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 represent a loop as a
195 /// pseudo-node once it's packaged.
197 typedef SmallVector<std::pair<BlockNode, BlockMass>, 4> ExitMap;
198 typedef SmallVector<BlockNode, 4> NodeList;
199 LoopData *Parent; ///< The parent loop.
200 bool IsPackaged; ///< Whether this has been packaged.
201 uint32_t NumHeaders; ///< Number of headers.
202 ExitMap Exits; ///< Successor edges (and weights).
203 NodeList Nodes; ///< Header and the members of the loop.
204 BlockMass BackedgeMass; ///< Mass returned to loop header.
208 LoopData(LoopData *Parent, const BlockNode &Header)
209 : Parent(Parent), IsPackaged(false), NumHeaders(1), Nodes(1, Header) {}
210 template <class It1, class It2>
211 LoopData(LoopData *Parent, It1 FirstHeader, It1 LastHeader, It2 FirstOther,
213 : Parent(Parent), IsPackaged(false), Nodes(FirstHeader, LastHeader) {
214 NumHeaders = Nodes.size();
215 Nodes.insert(Nodes.end(), FirstOther, LastOther);
217 bool isHeader(const BlockNode &Node) const {
219 return std::binary_search(Nodes.begin(), Nodes.begin() + NumHeaders,
221 return Node == Nodes[0];
223 BlockNode getHeader() const { return Nodes[0]; }
224 bool isIrreducible() const { return NumHeaders > 1; }
226 NodeList::const_iterator members_begin() const {
227 return Nodes.begin() + NumHeaders;
229 NodeList::const_iterator members_end() const { return Nodes.end(); }
230 iterator_range<NodeList::const_iterator> members() const {
231 return make_range(members_begin(), members_end());
235 /// \brief Index of loop information.
237 BlockNode Node; ///< This node.
238 LoopData *Loop; ///< The loop this block is inside.
239 BlockMass Mass; ///< Mass distribution from the entry block.
241 WorkingData(const BlockNode &Node) : Node(Node), Loop(nullptr) {}
243 bool isLoopHeader() const { return Loop && Loop->isHeader(Node); }
244 bool isDoubleLoopHeader() const {
245 return isLoopHeader() && Loop->Parent && Loop->Parent->isIrreducible() &&
246 Loop->Parent->isHeader(Node);
249 LoopData *getContainingLoop() const {
252 if (!isDoubleLoopHeader())
254 return Loop->Parent->Parent;
257 /// \brief Resolve a node to its representative.
259 /// Get the node currently representing Node, which could be a containing
262 /// This function should only be called when distributing mass. As long as
263 /// there are no irreducible edges to Node, then it will have complexity
264 /// O(1) in this context.
266 /// In general, the complexity is O(L), where L is the number of loop
267 /// headers Node has been packaged into. Since this method is called in
268 /// the context of distributing mass, L will be the number of loop headers
269 /// an early exit edge jumps out of.
270 BlockNode getResolvedNode() const {
271 auto L = getPackagedLoop();
272 return L ? L->getHeader() : Node;
274 LoopData *getPackagedLoop() const {
275 if (!Loop || !Loop->IsPackaged)
278 while (L->Parent && L->Parent->IsPackaged)
283 /// \brief Get the appropriate mass for a node.
285 /// Get appropriate mass for Node. If Node is a loop-header (whose loop
286 /// has been packaged), returns the mass of its pseudo-node. If it's a
287 /// node inside a packaged loop, it returns the loop's mass.
288 BlockMass &getMass() {
291 if (!isADoublePackage())
293 return Loop->Parent->Mass;
296 /// \brief Has ContainingLoop been packaged up?
297 bool isPackaged() const { return getResolvedNode() != Node; }
298 /// \brief Has Loop been packaged up?
299 bool isAPackage() const { return isLoopHeader() && Loop->IsPackaged; }
300 /// \brief Has Loop been packaged up twice?
301 bool isADoublePackage() const {
302 return isDoubleLoopHeader() && Loop->Parent->IsPackaged;
306 /// \brief Unscaled probability weight.
308 /// Probability weight for an edge in the graph (including the
309 /// successor/target node).
311 /// All edges in the original function are 32-bit. However, exit edges from
312 /// loop packages are taken from 64-bit exit masses, so we need 64-bits of
313 /// space in general.
315 /// In addition to the raw weight amount, Weight stores the type of the edge
316 /// in the current context (i.e., the context of the loop being processed).
317 /// Is this a local edge within the loop, an exit from the loop, or a
318 /// backedge to the loop header?
320 enum DistType { Local, Exit, Backedge };
322 BlockNode TargetNode;
324 Weight() : Type(Local), Amount(0) {}
325 Weight(DistType Type, BlockNode TargetNode, uint64_t Amount)
326 : Type(Type), TargetNode(TargetNode), Amount(Amount) {}
329 /// \brief Distribution of unscaled probability weight.
