LLVM 19.0.0git
BlockFrequencyInfoImpl.h
Go to the documentation of this file.
1//==- BlockFrequencyInfoImpl.h - Block Frequency Implementation --*- C++ -*-==//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// Shared implementation of BlockFrequency for IR and Machine Instructions.
10// See the documentation below for BlockFrequencyInfoImpl for details.
11//
12//===----------------------------------------------------------------------===//
13
14#ifndef LLVM_ANALYSIS_BLOCKFREQUENCYINFOIMPL_H
15#define LLVM_ANALYSIS_BLOCKFREQUENCYINFOIMPL_H
16
17#include "llvm/ADT/BitVector.h"
18#include "llvm/ADT/DenseMap.h"
19#include "llvm/ADT/DenseSet.h"
25#include "llvm/ADT/Twine.h"
27#include "llvm/IR/BasicBlock.h"
28#include "llvm/IR/Function.h"
29#include "llvm/IR/ValueHandle.h"
34#include "llvm/Support/Debug.h"
35#include "llvm/Support/Format.h"
38#include <algorithm>
39#include <cassert>
40#include <cstddef>
41#include <cstdint>
42#include <deque>
43#include <iterator>
44#include <limits>
45#include <list>
46#include <optional>
47#include <queue>
48#include <string>
49#include <utility>
50#include <vector>
51
52#define DEBUG_TYPE "block-freq"
53
54namespace llvm {
56
60
61class BranchProbabilityInfo;
62class Function;
63class Loop;
64class LoopInfo;
65class MachineBasicBlock;
66class MachineBranchProbabilityInfo;
67class MachineFunction;
68class MachineLoop;
69class MachineLoopInfo;
70
71namespace bfi_detail {
72
73struct IrreducibleGraph;
74
75// This is part of a workaround for a GCC 4.7 crash on lambdas.
76template <class BT> struct BlockEdgesAdder;
77
78/// Mass of a block.
79///
80/// This class implements a sort of fixed-point fraction always between 0.0 and
81/// 1.0. getMass() == std::numeric_limits<uint64_t>::max() indicates a value of
82/// 1.0.
83///
84/// Masses can be added and subtracted. Simple saturation arithmetic is used,
85/// so arithmetic operations never overflow or underflow.
86///
87/// Masses can be multiplied. Multiplication treats full mass as 1.0 and uses
88/// an inexpensive floating-point algorithm that's off-by-one (almost, but not
89/// quite, maximum precision).
90///
91/// Masses can be scaled by \a BranchProbability at maximum precision.
92class BlockMass {
93 uint64_t Mass = 0;
94
95public:
96 BlockMass() = default;
97 explicit BlockMass(uint64_t Mass) : Mass(Mass) {}
98
99 static BlockMass getEmpty() { return BlockMass(); }
100
102 return BlockMass(std::numeric_limits<uint64_t>::max());
103 }
104
105 uint64_t getMass() const { return Mass; }
106
107 bool isFull() const { return Mass == std::numeric_limits<uint64_t>::max(); }
108 bool isEmpty() const { return !Mass; }
109
110 bool operator!() const { return isEmpty(); }
111
112 /// Add another mass.
113 ///
114 /// Adds another mass, saturating at \a isFull() rather than overflowing.
116 uint64_t Sum = Mass + X.Mass;
117 Mass = Sum < Mass ? std::numeric_limits<uint64_t>::max() : Sum;
118 return *this;
119 }
120
121 /// Subtract another mass.
122 ///
123 /// Subtracts another mass, saturating at \a isEmpty() rather than
124 /// undeflowing.
126 uint64_t Diff = Mass - X.Mass;
127 Mass = Diff > Mass ? 0 : Diff;
128 return *this;
129 }
130
132 Mass = P.scale(Mass);
133 return *this;
134 }
135
136 bool operator==(BlockMass X) const { return Mass == X.Mass; }
137 bool operator!=(BlockMass X) const { return Mass != X.Mass; }
138 bool operator<=(BlockMass X) const { return Mass <= X.Mass; }
139 bool operator>=(BlockMass X) const { return Mass >= X.Mass; }
140 bool operator<(BlockMass X) const { return Mass < X.Mass; }
141 bool operator>(BlockMass X) const { return Mass > X.Mass; }
142
143 /// Convert to scaled number.
144 ///
145 /// Convert to \a ScaledNumber. \a isFull() gives 1.0, while \a isEmpty()
146 /// gives slightly above 0.0.
148
149 void dump() const;
151};
152
154 return BlockMass(L) += R;
155}
157 return BlockMass(L) -= R;
158}
160 return BlockMass(L) *= R;
161}
163 return BlockMass(R) *= L;
164}
165
167 return X.print(OS);
168}
169
170} // end namespace bfi_detail
171
172/// Base class for BlockFrequencyInfoImpl
173///
174/// BlockFrequencyInfoImplBase has supporting data structures and some
175/// algorithms for BlockFrequencyInfoImplBase. Only algorithms that depend on
176/// the block type (or that call such algorithms) are skipped here.
177///
178/// Nevertheless, the majority of the overall algorithm documentation lives with
179/// BlockFrequencyInfoImpl. See there for details.
181public:
184
185 /// Representative of a block.
186 ///
187 /// This is a simple wrapper around an index into the reverse-post-order
188 /// traversal of the blocks.
189 ///
190 /// Unlike a block pointer, its order has meaning (location in the
191 /// topological sort) and it's class is the same regardless of block type.
192 struct BlockNode {
194
196
197 BlockNode() : Index(std::numeric_limits<uint32_t>::max()) {}
199
200 bool operator==(const BlockNode &X) const { return Index == X.Index; }
201 bool operator!=(const BlockNode &X) const { return Index != X.Index; }
202 bool operator<=(const BlockNode &X) const { return Index <= X.Index; }
203 bool operator>=(const BlockNode &X) const { return Index >= X.Index; }
204 bool operator<(const BlockNode &X) const { return Index < X.Index; }
205 bool operator>(const BlockNode &X) const { return Index > X.Index; }
206
207 bool isValid() const { return Index <= getMaxIndex(); }
208
209 static size_t getMaxIndex() {
210 return std::numeric_limits<uint32_t>::max() - 1;
211 }
212 };
213
214 /// Stats about a block itself.
218 };
219
220 /// Data about a loop.
221 ///
222 /// Contains the data necessary to represent a loop as a pseudo-node once it's
223 /// packaged.
224 struct LoopData {
228
229 LoopData *Parent; ///< The parent loop.
230 bool IsPackaged = false; ///< Whether this has been packaged.
231 uint32_t NumHeaders = 1; ///< Number of headers.
232 ExitMap Exits; ///< Successor edges (and weights).
233 NodeList Nodes; ///< Header and the members of the loop.
234 HeaderMassList BackedgeMass; ///< Mass returned to each loop header.
237
239 : Parent(Parent), Nodes(1, Header), BackedgeMass(1) {}
240
241 template <class It1, class It2>
242 LoopData(LoopData *Parent, It1 FirstHeader, It1 LastHeader, It2 FirstOther,
243 It2 LastOther)
244 : Parent(Parent), Nodes(FirstHeader, LastHeader) {
246 Nodes.insert(Nodes.end(), FirstOther, LastOther);
248 }
249
250 bool isHeader(const BlockNode &Node) const {
251 if (isIrreducible())
252 return std::binary_search(Nodes.begin(), Nodes.begin() + NumHeaders,
253 Node);
254 return Node == Nodes[0];
255 }
256
257 BlockNode getHeader() const { return Nodes[0]; }
258 bool isIrreducible() const { return NumHeaders > 1; }
259
261 assert(isHeader(B) && "this is only valid on loop header blocks");
262 if (isIrreducible())
263 return std::lower_bound(Nodes.begin(), Nodes.begin() + NumHeaders, B) -
264 Nodes.begin();
265 return 0;
266 }
267
269 return Nodes.begin() + NumHeaders;
270 }
271
275 }
276 };
277
278 /// Index of loop information.
279 struct WorkingData {
280 BlockNode Node; ///< This node.
281 LoopData *Loop = nullptr; ///< The loop this block is inside.
282 BlockMass Mass; ///< Mass distribution from the entry block.
283
285
286 bool isLoopHeader() const { return Loop && Loop->isHeader(Node); }
287
288 bool isDoubleLoopHeader() const {
289 return isLoopHeader() && Loop->Parent && Loop->Parent->isIrreducible() &&
290 Loop->Parent->isHeader(Node);
291 }
292
294 if (!isLoopHeader())
295 return Loop;
296 if (!isDoubleLoopHeader())
297 return Loop->Parent;
298 return Loop->Parent->Parent;
299 }
300
301 /// Resolve a node to its representative.
302 ///
303 /// Get the node currently representing Node, which could be a containing
304 /// loop.
305 ///
306 /// This function should only be called when distributing mass. As long as
307 /// there are no irreducible edges to Node, then it will have complexity
308 /// O(1) in this context.
309 ///
310 /// In general, the complexity is O(L), where L is the number of loop
311 /// headers Node has been packaged into. Since this method is called in
312 /// the context of distributing mass, L will be the number of loop headers
313 /// an early exit edge jumps out of.
315 auto *L = getPackagedLoop();
316 return L ? L->getHeader() : Node;
317 }
318
320 if (!Loop || !Loop->IsPackaged)
321 return nullptr;
322 auto *L = Loop;
323 while (L->Parent && L->Parent->IsPackaged)
324 L = L->Parent;
325 return L;
326 }
327
328 /// Get the appropriate mass for a node.
329 ///
330 /// Get appropriate mass for Node. If Node is a loop-header (whose loop
331 /// has been packaged), returns the mass of its pseudo-node. If it's a
332 /// node inside a packaged loop, it returns the loop's mass.
334 if (!isAPackage())
335 return Mass;
336 if (!isADoublePackage())
337 return Loop->Mass;
338 return Loop->Parent->Mass;
339 }
340
341 /// Has ContainingLoop been packaged up?
342 bool isPackaged() const { return getResolvedNode() != Node; }
343
344 /// Has Loop been packaged up?
345 bool isAPackage() const { return isLoopHeader() && Loop->IsPackaged; }
346
347 /// Has Loop been packaged up twice?
348 bool isADoublePackage() const {
349 return isDoubleLoopHeader() && Loop->Parent->IsPackaged;
350 }
351 };
352
353 /// Unscaled probability weight.
354 ///
355 /// Probability weight for an edge in the graph (including the
356 /// successor/target node).
357 ///
358 /// All edges in the original function are 32-bit. However, exit edges from
359 /// loop packages are taken from 64-bit exit masses, so we need 64-bits of
360 /// space in general.
