1 //===---- MachineOutliner.cpp - Outline instructions -----------*- 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 //===----------------------------------------------------------------------===//
11 /// Replaces repeated sequences of instructions with function calls.
13 /// This works by placing every instruction from every basic block in a
14 /// suffix tree, and repeatedly querying that tree for repeated sequences of
15 /// instructions. If a sequence of instructions appears often, then it ought
16 /// to be beneficial to pull out into a function.
18 /// The MachineOutliner communicates with a given target using hooks defined in
19 /// TargetInstrInfo.h. The target supplies the outliner with information on how
20 /// a specific sequence of instructions should be outlined. This information
21 /// is used to deduce the number of instructions necessary to
23 /// * Create an outlined function
24 /// * Call that outlined function
26 /// Targets must implement
27 /// * getOutliningCandidateInfo
28 /// * insertOutlinerEpilogue
29 /// * insertOutlinedCall
30 /// * insertOutlinerPrologue
31 /// * isFunctionSafeToOutlineFrom
33 /// in order to make use of the MachineOutliner.
35 /// This was originally presented at the 2016 LLVM Developers' Meeting in the
36 /// talk "Reducing Code Size Using Outlining". For a high-level overview of
37 /// how this pass works, the talk is available on YouTube at
39 /// https://www.youtube.com/watch?v=yorld-WSOeU
41 /// The slides for the talk are available at
43 /// http://www.llvm.org/devmtg/2016-11/Slides/Paquette-Outliner.pdf
45 /// The talk provides an overview of how the outliner finds candidates and
46 /// ultimately outlines them. It describes how the main data structure for this
47 /// pass, the suffix tree, is queried and purged for candidates. It also gives
48 /// a simplified suffix tree construction algorithm for suffix trees based off
49 /// of the algorithm actually used here, Ukkonen's algorithm.
51 /// For the original RFC for this pass, please see
53 /// http://lists.llvm.org/pipermail/llvm-dev/2016-August/104170.html
55 /// For more information on the suffix tree data structure, please see
56 /// https://www.cs.helsinki.fi/u/ukkonen/SuffixT1withFigs.pdf
58 //===----------------------------------------------------------------------===//
59 #include "llvm/ADT/DenseMap.h"
60 #include "llvm/ADT/Statistic.h"
61 #include "llvm/ADT/Twine.h"
62 #include "llvm/CodeGen/MachineFunction.h"
63 #include "llvm/CodeGen/MachineModuleInfo.h"
64 #include "llvm/CodeGen/MachineOptimizationRemarkEmitter.h"
65 #include "llvm/CodeGen/Passes.h"
66 #include "llvm/CodeGen/TargetInstrInfo.h"
67 #include "llvm/CodeGen/TargetRegisterInfo.h"
68 #include "llvm/CodeGen/TargetSubtargetInfo.h"
69 #include "llvm/IR/IRBuilder.h"
70 #include "llvm/Support/Allocator.h"
71 #include "llvm/Support/Debug.h"
72 #include "llvm/Support/raw_ostream.h"
79 #define DEBUG_TYPE "machine-outliner"
84 STATISTIC(NumOutlined, "Number of candidates outlined");
85 STATISTIC(FunctionsCreated, "Number of functions created");
89 /// \brief An individual sequence of instructions to be replaced with a call to
90 /// an outlined function.
93 /// The start index of this \p Candidate in the instruction list.
96 /// The number of instructions in this \p Candidate.
100 /// Set to false if the candidate overlapped with another candidate.
101 bool InCandidateList = true;
103 /// \brief The index of this \p Candidate's \p OutlinedFunction in the list of
104 /// \p OutlinedFunctions.
105 unsigned FunctionIdx;
107 /// Contains all target-specific information for this \p Candidate.
108 TargetInstrInfo::MachineOutlinerInfo MInfo;
110 /// Return the number of instructions in this Candidate.
111 unsigned getLength() const { return Len; }
113 /// Return the start index of this candidate.
114 unsigned getStartIdx() const { return StartIdx; }
116 // Return the end index of this candidate.
117 unsigned getEndIdx() const { return StartIdx + Len - 1; }
119 /// \brief The number of instructions that would be saved by outlining every
120 /// candidate of this type.
122 /// This is a fixed value which is not updated during the candidate pruning
123 /// process. It is only used for deciding which candidate to keep if two
124 /// candidates overlap. The true benefit is stored in the OutlinedFunction
125 /// for some given candidate.
126 unsigned Benefit = 0;
128 Candidate(unsigned StartIdx, unsigned Len, unsigned FunctionIdx)
129 : StartIdx(StartIdx), Len(Len), FunctionIdx(FunctionIdx) {}
133 /// \brief Used to ensure that \p Candidates are outlined in an order that
134 /// preserves the start and end indices of other \p Candidates.
135 bool operator<(const Candidate &RHS) const {
136 return getStartIdx() > RHS.getStartIdx();
140 /// \brief The information necessary to create an outlined function for some
141 /// class of candidate.
142 struct OutlinedFunction {
145 /// The number of candidates for this \p OutlinedFunction.
146 unsigned OccurrenceCount = 0;
149 std::vector<std::shared_ptr<Candidate>> Candidates;
151 /// The actual outlined function created.
152 /// This is initialized after we go through and create the actual function.
153 MachineFunction *MF = nullptr;
155 /// A number assigned to this function which appears at the end of its name.
158 /// \brief The sequence of integers corresponding to the instructions in this
160 std::vector<unsigned> Sequence;
162 /// Contains all target-specific information for this \p OutlinedFunction.
163 TargetInstrInfo::MachineOutlinerInfo MInfo;
165 /// Return the number of candidates for this \p OutlinedFunction.
166 unsigned getOccurrenceCount() { return OccurrenceCount; }
168 /// Decrement the occurrence count of this OutlinedFunction and return the
170 unsigned decrement() {
171 assert(OccurrenceCount > 0 && "Can't decrement an empty function!");
173 return getOccurrenceCount();
176 /// \brief Return the number of instructions it would take to outline this
178 unsigned getOutliningCost() {
179 return (OccurrenceCount * MInfo.CallOverhead) + Sequence.size() +
183 /// \brief Return the number of instructions that would be saved by outlining
185 unsigned getBenefit() {
186 unsigned NotOutlinedCost = OccurrenceCount * Sequence.size();
187 unsigned OutlinedCost = getOutliningCost();
188 return (NotOutlinedCost < OutlinedCost) ? 0
189 : NotOutlinedCost - OutlinedCost;
192 OutlinedFunction(unsigned Name, unsigned OccurrenceCount,
193 const std::vector<unsigned> &Sequence,
194 TargetInstrInfo::MachineOutlinerInfo &MInfo)
195 : OccurrenceCount(OccurrenceCount), Name(Name), Sequence(Sequence),
199 /// Represents an undefined index in the suffix tree.
