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 BlockMass &BlockMass::operator*=(const BranchProbability &P) {
315 uint32_t N = P.getNumerator(), D = P.getDenominator();
316 assert(D && "divide by 0");
317 assert(N <= D && "fraction greater than 1");
319 // Fast path for multiplying by 1.0.
323 // Get as much precision as we can.
324 int Shift = countLeadingZeros(Mass);
325 uint64_t ShiftedQuotient = (Mass << Shift) / D;
326 uint64_t Product = ShiftedQuotient * N >> Shift;
328 // Now check for what's lost.
329 uint64_t Left = ShiftedQuotient * (D - N) >> Shift;
330 uint64_t Lost = Mass - Product - Left;
332 // TODO: prove this assertion.
333 assert(Lost <= UINT32_MAX);
335 // Take the product plus a portion of the spoils.
336 Mass = Product + Lost * N / D;
340 UnsignedFloat<uint64_t> BlockMass::toFloat() const {
342 return UnsignedFloat<uint64_t>(1, 0);
343 return UnsignedFloat<uint64_t>(getMass() + 1, -64);
346 void BlockMass::dump() const { print(dbgs()); }
348 static char getHexDigit(int N) {
354 raw_ostream &BlockMass::print(raw_ostream &OS) const {
355 for (int Digits = 0; Digits < 16; ++Digits)
356 OS << getHexDigit(Mass >> (60 - Digits * 4) & 0xf);
360 //===----------------------------------------------------------------------===//
362 // BlockFrequencyInfoImpl implementation.
364 //===----------------------------------------------------------------------===//
367 typedef BlockFrequencyInfoImplBase::BlockNode BlockNode;
368 typedef BlockFrequencyInfoImplBase::Distribution Distribution;
369 typedef BlockFrequencyInfoImplBase::Distribution::WeightList WeightList;
370 typedef BlockFrequencyInfoImplBase::Float Float;
371 typedef BlockFrequencyInfoImplBase::LoopData LoopData;
372 typedef BlockFrequencyInfoImplBase::Weight Weight;
373 typedef BlockFrequencyInfoImplBase::FrequencyData FrequencyData;
375 /// \brief Dithering mass distributer.
377 /// This class splits up a single mass into portions by weight, dithering to
378 /// spread out error. No mass is lost. The dithering precision depends on the
379 /// precision of the product of \a BlockMass and \a BranchProbability.
381 /// The distribution algorithm follows.
383 /// 1. Initialize by saving the sum of the weights in \a RemWeight and the
384 /// mass to distribute in \a RemMass.
386 /// 2. For each portion:
388 /// 1. Construct a branch probability, P, as the portion's weight divided
389 /// by the current value of \a RemWeight.
390 /// 2. Calculate the portion's mass as \a RemMass times P.
391 /// 3. Update \a RemWeight and \a RemMass at each portion by subtracting
392 /// the current portion's weight and mass.
393 struct DitheringDistributer {
397 DitheringDistributer(Distribution &Dist, const BlockMass &Mass);
399 BlockMass takeMass(uint32_t Weight);
403 DitheringDistributer::DitheringDistributer(Distribution &Dist,
404 const BlockMass &Mass) {
406 RemWeight = Dist.Total;
410 BlockMass DitheringDistributer::takeMass(uint32_t Weight) {
411 assert(Weight && "invalid weight");
412 assert(Weight <= RemWeight);
413 BlockMass Mass = RemMass * BranchProbability(Weight, RemWeight);
415 // Decrement totals (dither).
421 void Distribution::add(const BlockNode &Node, uint64_t Amount,
422 Weight::DistType Type) {
423 assert(Amount && "invalid weight of 0");
424 uint64_t NewTotal = Total + Amount;
426 // Check for overflow. It should be impossible to overflow twice.
