1 //===- BlockFrequencyImplInfo.cpp - Block Frequency Info Implementation ---===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Loops should be simplified before this analysis.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Analysis/BlockFrequencyInfoImpl.h"
15 #include "llvm/ADT/APFloat.h"
16 #include "llvm/Support/raw_ostream.h"
20 using namespace llvm::bfi_detail;
22 #define DEBUG_TYPE "block-freq"
24 //===----------------------------------------------------------------------===//
26 // UnsignedFloat implementation.
28 //===----------------------------------------------------------------------===//
30 const int32_t UnsignedFloatBase::MaxExponent;
31 const int32_t UnsignedFloatBase::MinExponent;
34 static void appendDigit(std::string &Str, unsigned D) {
39 static void appendNumber(std::string &Str, uint64_t N) {
41 appendDigit(Str, N % 10);
46 static bool doesRoundUp(char Digit) {
59 static std::string toStringAPFloat(uint64_t D, int E, unsigned Precision) {
60 assert(E >= UnsignedFloatBase::MinExponent);
61 assert(E <= UnsignedFloatBase::MaxExponent);
63 // Find a new E, but don't let it increase past MaxExponent.
64 int LeadingZeros = UnsignedFloatBase::countLeadingZeros64(D);
65 int NewE = std::min(UnsignedFloatBase::MaxExponent, E + 63 - LeadingZeros);
66 int Shift = 63 - (NewE - E);
67 assert(Shift <= LeadingZeros);
68 assert(Shift == LeadingZeros || NewE == UnsignedFloatBase::MaxExponent);
72 // Check for a denormal.
73 unsigned AdjustedE = E + 16383;
75 assert(E == UnsignedFloatBase::MaxExponent);
79 // Build the float and print it.
80 uint64_t RawBits[2] = {D, AdjustedE};
81 APFloat Float(APFloat::x87DoubleExtended, APInt(80, RawBits));
82 SmallVector<char, 24> Chars;
83 Float.toString(Chars, Precision, 0);
84 return std::string(Chars.begin(), Chars.end());
87 static std::string stripTrailingZeros(const std::string &Float) {
88 size_t NonZero = Float.find_last_not_of('0');
89 assert(NonZero != std::string::npos && "no . in floating point string");
91 if (Float[NonZero] == '.')
94 return Float.substr(0, NonZero + 1);
97 std::string UnsignedFloatBase::toString(uint64_t D, int16_t E, int Width,
102 // Canonicalize exponent and digits.
110 if (int Shift = std::min(int16_t(countLeadingZeros64(D)), E)) {
117 } else if (E > -64) {
119 Below0 = D << (64 + E);
120 } else if (E > -120) {
121 Below0 = D >> (-E - 64);
122 Extra = D << (128 + E);
123 ExtraShift = -64 - E;
126 // Fall back on APFloat for very small and very large numbers.
127 if (!Above0 && !Below0)
128 return toStringAPFloat(D, E, Precision);
130 // Append the digits before the decimal.
132 size_t DigitsOut = 0;
134 appendNumber(Str, Above0);
135 DigitsOut = Str.size();
138 std::reverse(Str.begin(), Str.end());
140 // Return early if there's nothing after the decimal.
144 // Append the decimal and beyond.
146 uint64_t Error = UINT64_C(1) << (64 - Width);
148 // We need to shift Below0 to the right to make space for calculating
149 // digits. Save the precision we're losing in Extra.
150 Extra = (Below0 & 0xf) << 56 | (Extra >> 8);
153 size_t AfterDot = Str.size();
163 Below0 += (Extra >> 60);
164 Extra = Extra & (UINT64_MAX >> 4);
165 appendDigit(Str, Below0 >> 60);
166 Below0 = Below0 & (UINT64_MAX >> 4);
167 if (DigitsOut || Str.back() != '0')
170 } while (Error && (Below0 << 4 | Extra >> 60) >= Error / 2 &&
171 (!Precision || DigitsOut <= Precision || SinceDot < 2));
173 // Return early for maximum precision.
