Coverage Report

Created: 2018-09-17 19:50

/Users/buildslave/jenkins/workspace/clang-stage2-coverage-R/llvm/include/llvm/Analysis/BlockFrequencyInfoImpl.h
Line
Count
Source (jump to first uncovered line)
1
//==- BlockFrequencyInfoImpl.h - Block Frequency Implementation --*- C++ -*-==//
2
//
3
//                     The LLVM Compiler Infrastructure
4
//
5
// This file is distributed under the University of Illinois Open Source
6
// License. See LICENSE.TXT for details.
7
//
8
//===----------------------------------------------------------------------===//
9
//
10
// Shared implementation of BlockFrequency for IR and Machine Instructions.
11
// See the documentation below for BlockFrequencyInfoImpl for details.
12
//
13
//===----------------------------------------------------------------------===//
14
15
#ifndef LLVM_ANALYSIS_BLOCKFREQUENCYINFOIMPL_H
16
#define LLVM_ANALYSIS_BLOCKFREQUENCYINFOIMPL_H
17
18
#include "llvm/ADT/DenseMap.h"
19
#include "llvm/ADT/DenseSet.h"
20
#include "llvm/ADT/GraphTraits.h"
21
#include "llvm/ADT/Optional.h"
22
#include "llvm/ADT/PostOrderIterator.h"
23
#include "llvm/ADT/SmallVector.h"
24
#include "llvm/ADT/SparseBitVector.h"
25
#include "llvm/ADT/Twine.h"
26
#include "llvm/ADT/iterator_range.h"
27
#include "llvm/IR/BasicBlock.h"
28
#include "llvm/Support/BlockFrequency.h"
29
#include "llvm/Support/BranchProbability.h"
30
#include "llvm/Support/DOTGraphTraits.h"
31
#include "llvm/Support/Debug.h"
32
#include "llvm/Support/ErrorHandling.h"
33
#include "llvm/Support/Format.h"
34
#include "llvm/Support/ScaledNumber.h"
35
#include "llvm/Support/raw_ostream.h"
36
#include <algorithm>
37
#include <cassert>
38
#include <cstddef>
39
#include <cstdint>
40
#include <deque>
41
#include <iterator>
42
#include <limits>
43
#include <list>
44
#include <string>
45
#include <utility>
46
#include <vector>
47
48
#define DEBUG_TYPE "block-freq"
49
50
namespace llvm {
51
52
class BranchProbabilityInfo;
53
class Function;
54
class Loop;
55
class LoopInfo;
56
class MachineBasicBlock;
57
class MachineBranchProbabilityInfo;
58
class MachineFunction;
59
class MachineLoop;
60
class MachineLoopInfo;
61
62
namespace bfi_detail {
63
64
struct IrreducibleGraph;
65
66
// This is part of a workaround for a GCC 4.7 crash on lambdas.
67
template <class BT> struct BlockEdgesAdder;
68
69
/// Mass of a block.
70
///
71
/// This class implements a sort of fixed-point fraction always between 0.0 and
72
/// 1.0.  getMass() == std::numeric_limits<uint64_t>::max() indicates a value of
73
/// 1.0.
74
///
75
/// Masses can be added and subtracted.  Simple saturation arithmetic is used,
76
/// so arithmetic operations never overflow or underflow.
77
///
78
/// Masses can be multiplied.  Multiplication treats full mass as 1.0 and uses
79
/// an inexpensive floating-point algorithm that's off-by-one (almost, but not
80
/// quite, maximum precision).
81
///
82
/// Masses can be scaled by \a BranchProbability at maximum precision.
83
0
class BlockMass {
84
  uint64_t Mass = 0;
85
86
public:
87
62.4M
  BlockMass() = default;
88
8.19M
  explicit BlockMass(uint64_t Mass) : Mass(Mass) {}
89
90
26.3k
  static BlockMass getEmpty() { return BlockMass(); }
91
92
8.19M
  static BlockMass getFull() {
93
8.19M
    return BlockMass(std::numeric_limits<uint64_t>::max());
94
8.19M
  }
95
96
25.4M
  uint64_t getMass() const { return Mass; }
97
98
31.2M
  bool isFull() const { return Mass == std::numeric_limits<uint64_t>::max(); }
99
2.04M
  bool isEmpty() const { return !Mass; }
100
101
  bool operator!() const { return isEmpty(); }
102
103
  /// Add another mass.
104
  ///
105
  /// Adds another mass, saturating at \a isFull() rather than overflowing.
106
37.0M
  BlockMass &operator+=(BlockMass X) {
107
37.0M
    uint64_t Sum = Mass + X.Mass;
108
37.0M
    Mass = Sum < Mass ? 
std::numeric_limits<uint64_t>::max()0
: Sum;
109
37.0M
    return *this;
110
37.0M
  }
111
112
  /// Subtract another mass.
113
  ///
114
  /// Subtracts another mass, saturating at \a isEmpty() rather than
115
  /// undeflowing.
116
39.8M
  BlockMass &operator-=(BlockMass X) {
117
39.8M
    uint64_t Diff = Mass - X.Mass;
118
39.8M
    Mass = Diff > Mass ? 
00
: Diff;
119
39.8M
    return *this;
120
39.8M
  }
121
122
37.8M
  BlockMass &operator*=(BranchProbability P) {
123
37.8M
    Mass = P.scale(Mass);
124
37.8M
    return *this;
125
37.8M
  }
126
127
  bool operator==(BlockMass X) const { return Mass == X.Mass; }
128
  bool operator!=(BlockMass X) const { return Mass != X.Mass; }
129
  bool operator<=(BlockMass X) const { return Mass <= X.Mass; }
130
  bool operator>=(BlockMass X) const { return Mass >= X.Mass; }
131
  bool operator<(BlockMass X) const { return Mass < X.Mass; }
132
  bool operator>(BlockMass X) const { return Mass > X.Mass; }
133
134
  /// Convert to scaled number.
135
  ///
136
  /// Convert to \a ScaledNumber.  \a isFull() gives 1.0, while \a isEmpty()
137
  /// gives slightly above 0.0.
138
  ScaledNumber<uint64_t> toScaled() const;
139
140
  void dump() const;
141
  raw_ostream &print(raw_ostream &OS) const;
142
};
143
144
0
inline BlockMass operator+(BlockMass L, BlockMass R) {
145
0
  return BlockMass(L) += R;
146
0
}
147
2.04M
inline BlockMass operator-(BlockMass L, BlockMass R) {
148
2.04M
  return BlockMass(L) -= R;
149
2.04M
}
150
37.8M
inline BlockMass operator*(BlockMass L, BranchProbability R) {
151
37.8M
  return BlockMass(L) *= R;
152
37.8M
}
153
0
inline BlockMass operator*(BranchProbability L, BlockMass R) {
154
0
  return BlockMass(R) *= L;
155
0
}
156
157
0
inline raw_ostream &operator<<(raw_ostream &OS, BlockMass X) {
158
0
  return X.print(OS);
159
0
}
160
161
} // end namespace bfi_detail
162
163
template <> struct isPodLike<bfi_detail::BlockMass> {
164
  static const bool value = true;
165
};
166
167
/// Base class for BlockFrequencyInfoImpl
168
///
169
/// BlockFrequencyInfoImplBase has supporting data structures and some
170
/// algorithms for BlockFrequencyInfoImplBase.  Only algorithms that depend on
171
/// the block type (or that call such algorithms) are skipped here.
172
///
173
/// Nevertheless, the majority of the overall algorithm documention lives with
174
/// BlockFrequencyInfoImpl.  See there for details.
175
class BlockFrequencyInfoImplBase {
176
public:
177
  using Scaled64 = ScaledNumber<uint64_t>;
178
  using BlockMass = bfi_detail::BlockMass;
179
180
  /// Representative of a block.
181
  ///
182
  /// This is a simple wrapper around an index into the reverse-post-order
183
  /// traversal of the blocks.
184
  ///
185
  /// Unlike a block pointer, its order has meaning (location in the
186
  /// topological sort) and it's class is the same regardless of block type.
187
  struct BlockNode {
188
    using IndexType = uint32_t;
189
190
    IndexType Index = std::numeric_limits<uint32_t>::max();
191
192
27.3M
    BlockNode() = default;
193
86.8M
    BlockNode(IndexType Index) : Index(Index) {}
194
195
75.8M
    bool operator==(const BlockNode &X) const { return Index == X.Index; }
196
27.1M
    bool operator!=(const BlockNode &X) const { return Index != X.Index; }
197
0
    bool operator<=(const BlockNode &X) const { return Index <= X.Index; }
198
0
    bool operator>=(const BlockNode &X) const { return Index >= X.Index; }
199
48.5M
    bool operator<(const BlockNode &X) const { return Index < X.Index; }
200
0
    bool operator>(const BlockNode &X) const { return Index > X.Index; }
201
202
46.4M
    bool isValid() const { return Index <= getMaxIndex(); }
203
204
46.4M
    static size_t getMaxIndex() {
205
46.4M
       return std::numeric_limits<uint32_t>::max() - 1;
206
46.4M
    }
207
  };
208
209
  /// Stats about a block itself.
210
  struct FrequencyData {
211
    Scaled64 Scaled;
212
    uint64_t Integer;
213
  };
214
215
  /// Data about a loop.
216
  ///
217
  /// Contains the data necessary to represent a loop as a pseudo-node once it's
218
  /// packaged.
219
  struct LoopData {
220
    using ExitMap = SmallVector<std::pair<BlockNode, BlockMass>, 4>;
221
    using NodeList = SmallVector<BlockNode, 4>;
222
    using HeaderMassList = SmallVector<BlockMass, 1>;
223
224
    LoopData *Parent;            ///< The parent loop.
225
    bool IsPackaged = false;     ///< Whether this has been packaged.
226
    uint32_t NumHeaders = 1;     ///< Number of headers.
227
    ExitMap Exits;               ///< Successor edges (and weights).
228
    NodeList Nodes;              ///< Header and the members of the loop.
229
    HeaderMassList BackedgeMass; ///< Mass returned to each loop header.
230
    BlockMass Mass;
231
    Scaled64 Scale;
232
233
    LoopData(LoopData *Parent, const BlockNode &Header)
234
2.03M
      : Parent(Parent), Nodes(1, Header), BackedgeMass(1) {}
235
236
    template <class It1, class It2>
237
    LoopData(LoopData *Parent, It1 FirstHeader, It1 LastHeader, It2 FirstOther,
238
             It2 LastOther)
239
789
        : Parent(Parent), Nodes(FirstHeader, LastHeader) {
240
789
      NumHeaders = Nodes.size();
241
789
      Nodes.insert(Nodes.end(), FirstOther, LastOther);
242
789
      BackedgeMass.resize(NumHeaders);
243
789
    }
244
245
62.0M
    bool isHeader(const BlockNode &Node) const {
246
62.0M
      if (isIrreducible())
247
92.7k
        return std::binary_search(Nodes.begin(), Nodes.begin() + NumHeaders,
248
92.7k
                                  Node);
249
62.0M
      return Node == Nodes[0];
250
62.0M
    }
251
252
15.8M
    BlockNode getHeader() const { return Nodes[0]; }
253
68.4M
    bool isIrreducible() const { return NumHeaders > 1; }
254
255
2.07M
    HeaderMassList::difference_type getHeaderIndex(const BlockNode &B) {
256
2.07M
      assert(isHeader(B) && "this is only valid on loop header blocks");
257
2.07M
      if (isIrreducible())
258
7.59k
        return std::lower_bound(Nodes.begin(), Nodes.begin() + NumHeaders, B) -
259
7.59k
               Nodes.begin();
260
2.07M
      return 0;
261
2.07M
    }
262
263
2.03M
    NodeList::const_iterator members_begin() const {
264
2.03M
      return Nodes.begin() + NumHeaders;
265
2.03M
    }
266
267
2.03M
    NodeList::const_iterator members_end() const { return Nodes.end(); }
268
2.03M
    iterator_range<NodeList::const_iterator> members() const {
269
2.03M
      return make_range(members_begin(), members_end());
270
2.03M
    }
271
  };
272
273
  /// Index of loop information.
274
  struct WorkingData {
275
    BlockNode Node;           ///< This node.
276
    LoopData *Loop = nullptr; ///< The loop this block is inside.
277
    BlockMass Mass;           ///< Mass distribution from the entry block.
278
279
27.1M
    WorkingData(const BlockNode &Node) : Node(Node) {}
280
281
139M
    bool isLoopHeader() const { return Loop && 
Loop->isHeader(Node)49.5M
; }
282
283
8.23M
    bool isDoubleLoopHeader() const {
284
8.23M
      return isLoopHeader() && Loop->Parent && 
Loop->Parent->isIrreducible()2.18M
&&
285
8.23M
             
Loop->Parent->isHeader(Node)4.30k
;
286
8.23M
    }
287
288
38.0M
    LoopData *getContainingLoop() const {
289
38.0M
      if (!isLoopHeader())
290
33.9M
        return Loop;
291
4.13M
      if (!isDoubleLoopHeader())
292
4.12M
        return Loop->Parent;
293
175
      return Loop->Parent->Parent;
294
175
    }
295
296
    /// Resolve a node to its representative.
297
    ///
298
    /// Get the node currently representing Node, which could be a containing
299
    /// loop.
300
    ///
301
    /// This function should only be called when distributing mass.  As long as
302
    /// there are no irreducible edges to Node, then it will have complexity
303
    /// O(1) in this context.
304
    ///
305
    /// In general, the complexity is O(L), where L is the number of loop
306
    /// headers Node has been packaged into.  Since this method is called in
307
    /// the context of distributing mass, L will be the number of loop headers
308
    /// an early exit edge jumps out of.
309
65.2M
    BlockNode getResolvedNode() const {
310
65.2M
      auto L = getPackagedLoop();
311
65.2M
      return L ? 
L->getHeader()8.95M
:
Node56.3M
;
312
65.2M
    }
313
314
102M
    LoopData *getPackagedLoop() const {
315
102M
      if (!Loop || 
!Loop->IsPackaged35.6M
)
316
90.3M
        return nullptr;
317
12.0M
      auto L = Loop;
318
14.4M
      while (L->Parent && 
L->Parent->IsPackaged4.50M
)
319
2.31M
        L = L->Parent;
320
12.0M
      return L;
321
12.0M
    }
322
323
    /// Get the appropriate mass for a node.
324
    ///
325
    /// Get appropriate mass for Node.  If Node is a loop-header (whose loop
326
    /// has been packaged), returns the mass of its pseudo-node.  If it's a
327
    /// node inside a packaged loop, it returns the loop's mass.
328
68.3M
    BlockMass &getMass() {
329
68.3M
      if (!isAPackage())
330
64.2M
        return Mass;
331
4.10M
      if (!isADoublePackage())
332
4.10M
        return Loop->Mass;
333
110
      return Loop->Parent->Mass;
334
110
    }
335
336
    /// Has ContainingLoop been packaged up?
337
27.1M
    bool isPackaged() const { return getResolvedNode() != Node; }
338
339
    /// Has Loop been packaged up?
