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 //===----------------------------------------------------------------------===//
28 const int PositiveFloatBase::MaxExponent;
29 const int PositiveFloatBase::MinExponent;
32 static void appendDigit(std::string &Str, unsigned D) {
37 static void appendNumber(std::string &Str, uint64_t N) {
39 appendDigit(Str, N % 10);
44 static bool doesRoundUp(char Digit) {
57 static std::string toStringAPFloat(uint64_t D, int E, unsigned Precision) {
58 assert(E >= PositiveFloatBase::MinExponent);
59 assert(E <= PositiveFloatBase::MaxExponent);
61 // Find a new E, but don't let it increase past MaxExponent.
62 int LeadingZeros = PositiveFloatBase::countLeadingZeros64(D);
63 int NewE = std::min(PositiveFloatBase::MaxExponent, E + 63 - LeadingZeros);
64 int Shift = 63 - (NewE - E);
65 assert(Shift <= LeadingZeros);
66 assert(Shift == LeadingZeros || NewE == PositiveFloatBase::MaxExponent);
70 // Check for a denormal.
71 unsigned AdjustedE = E + 16383;
73 assert(E == PositiveFloatBase::MaxExponent);
77 // Build the float and print it.
78 uint64_t RawBits[2] = {D, AdjustedE};
79 APFloat Float(APFloat::x87DoubleExtended, APInt(80, RawBits));
80 SmallVector<char, 24> Chars;
81 Float.toString(Chars, Precision, 0);
82 return std::string(Chars.begin(), Chars.end());
85 static std::string stripTrailingZeros(std::string Float) {
86 size_t NonZero = Float.find_last_not_of('0');
87 assert(NonZero != std::string::npos && "no . in floating point string");
89 if (Float[NonZero] == '.')
92 return Float.substr(0, NonZero + 1);
95 std::string PositiveFloatBase::toString(uint64_t D, int16_t E, int Width,
100 // Canonicalize exponent and digits.
108 if (int Shift = std::min(int16_t(countLeadingZeros64(D)), E)) {
115 } else if (E > -64) {
117 Below0 = D << (64 + E);
118 } else if (E > -120) {
119 Below0 = D >> (-E - 64);
120 Extra = D << (128 + E);
121 ExtraShift = -64 - E;
124 // Fall back on APFloat for very small and very large numbers.
125 if (!Above0 && !Below0)
126 return toStringAPFloat(D, E, Precision);
128 // Append the digits before the decimal.
130 size_t DigitsOut = 0;
132 appendNumber(Str, Above0);
133 DigitsOut = Str.size();
136 std::reverse(Str.begin(), Str.end());
138 // Return early if there's nothing after the decimal.
142 // Append the decimal and beyond.
144 uint64_t Error = UINT64_C(1) << (64 - Width);
146 // We need to shift Below0 to the right to make space for calculating
147 // digits. Save the precision we're losing in Extra.
148 Extra = (Below0 & 0xf) << 56 | (Extra >> 8);
151 size_t AfterDot = Str.size();
161 Below0 += (Extra >> 60);
162 Extra = Extra & (UINT64_MAX >> 4);
163 appendDigit(Str, Below0 >> 60);
164 Below0 = Below0 & (UINT64_MAX >> 4);
165 if (DigitsOut || Str.back() != '0')
168 } while (Error && (Below0 << 4 | Extra >> 60) >= Error / 2 &&
169 (!Precision || DigitsOut <= Precision || SinceDot < 2));
171 // Return early for maximum precision.
172 if (!Precision || DigitsOut <= Precision)
173 return stripTrailingZeros(Str);
175 // Find where to truncate.
177 std::max(Str.size() - (DigitsOut - Precision), AfterDot + 1);
179 // Check if there's anything to truncate.
180 if (Truncate >= Str.size())
181 return stripTrailingZeros(Str);
183 bool Carry = doesRoundUp(Str[Truncate]);
185 return stripTrailingZeros(Str.substr(0, Truncate));
187 // Round with the first truncated digit.
188 for (std::string::reverse_iterator I(Str.begin() + Truncate), E = Str.rend();
202 // Add "1" in front if we still need to carry.
