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"
34 //===----------------------------------------------------------------------===//
36 // BlockMass definition.
38 // TODO: Make this private to BlockFrequencyInfoImpl or delete.
40 //===----------------------------------------------------------------------===//
43 /// \brief Mass of a block.
45 /// This class implements a sort of fixed-point fraction always between 0.0 and
46 /// 1.0. getMass() == UINT64_MAX indicates a value of 1.0.
48 /// Masses can be added and subtracted. Simple saturation arithmetic is used,
49 /// so arithmetic operations never overflow or underflow.
51 /// Masses can be multiplied. Multiplication treats full mass as 1.0 and uses
52 /// an inexpensive floating-point algorithm that's off-by-one (almost, but not
53 /// quite, maximum precision).
55 /// Masses can be scaled by \a BranchProbability at maximum precision.
60 BlockMass() : Mass(0) {}
61 explicit BlockMass(uint64_t Mass) : Mass(Mass) {}
63 static BlockMass getEmpty() { return BlockMass(); }
64 static BlockMass getFull() { return BlockMass(UINT64_MAX); }
66 uint64_t getMass() const { return Mass; }
68 bool isFull() const { return Mass == UINT64_MAX; }
69 bool isEmpty() const { return !Mass; }
71 bool operator!() const { return isEmpty(); }
73 /// \brief Add another mass.
75 /// Adds another mass, saturating at \a isFull() rather than overflowing.
76 BlockMass &operator+=(const BlockMass &X) {
77 uint64_t Sum = Mass + X.Mass;
78 Mass = Sum < Mass ? UINT64_MAX : Sum;
82 /// \brief Subtract another mass.
84 /// Subtracts another mass, saturating at \a isEmpty() rather than
86 BlockMass &operator-=(const BlockMass &X) {
87 uint64_t Diff = Mass - X.Mass;
88 Mass = Diff > Mass ? 0 : Diff;
92 BlockMass &operator*=(const BranchProbability &P) {
97 bool operator==(const BlockMass &X) const { return Mass == X.Mass; }
98 bool operator!=(const BlockMass &X) const { return Mass != X.Mass; }
99 bool operator<=(const BlockMass &X) const { return Mass <= X.Mass; }
100 bool operator>=(const BlockMass &X) const { return Mass >= X.Mass; }
101 bool operator<(const BlockMass &X) const { return Mass < X.Mass; }
102 bool operator>(const BlockMass &X) const { return Mass > X.Mass; }
104 /// \brief Convert to scaled number.
106 /// Convert to \a ScaledNumber. \a isFull() gives 1.0, while \a isEmpty()
107 /// gives slightly above 0.0.
108 ScaledNumber<uint64_t> toScaled() const;
111 raw_ostream &print(raw_ostream &OS) const;
114 inline BlockMass operator+(const BlockMass &L, const BlockMass &R) {
115 return BlockMass(L) += R;
117 inline BlockMass operator-(const BlockMass &L, const BlockMass &R) {
118 return BlockMass(L) -= R;
120 inline BlockMass operator*(const BlockMass &L, const BranchProbability &R) {
121 return BlockMass(L) *= R;
123 inline BlockMass operator*(const BranchProbability &L, const BlockMass &R) {
124 return BlockMass(R) *= L;
127 inline raw_ostream &operator<<(raw_ostream &OS, const BlockMass &X) {
131 template <> struct isPodLike<BlockMass> {
132 static const bool value = true;
135 } // end namespace llvm
137 //===----------------------------------------------------------------------===//
139 // BlockFrequencyInfoImpl definition.
141 //===----------------------------------------------------------------------===//
145 class BranchProbabilityInfo;
149 class MachineBasicBlock;
150 class MachineBranchProbabilityInfo;
151 class MachineFunction;
153 class MachineLoopInfo;
155 namespace bfi_detail {
156 struct IrreducibleGraph;
158 // This is part of a workaround for a GCC 4.7 crash on lambdas.
159 template <class BT> struct BlockEdgesAdder;
162 /// \brief Base class for BlockFrequencyInfoImpl
164 /// BlockFrequencyInfoImplBase has supporting data structures and some
165 /// algorithms for BlockFrequencyInfoImplBase. Only algorithms that depend on
166 /// the block type (or that call such algorithms) are skipped here.
168 /// Nevertheless, the majority of the overall algorithm documention lives with
169 /// BlockFrequencyInfoImpl. See there for details.
170 class BlockFrequencyInfoImplBase {
172 typedef ScaledNumber<uint64_t> Scaled64;
174 /// \brief Representative of a block.
176 /// This is a simple wrapper around an index into the reverse-post-order
177 /// traversal of the blocks.
179 /// Unlike a block pointer, its order has meaning (location in the
180 /// topological sort) and it's class is the same regardless of block type.
