1 //==- BlockFrequencyInfoImpl.h - Block Frequency Implementation -*- C++ -*-===//
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 // Shared implementation of BlockFrequency for IR and Machine Instructions.
11 // See the documentation below for BlockFrequencyInfoImpl for details.
13 //===----------------------------------------------------------------------===//
15 #ifndef LLVM_ANALYSIS_BLOCKFREQUENCYINFOIMPL_H
16 #define LLVM_ANALYSIS_BLOCKFREQUENCYINFOIMPL_H
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/ADT/PostOrderIterator.h"
20 #include "llvm/ADT/iterator_range.h"
21 #include "llvm/IR/BasicBlock.h"
22 #include "llvm/Support/BlockFrequency.h"
23 #include "llvm/Support/BranchProbability.h"
24 #include "llvm/Support/Debug.h"
25 #include "llvm/Support/ScaledNumber.h"
26 #include "llvm/Support/raw_ostream.h"
32 #define DEBUG_TYPE "block-freq"
34 //===----------------------------------------------------------------------===//
36 // ScaledNumber definition.
38 // TODO: Move to include/llvm/Support/ScaledNumber.h
40 //===----------------------------------------------------------------------===//
43 class ScaledNumberBase {
45 static const int32_t MaxExponent = 16383;
46 static const int32_t MinExponent = -16382;
47 static const int DefaultPrecision = 10;
49 static void dump(uint64_t D, int16_t E, int Width);
50 static raw_ostream &print(raw_ostream &OS, uint64_t D, int16_t E, int Width,
52 static std::string toString(uint64_t D, int16_t E, int Width,
54 static int countLeadingZeros32(uint32_t N) { return countLeadingZeros(N); }
55 static int countLeadingZeros64(uint64_t N) { return countLeadingZeros(N); }
56 static uint64_t getHalf(uint64_t N) { return (N >> 1) + (N & 1); }
58 static std::pair<uint64_t, bool> splitSigned(int64_t N) {
60 return std::make_pair(N, false);
61 uint64_t Unsigned = N == INT64_MIN ? UINT64_C(1) << 63 : uint64_t(-N);
62 return std::make_pair(Unsigned, true);
64 static int64_t joinSigned(uint64_t U, bool IsNeg) {
65 if (U > uint64_t(INT64_MAX))
66 return IsNeg ? INT64_MIN : INT64_MAX;
67 return IsNeg ? -int64_t(U) : int64_t(U);
71 /// \brief Simple representation of an unsigned floating point.
73 /// ScaledNumber is a unsigned floating point number. It uses simple
74 /// saturation arithmetic, and every operation is well-defined for every value.
76 /// The number is split into a signed exponent and unsigned digits. The number
77 /// represented is \c getDigits()*2^getExponent(). In this way, the digits are
78 /// much like the mantissa in the x87 long double, but there is no canonical
79 /// form, so the same number can be represented by many bit representations
80 /// (it's always in "denormal" mode).
82 /// ScaledNumber is templated on the underlying integer type for digits, which
83 /// is expected to be one of uint64_t, uint32_t, uint16_t or uint8_t.
85 /// Unlike builtin floating point types, ScaledNumber is portable.
87 /// Unlike APFloat, ScaledNumber does not model architecture floating point
88 /// behaviour (this should make it a little faster), and implements most
89 /// operators (this makes it usable).
91 /// ScaledNumber is totally ordered. However, there is no canonical form, so
92 /// there are multiple representations of most scalars. E.g.:
94 /// ScaledNumber(8u, 0) == ScaledNumber(4u, 1)
95 /// ScaledNumber(4u, 1) == ScaledNumber(2u, 2)
96 /// ScaledNumber(2u, 2) == ScaledNumber(1u, 3)
98 /// ScaledNumber implements most arithmetic operations. Precision is kept
99 /// where possible. Uses simple saturation arithmetic, so that operations
100 /// saturate to 0.0 or getLargest() rather than under or overflowing. It has
101 /// some extra arithmetic for unit inversion. 0.0/0.0 is defined to be 0.0.
102 /// Any other division by 0.0 is defined to be getLargest().
104 /// As a convenience for modifying the exponent, left and right shifting are
105 /// both implemented, and both interpret negative shifts as positive shifts in
106 /// the opposite direction.
108 /// Exponents are limited to the range accepted by x87 long double. This makes
109 /// it trivial to add functionality to convert to APFloat (this is already
110 /// relied on for the implementation of printing).
112 /// The current plan is to gut this and make the necessary parts of it (even
113 /// more) private to BlockFrequencyInfo.
114 template <class DigitsT> class ScaledNumber : ScaledNumberBase {
116 static_assert(!std::numeric_limits<DigitsT>::is_signed,
117 "only unsigned floats supported");
119 typedef DigitsT DigitsType;
122 typedef std::numeric_limits<DigitsType> DigitsLimits;
124 static const int Width = sizeof(DigitsType) * 8;
125 static_assert(Width <= 64, "invalid integer width for digits");
132 ScaledNumber() : Digits(0), Exponent(0) {}
134 ScaledNumber(DigitsType Digits, int16_t Exponent)
135 : Digits(Digits), Exponent(Exponent) {}
138 ScaledNumber(const std::pair<uint64_t, int16_t> &X)
139 : Digits(X.first), Exponent(X.second) {}
142 static ScaledNumber getZero() { return ScaledNumber(0, 0); }
143 static ScaledNumber getOne() { return ScaledNumber(1, 0); }
144 static ScaledNumber getLargest() {
145 return ScaledNumber(DigitsLimits::max(), MaxExponent);
147 static ScaledNumber getFloat(uint64_t N) { return adjustToWidth(N, 0); }
148 static ScaledNumber getInverseFloat(uint64_t N) {
149 return getFloat(N).invert();
151 static ScaledNumber getFraction(DigitsType N, DigitsType D) {
152 return getQuotient(N, D);
155 int16_t getExponent() const { return Exponent; }
156 DigitsType getDigits() const { return Digits; }
158 /// \brief Convert to the given integer type.