331 /// Distribution of unscaled probability weight to a set of successors.
333 /// This class collates the successor edge weights for later processing.
335 /// \a DidOverflow indicates whether \a Total did overflow while adding to
336 /// the distribution. It should never overflow twice.
337 struct Distribution {
338 typedef SmallVector<Weight, 4> WeightList;
339 WeightList Weights; ///< Individual successor weights.
340 uint64_t Total; ///< Sum of all weights.
341 bool DidOverflow; ///< Whether \a Total did overflow.
343 Distribution() : Total(0), DidOverflow(false) {}
344 void addLocal(const BlockNode &Node, uint64_t Amount) {
345 add(Node, Amount, Weight::Local);
347 void addExit(const BlockNode &Node, uint64_t Amount) {
348 add(Node, Amount, Weight::Exit);
350 void addBackedge(const BlockNode &Node, uint64_t Amount) {
351 add(Node, Amount, Weight::Backedge);
354 /// \brief Normalize the distribution.
356 /// Combines multiple edges to the same \a Weight::TargetNode and scales
357 /// down so that \a Total fits into 32-bits.
359 /// This is linear in the size of \a Weights. For the vast majority of
360 /// cases, adjacent edge weights are combined by sorting WeightList and
361 /// combining adjacent weights. However, for very large edge lists an
362 /// auxiliary hash table is used.
366 void add(const BlockNode &Node, uint64_t Amount, Weight::DistType Type);
369 /// \brief Data about each block. This is used downstream.
370 std::vector<FrequencyData> Freqs;
372 /// \brief Loop data: see initializeLoops().
373 std::vector<WorkingData> Working;
375 /// \brief Indexed information about loops.
376 std::list<LoopData> Loops;
378 /// \brief Add all edges out of a packaged loop to the distribution.
380 /// Adds all edges from LocalLoopHead to Dist. Calls addToDist() to add each
383 /// \return \c true unless there's an irreducible backedge.
384 bool addLoopSuccessorsToDist(const LoopData *OuterLoop, LoopData &Loop,
387 /// \brief Add an edge to the distribution.
389 /// Adds an edge to Succ to Dist. If \c LoopHead.isValid(), then whether the
390 /// edge is local/exit/backedge is in the context of LoopHead. Otherwise,
391 /// every edge should be a local edge (since all the loops are packaged up).
393 /// \return \c true unless aborted due to an irreducible backedge.
394 bool addToDist(Distribution &Dist, const LoopData *OuterLoop,
395 const BlockNode &Pred, const BlockNode &Succ, uint64_t Weight);
397 LoopData &getLoopPackage(const BlockNode &Head) {
398 assert(Head.Index < Working.size());
399 assert(Working[Head.Index].isLoopHeader());
400 return *Working[Head.Index].Loop;
403 /// \brief Analyze irreducible SCCs.
405 /// Separate irreducible SCCs from \c G, which is an explict graph of \c
406 /// OuterLoop (or the top-level function, if \c OuterLoop is \c nullptr).
407 /// Insert them into \a Loops before \c Insert.
409 /// \return the \c LoopData nodes representing the irreducible SCCs.
410 iterator_range<std::list<LoopData>::iterator>
411 analyzeIrreducible(const bfi_detail::IrreducibleGraph &G, LoopData *OuterLoop,
412 std::list<LoopData>::iterator Insert);
414 /// \brief Update a loop after packaging irreducible SCCs inside of it.
416 /// Update \c OuterLoop. Before finding irreducible control flow, it was
417 /// partway through \a computeMassInLoop(), so \a LoopData::Exits and \a
418 /// LoopData::BackedgeMass need to be reset. Also, nodes that were packaged
419 /// up need to be removed from \a OuterLoop::Nodes.
420 void updateLoopWithIrreducible(LoopData &OuterLoop);
422 /// \brief Distribute mass according to a distribution.
424 /// Distributes the mass in Source according to Dist. If LoopHead.isValid(),
425 /// backedges and exits are stored in its entry in Loops.
427 /// Mass is distributed in parallel from two copies of the source mass.
428 void distributeMass(const BlockNode &Source, LoopData *OuterLoop,
431 /// \brief Compute the loop scale for a loop.
432 void computeLoopScale(LoopData &Loop);
434 /// \brief Package up a loop.
435 void packageLoop(LoopData &Loop);
437 /// \brief Unwrap loops.
440 /// \brief Finalize frequency metrics.
442 /// Calculates final frequencies and cleans up no-longer-needed data
444 void finalizeMetrics();
446 /// \brief Clear all memory.