200 const unsigned EmptyIdx = -1;
202 /// A node in a suffix tree which represents a substring or suffix.
204 /// Each node has either no children or at least two children, with the root
205 /// being a exception in the empty tree.
207 /// Children are represented as a map between unsigned integers and nodes. If
208 /// a node N has a child M on unsigned integer k, then the mapping represented
209 /// by N is a proper prefix of the mapping represented by M. Note that this,
210 /// although similar to a trie is somewhat different: each node stores a full
211 /// substring of the full mapping rather than a single character state.
213 /// Each internal node contains a pointer to the internal node representing
214 /// the same string, but with the first character chopped off. This is stored
215 /// in \p Link. Each leaf node stores the start index of its respective
216 /// suffix in \p SuffixIdx.
217 struct SuffixTreeNode {
219 /// The children of this node.
221 /// A child existing on an unsigned integer implies that from the mapping
222 /// represented by the current node, there is a way to reach another
223 /// mapping by tacking that character on the end of the current string.
224 DenseMap<unsigned, SuffixTreeNode *> Children;
226 /// A flag set to false if the node has been pruned from the tree.
227 bool IsInTree = true;
229 /// The start index of this node's substring in the main string.
230 unsigned StartIdx = EmptyIdx;
232 /// The end index of this node's substring in the main string.
234 /// Every leaf node must have its \p EndIdx incremented at the end of every
235 /// step in the construction algorithm. To avoid having to update O(N)
236 /// nodes individually at the end of every step, the end index is stored
238 unsigned *EndIdx = nullptr;
240 /// For leaves, the start index of the suffix represented by this node.
242 /// For all other nodes, this is ignored.
243 unsigned SuffixIdx = EmptyIdx;
245 /// \brief For internal nodes, a pointer to the internal node representing
246 /// the same sequence with the first character chopped off.
248 /// This acts as a shortcut in Ukkonen's algorithm. One of the things that
249 /// Ukkonen's algorithm does to achieve linear-time construction is
250 /// keep track of which node the next insert should be at. This makes each
251 /// insert O(1), and there are a total of O(N) inserts. The suffix link
252 /// helps with inserting children of internal nodes.
254 /// Say we add a child to an internal node with associated mapping S. The
255 /// next insertion must be at the node representing S - its first character.
256 /// This is given by the way that we iteratively build the tree in Ukkonen's
257 /// algorithm. The main idea is to look at the suffixes of each prefix in the
258 /// string, starting with the longest suffix of the prefix, and ending with
259 /// the shortest. Therefore, if we keep pointers between such nodes, we can
260 /// move to the next insertion point in O(1) time. If we don't, then we'd
261 /// have to query from the root, which takes O(N) time. This would make the
262 /// construction algorithm O(N^2) rather than O(N).
263 SuffixTreeNode *Link = nullptr;
265 /// The parent of this node. Every node except for the root has a parent.
266 SuffixTreeNode *Parent = nullptr;
268 /// The number of times this node's string appears in the tree.
270 /// This is equal to the number of leaf children of the string. It represents
271 /// the number of suffixes that the node's string is a prefix of.
272 unsigned OccurrenceCount = 0;
274 /// The length of the string formed by concatenating the edge labels from the
275 /// root to this node.
276 unsigned ConcatLen = 0;
278 /// Returns true if this node is a leaf.
279 bool isLeaf() const { return SuffixIdx != EmptyIdx; }
281 /// Returns true if this node is the root of its owning \p SuffixTree.
282 bool isRoot() const { return StartIdx == EmptyIdx; }
284 /// Return the number of elements in the substring associated with this node.
285 size_t size() const {
287 // Is it the root? If so, it's the empty string so return 0.
291 assert(*EndIdx != EmptyIdx && "EndIdx is undefined!");
293 // Size = the number of elements in the string.
294 // For example, [0 1 2 3] has length 4, not 3. 3-0 = 3, so we have 3-0+1.
295 return *EndIdx - StartIdx + 1;
298 SuffixTreeNode(unsigned StartIdx, unsigned *EndIdx, SuffixTreeNode *Link,
299 SuffixTreeNode *Parent)
300 : StartIdx(StartIdx), EndIdx(EndIdx), Link(Link), Parent(Parent) {}
305 /// A data structure for fast substring queries.
307 /// Suffix trees represent the suffixes of their input strings in their leaves.
308 /// A suffix tree is a type of compressed trie structure where each node
309 /// represents an entire substring rather than a single character. Each leaf
310 /// of the tree is a suffix.
312 /// A suffix tree can be seen as a type of state machine where each state is a
313 /// substring of the full string. The tree is structured so that, for a string
314 /// of length N, there are exactly N leaves in the tree. This structure allows
315 /// us to quickly find repeated substrings of the input string.
317 /// In this implementation, a "string" is a vector of unsigned integers.
318 /// These integers may result from hashing some data type. A suffix tree can
319 /// contain 1 or many strings, which can then be queried as one large string.
321 /// The suffix tree is implemented using Ukkonen's algorithm for linear-time
322 /// suffix tree construction. Ukkonen's algorithm is explained in more detail
323 /// in the paper by Esko Ukkonen "On-line construction of suffix trees. The
324 /// paper is available at
326 /// https://www.cs.helsinki.fi/u/ukkonen/SuffixT1withFigs.pdf
329 /// Stores each leaf node in the tree.
331 /// This is used for finding outlining candidates.
332 std::vector<SuffixTreeNode *> LeafVector;
334 /// Each element is an integer representing an instruction in the module.
335 ArrayRef<unsigned> Str;
338 /// Maintains each node in the tree.
339 SpecificBumpPtrAllocator<SuffixTreeNode> NodeAllocator;
341 /// The root of the suffix tree.
343 /// The root represents the empty string. It is maintained by the
344 /// \p NodeAllocator like every other node in the tree.
345 SuffixTreeNode *Root = nullptr;
347 /// Maintains the end indices of the internal nodes in the tree.
349 /// Each internal node is guaranteed to never have its end index change
350 /// during the construction algorithm; however, leaves must be updated at
351 /// every step. Therefore, we need to store leaf end indices by reference
352 /// to avoid updating O(N) leaves at every step of construction. Thus,
353 /// every internal node must be allocated its own end index.
