xref: /llvm-project/llvm/lib/Analysis/BlockFrequencyInfoImpl.cpp (revision d34e60ca8532511acb8c93ef26297e349fbec86a)
1 //===- BlockFrequencyImplInfo.cpp - Block Frequency Info Implementation ---===//
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 // Loops should be simplified before this analysis.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/Analysis/BlockFrequencyInfoImpl.h"
15 #include "llvm/ADT/APInt.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/GraphTraits.h"
18 #include "llvm/ADT/None.h"
19 #include "llvm/ADT/SCCIterator.h"
20 #include "llvm/Config/llvm-config.h"
21 #include "llvm/IR/Function.h"
22 #include "llvm/Support/BlockFrequency.h"
23 #include "llvm/Support/BranchProbability.h"
24 #include "llvm/Support/Compiler.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/Support/ScaledNumber.h"
27 #include "llvm/Support/MathExtras.h"
28 #include "llvm/Support/raw_ostream.h"
29 #include <algorithm>
30 #include <cassert>
31 #include <cstddef>
32 #include <cstdint>
33 #include <iterator>
34 #include <list>
35 #include <numeric>
36 #include <utility>
37 #include <vector>
38 
39 using namespace llvm;
40 using namespace llvm::bfi_detail;
41 
42 #define DEBUG_TYPE "block-freq"
43 
44 ScaledNumber<uint64_t> BlockMass::toScaled() const {
45   if (isFull())
46     return ScaledNumber<uint64_t>(1, 0);
47   return ScaledNumber<uint64_t>(getMass() + 1, -64);
48 }
49 
50 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
51 LLVM_DUMP_METHOD void BlockMass::dump() const { print(dbgs()); }
52 #endif
53 
54 static char getHexDigit(int N) {
55   assert(N < 16);
56   if (N < 10)
57     return '0' + N;
58   return 'a' + N - 10;
59 }
60 
61 raw_ostream &BlockMass::print(raw_ostream &OS) const {
62   for (int Digits = 0; Digits < 16; ++Digits)
63     OS << getHexDigit(Mass >> (60 - Digits * 4) & 0xf);
64   return OS;
65 }
66 
67 namespace {
68 
69 using BlockNode = BlockFrequencyInfoImplBase::BlockNode;
70 using Distribution = BlockFrequencyInfoImplBase::Distribution;
71 using WeightList = BlockFrequencyInfoImplBase::Distribution::WeightList;
72 using Scaled64 = BlockFrequencyInfoImplBase::Scaled64;
73 using LoopData = BlockFrequencyInfoImplBase::LoopData;
74 using Weight = BlockFrequencyInfoImplBase::Weight;
75 using FrequencyData = BlockFrequencyInfoImplBase::FrequencyData;
76 
77 /// Dithering mass distributer.
78 ///
79 /// This class splits up a single mass into portions by weight, dithering to
80 /// spread out error.  No mass is lost.  The dithering precision depends on the
81 /// precision of the product of \a BlockMass and \a BranchProbability.
82 ///
83 /// The distribution algorithm follows.
84 ///
85 ///  1. Initialize by saving the sum of the weights in \a RemWeight and the
86 ///     mass to distribute in \a RemMass.
87 ///
88 ///  2. For each portion:
89 ///
90 ///      1. Construct a branch probability, P, as the portion's weight divided
91 ///         by the current value of \a RemWeight.
92 ///      2. Calculate the portion's mass as \a RemMass times P.
93 ///      3. Update \a RemWeight and \a RemMass at each portion by subtracting
94 ///         the current portion's weight and mass.
95 struct DitheringDistributer {
96   uint32_t RemWeight;
97   BlockMass RemMass;
98 
99   DitheringDistributer(Distribution &Dist, const BlockMass &Mass);
100 
101   BlockMass takeMass(uint32_t Weight);
102 };
103 
104 } // end anonymous namespace
105 
106 DitheringDistributer::DitheringDistributer(Distribution &Dist,
107                                            const BlockMass &Mass) {
108   Dist.normalize();
109   RemWeight = Dist.Total;
110   RemMass = Mass;
111 }
112 
113 BlockMass DitheringDistributer::takeMass(uint32_t Weight) {
114   assert(Weight && "invalid weight");
115   assert(Weight <= RemWeight);
116   BlockMass Mass = RemMass * BranchProbability(Weight, RemWeight);
117 
118   // Decrement totals (dither).
