• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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/SCCIterator.h"
16 #include "llvm/IR/Function.h"
17 #include "llvm/Support/raw_ostream.h"
18 #include <numeric>
19 
20 using namespace llvm;
21 using namespace llvm::bfi_detail;
22 
23 #define DEBUG_TYPE "block-freq"
24 
toScaled() const25 ScaledNumber<uint64_t> BlockMass::toScaled() const {
26   if (isFull())
27     return ScaledNumber<uint64_t>(1, 0);
28   return ScaledNumber<uint64_t>(getMass() + 1, -64);
29 }
30 
dump() const31 LLVM_DUMP_METHOD void BlockMass::dump() const { print(dbgs()); }
32 
getHexDigit(int N)33 static char getHexDigit(int N) {
34   assert(N < 16);
35   if (N < 10)
36     return '0' + N;
37   return 'a' + N - 10;
38 }
39 
print(raw_ostream & OS) const40 raw_ostream &BlockMass::print(raw_ostream &OS) const {
41   for (int Digits = 0; Digits < 16; ++Digits)
42     OS << getHexDigit(Mass >> (60 - Digits * 4) & 0xf);
43   return OS;
44 }
45 
46 namespace {
47 
48 typedef BlockFrequencyInfoImplBase::BlockNode BlockNode;
49 typedef BlockFrequencyInfoImplBase::Distribution Distribution;
50 typedef BlockFrequencyInfoImplBase::Distribution::WeightList WeightList;
51 typedef BlockFrequencyInfoImplBase::Scaled64 Scaled64;
52 typedef BlockFrequencyInfoImplBase::LoopData LoopData;
53 typedef BlockFrequencyInfoImplBase::Weight Weight;
54 typedef BlockFrequencyInfoImplBase::FrequencyData FrequencyData;
55 
56 /// \brief Dithering mass distributer.
57 ///
58 /// This class splits up a single mass into portions by weight, dithering to
59 /// spread out error.  No mass is lost.  The dithering precision depends on the
60 /// precision of the product of \a BlockMass and \a BranchProbability.
61 ///
62 /// The distribution algorithm follows.
63 ///
64 ///  1. Initialize by saving the sum of the weights in \a RemWeight and the
65 ///     mass to distribute in \a RemMass.
66 ///
67 ///  2. For each portion:
68 ///
69 ///      1. Construct a branch probability, P, as the portion's weight divided
70 ///         by the current value of \a RemWeight.
71 ///      2. Calculate the portion's mass as \a RemMass times P.
72 ///      3. Update \a RemWeight and \a RemMass at each portion by subtracting
73 ///         the current portion's weight and mass.
74 struct DitheringDistributer {
75   uint32_t RemWeight;
76   BlockMass RemMass;
77 
78   DitheringDistributer(Distribution &Dist, const BlockMass &Mass);
79 
80   BlockMass takeMass(uint32_t Weight);
81 };
82 
83 } // end anonymous namespace
84 
DitheringDistributer(Distribution & Dist,const BlockMass & Mass)85 DitheringDistributer::DitheringDistributer(Distribution &Dist,
86                                            const BlockMass &Mass) {
87   Dist.normalize();
88   RemWeight = Dist.Total;
89   RemMass = Mass;
90 }
91 
takeMass(uint32_t Weight)92 BlockMass DitheringDistributer::takeMass(uint32_t Weight) {
93   assert(Weight && "invalid weight");
94   assert(Weight <= RemWeight);
95   BlockMass Mass = RemMass * BranchProbability(Weight, RemWeight);
96 
97   // Decrement totals (dither).
98   RemWeight -= Weight;
99   RemMass -= Mass;
100   return Mass;
101 }
102 
add(const BlockNode & Node,uint64_t Amount,Weight::DistType Type)103 void Distribution::add(const BlockNode &Node, uint64_t Amount,
104                        Weight::DistType Type) {
105   assert(Amount && "invalid weight of 0");
106   uint64_t NewTotal = Total + Amount;
107 
108   // Check for overflow.  It should be impossible to overflow twice.
109   bool IsOverflow = NewTotal < Total;
110   assert(!(DidOverflow && IsOverflow) && "unexpected repeated overflow");
111   DidOverflow |= IsOverflow;
112 
113   // Update the total.
114   Total = NewTotal;
115 
116   // Save the weight.