361 ///
362 /// In addition to the raw weight amount, Weight stores the type of the edge
363 /// in the current context (i.e., the context of the loop being processed).
364 /// Is this a local edge within the loop, an exit from the loop, or a
365 /// backedge to the loop header?
366 struct Weight {
371
372 Weight() = default;
375 };
376
377 /// Distribution of unscaled probability weight.
378 ///
379 /// Distribution of unscaled probability weight to a set of successors.
380 ///
381 /// This class collates the successor edge weights for later processing.
382 ///
383 /// \a DidOverflow indicates whether \a Total did overflow while adding to
384 /// the distribution. It should never overflow twice.
387
388 WeightList Weights; ///< Individual successor weights.
389 uint64_t Total = 0; ///< Sum of all weights.
390 bool DidOverflow = false; ///< Whether \a Total did overflow.
391
392 Distribution() = default;
393
394 void addLocal(const BlockNode &Node, uint64_t Amount) {
395 add(Node, Amount, Weight::Local);
396 }
397
398 void addExit(const BlockNode &Node, uint64_t Amount) {
399 add(Node, Amount, Weight::Exit);
400 }
401
402 void addBackedge(const BlockNode &Node, uint64_t Amount) {
403 add(Node, Amount, Weight::Backedge);
404 }
405
406 /// Normalize the distribution.
407 ///
408 /// Combines multiple edges to the same \a Weight::TargetNode and scales
409 /// down so that \a Total fits into 32-bits.
410 ///
411 /// This is linear in the size of \a Weights. For the vast majority of
412 /// cases, adjacent edge weights are combined by sorting WeightList and
413 /// combining adjacent weights. However, for very large edge lists an
414 /// auxiliary hash table is used.
415 void normalize();
416
417 private:
418 void add(const BlockNode &Node, uint64_t Amount, Weight::DistType Type);
419 };
420
421 /// Data about each block. This is used downstream.
422 std::vector<FrequencyData> Freqs;
423
424 /// Whether each block is an irreducible loop header.
425 /// This is used downstream.
427
428 /// Loop data: see initializeLoops().
429 std::vector<WorkingData> Working;
430
431 /// Indexed information about loops.
432 std::list<LoopData> Loops;
433
434 /// Virtual destructor.
435 ///
436 /// Need a virtual destructor to mask the compiler warning about
437 /// getBlockName().
438 virtual ~BlockFrequencyInfoImplBase() = default;
439
440 /// Add all edges out of a packaged loop to the distribution.
441 ///
442 /// Adds all edges from LocalLoopHead to Dist. Calls addToDist() to add each
443 /// successor edge.
444 ///
445 /// \return \c true unless there's an irreducible backedge.
446 bool addLoopSuccessorsToDist(const LoopData *OuterLoop, LoopData &Loop,
447 Distribution &Dist);
448
449 /// Add an edge to the distribution.
450 ///
451 /// Adds an edge to Succ to Dist. If \c LoopHead.isValid(), then whether the
452 /// edge is local/exit/backedge is in the context of LoopHead. Otherwise,
453 /// every edge should be a local edge (since all the loops are packaged up).
454 ///
455 /// \return \c true unless aborted due to an irreducible backedge.
456 bool addToDist(Distribution &Dist, const LoopData *OuterLoop,
457 const BlockNode &Pred, const BlockNode &Succ, uint64_t Weight);
458
459 /// Analyze irreducible SCCs.
460 ///
461 /// Separate irreducible SCCs from \c G, which is an explicit graph of \c
462 /// OuterLoop (or the top-level function, if \c OuterLoop is \c nullptr).
463 /// Insert them into \a Loops before \c Insert.
464 ///
465 /// \return the \c LoopData nodes representing the irreducible SCCs.
468 std::list<LoopData>::iterator Insert);
469
470 /// Update a loop after packaging irreducible SCCs inside of it.
471 ///
472 /// Update \c OuterLoop. Before finding irreducible control flow, it was
473 /// partway through \a computeMassInLoop(), so \a LoopData::Exits and \a
474 /// LoopData::BackedgeMass need to be reset. Also, nodes that were packaged
475 /// up need to be removed from \a OuterLoop::Nodes.
476 void updateLoopWithIrreducible(LoopData &OuterLoop);
477
478 /// Distribute mass according to a distribution.
479 ///
480 /// Distributes the mass in Source according to Dist. If LoopHead.isValid(),
481 /// backedges and exits are stored in its entry in Loops.
482 ///
483 /// Mass is distributed in parallel from two copies of the source mass.
484 void distributeMass(const BlockNode &Source, LoopData *OuterLoop,
485 Distribution &Dist);
486
487 /// Compute the loop scale for a loop.
489
490 /// Adjust the mass of all headers in an irreducible loop.
491 ///
492 /// Initially, irreducible loops are assumed to distribute their mass
493 /// equally among its headers. This can lead to wrong frequency estimates
494 /// since some headers may be executed more frequently than others.
495 ///
496 /// This adjusts header mass distribution so it matches the weights of
497 /// the backedges going into each of the loop headers.
499
501
502 /// Package up a loop.
504
505 /// Unwrap loops.
506 void unwrapLoops();
507
508 /// Finalize frequency metrics.
509 ///
510 /// Calculates final frequencies and cleans up no-longer-needed data
511 /// structures.
512 void finalizeMetrics();
513
514 /// Clear all memory.
515 void clear();
516
517 virtual std::string getBlockName(const BlockNode &Node) const;
518 std::string getLoopName(const LoopData &Loop) const;
519
520 virtual raw_ostream &print(raw_ostream &OS) const { return OS; }
521 void dump() const { print(dbgs()); }
522
523 Scaled64 getFloatingBlockFreq(const BlockNode &Node) const;
524
525 BlockFrequency getBlockFreq(const BlockNode &Node) const;
526 std::optional<uint64_t>
527 getBlockProfileCount(const Function &F, const BlockNode &Node,
528 bool AllowSynthetic = false) const;
529 std::optional<uint64_t>
531 bool AllowSynthetic = false) const;
532 bool isIrrLoopHeader(const BlockNode &Node);
533
534 void setBlockFreq(const BlockNode &Node, BlockFrequency Freq);
535
537 assert(!Freqs.empty());
538 return BlockFrequency(Freqs[0].Integer);
539 }
540};
541
542namespace bfi_detail {
543
544template <class BlockT> struct TypeMap {};
545template <> struct TypeMap<BasicBlock> {
550 using LoopT = Loop;
552};
553template <> struct TypeMap<MachineBasicBlock> {
560};
561
562template <class BlockT, class BFIImplT>
564
565/// Get the name of a MachineBasicBlock.
566///
567/// Get the name of a MachineBasicBlock. It's templated so that including from
568/// CodeGen is unnecessary (that would be a layering issue).
569///
570/// This is used mainly for debug output. The name is similar to
571/// MachineBasicBlock::getFullName(), but skips the name of the function.
572template <class BlockT> std::string getBlockName(const BlockT *BB) {
573 assert(BB && "Unexpected nullptr");
574 auto MachineName = "BB" + Twine(BB->getNumber());
575 if (BB->getBasicBlock())
576 return (MachineName + "[" + BB->getName() + "]").str();
577 return MachineName.str();
578}
579/// Get the name of a BasicBlock.
580template <> inline std::string getBlockName(const BasicBlock *BB) {
581 assert(BB && "Unexpected nullptr");
582 return BB->getName().str();
583}
584
585/// Graph of irreducible control flow.
586///
587/// This graph is used for determining the SCCs in a loop (or top-level
588/// function) that has irreducible control flow.
589///
590/// During the block frequency algorithm, the local graphs are defined in a
591/// light-weight way, deferring to the \a BasicBlock or \a MachineBasicBlock
592/// graphs for most edges, but getting others from \a LoopData::ExitMap. The
593/// latter only has successor information.
594///
595/// \a IrreducibleGraph makes this graph explicit. It's in a form that can use
596/// \a GraphTraits (so that \a analyzeIrreducible() can use \a scc_iterator),
597/// and it explicitly lists predecessors and successors. The initialization
598/// that relies on \c MachineBasicBlock is defined in the header.
601
603
605 struct IrrNode {
607 unsigned NumIn = 0;
608 std::deque<const IrrNode *> Edges;
609
611
612 using iterator = std::deque<const IrrNode *>::const_iterator;
613
614 iterator pred_begin() const { return Edges.begin(); }
615 iterator succ_begin() const { return Edges.begin() + NumIn; }
616 iterator pred_end() const { return succ_begin(); }
617 iterator succ_end() const { return Edges.end(); }
618 };
620 const IrrNode *StartIrr = nullptr;
621 std::vector<IrrNode> Nodes;
623
624 /// Construct an explicit graph containing irreducible control flow.
625 ///
626 /// Construct an explicit graph of the control flow in \c OuterLoop (or the
627 /// top-level function, if \c OuterLoop is \c nullptr). Uses \c
628 /// addBlockEdges to add block successors that have not been packaged into
629 /// loops.
630 ///
631 /// \a BlockFrequencyInfoImpl::computeIrreducibleMass() is the only expected
632 /// user of this.
633 template <class BlockEdgesAdder>
635 BlockEdgesAdder addBlockEdges) : BFI(BFI) {
636 initialize(OuterLoop, addBlockEdges);
637 }
638
639 template <class BlockEdgesAdder>
640 void initialize(const BFIBase::LoopData *OuterLoop,
641 BlockEdgesAdder addBlockEdges);
642 void addNodesInLoop(const BFIBase::LoopData &OuterLoop);
643 void addNodesInFunction();
644
645 void addNode(const BlockNode &Node) {
646 Nodes.emplace_back(Node);
647 BFI.Working[Node.Index].getMass() = BlockMass::getEmpty();
648 }
649
650 void indexNodes();
651 template <class BlockEdgesAdder>
652 void addEdges(const BlockNode &Node, const BFIBase::LoopData *OuterLoop,
653 BlockEdgesAdder addBlockEdges);
654 void addEdge(IrrNode &Irr, const BlockNode &Succ,
655 const BFIBase::LoopData *OuterLoop);
656};
657
658template <class BlockEdgesAdder>
660 BlockEdgesAdder addBlockEdges) {
661 if (OuterLoop) {
662 addNodesInLoop(*OuterLoop);
663 for (auto N : OuterLoop->Nodes)
664 addEdges(N, OuterLoop, addBlockEdges);
665 } else {
667 for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index)
668 addEdges(Index, OuterLoop, addBlockEdges);
669 }
671}
672
673template <class BlockEdgesAdder>
675 const BFIBase::LoopData *OuterLoop,
676 BlockEdgesAdder addBlockEdges) {
677 auto L = Lookup.find(Node.Index);
678 if (L == Lookup.end())
679 return;
680 IrrNode &Irr = *L->second;
681 const auto &Working = BFI.Working[Node.Index];
682
683 if (Working.isAPackage())
684 for (const auto &I : Working.Loop->Exits)
685 addEdge(Irr, I.first, OuterLoop);
686 else
687 addBlockEdges(*this, Irr, OuterLoop);
688}
689
690} // end namespace bfi_detail
691
692/// Shared implementation for block frequency analysis.
