1 //==- BlockFrequencyInfoImpl.h - Block Frequency Implementation --*- 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 // Shared implementation of BlockFrequency for IR and Machine Instructions.
11 // See the documentation below for BlockFrequencyInfoImpl for details.
13 //===----------------------------------------------------------------------===//
15 #ifndef LLVM_ANALYSIS_BLOCKFREQUENCYINFOIMPL_H
16 #define LLVM_ANALYSIS_BLOCKFREQUENCYINFOIMPL_H
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/ADT/DenseSet.h"
20 #include "llvm/ADT/GraphTraits.h"
21 #include "llvm/ADT/Optional.h"
22 #include "llvm/ADT/PostOrderIterator.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/ADT/SparseBitVector.h"
25 #include "llvm/ADT/Twine.h"
26 #include "llvm/ADT/iterator_range.h"
27 #include "llvm/IR/BasicBlock.h"
28 #include "llvm/Support/BlockFrequency.h"
29 #include "llvm/Support/BranchProbability.h"
30 #include "llvm/Support/DOTGraphTraits.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/Format.h"
34 #include "llvm/Support/ScaledNumber.h"
35 #include "llvm/Support/raw_ostream.h"
48 #define DEBUG_TYPE "block-freq"
52 class BranchProbabilityInfo;
56 class MachineBasicBlock;
57 class MachineBranchProbabilityInfo;
58 class MachineFunction;
60 class MachineLoopInfo;
62 namespace bfi_detail {
64 struct IrreducibleGraph;
66 // This is part of a workaround for a GCC 4.7 crash on lambdas.
67 template <class BT> struct BlockEdgesAdder;
69 /// \brief Mass of a block.
71 /// This class implements a sort of fixed-point fraction always between 0.0 and
72 /// 1.0. getMass() == std::numeric_limits<uint64_t>::max() indicates a value of
75 /// Masses can be added and subtracted. Simple saturation arithmetic is used,
76 /// so arithmetic operations never overflow or underflow.
78 /// Masses can be multiplied. Multiplication treats full mass as 1.0 and uses
79 /// an inexpensive floating-point algorithm that's off-by-one (almost, but not
80 /// quite, maximum precision).
82 /// Masses can be scaled by \a BranchProbability at maximum precision.
87 BlockMass() = default;
88 explicit BlockMass(uint64_t Mass) : Mass(Mass) {}
90 static BlockMass getEmpty() { return BlockMass(); }
92 static BlockMass getFull() {
93 return BlockMass(std::numeric_limits<uint64_t>::max());
96 uint64_t getMass() const { return Mass; }
98 bool isFull() const { return Mass == std::numeric_limits<uint64_t>::max(); }
99 bool isEmpty() const { return !Mass; }
101 bool operator!() const { return isEmpty(); }
103 /// \brief Add another mass.
105 /// Adds another mass, saturating at \a isFull() rather than overflowing.
106 BlockMass &operator+=(BlockMass X) {
107 uint64_t Sum = Mass + X.Mass;
108 Mass = Sum < Mass ? std::numeric_limits<uint64_t>::max() : Sum;
112 /// \brief Subtract another mass.
114 /// Subtracts another mass, saturating at \a isEmpty() rather than
116 BlockMass &operator-=(BlockMass X) {
117 uint64_t Diff = Mass - X.Mass;
118 Mass = Diff > Mass ? 0 : Diff;
122 BlockMass &operator*=(BranchProbability P) {
123 Mass = P.scale(Mass);
127 bool operator==(BlockMass X) const { return Mass == X.Mass; }
128 bool operator!=(BlockMass X) const { return Mass != X.Mass; }
129 bool operator<=(BlockMass X) const { return Mass <= X.Mass; }
130 bool operator>=(BlockMass X) const { return Mass >= X.Mass; }
131 bool operator<(BlockMass X) const { return Mass < X.Mass; }
132 bool operator>(BlockMass X) const { return Mass > X.Mass; }
134 /// \brief Convert to scaled number.
136 /// Convert to \a ScaledNumber. \a isFull() gives 1.0, while \a isEmpty()
137 /// gives slightly above 0.0.
138 ScaledNumber<uint64_t> toScaled() const;
141 raw_ostream &print(raw_ostream &OS) const;
144 inline BlockMass operator+(BlockMass L, BlockMass R) {
145 return BlockMass(L) += R;
147 inline BlockMass operator-(BlockMass L, BlockMass R) {
148 return BlockMass(L) -= R;
150 inline BlockMass operator*(BlockMass L, BranchProbability R) {
151 return BlockMass(L) *= R;
153 inline BlockMass operator*(BranchProbability L, BlockMass R) {
154 return BlockMass(R) *= L;
157 inline raw_ostream &operator<<(raw_ostream &OS, BlockMass X) {
161 } // end namespace bfi_detail
163 template <> struct isPodLike<bfi_detail::BlockMass> {
164 static const bool value = true;
167 /// \brief Base class for BlockFrequencyInfoImpl
169 /// BlockFrequencyInfoImplBase has supporting data structures and some
170 /// algorithms for BlockFrequencyInfoImplBase. Only algorithms that depend on
171 /// the block type (or that call such algorithms) are skipped here.
173 /// Nevertheless, the majority of the overall algorithm documention lives with
174 /// BlockFrequencyInfoImpl. See there for details.
175 class BlockFrequencyInfoImplBase {
177 using Scaled64 = ScaledNumber<uint64_t>;
178 using BlockMass = bfi_detail::BlockMass;
180 /// \brief Representative of a block.
182 /// This is a simple wrapper around an index into the reverse-post-order
183 /// traversal of the blocks.
185 /// Unlike a block pointer, its order has meaning (location in the
186 /// topological sort) and it's class is the same regardless of block type.
188 using IndexType = uint32_t;
190 IndexType Index = std::numeric_limits<uint32_t>::max();
192 BlockNode() = default;
193 BlockNode(IndexType Index) : Index(Index) {}
195 bool operator==(const BlockNode &X) const { return Index == X.Index; }
196 bool operator!=(const BlockNode &X) const { return Index != X.Index; }
197 bool operator<=(const BlockNode &X) const { return Index <= X.Index; }
198 bool operator>=(const BlockNode &X) const { return Index >= X.Index; }
199 bool operator<(const BlockNode &X) const { return Index < X.Index; }
200 bool operator>(const BlockNode &X) const { return Index > X.Index; }
202 bool isValid() const { return Index <= getMaxIndex(); }
204 static size_t getMaxIndex() {
205 return std::numeric_limits<uint32_t>::max() - 1;
209 /// \brief Stats about a block itself.
210 struct FrequencyData {
215 /// \brief Data about a loop.
