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+=(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-=(BlockMass X) {
98 uint64_t Diff = Mass - X.Mass;
99 Mass = Diff > Mass ? 0 : Diff;
103 BlockMass &operator*=(BranchProbability P) {
104 Mass = P.scale(Mass);
108 bool operator==(BlockMass X) const { return Mass == X.Mass; }
109 bool operator!=(BlockMass X) const { return Mass != X.Mass; }
110 bool operator<=(BlockMass X) const { return Mass <= X.Mass; }
111 bool operator>=(BlockMass X) const { return Mass >= X.Mass; }
112 bool operator<(BlockMass X) const { return Mass < X.Mass; }
113 bool operator>(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+(BlockMass L, BlockMass R) {
126 return BlockMass(L) += R;
128 inline BlockMass operator-(BlockMass L, BlockMass R) {
129 return BlockMass(L) -= R;
131 inline BlockMass operator*(BlockMass L, BranchProbability R) {
132 return BlockMass(L) *= R;
134 inline BlockMass operator*(BranchProbability L, BlockMass R) {
135 return BlockMass(R) *= L;
138 inline raw_ostream &operator<<(raw_ostream &OS, 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 void setBlockFreq(const BlockNode &Node, uint64_t Freq);
482 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockNode &Node) const;
483 raw_ostream &printBlockFreq(raw_ostream &OS,
484 const BlockFrequency &Freq) const;
486 uint64_t getEntryFreq() const {
487 assert(!Freqs.empty());
488 return Freqs[0].Integer;
490 /// \brief Virtual destructor.
492 /// Need a virtual destructor to mask the compiler warning about
494 virtual ~BlockFrequencyInfoImplBase() {}
497 namespace bfi_detail {
498 template <class BlockT> struct TypeMap {};
499 template <> struct TypeMap<BasicBlock> {
500 typedef BasicBlock BlockT;
501 typedef Function FunctionT;
502 typedef BranchProbabilityInfo BranchProbabilityInfoT;
504 typedef LoopInfo LoopInfoT;
506 template <> struct TypeMap<MachineBasicBlock> {
507 typedef MachineBasicBlock BlockT;
508 typedef MachineFunction FunctionT;
509 typedef MachineBranchProbabilityInfo BranchProbabilityInfoT;
510 typedef MachineLoop LoopT;
511 typedef MachineLoopInfo LoopInfoT;
514 /// \brief Get the name of a MachineBasicBlock.
516 /// Get the name of a MachineBasicBlock. It's templated so that including from
517 /// CodeGen is unnecessary (that would be a layering issue).
519 /// This is used mainly for debug output. The name is similar to
520 /// MachineBasicBlock::getFullName(), but skips the name of the function.
521 template <class BlockT> std::string getBlockName(const BlockT *BB) {
522 assert(BB && "Unexpected nullptr");
523 auto MachineName = "BB" + Twine(BB->getNumber());
524 if (BB->getBasicBlock())
525 return (MachineName + "[" + BB->getName() + "]").str();
526 return MachineName.str();
528 /// \brief Get the name of a BasicBlock.
529 template <> inline std::string getBlockName(const BasicBlock *BB) {
530 assert(BB && "Unexpected nullptr");
531 return BB->getName().str();
534 /// \brief Graph of irreducible control flow.
536 /// This graph is used for determining the SCCs in a loop (or top-level
537 /// function) that has irreducible control flow.
539 /// During the block frequency algorithm, the local graphs are defined in a
540 /// light-weight way, deferring to the \a BasicBlock or \a MachineBasicBlock
541 /// graphs for most edges, but getting others from \a LoopData::ExitMap. The
542 /// latter only has successor information.
544 /// \a IrreducibleGraph makes this graph explicit. It's in a form that can use
545 /// \a GraphTraits (so that \a analyzeIrreducible() can use \a scc_iterator),
546 /// and it explicitly lists predecessors and successors. The initialization
547 /// that relies on \c MachineBasicBlock is defined in the header.