119   RemWeight -= Weight;
120   RemMass -= Mass;
121   return Mass;
122 }
123 
124 void Distribution::add(const BlockNode &Node, uint64_t Amount,
125                        Weight::DistType Type) {
126   assert(Amount && "invalid weight of 0");
127   uint64_t NewTotal = Total + Amount;
128 
129   // Check for overflow.  It should be impossible to overflow twice.
130   bool IsOverflow = NewTotal < Total;
131   assert(!(DidOverflow && IsOverflow) && "unexpected repeated overflow");
132   DidOverflow |= IsOverflow;
133 
134   // Update the total.
135   Total = NewTotal;
136 
137   // Save the weight.
138   Weights.push_back(Weight(Type, Node, Amount));
139 }
140 
141 static void combineWeight(Weight &W, const Weight &OtherW) {
142   assert(OtherW.TargetNode.isValid());
143   if (!W.Amount) {
144     W = OtherW;
145     return;
146   }
147   assert(W.Type == OtherW.Type);
148   assert(W.TargetNode == OtherW.TargetNode);
149   assert(OtherW.Amount && "Expected non-zero weight");
150   if (W.Amount > W.Amount + OtherW.Amount)
151     // Saturate on overflow.
152     W.Amount = UINT64_MAX;
153   else
154     W.Amount += OtherW.Amount;
155 }
156 
157 static void combineWeightsBySorting(WeightList &Weights) {
158   // Sort so edges to the same node are adjacent.
159   llvm::sort(Weights.begin(), Weights.end(),
160              [](const Weight &L,
161                 const Weight &R) { return L.TargetNode < R.TargetNode; });
162 
163   // Combine adjacent edges.
164   WeightList::iterator O = Weights.begin();
165   for (WeightList::const_iterator I = O, L = O, E = Weights.end(); I != E;
166        ++O, (I = L)) {
167     *O = *I;
168 
169     // Find the adjacent weights to the same node.
170     for (++L; L != E && I->TargetNode == L->TargetNode; ++L)
171       combineWeight(*O, *L);
172   }
173 
174   // Erase extra entries.
175   Weights.erase(O, Weights.end());
176 }
177 
178 static void combineWeightsByHashing(WeightList &Weights) {
179   // Collect weights into a DenseMap.
180   using HashTable = DenseMap<BlockNode::IndexType, Weight>;
181 
182   HashTable Combined(NextPowerOf2(2 * Weights.size()));
183   for (const Weight &W : Weights)
184     combineWeight(Combined[W.TargetNode.Index], W);
185 
186   // Check whether anything changed.
187   if (Weights.size() == Combined.size())
188     return;
189 
190   // Fill in the new weights.
191   Weights.clear();
192   Weights.reserve(Combined.size());
193   for (const auto &I : Combined)
194     Weights.push_back(I.second);
195 }
196 
197 static void combineWeights(WeightList &Weights) {
198   // Use a hash table for many successors to keep this linear.
199   if (Weights.size() > 128) {
200     combineWeightsByHashing(Weights);
201     return;
202   }
203 
204   combineWeightsBySorting(Weights);
205 }
206 
207 static uint64_t shiftRightAndRound(uint64_t N, int Shift) {
208   assert(Shift >= 0);
209   assert(Shift < 64);
210   if (!Shift)
211     return N;
212   return (N >> Shift) + (UINT64_C(1) & N >> (Shift - 1));
213 }
214 
215 void Distribution::normalize() {
216   // Early exit for termination nodes.
217   if (Weights.empty())
218     return;
219 
220   // Only bother if there are multiple successors.
221   if (Weights.size() > 1)
222     combineWeights(Weights);
223 
224   // Early exit when combined into a single successor.
225   if (Weights.size() == 1) {
226     Total = 1;
227     Weights.front().Amount = 1;
228     return;
229   }
230 
231   // Determine how much to shift right so that the total fits into 32-bits.
232   //
233   // If we shift at all, shift by 1 extra.  Otherwise, the lower limit of 1
234   // for each weight can cause a 32-bit overflow.
235   int Shift = 0;
236   if (DidOverflow)
237     Shift = 33;
238   else if (Total > UINT32_MAX)
239     Shift = 33 - countLeadingZeros(Total);
240 
241   // Early exit if nothing needs to be scaled.
242   if (!Shift) {
243     // If we didn't overflow then combineWeights() shouldn't have changed the
244     // sum of the weights, but let's double-check.
