xref: /freebsd/contrib/llvm-project/llvm/lib/Analysis/BlockFrequencyInfoImpl.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
1 //===- BlockFrequencyImplInfo.cpp - Block Frequency Info Implementation ---===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // Loops should be simplified before this analysis.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/Analysis/BlockFrequencyInfoImpl.h"
14 #include "llvm/ADT/APInt.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/SCCIterator.h"
17 #include "llvm/ADT/SmallString.h"
18 #include "llvm/Config/llvm-config.h"
19 #include "llvm/IR/Function.h"
20 #include "llvm/Support/BlockFrequency.h"
21 #include "llvm/Support/BranchProbability.h"
22 #include "llvm/Support/Compiler.h"
23 #include "llvm/Support/Debug.h"
24 #include "llvm/Support/MathExtras.h"
25 #include "llvm/Support/ScaledNumber.h"
26 #include "llvm/Support/raw_ostream.h"
27 #include <algorithm>
28 #include <cassert>
29 #include <cstddef>
30 #include <cstdint>
31 #include <iterator>
32 #include <list>
33 #include <numeric>
34 #include <optional>
35 #include <utility>
36 #include <vector>
37 
38 using namespace llvm;
39 using namespace llvm::bfi_detail;
40 
41 #define DEBUG_TYPE "block-freq"
42 
43 namespace llvm {
44 cl::opt<bool> CheckBFIUnknownBlockQueries(
45     "check-bfi-unknown-block-queries",
46     cl::init(false), cl::Hidden,
47     cl::desc("Check if block frequency is queried for an unknown block "
48              "for debugging missed BFI updates"));
49 
50 cl::opt<bool> UseIterativeBFIInference(
51     "use-iterative-bfi-inference", cl::Hidden,
52     cl::desc("Apply an iterative post-processing to infer correct BFI counts"));
53 
54 cl::opt<unsigned> IterativeBFIMaxIterationsPerBlock(
55     "iterative-bfi-max-iterations-per-block", cl::init(1000), cl::Hidden,
56     cl::desc("Iterative inference: maximum number of update iterations "
57              "per block"));
58 
59 cl::opt<double> IterativeBFIPrecision(
60     "iterative-bfi-precision", cl::init(1e-12), cl::Hidden,
61     cl::desc("Iterative inference: delta convergence precision; smaller values "
62              "typically lead to better results at the cost of worsen runtime"));
63 } // namespace llvm
64 
toScaled() const65 ScaledNumber<uint64_t> BlockMass::toScaled() const {
66   if (isFull())
67     return ScaledNumber<uint64_t>(1, 0);
68   return ScaledNumber<uint64_t>(getMass() + 1, -64);
69 }
70 
71 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
dump() const72 LLVM_DUMP_METHOD void BlockMass::dump() const { print(dbgs()); }
73 #endif
74 
getHexDigit(int N)75 static char getHexDigit(int N) {
76   assert(N < 16);
77   if (N < 10)
78     return '0' + N;
79   return 'a' + N - 10;
80 }
81 
print(raw_ostream & OS) const82 raw_ostream &BlockMass::print(raw_ostream &OS) const {
83   for (int Digits = 0; Digits < 16; ++Digits)
84     OS << getHexDigit(Mass >> (60 - Digits * 4) & 0xf);
85   return OS;
86 }
87 
88 namespace {
89 
90 using BlockNode = BlockFrequencyInfoImplBase::BlockNode;
91 using Distribution = BlockFrequencyInfoImplBase::Distribution;
92 using WeightList = BlockFrequencyInfoImplBase::Distribution::WeightList;
93 using Scaled64 = BlockFrequencyInfoImplBase::Scaled64;
94 using LoopData = BlockFrequencyInfoImplBase::LoopData;
95 using Weight = BlockFrequencyInfoImplBase::Weight;
96 using FrequencyData = BlockFrequencyInfoImplBase::FrequencyData;
97 
98 /// Dithering mass distributer.
99 ///
100 /// This class splits up a single mass into portions by weight, dithering to
101 /// spread out error.  No mass is lost.  The dithering precision depends on the
102 /// precision of the product of \a BlockMass and \a BranchProbability.
103 ///
104 /// The distribution algorithm follows.
105 ///
106 ///  1. Initialize by saving the sum of the weights in \a RemWeight and the
107 ///     mass to distribute in \a RemMass.
108 ///
109 ///  2. For each portion:
110 ///
111 ///      1. Construct a branch probability, P, as the portion's weight divided
112 ///         by the current value of \a RemWeight.
113 ///      2. Calculate the portion's mass as \a RemMass times P.
114 ///      3. Update \a RemWeight and \a RemMass at each portion by subtracting
115 ///         the current portion's weight and mass.