548 struct IrreducibleGraph {
549 typedef BlockFrequencyInfoImplBase BFIBase;
553 typedef BFIBase::BlockNode BlockNode;
557 std::deque<const IrrNode *> Edges;
558 IrrNode(const BlockNode &Node) : Node(Node), NumIn(0) {}
560 typedef std::deque<const IrrNode *>::const_iterator iterator;
561 iterator pred_begin() const { return Edges.begin(); }
562 iterator succ_begin() const { return Edges.begin() + NumIn; }
563 iterator pred_end() const { return succ_begin(); }
564 iterator succ_end() const { return Edges.end(); }
567 const IrrNode *StartIrr;
568 std::vector<IrrNode> Nodes;
569 SmallDenseMap<uint32_t, IrrNode *, 4> Lookup;
571 /// \brief Construct an explicit graph containing irreducible control flow.
573 /// Construct an explicit graph of the control flow in \c OuterLoop (or the
574 /// top-level function, if \c OuterLoop is \c nullptr). Uses \c
575 /// addBlockEdges to add block successors that have not been packaged into
578 /// \a BlockFrequencyInfoImpl::computeIrreducibleMass() is the only expected
580 template <class BlockEdgesAdder>
581 IrreducibleGraph(BFIBase &BFI, const BFIBase::LoopData *OuterLoop,
582 BlockEdgesAdder addBlockEdges)
583 : BFI(BFI), StartIrr(nullptr) {
584 initialize(OuterLoop, addBlockEdges);
587 template <class BlockEdgesAdder>
588 void initialize(const BFIBase::LoopData *OuterLoop,
589 BlockEdgesAdder addBlockEdges);
590 void addNodesInLoop(const BFIBase::LoopData &OuterLoop);
591 void addNodesInFunction();
592 void addNode(const BlockNode &Node) {
593 Nodes.emplace_back(Node);
594 BFI.Working[Node.Index].getMass() = BlockMass::getEmpty();
597 template <class BlockEdgesAdder>
598 void addEdges(const BlockNode &Node, const BFIBase::LoopData *OuterLoop,
599 BlockEdgesAdder addBlockEdges);
600 void addEdge(IrrNode &Irr, const BlockNode &Succ,
601 const BFIBase::LoopData *OuterLoop);
603 template <class BlockEdgesAdder>
604 void IrreducibleGraph::initialize(const BFIBase::LoopData *OuterLoop,
605 BlockEdgesAdder addBlockEdges) {
607 addNodesInLoop(*OuterLoop);
608 for (auto N : OuterLoop->Nodes)
609 addEdges(N, OuterLoop, addBlockEdges);
611 addNodesInFunction();
612 for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index)
613 addEdges(Index, OuterLoop, addBlockEdges);
615 StartIrr = Lookup[Start.Index];
617 template <class BlockEdgesAdder>
618 void IrreducibleGraph::addEdges(const BlockNode &Node,
619 const BFIBase::LoopData *OuterLoop,
620 BlockEdgesAdder addBlockEdges) {
621 auto L = Lookup.find(Node.Index);
622 if (L == Lookup.end())
624 IrrNode &Irr = *L->second;
625 const auto &Working = BFI.Working[Node.Index];
627 if (Working.isAPackage())
628 for (const auto &I : Working.Loop->Exits)
629 addEdge(Irr, I.first, OuterLoop);
631 addBlockEdges(*this, Irr, OuterLoop);
635 /// \brief Shared implementation for block frequency analysis.
637 /// This is a shared implementation of BlockFrequencyInfo and
638 /// MachineBlockFrequencyInfo, and calculates the relative frequencies of
641 /// LoopInfo defines a loop as a "non-trivial" SCC dominated by a single block,
642 /// which is called the header. A given loop, L, can have sub-loops, which are
643 /// loops within the subgraph of L that exclude its header. (A "trivial" SCC
644 /// consists of a single block that does not have a self-edge.)