340
75.7M
    bool isAPackage() const { return isLoopHeader() && 
Loop->IsPackaged10.7M
; }
341
342
    /// Has Loop been packaged up twice?
343
4.10M
    bool isADoublePackage() const {
344
4.10M
      return isDoubleLoopHeader() && 
Loop->Parent->IsPackaged766
;
345
4.10M
    }
346
  };
347
348
  /// Unscaled probability weight.
349
  ///
350
  /// Probability weight for an edge in the graph (including the
351
  /// successor/target node).
352
  ///
353
  /// All edges in the original function are 32-bit.  However, exit edges from
354
  /// loop packages are taken from 64-bit exit masses, so we need 64-bits of
355
  /// space in general.
356
  ///
357
  /// In addition to the raw weight amount, Weight stores the type of the edge
358
  /// in the current context (i.e., the context of the loop being processed).
359
  /// Is this a local edge within the loop, an exit from the loop, or a
360
  /// backedge to the loop header?
361
  struct Weight {
362
    enum DistType { Local, Exit, Backedge };
363
    DistType Type = Local;
364
    BlockNode TargetNode;
365
    uint64_t Amount = 0;
366
367
27.1k
    Weight() = default;
368
    Weight(DistType Type, BlockNode TargetNode, uint64_t Amount)
369
38.1M
        : Type(Type), TargetNode(TargetNode), Amount(Amount) {}
370
  };
371
372
  /// Distribution of unscaled probability weight.
373
  ///
374
  /// Distribution of unscaled probability weight to a set of successors.
375
  ///
376
  /// This class collates the successor edge weights for later processing.
377
  ///
378
  /// \a DidOverflow indicates whether \a Total did overflow while adding to
379
  /// the distribution.  It should never overflow twice.
380
  struct Distribution {
381
    using WeightList = SmallVector<Weight, 4>;
382
383
    WeightList Weights;       ///< Individual successor weights.
384
    uint64_t Total = 0;       ///< Sum of all weights.
385
    bool DidOverflow = false; ///< Whether \a Total did overflow.
386
387
29.1M
    Distribution() = default;
388
389
33.2M
    void addLocal(const BlockNode &Node, uint64_t Amount) {
390
33.2M
      add(Node, Amount, Weight::Local);
391
33.2M
    }
392
393
2.81M
    void addExit(const BlockNode &Node, uint64_t Amount) {
394
2.81M
      add(Node, Amount, Weight::Exit);
395
2.81M
    }
396
397
2.07M
    void addBackedge(const BlockNode &Node, uint64_t Amount) {
398
2.07M
      add(Node, Amount, Weight::Backedge);
399
2.07M
    }
400
401
    /// Normalize the distribution.
402
    ///
403
    /// Combines multiple edges to the same \a Weight::TargetNode and scales
404
    /// down so that \a Total fits into 32-bits.
405
    ///
406
    /// This is linear in the size of \a Weights.  For the vast majority of
407
    /// cases, adjacent edge weights are combined by sorting WeightList and
408
    /// combining adjacent weights.  However, for very large edge lists an
409
    /// auxiliary hash table is used.
410
    void normalize();
411
412
  private:
413
    void add(const BlockNode &Node, uint64_t Amount, Weight::DistType Type);
414
  };
415
416
  /// Data about each block.  This is used downstream.
417
  std::vector<FrequencyData> Freqs;
418
419
  /// Whether each block is an irreducible loop header.
420
  /// This is used downstream.
421
  SparseBitVector<> IsIrrLoopHeader;
422
423
  /// Loop data: see initializeLoops().
424
  std::vector<WorkingData> Working;
425
426
  /// Indexed information about loops.
427
  std::list<LoopData> Loops;
428
429
  /// Virtual destructor.
430
  ///
431
  /// Need a virtual destructor to mask the compiler warning about
432
  /// getBlockName().
433
4.11M
  virtual ~BlockFrequencyInfoImplBase() = default;
434
435
  /// Add all edges out of a packaged loop to the distribution.
436
  ///
437
  /// Adds all edges from LocalLoopHead to Dist.  Calls addToDist() to add each
438
  /// successor edge.
439
  ///
440
  /// \return \c true unless there's an irreducible backedge.
441
  bool addLoopSuccessorsToDist(const LoopData *OuterLoop, LoopData &Loop,
442
                               Distribution &Dist);
443
444
  /// Add an edge to the distribution.
445
  ///
446
  /// Adds an edge to Succ to Dist.  If \c LoopHead.isValid(), then whether the
447
  /// edge is local/exit/backedge is in the context of LoopHead.  Otherwise,
448
  /// every edge should be a local edge (since all the loops are packaged up).
449
  ///
450
  /// \return \c true unless aborted due to an irreducible backedge.
451
  bool addToDist(Distribution &Dist, const LoopData *OuterLoop,
452
                 const BlockNode &Pred, const BlockNode &Succ, uint64_t Weight);
453
454
0
  LoopData &getLoopPackage(const BlockNode &Head) {
455
0
    assert(Head.Index < Working.size());
456
0
    assert(Working[Head.Index].isLoopHeader());
457
0
    return *Working[Head.Index].Loop;
458
0
  }
459
460
  /// Analyze irreducible SCCs.
461
  ///
462
  /// Separate irreducible SCCs from \c G, which is an explict graph of \c
463
  /// OuterLoop (or the top-level function, if \c OuterLoop is \c nullptr).
464
  /// Insert them into \a Loops before \c Insert.
465
  ///
466
  /// \return the \c LoopData nodes representing the irreducible SCCs.
467
  iterator_range<std::list<LoopData>::iterator>
468
  analyzeIrreducible(const bfi_detail::IrreducibleGraph &G, LoopData *OuterLoop,
469
                     std::list<LoopData>::iterator Insert);
470
471
  /// Update a loop after packaging irreducible SCCs inside of it.
472
  ///
473
  /// Update \c OuterLoop.  Before finding irreducible control flow, it was
474
  /// partway through \a computeMassInLoop(), so \a LoopData::Exits and \a
475
  /// LoopData::BackedgeMass need to be reset.  Also, nodes that were packaged
476
  /// up need to be removed from \a OuterLoop::Nodes.
477
  void updateLoopWithIrreducible(LoopData &OuterLoop);
478
479
  /// Distribute mass according to a distribution.
480
  ///
481
  /// Distributes the mass in Source according to Dist.  If LoopHead.isValid(),
482
  /// backedges and exits are stored in its entry in Loops.
483
  ///
484
  /// Mass is distributed in parallel from two copies of the source mass.
485
  void distributeMass(const BlockNode &Source, LoopData *OuterLoop,
486
                      Distribution &Dist);
487
488
  /// Compute the loop scale for a loop.
489
  void computeLoopScale(LoopData &Loop);
490
491
  /// Adjust the mass of all headers in an irreducible loop.
492
  ///
493
  /// Initially, irreducible loops are assumed to distribute their mass
494
  /// equally among its headers. This can lead to wrong frequency estimates
495
  /// since some headers may be executed more frequently than others.
496
  ///
497
  /// This adjusts header mass distribution so it matches the weights of
498
  /// the backedges going into each of the loop headers.
499
  void adjustLoopHeaderMass(LoopData &Loop);
500
501
  void distributeIrrLoopHeaderMass(Distribution &Dist);
502
503
  /// Package up a loop.
504
  void packageLoop(LoopData &Loop);
505
506
  /// Unwrap loops.
507
  void unwrapLoops();
508
509
  /// Finalize frequency metrics.
510
  ///
511
  /// Calculates final frequencies and cleans up no-longer-needed data
512
  /// structures.
513
  void finalizeMetrics();
514
515
  /// Clear all memory.
516
  void clear();
517
518
  virtual std::string getBlockName(const BlockNode &Node) const;
519
  std::string getLoopName(const LoopData &Loop) const;
520
521
0
  virtual raw_ostream &print(raw_ostream &OS) const { return OS; }
522
0
  void dump() const { print(dbgs()); }
523
524
  Scaled64 getFloatingBlockFreq(const BlockNode &Node) const;
525
526
  BlockFrequency getBlockFreq(const BlockNode &Node) const;
527
  Optional<uint64_t> getBlockProfileCount(const Function &F,
528
                                          const BlockNode &Node) const;
529
  Optional<uint64_t> getProfileCountFromFreq(const Function &F,
530
                                             uint64_t Freq) const;
531
  bool isIrrLoopHeader(const BlockNode &Node);
532
533
  void setBlockFreq(const BlockNode &Node, uint64_t Freq);
534
535
  raw_ostream &printBlockFreq(raw_ostream &OS, const BlockNode &Node) const;
536
  raw_ostream &printBlockFreq(raw_ostream &OS,
537
                              const BlockFrequency &Freq) const;
538
539
24.7M
  uint64_t getEntryFreq() const {
540
24.7M
    assert(!Freqs.empty());
541
24.7M
    return Freqs[0].Integer;
542
24.7M
  }
543
};
544
545
namespace bfi_detail {
546
547
template <class BlockT> struct TypeMap {};
548
template <> struct TypeMap<BasicBlock> {
549
  using BlockT = BasicBlock;
550
  using FunctionT = Function;
551
  using BranchProbabilityInfoT = BranchProbabilityInfo;
552
  using LoopT = Loop;
553
  using LoopInfoT = LoopInfo;
554
};
555
template <> struct TypeMap<MachineBasicBlock> {
556
  using BlockT = MachineBasicBlock;
557
  using FunctionT = MachineFunction;
558
  using BranchProbabilityInfoT = MachineBranchProbabilityInfo;
559
  using LoopT = MachineLoop;
560
  using LoopInfoT = MachineLoopInfo;
561
};
562
563
/// Get the name of a MachineBasicBlock.
564
///
565
/// Get the name of a MachineBasicBlock.  It's templated so that including from
566
/// CodeGen is unnecessary (that would be a layering issue).
567
///
568
/// This is used mainly for debug output.  The name is similar to
569
/// MachineBasicBlock::getFullName(), but skips the name of the function.
570
0
template <class BlockT> std::string getBlockName(const BlockT *BB) {
571
0
  assert(BB && "Unexpected nullptr");
572
0
  auto MachineName = "BB" + Twine(BB->getNumber());
573
0
  if (BB->getBasicBlock())
574
0
    return (MachineName + "[" + BB->getName() + "]").str();
575
0
  return MachineName.str();
576
0
}
577
/// Get the name of a BasicBlock.
578
532
template <> inline std::string getBlockName(const BasicBlock *BB) {
579
532
  assert(BB && "Unexpected nullptr");
580
532
  return BB->getName().str();
581
532
}
582
583
/// Graph of irreducible control flow.
584
///
585
/// This graph is used for determining the SCCs in a loop (or top-level
586
/// function) that has irreducible control flow.
587
///
588
/// During the block frequency algorithm, the local graphs are defined in a
589
/// light-weight way, deferring to the \a BasicBlock or \a MachineBasicBlock
590
/// graphs for most edges, but getting others from \a LoopData::ExitMap.  The
591
/// latter only has successor information.
592
///
593
/// \a IrreducibleGraph makes this graph explicit.  It's in a form that can use
594
/// \a GraphTraits (so that \a analyzeIrreducible() can use \a scc_iterator),
595
/// and it explicitly lists predecessors and successors.  The initialization
596
/// that relies on \c MachineBasicBlock is defined in the header.
597
struct IrreducibleGraph {
598
  using BFIBase = BlockFrequencyInfoImplBase;
599
600
  BFIBase &BFI;
601
602
  using BlockNode = BFIBase::BlockNode;
603
  struct IrrNode {
604
    BlockNode Node;
605
    unsigned NumIn = 0;
606
    std::deque<const IrrNode *> Edges;
607
608
26.2k
    IrrNode(const BlockNode &Node) : Node(Node) {}
609
610
    using iterator = std::deque<const IrrNode *>::const_iterator;
611
612
20.8k
    iterator pred_begin() const { return Edges.begin(); }
613
47.0k
    iterator succ_begin() const { return Edges.begin() + NumIn; }
614
20.8k
    iterator pred_end() const { return succ_begin(); }
615
69.7k
    iterator succ_end() const { return Edges.end(); }
616
  };
617
  BlockNode Start;
618
  const IrrNode *StartIrr = nullptr;
619
  std::vector<IrrNode> Nodes;
620
  SmallDenseMap<uint32_t, IrrNode *, 4> Lookup;
621
622
  /// Construct an explicit graph containing irreducible control flow.
623
  ///
624
  /// Construct an explicit graph of the control flow in \c OuterLoop (or the
625
  /// top-level function, if \c OuterLoop is \c nullptr).  Uses \c
626
  /// addBlockEdges to add block successors that have not been packaged into
627
  /// loops.
628
  ///
629
  /// \a BlockFrequencyInfoImpl::computeIrreducibleMass() is the only expected
630
  /// user of this.
