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/Support/raw_ostream.h"
17 #include <numeric>
18 
19 using namespace llvm;
20 using namespace llvm::bfi_detail;
21 
22 #define DEBUG_TYPE "block-freq"
23 
toScaled() const24 ScaledNumber<uint64_t> BlockMass::toScaled() const {
25   if (isFull())
26     return ScaledNumber<uint64_t>(1, 0);
27   return ScaledNumber<uint64_t>(getMass() + 1, -64);
28 }
29 
dump() const30 void BlockMass::dump() const { print(dbgs()); }
31 
getHexDigit(int N)32 static char getHexDigit(int N) {
33   assert(N < 16);
34   if (N < 10)
35     return '0' + N;
36   return 'a' + N - 10;
37 }
print(raw_ostream & OS) const38 raw_ostream &BlockMass::print(raw_ostream &OS) const {
39   for (int Digits = 0; Digits < 16; ++Digits)
40     OS << getHexDigit(Mass >> (60 - Digits * 4) & 0xf);
41   return OS;
42 }
43 
44 namespace {
45 
46 typedef BlockFrequencyInfoImplBase::BlockNode BlockNode;
47 typedef BlockFrequencyInfoImplBase::Distribution Distribution;
48 typedef BlockFrequencyInfoImplBase::Distribution::WeightList WeightList;
49 typedef BlockFrequencyInfoImplBase::Scaled64 Scaled64;
50 typedef BlockFrequencyInfoImplBase::LoopData LoopData;
51 typedef BlockFrequencyInfoImplBase::Weight Weight;
52 typedef BlockFrequencyInfoImplBase::FrequencyData FrequencyData;
53 
54 /// \brief Dithering mass distributer.
55 ///
56 /// This class splits up a single mass into portions by weight, dithering to
57 /// spread out error.  No mass is lost.  The dithering precision depends on the
58 /// precision of the product of \a BlockMass and \a BranchProbability.
59 ///
60 /// The distribution algorithm follows.
61 ///
62 ///  1. Initialize by saving the sum of the weights in \a RemWeight and the
63 ///     mass to distribute in \a RemMass.
64 ///
65 ///  2. For each portion:
66 ///
67 ///      1. Construct a branch probability, P, as the portion's weight divided
68 ///         by the current value of \a RemWeight.
69 ///      2. Calculate the portion's mass as \a RemMass times P.
70 ///      3. Update \a RemWeight and \a RemMass at each portion by subtracting
71 ///         the current portion's weight and mass.
72 struct DitheringDistributer {
73   uint32_t RemWeight;
74   BlockMass RemMass;
75 
76   DitheringDistributer(Distribution &Dist, const BlockMass &Mass);
77 
78   BlockMass takeMass(uint32_t Weight);
79 };
80 
81 } // end namespace
82 
DitheringDistributer(Distribution & Dist,const BlockMass & Mass)83 DitheringDistributer::DitheringDistributer(Distribution &Dist,
84                                            const BlockMass &Mass) {
85   Dist.normalize();
86   RemWeight = Dist.Total;
87   RemMass = Mass;
88 }
89 
takeMass(uint32_t Weight)90 BlockMass DitheringDistributer::takeMass(uint32_t Weight) {
91   assert(Weight && "invalid weight");
92   assert(Weight <= RemWeight);
93   BlockMass Mass = RemMass * BranchProbability(Weight, RemWeight);
94 
95   // Decrement totals (dither).
96   RemWeight -= Weight;
97   RemMass -= Mass;
98   return Mass;
99 }
100 
add(const BlockNode & Node,uint64_t Amount,Weight::DistType Type)101 void Distribution::add(const BlockNode &Node, uint64_t Amount,
102                        Weight::DistType Type) {
103   assert(Amount && "invalid weight of 0");
104   uint64_t NewTotal = Total + Amount;
105 
106   // Check for overflow.  It should be impossible to overflow twice.
107   bool IsOverflow = NewTotal < Total;
108   assert(!(DidOverflow && IsOverflow) && "unexpected repeated overflow");
109   DidOverflow |= IsOverflow;
110 
111   // Update the total.
112   Total = NewTotal;
113 
114   // Save the weight.
