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/PostOrderIterator.h"
20 #include "llvm/ADT/iterator_range.h"
21 #include "llvm/IR/BasicBlock.h"
22 #include "llvm/Support/BlockFrequency.h"
23 #include "llvm/Support/BranchProbability.h"
24 #include "llvm/Support/Debug.h"
25 #include "llvm/Support/ScaledNumber.h"
26 #include "llvm/Support/raw_ostream.h"
32 #define DEBUG_TYPE "block-freq"
37 class BranchProbabilityInfo;
41 class MachineBasicBlock;
42 class MachineBranchProbabilityInfo;
43 class MachineFunction;
45 class MachineLoopInfo;
47 namespace bfi_detail {
49 struct IrreducibleGraph;
51 // This is part of a workaround for a GCC 4.7 crash on lambdas.
52 template <class BT> struct BlockEdgesAdder;
54 /// \brief Mass of a block.
56 /// This class implements a sort of fixed-point fraction always between 0.0 and
57 /// 1.0. getMass() == UINT64_MAX indicates a value of 1.0.
59 /// Masses can be added and subtracted. Simple saturation arithmetic is used,
60 /// so arithmetic operations never overflow or underflow.
62 /// Masses can be multiplied. Multiplication treats full mass as 1.0 and uses
63 /// an inexpensive floating-point algorithm that's off-by-one (almost, but not
64 /// quite, maximum precision).
66 /// Masses can be scaled by \a BranchProbability at maximum precision.
71 BlockMass() : Mass(0) {}
72 explicit BlockMass(uint64_t Mass) : Mass(Mass) {}
74 static BlockMass getEmpty() { return BlockMass(); }
75 static BlockMass getFull() { return BlockMass(UINT64_MAX); }
77 uint64_t getMass() const { return Mass; }
79 bool isFull() const { return Mass == UINT64_MAX; }
80 bool isEmpty() const { return !Mass; }
82 bool operator!() const { return isEmpty(); }
84 /// \brief Add another mass.
86 /// Adds another mass, saturating at \a isFull() rather than overflowing.
87 BlockMass &operator+=(const BlockMass &X) {
88 uint64_t Sum = Mass + X.Mass;
89 Mass = Sum < Mass ? UINT64_MAX : Sum;
93 /// \brief Subtract another mass.
95 /// Subtracts another mass, saturating at \a isEmpty() rather than
97 BlockMass &operator-=(const BlockMass &X) {
98 uint64_t Diff = Mass - X.Mass;
99 Mass = Diff > Mass ? 0 : Diff;
103 BlockMass &operator*=(const BranchProbability &P) {
104 Mass = P.scale(Mass);
108 bool operator==(const BlockMass &X) const { return Mass == X.Mass; }
109 bool operator!=(const BlockMass &X) const { return Mass != X.Mass; }
110 bool operator<=(const BlockMass &X) const { return Mass <= X.Mass; }
111 bool operator>=(const BlockMass &X) const { return Mass >= X.Mass; }
112 bool operator<(const BlockMass &X) const { return Mass < X.Mass; }
113 bool operator>(const BlockMass &X) const { return Mass > X.Mass; }
115 /// \brief Convert to scaled number.
117 /// Convert to \a ScaledNumber. \a isFull() gives 1.0, while \a isEmpty()
118 /// gives slightly above 0.0.
119 ScaledNumber<uint64_t> toScaled() const;
122 raw_ostream &print(raw_ostream &OS) const;
125 inline BlockMass operator+(const BlockMass &L, const BlockMass &R) {
126 return BlockMass(L) += R;
128 inline BlockMass operator-(const BlockMass &L, const BlockMass &R) {
129 return BlockMass(L) -= R;
131 inline BlockMass operator*(const BlockMass &L, const BranchProbability &R) {
132 return BlockMass(L) *= R;
134 inline BlockMass operator*(const BranchProbability &L, const BlockMass &R) {
135 return BlockMass(R) *= L;
138 inline raw_ostream &operator<<(raw_ostream &OS, const BlockMass &X) {
142 } // end namespace bfi_detail
144 template <> struct isPodLike<bfi_detail::BlockMass> {
145 static const bool value = true;
148 /// \brief Base class for BlockFrequencyInfoImpl
150 /// BlockFrequencyInfoImplBase has supporting data structures and some
151 /// algorithms for BlockFrequencyInfoImplBase. Only algorithms that depend on
152 /// the block type (or that call such algorithms) are skipped here.
154 /// Nevertheless, the majority of the overall algorithm documention lives with
155 /// BlockFrequencyInfoImpl. See there for details.
156 class BlockFrequencyInfoImplBase {
158 typedef ScaledNumber<uint64_t> Scaled64;
159 typedef bfi_detail::BlockMass BlockMass;
161 /// \brief Representative of a block.
163 /// This is a simple wrapper around an index into the reverse-post-order
164 /// traversal of the blocks.
166 /// Unlike a block pointer, its order has meaning (location in the
167 /// topological sort) and it's class is the same regardless of block type.
169 typedef uint32_t IndexType;
172 bool operator==(const BlockNode &X) const { return Index == X.Index; }
173 bool operator!=(const BlockNode &X) const { return Index != X.Index; }
174 bool operator<=(const BlockNode &X) const { return Index <= X.Index; }
175 bool operator>=(const BlockNode &X) const { return Index >= X.Index; }
176 bool operator<(const BlockNode &X) const { return Index < X.Index; }
177 bool operator>(const BlockNode &X) const { return Index > X.Index; }
179 BlockNode() : Index(UINT32_MAX) {}
180 BlockNode(IndexType Index) : Index(Index) {}
182 bool isValid() const { return Index <= getMaxIndex(); }
183 static size_t getMaxIndex() { return UINT32_MAX - 1; }
186 /// \brief Stats about a block itself.