427 bool IsOverflow = NewTotal < Total;
428 assert(!(DidOverflow && IsOverflow) && "unexpected repeated overflow");
429 DidOverflow |= IsOverflow;
439 Weights.push_back(W);
442 static void combineWeight(Weight &W, const Weight &OtherW) {
443 assert(OtherW.TargetNode.isValid());
448 assert(W.Type == OtherW.Type);
449 assert(W.TargetNode == OtherW.TargetNode);
450 assert(W.Amount < W.Amount + OtherW.Amount && "Unexpected overflow");
451 W.Amount += OtherW.Amount;
453 static void combineWeightsBySorting(WeightList &Weights) {
454 // Sort so edges to the same node are adjacent.
455 std::sort(Weights.begin(), Weights.end(),
457 const Weight &R) { return L.TargetNode < R.TargetNode; });
459 // Combine adjacent edges.
460 WeightList::iterator O = Weights.begin();
461 for (WeightList::const_iterator I = O, L = O, E = Weights.end(); I != E;
465 // Find the adjacent weights to the same node.
466 for (++L; L != E && I->TargetNode == L->TargetNode; ++L)
467 combineWeight(*O, *L);
470 // Erase extra entries.
471 Weights.erase(O, Weights.end());
474 static void combineWeightsByHashing(WeightList &Weights) {
475 // Collect weights into a DenseMap.
476 typedef DenseMap<BlockNode::IndexType, Weight> HashTable;
477 HashTable Combined(NextPowerOf2(2 * Weights.size()));
478 for (const Weight &W : Weights)
479 combineWeight(Combined[W.TargetNode.Index], W);
481 // Check whether anything changed.
482 if (Weights.size() == Combined.size())
485 // Fill in the new weights.
487 Weights.reserve(Combined.size());
488 for (const auto &I : Combined)
489 Weights.push_back(I.second);
491 static void combineWeights(WeightList &Weights) {
492 // Use a hash table for many successors to keep this linear.
493 if (Weights.size() > 128) {
494 combineWeightsByHashing(Weights);
498 combineWeightsBySorting(Weights);
500 static uint64_t shiftRightAndRound(uint64_t N, int Shift) {
505 return (N >> Shift) + (UINT64_C(1) & N >> (Shift - 1));
507 void Distribution::normalize() {
508 // Early exit for termination nodes.
512 // Only bother if there are multiple successors.
513 if (Weights.size() > 1)
514 combineWeights(Weights);
516 // Early exit when combined into a single successor.
517 if (Weights.size() == 1) {
519 Weights.front().Amount = 1;
523 // Determine how much to shift right so that the total fits into 32-bits.
525 // If we shift at all, shift by 1 extra. Otherwise, the lower limit of 1
526 // for each weight can cause a 32-bit overflow.
530 else if (Total > UINT32_MAX)
531 Shift = 33 - countLeadingZeros(Total);
533 // Early exit if nothing needs to be scaled.
537 // Recompute the total through accumulation (rather than shifting it) so that
538 // it's accurate after shifting.
541 // Sum the weights to each node and shift right if necessary.
542 for (Weight &W : Weights) {
543 // Scale down below UINT32_MAX. Since Shift is larger than necessary, we
544 // can round here without concern about overflow.
545 assert(W.TargetNode.isValid());
546 W.Amount = std::max(UINT64_C(1), shiftRightAndRound(W.Amount, Shift));
547 assert(W.Amount <= UINT32_MAX);
552 assert(Total <= UINT32_MAX);
555 void BlockFrequencyInfoImplBase::clear() {
556 // Swap with a default-constructed std::vector, since std::vector<>::clear()
557 // does not actually clear heap storage.
558 std::vector<FrequencyData>().swap(Freqs);
559 std::vector<WorkingData>().swap(Working);
563 /// \brief Clear all memory not needed downstream.
565 /// Releases all memory not used downstream. In particular, saves Freqs.