160 /// Convert to \c IntT using simple saturating arithmetic, truncating if
162 template <class IntT> IntT toInt() const;
164 bool isZero() const { return !Digits; }
165 bool isLargest() const { return *this == getLargest(); }
167 if (Exponent > 0 || Exponent <= -Width)
169 return Digits == DigitsType(1) << -Exponent;
172 /// \brief The log base 2, rounded.
174 /// Get the lg of the scalar. lg 0 is defined to be INT32_MIN.
175 int32_t lg() const { return ScaledNumbers::getLg(Digits, Exponent); }
177 /// \brief The log base 2, rounded towards INT32_MIN.
179 /// Get the lg floor. lg 0 is defined to be INT32_MIN.
180 int32_t lgFloor() const {
181 return ScaledNumbers::getLgFloor(Digits, Exponent);
184 /// \brief The log base 2, rounded towards INT32_MAX.
186 /// Get the lg ceiling. lg 0 is defined to be INT32_MIN.
187 int32_t lgCeiling() const {
188 return ScaledNumbers::getLgCeiling(Digits, Exponent);
191 bool operator==(const ScaledNumber &X) const { return compare(X) == 0; }
192 bool operator<(const ScaledNumber &X) const { return compare(X) < 0; }
193 bool operator!=(const ScaledNumber &X) const { return compare(X) != 0; }
194 bool operator>(const ScaledNumber &X) const { return compare(X) > 0; }
195 bool operator<=(const ScaledNumber &X) const { return compare(X) <= 0; }
196 bool operator>=(const ScaledNumber &X) const { return compare(X) >= 0; }
198 bool operator!() const { return isZero(); }
200 /// \brief Convert to a decimal representation in a string.
202 /// Convert to a string. Uses scientific notation for very large/small
203 /// numbers. Scientific notation is used roughly for numbers outside of the
204 /// range 2^-64 through 2^64.
206 /// \c Precision indicates the number of decimal digits of precision to use;
207 /// 0 requests the maximum available.
209 /// As a special case to make debugging easier, if the number is small enough
210 /// to convert without scientific notation and has more than \c Precision
211 /// digits before the decimal place, it's printed accurately to the first
212 /// digit past zero. E.g., assuming 10 digits of precision:
214 /// 98765432198.7654... => 98765432198.8
215 /// 8765432198.7654... => 8765432198.8
216 /// 765432198.7654... => 765432198.8
217 /// 65432198.7654... => 65432198.77
218 /// 5432198.7654... => 5432198.765
219 std::string toString(unsigned Precision = DefaultPrecision) {
220 return ScaledNumberBase::toString(Digits, Exponent, Width, Precision);
223 /// \brief Print a decimal representation.
225 /// Print a string. See toString for documentation.
226 raw_ostream &print(raw_ostream &OS,
227 unsigned Precision = DefaultPrecision) const {
228 return ScaledNumberBase::print(OS, Digits, Exponent, Width, Precision);
230 void dump() const { return ScaledNumberBase::dump(Digits, Exponent, Width); }
232 ScaledNumber &operator+=(const ScaledNumber &X) {
233 std::tie(Digits, Exponent) =
234 ScaledNumbers::getSum(Digits, Exponent, X.Digits, X.Exponent);
235 // Check for exponent past MaxExponent.
236 if (Exponent > MaxExponent)
237 *this = getLargest();
240 ScaledNumber &operator-=(const ScaledNumber &X) {
241 std::tie(Digits, Exponent) =
242 ScaledNumbers::getDifference(Digits, Exponent, X.Digits, X.Exponent);
245 ScaledNumber &operator*=(const ScaledNumber &X);
246 ScaledNumber &operator/=(const ScaledNumber &X);
247 ScaledNumber &operator<<=(int16_t Shift) {
251 ScaledNumber &operator>>=(int16_t Shift) {
257 void shiftLeft(int32_t Shift);
258 void shiftRight(int32_t Shift);
260 /// \brief Adjust two floats to have matching exponents.
262 /// Adjust \c this and \c X to have matching exponents. Returns the new \c X
263 /// by value. Does nothing if \a isZero() for either.
265 /// The value that compares smaller will lose precision, and possibly become
267 ScaledNumber matchExponents(ScaledNumber X) {
268 ScaledNumbers::matchScales(Digits, Exponent, X.Digits, X.Exponent);
273 /// \brief Scale a large number accurately.
275 /// Scale N (multiply it by this). Uses full precision multiplication, even
276 /// if Width is smaller than 64, so information is not lost.
277 uint64_t scale(uint64_t N) const;
278 uint64_t scaleByInverse(uint64_t N) const {
279 // TODO: implement directly, rather than relying on inverse. Inverse is
281 return inverse().scale(N);
283 int64_t scale(int64_t N) const {
284 std::pair<uint64_t, bool> Unsigned = splitSigned(N);
285 return joinSigned(scale(Unsigned.first), Unsigned.second);
287 int64_t scaleByInverse(int64_t N) const {
288 std::pair<uint64_t, bool> Unsigned = splitSigned(N);
289 return joinSigned(scaleByInverse(Unsigned.first), Unsigned.second);
292 int compare(const ScaledNumber &X) const {
293 return ScaledNumbers::compare(Digits, Exponent, X.Digits, X.Exponent);
295 int compareTo(uint64_t N) const {
296 ScaledNumber Float = getFloat(N);
297 int Compare = compare(Float);
298 if (Width == 64 || Compare != 0)
301 // Check for precision loss. We know *this == RoundTrip.