449 virtual std::string getBlockName(const BlockNode &Node) const;
450 std::string getLoopName(const LoopData &Loop) const;
452 virtual raw_ostream &print(raw_ostream &OS) const { return OS; }
453 void dump() const { print(dbgs()); }
455 Scaled64 getFloatingBlockFreq(const BlockNode &Node) const;
457 BlockFrequency getBlockFreq(const BlockNode &Node) const;
459 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockNode &Node) const;
460 raw_ostream &printBlockFreq(raw_ostream &OS,
461 const BlockFrequency &Freq) const;
463 uint64_t getEntryFreq() const {
464 assert(!Freqs.empty());
465 return Freqs[0].Integer;
467 /// \brief Virtual destructor.
469 /// Need a virtual destructor to mask the compiler warning about
471 virtual ~BlockFrequencyInfoImplBase() {}
474 namespace bfi_detail {
475 template <class BlockT> struct TypeMap {};
476 template <> struct TypeMap<BasicBlock> {
477 typedef BasicBlock BlockT;
478 typedef Function FunctionT;
479 typedef BranchProbabilityInfo BranchProbabilityInfoT;
481 typedef LoopInfo LoopInfoT;
483 template <> struct TypeMap<MachineBasicBlock> {
484 typedef MachineBasicBlock BlockT;
485 typedef MachineFunction FunctionT;
486 typedef MachineBranchProbabilityInfo BranchProbabilityInfoT;
487 typedef MachineLoop LoopT;
488 typedef MachineLoopInfo LoopInfoT;
491 /// \brief Get the name of a MachineBasicBlock.
493 /// Get the name of a MachineBasicBlock. It's templated so that including from
494 /// CodeGen is unnecessary (that would be a layering issue).
496 /// This is used mainly for debug output. The name is similar to
497 /// MachineBasicBlock::getFullName(), but skips the name of the function.
498 template <class BlockT> std::string getBlockName(const BlockT *BB) {
499 assert(BB && "Unexpected nullptr");
500 auto MachineName = "BB" + Twine(BB->getNumber());
501 if (BB->getBasicBlock())
502 return (MachineName + "[" + BB->getName() + "]").str();
503 return MachineName.str();
505 /// \brief Get the name of a BasicBlock.
506 template <> inline std::string getBlockName(const BasicBlock *BB) {
507 assert(BB && "Unexpected nullptr");
508 return BB->getName().str();
511 /// \brief Graph of irreducible control flow.
513 /// This graph is used for determining the SCCs in a loop (or top-level
514 /// function) that has irreducible control flow.
516 /// During the block frequency algorithm, the local graphs are defined in a
517 /// light-weight way, deferring to the \a BasicBlock or \a MachineBasicBlock
518 /// graphs for most edges, but getting others from \a LoopData::ExitMap. The
519 /// latter only has successor information.
521 /// \a IrreducibleGraph makes this graph explicit. It's in a form that can use
522 /// \a GraphTraits (so that \a analyzeIrreducible() can use \a scc_iterator),
523 /// and it explicitly lists predecessors and successors. The initialization
524 /// that relies on \c MachineBasicBlock is defined in the header.
525 struct IrreducibleGraph {
526 typedef BlockFrequencyInfoImplBase BFIBase;
530 typedef BFIBase::BlockNode BlockNode;
534 std::deque<const IrrNode *> Edges;
535 IrrNode(const BlockNode &Node) : Node(Node), NumIn(0) {}
537 typedef std::deque<const IrrNode *>::const_iterator iterator;
538 iterator pred_begin() const { return Edges.begin(); }
539 iterator succ_begin() const { return Edges.begin() + NumIn; }
540 iterator pred_end() const { return succ_begin(); }
541 iterator succ_end() const { return Edges.end(); }
544 const IrrNode *StartIrr;
545 std::vector<IrrNode> Nodes;
546 SmallDenseMap<uint32_t, IrrNode *, 4> Lookup;
548 /// \brief Construct an explicit graph containing irreducible control flow.