174 if (!Precision || DigitsOut <= Precision)
175 return stripTrailingZeros(Str);
177 // Find where to truncate.
179 std::max(Str.size() - (DigitsOut - Precision), AfterDot + 1);
181 // Check if there's anything to truncate.
182 if (Truncate >= Str.size())
183 return stripTrailingZeros(Str);
185 bool Carry = doesRoundUp(Str[Truncate]);
187 return stripTrailingZeros(Str.substr(0, Truncate));
189 // Round with the first truncated digit.
190 for (std::string::reverse_iterator I(Str.begin() + Truncate), E = Str.rend();
204 // Add "1" in front if we still need to carry.
205 return stripTrailingZeros(std::string(Carry, '1') + Str.substr(0, Truncate));
208 raw_ostream &UnsignedFloatBase::print(raw_ostream &OS, uint64_t D, int16_t E,
209 int Width, unsigned Precision) {
210 return OS << toString(D, E, Width, Precision);
213 void UnsignedFloatBase::dump(uint64_t D, int16_t E, int Width) {
214 print(dbgs(), D, E, Width, 0) << "[" << Width << ":" << D << "*2^" << E
218 static std::pair<uint64_t, int16_t>
219 getRoundedFloat(uint64_t N, bool ShouldRound, int64_t Shift) {
222 // Rounding caused an overflow.
223 return std::make_pair(UINT64_C(1), Shift + 64);
224 return std::make_pair(N, Shift);
227 std::pair<uint64_t, int16_t> UnsignedFloatBase::divide64(uint64_t Dividend,
229 // Input should be sanitized.
233 // Minimize size of divisor.
235 if (int Zeros = countTrailingZeros(Divisor)) {
240 // Check for powers of two.
242 return std::make_pair(Dividend, Shift);
244 // Maximize size of dividend.
245 if (int Zeros = countLeadingZeros64(Dividend)) {
250 // Start with the result of a divide.
251 uint64_t Quotient = Dividend / Divisor;
254 // Continue building the quotient with long division.
256 // TODO: continue with largers digits.
257 while (!(Quotient >> 63) && Dividend) {
258 // Shift Dividend, and check for overflow.
259 bool IsOverflow = Dividend >> 63;
264 bool DoesDivide = IsOverflow || Divisor <= Dividend;
265 Quotient = (Quotient << 1) | uint64_t(DoesDivide);
266 Dividend -= DoesDivide ? Divisor : 0;
270 if (Dividend >= getHalf(Divisor))
272 // Rounding caused an overflow in Quotient.
273 return std::make_pair(UINT64_C(1), Shift + 64);
275 return getRoundedFloat(Quotient, Dividend >= getHalf(Divisor), Shift);
278 std::pair<uint64_t, int16_t> UnsignedFloatBase::multiply64(uint64_t L,
280 // Separate into two 32-bit digits (U.L).
281 uint64_t UL = L >> 32, LL = L & UINT32_MAX, UR = R >> 32, LR = R & UINT32_MAX;
283 // Compute cross products.
284 uint64_t P1 = UL * UR, P2 = UL * LR, P3 = LL * UR, P4 = LL * LR;
286 // Sum into two 64-bit digits.
287 uint64_t Upper = P1, Lower = P4;
288 auto addWithCarry = [&](uint64_t N) {
289 uint64_t NewLower = Lower + (N << 32);
290 Upper += (N >> 32) + (NewLower < Lower);
296 // Check whether the upper digit is empty.
298 return std::make_pair(Lower, 0);
300 // Shift as little as possible to maximize precision.
301 unsigned LeadingZeros = countLeadingZeros64(Upper);
302 int16_t Shift = 64 - LeadingZeros;
304 Upper = Upper << LeadingZeros | Lower >> Shift;
305 bool ShouldRound = Shift && (Lower & UINT64_C(1) << (Shift - 1));
306 return getRoundedFloat(Upper, ShouldRound, Shift);
309 //===----------------------------------------------------------------------===//
311 // BlockMass implementation.