631
  template <class BlockEdgesAdder>
632
  IrreducibleGraph(BFIBase &BFI, const BFIBase::LoopData *OuterLoop,
633
658
                   BlockEdgesAdder addBlockEdges) : BFI(BFI) {
634
658
    initialize(OuterLoop, addBlockEdges);
635
658
  }
llvm::bfi_detail::IrreducibleGraph::IrreducibleGraph<llvm::bfi_detail::BlockEdgesAdder<llvm::BasicBlock> >(llvm::BlockFrequencyInfoImplBase&, llvm::BlockFrequencyInfoImplBase::LoopData const*, llvm::bfi_detail::BlockEdgesAdder<llvm::BasicBlock>)
Line
Count
Source
633
305
                   BlockEdgesAdder addBlockEdges) : BFI(BFI) {
634
305
    initialize(OuterLoop, addBlockEdges);
635
305
  }
llvm::bfi_detail::IrreducibleGraph::IrreducibleGraph<llvm::bfi_detail::BlockEdgesAdder<llvm::MachineBasicBlock> >(llvm::BlockFrequencyInfoImplBase&, llvm::BlockFrequencyInfoImplBase::LoopData const*, llvm::bfi_detail::BlockEdgesAdder<llvm::MachineBasicBlock>)
Line
Count
Source
633
353
                   BlockEdgesAdder addBlockEdges) : BFI(BFI) {
634
353
    initialize(OuterLoop, addBlockEdges);
635
353
  }
636
637
  template <class BlockEdgesAdder>
638
  void initialize(const BFIBase::LoopData *OuterLoop,
639
                  BlockEdgesAdder addBlockEdges);
640
  void addNodesInLoop(const BFIBase::LoopData &OuterLoop);
641
  void addNodesInFunction();
642
643
26.2k
  void addNode(const BlockNode &Node) {
644
26.2k
    Nodes.emplace_back(Node);
645
26.2k
    BFI.Working[Node.Index].getMass() = BlockMass::getEmpty();
646
26.2k
  }
647
648
  void indexNodes();
649
  template <class BlockEdgesAdder>
650
  void addEdges(const BlockNode &Node, const BFIBase::LoopData *OuterLoop,
651
                BlockEdgesAdder addBlockEdges);
652
  void addEdge(IrrNode &Irr, const BlockNode &Succ,
653
               const BFIBase::LoopData *OuterLoop);
654
};
655
656
template <class BlockEdgesAdder>
657
void IrreducibleGraph::initialize(const BFIBase::LoopData *OuterLoop,
658
658
                                  BlockEdgesAdder addBlockEdges) {
659
658
  if (OuterLoop) {
660
141
    addNodesInLoop(*OuterLoop);
661
141
    for (auto N : OuterLoop->Nodes)
662
6.64k
      addEdges(N, OuterLoop, addBlockEdges);
663
517
  } else {
664
517
    addNodesInFunction();
665
23.3k
    for (uint32_t Index = 0; Index < BFI.Working.size(); 
++Index22.8k
)
666
22.8k
      addEdges(Index, OuterLoop, addBlockEdges);
667
517
  }
668
658
  StartIrr = Lookup[Start.Index];
669
658
}
void llvm::bfi_detail::IrreducibleGraph::initialize<llvm::bfi_detail::BlockEdgesAdder<llvm::BasicBlock> >(llvm::BlockFrequencyInfoImplBase::LoopData const*, llvm::bfi_detail::BlockEdgesAdder<llvm::BasicBlock>)
Line
Count
Source
658
305
                                  BlockEdgesAdder addBlockEdges) {
659
305
  if (OuterLoop) {
660
70
    addNodesInLoop(*OuterLoop);
661
70
    for (auto N : OuterLoop->Nodes)
662
3.23k
      addEdges(N, OuterLoop, addBlockEdges);
663
235
  } else {
664
235
    addNodesInFunction();
665
11.2k
    for (uint32_t Index = 0; Index < BFI.Working.size(); 
++Index11.0k
)
666
11.0k
      addEdges(Index, OuterLoop, addBlockEdges);
667
235
  }
668
305
  StartIrr = Lookup[Start.Index];
669
305
}
void llvm::bfi_detail::IrreducibleGraph::initialize<llvm::bfi_detail::BlockEdgesAdder<llvm::MachineBasicBlock> >(llvm::BlockFrequencyInfoImplBase::LoopData const*, llvm::bfi_detail::BlockEdgesAdder<llvm::MachineBasicBlock>)
Line
Count
Source
658
353
                                  BlockEdgesAdder addBlockEdges) {
659
353
  if (OuterLoop) {
660
71
    addNodesInLoop(*OuterLoop);
661
71
    for (auto N : OuterLoop->Nodes)
662
3.41k
      addEdges(N, OuterLoop, addBlockEdges);
663
282
  } else {
664
282
    addNodesInFunction();
665
12.1k
    for (uint32_t Index = 0; Index < BFI.Working.size(); 
++Index11.8k
)
666
11.8k
      addEdges(Index, OuterLoop, addBlockEdges);
667
282
  }
668
353
  StartIrr = Lookup[Start.Index];
669
353
}
670
671
template <class BlockEdgesAdder>
672
void IrreducibleGraph::addEdges(const BlockNode &Node,
673
                                const BFIBase::LoopData *OuterLoop,
674
29.4k
                                BlockEdgesAdder addBlockEdges) {
675
29.4k
  auto L = Lookup.find(Node.Index);
676
29.4k
  if (L == Lookup.end())
677
3.28k
    return;
678
26.2k
  IrrNode &Irr = *L->second;
679
26.2k
  const auto &Working = BFI.Working[Node.Index];
680
26.2k
681
26.2k
  if (Working.isAPackage())
682
2.42k
    for (const auto &I : Working.Loop->Exits)
683
6.97k
      addEdge(Irr, I.first, OuterLoop);
684
23.7k
  else
685
23.7k
    addBlockEdges(*this, Irr, OuterLoop);
686
26.2k
}
void llvm::bfi_detail::IrreducibleGraph::addEdges<llvm::bfi_detail::BlockEdgesAdder<llvm::BasicBlock> >(llvm::BlockFrequencyInfoImplBase::BlockNode const&, llvm::BlockFrequencyInfoImplBase::LoopData const*, llvm::bfi_detail::BlockEdgesAdder<llvm::BasicBlock>)
Line
Count
Source
674
14.2k
                                BlockEdgesAdder addBlockEdges) {
675
14.2k
  auto L = Lookup.find(Node.Index);
676
14.2k
  if (L == Lookup.end())
677
1.66k
    return;
678
12.5k
  IrrNode &Irr = *L->second;
679
12.5k
  const auto &Working = BFI.Working[Node.Index];
680
12.5k
681
12.5k
  if (Working.isAPackage())
682
1.02k
    for (const auto &I : Working.Loop->Exits)
683
2.52k
      addEdge(Irr, I.first, OuterLoop);
684
11.5k
  else
685
11.5k
    addBlockEdges(*this, Irr, OuterLoop);
686
12.5k
}
void llvm::bfi_detail::IrreducibleGraph::addEdges<llvm::bfi_detail::BlockEdgesAdder<llvm::MachineBasicBlock> >(llvm::BlockFrequencyInfoImplBase::BlockNode const&, llvm::BlockFrequencyInfoImplBase::LoopData const*, llvm::bfi_detail::BlockEdgesAdder<llvm::MachineBasicBlock>)
Line
Count
Source
674
15.2k
                                BlockEdgesAdder addBlockEdges) {
675
15.2k
  auto L = Lookup.find(Node.Index);
676
15.2k
  if (L == Lookup.end())
677
1.62k
    return;
678
13.6k
  IrrNode &Irr = *L->second;
679
13.6k
  const auto &Working = BFI.Working[Node.Index];
680
13.6k
681
13.6k
  if (Working.isAPackage())
682
1.40k
    for (const auto &I : Working.Loop->Exits)
683
4.45k
      addEdge(Irr, I.first, OuterLoop);
684
12.1k
  else
685
12.1k
    addBlockEdges(*this, Irr, OuterLoop);
686
13.6k
}
687
688
} // end namespace bfi_detail
689
690
/// Shared implementation for block frequency analysis.
691
///
692
/// This is a shared implementation of BlockFrequencyInfo and
693
/// MachineBlockFrequencyInfo, and calculates the relative frequencies of
694
/// blocks.
695
///
696
/// LoopInfo defines a loop as a "non-trivial" SCC dominated by a single block,
697
/// which is called the header.  A given loop, L, can have sub-loops, which are
698
/// loops within the subgraph of L that exclude its header.  (A "trivial" SCC
699
/// consists of a single block that does not have a self-edge.)
700
///
701
/// In addition to loops, this algorithm has limited support for irreducible
702
/// SCCs, which are SCCs with multiple entry blocks.  Irreducible SCCs are
703
/// discovered on they fly, and modelled as loops with multiple headers.
704
///
705
/// The headers of irreducible sub-SCCs consist of its entry blocks and all
706
/// nodes that are targets of a backedge within it (excluding backedges within
707
/// true sub-loops).  Block frequency calculations act as if a block is
708
/// inserted that intercepts all the edges to the headers.  All backedges and
709
/// entries point to this block.  Its successors are the headers, which split
710
/// the frequency evenly.
711
///
712
/// This algorithm leverages BlockMass and ScaledNumber to maintain precision,
713
/// separates mass distribution from loop scaling, and dithers to eliminate
714
/// probability mass loss.
715
///
716
/// The implementation is split between BlockFrequencyInfoImpl, which knows the
717
/// type of graph being modelled (BasicBlock vs. MachineBasicBlock), and
718
/// BlockFrequencyInfoImplBase, which doesn't.  The base class uses \a
719
/// BlockNode, a wrapper around a uint32_t.  BlockNode is numbered from 0 in
720
/// reverse-post order.  This gives two advantages:  it's easy to compare the
721
/// relative ordering of two nodes, and maps keyed on BlockT can be represented
722
/// by vectors.
723
///
724
/// This algorithm is O(V+E), unless there is irreducible control flow, in
725
/// which case it's O(V*E) in the worst case.
726
///
727
/// These are the main stages:
728
///
729
///  0. Reverse post-order traversal (\a initializeRPOT()).
730
///
731
///     Run a single post-order traversal and save it (in reverse) in RPOT.
732
///     All other stages make use of this ordering.  Save a lookup from BlockT
733
///     to BlockNode (the index into RPOT) in Nodes.
734
///
735
///  1. Loop initialization (\a initializeLoops()).
736
///
737
///     Translate LoopInfo/MachineLoopInfo into a form suitable for the rest of
738
///     the algorithm.  In particular, store the immediate members of each loop
739
///     in reverse post-order.
740
///
741
///  2. Calculate mass and scale in loops (\a computeMassInLoops()).
742
///
743
///     For each loop (bottom-up), distribute mass through the DAG resulting
744
///     from ignoring backedges and treating sub-loops as a single pseudo-node.
745
///     Track the backedge mass distributed to the loop header, and use it to
746
///     calculate the loop scale (number of loop iterations).  Immediate
747
///     members that represent sub-loops will already have been visited and
748
///     packaged into a pseudo-node.
749
///
750
///     Distributing mass in a loop is a reverse-post-order traversal through
751
///     the loop.  Start by assigning full mass to the Loop header.  For each
752
///     node in the loop:
753
///
754
///         - Fetch and categorize the weight distribution for its successors.
755
///           If this is a packaged-subloop, the weight distribution is stored
756
///           in \a LoopData::Exits.  Otherwise, fetch it from
757
///           BranchProbabilityInfo.
758
///
759
///         - Each successor is categorized as \a Weight::Local, a local edge
760
///           within the current loop, \a Weight::Backedge, a backedge to the
761
///           loop header, or \a Weight::Exit, any successor outside the loop.
762
///           The weight, the successor, and its category are stored in \a
763
///           Distribution.  There can be multiple edges to each successor.
764
///
765
///         - If there's a backedge to a non-header, there's an irreducible SCC.
766
///           The usual flow is temporarily aborted.  \a
767
///           computeIrreducibleMass() finds the irreducible SCCs within the
768
///           loop, packages them up, and restarts the flow.
769
///
770
///         - Normalize the distribution:  scale weights down so that their sum
771
///           is 32-bits, and coalesce multiple edges to the same node.
772
///
773
///         - Distribute the mass accordingly, dithering to minimize mass loss,
774
///           as described in \a distributeMass().
775
///
776
///     In the case of irreducible loops, instead of a single loop header,
777
///     there will be several. The computation of backedge masses is similar
778
///     but instead of having a single backedge mass, there will be one
779
///     backedge per loop header. In these cases, each backedge will carry
780
///     a mass proportional to the edge weights along the corresponding
781
///     path.
782
///
783
///     At the end of propagation, the full mass assigned to the loop will be
784
///     distributed among the loop headers proportionally according to the
785
///     mass flowing through their backedges.
786
///
787
///     Finally, calculate the loop scale from the accumulated backedge mass.
788
///
789
///  3. Distribute mass in the function (\a computeMassInFunction()).
790
///
791
///     Finally, distribute mass through the DAG resulting from packaging all
792
///     loops in the function.  This uses the same algorithm as distributing
793
///     mass in a loop, except that there are no exit or backedge edges.
794
///
795
///  4. Unpackage loops (\a unwrapLoops()).
796
///
797
///     Initialize each block's frequency to a floating point representation of
798
///     its mass.
799
///
800
///     Visit loops top-down, scaling the frequencies of its immediate members
801
///     by the loop's pseudo-node's frequency.
802
///
803
///  5. Convert frequencies to a 64-bit range (\a finalizeMetrics()).
804
///
805
///     Using the min and max frequencies as a guide, translate floating point
806
///     frequencies to an appropriate range in uint64_t.
807
///
808
/// It has some known flaws.
809
///
810
///   - The model of irreducible control flow is a rough approximation.
811
///
812
///     Modelling irreducible control flow exactly involves setting up and
813
///     solving a group of infinite geometric series.  Such precision is
814
///     unlikely to be worthwhile, since most of our algorithms give up on
815
///     irreducible control flow anyway.
816
///
817
///     Nevertheless, we might find that we need to get closer.  Here's a sort
818
///     of TODO list for the model with diminishing returns, to be completed as
819
///     necessary.
820
///
821
///       - The headers for the \a LoopData representing an irreducible SCC
822
///         include non-entry blocks.  When these extra blocks exist, they
823
///         indicate a self-contained irreducible sub-SCC.  We could treat them
824
///         as sub-loops, rather than arbitrarily shoving the problematic
825
///         blocks into the headers of the main irreducible SCC.
826
///
827
///       - Entry frequencies are assumed to be evenly split between the
828
///         headers of a given irreducible SCC, which is the only option if we
829
///         need to compute mass in the SCC before its parent loop.  Instead,
830
///         we could partially compute mass in the parent loop, and stop when
831
///         we get to the SCC.  Here, we have the correct ratio of entry
832
///         masses, which we can use to adjust their relative frequencies.
833
///         Compute mass in the SCC, and then continue propagation in the
834
///         parent.
835
///
836
///       - We can propagate mass iteratively through the SCC, for some fixed
837
///         number of iterations.  Each iteration starts by assigning the entry
838
///         blocks their backedge mass from the prior iteration.  The final
839
///         mass for each block (and each exit, and the total backedge mass
840
///         used for computing loop scale) is the sum of all iterations.
841
///         (Running this until fixed point would "solve" the geometric
842
///         series by simulation.)
843
template <class BT> class BlockFrequencyInfoImpl : BlockFrequencyInfoImplBase {
844
  // This is part of a workaround for a GCC 4.7 crash on lambdas.
845
  friend struct bfi_detail::BlockEdgesAdder<BT>;
846
847
  using BlockT = typename bfi_detail::TypeMap<BT>::BlockT;
848
  using FunctionT = typename bfi_detail::TypeMap<BT>::FunctionT;
849
  using BranchProbabilityInfoT =
850
      typename bfi_detail::TypeMap<BT>::BranchProbabilityInfoT;
851
  using LoopT = typename bfi_detail::TypeMap<BT>::LoopT;
852
  using LoopInfoT = typename bfi_detail::TypeMap<BT>::LoopInfoT;
853
  using Successor = GraphTraits<const BlockT *>;
854
  using Predecessor = GraphTraits<Inverse<const BlockT *>>;
855
856
  const BranchProbabilityInfoT *BPI = nullptr;
857
  const LoopInfoT *LI = nullptr;
858
  const FunctionT *F = nullptr;
859
860
  // All blocks in reverse postorder.