187 struct FrequencyData {
192 /// \brief Data about a loop.
194 /// Contains the data necessary to represent a loop as a pseudo-node once it's
197 typedef SmallVector<std::pair<BlockNode, BlockMass>, 4> ExitMap;
198 typedef SmallVector<BlockNode, 4> NodeList;
199 typedef SmallVector<BlockMass, 1> HeaderMassList;
200 LoopData *Parent; ///< The parent loop.
201 bool IsPackaged; ///< Whether this has been packaged.
202 uint32_t NumHeaders; ///< Number of headers.
203 ExitMap Exits; ///< Successor edges (and weights).
204 NodeList Nodes; ///< Header and the members of the loop.
205 HeaderMassList BackedgeMass; ///< Mass returned to each loop header.
209 LoopData(LoopData *Parent, const BlockNode &Header)
210 : Parent(Parent), IsPackaged(false), NumHeaders(1), Nodes(1, Header),
212 template <class It1, class It2>
213 LoopData(LoopData *Parent, It1 FirstHeader, It1 LastHeader, It2 FirstOther,
215 : Parent(Parent), IsPackaged(false), Nodes(FirstHeader, LastHeader) {
216 NumHeaders = Nodes.size();
217 Nodes.insert(Nodes.end(), FirstOther, LastOther);
218 BackedgeMass.resize(NumHeaders);
220 bool isHeader(const BlockNode &Node) const {
222 return std::binary_search(Nodes.begin(), Nodes.begin() + NumHeaders,
224 return Node == Nodes[0];
226 BlockNode getHeader() const { return Nodes[0]; }
227 bool isIrreducible() const { return NumHeaders > 1; }
229 HeaderMassList::difference_type getHeaderIndex(const BlockNode &B) {
230 assert(isHeader(B) && "this is only valid on loop header blocks");
232 return std::lower_bound(Nodes.begin(), Nodes.begin() + NumHeaders, B) -
237 NodeList::const_iterator members_begin() const {
238 return Nodes.begin() + NumHeaders;
240 NodeList::const_iterator members_end() const { return Nodes.end(); }
241 iterator_range<NodeList::const_iterator> members() const {
242 return make_range(members_begin(), members_end());
246 /// \brief Index of loop information.
248 BlockNode Node; ///< This node.
249 LoopData *Loop; ///< The loop this block is inside.
250 BlockMass Mass; ///< Mass distribution from the entry block.
252 WorkingData(const BlockNode &Node) : Node(Node), Loop(nullptr) {}
254 bool isLoopHeader() const { return Loop && Loop->isHeader(Node); }
255 bool isDoubleLoopHeader() const {
256 return isLoopHeader() && Loop->Parent && Loop->Parent->isIrreducible() &&
257 Loop->Parent->isHeader(Node);
260 LoopData *getContainingLoop() const {
263 if (!isDoubleLoopHeader())
265 return Loop->Parent->Parent;
268 /// \brief Resolve a node to its representative.
270 /// Get the node currently representing Node, which could be a containing
273 /// This function should only be called when distributing mass. As long as
274 /// there are no irreducible edges to Node, then it will have complexity
275 /// O(1) in this context.
277 /// In general, the complexity is O(L), where L is the number of loop
278 /// headers Node has been packaged into. Since this method is called in
279 /// the context of distributing mass, L will be the number of loop headers
280 /// an early exit edge jumps out of.
281 BlockNode getResolvedNode() const {
282 auto L = getPackagedLoop();
283 return L ? L->getHeader() : Node;
285 LoopData *getPackagedLoop() const {
286 if (!Loop || !Loop->IsPackaged)
289 while (L->Parent && L->Parent->IsPackaged)
294 /// \brief Get the appropriate mass for a node.
296 /// Get appropriate mass for Node. If Node is a loop-header (whose loop
297 /// has been packaged), returns the mass of its pseudo-node. If it's a
298 /// node inside a packaged loop, it returns the loop's mass.
299 BlockMass &getMass() {
302 if (!isADoublePackage())
304 return Loop->Parent->Mass;
307 /// \brief Has ContainingLoop been packaged up?
308 bool isPackaged() const { return getResolvedNode() != Node; }
309 /// \brief Has Loop been packaged up?
310 bool isAPackage() const { return isLoopHeader() && Loop->IsPackaged; }
311 /// \brief Has Loop been packaged up twice?
312 bool isADoublePackage() const {
313 return isDoubleLoopHeader() && Loop->Parent->IsPackaged;
317 /// \brief Unscaled probability weight.
319 /// Probability weight for an edge in the graph (including the
320 /// successor/target node).
322 /// All edges in the original function are 32-bit. However, exit edges from
323 /// loop packages are taken from 64-bit exit masses, so we need 64-bits of
324 /// space in general.
326 /// In addition to the raw weight amount, Weight stores the type of the edge
327 /// in the current context (i.e., the context of the loop being processed).
328 /// Is this a local edge within the loop, an exit from the loop, or a
329 /// backedge to the loop header?
331 enum DistType { Local, Exit, Backedge };
333 BlockNode TargetNode;
335 Weight() : Type(Local), Amount(0) {}
336 Weight(DistType Type, BlockNode TargetNode, uint64_t Amount)
337 : Type(Type), TargetNode(TargetNode), Amount(Amount) {}
340 /// \brief Distribution of unscaled probability weight.
342 /// Distribution of unscaled probability weight to a set of successors.
344 /// This class collates the successor edge weights for later processing.
346 /// \a DidOverflow indicates whether \a Total did overflow while adding to
347 /// the distribution. It should never overflow twice.