182 typedef uint32_t IndexType;
185 bool operator==(const BlockNode &X) const { return Index == X.Index; }
186 bool operator!=(const BlockNode &X) const { return Index != X.Index; }
187 bool operator<=(const BlockNode &X) const { return Index <= X.Index; }
188 bool operator>=(const BlockNode &X) const { return Index >= X.Index; }
189 bool operator<(const BlockNode &X) const { return Index < X.Index; }
190 bool operator>(const BlockNode &X) const { return Index > X.Index; }
192 BlockNode() : Index(UINT32_MAX) {}
193 BlockNode(IndexType Index) : Index(Index) {}
195 bool isValid() const { return Index <= getMaxIndex(); }
196 static size_t getMaxIndex() { return UINT32_MAX - 1; }
199 /// \brief Stats about a block itself.
200 struct FrequencyData {
205 /// \brief Data about a loop.
207 /// Contains the data necessary to represent represent a loop as a
208 /// pseudo-node once it's packaged.
210 typedef SmallVector<std::pair<BlockNode, BlockMass>, 4> ExitMap;
211 typedef SmallVector<BlockNode, 4> NodeList;
212 LoopData *Parent; ///< The parent loop.
213 bool IsPackaged; ///< Whether this has been packaged.
214 uint32_t NumHeaders; ///< Number of headers.
215 ExitMap Exits; ///< Successor edges (and weights).
216 NodeList Nodes; ///< Header and the members of the loop.
217 BlockMass BackedgeMass; ///< Mass returned to loop header.
221 LoopData(LoopData *Parent, const BlockNode &Header)
222 : Parent(Parent), IsPackaged(false), NumHeaders(1), Nodes(1, Header) {}
223 template <class It1, class It2>
224 LoopData(LoopData *Parent, It1 FirstHeader, It1 LastHeader, It2 FirstOther,
226 : Parent(Parent), IsPackaged(false), Nodes(FirstHeader, LastHeader) {
227 NumHeaders = Nodes.size();
228 Nodes.insert(Nodes.end(), FirstOther, LastOther);
230 bool isHeader(const BlockNode &Node) const {
232 return std::binary_search(Nodes.begin(), Nodes.begin() + NumHeaders,
234 return Node == Nodes[0];
236 BlockNode getHeader() const { return Nodes[0]; }
237 bool isIrreducible() const { return NumHeaders > 1; }
239 NodeList::const_iterator members_begin() const {
240 return Nodes.begin() + NumHeaders;
242 NodeList::const_iterator members_end() const { return Nodes.end(); }
243 iterator_range<NodeList::const_iterator> members() const {
244 return make_range(members_begin(), members_end());
248 /// \brief Index of loop information.
250 BlockNode Node; ///< This node.
251 LoopData *Loop; ///< The loop this block is inside.
252 BlockMass Mass; ///< Mass distribution from the entry block.
254 WorkingData(const BlockNode &Node) : Node(Node), Loop(nullptr) {}
256 bool isLoopHeader() const { return Loop && Loop->isHeader(Node); }
257 bool isDoubleLoopHeader() const {
258 return isLoopHeader() && Loop->Parent && Loop->Parent->isIrreducible() &&
259 Loop->Parent->isHeader(Node);
262 LoopData *getContainingLoop() const {
265 if (!isDoubleLoopHeader())
267 return Loop->Parent->Parent;
270 /// \brief Resolve a node to its representative.
272 /// Get the node currently representing Node, which could be a containing
275 /// This function should only be called when distributing mass. As long as
276 /// there are no irreducilbe edges to Node, then it will have complexity
277 /// O(1) in this context.
279 /// In general, the complexity is O(L), where L is the number of loop
280 /// headers Node has been packaged into. Since this method is called in
281 /// the context of distributing mass, L will be the number of loop headers
282 /// an early exit edge jumps out of.
283 BlockNode getResolvedNode() const {
284 auto L = getPackagedLoop();
285 return L ? L->getHeader() : Node;
287 LoopData *getPackagedLoop() const {
288 if (!Loop || !Loop->IsPackaged)
291 while (L->Parent && L->Parent->IsPackaged)
296 /// \brief Get the appropriate mass for a node.
298 /// Get appropriate mass for Node. If Node is a loop-header (whose loop
299 /// has been packaged), returns the mass of its pseudo-node. If it's a
300 /// node inside a packaged loop, it returns the loop's mass.
301 BlockMass &getMass() {
304 if (!isADoublePackage())
306 return Loop->Parent->Mass;
309 /// \brief Has ContainingLoop been packaged up?
310 bool isPackaged() const { return getResolvedNode() != Node; }
311 /// \brief Has Loop been packaged up?
312 bool isAPackage() const { return isLoopHeader() && Loop->IsPackaged; }
313 /// \brief Has Loop been packaged up twice?
314 bool isADoublePackage() const {
315 return isDoubleLoopHeader() && Loop->Parent->IsPackaged;
319 /// \brief Unscaled probability weight.
321 /// Probability weight for an edge in the graph (including the
322 /// successor/target node).
324 /// All edges in the original function are 32-bit. However, exit edges from
325 /// loop packages are taken from 64-bit exit masses, so we need 64-bits of
326 /// space in general.