861
  std::vector<const BlockT *> RPOT;
862
  DenseMap<const BlockT *, BlockNode> Nodes;
863
864
  using rpot_iterator = typename std::vector<const BlockT *>::const_iterator;
865
866
62.5M
  rpot_iterator rpot_begin() const { return RPOT.begin(); }
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::rpot_begin() const
Line
Count
Source
866
36.4M
  rpot_iterator rpot_begin() const { return RPOT.begin(); }
llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::rpot_begin() const
Line
Count
Source
866
26.1M
  rpot_iterator rpot_begin() const { return RPOT.begin(); }
867
8.23M
  rpot_iterator rpot_end() const { return RPOT.end(); }
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::rpot_end() const
Line
Count
Source
867
4.49M
  rpot_iterator rpot_end() const { return RPOT.end(); }
llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::rpot_end() const
Line
Count
Source
867
3.74M
  rpot_iterator rpot_end() const { return RPOT.end(); }
868
869
54.2M
  size_t getIndex(const rpot_iterator &I) const { return I - rpot_begin(); }
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::getIndex(std::__1::__wrap_iter<llvm::BasicBlock const* const*> const&) const
Line
Count
Source
869
31.9M
  size_t getIndex(const rpot_iterator &I) const { return I - rpot_begin(); }
llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::getIndex(std::__1::__wrap_iter<llvm::MachineBasicBlock const* const*> const&) const
Line
Count
Source
869
22.3M
  size_t getIndex(const rpot_iterator &I) const { return I - rpot_begin(); }
870
871
54.2M
  BlockNode getNode(const rpot_iterator &I) const {
872
54.2M
    return BlockNode(getIndex(I));
873
54.2M
  }
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::getNode(std::__1::__wrap_iter<llvm::BasicBlock const* const*> const&) const
Line
Count
Source
871
31.9M
  BlockNode getNode(const rpot_iterator &I) const {
872
31.9M
    return BlockNode(getIndex(I));
873
31.9M
  }
llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::getNode(std::__1::__wrap_iter<llvm::MachineBasicBlock const* const*> const&) const
Line
Count
Source
871
22.3M
  BlockNode getNode(const rpot_iterator &I) const {
872
22.3M
    return BlockNode(getIndex(I));
873
22.3M
  }
874
88.6M
  BlockNode getNode(const BlockT *BB) const { return Nodes.lookup(BB); }
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::getNode(llvm::BasicBlock const*) const
Line
Count
Source
874
28.4M
  BlockNode getNode(const BlockT *BB) const { return Nodes.lookup(BB); }
llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::getNode(llvm::MachineBasicBlock const*) const
Line
Count
Source
874
60.2M
  BlockNode getNode(const BlockT *BB) const { return Nodes.lookup(BB); }
875
876
27.1M
  const BlockT *getBlock(const BlockNode &Node) const {
877
27.1M
    assert(Node.Index < RPOT.size());
878
27.1M
    return RPOT[Node.Index];
879
27.1M
  }
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::getBlock(llvm::BlockFrequencyInfoImplBase::BlockNode const&) const
Line
Count
Source
876
15.9M
  const BlockT *getBlock(const BlockNode &Node) const {
877
15.9M
    assert(Node.Index < RPOT.size());
878
15.9M
    return RPOT[Node.Index];
879
15.9M
  }
llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::getBlock(llvm::BlockFrequencyInfoImplBase::BlockNode const&) const
Line
Count
Source
876
11.1M
  const BlockT *getBlock(const BlockNode &Node) const {
877
11.1M
    assert(Node.Index < RPOT.size());
878
11.1M
    return RPOT[Node.Index];
879
11.1M
  }
880
881
  /// Run (and save) a post-order traversal.
882
  ///
883
  /// Saves a reverse post-order traversal of all the nodes in \a F.
884
  void initializeRPOT();
885
886
  /// Initialize loop data.
887
  ///
888
  /// Build up \a Loops using \a LoopInfo.  \a LoopInfo gives us a mapping from
889
  /// each block to the deepest loop it's in, but we need the inverse.  For each
890
  /// loop, we store in reverse post-order its "immediate" members, defined as
891
  /// the header, the headers of immediate sub-loops, and all other blocks in
892
  /// the loop that are not in sub-loops.
893
  void initializeLoops();
894
895
  /// Propagate to a block's successors.
896
  ///
897
  /// In the context of distributing mass through \c OuterLoop, divide the mass
898
  /// currently assigned to \c Node between its successors.
899
  ///
900
  /// \return \c true unless there's an irreducible backedge.
901
  bool propagateMassToSuccessors(LoopData *OuterLoop, const BlockNode &Node);
902
903
  /// Compute mass in a particular loop.
904
  ///
905
  /// Assign mass to \c Loop's header, and then for each block in \c Loop in
906
  /// reverse post-order, distribute mass to its successors.  Only visits nodes
907
  /// that have not been packaged into sub-loops.
908
  ///
909
  /// \pre \a computeMassInLoop() has been called for each subloop of \c Loop.
910
  /// \return \c true unless there's an irreducible backedge.
911
  bool computeMassInLoop(LoopData &Loop);
912
913
  /// Try to compute mass in the top-level function.
914
  ///
915
  /// Assign mass to the entry block, and then for each block in reverse
916
  /// post-order, distribute mass to its successors.  Skips nodes that have
917
  /// been packaged into loops.
918
  ///
919
  /// \pre \a computeMassInLoops() has been called.
920
  /// \return \c true unless there's an irreducible backedge.
921
  bool tryToComputeMassInFunction();
922
923
  /// Compute mass in (and package up) irreducible SCCs.
924
  ///
925
  /// Find the irreducible SCCs in \c OuterLoop, add them to \a Loops (in front
926
  /// of \c Insert), and call \a computeMassInLoop() on each of them.
927
  ///
928
  /// If \c OuterLoop is \c nullptr, it refers to the top-level function.
929
  ///
930
  /// \pre \a computeMassInLoop() has been called for each subloop of \c
931
  /// OuterLoop.
932
  /// \pre \c Insert points at the last loop successfully processed by \a
933
  /// computeMassInLoop().
934
  /// \pre \c OuterLoop has irreducible SCCs.
935
  void computeIrreducibleMass(LoopData *OuterLoop,
936
                              std::list<LoopData>::iterator Insert);
937
938
  /// Compute mass in all loops.
939
  ///
940
  /// For each loop bottom-up, call \a computeMassInLoop().
941
  ///
942
  /// \a computeMassInLoop() aborts (and returns \c false) on loops that
943
  /// contain a irreducible sub-SCCs.  Use \a computeIrreducibleMass() and then
944
  /// re-enter \a computeMassInLoop().
945
  ///
946
  /// \post \a computeMassInLoop() has returned \c true for every loop.
947
  void computeMassInLoops();
948
949
  /// Compute mass in the top-level function.
950
  ///
951
  /// Uses \a tryToComputeMassInFunction() and \a computeIrreducibleMass() to
952
  /// compute mass in the top-level function.
953
  ///
954
  /// \post \a tryToComputeMassInFunction() has returned \c true.
955
  void computeMassInFunction();
956
957
0
  std::string getBlockName(const BlockNode &Node) const override {
958
0
    return bfi_detail::getBlockName(getBlock(Node));
959
0
  }
Unexecuted instantiation: llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::getBlockName(llvm::BlockFrequencyInfoImplBase::BlockNode const&) const
Unexecuted instantiation: llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::getBlockName(llvm::BlockFrequencyInfoImplBase::BlockNode const&) const
960
961
public:
962
4.11M
  BlockFrequencyInfoImpl() = default;
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::BlockFrequencyInfoImpl()
Line
Count
Source
962
2.24M
  BlockFrequencyInfoImpl() = default;
llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::BlockFrequencyInfoImpl()
Line
Count
Source
962
1.87M
  BlockFrequencyInfoImpl() = default;
963
964
1.04k
  const FunctionT *getFunction() const { return F; }
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::getFunction() const
Line
Count
Source
964
991
  const FunctionT *getFunction() const { return F; }
llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::getFunction() const
Line
Count
Source
964
58
  const FunctionT *getFunction() const { return F; }
965
966
  void calculate(const FunctionT &F, const BranchProbabilityInfoT &BPI,
967
                 const LoopInfoT &LI);
968
969
  using BlockFrequencyInfoImplBase::getEntryFreq;
970
971
46.4M
  BlockFrequency getBlockFreq(const BlockT *BB) const {
972
46.4M
    return BlockFrequencyInfoImplBase::getBlockFreq(getNode(BB));
973
46.4M
  }
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::getBlockFreq(llvm::BasicBlock const*) const
Line
Count
Source
971
2.93M
  BlockFrequency getBlockFreq(const BlockT *BB) const {
972
2.93M
    return BlockFrequencyInfoImplBase::getBlockFreq(getNode(BB));
973
2.93M
  }
llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::getBlockFreq(llvm::MachineBasicBlock const*) const
Line
Count
Source
971
43.4M
  BlockFrequency getBlockFreq(const BlockT *BB) const {
972
43.4M
    return BlockFrequencyInfoImplBase::getBlockFreq(getNode(BB));
973
43.4M
  }
974
975
  Optional<uint64_t> getBlockProfileCount(const Function &F,
976
957
                                          const BlockT *BB) const {
977
957
    return BlockFrequencyInfoImplBase::getBlockProfileCount(F, getNode(BB));
978
957
  }
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::getBlockProfileCount(llvm::Function const&, llvm::BasicBlock const*) const
Line
Count
Source
976
899
                                          const BlockT *BB) const {
977
899
    return BlockFrequencyInfoImplBase::getBlockProfileCount(F, getNode(BB));
978
899
  }
llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::getBlockProfileCount(llvm::Function const&, llvm::MachineBasicBlock const*) const
Line
Count
Source
976
58
                                          const BlockT *BB) const {
977
58
    return BlockFrequencyInfoImplBase::getBlockProfileCount(F, getNode(BB));
978
58
  }
979
980
  Optional<uint64_t> getProfileCountFromFreq(const Function &F,
981
92
                                             uint64_t Freq) const {
982
92
    return BlockFrequencyInfoImplBase::getProfileCountFromFreq(F, Freq);
983
92
  }
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::getProfileCountFromFreq(llvm::Function const&, unsigned long long) const
Line
Count
Source
981
92
                                             uint64_t Freq) const {
982
92
    return BlockFrequencyInfoImplBase::getProfileCountFromFreq(F, Freq);
983
92
  }
Unexecuted instantiation: llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::getProfileCountFromFreq(llvm::Function const&, unsigned long long) const
984
985
265
  bool isIrrLoopHeader(const BlockT *BB) {
986
265
    return BlockFrequencyInfoImplBase::isIrrLoopHeader(getNode(BB));
987
265
  }
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::isIrrLoopHeader(llvm::BasicBlock const*)
Line
Count
Source
985
265
  bool isIrrLoopHeader(const BlockT *BB) {
986
265
    return BlockFrequencyInfoImplBase::isIrrLoopHeader(getNode(BB));
987
265
  }
Unexecuted instantiation: llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::isIrrLoopHeader(llvm::MachineBasicBlock const*)
988
989
  void setBlockFreq(const BlockT *BB, uint64_t Freq);
990
991
532
  Scaled64 getFloatingBlockFreq(const BlockT *BB) const {
992
532
    return BlockFrequencyInfoImplBase::getFloatingBlockFreq(getNode(BB));
993
532
  }
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::getFloatingBlockFreq(llvm::BasicBlock const*) const
Line
Count
Source
991
532
  Scaled64 getFloatingBlockFreq(const BlockT *BB) const {
992
532
    return BlockFrequencyInfoImplBase::getFloatingBlockFreq(getNode(BB));
993
532
  }
Unexecuted instantiation: llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::getFloatingBlockFreq(llvm::MachineBasicBlock const*) const
994
995
0
  const BranchProbabilityInfoT &getBPI() const { return *BPI; }
Unexecuted instantiation: llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::getBPI() const
Unexecuted instantiation: llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::getBPI() const
996
997
  /// Print the frequencies for the current function.
998
  ///
999
  /// Prints the frequencies for the blocks in the current function.
1000
  ///
1001
  /// Blocks are printed in the natural iteration order of the function, rather
1002
  /// than reverse post-order.  This provides two advantages:  writing -analyze
1003
  /// tests is easier (since blocks come out in source order), and even
1004
  /// unreachable blocks are printed.
1005
  ///
1006
  /// \a BlockFrequencyInfoImplBase::print() only knows reverse post-order, so
1007
  /// we need to override it here.
1008
  raw_ostream &print(raw_ostream &OS) const override;
1009
1010
  using BlockFrequencyInfoImplBase::dump;
1011
  using BlockFrequencyInfoImplBase::printBlockFreq;
1012
1013
0
  raw_ostream &printBlockFreq(raw_ostream &OS, const BlockT *BB) const {
1014
0
    return BlockFrequencyInfoImplBase::printBlockFreq(OS, getNode(BB));
1015
0
  }
Unexecuted instantiation: llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::printBlockFreq(llvm::raw_ostream&, llvm::BasicBlock const*) const
Unexecuted instantiation: llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::printBlockFreq(llvm::raw_ostream&, llvm::MachineBasicBlock const*) const
1016
};
1017
1018
template <class BT>
1019
void BlockFrequencyInfoImpl<BT>::calculate(const FunctionT &F,
1020
                                           const BranchProbabilityInfoT &BPI,
1021
4.11M
                                           const LoopInfoT &LI) {
1022
4.11M
  // Save the parameters.
1023
4.11M
  this->BPI = &BPI;
1024
4.11M
  this->LI = &LI;
1025
4.11M
  this->F = &F;
1026
4.11M
1027
4.11M
  // Clean up left-over data structures.
1028
4.11M
  BlockFrequencyInfoImplBase::clear();
1029
4.11M
  RPOT.clear();
1030
4.11M
  Nodes.clear();
1031
4.11M
1032
4.11M
  // Initialize.
1033
4.11M
  LLVM_DEBUG(dbgs() << "\nblock-frequency: " << F.getName()
1034
4.11M
                    << "\n================="
1035
4.11M
                    << std::string(F.getName().size(), '=') << "\n");
1036
4.11M
  initializeRPOT();
1037
4.11M
  initializeLoops();
1038
4.11M
1039
4.11M
  // Visit loops in post-order to find the local mass distribution, and then do
1040
4.11M
  // the full function.
1041
4.11M
  computeMassInLoops();
1042
4.11M
  computeMassInFunction();
1043
4.11M
  unwrapLoops();
1044
4.11M
  finalizeMetrics();
1045
4.11M
}
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::calculate(llvm::Function const&, llvm::BranchProbabilityInfo const&, llvm::LoopInfo const&)
Line
Count
Source
1021
2.24M
                                           const LoopInfoT &LI) {
1022
2.24M
  // Save the parameters.
1023
2.24M
  this->BPI = &BPI;
1024
2.24M
  this->LI = &LI;
1025
2.24M
  this->F = &F;
1026
2.24M
1027
2.24M
  // Clean up left-over data structures.
1028
2.24M
  BlockFrequencyInfoImplBase::clear();
1029
2.24M
  RPOT.clear();
1030
2.24M
  Nodes.clear();
1031
2.24M
1032
2.24M
  // Initialize.
1033
2.24M
  LLVM_DEBUG(dbgs() << "\nblock-frequency: " << F.getName()
1034
2.24M
                    << "\n================="
1035
2.24M
                    << std::string(F.getName().size(), '=') << "\n");
1036
2.24M
  initializeRPOT();
1037
2.24M
  initializeLoops();
1038
2.24M
1039
2.24M
  // Visit loops in post-order to find the local mass distribution, and then do
1040
2.24M
  // the full function.