117   Weights.push_back(Weight(Type, Node, Amount));
118 }
119 
combineWeight(Weight & W,const Weight & OtherW)120 static void combineWeight(Weight &W, const Weight &OtherW) {
121   assert(OtherW.TargetNode.isValid());
122   if (!W.Amount) {
123     W = OtherW;
124     return;
125   }
126   assert(W.Type == OtherW.Type);
127   assert(W.TargetNode == OtherW.TargetNode);
128   assert(OtherW.Amount && "Expected non-zero weight");
129   if (W.Amount > W.Amount + OtherW.Amount)
130     // Saturate on overflow.
131     W.Amount = UINT64_MAX;
132   else
133     W.Amount += OtherW.Amount;
134 }
135 
combineWeightsBySorting(WeightList & Weights)136 static void combineWeightsBySorting(WeightList &Weights) {
137   // Sort so edges to the same node are adjacent.
138   std::sort(Weights.begin(), Weights.end(),
139             [](const Weight &L,
140                const Weight &R) { return L.TargetNode < R.TargetNode; });
141 
142   // Combine adjacent edges.
143   WeightList::iterator O = Weights.begin();
144   for (WeightList::const_iterator I = O, L = O, E = Weights.end(); I != E;
145        ++O, (I = L)) {
146     *O = *I;
147 
148     // Find the adjacent weights to the same node.
149     for (++L; L != E && I->TargetNode == L->TargetNode; ++L)
150       combineWeight(*O, *L);
151   }
152 
153   // Erase extra entries.
154   Weights.erase(O, Weights.end());
155 }
156 
combineWeightsByHashing(WeightList & Weights)157 static void combineWeightsByHashing(WeightList &Weights) {
158   // Collect weights into a DenseMap.
159   typedef DenseMap<BlockNode::IndexType, Weight> HashTable;
160   HashTable Combined(NextPowerOf2(2 * Weights.size()));
161   for (const Weight &W : Weights)
162     combineWeight(Combined[W.TargetNode.Index], W);
163 
164   // Check whether anything changed.
165   if (Weights.size() == Combined.size())
166     return;
167 
168   // Fill in the new weights.
169   Weights.clear();
170   Weights.reserve(Combined.size());
171   for (const auto &I : Combined)
172     Weights.push_back(I.second);
173 }
174 
combineWeights(WeightList & Weights)175 static void combineWeights(WeightList &Weights) {
176   // Use a hash table for many successors to keep this linear.
177   if (Weights.size() > 128) {
178     combineWeightsByHashing(Weights);
179     return;
180   }
181 
182   combineWeightsBySorting(Weights);
183 }
184 
shiftRightAndRound(uint64_t N,int Shift)185 static uint64_t shiftRightAndRound(uint64_t N, int Shift) {
186   assert(Shift >= 0);
187   assert(Shift < 64);
188   if (!Shift)
189     return N;
190   return (N >> Shift) + (UINT64_C(1) & N >> (Shift - 1));
191 }
192 
normalize()193 void Distribution::normalize() {
194   // Early exit for termination nodes.
195   if (Weights.empty())
196     return;
197 
198   // Only bother if there are multiple successors.
199   if (Weights.size() > 1)
200     combineWeights(Weights);
201 
202   // Early exit when combined into a single successor.
203   if (Weights.size() == 1) {
204     Total = 1;
205     Weights.front().Amount = 1;
206     return;
207   }
208 
209   // Determine how much to shift right so that the total fits into 32-bits.
210   //
211   // If we shift at all, shift by 1 extra.  Otherwise, the lower limit of 1
212   // for each weight can cause a 32-bit overflow.
213   int Shift = 0;
214   if (DidOverflow)
215     Shift = 33;
216   else if (Total > UINT32_MAX)
217     Shift = 33 - countLeadingZeros(Total);
218 
219   // Early exit if nothing needs to be scaled.
220   if (!Shift) {
221     // If we didn't overflow then combineWeights() shouldn't have changed the
222     // sum of the weights, but let's double-check.
223     assert(Total == std::accumulate(Weights.begin(), Weights.end(), UINT64_C(0),
224                                     [](uint64_t Sum, const Weight &W) {
225                       return Sum + W.Amount;
226                     }) &&
227            "Expected total to be correct");
228     return;
229   }
230 
231   // Recompute the total through accumulation (rather than shifting it) so that
232   // it's accurate after shifting and any changes combineWeights() made above.
233   Total = 0;
234 
235   // Sum the weights to each node and shift right if necessary.
236   for (Weight &W : Weights) {
237     // Scale down below UINT32_MAX.  Since Shift is larger than necessary, we
238     // can round here without concern about overflow.
