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 #define DEBUG_TYPE "block-freq"
15 #include "llvm/Analysis/BlockFrequencyInfoImpl.h"
16 #include "llvm/ADT/APFloat.h"
17 #include "llvm/Support/raw_ostream.h"
22 //===----------------------------------------------------------------------===//
24 // PositiveFloat implementation.
26 //===----------------------------------------------------------------------===//
27 const int PositiveFloatBase::MaxExponent;
28 const int PositiveFloatBase::MinExponent;
30 static void appendDigit(std::string &Str, unsigned D) {
35 static void appendNumber(std::string &Str, uint64_t N) {
37 appendDigit(Str, N % 10);
42 static bool doesRoundUp(char Digit) {
55 static std::string toStringAPFloat(uint64_t D, int E, unsigned Precision) {
56 assert(E >= PositiveFloatBase::MinExponent);
57 assert(E <= PositiveFloatBase::MaxExponent);
59 // Find a new E, but don't let it increase past MaxExponent.
60 int LeadingZeros = PositiveFloatBase::countLeadingZeros64(D);
61 int NewE = std::min(PositiveFloatBase::MaxExponent, E + 63 - LeadingZeros);
62 int Shift = 63 - (NewE - E);
63 assert(Shift <= LeadingZeros);
64 assert(Shift == LeadingZeros || NewE == PositiveFloatBase::MaxExponent);
68 // Check for a denormal.
69 unsigned AdjustedE = E + 16383;
71 assert(E == PositiveFloatBase::MaxExponent);
75 // Build the float and print it.
76 uint64_t RawBits[2] = {D, AdjustedE};
77 APFloat Float(APFloat::x87DoubleExtended, APInt(80, RawBits));
78 SmallVector<char, 24> Chars;
79 Float.toString(Chars, Precision, 0);
80 return std::string(Chars.begin(), Chars.end());
83 static std::string stripTrailingZeros(std::string Float) {
84 size_t NonZero = Float.find_last_not_of('0');
85 assert(NonZero != std::string::npos && "no . in floating point string");
87 if (Float[NonZero] == '.')
90 return Float.substr(0, NonZero + 1);
93 std::string PositiveFloatBase::toString(uint64_t D, int16_t E, int Width,
98 // Canonicalize exponent and digits.
106 if (int Shift = std::min(int16_t(countLeadingZeros64(D)), E)) {
113 } else if (E > -64) {
115 Below0 = D << (64 + E);
116 } else if (E > -120) {
117 Below0 = D >> (-E - 64);
118 Extra = D << (128 + E);
119 ExtraShift = -64 - E;
122 // Fall back on APFloat for very small and very large numbers.
123 if (!Above0 && !Below0)
124 return toStringAPFloat(D, E, Precision);
126 // Append the digits before the decimal.
128 size_t DigitsOut = 0;
130 appendNumber(Str, Above0);
131 DigitsOut = Str.size();
134 std::reverse(Str.begin(), Str.end());
136 // Return early if there's nothing after the decimal.
140 // Append the decimal and beyond.
142 uint64_t Error = UINT64_C(1) << (64 - Width);
144 // We need to shift Below0 to the right to make space for calculating
145 // digits. Save the precision we're losing in Extra.
146 Extra = (Below0 & 0xf) << 56 | (Extra >> 8);
149 size_t AfterDot = Str.size();
159 Below0 += (Extra >> 60);
160 Extra = Extra & (UINT64_MAX >> 4);
161 appendDigit(Str, Below0 >> 60);
162 Below0 = Below0 & (UINT64_MAX >> 4);
163 if (DigitsOut || Str.back() != '0')
166 } while (Error && (Below0 << 4 | Extra >> 60) >= Error / 2 &&
167 (!Precision || DigitsOut <= Precision || SinceDot < 2));
169 // Return early for maximum precision.