693///
694/// This is a shared implementation of BlockFrequencyInfo and
695/// MachineBlockFrequencyInfo, and calculates the relative frequencies of
696/// blocks.
697///
698/// LoopInfo defines a loop as a "non-trivial" SCC dominated by a single block,
699/// which is called the header. A given loop, L, can have sub-loops, which are
700/// loops within the subgraph of L that exclude its header. (A "trivial" SCC
701/// consists of a single block that does not have a self-edge.)
702///
703/// In addition to loops, this algorithm has limited support for irreducible
704/// SCCs, which are SCCs with multiple entry blocks. Irreducible SCCs are
705/// discovered on the fly, and modelled as loops with multiple headers.
706///
707/// The headers of irreducible sub-SCCs consist of its entry blocks and all
708/// nodes that are targets of a backedge within it (excluding backedges within
709/// true sub-loops). Block frequency calculations act as if a block is
710/// inserted that intercepts all the edges to the headers. All backedges and
711/// entries point to this block. Its successors are the headers, which split
712/// the frequency evenly.
713///
714/// This algorithm leverages BlockMass and ScaledNumber to maintain precision,
715/// separates mass distribution from loop scaling, and dithers to eliminate
716/// probability mass loss.
717///
718/// The implementation is split between BlockFrequencyInfoImpl, which knows the
719/// type of graph being modelled (BasicBlock vs. MachineBasicBlock), and
720/// BlockFrequencyInfoImplBase, which doesn't. The base class uses \a
721/// BlockNode, a wrapper around a uint32_t. BlockNode is numbered from 0 in
722/// reverse-post order. This gives two advantages: it's easy to compare the
723/// relative ordering of two nodes, and maps keyed on BlockT can be represented
724/// by vectors.
725///
726/// This algorithm is O(V+E), unless there is irreducible control flow, in
727/// which case it's O(V*E) in the worst case.
728///
729/// These are the main stages:
730///
731/// 0. Reverse post-order traversal (\a initializeRPOT()).
732///
733/// Run a single post-order traversal and save it (in reverse) in RPOT.
734/// All other stages make use of this ordering. Save a lookup from BlockT
735/// to BlockNode (the index into RPOT) in Nodes.
736///
737/// 1. Loop initialization (\a initializeLoops()).
738///
739/// Translate LoopInfo/MachineLoopInfo into a form suitable for the rest of
740/// the algorithm. In particular, store the immediate members of each loop
741/// in reverse post-order.
742///
743/// 2. Calculate mass and scale in loops (\a computeMassInLoops()).
744///
745/// For each loop (bottom-up), distribute mass through the DAG resulting
746/// from ignoring backedges and treating sub-loops as a single pseudo-node.
747/// Track the backedge mass distributed to the loop header, and use it to
748/// calculate the loop scale (number of loop iterations). Immediate
749/// members that represent sub-loops will already have been visited and
750/// packaged into a pseudo-node.
751///
752/// Distributing mass in a loop is a reverse-post-order traversal through
753/// the loop. Start by assigning full mass to the Loop header. For each
754/// node in the loop:
755///
756/// - Fetch and categorize the weight distribution for its successors.
757/// If this is a packaged-subloop, the weight distribution is stored
758/// in \a LoopData::Exits. Otherwise, fetch it from
759/// BranchProbabilityInfo.
760///
761/// - Each successor is categorized as \a Weight::Local, a local edge
762/// within the current loop, \a Weight::Backedge, a backedge to the
763/// loop header, or \a Weight::Exit, any successor outside the loop.
764/// The weight, the successor, and its category are stored in \a
765/// Distribution. There can be multiple edges to each successor.
766///
767/// - If there's a backedge to a non-header, there's an irreducible SCC.
768/// The usual flow is temporarily aborted. \a
769/// computeIrreducibleMass() finds the irreducible SCCs within the
770/// loop, packages them up, and restarts the flow.
771///
772/// - Normalize the distribution: scale weights down so that their sum
773/// is 32-bits, and coalesce multiple edges to the same node.
774///
775/// - Distribute the mass accordingly, dithering to minimize mass loss,
776/// as described in \a distributeMass().
777///
778/// In the case of irreducible loops, instead of a single loop header,
779/// there will be several. The computation of backedge masses is similar
780/// but instead of having a single backedge mass, there will be one
781/// backedge per loop header. In these cases, each backedge will carry
782/// a mass proportional to the edge weights along the corresponding
783/// path.
784///
785/// At the end of propagation, the full mass assigned to the loop will be
786/// distributed among the loop headers proportionally according to the
787/// mass flowing through their backedges.
788///
789/// Finally, calculate the loop scale from the accumulated backedge mass.
790///
791/// 3. Distribute mass in the function (\a computeMassInFunction()).
792///
793/// Finally, distribute mass through the DAG resulting from packaging all
794/// loops in the function. This uses the same algorithm as distributing
795/// mass in a loop, except that there are no exit or backedge edges.
796///
797/// 4. Unpackage loops (\a unwrapLoops()).
798///
799/// Initialize each block's frequency to a floating point representation of
800/// its mass.
801///
802/// Visit loops top-down, scaling the frequencies of its immediate members
803/// by the loop's pseudo-node's frequency.
804///
805/// 5. Convert frequencies to a 64-bit range (\a finalizeMetrics()).
806///
807/// Using the min and max frequencies as a guide, translate floating point
808/// frequencies to an appropriate range in uint64_t.
809///
810/// It has some known flaws.
811///
812/// - The model of irreducible control flow is a rough approximation.
813///
814/// Modelling irreducible control flow exactly involves setting up and
815/// solving a group of infinite geometric series. Such precision is
816/// unlikely to be worthwhile, since most of our algorithms give up on
817/// irreducible control flow anyway.
818///
819/// Nevertheless, we might find that we need to get closer. Here's a sort
820/// of TODO list for the model with diminishing returns, to be completed as
821/// necessary.
822///
823/// - The headers for the \a LoopData representing an irreducible SCC
824/// include non-entry blocks. When these extra blocks exist, they
825/// indicate a self-contained irreducible sub-SCC. We could treat them
826/// as sub-loops, rather than arbitrarily shoving the problematic
827/// blocks into the headers of the main irreducible SCC.
828///
829/// - Entry frequencies are assumed to be evenly split between the
830/// headers of a given irreducible SCC, which is the only option if we
831/// need to compute mass in the SCC before its parent loop. Instead,
832/// we could partially compute mass in the parent loop, and stop when
833/// we get to the SCC. Here, we have the correct ratio of entry
834/// masses, which we can use to adjust their relative frequencies.
835/// Compute mass in the SCC, and then continue propagation in the
836/// parent.
837///
838/// - We can propagate mass iteratively through the SCC, for some fixed
839/// number of iterations. Each iteration starts by assigning the entry
840/// blocks their backedge mass from the prior iteration. The final
841/// mass for each block (and each exit, and the total backedge mass
842/// used for computing loop scale) is the sum of all iterations.
843/// (Running this until fixed point would "solve" the geometric
844/// series by simulation.)
846 // This is part of a workaround for a GCC 4.7 crash on lambdas.
847 friend struct bfi_detail::BlockEdgesAdder<BT>;
848
849 using BlockT = typename bfi_detail::TypeMap<BT>::BlockT;
850 using BlockKeyT = typename bfi_detail::TypeMap<BT>::BlockKeyT;
851 using FunctionT = typename bfi_detail::TypeMap<BT>::FunctionT;
852 using BranchProbabilityInfoT =
854 using LoopT = typename bfi_detail::TypeMap<BT>::LoopT;
855 using LoopInfoT = typename bfi_detail::TypeMap<BT>::LoopInfoT;
858 using BFICallbackVH =
860
861 const BranchProbabilityInfoT *BPI = nullptr;
862 const LoopInfoT *LI = nullptr;
863 const FunctionT *F = nullptr;
864
865 // All blocks in reverse postorder.
866 std::vector<const BlockT *> RPOT;
868
869 using rpot_iterator = typename std::vector<const BlockT *>::const_iterator;
870
871 rpot_iterator rpot_begin() const { return RPOT.begin(); }
872 rpot_iterator rpot_end() const { return RPOT.end(); }
873
874 size_t getIndex(const rpot_iterator &I) const { return I - rpot_begin(); }
875
876 BlockNode getNode(const rpot_iterator &I) const {
877 return BlockNode(getIndex(I));
878 }
879
880 BlockNode getNode(const BlockT *BB) const { return Nodes.lookup(BB).first; }
881
882 const BlockT *getBlock(const BlockNode &Node) const {
883 assert(Node.Index < RPOT.size());
884 return RPOT[Node.Index];
885 }
886
887 /// Run (and save) a post-order traversal.
888 ///
889 /// Saves a reverse post-order traversal of all the nodes in \a F.
890 void initializeRPOT();
891
892 /// Initialize loop data.
893 ///
894 /// Build up \a Loops using \a LoopInfo. \a LoopInfo gives us a mapping from
895 /// each block to the deepest loop it's in, but we need the inverse. For each
896 /// loop, we store in reverse post-order its "immediate" members, defined as
897 /// the header, the headers of immediate sub-loops, and all other blocks in
898 /// the loop that are not in sub-loops.
899 void initializeLoops();
900
901 /// Propagate to a block's successors.
902 ///
903 /// In the context of distributing mass through \c OuterLoop, divide the mass
904 /// currently assigned to \c Node between its successors.
905 ///
906 /// \return \c true unless there's an irreducible backedge.
907 bool propagateMassToSuccessors(LoopData *OuterLoop, const BlockNode &Node);
908
909 /// Compute mass in a particular loop.
910 ///
911 /// Assign mass to \c Loop's header, and then for each block in \c Loop in
912 /// reverse post-order, distribute mass to its successors. Only visits nodes
913 /// that have not been packaged into sub-loops.
914 ///
915 /// \pre \a computeMassInLoop() has been called for each subloop of \c Loop.