550 /// Construct an explicit graph of the control flow in \c OuterLoop (or the
551 /// top-level function, if \c OuterLoop is \c nullptr). Uses \c
552 /// addBlockEdges to add block successors that have not been packaged into
555 /// \a BlockFrequencyInfoImpl::computeIrreducibleMass() is the only expected
557 template <class BlockEdgesAdder>
558 IrreducibleGraph(BFIBase &BFI, const BFIBase::LoopData *OuterLoop,
559 BlockEdgesAdder addBlockEdges)
560 : BFI(BFI), StartIrr(nullptr) {
561 initialize(OuterLoop, addBlockEdges);
564 template <class BlockEdgesAdder>
565 void initialize(const BFIBase::LoopData *OuterLoop,
566 BlockEdgesAdder addBlockEdges);
567 void addNodesInLoop(const BFIBase::LoopData &OuterLoop);
568 void addNodesInFunction();
569 void addNode(const BlockNode &Node) {
570 Nodes.emplace_back(Node);
571 BFI.Working[Node.Index].getMass() = BlockMass::getEmpty();
574 template <class BlockEdgesAdder>
575 void addEdges(const BlockNode &Node, const BFIBase::LoopData *OuterLoop,
576 BlockEdgesAdder addBlockEdges);
577 void addEdge(IrrNode &Irr, const BlockNode &Succ,
578 const BFIBase::LoopData *OuterLoop);
580 template <class BlockEdgesAdder>
581 void IrreducibleGraph::initialize(const BFIBase::LoopData *OuterLoop,
582 BlockEdgesAdder addBlockEdges) {
584 addNodesInLoop(*OuterLoop);
585 for (auto N : OuterLoop->Nodes)
586 addEdges(N, OuterLoop, addBlockEdges);
588 addNodesInFunction();
589 for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index)
590 addEdges(Index, OuterLoop, addBlockEdges);
592 StartIrr = Lookup[Start.Index];
594 template <class BlockEdgesAdder>
595 void IrreducibleGraph::addEdges(const BlockNode &Node,
596 const BFIBase::LoopData *OuterLoop,
597 BlockEdgesAdder addBlockEdges) {
598 auto L = Lookup.find(Node.Index);
599 if (L == Lookup.end())
601 IrrNode &Irr = *L->second;
602 const auto &Working = BFI.Working[Node.Index];
604 if (Working.isAPackage())
605 for (const auto &I : Working.Loop->Exits)
606 addEdge(Irr, I.first, OuterLoop);
608 addBlockEdges(*this, Irr, OuterLoop);
612 /// \brief Shared implementation for block frequency analysis.
614 /// This is a shared implementation of BlockFrequencyInfo and
615 /// MachineBlockFrequencyInfo, and calculates the relative frequencies of
618 /// LoopInfo defines a loop as a "non-trivial" SCC dominated by a single block,
619 /// which is called the header. A given loop, L, can have sub-loops, which are
620 /// loops within the subgraph of L that exclude its header. (A "trivial" SCC
621 /// consists of a single block that does not have a self-edge.)
623 /// In addition to loops, this algorithm has limited support for irreducible
624 /// SCCs, which are SCCs with multiple entry blocks. Irreducible SCCs are
625 /// discovered on they fly, and modelled as loops with multiple headers.
627 /// The headers of irreducible sub-SCCs consist of its entry blocks and all
628 /// nodes that are targets of a backedge within it (excluding backedges within
629 /// true sub-loops). Block frequency calculations act as if a block is
630 /// inserted that intercepts all the edges to the headers. All backedges and
631 /// entries point to this block. Its successors are the headers, which split
632 /// the frequency evenly.
634 /// This algorithm leverages BlockMass and ScaledNumber to maintain precision,
635 /// separates mass distribution from loop scaling, and dithers to eliminate
636 /// probability mass loss.
638 /// The implementation is split between BlockFrequencyInfoImpl, which knows the
639 /// type of graph being modelled (BasicBlock vs. MachineBasicBlock), and
640 /// BlockFrequencyInfoImplBase, which doesn't. The base class uses \a
641 /// BlockNode, a wrapper around a uint32_t. BlockNode is numbered from 0 in
642 /// reverse-post order. This gives two advantages: it's easy to compare the
643 /// relative ordering of two nodes, and maps keyed on BlockT can be represented
646 /// This algorithm is O(V+E), unless there is irreducible control flow, in
647 /// which case it's O(V*E) in the worst case.
649 /// These are the main stages:
651 /// 0. Reverse post-order traversal (\a initializeRPOT()).
653 /// Run a single post-order traversal and save it (in reverse) in RPOT.
654 /// All other stages make use of this ordering. Save a lookup from BlockT
655 /// to BlockNode (the index into RPOT) in Nodes.
657 /// 1. Loop initialization (\a initializeLoops()).
659 /// Translate LoopInfo/MachineLoopInfo into a form suitable for the rest of
660 /// the algorithm. In particular, store the immediate members of each loop
661 /// in reverse post-order.
663 /// 2. Calculate mass and scale in loops (\a computeMassInLoops()).
665 /// For each loop (bottom-up), distribute mass through the DAG resulting
666 /// from ignoring backedges and treating sub-loops as a single pseudo-node.
667 /// Track the backedge mass distributed to the loop header, and use it to
668 /// calculate the loop scale (number of loop iterations). Immediate
669 /// members that represent sub-loops will already have been visited and
670 /// packaged into a pseudo-node.
672 /// Distributing mass in a loop is a reverse-post-order traversal through
673 /// the loop. Start by assigning full mass to the Loop header. For each
674 /// node in the loop:
676 /// - Fetch and categorize the weight distribution for its successors.
677 /// If this is a packaged-subloop, the weight distribution is stored
678 /// in \a LoopData::Exits. Otherwise, fetch it from
679 /// BranchProbabilityInfo.