302 uint64_t RoundTrip = Float.template toInt<uint64_t>();
303 return N == RoundTrip ? 0 : RoundTrip < N ? -1 : 1;
305 int compareTo(int64_t N) const { return N < 0 ? 1 : compareTo(uint64_t(N)); }
307 ScaledNumber &invert() { return *this = ScaledNumber::getFloat(1) / *this; }
308 ScaledNumber inverse() const { return ScaledNumber(*this).invert(); }
311 static ScaledNumber getProduct(DigitsType LHS, DigitsType RHS) {
312 return ScaledNumbers::getProduct(LHS, RHS);
314 static ScaledNumber getQuotient(DigitsType Dividend, DigitsType Divisor) {
315 return ScaledNumbers::getQuotient(Dividend, Divisor);
318 static int countLeadingZerosWidth(DigitsType Digits) {
320 return countLeadingZeros64(Digits);
322 return countLeadingZeros32(Digits);
323 return countLeadingZeros32(Digits) + Width - 32;
326 /// \brief Adjust a number to width, rounding up if necessary.
328 /// Should only be called for \c Shift close to zero.
330 /// \pre Shift >= MinExponent && Shift + 64 <= MaxExponent.
331 static ScaledNumber adjustToWidth(uint64_t N, int32_t Shift) {
332 assert(Shift >= MinExponent && "Shift should be close to 0");
333 assert(Shift <= MaxExponent - 64 && "Shift should be close to 0");
334 auto Adjusted = ScaledNumbers::getAdjusted<DigitsT>(N, Shift);
338 static ScaledNumber getRounded(ScaledNumber P, bool Round) {
343 return ScaledNumbers::getRounded(P.Digits, P.Exponent, Round);
347 #define SCALED_NUMBER_BOP(op, base) \
348 template <class DigitsT> \
349 ScaledNumber<DigitsT> operator op(const ScaledNumber<DigitsT> &L, \
350 const ScaledNumber<DigitsT> &R) { \
351 return ScaledNumber<DigitsT>(L) base R; \
353 SCALED_NUMBER_BOP(+, += )
354 SCALED_NUMBER_BOP(-, -= )
355 SCALED_NUMBER_BOP(*, *= )
356 SCALED_NUMBER_BOP(/, /= )
357 SCALED_NUMBER_BOP(<<, <<= )
358 SCALED_NUMBER_BOP(>>, >>= )
359 #undef SCALED_NUMBER_BOP
361 template <class DigitsT>
362 raw_ostream &operator<<(raw_ostream &OS, const ScaledNumber<DigitsT> &X) {
363 return X.print(OS, 10);
366 #define SCALED_NUMBER_COMPARE_TO_TYPE(op, T1, T2) \
367 template <class DigitsT> \
368 bool operator op(const ScaledNumber<DigitsT> &L, T1 R) { \
369 return L.compareTo(T2(R)) op 0; \
371 template <class DigitsT> \
372 bool operator op(T1 L, const ScaledNumber<DigitsT> &R) { \
373 return 0 op R.compareTo(T2(L)); \
375 #define SCALED_NUMBER_COMPARE_TO(op) \
376 SCALED_NUMBER_COMPARE_TO_TYPE(op, uint64_t, uint64_t) \
377 SCALED_NUMBER_COMPARE_TO_TYPE(op, uint32_t, uint64_t) \
378 SCALED_NUMBER_COMPARE_TO_TYPE(op, int64_t, int64_t) \
379 SCALED_NUMBER_COMPARE_TO_TYPE(op, int32_t, int64_t)
380 SCALED_NUMBER_COMPARE_TO(< )
381 SCALED_NUMBER_COMPARE_TO(> )
382 SCALED_NUMBER_COMPARE_TO(== )
383 SCALED_NUMBER_COMPARE_TO(!= )
384 SCALED_NUMBER_COMPARE_TO(<= )
385 SCALED_NUMBER_COMPARE_TO(>= )
386 #undef SCALED_NUMBER_COMPARE_TO
387 #undef SCALED_NUMBER_COMPARE_TO_TYPE
389 template <class DigitsT>
390 uint64_t ScaledNumber<DigitsT>::scale(uint64_t N) const {
391 if (Width == 64 || N <= DigitsLimits::max())
392 return (getFloat(N) * *this).template toInt<uint64_t>();
394 // Defer to the 64-bit version.
395 return ScaledNumber<uint64_t>(Digits, Exponent).scale(N);
398 template <class DigitsT>
399 template <class IntT>
400 IntT ScaledNumber<DigitsT>::toInt() const {
401 typedef std::numeric_limits<IntT> Limits;
404 if (*this >= Limits::max())
405 return Limits::max();
409 assert(size_t(Exponent) < sizeof(IntT) * 8);
410 return N << Exponent;
413 assert(size_t(-Exponent) < sizeof(IntT) * 8);
414 return N >> -Exponent;
419 template <class DigitsT>
420 ScaledNumber<DigitsT> &ScaledNumber<DigitsT>::
421 operator*=(const ScaledNumber &X) {
427 // Save the exponents.
428 int32_t Exponents = int32_t(Exponent) + int32_t(X.Exponent);
430 // Get the raw product.
431 *this = getProduct(Digits, X.Digits);
433 // Combine with exponents.