566 static void cleanup(BlockFrequencyInfoImplBase &BFI) {
567 std::vector<FrequencyData> SavedFreqs(std::move(BFI.Freqs));
569 BFI.Freqs = std::move(SavedFreqs);
572 bool BlockFrequencyInfoImplBase::addToDist(Distribution &Dist,
573 const LoopData *OuterLoop,
574 const BlockNode &Pred,
575 const BlockNode &Succ,
580 auto isLoopHeader = [&OuterLoop](const BlockNode &Node) {
581 return OuterLoop && OuterLoop->isHeader(Node);
584 BlockNode Resolved = Working[Succ.Index].getResolvedNode();
587 auto debugSuccessor = [&](const char *Type) {
589 << " [" << Type << "] weight = " << Weight;
590 if (!isLoopHeader(Resolved))
591 dbgs() << ", succ = " << getBlockName(Succ);
592 if (Resolved != Succ)
593 dbgs() << ", resolved = " << getBlockName(Resolved);
596 (void)debugSuccessor;
599 if (isLoopHeader(Resolved)) {
600 DEBUG(debugSuccessor("backedge"));
601 Dist.addBackedge(OuterLoop->getHeader(), Weight);
605 if (Working[Resolved.Index].getContainingLoop() != OuterLoop) {
606 DEBUG(debugSuccessor(" exit "));
607 Dist.addExit(Resolved, Weight);
611 if (Resolved < Pred) {
612 if (!isLoopHeader(Pred)) {
613 // If OuterLoop is an irreducible loop, we can't actually handle this.
614 assert((!OuterLoop || !OuterLoop->isIrreducible()) &&
615 "unhandled irreducible control flow");
617 // Irreducible backedge. Abort.
618 DEBUG(debugSuccessor("abort!!!"));
622 // If "Pred" is a loop header, then this isn't really a backedge; rather,
623 // OuterLoop must be irreducible. These false backedges can come only from
624 // secondary loop headers.
625 assert(OuterLoop && OuterLoop->isIrreducible() && !isLoopHeader(Resolved) &&
626 "unhandled irreducible control flow");
629 DEBUG(debugSuccessor(" local "));
630 Dist.addLocal(Resolved, Weight);
634 bool BlockFrequencyInfoImplBase::addLoopSuccessorsToDist(
635 const LoopData *OuterLoop, LoopData &Loop, Distribution &Dist) {
636 // Copy the exit map into Dist.
637 for (const auto &I : Loop.Exits)
638 if (!addToDist(Dist, OuterLoop, Loop.getHeader(), I.first,
640 // Irreducible backedge.
646 /// \brief Get the maximum allowed loop scale.
648 /// Gives the maximum number of estimated iterations allowed for a loop. Very
649 /// large numbers cause problems downstream (even within 64-bits).
650 static Float getMaxLoopScale() { return Float(1, 12); }
652 /// \brief Compute the loop scale for a loop.
653 void BlockFrequencyInfoImplBase::computeLoopScale(LoopData &Loop) {
654 // Compute loop scale.
655 DEBUG(dbgs() << "compute-loop-scale: " << getLoopName(Loop) << "\n");
657 // LoopScale == 1 / ExitMass
658 // ExitMass == HeadMass - BackedgeMass
659 BlockMass ExitMass = BlockMass::getFull() - Loop.BackedgeMass;
661 // Block scale stores the inverse of the scale.
662 Loop.Scale = ExitMass.toFloat().inverse();
664 DEBUG(dbgs() << " - exit-mass = " << ExitMass << " (" << BlockMass::getFull()
665 << " - " << Loop.BackedgeMass << ")\n"
666 << " - scale = " << Loop.Scale << "\n");
668 if (Loop.Scale > getMaxLoopScale()) {
669 Loop.Scale = getMaxLoopScale();
670 DEBUG(dbgs() << " - reduced-to-max-scale: " << getMaxLoopScale() << "\n");
674 /// \brief Package up a loop.
675 void BlockFrequencyInfoImplBase::packageLoop(LoopData &Loop) {
676 DEBUG(dbgs() << "packaging-loop: " << getLoopName(Loop) << "\n");
678 // Clear the subloop exits to prevent quadratic memory usage.
679 for (const BlockNode &M : Loop.Nodes) {
680 if (auto *Loop = Working[M.Index].getPackagedLoop())
682 DEBUG(dbgs() << " - node: " << getBlockName(M.Index) << "\n");
684 Loop.IsPackaged = true;
687 void BlockFrequencyInfoImplBase::distributeMass(const BlockNode &Source,
689 Distribution &Dist) {
690 BlockMass Mass = Working[Source.Index].getMass();
691 DEBUG(dbgs() << " => mass: " << Mass << "\n");
693 // Distribute mass to successors as laid out in Dist.