170 if (!Precision || DigitsOut <= Precision)
171 return stripTrailingZeros(Str);
173 // Find where to truncate.
175 std::max(Str.size() - (DigitsOut - Precision), AfterDot + 1);
177 // Check if there's anything to truncate.
178 if (Truncate >= Str.size())
179 return stripTrailingZeros(Str);
181 bool Carry = doesRoundUp(Str[Truncate]);
183 return stripTrailingZeros(Str.substr(0, Truncate));
185 // Round with the first truncated digit.
186 for (std::string::reverse_iterator I(Str.begin() + Truncate), E = Str.rend();
200 // Add "1" in front if we still need to carry.
201 return stripTrailingZeros(std::string(Carry, '1') + Str.substr(0, Truncate));
204 raw_ostream &PositiveFloatBase::print(raw_ostream &OS, uint64_t D, int16_t E,
205 int Width, unsigned Precision) {
206 return OS << toString(D, E, Width, Precision);
209 void PositiveFloatBase::dump(uint64_t D, int16_t E, int Width) {
210 print(dbgs(), D, E, Width, 0) << "[" << Width << ":" << D << "*2^" << E
214 static std::pair<uint64_t, int16_t>
215 getRoundedFloat(uint64_t N, bool ShouldRound, int64_t Shift) {
218 // Rounding caused an overflow.
219 return std::make_pair(UINT64_C(1), Shift + 64);
220 return std::make_pair(N, Shift);
223 std::pair<uint64_t, int16_t> PositiveFloatBase::divide64(uint64_t Dividend,
225 // Input should be sanitized.
229 // Minimize size of divisor.
231 if (int Zeros = countTrailingZeros(Divisor)) {
236 // Check for powers of two.
238 return std::make_pair(Dividend, Shift);
240 // Maximize size of dividend.
241 if (int Zeros = countLeadingZeros64(Dividend)) {
246 // Start with the result of a divide.
247 uint64_t Quotient = Dividend / Divisor;
250 // Continue building the quotient with long division.
252 // TODO: continue with largers digits.
253 while (!(Quotient >> 63) && Dividend) {
254 // Shift Dividend, and check for overflow.
255 bool IsOverflow = Dividend >> 63;
260 bool DoesDivide = IsOverflow || Divisor <= Dividend;
261 Quotient = (Quotient << 1) | DoesDivide;
262 Dividend -= DoesDivide ? Divisor : 0;
266 if (Dividend >= getHalf(Divisor))
268 // Rounding caused an overflow in Quotient.
269 return std::make_pair(UINT64_C(1), Shift + 64);
271 return getRoundedFloat(Quotient, Dividend >= getHalf(Divisor), Shift);
274 static void addWithCarry(uint64_t &Upper, uint64_t &Lower, uint64_t N) {
275 uint64_t NewLower = Lower + (N << 32);
276 Upper += (N >> 32) + (NewLower < Lower);
280 std::pair<uint64_t, int16_t> PositiveFloatBase::multiply64(uint64_t L,
282 // Separate into two 32-bit digits (U.L).
283 uint64_t UL = L >> 32, LL = L & UINT32_MAX, UR = R >> 32, LR = R & UINT32_MAX;
285 // Compute cross products.
286 uint64_t P1 = UL * UR, P2 = UL * LR, P3 = LL * UR, P4 = LL * LR;
288 // Sum into two 64-bit digits.
289 uint64_t Upper = P1, Lower = P4;
290 addWithCarry(Upper, Lower, P2);
291 addWithCarry(Upper, Lower, P3);
293 // Check for the lower 32 bits.
295 return std::make_pair(Lower, 0);
297 // Shift as little as possible to maximize precision.
298 unsigned LeadingZeros = countLeadingZeros64(Upper);
299 int16_t Shift = 64 - LeadingZeros;
301 Upper = Upper << LeadingZeros | Lower >> Shift;
302 bool ShouldRound = Shift && (Lower & UINT64_C(1) << (Shift - 1));
303 return getRoundedFloat(Upper, ShouldRound, Shift);
306 //===----------------------------------------------------------------------===//
308 // BlockMass implementation.