245     assert(Total == std::accumulate(Weights.begin(), Weights.end(), UINT64_C(0),
246                                     [](uint64_t Sum, const Weight &W) {
247                       return Sum + W.Amount;
248                     }) &&
249            "Expected total to be correct");
250     return;
251   }
252 
253   // Recompute the total through accumulation (rather than shifting it) so that
254   // it's accurate after shifting and any changes combineWeights() made above.
255   Total = 0;
256 
257   // Sum the weights to each node and shift right if necessary.
258   for (Weight &W : Weights) {
259     // Scale down below UINT32_MAX.  Since Shift is larger than necessary, we
260     // can round here without concern about overflow.
261     assert(W.TargetNode.isValid());
262     W.Amount = std::max(UINT64_C(1), shiftRightAndRound(W.Amount, Shift));
263     assert(W.Amount <= UINT32_MAX);
264 
265     // Update the total.
266     Total += W.Amount;
267   }
268   assert(Total <= UINT32_MAX);
269 }
270 
271 void BlockFrequencyInfoImplBase::clear() {
272   // Swap with a default-constructed std::vector, since std::vector<>::clear()
273   // does not actually clear heap storage.
274   std::vector<FrequencyData>().swap(Freqs);
275   IsIrrLoopHeader.clear();
276   std::vector<WorkingData>().swap(Working);
277   Loops.clear();
278 }
279 
280 /// Clear all memory not needed downstream.
281 ///
282 /// Releases all memory not used downstream.  In particular, saves Freqs.
283 static void cleanup(BlockFrequencyInfoImplBase &BFI) {
284   std::vector<FrequencyData> SavedFreqs(std::move(BFI.Freqs));
285   SparseBitVector<> SavedIsIrrLoopHeader(std::move(BFI.IsIrrLoopHeader));
286   BFI.clear();
287   BFI.Freqs = std::move(SavedFreqs);
288   BFI.IsIrrLoopHeader = std::move(SavedIsIrrLoopHeader);
289 }
290 
291 bool BlockFrequencyInfoImplBase::addToDist(Distribution &Dist,
292                                            const LoopData *OuterLoop,
293                                            const BlockNode &Pred,
294                                            const BlockNode &Succ,
295                                            uint64_t Weight) {
296   if (!Weight)
297     Weight = 1;
298 
299   auto isLoopHeader = [&OuterLoop](const BlockNode &Node) {
300     return OuterLoop && OuterLoop->isHeader(Node);
301   };
302 
303   BlockNode Resolved = Working[Succ.Index].getResolvedNode();
304 
305 #ifndef NDEBUG
306   auto debugSuccessor = [&](const char *Type) {
307     dbgs() << "  =>"
308            << " [" << Type << "] weight = " << Weight;
309     if (!isLoopHeader(Resolved))
310       dbgs() << ", succ = " << getBlockName(Succ);
311     if (Resolved != Succ)
312       dbgs() << ", resolved = " << getBlockName(Resolved);
313     dbgs() << "\n";
314   };
315   (void)debugSuccessor;
316 #endif
317 
318   if (isLoopHeader(Resolved)) {
319     LLVM_DEBUG(debugSuccessor("backedge"));
320     Dist.addBackedge(Resolved, Weight);
321     return true;
322   }
323 
324   if (Working[Resolved.Index].getContainingLoop() != OuterLoop) {
325     LLVM_DEBUG(debugSuccessor("  exit  "));
326     Dist.addExit(Resolved, Weight);
327     return true;
328   }
329 
330   if (Resolved < Pred) {
331     if (!isLoopHeader(Pred)) {
332       // If OuterLoop is an irreducible loop, we can't actually handle this.
333       assert((!OuterLoop || !OuterLoop->isIrreducible()) &&
334              "unhandled irreducible control flow");
335 
336       // Irreducible backedge.  Abort.
337       LLVM_DEBUG(debugSuccessor("abort!!!"));
338       return false;
339     }
340 
341     // If "Pred" is a loop header, then this isn't really a backedge; rather,
342     // OuterLoop must be irreducible.  These false backedges can come only from
343     // secondary loop headers.
344     assert(OuterLoop && OuterLoop->isIrreducible() && !isLoopHeader(Resolved) &&
345            "unhandled irreducible control flow");
346   }
347 
348   LLVM_DEBUG(debugSuccessor(" local  "));
349   Dist.addLocal(Resolved, Weight);
350   return true;
351 }
352 
353 bool BlockFrequencyInfoImplBase::addLoopSuccessorsToDist(
354     const LoopData *OuterLoop, LoopData &Loop, Distribution &Dist) {
355   // Copy the exit map into Dist.