681 /// - Each successor is categorized as \a Weight::Local, a local edge
682 /// within the current loop, \a Weight::Backedge, a backedge to the
683 /// loop header, or \a Weight::Exit, any successor outside the loop.
684 /// The weight, the successor, and its category are stored in \a
685 /// Distribution. There can be multiple edges to each successor.
687 /// - If there's a backedge to a non-header, there's an irreducible SCC.
688 /// The usual flow is temporarily aborted. \a
689 /// computeIrreducibleMass() finds the irreducible SCCs within the
690 /// loop, packages them up, and restarts the flow.
692 /// - Normalize the distribution: scale weights down so that their sum
693 /// is 32-bits, and coalesce multiple edges to the same node.
695 /// - Distribute the mass accordingly, dithering to minimize mass loss,
696 /// as described in \a distributeMass().
698 /// Finally, calculate the loop scale from the accumulated backedge mass.
700 /// 3. Distribute mass in the function (\a computeMassInFunction()).
702 /// Finally, distribute mass through the DAG resulting from packaging all
703 /// loops in the function. This uses the same algorithm as distributing
704 /// mass in a loop, except that there are no exit or backedge edges.
706 /// 4. Unpackage loops (\a unwrapLoops()).
708 /// Initialize each block's frequency to a floating point representation of
711 /// Visit loops top-down, scaling the frequencies of its immediate members
712 /// by the loop's pseudo-node's frequency.
714 /// 5. Convert frequencies to a 64-bit range (\a finalizeMetrics()).
716 /// Using the min and max frequencies as a guide, translate floating point
717 /// frequencies to an appropriate range in uint64_t.
719 /// It has some known flaws.
721 /// - The model of irreducible control flow is a rough approximation.
723 /// Modelling irreducible control flow exactly involves setting up and
724 /// solving a group of infinite geometric series. Such precision is
725 /// unlikely to be worthwhile, since most of our algorithms give up on
726 /// irreducible control flow anyway.
728 /// Nevertheless, we might find that we need to get closer. Here's a sort
729 /// of TODO list for the model with diminishing returns, to be completed as
732 /// - The headers for the \a LoopData representing an irreducible SCC
733 /// include non-entry blocks. When these extra blocks exist, they
734 /// indicate a self-contained irreducible sub-SCC. We could treat them
735 /// as sub-loops, rather than arbitrarily shoving the problematic
736 /// blocks into the headers of the main irreducible SCC.
738 /// - Backedge frequencies are assumed to be evenly split between the
739 /// headers of a given irreducible SCC. Instead, we could track the
740 /// backedge mass separately for each header, and adjust their relative
743 /// - Entry frequencies are assumed to be evenly split between the
744 /// headers of a given irreducible SCC, which is the only option if we
745 /// need to compute mass in the SCC before its parent loop. Instead,
746 /// we could partially compute mass in the parent loop, and stop when
747 /// we get to the SCC. Here, we have the correct ratio of entry
748 /// masses, which we can use to adjust their relative frequencies.
749 /// Compute mass in the SCC, and then continue propagation in the
752 /// - We can propagate mass iteratively through the SCC, for some fixed
753 /// number of iterations. Each iteration starts by assigning the entry
754 /// blocks their backedge mass from the prior iteration. The final
755 /// mass for each block (and each exit, and the total backedge mass
756 /// used for computing loop scale) is the sum of all iterations.
757 /// (Running this until fixed point would "solve" the geometric
758 /// series by simulation.)
759 template <class BT> class BlockFrequencyInfoImpl : BlockFrequencyInfoImplBase {
760 typedef typename bfi_detail::TypeMap<BT>::BlockT BlockT;
761 typedef typename bfi_detail::TypeMap<BT>::FunctionT FunctionT;
762 typedef typename bfi_detail::TypeMap<BT>::BranchProbabilityInfoT
763 BranchProbabilityInfoT;
764 typedef typename bfi_detail::TypeMap<BT>::LoopT LoopT;
765 typedef typename bfi_detail::TypeMap<BT>::LoopInfoT LoopInfoT;
767 // This is part of a workaround for a GCC 4.7 crash on lambdas.
768 friend struct bfi_detail::BlockEdgesAdder<BT>;
770 typedef GraphTraits<const BlockT *> Successor;
771 typedef GraphTraits<Inverse<const BlockT *>> Predecessor;
773 const BranchProbabilityInfoT *BPI;
777 // All blocks in reverse postorder.
778 std::vector<const BlockT *> RPOT;
779 DenseMap<const BlockT *, BlockNode> Nodes;
781 typedef typename std::vector<const BlockT *>::const_iterator rpot_iterator;
783 rpot_iterator rpot_begin() const { return RPOT.begin(); }
784 rpot_iterator rpot_end() const { return RPOT.end(); }
786 size_t getIndex(const rpot_iterator &I) const { return I - rpot_begin(); }
788 BlockNode getNode(const rpot_iterator &I) const {
789 return BlockNode(getIndex(I));
791 BlockNode getNode(const BlockT *BB) const { return Nodes.lookup(BB); }
793 const BlockT *getBlock(const BlockNode &Node) const {
794 assert(Node.Index < RPOT.size());
795 return RPOT[Node.Index];
798 /// \brief Run (and save) a post-order traversal.