310 //===----------------------------------------------------------------------===//
311 BlockMass &BlockMass::operator*=(const BranchProbability &P) {
312 uint32_t N = P.getNumerator(), D = P.getDenominator();
313 assert(D || "divide by 0");
314 assert(N <= D || "fraction greater than 1");
316 // Fast path for multiplying by 1.0.
320 // Get as much precision as we can.
321 int Shift = countLeadingZeros(Mass);
322 uint64_t ShiftedQuotient = (Mass << Shift) / D;
323 uint64_t Product = ShiftedQuotient * N >> Shift;
325 // Now check for what's lost.
326 uint64_t Left = ShiftedQuotient * (D - N) >> Shift;
327 uint64_t Lost = Mass - Product - Left;
329 // TODO: prove this assertion.
330 assert(Lost <= UINT32_MAX);
332 // Take the product plus a portion of the spoils.
333 Mass = Product + Lost * N / D;
337 PositiveFloat<uint64_t> BlockMass::toFloat() const {
339 return PositiveFloat<uint64_t>(1, 0);
340 return PositiveFloat<uint64_t>(getMass() + 1, -64);
343 void BlockMass::dump() const { print(dbgs()); }
345 static char getHexDigit(int N) {
351 raw_ostream &BlockMass::print(raw_ostream &OS) const {
352 for (int Digits = 0; Digits < 16; ++Digits)
353 OS << getHexDigit(Mass >> (60 - Digits * 4) & 0xf);
357 //===----------------------------------------------------------------------===//
359 // BlockFrequencyInfoImpl implementation.
361 //===----------------------------------------------------------------------===//
364 typedef BlockFrequencyInfoImplBase::BlockNode BlockNode;
365 typedef BlockFrequencyInfoImplBase::Distribution Distribution;
366 typedef BlockFrequencyInfoImplBase::Distribution::WeightList WeightList;
367 typedef BlockFrequencyInfoImplBase::Float Float;
368 typedef BlockFrequencyInfoImplBase::PackagedLoopData PackagedLoopData;
369 typedef BlockFrequencyInfoImplBase::Weight Weight;
370 typedef BlockFrequencyInfoImplBase::FrequencyData FrequencyData;
372 /// \brief Stack entry describing a loop.
373 struct LoopStackEntry {
375 BlockNode LatestBackedge;
378 /// \brief Stack describing currently open loops.
380 std::vector<LoopStackEntry> OpenLoops;
382 void push(const BlockNode &LoopHead, const BlockNode &LatestBackedge) {
383 assert(LoopHead.isValid());
384 assert(LatestBackedge.isValid());
385 OpenLoops.push_back({LoopHead, LatestBackedge});
387 void pop(const BlockNode &FinishedNode) {
388 while (!empty() && top().LatestBackedge <= FinishedNode)
389 OpenLoops.pop_back();
391 bool empty() const { return OpenLoops.empty(); }
392 const LoopStackEntry &top() const {
393 assert(!OpenLoops.empty());
394 return OpenLoops.back();
396 void adjustAfterFinishing(const BlockNode &Current,
397 const BlockNode &LatestBackedge) {
399 if (LatestBackedge.isValid() && LatestBackedge > Current)
400 push(Current, LatestBackedge);
404 /// \brief Dithering mass distributer.
406 /// This class splits up a single mass into portions by weight, dithering to
407 /// spread out error. No mass is lost. The dithering precision depends on the
408 /// precision of the product of \a BlockMass and \a BranchProbability.
410 /// The distribution algorithm follows.
412 /// 1. Initialize by saving the sum of the weights in \a RemWeight and the
413 /// mass to distribute in \a RemMass.