313 //===----------------------------------------------------------------------===//
314 UnsignedFloat<uint64_t> BlockMass::toFloat() const {
316 return UnsignedFloat<uint64_t>(1, 0);
317 return UnsignedFloat<uint64_t>(getMass() + 1, -64);
320 void BlockMass::dump() const { print(dbgs()); }
322 static char getHexDigit(int N) {
328 raw_ostream &BlockMass::print(raw_ostream &OS) const {
329 for (int Digits = 0; Digits < 16; ++Digits)
330 OS << getHexDigit(Mass >> (60 - Digits * 4) & 0xf);
334 //===----------------------------------------------------------------------===//
336 // BlockFrequencyInfoImpl implementation.
338 //===----------------------------------------------------------------------===//
341 typedef BlockFrequencyInfoImplBase::BlockNode BlockNode;
342 typedef BlockFrequencyInfoImplBase::Distribution Distribution;
343 typedef BlockFrequencyInfoImplBase::Distribution::WeightList WeightList;
344 typedef BlockFrequencyInfoImplBase::Float Float;
345 typedef BlockFrequencyInfoImplBase::LoopData LoopData;
346 typedef BlockFrequencyInfoImplBase::Weight Weight;
347 typedef BlockFrequencyInfoImplBase::FrequencyData FrequencyData;
349 /// \brief Dithering mass distributer.
351 /// This class splits up a single mass into portions by weight, dithering to
352 /// spread out error. No mass is lost. The dithering precision depends on the
353 /// precision of the product of \a BlockMass and \a BranchProbability.
355 /// The distribution algorithm follows.
357 /// 1. Initialize by saving the sum of the weights in \a RemWeight and the
358 /// mass to distribute in \a RemMass.
360 /// 2. For each portion:
362 /// 1. Construct a branch probability, P, as the portion's weight divided
363 /// by the current value of \a RemWeight.
364 /// 2. Calculate the portion's mass as \a RemMass times P.
365 /// 3. Update \a RemWeight and \a RemMass at each portion by subtracting
366 /// the current portion's weight and mass.
367 struct DitheringDistributer {
371 DitheringDistributer(Distribution &Dist, const BlockMass &Mass);
373 BlockMass takeMass(uint32_t Weight);
377 DitheringDistributer::DitheringDistributer(Distribution &Dist,
378 const BlockMass &Mass) {
380 RemWeight = Dist.Total;
384 BlockMass DitheringDistributer::takeMass(uint32_t Weight) {
385 assert(Weight && "invalid weight");
386 assert(Weight <= RemWeight);
387 BlockMass Mass = RemMass * BranchProbability(Weight, RemWeight);
389 // Decrement totals (dither).
395 void Distribution::add(const BlockNode &Node, uint64_t Amount,
396 Weight::DistType Type) {
397 assert(Amount && "invalid weight of 0");
398 uint64_t NewTotal = Total + Amount;
400 // Check for overflow. It should be impossible to overflow twice.
401 bool IsOverflow = NewTotal < Total;
402 assert(!(DidOverflow && IsOverflow) && "unexpected repeated overflow");
403 DidOverflow |= IsOverflow;
413 Weights.push_back(W);
416 static void combineWeight(Weight &W, const Weight &OtherW) {
417 assert(OtherW.TargetNode.isValid());
422 assert(W.Type == OtherW.Type);
423 assert(W.TargetNode == OtherW.TargetNode);
424 assert(W.Amount < W.Amount + OtherW.Amount && "Unexpected overflow");
425 W.Amount += OtherW.Amount;
427 static void combineWeightsBySorting(WeightList &Weights) {
428 // Sort so edges to the same node are adjacent.
429 std::sort(Weights.begin(), Weights.end(),
431 const Weight &R) { return L.TargetNode < R.TargetNode; });
433 // Combine adjacent edges.