217 /// Contains the data necessary to represent a loop as a pseudo-node once it's
220 using ExitMap = SmallVector<std::pair<BlockNode, BlockMass>, 4>;
221 using NodeList = SmallVector<BlockNode, 4>;
222 using HeaderMassList = SmallVector<BlockMass, 1>;
224 LoopData *Parent; ///< The parent loop.
225 bool IsPackaged = false; ///< Whether this has been packaged.
226 uint32_t NumHeaders = 1; ///< Number of headers.
227 ExitMap Exits; ///< Successor edges (and weights).
228 NodeList Nodes; ///< Header and the members of the loop.
229 HeaderMassList BackedgeMass; ///< Mass returned to each loop header.
233 LoopData(LoopData *Parent, const BlockNode &Header)
234 : Parent(Parent), Nodes(1, Header), BackedgeMass(1) {}
236 template <class It1, class It2>
237 LoopData(LoopData *Parent, It1 FirstHeader, It1 LastHeader, It2 FirstOther,
239 : Parent(Parent), Nodes(FirstHeader, LastHeader) {
240 NumHeaders = Nodes.size();
241 Nodes.insert(Nodes.end(), FirstOther, LastOther);
242 BackedgeMass.resize(NumHeaders);
245 bool isHeader(const BlockNode &Node) const {
247 return std::binary_search(Nodes.begin(), Nodes.begin() + NumHeaders,
249 return Node == Nodes[0];
252 BlockNode getHeader() const { return Nodes[0]; }
253 bool isIrreducible() const { return NumHeaders > 1; }
255 HeaderMassList::difference_type getHeaderIndex(const BlockNode &B) {
256 assert(isHeader(B) && "this is only valid on loop header blocks");
258 return std::lower_bound(Nodes.begin(), Nodes.begin() + NumHeaders, B) -
263 NodeList::const_iterator members_begin() const {
264 return Nodes.begin() + NumHeaders;
267 NodeList::const_iterator members_end() const { return Nodes.end(); }
268 iterator_range<NodeList::const_iterator> members() const {
269 return make_range(members_begin(), members_end());
273 /// \brief Index of loop information.
275 BlockNode Node; ///< This node.
276 LoopData *Loop = nullptr; ///< The loop this block is inside.
277 BlockMass Mass; ///< Mass distribution from the entry block.
279 WorkingData(const BlockNode &Node) : Node(Node) {}
281 bool isLoopHeader() const { return Loop && Loop->isHeader(Node); }
283 bool isDoubleLoopHeader() const {
284 return isLoopHeader() && Loop->Parent && Loop->Parent->isIrreducible() &&
285 Loop->Parent->isHeader(Node);
288 LoopData *getContainingLoop() const {
291 if (!isDoubleLoopHeader())
293 return Loop->Parent->Parent;
296 /// \brief Resolve a node to its representative.
298 /// Get the node currently representing Node, which could be a containing
301 /// This function should only be called when distributing mass. As long as
302 /// there are no irreducible edges to Node, then it will have complexity
303 /// O(1) in this context.
305 /// In general, the complexity is O(L), where L is the number of loop
306 /// headers Node has been packaged into. Since this method is called in
307 /// the context of distributing mass, L will be the number of loop headers
308 /// an early exit edge jumps out of.
309 BlockNode getResolvedNode() const {
310 auto L = getPackagedLoop();
311 return L ? L->getHeader() : Node;
314 LoopData *getPackagedLoop() const {
315 if (!Loop || !Loop->IsPackaged)
318 while (L->Parent && L->Parent->IsPackaged)
323 /// \brief Get the appropriate mass for a node.
325 /// Get appropriate mass for Node. If Node is a loop-header (whose loop
326 /// has been packaged), returns the mass of its pseudo-node. If it's a
327 /// node inside a packaged loop, it returns the loop's mass.
328 BlockMass &getMass() {
331 if (!isADoublePackage())
333 return Loop->Parent->Mass;
336 /// \brief Has ContainingLoop been packaged up?
337 bool isPackaged() const { return getResolvedNode() != Node; }
339 /// \brief Has Loop been packaged up?
340 bool isAPackage() const { return isLoopHeader() && Loop->IsPackaged; }
342 /// \brief Has Loop been packaged up twice?
343 bool isADoublePackage() const {
344 return isDoubleLoopHeader() && Loop->Parent->IsPackaged;
348 /// \brief Unscaled probability weight.
350 /// Probability weight for an edge in the graph (including the
351 /// successor/target node).
353 /// All edges in the original function are 32-bit. However, exit edges from
354 /// loop packages are taken from 64-bit exit masses, so we need 64-bits of
355 /// space in general.
357 /// In addition to the raw weight amount, Weight stores the type of the edge
358 /// in the current context (i.e., the context of the loop being processed).
359 /// Is this a local edge within the loop, an exit from the loop, or a
360 /// backedge to the loop header?
362 enum DistType { Local, Exit, Backedge };
363 DistType Type = Local;
364 BlockNode TargetNode;
368 Weight(DistType Type, BlockNode TargetNode, uint64_t Amount)
369 : Type(Type), TargetNode(TargetNode), Amount(Amount) {}
372 /// \brief Distribution of unscaled probability weight.
374 /// Distribution of unscaled probability weight to a set of successors.
376 /// This class collates the successor edge weights for later processing.
378 /// \a DidOverflow indicates whether \a Total did overflow while adding to
379 /// the distribution. It should never overflow twice.
380 struct Distribution {
381 using WeightList = SmallVector<Weight, 4>;
383 WeightList Weights; ///< Individual successor weights.
384 uint64_t Total = 0; ///< Sum of all weights.
385 bool DidOverflow = false; ///< Whether \a Total did overflow.
387 Distribution() = default;
389 void addLocal(const BlockNode &Node, uint64_t Amount) {
390 add(Node, Amount, Weight::Local);
393 void addExit(const BlockNode &Node, uint64_t Amount) {
394 add(Node, Amount, Weight::Exit);
397 void addBackedge(const BlockNode &Node, uint64_t Amount) {
398 add(Node, Amount, Weight::Backedge);
401 /// \brief Normalize the distribution.
403 /// Combines multiple edges to the same \a Weight::TargetNode and scales
404 /// down so that \a Total fits into 32-bits.
406 /// This is linear in the size of \a Weights. For the vast majority of
407 /// cases, adjacent edge weights are combined by sorting WeightList and
408 /// combining adjacent weights. However, for very large edge lists an
409 /// auxiliary hash table is used.
413 void add(const BlockNode &Node, uint64_t Amount, Weight::DistType Type);
416 /// \brief Data about each block. This is used downstream.
417 std::vector<FrequencyData> Freqs;
419 /// \brief Whether each block is an irreducible loop header.
420 /// This is used downstream.
421 SparseBitVector<> IsIrrLoopHeader;
423 /// \brief Loop data: see initializeLoops().
424 std::vector<WorkingData> Working;
426 /// \brief Indexed information about loops.
427 std::list<LoopData> Loops;
429 /// \brief Virtual destructor.