434 return *this <<= Exponents;
436 template <class DigitsT>
437 ScaledNumber<DigitsT> &ScaledNumber<DigitsT>::
438 operator/=(const ScaledNumber &X) {
442 return *this = getLargest();
444 // Save the exponents.
445 int32_t Exponents = int32_t(Exponent) - int32_t(X.Exponent);
447 // Get the raw quotient.
448 *this = getQuotient(Digits, X.Digits);
450 // Combine with exponents.
451 return *this <<= Exponents;
453 template <class DigitsT> void ScaledNumber<DigitsT>::shiftLeft(int32_t Shift) {
454 if (!Shift || isZero())
456 assert(Shift != INT32_MIN);
462 // Shift as much as we can in the exponent.
463 int32_t ExponentShift = std::min(Shift, MaxExponent - Exponent);
464 Exponent += ExponentShift;
465 if (ExponentShift == Shift)
468 // Check this late, since it's rare.
472 // Shift the digits themselves.
473 Shift -= ExponentShift;
474 if (Shift > countLeadingZerosWidth(Digits)) {
476 *this = getLargest();
484 template <class DigitsT> void ScaledNumber<DigitsT>::shiftRight(int32_t Shift) {
485 if (!Shift || isZero())
487 assert(Shift != INT32_MIN);
493 // Shift as much as we can in the exponent.
494 int32_t ExponentShift = std::min(Shift, Exponent - MinExponent);
495 Exponent -= ExponentShift;
496 if (ExponentShift == Shift)
499 // Shift the digits themselves.
500 Shift -= ExponentShift;
501 if (Shift >= Width) {
511 template <class T> struct isPodLike<ScaledNumber<T>> {
512 static const bool value = true;
516 //===----------------------------------------------------------------------===//
518 // BlockMass definition.
520 // TODO: Make this private to BlockFrequencyInfoImpl or delete.
522 //===----------------------------------------------------------------------===//
525 /// \brief Mass of a block.
527 /// This class implements a sort of fixed-point fraction always between 0.0 and
528 /// 1.0. getMass() == UINT64_MAX indicates a value of 1.0.
530 /// Masses can be added and subtracted. Simple saturation arithmetic is used,
531 /// so arithmetic operations never overflow or underflow.
533 /// Masses can be multiplied. Multiplication treats full mass as 1.0 and uses
534 /// an inexpensive floating-point algorithm that's off-by-one (almost, but not
535 /// quite, maximum precision).
537 /// Masses can be scaled by \a BranchProbability at maximum precision.
542 BlockMass() : Mass(0) {}
543 explicit BlockMass(uint64_t Mass) : Mass(Mass) {}
545 static BlockMass getEmpty() { return BlockMass(); }
546 static BlockMass getFull() { return BlockMass(UINT64_MAX); }
548 uint64_t getMass() const { return Mass; }
550 bool isFull() const { return Mass == UINT64_MAX; }
551 bool isEmpty() const { return !Mass; }
553 bool operator!() const { return isEmpty(); }
555 /// \brief Add another mass.
557 /// Adds another mass, saturating at \a isFull() rather than overflowing.
558 BlockMass &operator+=(const BlockMass &X) {
559 uint64_t Sum = Mass + X.Mass;
560 Mass = Sum < Mass ? UINT64_MAX : Sum;
564 /// \brief Subtract another mass.
566 /// Subtracts another mass, saturating at \a isEmpty() rather than
568 BlockMass &operator-=(const BlockMass &X) {
569 uint64_t Diff = Mass - X.Mass;
570 Mass = Diff > Mass ? 0 : Diff;
574 BlockMass &operator*=(const BranchProbability &P) {
575 Mass = P.scale(Mass);
579 bool operator==(const BlockMass &X) const { return Mass == X.Mass; }
580 bool operator!=(const BlockMass &X) const { return Mass != X.Mass; }
581 bool operator<=(const BlockMass &X) const { return Mass <= X.Mass; }
582 bool operator>=(const BlockMass &X) const { return Mass >= X.Mass; }
583 bool operator<(const BlockMass &X) const { return Mass < X.Mass; }
584 bool operator>(const BlockMass &X) const { return Mass > X.Mass; }
586 /// \brief Convert to floating point.
588 /// Convert to a float. \a isFull() gives 1.0, while \a isEmpty() gives
589 /// slightly above 0.0.
590 ScaledNumber<uint64_t> toFloat() const;
593 raw_ostream &print(raw_ostream &OS) const;
596 inline BlockMass operator+(const BlockMass &L, const BlockMass &R) {
597 return BlockMass(L) += R;
599 inline BlockMass operator-(const BlockMass &L, const BlockMass &R) {
600 return BlockMass(L) -= R;
602 inline BlockMass operator*(const BlockMass &L, const BranchProbability &R) {
603 return BlockMass(L) *= R;
605 inline BlockMass operator*(const BranchProbability &L, const BlockMass &R) {
606 return BlockMass(R) *= L;
609 inline raw_ostream &operator<<(raw_ostream &OS, const BlockMass &X) {
613 template <> struct isPodLike<BlockMass> {
614 static const bool value = true;
618 //===----------------------------------------------------------------------===//
620 // BlockFrequencyInfoImpl definition.
622 //===----------------------------------------------------------------------===//
626 class BranchProbabilityInfo;
630 class MachineBasicBlock;
631 class MachineBranchProbabilityInfo;
632 class MachineFunction;
634 class MachineLoopInfo;
636 namespace bfi_detail {
637 struct IrreducibleGraph;
639 // This is part of a workaround for a GCC 4.7 crash on lambdas.