434 WeightList::iterator O = Weights.begin();
435 for (WeightList::const_iterator I = O, L = O, E = Weights.end(); I != E;
439 // Find the adjacent weights to the same node.
440 for (++L; L != E && I->TargetNode == L->TargetNode; ++L)
441 combineWeight(*O, *L);
444 // Erase extra entries.
445 Weights.erase(O, Weights.end());
448 static void combineWeightsByHashing(WeightList &Weights) {
449 // Collect weights into a DenseMap.
450 typedef DenseMap<BlockNode::IndexType, Weight> HashTable;
451 HashTable Combined(NextPowerOf2(2 * Weights.size()));
452 for (const Weight &W : Weights)
453 combineWeight(Combined[W.TargetNode.Index], W);
455 // Check whether anything changed.
456 if (Weights.size() == Combined.size())
459 // Fill in the new weights.
461 Weights.reserve(Combined.size());
462 for (const auto &I : Combined)
463 Weights.push_back(I.second);
465 static void combineWeights(WeightList &Weights) {
466 // Use a hash table for many successors to keep this linear.
467 if (Weights.size() > 128) {
468 combineWeightsByHashing(Weights);
472 combineWeightsBySorting(Weights);
474 static uint64_t shiftRightAndRound(uint64_t N, int Shift) {
479 return (N >> Shift) + (UINT64_C(1) & N >> (Shift - 1));
481 void Distribution::normalize() {
482 // Early exit for termination nodes.
486 // Only bother if there are multiple successors.
487 if (Weights.size() > 1)
488 combineWeights(Weights);
490 // Early exit when combined into a single successor.
491 if (Weights.size() == 1) {
493 Weights.front().Amount = 1;
497 // Determine how much to shift right so that the total fits into 32-bits.
499 // If we shift at all, shift by 1 extra. Otherwise, the lower limit of 1
500 // for each weight can cause a 32-bit overflow.
504 else if (Total > UINT32_MAX)
505 Shift = 33 - countLeadingZeros(Total);
507 // Early exit if nothing needs to be scaled.
511 // Recompute the total through accumulation (rather than shifting it) so that
512 // it's accurate after shifting.
515 // Sum the weights to each node and shift right if necessary.
516 for (Weight &W : Weights) {
517 // Scale down below UINT32_MAX. Since Shift is larger than necessary, we
518 // can round here without concern about overflow.
519 assert(W.TargetNode.isValid());
520 W.Amount = std::max(UINT64_C(1), shiftRightAndRound(W.Amount, Shift));
521 assert(W.Amount <= UINT32_MAX);
526 assert(Total <= UINT32_MAX);
529 void BlockFrequencyInfoImplBase::clear() {
530 // Swap with a default-constructed std::vector, since std::vector<>::clear()
531 // does not actually clear heap storage.
532 std::vector<FrequencyData>().swap(Freqs);
533 std::vector<WorkingData>().swap(Working);
537 /// \brief Clear all memory not needed downstream.
539 /// Releases all memory not used downstream. In particular, saves Freqs.
540 static void cleanup(BlockFrequencyInfoImplBase &BFI) {
541 std::vector<FrequencyData> SavedFreqs(std::move(BFI.Freqs));
543 BFI.Freqs = std::move(SavedFreqs);
546 bool BlockFrequencyInfoImplBase::addToDist(Distribution &Dist,
547 const LoopData *OuterLoop,
548 const BlockNode &Pred,
549 const BlockNode &Succ,
554 auto isLoopHeader = [&OuterLoop](const BlockNode &Node) {
555 return OuterLoop && OuterLoop->isHeader(Node);
558 BlockNode Resolved = Working[Succ.Index].getResolvedNode();
561 auto debugSuccessor = [&](const char *Type) {
563 << " [" << Type << "] weight = " << Weight;
564 if (!isLoopHeader(Resolved))
565 dbgs() << ", succ = " << getBlockName(Succ);
566 if (Resolved != Succ)
567 dbgs() << ", resolved = " << getBlockName(Resolved);
570 (void)debugSuccessor;
573 if (isLoopHeader(Resolved)) {
574 DEBUG(debugSuccessor("backedge"));
575 Dist.addBackedge(OuterLoop->getHeader(), Weight);
579 if (Working[Resolved.Index].getContainingLoop() != OuterLoop) {
580 DEBUG(debugSuccessor(" exit "));
581 Dist.addExit(Resolved, Weight);
585 if (Resolved < Pred) {
586 if (!isLoopHeader(Pred)) {
587 // If OuterLoop is an irreducible loop, we can't actually handle this.