239     assert(W.TargetNode.isValid());
240     W.Amount = std::max(UINT64_C(1), shiftRightAndRound(W.Amount, Shift));
241     assert(W.Amount <= UINT32_MAX);
242 
243     // Update the total.
244     Total += W.Amount;
245   }
246   assert(Total <= UINT32_MAX);
247 }
248 
clear()249 void BlockFrequencyInfoImplBase::clear() {
250   // Swap with a default-constructed std::vector, since std::vector<>::clear()
251   // does not actually clear heap storage.
252   std::vector<FrequencyData>().swap(Freqs);
253   std::vector<WorkingData>().swap(Working);
254   Loops.clear();
255 }
256 
257 /// \brief Clear all memory not needed downstream.
258 ///
259 /// Releases all memory not used downstream.  In particular, saves Freqs.
cleanup(BlockFrequencyInfoImplBase & BFI)260 static void cleanup(BlockFrequencyInfoImplBase &BFI) {
261   std::vector<FrequencyData> SavedFreqs(std::move(BFI.Freqs));
262   BFI.clear();
263   BFI.Freqs = std::move(SavedFreqs);
264 }
265 
addToDist(Distribution & Dist,const LoopData * OuterLoop,const BlockNode & Pred,const BlockNode & Succ,uint64_t Weight)266 bool BlockFrequencyInfoImplBase::addToDist(Distribution &Dist,
267                                            const LoopData *OuterLoop,
268                                            const BlockNode &Pred,
269                                            const BlockNode &Succ,
270                                            uint64_t Weight) {
271   if (!Weight)
272     Weight = 1;
273 
274   auto isLoopHeader = [&OuterLoop](const BlockNode &Node) {
275     return OuterLoop && OuterLoop->isHeader(Node);
276   };
277 
278   BlockNode Resolved = Working[Succ.Index].getResolvedNode();
279 
280 #ifndef NDEBUG
281   auto debugSuccessor = [&](const char *Type) {
282     dbgs() << "  =>"
283            << " [" << Type << "] weight = " << Weight;
284     if (!isLoopHeader(Resolved))
285       dbgs() << ", succ = " << getBlockName(Succ);
286     if (Resolved != Succ)
287       dbgs() << ", resolved = " << getBlockName(Resolved);
288     dbgs() << "\n";
289   };
290   (void)debugSuccessor;
291 #endif
292 
293   if (isLoopHeader(Resolved)) {
294     DEBUG(debugSuccessor("backedge"));
295     Dist.addBackedge(Resolved, Weight);
296     return true;
297   }
298 
299   if (Working[Resolved.Index].getContainingLoop() != OuterLoop) {
300     DEBUG(debugSuccessor("  exit  "));
301     Dist.addExit(Resolved, Weight);
302     return true;
303   }
304 
305   if (Resolved < Pred) {
306     if (!isLoopHeader(Pred)) {
307       // If OuterLoop is an irreducible loop, we can't actually handle this.
308       assert((!OuterLoop || !OuterLoop->isIrreducible()) &&
309              "unhandled irreducible control flow");
310 
311       // Irreducible backedge.  Abort.
312       DEBUG(debugSuccessor("abort!!!"));
313       return false;
314     }
315 
316     // If "Pred" is a loop header, then this isn't really a backedge; rather,
317     // OuterLoop must be irreducible.  These false backedges can come only from
318     // secondary loop headers.
319     assert(OuterLoop && OuterLoop->isIrreducible() && !isLoopHeader(Resolved) &&
320            "unhandled irreducible control flow");
321   }
322 
323   DEBUG(debugSuccessor(" local  "));
324   Dist.addLocal(Resolved, Weight);
325   return true;
326 }
327 
addLoopSuccessorsToDist(const LoopData * OuterLoop,LoopData & Loop,Distribution & Dist)328 bool BlockFrequencyInfoImplBase::addLoopSuccessorsToDist(
329     const LoopData *OuterLoop, LoopData &Loop, Distribution &Dist) {
330   // Copy the exit map into Dist.
331   for (const auto &I : Loop.Exits)
332     if (!addToDist(Dist, OuterLoop, Loop.getHeader(), I.first,
333                    I.second.getMass()))
334       // Irreducible backedge.
335       return false;
336 
337   return true;
338 }
339 
340 /// \brief Compute the loop scale for a loop.
computeLoopScale(LoopData & Loop)341 void BlockFrequencyInfoImplBase::computeLoopScale(LoopData &Loop) {
342   // Compute loop scale.