800 /// Saves a reverse post-order traversal of all the nodes in \a F.
801 void initializeRPOT();
803 /// \brief Initialize loop data.
805 /// Build up \a Loops using \a LoopInfo. \a LoopInfo gives us a mapping from
806 /// each block to the deepest loop it's in, but we need the inverse. For each
807 /// loop, we store in reverse post-order its "immediate" members, defined as
808 /// the header, the headers of immediate sub-loops, and all other blocks in
809 /// the loop that are not in sub-loops.
810 void initializeLoops();
812 /// \brief Propagate to a block's successors.
814 /// In the context of distributing mass through \c OuterLoop, divide the mass
815 /// currently assigned to \c Node between its successors.
817 /// \return \c true unless there's an irreducible backedge.
818 bool propagateMassToSuccessors(LoopData *OuterLoop, const BlockNode &Node);
820 /// \brief Compute mass in a particular loop.
822 /// Assign mass to \c Loop's header, and then for each block in \c Loop in
823 /// reverse post-order, distribute mass to its successors. Only visits nodes
824 /// that have not been packaged into sub-loops.
826 /// \pre \a computeMassInLoop() has been called for each subloop of \c Loop.
827 /// \return \c true unless there's an irreducible backedge.
828 bool computeMassInLoop(LoopData &Loop);
830 /// \brief Try to compute mass in the top-level function.
832 /// Assign mass to the entry block, and then for each block in reverse
833 /// post-order, distribute mass to its successors. Skips nodes that have
834 /// been packaged into loops.
836 /// \pre \a computeMassInLoops() has been called.
837 /// \return \c true unless there's an irreducible backedge.
838 bool tryToComputeMassInFunction();
840 /// \brief Compute mass in (and package up) irreducible SCCs.
842 /// Find the irreducible SCCs in \c OuterLoop, add them to \a Loops (in front
843 /// of \c Insert), and call \a computeMassInLoop() on each of them.
845 /// If \c OuterLoop is \c nullptr, it refers to the top-level function.
847 /// \pre \a computeMassInLoop() has been called for each subloop of \c
849 /// \pre \c Insert points at the the last loop successfully processed by \a
850 /// computeMassInLoop().
851 /// \pre \c OuterLoop has irreducible SCCs.
852 void computeIrreducibleMass(LoopData *OuterLoop,
853 std::list<LoopData>::iterator Insert);
855 /// \brief Compute mass in all loops.
857 /// For each loop bottom-up, call \a computeMassInLoop().
859 /// \a computeMassInLoop() aborts (and returns \c false) on loops that
860 /// contain a irreducible sub-SCCs. Use \a computeIrreducibleMass() and then
861 /// re-enter \a computeMassInLoop().
863 /// \post \a computeMassInLoop() has returned \c true for every loop.
864 void computeMassInLoops();
866 /// \brief Compute mass in the top-level function.
868 /// Uses \a tryToComputeMassInFunction() and \a computeIrreducibleMass() to
869 /// compute mass in the top-level function.
871 /// \post \a tryToComputeMassInFunction() has returned \c true.
872 void computeMassInFunction();
874 std::string getBlockName(const BlockNode &Node) const override {
875 return bfi_detail::getBlockName(getBlock(Node));
879 const FunctionT *getFunction() const { return F; }
881 void doFunction(const FunctionT *F, const BranchProbabilityInfoT *BPI,
882 const LoopInfoT *LI);
883 BlockFrequencyInfoImpl() : BPI(nullptr), LI(nullptr), F(nullptr) {}
885 using BlockFrequencyInfoImplBase::getEntryFreq;
886 BlockFrequency getBlockFreq(const BlockT *BB) const {
887 return BlockFrequencyInfoImplBase::getBlockFreq(getNode(BB));
889 Scaled64 getFloatingBlockFreq(const BlockT *BB) const {
890 return BlockFrequencyInfoImplBase::getFloatingBlockFreq(getNode(BB));
893 /// \brief Print the frequencies for the current function.
895 /// Prints the frequencies for the blocks in the current function.
897 /// Blocks are printed in the natural iteration order of the function, rather
898 /// than reverse post-order. This provides two advantages: writing -analyze
899 /// tests is easier (since blocks come out in source order), and even
900 /// unreachable blocks are printed.
902 /// \a BlockFrequencyInfoImplBase::print() only knows reverse post-order, so
903 /// we need to override it here.