203 return stripTrailingZeros(std::string(Carry, '1') + Str.substr(0, Truncate));
206 raw_ostream &PositiveFloatBase::print(raw_ostream &OS, uint64_t D, int16_t E,
207 int Width, unsigned Precision) {
208 return OS << toString(D, E, Width, Precision);
211 void PositiveFloatBase::dump(uint64_t D, int16_t E, int Width) {
212 print(dbgs(), D, E, Width, 0) << "[" << Width << ":" << D << "*2^" << E
216 static std::pair<uint64_t, int16_t>
217 getRoundedFloat(uint64_t N, bool ShouldRound, int64_t Shift) {
220 // Rounding caused an overflow.
221 return std::make_pair(UINT64_C(1), Shift + 64);
222 return std::make_pair(N, Shift);
225 std::pair<uint64_t, int16_t> PositiveFloatBase::divide64(uint64_t Dividend,
227 // Input should be sanitized.
231 // Minimize size of divisor.
233 if (int Zeros = countTrailingZeros(Divisor)) {
238 // Check for powers of two.
240 return std::make_pair(Dividend, Shift);
242 // Maximize size of dividend.
243 if (int Zeros = countLeadingZeros64(Dividend)) {
248 // Start with the result of a divide.
249 uint64_t Quotient = Dividend / Divisor;
252 // Continue building the quotient with long division.
254 // TODO: continue with largers digits.
255 while (!(Quotient >> 63) && Dividend) {
256 // Shift Dividend, and check for overflow.
257 bool IsOverflow = Dividend >> 63;
262 bool DoesDivide = IsOverflow || Divisor <= Dividend;
263 Quotient = (Quotient << 1) | uint64_t(DoesDivide);
264 Dividend -= DoesDivide ? Divisor : 0;
268 if (Dividend >= getHalf(Divisor))
270 // Rounding caused an overflow in Quotient.
271 return std::make_pair(UINT64_C(1), Shift + 64);
273 return getRoundedFloat(Quotient, Dividend >= getHalf(Divisor), Shift);
276 static void addWithCarry(uint64_t &Upper, uint64_t &Lower, uint64_t N) {
277 uint64_t NewLower = Lower + (N << 32);
278 Upper += (N >> 32) + (NewLower < Lower);
282 std::pair<uint64_t, int16_t> PositiveFloatBase::multiply64(uint64_t L,
284 // Separate into two 32-bit digits (U.L).
285 uint64_t UL = L >> 32, LL = L & UINT32_MAX, UR = R >> 32, LR = R & UINT32_MAX;
287 // Compute cross products.
288 uint64_t P1 = UL * UR, P2 = UL * LR, P3 = LL * UR, P4 = LL * LR;
290 // Sum into two 64-bit digits.
291 uint64_t Upper = P1, Lower = P4;
292 addWithCarry(Upper, Lower, P2);
293 addWithCarry(Upper, Lower, P3);
295 // Check for the lower 32 bits.
297 return std::make_pair(Lower, 0);
299 // Shift as little as possible to maximize precision.
300 unsigned LeadingZeros = countLeadingZeros64(Upper);
301 int16_t Shift = 64 - LeadingZeros;
303 Upper = Upper << LeadingZeros | Lower >> Shift;
304 bool ShouldRound = Shift && (Lower & UINT64_C(1) << (Shift - 1));
305 return getRoundedFloat(Upper, ShouldRound, Shift);
308 //===----------------------------------------------------------------------===//
310 // BlockMass implementation.
312 //===----------------------------------------------------------------------===//
313 BlockMass &BlockMass::operator*=(const BranchProbability &P) {
314 uint32_t N = P.getNumerator(), D = P.getDenominator();
315 assert(D || "divide by 0");
316 assert(N <= D || "fraction greater than 1");
318 // Fast path for multiplying by 1.0.
322 // Get as much precision as we can.
323 int Shift = countLeadingZeros(Mass);
324 uint64_t ShiftedQuotient = (Mass << Shift) / D;
325 uint64_t Product = ShiftedQuotient * N >> Shift;
327 // Now check for what's lost.
328 uint64_t Left = ShiftedQuotient * (D - N) >> Shift;
329 uint64_t Lost = Mass - Product - Left;
331 // TODO: prove this assertion.
332 assert(Lost <= UINT32_MAX);
334 // Take the product plus a portion of the spoils.