328 /// In addition to the raw weight amount, Weight stores the type of the edge
329 /// in the current context (i.e., the context of the loop being processed).
330 /// Is this a local edge within the loop, an exit from the loop, or a
331 /// backedge to the loop header?
333 enum DistType { Local, Exit, Backedge };
335 BlockNode TargetNode;
337 Weight() : Type(Local), Amount(0) {}
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 /// \brief Package up a loop.
446 void packageLoop(LoopData &Loop);
448 /// \brief Unwrap loops.
451 /// \brief Finalize frequency metrics.
453 /// Calculates final frequencies and cleans up no-longer-needed data
455 void finalizeMetrics();
457 /// \brief Clear all memory.
460 virtual std::string getBlockName(const BlockNode &Node) const;
461 std::string getLoopName(const LoopData &Loop) const;
463 virtual raw_ostream &print(raw_ostream &OS) const { return OS; }
464 void dump() const { print(dbgs()); }
466 Scaled64 getFloatingBlockFreq(const BlockNode &Node) const;
468 BlockFrequency getBlockFreq(const BlockNode &Node) const;
470 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockNode &Node) const;
471 raw_ostream &printBlockFreq(raw_ostream &OS,
472 const BlockFrequency &Freq) const;
474 uint64_t getEntryFreq() const {
475 assert(!Freqs.empty());
476 return Freqs[0].Integer;
478 /// \brief Virtual destructor.
480 /// Need a virtual destructor to mask the compiler warning about
482 virtual ~BlockFrequencyInfoImplBase() {}
485 namespace bfi_detail {
486 template <class BlockT> struct TypeMap {};
487 template <> struct TypeMap<BasicBlock> {
488 typedef BasicBlock BlockT;
489 typedef Function FunctionT;
490 typedef BranchProbabilityInfo BranchProbabilityInfoT;
492 typedef LoopInfo LoopInfoT;
494 template <> struct TypeMap<MachineBasicBlock> {
495 typedef MachineBasicBlock BlockT;
496 typedef MachineFunction FunctionT;
497 typedef MachineBranchProbabilityInfo BranchProbabilityInfoT;
498 typedef MachineLoop LoopT;
499 typedef MachineLoopInfo LoopInfoT;
502 /// \brief Get the name of a MachineBasicBlock.
504 /// Get the name of a MachineBasicBlock. It's templated so that including from
505 /// CodeGen is unnecessary (that would be a layering issue).
507 /// This is used mainly for debug output. The name is similar to
508 /// MachineBasicBlock::getFullName(), but skips the name of the function.
509 template <class BlockT> std::string getBlockName(const BlockT *BB) {
510 assert(BB && "Unexpected nullptr");
511 auto MachineName = "BB" + Twine(BB->getNumber());
512 if (BB->getBasicBlock())
513 return (MachineName + "[" + BB->getName() + "]").str();
514 return MachineName.str();
516 /// \brief Get the name of a BasicBlock.
517 template <> inline std::string getBlockName(const BasicBlock *BB) {
518 assert(BB && "Unexpected nullptr");
519 return BB->getName().str();
522 /// \brief Graph of irreducible control flow.
524 /// This graph is used for determining the SCCs in a loop (or top-level
525 /// function) that has irreducible control flow.
527 /// During the block frequency algorithm, the local graphs are defined in a
528 /// light-weight way, deferring to the \a BasicBlock or \a MachineBasicBlock
529 /// graphs for most edges, but getting others from \a LoopData::ExitMap. The
530 /// latter only has successor information.
532 /// \a IrreducibleGraph makes this graph explicit. It's in a form that can use
533 /// \a GraphTraits (so that \a analyzeIrreducible() can use \a scc_iterator),
534 /// and it explicitly lists predecessors and successors. The initialization
535 /// that relies on \c MachineBasicBlock is defined in the header.
536 struct IrreducibleGraph {
537 typedef BlockFrequencyInfoImplBase BFIBase;
541 typedef BFIBase::BlockNode BlockNode;
545 std::deque<const IrrNode *> Edges;
546 IrrNode(const BlockNode &Node) : Node(Node), NumIn(0) {}
548 typedef std::deque<const IrrNode *>::const_iterator iterator;
549 iterator pred_begin() const { return Edges.begin(); }
550 iterator succ_begin() const { return Edges.begin() + NumIn; }
551 iterator pred_end() const { return succ_begin(); }
552 iterator succ_end() const { return Edges.end(); }
555 const IrrNode *StartIrr;
556 std::vector<IrrNode> Nodes;
557 SmallDenseMap<uint32_t, IrrNode *, 4> Lookup;
559 /// \brief Construct an explicit graph containing irreducible control flow.