343   DEBUG(dbgs() << "compute-loop-scale: " << getLoopName(Loop) << "\n");
344 
345   // Infinite loops need special handling. If we give the back edge an infinite
346   // mass, they may saturate all the other scales in the function down to 1,
347   // making all the other region temperatures look exactly the same. Choose an
348   // arbitrary scale to avoid these issues.
349   //
350   // FIXME: An alternate way would be to select a symbolic scale which is later
351   // replaced to be the maximum of all computed scales plus 1. This would
352   // appropriately describe the loop as having a large scale, without skewing
353   // the final frequency computation.
354   const Scaled64 InfiniteLoopScale(1, 12);
355 
356   // LoopScale == 1 / ExitMass
357   // ExitMass == HeadMass - BackedgeMass
358   BlockMass TotalBackedgeMass;
359   for (auto &Mass : Loop.BackedgeMass)
360     TotalBackedgeMass += Mass;
361   BlockMass ExitMass = BlockMass::getFull() - TotalBackedgeMass;
362 
363   // Block scale stores the inverse of the scale. If this is an infinite loop,
364   // its exit mass will be zero. In this case, use an arbitrary scale for the
365   // loop scale.
366   Loop.Scale =
367       ExitMass.isEmpty() ? InfiniteLoopScale : ExitMass.toScaled().inverse();
368 
369   DEBUG(dbgs() << " - exit-mass = " << ExitMass << " (" << BlockMass::getFull()
370                << " - " << TotalBackedgeMass << ")\n"
371                << " - scale = " << Loop.Scale << "\n");
372 }
373 
374 /// \brief Package up a loop.
packageLoop(LoopData & Loop)375 void BlockFrequencyInfoImplBase::packageLoop(LoopData &Loop) {
376   DEBUG(dbgs() << "packaging-loop: " << getLoopName(Loop) << "\n");
377 
378   // Clear the subloop exits to prevent quadratic memory usage.
379   for (const BlockNode &M : Loop.Nodes) {
380     if (auto *Loop = Working[M.Index].getPackagedLoop())
381       Loop->Exits.clear();
382     DEBUG(dbgs() << " - node: " << getBlockName(M.Index) << "\n");
383   }
384   Loop.IsPackaged = true;
385 }
386 
387 #ifndef NDEBUG
debugAssign(const BlockFrequencyInfoImplBase & BFI,const DitheringDistributer & D,const BlockNode & T,const BlockMass & M,const char * Desc)388 static void debugAssign(const BlockFrequencyInfoImplBase &BFI,
389                         const DitheringDistributer &D, const BlockNode &T,
390                         const BlockMass &M, const char *Desc) {
391   dbgs() << "  => assign " << M << " (" << D.RemMass << ")";
392   if (Desc)
393     dbgs() << " [" << Desc << "]";
394   if (T.isValid())
395     dbgs() << " to " << BFI.getBlockName(T);
396   dbgs() << "\n";
397 }
398 #endif
399 
distributeMass(const BlockNode & Source,LoopData * OuterLoop,Distribution & Dist)400 void BlockFrequencyInfoImplBase::distributeMass(const BlockNode &Source,
401                                                 LoopData *OuterLoop,
402                                                 Distribution &Dist) {
403   BlockMass Mass = Working[Source.Index].getMass();
404   DEBUG(dbgs() << "  => mass:  " << Mass << "\n");
405 
406   // Distribute mass to successors as laid out in Dist.
407   DitheringDistributer D(Dist, Mass);
408 
409   for (const Weight &W : Dist.Weights) {
410     // Check for a local edge (non-backedge and non-exit).
411     BlockMass Taken = D.takeMass(W.Amount);
412     if (W.Type == Weight::Local) {
413       Working[W.TargetNode.Index].getMass() += Taken;
414       DEBUG(debugAssign(*this, D, W.TargetNode, Taken, nullptr));
415       continue;
416     }
417 
418     // Backedges and exits only make sense if we're processing a loop.
419     assert(OuterLoop && "backedge or exit outside of loop");
420 
421     // Check for a backedge.
422     if (W.Type == Weight::Backedge) {
423       OuterLoop->BackedgeMass[OuterLoop->getHeaderIndex(W.TargetNode)] += Taken;
424       DEBUG(debugAssign(*this, D, W.TargetNode, Taken, "back"));
425       continue;
426     }
427 
428     // This must be an exit.