335 Mass = Product + Lost * N / D;
339 PositiveFloat<uint64_t> BlockMass::toFloat() const {
341 return PositiveFloat<uint64_t>(1, 0);
342 return PositiveFloat<uint64_t>(getMass() + 1, -64);
345 void BlockMass::dump() const { print(dbgs()); }
347 static char getHexDigit(int N) {
353 raw_ostream &BlockMass::print(raw_ostream &OS) const {
354 for (int Digits = 0; Digits < 16; ++Digits)
355 OS << getHexDigit(Mass >> (60 - Digits * 4) & 0xf);
359 //===----------------------------------------------------------------------===//
361 // BlockFrequencyInfoImpl implementation.
363 //===----------------------------------------------------------------------===//
366 typedef BlockFrequencyInfoImplBase::BlockNode BlockNode;
367 typedef BlockFrequencyInfoImplBase::Distribution Distribution;
368 typedef BlockFrequencyInfoImplBase::Distribution::WeightList WeightList;
369 typedef BlockFrequencyInfoImplBase::Float Float;
370 typedef BlockFrequencyInfoImplBase::PackagedLoopData PackagedLoopData;
371 typedef BlockFrequencyInfoImplBase::Weight Weight;
372 typedef BlockFrequencyInfoImplBase::FrequencyData FrequencyData;
374 /// \brief Dithering mass distributer.
376 /// This class splits up a single mass into portions by weight, dithering to
377 /// spread out error. No mass is lost. The dithering precision depends on the
378 /// precision of the product of \a BlockMass and \a BranchProbability.
380 /// The distribution algorithm follows.
382 /// 1. Initialize by saving the sum of the weights in \a RemWeight and the
383 /// mass to distribute in \a RemMass.
385 /// 2. For each portion:
387 /// 1. Construct a branch probability, P, as the portion's weight divided
388 /// by the current value of \a RemWeight.
389 /// 2. Calculate the portion's mass as \a RemMass times P.
390 /// 3. Update \a RemWeight and \a RemMass at each portion by subtracting
391 /// the current portion's weight and mass.
393 /// Mass is distributed in two ways: full distribution and forward
394 /// distribution. The latter ignores backedges, and uses the parallel fields
395 /// \a RemForwardWeight and \a RemForwardMass.
396 struct DitheringDistributer {
398 uint32_t RemForwardWeight;
401 BlockMass RemForwardMass;
403 DitheringDistributer(Distribution &Dist, const BlockMass &Mass);
405 BlockMass takeLocalMass(uint32_t Weight) {
406 (void)takeMass(Weight);
407 return takeForwardMass(Weight);
409 BlockMass takeExitMass(uint32_t Weight) {
410 (void)takeForwardMass(Weight);
411 return takeMass(Weight);
413 BlockMass takeBackedgeMass(uint32_t Weight) { return takeMass(Weight); }
416 BlockMass takeForwardMass(uint32_t Weight);
417 BlockMass takeMass(uint32_t Weight);
421 DitheringDistributer::DitheringDistributer(Distribution &Dist,
422 const BlockMass &Mass) {
424 RemWeight = Dist.Total;
425 RemForwardWeight = Dist.ForwardTotal;
427 RemForwardMass = Dist.ForwardTotal ? Mass : BlockMass();
430 BlockMass DitheringDistributer::takeForwardMass(uint32_t Weight) {
431 // Compute the amount of mass to take.
432 assert(Weight && "invalid weight");
433 assert(Weight <= RemForwardWeight);
434 BlockMass Mass = RemForwardMass * BranchProbability(Weight, RemForwardWeight);
436 // Decrement totals (dither).
437 RemForwardWeight -= Weight;
438 RemForwardMass -= Mass;
441 BlockMass DitheringDistributer::takeMass(uint32_t Weight) {
442 assert(Weight && "invalid weight");
443 assert(Weight <= RemWeight);
444 BlockMass Mass = RemMass * BranchProbability(Weight, RemWeight);
446 // Decrement totals (dither).
452 void Distribution::add(const BlockNode &Node, uint64_t Amount,
453 Weight::DistType Type) {
454 assert(Amount && "invalid weight of 0");
455 uint64_t NewTotal = Total + Amount;
457 // Check for overflow. It should be impossible to overflow twice.