354 BumpPtrAllocator InternalEndIdxAllocator;
356 /// The end index of each leaf in the tree.
357 unsigned LeafEndIdx = -1;
359 /// \brief Helper struct which keeps track of the next insertion point in
360 /// Ukkonen's algorithm.
362 /// The next node to insert at.
363 SuffixTreeNode *Node;
365 /// The index of the first character in the substring currently being added.
366 unsigned Idx = EmptyIdx;
368 /// The length of the substring we have to add at the current step.
372 /// \brief The point the next insertion will take place at in the
373 /// construction algorithm.
376 /// Allocate a leaf node and add it to the tree.
378 /// \param Parent The parent of this node.
379 /// \param StartIdx The start index of this node's associated string.
380 /// \param Edge The label on the edge leaving \p Parent to this node.
382 /// \returns A pointer to the allocated leaf node.
383 SuffixTreeNode *insertLeaf(SuffixTreeNode &Parent, unsigned StartIdx,
386 assert(StartIdx <= LeafEndIdx && "String can't start after it ends!");
388 SuffixTreeNode *N = new (NodeAllocator.Allocate())
389 SuffixTreeNode(StartIdx, &LeafEndIdx, nullptr, &Parent);
390 Parent.Children[Edge] = N;
395 /// Allocate an internal node and add it to the tree.
397 /// \param Parent The parent of this node. Only null when allocating the root.
398 /// \param StartIdx The start index of this node's associated string.
399 /// \param EndIdx The end index of this node's associated string.
400 /// \param Edge The label on the edge leaving \p Parent to this node.
402 /// \returns A pointer to the allocated internal node.
403 SuffixTreeNode *insertInternalNode(SuffixTreeNode *Parent, unsigned StartIdx,
404 unsigned EndIdx, unsigned Edge) {
406 assert(StartIdx <= EndIdx && "String can't start after it ends!");
407 assert(!(!Parent && StartIdx != EmptyIdx) &&
408 "Non-root internal nodes must have parents!");
410 unsigned *E = new (InternalEndIdxAllocator) unsigned(EndIdx);
411 SuffixTreeNode *N = new (NodeAllocator.Allocate())
412 SuffixTreeNode(StartIdx, E, Root, Parent);
414 Parent->Children[Edge] = N;
419 /// \brief Set the suffix indices of the leaves to the start indices of their
420 /// respective suffixes. Also stores each leaf in \p LeafVector at its
421 /// respective suffix index.
423 /// \param[in] CurrNode The node currently being visited.
424 /// \param CurrIdx The current index of the string being visited.
425 void setSuffixIndices(SuffixTreeNode &CurrNode, unsigned CurrIdx) {
427 bool IsLeaf = CurrNode.Children.size() == 0 && !CurrNode.isRoot();
429 // Store the length of the concatenation of all strings from the root to
431 if (!CurrNode.isRoot()) {
432 if (CurrNode.ConcatLen == 0)
433 CurrNode.ConcatLen = CurrNode.size();
436 CurrNode.ConcatLen += CurrNode.Parent->ConcatLen;
439 // Traverse the tree depth-first.
440 for (auto &ChildPair : CurrNode.Children) {
441 assert(ChildPair.second && "Node had a null child!");
442 setSuffixIndices(*ChildPair.second, CurrIdx + ChildPair.second->size());
445 // Is this node a leaf?
447 // If yes, give it a suffix index and bump its parent's occurrence count.
448 CurrNode.SuffixIdx = Str.size() - CurrIdx;
449 assert(CurrNode.Parent && "CurrNode had no parent!");
450 CurrNode.Parent->OccurrenceCount++;
452 // Store the leaf in the leaf vector for pruning later.
453 LeafVector[CurrNode.SuffixIdx] = &CurrNode;
457 /// \brief Construct the suffix tree for the prefix of the input ending at
460 /// Used to construct the full suffix tree iteratively. At the end of each
461 /// step, the constructed suffix tree is either a valid suffix tree, or a
462 /// suffix tree with implicit suffixes. At the end of the final step, the
463 /// suffix tree is a valid tree.
465 /// \param EndIdx The end index of the current prefix in the main string.
466 /// \param SuffixesToAdd The number of suffixes that must be added
467 /// to complete the suffix tree at the current phase.
469 /// \returns The number of suffixes that have not been added at the end of
471 unsigned extend(unsigned EndIdx, unsigned SuffixesToAdd) {
472 SuffixTreeNode *NeedsLink = nullptr;
474 while (SuffixesToAdd > 0) {
476 // Are we waiting to add anything other than just the last character?
477 if (Active.Len == 0) {
478 // If not, then say the active index is the end index.
482 assert(Active.Idx <= EndIdx && "Start index can't be after end index!");
484 // The first character in the current substring we're looking at.
485 unsigned FirstChar = Str[Active.Idx];
487 // Have we inserted anything starting with FirstChar at the current node?
488 if (Active.Node->Children.count(FirstChar) == 0) {
489 // If not, then we can just insert a leaf and move too the next step.
490 insertLeaf(*Active.Node, EndIdx, FirstChar);
492 // The active node is an internal node, and we visited it, so it must
493 // need a link if it doesn't have one.
495 NeedsLink->Link = Active.Node;
499 // There's a match with FirstChar, so look for the point in the tree to
500 // insert a new node.
501 SuffixTreeNode *NextNode = Active.Node->Children[FirstChar];
503 unsigned SubstringLen = NextNode->size();
505 // Is the current suffix we're trying to insert longer than the size of
506 // the child we want to move to?
507 if (Active.Len >= SubstringLen) {
508 // If yes, then consume the characters we've seen and move to the next
510 Active.Idx += SubstringLen;
511 Active.Len -= SubstringLen;
512 Active.Node = NextNode;
516 // Otherwise, the suffix we're trying to insert must be contained in the
517 // next node we want to move to.
518 unsigned LastChar = Str[EndIdx];
520 // Is the string we're trying to insert a substring of the next node?
521 if (Str[NextNode->StartIdx + Active.Len] == LastChar) {
522 // If yes, then we're done for this step. Remember our insertion point
523 // and move to the next end index. At this point, we have an implicit
525 if (NeedsLink && !Active.Node->isRoot()) {
526 NeedsLink->Link = Active.Node;
534 // The string we're trying to insert isn't a substring of the next node,
535 // but matches up to a point. Split the node.