694 DitheringDistributer D(Dist, Mass);
697 auto debugAssign = [&](const BlockNode &T, const BlockMass &M,
699 dbgs() << " => assign " << M << " (" << D.RemMass << ")";
701 dbgs() << " [" << Desc << "]";
703 dbgs() << " to " << getBlockName(T);
709 for (const Weight &W : Dist.Weights) {
710 // Check for a local edge (non-backedge and non-exit).
711 BlockMass Taken = D.takeMass(W.Amount);
712 if (W.Type == Weight::Local) {
713 Working[W.TargetNode.Index].getMass() += Taken;
714 DEBUG(debugAssign(W.TargetNode, Taken, nullptr));
718 // Backedges and exits only make sense if we're processing a loop.
719 assert(OuterLoop && "backedge or exit outside of loop");
721 // Check for a backedge.
722 if (W.Type == Weight::Backedge) {
723 OuterLoop->BackedgeMass += Taken;
724 DEBUG(debugAssign(BlockNode(), Taken, "back"));
728 // This must be an exit.
729 assert(W.Type == Weight::Exit);
730 OuterLoop->Exits.push_back(std::make_pair(W.TargetNode, Taken));
731 DEBUG(debugAssign(W.TargetNode, Taken, "exit"));
735 static void convertFloatingToInteger(BlockFrequencyInfoImplBase &BFI,
736 const Float &Min, const Float &Max) {
737 // Scale the Factor to a size that creates integers. Ideally, integers would
738 // be scaled so that Max == UINT64_MAX so that they can be best
739 // differentiated. However, the register allocator currently deals poorly
740 // with large numbers. Instead, push Min up a little from 1 to give some
741 // room to differentiate small, unequal numbers.
743 // TODO: fix issues downstream so that ScalingFactor can be Float(1,64)/Max.
744 Float ScalingFactor = Min.inverse();
745 if ((Max / Min).lg() < 60)
748 // Translate the floats to integers.
749 DEBUG(dbgs() << "float-to-int: min = " << Min << ", max = " << Max
750 << ", factor = " << ScalingFactor << "\n");
751 for (size_t Index = 0; Index < BFI.Freqs.size(); ++Index) {
752 Float Scaled = BFI.Freqs[Index].Floating * ScalingFactor;
753 BFI.Freqs[Index].Integer = std::max(UINT64_C(1), Scaled.toInt<uint64_t>());
754 DEBUG(dbgs() << " - " << BFI.getBlockName(Index) << ": float = "
755 << BFI.Freqs[Index].Floating << ", scaled = " << Scaled
756 << ", int = " << BFI.Freqs[Index].Integer << "\n");
760 /// \brief Unwrap a loop package.
762 /// Visits all the members of a loop, adjusting their BlockData according to
763 /// the loop's pseudo-node.
764 static void unwrapLoop(BlockFrequencyInfoImplBase &BFI, LoopData &Loop) {
765 DEBUG(dbgs() << "unwrap-loop-package: " << BFI.getLoopName(Loop)
766 << ": mass = " << Loop.Mass << ", scale = " << Loop.Scale
768 Loop.Scale *= Loop.Mass.toFloat();
769 Loop.IsPackaged = false;
770 DEBUG(dbgs() << " => combined-scale = " << Loop.Scale << "\n");
772 // Propagate the head scale through the loop. Since members are visited in
773 // RPO, the head scale will be updated by the loop scale first, and then the
774 // final head scale will be used for updated the rest of the members.
775 for (const BlockNode &N : Loop.Nodes) {
776 const auto &Working = BFI.Working[N.Index];
777 Float &F = Working.isAPackage() ? Working.getPackagedLoop()->Scale
778 : BFI.Freqs[N.Index].Floating;
779 Float New = Loop.Scale * F;
780 DEBUG(dbgs() << " - " << BFI.getBlockName(N) << ": " << F << " => " << New
786 void BlockFrequencyInfoImplBase::unwrapLoops() {
787 // Set initial frequencies from loop-local masses.