115   Weights.push_back(Weight(Type, Node, Amount));
116 }
117 
combineWeight(Weight & W,const Weight & OtherW)118 static void combineWeight(Weight &W, const Weight &OtherW) {
119   assert(OtherW.TargetNode.isValid());
120   if (!W.Amount) {
121     W = OtherW;
122     return;
123   }
124   assert(W.Type == OtherW.Type);
125   assert(W.TargetNode == OtherW.TargetNode);
126   assert(OtherW.Amount && "Expected non-zero weight");
127   if (W.Amount > W.Amount + OtherW.Amount)
128     // Saturate on overflow.
129     W.Amount = UINT64_MAX;
130   else
131     W.Amount += OtherW.Amount;
132 }
combineWeightsBySorting(WeightList & Weights)133 static void combineWeightsBySorting(WeightList &Weights) {
134   // Sort so edges to the same node are adjacent.
135   std::sort(Weights.begin(), Weights.end(),
136             [](const Weight &L,
137                const Weight &R) { return L.TargetNode < R.TargetNode; });
138 
139   // Combine adjacent edges.
140   WeightList::iterator O = Weights.begin();
141   for (WeightList::const_iterator I = O, L = O, E = Weights.end(); I != E;
142        ++O, (I = L)) {
143     *O = *I;
144 
145     // Find the adjacent weights to the same node.
146     for (++L; L != E && I->TargetNode == L->TargetNode; ++L)
147       combineWeight(*O, *L);
148   }
149 
150   // Erase extra entries.
151   Weights.erase(O, Weights.end());
152   return;
153 }
combineWeightsByHashing(WeightList & Weights)154 static void combineWeightsByHashing(WeightList &Weights) {
155   // Collect weights into a DenseMap.
156   typedef DenseMap<BlockNode::IndexType, Weight> HashTable;
157   HashTable Combined(NextPowerOf2(2 * Weights.size()));
158   for (const Weight &W : Weights)
159     combineWeight(Combined[W.TargetNode.Index], W);
160 
161   // Check whether anything changed.
162   if (Weights.size() == Combined.size())
163     return;
164 
165   // Fill in the new weights.
166   Weights.clear();
167   Weights.reserve(Combined.size());
168   for (const auto &I : Combined)
169     Weights.push_back(I.second);
170 }
combineWeights(WeightList & Weights)171 static void combineWeights(WeightList &Weights) {
172   // Use a hash table for many successors to keep this linear.
173   if (Weights.size() > 128) {
174     combineWeightsByHashing(Weights);
175     return;
176   }
177 
178   combineWeightsBySorting(Weights);
179 }
shiftRightAndRound(uint64_t N,int Shift)180 static uint64_t shiftRightAndRound(uint64_t N, int Shift) {
181   assert(Shift >= 0);
182   assert(Shift < 64);
183   if (!Shift)
184     return N;
185   return (N >> Shift) + (UINT64_C(1) & N >> (Shift - 1));
186 }
normalize()187 void Distribution::normalize() {
188   // Early exit for termination nodes.
189   if (Weights.empty())
190     return;
191 
192   // Only bother if there are multiple successors.
193   if (Weights.size() > 1)
194     combineWeights(Weights);
195 
196   // Early exit when combined into a single successor.
197   if (Weights.size() == 1) {
198     Total = 1;
199     Weights.front().Amount = 1;
200     return;
201   }
202 
203   // Determine how much to shift right so that the total fits into 32-bits.
204   //
205   // If we shift at all, shift by 1 extra.  Otherwise, the lower limit of 1
206   // for each weight can cause a 32-bit overflow.
207   int Shift = 0;
208   if (DidOverflow)
209     Shift = 33;
210   else if (Total > UINT32_MAX)
211     Shift = 33 - countLeadingZeros(Total);
212 
213   // Early exit if nothing needs to be scaled.
214   if (!Shift) {
215     // If we didn't overflow then combineWeights() shouldn't have changed the
216     // sum of the weights, but let's double-check.
217     assert(Total == std::accumulate(Weights.begin(), Weights.end(), UINT64_C(0),
218                                     [](uint64_t Sum, const Weight &W) {
219                       return Sum + W.Amount;
220                     }) &&
221            "Expected total to be correct");
222     return;
223   }
224 
225   // Recompute the total through accumulation (rather than shifting it) so that
226   // it's accurate after shifting and any changes combineWeights() made above.
227   Total = 0;
228 
229   // Sum the weights to each node and shift right if necessary.
