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(OuterLoop->getHeader(), 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 Get the maximum allowed loop scale.
335 ///
336 /// Gives the maximum number of estimated iterations allowed for a loop. Very
337 /// large numbers cause problems downstream (even within 64-bits).
getMaxLoopScale()338 static Scaled64 getMaxLoopScale() { return Scaled64(1, 12); }
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 // LoopScale == 1 / ExitMass
346 // ExitMass == HeadMass - BackedgeMass
347 BlockMass ExitMass = BlockMass::getFull() - Loop.BackedgeMass;
348
349 // Block scale stores the inverse of the scale.
350 Loop.Scale = ExitMass.toScaled().inverse();
351
352 DEBUG(dbgs() << " - exit-mass = " << ExitMass << " (" << BlockMass::getFull()
353 << " - " << Loop.BackedgeMass << ")\n"
354 << " - scale = " << Loop.Scale << "\n");
355
356 if (Loop.Scale > getMaxLoopScale()) {
357 Loop.Scale = getMaxLoopScale();
358 DEBUG(dbgs() << " - reduced-to-max-scale: " << getMaxLoopScale() << "\n");
359 }
360 }
361
362 /// \brief Package up a loop.
packageLoop(LoopData & Loop)363 void BlockFrequencyInfoImplBase::packageLoop(LoopData &Loop) {
364 DEBUG(dbgs() << "packaging-loop: " << getLoopName(Loop) << "\n");
365
366 // Clear the subloop exits to prevent quadratic memory usage.
367 for (const BlockNode &M : Loop.Nodes) {
368 if (auto *Loop = Working[M.Index].getPackagedLoop())
369 Loop->Exits.clear();
370 DEBUG(dbgs() << " - node: " << getBlockName(M.Index) << "\n");
371 }
372 Loop.IsPackaged = true;
373 }
374
distributeMass(const BlockNode & Source,LoopData * OuterLoop,Distribution & Dist)375 void BlockFrequencyInfoImplBase::distributeMass(const BlockNode &Source,
376 LoopData *OuterLoop,
377 Distribution &Dist) {
378 BlockMass Mass = Working[Source.Index].getMass();
379 DEBUG(dbgs() << " => mass: " << Mass << "\n");
380
381 // Distribute mass to successors as laid out in Dist.
382 DitheringDistributer D(Dist, Mass);
383
384 #ifndef NDEBUG
385 auto debugAssign = [&](const BlockNode &T, const BlockMass &M,
386 const char *Desc) {
387 dbgs() << " => assign " << M << " (" << D.RemMass << ")";
388 if (Desc)
389 dbgs() << " [" << Desc << "]";
390 if (T.isValid())
391 dbgs() << " to " << getBlockName(T);
392 dbgs() << "\n";
393 };
394 (void)debugAssign;
395 #endif
396
397 for (const Weight &W : Dist.Weights) {
398 // Check for a local edge (non-backedge and non-exit).
399 BlockMass Taken = D.takeMass(W.Amount);
400 if (W.Type == Weight::Local) {
401 Working[W.TargetNode.Index].getMass() += Taken;
402 DEBUG(debugAssign(W.TargetNode, Taken, nullptr));
403 continue;
404 }
405
406 // Backedges and exits only make sense if we're processing a loop.
407 assert(OuterLoop && "backedge or exit outside of loop");
408
409 // Check for a backedge.
410 if (W.Type == Weight::Backedge) {
411 OuterLoop->BackedgeMass += Taken;
412 DEBUG(debugAssign(BlockNode(), Taken, "back"));
413 continue;
414 }
415
416 // This must be an exit.
417 assert(W.Type == Weight::Exit);
418 OuterLoop->Exits.push_back(std::make_pair(W.TargetNode, Taken));
419 DEBUG(debugAssign(W.TargetNode, Taken, "exit"));
420 }
421 }
422
convertFloatingToInteger(BlockFrequencyInfoImplBase & BFI,const Scaled64 & Min,const Scaled64 & Max)423 static void convertFloatingToInteger(BlockFrequencyInfoImplBase &BFI,
424 const Scaled64 &Min, const Scaled64 &Max) {
425 // Scale the Factor to a size that creates integers. Ideally, integers would
426 // be scaled so that Max == UINT64_MAX so that they can be best
427 // differentiated. However, the register allocator currently deals poorly
428 // with large numbers. Instead, push Min up a little from 1 to give some
429 // room to differentiate small, unequal numbers.