788 for (size_t Index = 0; Index < Working.size(); ++Index)
789 Freqs[Index].Floating = Working[Index].Mass.toFloat();
791 for (LoopData &Loop : Loops)
792 unwrapLoop(*this, Loop);
795 void BlockFrequencyInfoImplBase::finalizeMetrics() {
796 // Unwrap loop packages in reverse post-order, tracking min and max
798 auto Min = Float::getLargest();
799 auto Max = Float::getZero();
800 for (size_t Index = 0; Index < Working.size(); ++Index) {
801 // Update min/max scale.
802 Min = std::min(Min, Freqs[Index].Floating);
803 Max = std::max(Max, Freqs[Index].Floating);
806 // Convert to integers.
807 convertFloatingToInteger(*this, Min, Max);
809 // Clean up data structures.
812 // Print out the final stats.
817 BlockFrequencyInfoImplBase::getBlockFreq(const BlockNode &Node) const {
820 return Freqs[Node.Index].Integer;
823 BlockFrequencyInfoImplBase::getFloatingBlockFreq(const BlockNode &Node) const {
825 return Float::getZero();
826 return Freqs[Node.Index].Floating;
830 BlockFrequencyInfoImplBase::getBlockName(const BlockNode &Node) const {
831 return std::string();
834 BlockFrequencyInfoImplBase::getLoopName(const LoopData &Loop) const {
835 return getBlockName(Loop.getHeader()) + (Loop.isIrreducible() ? "**" : "*");
839 BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
840 const BlockNode &Node) const {
841 return OS << getFloatingBlockFreq(Node);
845 BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
846 const BlockFrequency &Freq) const {
847 Float Block(Freq.getFrequency(), 0);
848 Float Entry(getEntryFreq(), 0);
850 return OS << Block / Entry;
853 void IrreducibleGraph::addNodesInLoop(const BFIBase::LoopData &OuterLoop) {
854 Start = OuterLoop.getHeader();
855 Nodes.reserve(OuterLoop.Nodes.size());
856 for (auto N : OuterLoop.Nodes)
860 void IrreducibleGraph::addNodesInFunction() {
862 for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index)
863 if (!BFI.Working[Index].isPackaged())
867 void IrreducibleGraph::indexNodes() {
868 for (auto &I : Nodes)
869 Lookup[I.Node.Index] = &I;
871 void IrreducibleGraph::addEdge(IrrNode &Irr, const BlockNode &Succ,
872 const BFIBase::LoopData *OuterLoop) {
873 if (OuterLoop && OuterLoop->isHeader(Succ))
875 auto L = Lookup.find(Succ.Index);
876 if (L == Lookup.end())
878 IrrNode &SuccIrr = *L->second;
879 Irr.Edges.push_back(&SuccIrr);
880 SuccIrr.Edges.push_front(&Irr);
885 template <> struct GraphTraits<IrreducibleGraph> {
886 typedef bfi_detail::IrreducibleGraph GraphT;
888 typedef const typename GraphT::IrrNode NodeType;
889 typedef typename GraphT::IrrNode::iterator ChildIteratorType;
891 static const NodeType *getEntryNode(const GraphT &G) {
894 static ChildIteratorType child_begin(NodeType *N) { return N->succ_begin(); }
895 static ChildIteratorType child_end(NodeType *N) { return N->succ_end(); }
899 /// \brief Find extra irreducible headers.
901 /// Find entry blocks and other blocks with backedges, which exist when \c G
902 /// contains irreducible sub-SCCs.
903 static void findIrreducibleHeaders(
904 const BlockFrequencyInfoImplBase &BFI,
905 const IrreducibleGraph &G,
906 const std::vector<const IrreducibleGraph::IrrNode *> &SCC,
907 LoopData::NodeList &Headers, LoopData::NodeList &Others) {
908 // Map from nodes in the SCC to whether it's an entry block.