356   for (const auto &I : Loop.Exits)
357     if (!addToDist(Dist, OuterLoop, Loop.getHeader(), I.first,
358                    I.second.getMass()))
359       // Irreducible backedge.
360       return false;
361 
362   return true;
363 }
364 
365 /// Compute the loop scale for a loop.
366 void BlockFrequencyInfoImplBase::computeLoopScale(LoopData &Loop) {
367   // Compute loop scale.
368   LLVM_DEBUG(dbgs() << "compute-loop-scale: " << getLoopName(Loop) << "\n");
369 
370   // Infinite loops need special handling. If we give the back edge an infinite
371   // mass, they may saturate all the other scales in the function down to 1,
372   // making all the other region temperatures look exactly the same. Choose an
373   // arbitrary scale to avoid these issues.
374   //
375   // FIXME: An alternate way would be to select a symbolic scale which is later
376   // replaced to be the maximum of all computed scales plus 1. This would
377   // appropriately describe the loop as having a large scale, without skewing
378   // the final frequency computation.
379   const Scaled64 InfiniteLoopScale(1, 12);
380 
381   // LoopScale == 1 / ExitMass
382   // ExitMass == HeadMass - BackedgeMass
383   BlockMass TotalBackedgeMass;
384   for (auto &Mass : Loop.BackedgeMass)
385     TotalBackedgeMass += Mass;
386   BlockMass ExitMass = BlockMass::getFull() - TotalBackedgeMass;
387 
388   // Block scale stores the inverse of the scale. If this is an infinite loop,
389   // its exit mass will be zero. In this case, use an arbitrary scale for the
390   // loop scale.
391   Loop.Scale =
392       ExitMass.isEmpty() ? InfiniteLoopScale : ExitMass.toScaled().inverse();
393 
394   LLVM_DEBUG(dbgs() << " - exit-mass = " << ExitMass << " ("
395                     << BlockMass::getFull() << " - " << TotalBackedgeMass
396                     << ")\n"
397                     << " - scale = " << Loop.Scale << "\n");
398 }
399 
400 /// Package up a loop.
401 void BlockFrequencyInfoImplBase::packageLoop(LoopData &Loop) {
402   LLVM_DEBUG(dbgs() << "packaging-loop: " << getLoopName(Loop) << "\n");
403 
404   // Clear the subloop exits to prevent quadratic memory usage.
405   for (const BlockNode &M : Loop.Nodes) {
406     if (auto *Loop = Working[M.Index].getPackagedLoop())
407       Loop->Exits.clear();
408     LLVM_DEBUG(dbgs() << " - node: " << getBlockName(M.Index) << "\n");
409   }
410   Loop.IsPackaged = true;
411 }
412 
413 #ifndef NDEBUG
414 static void debugAssign(const BlockFrequencyInfoImplBase &BFI,
415                         const DitheringDistributer &D, const BlockNode &T,
416                         const BlockMass &M, const char *Desc) {
417   dbgs() << "  => assign " << M << " (" << D.RemMass << ")";
418   if (Desc)
419     dbgs() << " [" << Desc << "]";
420   if (T.isValid())
421     dbgs() << " to " << BFI.getBlockName(T);
422   dbgs() << "\n";
423 }
424 #endif
425 
426 void BlockFrequencyInfoImplBase::distributeMass(const BlockNode &Source,
427                                                 LoopData *OuterLoop,
428                                                 Distribution &Dist) {
429   BlockMass Mass = Working[Source.Index].getMass();
430   LLVM_DEBUG(dbgs() << "  => mass:  " << Mass << "\n");
431 
432   // Distribute mass to successors as laid out in Dist.
433   DitheringDistributer D(Dist, Mass);
434 
435   for (const Weight &W : Dist.Weights) {
436     // Check for a local edge (non-backedge and non-exit).
437     BlockMass Taken = D.takeMass(W.Amount);
438     if (W.Type == Weight::Local) {
439       Working[W.TargetNode.Index].getMass() += Taken;
440       LLVM_DEBUG(debugAssign(*this, D, W.TargetNode, Taken, nullptr));
441       continue;
442     }
443 
444     // Backedges and exits only make sense if we're processing a loop.
445     assert(OuterLoop && "backedge or exit outside of loop");
446 
447     // Check for a backedge.