429     assert(W.Type == Weight::Exit);
430     OuterLoop->Exits.push_back(std::make_pair(W.TargetNode, Taken));
431     DEBUG(debugAssign(*this, D, W.TargetNode, Taken, "exit"));
432   }
433 }
434 
convertFloatingToInteger(BlockFrequencyInfoImplBase & BFI,const Scaled64 & Min,const Scaled64 & Max)435 static void convertFloatingToInteger(BlockFrequencyInfoImplBase &BFI,
436                                      const Scaled64 &Min, const Scaled64 &Max) {
437   // Scale the Factor to a size that creates integers.  Ideally, integers would
438   // be scaled so that Max == UINT64_MAX so that they can be best
439   // differentiated.  However, in the presence of large frequency values, small
440   // frequencies are scaled down to 1, making it impossible to differentiate
441   // small, unequal numbers. When the spread between Min and Max frequencies
442   // fits well within MaxBits, we make the scale be at least 8.
443   const unsigned MaxBits = 64;
444   const unsigned SpreadBits = (Max / Min).lg();
445   Scaled64 ScalingFactor;
446   if (SpreadBits <= MaxBits - 3) {
447     // If the values are small enough, make the scaling factor at least 8 to
448     // allow distinguishing small values.
449     ScalingFactor = Min.inverse();
450     ScalingFactor <<= 3;
451   } else {
452     // If the values need more than MaxBits to be represented, saturate small
453     // frequency values down to 1 by using a scaling factor that benefits large
454     // frequency values.
455     ScalingFactor = Scaled64(1, MaxBits) / Max;
456   }
457 
458   // Translate the floats to integers.
459   DEBUG(dbgs() << "float-to-int: min = " << Min << ", max = " << Max
460                << ", factor = " << ScalingFactor << "\n");
461   for (size_t Index = 0; Index < BFI.Freqs.size(); ++Index) {
462     Scaled64 Scaled = BFI.Freqs[Index].Scaled * ScalingFactor;
463     BFI.Freqs[Index].Integer = std::max(UINT64_C(1), Scaled.toInt<uint64_t>());
464     DEBUG(dbgs() << " - " << BFI.getBlockName(Index) << ": float = "
465                  << BFI.Freqs[Index].Scaled << ", scaled = " << Scaled
466                  << ", int = " << BFI.Freqs[Index].Integer << "\n");
467   }
468 }
469 
470 /// \brief Unwrap a loop package.
471 ///
472 /// Visits all the members of a loop, adjusting their BlockData according to
473 /// the loop's pseudo-node.
unwrapLoop(BlockFrequencyInfoImplBase & BFI,LoopData & Loop)474 static void unwrapLoop(BlockFrequencyInfoImplBase &BFI, LoopData &Loop) {
475   DEBUG(dbgs() << "unwrap-loop-package: " << BFI.getLoopName(Loop)
476                << ": mass = " << Loop.Mass << ", scale = " << Loop.Scale
477                << "\n");
478   Loop.Scale *= Loop.Mass.toScaled();
479   Loop.IsPackaged = false;
480   DEBUG(dbgs() << "  => combined-scale = " << Loop.Scale << "\n");
481 
482   // Propagate the head scale through the loop.  Since members are visited in
483   // RPO, the head scale will be updated by the loop scale first, and then the
484   // final head scale will be used for updated the rest of the members.
485   for (const BlockNode &N : Loop.Nodes) {
486     const auto &Working = BFI.Working[N.Index];
487     Scaled64 &F = Working.isAPackage() ? Working.getPackagedLoop()->Scale
488                                        : BFI.Freqs[N.Index].Scaled;
489     Scaled64 New = Loop.Scale * F;
490     DEBUG(dbgs() << " - " << BFI.getBlockName(N) << ": " << F << " => " << New
491                  << "\n");
492     F = New;
493   }
494 }
495 
unwrapLoops()496 void BlockFrequencyInfoImplBase::unwrapLoops() {
497   // Set initial frequencies from loop-local masses.
498   for (size_t Index = 0; Index < Working.size(); ++Index)
499     Freqs[Index].Scaled = Working[Index].Mass.toScaled();
500 
501   for (LoopData &Loop : Loops)
502     unwrapLoop(*this, Loop);
503 }
504 
finalizeMetrics()505 void BlockFrequencyInfoImplBase::finalizeMetrics() {
506   // Unwrap loop packages in reverse post-order, tracking min and max
507   // frequencies.
508   auto Min = Scaled64::getLargest();
509   auto Max = Scaled64::getZero();
510   for (size_t Index = 0; Index < Working.size(); ++Index) {
511     // Update min/max scale.