348 struct Distribution {
349 typedef SmallVector<Weight, 4> WeightList;
350 WeightList Weights; ///< Individual successor weights.
351 uint64_t Total; ///< Sum of all weights.
352 bool DidOverflow; ///< Whether \a Total did overflow.
354 Distribution() : Total(0), DidOverflow(false) {}
355 void addLocal(const BlockNode &Node, uint64_t Amount) {
356 add(Node, Amount, Weight::Local);
358 void addExit(const BlockNode &Node, uint64_t Amount) {
359 add(Node, Amount, Weight::Exit);
361 void addBackedge(const BlockNode &Node, uint64_t Amount) {
362 add(Node, Amount, Weight::Backedge);
365 /// \brief Normalize the distribution.
367 /// Combines multiple edges to the same \a Weight::TargetNode and scales
368 /// down so that \a Total fits into 32-bits.
370 /// This is linear in the size of \a Weights. For the vast majority of
371 /// cases, adjacent edge weights are combined by sorting WeightList and
372 /// combining adjacent weights. However, for very large edge lists an
373 /// auxiliary hash table is used.
377 void add(const BlockNode &Node, uint64_t Amount, Weight::DistType Type);
380 /// \brief Data about each block. This is used downstream.
381 std::vector<FrequencyData> Freqs;
383 /// \brief Loop data: see initializeLoops().
384 std::vector<WorkingData> Working;
386 /// \brief Indexed information about loops.
387 std::list<LoopData> Loops;
389 /// \brief Add all edges out of a packaged loop to the distribution.
391 /// Adds all edges from LocalLoopHead to Dist. Calls addToDist() to add each
394 /// \return \c true unless there's an irreducible backedge.
395 bool addLoopSuccessorsToDist(const LoopData *OuterLoop, LoopData &Loop,
398 /// \brief Add an edge to the distribution.
400 /// Adds an edge to Succ to Dist. If \c LoopHead.isValid(), then whether the
401 /// edge is local/exit/backedge is in the context of LoopHead. Otherwise,
402 /// every edge should be a local edge (since all the loops are packaged up).
404 /// \return \c true unless aborted due to an irreducible backedge.
405 bool addToDist(Distribution &Dist, const LoopData *OuterLoop,
406 const BlockNode &Pred, const BlockNode &Succ, uint64_t Weight);
408 LoopData &getLoopPackage(const BlockNode &Head) {
409 assert(Head.Index < Working.size());
410 assert(Working[Head.Index].isLoopHeader());
411 return *Working[Head.Index].Loop;
414 /// \brief Analyze irreducible SCCs.
416 /// Separate irreducible SCCs from \c G, which is an explict graph of \c
417 /// OuterLoop (or the top-level function, if \c OuterLoop is \c nullptr).
418 /// Insert them into \a Loops before \c Insert.
420 /// \return the \c LoopData nodes representing the irreducible SCCs.
421 iterator_range<std::list<LoopData>::iterator>
422 analyzeIrreducible(const bfi_detail::IrreducibleGraph &G, LoopData *OuterLoop,
423 std::list<LoopData>::iterator Insert);
425 /// \brief Update a loop after packaging irreducible SCCs inside of it.
427 /// Update \c OuterLoop. Before finding irreducible control flow, it was
428 /// partway through \a computeMassInLoop(), so \a LoopData::Exits and \a
429 /// LoopData::BackedgeMass need to be reset. Also, nodes that were packaged
430 /// up need to be removed from \a OuterLoop::Nodes.
431 void updateLoopWithIrreducible(LoopData &OuterLoop);
433 /// \brief Distribute mass according to a distribution.
435 /// Distributes the mass in Source according to Dist. If LoopHead.isValid(),
436 /// backedges and exits are stored in its entry in Loops.
438 /// Mass is distributed in parallel from two copies of the source mass.
439 void distributeMass(const BlockNode &Source, LoopData *OuterLoop,
442 /// \brief Compute the loop scale for a loop.
443 void computeLoopScale(LoopData &Loop);
445 /// Adjust the mass of all headers in an irreducible loop.
447 /// Initially, irreducible loops are assumed to distribute their mass
448 /// equally among its headers. This can lead to wrong frequency estimates
449 /// since some headers may be executed more frequently than others.
451 /// This adjusts header mass distribution so it matches the weights of
452 /// the backedges going into each of the loop headers.
453 void adjustLoopHeaderMass(LoopData &Loop);
455 /// \brief Package up a loop.
456 void packageLoop(LoopData &Loop);
458 /// \brief Unwrap loops.
461 /// \brief Finalize frequency metrics.
463 /// Calculates final frequencies and cleans up no-longer-needed data
465 void finalizeMetrics();
467 /// \brief Clear all memory.
470 virtual std::string getBlockName(const BlockNode &Node) const;
471 std::string getLoopName(const LoopData &Loop) const;
473 virtual raw_ostream &print(raw_ostream &OS) const { return OS; }
474 void dump() const { print(dbgs()); }
476 Scaled64 getFloatingBlockFreq(const BlockNode &Node) const;
478 BlockFrequency getBlockFreq(const BlockNode &Node) const;
480 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockNode &Node) const;
481 raw_ostream &printBlockFreq(raw_ostream &OS,
482 const BlockFrequency &Freq) const;
484 uint64_t getEntryFreq() const {
485 assert(!Freqs.empty());
486 return Freqs[0].Integer;
488 /// \brief Virtual destructor.