458 bool IsOverflow = NewTotal < Total;
459 assert(!(DidOverflow && IsOverflow) && "unexpected repeated overflow");
460 DidOverflow |= IsOverflow;
470 Weights.push_back(W);
472 if (Type == Weight::Backedge)
475 // Update forward total. Don't worry about overflow here, since then Total
476 // will exceed 32-bits and they'll both be recomputed in normalize().
477 ForwardTotal += Amount;
480 static void combineWeight(Weight &W, const Weight &OtherW) {
481 assert(OtherW.TargetNode.isValid());
486 assert(W.Type == OtherW.Type);
487 assert(W.TargetNode == OtherW.TargetNode);
488 assert(W.Amount < W.Amount + OtherW.Amount);
489 W.Amount += OtherW.Amount;
491 static void combineWeightsBySorting(WeightList &Weights) {
492 // Sort so edges to the same node are adjacent.
493 std::sort(Weights.begin(), Weights.end(),
495 const Weight &R) { return L.TargetNode < R.TargetNode; });
497 // Combine adjacent edges.
498 WeightList::iterator O = Weights.begin();
499 for (WeightList::const_iterator I = O, L = O, E = Weights.end(); I != E;
503 // Find the adjacent weights to the same node.
504 for (++L; L != E && I->TargetNode == L->TargetNode; ++L)
505 combineWeight(*O, *L);
508 // Erase extra entries.
509 Weights.erase(O, Weights.end());
512 static void combineWeightsByHashing(WeightList &Weights) {
513 // Collect weights into a DenseMap.
514 typedef DenseMap<BlockNode::IndexType, Weight> HashTable;
515 HashTable Combined(NextPowerOf2(2 * Weights.size()));
516 for (const Weight &W : Weights)
517 combineWeight(Combined[W.TargetNode.Index], W);
519 // Check whether anything changed.
520 if (Weights.size() == Combined.size())
523 // Fill in the new weights.
525 Weights.reserve(Combined.size());
526 for (const auto &I : Combined)
527 Weights.push_back(I.second);
529 static void combineWeights(WeightList &Weights) {
530 // Use a hash table for many successors to keep this linear.
531 if (Weights.size() > 128) {
532 combineWeightsByHashing(Weights);
536 combineWeightsBySorting(Weights);
538 static uint64_t shiftRightAndRound(uint64_t N, int Shift) {
543 return (N >> Shift) + (UINT64_C(1) & N >> (Shift - 1));
545 void Distribution::normalize() {
546 // Early exit for termination nodes.
550 // Only bother if there are multiple successors.
551 if (Weights.size() > 1)
552 combineWeights(Weights);
554 // Early exit when combined into a single successor.
555 if (Weights.size() == 1) {
557 ForwardTotal = Weights.front().Type != Weight::Backedge;
558 Weights.front().Amount = 1;
562 // Determine how much to shift right so that the total fits into 32-bits.
564 // If we shift at all, shift by 1 extra. Otherwise, the lower limit of 1
565 // for each weight can cause a 32-bit overflow.
569 else if (Total > UINT32_MAX)
570 Shift = 33 - countLeadingZeros(Total);
572 // Early exit if nothing needs to be scaled.
576 // Recompute the total through accumulation (rather than shifting it) so that
577 // it's accurate after shifting. ForwardTotal is dirty here anyway.
581 // Sum the weights to each node and shift right if necessary.
582 for (Weight &W : Weights) {
583 // Scale down below UINT32_MAX. Since Shift is larger than necessary, we
584 // can round here without concern about overflow.
585 assert(W.TargetNode.isValid());
586 W.Amount = std::max(UINT64_C(1), shiftRightAndRound(W.Amount, Shift));
587 assert(W.Amount <= UINT32_MAX);
591 if (W.Type == Weight::Backedge)
594 // Update the forward total.
595 ForwardTotal += W.Amount;
597 assert(Total <= UINT32_MAX);
600 void BlockFrequencyInfoImplBase::clear() {
601 *this = BlockFrequencyInfoImplBase();
604 /// \brief Clear all memory not needed downstream.
606 /// Releases all memory not used downstream. In particular, saves Freqs.