431 /// Need a virtual destructor to mask the compiler warning about
433 virtual ~BlockFrequencyInfoImplBase() = default;
435 /// \brief Add all edges out of a packaged loop to the distribution.
437 /// Adds all edges from LocalLoopHead to Dist. Calls addToDist() to add each
440 /// \return \c true unless there's an irreducible backedge.
441 bool addLoopSuccessorsToDist(const LoopData *OuterLoop, LoopData &Loop,
444 /// \brief Add an edge to the distribution.
446 /// Adds an edge to Succ to Dist. If \c LoopHead.isValid(), then whether the
447 /// edge is local/exit/backedge is in the context of LoopHead. Otherwise,
448 /// every edge should be a local edge (since all the loops are packaged up).
450 /// \return \c true unless aborted due to an irreducible backedge.
451 bool addToDist(Distribution &Dist, const LoopData *OuterLoop,
452 const BlockNode &Pred, const BlockNode &Succ, uint64_t Weight);
454 LoopData &getLoopPackage(const BlockNode &Head) {
455 assert(Head.Index < Working.size());
456 assert(Working[Head.Index].isLoopHeader());
457 return *Working[Head.Index].Loop;
460 /// \brief Analyze irreducible SCCs.
462 /// Separate irreducible SCCs from \c G, which is an explict graph of \c
463 /// OuterLoop (or the top-level function, if \c OuterLoop is \c nullptr).
464 /// Insert them into \a Loops before \c Insert.
466 /// \return the \c LoopData nodes representing the irreducible SCCs.
467 iterator_range<std::list<LoopData>::iterator>
468 analyzeIrreducible(const bfi_detail::IrreducibleGraph &G, LoopData *OuterLoop,
469 std::list<LoopData>::iterator Insert);
471 /// \brief Update a loop after packaging irreducible SCCs inside of it.
473 /// Update \c OuterLoop. Before finding irreducible control flow, it was
474 /// partway through \a computeMassInLoop(), so \a LoopData::Exits and \a
475 /// LoopData::BackedgeMass need to be reset. Also, nodes that were packaged
476 /// up need to be removed from \a OuterLoop::Nodes.
477 void updateLoopWithIrreducible(LoopData &OuterLoop);
479 /// \brief Distribute mass according to a distribution.
481 /// Distributes the mass in Source according to Dist. If LoopHead.isValid(),
482 /// backedges and exits are stored in its entry in Loops.
484 /// Mass is distributed in parallel from two copies of the source mass.
485 void distributeMass(const BlockNode &Source, LoopData *OuterLoop,
488 /// \brief Compute the loop scale for a loop.
489 void computeLoopScale(LoopData &Loop);
491 /// Adjust the mass of all headers in an irreducible loop.
493 /// Initially, irreducible loops are assumed to distribute their mass
494 /// equally among its headers. This can lead to wrong frequency estimates
495 /// since some headers may be executed more frequently than others.
497 /// This adjusts header mass distribution so it matches the weights of
498 /// the backedges going into each of the loop headers.
499 void adjustLoopHeaderMass(LoopData &Loop);
501 void distributeIrrLoopHeaderMass(Distribution &Dist);
503 /// \brief Package up a loop.
504 void packageLoop(LoopData &Loop);
506 /// \brief Unwrap loops.
509 /// \brief Finalize frequency metrics.
511 /// Calculates final frequencies and cleans up no-longer-needed data
513 void finalizeMetrics();
515 /// \brief Clear all memory.
518 virtual std::string getBlockName(const BlockNode &Node) const;
519 std::string getLoopName(const LoopData &Loop) const;
521 virtual raw_ostream &print(raw_ostream &OS) const { return OS; }
522 void dump() const { print(dbgs()); }
524 Scaled64 getFloatingBlockFreq(const BlockNode &Node) const;
526 BlockFrequency getBlockFreq(const BlockNode &Node) const;
527 Optional<uint64_t> getBlockProfileCount(const Function &F,
528 const BlockNode &Node) const;
529 Optional<uint64_t> getProfileCountFromFreq(const Function &F,
530 uint64_t Freq) const;
531 bool isIrrLoopHeader(const BlockNode &Node);
533 void setBlockFreq(const BlockNode &Node, uint64_t Freq);
535 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockNode &Node) const;
536 raw_ostream &printBlockFreq(raw_ostream &OS,
537 const BlockFrequency &Freq) const;
539 uint64_t getEntryFreq() const {
540 assert(!Freqs.empty());
541 return Freqs[0].Integer;
545 namespace bfi_detail {
547 template <class BlockT> struct TypeMap {};
548 template <> struct TypeMap<BasicBlock> {
549 using BlockT = BasicBlock;
550 using FunctionT = Function;
551 using BranchProbabilityInfoT = BranchProbabilityInfo;
553 using LoopInfoT = LoopInfo;
555 template <> struct TypeMap<MachineBasicBlock> {
556 using BlockT = MachineBasicBlock;
557 using FunctionT = MachineFunction;
558 using BranchProbabilityInfoT = MachineBranchProbabilityInfo;
559 using LoopT = MachineLoop;
560 using LoopInfoT = MachineLoopInfo;
563 /// \brief Get the name of a MachineBasicBlock.
565 /// Get the name of a MachineBasicBlock. It's templated so that including from
566 /// CodeGen is unnecessary (that would be a layering issue).
568 /// This is used mainly for debug output. The name is similar to
569 /// MachineBasicBlock::getFullName(), but skips the name of the function.
570 template <class BlockT> std::string getBlockName(const BlockT *BB) {
571 assert(BB && "Unexpected nullptr");
572 auto MachineName = "BB" + Twine(BB->getNumber());
573 if (BB->getBasicBlock())
574 return (MachineName + "[" + BB->getName() + "]").str();
575 return MachineName.str();
577 /// \brief Get the name of a BasicBlock.
578 template <> inline std::string getBlockName(const BasicBlock *BB) {
579 assert(BB && "Unexpected nullptr");
580 return BB->getName().str();
583 /// \brief Graph of irreducible control flow.
585 /// This graph is used for determining the SCCs in a loop (or top-level
586 /// function) that has irreducible control flow.
588 /// During the block frequency algorithm, the local graphs are defined in a
589 /// light-weight way, deferring to the \a BasicBlock or \a MachineBasicBlock
590 /// graphs for most edges, but getting others from \a LoopData::ExitMap. The
591 /// latter only has successor information.
593 /// \a IrreducibleGraph makes this graph explicit. It's in a form that can use
594 /// \a GraphTraits (so that \a analyzeIrreducible() can use \a scc_iterator),
595 /// and it explicitly lists predecessors and successors. The initialization
596 /// that relies on \c MachineBasicBlock is defined in the header.