646 /// In addition to loops, this algorithm has limited support for irreducible
647 /// SCCs, which are SCCs with multiple entry blocks. Irreducible SCCs are
648 /// discovered on they fly, and modelled as loops with multiple headers.
650 /// The headers of irreducible sub-SCCs consist of its entry blocks and all
651 /// nodes that are targets of a backedge within it (excluding backedges within
652 /// true sub-loops). Block frequency calculations act as if a block is
653 /// inserted that intercepts all the edges to the headers. All backedges and
654 /// entries point to this block. Its successors are the headers, which split
655 /// the frequency evenly.
657 /// This algorithm leverages BlockMass and ScaledNumber to maintain precision,
658 /// separates mass distribution from loop scaling, and dithers to eliminate
659 /// probability mass loss.
661 /// The implementation is split between BlockFrequencyInfoImpl, which knows the
662 /// type of graph being modelled (BasicBlock vs. MachineBasicBlock), and
663 /// BlockFrequencyInfoImplBase, which doesn't. The base class uses \a
664 /// BlockNode, a wrapper around a uint32_t. BlockNode is numbered from 0 in
665 /// reverse-post order. This gives two advantages: it's easy to compare the
666 /// relative ordering of two nodes, and maps keyed on BlockT can be represented
669 /// This algorithm is O(V+E), unless there is irreducible control flow, in
670 /// which case it's O(V*E) in the worst case.
672 /// These are the main stages:
674 /// 0. Reverse post-order traversal (\a initializeRPOT()).
676 /// Run a single post-order traversal and save it (in reverse) in RPOT.
677 /// All other stages make use of this ordering. Save a lookup from BlockT
678 /// to BlockNode (the index into RPOT) in Nodes.
680 /// 1. Loop initialization (\a initializeLoops()).
682 /// Translate LoopInfo/MachineLoopInfo into a form suitable for the rest of
683 /// the algorithm. In particular, store the immediate members of each loop
684 /// in reverse post-order.
686 /// 2. Calculate mass and scale in loops (\a computeMassInLoops()).
688 /// For each loop (bottom-up), distribute mass through the DAG resulting
689 /// from ignoring backedges and treating sub-loops as a single pseudo-node.
690 /// Track the backedge mass distributed to the loop header, and use it to
691 /// calculate the loop scale (number of loop iterations). Immediate
692 /// members that represent sub-loops will already have been visited and
693 /// packaged into a pseudo-node.
695 /// Distributing mass in a loop is a reverse-post-order traversal through
696 /// the loop. Start by assigning full mass to the Loop header. For each
697 /// node in the loop:
699 /// - Fetch and categorize the weight distribution for its successors.
700 /// If this is a packaged-subloop, the weight distribution is stored
701 /// in \a LoopData::Exits. Otherwise, fetch it from
702 /// BranchProbabilityInfo.
704 /// - Each successor is categorized as \a Weight::Local, a local edge
705 /// within the current loop, \a Weight::Backedge, a backedge to the
706 /// loop header, or \a Weight::Exit, any successor outside the loop.
707 /// The weight, the successor, and its category are stored in \a
708 /// Distribution. There can be multiple edges to each successor.
710 /// - If there's a backedge to a non-header, there's an irreducible SCC.
711 /// The usual flow is temporarily aborted. \a
712 /// computeIrreducibleMass() finds the irreducible SCCs within the
713 /// loop, packages them up, and restarts the flow.
715 /// - Normalize the distribution: scale weights down so that their sum
716 /// is 32-bits, and coalesce multiple edges to the same node.
718 /// - Distribute the mass accordingly, dithering to minimize mass loss,
719 /// as described in \a distributeMass().