490 /// Need a virtual destructor to mask the compiler warning about
492 virtual ~BlockFrequencyInfoImplBase() {}
495 namespace bfi_detail {
496 template <class BlockT> struct TypeMap {};
497 template <> struct TypeMap<BasicBlock> {
498 typedef BasicBlock BlockT;
499 typedef Function FunctionT;
500 typedef BranchProbabilityInfo BranchProbabilityInfoT;
502 typedef LoopInfo LoopInfoT;
504 template <> struct TypeMap<MachineBasicBlock> {
505 typedef MachineBasicBlock BlockT;
506 typedef MachineFunction FunctionT;
507 typedef MachineBranchProbabilityInfo BranchProbabilityInfoT;
508 typedef MachineLoop LoopT;
509 typedef MachineLoopInfo LoopInfoT;
512 /// \brief Get the name of a MachineBasicBlock.
514 /// Get the name of a MachineBasicBlock. It's templated so that including from
515 /// CodeGen is unnecessary (that would be a layering issue).
517 /// This is used mainly for debug output. The name is similar to
518 /// MachineBasicBlock::getFullName(), but skips the name of the function.
519 template <class BlockT> std::string getBlockName(const BlockT *BB) {
520 assert(BB && "Unexpected nullptr");
521 auto MachineName = "BB" + Twine(BB->getNumber());
522 if (BB->getBasicBlock())
523 return (MachineName + "[" + BB->getName() + "]").str();
524 return MachineName.str();
526 /// \brief Get the name of a BasicBlock.
527 template <> inline std::string getBlockName(const BasicBlock *BB) {
528 assert(BB && "Unexpected nullptr");
529 return BB->getName().str();
532 /// \brief Graph of irreducible control flow.
534 /// This graph is used for determining the SCCs in a loop (or top-level
535 /// function) that has irreducible control flow.
537 /// During the block frequency algorithm, the local graphs are defined in a
538 /// light-weight way, deferring to the \a BasicBlock or \a MachineBasicBlock
539 /// graphs for most edges, but getting others from \a LoopData::ExitMap. The
540 /// latter only has successor information.
542 /// \a IrreducibleGraph makes this graph explicit. It's in a form that can use
543 /// \a GraphTraits (so that \a analyzeIrreducible() can use \a scc_iterator),
544 /// and it explicitly lists predecessors and successors. The initialization
545 /// that relies on \c MachineBasicBlock is defined in the header.
546 struct IrreducibleGraph {
547 typedef BlockFrequencyInfoImplBase BFIBase;
551 typedef BFIBase::BlockNode BlockNode;
555 std::deque<const IrrNode *> Edges;
556 IrrNode(const BlockNode &Node) : Node(Node), NumIn(0) {}
558 typedef std::deque<const IrrNode *>::const_iterator iterator;
559 iterator pred_begin() const { return Edges.begin(); }
560 iterator succ_begin() const { return Edges.begin() + NumIn; }
561 iterator pred_end() const { return succ_begin(); }
562 iterator succ_end() const { return Edges.end(); }
565 const IrrNode *StartIrr;
566 std::vector<IrrNode> Nodes;
567 SmallDenseMap<uint32_t, IrrNode *, 4> Lookup;
569 /// \brief Construct an explicit graph containing irreducible control flow.
571 /// Construct an explicit graph of the control flow in \c OuterLoop (or the
572 /// top-level function, if \c OuterLoop is \c nullptr). Uses \c
573 /// addBlockEdges to add block successors that have not been packaged into
576 /// \a BlockFrequencyInfoImpl::computeIrreducibleMass() is the only expected
578 template <class BlockEdgesAdder>
579 IrreducibleGraph(BFIBase &BFI, const BFIBase::LoopData *OuterLoop,
580 BlockEdgesAdder addBlockEdges)
581 : BFI(BFI), StartIrr(nullptr) {
582 initialize(OuterLoop, addBlockEdges);
585 template <class BlockEdgesAdder>
586 void initialize(const BFIBase::LoopData *OuterLoop,
587 BlockEdgesAdder addBlockEdges);
588 void addNodesInLoop(const BFIBase::LoopData &OuterLoop);
589 void addNodesInFunction();
590 void addNode(const BlockNode &Node) {
591 Nodes.emplace_back(Node);
592 BFI.Working[Node.Index].getMass() = BlockMass::getEmpty();
595 template <class BlockEdgesAdder>
596 void addEdges(const BlockNode &Node, const BFIBase::LoopData *OuterLoop,
597 BlockEdgesAdder addBlockEdges);
598 void addEdge(IrrNode &Irr, const BlockNode &Succ,
599 const BFIBase::LoopData *OuterLoop);
601 template <class BlockEdgesAdder>
602 void IrreducibleGraph::initialize(const BFIBase::LoopData *OuterLoop,
603 BlockEdgesAdder addBlockEdges) {
605 addNodesInLoop(*OuterLoop);
606 for (auto N : OuterLoop->Nodes)
607 addEdges(N, OuterLoop, addBlockEdges);
609 addNodesInFunction();
610 for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index)
611 addEdges(Index, OuterLoop, addBlockEdges);
613 StartIrr = Lookup[Start.Index];
615 template <class BlockEdgesAdder>
616 void IrreducibleGraph::addEdges(const BlockNode &Node,
617 const BFIBase::LoopData *OuterLoop,
618 BlockEdgesAdder addBlockEdges) {
619 auto L = Lookup.find(Node.Index);
620 if (L == Lookup.end())
622 IrrNode &Irr = *L->second;
623 const auto &Working = BFI.Working[Node.Index];
625 if (Working.isAPackage())
626 for (const auto &I : Working.Loop->Exits)
627 addEdge(Irr, I.first, OuterLoop);
629 addBlockEdges(*this, Irr, OuterLoop);
633 /// \brief Shared implementation for block frequency analysis.