1041
2.24M
  computeMassInLoops();
1042
2.24M
  computeMassInFunction();
1043
2.24M
  unwrapLoops();
1044
2.24M
  finalizeMetrics();
1045
2.24M
}
llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::calculate(llvm::MachineFunction const&, llvm::MachineBranchProbabilityInfo const&, llvm::MachineLoopInfo const&)
Line
Count
Source
1021
1.87M
                                           const LoopInfoT &LI) {
1022
1.87M
  // Save the parameters.
1023
1.87M
  this->BPI = &BPI;
1024
1.87M
  this->LI = &LI;
1025
1.87M
  this->F = &F;
1026
1.87M
1027
1.87M
  // Clean up left-over data structures.
1028
1.87M
  BlockFrequencyInfoImplBase::clear();
1029
1.87M
  RPOT.clear();
1030
1.87M
  Nodes.clear();
1031
1.87M
1032
1.87M
  // Initialize.
1033
1.87M
  LLVM_DEBUG(dbgs() << "\nblock-frequency: " << F.getName()
1034
1.87M
                    << "\n================="
1035
1.87M
                    << std::string(F.getName().size(), '=') << "\n");
1036
1.87M
  initializeRPOT();
1037
1.87M
  initializeLoops();
1038
1.87M
1039
1.87M
  // Visit loops in post-order to find the local mass distribution, and then do
1040
1.87M
  // the full function.
1041
1.87M
  computeMassInLoops();
1042
1.87M
  computeMassInFunction();
1043
1.87M
  unwrapLoops();
1044
1.87M
  finalizeMetrics();
1045
1.87M
}
1046
1047
template <class BT>
1048
2.12k
void BlockFrequencyInfoImpl<BT>::setBlockFreq(const BlockT *BB, uint64_t Freq) {
1049
2.12k
  if (Nodes.count(BB))
1050
1.21k
    BlockFrequencyInfoImplBase::setBlockFreq(getNode(BB), Freq);
1051
911
  else {
1052
911
    // If BB is a newly added block after BFI is done, we need to create a new
1053
911
    // BlockNode for it assigned with a new index. The index can be determined
1054
911
    // by the size of Freqs.
1055
911
    BlockNode NewNode(Freqs.size());
1056
911
    Nodes[BB] = NewNode;
1057
911
    Freqs.emplace_back();
1058
911
    BlockFrequencyInfoImplBase::setBlockFreq(NewNode, Freq);
1059
911
  }
1060
2.12k
}
1061
1062
4.11M
template <class BT> void BlockFrequencyInfoImpl<BT>::initializeRPOT() {
1063
4.11M
  const BlockT *Entry = &F->front();
1064
4.11M
  RPOT.reserve(F->size());
1065
4.11M
  std::copy(po_begin(Entry), po_end(Entry), std::back_inserter(RPOT));
1066
4.11M
  std::reverse(RPOT.begin(), RPOT.end());
1067
4.11M
1068
4.11M
  assert(RPOT.size() - 1 <= BlockNode::getMaxIndex() &&
1069
4.11M
         "More nodes in function than Block Frequency Info supports");
1070
4.11M
1071
4.11M
  LLVM_DEBUG(dbgs() << "reverse-post-order-traversal\n");
1072
31.2M
  for (rpot_iterator I = rpot_begin(), E = rpot_end(); I != E; 
++I27.1M
) {
1073
27.1M
    BlockNode Node = getNode(I);
1074
27.1M
    LLVM_DEBUG(dbgs() << " - " << getIndex(I) << ": " << getBlockName(Node)
1075
27.1M
                      << "\n");
1076
27.1M
    Nodes[*I] = Node;
1077
27.1M
  }
1078
4.11M
1079
4.11M
  Working.reserve(RPOT.size());
1080
31.2M
  for (size_t Index = 0; Index < RPOT.size(); 
++Index27.1M
)
1081
27.1M
    Working.emplace_back(Index);
1082
4.11M
  Freqs.resize(RPOT.size());
1083
4.11M
}
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::initializeRPOT()
Line
Count
Source
1062
2.24M
template <class BT> void BlockFrequencyInfoImpl<BT>::initializeRPOT() {
1063
2.24M
  const BlockT *Entry = &F->front();
1064
2.24M
  RPOT.reserve(F->size());
1065
2.24M
  std::copy(po_begin(Entry), po_end(Entry), std::back_inserter(RPOT));
1066
2.24M
  std::reverse(RPOT.begin(), RPOT.end());
1067
2.24M
1068
2.24M
  assert(RPOT.size() - 1 <= BlockNode::getMaxIndex() &&
1069
2.24M
         "More nodes in function than Block Frequency Info supports");
1070
2.24M
1071
2.24M
  LLVM_DEBUG(dbgs() << "reverse-post-order-traversal\n");
1072
18.2M
  for (rpot_iterator I = rpot_begin(), E = rpot_end(); I != E; 
++I15.9M
) {
1073
15.9M
    BlockNode Node = getNode(I);
1074
15.9M
    LLVM_DEBUG(dbgs() << " - " << getIndex(I) << ": " << getBlockName(Node)
1075
15.9M
                      << "\n");
1076
15.9M
    Nodes[*I] = Node;
1077
15.9M
  }
1078
2.24M
1079
2.24M
  Working.reserve(RPOT.size());
1080
18.2M
  for (size_t Index = 0; Index < RPOT.size(); 
++Index15.9M
)
1081
15.9M
    Working.emplace_back(Index);
1082
2.24M
  Freqs.resize(RPOT.size());
1083
2.24M
}
llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::initializeRPOT()
Line
Count
Source
1062
1.87M
template <class BT> void BlockFrequencyInfoImpl<BT>::initializeRPOT() {
1063
1.87M
  const BlockT *Entry = &F->front();
1064
1.87M
  RPOT.reserve(F->size());
1065
1.87M
  std::copy(po_begin(Entry), po_end(Entry), std::back_inserter(RPOT));
1066
1.87M
  std::reverse(RPOT.begin(), RPOT.end());
1067
1.87M
1068
1.87M
  assert(RPOT.size() - 1 <= BlockNode::getMaxIndex() &&
1069
1.87M
         "More nodes in function than Block Frequency Info supports");
1070
1.87M
1071
1.87M
  LLVM_DEBUG(dbgs() << "reverse-post-order-traversal\n");
1072
13.0M
  for (rpot_iterator I = rpot_begin(), E = rpot_end(); I != E; 
++I11.1M
) {
1073
11.1M
    BlockNode Node = getNode(I);
1074
11.1M
    LLVM_DEBUG(dbgs() << " - " << getIndex(I) << ": " << getBlockName(Node)
1075
11.1M
                      << "\n");
1076
11.1M
    Nodes[*I] = Node;
1077
11.1M
  }
1078
1.87M
1079
1.87M
  Working.reserve(RPOT.size());
1080
13.0M
  for (size_t Index = 0; Index < RPOT.size(); 
++Index11.1M
)
1081
11.1M
    Working.emplace_back(Index);
1082
1.87M
  Freqs.resize(RPOT.size());
1083
1.87M
}
1084
1085
4.11M
template <class BT> void BlockFrequencyInfoImpl<BT>::initializeLoops() {
1086
4.11M
  LLVM_DEBUG(dbgs() << "loop-detection\n");
1087
4.11M
  if (LI->empty())
1088
3.44M
    return;
1089
677k
1090
677k
  // Visit loops top down and assign them an index.
1091
677k
  std::deque<std::pair<const LoopT *, LoopData *>> Q;
1092
677k
  for (const LoopT *L : *LI)
1093
1.49M
    Q.emplace_back(L, nullptr);
1094
2.71M
  while (!Q.empty()) {
1095
2.03M
    const LoopT *Loop = Q.front().first;
1096
2.03M
    LoopData *Parent = Q.front().second;
1097
2.03M
    Q.pop_front();
1098
2.03M
1099
2.03M
    BlockNode Header = getNode(Loop->getHeader());
1100
2.03M
    assert(Header.isValid());
1101
2.03M
1102
2.03M
    Loops.emplace_back(Parent, Header);
1103
2.03M
    Working[Header.Index].Loop = &Loops.back();
1104
2.03M
    LLVM_DEBUG(dbgs() << " - loop = " << getBlockName(Header) << "\n");
1105
2.03M
1106
2.03M
    for (const LoopT *L : *Loop)
1107
544k
      Q.emplace_back(L, &Loops.back());
1108
2.03M
  }
1109
677k
1110
677k
  // Visit nodes in reverse post-order and add them to their deepest containing
1111
677k
  // loop.
1112
17.9M
  for (size_t Index = 0; Index < RPOT.size(); 
++Index17.2M
) {
1113
17.2M
    // Loop headers have already been mostly mapped.
1114
17.2M
    if (Working[Index].isLoopHeader()) {
1115
2.03M
      LoopData *ContainingLoop = Working[Index].getContainingLoop();
1116
2.03M
      if (ContainingLoop)
1117
544k
        ContainingLoop->Nodes.push_back(Index);
1118
2.03M
      continue;
1119
2.03M
    }
1120
15.2M
1121
15.2M
    const LoopT *Loop = LI->getLoopFor(RPOT[Index]);
1122
15.2M
    if (!Loop)
1123
10.3M
      continue;
1124
4.83M
1125
4.83M
    // Add this node to its containing loop's member list.
1126
4.83M
    BlockNode Header = getNode(Loop->getHeader());
1127
4.83M
    assert(Header.isValid());
1128
4.83M
    const auto &HeaderData = Working[Header.Index];
1129
4.83M
    assert(HeaderData.isLoopHeader());
1130
4.83M
1131
4.83M
    Working[Index].Loop = HeaderData.Loop;
1132
4.83M
    HeaderData.Loop->Nodes.push_back(Index);
1133
4.83M
    LLVM_DEBUG(dbgs() << " - loop = " << getBlockName(Header)
1134
4.83M
                      << ": member = " << getBlockName(Index) << "\n");
1135
4.83M
  }
1136
677k
}
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::initializeLoops()
Line
Count
Source
1085
2.24M
template <class BT> void BlockFrequencyInfoImpl<BT>::initializeLoops() {
1086
2.24M
  LLVM_DEBUG(dbgs() << "loop-detection\n");
1087
2.24M
  if (LI->empty())
1088
1.84M
    return;
1089
402k
1090
402k
  // Visit loops top down and assign them an index.
1091
402k
  std::deque<std::pair<const LoopT *, LoopData *>> Q;
1092
402k
  for (const LoopT *L : *LI)
1093
878k
    Q.emplace_back(L, nullptr);
1094
1.60M
  while (!Q.empty()) {
1095
1.20M
    const LoopT *Loop = Q.front().first;
1096
1.20M
    LoopData *Parent = Q.front().second;
1097
1.20M
    Q.pop_front();
1098
1.20M
1099
1.20M
    BlockNode Header = getNode(Loop->getHeader());
1100
1.20M
    assert(Header.isValid());
1101
1.20M
1102
1.20M
    Loops.emplace_back(Parent, Header);
1103
1.20M
    Working[Header.Index].Loop = &Loops.back();
1104
1.20M
    LLVM_DEBUG(dbgs() << " - loop = " << getBlockName(Header) << "\n");
1105
1.20M
1106
1.20M
    for (const LoopT *L : *Loop)
1107
324k
      Q.emplace_back(L, &Loops.back());
1108
1.20M
  }
1109
402k
1110
402k
  // Visit nodes in reverse post-order and add them to their deepest containing
1111
402k
  // loop.
1112
10.6M
  for (size_t Index = 0; Index < RPOT.size(); 
++Index10.2M
) {
1113
10.2M
    // Loop headers have already been mostly mapped.
1114
10.2M
    if (Working[Index].isLoopHeader()) {
1115
1.20M
      LoopData *ContainingLoop = Working[Index].getContainingLoop();
1116
1.20M
      if (ContainingLoop)
1117
324k
        ContainingLoop->Nodes.push_back(Index);
1118
1.20M
      continue;
1119
1.20M
    }
1120
9.01M
1121
9.01M
    const LoopT *Loop = LI->getLoopFor(RPOT[Index]);
1122
9.01M
    if (!Loop)
1123
6.09M
      continue;
1124
2.91M
1125
2.91M
    // Add this node to its containing loop's member list.
1126
2.91M
    BlockNode Header = getNode(Loop->getHeader());
1127
2.91M
    assert(Header.isValid());
1128
2.91M
    const auto &HeaderData = Working[Header.Index];
1129
2.91M
    assert(HeaderData.isLoopHeader());
1130
2.91M
1131
2.91M
    Working[Index].Loop = HeaderData.Loop;
1132
2.91M
    HeaderData.Loop->Nodes.push_back(Index);
1133
2.91M
    LLVM_DEBUG(dbgs() << " - loop = " << getBlockName(Header)
1134
2.91M
                      << ": member = " << getBlockName(Index) << "\n");
1135
2.91M
  }
1136
402k
}
llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::initializeLoops()
Line
Count
Source
1085
1.87M
template <class BT> void BlockFrequencyInfoImpl<BT>::initializeLoops() {
1086
1.87M
  LLVM_DEBUG(dbgs() << "loop-detection\n");
1087
1.87M
  if (LI->empty())
1088
1.59M
    return;
1089
274k
1090
274k
  // Visit loops top down and assign them an index.
1091
274k
  std::deque<std::pair<const LoopT *, LoopData *>> Q;
1092
274k
  for (const LoopT *L : *LI)
1093
616k
    Q.emplace_back(L, nullptr);
1094
1.11M
  while (!Q.empty()) {
1095
836k
    const LoopT *Loop = Q.front().first;
1096
836k
    LoopData *Parent = Q.front().second;
1097
836k
    Q.pop_front();
1098
836k
1099
836k
    BlockNode Header = getNode(Loop->getHeader());
1100
836k
    assert(Header.isValid());
1101
836k
1102
836k
    Loops.emplace_back(Parent, Header);
1103
836k
    Working[Header.Index].Loop = &Loops.back();
1104
836k
    LLVM_DEBUG(dbgs() << " - loop = " << getBlockName(Header) << "\n");
1105
836k
1106
836k
    for (const LoopT *L : *Loop)
1107
219k
      Q.emplace_back(L, &Loops.back());
1108
836k
  }
1109
274k
1110
274k
  // Visit nodes in reverse post-order and add them to their deepest containing
1111
274k
  // loop.
1112
7.31M
  for (size_t Index = 0; Index < RPOT.size(); 
++Index7.03M
) {
1113
7.03M
    // Loop headers have already been mostly mapped.
1114
7.03M
    if (Working[Index].isLoopHeader()) {
1115
836k
      LoopData *ContainingLoop = Working[Index].getContainingLoop();
1116
836k
      if (ContainingLoop)
1117
219k
        ContainingLoop->Nodes.push_back(Index);
1118
836k
      continue;
1119
836k
    }
1120
6.20M
1121
6.20M
    const LoopT *Loop = LI->getLoopFor(RPOT[Index]);
1122
6.20M
    if (!Loop)
1123
4.28M
      continue;
1124
1.91M
1125
1.91M
    // Add this node to its containing loop's member list.