721 /// In the case of irreducible loops, instead of a single loop header,
722 /// there will be several. The computation of backedge masses is similar
723 /// but instead of having a single backedge mass, there will be one
724 /// backedge per loop header. In these cases, each backedge will carry
725 /// a mass proportional to the edge weights along the corresponding
728 /// At the end of propagation, the full mass assigned to the loop will be
729 /// distributed among the loop headers proportionally according to the
730 /// mass flowing through their backedges.
732 /// Finally, calculate the loop scale from the accumulated backedge mass.
734 /// 3. Distribute mass in the function (\a computeMassInFunction()).
736 /// Finally, distribute mass through the DAG resulting from packaging all
737 /// loops in the function. This uses the same algorithm as distributing
738 /// mass in a loop, except that there are no exit or backedge edges.
740 /// 4. Unpackage loops (\a unwrapLoops()).
742 /// Initialize each block's frequency to a floating point representation of
745 /// Visit loops top-down, scaling the frequencies of its immediate members
746 /// by the loop's pseudo-node's frequency.
748 /// 5. Convert frequencies to a 64-bit range (\a finalizeMetrics()).
750 /// Using the min and max frequencies as a guide, translate floating point
751 /// frequencies to an appropriate range in uint64_t.
753 /// It has some known flaws.
755 /// - The model of irreducible control flow is a rough approximation.
757 /// Modelling irreducible control flow exactly involves setting up and
758 /// solving a group of infinite geometric series. Such precision is
759 /// unlikely to be worthwhile, since most of our algorithms give up on
760 /// irreducible control flow anyway.
762 /// Nevertheless, we might find that we need to get closer. Here's a sort
763 /// of TODO list for the model with diminishing returns, to be completed as
766 /// - The headers for the \a LoopData representing an irreducible SCC
767 /// include non-entry blocks. When these extra blocks exist, they
768 /// indicate a self-contained irreducible sub-SCC. We could treat them
769 /// as sub-loops, rather than arbitrarily shoving the problematic
770 /// blocks into the headers of the main irreducible SCC.
772 /// - Entry frequencies are assumed to be evenly split between the
773 /// headers of a given irreducible SCC, which is the only option if we
774 /// need to compute mass in the SCC before its parent loop. Instead,
775 /// we could partially compute mass in the parent loop, and stop when
776 /// we get to the SCC. Here, we have the correct ratio of entry
777 /// masses, which we can use to adjust their relative frequencies.
778 /// Compute mass in the SCC, and then continue propagation in the
781 /// - We can propagate mass iteratively through the SCC, for some fixed
782 /// number of iterations. Each iteration starts by assigning the entry
783 /// blocks their backedge mass from the prior iteration. The final
784 /// mass for each block (and each exit, and the total backedge mass
785 /// used for computing loop scale) is the sum of all iterations.
786 /// (Running this until fixed point would "solve" the geometric
787 /// series by simulation.)
788 template <class BT> class BlockFrequencyInfoImpl : BlockFrequencyInfoImplBase {
789 typedef typename bfi_detail::TypeMap<BT>::BlockT BlockT;
790 typedef typename bfi_detail::TypeMap<BT>::FunctionT FunctionT;
791 typedef typename bfi_detail::TypeMap<BT>::BranchProbabilityInfoT
792 BranchProbabilityInfoT;
793 typedef typename bfi_detail::TypeMap<BT>::LoopT LoopT;
794 typedef typename bfi_detail::TypeMap<BT>::LoopInfoT LoopInfoT;
796 // This is part of a workaround for a GCC 4.7 crash on lambdas.
797 friend struct bfi_detail::BlockEdgesAdder<BT>;
799 typedef GraphTraits<const BlockT *> Successor;
800 typedef GraphTraits<Inverse<const BlockT *>> Predecessor;
802 const BranchProbabilityInfoT *BPI;
806 // All blocks in reverse postorder.