415 /// 2. For each portion:
417 /// 1. Construct a branch probability, P, as the portion's weight divided
418 /// by the current value of \a RemWeight.
419 /// 2. Calculate the portion's mass as \a RemMass times P.
420 /// 3. Update \a RemWeight and \a RemMass at each portion by subtracting
421 /// the current portion's weight and mass.
423 /// Mass is distributed in two ways: full distribution and forward
424 /// distribution. The latter ignores backedges, and uses the parallel fields
425 /// \a RemForwardWeight and \a RemForwardMass.
426 struct DitheringDistributer {
428 uint32_t RemForwardWeight;
431 BlockMass RemForwardMass;
433 DitheringDistributer(Distribution &Dist, const BlockMass &Mass);
435 BlockMass takeLocalMass(uint32_t Weight) {
436 (void)takeMass(Weight);
437 return takeForwardMass(Weight);
439 BlockMass takeExitMass(uint32_t Weight) {
440 (void)takeForwardMass(Weight);
441 return takeMass(Weight);
443 BlockMass takeBackedgeMass(uint32_t Weight) { return takeMass(Weight); }
446 BlockMass takeForwardMass(uint32_t Weight);
447 BlockMass takeMass(uint32_t Weight);
451 DitheringDistributer::DitheringDistributer(Distribution &Dist,
452 const BlockMass &Mass) {
454 RemWeight = Dist.Total;
455 RemForwardWeight = Dist.ForwardTotal;
457 RemForwardMass = Dist.ForwardTotal ? Mass : BlockMass();
460 BlockMass DitheringDistributer::takeForwardMass(uint32_t Weight) {
461 // Compute the amount of mass to take.
462 assert(Weight && "invalid weight");
463 assert(Weight <= RemForwardWeight);
464 BlockMass Mass = RemForwardMass * BranchProbability(Weight, RemForwardWeight);
466 // Decrement totals (dither).
467 RemForwardWeight -= Weight;
468 RemForwardMass -= Mass;
471 BlockMass DitheringDistributer::takeMass(uint32_t Weight) {
472 assert(Weight && "invalid weight");
473 assert(Weight <= RemWeight);
474 BlockMass Mass = RemMass * BranchProbability(Weight, RemWeight);
476 // Decrement totals (dither).
482 void Distribution::add(const BlockNode &Node, uint64_t Amount,
483 Weight::DistType Type) {
484 assert(Amount && "invalid weight of 0");
485 uint64_t NewTotal = Total + Amount;
487 // Check for overflow. It should be impossible to overflow twice.
488 bool IsOverflow = NewTotal < Total;
489 assert(!(DidOverflow && IsOverflow) && "unexpected repeated overflow");
490 DidOverflow |= IsOverflow;
500 Weights.push_back(W);
502 if (Type == Weight::Backedge)
505 // Update forward total. Don't worry about overflow here, since then Total
506 // will exceed 32-bits and they'll both be recomputed in normalize().
507 ForwardTotal += Amount;
510 static void combineWeight(Weight &W, const Weight &OtherW) {
511 assert(OtherW.TargetNode.isValid());
516 assert(W.Type == OtherW.Type);
517 assert(W.TargetNode == OtherW.TargetNode);
518 assert(W.Amount < W.Amount + OtherW.Amount);
519 W.Amount += OtherW.Amount;
521 static void combineWeightsBySorting(WeightList &Weights) {
522 // Sort so edges to the same node are adjacent.
523 std::sort(Weights.begin(), Weights.end(),
525 const Weight &R) { return L.TargetNode < R.TargetNode; });
527 // Combine adjacent edges.
528 WeightList::iterator O = Weights.begin();
529 for (WeightList::const_iterator I = O, L = O, E = Weights.end(); I != E;
533 // Find the adjacent weights to the same node.
534 for (++L; L != E && I->TargetNode == L->TargetNode; ++L)
535 combineWeight(*O, *L);
538 // Erase extra entries.