116 struct DitheringDistributer {
117   uint32_t RemWeight;
118   BlockMass RemMass;
119 
120   DitheringDistributer(Distribution &Dist, const BlockMass &Mass);
121 
122   BlockMass takeMass(uint32_t Weight);
123 };
124 
125 } // end anonymous namespace
126 
DitheringDistributer(Distribution & Dist,const BlockMass & Mass)127 DitheringDistributer::DitheringDistributer(Distribution &Dist,
128                                            const BlockMass &Mass) {
129   Dist.normalize();
130   RemWeight = Dist.Total;
131   RemMass = Mass;
132 }
133 
takeMass(uint32_t Weight)134 BlockMass DitheringDistributer::takeMass(uint32_t Weight) {
135   assert(Weight && "invalid weight");
136   assert(Weight <= RemWeight);
137   BlockMass Mass = RemMass * BranchProbability(Weight, RemWeight);
138 
139   // Decrement totals (dither).
140   RemWeight -= Weight;
141   RemMass -= Mass;
142   return Mass;
143 }
144 
add(const BlockNode & Node,uint64_t Amount,Weight::DistType Type)145 void Distribution::add(const BlockNode &Node, uint64_t Amount,
146                        Weight::DistType Type) {
147   assert(Amount && "invalid weight of 0");
148   uint64_t NewTotal = Total + Amount;
149 
150   // Check for overflow.  It should be impossible to overflow twice.
151   bool IsOverflow = NewTotal < Total;
152   assert(!(DidOverflow && IsOverflow) && "unexpected repeated overflow");
153   DidOverflow |= IsOverflow;
154 
155   // Update the total.
156   Total = NewTotal;
157 
158   // Save the weight.
159   Weights.push_back(Weight(Type, Node, Amount));
160 }
161 
combineWeight(Weight & W,const Weight & OtherW)162 static void combineWeight(Weight &W, const Weight &OtherW) {
163   assert(OtherW.TargetNode.isValid());
164   if (!W.Amount) {
165     W = OtherW;
166     return;
167   }
168   assert(W.Type == OtherW.Type);
169   assert(W.TargetNode == OtherW.TargetNode);
170   assert(OtherW.Amount && "Expected non-zero weight");
171   if (W.Amount > W.Amount + OtherW.Amount)
172     // Saturate on overflow.
173     W.Amount = UINT64_MAX;
174   else
175     W.Amount += OtherW.Amount;
176 }
177 
combineWeightsBySorting(WeightList & Weights)178 static void combineWeightsBySorting(WeightList &Weights) {
179   // Sort so edges to the same node are adjacent.
180   llvm::sort(Weights, [](const Weight &L, const Weight &R) {
181     return L.TargetNode < R.TargetNode;
182   });
183 
184   // Combine adjacent edges.
185   WeightList::iterator O = Weights.begin();
186   for (WeightList::const_iterator I = O, L = O, E = Weights.end(); I != E;
187        ++O, (I = L)) {
188     *O = *I;
189 
190     // Find the adjacent weights to the same node.
191     for (++L; L != E && I->TargetNode == L->TargetNode; ++L)
192       combineWeight(*O, *L);
193   }
194 
195   // Erase extra entries.
196   Weights.erase(O, Weights.end());
197 }
198 
combineWeightsByHashing(WeightList & Weights)199 static void combineWeightsByHashing(WeightList &Weights) {
200   // Collect weights into a DenseMap.
201   using HashTable = DenseMap<BlockNode::IndexType, Weight>;
202 
203   HashTable Combined(NextPowerOf2(2 * Weights.size()));
204   for (const Weight &W : Weights)
205     combineWeight(Combined[W.TargetNode.Index], W);
206 
207   // Check whether anything changed.
208   if (Weights.size() == Combined.size())
209     return;
210 
211   // Fill in the new weights.
212   Weights.clear();
213   Weights.reserve(Combined.size());
214   for (const auto &I : Combined)
215     Weights.push_back(I.second);
216 }
217 
combineWeights(WeightList & Weights)218 static void combineWeights(WeightList &Weights) {
219   // Use a hash table for many successors to keep this linear.
220   if (Weights.size() > 128) {
221     combineWeightsByHashing(Weights);
222     return;
223   }
224 
225   combineWeightsBySorting(Weights);
226 }
227 
shiftRightAndRound(uint64_t N,int Shift)228 static uint64_t shiftRightAndRound(uint64_t N, int Shift) {
229   assert(Shift >= 0);
230   assert(Shift < 64);
231   if (!Shift)
232     return N;
233   return (N >> Shift) + (UINT64_C(1) & N >> (Shift - 1));
234 }
235 
normalize()236 void Distribution::normalize() {
237   // Early exit for termination nodes.
238   if (Weights.empty())
239     return;
240 
241   // Only bother if there are multiple successors.
242   if (Weights.size() > 1)
243     combineWeights(Weights);
244 
245   // Early exit when combined into a single successor.
246   if (Weights.size() == 1) {
247     Total = 1;
248     Weights.front().Amount = 1;
249     return;
250   }
251 
252   // Determine how much to shift right so that the total fits into 32-bits.