597 struct IrreducibleGraph {
598 using BFIBase = BlockFrequencyInfoImplBase;
602 using BlockNode = BFIBase::BlockNode;
606 std::deque<const IrrNode *> Edges;
608 IrrNode(const BlockNode &Node) : Node(Node) {}
610 using iterator = std::deque<const IrrNode *>::const_iterator;
612 iterator pred_begin() const { return Edges.begin(); }
613 iterator succ_begin() const { return Edges.begin() + NumIn; }
614 iterator pred_end() const { return succ_begin(); }
615 iterator succ_end() const { return Edges.end(); }
618 const IrrNode *StartIrr = nullptr;
619 std::vector<IrrNode> Nodes;
620 SmallDenseMap<uint32_t, IrrNode *, 4> Lookup;
622 /// \brief Construct an explicit graph containing irreducible control flow.
624 /// Construct an explicit graph of the control flow in \c OuterLoop (or the
625 /// top-level function, if \c OuterLoop is \c nullptr). Uses \c
626 /// addBlockEdges to add block successors that have not been packaged into
629 /// \a BlockFrequencyInfoImpl::computeIrreducibleMass() is the only expected
631 template <class BlockEdgesAdder>
632 IrreducibleGraph(BFIBase &BFI, const BFIBase::LoopData *OuterLoop,
633 BlockEdgesAdder addBlockEdges) : BFI(BFI) {
634 initialize(OuterLoop, addBlockEdges);
637 template <class BlockEdgesAdder>
638 void initialize(const BFIBase::LoopData *OuterLoop,
639 BlockEdgesAdder addBlockEdges);
640 void addNodesInLoop(const BFIBase::LoopData &OuterLoop);
641 void addNodesInFunction();
643 void addNode(const BlockNode &Node) {
644 Nodes.emplace_back(Node);
645 BFI.Working[Node.Index].getMass() = BlockMass::getEmpty();
649 template <class BlockEdgesAdder>
650 void addEdges(const BlockNode &Node, const BFIBase::LoopData *OuterLoop,
651 BlockEdgesAdder addBlockEdges);
652 void addEdge(IrrNode &Irr, const BlockNode &Succ,
653 const BFIBase::LoopData *OuterLoop);
656 template <class BlockEdgesAdder>
657 void IrreducibleGraph::initialize(const BFIBase::LoopData *OuterLoop,
658 BlockEdgesAdder addBlockEdges) {
660 addNodesInLoop(*OuterLoop);
661 for (auto N : OuterLoop->Nodes)
662 addEdges(N, OuterLoop, addBlockEdges);
664 addNodesInFunction();
665 for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index)
666 addEdges(Index, OuterLoop, addBlockEdges);
668 StartIrr = Lookup[Start.Index];
671 template <class BlockEdgesAdder>
672 void IrreducibleGraph::addEdges(const BlockNode &Node,
673 const BFIBase::LoopData *OuterLoop,
674 BlockEdgesAdder addBlockEdges) {
675 auto L = Lookup.find(Node.Index);
676 if (L == Lookup.end())
678 IrrNode &Irr = *L->second;
679 const auto &Working = BFI.Working[Node.Index];
681 if (Working.isAPackage())
682 for (const auto &I : Working.Loop->Exits)
683 addEdge(Irr, I.first, OuterLoop);
685 addBlockEdges(*this, Irr, OuterLoop);
688 } // end namespace bfi_detail
690 /// \brief Shared implementation for block frequency analysis.
692 /// This is a shared implementation of BlockFrequencyInfo and
693 /// MachineBlockFrequencyInfo, and calculates the relative frequencies of
696 /// LoopInfo defines a loop as a "non-trivial" SCC dominated by a single block,
697 /// which is called the header. A given loop, L, can have sub-loops, which are
698 /// loops within the subgraph of L that exclude its header. (A "trivial" SCC
699 /// consists of a single block that does not have a self-edge.)
701 /// In addition to loops, this algorithm has limited support for irreducible
702 /// SCCs, which are SCCs with multiple entry blocks. Irreducible SCCs are
703 /// discovered on they fly, and modelled as loops with multiple headers.
705 /// The headers of irreducible sub-SCCs consist of its entry blocks and all
706 /// nodes that are targets of a backedge within it (excluding backedges within
707 /// true sub-loops). Block frequency calculations act as if a block is
708 /// inserted that intercepts all the edges to the headers. All backedges and
709 /// entries point to this block. Its successors are the headers, which split
710 /// the frequency evenly.
712 /// This algorithm leverages BlockMass and ScaledNumber to maintain precision,
713 /// separates mass distribution from loop scaling, and dithers to eliminate
714 /// probability mass loss.
716 /// The implementation is split between BlockFrequencyInfoImpl, which knows the
717 /// type of graph being modelled (BasicBlock vs. MachineBasicBlock), and
718 /// BlockFrequencyInfoImplBase, which doesn't. The base class uses \a
719 /// BlockNode, a wrapper around a uint32_t. BlockNode is numbered from 0 in
720 /// reverse-post order. This gives two advantages: it's easy to compare the
721 /// relative ordering of two nodes, and maps keyed on BlockT can be represented
724 /// This algorithm is O(V+E), unless there is irreducible control flow, in
725 /// which case it's O(V*E) in the worst case.
727 /// These are the main stages:
729 /// 0. Reverse post-order traversal (\a initializeRPOT()).
731 /// Run a single post-order traversal and save it (in reverse) in RPOT.
732 /// All other stages make use of this ordering. Save a lookup from BlockT
733 /// to BlockNode (the index into RPOT) in Nodes.
735 /// 1. Loop initialization (\a initializeLoops()).
737 /// Translate LoopInfo/MachineLoopInfo into a form suitable for the rest of
738 /// the algorithm. In particular, store the immediate members of each loop
739 /// in reverse post-order.
741 /// 2. Calculate mass and scale in loops (\a computeMassInLoops()).
743 /// For each loop (bottom-up), distribute mass through the DAG resulting
744 /// from ignoring backedges and treating sub-loops as a single pseudo-node.
745 /// Track the backedge mass distributed to the loop header, and use it to
746 /// calculate the loop scale (number of loop iterations). Immediate
747 /// members that represent sub-loops will already have been visited and
748 /// packaged into a pseudo-node.
750 /// Distributing mass in a loop is a reverse-post-order traversal through
751 /// the loop. Start by assigning full mass to the Loop header. For each
752 /// node in the loop:
754 /// - Fetch and categorize the weight distribution for its successors.
755 /// If this is a packaged-subloop, the weight distribution is stored
756 /// in \a LoopData::Exits. Otherwise, fetch it from
757 /// BranchProbabilityInfo.
759 /// - Each successor is categorized as \a Weight::Local, a local edge
760 /// within the current loop, \a Weight::Backedge, a backedge to the
761 /// loop header, or \a Weight::Exit, any successor outside the loop.