635 /// This is a shared implementation of BlockFrequencyInfo and
636 /// MachineBlockFrequencyInfo, and calculates the relative frequencies of
639 /// LoopInfo defines a loop as a "non-trivial" SCC dominated by a single block,
640 /// which is called the header. A given loop, L, can have sub-loops, which are
641 /// loops within the subgraph of L that exclude its header. (A "trivial" SCC
642 /// consists of a single block that does not have a self-edge.)
644 /// In addition to loops, this algorithm has limited support for irreducible
645 /// SCCs, which are SCCs with multiple entry blocks. Irreducible SCCs are
646 /// discovered on they fly, and modelled as loops with multiple headers.
648 /// The headers of irreducible sub-SCCs consist of its entry blocks and all
649 /// nodes that are targets of a backedge within it (excluding backedges within
650 /// true sub-loops). Block frequency calculations act as if a block is
651 /// inserted that intercepts all the edges to the headers. All backedges and
652 /// entries point to this block. Its successors are the headers, which split
653 /// the frequency evenly.
655 /// This algorithm leverages BlockMass and ScaledNumber to maintain precision,
656 /// separates mass distribution from loop scaling, and dithers to eliminate
657 /// probability mass loss.
659 /// The implementation is split between BlockFrequencyInfoImpl, which knows the
660 /// type of graph being modelled (BasicBlock vs. MachineBasicBlock), and
661 /// BlockFrequencyInfoImplBase, which doesn't. The base class uses \a
662 /// BlockNode, a wrapper around a uint32_t. BlockNode is numbered from 0 in
663 /// reverse-post order. This gives two advantages: it's easy to compare the
664 /// relative ordering of two nodes, and maps keyed on BlockT can be represented
667 /// This algorithm is O(V+E), unless there is irreducible control flow, in
668 /// which case it's O(V*E) in the worst case.
670 /// These are the main stages:
672 /// 0. Reverse post-order traversal (\a initializeRPOT()).
674 /// Run a single post-order traversal and save it (in reverse) in RPOT.
675 /// All other stages make use of this ordering. Save a lookup from BlockT
676 /// to BlockNode (the index into RPOT) in Nodes.
678 /// 1. Loop initialization (\a initializeLoops()).
680 /// Translate LoopInfo/MachineLoopInfo into a form suitable for the rest of
681 /// the algorithm. In particular, store the immediate members of each loop
682 /// in reverse post-order.
684 /// 2. Calculate mass and scale in loops (\a computeMassInLoops()).
686 /// For each loop (bottom-up), distribute mass through the DAG resulting
687 /// from ignoring backedges and treating sub-loops as a single pseudo-node.
688 /// Track the backedge mass distributed to the loop header, and use it to
689 /// calculate the loop scale (number of loop iterations). Immediate
690 /// members that represent sub-loops will already have been visited and
691 /// packaged into a pseudo-node.
693 /// Distributing mass in a loop is a reverse-post-order traversal through
694 /// the loop. Start by assigning full mass to the Loop header. For each
695 /// node in the loop:
697 /// - Fetch and categorize the weight distribution for its successors.
698 /// If this is a packaged-subloop, the weight distribution is stored
699 /// in \a LoopData::Exits. Otherwise, fetch it from
700 /// BranchProbabilityInfo.
702 /// - Each successor is categorized as \a Weight::Local, a local edge
703 /// within the current loop, \a Weight::Backedge, a backedge to the
704 /// loop header, or \a Weight::Exit, any successor outside the loop.
705 /// The weight, the successor, and its category are stored in \a
706 /// Distribution. There can be multiple edges to each successor.
708 /// - If there's a backedge to a non-header, there's an irreducible SCC.
709 /// The usual flow is temporarily aborted. \a
710 /// computeIrreducibleMass() finds the irreducible SCCs within the
711 /// loop, packages them up, and restarts the flow.
713 /// - Normalize the distribution: scale weights down so that their sum
714 /// is 32-bits, and coalesce multiple edges to the same node.
716 /// - Distribute the mass accordingly, dithering to minimize mass loss,
717 /// as described in \a distributeMass().
719 /// In the case of irreducible loops, instead of a single loop header,
720 /// there will be several. The computation of backedge masses is similar
721 /// but instead of having a single backedge mass, there will be one
722 /// backedge per loop header. In these cases, each backedge will carry
723 /// a mass proportional to the edge weights along the corresponding
726 /// At the end of propagation, the full mass assigned to the loop will be
727 /// distributed among the loop headers proportionally according to the
728 /// mass flowing through their backedges.
730 /// Finally, calculate the loop scale from the accumulated backedge mass.
732 /// 3. Distribute mass in the function (\a computeMassInFunction()).
734 /// Finally, distribute mass through the DAG resulting from packaging all
735 /// loops in the function. This uses the same algorithm as distributing
736 /// mass in a loop, except that there are no exit or backedge edges.
738 /// 4. Unpackage loops (\a unwrapLoops()).
740 /// Initialize each block's frequency to a floating point representation of
743 /// Visit loops top-down, scaling the frequencies of its immediate members
744 /// by the loop's pseudo-node's frequency.
746 /// 5. Convert frequencies to a 64-bit range (\a finalizeMetrics()).
748 /// Using the min and max frequencies as a guide, translate floating point
749 /// frequencies to an appropriate range in uint64_t.
751 /// It has some known flaws.
753 /// - The model of irreducible control flow is a rough approximation.
755 /// Modelling irreducible control flow exactly involves setting up and
756 /// solving a group of infinite geometric series. Such precision is
757 /// unlikely to be worthwhile, since most of our algorithms give up on
758 /// irreducible control flow anyway.