253   //
254   // If we shift at all, shift by 1 extra.  Otherwise, the lower limit of 1
255   // for each weight can cause a 32-bit overflow.
256   int Shift = 0;
257   if (DidOverflow)
258     Shift = 33;
259   else if (Total > UINT32_MAX)
260     Shift = 33 - llvm::countl_zero(Total);
261 
262   // Early exit if nothing needs to be scaled.
263   if (!Shift) {
264     // If we didn't overflow then combineWeights() shouldn't have changed the
265     // sum of the weights, but let's double-check.
266     assert(Total == std::accumulate(Weights.begin(), Weights.end(), UINT64_C(0),
267                                     [](uint64_t Sum, const Weight &W) {
268                       return Sum + W.Amount;
269                     }) &&
270            "Expected total to be correct");
271     return;
272   }
273 
274   // Recompute the total through accumulation (rather than shifting it) so that
275   // it's accurate after shifting and any changes combineWeights() made above.
276   Total = 0;
277 
278   // Sum the weights to each node and shift right if necessary.
279   for (Weight &W : Weights) {
280     // Scale down below UINT32_MAX.  Since Shift is larger than necessary, we
281     // can round here without concern about overflow.
282     assert(W.TargetNode.isValid());
283     W.Amount = std::max(UINT64_C(1), shiftRightAndRound(W.Amount, Shift));
284     assert(W.Amount <= UINT32_MAX);
285 
286     // Update the total.
287     Total += W.Amount;
288   }
289   assert(Total <= UINT32_MAX);
290 }
291 
clear()292 void BlockFrequencyInfoImplBase::clear() {
293   // Swap with a default-constructed std::vector, since std::vector<>::clear()
294   // does not actually clear heap storage.
295   std::vector<FrequencyData>().swap(Freqs);
296   IsIrrLoopHeader.clear();
297   std::vector<WorkingData>().swap(Working);
298   Loops.clear();
299 }
300 
301 /// Clear all memory not needed downstream.
302 ///
303 /// Releases all memory not used downstream.  In particular, saves Freqs.
cleanup(BlockFrequencyInfoImplBase & BFI)304 static void cleanup(BlockFrequencyInfoImplBase &BFI) {
305   std::vector<FrequencyData> SavedFreqs(std::move(BFI.Freqs));
306   SparseBitVector<> SavedIsIrrLoopHeader(std::move(BFI.IsIrrLoopHeader));
307   BFI.clear();
308   BFI.Freqs = std::move(SavedFreqs);
309   BFI.IsIrrLoopHeader = std::move(SavedIsIrrLoopHeader);
310 }
311 
addToDist(Distribution & Dist,const LoopData * OuterLoop,const BlockNode & Pred,const BlockNode & Succ,uint64_t Weight)312 bool BlockFrequencyInfoImplBase::addToDist(Distribution &Dist,
313                                            const LoopData *OuterLoop,
314                                            const BlockNode &Pred,
315                                            const BlockNode &Succ,
316                                            uint64_t Weight) {
317   if (!Weight)
318     Weight = 1;
319 
320   auto isLoopHeader = [&OuterLoop](const BlockNode &Node) {
321     return OuterLoop && OuterLoop->isHeader(Node);
322   };
323 
324   BlockNode Resolved = Working[Succ.Index].getResolvedNode();
325 
326 #ifndef NDEBUG
327   auto debugSuccessor = [&](const char *Type) {
328     dbgs() << "  =>"
329            << " [" << Type << "] weight = " << Weight;
330     if (!isLoopHeader(Resolved))
331       dbgs() << ", succ = " << getBlockName(Succ);
332     if (Resolved != Succ)
333       dbgs() << ", resolved = " << getBlockName(Resolved);
334     dbgs() << "\n";
335   };
336   (void)debugSuccessor;
337 #endif
338 
339   if (isLoopHeader(Resolved)) {
340     LLVM_DEBUG(debugSuccessor("backedge"));
341     Dist.addBackedge(Resolved, Weight);
342     return true;
343   }
344 
345   if (Working[Resolved.Index].getContainingLoop() != OuterLoop) {
346     LLVM_DEBUG(debugSuccessor("  exit  "));
347     Dist.addExit(Resolved, Weight);
348     return true;
349   }
350 
351   if (Resolved < Pred) {
352     if (!isLoopHeader(Pred)) {
353       // If OuterLoop is an irreducible loop, we can't actually handle this.
354       assert((!OuterLoop || !OuterLoop->isIrreducible()) &&
355              "unhandled irreducible control flow");
356 
357       // Irreducible backedge.  Abort.
358       LLVM_DEBUG(debugSuccessor("abort!!!"));
359       return false;
360     }
361 
362     // If "Pred" is a loop header, then this isn't really a backedge; rather,
363     // OuterLoop must be irreducible.  These false backedges can come only from
364     // secondary loop headers.