762 /// The weight, the successor, and its category are stored in \a
763 /// Distribution. There can be multiple edges to each successor.
765 /// - If there's a backedge to a non-header, there's an irreducible SCC.
766 /// The usual flow is temporarily aborted. \a
767 /// computeIrreducibleMass() finds the irreducible SCCs within the
768 /// loop, packages them up, and restarts the flow.
770 /// - Normalize the distribution: scale weights down so that their sum
771 /// is 32-bits, and coalesce multiple edges to the same node.
773 /// - Distribute the mass accordingly, dithering to minimize mass loss,
774 /// as described in \a distributeMass().
776 /// In the case of irreducible loops, instead of a single loop header,
777 /// there will be several. The computation of backedge masses is similar
778 /// but instead of having a single backedge mass, there will be one
779 /// backedge per loop header. In these cases, each backedge will carry
780 /// a mass proportional to the edge weights along the corresponding
783 /// At the end of propagation, the full mass assigned to the loop will be
784 /// distributed among the loop headers proportionally according to the
785 /// mass flowing through their backedges.
787 /// Finally, calculate the loop scale from the accumulated backedge mass.
789 /// 3. Distribute mass in the function (\a computeMassInFunction()).
791 /// Finally, distribute mass through the DAG resulting from packaging all
792 /// loops in the function. This uses the same algorithm as distributing
793 /// mass in a loop, except that there are no exit or backedge edges.
795 /// 4. Unpackage loops (\a unwrapLoops()).
797 /// Initialize each block's frequency to a floating point representation of
800 /// Visit loops top-down, scaling the frequencies of its immediate members
801 /// by the loop's pseudo-node's frequency.
803 /// 5. Convert frequencies to a 64-bit range (\a finalizeMetrics()).
805 /// Using the min and max frequencies as a guide, translate floating point
806 /// frequencies to an appropriate range in uint64_t.
808 /// It has some known flaws.
810 /// - The model of irreducible control flow is a rough approximation.
812 /// Modelling irreducible control flow exactly involves setting up and
813 /// solving a group of infinite geometric series. Such precision is
814 /// unlikely to be worthwhile, since most of our algorithms give up on
815 /// irreducible control flow anyway.
817 /// Nevertheless, we might find that we need to get closer. Here's a sort
818 /// of TODO list for the model with diminishing returns, to be completed as
821 /// - The headers for the \a LoopData representing an irreducible SCC
822 /// include non-entry blocks. When these extra blocks exist, they
823 /// indicate a self-contained irreducible sub-SCC. We could treat them
824 /// as sub-loops, rather than arbitrarily shoving the problematic
825 /// blocks into the headers of the main irreducible SCC.
827 /// - Entry frequencies are assumed to be evenly split between the
828 /// headers of a given irreducible SCC, which is the only option if we
829 /// need to compute mass in the SCC before its parent loop. Instead,
830 /// we could partially compute mass in the parent loop, and stop when
831 /// we get to the SCC. Here, we have the correct ratio of entry
832 /// masses, which we can use to adjust their relative frequencies.
833 /// Compute mass in the SCC, and then continue propagation in the
836 /// - We can propagate mass iteratively through the SCC, for some fixed
837 /// number of iterations. Each iteration starts by assigning the entry
838 /// blocks their backedge mass from the prior iteration. The final
839 /// mass for each block (and each exit, and the total backedge mass
840 /// used for computing loop scale) is the sum of all iterations.
841 /// (Running this until fixed point would "solve" the geometric
842 /// series by simulation.)
843 template <class BT> class BlockFrequencyInfoImpl : BlockFrequencyInfoImplBase {
844 // This is part of a workaround for a GCC 4.7 crash on lambdas.
845 friend struct bfi_detail::BlockEdgesAdder<BT>;
847 using BlockT = typename bfi_detail::TypeMap<BT>::BlockT;
848 using FunctionT = typename bfi_detail::TypeMap<BT>::FunctionT;
849 using BranchProbabilityInfoT =
850 typename bfi_detail::TypeMap<BT>::BranchProbabilityInfoT;
851 using LoopT = typename bfi_detail::TypeMap<BT>::LoopT;
852 using LoopInfoT = typename bfi_detail::TypeMap<BT>::LoopInfoT;
853 using Successor = GraphTraits<const BlockT *>;
854 using Predecessor = GraphTraits<Inverse<const BlockT *>>;
856 const BranchProbabilityInfoT *BPI = nullptr;
857 const LoopInfoT *LI = nullptr;
858 const FunctionT *F = nullptr;
860 // All blocks in reverse postorder.
861 std::vector<const BlockT *> RPOT;
862 DenseMap<const BlockT *, BlockNode> Nodes;
864 using rpot_iterator = typename std::vector<const BlockT *>::const_iterator;
866 rpot_iterator rpot_begin() const { return RPOT.begin(); }
867 rpot_iterator rpot_end() const { return RPOT.end(); }
869 size_t getIndex(const rpot_iterator &I) const { return I - rpot_begin(); }
871 BlockNode getNode(const rpot_iterator &I) const {
872 return BlockNode(getIndex(I));
874 BlockNode getNode(const BlockT *BB) const { return Nodes.lookup(BB); }
876 const BlockT *getBlock(const BlockNode &Node) const {
877 assert(Node.Index < RPOT.size());
878 return RPOT[Node.Index];
881 /// \brief Run (and save) a post-order traversal.
883 /// Saves a reverse post-order traversal of all the nodes in \a F.
884 void initializeRPOT();
886 /// \brief Initialize loop data.
888 /// Build up \a Loops using \a LoopInfo. \a LoopInfo gives us a mapping from
889 /// each block to the deepest loop it's in, but we need the inverse. For each
890 /// loop, we store in reverse post-order its "immediate" members, defined as
891 /// the header, the headers of immediate sub-loops, and all other blocks in
892 /// the loop that are not in sub-loops.
893 void initializeLoops();
895 /// \brief Propagate to a block's successors.
897 /// In the context of distributing mass through \c OuterLoop, divide the mass
898 /// currently assigned to \c Node between its successors.
900 /// \return \c true unless there's an irreducible backedge.
901 bool propagateMassToSuccessors(LoopData *OuterLoop, const BlockNode &Node);
903 /// \brief Compute mass in a particular loop.
905 /// Assign mass to \c Loop's header, and then for each block in \c Loop in
906 /// reverse post-order, distribute mass to its successors. Only visits nodes
907 /// that have not been packaged into sub-loops.
909 /// \pre \a computeMassInLoop() has been called for each subloop of \c Loop.
910 /// \return \c true unless there's an irreducible backedge.
911 bool computeMassInLoop(LoopData &Loop);
913 /// \brief Try to compute mass in the top-level function.
915 /// Assign mass to the entry block, and then for each block in reverse
916 /// post-order, distribute mass to its successors. Skips nodes that have
917 /// been packaged into loops.