1126
1.91M
    BlockNode Header = getNode(Loop->getHeader());
1127
1.91M
    assert(Header.isValid());
1128
1.91M
    const auto &HeaderData = Working[Header.Index];
1129
1.91M
    assert(HeaderData.isLoopHeader());
1130
1.91M
1131
1.91M
    Working[Index].Loop = HeaderData.Loop;
1132
1.91M
    HeaderData.Loop->Nodes.push_back(Index);
1133
1.91M
    LLVM_DEBUG(dbgs() << " - loop = " << getBlockName(Header)
1134
1.91M
                      << ": member = " << getBlockName(Index) << "\n");
1135
1.91M
  }
1136
274k
}
1137
1138
4.11M
template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInLoops() {
1139
4.11M
  // Visit loops with the deepest first, and the top-level loops last.
1140
6.15M
  for (auto L = Loops.rbegin(), E = Loops.rend(); L != E; 
++L2.03M
) {
1141
2.03M
    if (computeMassInLoop(*L))
1142
2.03M
      continue;
1143
141
    auto Next = std::next(L);
1144
141
    computeIrreducibleMass(&*L, L.base());
1145
141
    L = std::prev(Next);
1146
141
    if (computeMassInLoop(*L))
1147
141
      continue;
1148
0
    llvm_unreachable("unhandled irreducible control flow");
1149
0
  }
1150
4.11M
}
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::computeMassInLoops()
Line
Count
Source
1138
2.24M
template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInLoops() {
1139
2.24M
  // Visit loops with the deepest first, and the top-level loops last.
1140
3.44M
  for (auto L = Loops.rbegin(), E = Loops.rend(); L != E; 
++L1.20M
) {
1141
1.20M
    if (computeMassInLoop(*L))
1142
1.20M
      continue;
1143
70
    auto Next = std::next(L);
1144
70
    computeIrreducibleMass(&*L, L.base());
1145
70
    L = std::prev(Next);
1146
70
    if (computeMassInLoop(*L))
1147
70
      continue;
1148
0
    llvm_unreachable("unhandled irreducible control flow");
1149
0
  }
1150
2.24M
}
llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::computeMassInLoops()
Line
Count
Source
1138
1.87M
template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInLoops() {
1139
1.87M
  // Visit loops with the deepest first, and the top-level loops last.
1140
2.70M
  for (auto L = Loops.rbegin(), E = Loops.rend(); L != E; 
++L836k
) {
1141
836k
    if (computeMassInLoop(*L))
1142
836k
      continue;
1143
71
    auto Next = std::next(L);
1144
71
    computeIrreducibleMass(&*L, L.base());
1145
71
    L = std::prev(Next);
1146
71
    if (computeMassInLoop(*L))
1147
71
      continue;
1148
0
    llvm_unreachable("unhandled irreducible control flow");
1149
0
  }
1150
1.87M
}
1151
1152
template <class BT>
1153
2.04M
bool BlockFrequencyInfoImpl<BT>::computeMassInLoop(LoopData &Loop) {
1154
2.04M
  // Compute mass in loop.
1155
2.04M
  LLVM_DEBUG(dbgs() << "compute-mass-in-loop: " << getLoopName(Loop) << "\n");
1156
2.04M
1157
2.04M
  if (Loop.isIrreducible()) {
1158
789
    LLVM_DEBUG(dbgs() << "isIrreducible = true\n");
1159
789
    Distribution Dist;
1160
789
    unsigned NumHeadersWithWeight = 0;
1161
789
    Optional<uint64_t> MinHeaderWeight;
1162
789
    DenseSet<uint32_t> HeadersWithoutWeight;
1163
789
    HeadersWithoutWeight.reserve(Loop.NumHeaders);
1164
3.24k
    for (uint32_t H = 0; H < Loop.NumHeaders; 
++H2.45k
) {
1165
2.45k
      auto &HeaderNode = Loop.Nodes[H];
1166
2.45k
      const BlockT *Block = getBlock(HeaderNode);
1167
2.45k
      IsIrrLoopHeader.set(Loop.Nodes[H].Index);
1168
2.45k
      Optional<uint64_t> HeaderWeight = Block->getIrrLoopHeaderWeight();
1169
2.45k
      if (!HeaderWeight) {
1170
2.44k
        LLVM_DEBUG(dbgs() << "Missing irr loop header metadata on "
1171
2.44k
                          << getBlockName(HeaderNode) << "\n");
1172
2.44k
        HeadersWithoutWeight.insert(H);
1173
2.44k
        continue;
1174
2.44k
      }
1175
16
      LLVM_DEBUG(dbgs() << getBlockName(HeaderNode)
1176
16
                        << " has irr loop header weight "
1177
16
                        << HeaderWeight.getValue() << "\n");
1178
16
      NumHeadersWithWeight++;
1179
16
      uint64_t HeaderWeightValue = HeaderWeight.getValue();
1180
16
      if (!MinHeaderWeight || 
HeaderWeightValue < MinHeaderWeight10
)
1181
12
        MinHeaderWeight = HeaderWeightValue;
1182
16
      if (HeaderWeightValue) {
1183
16
        Dist.addLocal(HeaderNode, HeaderWeightValue);
1184
16
      }
1185
16
    }
1186
789
    // As a heuristic, if some headers don't have a weight, give them the
1187
789
    // minimium weight seen (not to disrupt the existing trends too much by
1188
789
    // using a weight that's in the general range of the other headers' weights,
1189
789
    // and the minimum seems to perform better than the average.)
1190
789
    // FIXME: better update in the passes that drop the header weight.
1191
789
    // If no headers have a weight, give them even weight (use weight 1).
1192
789
    if (!MinHeaderWeight)
1193
783
      MinHeaderWeight = 1;
1194
2.44k
    for (uint32_t H : HeadersWithoutWeight) {
1195
2.44k
      auto &HeaderNode = Loop.Nodes[H];
1196
2.44k
      assert(!getBlock(HeaderNode)->getIrrLoopHeaderWeight() &&
1197
2.44k
             "Shouldn't have a weight metadata");
1198
2.44k
      uint64_t MinWeight = MinHeaderWeight.getValue();
1199
2.44k
      LLVM_DEBUG(dbgs() << "Giving weight " << MinWeight << " to "
1200
2.44k
                        << getBlockName(HeaderNode) << "\n");
1201
2.44k
      if (MinWeight)
1202
2.44k
        Dist.addLocal(HeaderNode, MinWeight);
1203
2.44k
    }
1204
789
    distributeIrrLoopHeaderMass(Dist);
1205
789
    for (const BlockNode &M : Loop.Nodes)
1206
11.4k
      if (!propagateMassToSuccessors(&Loop, M))
1207
11.4k
        
llvm_unreachable0
("unhandled irreducible control flow");
1208
789
    if (NumHeadersWithWeight == 0)
1209
783
      // No headers have a metadata. Adjust header mass.
1210
783
      adjustLoopHeaderMass(Loop);
1211
2.03M
  } else {
1212
2.03M
    Working[Loop.getHeader().Index].getMass() = BlockMass::getFull();
1213
2.03M
    if (!propagateMassToSuccessors(&Loop, Loop.getHeader()))
1214
2.03M
      
llvm_unreachable0
("irreducible control flow to loop header!?");
1215
2.03M
    for (const BlockNode &M : Loop.members())
1216
5.37M
      if (!propagateMassToSuccessors(&Loop, M))
1217
141
        // Irreducible backedge.
1218
141
        return false;
1219
2.03M
  }
1220
2.04M
1221
2.04M
  computeLoopScale(Loop);
1222
2.04M
  packageLoop(Loop);
1223
2.04M
  return true;
1224
2.04M
}
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::computeMassInLoop(llvm::BlockFrequencyInfoImplBase::LoopData&)
Line
Count
Source
1153
1.20M
bool BlockFrequencyInfoImpl<BT>::computeMassInLoop(LoopData &Loop) {
1154
1.20M
  // Compute mass in loop.
1155
1.20M
  LLVM_DEBUG(dbgs() << "compute-mass-in-loop: " << getLoopName(Loop) << "\n");
1156
1.20M
1157
1.20M
  if (Loop.isIrreducible()) {
1158
395
    LLVM_DEBUG(dbgs() << "isIrreducible = true\n");
1159
395
    Distribution Dist;
1160
395
    unsigned NumHeadersWithWeight = 0;
1161
395
    Optional<uint64_t> MinHeaderWeight;
1162
395
    DenseSet<uint32_t> HeadersWithoutWeight;
1163
395
    HeadersWithoutWeight.reserve(Loop.NumHeaders);
1164
1.63k
    for (uint32_t H = 0; H < Loop.NumHeaders; 
++H1.23k
) {
1165
1.23k
      auto &HeaderNode = Loop.Nodes[H];
1166
1.23k
      const BlockT *Block = getBlock(HeaderNode);
1167
1.23k
      IsIrrLoopHeader.set(Loop.Nodes[H].Index);
1168
1.23k
      Optional<uint64_t> HeaderWeight = Block->getIrrLoopHeaderWeight();
1169
1.23k
      if (!HeaderWeight) {
1170
1.22k
        LLVM_DEBUG(dbgs() << "Missing irr loop header metadata on "
1171
1.22k
                          << getBlockName(HeaderNode) << "\n");
1172
1.22k
        HeadersWithoutWeight.insert(H);
1173
1.22k
        continue;
1174
1.22k
      }
1175
16
      LLVM_DEBUG(dbgs() << getBlockName(HeaderNode)
1176
16
                        << " has irr loop header weight "
1177
16
                        << HeaderWeight.getValue() << "\n");
1178
16
      NumHeadersWithWeight++;
1179
16
      uint64_t HeaderWeightValue = HeaderWeight.getValue();
1180
16
      if (!MinHeaderWeight || 
HeaderWeightValue < MinHeaderWeight10
)
1181
12
        MinHeaderWeight = HeaderWeightValue;
1182
16
      if (HeaderWeightValue) {
1183
16
        Dist.addLocal(HeaderNode, HeaderWeightValue);
1184
16
      }
1185
16
    }
1186
395
    // As a heuristic, if some headers don't have a weight, give them the
1187
395
    // minimium weight seen (not to disrupt the existing trends too much by
1188
395
    // using a weight that's in the general range of the other headers' weights,
1189
395
    // and the minimum seems to perform better than the average.)
1190
395
    // FIXME: better update in the passes that drop the header weight.
1191
395
    // If no headers have a weight, give them even weight (use weight 1).
1192
395
    if (!MinHeaderWeight)
1193
389
      MinHeaderWeight = 1;
1194
1.22k
    for (uint32_t H : HeadersWithoutWeight) {
1195
1.22k
      auto &HeaderNode = Loop.Nodes[H];
1196
1.22k
      assert(!getBlock(HeaderNode)->getIrrLoopHeaderWeight() &&
1197
1.22k
             "Shouldn't have a weight metadata");
1198
1.22k
      uint64_t MinWeight = MinHeaderWeight.getValue();
1199
1.22k
      LLVM_DEBUG(dbgs() << "Giving weight " << MinWeight << " to "
1200
1.22k
                        << getBlockName(HeaderNode) << "\n");
1201
1.22k
      if (MinWeight)
1202
1.22k
        Dist.addLocal(HeaderNode, MinWeight);
1203
1.22k
    }
1204
395
    distributeIrrLoopHeaderMass(Dist);
1205
395
    for (const BlockNode &M : Loop.Nodes)
1206
6.20k
      if (!propagateMassToSuccessors(&Loop, M))
1207
6.20k
        
llvm_unreachable0
("unhandled irreducible control flow");
1208
395
    if (NumHeadersWithWeight == 0)
1209
389
      // No headers have a metadata. Adjust header mass.
1210
389
      adjustLoopHeaderMass(Loop);
1211
1.20M
  } else {
1212
1.20M
    Working[Loop.getHeader().Index].getMass() = BlockMass::getFull();
1213
1.20M
    if (!propagateMassToSuccessors(&Loop, Loop.getHeader()))
1214
1.20M
      
llvm_unreachable0
("irreducible control flow to loop header!?");
1215
1.20M
    for (const BlockNode &M : Loop.members())
1216
3.24M
      if (!propagateMassToSuccessors(&Loop, M))
1217
70
        // Irreducible backedge.
1218
70
        return false;
1219
1.20M
  }
1220
1.20M
1221
1.20M
  computeLoopScale(Loop);
1222
1.20M
  packageLoop(Loop);
1223
1.20M
  return true;
1224
1.20M
}
llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::computeMassInLoop(llvm::BlockFrequencyInfoImplBase::LoopData&)
Line
Count
Source
1153
836k
bool BlockFrequencyInfoImpl<BT>::computeMassInLoop(LoopData &Loop) {
1154
836k
  // Compute mass in loop.
1155
836k
  LLVM_DEBUG(dbgs() << "compute-mass-in-loop: " << getLoopName(Loop) << "\n");
1156
836k
1157
836k
  if (Loop.isIrreducible()) {
1158
394
    LLVM_DEBUG(dbgs() << "isIrreducible = true\n");
1159
394
    Distribution Dist;
1160
394
    unsigned NumHeadersWithWeight = 0;
1161
394
    Optional<uint64_t> MinHeaderWeight;
1162
394
    DenseSet<uint32_t> HeadersWithoutWeight;
1163
394
    HeadersWithoutWeight.reserve(Loop.NumHeaders);
1164
1.61k
    for (uint32_t H = 0; H < Loop.NumHeaders; 
++H1.21k
) {
1165
1.21k
      auto &HeaderNode = Loop.Nodes[H];
1166
1.21k
      const BlockT *Block = getBlock(HeaderNode);
1167
1.21k
      IsIrrLoopHeader.set(Loop.Nodes[H].Index);
1168
1.21k
      Optional<uint64_t> HeaderWeight = Block->getIrrLoopHeaderWeight();
1169
1.21k
      if (!HeaderWeight) {
1170
1.21k
        LLVM_DEBUG(dbgs() << "Missing irr loop header metadata on "
1171
1.21k
                          << getBlockName(HeaderNode) << "\n");
1172
1.21k
        HeadersWithoutWeight.insert(H);
1173
1.21k
        continue;
1174
1.21k
      }
1175
0
      LLVM_DEBUG(dbgs() << getBlockName(HeaderNode)
1176
0
                        << " has irr loop header weight "
1177
0
                        << HeaderWeight.getValue() << "\n");
1178
0
      NumHeadersWithWeight++;
1179
0
      uint64_t HeaderWeightValue = HeaderWeight.getValue();
1180
0
      if (!MinHeaderWeight || HeaderWeightValue < MinHeaderWeight)
1181
0
        MinHeaderWeight = HeaderWeightValue;
1182
0
      if (HeaderWeightValue) {
1183
0
        Dist.addLocal(HeaderNode, HeaderWeightValue);
1184
0
      }
1185
0
    }
1186
394
    // As a heuristic, if some headers don't have a weight, give them the
1187
394
    // minimium weight seen (not to disrupt the existing trends too much by
1188
394
    // using a weight that's in the general range of the other headers' weights,
1189
394
    // and the minimum seems to perform better than the average.)