561 /// Construct an explicit graph of the control flow in \c OuterLoop (or the
562 /// top-level function, if \c OuterLoop is \c nullptr). Uses \c
563 /// addBlockEdges to add block successors that have not been packaged into
566 /// \a BlockFrequencyInfoImpl::computeIrreducibleMass() is the only expected
568 template <class BlockEdgesAdder>
569 IrreducibleGraph(BFIBase &BFI, const BFIBase::LoopData *OuterLoop,
570 BlockEdgesAdder addBlockEdges)
571 : BFI(BFI), StartIrr(nullptr) {
572 initialize(OuterLoop, addBlockEdges);
575 template <class BlockEdgesAdder>
576 void initialize(const BFIBase::LoopData *OuterLoop,
577 BlockEdgesAdder addBlockEdges);
578 void addNodesInLoop(const BFIBase::LoopData &OuterLoop);
579 void addNodesInFunction();
580 void addNode(const BlockNode &Node) {
581 Nodes.emplace_back(Node);
582 BFI.Working[Node.Index].getMass() = BlockMass::getEmpty();
585 template <class BlockEdgesAdder>
586 void addEdges(const BlockNode &Node, const BFIBase::LoopData *OuterLoop,
587 BlockEdgesAdder addBlockEdges);
588 void addEdge(IrrNode &Irr, const BlockNode &Succ,
589 const BFIBase::LoopData *OuterLoop);
591 template <class BlockEdgesAdder>
592 void IrreducibleGraph::initialize(const BFIBase::LoopData *OuterLoop,
593 BlockEdgesAdder addBlockEdges) {
595 addNodesInLoop(*OuterLoop);
596 for (auto N : OuterLoop->Nodes)
597 addEdges(N, OuterLoop, addBlockEdges);
599 addNodesInFunction();
600 for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index)
601 addEdges(Index, OuterLoop, addBlockEdges);
603 StartIrr = Lookup[Start.Index];
605 template <class BlockEdgesAdder>
606 void IrreducibleGraph::addEdges(const BlockNode &Node,
607 const BFIBase::LoopData *OuterLoop,
608 BlockEdgesAdder addBlockEdges) {
609 auto L = Lookup.find(Node.Index);
610 if (L == Lookup.end())
612 IrrNode &Irr = *L->second;
613 const auto &Working = BFI.Working[Node.Index];
615 if (Working.isAPackage())
616 for (const auto &I : Working.Loop->Exits)
617 addEdge(Irr, I.first, OuterLoop);
619 addBlockEdges(*this, Irr, OuterLoop);
623 /// \brief Shared implementation for block frequency analysis.
625 /// This is a shared implementation of BlockFrequencyInfo and
626 /// MachineBlockFrequencyInfo, and calculates the relative frequencies of
629 /// LoopInfo defines a loop as a "non-trivial" SCC dominated by a single block,
630 /// which is called the header. A given loop, L, can have sub-loops, which are
631 /// loops within the subgraph of L that exclude its header. (A "trivial" SCC
632 /// consists of a single block that does not have a self-edge.)
634 /// In addition to loops, this algorithm has limited support for irreducible
635 /// SCCs, which are SCCs with multiple entry blocks. Irreducible SCCs are
636 /// discovered on they fly, and modelled as loops with multiple headers.
638 /// The headers of irreducible sub-SCCs consist of its entry blocks and all
639 /// nodes that are targets of a backedge within it (excluding backedges within
640 /// true sub-loops). Block frequency calculations act as if a block is
641 /// inserted that intercepts all the edges to the headers. All backedges and
642 /// entries point to this block. Its successors are the headers, which split
643 /// the frequency evenly.
645 /// This algorithm leverages BlockMass and ScaledNumber to maintain precision,
646 /// separates mass distribution from loop scaling, and dithers to eliminate
647 /// probability mass loss.
649 /// The implementation is split between BlockFrequencyInfoImpl, which knows the
650 /// type of graph being modelled (BasicBlock vs. MachineBasicBlock), and
651 /// BlockFrequencyInfoImplBase, which doesn't. The base class uses \a
652 /// BlockNode, a wrapper around a uint32_t. BlockNode is numbered from 0 in
653 /// reverse-post order. This gives two advantages: it's easy to compare the
654 /// relative ordering of two nodes, and maps keyed on BlockT can be represented
657 /// This algorithm is O(V+E), unless there is irreducible control flow, in
658 /// which case it's O(V*E) in the worst case.
660 /// These are the main stages:
662 /// 0. Reverse post-order traversal (\a initializeRPOT()).
664 /// Run a single post-order traversal and save it (in reverse) in RPOT.
665 /// All other stages make use of this ordering. Save a lookup from BlockT
666 /// to BlockNode (the index into RPOT) in Nodes.
668 /// 1. Loop initialization (\a initializeLoops()).
670 /// Translate LoopInfo/MachineLoopInfo into a form suitable for the rest of
671 /// the algorithm. In particular, store the immediate members of each loop
672 /// in reverse post-order.
674 /// 2. Calculate mass and scale in loops (\a computeMassInLoops()).
676 /// For each loop (bottom-up), distribute mass through the DAG resulting
677 /// from ignoring backedges and treating sub-loops as a single pseudo-node.
678 /// Track the backedge mass distributed to the loop header, and use it to
679 /// calculate the loop scale (number of loop iterations). Immediate
680 /// members that represent sub-loops will already have been visited and
681 /// packaged into a pseudo-node.