448     if (W.Type == Weight::Backedge) {
449       OuterLoop->BackedgeMass[OuterLoop->getHeaderIndex(W.TargetNode)] += Taken;
450       LLVM_DEBUG(debugAssign(*this, D, W.TargetNode, Taken, "back"));
451       continue;
452     }
453 
454     // This must be an exit.
455     assert(W.Type == Weight::Exit);
456     OuterLoop->Exits.push_back(std::make_pair(W.TargetNode, Taken));
457     LLVM_DEBUG(debugAssign(*this, D, W.TargetNode, Taken, "exit"));
458   }
459 }
460 
461 static void convertFloatingToInteger(BlockFrequencyInfoImplBase &BFI,
462                                      const Scaled64 &Min, const Scaled64 &Max) {
463   // Scale the Factor to a size that creates integers.  Ideally, integers would
464   // be scaled so that Max == UINT64_MAX so that they can be best
465   // differentiated.  However, in the presence of large frequency values, small
466   // frequencies are scaled down to 1, making it impossible to differentiate
467   // small, unequal numbers. When the spread between Min and Max frequencies
468   // fits well within MaxBits, we make the scale be at least 8.
469   const unsigned MaxBits = 64;
470   const unsigned SpreadBits = (Max / Min).lg();
471   Scaled64 ScalingFactor;
472   if (SpreadBits <= MaxBits - 3) {
473     // If the values are small enough, make the scaling factor at least 8 to
474     // allow distinguishing small values.
475     ScalingFactor = Min.inverse();
476     ScalingFactor <<= 3;
477   } else {
478     // If the values need more than MaxBits to be represented, saturate small
479     // frequency values down to 1 by using a scaling factor that benefits large
480     // frequency values.
481     ScalingFactor = Scaled64(1, MaxBits) / Max;
482   }
483 
484   // Translate the floats to integers.
485   LLVM_DEBUG(dbgs() << "float-to-int: min = " << Min << ", max = " << Max
486                     << ", factor = " << ScalingFactor << "\n");
487   for (size_t Index = 0; Index < BFI.Freqs.size(); ++Index) {
488     Scaled64 Scaled = BFI.Freqs[Index].Scaled * ScalingFactor;
489     BFI.Freqs[Index].Integer = std::max(UINT64_C(1), Scaled.toInt<uint64_t>());
490     LLVM_DEBUG(dbgs() << " - " << BFI.getBlockName(Index) << ": float = "
491                       << BFI.Freqs[Index].Scaled << ", scaled = " << Scaled
492                       << ", int = " << BFI.Freqs[Index].Integer << "\n");
493   }
494 }
495 
496 /// Unwrap a loop package.
497 ///
498 /// Visits all the members of a loop, adjusting their BlockData according to
499 /// the loop's pseudo-node.
500 static void unwrapLoop(BlockFrequencyInfoImplBase &BFI, LoopData &Loop) {
501   LLVM_DEBUG(dbgs() << "unwrap-loop-package: " << BFI.getLoopName(Loop)
502                     << ": mass = " << Loop.Mass << ", scale = " << Loop.Scale
503                     << "\n");
504   Loop.Scale *= Loop.Mass.toScaled();
505   Loop.IsPackaged = false;
506   LLVM_DEBUG(dbgs() << "  => combined-scale = " << Loop.Scale << "\n");
507 
508   // Propagate the head scale through the loop.  Since members are visited in
509   // RPO, the head scale will be updated by the loop scale first, and then the
510   // final head scale will be used for updated the rest of the members.
511   for (const BlockNode &N : Loop.Nodes) {
512     const auto &Working = BFI.Working[N.Index];
513     Scaled64 &F = Working.isAPackage() ? Working.getPackagedLoop()->Scale
514                                        : BFI.Freqs[N.Index].Scaled;
515     Scaled64 New = Loop.Scale * F;
516     LLVM_DEBUG(dbgs() << " - " << BFI.getBlockName(N) << ": " << F << " => "
517                       << New << "\n");
518     F = New;
519   }
520 }
521 
522 void BlockFrequencyInfoImplBase::unwrapLoops() {
523   // Set initial frequencies from loop-local masses.
524   for (size_t Index = 0; Index < Working.size(); ++Index)
525     Freqs[Index].Scaled = Working[Index].Mass.toScaled();
526 
527   for (LoopData &Loop : Loops)
528     unwrapLoop(*this, Loop);
529 }
530 
531 void BlockFrequencyInfoImplBase::finalizeMetrics() {
532   // Unwrap loop packages in reverse post-order, tracking min and max
533   // frequencies.