365     assert(OuterLoop && OuterLoop->isIrreducible() && !isLoopHeader(Resolved) &&
366            "unhandled irreducible control flow");
367   }
368 
369   LLVM_DEBUG(debugSuccessor(" local  "));
370   Dist.addLocal(Resolved, Weight);
371   return true;
372 }
373 
addLoopSuccessorsToDist(const LoopData * OuterLoop,LoopData & Loop,Distribution & Dist)374 bool BlockFrequencyInfoImplBase::addLoopSuccessorsToDist(
375     const LoopData *OuterLoop, LoopData &Loop, Distribution &Dist) {
376   // Copy the exit map into Dist.
377   for (const auto &I : Loop.Exits)
378     if (!addToDist(Dist, OuterLoop, Loop.getHeader(), I.first,
379                    I.second.getMass()))
380       // Irreducible backedge.
381       return false;
382 
383   return true;
384 }
385 
386 /// Compute the loop scale for a loop.
computeLoopScale(LoopData & Loop)387 void BlockFrequencyInfoImplBase::computeLoopScale(LoopData &Loop) {
388   // Compute loop scale.
389   LLVM_DEBUG(dbgs() << "compute-loop-scale: " << getLoopName(Loop) << "\n");
390 
391   // Infinite loops need special handling. If we give the back edge an infinite
392   // mass, they may saturate all the other scales in the function down to 1,
393   // making all the other region temperatures look exactly the same. Choose an
394   // arbitrary scale to avoid these issues.
395   //
396   // FIXME: An alternate way would be to select a symbolic scale which is later
397   // replaced to be the maximum of all computed scales plus 1. This would
398   // appropriately describe the loop as having a large scale, without skewing
399   // the final frequency computation.
400   const Scaled64 InfiniteLoopScale(1, 12);
401 
402   // LoopScale == 1 / ExitMass
403   // ExitMass == HeadMass - BackedgeMass
404   BlockMass TotalBackedgeMass;
405   for (auto &Mass : Loop.BackedgeMass)
406     TotalBackedgeMass += Mass;
407   BlockMass ExitMass = BlockMass::getFull() - TotalBackedgeMass;
408 
409   // Block scale stores the inverse of the scale. If this is an infinite loop,
410   // its exit mass will be zero. In this case, use an arbitrary scale for the
411   // loop scale.
412   Loop.Scale =
413       ExitMass.isEmpty() ? InfiniteLoopScale : ExitMass.toScaled().inverse();
414 
415   LLVM_DEBUG(dbgs() << " - exit-mass = " << ExitMass << " ("
416                     << BlockMass::getFull() << " - " << TotalBackedgeMass
417                     << ")\n"
418                     << " - scale = " << Loop.Scale << "\n");
419 }
420 
421 /// Package up a loop.
packageLoop(LoopData & Loop)422 void BlockFrequencyInfoImplBase::packageLoop(LoopData &Loop) {
423   LLVM_DEBUG(dbgs() << "packaging-loop: " << getLoopName(Loop) << "\n");
424 
425   // Clear the subloop exits to prevent quadratic memory usage.
426   for (const BlockNode &M : Loop.Nodes) {
427     if (auto *Loop = Working[M.Index].getPackagedLoop())
428       Loop->Exits.clear();
429     LLVM_DEBUG(dbgs() << " - node: " << getBlockName(M.Index) << "\n");
430   }
431   Loop.IsPackaged = true;
432 }
433 
434 #ifndef NDEBUG
debugAssign(const BlockFrequencyInfoImplBase & BFI,const DitheringDistributer & D,const BlockNode & T,const BlockMass & M,const char * Desc)435 static void debugAssign(const BlockFrequencyInfoImplBase &BFI,
436                         const DitheringDistributer &D, const BlockNode &T,
437                         const BlockMass &M, const char *Desc) {
438   dbgs() << "  => assign " << M << " (" << D.RemMass << ")";
439   if (Desc)
440     dbgs() << " [" << Desc << "]";
441   if (T.isValid())
442     dbgs() << " to " << BFI.getBlockName(T);
443   dbgs() << "\n";
444 }
445 #endif
446 
distributeMass(const BlockNode & Source,LoopData * OuterLoop,Distribution & Dist)447 void BlockFrequencyInfoImplBase::distributeMass(const BlockNode &Source,
448                                                 LoopData *OuterLoop,
449                                                 Distribution &Dist) {
450   BlockMass Mass = Working[Source.Index].getMass();
451   LLVM_DEBUG(dbgs() << "  => mass:  " << Mass << "\n");
452 
453   // Distribute mass to successors as laid out in Dist.
454   DitheringDistributer D(Dist, Mass);
455 
456   for (const Weight &W : Dist.Weights) {
457     // Check for a local edge (non-backedge and non-exit).
458     BlockMass Taken = D.takeMass(W.Amount);
459     if (W.Type == Weight::Local) {
460       Working[W.TargetNode.Index].getMass() += Taken;
461       LLVM_DEBUG(debugAssign(*this, D, W.TargetNode, Taken, nullptr));
462       continue;
463     }
464 
465     // Backedges and exits only make sense if we're processing a loop.