683 /// Distributing mass in a loop is a reverse-post-order traversal through
684 /// the loop. Start by assigning full mass to the Loop header. For each
685 /// node in the loop:
687 /// - Fetch and categorize the weight distribution for its successors.
688 /// If this is a packaged-subloop, the weight distribution is stored
689 /// in \a LoopData::Exits. Otherwise, fetch it from
690 /// BranchProbabilityInfo.
692 /// - Each successor is categorized as \a Weight::Local, a local edge
693 /// within the current loop, \a Weight::Backedge, a backedge to the
694 /// loop header, or \a Weight::Exit, any successor outside the loop.
695 /// The weight, the successor, and its category are stored in \a
696 /// Distribution. There can be multiple edges to each successor.
698 /// - If there's a backedge to a non-header, there's an irreducible SCC.
699 /// The usual flow is temporarily aborted. \a
700 /// computeIrreducibleMass() finds the irreducible SCCs within the
701 /// loop, packages them up, and restarts the flow.
703 /// - Normalize the distribution: scale weights down so that their sum
704 /// is 32-bits, and coalesce multiple edges to the same node.
706 /// - Distribute the mass accordingly, dithering to minimize mass loss,
707 /// as described in \a distributeMass().
709 /// Finally, calculate the loop scale from the accumulated backedge mass.
711 /// 3. Distribute mass in the function (\a computeMassInFunction()).
713 /// Finally, distribute mass through the DAG resulting from packaging all
714 /// loops in the function. This uses the same algorithm as distributing
715 /// mass in a loop, except that there are no exit or backedge edges.
717 /// 4. Unpackage loops (\a unwrapLoops()).
719 /// Initialize each block's frequency to a floating point representation of
722 /// Visit loops top-down, scaling the frequencies of its immediate members
723 /// by the loop's pseudo-node's frequency.
725 /// 5. Convert frequencies to a 64-bit range (\a finalizeMetrics()).
727 /// Using the min and max frequencies as a guide, translate floating point
728 /// frequencies to an appropriate range in uint64_t.
730 /// It has some known flaws.
732 /// - Loop scale is limited to 4096 per loop (2^12) to avoid exhausting
733 /// BlockFrequency's 64-bit integer precision.
735 /// - The model of irreducible control flow is a rough approximation.
737 /// Modelling irreducible control flow exactly involves setting up and
738 /// solving a group of infinite geometric series. Such precision is
739 /// unlikely to be worthwhile, since most of our algorithms give up on
740 /// irreducible control flow anyway.
742 /// Nevertheless, we might find that we need to get closer. Here's a sort
743 /// of TODO list for the model with diminishing returns, to be completed as
746 /// - The headers for the \a LoopData representing an irreducible SCC
747 /// include non-entry blocks. When these extra blocks exist, they
748 /// indicate a self-contained irreducible sub-SCC. We could treat them
749 /// as sub-loops, rather than arbitrarily shoving the problematic
750 /// blocks into the headers of the main irreducible SCC.
752 /// - Backedge frequencies are assumed to be evenly split between the
753 /// headers of a given irreducible SCC. Instead, we could track the
754 /// backedge mass separately for each header, and adjust their relative
757 /// - Entry frequencies are assumed to be evenly split between the
758 /// headers of a given irreducible SCC, which is the only option if we
759 /// need to compute mass in the SCC before its parent loop. Instead,
760 /// we could partially compute mass in the parent loop, and stop when
761 /// we get to the SCC. Here, we have the correct ratio of entry
762 /// masses, which we can use to adjust their relative frequencies.
763 /// Compute mass in the SCC, and then continue propagation in the
766 /// - We can propagate mass iteratively through the SCC, for some fixed
767 /// number of iterations. Each iteration starts by assigning the entry
768 /// blocks their backedge mass from the prior iteration. The final
769 /// mass for each block (and each exit, and the total backedge mass
770 /// used for computing loop scale) is the sum of all iterations.
771 /// (Running this until fixed point would "solve" the geometric
772 /// series by simulation.)
773 template <class BT> class BlockFrequencyInfoImpl : BlockFrequencyInfoImplBase {
774 typedef typename bfi_detail::TypeMap<BT>::BlockT BlockT;
775 typedef typename bfi_detail::TypeMap<BT>::FunctionT FunctionT;
776 typedef typename bfi_detail::TypeMap<BT>::BranchProbabilityInfoT
777 BranchProbabilityInfoT;
778 typedef typename bfi_detail::TypeMap<BT>::LoopT LoopT;
779 typedef typename bfi_detail::TypeMap<BT>::LoopInfoT LoopInfoT;
781 // This is part of a workaround for a GCC 4.7 crash on lambdas.
782 friend struct bfi_detail::BlockEdgesAdder<BT>;
784 typedef GraphTraits<const BlockT *> Successor;
785 typedef GraphTraits<Inverse<const BlockT *>> Predecessor;
787 const BranchProbabilityInfoT *BPI;
791 // All blocks in reverse postorder.