1190
394
    // FIXME: better update in the passes that drop the header weight.
1191
394
    // If no headers have a weight, give them even weight (use weight 1).
1192
394
    if (!MinHeaderWeight)
1193
394
      MinHeaderWeight = 1;
1194
1.21k
    for (uint32_t H : HeadersWithoutWeight) {
1195
1.21k
      auto &HeaderNode = Loop.Nodes[H];
1196
1.21k
      assert(!getBlock(HeaderNode)->getIrrLoopHeaderWeight() &&
1197
1.21k
             "Shouldn't have a weight metadata");
1198
1.21k
      uint64_t MinWeight = MinHeaderWeight.getValue();
1199
1.21k
      LLVM_DEBUG(dbgs() << "Giving weight " << MinWeight << " to "
1200
1.21k
                        << getBlockName(HeaderNode) << "\n");
1201
1.21k
      if (MinWeight)
1202
1.21k
        Dist.addLocal(HeaderNode, MinWeight);
1203
1.21k
    }
1204
394
    distributeIrrLoopHeaderMass(Dist);
1205
394
    for (const BlockNode &M : Loop.Nodes)
1206
5.28k
      if (!propagateMassToSuccessors(&Loop, M))
1207
5.28k
        
llvm_unreachable0
("unhandled irreducible control flow");
1208
394
    if (NumHeadersWithWeight == 0)
1209
394
      // No headers have a metadata. Adjust header mass.
1210
394
      adjustLoopHeaderMass(Loop);
1211
836k
  } else {
1212
836k
    Working[Loop.getHeader().Index].getMass() = BlockMass::getFull();
1213
836k
    if (!propagateMassToSuccessors(&Loop, Loop.getHeader()))
1214
836k
      
llvm_unreachable0
("irreducible control flow to loop header!?");
1215
836k
    for (const BlockNode &M : Loop.members())
1216
2.13M
      if (!propagateMassToSuccessors(&Loop, M))
1217
71
        // Irreducible backedge.
1218
71
        return false;
1219
836k
  }
1220
836k
1221
836k
  computeLoopScale(Loop);
1222
836k
  packageLoop(Loop);
1223
836k
  return true;
1224
836k
}
1225
1226
template <class BT>
1227
4.11M
bool BlockFrequencyInfoImpl<BT>::tryToComputeMassInFunction() {
1228
4.11M
  // Compute mass in function.
1229
4.11M
  LLVM_DEBUG(dbgs() << "compute-mass-in-function\n");
1230
4.11M
  assert(!Working.empty() && "no blocks in function");
1231
4.11M
  assert(!Working[0].isLoopHeader() && "entry block is a loop header");
1232
4.11M
1233
4.11M
  Working[0].getMass() = BlockMass::getFull();
1234
31.2M
  for (rpot_iterator I = rpot_begin(), IE = rpot_end(); I != IE; 
++I27.1M
) {
1235
27.1M
    // Check for nodes that have been packaged.
1236
27.1M
    BlockNode Node = getNode(I);
1237
27.1M
    if (Working[Node.Index].isPackaged())
1238
5.38M
      continue;
1239
21.7M
1240
21.7M
    if (!propagateMassToSuccessors(nullptr, Node))
1241
517
      return false;
1242
21.7M
  }
1243
4.11M
  
return true4.11M
;
1244
4.11M
}
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::tryToComputeMassInFunction()
Line
Count
Source
1227
2.24M
bool BlockFrequencyInfoImpl<BT>::tryToComputeMassInFunction() {
1228
2.24M
  // Compute mass in function.
1229
2.24M
  LLVM_DEBUG(dbgs() << "compute-mass-in-function\n");
1230
2.24M
  assert(!Working.empty() && "no blocks in function");
1231
2.24M
  assert(!Working[0].isLoopHeader() && "entry block is a loop header");
1232
2.24M
1233
2.24M
  Working[0].getMass() = BlockMass::getFull();
1234
18.2M
  for (rpot_iterator I = rpot_begin(), IE = rpot_end(); I != IE; 
++I15.9M
) {
1235
15.9M
    // Check for nodes that have been packaged.
1236
15.9M
    BlockNode Node = getNode(I);
1237
15.9M
    if (Working[Node.Index].isPackaged())
1238
3.25M
      continue;
1239
12.7M
1240
12.7M
    if (!propagateMassToSuccessors(nullptr, Node))
1241
235
      return false;
1242
12.7M
  }
1243
2.24M
  
return true2.24M
;
1244
2.24M
}
llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::tryToComputeMassInFunction()
Line
Count
Source
1227
1.87M
bool BlockFrequencyInfoImpl<BT>::tryToComputeMassInFunction() {
1228
1.87M
  // Compute mass in function.
1229
1.87M
  LLVM_DEBUG(dbgs() << "compute-mass-in-function\n");
1230
1.87M
  assert(!Working.empty() && "no blocks in function");
1231
1.87M
  assert(!Working[0].isLoopHeader() && "entry block is a loop header");
1232
1.87M
1233
1.87M
  Working[0].getMass() = BlockMass::getFull();
1234
13.0M
  for (rpot_iterator I = rpot_begin(), IE = rpot_end(); I != IE; 
++I11.1M
) {
1235
11.1M
    // Check for nodes that have been packaged.
1236
11.1M
    BlockNode Node = getNode(I);
1237
11.1M
    if (Working[Node.Index].isPackaged())
1238
2.13M
      continue;
1239
9.04M
1240
9.04M
    if (!propagateMassToSuccessors(nullptr, Node))
1241
282
      return false;
1242
9.04M
  }
1243
1.87M
  
return true1.87M
;
1244
1.87M
}
1245
1246
4.11M
template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInFunction() {
1247
4.11M
  if (tryToComputeMassInFunction())
1248
4.11M
    return;
1249
499
  computeIrreducibleMass(nullptr, Loops.begin());
1250
499
  if (tryToComputeMassInFunction())
1251
517
    return;
1252
18.4E
  llvm_unreachable("unhandled irreducible control flow");
1253
18.4E
}
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::computeMassInFunction()
Line
Count
Source
1246
2.24M
template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInFunction() {
1247
2.24M
  if (tryToComputeMassInFunction())
1248
2.24M
    return;
1249
219
  computeIrreducibleMass(nullptr, Loops.begin());
1250
219
  if (tryToComputeMassInFunction())
1251
235
    return;
1252
18.4E
  llvm_unreachable("unhandled irreducible control flow");
1253
18.4E
}
llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::computeMassInFunction()
Line
Count
Source
1246
1.87M
template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInFunction() {
1247
1.87M
  if (tryToComputeMassInFunction())
1248
1.87M
    return;
1249
280
  computeIrreducibleMass(nullptr, Loops.begin());
1250
280
  if (tryToComputeMassInFunction())
1251
282
    return;
1252
18.4E
  llvm_unreachable("unhandled irreducible control flow");
1253
18.4E
}
1254
1255
/// \note This should be a lambda, but that crashes GCC 4.7.
1256
namespace bfi_detail {
1257
1258
template <class BT> struct BlockEdgesAdder {
1259
  using BlockT = BT;
1260
  using LoopData = BlockFrequencyInfoImplBase::LoopData;
1261
  using Successor = GraphTraits<const BlockT *>;
1262
1263
  const BlockFrequencyInfoImpl<BT> &BFI;
1264
1265
  explicit BlockEdgesAdder(const BlockFrequencyInfoImpl<BT> &BFI)
1266
658
      : BFI(BFI) {}
llvm::bfi_detail::BlockEdgesAdder<llvm::BasicBlock>::BlockEdgesAdder(llvm::BlockFrequencyInfoImpl<llvm::BasicBlock> const&)
Line
Count
Source
1266
305
      : BFI(BFI) {}
llvm::bfi_detail::BlockEdgesAdder<llvm::MachineBasicBlock>::BlockEdgesAdder(llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock> const&)
Line
Count
Source
1266
353
      : BFI(BFI) {}
1267
1268
  void operator()(IrreducibleGraph &G, IrreducibleGraph::IrrNode &Irr,
1269
23.7k
                  const LoopData *OuterLoop) {
1270
23.7k
    const BlockT *BB = BFI.RPOT[Irr.Node.Index];
1271
23.7k
    for (const auto Succ : children<const BlockT *>(BB))
1272
38.2k
      G.addEdge(Irr, BFI.getNode(Succ), OuterLoop);
1273
23.7k
  }
llvm::bfi_detail::BlockEdgesAdder<llvm::BasicBlock>::operator()(llvm::bfi_detail::IrreducibleGraph&, llvm::bfi_detail::IrreducibleGraph::IrrNode&, llvm::BlockFrequencyInfoImplBase::LoopData const*)
Line
Count
Source
1269
11.5k
                  const LoopData *OuterLoop) {
1270
11.5k
    const BlockT *BB = BFI.RPOT[Irr.Node.Index];
1271
11.5k
    for (const auto Succ : children<const BlockT *>(BB))
1272
16.8k
      G.addEdge(Irr, BFI.getNode(Succ), OuterLoop);
1273
11.5k
  }
llvm::bfi_detail::BlockEdgesAdder<llvm::MachineBasicBlock>::operator()(llvm::bfi_detail::IrreducibleGraph&, llvm::bfi_detail::IrreducibleGraph::IrrNode&, llvm::BlockFrequencyInfoImplBase::LoopData const*)
Line
Count
Source
1269
12.1k
                  const LoopData *OuterLoop) {
1270
12.1k
    const BlockT *BB = BFI.RPOT[Irr.Node.Index];
1271
12.1k
    for (const auto Succ : children<const BlockT *>(BB))
1272
21.4k
      G.addEdge(Irr, BFI.getNode(Succ), OuterLoop);
1273
12.1k
  }
1274
};
1275
1276
} // end namespace bfi_detail
1277
1278
template <class BT>
1279
void BlockFrequencyInfoImpl<BT>::computeIrreducibleMass(
1280
658
    LoopData *OuterLoop, std::list<LoopData>::iterator Insert) {
1281
658
  LLVM_DEBUG(dbgs() << "analyze-irreducible-in-";
1282
658
             if (OuterLoop) dbgs()
1283
658
             << "loop: " << getLoopName(*OuterLoop) << "\n";
1284
658
             else dbgs() << "function\n");
1285
658
1286
658
  using namespace bfi_detail;
1287
658
1288
658
  // Ideally, addBlockEdges() would be declared here as a lambda, but that
1289
658
  // crashes GCC 4.7.
1290
658
  BlockEdgesAdder<BT> addBlockEdges(*this);
1291
658
  IrreducibleGraph G(*this, OuterLoop, addBlockEdges);
1292
658
1293
658
  for (auto &L : analyzeIrreducible(G, OuterLoop, Insert))
1294
789
    computeMassInLoop(L);
1295
658
1296
658
  if (!OuterLoop)
1297
517
    return;
1298
141
  updateLoopWithIrreducible(*OuterLoop);
1299
141
}
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::computeIrreducibleMass(llvm::BlockFrequencyInfoImplBase::LoopData*, std::__1::__list_iterator<llvm::BlockFrequencyInfoImplBase::LoopData, void*>)
Line
Count
Source
1280
305
    LoopData *OuterLoop, std::list<LoopData>::iterator Insert) {
1281
305
  LLVM_DEBUG(dbgs() << "analyze-irreducible-in-";
1282
305
             if (OuterLoop) dbgs()
1283
305
             << "loop: " << getLoopName(*OuterLoop) << "\n";
1284
305
             else dbgs() << "function\n");
1285
305
1286
305
  using namespace bfi_detail;
1287
305
1288
305
  // Ideally, addBlockEdges() would be declared here as a lambda, but that
1289
305
  // crashes GCC 4.7.
1290
305
  BlockEdgesAdder<BT> addBlockEdges(*this);
1291
305
  IrreducibleGraph G(*this, OuterLoop, addBlockEdges);
1292
305
1293
305
  for (auto &L : analyzeIrreducible(G, OuterLoop, Insert))
1294
395
    computeMassInLoop(L);
1295
305
1296
305
  if (!OuterLoop)
1297
235
    return;
1298
70
  updateLoopWithIrreducible(*OuterLoop);
1299
70
}
llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::computeIrreducibleMass(llvm::BlockFrequencyInfoImplBase::LoopData*, std::__1::__list_iterator<llvm::BlockFrequencyInfoImplBase::LoopData, void*>)
Line
Count
Source
1280
353
    LoopData *OuterLoop, std::list<LoopData>::iterator Insert) {
1281
353
  LLVM_DEBUG(dbgs() << "analyze-irreducible-in-";
1282
353
             if (OuterLoop) dbgs()
1283
353
             << "loop: " << getLoopName(*OuterLoop) << "\n";
1284
353
             else dbgs() << "function\n");
1285
353
1286
353
  using namespace bfi_detail;
1287
353
1288
353
  // Ideally, addBlockEdges() would be declared here as a lambda, but that
1289
353
  // crashes GCC 4.7.
1290
353
  BlockEdgesAdder<BT> addBlockEdges(*this);
1291
353
  IrreducibleGraph G(*this, OuterLoop, addBlockEdges);
1292
353
1293
353
  for (auto &L : analyzeIrreducible(G, OuterLoop, Insert))
1294
394
    computeMassInLoop(L);
1295
353
1296
353
  if (!OuterLoop)
1297
282
    return;
1298
71
  updateLoopWithIrreducible(*OuterLoop);
1299
71
}
1300
1301
// A helper function that converts a branch probability into weight.
1302
35.3M
inline uint32_t getWeightFromBranchProb(const BranchProbability Prob) {
1303
35.3M
  return Prob.getNumerator();
1304
35.3M
}
1305
1306
template <class BT>
1307
bool
1308
BlockFrequencyInfoImpl<BT>::propagateMassToSuccessors(LoopData *OuterLoop,
1309
29.1M
                                                      const BlockNode &Node) {
1310
29.1M
  LLVM_DEBUG(dbgs() << " - node: " << getBlockName(Node) << "\n");
1311
29.1M
  // Calculate probability for successors.
1312
29.1M
  Distribution Dist;
1313
29.1M
  if (auto *Loop = Working[Node.Index].getPackagedLoop()) {
1314
2.04M
    assert(Loop != OuterLoop && "Cannot propagate mass in a packaged loop");
1315
2.04M
    if (!addLoopSuccessorsToDist(OuterLoop, *Loop, Dist))
1316
57
      // Irreducible backedge.
1317
57
      return false;
1318
27.1M
  } else {
1319
27.1M
    const BlockT *BB = getBlock(Node);
1320
27.1M
    for (auto SI = GraphTraits<const BlockT *>::child_begin(BB),
1321
27.1M
              SE = GraphTraits<const BlockT *>::child_end(BB);
1322
62.4M
         SI != SE; 
++SI35.3M
)
1323
35.3M
      if (!addToDist(
1324
35.3M
              Dist, OuterLoop, Node, getNode(*SI),
1325
35.3M
              getWeightFromBranchProb(BPI->getEdgeProbability(BB, SI))))
1326
601
        // Irreducible backedge.