230   for (Weight &W : Weights) {
231     // Scale down below UINT32_MAX.  Since Shift is larger than necessary, we
232     // can round here without concern about overflow.
233     assert(W.TargetNode.isValid());
234     W.Amount = std::max(UINT64_C(1), shiftRightAndRound(W.Amount, Shift));
235     assert(W.Amount <= UINT32_MAX);
236 
237     // Update the total.
238     Total += W.Amount;
239   }
240   assert(Total <= UINT32_MAX);
241 }
242 
clear()243 void BlockFrequencyInfoImplBase::clear() {
244   // Swap with a default-constructed std::vector, since std::vector<>::clear()
245   // does not actually clear heap storage.
246   std::vector<FrequencyData>().swap(Freqs);
247   std::vector<WorkingData>().swap(Working);
248   Loops.clear();
249 }
250 
251 /// \brief Clear all memory not needed downstream.
252 ///
253 /// Releases all memory not used downstream.  In particular, saves Freqs.
cleanup(BlockFrequencyInfoImplBase & BFI)254 static void cleanup(BlockFrequencyInfoImplBase &BFI) {
255   std::vector<FrequencyData> SavedFreqs(std::move(BFI.Freqs));
256   BFI.clear();
257   BFI.Freqs = std::move(SavedFreqs);
258 }
259 
addToDist(Distribution & Dist,const LoopData * OuterLoop,const BlockNode & Pred,const BlockNode & Succ,uint64_t Weight)260 bool BlockFrequencyInfoImplBase::addToDist(Distribution &Dist,
261                                            const LoopData *OuterLoop,
262                                            const BlockNode &Pred,
263                                            const BlockNode &Succ,
264                                            uint64_t Weight) {
265   if (!Weight)
266     Weight = 1;
267 
268   auto isLoopHeader = [&OuterLoop](const BlockNode &Node) {
269     return OuterLoop && OuterLoop->isHeader(Node);
270   };
271 
272   BlockNode Resolved = Working[Succ.Index].getResolvedNode();
273 
274 #ifndef NDEBUG
275   auto debugSuccessor = [&](const char *Type) {
276     dbgs() << "  =>"
277            << " [" << Type << "] weight = " << Weight;
278     if (!isLoopHeader(Resolved))
279       dbgs() << ", succ = " << getBlockName(Succ);
280     if (Resolved != Succ)
281       dbgs() << ", resolved = " << getBlockName(Resolved);
282     dbgs() << "\n";
283   };
284   (void)debugSuccessor;
285 #endif
286 
287   if (isLoopHeader(Resolved)) {
288     DEBUG(debugSuccessor("backedge"));
289     Dist.addBackedge(Resolved, Weight);
290     return true;
291   }
292 
293   if (Working[Resolved.Index].getContainingLoop() != OuterLoop) {
294     DEBUG(debugSuccessor("  exit  "));
295     Dist.addExit(Resolved, Weight);
296     return true;
297   }
298 
299   if (Resolved < Pred) {
300     if (!isLoopHeader(Pred)) {
301       // If OuterLoop is an irreducible loop, we can't actually handle this.
302       assert((!OuterLoop || !OuterLoop->isIrreducible()) &&
303              "unhandled irreducible control flow");
304 
305       // Irreducible backedge.  Abort.
306       DEBUG(debugSuccessor("abort!!!"));
307       return false;
308     }
309 
310     // If "Pred" is a loop header, then this isn't really a backedge; rather,
311     // OuterLoop must be irreducible.  These false backedges can come only from
312     // secondary loop headers.
313     assert(OuterLoop && OuterLoop->isIrreducible() && !isLoopHeader(Resolved) &&
314            "unhandled irreducible control flow");
315   }
316 
317   DEBUG(debugSuccessor(" local  "));
318   Dist.addLocal(Resolved, Weight);
319   return true;
320 }
321 
addLoopSuccessorsToDist(const LoopData * OuterLoop,LoopData & Loop,Distribution & Dist)322 bool BlockFrequencyInfoImplBase::addLoopSuccessorsToDist(
323     const LoopData *OuterLoop, LoopData &Loop, Distribution &Dist) {
324   // Copy the exit map into Dist.
325   for (const auto &I : Loop.Exits)
326     if (!addToDist(Dist, OuterLoop, Loop.getHeader(), I.first,
327                    I.second.getMass()))
328       // Irreducible backedge.