430 //
431 // TODO: fix issues downstream so that ScalingFactor can be
432 // Scaled64(1,64)/Max.
433 Scaled64 ScalingFactor = Min.inverse();
434 if ((Max / Min).lg() < 60)
435 ScalingFactor <<= 3;
436
437 // Translate the floats to integers.
438 DEBUG(dbgs() << "float-to-int: min = " << Min << ", max = " << Max
439 << ", factor = " << ScalingFactor << "\n");
440 for (size_t Index = 0; Index < BFI.Freqs.size(); ++Index) {
441 Scaled64 Scaled = BFI.Freqs[Index].Scaled * ScalingFactor;
442 BFI.Freqs[Index].Integer = std::max(UINT64_C(1), Scaled.toInt<uint64_t>());
443 DEBUG(dbgs() << " - " << BFI.getBlockName(Index) << ": float = "
444 << BFI.Freqs[Index].Scaled << ", scaled = " << Scaled
445 << ", int = " << BFI.Freqs[Index].Integer << "\n");
446 }
447 }
448
449 /// \brief Unwrap a loop package.
450 ///
451 /// Visits all the members of a loop, adjusting their BlockData according to
452 /// the loop's pseudo-node.
unwrapLoop(BlockFrequencyInfoImplBase & BFI,LoopData & Loop)453 static void unwrapLoop(BlockFrequencyInfoImplBase &BFI, LoopData &Loop) {
454 DEBUG(dbgs() << "unwrap-loop-package: " << BFI.getLoopName(Loop)
455 << ": mass = " << Loop.Mass << ", scale = " << Loop.Scale
456 << "\n");
457 Loop.Scale *= Loop.Mass.toScaled();
458 Loop.IsPackaged = false;
459 DEBUG(dbgs() << " => combined-scale = " << Loop.Scale << "\n");
460
461 // Propagate the head scale through the loop. Since members are visited in
462 // RPO, the head scale will be updated by the loop scale first, and then the
463 // final head scale will be used for updated the rest of the members.
464 for (const BlockNode &N : Loop.Nodes) {
465 const auto &Working = BFI.Working[N.Index];
466 Scaled64 &F = Working.isAPackage() ? Working.getPackagedLoop()->Scale
467 : BFI.Freqs[N.Index].Scaled;
468 Scaled64 New = Loop.Scale * F;
469 DEBUG(dbgs() << " - " << BFI.getBlockName(N) << ": " << F << " => " << New
470 << "\n");
471 F = New;
472 }
473 }
474
unwrapLoops()475 void BlockFrequencyInfoImplBase::unwrapLoops() {
476 // Set initial frequencies from loop-local masses.
477 for (size_t Index = 0; Index < Working.size(); ++Index)
478 Freqs[Index].Scaled = Working[Index].Mass.toScaled();
479
480 for (LoopData &Loop : Loops)
481 unwrapLoop(*this, Loop);
482 }
483
finalizeMetrics()484 void BlockFrequencyInfoImplBase::finalizeMetrics() {
485 // Unwrap loop packages in reverse post-order, tracking min and max
486 // frequencies.
487 auto Min = Scaled64::getLargest();
488 auto Max = Scaled64::getZero();
489 for (size_t Index = 0; Index < Working.size(); ++Index) {
490 // Update min/max scale.
491 Min = std::min(Min, Freqs[Index].Scaled);
492 Max = std::max(Max, Freqs[Index].Scaled);
493 }
494
495 // Convert to integers.
496 convertFloatingToInteger(*this, Min, Max);
497
498 // Clean up data structures.
499 cleanup(*this);
500
501 // Print out the final stats.