539 Weights.erase(O, Weights.end());
542 static void combineWeightsByHashing(WeightList &Weights) {
543 // Collect weights into a DenseMap.
544 typedef DenseMap<BlockNode::IndexType, Weight> HashTable;
545 HashTable Combined(NextPowerOf2(2 * Weights.size()));
546 for (const Weight &W : Weights)
547 combineWeight(Combined[W.TargetNode.Index], W);
549 // Check whether anything changed.
550 if (Weights.size() == Combined.size())
553 // Fill in the new weights.
555 Weights.reserve(Combined.size());
556 for (const auto &I : Combined)
557 Weights.push_back(I.second);
559 static void combineWeights(WeightList &Weights) {
560 // Use a hash table for many successors to keep this linear.
561 if (Weights.size() > 128) {
562 combineWeightsByHashing(Weights);
566 combineWeightsBySorting(Weights);
568 static uint64_t shiftRightAndRound(uint64_t N, int Shift) {
573 return (N >> Shift) + (UINT64_C(1) & N >> (Shift - 1));
575 void Distribution::normalize() {
576 // Early exit for termination nodes.
580 // Only bother if there are multiple successors.
581 if (Weights.size() > 1)
582 combineWeights(Weights);
584 // Early exit when combined into a single successor.
585 if (Weights.size() == 1) {
587 ForwardTotal = Weights.front().Type != Weight::Backedge;
588 Weights.front().Amount = 1;
592 // Determine how much to shift right so that the total fits into 32-bits.
594 // If we shift at all, shift by 1 extra. Otherwise, the lower limit of 1
595 // for each weight can cause a 32-bit overflow.
599 else if (Total > UINT32_MAX)
600 Shift = 33 - countLeadingZeros(Total);
602 // Early exit if nothing needs to be scaled.
606 // Recompute the total through accumulation (rather than shifting it) so that
607 // it's accurate after shifting. ForwardTotal is dirty here anyway.
611 // Sum the weights to each node and shift right if necessary.
612 for (Weight &W : Weights) {
613 // Scale down below UINT32_MAX. Since Shift is larger than necessary, we
614 // can round here without concern about overflow.
615 assert(W.TargetNode.isValid());
616 W.Amount = std::max(UINT64_C(1), shiftRightAndRound(W.Amount, Shift));
617 assert(W.Amount <= UINT32_MAX);
621 if (W.Type == Weight::Backedge)
624 // Update the forward total.
625 ForwardTotal += W.Amount;
627 assert(Total <= UINT32_MAX);
630 void BlockFrequencyInfoImplBase::clear() {
631 *this = BlockFrequencyInfoImplBase();
634 /// \brief Clear all memory not needed downstream.
636 /// Releases all memory not used downstream. In particular, saves Freqs.
637 static void cleanup(BlockFrequencyInfoImplBase &BFI) {
638 std::vector<FrequencyData> SavedFreqs(std::move(BFI.Freqs));
640 BFI.Freqs = std::move(SavedFreqs);
643 /// \brief Get a possibly packaged node.
645 /// Get the node currently representing Node, which could be a containing
648 /// This function should only be called when distributing mass. As long as
649 /// there are no irreducilbe edges to Node, then it will have complexity O(1)
652 /// In general, the complexity is O(L), where L is the number of loop headers
653 /// Node has been packaged into. Since this method is called in the context
654 /// of distributing mass, L will be the number of loop headers an early exit
655 /// edge jumps out of.
656 static BlockNode getPackagedNode(const BlockFrequencyInfoImplBase &BFI,
657 const BlockNode &Node) {
658 assert(Node.isValid());
659 if (!BFI.Working[Node.Index].IsPackaged)
661 if (!BFI.Working[Node.Index].ContainingLoop.isValid())
663 return getPackagedNode(BFI, BFI.Working[Node.Index].ContainingLoop);
666 /// \brief Get the appropriate mass for a possible pseudo-node loop package.
