239 /// @brief Create a new region.
241 /// @param Entry The entry basic block of the region.
242 /// @param Exit The exit basic block of the region.
243 /// @param RI The region info object that is managing this region.
244 /// @param DT The dominator tree of the current function.
245 /// @param Parent The surrounding region or NULL if this is a top level
247 Region(BasicBlock *Entry, BasicBlock *Exit, RegionInfo* RI,
248 DominatorTree *DT, Region *Parent = 0);
250 /// Delete the Region and all its subregions.
253 /// @brief Get the entry BasicBlock of the Region.
254 /// @return The entry BasicBlock of the region.
255 BasicBlock *getEntry() const { return RegionNode::getEntry(); }
257 /// @brief Replace the entry basic block of the region with the new basic
260 /// @param BB The new entry basic block of the region.
261 void replaceEntry(BasicBlock *BB);
263 /// @brief Replace the exit basic block of the region with the new basic
266 /// @param BB The new exit basic block of the region.
267 void replaceExit(BasicBlock *BB);
269 /// @brief Get the exit BasicBlock of the Region.
270 /// @return The exit BasicBlock of the Region, NULL if this is the TopLevel
272 BasicBlock *getExit() const { return exit; }
274 /// @brief Get the parent of the Region.
275 /// @return The parent of the Region or NULL if this is a top level
277 Region *getParent() const { return RegionNode::getParent(); }
279 /// @brief Get the RegionNode representing the current Region.
280 /// @return The RegionNode representing the current Region.
281 RegionNode* getNode() const {
282 return const_cast<RegionNode*>(reinterpret_cast<const RegionNode*>(this));
285 /// @brief Get the nesting level of this Region.
287 /// An toplevel Region has depth 0.
289 /// @return The depth of the region.
290 unsigned getDepth() const;
292 /// @brief Check if a Region is the TopLevel region.
294 /// The toplevel region represents the whole function.
295 bool isTopLevelRegion() const { return exit == NULL; }
297 /// @brief Return a new (non canonical) region, that is obtained by joining
298 /// this region with its predecessors.
300 /// @return A region also starting at getEntry(), but reaching to the next
301 /// basic block that forms with getEntry() a (non canonical) region.
302 /// NULL if such a basic block does not exist.
303 Region *getExpandedRegion() const;
305 /// @brief Return the first block of this region's single entry edge,
308 /// @return The BasicBlock starting this region's single entry edge,
310 BasicBlock *getEnteringBlock() const;
312 /// @brief Return the first block of this region's single exit edge,
315 /// @return The BasicBlock starting this region's single exit edge,
317 BasicBlock *getExitingBlock() const;
319 /// @brief Is this a simple region?
321 /// A region is simple if it has exactly one exit and one entry edge.
323 /// @return True if the Region is simple.
324 bool isSimple() const;
326 /// @brief Returns the name of the Region.
327 /// @return The Name of the Region.
328 std::string getNameStr() const;
330 /// @brief Return the RegionInfo object, that belongs to this Region.
331 RegionInfo *getRegionInfo() const {
335 /// PrintStyle - Print region in difference ways.
336 enum PrintStyle { PrintNone, PrintBB, PrintRN };
338 /// @brief Print the region.
340 /// @param OS The output stream the Region is printed to.
341 /// @param printTree Print also the tree of subregions.
342 /// @param level The indentation level used for printing.
343 void print(raw_ostream& OS, bool printTree = true, unsigned level = 0,
344 enum PrintStyle Style = PrintNone) const;
346 /// @brief Print the region to stderr.
349 /// @brief Check if the region contains a BasicBlock.
351 /// @param BB The BasicBlock that might be contained in this Region.
352 /// @return True if the block is contained in the region otherwise false.
353 bool contains(const BasicBlock *BB) const;
355 /// @brief Check if the region contains another region.
357 /// @param SubRegion The region that might be contained in this Region.
358 /// @return True if SubRegion is contained in the region otherwise false.
359 bool contains(const Region *SubRegion) const {
364 return contains(SubRegion->getEntry())
365 && (contains(SubRegion->getExit()) || SubRegion->getExit() == getExit());
368 /// @brief Check if the region contains an Instruction.
370 /// @param Inst The Instruction that might be contained in this region.