537 // For example, say we ended our search at a node n and we're trying to
538 // insert ABD. Then we'll create a new node s for AB, reduce n to just
539 // representing C, and insert a new leaf node l to represent d. This
540 // allows us to ensure that if n was a leaf, it remains a leaf.
542 // | ABC ---split---> | AB
547 // The node s from the diagram
548 SuffixTreeNode *SplitNode =
549 insertInternalNode(Active.Node, NextNode->StartIdx,
550 NextNode->StartIdx + Active.Len - 1, FirstChar);
552 // Insert the new node representing the new substring into the tree as
553 // a child of the split node. This is the node l from the diagram.
554 insertLeaf(*SplitNode, EndIdx, LastChar);
556 // Make the old node a child of the split node and update its start
557 // index. This is the node n from the diagram.
558 NextNode->StartIdx += Active.Len;
559 NextNode->Parent = SplitNode;
560 SplitNode->Children[Str[NextNode->StartIdx]] = NextNode;
562 // SplitNode is an internal node, update the suffix link.
564 NeedsLink->Link = SplitNode;
566 NeedsLink = SplitNode;
569 // We've added something new to the tree, so there's one less suffix to
573 if (Active.Node->isRoot()) {
574 if (Active.Len > 0) {
576 Active.Idx = EndIdx - SuffixesToAdd + 1;
579 // Start the next phase at the next smallest suffix.
580 Active.Node = Active.Node->Link;
584 return SuffixesToAdd;
588 /// Construct a suffix tree from a sequence of unsigned integers.
590 /// \param Str The string to construct the suffix tree for.
591 SuffixTree(const std::vector<unsigned> &Str) : Str(Str) {
592 Root = insertInternalNode(nullptr, EmptyIdx, EmptyIdx, 0);
593 Root->IsInTree = true;
595 LeafVector = std::vector<SuffixTreeNode *>(Str.size());
597 // Keep track of the number of suffixes we have to add of the current
599 unsigned SuffixesToAdd = 0;
602 // Construct the suffix tree iteratively on each prefix of the string.
603 // PfxEndIdx is the end index of the current prefix.
604 // End is one past the last element in the string.
605 for (unsigned PfxEndIdx = 0, End = Str.size(); PfxEndIdx < End;
608 LeafEndIdx = PfxEndIdx; // Extend each of the leaves.
609 SuffixesToAdd = extend(PfxEndIdx, SuffixesToAdd);
612 // Set the suffix indices of each leaf.
613 assert(Root && "Root node can't be nullptr!");
614 setSuffixIndices(*Root, 0);
618 /// \brief Maps \p MachineInstrs to unsigned integers and stores the mappings.
619 struct InstructionMapper {
621 /// \brief The next available integer to assign to a \p MachineInstr that
622 /// cannot be outlined.
624 /// Set to -3 for compatability with \p DenseMapInfo<unsigned>.
625 unsigned IllegalInstrNumber = -3;
627 /// \brief The next available integer to assign to a \p MachineInstr that can
629 unsigned LegalInstrNumber = 0;
631 /// Correspondence from \p MachineInstrs to unsigned integers.
632 DenseMap<MachineInstr *, unsigned, MachineInstrExpressionTrait>
633 InstructionIntegerMap;
635 /// Corresponcence from unsigned integers to \p MachineInstrs.
636 /// Inverse of \p InstructionIntegerMap.
637 DenseMap<unsigned, MachineInstr *> IntegerInstructionMap;
639 /// The vector of unsigned integers that the module is mapped to.
640 std::vector<unsigned> UnsignedVec;
642 /// \brief Stores the location of the instruction associated with the integer
643 /// at index i in \p UnsignedVec for each index i.
644 std::vector<MachineBasicBlock::iterator> InstrList;
646 /// \brief Maps \p *It to a legal integer.
648 /// Updates \p InstrList, \p UnsignedVec, \p InstructionIntegerMap,
649 /// \p IntegerInstructionMap, and \p LegalInstrNumber.
651 /// \returns The integer that \p *It was mapped to.
652 unsigned mapToLegalUnsigned(MachineBasicBlock::iterator &It) {
654 // Get the integer for this instruction or give it the current
656 InstrList.push_back(It);
657 MachineInstr &MI = *It;
659 DenseMap<MachineInstr *, unsigned, MachineInstrExpressionTrait>::iterator
661 std::tie(ResultIt, WasInserted) =
662 InstructionIntegerMap.insert(std::make_pair(&MI, LegalInstrNumber));
663 unsigned MINumber = ResultIt->second;
665 // There was an insertion.
668 IntegerInstructionMap.insert(std::make_pair(MINumber, &MI));
671 UnsignedVec.push_back(MINumber);
673 // Make sure we don't overflow or use any integers reserved by the DenseMap.
674 if (LegalInstrNumber >= IllegalInstrNumber)
675 report_fatal_error("Instruction mapping overflow!");
677 assert(LegalInstrNumber != DenseMapInfo<unsigned>::getEmptyKey() &&
678 "Tried to assign DenseMap tombstone or empty key to instruction.");
679 assert(LegalInstrNumber != DenseMapInfo<unsigned>::getTombstoneKey() &&
680 "Tried to assign DenseMap tombstone or empty key to instruction.");
685 /// Maps \p *It to an illegal integer.
687 /// Updates \p InstrList, \p UnsignedVec, and \p IllegalInstrNumber.
689 /// \returns The integer that \p *It was mapped to.
690 unsigned mapToIllegalUnsigned(MachineBasicBlock::iterator &It) {
691 unsigned MINumber = IllegalInstrNumber;
693 InstrList.push_back(It);
694 UnsignedVec.push_back(IllegalInstrNumber);
695 IllegalInstrNumber--;
697 assert(LegalInstrNumber < IllegalInstrNumber &&
698 "Instruction mapping overflow!");
700 assert(IllegalInstrNumber != DenseMapInfo<unsigned>::getEmptyKey() &&
701 "IllegalInstrNumber cannot be DenseMap tombstone or empty key!");
703 assert(IllegalInstrNumber != DenseMapInfo<unsigned>::getTombstoneKey() &&
704 "IllegalInstrNumber cannot be DenseMap tombstone or empty key!");
709 /// \brief Transforms a \p MachineBasicBlock into a \p vector of \p unsigneds
710 /// and appends it to \p UnsignedVec and \p InstrList.
712 /// Two instructions are assigned the same integer if they are identical.
713 /// If an instruction is deemed unsafe to outline, then it will be assigned an
714 /// unique integer. The resulting mapping is placed into a suffix tree and
715 /// queried for candidates.
717 /// \param MBB The \p MachineBasicBlock to be translated into integers.