668 /// Get appropriate mass for Node. If Node is a loop-header (whose loop has
669 /// been packaged), returns the mass of its pseudo-node. If it's a node inside
670 /// a packaged loop, it returns the loop's pseudo-node.
671 static BlockMass &getPackageMass(BlockFrequencyInfoImplBase &BFI,
672 const BlockNode &Node) {
673 assert(Node.isValid());
674 assert(!BFI.Working[Node.Index].IsPackaged);
675 if (!BFI.Working[Node.Index].IsAPackage)
676 return BFI.Working[Node.Index].Mass;
678 return BFI.getLoopPackage(Node).Mass;
681 void BlockFrequencyInfoImplBase::addToDist(Distribution &Dist,
682 const BlockNode &LoopHead,
683 const BlockNode &Pred,
684 const BlockNode &Succ,
690 auto debugSuccessor = [&](const char *Type, const BlockNode &Resolved) {
692 << " [" << Type << "] weight = " << Weight;
693 if (Succ != LoopHead)
694 dbgs() << ", succ = " << getBlockName(Succ);
695 if (Resolved != Succ)
696 dbgs() << ", resolved = " << getBlockName(Resolved);
699 (void)debugSuccessor;
702 if (Succ == LoopHead) {
703 DEBUG(debugSuccessor("backedge", Succ));
704 Dist.addBackedge(LoopHead, Weight);
707 BlockNode Resolved = getPackagedNode(*this, Succ);
708 assert(Resolved != LoopHead);
710 if (Working[Resolved.Index].ContainingLoop != LoopHead) {
711 DEBUG(debugSuccessor(" exit ", Resolved));
712 Dist.addExit(Resolved, Weight);
716 if (!LoopHead.isValid() && Resolved < Pred) {
717 // Irreducible backedge. Skip this edge in the distribution.
718 DEBUG(debugSuccessor("skipped ", Resolved));
722 DEBUG(debugSuccessor(" local ", Resolved));
723 Dist.addLocal(Resolved, Weight);
726 void BlockFrequencyInfoImplBase::addLoopSuccessorsToDist(
727 const BlockNode &LoopHead, const BlockNode &LocalLoopHead,
728 Distribution &Dist) {
729 PackagedLoopData &LoopPackage = getLoopPackage(LocalLoopHead);
730 const PackagedLoopData::ExitMap &Exits = LoopPackage.Exits;
732 // Copy the exit map into Dist.
733 for (const auto &I : Exits)
734 addToDist(Dist, LoopHead, LocalLoopHead, I.first, I.second.getMass());
736 // We don't need this map any more. Clear it to prevent quadratic memory
737 // usage in deeply nested loops with irreducible control flow.
738 LoopPackage.Exits.clear();
741 /// \brief Get the maximum allowed loop scale.
743 /// Gives the maximum number of estimated iterations allowed for a loop.
744 /// Downstream users have trouble with very large numbers (even within
745 /// 64-bits). Perhaps they can be changed to use PositiveFloat.
747 /// TODO: change downstream users so that this can be increased or removed.
748 static Float getMaxLoopScale() { return Float(1, 12); }
750 /// \brief Compute the loop scale for a loop.
751 void BlockFrequencyInfoImplBase::computeLoopScale(const BlockNode &LoopHead) {
752 // Compute loop scale.
753 DEBUG(dbgs() << "compute-loop-scale: " << getBlockName(LoopHead) << "\n");
755 // LoopScale == 1 / ExitMass
756 // ExitMass == HeadMass - BackedgeMass
757 PackagedLoopData &LoopPackage = getLoopPackage(LoopHead);
758 BlockMass ExitMass = BlockMass::getFull() - LoopPackage.BackedgeMass;
760 // Block scale stores the inverse of the scale.