792 std::vector<const BlockT *> RPOT;
793 DenseMap<const BlockT *, BlockNode> Nodes;
795 typedef typename std::vector<const BlockT *>::const_iterator rpot_iterator;
797 rpot_iterator rpot_begin() const { return RPOT.begin(); }
798 rpot_iterator rpot_end() const { return RPOT.end(); }
800 size_t getIndex(const rpot_iterator &I) const { return I - rpot_begin(); }
802 BlockNode getNode(const rpot_iterator &I) const {
803 return BlockNode(getIndex(I));
805 BlockNode getNode(const BlockT *BB) const { return Nodes.lookup(BB); }
807 const BlockT *getBlock(const BlockNode &Node) const {
808 assert(Node.Index < RPOT.size());
809 return RPOT[Node.Index];
812 /// \brief Run (and save) a post-order traversal.
814 /// Saves a reverse post-order traversal of all the nodes in \a F.
815 void initializeRPOT();
817 /// \brief Initialize loop data.
819 /// Build up \a Loops using \a LoopInfo. \a LoopInfo gives us a mapping from
820 /// each block to the deepest loop it's in, but we need the inverse. For each
821 /// loop, we store in reverse post-order its "immediate" members, defined as
822 /// the header, the headers of immediate sub-loops, and all other blocks in
823 /// the loop that are not in sub-loops.
824 void initializeLoops();
826 /// \brief Propagate to a block's successors.
828 /// In the context of distributing mass through \c OuterLoop, divide the mass
829 /// currently assigned to \c Node between its successors.
831 /// \return \c true unless there's an irreducible backedge.
832 bool propagateMassToSuccessors(LoopData *OuterLoop, const BlockNode &Node);
834 /// \brief Compute mass in a particular loop.
836 /// Assign mass to \c Loop's header, and then for each block in \c Loop in
837 /// reverse post-order, distribute mass to its successors. Only visits nodes
838 /// that have not been packaged into sub-loops.
840 /// \pre \a computeMassInLoop() has been called for each subloop of \c Loop.
841 /// \return \c true unless there's an irreducible backedge.
842 bool computeMassInLoop(LoopData &Loop);
844 /// \brief Try to compute mass in the top-level function.
846 /// Assign mass to the entry block, and then for each block in reverse
847 /// post-order, distribute mass to its successors. Skips nodes that have
848 /// been packaged into loops.
850 /// \pre \a computeMassInLoops() has been called.
851 /// \return \c true unless there's an irreducible backedge.
852 bool tryToComputeMassInFunction();
854 /// \brief Compute mass in (and package up) irreducible SCCs.
856 /// Find the irreducible SCCs in \c OuterLoop, add them to \a Loops (in front
857 /// of \c Insert), and call \a computeMassInLoop() on each of them.
859 /// If \c OuterLoop is \c nullptr, it refers to the top-level function.
861 /// \pre \a computeMassInLoop() has been called for each subloop of \c
863 /// \pre \c Insert points at the the last loop successfully processed by \a
864 /// computeMassInLoop().
865 /// \pre \c OuterLoop has irreducible SCCs.
866 void computeIrreducibleMass(LoopData *OuterLoop,
867 std::list<LoopData>::iterator Insert);
869 /// \brief Compute mass in all loops.
871 /// For each loop bottom-up, call \a computeMassInLoop().
873 /// \a computeMassInLoop() aborts (and returns \c false) on loops that
874 /// contain a irreducible sub-SCCs. Use \a computeIrreducibleMass() and then
875 /// re-enter \a computeMassInLoop().
877 /// \post \a computeMassInLoop() has returned \c true for every loop.
878 void computeMassInLoops();
880 /// \brief Compute mass in the top-level function.
882 /// Uses \a tryToComputeMassInFunction() and \a computeIrreducibleMass() to
883 /// compute mass in the top-level function.
885 /// \post \a tryToComputeMassInFunction() has returned \c true.
886 void computeMassInFunction();
888 std::string getBlockName(const BlockNode &Node) const override {
889 return bfi_detail::getBlockName(getBlock(Node));
893 const FunctionT *getFunction() const { return F; }
895 void doFunction(const FunctionT *F, const BranchProbabilityInfoT *BPI,
896 const LoopInfoT *LI);
897 BlockFrequencyInfoImpl() : BPI(nullptr), LI(nullptr), F(nullptr) {}
899 using BlockFrequencyInfoImplBase::getEntryFreq;
900 BlockFrequency getBlockFreq(const BlockT *BB) const {
901 return BlockFrequencyInfoImplBase::getBlockFreq(getNode(BB));
903 Scaled64 getFloatingBlockFreq(const BlockT *BB) const {
904 return BlockFrequencyInfoImplBase::getFloatingBlockFreq(getNode(BB));
907 /// \brief Print the frequencies for the current function.
909 /// Prints the frequencies for the blocks in the current function.