534   auto Min = Scaled64::getLargest();
535   auto Max = Scaled64::getZero();
536   for (size_t Index = 0; Index < Working.size(); ++Index) {
537     // Update min/max scale.
538     Min = std::min(Min, Freqs[Index].Scaled);
539     Max = std::max(Max, Freqs[Index].Scaled);
540   }
541 
542   // Convert to integers.
543   convertFloatingToInteger(*this, Min, Max);
544 
545   // Clean up data structures.
546   cleanup(*this);
547 
548   // Print out the final stats.
549   LLVM_DEBUG(dump());
550 }
551 
552 BlockFrequency
553 BlockFrequencyInfoImplBase::getBlockFreq(const BlockNode &Node) const {
554   if (!Node.isValid())
555     return 0;
556   return Freqs[Node.Index].Integer;
557 }
558 
559 Optional<uint64_t>
560 BlockFrequencyInfoImplBase::getBlockProfileCount(const Function &F,
561                                                  const BlockNode &Node) const {
562   return getProfileCountFromFreq(F, getBlockFreq(Node).getFrequency());
563 }
564 
565 Optional<uint64_t>
566 BlockFrequencyInfoImplBase::getProfileCountFromFreq(const Function &F,
567                                                     uint64_t Freq) const {
568   auto EntryCount = F.getEntryCount();
569   if (!EntryCount)
570     return None;
571   // Use 128 bit APInt to do the arithmetic to avoid overflow.
572   APInt BlockCount(128, EntryCount.getCount());
573   APInt BlockFreq(128, Freq);
574   APInt EntryFreq(128, getEntryFreq());
575   BlockCount *= BlockFreq;
576   BlockCount = BlockCount.udiv(EntryFreq);
577   return BlockCount.getLimitedValue();
578 }
579 
580 bool
581 BlockFrequencyInfoImplBase::isIrrLoopHeader(const BlockNode &Node) {
582   if (!Node.isValid())
583     return false;
584   return IsIrrLoopHeader.test(Node.Index);
585 }
586 
587 Scaled64
588 BlockFrequencyInfoImplBase::getFloatingBlockFreq(const BlockNode &Node) const {
589   if (!Node.isValid())
590     return Scaled64::getZero();
591   return Freqs[Node.Index].Scaled;
592 }
593 
594 void BlockFrequencyInfoImplBase::setBlockFreq(const BlockNode &Node,
595                                               uint64_t Freq) {
596   assert(Node.isValid() && "Expected valid node");
597   assert(Node.Index < Freqs.size() && "Expected legal index");
598   Freqs[Node.Index].Integer = Freq;
599 }
600 
601 std::string
602 BlockFrequencyInfoImplBase::getBlockName(const BlockNode &Node) const {
603   return {};
604 }
605 
606 std::string
607 BlockFrequencyInfoImplBase::getLoopName(const LoopData &Loop) const {
608   return getBlockName(Loop.getHeader()) + (Loop.isIrreducible() ? "**" : "*");
609 }
610 
611 raw_ostream &
612 BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
613                                            const BlockNode &Node) const {
614   return OS << getFloatingBlockFreq(Node);
615 }
616 
617 raw_ostream &
618 BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
619                                            const BlockFrequency &Freq) const {
620   Scaled64 Block(Freq.getFrequency(), 0);
621   Scaled64 Entry(getEntryFreq(), 0);
622 
623   return OS << Block / Entry;
624 }
625 
626 void IrreducibleGraph::addNodesInLoop(const BFIBase::LoopData &OuterLoop) {
627   Start = OuterLoop.getHeader();
628   Nodes.reserve(OuterLoop.Nodes.size());
629   for (auto N : OuterLoop.Nodes)
630     addNode(N);
631   indexNodes();
632 }
633 
634 void IrreducibleGraph::addNodesInFunction() {
635   Start = 0;
636   for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index)
637     if (!BFI.Working[Index].isPackaged())
638       addNode(Index);
639   indexNodes();
640 }
641 
642 void IrreducibleGraph::indexNodes() {
643   for (auto &I : Nodes)
644     Lookup[I.Node.Index] = &I;
645 }
646 
647 void IrreducibleGraph::addEdge(IrrNode &Irr, const BlockNode &Succ,
648                                const BFIBase::LoopData *OuterLoop) {
649   if (OuterLoop && OuterLoop->isHeader(Succ))
650     return;
651   auto L = Lookup.find(Succ.Index);
652   if (L == Lookup.end())
653     return;
654   IrrNode &SuccIrr = *L->second;
655   Irr.Edges.push_back(&SuccIrr);
656   SuccIrr.Edges.push_front(&Irr);
657   ++SuccIrr.NumIn;
658 }
659 
660 namespace llvm {
661 
662 template <> struct GraphTraits<IrreducibleGraph> {
663   using GraphT = bfi_detail::IrreducibleGraph;
664   using NodeRef = const GraphT::IrrNode *;
665   using ChildIteratorType = GraphT::IrrNode::iterator;
666 
667   static NodeRef getEntryNode(const GraphT &G) { return G.StartIrr; }
668   static ChildIteratorType child_begin(NodeRef N) { return N->succ_begin(); }
669   static ChildIteratorType child_end(NodeRef N) { return N->succ_end(); }
670 };
671 
672 } // end namespace llvm
673 
674 /// Find extra irreducible headers.