760 /// Nevertheless, we might find that we need to get closer. Here's a sort
761 /// of TODO list for the model with diminishing returns, to be completed as
764 /// - The headers for the \a LoopData representing an irreducible SCC
765 /// include non-entry blocks. When these extra blocks exist, they
766 /// indicate a self-contained irreducible sub-SCC. We could treat them
767 /// as sub-loops, rather than arbitrarily shoving the problematic
768 /// blocks into the headers of the main irreducible SCC.
770 /// - Entry frequencies are assumed to be evenly split between the
771 /// headers of a given irreducible SCC, which is the only option if we
772 /// need to compute mass in the SCC before its parent loop. Instead,
773 /// we could partially compute mass in the parent loop, and stop when
774 /// we get to the SCC. Here, we have the correct ratio of entry
775 /// masses, which we can use to adjust their relative frequencies.
776 /// Compute mass in the SCC, and then continue propagation in the
779 /// - We can propagate mass iteratively through the SCC, for some fixed
780 /// number of iterations. Each iteration starts by assigning the entry
781 /// blocks their backedge mass from the prior iteration. The final
782 /// mass for each block (and each exit, and the total backedge mass
783 /// used for computing loop scale) is the sum of all iterations.
784 /// (Running this until fixed point would "solve" the geometric
785 /// series by simulation.)
786 template <class BT> class BlockFrequencyInfoImpl : BlockFrequencyInfoImplBase {
787 typedef typename bfi_detail::TypeMap<BT>::BlockT BlockT;
788 typedef typename bfi_detail::TypeMap<BT>::FunctionT FunctionT;
789 typedef typename bfi_detail::TypeMap<BT>::BranchProbabilityInfoT
790 BranchProbabilityInfoT;
791 typedef typename bfi_detail::TypeMap<BT>::LoopT LoopT;
792 typedef typename bfi_detail::TypeMap<BT>::LoopInfoT LoopInfoT;
794 // This is part of a workaround for a GCC 4.7 crash on lambdas.
795 friend struct bfi_detail::BlockEdgesAdder<BT>;
797 typedef GraphTraits<const BlockT *> Successor;
798 typedef GraphTraits<Inverse<const BlockT *>> Predecessor;
800 const BranchProbabilityInfoT *BPI;
804 // All blocks in reverse postorder.
805 std::vector<const BlockT *> RPOT;
806 DenseMap<const BlockT *, BlockNode> Nodes;
808 typedef typename std::vector<const BlockT *>::const_iterator rpot_iterator;
810 rpot_iterator rpot_begin() const { return RPOT.begin(); }
811 rpot_iterator rpot_end() const { return RPOT.end(); }
813 size_t getIndex(const rpot_iterator &I) const { return I - rpot_begin(); }
815 BlockNode getNode(const rpot_iterator &I) const {
816 return BlockNode(getIndex(I));
818 BlockNode getNode(const BlockT *BB) const { return Nodes.lookup(BB); }
820 const BlockT *getBlock(const BlockNode &Node) const {
821 assert(Node.Index < RPOT.size());
822 return RPOT[Node.Index];
825 /// \brief Run (and save) a post-order traversal.
827 /// Saves a reverse post-order traversal of all the nodes in \a F.
828 void initializeRPOT();
830 /// \brief Initialize loop data.
832 /// Build up \a Loops using \a LoopInfo. \a LoopInfo gives us a mapping from
833 /// each block to the deepest loop it's in, but we need the inverse. For each
834 /// loop, we store in reverse post-order its "immediate" members, defined as
835 /// the header, the headers of immediate sub-loops, and all other blocks in
836 /// the loop that are not in sub-loops.
837 void initializeLoops();
839 /// \brief Propagate to a block's successors.
841 /// In the context of distributing mass through \c OuterLoop, divide the mass
842 /// currently assigned to \c Node between its successors.
844 /// \return \c true unless there's an irreducible backedge.
845 bool propagateMassToSuccessors(LoopData *OuterLoop, const BlockNode &Node);
847 /// \brief Compute mass in a particular loop.
849 /// Assign mass to \c Loop's header, and then for each block in \c Loop in
850 /// reverse post-order, distribute mass to its successors. Only visits nodes
851 /// that have not been packaged into sub-loops.
853 /// \pre \a computeMassInLoop() has been called for each subloop of \c Loop.
854 /// \return \c true unless there's an irreducible backedge.
855 bool computeMassInLoop(LoopData &Loop);
857 /// \brief Try to compute mass in the top-level function.
859 /// Assign mass to the entry block, and then for each block in reverse
860 /// post-order, distribute mass to its successors. Skips nodes that have
861 /// been packaged into loops.
863 /// \pre \a computeMassInLoops() has been called.
864 /// \return \c true unless there's an irreducible backedge.
865 bool tryToComputeMassInFunction();
867 /// \brief Compute mass in (and package up) irreducible SCCs.
869 /// Find the irreducible SCCs in \c OuterLoop, add them to \a Loops (in front
870 /// of \c Insert), and call \a computeMassInLoop() on each of them.
872 /// If \c OuterLoop is \c nullptr, it refers to the top-level function.
874 /// \pre \a computeMassInLoop() has been called for each subloop of \c
876 /// \pre \c Insert points at the last loop successfully processed by \a
877 /// computeMassInLoop().
878 /// \pre \c OuterLoop has irreducible SCCs.
879 void computeIrreducibleMass(LoopData *OuterLoop,
880 std::list<LoopData>::iterator Insert);
882 /// \brief Compute mass in all loops.
884 /// For each loop bottom-up, call \a computeMassInLoop().
886 /// \a computeMassInLoop() aborts (and returns \c false) on loops that
887 /// contain a irreducible sub-SCCs. Use \a computeIrreducibleMass() and then
888 /// re-enter \a computeMassInLoop().
890 /// \post \a computeMassInLoop() has returned \c true for every loop.