919 /// \pre \a computeMassInLoops() has been called.
920 /// \return \c true unless there's an irreducible backedge.
921 bool tryToComputeMassInFunction();
923 /// \brief Compute mass in (and package up) irreducible SCCs.
925 /// Find the irreducible SCCs in \c OuterLoop, add them to \a Loops (in front
926 /// of \c Insert), and call \a computeMassInLoop() on each of them.
928 /// If \c OuterLoop is \c nullptr, it refers to the top-level function.
930 /// \pre \a computeMassInLoop() has been called for each subloop of \c
932 /// \pre \c Insert points at the last loop successfully processed by \a
933 /// computeMassInLoop().
934 /// \pre \c OuterLoop has irreducible SCCs.
935 void computeIrreducibleMass(LoopData *OuterLoop,
936 std::list<LoopData>::iterator Insert);
938 /// \brief Compute mass in all loops.
940 /// For each loop bottom-up, call \a computeMassInLoop().
942 /// \a computeMassInLoop() aborts (and returns \c false) on loops that
943 /// contain a irreducible sub-SCCs. Use \a computeIrreducibleMass() and then
944 /// re-enter \a computeMassInLoop().
946 /// \post \a computeMassInLoop() has returned \c true for every loop.
947 void computeMassInLoops();
949 /// \brief Compute mass in the top-level function.
951 /// Uses \a tryToComputeMassInFunction() and \a computeIrreducibleMass() to
952 /// compute mass in the top-level function.
954 /// \post \a tryToComputeMassInFunction() has returned \c true.
955 void computeMassInFunction();
957 std::string getBlockName(const BlockNode &Node) const override {
958 return bfi_detail::getBlockName(getBlock(Node));
962 BlockFrequencyInfoImpl() = default;
964 const FunctionT *getFunction() const { return F; }
966 void calculate(const FunctionT &F, const BranchProbabilityInfoT &BPI,
967 const LoopInfoT &LI);
969 using BlockFrequencyInfoImplBase::getEntryFreq;
971 BlockFrequency getBlockFreq(const BlockT *BB) const {
972 return BlockFrequencyInfoImplBase::getBlockFreq(getNode(BB));
975 Optional<uint64_t> getBlockProfileCount(const Function &F,
976 const BlockT *BB) const {
977 return BlockFrequencyInfoImplBase::getBlockProfileCount(F, getNode(BB));
980 Optional<uint64_t> getProfileCountFromFreq(const Function &F,
981 uint64_t Freq) const {
982 return BlockFrequencyInfoImplBase::getProfileCountFromFreq(F, Freq);
985 bool isIrrLoopHeader(const BlockT *BB) {
986 return BlockFrequencyInfoImplBase::isIrrLoopHeader(getNode(BB));
989 void setBlockFreq(const BlockT *BB, uint64_t Freq);
991 Scaled64 getFloatingBlockFreq(const BlockT *BB) const {
992 return BlockFrequencyInfoImplBase::getFloatingBlockFreq(getNode(BB));
995 const BranchProbabilityInfoT &getBPI() const { return *BPI; }
997 /// \brief Print the frequencies for the current function.
999 /// Prints the frequencies for the blocks in the current function.
1001 /// Blocks are printed in the natural iteration order of the function, rather
1002 /// than reverse post-order. This provides two advantages: writing -analyze
1003 /// tests is easier (since blocks come out in source order), and even
1004 /// unreachable blocks are printed.
1006 /// \a BlockFrequencyInfoImplBase::print() only knows reverse post-order, so
1007 /// we need to override it here.
1008 raw_ostream &print(raw_ostream &OS) const override;
1010 using BlockFrequencyInfoImplBase::dump;
1011 using BlockFrequencyInfoImplBase::printBlockFreq;
1013 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockT *BB) const {
1014 return BlockFrequencyInfoImplBase::printBlockFreq(OS, getNode(BB));
1019 void BlockFrequencyInfoImpl<BT>::calculate(const FunctionT &F,
1020 const BranchProbabilityInfoT &BPI,
1021 const LoopInfoT &LI) {
1022 // Save the parameters.
1027 // Clean up left-over data structures.
1028 BlockFrequencyInfoImplBase::clear();
1033 DEBUG(dbgs() << "\nblock-frequency: " << F.getName() << "\n================="
1034 << std::string(F.getName().size(), '=') << "\n");
1038 // Visit loops in post-order to find the local mass distribution, and then do
1039 // the full function.
1040 computeMassInLoops();
1041 computeMassInFunction();
1047 void BlockFrequencyInfoImpl<BT>::setBlockFreq(const BlockT *BB, uint64_t Freq) {
1048 if (Nodes.count(BB))
1049 BlockFrequencyInfoImplBase::setBlockFreq(getNode(BB), Freq);
1051 // If BB is a newly added block after BFI is done, we need to create a new
1052 // BlockNode for it assigned with a new index. The index can be determined
1053 // by the size of Freqs.
1054 BlockNode NewNode(Freqs.size());
1055 Nodes[BB] = NewNode;
1056 Freqs.emplace_back();
1057 BlockFrequencyInfoImplBase::setBlockFreq(NewNode, Freq);
1061 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeRPOT() {
1062 const BlockT *Entry = &F->front();
1063 RPOT.reserve(F->size());
1064 std::copy(po_begin(Entry), po_end(Entry), std::back_inserter(RPOT));
1065 std::reverse(RPOT.begin(), RPOT.end());
1067 assert(RPOT.size() - 1 <= BlockNode::getMaxIndex() &&
1068 "More nodes in function than Block Frequency Info supports");
1070 DEBUG(dbgs() << "reverse-post-order-traversal\n");
1071 for (rpot_iterator I = rpot_begin(), E = rpot_end(); I != E; ++I) {
1072 BlockNode Node = getNode(I);
1073 DEBUG(dbgs() << " - " << getIndex(I) << ": " << getBlockName(Node) << "\n");
1077 Working.reserve(RPOT.size());
1078 for (size_t Index = 0; Index < RPOT.size(); ++Index)
1079 Working.emplace_back(Index);
1080 Freqs.resize(RPOT.size());
1083 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeLoops() {
1084 DEBUG(dbgs() << "loop-detection\n");
1088 // Visit loops top down and assign them an index.
1089 std::deque<std::pair<const LoopT *, LoopData *>> Q;
1090 for (const LoopT *L : *LI)
1091 Q.emplace_back(L, nullptr);
1092 while (!Q.empty()) {
1093 const LoopT *Loop = Q.front().first;
1094 LoopData *Parent = Q.front().second;
1097 BlockNode Header = getNode(Loop->getHeader());
1098 assert(Header.isValid());
1100 Loops.emplace_back(Parent, Header);
1101 Working[Header.Index].Loop = &Loops.back();
1102 DEBUG(dbgs() << " - loop = " << getBlockName(Header) << "\n");
1104 for (const LoopT *L : *Loop)
1105 Q.emplace_back(L, &Loops.back());
1108 // Visit nodes in reverse post-order and add them to their deepest containing
1110 for (size_t Index = 0; Index < RPOT.size(); ++Index) {
1111 // Loop headers have already been mostly mapped.