588 assert((!OuterLoop || !OuterLoop->isIrreducible()) &&
589 "unhandled irreducible control flow");
591 // Irreducible backedge. Abort.
592 DEBUG(debugSuccessor("abort!!!"));
596 // If "Pred" is a loop header, then this isn't really a backedge; rather,
597 // OuterLoop must be irreducible. These false backedges can come only from
598 // secondary loop headers.
599 assert(OuterLoop && OuterLoop->isIrreducible() && !isLoopHeader(Resolved) &&
600 "unhandled irreducible control flow");
603 DEBUG(debugSuccessor(" local "));
604 Dist.addLocal(Resolved, Weight);
608 bool BlockFrequencyInfoImplBase::addLoopSuccessorsToDist(
609 const LoopData *OuterLoop, LoopData &Loop, Distribution &Dist) {
610 // Copy the exit map into Dist.
611 for (const auto &I : Loop.Exits)
612 if (!addToDist(Dist, OuterLoop, Loop.getHeader(), I.first,
614 // Irreducible backedge.
620 /// \brief Get the maximum allowed loop scale.
622 /// Gives the maximum number of estimated iterations allowed for a loop. Very
623 /// large numbers cause problems downstream (even within 64-bits).
624 static Float getMaxLoopScale() { return Float(1, 12); }
626 /// \brief Compute the loop scale for a loop.
627 void BlockFrequencyInfoImplBase::computeLoopScale(LoopData &Loop) {
628 // Compute loop scale.
629 DEBUG(dbgs() << "compute-loop-scale: " << getLoopName(Loop) << "\n");
631 // LoopScale == 1 / ExitMass
632 // ExitMass == HeadMass - BackedgeMass
633 BlockMass ExitMass = BlockMass::getFull() - Loop.BackedgeMass;
635 // Block scale stores the inverse of the scale.
636 Loop.Scale = ExitMass.toFloat().inverse();
638 DEBUG(dbgs() << " - exit-mass = " << ExitMass << " (" << BlockMass::getFull()
639 << " - " << Loop.BackedgeMass << ")\n"
640 << " - scale = " << Loop.Scale << "\n");
642 if (Loop.Scale > getMaxLoopScale()) {
643 Loop.Scale = getMaxLoopScale();
644 DEBUG(dbgs() << " - reduced-to-max-scale: " << getMaxLoopScale() << "\n");
648 /// \brief Package up a loop.
649 void BlockFrequencyInfoImplBase::packageLoop(LoopData &Loop) {
650 DEBUG(dbgs() << "packaging-loop: " << getLoopName(Loop) << "\n");
652 // Clear the subloop exits to prevent quadratic memory usage.
653 for (const BlockNode &M : Loop.Nodes) {
654 if (auto *Loop = Working[M.Index].getPackagedLoop())
656 DEBUG(dbgs() << " - node: " << getBlockName(M.Index) << "\n");
658 Loop.IsPackaged = true;
661 void BlockFrequencyInfoImplBase::distributeMass(const BlockNode &Source,
663 Distribution &Dist) {
664 BlockMass Mass = Working[Source.Index].getMass();
665 DEBUG(dbgs() << " => mass: " << Mass << "\n");
667 // Distribute mass to successors as laid out in Dist.