466     assert(OuterLoop && "backedge or exit outside of loop");
467 
468     // Check for a backedge.
469     if (W.Type == Weight::Backedge) {
470       OuterLoop->BackedgeMass[OuterLoop->getHeaderIndex(W.TargetNode)] += Taken;
471       LLVM_DEBUG(debugAssign(*this, D, W.TargetNode, Taken, "back"));
472       continue;
473     }
474 
475     // This must be an exit.
476     assert(W.Type == Weight::Exit);
477     OuterLoop->Exits.push_back(std::make_pair(W.TargetNode, Taken));
478     LLVM_DEBUG(debugAssign(*this, D, W.TargetNode, Taken, "exit"));
479   }
480 }
481 
convertFloatingToInteger(BlockFrequencyInfoImplBase & BFI,const Scaled64 & Min,const Scaled64 & Max)482 static void convertFloatingToInteger(BlockFrequencyInfoImplBase &BFI,
483                                      const Scaled64 &Min, const Scaled64 &Max) {
484   // Scale the Factor to a size that creates integers.  If possible scale
485   // integers so that Max == UINT64_MAX so that they can be best differentiated.
486   // Is is possible that the range between min and max cannot be accurately
487   // represented in a 64bit integer without either loosing precision for small
488   // values (so small unequal numbers all map to 1) or saturaturing big numbers
489   // loosing precision for big numbers (so unequal big numbers may map to
490   // UINT64_MAX). We choose to loose precision for small numbers.
491   const unsigned MaxBits = sizeof(Scaled64::DigitsType) * CHAR_BIT;
492   // Users often add up multiple BlockFrequency values or multiply them with
493   // things like instruction costs. Leave some room to avoid saturating
494   // operations reaching UIN64_MAX too early.
495   const unsigned Slack = 10;
496   Scaled64 ScalingFactor = Scaled64(1, MaxBits - Slack) / Max;
497 
498   // Translate the floats to integers.
499   LLVM_DEBUG(dbgs() << "float-to-int: min = " << Min << ", max = " << Max
500                     << ", factor = " << ScalingFactor << "\n");
501   (void)Min;
502   for (size_t Index = 0; Index < BFI.Freqs.size(); ++Index) {
503     Scaled64 Scaled = BFI.Freqs[Index].Scaled * ScalingFactor;
504     BFI.Freqs[Index].Integer = std::max(UINT64_C(1), Scaled.toInt<uint64_t>());
505     LLVM_DEBUG(dbgs() << " - " << BFI.getBlockName(Index) << ": float = "
506                       << BFI.Freqs[Index].Scaled << ", scaled = " << Scaled
507                       << ", int = " << BFI.Freqs[Index].Integer << "\n");
508   }
509 }
510 
511 /// Unwrap a loop package.
512 ///
513 /// Visits all the members of a loop, adjusting their BlockData according to
514 /// the loop's pseudo-node.
unwrapLoop(BlockFrequencyInfoImplBase & BFI,LoopData & Loop)515 static void unwrapLoop(BlockFrequencyInfoImplBase &BFI, LoopData &Loop) {
516   LLVM_DEBUG(dbgs() << "unwrap-loop-package: " << BFI.getLoopName(Loop)
517                     << ": mass = " << Loop.Mass << ", scale = " << Loop.Scale
518                     << "\n");
519   Loop.Scale *= Loop.Mass.toScaled();
520   Loop.IsPackaged = false;
521   LLVM_DEBUG(dbgs() << "  => combined-scale = " << Loop.Scale << "\n");
522 
523   // Propagate the head scale through the loop.  Since members are visited in
524   // RPO, the head scale will be updated by the loop scale first, and then the
525   // final head scale will be used for updated the rest of the members.
526   for (const BlockNode &N : Loop.Nodes) {
527     const auto &Working = BFI.Working[N.Index];
528     Scaled64 &F = Working.isAPackage() ? Working.getPackagedLoop()->Scale
529                                        : BFI.Freqs[N.Index].Scaled;
530     Scaled64 New = Loop.Scale * F;
531     LLVM_DEBUG(dbgs() << " - " << BFI.getBlockName(N) << ": " << F << " => "
532                       << New << "\n");
533     F = New;
534   }
535 }
536 
unwrapLoops()537 void BlockFrequencyInfoImplBase::unwrapLoops() {
538   // Set initial frequencies from loop-local masses.
539   for (size_t Index = 0; Index < Working.size(); ++Index)
540     Freqs[Index].Scaled = Working[Index].Mass.toScaled();
541 
542   for (LoopData &Loop : Loops)
543     unwrapLoop(*this, Loop);
544 }
545 
finalizeMetrics()546 void BlockFrequencyInfoImplBase::finalizeMetrics() {
547   // Unwrap loop packages in reverse post-order, tracking min and max
548   // frequencies.