675 ///
676 /// Find entry blocks and other blocks with backedges, which exist when \c G
677 /// contains irreducible sub-SCCs.
678 static void findIrreducibleHeaders(
679     const BlockFrequencyInfoImplBase &BFI,
680     const IrreducibleGraph &G,
681     const std::vector<const IrreducibleGraph::IrrNode *> &SCC,
682     LoopData::NodeList &Headers, LoopData::NodeList &Others) {
683   // Map from nodes in the SCC to whether it's an entry block.
684   SmallDenseMap<const IrreducibleGraph::IrrNode *, bool, 8> InSCC;
685 
686   // InSCC also acts the set of nodes in the graph.  Seed it.
687   for (const auto *I : SCC)
688     InSCC[I] = false;
689 
690   for (auto I = InSCC.begin(), E = InSCC.end(); I != E; ++I) {
691     auto &Irr = *I->first;
692     for (const auto *P : make_range(Irr.pred_begin(), Irr.pred_end())) {
693       if (InSCC.count(P))
694         continue;
695 
696       // This is an entry block.
697       I->second = true;
698       Headers.push_back(Irr.Node);
699       LLVM_DEBUG(dbgs() << "  => entry = " << BFI.getBlockName(Irr.Node)
700                         << "\n");
701       break;
702     }
703   }
704   assert(Headers.size() >= 2 &&
705          "Expected irreducible CFG; -loop-info is likely invalid");
706   if (Headers.size() == InSCC.size()) {
707     // Every block is a header.
708     llvm::sort(Headers.begin(), Headers.end());
709     return;
710   }
711 
712   // Look for extra headers from irreducible sub-SCCs.
713   for (const auto &I : InSCC) {
714     // Entry blocks are already headers.
715     if (I.second)
716       continue;
717 
718     auto &Irr = *I.first;
719     for (const auto *P : make_range(Irr.pred_begin(), Irr.pred_end())) {
720       // Skip forward edges.
721       if (P->Node < Irr.Node)
722         continue;
723 
724       // Skip predecessors from entry blocks.  These can have inverted
725       // ordering.
726       if (InSCC.lookup(P))
727         continue;
728 
729       // Store the extra header.
730       Headers.push_back(Irr.Node);
731       LLVM_DEBUG(dbgs() << "  => extra = " << BFI.getBlockName(Irr.Node)
732                         << "\n");
733       break;
734     }
735     if (Headers.back() == Irr.Node)
736       // Added this as a header.
737       continue;
738 
739     // This is not a header.
740     Others.push_back(Irr.Node);
741     LLVM_DEBUG(dbgs() << "  => other = " << BFI.getBlockName(Irr.Node) << "\n");
742   }
743   llvm::sort(Headers.begin(), Headers.end());
744   llvm::sort(Others.begin(), Others.end());
745 }
746 
747 static void createIrreducibleLoop(
748     BlockFrequencyInfoImplBase &BFI, const IrreducibleGraph &G,
749     LoopData *OuterLoop, std::list<LoopData>::iterator Insert,
750     const std::vector<const IrreducibleGraph::IrrNode *> &SCC) {
751   // Translate the SCC into RPO.
752   LLVM_DEBUG(dbgs() << " - found-scc\n");
753 
754   LoopData::NodeList Headers;
755   LoopData::NodeList Others;
756   findIrreducibleHeaders(BFI, G, SCC, Headers, Others);
757 
758   auto Loop = BFI.Loops.emplace(Insert, OuterLoop, Headers.begin(),
759                                 Headers.end(), Others.begin(), Others.end());
760 
761   // Update loop hierarchy.