1112 if (Working[Index].isLoopHeader()) {
1113 LoopData *ContainingLoop = Working[Index].getContainingLoop();
1115 ContainingLoop->Nodes.push_back(Index);
1119 const LoopT *Loop = LI->getLoopFor(RPOT[Index]);
1123 // Add this node to its containing loop's member list.
1124 BlockNode Header = getNode(Loop->getHeader());
1125 assert(Header.isValid());
1126 const auto &HeaderData = Working[Header.Index];
1127 assert(HeaderData.isLoopHeader());
1129 Working[Index].Loop = HeaderData.Loop;
1130 HeaderData.Loop->Nodes.push_back(Index);
1131 DEBUG(dbgs() << " - loop = " << getBlockName(Header)
1132 << ": member = " << getBlockName(Index) << "\n");
1136 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInLoops() {
1137 // Visit loops with the deepest first, and the top-level loops last.
1138 for (auto L = Loops.rbegin(), E = Loops.rend(); L != E; ++L) {
1139 if (computeMassInLoop(*L))
1141 auto Next = std::next(L);
1142 computeIrreducibleMass(&*L, L.base());
1143 L = std::prev(Next);
1144 if (computeMassInLoop(*L))
1146 llvm_unreachable("unhandled irreducible control flow");
1151 bool BlockFrequencyInfoImpl<BT>::computeMassInLoop(LoopData &Loop) {
1152 // Compute mass in loop.
1153 DEBUG(dbgs() << "compute-mass-in-loop: " << getLoopName(Loop) << "\n");
1155 if (Loop.isIrreducible()) {
1156 DEBUG(dbgs() << "isIrreducible = true\n");
1158 unsigned NumHeadersWithWeight = 0;
1159 Optional<uint64_t> MinHeaderWeight;
1160 DenseSet<uint32_t> HeadersWithoutWeight;
1161 HeadersWithoutWeight.reserve(Loop.NumHeaders);
1162 for (uint32_t H = 0; H < Loop.NumHeaders; ++H) {
1163 auto &HeaderNode = Loop.Nodes[H];
1164 const BlockT *Block = getBlock(HeaderNode);
1165 IsIrrLoopHeader.set(Loop.Nodes[H].Index);
1166 Optional<uint64_t> HeaderWeight = Block->getIrrLoopHeaderWeight();
1167 if (!HeaderWeight) {
1168 DEBUG(dbgs() << "Missing irr loop header metadata on "
1169 << getBlockName(HeaderNode) << "\n");
1170 HeadersWithoutWeight.insert(H);
1173 DEBUG(dbgs() << getBlockName(HeaderNode)
1174 << " has irr loop header weight " << HeaderWeight.getValue()
1176 NumHeadersWithWeight++;
1177 uint64_t HeaderWeightValue = HeaderWeight.getValue();
1178 if (!MinHeaderWeight || HeaderWeightValue < MinHeaderWeight)
1179 MinHeaderWeight = HeaderWeightValue;
1180 if (HeaderWeightValue) {
1181 Dist.addLocal(HeaderNode, HeaderWeightValue);
1184 // As a heuristic, if some headers don't have a weight, give them the
1185 // minimium weight seen (not to disrupt the existing trends too much by
1186 // using a weight that's in the general range of the other headers' weights,
1187 // and the minimum seems to perform better than the average.)
1188 // FIXME: better update in the passes that drop the header weight.
1189 // If no headers have a weight, give them even weight (use weight 1).
1190 if (!MinHeaderWeight)
1191 MinHeaderWeight = 1;
1192 for (uint32_t H : HeadersWithoutWeight) {
1193 auto &HeaderNode = Loop.Nodes[H];
1194 assert(!getBlock(HeaderNode)->getIrrLoopHeaderWeight() &&
1195 "Shouldn't have a weight metadata");
1196 uint64_t MinWeight = MinHeaderWeight.getValue();
1197 DEBUG(dbgs() << "Giving weight " << MinWeight
1198 << " to " << getBlockName(HeaderNode) << "\n");
1200 Dist.addLocal(HeaderNode, MinWeight);
1202 distributeIrrLoopHeaderMass(Dist);
1203 for (const BlockNode &M : Loop.Nodes)
1204 if (!propagateMassToSuccessors(&Loop, M))
1205 llvm_unreachable("unhandled irreducible control flow");
1206 if (NumHeadersWithWeight == 0)
1207 // No headers have a metadata. Adjust header mass.
1208 adjustLoopHeaderMass(Loop);
1210 Working[Loop.getHeader().Index].getMass() = BlockMass::getFull();
1211 if (!propagateMassToSuccessors(&Loop, Loop.getHeader()))
1212 llvm_unreachable("irreducible control flow to loop header!?");
1213 for (const BlockNode &M : Loop.members())
1214 if (!propagateMassToSuccessors(&Loop, M))
1215 // Irreducible backedge.
1219 computeLoopScale(Loop);
1225 bool BlockFrequencyInfoImpl<BT>::tryToComputeMassInFunction() {
1226 // Compute mass in function.
1227 DEBUG(dbgs() << "compute-mass-in-function\n");
1228 assert(!Working.empty() && "no blocks in function");
1229 assert(!Working[0].isLoopHeader() && "entry block is a loop header");
1231 Working[0].getMass() = BlockMass::getFull();
1232 for (rpot_iterator I = rpot_begin(), IE = rpot_end(); I != IE; ++I) {
1233 // Check for nodes that have been packaged.
1234 BlockNode Node = getNode(I);
1235 if (Working[Node.Index].isPackaged())
1238 if (!propagateMassToSuccessors(nullptr, Node))
1244 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInFunction() {
1245 if (tryToComputeMassInFunction())
1247 computeIrreducibleMass(nullptr, Loops.begin());
1248 if (tryToComputeMassInFunction())
1250 llvm_unreachable("unhandled irreducible control flow");
1253 /// \note This should be a lambda, but that crashes GCC 4.7.