549   auto Min = Scaled64::getLargest();
550   auto Max = Scaled64::getZero();
551   for (size_t Index = 0; Index < Working.size(); ++Index) {
552     // Update min/max scale.
553     Min = std::min(Min, Freqs[Index].Scaled);
554     Max = std::max(Max, Freqs[Index].Scaled);
555   }
556 
557   // Convert to integers.
558   convertFloatingToInteger(*this, Min, Max);
559 
560   // Clean up data structures.
561   cleanup(*this);
562 
563   // Print out the final stats.
564   LLVM_DEBUG(dump());
565 }
566 
567 BlockFrequency
getBlockFreq(const BlockNode & Node) const568 BlockFrequencyInfoImplBase::getBlockFreq(const BlockNode &Node) const {
569   if (!Node.isValid()) {
570 #ifndef NDEBUG
571     if (CheckBFIUnknownBlockQueries) {
572       SmallString<256> Msg;
573       raw_svector_ostream OS(Msg);
574       OS << "*** Detected BFI query for unknown block " << getBlockName(Node);
575       report_fatal_error(OS.str());
576     }
577 #endif
578     return BlockFrequency(0);
579   }
580   return BlockFrequency(Freqs[Node.Index].Integer);
581 }
582 
583 std::optional<uint64_t>
getBlockProfileCount(const Function & F,const BlockNode & Node,bool AllowSynthetic) const584 BlockFrequencyInfoImplBase::getBlockProfileCount(const Function &F,
585                                                  const BlockNode &Node,
586                                                  bool AllowSynthetic) const {
587   return getProfileCountFromFreq(F, getBlockFreq(Node), AllowSynthetic);
588 }
589 
getProfileCountFromFreq(const Function & F,BlockFrequency Freq,bool AllowSynthetic) const590 std::optional<uint64_t> BlockFrequencyInfoImplBase::getProfileCountFromFreq(
591     const Function &F, BlockFrequency Freq, bool AllowSynthetic) const {
592   auto EntryCount = F.getEntryCount(AllowSynthetic);
593   if (!EntryCount)
594     return std::nullopt;
595   // Use 128 bit APInt to do the arithmetic to avoid overflow.
596   APInt BlockCount(128, EntryCount->getCount());
597   APInt BlockFreq(128, Freq.getFrequency());
598   APInt EntryFreq(128, getEntryFreq().getFrequency());
599   BlockCount *= BlockFreq;
600   // Rounded division of BlockCount by EntryFreq. Since EntryFreq is unsigned
601   // lshr by 1 gives EntryFreq/2.
602   BlockCount = (BlockCount + EntryFreq.lshr(1)).udiv(EntryFreq);
603   return BlockCount.getLimitedValue();
604 }
605 
606 bool
isIrrLoopHeader(const BlockNode & Node)607 BlockFrequencyInfoImplBase::isIrrLoopHeader(const BlockNode &Node) {
608   if (!Node.isValid())
609     return false;
610   return IsIrrLoopHeader.test(Node.Index);
611 }
612 
613 Scaled64
getFloatingBlockFreq(const BlockNode & Node) const614 BlockFrequencyInfoImplBase::getFloatingBlockFreq(const BlockNode &Node) const {
615   if (!Node.isValid())
616     return Scaled64::getZero();
617   return Freqs[Node.Index].Scaled;
618 }
619 
setBlockFreq(const BlockNode & Node,BlockFrequency Freq)620 void BlockFrequencyInfoImplBase::setBlockFreq(const BlockNode &Node,
621                                               BlockFrequency Freq) {
622   assert(Node.isValid() && "Expected valid node");
623   assert(Node.Index < Freqs.size() && "Expected legal index");
624   Freqs[Node.Index].Integer = Freq.getFrequency();
625 }
626 
627 std::string
getBlockName(const BlockNode & Node) const628 BlockFrequencyInfoImplBase::getBlockName(const BlockNode &Node) const {
629   return {};
630 }
631 
632 std::string
getLoopName(const LoopData & Loop) const633 BlockFrequencyInfoImplBase::getLoopName(const LoopData &Loop) const {
634   return getBlockName(Loop.getHeader()) + (Loop.isIrreducible() ? "**" : "*");
635 }
636 
addNodesInLoop(const BFIBase::LoopData & OuterLoop)637 void IrreducibleGraph::addNodesInLoop(const BFIBase::LoopData &OuterLoop) {
638   Start = OuterLoop.getHeader();
639   Nodes.reserve(OuterLoop.Nodes.size());
640   for (auto N : OuterLoop.Nodes)
641     addNode(N);
642   indexNodes();
643 }
644 
addNodesInFunction()645 void IrreducibleGraph::addNodesInFunction() {
646   Start = 0;
647   for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index)
648     if (!BFI.Working[Index].isPackaged())
649       addNode(Index);
650   indexNodes();
651 }
652 
indexNodes()653 void IrreducibleGraph::indexNodes() {
654   for (auto &I : Nodes)
655     Lookup[I.Node.Index] = &I;
656 }
657 