607 static void cleanup(BlockFrequencyInfoImplBase &BFI) {
608 std::vector<FrequencyData> SavedFreqs(std::move(BFI.Freqs));
610 BFI.Freqs = std::move(SavedFreqs);
613 /// \brief Get a possibly packaged node.
615 /// Get the node currently representing Node, which could be a containing
618 /// This function should only be called when distributing mass. As long as
619 /// there are no irreducilbe edges to Node, then it will have complexity O(1)
622 /// In general, the complexity is O(L), where L is the number of loop headers
623 /// Node has been packaged into. Since this method is called in the context
624 /// of distributing mass, L will be the number of loop headers an early exit
625 /// edge jumps out of.
626 static BlockNode getPackagedNode(const BlockFrequencyInfoImplBase &BFI,
627 const BlockNode &Node) {
628 assert(Node.isValid());
629 if (!BFI.Working[Node.Index].IsPackaged)
631 if (!BFI.Working[Node.Index].ContainingLoop.isValid())
633 return getPackagedNode(BFI, BFI.Working[Node.Index].ContainingLoop);
636 /// \brief Get the appropriate mass for a possible pseudo-node loop package.
638 /// Get appropriate mass for Node. If Node is a loop-header (whose loop has
639 /// been packaged), returns the mass of its pseudo-node. If it's a node inside
640 /// a packaged loop, it returns the loop's pseudo-node.
641 static BlockMass &getPackageMass(BlockFrequencyInfoImplBase &BFI,
642 const BlockNode &Node) {
643 assert(Node.isValid());
644 assert(!BFI.Working[Node.Index].IsPackaged);
645 if (!BFI.Working[Node.Index].IsAPackage)
646 return BFI.Working[Node.Index].Mass;
648 return BFI.getLoopPackage(Node).Mass;
651 void BlockFrequencyInfoImplBase::addToDist(Distribution &Dist,
652 const BlockNode &LoopHead,
653 const BlockNode &Pred,
654 const BlockNode &Succ,
660 auto debugSuccessor = [&](const char *Type, const BlockNode &Resolved) {
662 << " [" << Type << "] weight = " << Weight;
663 if (Succ != LoopHead)
664 dbgs() << ", succ = " << getBlockName(Succ);
665 if (Resolved != Succ)
666 dbgs() << ", resolved = " << getBlockName(Resolved);
669 (void)debugSuccessor;
672 if (Succ == LoopHead) {
673 DEBUG(debugSuccessor("backedge", Succ));
674 Dist.addBackedge(LoopHead, Weight);
677 BlockNode Resolved = getPackagedNode(*this, Succ);
678 assert(Resolved != LoopHead);
680 if (Working[Resolved.Index].ContainingLoop != LoopHead) {
681 DEBUG(debugSuccessor(" exit ", Resolved));
682 Dist.addExit(Resolved, Weight);
686 if (!LoopHead.isValid() && Resolved < Pred) {
687 // Irreducible backedge. Skip this edge in the distribution.
688 DEBUG(debugSuccessor("skipped ", Resolved));
692 DEBUG(debugSuccessor(" local ", Resolved));
693 Dist.addLocal(Resolved, Weight);
696 void BlockFrequencyInfoImplBase::addLoopSuccessorsToDist(
697 const BlockNode &LoopHead, const BlockNode &LocalLoopHead,
698 Distribution &Dist) {
699 PackagedLoopData &LoopPackage = getLoopPackage(LocalLoopHead);
700 const PackagedLoopData::ExitMap &Exits = LoopPackage.Exits;
702 // Copy the exit map into Dist.
703 for (const auto &I : Exits)
704 addToDist(Dist, LoopHead, LocalLoopHead, I.first, I.second.getMass());
706 // We don't need this map any more. Clear it to prevent quadratic memory
707 // usage in deeply nested loops with irreducible control flow.
708 LoopPackage.Exits.clear();
711 /// \brief Get the maximum allowed loop scale.
713 /// Gives the maximum number of estimated iterations allowed for a loop.
714 /// Downstream users have trouble with very large numbers (even within
715 /// 64-bits). Perhaps they can be changed to use PositiveFloat.
717 /// TODO: change downstream users so that this can be increased or removed.
718 static Float getMaxLoopScale() { return Float(1, 12); }
720 /// \brief Compute the loop scale for a loop.