807 std::vector<const BlockT *> RPOT;
808 DenseMap<const BlockT *, BlockNode> Nodes;
810 typedef typename std::vector<const BlockT *>::const_iterator rpot_iterator;
812 rpot_iterator rpot_begin() const { return RPOT.begin(); }
813 rpot_iterator rpot_end() const { return RPOT.end(); }
815 size_t getIndex(const rpot_iterator &I) const { return I - rpot_begin(); }
817 BlockNode getNode(const rpot_iterator &I) const {
818 return BlockNode(getIndex(I));
820 BlockNode getNode(const BlockT *BB) const { return Nodes.lookup(BB); }
822 const BlockT *getBlock(const BlockNode &Node) const {
823 assert(Node.Index < RPOT.size());
824 return RPOT[Node.Index];
827 /// \brief Run (and save) a post-order traversal.
829 /// Saves a reverse post-order traversal of all the nodes in \a F.
830 void initializeRPOT();
832 /// \brief Initialize loop data.
834 /// Build up \a Loops using \a LoopInfo. \a LoopInfo gives us a mapping from
835 /// each block to the deepest loop it's in, but we need the inverse. For each
836 /// loop, we store in reverse post-order its "immediate" members, defined as
837 /// the header, the headers of immediate sub-loops, and all other blocks in
838 /// the loop that are not in sub-loops.
839 void initializeLoops();
841 /// \brief Propagate to a block's successors.
843 /// In the context of distributing mass through \c OuterLoop, divide the mass
844 /// currently assigned to \c Node between its successors.
846 /// \return \c true unless there's an irreducible backedge.
847 bool propagateMassToSuccessors(LoopData *OuterLoop, const BlockNode &Node);
849 /// \brief Compute mass in a particular loop.
851 /// Assign mass to \c Loop's header, and then for each block in \c Loop in
852 /// reverse post-order, distribute mass to its successors. Only visits nodes
853 /// that have not been packaged into sub-loops.
855 /// \pre \a computeMassInLoop() has been called for each subloop of \c Loop.
856 /// \return \c true unless there's an irreducible backedge.
857 bool computeMassInLoop(LoopData &Loop);
859 /// \brief Try to compute mass in the top-level function.
861 /// Assign mass to the entry block, and then for each block in reverse
862 /// post-order, distribute mass to its successors. Skips nodes that have
863 /// been packaged into loops.
865 /// \pre \a computeMassInLoops() has been called.
866 /// \return \c true unless there's an irreducible backedge.
867 bool tryToComputeMassInFunction();
869 /// \brief Compute mass in (and package up) irreducible SCCs.
871 /// Find the irreducible SCCs in \c OuterLoop, add them to \a Loops (in front
872 /// of \c Insert), and call \a computeMassInLoop() on each of them.
874 /// If \c OuterLoop is \c nullptr, it refers to the top-level function.
876 /// \pre \a computeMassInLoop() has been called for each subloop of \c
878 /// \pre \c Insert points at the last loop successfully processed by \a
879 /// computeMassInLoop().
880 /// \pre \c OuterLoop has irreducible SCCs.
881 void computeIrreducibleMass(LoopData *OuterLoop,
882 std::list<LoopData>::iterator Insert);
884 /// \brief Compute mass in all loops.
886 /// For each loop bottom-up, call \a computeMassInLoop().
888 /// \a computeMassInLoop() aborts (and returns \c false) on loops that
889 /// contain a irreducible sub-SCCs. Use \a computeIrreducibleMass() and then
890 /// re-enter \a computeMassInLoop().
892 /// \post \a computeMassInLoop() has returned \c true for every loop.
893 void computeMassInLoops();
895 /// \brief Compute mass in the top-level function.
897 /// Uses \a tryToComputeMassInFunction() and \a computeIrreducibleMass() to
898 /// compute mass in the top-level function.
900 /// \post \a tryToComputeMassInFunction() has returned \c true.