911 /// Blocks are printed in the natural iteration order of the function, rather
912 /// than reverse post-order. This provides two advantages: writing -analyze
913 /// tests is easier (since blocks come out in source order), and even
914 /// unreachable blocks are printed.
916 /// \a BlockFrequencyInfoImplBase::print() only knows reverse post-order, so
917 /// we need to override it here.
918 raw_ostream &print(raw_ostream &OS) const override;
919 using BlockFrequencyInfoImplBase::dump;
921 using BlockFrequencyInfoImplBase::printBlockFreq;
922 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockT *BB) const {
923 return BlockFrequencyInfoImplBase::printBlockFreq(OS, getNode(BB));
928 void BlockFrequencyInfoImpl<BT>::doFunction(const FunctionT *F,
929 const BranchProbabilityInfoT *BPI,
930 const LoopInfoT *LI) {
931 // Save the parameters.
936 // Clean up left-over data structures.
937 BlockFrequencyInfoImplBase::clear();
942 DEBUG(dbgs() << "\nblock-frequency: " << F->getName() << "\n================="
943 << std::string(F->getName().size(), '=') << "\n");
947 // Visit loops in post-order to find thelocal mass distribution, and then do
948 // the full function.
949 computeMassInLoops();
950 computeMassInFunction();
955 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeRPOT() {
956 const BlockT *Entry = F->begin();
957 RPOT.reserve(F->size());
958 std::copy(po_begin(Entry), po_end(Entry), std::back_inserter(RPOT));
959 std::reverse(RPOT.begin(), RPOT.end());
961 assert(RPOT.size() - 1 <= BlockNode::getMaxIndex() &&
962 "More nodes in function than Block Frequency Info supports");
964 DEBUG(dbgs() << "reverse-post-order-traversal\n");
965 for (rpot_iterator I = rpot_begin(), E = rpot_end(); I != E; ++I) {
966 BlockNode Node = getNode(I);
967 DEBUG(dbgs() << " - " << getIndex(I) << ": " << getBlockName(Node) << "\n");
971 Working.reserve(RPOT.size());
972 for (size_t Index = 0; Index < RPOT.size(); ++Index)
973 Working.emplace_back(Index);
974 Freqs.resize(RPOT.size());
977 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeLoops() {
978 DEBUG(dbgs() << "loop-detection\n");
982 // Visit loops top down and assign them an index.
983 std::deque<std::pair<const LoopT *, LoopData *>> Q;
984 for (const LoopT *L : *LI)
985 Q.emplace_back(L, nullptr);
987 const LoopT *Loop = Q.front().first;
988 LoopData *Parent = Q.front().second;
991 BlockNode Header = getNode(Loop->getHeader());
992 assert(Header.isValid());
994 Loops.emplace_back(Parent, Header);
995 Working[Header.Index].Loop = &Loops.back();
996 DEBUG(dbgs() << " - loop = " << getBlockName(Header) << "\n");
998 for (const LoopT *L : *Loop)
999 Q.emplace_back(L, &Loops.back());
1002 // Visit nodes in reverse post-order and add them to their deepest containing
1004 for (size_t Index = 0; Index < RPOT.size(); ++Index) {
1005 // Loop headers have already been mostly mapped.
1006 if (Working[Index].isLoopHeader()) {
1007 LoopData *ContainingLoop = Working[Index].getContainingLoop();
1009 ContainingLoop->Nodes.push_back(Index);
1013 const LoopT *Loop = LI->getLoopFor(RPOT[Index]);
1017 // Add this node to its containing loop's member list.
1018 BlockNode Header = getNode(Loop->getHeader());
1019 assert(Header.isValid());
1020 const auto &HeaderData = Working[Header.Index];
1021 assert(HeaderData.isLoopHeader());
1023 Working[Index].Loop = HeaderData.Loop;
1024 HeaderData.Loop->Nodes.push_back(Index);
1025 DEBUG(dbgs() << " - loop = " << getBlockName(Header)
1026 << ": member = " << getBlockName(Index) << "\n");
1030 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInLoops() {
1031 // Visit loops with the deepest first, and the top-level loops last.
1032 for (auto L = Loops.rbegin(), E = Loops.rend(); L != E; ++L) {
1033 if (computeMassInLoop(*L))
1035 auto Next = std::next(L);
1036 computeIrreducibleMass(&*L, L.base());
1037 L = std::prev(Next);
1038 if (computeMassInLoop(*L))
1040 llvm_unreachable("unhandled irreducible control flow");
1045 bool BlockFrequencyInfoImpl<BT>::computeMassInLoop(LoopData &Loop) {
1046 // Compute mass in loop.