904 raw_ostream &print(raw_ostream &OS) const override;
905 using BlockFrequencyInfoImplBase::dump;
907 using BlockFrequencyInfoImplBase::printBlockFreq;
908 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockT *BB) const {
909 return BlockFrequencyInfoImplBase::printBlockFreq(OS, getNode(BB));
914 void BlockFrequencyInfoImpl<BT>::doFunction(const FunctionT *F,
915 const BranchProbabilityInfoT *BPI,
916 const LoopInfoT *LI) {
917 // Save the parameters.
922 // Clean up left-over data structures.
923 BlockFrequencyInfoImplBase::clear();
928 DEBUG(dbgs() << "\nblock-frequency: " << F->getName() << "\n================="
929 << std::string(F->getName().size(), '=') << "\n");
933 // Visit loops in post-order to find thelocal mass distribution, and then do
934 // the full function.
935 computeMassInLoops();
936 computeMassInFunction();
941 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeRPOT() {
942 const BlockT *Entry = F->begin();
943 RPOT.reserve(F->size());
944 std::copy(po_begin(Entry), po_end(Entry), std::back_inserter(RPOT));
945 std::reverse(RPOT.begin(), RPOT.end());
947 assert(RPOT.size() - 1 <= BlockNode::getMaxIndex() &&
948 "More nodes in function than Block Frequency Info supports");
950 DEBUG(dbgs() << "reverse-post-order-traversal\n");
951 for (rpot_iterator I = rpot_begin(), E = rpot_end(); I != E; ++I) {
952 BlockNode Node = getNode(I);
953 DEBUG(dbgs() << " - " << getIndex(I) << ": " << getBlockName(Node) << "\n");
957 Working.reserve(RPOT.size());
958 for (size_t Index = 0; Index < RPOT.size(); ++Index)
959 Working.emplace_back(Index);
960 Freqs.resize(RPOT.size());
963 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeLoops() {
964 DEBUG(dbgs() << "loop-detection\n");
968 // Visit loops top down and assign them an index.
969 std::deque<std::pair<const LoopT *, LoopData *>> Q;
970 for (const LoopT *L : *LI)
971 Q.emplace_back(L, nullptr);
973 const LoopT *Loop = Q.front().first;
974 LoopData *Parent = Q.front().second;
977 BlockNode Header = getNode(Loop->getHeader());
978 assert(Header.isValid());
980 Loops.emplace_back(Parent, Header);
981 Working[Header.Index].Loop = &Loops.back();
982 DEBUG(dbgs() << " - loop = " << getBlockName(Header) << "\n");
984 for (const LoopT *L : *Loop)
985 Q.emplace_back(L, &Loops.back());
988 // Visit nodes in reverse post-order and add them to their deepest containing
990 for (size_t Index = 0; Index < RPOT.size(); ++Index) {
991 // Loop headers have already been mostly mapped.
992 if (Working[Index].isLoopHeader()) {
993 LoopData *ContainingLoop = Working[Index].getContainingLoop();
995 ContainingLoop->Nodes.push_back(Index);
999 const LoopT *Loop = LI->getLoopFor(RPOT[Index]);
1003 // Add this node to its containing loop's member list.
1004 BlockNode Header = getNode(Loop->getHeader());
1005 assert(Header.isValid());
1006 const auto &HeaderData = Working[Header.Index];
1007 assert(HeaderData.isLoopHeader());
1009 Working[Index].Loop = HeaderData.Loop;
1010 HeaderData.Loop->Nodes.push_back(Index);
1011 DEBUG(dbgs() << " - loop = " << getBlockName(Header)
1012 << ": member = " << getBlockName(Index) << "\n");
1016 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInLoops() {
1017 // Visit loops with the deepest first, and the top-level loops last.
1018 for (auto L = Loops.rbegin(), E = Loops.rend(); L != E; ++L) {
1019 if (computeMassInLoop(*L))
1021 auto Next = std::next(L);
1022 computeIrreducibleMass(&*L, L.base());
1023 L = std::prev(Next);
1024 if (computeMassInLoop(*L))
1026 llvm_unreachable("unhandled irreducible control flow");
1031 bool BlockFrequencyInfoImpl<BT>::computeMassInLoop(LoopData &Loop) {
1032 // Compute mass in loop.
1033 DEBUG(dbgs() << "compute-mass-in-loop: " << getLoopName(Loop) << "\n");
1035 if (Loop.isIrreducible()) {
1036 BlockMass Remaining = BlockMass::getFull();
1037 for (uint32_t H = 0; H < Loop.NumHeaders; ++H) {
1038 auto &Mass = Working[Loop.Nodes[H].Index].getMass();
1039 Mass = Remaining * BranchProbability(1, Loop.NumHeaders - H);
1042 for (const BlockNode &M : Loop.Nodes)
1043 if (!propagateMassToSuccessors(&Loop, M))
1044 llvm_unreachable("unhandled irreducible control flow");
1046 Working[Loop.getHeader().Index].getMass() = BlockMass::getFull();
1047 if (!propagateMassToSuccessors(&Loop, Loop.getHeader()))
1048 llvm_unreachable("irreducible control flow to loop header!?");
1049 for (const BlockNode &M : Loop.members())
1050 if (!propagateMassToSuccessors(&Loop, M))
1051 // Irreducible backedge.