addEdge(IrrNode & Irr,const BlockNode & Succ,const BFIBase::LoopData * OuterLoop)658 void IrreducibleGraph::addEdge(IrrNode &Irr, const BlockNode &Succ,
659                                const BFIBase::LoopData *OuterLoop) {
660   if (OuterLoop && OuterLoop->isHeader(Succ))
661     return;
662   auto L = Lookup.find(Succ.Index);
663   if (L == Lookup.end())
664     return;
665   IrrNode &SuccIrr = *L->second;
666   Irr.Edges.push_back(&SuccIrr);
667   SuccIrr.Edges.push_front(&Irr);
668   ++SuccIrr.NumIn;
669 }
670 
671 namespace llvm {
672 
673 template <> struct GraphTraits<IrreducibleGraph> {
674   using GraphT = bfi_detail::IrreducibleGraph;
675   using NodeRef = const GraphT::IrrNode *;
676   using ChildIteratorType = GraphT::IrrNode::iterator;
677 
getEntryNodellvm::GraphTraits678   static NodeRef getEntryNode(const GraphT &G) { return G.StartIrr; }
child_beginllvm::GraphTraits679   static ChildIteratorType child_begin(NodeRef N) { return N->succ_begin(); }
child_endllvm::GraphTraits680   static ChildIteratorType child_end(NodeRef N) { return N->succ_end(); }
681 };
682 
683 } // end namespace llvm
684 
685 /// Find extra irreducible headers.
686 ///
687 /// Find entry blocks and other blocks with backedges, which exist when \c G
688 /// contains irreducible sub-SCCs.
findIrreducibleHeaders(const BlockFrequencyInfoImplBase & BFI,const IrreducibleGraph & G,const std::vector<const IrreducibleGraph::IrrNode * > & SCC,LoopData::NodeList & Headers,LoopData::NodeList & Others)689 static void findIrreducibleHeaders(
690     const BlockFrequencyInfoImplBase &BFI,
691     const IrreducibleGraph &G,
692     const std::vector<const IrreducibleGraph::IrrNode *> &SCC,
693     LoopData::NodeList &Headers, LoopData::NodeList &Others) {
694   // Map from nodes in the SCC to whether it's an entry block.
695   SmallDenseMap<const IrreducibleGraph::IrrNode *, bool, 8> InSCC;
696 
697   // InSCC also acts the set of nodes in the graph.  Seed it.
698   for (const auto *I : SCC)
699     InSCC[I] = false;
700 
701   for (auto I = InSCC.begin(), E = InSCC.end(); I != E; ++I) {
702     auto &Irr = *I->first;
703     for (const auto *P : make_range(Irr.pred_begin(), Irr.pred_end())) {
704       if (InSCC.count(P))
705         continue;
706 
707       // This is an entry block.
708       I->second = true;
709       Headers.push_back(Irr.Node);
710       LLVM_DEBUG(dbgs() << "  => entry = " << BFI.getBlockName(Irr.Node)
711                         << "\n");
712       break;
713     }
714   }
715   assert(Headers.size() >= 2 &&
716          "Expected irreducible CFG; -loop-info is likely invalid");
717   if (Headers.size() == InSCC.size()) {
718     // Every block is a header.
719     llvm::sort(Headers);
720     return;
721   }
722 
723   // Look for extra headers from irreducible sub-SCCs.
724   for (const auto &I : InSCC) {
725     // Entry blocks are already headers.
726     if (I.second)
727       continue;
728 
729     auto &Irr = *I.first;
730     for (const auto *P : make_range(Irr.pred_begin(), Irr.pred_end())) {
731       // Skip forward edges.
732       if (P->Node < Irr.Node)
733         continue;
734 
735       // Skip predecessors from entry blocks.  These can have inverted
736       // ordering.
737       if (InSCC.lookup(P))
738         continue;
739 
740       // Store the extra header.
741       Headers.push_back(Irr.Node);
742       LLVM_DEBUG(dbgs() << "  => extra = " << BFI.getBlockName(Irr.Node)
743                         << "\n");
744       break;
745     }
746     if (Headers.back() == Irr.Node)
747       // Added this as a header.
748       continue;
749 
750     // This is not a header.
751     Others.push_back(Irr.Node);
752     LLVM_DEBUG(dbgs() << "  => other = " << BFI.getBlockName(Irr.Node) << "\n");
753   }
754   llvm::sort(Headers);
755   llvm::sort(Others);
756 }
757 
createIrreducibleLoop(BlockFrequencyInfoImplBase & BFI,const IrreducibleGraph & G,LoopData * OuterLoop,std::list<LoopData>::iterator Insert,const std::vector<const IrreducibleGraph::IrrNode * > & SCC)758 static void createIrreducibleLoop(
759     BlockFrequencyInfoImplBase &BFI, const IrreducibleGraph &G,
760     LoopData *OuterLoop, std::list<LoopData>::iterator Insert,
761     const std::vector<const IrreducibleGraph::IrrNode *> &SCC) {
762   // Translate the SCC into RPO.