668 DitheringDistributer D(Dist, Mass);
671 auto debugAssign = [&](const BlockNode &T, const BlockMass &M,
673 dbgs() << " => assign " << M << " (" << D.RemMass << ")";
675 dbgs() << " [" << Desc << "]";
677 dbgs() << " to " << getBlockName(T);
683 for (const Weight &W : Dist.Weights) {
684 // Check for a local edge (non-backedge and non-exit).
685 BlockMass Taken = D.takeMass(W.Amount);
686 if (W.Type == Weight::Local) {
687 Working[W.TargetNode.Index].getMass() += Taken;
688 DEBUG(debugAssign(W.TargetNode, Taken, nullptr));
692 // Backedges and exits only make sense if we're processing a loop.
693 assert(OuterLoop && "backedge or exit outside of loop");
695 // Check for a backedge.
696 if (W.Type == Weight::Backedge) {
697 OuterLoop->BackedgeMass += Taken;
698 DEBUG(debugAssign(BlockNode(), Taken, "back"));
702 // This must be an exit.
703 assert(W.Type == Weight::Exit);
704 OuterLoop->Exits.push_back(std::make_pair(W.TargetNode, Taken));
705 DEBUG(debugAssign(W.TargetNode, Taken, "exit"));
709 static void convertFloatingToInteger(BlockFrequencyInfoImplBase &BFI,
710 const Float &Min, const Float &Max) {
711 // Scale the Factor to a size that creates integers. Ideally, integers would
712 // be scaled so that Max == UINT64_MAX so that they can be best
713 // differentiated. However, the register allocator currently deals poorly
714 // with large numbers. Instead, push Min up a little from 1 to give some
715 // room to differentiate small, unequal numbers.
717 // TODO: fix issues downstream so that ScalingFactor can be Float(1,64)/Max.
718 Float ScalingFactor = Min.inverse();
719 if ((Max / Min).lg() < 60)
722 // Translate the floats to integers.
723 DEBUG(dbgs() << "float-to-int: min = " << Min << ", max = " << Max
724 << ", factor = " << ScalingFactor << "\n");
725 for (size_t Index = 0; Index < BFI.Freqs.size(); ++Index) {
726 Float Scaled = BFI.Freqs[Index].Floating * ScalingFactor;
727 BFI.Freqs[Index].Integer = std::max(UINT64_C(1), Scaled.toInt<uint64_t>());
728 DEBUG(dbgs() << " - " << BFI.getBlockName(Index) << ": float = "
729 << BFI.Freqs[Index].Floating << ", scaled = " << Scaled
730 << ", int = " << BFI.Freqs[Index].Integer << "\n");
734 /// \brief Unwrap a loop package.
736 /// Visits all the members of a loop, adjusting their BlockData according to
737 /// the loop's pseudo-node.
738 static void unwrapLoop(BlockFrequencyInfoImplBase &BFI, LoopData &Loop) {
739 DEBUG(dbgs() << "unwrap-loop-package: " << BFI.getLoopName(Loop)
740 << ": mass = " << Loop.Mass << ", scale = " << Loop.Scale
742 Loop.Scale *= Loop.Mass.toFloat();
743 Loop.IsPackaged = false;
744 DEBUG(dbgs() << " => combined-scale = " << Loop.Scale << "\n");
746 // Propagate the head scale through the loop. Since members are visited in
747 // RPO, the head scale will be updated by the loop scale first, and then the
748 // final head scale will be used for updated the rest of the members.
749 for (const BlockNode &N : Loop.Nodes) {
750 const auto &Working = BFI.Working[N.Index];
751 Float &F = Working.isAPackage() ? Working.getPackagedLoop()->Scale
752 : BFI.Freqs[N.Index].Floating;
753 Float New = Loop.Scale * F;
754 DEBUG(dbgs() << " - " << BFI.getBlockName(N) << ": " << F << " => " << New
760 void BlockFrequencyInfoImplBase::unwrapLoops() {
761 // Set initial frequencies from loop-local masses.