891 void computeMassInLoops();
893 /// \brief Compute mass in the top-level function.
895 /// Uses \a tryToComputeMassInFunction() and \a computeIrreducibleMass() to
896 /// compute mass in the top-level function.
898 /// \post \a tryToComputeMassInFunction() has returned \c true.
899 void computeMassInFunction();
901 std::string getBlockName(const BlockNode &Node) const override {
902 return bfi_detail::getBlockName(getBlock(Node));
906 const FunctionT *getFunction() const { return F; }
908 void calculate(const FunctionT &F, const BranchProbabilityInfoT &BPI,
909 const LoopInfoT &LI);
910 BlockFrequencyInfoImpl() : BPI(nullptr), LI(nullptr), F(nullptr) {}
912 using BlockFrequencyInfoImplBase::getEntryFreq;
913 BlockFrequency getBlockFreq(const BlockT *BB) const {
914 return BlockFrequencyInfoImplBase::getBlockFreq(getNode(BB));
916 Scaled64 getFloatingBlockFreq(const BlockT *BB) const {
917 return BlockFrequencyInfoImplBase::getFloatingBlockFreq(getNode(BB));
920 /// \brief Print the frequencies for the current function.
922 /// Prints the frequencies for the blocks in the current function.
924 /// Blocks are printed in the natural iteration order of the function, rather
925 /// than reverse post-order. This provides two advantages: writing -analyze
926 /// tests is easier (since blocks come out in source order), and even
927 /// unreachable blocks are printed.
929 /// \a BlockFrequencyInfoImplBase::print() only knows reverse post-order, so
930 /// we need to override it here.
931 raw_ostream &print(raw_ostream &OS) const override;
932 using BlockFrequencyInfoImplBase::dump;
934 using BlockFrequencyInfoImplBase::printBlockFreq;
935 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockT *BB) const {
936 return BlockFrequencyInfoImplBase::printBlockFreq(OS, getNode(BB));
941 void BlockFrequencyInfoImpl<BT>::calculate(const FunctionT &F,
942 const BranchProbabilityInfoT &BPI,
943 const LoopInfoT &LI) {
944 // Save the parameters.
949 // Clean up left-over data structures.
950 BlockFrequencyInfoImplBase::clear();
955 DEBUG(dbgs() << "\nblock-frequency: " << F.getName() << "\n================="
956 << std::string(F.getName().size(), '=') << "\n");
960 // Visit loops in post-order to find the local mass distribution, and then do
961 // the full function.
962 computeMassInLoops();
963 computeMassInFunction();
968 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeRPOT() {
969 const BlockT *Entry = F->begin();
970 RPOT.reserve(F->size());
971 std::copy(po_begin(Entry), po_end(Entry), std::back_inserter(RPOT));
972 std::reverse(RPOT.begin(), RPOT.end());
974 assert(RPOT.size() - 1 <= BlockNode::getMaxIndex() &&
975 "More nodes in function than Block Frequency Info supports");
977 DEBUG(dbgs() << "reverse-post-order-traversal\n");
978 for (rpot_iterator I = rpot_begin(), E = rpot_end(); I != E; ++I) {
979 BlockNode Node = getNode(I);
980 DEBUG(dbgs() << " - " << getIndex(I) << ": " << getBlockName(Node) << "\n");
984 Working.reserve(RPOT.size());
985 for (size_t Index = 0; Index < RPOT.size(); ++Index)
986 Working.emplace_back(Index);
987 Freqs.resize(RPOT.size());
990 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeLoops() {
991 DEBUG(dbgs() << "loop-detection\n");
995 // Visit loops top down and assign them an index.
996 std::deque<std::pair<const LoopT *, LoopData *>> Q;
997 for (const LoopT *L : *LI)
998 Q.emplace_back(L, nullptr);
1000 const LoopT *Loop = Q.front().first;
1001 LoopData *Parent = Q.front().second;
1004 BlockNode Header = getNode(Loop->getHeader());
1005 assert(Header.isValid());
1007 Loops.emplace_back(Parent, Header);
1008 Working[Header.Index].Loop = &Loops.back();
1009 DEBUG(dbgs() << " - loop = " << getBlockName(Header) << "\n");
1011 for (const LoopT *L : *Loop)
1012 Q.emplace_back(L, &Loops.back());
1015 // Visit nodes in reverse post-order and add them to their deepest containing
1017 for (size_t Index = 0; Index < RPOT.size(); ++Index) {
1018 // Loop headers have already been mostly mapped.
1019 if (Working[Index].isLoopHeader()) {
1020 LoopData *ContainingLoop = Working[Index].getContainingLoop();
1022 ContainingLoop->Nodes.push_back(Index);
1026 const LoopT *Loop = LI->getLoopFor(RPOT[Index]);
1030 // Add this node to its containing loop's member list.
1031 BlockNode Header = getNode(Loop->getHeader());
1032 assert(Header.isValid());
1033 const auto &HeaderData = Working[Header.Index];
1034 assert(HeaderData.isLoopHeader());
1036 Working[Index].Loop = HeaderData.Loop;
1037 HeaderData.Loop->Nodes.push_back(Index);
1038 DEBUG(dbgs() << " - loop = " << getBlockName(Header)
1039 << ": member = " << getBlockName(Index) << "\n");
1043 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInLoops() {
1044 // Visit loops with the deepest first, and the top-level loops last.
1045 for (auto L = Loops.rbegin(), E = Loops.rend(); L != E; ++L) {
1046 if (computeMassInLoop(*L))
1048 auto Next = std::next(L);
1049 computeIrreducibleMass(&*L, L.base());
1050 L = std::prev(Next);
1051 if (computeMassInLoop(*L))
1053 llvm_unreachable("unhandled irreducible control flow");
1058 bool BlockFrequencyInfoImpl<BT>::computeMassInLoop(LoopData &Loop) {
1059 // Compute mass in loop.