329       return false;
330 
331   return true;
332 }
333 
334 /// \brief Compute the loop scale for a loop.
computeLoopScale(LoopData & Loop)335 void BlockFrequencyInfoImplBase::computeLoopScale(LoopData &Loop) {
336   // Compute loop scale.
337   DEBUG(dbgs() << "compute-loop-scale: " << getLoopName(Loop) << "\n");
338 
339   // Infinite loops need special handling. If we give the back edge an infinite
340   // mass, they may saturate all the other scales in the function down to 1,
341   // making all the other region temperatures look exactly the same. Choose an
342   // arbitrary scale to avoid these issues.
343   //
344   // FIXME: An alternate way would be to select a symbolic scale which is later
345   // replaced to be the maximum of all computed scales plus 1. This would
346   // appropriately describe the loop as having a large scale, without skewing
347   // the final frequency computation.
348   const Scaled64 InifiniteLoopScale(1, 12);
349 
350   // LoopScale == 1 / ExitMass
351   // ExitMass == HeadMass - BackedgeMass
352   BlockMass TotalBackedgeMass;
353   for (auto &Mass : Loop.BackedgeMass)
354     TotalBackedgeMass += Mass;
355   BlockMass ExitMass = BlockMass::getFull() - TotalBackedgeMass;
356 
357   // Block scale stores the inverse of the scale. If this is an infinite loop,
358   // its exit mass will be zero. In this case, use an arbitrary scale for the
359   // loop scale.
360   Loop.Scale =
361       ExitMass.isEmpty() ? InifiniteLoopScale : ExitMass.toScaled().inverse();
362 
363   DEBUG(dbgs() << " - exit-mass = " << ExitMass << " (" << BlockMass::getFull()
364                << " - " << TotalBackedgeMass << ")\n"
365                << " - scale = " << Loop.Scale << "\n");
366 }
367 
368 /// \brief Package up a loop.
packageLoop(LoopData & Loop)369 void BlockFrequencyInfoImplBase::packageLoop(LoopData &Loop) {
370   DEBUG(dbgs() << "packaging-loop: " << getLoopName(Loop) << "\n");
371 
372   // Clear the subloop exits to prevent quadratic memory usage.
373   for (const BlockNode &M : Loop.Nodes) {
374     if (auto *Loop = Working[M.Index].getPackagedLoop())
375       Loop->Exits.clear();
376     DEBUG(dbgs() << " - node: " << getBlockName(M.Index) << "\n");
377   }
378   Loop.IsPackaged = true;
379 }
380 
381 #ifndef NDEBUG
debugAssign(const BlockFrequencyInfoImplBase & BFI,const DitheringDistributer & D,const BlockNode & T,const BlockMass & M,const char * Desc)382 static void debugAssign(const BlockFrequencyInfoImplBase &BFI,
383                         const DitheringDistributer &D, const BlockNode &T,
384                         const BlockMass &M, const char *Desc) {
385   dbgs() << "  => assign " << M << " (" << D.RemMass << ")";
386   if (Desc)
387     dbgs() << " [" << Desc << "]";
388   if (T.isValid())
389     dbgs() << " to " << BFI.getBlockName(T);
390   dbgs() << "\n";
391 }
392 #endif
393 
distributeMass(const BlockNode & Source,LoopData * OuterLoop,Distribution & Dist)394 void BlockFrequencyInfoImplBase::distributeMass(const BlockNode &Source,
395                                                 LoopData *OuterLoop,
396                                                 Distribution &Dist) {
397   BlockMass Mass = Working[Source.Index].getMass();
398   DEBUG(dbgs() << "  => mass:  " << Mass << "\n");
399 
400   // Distribute mass to successors as laid out in Dist.
401   DitheringDistributer D(Dist, Mass);
402 
403   for (const Weight &W : Dist.Weights) {
404     // Check for a local edge (non-backedge and non-exit).
405     BlockMass Taken = D.takeMass(W.Amount);
406     if (W.Type == Weight::Local) {
407       Working[W.TargetNode.Index].getMass() += Taken;
408       DEBUG(debugAssign(*this, D, W.TargetNode, Taken, nullptr));
409       continue;
410     }
411 
412     // Backedges and exits only make sense if we're processing a loop.
413     assert(OuterLoop && "backedge or exit outside of loop");
414 
415     // Check for a backedge.