640 template <class BT> struct BlockEdgesAdder;
643 /// \brief Base class for BlockFrequencyInfoImpl
645 /// BlockFrequencyInfoImplBase has supporting data structures and some
646 /// algorithms for BlockFrequencyInfoImplBase. Only algorithms that depend on
647 /// the block type (or that call such algorithms) are skipped here.
649 /// Nevertheless, the majority of the overall algorithm documention lives with
650 /// BlockFrequencyInfoImpl. See there for details.
651 class BlockFrequencyInfoImplBase {
653 typedef ScaledNumber<uint64_t> Float;
655 /// \brief Representative of a block.
657 /// This is a simple wrapper around an index into the reverse-post-order
658 /// traversal of the blocks.
660 /// Unlike a block pointer, its order has meaning (location in the
661 /// topological sort) and it's class is the same regardless of block type.
663 typedef uint32_t IndexType;
666 bool operator==(const BlockNode &X) const { return Index == X.Index; }
667 bool operator!=(const BlockNode &X) const { return Index != X.Index; }
668 bool operator<=(const BlockNode &X) const { return Index <= X.Index; }
669 bool operator>=(const BlockNode &X) const { return Index >= X.Index; }
670 bool operator<(const BlockNode &X) const { return Index < X.Index; }
671 bool operator>(const BlockNode &X) const { return Index > X.Index; }
673 BlockNode() : Index(UINT32_MAX) {}
674 BlockNode(IndexType Index) : Index(Index) {}
676 bool isValid() const { return Index <= getMaxIndex(); }
677 static size_t getMaxIndex() { return UINT32_MAX - 1; }
680 /// \brief Stats about a block itself.
681 struct FrequencyData {
686 /// \brief Data about a loop.
688 /// Contains the data necessary to represent represent a loop as a
689 /// pseudo-node once it's packaged.
691 typedef SmallVector<std::pair<BlockNode, BlockMass>, 4> ExitMap;
692 typedef SmallVector<BlockNode, 4> NodeList;
693 LoopData *Parent; ///< The parent loop.
694 bool IsPackaged; ///< Whether this has been packaged.
695 uint32_t NumHeaders; ///< Number of headers.
696 ExitMap Exits; ///< Successor edges (and weights).
697 NodeList Nodes; ///< Header and the members of the loop.
698 BlockMass BackedgeMass; ///< Mass returned to loop header.
702 LoopData(LoopData *Parent, const BlockNode &Header)
703 : Parent(Parent), IsPackaged(false), NumHeaders(1), Nodes(1, Header) {}
704 template <class It1, class It2>
705 LoopData(LoopData *Parent, It1 FirstHeader, It1 LastHeader, It2 FirstOther,
707 : Parent(Parent), IsPackaged(false), Nodes(FirstHeader, LastHeader) {
708 NumHeaders = Nodes.size();
709 Nodes.insert(Nodes.end(), FirstOther, LastOther);
711 bool isHeader(const BlockNode &Node) const {
713 return std::binary_search(Nodes.begin(), Nodes.begin() + NumHeaders,
715 return Node == Nodes[0];
717 BlockNode getHeader() const { return Nodes[0]; }
718 bool isIrreducible() const { return NumHeaders > 1; }
720 NodeList::const_iterator members_begin() const {
721 return Nodes.begin() + NumHeaders;
723 NodeList::const_iterator members_end() const { return Nodes.end(); }
724 iterator_range<NodeList::const_iterator> members() const {
725 return make_range(members_begin(), members_end());
729 /// \brief Index of loop information.
731 BlockNode Node; ///< This node.
732 LoopData *Loop; ///< The loop this block is inside.
733 BlockMass Mass; ///< Mass distribution from the entry block.
735 WorkingData(const BlockNode &Node) : Node(Node), Loop(nullptr) {}
737 bool isLoopHeader() const { return Loop && Loop->isHeader(Node); }
738 bool isDoubleLoopHeader() const {
739 return isLoopHeader() && Loop->Parent && Loop->Parent->isIrreducible() &&
740 Loop->Parent->isHeader(Node);
743 LoopData *getContainingLoop() const {
746 if (!isDoubleLoopHeader())
748 return Loop->Parent->Parent;
751 /// \brief Resolve a node to its representative.
753 /// Get the node currently representing Node, which could be a containing
756 /// This function should only be called when distributing mass. As long as
757 /// there are no irreducilbe edges to Node, then it will have complexity
758 /// O(1) in this context.
760 /// In general, the complexity is O(L), where L is the number of loop
761 /// headers Node has been packaged into. Since this method is called in
762 /// the context of distributing mass, L will be the number of loop headers
763 /// an early exit edge jumps out of.
764 BlockNode getResolvedNode() const {
765 auto L = getPackagedLoop();
766 return L ? L->getHeader() : Node;
768 LoopData *getPackagedLoop() const {
769 if (!Loop || !Loop->IsPackaged)
772 while (L->Parent && L->Parent->IsPackaged)
777 /// \brief Get the appropriate mass for a node.
779 /// Get appropriate mass for Node. If Node is a loop-header (whose loop
780 /// has been packaged), returns the mass of its pseudo-node. If it's a
781 /// node inside a packaged loop, it returns the loop's mass.
782 BlockMass &getMass() {
785 if (!isADoublePackage())
787 return Loop->Parent->Mass;
790 /// \brief Has ContainingLoop been packaged up?
791 bool isPackaged() const { return getResolvedNode() != Node; }
792 /// \brief Has Loop been packaged up?
793 bool isAPackage() const { return isLoopHeader() && Loop->IsPackaged; }
794 /// \brief Has Loop been packaged up twice?