762   for (const auto &N : Loop->Nodes)
763     if (BFI.Working[N.Index].isLoopHeader())
764       BFI.Working[N.Index].Loop->Parent = &*Loop;
765     else
766       BFI.Working[N.Index].Loop = &*Loop;
767 }
768 
769 iterator_range<std::list<LoopData>::iterator>
770 BlockFrequencyInfoImplBase::analyzeIrreducible(
771     const IrreducibleGraph &G, LoopData *OuterLoop,
772     std::list<LoopData>::iterator Insert) {
773   assert((OuterLoop == nullptr) == (Insert == Loops.begin()));
774   auto Prev = OuterLoop ? std::prev(Insert) : Loops.end();
775 
776   for (auto I = scc_begin(G); !I.isAtEnd(); ++I) {
777     if (I->size() < 2)
778       continue;
779 
780     // Translate the SCC into RPO.
781     createIrreducibleLoop(*this, G, OuterLoop, Insert, *I);
782   }
783 
784   if (OuterLoop)
785     return make_range(std::next(Prev), Insert);
786   return make_range(Loops.begin(), Insert);
787 }
788 
789 void
790 BlockFrequencyInfoImplBase::updateLoopWithIrreducible(LoopData &OuterLoop) {
791   OuterLoop.Exits.clear();
792   for (auto &Mass : OuterLoop.BackedgeMass)
793     Mass = BlockMass::getEmpty();
794   auto O = OuterLoop.Nodes.begin() + 1;
795   for (auto I = O, E = OuterLoop.Nodes.end(); I != E; ++I)
796     if (!Working[I->Index].isPackaged())
797       *O++ = *I;
798   OuterLoop.Nodes.erase(O, OuterLoop.Nodes.end());
799 }
800 
801 void BlockFrequencyInfoImplBase::adjustLoopHeaderMass(LoopData &Loop) {
802   assert(Loop.isIrreducible() && "this only makes sense on irreducible loops");
803 
804   // Since the loop has more than one header block, the mass flowing back into
805   // each header will be different. Adjust the mass in each header loop to
806   // reflect the masses flowing through back edges.
807   //
808   // To do this, we distribute the initial mass using the backedge masses
809   // as weights for the distribution.
810   BlockMass LoopMass = BlockMass::getFull();
811   Distribution Dist;
812 
813   LLVM_DEBUG(dbgs() << "adjust-loop-header-mass:\n");
814   for (uint32_t H = 0; H < Loop.NumHeaders; ++H) {
815     auto &HeaderNode = Loop.Nodes[H];
816     auto &BackedgeMass = Loop.BackedgeMass[Loop.getHeaderIndex(HeaderNode)];
817     LLVM_DEBUG(dbgs() << " - Add back edge mass for node "
818                       << getBlockName(HeaderNode) << ": " << BackedgeMass
819                       << "\n");
820     if (BackedgeMass.getMass() > 0)
821       Dist.addLocal(HeaderNode, BackedgeMass.getMass());
822     else
823       LLVM_DEBUG(dbgs() << "   Nothing added. Back edge mass is zero\n");
824   }
825 
826   DitheringDistributer D(Dist, LoopMass);
827 
828   LLVM_DEBUG(dbgs() << " Distribute loop mass " << LoopMass
829                     << " to headers using above weights\n");
830   for (const Weight &W : Dist.Weights) {
831     BlockMass Taken = D.takeMass(W.Amount);
832     assert(W.Type == Weight::Local && "all weights should be local");
833     Working[W.TargetNode.Index].getMass() = Taken;
834     LLVM_DEBUG(debugAssign(*this, D, W.TargetNode, Taken, nullptr));
835   }
836 }
837 
838 void BlockFrequencyInfoImplBase::distributeIrrLoopHeaderMass(Distribution &Dist) {
839   BlockMass LoopMass = BlockMass::getFull();
840   DitheringDistributer D(Dist, LoopMass);
841   for (const Weight &W : Dist.Weights) {
842     BlockMass Taken = D.takeMass(W.Amount);
843     assert(W.Type == Weight::Local && "all weights should be local");
844     Working[W.TargetNode.Index].getMass() = Taken;
845     LLVM_DEBUG(debugAssign(*this, D, W.TargetNode, Taken, nullptr));
846   }
847 }
848