1055 computeLoopScale(Loop);
1061 bool BlockFrequencyInfoImpl<BT>::tryToComputeMassInFunction() {
1062 // Compute mass in function.
1063 DEBUG(dbgs() << "compute-mass-in-function\n");
1064 assert(!Working.empty() && "no blocks in function");
1065 assert(!Working[0].isLoopHeader() && "entry block is a loop header");
1067 Working[0].getMass() = BlockMass::getFull();
1068 for (rpot_iterator I = rpot_begin(), IE = rpot_end(); I != IE; ++I) {
1069 // Check for nodes that have been packaged.
1070 BlockNode Node = getNode(I);
1071 if (Working[Node.Index].isPackaged())
1074 if (!propagateMassToSuccessors(nullptr, Node))
1080 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInFunction() {
1081 if (tryToComputeMassInFunction())
1083 computeIrreducibleMass(nullptr, Loops.begin());
1084 if (tryToComputeMassInFunction())
1086 llvm_unreachable("unhandled irreducible control flow");
1089 /// \note This should be a lambda, but that crashes GCC 4.7.
1090 namespace bfi_detail {
1091 template <class BT> struct BlockEdgesAdder {
1093 typedef BlockFrequencyInfoImplBase::LoopData LoopData;
1094 typedef GraphTraits<const BlockT *> Successor;
1096 const BlockFrequencyInfoImpl<BT> &BFI;
1097 explicit BlockEdgesAdder(const BlockFrequencyInfoImpl<BT> &BFI)
1099 void operator()(IrreducibleGraph &G, IrreducibleGraph::IrrNode &Irr,
1100 const LoopData *OuterLoop) {
1101 const BlockT *BB = BFI.RPOT[Irr.Node.Index];
1102 for (auto I = Successor::child_begin(BB), E = Successor::child_end(BB);
1104 G.addEdge(Irr, BFI.getNode(*I), OuterLoop);
1109 void BlockFrequencyInfoImpl<BT>::computeIrreducibleMass(
1110 LoopData *OuterLoop, std::list<LoopData>::iterator Insert) {
1111 DEBUG(dbgs() << "analyze-irreducible-in-";
1112 if (OuterLoop) dbgs() << "loop: " << getLoopName(*OuterLoop) << "\n";
1113 else dbgs() << "function\n");
1115 using namespace bfi_detail;
1116 // Ideally, addBlockEdges() would be declared here as a lambda, but that
1118 BlockEdgesAdder<BT> addBlockEdges(*this);
1119 IrreducibleGraph G(*this, OuterLoop, addBlockEdges);
1121 for (auto &L : analyzeIrreducible(G, OuterLoop, Insert))
1122 computeMassInLoop(L);
1126 updateLoopWithIrreducible(*OuterLoop);
1131 BlockFrequencyInfoImpl<BT>::propagateMassToSuccessors(LoopData *OuterLoop,
1132 const BlockNode &Node) {
1133 DEBUG(dbgs() << " - node: " << getBlockName(Node) << "\n");
1134 // Calculate probability for successors.
1136 if (auto *Loop = Working[Node.Index].getPackagedLoop()) {
1137 assert(Loop != OuterLoop && "Cannot propagate mass in a packaged loop");
1138 if (!addLoopSuccessorsToDist(OuterLoop, *Loop, Dist))
1139 // Irreducible backedge.
1142 const BlockT *BB = getBlock(Node);
1143 for (auto SI = Successor::child_begin(BB), SE = Successor::child_end(BB);
1145 // Do not dereference SI, or getEdgeWeight() is linear in the number of
1147 if (!addToDist(Dist, OuterLoop, Node, getNode(*SI),
1148 BPI->getEdgeWeight(BB, SI)))
1149 // Irreducible backedge.
1153 // Distribute mass to successors, saving exit and backedge data in the
1155 distributeMass(Node, OuterLoop, Dist);
1160 raw_ostream &BlockFrequencyInfoImpl<BT>::print(raw_ostream &OS) const {
1163 OS << "block-frequency-info: " << F->getName() << "\n";
1164 for (const BlockT &BB : *F)
1165 OS << " - " << bfi_detail::getBlockName(&BB)
1166 << ": float = " << getFloatingBlockFreq(&BB)
1167 << ", int = " << getBlockFreq(&BB).getFrequency() << "\n";
1169 // Add an extra newline for readability.
1174 } // end namespace llvm