901 void computeMassInFunction();
903 std::string getBlockName(const BlockNode &Node) const override {
904 return bfi_detail::getBlockName(getBlock(Node));
908 const FunctionT *getFunction() const { return F; }
910 void calculate(const FunctionT &F, const BranchProbabilityInfoT &BPI,
911 const LoopInfoT &LI);
912 BlockFrequencyInfoImpl() : BPI(nullptr), LI(nullptr), F(nullptr) {}
914 using BlockFrequencyInfoImplBase::getEntryFreq;
915 BlockFrequency getBlockFreq(const BlockT *BB) const {
916 return BlockFrequencyInfoImplBase::getBlockFreq(getNode(BB));
918 void setBlockFreq(const BlockT *BB, uint64_t Freq);
919 Scaled64 getFloatingBlockFreq(const BlockT *BB) const {
920 return BlockFrequencyInfoImplBase::getFloatingBlockFreq(getNode(BB));
923 /// \brief Print the frequencies for the current function.
925 /// Prints the frequencies for the blocks in the current function.
927 /// Blocks are printed in the natural iteration order of the function, rather
928 /// than reverse post-order. This provides two advantages: writing -analyze
929 /// tests is easier (since blocks come out in source order), and even
930 /// unreachable blocks are printed.
932 /// \a BlockFrequencyInfoImplBase::print() only knows reverse post-order, so
933 /// we need to override it here.
934 raw_ostream &print(raw_ostream &OS) const override;
935 using BlockFrequencyInfoImplBase::dump;
937 using BlockFrequencyInfoImplBase::printBlockFreq;
938 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockT *BB) const {
939 return BlockFrequencyInfoImplBase::printBlockFreq(OS, getNode(BB));
944 void BlockFrequencyInfoImpl<BT>::calculate(const FunctionT &F,
945 const BranchProbabilityInfoT &BPI,
946 const LoopInfoT &LI) {
947 // Save the parameters.
952 // Clean up left-over data structures.
953 BlockFrequencyInfoImplBase::clear();
958 DEBUG(dbgs() << "\nblock-frequency: " << F.getName() << "\n================="
959 << std::string(F.getName().size(), '=') << "\n");
963 // Visit loops in post-order to find the local mass distribution, and then do
964 // the full function.
965 computeMassInLoops();
966 computeMassInFunction();
972 void BlockFrequencyInfoImpl<BT>::setBlockFreq(const BlockT *BB, uint64_t Freq) {
974 BlockFrequencyInfoImplBase::setBlockFreq(getNode(BB), Freq);
976 // If BB is a newly added block after BFI is done, we need to create a new
977 // BlockNode for it assigned with a new index. The index can be determined
978 // by the size of Freqs.
979 BlockNode NewNode(Freqs.size());
981 Freqs.emplace_back();
982 BlockFrequencyInfoImplBase::setBlockFreq(NewNode, Freq);
986 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeRPOT() {
987 const BlockT *Entry = &F->front();
988 RPOT.reserve(F->size());
989 std::copy(po_begin(Entry), po_end(Entry), std::back_inserter(RPOT));
990 std::reverse(RPOT.begin(), RPOT.end());
992 assert(RPOT.size() - 1 <= BlockNode::getMaxIndex() &&
993 "More nodes in function than Block Frequency Info supports");
995 DEBUG(dbgs() << "reverse-post-order-traversal\n");
996 for (rpot_iterator I = rpot_begin(), E = rpot_end(); I != E; ++I) {
997 BlockNode Node = getNode(I);
998 DEBUG(dbgs() << " - " << getIndex(I) << ": " << getBlockName(Node) << "\n");
1002 Working.reserve(RPOT.size());
1003 for (size_t Index = 0; Index < RPOT.size(); ++Index)
1004 Working.emplace_back(Index);
1005 Freqs.resize(RPOT.size());
1008 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeLoops() {
1009 DEBUG(dbgs() << "loop-detection\n");
1013 // Visit loops top down and assign them an index.