1254 namespace bfi_detail {
1256 template <class BT> struct BlockEdgesAdder {
1258 using LoopData = BlockFrequencyInfoImplBase::LoopData;
1259 using Successor = GraphTraits<const BlockT *>;
1261 const BlockFrequencyInfoImpl<BT> &BFI;
1263 explicit BlockEdgesAdder(const BlockFrequencyInfoImpl<BT> &BFI)
1266 void operator()(IrreducibleGraph &G, IrreducibleGraph::IrrNode &Irr,
1267 const LoopData *OuterLoop) {
1268 const BlockT *BB = BFI.RPOT[Irr.Node.Index];
1269 for (const auto Succ : children<const BlockT *>(BB))
1270 G.addEdge(Irr, BFI.getNode(Succ), OuterLoop);
1274 } // end namespace bfi_detail
1277 void BlockFrequencyInfoImpl<BT>::computeIrreducibleMass(
1278 LoopData *OuterLoop, std::list<LoopData>::iterator Insert) {
1279 DEBUG(dbgs() << "analyze-irreducible-in-";
1280 if (OuterLoop) dbgs() << "loop: " << getLoopName(*OuterLoop) << "\n";
1281 else dbgs() << "function\n");
1283 using namespace bfi_detail;
1285 // Ideally, addBlockEdges() would be declared here as a lambda, but that
1287 BlockEdgesAdder<BT> addBlockEdges(*this);
1288 IrreducibleGraph G(*this, OuterLoop, addBlockEdges);
1290 for (auto &L : analyzeIrreducible(G, OuterLoop, Insert))
1291 computeMassInLoop(L);
1295 updateLoopWithIrreducible(*OuterLoop);
1298 // A helper function that converts a branch probability into weight.
1299 inline uint32_t getWeightFromBranchProb(const BranchProbability Prob) {
1300 return Prob.getNumerator();
1305 BlockFrequencyInfoImpl<BT>::propagateMassToSuccessors(LoopData *OuterLoop,
1306 const BlockNode &Node) {
1307 DEBUG(dbgs() << " - node: " << getBlockName(Node) << "\n");
1308 // Calculate probability for successors.
1310 if (auto *Loop = Working[Node.Index].getPackagedLoop()) {
1311 assert(Loop != OuterLoop && "Cannot propagate mass in a packaged loop");
1312 if (!addLoopSuccessorsToDist(OuterLoop, *Loop, Dist))
1313 // Irreducible backedge.
1316 const BlockT *BB = getBlock(Node);
1317 for (auto SI = GraphTraits<const BlockT *>::child_begin(BB),
1318 SE = GraphTraits<const BlockT *>::child_end(BB);
1321 Dist, OuterLoop, Node, getNode(*SI),
1322 getWeightFromBranchProb(BPI->getEdgeProbability(BB, SI))))
1323 // Irreducible backedge.
1327 // Distribute mass to successors, saving exit and backedge data in the
1329 distributeMass(Node, OuterLoop, Dist);
1334 raw_ostream &BlockFrequencyInfoImpl<BT>::print(raw_ostream &OS) const {
1337 OS << "block-frequency-info: " << F->getName() << "\n";
1338 for (const BlockT &BB : *F) {
1339 OS << " - " << bfi_detail::getBlockName(&BB) << ": float = ";
1340 getFloatingBlockFreq(&BB).print(OS, 5)
1341 << ", int = " << getBlockFreq(&BB).getFrequency();
1342 if (Optional<uint64_t> ProfileCount =
1343 BlockFrequencyInfoImplBase::getBlockProfileCount(
1344 F->getFunction(), getNode(&BB)))
1345 OS << ", count = " << ProfileCount.getValue();
1346 if (Optional<uint64_t> IrrLoopHeaderWeight =
1347 BB.getIrrLoopHeaderWeight())
1348 OS << ", irr_loop_header_weight = " << IrrLoopHeaderWeight.getValue();
1352 // Add an extra newline for readability.
1357 // Graph trait base class for block frequency information graph
1360 enum GVDAGType { GVDT_None, GVDT_Fraction, GVDT_Integer, GVDT_Count };
1362 template <class BlockFrequencyInfoT, class BranchProbabilityInfoT>
1363 struct BFIDOTGraphTraitsBase : public DefaultDOTGraphTraits {
1364 using GTraits = GraphTraits<BlockFrequencyInfoT *>;
1365 using NodeRef = typename GTraits::NodeRef;
1366 using EdgeIter = typename GTraits::ChildIteratorType;
1367 using NodeIter = typename GTraits::nodes_iterator;
1369 uint64_t MaxFrequency = 0;
1371 explicit BFIDOTGraphTraitsBase(bool isSimple = false)
1372 : DefaultDOTGraphTraits(isSimple) {}
1374 static std::string getGraphName(const BlockFrequencyInfoT *G) {
1375 return G->getFunction()->getName();
1378 std::string getNodeAttributes(NodeRef Node, const BlockFrequencyInfoT *Graph,
1379 unsigned HotPercentThreshold = 0) {
1381 if (!HotPercentThreshold)
1384 // Compute MaxFrequency on the fly:
1385 if (!MaxFrequency) {
1386 for (NodeIter I = GTraits::nodes_begin(Graph),
1387 E = GTraits::nodes_end(Graph);
1391 std::max(MaxFrequency, Graph->getBlockFreq(N).getFrequency());
1394 BlockFrequency Freq = Graph->getBlockFreq(Node);
1395 BlockFrequency HotFreq =
1396 (BlockFrequency(MaxFrequency) *
1397 BranchProbability::getBranchProbability(HotPercentThreshold, 100));
1402 raw_string_ostream OS(Result);
1403 OS << "color=\"red\"";
1408 std::string getNodeLabel(NodeRef Node, const BlockFrequencyInfoT *Graph,
1409 GVDAGType GType, int layout_order = -1) {
1411 raw_string_ostream OS(Result);
1413 if (layout_order != -1)
1414 OS << Node->getName() << "[" << layout_order << "] : ";
1416 OS << Node->getName() << " : ";
1419 Graph->printBlockFreq(OS, Node);
1422 OS << Graph->getBlockFreq(Node).getFrequency();
1425 auto Count = Graph->getBlockProfileCount(Node);
1427 OS << Count.getValue();
1433 llvm_unreachable("If we are not supposed to render a graph we should "
1434 "never reach this point.");
1439 std::string getEdgeAttributes(NodeRef Node, EdgeIter EI,
1440 const BlockFrequencyInfoT *BFI,
1441 const BranchProbabilityInfoT *BPI,
1442 unsigned HotPercentThreshold = 0) {
1447 BranchProbability BP = BPI->getEdgeProbability(Node, EI);
1448 uint32_t N = BP.getNumerator();
1449 uint32_t D = BP.getDenominator();
1450 double Percent = 100.0 * N / D;
1451 raw_string_ostream OS(Str);
1452 OS << format("label=\"%.1f%%\"", Percent);
1454 if (HotPercentThreshold) {
1455 BlockFrequency EFreq = BFI->getBlockFreq(Node) * BP;
1456 BlockFrequency HotFreq = BlockFrequency(MaxFrequency) *
1457 BranchProbability(HotPercentThreshold, 100);
1459 if (EFreq >= HotFreq) {
1460 OS << ",color=\"red\"";
1469 } // end namespace llvm
1473 #endif // LLVM_ANALYSIS_BLOCKFREQUENCYINFOIMPL_H