371 /// @return True if the Instruction is contained in the region otherwise false.
372 bool contains(const Instruction *Inst) const {
373 return contains(Inst->getParent());
376 /// @brief Check if the region contains a loop.
378 /// @param L The loop that might be contained in this region.
379 /// @return True if the loop is contained in the region otherwise false.
380 /// In case a NULL pointer is passed to this function the result
381 /// is false, except for the region that describes the whole function.
382 /// In that case true is returned.
383 bool contains(const Loop *L) const;
385 /// @brief Get the outermost loop in the region that contains a loop.
387 /// Find for a Loop L the outermost loop OuterL that is a parent loop of L
388 /// and is itself contained in the region.
390 /// @param L The loop the lookup is started.
391 /// @return The outermost loop in the region, NULL if such a loop does not
392 /// exist or if the region describes the whole function.
393 Loop *outermostLoopInRegion(Loop *L) const;
395 /// @brief Get the outermost loop in the region that contains a basic block.
397 /// Find for a basic block BB the outermost loop L that contains BB and is
398 /// itself contained in the region.
400 /// @param LI A pointer to a LoopInfo analysis.
401 /// @param BB The basic block surrounded by the loop.
402 /// @return The outermost loop in the region, NULL if such a loop does not
403 /// exist or if the region describes the whole function.
404 Loop *outermostLoopInRegion(LoopInfo *LI, BasicBlock* BB) const;
406 /// @brief Get the subregion that starts at a BasicBlock
408 /// @param BB The BasicBlock the subregion should start.
409 /// @return The Subregion if available, otherwise NULL.
410 Region* getSubRegionNode(BasicBlock *BB) const;
412 /// @brief Get the RegionNode for a BasicBlock
414 /// @param BB The BasicBlock at which the RegionNode should start.
415 /// @return If available, the RegionNode that represents the subregion
416 /// starting at BB. If no subregion starts at BB, the RegionNode
418 RegionNode* getNode(BasicBlock *BB) const;
420 /// @brief Get the BasicBlock RegionNode for a BasicBlock
422 /// @param BB The BasicBlock for which the RegionNode is requested.
423 /// @return The RegionNode representing the BB.
424 RegionNode* getBBNode(BasicBlock *BB) const;
426 /// @brief Add a new subregion to this Region.
428 /// @param SubRegion The new subregion that will be added.
429 /// @param moveChildren Move the children of this region, that are also
430 /// contained in SubRegion into SubRegion.
431 void addSubRegion(Region *SubRegion, bool moveChildren = false);
433 /// @brief Remove a subregion from this Region.
435 /// The subregion is not deleted, as it will probably be inserted into another
437 /// @param SubRegion The SubRegion that will be removed.
438 Region *removeSubRegion(Region *SubRegion);
440 /// @brief Move all direct child nodes of this Region to another Region.
442 /// @param To The Region the child nodes will be transferred to.
443 void transferChildrenTo(Region *To);
445 /// @brief Verify if the region is a correct region.
447 /// Check if this is a correctly build Region. This is an expensive check, as
448 /// the complete CFG of the Region will be walked.
449 void verifyRegion() const;
451 /// @brief Clear the cache for BB RegionNodes.
453 /// After calling this function the BasicBlock RegionNodes will be stored at
454 /// different memory locations. RegionNodes obtained before this function is
455 /// called are therefore not comparable to RegionNodes abtained afterwords.
456 void clearNodeCache();
458 /// @name Subregion Iterators
460 /// These iterators iterator over all subregions of this Region.
462 typedef RegionSet::iterator iterator;
463 typedef RegionSet::const_iterator const_iterator;
465 iterator begin() { return children.begin(); }
466 iterator end() { return children.end(); }
468 const_iterator begin() const { return children.begin(); }
469 const_iterator end() const { return children.end(); }
472 /// @name BasicBlock Iterators
474 /// These iterators iterate over all BasicBlocks that are contained in this
475 /// Region. The iterator also iterates over BasicBlocks that are elements of
476 /// a subregion of this Region. It is therefore called a flat iterator.