502 DEBUG(dump());
503 }
504
505 BlockFrequency
getBlockFreq(const BlockNode & Node) const506 BlockFrequencyInfoImplBase::getBlockFreq(const BlockNode &Node) const {
507 if (!Node.isValid())
508 return 0;
509 return Freqs[Node.Index].Integer;
510 }
511 Scaled64
getFloatingBlockFreq(const BlockNode & Node) const512 BlockFrequencyInfoImplBase::getFloatingBlockFreq(const BlockNode &Node) const {
513 if (!Node.isValid())
514 return Scaled64::getZero();
515 return Freqs[Node.Index].Scaled;
516 }
517
518 std::string
getBlockName(const BlockNode & Node) const519 BlockFrequencyInfoImplBase::getBlockName(const BlockNode &Node) const {
520 return std::string();
521 }
522 std::string
getLoopName(const LoopData & Loop) const523 BlockFrequencyInfoImplBase::getLoopName(const LoopData &Loop) const {
524 return getBlockName(Loop.getHeader()) + (Loop.isIrreducible() ? "**" : "*");
525 }
526
527 raw_ostream &
printBlockFreq(raw_ostream & OS,const BlockNode & Node) const528 BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
529 const BlockNode &Node) const {
530 return OS << getFloatingBlockFreq(Node);
531 }
532
533 raw_ostream &
printBlockFreq(raw_ostream & OS,const BlockFrequency & Freq) const534 BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
535 const BlockFrequency &Freq) const {
536 Scaled64 Block(Freq.getFrequency(), 0);
537 Scaled64 Entry(getEntryFreq(), 0);
538
539 return OS << Block / Entry;
540 }
541
addNodesInLoop(const BFIBase::LoopData & OuterLoop)542 void IrreducibleGraph::addNodesInLoop(const BFIBase::LoopData &OuterLoop) {
543 Start = OuterLoop.getHeader();
544 Nodes.reserve(OuterLoop.Nodes.size());
545 for (auto N : OuterLoop.Nodes)
546 addNode(N);
547 indexNodes();
548 }
addNodesInFunction()549 void IrreducibleGraph::addNodesInFunction() {
550 Start = 0;
551 for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index)
552 if (!BFI.Working[Index].isPackaged())
553 addNode(Index);
554 indexNodes();
555 }
indexNodes()556 void IrreducibleGraph::indexNodes() {
557 for (auto &I : Nodes)
558 Lookup[I.Node.Index] = &I;
559 }
addEdge(IrrNode & Irr,const BlockNode & Succ,const BFIBase::LoopData * OuterLoop)560 void IrreducibleGraph::addEdge(IrrNode &Irr, const BlockNode &Succ,
561 const BFIBase::LoopData *OuterLoop) {
562 if (OuterLoop && OuterLoop->isHeader(Succ))
563 return;
564 auto L = Lookup.find(Succ.Index);
565 if (L == Lookup.end())
566 return;
567 IrrNode &SuccIrr = *L->second;
568 Irr.Edges.push_back(&SuccIrr);
569 SuccIrr.Edges.push_front(&Irr);
570 ++SuccIrr.NumIn;
571 }
572
573 namespace llvm {
574 template <> struct GraphTraits<IrreducibleGraph> {
575 typedef bfi_detail::IrreducibleGraph GraphT;
576
577 typedef const GraphT::IrrNode NodeType;
578 typedef GraphT::IrrNode::iterator ChildIteratorType;
579
getEntryNodellvm::GraphTraits580 static const NodeType *getEntryNode(const GraphT &G) {
581 return G.StartIrr;
582 }
child_beginllvm::GraphTraits583 static ChildIteratorType child_begin(NodeType *N) { return N->succ_begin(); }
child_endllvm::GraphTraits584 static ChildIteratorType child_end(NodeType *N) { return N->succ_end(); }
585 };
586 }
587
588 /// \brief Find extra irreducible headers.
589 ///
590 /// Find entry blocks and other blocks with backedges, which exist when \c G
591 /// contains irreducible sub-SCCs.
findIrreducibleHeaders(const BlockFrequencyInfoImplBase & BFI,const IrreducibleGraph & G,const std::vector<const IrreducibleGraph::IrrNode * > & SCC,LoopData::NodeList & Headers,LoopData::NodeList & Others)592 static void findIrreducibleHeaders(
593 const BlockFrequencyInfoImplBase &BFI,
594 const IrreducibleGraph &G,
595 const std::vector<const IrreducibleGraph::IrrNode *> &SCC,
596 LoopData::NodeList &Headers, LoopData::NodeList &Others) {
597 // Map from nodes in the SCC to whether it's an entry block.
598 SmallDenseMap<const IrreducibleGraph::IrrNode *, bool, 8> InSCC;
599
600 // InSCC also acts the set of nodes in the graph. Seed it.
601 for (const auto *I : SCC)
602 InSCC[I] = false;
603
604 for (auto I = InSCC.begin(), E = InSCC.end(); I != E; ++I) {
605 auto &Irr = *I->first;
606 for (const auto *P : make_range(Irr.pred_begin(), Irr.pred_end())) {
607 if (InSCC.count(P))
608 continue;
609
610 // This is an entry block.