762 for (size_t Index = 0; Index < Working.size(); ++Index)
763 Freqs[Index].Floating = Working[Index].Mass.toFloat();
765 for (LoopData &Loop : Loops)
766 unwrapLoop(*this, Loop);
769 void BlockFrequencyInfoImplBase::finalizeMetrics() {
770 // Unwrap loop packages in reverse post-order, tracking min and max
772 auto Min = Float::getLargest();
773 auto Max = Float::getZero();
774 for (size_t Index = 0; Index < Working.size(); ++Index) {
775 // Update min/max scale.
776 Min = std::min(Min, Freqs[Index].Floating);
777 Max = std::max(Max, Freqs[Index].Floating);
780 // Convert to integers.
781 convertFloatingToInteger(*this, Min, Max);
783 // Clean up data structures.
786 // Print out the final stats.
791 BlockFrequencyInfoImplBase::getBlockFreq(const BlockNode &Node) const {
794 return Freqs[Node.Index].Integer;
797 BlockFrequencyInfoImplBase::getFloatingBlockFreq(const BlockNode &Node) const {
799 return Float::getZero();
800 return Freqs[Node.Index].Floating;
804 BlockFrequencyInfoImplBase::getBlockName(const BlockNode &Node) const {
805 return std::string();
808 BlockFrequencyInfoImplBase::getLoopName(const LoopData &Loop) const {
809 return getBlockName(Loop.getHeader()) + (Loop.isIrreducible() ? "**" : "*");
813 BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
814 const BlockNode &Node) const {
815 return OS << getFloatingBlockFreq(Node);
819 BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
820 const BlockFrequency &Freq) const {
821 Float Block(Freq.getFrequency(), 0);
822 Float Entry(getEntryFreq(), 0);
824 return OS << Block / Entry;
827 void IrreducibleGraph::addNodesInLoop(const BFIBase::LoopData &OuterLoop) {
828 Start = OuterLoop.getHeader();
829 Nodes.reserve(OuterLoop.Nodes.size());
830 for (auto N : OuterLoop.Nodes)
834 void IrreducibleGraph::addNodesInFunction() {
836 for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index)
837 if (!BFI.Working[Index].isPackaged())
841 void IrreducibleGraph::indexNodes() {
842 for (auto &I : Nodes)
843 Lookup[I.Node.Index] = &I;
845 void IrreducibleGraph::addEdge(IrrNode &Irr, const BlockNode &Succ,
846 const BFIBase::LoopData *OuterLoop) {
847 if (OuterLoop && OuterLoop->isHeader(Succ))
849 auto L = Lookup.find(Succ.Index);
850 if (L == Lookup.end())
852 IrrNode &SuccIrr = *L->second;
853 Irr.Edges.push_back(&SuccIrr);
854 SuccIrr.Edges.push_front(&Irr);
859 template <> struct GraphTraits<IrreducibleGraph> {
860 typedef bfi_detail::IrreducibleGraph GraphT;
862 typedef const GraphT::IrrNode NodeType;
863 typedef GraphT::IrrNode::iterator ChildIteratorType;
865 static const NodeType *getEntryNode(const GraphT &G) {
868 static ChildIteratorType child_begin(NodeType *N) { return N->succ_begin(); }
869 static ChildIteratorType child_end(NodeType *N) { return N->succ_end(); }
873 /// \brief Find extra irreducible headers.
875 /// Find entry blocks and other blocks with backedges, which exist when \c G
876 /// contains irreducible sub-SCCs.
877 static void findIrreducibleHeaders(
878 const BlockFrequencyInfoImplBase &BFI,
879 const IrreducibleGraph &G,
880 const std::vector<const IrreducibleGraph::IrrNode *> &SCC,
881 LoopData::NodeList &Headers, LoopData::NodeList &Others) {
882 // Map from nodes in the SCC to whether it's an entry block.