478 template <bool IsConst>
479 class block_iterator_wrapper
480 : public df_iterator<typename conditional<IsConst,
482 BasicBlock>::type*> {
483 typedef df_iterator<typename conditional<IsConst,
488 typedef block_iterator_wrapper<IsConst> Self;
489 typedef typename super::pointer pointer;
491 // Construct the begin iterator.
492 block_iterator_wrapper(pointer Entry, pointer Exit) : super(df_begin(Entry))
494 // Mark the exit of the region as visited, so that the children of the
495 // exit and the exit itself, i.e. the block outside the region will never
497 super::Visited.insert(Exit);
500 // Construct the end iterator.
501 block_iterator_wrapper() : super(df_end<pointer>((BasicBlock *)0)) {}
503 /*implicit*/ block_iterator_wrapper(super I) : super(I) {}
505 // FIXME: Even a const_iterator returns a non-const BasicBlock pointer.
506 // This was introduced for backwards compatibility, but should
507 // be removed as soon as all users are fixed.
508 BasicBlock *operator*() const {
509 return const_cast<BasicBlock*>(super::operator*());
513 typedef block_iterator_wrapper<false> block_iterator;
514 typedef block_iterator_wrapper<true> const_block_iterator;
516 block_iterator block_begin() {
517 return block_iterator(getEntry(), getExit());
520 block_iterator block_end() {
521 return block_iterator();
524 const_block_iterator block_begin() const {
525 return const_block_iterator(getEntry(), getExit());
527 const_block_iterator block_end() const {
528 return const_block_iterator();
532 /// @name Element Iterators
534 /// These iterators iterate over all BasicBlock and subregion RegionNodes that
535 /// are direct children of this Region. It does not iterate over any
536 /// RegionNodes that are also element of a subregion of this Region.
538 typedef df_iterator<RegionNode*, SmallPtrSet<RegionNode*, 8>, false,
539 GraphTraits<RegionNode*> > element_iterator;
541 typedef df_iterator<const RegionNode*, SmallPtrSet<const RegionNode*, 8>,
542 false, GraphTraits<const RegionNode*> >
543 const_element_iterator;
545 element_iterator element_begin();
546 element_iterator element_end();
548 const_element_iterator element_begin() const;
549 const_element_iterator element_end() const;
553 //===----------------------------------------------------------------------===//
554 /// @brief Analysis that detects all canonical Regions.
556 /// The RegionInfo pass detects all canonical regions in a function. The Regions
557 /// are connected using the parent relation. This builds a Program Structure
559 class RegionInfo : public FunctionPass {
560 typedef DenseMap<BasicBlock*,BasicBlock*> BBtoBBMap;
561 typedef DenseMap<BasicBlock*, Region*> BBtoRegionMap;
562 typedef SmallPtrSet<Region*, 4> RegionSet;
564 RegionInfo(const RegionInfo &) LLVM_DELETED_FUNCTION;
565 const RegionInfo &operator=(const RegionInfo &) LLVM_DELETED_FUNCTION;
568 PostDominatorTree *PDT;
569 DominanceFrontier *DF;
571 /// The top level region.
572 Region *TopLevelRegion;
574 /// Map every BB to the smallest region, that contains BB.
575 BBtoRegionMap BBtoRegion;
577 // isCommonDomFrontier - Returns true if BB is in the dominance frontier of
578 // entry, because it was inherited from exit. In the other case there is an
579 // edge going from entry to BB without passing exit.
580 bool isCommonDomFrontier(BasicBlock* BB, BasicBlock* entry,
581 BasicBlock* exit) const;
583 // isRegion - Check if entry and exit surround a valid region, based on
584 // dominance tree and dominance frontier.
585 bool isRegion(BasicBlock* entry, BasicBlock* exit) const;
587 // insertShortCut - Saves a shortcut pointing from entry to exit.
588 // This function may extend this shortcut if possible.
589 void insertShortCut(BasicBlock* entry, BasicBlock* exit,
590 BBtoBBMap* ShortCut) const;
592 // getNextPostDom - Returns the next BB that postdominates N, while skipping
593 // all post dominators that cannot finish a canonical region.
594 DomTreeNode *getNextPostDom(DomTreeNode* N, BBtoBBMap *ShortCut) const;
596 // isTrivialRegion - A region is trivial, if it contains only one BB.