916 /// \return \c true unless there's an irreducible backedge.
917 bool computeMassInLoop(LoopData &Loop);
918
919 /// Try to compute mass in the top-level function.
920 ///
921 /// Assign mass to the entry block, and then for each block in reverse
922 /// post-order, distribute mass to its successors. Skips nodes that have
923 /// been packaged into loops.
924 ///
925 /// \pre \a computeMassInLoops() has been called.
926 /// \return \c true unless there's an irreducible backedge.
927 bool tryToComputeMassInFunction();
928
929 /// Compute mass in (and package up) irreducible SCCs.
930 ///
931 /// Find the irreducible SCCs in \c OuterLoop, add them to \a Loops (in front
932 /// of \c Insert), and call \a computeMassInLoop() on each of them.
933 ///
934 /// If \c OuterLoop is \c nullptr, it refers to the top-level function.
935 ///
936 /// \pre \a computeMassInLoop() has been called for each subloop of \c
937 /// OuterLoop.
938 /// \pre \c Insert points at the last loop successfully processed by \a
939 /// computeMassInLoop().
940 /// \pre \c OuterLoop has irreducible SCCs.
941 void computeIrreducibleMass(LoopData *OuterLoop,
942 std::list<LoopData>::iterator Insert);
943
944 /// Compute mass in all loops.
945 ///
946 /// For each loop bottom-up, call \a computeMassInLoop().
947 ///
948 /// \a computeMassInLoop() aborts (and returns \c false) on loops that
949 /// contain a irreducible sub-SCCs. Use \a computeIrreducibleMass() and then
950 /// re-enter \a computeMassInLoop().
951 ///
952 /// \post \a computeMassInLoop() has returned \c true for every loop.
953 void computeMassInLoops();
954
955 /// Compute mass in the top-level function.
956 ///
957 /// Uses \a tryToComputeMassInFunction() and \a computeIrreducibleMass() to
958 /// compute mass in the top-level function.
959 ///
960 /// \post \a tryToComputeMassInFunction() has returned \c true.
961 void computeMassInFunction();
962
963 std::string getBlockName(const BlockNode &Node) const override {
964 return bfi_detail::getBlockName(getBlock(Node));
965 }
966
967 /// The current implementation for computing relative block frequencies does
968 /// not handle correctly control-flow graphs containing irreducible loops. To
969 /// resolve the problem, we apply a post-processing step, which iteratively
970 /// updates block frequencies based on the frequencies of their predesessors.
971 /// This corresponds to finding the stationary point of the Markov chain by
972 /// an iterative method aka "PageRank computation".
973 /// The algorithm takes at most O(|E| * IterativeBFIMaxIterations) steps but
974 /// typically converges faster.
975 ///
976 /// Decide whether we want to apply iterative inference for a given function.
977 bool needIterativeInference() const;
978
979 /// Apply an iterative post-processing to infer correct counts for irr loops.
980 void applyIterativeInference();
981
982 using ProbMatrixType = std::vector<std::vector<std::pair<size_t, Scaled64>>>;
983
984 /// Run iterative inference for a probability matrix and initial frequencies.
985 void iterativeInference(const ProbMatrixType &ProbMatrix,
986 std::vector<Scaled64> &Freq) const;
987
988 /// Find all blocks to apply inference on, that is, reachable from the entry
989 /// and backward reachable from exists along edges with positive probability.
990 void findReachableBlocks(std::vector<const BlockT *> &Blocks) const;
991
992 /// Build a matrix of probabilities with transitions (edges) between the
993 /// blocks: ProbMatrix[I] holds pairs (J, P), where Pr[J -> I | J] = P
994 void initTransitionProbabilities(
995 const std::vector<const BlockT *> &Blocks,
996 const DenseMap<const BlockT *, size_t> &BlockIndex,
997 ProbMatrixType &ProbMatrix) const;
998
999#ifndef NDEBUG
1000 /// Compute the discrepancy between current block frequencies and the
1001 /// probability matrix.
1002 Scaled64 discrepancy(const ProbMatrixType &ProbMatrix,
1003 const std::vector<Scaled64> &Freq) const;
1004#endif
1005
1006public:
1008
1009 const FunctionT *getFunction() const { return F; }
1010
1011 void calculate(const FunctionT &F, const BranchProbabilityInfoT &BPI,
1012 const LoopInfoT &LI);
1013
1015
1016 BlockFrequency getBlockFreq(const BlockT *BB) const {
1017 return BlockFrequencyInfoImplBase::getBlockFreq(getNode(BB));
1018 }
1019
1020 std::optional<uint64_t>
1021 getBlockProfileCount(const Function &F, const BlockT *BB,
1022 bool AllowSynthetic = false) const {
1024 AllowSynthetic);
1025 }
1026
1027 std::optional<uint64_t>
1029 bool AllowSynthetic = false) const {
1031 AllowSynthetic);
1032 }
1033
1034 bool isIrrLoopHeader(const BlockT *BB) {
1036 }
1037
1038 void setBlockFreq(const BlockT *BB, BlockFrequency Freq);
1039
1040 void forgetBlock(const BlockT *BB) {
1041 // We don't erase corresponding items from `Freqs`, `RPOT` and other to
1042 // avoid invalidating indices. Doing so would have saved some memory, but
1043 // it's not worth it.
1044 Nodes.erase(BB);
1045 }
1046
1047 Scaled64 getFloatingBlockFreq(const BlockT *BB) const {
1049 }
1050
1051 const BranchProbabilityInfoT &getBPI() const { return *BPI; }
1052
1053 /// Print the frequencies for the current function.
1054 ///
1055 /// Prints the frequencies for the blocks in the current function.
1056 ///
1057 /// Blocks are printed in the natural iteration order of the function, rather
1058 /// than reverse post-order. This provides two advantages: writing -analyze
1059 /// tests is easier (since blocks come out in source order), and even
1060 /// unreachable blocks are printed.
1061 ///
1062 /// \a BlockFrequencyInfoImplBase::print() only knows reverse post-order, so
1063 /// we need to override it here.
1064 raw_ostream &print(raw_ostream &OS) const override;
1065
1067
1069};
1070
1071namespace bfi_detail {
1072
1073template <class BFIImplT>
1074class BFICallbackVH<BasicBlock, BFIImplT> : public CallbackVH {
1075 BFIImplT *BFIImpl;
1076
1077public:
1078 BFICallbackVH() = default;
1079
1080 BFICallbackVH(const BasicBlock *BB, BFIImplT *BFIImpl)
1081 : CallbackVH(BB), BFIImpl(BFIImpl) {}
1082
1083 virtual ~BFICallbackVH() = default;
1084
1085 void deleted() override {
1086 BFIImpl->forgetBlock(cast<BasicBlock>(getValPtr()));
1087 }
1088};
1089
1090/// Dummy implementation since MachineBasicBlocks aren't Values, so ValueHandles
1091/// don't apply to them.
1092template <class BFIImplT>
1094public:
1095 BFICallbackVH() = default;
1096 BFICallbackVH(const MachineBasicBlock *, BFIImplT *) {}
1097};
1098
1099} // end namespace bfi_detail
1100
1101template <class BT>
1103 const BranchProbabilityInfoT &BPI,
1104 const LoopInfoT &LI) {
1105 // Save the parameters.
1106 this->BPI = &BPI;
1107 this->LI = &LI;
1108 this->F = &F;
1109
1110 // Clean up left-over data structures.
1112 RPOT.clear();
1113 Nodes.clear();
1114
1115 // Initialize.
1116 LLVM_DEBUG(dbgs() << "\nblock-frequency: " << F.getName()
1117 << "\n================="
1118 << std::string(F.getName().size(), '=') << "\n");
1119 initializeRPOT();
1120 initializeLoops();
1121
1122 // Visit loops in post-order to find the local mass distribution, and then do
1123 // the full function.
1124 computeMassInLoops();
1125 computeMassInFunction();
1126 unwrapLoops();
1127 // Apply a post-processing step improving computed frequencies for functions
1128 // with irreducible loops.
1129 if (needIterativeInference())
1130 applyIterativeInference();
1131 finalizeMetrics();
1132
1134 // To detect BFI queries for unknown blocks, add entries for unreachable
1135 // blocks, if any. This is to distinguish between known/existing unreachable
1136 // blocks and unknown blocks.
1137 for (const BlockT &BB : F)
1138 if (!Nodes.count(&BB))
1139 setBlockFreq(&BB, BlockFrequency());
1140 }
1141}
1142
1143template <class BT>
1145 BlockFrequency Freq) {
1146 if (Nodes.count(BB))
1148 else {
1149 // If BB is a newly added block after BFI is done, we need to create a new
1150 // BlockNode for it assigned with a new index. The index can be determined
1151 // by the size of Freqs.
1152 BlockNode NewNode(Freqs.size());
1153 Nodes[BB] = {NewNode, BFICallbackVH(BB, this)};
1154 Freqs.emplace_back();
1156 }
1157}
1158
1159template <class BT> void BlockFrequencyInfoImpl<BT>::initializeRPOT() {
1160 const BlockT *Entry = &F->front();
1161 RPOT.reserve(F->size());
1162 std::copy(po_begin(Entry), po_end(Entry), std::back_inserter(RPOT));
1163 std::reverse(RPOT.begin(), RPOT.end());
1164
1165 assert(RPOT.size() - 1 <= BlockNode::getMaxIndex() &&
1166 "More nodes in function than Block Frequency Info supports");
1167
1168 LLVM_DEBUG(dbgs() << "reverse-post-order-traversal\n");
1169 for (rpot_iterator I = rpot_begin(), E = rpot_end(); I != E; ++I) {
1170 BlockNode Node = getNode(I);
1171 LLVM_DEBUG(dbgs() << " - " << getIndex(I) << ": " << getBlockName(Node)
1172 << "\n");
1173 Nodes[*I] = {Node, BFICallbackVH(*I, this)};
1174 }
1175
1176 Working.reserve(RPOT.size());
1177 for (size_t Index = 0; Index < RPOT.size(); ++Index)
1178 Working.emplace_back(Index);
1179 Freqs.resize(RPOT.size());
1180}
1181
1182template <class BT> void BlockFrequencyInfoImpl<BT>::initializeLoops() {
1183 LLVM_DEBUG(dbgs() << "loop-detection\n");
1184 if (LI->empty())
1185 return;
1186
1187 // Visit loops top down and assign them an index.