1014 std::deque<std::pair<const LoopT *, LoopData *>> Q;
1015 for (const LoopT *L : *LI)
1016 Q.emplace_back(L, nullptr);
1017 while (!Q.empty()) {
1018 const LoopT *Loop = Q.front().first;
1019 LoopData *Parent = Q.front().second;
1022 BlockNode Header = getNode(Loop->getHeader());
1023 assert(Header.isValid());
1025 Loops.emplace_back(Parent, Header);
1026 Working[Header.Index].Loop = &Loops.back();
1027 DEBUG(dbgs() << " - loop = " << getBlockName(Header) << "\n");
1029 for (const LoopT *L : *Loop)
1030 Q.emplace_back(L, &Loops.back());
1033 // Visit nodes in reverse post-order and add them to their deepest containing
1035 for (size_t Index = 0; Index < RPOT.size(); ++Index) {
1036 // Loop headers have already been mostly mapped.
1037 if (Working[Index].isLoopHeader()) {
1038 LoopData *ContainingLoop = Working[Index].getContainingLoop();
1040 ContainingLoop->Nodes.push_back(Index);
1044 const LoopT *Loop = LI->getLoopFor(RPOT[Index]);
1048 // Add this node to its containing loop's member list.
1049 BlockNode Header = getNode(Loop->getHeader());
1050 assert(Header.isValid());
1051 const auto &HeaderData = Working[Header.Index];
1052 assert(HeaderData.isLoopHeader());
1054 Working[Index].Loop = HeaderData.Loop;
1055 HeaderData.Loop->Nodes.push_back(Index);
1056 DEBUG(dbgs() << " - loop = " << getBlockName(Header)
1057 << ": member = " << getBlockName(Index) << "\n");
1061 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInLoops() {
1062 // Visit loops with the deepest first, and the top-level loops last.
1063 for (auto L = Loops.rbegin(), E = Loops.rend(); L != E; ++L) {
1064 if (computeMassInLoop(*L))
1066 auto Next = std::next(L);
1067 computeIrreducibleMass(&*L, L.base());
1068 L = std::prev(Next);
1069 if (computeMassInLoop(*L))
1071 llvm_unreachable("unhandled irreducible control flow");
1076 bool BlockFrequencyInfoImpl<BT>::computeMassInLoop(LoopData &Loop) {
1077 // Compute mass in loop.
1078 DEBUG(dbgs() << "compute-mass-in-loop: " << getLoopName(Loop) << "\n");
1080 if (Loop.isIrreducible()) {
1081 BlockMass Remaining = BlockMass::getFull();
1082 for (uint32_t H = 0; H < Loop.NumHeaders; ++H) {
1083 auto &Mass = Working[Loop.Nodes[H].Index].getMass();
1084 Mass = Remaining * BranchProbability(1, Loop.NumHeaders - H);
1087 for (const BlockNode &M : Loop.Nodes)
1088 if (!propagateMassToSuccessors(&Loop, M))
1089 llvm_unreachable("unhandled irreducible control flow");
1091 adjustLoopHeaderMass(Loop);
1093 Working[Loop.getHeader().Index].getMass() = BlockMass::getFull();
1094 if (!propagateMassToSuccessors(&Loop, Loop.getHeader()))
1095 llvm_unreachable("irreducible control flow to loop header!?");
1096 for (const BlockNode &M : Loop.members())
1097 if (!propagateMassToSuccessors(&Loop, M))
1098 // Irreducible backedge.
1102 computeLoopScale(Loop);
1108 bool BlockFrequencyInfoImpl<BT>::tryToComputeMassInFunction() {
1109 // Compute mass in function.
1110 DEBUG(dbgs() << "compute-mass-in-function\n");
1111 assert(!Working.empty() && "no blocks in function");
1112 assert(!Working[0].isLoopHeader() && "entry block is a loop header");
1114 Working[0].getMass() = BlockMass::getFull();
1115 for (rpot_iterator I = rpot_begin(), IE = rpot_end(); I != IE; ++I) {
1116 // Check for nodes that have been packaged.