761 LoopPackage.Scale = ExitMass.toFloat().inverse();
763 DEBUG(dbgs() << " - exit-mass = " << ExitMass << " (" << BlockMass::getFull()
764 << " - " << LoopPackage.BackedgeMass << ")\n"
765 << " - scale = " << LoopPackage.Scale << "\n");
767 if (LoopPackage.Scale > getMaxLoopScale()) {
768 LoopPackage.Scale = getMaxLoopScale();
769 DEBUG(dbgs() << " - reduced-to-max-scale: " << getMaxLoopScale() << "\n");
773 /// \brief Package up a loop.
774 void BlockFrequencyInfoImplBase::packageLoop(const BlockNode &LoopHead) {
775 DEBUG(dbgs() << "packaging-loop: " << getBlockName(LoopHead) << "\n");
776 Working[LoopHead.Index].IsAPackage = true;
777 for (const BlockNode &M : getLoopPackage(LoopHead).Members) {
778 DEBUG(dbgs() << " - node: " << getBlockName(M.Index) << "\n");
779 Working[M.Index].IsPackaged = true;
783 void BlockFrequencyInfoImplBase::distributeMass(const BlockNode &Source,
784 const BlockNode &LoopHead,
785 Distribution &Dist) {
786 BlockMass Mass = getPackageMass(*this, Source);
787 DEBUG(dbgs() << " => mass: " << Mass
788 << " ( general | forward )\n");
790 // Distribute mass to successors as laid out in Dist.
791 DitheringDistributer D(Dist, Mass);
794 auto debugAssign = [&](const BlockNode &T, const BlockMass &M,
796 dbgs() << " => assign " << M << " (" << D.RemMass << "|"
797 << D.RemForwardMass << ")";
799 dbgs() << " [" << Desc << "]";
801 dbgs() << " to " << getBlockName(T);
807 PackagedLoopData *LoopPackage = 0;
808 if (LoopHead.isValid())
809 LoopPackage = &getLoopPackage(LoopHead);
810 for (const Weight &W : Dist.Weights) {
811 // Check for a local edge (forward and non-exit).
812 if (W.Type == Weight::Local) {
813 BlockMass Local = D.takeLocalMass(W.Amount);
814 getPackageMass(*this, W.TargetNode) += Local;
815 DEBUG(debugAssign(W.TargetNode, Local, nullptr));
819 // Backedges and exits only make sense if we're processing a loop.
820 assert(LoopPackage && "backedge or exit outside of loop");
822 // Check for a backedge.
823 if (W.Type == Weight::Backedge) {
824 BlockMass Back = D.takeBackedgeMass(W.Amount);
825 LoopPackage->BackedgeMass += Back;
826 DEBUG(debugAssign(BlockNode(), Back, "back"));
830 // This must be an exit.
831 assert(W.Type == Weight::Exit);
832 BlockMass Exit = D.takeExitMass(W.Amount);
833 LoopPackage->Exits.push_back(std::make_pair(W.TargetNode, Exit));
834 DEBUG(debugAssign(W.TargetNode, Exit, "exit"));
838 static void convertFloatingToInteger(BlockFrequencyInfoImplBase &BFI,
839 const Float &Min, const Float &Max) {
840 // Scale the Factor to a size that creates integers. Ideally, integers would
841 // be scaled so that Max == UINT64_MAX so that they can be best
842 // differentiated. However, the register allocator currently deals poorly
843 // with large numbers. Instead, push Min up a little from 1 to give some
844 // room to differentiate small, unequal numbers.
846 // TODO: fix issues downstream so that ScalingFactor can be Float(1,64)/Max.
847 Float ScalingFactor = Min.inverse();
848 if ((Max / Min).lg() < 60)
851 // Translate the floats to integers.