909 SmallDenseMap<const IrreducibleGraph::IrrNode *, bool, 8> InSCC;
911 // InSCC also acts the set of nodes in the graph. Seed it.
912 for (const auto *I : SCC)
915 for (auto I = InSCC.begin(), E = InSCC.end(); I != E; ++I) {
916 auto &Irr = *I->first;
917 for (const auto *P : make_range(Irr.pred_begin(), Irr.pred_end())) {
921 // This is an entry block.
923 Headers.push_back(Irr.Node);
924 DEBUG(dbgs() << " => entry = " << BFI.getBlockName(Irr.Node) << "\n");
928 assert(Headers.size() >= 2 && "Should be irreducible");
929 if (Headers.size() == InSCC.size()) {
930 // Every block is a header.
931 std::sort(Headers.begin(), Headers.end());
935 // Look for extra headers from irreducible sub-SCCs.
936 for (const auto &I : InSCC) {
937 // Entry blocks are already headers.
941 auto &Irr = *I.first;
942 for (const auto *P : make_range(Irr.pred_begin(), Irr.pred_end())) {
943 // Skip forward edges.
944 if (P->Node < Irr.Node)
947 // Skip predecessors from entry blocks. These can have inverted
952 // Store the extra header.
953 Headers.push_back(Irr.Node);
954 DEBUG(dbgs() << " => extra = " << BFI.getBlockName(Irr.Node) << "\n");
957 if (Headers.back() == Irr.Node)
958 // Added this as a header.
961 // This is not a header.
962 Others.push_back(Irr.Node);
963 DEBUG(dbgs() << " => other = " << BFI.getBlockName(Irr.Node) << "\n");
965 std::sort(Headers.begin(), Headers.end());
966 std::sort(Others.begin(), Others.end());
969 static void createIrreducibleLoop(
970 BlockFrequencyInfoImplBase &BFI, const IrreducibleGraph &G,
971 LoopData *OuterLoop, std::list<LoopData>::iterator Insert,
972 const std::vector<const IrreducibleGraph::IrrNode *> &SCC) {
973 // Translate the SCC into RPO.
974 DEBUG(dbgs() << " - found-scc\n");
976 LoopData::NodeList Headers;
977 LoopData::NodeList Others;
978 findIrreducibleHeaders(BFI, G, SCC, Headers, Others);
980 auto Loop = BFI.Loops.emplace(Insert, OuterLoop, Headers.begin(),
981 Headers.end(), Others.begin(), Others.end());
983 // Update loop hierarchy.
984 for (const auto &N : Loop->Nodes)
985 if (BFI.Working[N.Index].isLoopHeader())
986 BFI.Working[N.Index].Loop->Parent = &*Loop;
988 BFI.Working[N.Index].Loop = &*Loop;
991 iterator_range<std::list<LoopData>::iterator>
992 BlockFrequencyInfoImplBase::analyzeIrreducible(
993 const IrreducibleGraph &G, LoopData *OuterLoop,
994 std::list<LoopData>::iterator Insert) {
995 assert((OuterLoop == nullptr) == (Insert == Loops.begin()));
996 auto Prev = OuterLoop ? std::prev(Insert) : Loops.end();
998 for (auto I = scc_begin(G); !I.isAtEnd(); ++I) {
1002 // Translate the SCC into RPO.
1003 createIrreducibleLoop(*this, G, OuterLoop, Insert, *I);
1007 return make_range(std::next(Prev), Insert);
1008 return make_range(Loops.begin(), Insert);
1012 BlockFrequencyInfoImplBase::updateLoopWithIrreducible(LoopData &OuterLoop) {
1013 OuterLoop.Exits.clear();
1014 OuterLoop.BackedgeMass = BlockMass::getEmpty();
1015 auto O = OuterLoop.Nodes.begin() + 1;
1016 for (auto I = O, E = OuterLoop.Nodes.end(); I != E; ++I)
1017 if (!Working[I->Index].isPackaged())
1019 OuterLoop.Nodes.erase(O, OuterLoop.Nodes.end());