1188 std::deque<std::pair<const LoopT *, LoopData *>> Q;
1189 for (const LoopT *L : *LI)
1190 Q.emplace_back(L, nullptr);
1191 while (!Q.empty()) {
1192 const LoopT *Loop = Q.front().first;
1193 LoopData *Parent = Q.front().second;
1194 Q.pop_front();
1195
1196 BlockNode Header = getNode(Loop->getHeader());
1197 assert(Header.isValid());
1198
1199 Loops.emplace_back(Parent, Header);
1200 Working[Header.Index].Loop = &Loops.back();
1201 LLVM_DEBUG(dbgs() << " - loop = " << getBlockName(Header) << "\n");
1202
1203 for (const LoopT *L : *Loop)
1204 Q.emplace_back(L, &Loops.back());
1205 }
1206
1207 // Visit nodes in reverse post-order and add them to their deepest containing
1208 // loop.
1209 for (size_t Index = 0; Index < RPOT.size(); ++Index) {
1210 // Loop headers have already been mostly mapped.
1211 if (Working[Index].isLoopHeader()) {
1212 LoopData *ContainingLoop = Working[Index].getContainingLoop();
1213 if (ContainingLoop)
1214 ContainingLoop->Nodes.push_back(Index);
1215 continue;
1216 }
1217
1218 const LoopT *Loop = LI->getLoopFor(RPOT[Index]);
1219 if (!Loop)
1220 continue;
1221
1222 // Add this node to its containing loop's member list.
1223 BlockNode Header = getNode(Loop->getHeader());
1224 assert(Header.isValid());
1225 const auto &HeaderData = Working[Header.Index];
1226 assert(HeaderData.isLoopHeader());
1227
1228 Working[Index].Loop = HeaderData.Loop;
1229 HeaderData.Loop->Nodes.push_back(Index);
1230 LLVM_DEBUG(dbgs() << " - loop = " << getBlockName(Header)
1231 << ": member = " << getBlockName(Index) << "\n");
1232 }
1233}
1234
1235template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInLoops() {
1236 // Visit loops with the deepest first, and the top-level loops last.
1237 for (auto L = Loops.rbegin(), E = Loops.rend(); L != E; ++L) {
1238 if (computeMassInLoop(*L))
1239 continue;
1240 auto Next = std::next(L);
1241 computeIrreducibleMass(&*L, L.base());
1242 L = std::prev(Next);
1243 if (computeMassInLoop(*L))
1244 continue;
1245 llvm_unreachable("unhandled irreducible control flow");
1246 }
1247}
1248
1249template <class BT>
1250bool BlockFrequencyInfoImpl<BT>::computeMassInLoop(LoopData &Loop) {
1251 // Compute mass in loop.
1252 LLVM_DEBUG(dbgs() << "compute-mass-in-loop: " << getLoopName(Loop) << "\n");
1253
1254 if (Loop.isIrreducible()) {
1255 LLVM_DEBUG(dbgs() << "isIrreducible = true\n");
1256 Distribution Dist;
1257 unsigned NumHeadersWithWeight = 0;
1258 std::optional<uint64_t> MinHeaderWeight;
1259 DenseSet<uint32_t> HeadersWithoutWeight;
1260 HeadersWithoutWeight.reserve(Loop.NumHeaders);
1261 for (uint32_t H = 0; H < Loop.NumHeaders; ++H) {
1262 auto &HeaderNode = Loop.Nodes[H];
1263 const BlockT *Block = getBlock(HeaderNode);
1264 IsIrrLoopHeader.set(Loop.Nodes[H].Index);
1265 std::optional<uint64_t> HeaderWeight = Block->getIrrLoopHeaderWeight();
1266 if (!HeaderWeight) {
1267 LLVM_DEBUG(dbgs() << "Missing irr loop header metadata on "
1268 << getBlockName(HeaderNode) << "\n");
1269 HeadersWithoutWeight.insert(H);
1270 continue;
1271 }
1272 LLVM_DEBUG(dbgs() << getBlockName(HeaderNode)
1273 << " has irr loop header weight " << *HeaderWeight
1274 << "\n");
1275 NumHeadersWithWeight++;
1276 uint64_t HeaderWeightValue = *HeaderWeight;
1277 if (!MinHeaderWeight || HeaderWeightValue < MinHeaderWeight)
1278 MinHeaderWeight = HeaderWeightValue;
1279 if (HeaderWeightValue) {
1280 Dist.addLocal(HeaderNode, HeaderWeightValue);
1281 }
1282 }
1283 // As a heuristic, if some headers don't have a weight, give them the
1284 // minimum weight seen (not to disrupt the existing trends too much by
1285 // using a weight that's in the general range of the other headers' weights,
1286 // and the minimum seems to perform better than the average.)
1287 // FIXME: better update in the passes that drop the header weight.
1288 // If no headers have a weight, give them even weight (use weight 1).
1289 if (!MinHeaderWeight)
1290 MinHeaderWeight = 1;
1291 for (uint32_t H : HeadersWithoutWeight) {
1292 auto &HeaderNode = Loop.Nodes[H];
1293 assert(!getBlock(HeaderNode)->getIrrLoopHeaderWeight() &&
1294 "Shouldn't have a weight metadata");
1295 uint64_t MinWeight = *MinHeaderWeight;
1296 LLVM_DEBUG(dbgs() << "Giving weight " << MinWeight << " to "
1297 << getBlockName(HeaderNode) << "\n");
1298 if (MinWeight)
1299 Dist.addLocal(HeaderNode, MinWeight);
1300 }
1301 distributeIrrLoopHeaderMass(Dist);
1302 for (const BlockNode &M : Loop.Nodes)
1303 if (!propagateMassToSuccessors(&Loop, M))
1304 llvm_unreachable("unhandled irreducible control flow");
1305 if (NumHeadersWithWeight == 0)
1306 // No headers have a metadata. Adjust header mass.
1307 adjustLoopHeaderMass(Loop);
1308 } else {
1309 Working[Loop.getHeader().Index].getMass() = BlockMass::getFull();
1310 if (!propagateMassToSuccessors(&Loop, Loop.getHeader()))
1311 llvm_unreachable("irreducible control flow to loop header!?");
1312 for (const BlockNode &M : Loop.members())
1313 if (!propagateMassToSuccessors(&Loop, M))
1314 // Irreducible backedge.
1315 return false;
1316 }
1317
1318 computeLoopScale(Loop);
1319 packageLoop(Loop);
1320 return true;
1321}
1322
1323template <class BT>
1324bool BlockFrequencyInfoImpl<BT>::tryToComputeMassInFunction() {
1325 // Compute mass in function.
1326 LLVM_DEBUG(dbgs() << "compute-mass-in-function\n");
1327 assert(!Working.empty() && "no blocks in function");
1328 assert(!Working[0].isLoopHeader() && "entry block is a loop header");
1329
1330 Working[0].getMass() = BlockMass::getFull();
1331 for (rpot_iterator I = rpot_begin(), IE = rpot_end(); I != IE; ++I) {
1332 // Check for nodes that have been packaged.
1333 BlockNode Node = getNode(I);
1334 if (Working[Node.Index].isPackaged())
1335 continue;
1336
1337 if (!propagateMassToSuccessors(nullptr, Node))
1338 return false;
1339 }
1340 return true;
1341}
1342
1343template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInFunction() {
1344 if (tryToComputeMassInFunction())
1345 return;
1346 computeIrreducibleMass(nullptr, Loops.begin());
1347 if (tryToComputeMassInFunction())
1348 return;
1349 llvm_unreachable("unhandled irreducible control flow");
1350}
1351
1352template <class BT>
1353bool BlockFrequencyInfoImpl<BT>::needIterativeInference() const {
1355 return false;
1356 if (!F->getFunction().hasProfileData())
1357 return false;
1358 // Apply iterative inference only if the function contains irreducible loops;
1359 // otherwise, computed block frequencies are reasonably correct.
1360 for (auto L = Loops.rbegin(), E = Loops.rend(); L != E; ++L) {
1361 if (L->isIrreducible())
1362 return true;
1363 }
1364 return false;
1365}
1366
1367template <class BT> void BlockFrequencyInfoImpl<BT>::applyIterativeInference() {
1368 // Extract blocks for processing: a block is considered for inference iff it
1369 // can be reached from the entry by edges with a positive probability.
1370 // Non-processed blocks are assigned with the zero frequency and are ignored
1371 // in the computation
1372 std::vector<const BlockT *> ReachableBlocks;
1373 findReachableBlocks(ReachableBlocks);
1374 if (ReachableBlocks.empty())
1375 return;
1376
1377 // The map is used to index successors/predecessors of reachable blocks in
1378 // the ReachableBlocks vector
1379 DenseMap<const BlockT *, size_t> BlockIndex;
1380 // Extract initial frequencies for the reachable blocks
1381 auto Freq = std::vector<Scaled64>(ReachableBlocks.size());
1382 Scaled64 SumFreq;
1383 for (size_t I = 0; I < ReachableBlocks.size(); I++) {
1384 const BlockT *BB = ReachableBlocks[I];
1385 BlockIndex[BB] = I;
1386 Freq[I] = getFloatingBlockFreq(BB);
1387 SumFreq += Freq[I];
1388 }
1389 assert(!SumFreq.isZero() && "empty initial block frequencies");
1390
1391 LLVM_DEBUG(dbgs() << "Applying iterative inference for " << F->getName()
1392 << " with " << ReachableBlocks.size() << " blocks\n");
1393
1394 // Normalizing frequencies so they sum up to 1.0
1395 for (auto &Value : Freq) {
1396 Value /= SumFreq;
1397 }
1398
1399 // Setting up edge probabilities using sparse matrix representation:
1400 // ProbMatrix[I] holds a vector of pairs (J, P) where Pr[J -> I | J] = P
1401 ProbMatrixType ProbMatrix;
1402 initTransitionProbabilities(ReachableBlocks, BlockIndex, ProbMatrix);
1403
1404 // Run the propagation
1405 iterativeInference(ProbMatrix, Freq);
1406
1407 // Assign computed frequency values
1408 for (const BlockT &BB : *F) {
1409 auto Node = getNode(&BB);
1410 if (!Node.isValid())
1411 continue;
1412 if (BlockIndex.count(&BB)) {
1413 Freqs[Node.Index].Scaled = Freq[BlockIndex[&BB]];
1414 } else {
1415 Freqs[Node.Index].Scaled = Scaled64::getZero();
1416 }
1417 }
1418}
1419
1420template <class BT>
1421void BlockFrequencyInfoImpl<BT>::iterativeInference(
1422 const ProbMatrixType &ProbMatrix, std::vector<Scaled64> &Freq) const {
1424 "incorrectly specified precision");
1425 // Convert double precision to Scaled64
1426 const auto Precision =
1427 Scaled64::getInverse(static_cast<uint64_t>(1.0 / IterativeBFIPrecision));
1428 const size_t MaxIterations = IterativeBFIMaxIterationsPerBlock * Freq.size();
1429
1430#ifndef NDEBUG
1431 LLVM_DEBUG(dbgs() << " Initial discrepancy = "
1432 << discrepancy(ProbMatrix, Freq).toString() << "\n");
1433#endif
1434
1435 // Successors[I] holds unique sucessors of the I-th block
1436 auto Successors = std::vector<std::vector<size_t>>(Freq.size());
1437 for (size_t I = 0; I < Freq.size(); I++) {
1438 for (const auto &Jump : ProbMatrix[I]) {
1439 Successors[Jump.first].push_back(I);
1440 }
1441 }
1442
1443 // To speedup computation, we maintain a set of "active" blocks whose
1444 // frequencies need to be updated based on the incoming edges.