1060 DEBUG(dbgs() << "compute-mass-in-loop: " << getLoopName(Loop) << "\n");
1062 if (Loop.isIrreducible()) {
1063 BlockMass Remaining = BlockMass::getFull();
1064 for (uint32_t H = 0; H < Loop.NumHeaders; ++H) {
1065 auto &Mass = Working[Loop.Nodes[H].Index].getMass();
1066 Mass = Remaining * BranchProbability(1, Loop.NumHeaders - H);
1069 for (const BlockNode &M : Loop.Nodes)
1070 if (!propagateMassToSuccessors(&Loop, M))
1071 llvm_unreachable("unhandled irreducible control flow");
1073 adjustLoopHeaderMass(Loop);
1075 Working[Loop.getHeader().Index].getMass() = BlockMass::getFull();
1076 if (!propagateMassToSuccessors(&Loop, Loop.getHeader()))
1077 llvm_unreachable("irreducible control flow to loop header!?");
1078 for (const BlockNode &M : Loop.members())
1079 if (!propagateMassToSuccessors(&Loop, M))
1080 // Irreducible backedge.
1084 computeLoopScale(Loop);
1090 bool BlockFrequencyInfoImpl<BT>::tryToComputeMassInFunction() {
1091 // Compute mass in function.
1092 DEBUG(dbgs() << "compute-mass-in-function\n");
1093 assert(!Working.empty() && "no blocks in function");
1094 assert(!Working[0].isLoopHeader() && "entry block is a loop header");
1096 Working[0].getMass() = BlockMass::getFull();
1097 for (rpot_iterator I = rpot_begin(), IE = rpot_end(); I != IE; ++I) {
1098 // Check for nodes that have been packaged.
1099 BlockNode Node = getNode(I);
1100 if (Working[Node.Index].isPackaged())
1103 if (!propagateMassToSuccessors(nullptr, Node))
1109 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInFunction() {
1110 if (tryToComputeMassInFunction())
1112 computeIrreducibleMass(nullptr, Loops.begin());
1113 if (tryToComputeMassInFunction())
1115 llvm_unreachable("unhandled irreducible control flow");
1118 /// \note This should be a lambda, but that crashes GCC 4.7.
1119 namespace bfi_detail {
1120 template <class BT> struct BlockEdgesAdder {
1122 typedef BlockFrequencyInfoImplBase::LoopData LoopData;
1123 typedef GraphTraits<const BlockT *> Successor;
1125 const BlockFrequencyInfoImpl<BT> &BFI;
1126 explicit BlockEdgesAdder(const BlockFrequencyInfoImpl<BT> &BFI)
1128 void operator()(IrreducibleGraph &G, IrreducibleGraph::IrrNode &Irr,
1129 const LoopData *OuterLoop) {
1130 const BlockT *BB = BFI.RPOT[Irr.Node.Index];
1131 for (auto I = Successor::child_begin(BB), E = Successor::child_end(BB);
1133 G.addEdge(Irr, BFI.getNode(*I), OuterLoop);
1138 void BlockFrequencyInfoImpl<BT>::computeIrreducibleMass(
1139 LoopData *OuterLoop, std::list<LoopData>::iterator Insert) {
1140 DEBUG(dbgs() << "analyze-irreducible-in-";
1141 if (OuterLoop) dbgs() << "loop: " << getLoopName(*OuterLoop) << "\n";
1142 else dbgs() << "function\n");
1144 using namespace bfi_detail;
1145 // Ideally, addBlockEdges() would be declared here as a lambda, but that
1147 BlockEdgesAdder<BT> addBlockEdges(*this);
1148 IrreducibleGraph G(*this, OuterLoop, addBlockEdges);
1150 for (auto &L : analyzeIrreducible(G, OuterLoop, Insert))
1151 computeMassInLoop(L);
1155 updateLoopWithIrreducible(*OuterLoop);
1160 BlockFrequencyInfoImpl<BT>::propagateMassToSuccessors(LoopData *OuterLoop,
1161 const BlockNode &Node) {
1162 DEBUG(dbgs() << " - node: " << getBlockName(Node) << "\n");
1163 // Calculate probability for successors.
1165 if (auto *Loop = Working[Node.Index].getPackagedLoop()) {
1166 assert(Loop != OuterLoop && "Cannot propagate mass in a packaged loop");
1167 if (!addLoopSuccessorsToDist(OuterLoop, *Loop, Dist))
1168 // Irreducible backedge.
1171 const BlockT *BB = getBlock(Node);
1172 for (auto SI = Successor::child_begin(BB), SE = Successor::child_end(BB);
1174 // Do not dereference SI, or getEdgeWeight() is linear in the number of
1176 if (!addToDist(Dist, OuterLoop, Node, getNode(*SI),
1177 BPI->getEdgeWeight(BB, SI)))
1178 // Irreducible backedge.
1182 // Distribute mass to successors, saving exit and backedge data in the
1184 distributeMass(Node, OuterLoop, Dist);
1189 raw_ostream &BlockFrequencyInfoImpl<BT>::print(raw_ostream &OS) const {
1192 OS << "block-frequency-info: " << F->getName() << "\n";
1193 for (const BlockT &BB : *F)
1194 OS << " - " << bfi_detail::getBlockName(&BB)
1195 << ": float = " << getFloatingBlockFreq(&BB)
1196 << ", int = " << getBlockFreq(&BB).getFrequency() << "\n";
1198 // Add an extra newline for readability.
1203 } // end namespace llvm