611 I->second = true;
612 Headers.push_back(Irr.Node);
613 DEBUG(dbgs() << " => entry = " << BFI.getBlockName(Irr.Node) << "\n");
614 break;
615 }
616 }
617 assert(Headers.size() >= 2 &&
618 "Expected irreducible CFG; -loop-info is likely invalid");
619 if (Headers.size() == InSCC.size()) {
620 // Every block is a header.
621 std::sort(Headers.begin(), Headers.end());
622 return;
623 }
624
625 // Look for extra headers from irreducible sub-SCCs.
626 for (const auto &I : InSCC) {
627 // Entry blocks are already headers.
628 if (I.second)
629 continue;
630
631 auto &Irr = *I.first;
632 for (const auto *P : make_range(Irr.pred_begin(), Irr.pred_end())) {
633 // Skip forward edges.
634 if (P->Node < Irr.Node)
635 continue;
636
637 // Skip predecessors from entry blocks. These can have inverted
638 // ordering.
639 if (InSCC.lookup(P))
640 continue;
641
642 // Store the extra header.
643 Headers.push_back(Irr.Node);
644 DEBUG(dbgs() << " => extra = " << BFI.getBlockName(Irr.Node) << "\n");
645 break;
646 }
647 if (Headers.back() == Irr.Node)
648 // Added this as a header.
649 continue;
650
651 // This is not a header.
652 Others.push_back(Irr.Node);
653 DEBUG(dbgs() << " => other = " << BFI.getBlockName(Irr.Node) << "\n");
654 }
655 std::sort(Headers.begin(), Headers.end());
656 std::sort(Others.begin(), Others.end());
657 }
658
createIrreducibleLoop(BlockFrequencyInfoImplBase & BFI,const IrreducibleGraph & G,LoopData * OuterLoop,std::list<LoopData>::iterator Insert,const std::vector<const IrreducibleGraph::IrrNode * > & SCC)659 static void createIrreducibleLoop(
660 BlockFrequencyInfoImplBase &BFI, const IrreducibleGraph &G,
661 LoopData *OuterLoop, std::list<LoopData>::iterator Insert,
662 const std::vector<const IrreducibleGraph::IrrNode *> &SCC) {
663 // Translate the SCC into RPO.
664 DEBUG(dbgs() << " - found-scc\n");
665
666 LoopData::NodeList Headers;
667 LoopData::NodeList Others;
668 findIrreducibleHeaders(BFI, G, SCC, Headers, Others);
669
670 auto Loop = BFI.Loops.emplace(Insert, OuterLoop, Headers.begin(),
671 Headers.end(), Others.begin(), Others.end());
672
673 // Update loop hierarchy.
674 for (const auto &N : Loop->Nodes)
675 if (BFI.Working[N.Index].isLoopHeader())
676 BFI.Working[N.Index].Loop->Parent = &*Loop;
677 else
678 BFI.Working[N.Index].Loop = &*Loop;
679 }
680
681 iterator_range<std::list<LoopData>::iterator>
analyzeIrreducible(const IrreducibleGraph & G,LoopData * OuterLoop,std::list<LoopData>::iterator Insert)682 BlockFrequencyInfoImplBase::analyzeIrreducible(
683 const IrreducibleGraph &G, LoopData *OuterLoop,
684 std::list<LoopData>::iterator Insert) {
685 assert((OuterLoop == nullptr) == (Insert == Loops.begin()));
686 auto Prev = OuterLoop ? std::prev(Insert) : Loops.end();
687
688 for (auto I = scc_begin(G); !I.isAtEnd(); ++I) {
689 if (I->size() < 2)
690 continue;
691
692 // Translate the SCC into RPO.
693 createIrreducibleLoop(*this, G, OuterLoop, Insert, *I);
694 }
695
696 if (OuterLoop)
697 return make_range(std::next(Prev), Insert);
698 return make_range(Loops.begin(), Insert);
699 }
700
701 void
updateLoopWithIrreducible(LoopData & OuterLoop)702 BlockFrequencyInfoImplBase::updateLoopWithIrreducible(LoopData &OuterLoop) {
703 OuterLoop.Exits.clear();
704 OuterLoop.BackedgeMass = BlockMass::getEmpty();
705 auto O = OuterLoop.Nodes.begin() + 1;
706 for (auto I = O, E = OuterLoop.Nodes.end(); I != E; ++I)
707 if (!Working[I->Index].isPackaged())
708 *O++ = *I;
709 OuterLoop.Nodes.erase(O, OuterLoop.Nodes.end());
710 }
711