1047 DEBUG(dbgs() << "compute-mass-in-loop: " << getLoopName(Loop) << "\n");
1049 if (Loop.isIrreducible()) {
1050 BlockMass Remaining = BlockMass::getFull();
1051 for (uint32_t H = 0; H < Loop.NumHeaders; ++H) {
1052 auto &Mass = Working[Loop.Nodes[H].Index].getMass();
1053 Mass = Remaining * BranchProbability(1, Loop.NumHeaders - H);
1056 for (const BlockNode &M : Loop.Nodes)
1057 if (!propagateMassToSuccessors(&Loop, M))
1058 llvm_unreachable("unhandled irreducible control flow");
1060 Working[Loop.getHeader().Index].getMass() = BlockMass::getFull();
1061 if (!propagateMassToSuccessors(&Loop, Loop.getHeader()))
1062 llvm_unreachable("irreducible control flow to loop header!?");
1063 for (const BlockNode &M : Loop.members())
1064 if (!propagateMassToSuccessors(&Loop, M))
1065 // Irreducible backedge.
1069 computeLoopScale(Loop);
1075 bool BlockFrequencyInfoImpl<BT>::tryToComputeMassInFunction() {
1076 // Compute mass in function.
1077 DEBUG(dbgs() << "compute-mass-in-function\n");
1078 assert(!Working.empty() && "no blocks in function");
1079 assert(!Working[0].isLoopHeader() && "entry block is a loop header");
1081 Working[0].getMass() = BlockMass::getFull();
1082 for (rpot_iterator I = rpot_begin(), IE = rpot_end(); I != IE; ++I) {
1083 // Check for nodes that have been packaged.
1084 BlockNode Node = getNode(I);
1085 if (Working[Node.Index].isPackaged())
1088 if (!propagateMassToSuccessors(nullptr, Node))
1094 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInFunction() {
1095 if (tryToComputeMassInFunction())
1097 computeIrreducibleMass(nullptr, Loops.begin());
1098 if (tryToComputeMassInFunction())
1100 llvm_unreachable("unhandled irreducible control flow");
1103 /// \note This should be a lambda, but that crashes GCC 4.7.
1104 namespace bfi_detail {
1105 template <class BT> struct BlockEdgesAdder {
1107 typedef BlockFrequencyInfoImplBase::LoopData LoopData;
1108 typedef GraphTraits<const BlockT *> Successor;
1110 const BlockFrequencyInfoImpl<BT> &BFI;
1111 explicit BlockEdgesAdder(const BlockFrequencyInfoImpl<BT> &BFI)
1113 void operator()(IrreducibleGraph &G, IrreducibleGraph::IrrNode &Irr,
1114 const LoopData *OuterLoop) {
1115 const BlockT *BB = BFI.RPOT[Irr.Node.Index];
1116 for (auto I = Successor::child_begin(BB), E = Successor::child_end(BB);
1118 G.addEdge(Irr, BFI.getNode(*I), OuterLoop);
1123 void BlockFrequencyInfoImpl<BT>::computeIrreducibleMass(
1124 LoopData *OuterLoop, std::list<LoopData>::iterator Insert) {
1125 DEBUG(dbgs() << "analyze-irreducible-in-";
1126 if (OuterLoop) dbgs() << "loop: " << getLoopName(*OuterLoop) << "\n";
1127 else dbgs() << "function\n");
1129 using namespace bfi_detail;
1130 // Ideally, addBlockEdges() would be declared here as a lambda, but that
1132 BlockEdgesAdder<BT> addBlockEdges(*this);
1133 IrreducibleGraph G(*this, OuterLoop, addBlockEdges);
1135 for (auto &L : analyzeIrreducible(G, OuterLoop, Insert))
1136 computeMassInLoop(L);
1140 updateLoopWithIrreducible(*OuterLoop);
1145 BlockFrequencyInfoImpl<BT>::propagateMassToSuccessors(LoopData *OuterLoop,
1146 const BlockNode &Node) {
1147 DEBUG(dbgs() << " - node: " << getBlockName(Node) << "\n");
1148 // Calculate probability for successors.
1150 if (auto *Loop = Working[Node.Index].getPackagedLoop()) {
1151 assert(Loop != OuterLoop && "Cannot propagate mass in a packaged loop");
1152 if (!addLoopSuccessorsToDist(OuterLoop, *Loop, Dist))
1153 // Irreducible backedge.
1156 const BlockT *BB = getBlock(Node);
1157 for (auto SI = Successor::child_begin(BB), SE = Successor::child_end(BB);
1159 // Do not dereference SI, or getEdgeWeight() is linear in the number of
1161 if (!addToDist(Dist, OuterLoop, Node, getNode(*SI),
1162 BPI->getEdgeWeight(BB, SI)))
1163 // Irreducible backedge.
1167 // Distribute mass to successors, saving exit and backedge data in the
1169 distributeMass(Node, OuterLoop, Dist);
1174 raw_ostream &BlockFrequencyInfoImpl<BT>::print(raw_ostream &OS) const {
1177 OS << "block-frequency-info: " << F->getName() << "\n";
1178 for (const BlockT &BB : *F)
1179 OS << " - " << bfi_detail::getBlockName(&BB)
1180 << ": float = " << getFloatingBlockFreq(&BB)
1181 << ", int = " << getBlockFreq(&BB).getFrequency() << "\n";
1183 // Add an extra newline for readability.
1188 } // end namespace llvm