1445 // The set is dynamic and changes after every update. Initially all blocks
1446 // with a positive frequency are active
1447 auto IsActive = BitVector(Freq.size(), false);
1448 std::queue<size_t> ActiveSet;
1449 for (size_t I = 0; I < Freq.size(); I++) {
1450 if (Freq[I] > 0) {
1451 ActiveSet.push(I);
1452 IsActive[I] = true;
1453 }
1454 }
1455
1456 // Iterate over the blocks propagating frequencies
1457 size_t It = 0;
1458 while (It++ < MaxIterations && !ActiveSet.empty()) {
1459 size_t I = ActiveSet.front();
1460 ActiveSet.pop();
1461 IsActive[I] = false;
1462
1463 // Compute a new frequency for the block: NewFreq := Freq \times ProbMatrix.
1464 // A special care is taken for self-edges that needs to be scaled by
1465 // (1.0 - SelfProb), where SelfProb is the sum of probabilities on the edges
1466 Scaled64 NewFreq;
1467 Scaled64 OneMinusSelfProb = Scaled64::getOne();
1468 for (const auto &Jump : ProbMatrix[I]) {
1469 if (Jump.first == I) {
1470 OneMinusSelfProb -= Jump.second;
1471 } else {
1472 NewFreq += Freq[Jump.first] * Jump.second;
1473 }
1474 }
1475 if (OneMinusSelfProb != Scaled64::getOne())
1476 NewFreq /= OneMinusSelfProb;
1477
1478 // If the block's frequency has changed enough, then
1479 // make sure the block and its successors are in the active set
1480 auto Change = Freq[I] >= NewFreq ? Freq[I] - NewFreq : NewFreq - Freq[I];
1481 if (Change > Precision) {
1482 ActiveSet.push(I);
1483 IsActive[I] = true;
1484 for (size_t Succ : Successors[I]) {
1485 if (!IsActive[Succ]) {
1486 ActiveSet.push(Succ);
1487 IsActive[Succ] = true;
1488 }
1489 }
1490 }
1491
1492 // Update the frequency for the block
1493 Freq[I] = NewFreq;
1494 }
1495
1496 LLVM_DEBUG(dbgs() << " Completed " << It << " inference iterations"
1497 << format(" (%0.0f per block)", double(It) / Freq.size())
1498 << "\n");
1499#ifndef NDEBUG
1500 LLVM_DEBUG(dbgs() << " Final discrepancy = "
1501 << discrepancy(ProbMatrix, Freq).toString() << "\n");
1502#endif
1503}
1504
1505template <class BT>
1506void BlockFrequencyInfoImpl<BT>::findReachableBlocks(
1507 std::vector<const BlockT *> &Blocks) const {
1508 // Find all blocks to apply inference on, that is, reachable from the entry
1509 // along edges with non-zero probablities
1510 std::queue<const BlockT *> Queue;
1511 SmallPtrSet<const BlockT *, 8> Reachable;
1512 const BlockT *Entry = &F->front();
1513 Queue.push(Entry);
1514 Reachable.insert(Entry);
1515 while (!Queue.empty()) {
1516 const BlockT *SrcBB = Queue.front();
1517 Queue.pop();
1518 for (const BlockT *DstBB : children<const BlockT *>(SrcBB)) {
1519 auto EP = BPI->getEdgeProbability(SrcBB, DstBB);
1520 if (EP.isZero())
1521 continue;
1522 if (Reachable.insert(DstBB).second)
1523 Queue.push(DstBB);
1524 }
1525 }
1526
1527 // Find all blocks to apply inference on, that is, backward reachable from
1528 // the entry along (backward) edges with non-zero probablities
1529 SmallPtrSet<const BlockT *, 8> InverseReachable;
1530 for (const BlockT &BB : *F) {
1531 // An exit block is a block without any successors
1532 bool HasSucc = !llvm::children<const BlockT *>(&BB).empty();
1533 if (!HasSucc && Reachable.count(&BB)) {
1534 Queue.push(&BB);
1535 InverseReachable.insert(&BB);
1536 }
1537 }
1538 while (!Queue.empty()) {
1539 const BlockT *SrcBB = Queue.front();
1540 Queue.pop();
1541 for (const BlockT *DstBB : inverse_children<const BlockT *>(SrcBB)) {
1542 auto EP = BPI->getEdgeProbability(DstBB, SrcBB);
1543 if (EP.isZero())
1544 continue;
1545 if (InverseReachable.insert(DstBB).second)
1546 Queue.push(DstBB);
1547 }
1548 }
1549
1550 // Collect the result
1551 Blocks.reserve(F->size());
1552 for (const BlockT &BB : *F) {
1553 if (Reachable.count(&BB) && InverseReachable.count(&BB)) {
1554 Blocks.push_back(&BB);
1555 }
1556 }
1557}
1558
1559template <class BT>
1560void BlockFrequencyInfoImpl<BT>::initTransitionProbabilities(
1561 const std::vector<const BlockT *> &Blocks,
1562 const DenseMap<const BlockT *, size_t> &BlockIndex,
1563 ProbMatrixType &ProbMatrix) const {
1564 const size_t NumBlocks = Blocks.size();
1565 auto Succs = std::vector<std::vector<std::pair<size_t, Scaled64>>>(NumBlocks);
1566 auto SumProb = std::vector<Scaled64>(NumBlocks);
1567
1568 // Find unique successors and corresponding probabilities for every block
1569 for (size_t Src = 0; Src < NumBlocks; Src++) {
1570 const BlockT *BB = Blocks[Src];
1571 SmallPtrSet<const BlockT *, 2> UniqueSuccs;
1572 for (const auto SI : children<const BlockT *>(BB)) {
1573 // Ignore cold blocks
1574 if (!BlockIndex.contains(SI))
1575 continue;
1576 // Ignore parallel edges between BB and SI blocks
1577 if (!UniqueSuccs.insert(SI).second)
1578 continue;
1579 // Ignore jumps with zero probability
1580 auto EP = BPI->getEdgeProbability(BB, SI);
1581 if (EP.isZero())
1582 continue;
1583
1584 auto EdgeProb =
1585 Scaled64::getFraction(EP.getNumerator(), EP.getDenominator());
1586 size_t Dst = BlockIndex.find(SI)->second;
1587 Succs[Src].push_back(std::make_pair(Dst, EdgeProb));
1588 SumProb[Src] += EdgeProb;
1589 }
1590 }
1591
1592 // Add transitions for every jump with positive branch probability
1593 ProbMatrix = ProbMatrixType(NumBlocks);
1594 for (size_t Src = 0; Src < NumBlocks; Src++) {
1595 // Ignore blocks w/o successors
1596 if (Succs[Src].empty())
1597 continue;
1598
1599 assert(!SumProb[Src].isZero() && "Zero sum probability of non-exit block");
1600 for (auto &Jump : Succs[Src]) {
1601 size_t Dst = Jump.first;
1602 Scaled64 Prob = Jump.second;
1603 ProbMatrix[Dst].push_back(std::make_pair(Src, Prob / SumProb[Src]));
1604 }
1605 }
1606
1607 // Add transitions from sinks to the source
1608 size_t EntryIdx = BlockIndex.find(&F->front())->second;
1609 for (size_t Src = 0; Src < NumBlocks; Src++) {
1610 if (Succs[Src].empty()) {
1611 ProbMatrix[EntryIdx].push_back(std::make_pair(Src, Scaled64::getOne()));
1612 }
1613 }
1614}
1615
1616#ifndef NDEBUG
1617template <class BT>
1618BlockFrequencyInfoImplBase::Scaled64 BlockFrequencyInfoImpl<BT>::discrepancy(
1619 const ProbMatrixType &ProbMatrix, const std::vector<Scaled64> &Freq) const {
1620 assert(Freq[0] > 0 && "Incorrectly computed frequency of the entry block");
1621 Scaled64 Discrepancy;
1622 for (size_t I = 0; I < ProbMatrix.size(); I++) {
1623 Scaled64 Sum;
1624 for (const auto &Jump : ProbMatrix[I]) {
1625 Sum += Freq[Jump.first] * Jump.second;
1626 }
1627 Discrepancy += Freq[I] >= Sum ? Freq[I] - Sum : Sum - Freq[I];
1628 }
1629 // Normalizing by the frequency of the entry block
1630 return Discrepancy / Freq[0];
1631}
1632#endif
1633
1634/// \note This should be a lambda, but that crashes GCC 4.7.
1635namespace bfi_detail {
1636
1637template <class BT> struct BlockEdgesAdder {
1638 using BlockT = BT;
1641
1643
1645 : BFI(BFI) {}
1646
1648 const LoopData *OuterLoop) {
1649 const BlockT *BB = BFI.RPOT[Irr.Node.Index];
1650 for (const auto *Succ : children<const BlockT *>(BB))
1651 G.addEdge(Irr, BFI.getNode(Succ), OuterLoop);
1652 }
1653};
1654
1655} // end namespace bfi_detail
1656
1657template <class BT>
1658void BlockFrequencyInfoImpl<BT>::computeIrreducibleMass(
1659 LoopData *OuterLoop, std::list<LoopData>::iterator Insert) {
1660 LLVM_DEBUG(dbgs() << "analyze-irreducible-in-";
1661 if (OuterLoop) dbgs()
1662 << "loop: " << getLoopName(*OuterLoop) << "\n";
1663 else dbgs() << "function\n");
1664
1665 using namespace bfi_detail;
1666
1667 // Ideally, addBlockEdges() would be declared here as a lambda, but that
1668 // crashes GCC 4.7.