718 /// \param TRI \p TargetRegisterInfo for the module.
719 /// \param TII \p TargetInstrInfo for the module.
720 void convertToUnsignedVec(MachineBasicBlock &MBB,
721 const TargetRegisterInfo &TRI,
722 const TargetInstrInfo &TII) {
723 for (MachineBasicBlock::iterator It = MBB.begin(), Et = MBB.end(); It != Et;
726 // Keep track of where this instruction is in the module.
727 switch (TII.getOutliningType(*It)) {
728 case TargetInstrInfo::MachineOutlinerInstrType::Illegal:
729 mapToIllegalUnsigned(It);
732 case TargetInstrInfo::MachineOutlinerInstrType::Legal:
733 mapToLegalUnsigned(It);
736 case TargetInstrInfo::MachineOutlinerInstrType::Invisible:
741 // After we're done every insertion, uniquely terminate this part of the
742 // "string". This makes sure we won't match across basic block or function
743 // boundaries since the "end" is encoded uniquely and thus appears in no
744 // repeated substring.
745 InstrList.push_back(MBB.end());
746 UnsignedVec.push_back(IllegalInstrNumber);
747 IllegalInstrNumber--;
750 InstructionMapper() {
751 // Make sure that the implementation of DenseMapInfo<unsigned> hasn't
753 assert(DenseMapInfo<unsigned>::getEmptyKey() == (unsigned)-1 &&
754 "DenseMapInfo<unsigned>'s empty key isn't -1!");
755 assert(DenseMapInfo<unsigned>::getTombstoneKey() == (unsigned)-2 &&
756 "DenseMapInfo<unsigned>'s tombstone key isn't -2!");
760 /// \brief An interprocedural pass which finds repeated sequences of
761 /// instructions and replaces them with calls to functions.
763 /// Each instruction is mapped to an unsigned integer and placed in a string.
764 /// The resulting mapping is then placed in a \p SuffixTree. The \p SuffixTree
765 /// is then repeatedly queried for repeated sequences of instructions. Each
766 /// non-overlapping repeated sequence is then placed in its own
767 /// \p MachineFunction and each instance is then replaced with a call to that
769 struct MachineOutliner : public ModulePass {
773 /// \brief Set to true if the outliner should consider functions with
774 /// linkonceodr linkage.
775 bool OutlineFromLinkOnceODRs = false;
777 StringRef getPassName() const override { return "Machine Outliner"; }
779 void getAnalysisUsage(AnalysisUsage &AU) const override {
780 AU.addRequired<MachineModuleInfo>();
781 AU.addPreserved<MachineModuleInfo>();
782 AU.setPreservesAll();
783 ModulePass::getAnalysisUsage(AU);
786 MachineOutliner(bool OutlineFromLinkOnceODRs = false)
787 : ModulePass(ID), OutlineFromLinkOnceODRs(OutlineFromLinkOnceODRs) {
788 initializeMachineOutlinerPass(*PassRegistry::getPassRegistry());
791 /// Find all repeated substrings that satisfy the outlining cost model.
793 /// If a substring appears at least twice, then it must be represented by
794 /// an internal node which appears in at least two suffixes. Each suffix is
795 /// represented by a leaf node. To do this, we visit each internal node in
796 /// the tree, using the leaf children of each internal node. If an internal
797 /// node represents a beneficial substring, then we use each of its leaf
798 /// children to find the locations of its substring.
800 /// \param ST A suffix tree to query.
801 /// \param TII TargetInstrInfo for the target.
802 /// \param Mapper Contains outlining mapping information.
803 /// \param[out] CandidateList Filled with candidates representing each
804 /// beneficial substring.
805 /// \param[out] FunctionList Filled with a list of \p OutlinedFunctions each
806 /// type of candidate.
808 /// \returns The length of the longest candidate found.
810 findCandidates(SuffixTree &ST, const TargetInstrInfo &TII,
811 InstructionMapper &Mapper,
812 std::vector<std::shared_ptr<Candidate>> &CandidateList,
813 std::vector<OutlinedFunction> &FunctionList);
815 /// \brief Replace the sequences of instructions represented by the
816 /// \p Candidates in \p CandidateList with calls to \p MachineFunctions
817 /// described in \p FunctionList.
819 /// \param M The module we are outlining from.
820 /// \param CandidateList A list of candidates to be outlined.
821 /// \param FunctionList A list of functions to be inserted into the module.
822 /// \param Mapper Contains the instruction mappings for the module.
823 bool outline(Module &M,
824 const ArrayRef<std::shared_ptr<Candidate>> &CandidateList,
825 std::vector<OutlinedFunction> &FunctionList,
826 InstructionMapper &Mapper);
828 /// Creates a function for \p OF and inserts it into the module.
829 MachineFunction *createOutlinedFunction(Module &M, const OutlinedFunction &OF,
830 InstructionMapper &Mapper);
832 /// Find potential outlining candidates and store them in \p CandidateList.
834 /// For each type of potential candidate, also build an \p OutlinedFunction
835 /// struct containing the information to build the function for that
838 /// \param[out] CandidateList Filled with outlining candidates for the module.
839 /// \param[out] FunctionList Filled with functions corresponding to each type
841 /// \param ST The suffix tree for the module.
842 /// \param TII TargetInstrInfo for the module.
844 /// \returns The length of the longest candidate found. 0 if there are none.
846 buildCandidateList(std::vector<std::shared_ptr<Candidate>> &CandidateList,
847 std::vector<OutlinedFunction> &FunctionList,
848 SuffixTree &ST, InstructionMapper &Mapper,
849 const TargetInstrInfo &TII);
851 /// Helper function for pruneOverlaps.
852 /// Removes \p C from the candidate list, and updates its \p OutlinedFunction.
853 void prune(Candidate &C, std::vector<OutlinedFunction> &FunctionList);
855 /// \brief Remove any overlapping candidates that weren't handled by the
856 /// suffix tree's pruning method.
858 /// Pruning from the suffix tree doesn't necessarily remove all overlaps.
859 /// If a short candidate is chosen for outlining, then a longer candidate
860 /// which has that short candidate as a suffix is chosen, the tree's pruning
861 /// method will not find it. Thus, we need to prune before outlining as well.
863 /// \param[in,out] CandidateList A list of outlining candidates.
864 /// \param[in,out] FunctionList A list of functions to be outlined.
865 /// \param Mapper Contains instruction mapping info for outlining.
866 /// \param MaxCandidateLen The length of the longest candidate.
867 /// \param TII TargetInstrInfo for the module.