512     Min = std::min(Min, Freqs[Index].Scaled);
513     Max = std::max(Max, Freqs[Index].Scaled);
514   }
515 
516   // Convert to integers.
517   convertFloatingToInteger(*this, Min, Max);
518 
519   // Clean up data structures.
520   cleanup(*this);
521 
522   // Print out the final stats.
523   DEBUG(dump());
524 }
525 
526 BlockFrequency
getBlockFreq(const BlockNode & Node) const527 BlockFrequencyInfoImplBase::getBlockFreq(const BlockNode &Node) const {
528   if (!Node.isValid())
529     return 0;
530   return Freqs[Node.Index].Integer;
531 }
532 
533 Optional<uint64_t>
getBlockProfileCount(const Function & F,const BlockNode & Node) const534 BlockFrequencyInfoImplBase::getBlockProfileCount(const Function &F,
535                                                  const BlockNode &Node) const {
536   auto EntryCount = F.getEntryCount();
537   if (!EntryCount)
538     return None;
539   // Use 128 bit APInt to do the arithmetic to avoid overflow.
540   APInt BlockCount(128, EntryCount.getValue());
541   APInt BlockFreq(128, getBlockFreq(Node).getFrequency());
542   APInt EntryFreq(128, getEntryFreq());
543   BlockCount *= BlockFreq;
544   BlockCount = BlockCount.udiv(EntryFreq);
545   return BlockCount.getLimitedValue();
546 }
547 
548 Scaled64
getFloatingBlockFreq(const BlockNode & Node) const549 BlockFrequencyInfoImplBase::getFloatingBlockFreq(const BlockNode &Node) const {
550   if (!Node.isValid())
551     return Scaled64::getZero();
552   return Freqs[Node.Index].Scaled;
553 }
554 
setBlockFreq(const BlockNode & Node,uint64_t Freq)555 void BlockFrequencyInfoImplBase::setBlockFreq(const BlockNode &Node,
556                                               uint64_t Freq) {
557   assert(Node.isValid() && "Expected valid node");
558   assert(Node.Index < Freqs.size() && "Expected legal index");
559   Freqs[Node.Index].Integer = Freq;
560 }
561 
562 std::string
getBlockName(const BlockNode & Node) const563 BlockFrequencyInfoImplBase::getBlockName(const BlockNode &Node) const {
564   return std::string();
565 }
566 
567 std::string
getLoopName(const LoopData & Loop) const568 BlockFrequencyInfoImplBase::getLoopName(const LoopData &Loop) const {
569   return getBlockName(Loop.getHeader()) + (Loop.isIrreducible() ? "**" : "*");
570 }
571 
572 raw_ostream &
printBlockFreq(raw_ostream & OS,const BlockNode & Node) const573 BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
574                                            const BlockNode &Node) const {
575   return OS << getFloatingBlockFreq(Node);
576 }
577 
578 raw_ostream &
printBlockFreq(raw_ostream & OS,const BlockFrequency & Freq) const579 BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
580                                            const BlockFrequency &Freq) const {
581   Scaled64 Block(Freq.getFrequency(), 0);
582   Scaled64 Entry(getEntryFreq(), 0);
583 
584   return OS << Block / Entry;
585 }
586 
addNodesInLoop(const BFIBase::LoopData & OuterLoop)587 void IrreducibleGraph::addNodesInLoop(const BFIBase::LoopData &OuterLoop) {
588   Start = OuterLoop.getHeader();
589   Nodes.reserve(OuterLoop.Nodes.size());
590   for (auto N : OuterLoop.Nodes)
591     addNode(N);
592   indexNodes();
593 }
594 
addNodesInFunction()595 void IrreducibleGraph::addNodesInFunction() {
596   Start = 0;
597   for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index)
598     if (!BFI.Working[Index].isPackaged())
599       addNode(Index);
600   indexNodes();
601 }
602 
indexNodes()603 void IrreducibleGraph::indexNodes() {
604   for (auto &I : Nodes)
605     Lookup[I.Node.Index] = &I;
606 }
607 
addEdge(IrrNode & Irr,const BlockNode & Succ,const BFIBase::LoopData * OuterLoop)608 void IrreducibleGraph::addEdge(IrrNode &Irr, const BlockNode &Succ,
609                                const BFIBase::LoopData *OuterLoop) {
610   if (OuterLoop && OuterLoop->isHeader(Succ))
611     return;
612   auto L = Lookup.find(Succ.Index);
613   if (L == Lookup.end())
614     return;
615   IrrNode &SuccIrr = *L->second;
616   Irr.Edges.push_back(&SuccIrr);
617   SuccIrr.Edges.push_front(&Irr);
618   ++SuccIrr.NumIn;
619 }
620 
621 namespace llvm {
622 template <> struct GraphTraits<IrreducibleGraph> {
623   typedef bfi_detail::IrreducibleGraph GraphT;
624 
625   typedef const GraphT::IrrNode NodeType;
626   typedef GraphT::IrrNode::iterator ChildIteratorType;
627 
getEntryNodellvm::GraphTraits628   static const NodeType *getEntryNode(const GraphT &G) {
629     return G.StartIrr;
630   }
child_beginllvm::GraphTraits631   static ChildIteratorType child_begin(NodeType *N) { return N->succ_begin(); }
child_endllvm::GraphTraits632   static ChildIteratorType child_end(NodeType *N) { return N->succ_end(); }
633 };
634 } // end namespace llvm
635 
636 /// \brief Find extra irreducible headers.