1117 BlockNode Node = getNode(I);
1118 if (Working[Node.Index].isPackaged())
1121 if (!propagateMassToSuccessors(nullptr, Node))
1127 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInFunction() {
1128 if (tryToComputeMassInFunction())
1130 computeIrreducibleMass(nullptr, Loops.begin());
1131 if (tryToComputeMassInFunction())
1133 llvm_unreachable("unhandled irreducible control flow");
1136 /// \note This should be a lambda, but that crashes GCC 4.7.
1137 namespace bfi_detail {
1138 template <class BT> struct BlockEdgesAdder {
1140 typedef BlockFrequencyInfoImplBase::LoopData LoopData;
1141 typedef GraphTraits<const BlockT *> Successor;
1143 const BlockFrequencyInfoImpl<BT> &BFI;
1144 explicit BlockEdgesAdder(const BlockFrequencyInfoImpl<BT> &BFI)
1146 void operator()(IrreducibleGraph &G, IrreducibleGraph::IrrNode &Irr,
1147 const LoopData *OuterLoop) {
1148 const BlockT *BB = BFI.RPOT[Irr.Node.Index];
1149 for (auto I = Successor::child_begin(BB), E = Successor::child_end(BB);
1151 G.addEdge(Irr, BFI.getNode(*I), OuterLoop);
1156 void BlockFrequencyInfoImpl<BT>::computeIrreducibleMass(
1157 LoopData *OuterLoop, std::list<LoopData>::iterator Insert) {
1158 DEBUG(dbgs() << "analyze-irreducible-in-";
1159 if (OuterLoop) dbgs() << "loop: " << getLoopName(*OuterLoop) << "\n";
1160 else dbgs() << "function\n");
1162 using namespace bfi_detail;
1163 // Ideally, addBlockEdges() would be declared here as a lambda, but that
1165 BlockEdgesAdder<BT> addBlockEdges(*this);
1166 IrreducibleGraph G(*this, OuterLoop, addBlockEdges);
1168 for (auto &L : analyzeIrreducible(G, OuterLoop, Insert))
1169 computeMassInLoop(L);
1173 updateLoopWithIrreducible(*OuterLoop);
1177 // A helper function that converts a branch probability into weight.
1178 inline uint32_t getWeightFromBranchProb(const BranchProbability Prob) {
1179 return Prob.getNumerator();
1185 BlockFrequencyInfoImpl<BT>::propagateMassToSuccessors(LoopData *OuterLoop,
1186 const BlockNode &Node) {
1187 DEBUG(dbgs() << " - node: " << getBlockName(Node) << "\n");
1188 // Calculate probability for successors.
1190 if (auto *Loop = Working[Node.Index].getPackagedLoop()) {
1191 assert(Loop != OuterLoop && "Cannot propagate mass in a packaged loop");
1192 if (!addLoopSuccessorsToDist(OuterLoop, *Loop, Dist))
1193 // Irreducible backedge.
1196 const BlockT *BB = getBlock(Node);
1197 for (auto SI = Successor::child_begin(BB), SE = Successor::child_end(BB);
1199 if (!addToDist(Dist, OuterLoop, Node, getNode(*SI),
1200 getWeightFromBranchProb(BPI->getEdgeProbability(BB, SI))))
1201 // Irreducible backedge.
1205 // Distribute mass to successors, saving exit and backedge data in the
1207 distributeMass(Node, OuterLoop, Dist);
1212 raw_ostream &BlockFrequencyInfoImpl<BT>::print(raw_ostream &OS) const {
1215 OS << "block-frequency-info: " << F->getName() << "\n";
1216 for (const BlockT &BB : *F) {
1217 OS << " - " << bfi_detail::getBlockName(&BB) << ": float = ";
1218 getFloatingBlockFreq(&BB).print(OS, 5)
1219 << ", int = " << getBlockFreq(&BB).getFrequency() << "\n";
1222 // Add an extra newline for readability.
1227 } // end namespace llvm