795 bool isADoublePackage() const {
796 return isDoubleLoopHeader() && Loop->Parent->IsPackaged;
800 /// \brief Unscaled probability weight.
802 /// Probability weight for an edge in the graph (including the
803 /// successor/target node).
805 /// All edges in the original function are 32-bit. However, exit edges from
806 /// loop packages are taken from 64-bit exit masses, so we need 64-bits of
807 /// space in general.
809 /// In addition to the raw weight amount, Weight stores the type of the edge
810 /// in the current context (i.e., the context of the loop being processed).
811 /// Is this a local edge within the loop, an exit from the loop, or a
812 /// backedge to the loop header?
814 enum DistType { Local, Exit, Backedge };
816 BlockNode TargetNode;
818 Weight() : Type(Local), Amount(0) {}
821 /// \brief Distribution of unscaled probability weight.
823 /// Distribution of unscaled probability weight to a set of successors.
825 /// This class collates the successor edge weights for later processing.
827 /// \a DidOverflow indicates whether \a Total did overflow while adding to
828 /// the distribution. It should never overflow twice.
829 struct Distribution {
830 typedef SmallVector<Weight, 4> WeightList;
831 WeightList Weights; ///< Individual successor weights.
832 uint64_t Total; ///< Sum of all weights.
833 bool DidOverflow; ///< Whether \a Total did overflow.
835 Distribution() : Total(0), DidOverflow(false) {}
836 void addLocal(const BlockNode &Node, uint64_t Amount) {
837 add(Node, Amount, Weight::Local);
839 void addExit(const BlockNode &Node, uint64_t Amount) {
840 add(Node, Amount, Weight::Exit);
842 void addBackedge(const BlockNode &Node, uint64_t Amount) {
843 add(Node, Amount, Weight::Backedge);
846 /// \brief Normalize the distribution.
848 /// Combines multiple edges to the same \a Weight::TargetNode and scales
849 /// down so that \a Total fits into 32-bits.
851 /// This is linear in the size of \a Weights. For the vast majority of
852 /// cases, adjacent edge weights are combined by sorting WeightList and
853 /// combining adjacent weights. However, for very large edge lists an
854 /// auxiliary hash table is used.
858 void add(const BlockNode &Node, uint64_t Amount, Weight::DistType Type);
861 /// \brief Data about each block. This is used downstream.
862 std::vector<FrequencyData> Freqs;
864 /// \brief Loop data: see initializeLoops().
865 std::vector<WorkingData> Working;
867 /// \brief Indexed information about loops.
868 std::list<LoopData> Loops;
870 /// \brief Add all edges out of a packaged loop to the distribution.
872 /// Adds all edges from LocalLoopHead to Dist. Calls addToDist() to add each
875 /// \return \c true unless there's an irreducible backedge.
876 bool addLoopSuccessorsToDist(const LoopData *OuterLoop, LoopData &Loop,
879 /// \brief Add an edge to the distribution.
881 /// Adds an edge to Succ to Dist. If \c LoopHead.isValid(), then whether the
882 /// edge is local/exit/backedge is in the context of LoopHead. Otherwise,
883 /// every edge should be a local edge (since all the loops are packaged up).
885 /// \return \c true unless aborted due to an irreducible backedge.
886 bool addToDist(Distribution &Dist, const LoopData *OuterLoop,
887 const BlockNode &Pred, const BlockNode &Succ, uint64_t Weight);
889 LoopData &getLoopPackage(const BlockNode &Head) {
890 assert(Head.Index < Working.size());
891 assert(Working[Head.Index].isLoopHeader());
892 return *Working[Head.Index].Loop;
895 /// \brief Analyze irreducible SCCs.
897 /// Separate irreducible SCCs from \c G, which is an explict graph of \c
898 /// OuterLoop (or the top-level function, if \c OuterLoop is \c nullptr).
899 /// Insert them into \a Loops before \c Insert.
901 /// \return the \c LoopData nodes representing the irreducible SCCs.
902 iterator_range<std::list<LoopData>::iterator>
903 analyzeIrreducible(const bfi_detail::IrreducibleGraph &G, LoopData *OuterLoop,
904 std::list<LoopData>::iterator Insert);
906 /// \brief Update a loop after packaging irreducible SCCs inside of it.
908 /// Update \c OuterLoop. Before finding irreducible control flow, it was
909 /// partway through \a computeMassInLoop(), so \a LoopData::Exits and \a
910 /// LoopData::BackedgeMass need to be reset. Also, nodes that were packaged
911 /// up need to be removed from \a OuterLoop::Nodes.
912 void updateLoopWithIrreducible(LoopData &OuterLoop);
914 /// \brief Distribute mass according to a distribution.
916 /// Distributes the mass in Source according to Dist. If LoopHead.isValid(),
917 /// backedges and exits are stored in its entry in Loops.
919 /// Mass is distributed in parallel from two copies of the source mass.
920 void distributeMass(const BlockNode &Source, LoopData *OuterLoop,
923 /// \brief Compute the loop scale for a loop.
924 void computeLoopScale(LoopData &Loop);
926 /// \brief Package up a loop.
927 void packageLoop(LoopData &Loop);
929 /// \brief Unwrap loops.
932 /// \brief Finalize frequency metrics.
934 /// Calculates final frequencies and cleans up no-longer-needed data
936 void finalizeMetrics();
938 /// \brief Clear all memory.