868 void pruneOverlaps(std::vector<std::shared_ptr<Candidate>> &CandidateList,
869 std::vector<OutlinedFunction> &FunctionList,
870 InstructionMapper &Mapper, unsigned MaxCandidateLen,
871 const TargetInstrInfo &TII);
873 /// Construct a suffix tree on the instructions in \p M and outline repeated
874 /// strings from that tree.
875 bool runOnModule(Module &M) override;
878 } // Anonymous namespace.
880 char MachineOutliner::ID = 0;
883 ModulePass *createMachineOutlinerPass(bool OutlineFromLinkOnceODRs) {
884 return new MachineOutliner(OutlineFromLinkOnceODRs);
889 INITIALIZE_PASS(MachineOutliner, DEBUG_TYPE, "Machine Function Outliner", false,
892 unsigned MachineOutliner::findCandidates(
893 SuffixTree &ST, const TargetInstrInfo &TII, InstructionMapper &Mapper,
894 std::vector<std::shared_ptr<Candidate>> &CandidateList,
895 std::vector<OutlinedFunction> &FunctionList) {
896 CandidateList.clear();
897 FunctionList.clear();
900 // FIXME: Visit internal nodes instead of leaves.
901 for (SuffixTreeNode *Leaf : ST.LeafVector) {
902 assert(Leaf && "Leaves in LeafVector cannot be null!");
906 assert(Leaf->Parent && "All leaves must have parents!");
907 SuffixTreeNode &Parent = *(Leaf->Parent);
909 // If it doesn't appear enough, or we already outlined from it, skip it.
910 if (Parent.OccurrenceCount < 2 || Parent.isRoot() || !Parent.IsInTree)
913 // Figure out if this candidate is beneficial.
914 unsigned StringLen = Leaf->ConcatLen - (unsigned)Leaf->size();
916 // Too short to be beneficial; skip it.
917 // FIXME: This isn't necessarily true for, say, X86. If we factor in
918 // instruction lengths we need more information than this.
922 // If this is a beneficial class of candidate, then every one is stored in
924 std::vector<Candidate> CandidatesForRepeatedSeq;
926 // Describes the start and end point of each candidate. This allows the
927 // target to infer some information about each occurrence of each repeated
929 // FIXME: CandidatesForRepeatedSeq and this should be combined.
931 std::pair<MachineBasicBlock::iterator, MachineBasicBlock::iterator>>
932 RepeatedSequenceLocs;
934 // Figure out the call overhead for each instance of the sequence.
935 for (auto &ChildPair : Parent.Children) {
936 SuffixTreeNode *M = ChildPair.second;
938 if (M && M->IsInTree && M->isLeaf()) {
939 // Never visit this leaf again.
941 unsigned StartIdx = M->SuffixIdx;
942 unsigned EndIdx = StartIdx + StringLen - 1;
944 // Trick: Discard some candidates that would be incompatible with the
945 // ones we've already found for this sequence. This will save us some
946 // work in candidate selection.
948 // If two candidates overlap, then we can't outline them both. This
949 // happens when we have candidates that look like, say
951 // AA (where each "A" is an instruction).
953 // We might have some portion of the module that looks like this:
956 // In this case, there are 5 different copies of "AA" in this range, but
957 // at most 3 can be outlined. If only outlining 3 of these is going to
958 // be unbeneficial, then we ought to not bother.
960 // Note that two things DON'T overlap when they look like this:
961 // start1...end1 .... start2...end2
962 // That is, one must either
963 // * End before the other starts
964 // * Start after the other ends
965 if (std::all_of(CandidatesForRepeatedSeq.begin(),
966 CandidatesForRepeatedSeq.end(),
967 [&StartIdx, &EndIdx](const Candidate &C) {
968 return (EndIdx < C.getStartIdx() ||
969 StartIdx > C.getEndIdx());
971 // It doesn't overlap with anything, so we can outline it.
972 // Each sequence is over [StartIt, EndIt].
973 MachineBasicBlock::iterator StartIt = Mapper.InstrList[StartIdx];
974 MachineBasicBlock::iterator EndIt = Mapper.InstrList[EndIdx];
976 // Save the candidate and its location.
977 CandidatesForRepeatedSeq.emplace_back(StartIdx, StringLen,
978 FunctionList.size());
979 RepeatedSequenceLocs.emplace_back(std::make_pair(StartIt, EndIt));
984 // We've found something we might want to outline.
985 // Create an OutlinedFunction to store it and check if it'd be beneficial
987 TargetInstrInfo::MachineOutlinerInfo MInfo =
988 TII.getOutlininingCandidateInfo(RepeatedSequenceLocs);
989 std::vector<unsigned> Seq;
990 for (unsigned i = Leaf->SuffixIdx; i < Leaf->SuffixIdx + StringLen; i++)
991 Seq.push_back(ST.Str[i]);
992 OutlinedFunction OF(FunctionList.size(), CandidatesForRepeatedSeq.size(),
994 unsigned Benefit = OF.getBenefit();
996 // Is it better to outline this candidate than not?
998 // Outlining this candidate would take more instructions than not
1000 // Emit a remark explaining why we didn't outline this candidate.
1001 std::pair<MachineBasicBlock::iterator, MachineBasicBlock::iterator> C =
1002 RepeatedSequenceLocs[0];
1003 MachineOptimizationRemarkEmitter MORE(
1004 *(C.first->getParent()->getParent()), nullptr);
1006 MachineOptimizationRemarkMissed R(DEBUG_TYPE, "NotOutliningCheaper",
1007 C.first->getDebugLoc(),
1008 C.first->getParent());
1009 R << "Did not outline " << NV("Length", StringLen) << " instructions"
1010 << " from " << NV("NumOccurrences", RepeatedSequenceLocs.size())
1012 << " Instructions from outlining all occurrences ("
1013 << NV("OutliningCost", OF.getOutliningCost()) << ")"
1014 << " >= Unoutlined instruction count ("
1015 << NV("NotOutliningCost", StringLen * OF.getOccurrenceCount()) << ")"
1016 << " (Also found at: ";
1018 // Tell the user the other places the candidate was found.
1019 for (unsigned i = 1, e = RepeatedSequenceLocs.size(); i < e; i++) {
1020 R << NV((Twine("OtherStartLoc") + Twine(i)).str(),
1021 RepeatedSequenceLocs[i].first->getDebugLoc());
1030 // Move to the next candidate.
1034 if (StringLen > MaxLen)
1037 // At this point, the candidate class is seen as beneficial. Set their
1038 // benefit values and save them in the candidate list.