416     if (W.Type == Weight::Backedge) {
417       OuterLoop->BackedgeMass[OuterLoop->getHeaderIndex(W.TargetNode)] += Taken;
418       DEBUG(debugAssign(*this, D, W.TargetNode, Taken, "back"));
419       continue;
420     }
421 
422     // This must be an exit.
423     assert(W.Type == Weight::Exit);
424     OuterLoop->Exits.push_back(std::make_pair(W.TargetNode, Taken));
425     DEBUG(debugAssign(*this, D, W.TargetNode, Taken, "exit"));
426   }
427 }
428 
convertFloatingToInteger(BlockFrequencyInfoImplBase & BFI,const Scaled64 & Min,const Scaled64 & Max)429 static void convertFloatingToInteger(BlockFrequencyInfoImplBase &BFI,
430                                      const Scaled64 &Min, const Scaled64 &Max) {
431   // Scale the Factor to a size that creates integers.  Ideally, integers would
432   // be scaled so that Max == UINT64_MAX so that they can be best
433   // differentiated.  However, in the presence of large frequency values, small
434   // frequencies are scaled down to 1, making it impossible to differentiate
435   // small, unequal numbers. When the spread between Min and Max frequencies
436   // fits well within MaxBits, we make the scale be at least 8.
437   const unsigned MaxBits = 64;
438   const unsigned SpreadBits = (Max / Min).lg();
439   Scaled64 ScalingFactor;
440   if (SpreadBits <= MaxBits - 3) {
441     // If the values are small enough, make the scaling factor at least 8 to
442     // allow distinguishing small values.
443     ScalingFactor = Min.inverse();
444     ScalingFactor <<= 3;
445   } else {
446     // If the values need more than MaxBits to be represented, saturate small
447     // frequency values down to 1 by using a scaling factor that benefits large
448     // frequency values.
449     ScalingFactor = Scaled64(1, MaxBits) / Max;
450   }
451 
452   // Translate the floats to integers.
453   DEBUG(dbgs() << "float-to-int: min = " << Min << ", max = " << Max
454                << ", factor = " << ScalingFactor << "\n");
455   for (size_t Index = 0; Index < BFI.Freqs.size(); ++Index) {
456     Scaled64 Scaled = BFI.Freqs[Index].Scaled * ScalingFactor;
457     BFI.Freqs[Index].Integer = std::max(UINT64_C(1), Scaled.toInt<uint64_t>());
458     DEBUG(dbgs() << " - " << BFI.getBlockName(Index) << ": float = "
459                  << BFI.Freqs[Index].Scaled << ", scaled = " << Scaled
460                  << ", int = " << BFI.Freqs[Index].Integer << "\n");
461   }
462 }
463 
464 /// \brief Unwrap a loop package.
465 ///
466 /// Visits all the members of a loop, adjusting their BlockData according to
467 /// the loop's pseudo-node.
unwrapLoop(BlockFrequencyInfoImplBase & BFI,LoopData & Loop)468 static void unwrapLoop(BlockFrequencyInfoImplBase &BFI, LoopData &Loop) {
469   DEBUG(dbgs() << "unwrap-loop-package: " << BFI.getLoopName(Loop)
470                << ": mass = " << Loop.Mass << ", scale = " << Loop.Scale
471                << "\n");
472   Loop.Scale *= Loop.Mass.toScaled();
473   Loop.IsPackaged = false;
474   DEBUG(dbgs() << "  => combined-scale = " << Loop.Scale << "\n");
475 
476   // Propagate the head scale through the loop.  Since members are visited in
477   // RPO, the head scale will be updated by the loop scale first, and then the
478   // final head scale will be used for updated the rest of the members.
479   for (const BlockNode &N : Loop.Nodes) {
480     const auto &Working = BFI.Working[N.Index];
481     Scaled64 &F = Working.isAPackage() ? Working.getPackagedLoop()->Scale
482                                        : BFI.Freqs[N.Index].Scaled;
483     Scaled64 New = Loop.Scale * F;
484     DEBUG(dbgs() << " - " << BFI.getBlockName(N) << ": " << F << " => " << New
485                  << "\n");
486     F = New;
487   }
488 }
489 
unwrapLoops()490 void BlockFrequencyInfoImplBase::unwrapLoops() {
491   // Set initial frequencies from loop-local masses.