637 ///
638 /// Find entry blocks and other blocks with backedges, which exist when \c G
639 /// contains irreducible sub-SCCs.
findIrreducibleHeaders(const BlockFrequencyInfoImplBase & BFI,const IrreducibleGraph & G,const std::vector<const IrreducibleGraph::IrrNode * > & SCC,LoopData::NodeList & Headers,LoopData::NodeList & Others)640 static void findIrreducibleHeaders(
641     const BlockFrequencyInfoImplBase &BFI,
642     const IrreducibleGraph &G,
643     const std::vector<const IrreducibleGraph::IrrNode *> &SCC,
644     LoopData::NodeList &Headers, LoopData::NodeList &Others) {
645   // Map from nodes in the SCC to whether it's an entry block.
646   SmallDenseMap<const IrreducibleGraph::IrrNode *, bool, 8> InSCC;
647 
648   // InSCC also acts the set of nodes in the graph.  Seed it.
649   for (const auto *I : SCC)
650     InSCC[I] = false;
651 
652   for (auto I = InSCC.begin(), E = InSCC.end(); I != E; ++I) {
653     auto &Irr = *I->first;
654     for (const auto *P : make_range(Irr.pred_begin(), Irr.pred_end())) {
655       if (InSCC.count(P))
656         continue;
657 
658       // This is an entry block.
659       I->second = true;
660       Headers.push_back(Irr.Node);
661       DEBUG(dbgs() << "  => entry = " << BFI.getBlockName(Irr.Node) << "\n");
662       break;
663     }
664   }
665   assert(Headers.size() >= 2 &&
666          "Expected irreducible CFG; -loop-info is likely invalid");
667   if (Headers.size() == InSCC.size()) {
668     // Every block is a header.
669     std::sort(Headers.begin(), Headers.end());
670     return;
671   }
672 
673   // Look for extra headers from irreducible sub-SCCs.
674   for (const auto &I : InSCC) {
675     // Entry blocks are already headers.
676     if (I.second)
677       continue;
678 
679     auto &Irr = *I.first;
680     for (const auto *P : make_range(Irr.pred_begin(), Irr.pred_end())) {
681       // Skip forward edges.
682       if (P->Node < Irr.Node)
683         continue;
684 
685       // Skip predecessors from entry blocks.  These can have inverted
686       // ordering.
687       if (InSCC.lookup(P))
688         continue;
689 
690       // Store the extra header.
691       Headers.push_back(Irr.Node);
692       DEBUG(dbgs() << "  => extra = " << BFI.getBlockName(Irr.Node) << "\n");
693       break;
694     }
695     if (Headers.back() == Irr.Node)
696       // Added this as a header.
697       continue;
698 
699     // This is not a header.
700     Others.push_back(Irr.Node);
701     DEBUG(dbgs() << "  => other = " << BFI.getBlockName(Irr.Node) << "\n");
702   }
703   std::sort(Headers.begin(), Headers.end());
704   std::sort(Others.begin(), Others.end());
705 }
706 
createIrreducibleLoop(BlockFrequencyInfoImplBase & BFI,const IrreducibleGraph & G,LoopData * OuterLoop,std::list<LoopData>::iterator Insert,const std::vector<const IrreducibleGraph::IrrNode * > & SCC)707 static void createIrreducibleLoop(
708     BlockFrequencyInfoImplBase &BFI, const IrreducibleGraph &G,
709     LoopData *OuterLoop, std::list<LoopData>::iterator Insert,
710     const std::vector<const IrreducibleGraph::IrrNode *> &SCC) {
711   // Translate the SCC into RPO.