763   LLVM_DEBUG(dbgs() << " - found-scc\n");
764 
765   LoopData::NodeList Headers;
766   LoopData::NodeList Others;
767   findIrreducibleHeaders(BFI, G, SCC, Headers, Others);
768 
769   auto Loop = BFI.Loops.emplace(Insert, OuterLoop, Headers.begin(),
770                                 Headers.end(), Others.begin(), Others.end());
771 
772   // Update loop hierarchy.
773   for (const auto &N : Loop->Nodes)
774     if (BFI.Working[N.Index].isLoopHeader())
775       BFI.Working[N.Index].Loop->Parent = &*Loop;
776     else
777       BFI.Working[N.Index].Loop = &*Loop;
778 }
779 
780 iterator_range<std::list<LoopData>::iterator>
analyzeIrreducible(const IrreducibleGraph & G,LoopData * OuterLoop,std::list<LoopData>::iterator Insert)781 BlockFrequencyInfoImplBase::analyzeIrreducible(
782     const IrreducibleGraph &G, LoopData *OuterLoop,
783     std::list<LoopData>::iterator Insert) {
784   assert((OuterLoop == nullptr) == (Insert == Loops.begin()));
785   auto Prev = OuterLoop ? std::prev(Insert) : Loops.end();
786 
787   for (auto I = scc_begin(G); !I.isAtEnd(); ++I) {
788     if (I->size() < 2)
789       continue;
790 
791     // Translate the SCC into RPO.
792     createIrreducibleLoop(*this, G, OuterLoop, Insert, *I);
793   }
794 
795   if (OuterLoop)
796     return make_range(std::next(Prev), Insert);
797   return make_range(Loops.begin(), Insert);
798 }
799 
800 void
updateLoopWithIrreducible(LoopData & OuterLoop)801 BlockFrequencyInfoImplBase::updateLoopWithIrreducible(LoopData &OuterLoop) {
802   OuterLoop.Exits.clear();
803   for (auto &Mass : OuterLoop.BackedgeMass)
804     Mass = BlockMass::getEmpty();
805   auto O = OuterLoop.Nodes.begin() + 1;
806   for (auto I = O, E = OuterLoop.Nodes.end(); I != E; ++I)
807     if (!Working[I->Index].isPackaged())
808       *O++ = *I;
809   OuterLoop.Nodes.erase(O, OuterLoop.Nodes.end());
810 }
811 
adjustLoopHeaderMass(LoopData & Loop)812 void BlockFrequencyInfoImplBase::adjustLoopHeaderMass(LoopData &Loop) {
813   assert(Loop.isIrreducible() && "this only makes sense on irreducible loops");
814 
815   // Since the loop has more than one header block, the mass flowing back into
816   // each header will be different. Adjust the mass in each header loop to
817   // reflect the masses flowing through back edges.
818   //
819   // To do this, we distribute the initial mass using the backedge masses
820   // as weights for the distribution.
821   BlockMass LoopMass = BlockMass::getFull();
822   Distribution Dist;
823 
824   LLVM_DEBUG(dbgs() << "adjust-loop-header-mass:\n");
825   for (uint32_t H = 0; H < Loop.NumHeaders; ++H) {
826     auto &HeaderNode = Loop.Nodes[H];
827     auto &BackedgeMass = Loop.BackedgeMass[Loop.getHeaderIndex(HeaderNode)];
828     LLVM_DEBUG(dbgs() << " - Add back edge mass for node "
829                       << getBlockName(HeaderNode) << ": " << BackedgeMass
830                       << "\n");
831     if (BackedgeMass.getMass() > 0)
832       Dist.addLocal(HeaderNode, BackedgeMass.getMass());
833     else
834       LLVM_DEBUG(dbgs() << "   Nothing added. Back edge mass is zero\n");
835   }
836 
837   DitheringDistributer D(Dist, LoopMass);
838 
839   LLVM_DEBUG(dbgs() << " Distribute loop mass " << LoopMass
840                     << " to headers using above weights\n");
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 
distributeIrrLoopHeaderMass(Distribution & Dist)849 void BlockFrequencyInfoImplBase::distributeIrrLoopHeaderMass(Distribution &Dist) {
850   BlockMass LoopMass = BlockMass::getFull();
851   DitheringDistributer D(Dist, LoopMass);
852   for (const Weight &W : Dist.Weights) {
853     BlockMass Taken = D.takeMass(W.Amount);
854     assert(W.Type == Weight::Local && "all weights should be local");
855     Working[W.TargetNode.Index].getMass() = Taken;
856     LLVM_DEBUG(debugAssign(*this, D, W.TargetNode, Taken, nullptr));
857   }
858 }
859