721 void BlockFrequencyInfoImplBase::computeLoopScale(const BlockNode &LoopHead) {
722 // Compute loop scale.
723 DEBUG(dbgs() << "compute-loop-scale: " << getBlockName(LoopHead) << "\n");
725 // LoopScale == 1 / ExitMass
726 // ExitMass == HeadMass - BackedgeMass
727 PackagedLoopData &LoopPackage = getLoopPackage(LoopHead);
728 BlockMass ExitMass = BlockMass::getFull() - LoopPackage.BackedgeMass;
730 // Block scale stores the inverse of the scale.
731 LoopPackage.Scale = ExitMass.toFloat().inverse();
733 DEBUG(dbgs() << " - exit-mass = " << ExitMass << " (" << BlockMass::getFull()
734 << " - " << LoopPackage.BackedgeMass << ")\n"
735 << " - scale = " << LoopPackage.Scale << "\n");
737 if (LoopPackage.Scale > getMaxLoopScale()) {
738 LoopPackage.Scale = getMaxLoopScale();
739 DEBUG(dbgs() << " - reduced-to-max-scale: " << getMaxLoopScale() << "\n");
743 /// \brief Package up a loop.
744 void BlockFrequencyInfoImplBase::packageLoop(const BlockNode &LoopHead) {
745 DEBUG(dbgs() << "packaging-loop: " << getBlockName(LoopHead) << "\n");
746 Working[LoopHead.Index].IsAPackage = true;
747 for (const BlockNode &M : getLoopPackage(LoopHead).Members) {
748 DEBUG(dbgs() << " - node: " << getBlockName(M.Index) << "\n");
749 Working[M.Index].IsPackaged = true;
753 void BlockFrequencyInfoImplBase::distributeMass(const BlockNode &Source,
754 const BlockNode &LoopHead,
755 Distribution &Dist) {
756 BlockMass Mass = getPackageMass(*this, Source);
757 DEBUG(dbgs() << " => mass: " << Mass
758 << " ( general | forward )\n");
760 // Distribute mass to successors as laid out in Dist.
761 DitheringDistributer D(Dist, Mass);
764 auto debugAssign = [&](const BlockNode &T, const BlockMass &M,
766 dbgs() << " => assign " << M << " (" << D.RemMass << "|"
767 << D.RemForwardMass << ")";
769 dbgs() << " [" << Desc << "]";
771 dbgs() << " to " << getBlockName(T);
777 PackagedLoopData *LoopPackage = 0;
778 if (LoopHead.isValid())
779 LoopPackage = &getLoopPackage(LoopHead);
780 for (const Weight &W : Dist.Weights) {
781 // Check for a local edge (forward and non-exit).
782 if (W.Type == Weight::Local) {
783 BlockMass Local = D.takeLocalMass(W.Amount);
784 getPackageMass(*this, W.TargetNode) += Local;
785 DEBUG(debugAssign(W.TargetNode, Local, nullptr));
789 // Backedges and exits only make sense if we're processing a loop.
790 assert(LoopPackage && "backedge or exit outside of loop");
792 // Check for a backedge.
793 if (W.Type == Weight::Backedge) {
794 BlockMass Back = D.takeBackedgeMass(W.Amount);
795 LoopPackage->BackedgeMass += Back;
796 DEBUG(debugAssign(BlockNode(), Back, "back"));
800 // This must be an exit.
801 assert(W.Type == Weight::Exit);
802 BlockMass Exit = D.takeExitMass(W.Amount);
803 LoopPackage->Exits.push_back(std::make_pair(W.TargetNode, Exit));
804 DEBUG(debugAssign(W.TargetNode, Exit, "exit"));
808 static void convertFloatingToInteger(BlockFrequencyInfoImplBase &BFI,
809 const Float &Min, const Float &Max) {
810 // Scale the Factor to a size that creates integers. Ideally, integers would
811 // be scaled so that Max == UINT64_MAX so that they can be best
812 // differentiated. However, the register allocator currently deals poorly
813 // with large numbers. Instead, push Min up a little from 1 to give some
814 // room to differentiate small, unequal numbers.
816 // TODO: fix issues downstream so that ScalingFactor can be Float(1,64)/Max.
817 Float ScalingFactor = Min.inverse();
818 if ((Max / Min).lg() < 60)
821 // Translate the floats to integers.