883 SmallDenseMap<const IrreducibleGraph::IrrNode *, bool, 8> InSCC;
885 // InSCC also acts the set of nodes in the graph. Seed it.
886 for (const auto *I : SCC)
889 for (auto I = InSCC.begin(), E = InSCC.end(); I != E; ++I) {
890 auto &Irr = *I->first;
891 for (const auto *P : make_range(Irr.pred_begin(), Irr.pred_end())) {
895 // This is an entry block.
897 Headers.push_back(Irr.Node);
898 DEBUG(dbgs() << " => entry = " << BFI.getBlockName(Irr.Node) << "\n");
902 assert(Headers.size() >= 2 && "Should be irreducible");
903 if (Headers.size() == InSCC.size()) {
904 // Every block is a header.
905 std::sort(Headers.begin(), Headers.end());
909 // Look for extra headers from irreducible sub-SCCs.
910 for (const auto &I : InSCC) {
911 // Entry blocks are already headers.
915 auto &Irr = *I.first;
916 for (const auto *P : make_range(Irr.pred_begin(), Irr.pred_end())) {
917 // Skip forward edges.
918 if (P->Node < Irr.Node)
921 // Skip predecessors from entry blocks. These can have inverted
926 // Store the extra header.
927 Headers.push_back(Irr.Node);
928 DEBUG(dbgs() << " => extra = " << BFI.getBlockName(Irr.Node) << "\n");
931 if (Headers.back() == Irr.Node)
932 // Added this as a header.
935 // This is not a header.
936 Others.push_back(Irr.Node);
937 DEBUG(dbgs() << " => other = " << BFI.getBlockName(Irr.Node) << "\n");
939 std::sort(Headers.begin(), Headers.end());
940 std::sort(Others.begin(), Others.end());
943 static void createIrreducibleLoop(
944 BlockFrequencyInfoImplBase &BFI, const IrreducibleGraph &G,
945 LoopData *OuterLoop, std::list<LoopData>::iterator Insert,
946 const std::vector<const IrreducibleGraph::IrrNode *> &SCC) {
947 // Translate the SCC into RPO.
948 DEBUG(dbgs() << " - found-scc\n");
950 LoopData::NodeList Headers;
951 LoopData::NodeList Others;
952 findIrreducibleHeaders(BFI, G, SCC, Headers, Others);
954 auto Loop = BFI.Loops.emplace(Insert, OuterLoop, Headers.begin(),
955 Headers.end(), Others.begin(), Others.end());
957 // Update loop hierarchy.
958 for (const auto &N : Loop->Nodes)
959 if (BFI.Working[N.Index].isLoopHeader())
960 BFI.Working[N.Index].Loop->Parent = &*Loop;
962 BFI.Working[N.Index].Loop = &*Loop;
965 iterator_range<std::list<LoopData>::iterator>
966 BlockFrequencyInfoImplBase::analyzeIrreducible(
967 const IrreducibleGraph &G, LoopData *OuterLoop,
968 std::list<LoopData>::iterator Insert) {
969 assert((OuterLoop == nullptr) == (Insert == Loops.begin()));
970 auto Prev = OuterLoop ? std::prev(Insert) : Loops.end();
972 for (auto I = scc_begin(G); !I.isAtEnd(); ++I) {
976 // Translate the SCC into RPO.
977 createIrreducibleLoop(*this, G, OuterLoop, Insert, *I);
981 return make_range(std::next(Prev), Insert);
982 return make_range(Loops.begin(), Insert);
986 BlockFrequencyInfoImplBase::updateLoopWithIrreducible(LoopData &OuterLoop) {
987 OuterLoop.Exits.clear();
988 OuterLoop.BackedgeMass = BlockMass::getEmpty();
989 auto O = OuterLoop.Nodes.begin() + 1;
990 for (auto I = O, E = OuterLoop.Nodes.end(); I != E; ++I)
991 if (!Working[I->Index].isPackaged())
993 OuterLoop.Nodes.erase(O, OuterLoop.Nodes.end());