712   DEBUG(dbgs() << " - found-scc\n");
713 
714   LoopData::NodeList Headers;
715   LoopData::NodeList Others;
716   findIrreducibleHeaders(BFI, G, SCC, Headers, Others);
717 
718   auto Loop = BFI.Loops.emplace(Insert, OuterLoop, Headers.begin(),
719                                 Headers.end(), Others.begin(), Others.end());
720 
721   // Update loop hierarchy.
722   for (const auto &N : Loop->Nodes)
723     if (BFI.Working[N.Index].isLoopHeader())
724       BFI.Working[N.Index].Loop->Parent = &*Loop;
725     else
726       BFI.Working[N.Index].Loop = &*Loop;
727 }
728 
729 iterator_range<std::list<LoopData>::iterator>
analyzeIrreducible(const IrreducibleGraph & G,LoopData * OuterLoop,std::list<LoopData>::iterator Insert)730 BlockFrequencyInfoImplBase::analyzeIrreducible(
731     const IrreducibleGraph &G, LoopData *OuterLoop,
732     std::list<LoopData>::iterator Insert) {
733   assert((OuterLoop == nullptr) == (Insert == Loops.begin()));
734   auto Prev = OuterLoop ? std::prev(Insert) : Loops.end();
735 
736   for (auto I = scc_begin(G); !I.isAtEnd(); ++I) {
737     if (I->size() < 2)
738       continue;
739 
740     // Translate the SCC into RPO.
741     createIrreducibleLoop(*this, G, OuterLoop, Insert, *I);
742   }
743 
744   if (OuterLoop)
745     return make_range(std::next(Prev), Insert);
746   return make_range(Loops.begin(), Insert);
747 }
748 
749 void
updateLoopWithIrreducible(LoopData & OuterLoop)750 BlockFrequencyInfoImplBase::updateLoopWithIrreducible(LoopData &OuterLoop) {
751   OuterLoop.Exits.clear();
752   for (auto &Mass : OuterLoop.BackedgeMass)
753     Mass = BlockMass::getEmpty();
754   auto O = OuterLoop.Nodes.begin() + 1;
755   for (auto I = O, E = OuterLoop.Nodes.end(); I != E; ++I)
756     if (!Working[I->Index].isPackaged())
757       *O++ = *I;
758   OuterLoop.Nodes.erase(O, OuterLoop.Nodes.end());
759 }
760 
adjustLoopHeaderMass(LoopData & Loop)761 void BlockFrequencyInfoImplBase::adjustLoopHeaderMass(LoopData &Loop) {
762   assert(Loop.isIrreducible() && "this only makes sense on irreducible loops");
763 
764   // Since the loop has more than one header block, the mass flowing back into
765   // each header will be different. Adjust the mass in each header loop to
766   // reflect the masses flowing through back edges.
767   //
768   // To do this, we distribute the initial mass using the backedge masses
769   // as weights for the distribution.
770   BlockMass LoopMass = BlockMass::getFull();
771   Distribution Dist;
772 
773   DEBUG(dbgs() << "adjust-loop-header-mass:\n");
774   for (uint32_t H = 0; H < Loop.NumHeaders; ++H) {
775     auto &HeaderNode = Loop.Nodes[H];
776     auto &BackedgeMass = Loop.BackedgeMass[Loop.getHeaderIndex(HeaderNode)];
777     DEBUG(dbgs() << " - Add back edge mass for node "
778                  << getBlockName(HeaderNode) << ": " << BackedgeMass << "\n");
779     if (BackedgeMass.getMass() > 0)
780       Dist.addLocal(HeaderNode, BackedgeMass.getMass());
781     else
782       DEBUG(dbgs() << "   Nothing added. Back edge mass is zero\n");
783   }
784 
785   DitheringDistributer D(Dist, LoopMass);
786 
787   DEBUG(dbgs() << " Distribute loop mass " << LoopMass
788                << " to headers using above weights\n");
789   for (const Weight &W : Dist.Weights) {
790     BlockMass Taken = D.takeMass(W.Amount);
791     assert(W.Type == Weight::Local && "all weights should be local");
792     Working[W.TargetNode.Index].getMass() = Taken;
793     DEBUG(debugAssign(*this, D, W.TargetNode, Taken, nullptr));
794   }
795 }
796