941 virtual std::string getBlockName(const BlockNode &Node) const;
942 std::string getLoopName(const LoopData &Loop) const;
944 virtual raw_ostream &print(raw_ostream &OS) const { return OS; }
945 void dump() const { print(dbgs()); }
947 Float getFloatingBlockFreq(const BlockNode &Node) const;
949 BlockFrequency getBlockFreq(const BlockNode &Node) const;
951 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockNode &Node) const;
952 raw_ostream &printBlockFreq(raw_ostream &OS,
953 const BlockFrequency &Freq) const;
955 uint64_t getEntryFreq() const {
956 assert(!Freqs.empty());
957 return Freqs[0].Integer;
959 /// \brief Virtual destructor.
961 /// Need a virtual destructor to mask the compiler warning about
963 virtual ~BlockFrequencyInfoImplBase() {}
966 namespace bfi_detail {
967 template <class BlockT> struct TypeMap {};
968 template <> struct TypeMap<BasicBlock> {
969 typedef BasicBlock BlockT;
970 typedef Function FunctionT;
971 typedef BranchProbabilityInfo BranchProbabilityInfoT;
973 typedef LoopInfo LoopInfoT;
975 template <> struct TypeMap<MachineBasicBlock> {
976 typedef MachineBasicBlock BlockT;
977 typedef MachineFunction FunctionT;
978 typedef MachineBranchProbabilityInfo BranchProbabilityInfoT;
979 typedef MachineLoop LoopT;
980 typedef MachineLoopInfo LoopInfoT;
983 /// \brief Get the name of a MachineBasicBlock.
985 /// Get the name of a MachineBasicBlock. It's templated so that including from
986 /// CodeGen is unnecessary (that would be a layering issue).
988 /// This is used mainly for debug output. The name is similar to
989 /// MachineBasicBlock::getFullName(), but skips the name of the function.
990 template <class BlockT> std::string getBlockName(const BlockT *BB) {
991 assert(BB && "Unexpected nullptr");
992 auto MachineName = "BB" + Twine(BB->getNumber());
993 if (BB->getBasicBlock())
994 return (MachineName + "[" + BB->getName() + "]").str();
995 return MachineName.str();
997 /// \brief Get the name of a BasicBlock.
998 template <> inline std::string getBlockName(const BasicBlock *BB) {
999 assert(BB && "Unexpected nullptr");
1000 return BB->getName().str();
1003 /// \brief Graph of irreducible control flow.
1005 /// This graph is used for determining the SCCs in a loop (or top-level
1006 /// function) that has irreducible control flow.
1008 /// During the block frequency algorithm, the local graphs are defined in a
1009 /// light-weight way, deferring to the \a BasicBlock or \a MachineBasicBlock
1010 /// graphs for most edges, but getting others from \a LoopData::ExitMap. The
1011 /// latter only has successor information.
1013 /// \a IrreducibleGraph makes this graph explicit. It's in a form that can use
1014 /// \a GraphTraits (so that \a analyzeIrreducible() can use \a scc_iterator),
1015 /// and it explicitly lists predecessors and successors. The initialization
1016 /// that relies on \c MachineBasicBlock is defined in the header.
1017 struct IrreducibleGraph {
1018 typedef BlockFrequencyInfoImplBase BFIBase;
1022 typedef BFIBase::BlockNode BlockNode;
1026 std::deque<const IrrNode *> Edges;
1027 IrrNode(const BlockNode &Node) : Node(Node), NumIn(0) {}
1029 typedef std::deque<const IrrNode *>::const_iterator iterator;
1030 iterator pred_begin() const { return Edges.begin(); }
1031 iterator succ_begin() const { return Edges.begin() + NumIn; }
1032 iterator pred_end() const { return succ_begin(); }
1033 iterator succ_end() const { return Edges.end(); }
1036 const IrrNode *StartIrr;
1037 std::vector<IrrNode> Nodes;
1038 SmallDenseMap<uint32_t, IrrNode *, 4> Lookup;
1040 /// \brief Construct an explicit graph containing irreducible control flow.
1042 /// Construct an explicit graph of the control flow in \c OuterLoop (or the
1043 /// top-level function, if \c OuterLoop is \c nullptr). Uses \c
1044 /// addBlockEdges to add block successors that have not been packaged into
1047 /// \a BlockFrequencyInfoImpl::computeIrreducibleMass() is the only expected
1049 template <class BlockEdgesAdder>
1050 IrreducibleGraph(BFIBase &BFI, const BFIBase::LoopData *OuterLoop,
1051 BlockEdgesAdder addBlockEdges)
1052 : BFI(BFI), StartIrr(nullptr) {
1053 initialize(OuterLoop, addBlockEdges);
1056 template <class BlockEdgesAdder>
1057 void initialize(const BFIBase::LoopData *OuterLoop,
1058 BlockEdgesAdder addBlockEdges);
1059 void addNodesInLoop(const BFIBase::LoopData &OuterLoop);
1060 void addNodesInFunction();
1061 void addNode(const BlockNode &Node) {
1062 Nodes.emplace_back(Node);
1063 BFI.Working[Node.Index].getMass() = BlockMass::getEmpty();
1066 template <class BlockEdgesAdder>
1067 void addEdges(const BlockNode &Node, const BFIBase::LoopData *OuterLoop,
1068 BlockEdgesAdder addBlockEdges);
1069 void addEdge(IrrNode &Irr, const BlockNode &Succ,
1070 const BFIBase::LoopData *OuterLoop);
1072 template <class BlockEdgesAdder>
1073 void IrreducibleGraph::initialize(const BFIBase::LoopData *OuterLoop,
1074 BlockEdgesAdder addBlockEdges) {
1076 addNodesInLoop(*OuterLoop);
1077 for (auto N : OuterLoop->Nodes)
1078 addEdges(N, OuterLoop, addBlockEdges);
1080 addNodesInFunction();
1081 for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index)
1082 addEdges(Index, OuterLoop, addBlockEdges);
1084 StartIrr = Lookup[Start.Index];
1086 template <class BlockEdgesAdder>
1087 void IrreducibleGraph::addEdges(const BlockNode &Node,
1088 const BFIBase::LoopData *OuterLoop,
1089 BlockEdgesAdder addBlockEdges) {
1090 auto L = Lookup.find(Node.Index);
1091 if (L == Lookup.end())
1093 IrrNode &Irr = *L->second;
1094 const auto &Working = BFI.Working[Node.Index];
1096 if (Working.isAPackage())
1097 for (const auto &I : Working.Loop->Exits)
1098 addEdge(Irr, I.first, OuterLoop);
1100 addBlockEdges(*this, Irr, OuterLoop);
1104 /// \brief Shared implementation for block frequency analysis.