492   for (size_t Index = 0; Index < Working.size(); ++Index)
493     Freqs[Index].Scaled = Working[Index].Mass.toScaled();
494 
495   for (LoopData &Loop : Loops)
496     unwrapLoop(*this, Loop);
497 }
498 
finalizeMetrics()499 void BlockFrequencyInfoImplBase::finalizeMetrics() {
500   // Unwrap loop packages in reverse post-order, tracking min and max
501   // frequencies.
502   auto Min = Scaled64::getLargest();
503   auto Max = Scaled64::getZero();
504   for (size_t Index = 0; Index < Working.size(); ++Index) {
505     // Update min/max scale.
506     Min = std::min(Min, Freqs[Index].Scaled);
507     Max = std::max(Max, Freqs[Index].Scaled);
508   }
509 
510   // Convert to integers.
511   convertFloatingToInteger(*this, Min, Max);
512 
513   // Clean up data structures.
514   cleanup(*this);
515 
516   // Print out the final stats.
517   DEBUG(dump());
518 }
519 
520 BlockFrequency
getBlockFreq(const BlockNode & Node) const521 BlockFrequencyInfoImplBase::getBlockFreq(const BlockNode &Node) const {
522   if (!Node.isValid())
523     return 0;
524   return Freqs[Node.Index].Integer;
525 }
526 Scaled64
getFloatingBlockFreq(const BlockNode & Node) const527 BlockFrequencyInfoImplBase::getFloatingBlockFreq(const BlockNode &Node) const {
528   if (!Node.isValid())
529     return Scaled64::getZero();
530   return Freqs[Node.Index].Scaled;
531 }
532 
setBlockFreq(const BlockNode & Node,uint64_t Freq)533 void BlockFrequencyInfoImplBase::setBlockFreq(const BlockNode &Node,
534                                               uint64_t Freq) {
535   assert(Node.isValid() && "Expected valid node");
536   assert(Node.Index < Freqs.size() && "Expected legal index");
537   Freqs[Node.Index].Integer = Freq;
538 }
539 
540 std::string
getBlockName(const BlockNode & Node) const541 BlockFrequencyInfoImplBase::getBlockName(const BlockNode &Node) const {
542   return std::string();
543 }
544 std::string
getLoopName(const LoopData & Loop) const545 BlockFrequencyInfoImplBase::getLoopName(const LoopData &Loop) const {
546   return getBlockName(Loop.getHeader()) + (Loop.isIrreducible() ? "**" : "*");
547 }
548 
549 raw_ostream &
printBlockFreq(raw_ostream & OS,const BlockNode & Node) const550 BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
551                                            const BlockNode &Node) const {
552   return OS << getFloatingBlockFreq(Node);
553 }
554 
555 raw_ostream &
printBlockFreq(raw_ostream & OS,const BlockFrequency & Freq) const556 BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
557                                            const BlockFrequency &Freq) const {
558   Scaled64 Block(Freq.getFrequency(), 0);
559   Scaled64 Entry(getEntryFreq(), 0);
560 
561   return OS << Block / Entry;
562 }
563 
addNodesInLoop(const BFIBase::LoopData & OuterLoop)564 void IrreducibleGraph::addNodesInLoop(const BFIBase::LoopData &OuterLoop) {
565   Start = OuterLoop.getHeader();
566   Nodes.reserve(OuterLoop.Nodes.size());
567   for (auto N : OuterLoop.Nodes)
568     addNode(N);
569   indexNodes();
570 }
addNodesInFunction()571 void IrreducibleGraph::addNodesInFunction() {
572   Start = 0;
573   for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index)
574     if (!BFI.Working[Index].isPackaged())
575       addNode(Index);
576   indexNodes();
577 }
indexNodes()578 void IrreducibleGraph::indexNodes() {
579   for (auto &I : Nodes)
580     Lookup[I.Node.Index] = &I;
581 }
addEdge(IrrNode & Irr,const BlockNode & Succ,const BFIBase::LoopData * OuterLoop)582 void IrreducibleGraph::addEdge(IrrNode &Irr, const BlockNode &Succ,
583                                const BFIBase::LoopData *OuterLoop) {
584   if (OuterLoop && OuterLoop->isHeader(Succ))
585     return;
586   auto L = Lookup.find(Succ.Index);
587   if (L == Lookup.end())
588     return;
589   IrrNode &SuccIrr = *L->second;
590   Irr.Edges.push_back(&SuccIrr);
591   SuccIrr.Edges.push_front(&Irr);
592   ++SuccIrr.NumIn;
593 }
594 
595 namespace llvm {
596 template <> struct GraphTraits<IrreducibleGraph> {
597   typedef bfi_detail::IrreducibleGraph GraphT;
598 
599   typedef const GraphT::IrrNode NodeType;
600   typedef GraphT::IrrNode::iterator ChildIteratorType;
601 
getEntryNodellvm::GraphTraits602   static const NodeType *getEntryNode(const GraphT &G) {
603     return G.StartIrr;
604   }
child_beginllvm::GraphTraits605   static ChildIteratorType child_begin(NodeType *N) { return N->succ_begin(); }
child_endllvm::GraphTraits606   static ChildIteratorType child_end(NodeType *N) { return N->succ_end(); }
607 };
608 }
609 
610 /// \brief Find extra irreducible headers.