852 DEBUG(dbgs() << "float-to-int: min = " << Min << ", max = " << Max
853 << ", factor = " << ScalingFactor << "\n");
854 for (size_t Index = 0; Index < BFI.Freqs.size(); ++Index) {
855 Float Scaled = BFI.Freqs[Index].Floating * ScalingFactor;
856 BFI.Freqs[Index].Integer = std::max(UINT64_C(1), Scaled.toInt<uint64_t>());
857 DEBUG(dbgs() << " - " << BFI.getBlockName(Index) << ": float = "
858 << BFI.Freqs[Index].Floating << ", scaled = " << Scaled
859 << ", int = " << BFI.Freqs[Index].Integer << "\n");
863 static void scaleBlockData(BlockFrequencyInfoImplBase &BFI,
864 const BlockNode &Node,
865 const PackagedLoopData &Loop) {
866 Float F = Loop.Mass.toFloat() * Loop.Scale;
868 Float &Current = BFI.Freqs[Node.Index].Floating;
869 Float Updated = Current * F;
871 DEBUG(dbgs() << " - " << BFI.getBlockName(Node) << ": " << Current << " => "
877 /// \brief Unwrap a loop package.
879 /// Visits all the members of a loop, adjusting their BlockData according to
880 /// the loop's pseudo-node.
881 static void unwrapLoopPackage(BlockFrequencyInfoImplBase &BFI,
882 const BlockNode &Head) {
883 assert(Head.isValid());
885 PackagedLoopData &LoopPackage = BFI.getLoopPackage(Head);
886 DEBUG(dbgs() << "unwrap-loop-package: " << BFI.getBlockName(Head)
887 << ": mass = " << LoopPackage.Mass
888 << ", scale = " << LoopPackage.Scale << "\n");
889 scaleBlockData(BFI, Head, LoopPackage);
891 // Propagate the head scale through the loop. Since members are visited in
892 // RPO, the head scale will be updated by the loop scale first, and then the
893 // final head scale will be used for updated the rest of the members.
894 for (const BlockNode &M : LoopPackage.Members) {
895 const FrequencyData &HeadData = BFI.Freqs[Head.Index];
896 FrequencyData &Freqs = BFI.Freqs[M.Index];
897 Float NewFreq = Freqs.Floating * HeadData.Floating;
898 DEBUG(dbgs() << " - " << BFI.getBlockName(M) << ": " << Freqs.Floating
899 << " => " << NewFreq << "\n");
900 Freqs.Floating = NewFreq;
904 void BlockFrequencyInfoImplBase::finalizeMetrics() {
905 // Set initial frequencies from loop-local masses.
906 for (size_t Index = 0; Index < Working.size(); ++Index)
907 Freqs[Index].Floating = Working[Index].Mass.toFloat();
909 // Unwrap loop packages in reverse post-order, tracking min and max
911 auto Min = Float::getLargest();
912 auto Max = Float::getZero();
913 for (size_t Index = 0; Index < Working.size(); ++Index) {
914 if (Working[Index].isLoopHeader())
915 unwrapLoopPackage(*this, BlockNode(Index));
918 Min = std::min(Min, Freqs[Index].Floating);
919 Max = std::max(Max, Freqs[Index].Floating);
922 // Convert to integers.
923 convertFloatingToInteger(*this, Min, Max);
925 // Clean up data structures.
928 // Print out the final stats.
933 BlockFrequencyInfoImplBase::getBlockFreq(const BlockNode &Node) const {
936 return Freqs[Node.Index].Integer;
939 BlockFrequencyInfoImplBase::getFloatingBlockFreq(const BlockNode &Node) const {
941 return Float::getZero();
942 return Freqs[Node.Index].Floating;
946 BlockFrequencyInfoImplBase::getBlockName(const BlockNode &Node) const {
947 return std::string();
951 BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
952 const BlockNode &Node) const {
953 return OS << getFloatingBlockFreq(Node);
957 BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
958 const BlockFrequency &Freq) const {
959 Float Block(Freq.getFrequency(), 0);
960 Float Entry(getEntryFreq(), 0);
962 return OS << Block / Entry;