822 DEBUG(dbgs() << "float-to-int: min = " << Min << ", max = " << Max
823 << ", factor = " << ScalingFactor << "\n");
824 for (size_t Index = 0; Index < BFI.Freqs.size(); ++Index) {
825 Float Scaled = BFI.Freqs[Index].Floating * ScalingFactor;
826 BFI.Freqs[Index].Integer = std::max(UINT64_C(1), Scaled.toInt<uint64_t>());
827 DEBUG(dbgs() << " - " << BFI.getBlockName(Index) << ": float = "
828 << BFI.Freqs[Index].Floating << ", scaled = " << Scaled
829 << ", int = " << BFI.Freqs[Index].Integer << "\n");
833 static void scaleBlockData(BlockFrequencyInfoImplBase &BFI,
834 const BlockNode &Node,
835 const PackagedLoopData &Loop) {
836 Float F = Loop.Mass.toFloat() * Loop.Scale;
838 Float &Current = BFI.Freqs[Node.Index].Floating;
839 Float Updated = Current * F;
841 DEBUG(dbgs() << " - " << BFI.getBlockName(Node) << ": " << Current << " => "
847 /// \brief Unwrap a loop package.
849 /// Visits all the members of a loop, adjusting their BlockData according to
850 /// the loop's pseudo-node.
851 static void unwrapLoopPackage(BlockFrequencyInfoImplBase &BFI,
852 const BlockNode &Head) {
853 assert(Head.isValid());
855 PackagedLoopData &LoopPackage = BFI.getLoopPackage(Head);
856 DEBUG(dbgs() << "unwrap-loop-package: " << BFI.getBlockName(Head)
857 << ": mass = " << LoopPackage.Mass
858 << ", scale = " << LoopPackage.Scale << "\n");
859 scaleBlockData(BFI, Head, LoopPackage);
861 // Propagate the head scale through the loop. Since members are visited in
862 // RPO, the head scale will be updated by the loop scale first, and then the
863 // final head scale will be used for updated the rest of the members.
864 for (const BlockNode &M : LoopPackage.Members) {
865 const FrequencyData &HeadData = BFI.Freqs[Head.Index];
866 FrequencyData &Freqs = BFI.Freqs[M.Index];
867 Float NewFreq = Freqs.Floating * HeadData.Floating;
868 DEBUG(dbgs() << " - " << BFI.getBlockName(M) << ": " << Freqs.Floating
869 << " => " << NewFreq << "\n");
870 Freqs.Floating = NewFreq;
874 void BlockFrequencyInfoImplBase::finalizeMetrics() {
875 // Set initial frequencies from loop-local masses.
876 for (size_t Index = 0; Index < Working.size(); ++Index)
877 Freqs[Index].Floating = Working[Index].Mass.toFloat();
879 // Unwrap loop packages in reverse post-order, tracking min and max
881 auto Min = Float::getLargest();
882 auto Max = Float::getZero();
883 for (size_t Index = 0; Index < Working.size(); ++Index) {
884 if (Working[Index].isLoopHeader())
885 unwrapLoopPackage(*this, BlockNode(Index));
888 Min = std::min(Min, Freqs[Index].Floating);
889 Max = std::max(Max, Freqs[Index].Floating);
892 // Convert to integers.
893 convertFloatingToInteger(*this, Min, Max);
895 // Clean up data structures.
898 // Print out the final stats.
903 BlockFrequencyInfoImplBase::getBlockFreq(const BlockNode &Node) const {
906 return Freqs[Node.Index].Integer;
909 BlockFrequencyInfoImplBase::getFloatingBlockFreq(const BlockNode &Node) const {
911 return Float::getZero();
912 return Freqs[Node.Index].Floating;
916 BlockFrequencyInfoImplBase::getBlockName(const BlockNode &Node) const {
917 return std::string();
921 BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
922 const BlockNode &Node) const {
923 return OS << getFloatingBlockFreq(Node);
927 BlockFrequencyInfoImplBase::printBlockFreq(raw_ostream &OS,
928 const BlockFrequency &Freq) const {
929 Float Block(Freq.getFrequency(), 0);
930 Float Entry(getEntryFreq(), 0);
932 return OS << Block / Entry;