611 ///
612 /// Find entry blocks and other blocks with backedges, which exist when \c G
613 /// contains irreducible sub-SCCs.
findIrreducibleHeaders(const BlockFrequencyInfoImplBase & BFI,const IrreducibleGraph & G,const std::vector<const IrreducibleGraph::IrrNode * > & SCC,LoopData::NodeList & Headers,LoopData::NodeList & Others)614 static void findIrreducibleHeaders(
615     const BlockFrequencyInfoImplBase &BFI,
616     const IrreducibleGraph &G,
617     const std::vector<const IrreducibleGraph::IrrNode *> &SCC,
618     LoopData::NodeList &Headers, LoopData::NodeList &Others) {
619   // Map from nodes in the SCC to whether it's an entry block.
620   SmallDenseMap<const IrreducibleGraph::IrrNode *, bool, 8> InSCC;
621 
622   // InSCC also acts the set of nodes in the graph.  Seed it.
623   for (const auto *I : SCC)
624     InSCC[I] = false;
625 
626   for (auto I = InSCC.begin(), E = InSCC.end(); I != E; ++I) {
627     auto &Irr = *I->first;
628     for (const auto *P : make_range(Irr.pred_begin(), Irr.pred_end())) {
629       if (InSCC.count(P))
630         continue;
631 
632       // This is an entry block.
633       I->second = true;
634       Headers.push_back(Irr.Node);
635       DEBUG(dbgs() << "  => entry = " << BFI.getBlockName(Irr.Node) << "\n");
636       break;
637     }
638   }
639   assert(Headers.size() >= 2 &&
640          "Expected irreducible CFG; -loop-info is likely invalid");
641   if (Headers.size() == InSCC.size()) {
642     // Every block is a header.
643     std::sort(Headers.begin(), Headers.end());
644     return;
645   }
646 
647   // Look for extra headers from irreducible sub-SCCs.
648   for (const auto &I : InSCC) {
649     // Entry blocks are already headers.
650     if (I.second)
651       continue;
652 
653     auto &Irr = *I.first;
654     for (const auto *P : make_range(Irr.pred_begin(), Irr.pred_end())) {
655       // Skip forward edges.
656       if (P->Node < Irr.Node)
657         continue;
658 
659       // Skip predecessors from entry blocks.  These can have inverted
660       // ordering.
661       if (InSCC.lookup(P))
662         continue;
663 
664       // Store the extra header.
665       Headers.push_back(Irr.Node);
666       DEBUG(dbgs() << "  => extra = " << BFI.getBlockName(Irr.Node) << "\n");
667       break;
668     }
669     if (Headers.back() == Irr.Node)
670       // Added this as a header.
671       continue;
672 
673     // This is not a header.
674     Others.push_back(Irr.Node);
675     DEBUG(dbgs() << "  => other = " << BFI.getBlockName(Irr.Node) << "\n");
676   }
677   std::sort(Headers.begin(), Headers.end());
678   std::sort(Others.begin(), Others.end());
679 }
680 
createIrreducibleLoop(BlockFrequencyInfoImplBase & BFI,const IrreducibleGraph & G,LoopData * OuterLoop,std::list<LoopData>::iterator Insert,const std::vector<const IrreducibleGraph::IrrNode * > & SCC)681 static void createIrreducibleLoop(
682     BlockFrequencyInfoImplBase &BFI, const IrreducibleGraph &G,
683     LoopData *OuterLoop, std::list<LoopData>::iterator Insert,
684     const std::vector<const IrreducibleGraph::IrrNode *> &SCC) {
685   // Translate the SCC into RPO.