1039 std::vector<std::shared_ptr<Candidate>> CandidatesForFn;
1040 for (Candidate &C : CandidatesForRepeatedSeq) {
1041 C.Benefit = Benefit;
1043 std::shared_ptr<Candidate> Cptr = std::make_shared<Candidate>(C);
1044 CandidateList.push_back(Cptr);
1045 CandidatesForFn.push_back(Cptr);
1048 FunctionList.push_back(OF);
1049 FunctionList.back().Candidates = CandidatesForFn;
1051 // Move to the next function.
1052 Parent.IsInTree = false;
1058 // Remove C from the candidate space, and update its OutlinedFunction.
1059 void MachineOutliner::prune(Candidate &C,
1060 std::vector<OutlinedFunction> &FunctionList) {
1061 // Get the OutlinedFunction associated with this Candidate.
1062 OutlinedFunction &F = FunctionList[C.FunctionIdx];
1064 // Update C's associated function's occurrence count.
1067 // Remove C from the CandidateList.
1068 C.InCandidateList = false;
1070 DEBUG(dbgs() << "- Removed a Candidate \n";
1071 dbgs() << "--- Num fns left for candidate: " << F.getOccurrenceCount()
1073 dbgs() << "--- Candidate's functions's benefit: " << F.getBenefit()
1077 void MachineOutliner::pruneOverlaps(
1078 std::vector<std::shared_ptr<Candidate>> &CandidateList,
1079 std::vector<OutlinedFunction> &FunctionList, InstructionMapper &Mapper,
1080 unsigned MaxCandidateLen, const TargetInstrInfo &TII) {
1082 // Return true if this candidate became unbeneficial for outlining in a
1084 auto ShouldSkipCandidate = [&FunctionList, this](Candidate &C) {
1086 // Check if the candidate was removed in a previous step.
1087 if (!C.InCandidateList)
1090 // C must be alive. Check if we should remove it.
1091 if (FunctionList[C.FunctionIdx].getBenefit() < 1) {
1092 prune(C, FunctionList);
1096 // C is in the list, and F is still beneficial.
1100 // TODO: Experiment with interval trees or other interval-checking structures
1101 // to lower the time complexity of this function.
1102 // TODO: Can we do better than the simple greedy choice?
1103 // Check for overlaps in the range.
1104 // This is O(MaxCandidateLen * CandidateList.size()).
1105 for (auto It = CandidateList.begin(), Et = CandidateList.end(); It != Et;
1107 Candidate &C1 = **It;
1109 // If C1 was already pruned, or its function is no longer beneficial for
1110 // outlining, move to the next candidate.
1111 if (ShouldSkipCandidate(C1))
1114 // The minimum start index of any candidate that could overlap with this
1116 unsigned FarthestPossibleIdx = 0;
1118 // Either the index is 0, or it's at most MaxCandidateLen indices away.
1119 if (C1.getStartIdx() > MaxCandidateLen)
1120 FarthestPossibleIdx = C1.getStartIdx() - MaxCandidateLen;
1122 // Compare against the candidates in the list that start at at most
1123 // FarthestPossibleIdx indices away from C1. There are at most
1124 // MaxCandidateLen of these.
1125 for (auto Sit = It + 1; Sit != Et; Sit++) {
1126 Candidate &C2 = **Sit;
1128 // Is this candidate too far away to overlap?
1129 if (C2.getStartIdx() < FarthestPossibleIdx)
1132 // If C2 was already pruned, or its function is no longer beneficial for
1133 // outlining, move to the next candidate.
1134 if (ShouldSkipCandidate(C2))
1137 // Do C1 and C2 overlap?
1140 // High indices... [C1End ... C1Start][C2End ... C2Start] ...Low indices
1142 // We sorted our candidate list so C2Start <= C1Start. We know that
1143 // C2End > C2Start since each candidate has length >= 2. Therefore, all we
1144 // have to check is C2End < C2Start to see if we overlap.
1145 if (C2.getEndIdx() < C1.getStartIdx())
1148 // C1 and C2 overlap.
1149 // We need to choose the better of the two.
1151 // Approximate this by picking the one which would have saved us the
1152 // most instructions before any pruning.
1154 // Is C2 a better candidate?
1155 if (C2.Benefit > C1.Benefit) {
1156 // Yes, so prune C1. Since C1 is dead, we don't have to compare it
1157 // against anything anymore, so break.
1158 prune(C1, FunctionList);
1162 // Prune C2 and move on to the next candidate.
1163 prune(C2, FunctionList);
1168 unsigned MachineOutliner::buildCandidateList(
1169 std::vector<std::shared_ptr<Candidate>> &CandidateList,
1170 std::vector<OutlinedFunction> &FunctionList, SuffixTree &ST,
1171 InstructionMapper &Mapper, const TargetInstrInfo &TII) {
1173 std::vector<unsigned> CandidateSequence; // Current outlining candidate.
1174 unsigned MaxCandidateLen = 0; // Length of the longest candidate.
1177 findCandidates(ST, TII, Mapper, CandidateList, FunctionList);
1179 // Sort the candidates in decending order. This will simplify the outlining
1180 // process when we have to remove the candidates from the mapping by
1181 // allowing us to cut them out without keeping track of an offset.
1183 CandidateList.begin(), CandidateList.end(),
1184 [](const std::shared_ptr<Candidate> &LHS,
1185 const std::shared_ptr<Candidate> &RHS) { return *LHS < *RHS; });
1187 return MaxCandidateLen;
1191 MachineOutliner::createOutlinedFunction(Module &M, const OutlinedFunction &OF,
1192 InstructionMapper &Mapper) {
1194 // Create the function name. This should be unique. For now, just hash the
1195 // module name and include it in the function name plus the number of this
1197 std::ostringstream NameStream;
1198 NameStream << "OUTLINED_FUNCTION_" << OF.Name;
1200 // Create the function using an IR-level function.
1201 LLVMContext &C = M.getContext();
1202 Function *F = dyn_cast<Function>(
1203 M.getOrInsertFunction(NameStream.str(), Type::getVoidTy(C)));
1204 assert(F && "Function was null!");
1206 // NOTE: If this is linkonceodr, then we can take advantage of linker deduping
1207 // which gives us better results when we outline from linkonceodr functions.