1327
601
        return false;
1328
27.1M
  }
1329
29.1M
1330
29.1M
  // Distribute mass to successors, saving exit and backedge data in the
1331
29.1M
  // loop header.
1332
29.1M
  distributeMass(Node, OuterLoop, Dist);
1333
29.1M
  return true;
1334
29.1M
}
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::propagateMassToSuccessors(llvm::BlockFrequencyInfoImplBase::LoopData*, llvm::BlockFrequencyInfoImplBase::BlockNode const&)
Line
Count
Source
1309
17.1M
                                                      const BlockNode &Node) {
1310
17.1M
  LLVM_DEBUG(dbgs() << " - node: " << getBlockName(Node) << "\n");
1311
17.1M
  // Calculate probability for successors.
1312
17.1M
  Distribution Dist;
1313
17.1M
  if (auto *Loop = Working[Node.Index].getPackagedLoop()) {
1314
1.20M
    assert(Loop != OuterLoop && "Cannot propagate mass in a packaged loop");
1315
1.20M
    if (!addLoopSuccessorsToDist(OuterLoop, *Loop, Dist))
1316
12
      // Irreducible backedge.
1317
12
      return false;
1318
15.9M
  } else {
1319
15.9M
    const BlockT *BB = getBlock(Node);
1320
15.9M
    for (auto SI = GraphTraits<const BlockT *>::child_begin(BB),
1321
15.9M
              SE = GraphTraits<const BlockT *>::child_end(BB);
1322
37.2M
         SI != SE; 
++SI21.3M
)
1323
21.3M
      if (!addToDist(
1324
21.3M
              Dist, OuterLoop, Node, getNode(*SI),
1325
21.3M
              getWeightFromBranchProb(BPI->getEdgeProbability(BB, SI))))
1326
293
        // Irreducible backedge.
1327
293
        return false;
1328
15.9M
  }
1329
17.1M
1330
17.1M
  // Distribute mass to successors, saving exit and backedge data in the
1331
17.1M
  // loop header.
1332
17.1M
  distributeMass(Node, OuterLoop, Dist);
1333
17.1M
  return true;
1334
17.1M
}
llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::propagateMassToSuccessors(llvm::BlockFrequencyInfoImplBase::LoopData*, llvm::BlockFrequencyInfoImplBase::BlockNode const&)
Line
Count
Source
1309
12.0M
                                                      const BlockNode &Node) {
1310
12.0M
  LLVM_DEBUG(dbgs() << " - node: " << getBlockName(Node) << "\n");
1311
12.0M
  // Calculate probability for successors.
1312
12.0M
  Distribution Dist;
1313
12.0M
  if (auto *Loop = Working[Node.Index].getPackagedLoop()) {
1314
837k
    assert(Loop != OuterLoop && "Cannot propagate mass in a packaged loop");
1315
837k
    if (!addLoopSuccessorsToDist(OuterLoop, *Loop, Dist))
1316
45
      // Irreducible backedge.
1317
45
      return false;
1318
11.1M
  } else {
1319
11.1M
    const BlockT *BB = getBlock(Node);
1320
11.1M
    for (auto SI = GraphTraits<const BlockT *>::child_begin(BB),
1321
11.1M
              SE = GraphTraits<const BlockT *>::child_end(BB);
1322
25.1M
         SI != SE; 
++SI13.9M
)
1323
13.9M
      if (!addToDist(
1324
13.9M
              Dist, OuterLoop, Node, getNode(*SI),
1325
13.9M
              getWeightFromBranchProb(BPI->getEdgeProbability(BB, SI))))
1326
308
        // Irreducible backedge.
1327
308
        return false;
1328
11.1M
  }
1329
12.0M
1330
12.0M
  // Distribute mass to successors, saving exit and backedge data in the
1331
12.0M
  // loop header.
1332
12.0M
  distributeMass(Node, OuterLoop, Dist);
1333
12.0M
  return true;
1334
12.0M
}
1335
1336
template <class BT>
1337
63
raw_ostream &BlockFrequencyInfoImpl<BT>::print(raw_ostream &OS) const {
1338
63
  if (!F)
1339
0
    return OS;
1340
63
  OS << "block-frequency-info: " << F->getName() << "\n";
1341
532
  for (const BlockT &BB : *F) {
1342
532
    OS << " - " << bfi_detail::getBlockName(&BB) << ": float = ";
1343
532
    getFloatingBlockFreq(&BB).print(OS, 5)
1344
532
        << ", int = " << getBlockFreq(&BB).getFrequency();
1345
532
    if (Optional<uint64_t> ProfileCount =
1346
78
        BlockFrequencyInfoImplBase::getBlockProfileCount(
1347
78
            F->getFunction(), getNode(&BB)))
1348
78
      OS << ", count = " << ProfileCount.getValue();
1349
532
    if (Optional<uint64_t> IrrLoopHeaderWeight =
1350
16
        BB.getIrrLoopHeaderWeight())
1351
16
      OS << ", irr_loop_header_weight = " << IrrLoopHeaderWeight.getValue();
1352
532
    OS << "\n";
1353
532
  }
1354
63
1355
63
  // Add an extra newline for readability.
1356
63
  OS << "\n";
1357
63
  return OS;
1358
63
}
llvm::BlockFrequencyInfoImpl<llvm::BasicBlock>::print(llvm::raw_ostream&) const
Line
Count
Source
1337
63
raw_ostream &BlockFrequencyInfoImpl<BT>::print(raw_ostream &OS) const {
1338
63
  if (!F)
1339
0
    return OS;
1340
63
  OS << "block-frequency-info: " << F->getName() << "\n";
1341
532
  for (const BlockT &BB : *F) {
1342
532
    OS << " - " << bfi_detail::getBlockName(&BB) << ": float = ";
1343
532
    getFloatingBlockFreq(&BB).print(OS, 5)
1344
532
        << ", int = " << getBlockFreq(&BB).getFrequency();
1345
532
    if (Optional<uint64_t> ProfileCount =
1346
78
        BlockFrequencyInfoImplBase::getBlockProfileCount(
1347
78
            F->getFunction(), getNode(&BB)))
1348
78
      OS << ", count = " << ProfileCount.getValue();
1349
532
    if (Optional<uint64_t> IrrLoopHeaderWeight =
1350
16
        BB.getIrrLoopHeaderWeight())
1351
16
      OS << ", irr_loop_header_weight = " << IrrLoopHeaderWeight.getValue();
1352
532
    OS << "\n";
1353
532
  }
1354
63
1355
63
  // Add an extra newline for readability.
1356
63
  OS << "\n";
1357
63
  return OS;
1358
63
}
Unexecuted instantiation: llvm::BlockFrequencyInfoImpl<llvm::MachineBasicBlock>::print(llvm::raw_ostream&) const
1359
1360
// Graph trait base class for block frequency information graph
1361
// viewer.
1362
1363
enum GVDAGType { GVDT_None, GVDT_Fraction, GVDT_Integer, GVDT_Count };
1364
1365
template <class BlockFrequencyInfoT, class BranchProbabilityInfoT>
1366
struct BFIDOTGraphTraitsBase : public DefaultDOTGraphTraits {
1367
  using GTraits = GraphTraits<BlockFrequencyInfoT *>;
1368
  using NodeRef = typename GTraits::NodeRef;
1369
  using EdgeIter = typename GTraits::ChildIteratorType;
1370
  using NodeIter = typename GTraits::nodes_iterator;
1371
1372
  uint64_t MaxFrequency = 0;
1373
1374
  explicit BFIDOTGraphTraitsBase(bool isSimple = false)
1375
0
      : DefaultDOTGraphTraits(isSimple) {}
Unexecuted instantiation: llvm::BFIDOTGraphTraitsBase<llvm::BlockFrequencyInfo, llvm::BranchProbabilityInfo>::BFIDOTGraphTraitsBase(bool)
Unexecuted instantiation: llvm::BFIDOTGraphTraitsBase<llvm::MachineBlockFrequencyInfo, llvm::MachineBranchProbabilityInfo>::BFIDOTGraphTraitsBase(bool)
1376
1377
0
  static std::string getGraphName(const BlockFrequencyInfoT *G) {
1378
0
    return G->getFunction()->getName();
1379
0
  }
Unexecuted instantiation: llvm::BFIDOTGraphTraitsBase<llvm::BlockFrequencyInfo, llvm::BranchProbabilityInfo>::getGraphName(llvm::BlockFrequencyInfo const*)
Unexecuted instantiation: llvm::BFIDOTGraphTraitsBase<llvm::MachineBlockFrequencyInfo, llvm::MachineBranchProbabilityInfo>::getGraphName(llvm::MachineBlockFrequencyInfo const*)
1380
1381
  std::string getNodeAttributes(NodeRef Node, const BlockFrequencyInfoT *Graph,
1382
0
                                unsigned HotPercentThreshold = 0) {
1383
0
    std::string Result;
1384
0
    if (!HotPercentThreshold)
1385
0
      return Result;
1386
0
1387
0
    // Compute MaxFrequency on the fly:
1388
0
    if (!MaxFrequency) {
1389
0
      for (NodeIter I = GTraits::nodes_begin(Graph),
1390
0
                    E = GTraits::nodes_end(Graph);
1391
0
           I != E; ++I) {
1392
0
        NodeRef N = *I;
1393
0
        MaxFrequency =
1394
0
            std::max(MaxFrequency, Graph->getBlockFreq(N).getFrequency());
1395
0
      }
1396
0
    }
1397
0
    BlockFrequency Freq = Graph->getBlockFreq(Node);
1398
0
    BlockFrequency HotFreq =
1399
0
        (BlockFrequency(MaxFrequency) *
1400
0
         BranchProbability::getBranchProbability(HotPercentThreshold, 100));
1401
0
1402
0
    if (Freq < HotFreq)
1403
0
      return Result;
1404
0
1405
0
    raw_string_ostream OS(Result);
1406
0
    OS << "color=\"red\"";
1407
0
    OS.flush();
1408
0
    return Result;
1409
0
  }
Unexecuted instantiation: llvm::BFIDOTGraphTraitsBase<llvm::BlockFrequencyInfo, llvm::BranchProbabilityInfo>::getNodeAttributes(llvm::BasicBlock const*, llvm::BlockFrequencyInfo const*, unsigned int)
Unexecuted instantiation: llvm::BFIDOTGraphTraitsBase<llvm::MachineBlockFrequencyInfo, llvm::MachineBranchProbabilityInfo>::getNodeAttributes(llvm::MachineBasicBlock const*, llvm::MachineBlockFrequencyInfo const*, unsigned int)
1410
1411
  std::string getNodeLabel(NodeRef Node, const BlockFrequencyInfoT *Graph,
1412
0
                           GVDAGType GType, int layout_order = -1) {
1413
0
    std::string Result;
1414
0
    raw_string_ostream OS(Result);
1415
0
1416
0
    if (layout_order != -1)
1417
0
      OS << Node->getName() << "[" << layout_order << "] : ";
1418
0
    else
1419
0
      OS << Node->getName() << " : ";
1420
0
    switch (GType) {
1421
0
    case GVDT_Fraction:
1422
0
      Graph->printBlockFreq(OS, Node);
1423
0
      break;
1424
0
    case GVDT_Integer:
1425
0
      OS << Graph->getBlockFreq(Node).getFrequency();
1426
0
      break;
1427
0
    case GVDT_Count: {
1428
0
      auto Count = Graph->getBlockProfileCount(Node);
1429
0
      if (Count)
1430
0
        OS << Count.getValue();
1431
0
      else
1432
0
        OS << "Unknown";
1433
0
      break;
1434
0
    }
1435
0
    case GVDT_None:
1436
0
      llvm_unreachable("If we are not supposed to render a graph we should "
1437
0
                       "never reach this point.");
1438
0
    }
1439
0
    return Result;
1440
0
  }
Unexecuted instantiation: llvm::BFIDOTGraphTraitsBase<llvm::BlockFrequencyInfo, llvm::BranchProbabilityInfo>::getNodeLabel(llvm::BasicBlock const*, llvm::BlockFrequencyInfo const*, llvm::GVDAGType, int)
Unexecuted instantiation: llvm::BFIDOTGraphTraitsBase<llvm::MachineBlockFrequencyInfo, llvm::MachineBranchProbabilityInfo>::getNodeLabel(llvm::MachineBasicBlock const*, llvm::MachineBlockFrequencyInfo const*, llvm::GVDAGType, int)
1441
1442
  std::string getEdgeAttributes(NodeRef Node, EdgeIter EI,
1443
                                const BlockFrequencyInfoT *BFI,
1444
                                const BranchProbabilityInfoT *BPI,
1445
0
                                unsigned HotPercentThreshold = 0) {
1446
0
    std::string Str;
1447
0
    if (!BPI)
1448
0
      return Str;
1449
0
1450
0
    BranchProbability BP = BPI->getEdgeProbability(Node, EI);
1451
0
    uint32_t N = BP.getNumerator();
1452
0
    uint32_t D = BP.getDenominator();
1453
0
    double Percent = 100.0 * N / D;
1454
0
    raw_string_ostream OS(Str);
1455
0
    OS << format("label=\"%.1f%%\"", Percent);
1456
0
1457
0
    if (HotPercentThreshold) {
1458
0
      BlockFrequency EFreq = BFI->getBlockFreq(Node) * BP;
1459
0
      BlockFrequency HotFreq = BlockFrequency(MaxFrequency) *
1460
0
                               BranchProbability(HotPercentThreshold, 100);
1461
0
1462
0
      if (EFreq >= HotFreq) {
1463
0
        OS << ",color=\"red\"";
1464
0
      }
1465
0
    }
1466
0
1467
0
    OS.flush();
1468
0
    return Str;
1469
0
  }
Unexecuted instantiation: llvm::BFIDOTGraphTraitsBase<llvm::BlockFrequencyInfo, llvm::BranchProbabilityInfo>::getEdgeAttributes(llvm::BasicBlock const*, llvm::SuccIterator<llvm::Instruction const, llvm::BasicBlock const>, llvm::BlockFrequencyInfo const*, llvm::BranchProbabilityInfo const*, unsigned int)
Unexecuted instantiation: llvm::BFIDOTGraphTraitsBase<llvm::MachineBlockFrequencyInfo, llvm::MachineBranchProbabilityInfo>::getEdgeAttributes(llvm::MachineBasicBlock const*, std::__1::__wrap_iter<llvm::MachineBasicBlock* const*>, llvm::MachineBlockFrequencyInfo const*, llvm::MachineBranchProbabilityInfo const*, unsigned int)
1470
};
1471
1472
} // end namespace llvm
1473
1474
#undef DEBUG_TYPE
1475
1476
#endif // LLVM_ANALYSIS_BLOCKFREQUENCYINFOIMPL_H