686   DEBUG(dbgs() << " - found-scc\n");
687 
688   LoopData::NodeList Headers;
689   LoopData::NodeList Others;
690   findIrreducibleHeaders(BFI, G, SCC, Headers, Others);
691 
692   auto Loop = BFI.Loops.emplace(Insert, OuterLoop, Headers.begin(),
693                                 Headers.end(), Others.begin(), Others.end());
694 
695   // Update loop hierarchy.
696   for (const auto &N : Loop->Nodes)
697     if (BFI.Working[N.Index].isLoopHeader())
698       BFI.Working[N.Index].Loop->Parent = &*Loop;
699     else
700       BFI.Working[N.Index].Loop = &*Loop;
701 }
702 
703 iterator_range<std::list<LoopData>::iterator>
analyzeIrreducible(const IrreducibleGraph & G,LoopData * OuterLoop,std::list<LoopData>::iterator Insert)704 BlockFrequencyInfoImplBase::analyzeIrreducible(
705     const IrreducibleGraph &G, LoopData *OuterLoop,
706     std::list<LoopData>::iterator Insert) {
707   assert((OuterLoop == nullptr) == (Insert == Loops.begin()));
708   auto Prev = OuterLoop ? std::prev(Insert) : Loops.end();
709 
710   for (auto I = scc_begin(G); !I.isAtEnd(); ++I) {
711     if (I->size() < 2)
712       continue;
713 
714     // Translate the SCC into RPO.
715     createIrreducibleLoop(*this, G, OuterLoop, Insert, *I);
716   }
717 
718   if (OuterLoop)
719     return make_range(std::next(Prev), Insert);
720   return make_range(Loops.begin(), Insert);
721 }
722 
723 void
updateLoopWithIrreducible(LoopData & OuterLoop)724 BlockFrequencyInfoImplBase::updateLoopWithIrreducible(LoopData &OuterLoop) {
725   OuterLoop.Exits.clear();
726   for (auto &Mass : OuterLoop.BackedgeMass)
727     Mass = BlockMass::getEmpty();
728   auto O = OuterLoop.Nodes.begin() + 1;
729   for (auto I = O, E = OuterLoop.Nodes.end(); I != E; ++I)
730     if (!Working[I->Index].isPackaged())
731       *O++ = *I;
732   OuterLoop.Nodes.erase(O, OuterLoop.Nodes.end());
733 }
734 
adjustLoopHeaderMass(LoopData & Loop)735 void BlockFrequencyInfoImplBase::adjustLoopHeaderMass(LoopData &Loop) {
736   assert(Loop.isIrreducible() && "this only makes sense on irreducible loops");
737 
738   // Since the loop has more than one header block, the mass flowing back into
739   // each header will be different. Adjust the mass in each header loop to
740   // reflect the masses flowing through back edges.
741   //
742   // To do this, we distribute the initial mass using the backedge masses
743   // as weights for the distribution.
744   BlockMass LoopMass = BlockMass::getFull();
745   Distribution Dist;
746 
747   DEBUG(dbgs() << "adjust-loop-header-mass:\n");
748   for (uint32_t H = 0; H < Loop.NumHeaders; ++H) {
749     auto &HeaderNode = Loop.Nodes[H];
750     auto &BackedgeMass = Loop.BackedgeMass[Loop.getHeaderIndex(HeaderNode)];
751     DEBUG(dbgs() << " - Add back edge mass for node "
752                  << getBlockName(HeaderNode) << ": " << BackedgeMass << "\n");
753     if (BackedgeMass.getMass() > 0)
754       Dist.addLocal(HeaderNode, BackedgeMass.getMass());
755     else
756       DEBUG(dbgs() << "   Nothing added. Back edge mass is zero\n");
757   }
758 
759   DitheringDistributer D(Dist, LoopMass);
760 
761   DEBUG(dbgs() << " Distribute loop mass " << LoopMass
762                << " to headers using above weights\n");
763   for (const Weight &W : Dist.Weights) {
764     BlockMass Taken = D.takeMass(W.Amount);
765     assert(W.Type == Weight::Local && "all weights should be local");
766     Working[W.TargetNode.Index].getMass() = Taken;
767     DEBUG(debugAssign(*this, D, W.TargetNode, Taken, nullptr));
768   }
769 }
770