1669 BlockEdgesAdder<BT> addBlockEdges(*this);
1670 IrreducibleGraph G(*this, OuterLoop, addBlockEdges);
1671
1672 for (auto &L : analyzeIrreducible(G, OuterLoop, Insert))
1673 computeMassInLoop(L);
1674
1675 if (!OuterLoop)
1676 return;
1677 updateLoopWithIrreducible(*OuterLoop);
1678}
1679
1680// A helper function that converts a branch probability into weight.
1682 return Prob.getNumerator();
1683}
1684
1685template <class BT>
1686bool
1687BlockFrequencyInfoImpl<BT>::propagateMassToSuccessors(LoopData *OuterLoop,
1688 const BlockNode &Node) {
1689 LLVM_DEBUG(dbgs() << " - node: " << getBlockName(Node) << "\n");
1690 // Calculate probability for successors.
1691 Distribution Dist;
1692 if (auto *Loop = Working[Node.Index].getPackagedLoop()) {
1693 assert(Loop != OuterLoop && "Cannot propagate mass in a packaged loop");
1694 if (!addLoopSuccessorsToDist(OuterLoop, *Loop, Dist))
1695 // Irreducible backedge.
1696 return false;
1697 } else {
1698 const BlockT *BB = getBlock(Node);
1699 for (auto SI = GraphTraits<const BlockT *>::child_begin(BB),
1700 SE = GraphTraits<const BlockT *>::child_end(BB);
1701 SI != SE; ++SI)
1702 if (!addToDist(
1703 Dist, OuterLoop, Node, getNode(*SI),
1704 getWeightFromBranchProb(BPI->getEdgeProbability(BB, SI))))
1705 // Irreducible backedge.
1706 return false;
1707 }
1708
1709 // Distribute mass to successors, saving exit and backedge data in the
1710 // loop header.
1711 distributeMass(Node, OuterLoop, Dist);
1712 return true;
1713}
1714
1715template <class BT>
1717 if (!F)
1718 return OS;
1719 OS << "block-frequency-info: " << F->getName() << "\n";
1720 for (const BlockT &BB : *F) {
1721 OS << " - " << bfi_detail::getBlockName(&BB) << ": float = ";
1722 getFloatingBlockFreq(&BB).print(OS, 5)
1723 << ", int = " << getBlockFreq(&BB).getFrequency();
1724 if (std::optional<uint64_t> ProfileCount =
1726 F->getFunction(), getNode(&BB)))
1727 OS << ", count = " << *ProfileCount;
1728 if (std::optional<uint64_t> IrrLoopHeaderWeight =
1729 BB.getIrrLoopHeaderWeight())
1730 OS << ", irr_loop_header_weight = " << *IrrLoopHeaderWeight;
1731 OS << "\n";
1732 }
1733
1734 // Add an extra newline for readability.
1735 OS << "\n";
1736 return OS;
1737}
1738
1739template <class BT>
1742 bool Match = true;
1745 for (auto &Entry : Nodes) {
1746 const BlockT *BB = Entry.first;
1747 if (BB) {
1748 ValidNodes[BB] = Entry.second.first;
1749 }
1750 }
1751 for (auto &Entry : Other.Nodes) {
1752 const BlockT *BB = Entry.first;
1753 if (BB) {
1754 OtherValidNodes[BB] = Entry.second.first;
1755 }
1756 }
1757 unsigned NumValidNodes = ValidNodes.size();
1758 unsigned NumOtherValidNodes = OtherValidNodes.size();
1759 if (NumValidNodes != NumOtherValidNodes) {
1760 Match = false;
1761 dbgs() << "Number of blocks mismatch: " << NumValidNodes << " vs "
1762 << NumOtherValidNodes << "\n";
1763 } else {
1764 for (auto &Entry : ValidNodes) {
1765 const BlockT *BB = Entry.first;
1766 BlockNode Node = Entry.second;
1767 if (OtherValidNodes.count(BB)) {
1768 BlockNode OtherNode = OtherValidNodes[BB];
1769 const auto &Freq = Freqs[Node.Index];
1770 const auto &OtherFreq = Other.Freqs[OtherNode.Index];
1771 if (Freq.Integer != OtherFreq.Integer) {
1772 Match = false;
1773 dbgs() << "Freq mismatch: " << bfi_detail::getBlockName(BB) << " "
1774 << Freq.Integer << " vs " << OtherFreq.Integer << "\n";
1775 }
1776 } else {
1777 Match = false;
1778 dbgs() << "Block " << bfi_detail::getBlockName(BB) << " index "
1779 << Node.Index << " does not exist in Other.\n";
1780 }
1781 }
1782 // If there's a valid node in OtherValidNodes that's not in ValidNodes,
1783 // either the above num check or the check on OtherValidNodes will fail.
1784 }
1785 if (!Match) {
1786 dbgs() << "This\n";
1787 print(dbgs());
1788 dbgs() << "Other\n";
1789 Other.print(dbgs());
1790 }
1791 assert(Match && "BFI mismatch");
1792}
1793
1794// Graph trait base class for block frequency information graph
1795// viewer.
1796
1798
1799template <class BlockFrequencyInfoT, class BranchProbabilityInfoT>
1802 using NodeRef = typename GTraits::NodeRef;
1803 using EdgeIter = typename GTraits::ChildIteratorType;
1804 using NodeIter = typename GTraits::nodes_iterator;
1805
1807
1808 explicit BFIDOTGraphTraitsBase(bool isSimple = false)
1810
1811 static StringRef getGraphName(const BlockFrequencyInfoT *G) {
1812 return G->getFunction()->getName();
1813 }
1814
1815 std::string getNodeAttributes(NodeRef Node, const BlockFrequencyInfoT *Graph,
1816 unsigned HotPercentThreshold = 0) {
1817 std::string Result;
1818 if (!HotPercentThreshold)
1819 return Result;
1820
1821 // Compute MaxFrequency on the fly:
1822 if (!MaxFrequency) {
1823 for (NodeIter I = GTraits::nodes_begin(Graph),
1824 E = GTraits::nodes_end(Graph);
1825 I != E; ++I) {
1826 NodeRef N = *I;
1827 MaxFrequency =
1828 std::max(MaxFrequency, Graph->getBlockFreq(N).getFrequency());
1829 }
1830 }
1831 BlockFrequency Freq = Graph->getBlockFreq(Node);
1832 BlockFrequency HotFreq =
1834 BranchProbability::getBranchProbability(HotPercentThreshold, 100));
1835
1836 if (Freq < HotFreq)
1837 return Result;
1838
1839 raw_string_ostream OS(Result);
1840 OS << "color=\"red\"";
1841 OS.flush();
1842 return Result;
1843 }
1844
1845 std::string getNodeLabel(NodeRef Node, const BlockFrequencyInfoT *Graph,
1846 GVDAGType GType, int layout_order = -1) {
1847 std::string Result;
1848 raw_string_ostream OS(Result);
1849
1850 if (layout_order != -1)
1851 OS << Node->getName() << "[" << layout_order << "] : ";
1852 else
1853 OS << Node->getName() << " : ";
1854 switch (GType) {
1855 case GVDT_Fraction:
1856 OS << printBlockFreq(*Graph, *Node);
1857 break;
1858 case GVDT_Integer:
1859 OS << Graph->getBlockFreq(Node).getFrequency();
1860 break;
1861 case GVDT_Count: {
1862 auto Count = Graph->getBlockProfileCount(Node);
1863 if (Count)
1864 OS << *Count;
1865 else
1866 OS << "Unknown";
1867 break;
1868 }
1869 case GVDT_None:
1870 llvm_unreachable("If we are not supposed to render a graph we should "
1871 "never reach this point.");
1872 }
1873 return Result;
1874 }
1875
1877 const BlockFrequencyInfoT *BFI,
1878 const BranchProbabilityInfoT *BPI,
1879 unsigned HotPercentThreshold = 0) {
1880 std::string Str;
1881 if (!BPI)
1882 return Str;
1883
1884 BranchProbability BP = BPI->getEdgeProbability(Node, EI);
1885 uint32_t N = BP.getNumerator();
1886 uint32_t D = BP.getDenominator();
1887 double Percent = 100.0 * N / D;
1889 OS << format("label=\"%.1f%%\"", Percent);
1890
1891 if (HotPercentThreshold) {
1892 BlockFrequency EFreq = BFI->getBlockFreq(Node) * BP;
1894 BranchProbability(HotPercentThreshold, 100);
1895
1896 if (EFreq >= HotFreq) {
1897 OS << ",color=\"red\"";
1898 }
1899 }
1900
1901 OS.flush();
1902 return Str;
1903 }
1904};
1905
1906} // end namespace llvm
1907
1908#undef DEBUG_TYPE
1909
1910#endif // LLVM_ANALYSIS_BLOCKFREQUENCYINFOIMPL_H
BitTracker BT
Definition: BitTracker.cpp:73
This file implements the BitVector class.
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
static GCRegistry::Add< StatepointGC > D("statepoint-example", "an example strategy for statepoint")
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
#define LLVM_DEBUG(X)
Definition: Debug.h:101
This file defines the DenseMap class.
This file defines the DenseSet and SmallDenseSet classes.
DenseMap< Block *, BlockRelaxAux > Blocks
Definition: ELF_riscv.cpp:507
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
This file defines the little GraphTraits<X> template class that should be specialized by classes that...
Hexagon Hardware Loops
static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT, AssumptionCache *AC)
Definition: Lint.cpp:531
#define F(x, y, z)
Definition: MD5.cpp:55
#define I(x, y, z)
Definition: MD5.cpp:58
#define G(x, y, z)
Definition: MD5.cpp:56
#define H(x, y, z)
Definition: MD5.cpp:57
Branch Probability Basic Block static false std::string getBlockName(const MachineBasicBlock *BB)
Helper to print the name of a MBB.
#define P(N)
This file builds on the ADT/GraphTraits.h file to build a generic graph post order iterator.
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
raw_pwrite_stream & OS
This file defines the SmallPtrSet class.
This file defines the SmallVector class.
This file defines the SparseBitVector class.
ScaledNumber< uint64_t > Scaled64
Value handle that asserts if the Value is deleted.
Definition: ValueHandle.h:264
LLVM Basic Block Representation.