597 bool isTrivialRegion(BasicBlock *entry, BasicBlock *exit) const;
599 // createRegion - Creates a single entry single exit region.
600 Region *createRegion(BasicBlock *entry, BasicBlock *exit);
602 // findRegionsWithEntry - Detect all regions starting with bb 'entry'.
603 void findRegionsWithEntry(BasicBlock *entry, BBtoBBMap *ShortCut);
605 // scanForRegions - Detects regions in F.
606 void scanForRegions(Function &F, BBtoBBMap *ShortCut);
608 // getTopMostParent - Get the top most parent with the same entry block.
609 Region *getTopMostParent(Region *region);
611 // buildRegionsTree - build the region hierarchy after all region detected.
612 void buildRegionsTree(DomTreeNode *N, Region *region);
614 // Calculate - detecte all regions in function and build the region tree.
615 void Calculate(Function& F);
617 void releaseMemory();
619 // updateStatistics - Update statistic about created regions.
620 void updateStatistics(Region *R);
622 // isSimple - Check if a region is a simple region with exactly one entry
623 // edge and exactly one exit edge.
624 bool isSimple(Region* R) const;
628 explicit RegionInfo();
632 /// @name FunctionPass interface
634 virtual bool runOnFunction(Function &F);
635 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
636 virtual void print(raw_ostream &OS, const Module *) const;
637 virtual void verifyAnalysis() const;
640 /// @brief Get the smallest region that contains a BasicBlock.
642 /// @param BB The basic block.
643 /// @return The smallest region, that contains BB or NULL, if there is no
644 /// region containing BB.
645 Region *getRegionFor(BasicBlock *BB) const;
647 /// @brief Set the smallest region that surrounds a basic block.
649 /// @param BB The basic block surrounded by a region.
650 /// @param R The smallest region that surrounds BB.
651 void setRegionFor(BasicBlock *BB, Region *R);
653 /// @brief A shortcut for getRegionFor().
655 /// @param BB The basic block.
656 /// @return The smallest region, that contains BB or NULL, if there is no
657 /// region containing BB.
658 Region *operator[](BasicBlock *BB) const;
660 /// @brief Return the exit of the maximal refined region, that starts at a
663 /// @param BB The BasicBlock the refined region starts.
664 BasicBlock *getMaxRegionExit(BasicBlock *BB) const;
666 /// @brief Find the smallest region that contains two regions.
668 /// @param A The first region.
669 /// @param B The second region.
670 /// @return The smallest region containing A and B.
671 Region *getCommonRegion(Region* A, Region *B) const;
673 /// @brief Find the smallest region that contains two basic blocks.
675 /// @param A The first basic block.
676 /// @param B The second basic block.
677 /// @return The smallest region that contains A and B.
678 Region* getCommonRegion(BasicBlock* A, BasicBlock *B) const {
679 return getCommonRegion(getRegionFor(A), getRegionFor(B));
682 /// @brief Find the smallest region that contains a set of regions.
684 /// @param Regions A vector of regions.
685 /// @return The smallest region that contains all regions in Regions.
686 Region* getCommonRegion(SmallVectorImpl<Region*> &Regions) const;
688 /// @brief Find the smallest region that contains a set of basic blocks.
690 /// @param BBs A vector of basic blocks.
691 /// @return The smallest region that contains all basic blocks in BBS.
692 Region* getCommonRegion(SmallVectorImpl<BasicBlock*> &BBs) const;
694 Region *getTopLevelRegion() const {
695 return TopLevelRegion;
698 /// @brief Update RegionInfo after a basic block was split.
700 /// @param NewBB The basic block that was created before OldBB.
701 /// @param OldBB The old basic block.
702 void splitBlock(BasicBlock* NewBB, BasicBlock *OldBB);
704 /// @brief Clear the Node Cache for all Regions.
706 /// @see Region::clearNodeCache()
707 void clearNodeCache() {
709 TopLevelRegion->clearNodeCache();
713 inline raw_ostream &operator<<(raw_ostream &OS, const RegionNode &Node) {
714 if (Node.isSubRegion())
715 return OS << Node.getNodeAs<Region>()->getNameStr();
717 return OS << Node.getNodeAs<BasicBlock>()->getName();
719 } // End llvm namespace