1208 F->setLinkage(GlobalValue::PrivateLinkage);
1209 F->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
1211 BasicBlock *EntryBB = BasicBlock::Create(C, "entry", F);
1212 IRBuilder<> Builder(EntryBB);
1213 Builder.CreateRetVoid();
1215 MachineModuleInfo &MMI = getAnalysis<MachineModuleInfo>();
1216 MachineFunction &MF = MMI.getOrCreateMachineFunction(*F);
1217 MachineBasicBlock &MBB = *MF.CreateMachineBasicBlock();
1218 const TargetSubtargetInfo &STI = MF.getSubtarget();
1219 const TargetInstrInfo &TII = *STI.getInstrInfo();
1221 // Insert the new function into the module.
1222 MF.insert(MF.begin(), &MBB);
1224 TII.insertOutlinerPrologue(MBB, MF, OF.MInfo);
1226 // Copy over the instructions for the function using the integer mappings in
1228 for (unsigned Str : OF.Sequence) {
1229 MachineInstr *NewMI =
1230 MF.CloneMachineInstr(Mapper.IntegerInstructionMap.find(Str)->second);
1231 NewMI->dropMemRefs();
1233 // Don't keep debug information for outlined instructions.
1234 // FIXME: This means outlined functions are currently undebuggable.
1235 NewMI->setDebugLoc(DebugLoc());
1236 MBB.insert(MBB.end(), NewMI);
1239 TII.insertOutlinerEpilogue(MBB, MF, OF.MInfo);
1244 bool MachineOutliner::outline(
1245 Module &M, const ArrayRef<std::shared_ptr<Candidate>> &CandidateList,
1246 std::vector<OutlinedFunction> &FunctionList, InstructionMapper &Mapper) {
1248 bool OutlinedSomething = false;
1249 // Replace the candidates with calls to their respective outlined functions.
1250 for (const std::shared_ptr<Candidate> &Cptr : CandidateList) {
1251 Candidate &C = *Cptr;
1252 // Was the candidate removed during pruneOverlaps?
1253 if (!C.InCandidateList)
1256 // If not, then look at its OutlinedFunction.
1257 OutlinedFunction &OF = FunctionList[C.FunctionIdx];
1259 // Was its OutlinedFunction made unbeneficial during pruneOverlaps?
1260 if (OF.getBenefit() < 1)
1263 // If not, then outline it.
1264 assert(C.getStartIdx() < Mapper.InstrList.size() &&
1265 "Candidate out of bounds!");
1266 MachineBasicBlock *MBB = (*Mapper.InstrList[C.getStartIdx()]).getParent();
1267 MachineBasicBlock::iterator StartIt = Mapper.InstrList[C.getStartIdx()];
1268 unsigned EndIdx = C.getEndIdx();
1270 assert(EndIdx < Mapper.InstrList.size() && "Candidate out of bounds!");
1271 MachineBasicBlock::iterator EndIt = Mapper.InstrList[EndIdx];
1272 assert(EndIt != MBB->end() && "EndIt out of bounds!");
1274 EndIt++; // Erase needs one past the end index.
1276 // Does this candidate have a function yet?
1278 OF.MF = createOutlinedFunction(M, OF, Mapper);
1279 MachineBasicBlock *MBB = &*OF.MF->begin();
1281 // Output a remark telling the user that an outlined function was created,
1282 // and explaining where it came from.
1283 MachineOptimizationRemarkEmitter MORE(*OF.MF, nullptr);
1284 MachineOptimizationRemark R(DEBUG_TYPE, "OutlinedFunction",
1285 MBB->findDebugLoc(MBB->begin()), MBB);
1286 R << "Saved " << NV("OutliningBenefit", OF.getBenefit())
1287 << " instructions by "
1288 << "outlining " << NV("Length", OF.Sequence.size()) << " instructions "
1289 << "from " << NV("NumOccurrences", OF.getOccurrenceCount())
1293 // Tell the user the other places the candidate was found.
1294 for (size_t i = 0, e = OF.Candidates.size(); i < e; i++) {
1296 // Skip over things that were pruned.
1297 if (!OF.Candidates[i]->InCandidateList)
1301 (Twine("StartLoc") + Twine(i)).str(),
1302 Mapper.InstrList[OF.Candidates[i]->getStartIdx()]->getDebugLoc());
1313 MachineFunction *MF = OF.MF;
1314 const TargetSubtargetInfo &STI = MF->getSubtarget();
1315 const TargetInstrInfo &TII = *STI.getInstrInfo();
1317 // Insert a call to the new function and erase the old sequence.
1318 TII.insertOutlinedCall(M, *MBB, StartIt, *MF, C.MInfo);
1319 StartIt = Mapper.InstrList[C.getStartIdx()];
1320 MBB->erase(StartIt, EndIt);
1322 OutlinedSomething = true;
1328 DEBUG(dbgs() << "OutlinedSomething = " << OutlinedSomething << "\n";);
1330 return OutlinedSomething;
1333 bool MachineOutliner::runOnModule(Module &M) {
1335 // Is there anything in the module at all?
1339 MachineModuleInfo &MMI = getAnalysis<MachineModuleInfo>();
1340 const TargetSubtargetInfo &STI =
1341 MMI.getOrCreateMachineFunction(*M.begin()).getSubtarget();
1342 const TargetRegisterInfo *TRI = STI.getRegisterInfo();
1343 const TargetInstrInfo *TII = STI.getInstrInfo();
1345 InstructionMapper Mapper;
1347 // Build instruction mappings for each function in the module.
1348 for (Function &F : M) {
1349 MachineFunction &MF = MMI.getOrCreateMachineFunction(F);
1351 // Is the function empty? Safe to outline from?
1353 !TII->isFunctionSafeToOutlineFrom(MF, OutlineFromLinkOnceODRs))
1356 // If it is, look at each MachineBasicBlock in the function.
1357 for (MachineBasicBlock &MBB : MF) {
1359 // Is there anything in MBB?
1364 Mapper.convertToUnsignedVec(MBB, *TRI, *TII);
1368 // Construct a suffix tree, use it to find candidates, and then outline them.
1369 SuffixTree ST(Mapper.UnsignedVec);
1370 std::vector<std::shared_ptr<Candidate>> CandidateList;
1371 std::vector<OutlinedFunction> FunctionList;
1373 // Find all of the outlining candidates.
1374 unsigned MaxCandidateLen =
1375 buildCandidateList(CandidateList, FunctionList, ST, Mapper, *TII);
1377 // Remove candidates that overlap with other candidates.
1378 pruneOverlaps(CandidateList, FunctionList, Mapper, MaxCandidateLen, *TII);
1380 // Outline each of the candidates and return true if something was outlined.
1381 return outline(M, CandidateList, FunctionList, Mapper);