Definition: BasicBlock.h:60
Base class for BlockFrequencyInfoImpl.
std::vector< WorkingData > Working
Loop data: see initializeLoops().
virtual ~BlockFrequencyInfoImplBase()=default
Virtual destructor.
std::list< LoopData > Loops
Indexed information about loops.
bool addLoopSuccessorsToDist(const LoopData *OuterLoop, LoopData &Loop, Distribution &Dist)
Add all edges out of a packaged loop to the distribution.
std::string getLoopName(const LoopData &Loop) const
bool isIrrLoopHeader(const BlockNode &Node)
void computeLoopScale(LoopData &Loop)
Compute the loop scale for a loop.
void packageLoop(LoopData &Loop)
Package up a loop.
virtual raw_ostream & print(raw_ostream &OS) const
virtual std::string getBlockName(const BlockNode &Node) const
void finalizeMetrics()
Finalize frequency metrics.
void setBlockFreq(const BlockNode &Node, BlockFrequency Freq)
void updateLoopWithIrreducible(LoopData &OuterLoop)
Update a loop after packaging irreducible SCCs inside of it.
std::optional< uint64_t > getBlockProfileCount(const Function &F, const BlockNode &Node, bool AllowSynthetic=false) const
BlockFrequency getBlockFreq(const BlockNode &Node) const
void distributeIrrLoopHeaderMass(Distribution &Dist)
iterator_range< std::list< LoopData >::iterator > analyzeIrreducible(const bfi_detail::IrreducibleGraph &G, LoopData *OuterLoop, std::list< LoopData >::iterator Insert)
Analyze irreducible SCCs.
bool addToDist(Distribution &Dist, const LoopData *OuterLoop, const BlockNode &Pred, const BlockNode &Succ, uint64_t Weight)
Add an edge to the distribution.
std::optional< uint64_t > getProfileCountFromFreq(const Function &F, BlockFrequency Freq, bool AllowSynthetic=false) const
Scaled64 getFloatingBlockFreq(const BlockNode &Node) const
void distributeMass(const BlockNode &Source, LoopData *OuterLoop, Distribution &Dist)
Distribute mass according to a distribution.
SparseBitVector IsIrrLoopHeader
Whether each block is an irreducible loop header.
std::vector< FrequencyData > Freqs
Data about each block. This is used downstream.
void adjustLoopHeaderMass(LoopData &Loop)
Adjust the mass of all headers in an irreducible loop.
Shared implementation for block frequency analysis.
bool isIrrLoopHeader(const BlockT *BB)
std::optional< uint64_t > getBlockProfileCount(const Function &F, const BlockT *BB, bool AllowSynthetic=false) const
const BranchProbabilityInfoT & getBPI() const
const FunctionT * getFunction() const
void verifyMatch(BlockFrequencyInfoImpl< BT > &Other) const
Scaled64 getFloatingBlockFreq(const BlockT *BB) const
std::optional< uint64_t > getProfileCountFromFreq(const Function &F, BlockFrequency Freq, bool AllowSynthetic=false) const
void calculate(const FunctionT &F, const BranchProbabilityInfoT &BPI, const LoopInfoT &LI)
void setBlockFreq(const BlockT *BB, BlockFrequency Freq)
raw_ostream & print(raw_ostream &OS) const override
Print the frequencies for the current function.
BlockFrequency getBlockFreq(const BlockT *BB) const
Analysis providing branch probability information.
static BranchProbability getBranchProbability(uint64_t Numerator, uint64_t Denominator)
static uint32_t getDenominator()
uint32_t getNumerator() const
Value handle with callbacks on RAUW and destruction.
Definition: ValueHandle.h:383
ValueT lookup(const_arg_type_t< KeyT > Val) const
lookup - Return the entry for the specified key, or a default constructed value if no such entry exis...
Definition: DenseMap.h:202
bool erase(const KeyT &Val)
Definition: DenseMap.h:329
unsigned size() const
Definition: DenseMap.h:99
iterator begin()
Definition: DenseMap.h:75
size_type count(const_arg_type_t< KeyT > Val) const
Return 1 if the specified key is in the map, 0 otherwise.
Definition: DenseMap.h:151
Class to represent profile counts.
Definition: Function.h:277
Represents a single loop in the control flow graph.
Definition: LoopInfo.h:44
Simple representation of a scaled number.
Definition: ScaledNumber.h:493
size_t size() const
Definition: SmallVector.h:91
iterator insert(iterator I, T &&Elt)
Definition: SmallVector.h:818
void resize(size_type N)
Definition: SmallVector.h:651
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:50
std::string str() const
str - Get the contents as an std::string.
Definition: StringRef.h:222
Twine - A lightweight data structure for efficiently representing the concatenation of temporary valu...
Definition: Twine.h:81
The instances of the Type class are immutable: once they are created, they are never changed.
Definition: Type.h:45
StringRef getName() const
Return a constant reference to the value's name.
Definition: Value.cpp:309
void deleted() override
Callback for Value destruction.
BFICallbackVH(const BasicBlock *BB, BFIImplT *BFIImpl)
bool operator<(BlockMass X) const
bool operator>(BlockMass X) const
raw_ostream & print(raw_ostream &OS) const
bool operator==(BlockMass X) const
BlockMass & operator-=(BlockMass X)
Subtract another mass.
bool operator<=(BlockMass X) const
BlockMass & operator*=(BranchProbability P)
bool operator!=(BlockMass X) const
BlockMass & operator+=(BlockMass X)
Add another mass.
bool operator>=(BlockMass X) const
ScaledNumber< uint64_t > toScaled() const
Convert to scaled number.
A range adaptor for a pair of iterators.
This class implements an extremely fast bulk output stream that can only output to a stream.
Definition: raw_ostream.h:52
A raw_ostream that writes to an std::string.
Definition: raw_ostream.h:660
This provides a very simple, boring adaptor for a begin and end iterator into a range type.
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
std::string getBlockName(const BlockT *BB)
Get the name of a MachineBasicBlock.
BlockMass operator*(BlockMass L, BranchProbability R)
BlockMass operator+(BlockMass L, BlockMass R)
raw_ostream & operator<<(raw_ostream &OS, BlockMass X)
BlockMass operator-(BlockMass L, BlockMass R)
std::optional< const char * > toString(const std::optional< DWARFFormValue > &V)
Take an optional DWARFFormValue and try to extract a string value from it.
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:18
GCNRegPressure max(const GCNRegPressure &P1, const GCNRegPressure &P2)
uint32_t getWeightFromBranchProb(const BranchProbability Prob)
Function::ProfileCount ProfileCount
iterator_range< T > make_range(T x, T y)
Convenience function for iterating over sub-ranges.
Printable print(const GCNRegPressure &RP, const GCNSubtarget *ST=nullptr)
llvm::cl::opt< unsigned > IterativeBFIMaxIterationsPerBlock
po_iterator< T > po_begin(const T &G)
llvm::cl::opt< bool > UseIterativeBFIInference
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:163
format_object< Ts... > format(const char *Fmt, const Ts &... Vals)
These are helper functions used to produce formatted output.
Definition: Format.h:125
llvm::cl::opt< bool > CheckBFIUnknownBlockQueries
@ Other
Any other memory.
Printable printBlockFreq(const BlockFrequencyInfo &BFI, BlockFrequency Freq)
Print the block frequency Freq relative to the current functions entry frequency.
po_iterator< T > po_end(const T &G)
llvm::cl::opt< double > IterativeBFIPrecision
Implement std::hash so that hash_code can be used in STL containers.
Definition: BitVector.h:858
#define N
std::string getNodeAttributes(NodeRef Node, const BlockFrequencyInfoT *Graph, unsigned HotPercentThreshold=0)
typename GTraits::nodes_iterator NodeIter
typename GTraits::NodeRef NodeRef
typename GTraits::ChildIteratorType EdgeIter
std::string getNodeLabel(NodeRef Node, const BlockFrequencyInfoT *Graph, GVDAGType GType, int layout_order=-1)
std::string getEdgeAttributes(NodeRef Node, EdgeIter EI, const BlockFrequencyInfoT *BFI, const BranchProbabilityInfoT *BPI, unsigned HotPercentThreshold=0)
BFIDOTGraphTraitsBase(bool isSimple=false)
static StringRef getGraphName(const BlockFrequencyInfoT *G)
Distribution of unscaled probability weight.
void addBackedge(const BlockNode &Node, uint64_t Amount)
WeightList Weights
Individual successor weights.
void addExit(const BlockNode &Node, uint64_t Amount)
void addLocal(const BlockNode &Node, uint64_t Amount)
bool isHeader(const BlockNode &Node) const
LoopData(LoopData *Parent, It1 FirstHeader, It1 LastHeader, It2 FirstOther, It2 LastOther)
ExitMap Exits
Successor edges (and weights).
bool IsPackaged
Whether this has been packaged.
LoopData(LoopData *Parent, const BlockNode &Header)
NodeList::const_iterator members_end() const
NodeList::const_iterator members_begin() const
NodeList Nodes
Header and the members of the loop.
HeaderMassList BackedgeMass
Mass returned to each loop header.
HeaderMassList::difference_type getHeaderIndex(const BlockNode &B)
iterator_range< NodeList::const_iterator > members() const
Weight(DistType Type, BlockNode TargetNode, uint64_t Amount)
bool isPackaged() const
Has ContainingLoop been packaged up?
BlockMass Mass
Mass distribution from the entry block.
BlockMass & getMass()
Get the appropriate mass for a node.
bool isAPackage() const
Has Loop been packaged up?
LoopData * Loop
The loop this block is inside.
BlockNode getResolvedNode() const
Resolve a node to its representative.
bool isADoublePackage() const
Has Loop been packaged up twice?
DefaultDOTGraphTraits - This class provides the default implementations of all of the DOTGraphTraits ...
typename GraphType::UnknownGraphTypeError NodeRef
Definition: GraphTraits.h:80
BlockEdgesAdder(const BlockFrequencyInfoImpl< BT > &BFI)
const BlockFrequencyInfoImpl< BT > & BFI
void operator()(IrreducibleGraph &G, IrreducibleGraph::IrrNode &Irr, const LoopData *OuterLoop)
std::deque< const IrrNode * >::const_iterator iterator
Graph of irreducible control flow.
IrreducibleGraph(BFIBase &BFI, const BFIBase::LoopData *OuterLoop, BlockEdgesAdder addBlockEdges)
Construct an explicit graph containing irreducible control flow.
void addEdge(IrrNode &Irr, const BlockNode &Succ, const BFIBase::LoopData *OuterLoop)
void addEdges(const BlockNode &Node, const BFIBase::LoopData *OuterLoop, BlockEdgesAdder addBlockEdges)
SmallDenseMap< uint32_t, IrrNode *, 4 > Lookup
void initialize(const BFIBase::LoopData *OuterLoop, BlockEdgesAdder addBlockEdges)
void addNodesInLoop(const BFIBase::LoopData &OuterLoop)