1106 /// This is a shared implementation of BlockFrequencyInfo and
1107 /// MachineBlockFrequencyInfo, and calculates the relative frequencies of
1110 /// LoopInfo defines a loop as a "non-trivial" SCC dominated by a single block,
1111 /// which is called the header. A given loop, L, can have sub-loops, which are
1112 /// loops within the subgraph of L that exclude its header. (A "trivial" SCC
1113 /// consists of a single block that does not have a self-edge.)
1115 /// In addition to loops, this algorithm has limited support for irreducible
1116 /// SCCs, which are SCCs with multiple entry blocks. Irreducible SCCs are
1117 /// discovered on they fly, and modelled as loops with multiple headers.
1119 /// The headers of irreducible sub-SCCs consist of its entry blocks and all
1120 /// nodes that are targets of a backedge within it (excluding backedges within
1121 /// true sub-loops). Block frequency calculations act as if a block is
1122 /// inserted that intercepts all the edges to the headers. All backedges and
1123 /// entries point to this block. Its successors are the headers, which split
1124 /// the frequency evenly.
1126 /// This algorithm leverages BlockMass and ScaledNumber to maintain precision,
1127 /// separates mass distribution from loop scaling, and dithers to eliminate
1128 /// probability mass loss.
1130 /// The implementation is split between BlockFrequencyInfoImpl, which knows the
1131 /// type of graph being modelled (BasicBlock vs. MachineBasicBlock), and
1132 /// BlockFrequencyInfoImplBase, which doesn't. The base class uses \a
1133 /// BlockNode, a wrapper around a uint32_t. BlockNode is numbered from 0 in
1134 /// reverse-post order. This gives two advantages: it's easy to compare the
1135 /// relative ordering of two nodes, and maps keyed on BlockT can be represented
1138 /// This algorithm is O(V+E), unless there is irreducible control flow, in
1139 /// which case it's O(V*E) in the worst case.
1141 /// These are the main stages:
1143 /// 0. Reverse post-order traversal (\a initializeRPOT()).
1145 /// Run a single post-order traversal and save it (in reverse) in RPOT.
1146 /// All other stages make use of this ordering. Save a lookup from BlockT
1147 /// to BlockNode (the index into RPOT) in Nodes.
1149 /// 1. Loop initialization (\a initializeLoops()).
1151 /// Translate LoopInfo/MachineLoopInfo into a form suitable for the rest of
1152 /// the algorithm. In particular, store the immediate members of each loop
1153 /// in reverse post-order.
1155 /// 2. Calculate mass and scale in loops (\a computeMassInLoops()).
1157 /// For each loop (bottom-up), distribute mass through the DAG resulting
1158 /// from ignoring backedges and treating sub-loops as a single pseudo-node.
1159 /// Track the backedge mass distributed to the loop header, and use it to
1160 /// calculate the loop scale (number of loop iterations). Immediate
1161 /// members that represent sub-loops will already have been visited and
1162 /// packaged into a pseudo-node.
1164 /// Distributing mass in a loop is a reverse-post-order traversal through
1165 /// the loop. Start by assigning full mass to the Loop header. For each
1166 /// node in the loop:
1168 /// - Fetch and categorize the weight distribution for its successors.
1169 /// If this is a packaged-subloop, the weight distribution is stored
1170 /// in \a LoopData::Exits. Otherwise, fetch it from
1171 /// BranchProbabilityInfo.
1173 /// - Each successor is categorized as \a Weight::Local, a local edge
1174 /// within the current loop, \a Weight::Backedge, a backedge to the
1175 /// loop header, or \a Weight::Exit, any successor outside the loop.
1176 /// The weight, the successor, and its category are stored in \a
1177 /// Distribution. There can be multiple edges to each successor.
1179 /// - If there's a backedge to a non-header, there's an irreducible SCC.
1180 /// The usual flow is temporarily aborted. \a
1181 /// computeIrreducibleMass() finds the irreducible SCCs within the
1182 /// loop, packages them up, and restarts the flow.
1184 /// - Normalize the distribution: scale weights down so that their sum
1185 /// is 32-bits, and coalesce multiple edges to the same node.
1187 /// - Distribute the mass accordingly, dithering to minimize mass loss,
1188 /// as described in \a distributeMass().
1190 /// Finally, calculate the loop scale from the accumulated backedge mass.
1192 /// 3. Distribute mass in the function (\a computeMassInFunction()).
1194 /// Finally, distribute mass through the DAG resulting from packaging all
1195 /// loops in the function. This uses the same algorithm as distributing
1196 /// mass in a loop, except that there are no exit or backedge edges.
1198 /// 4. Unpackage loops (\a unwrapLoops()).
1200 /// Initialize each block's frequency to a floating point representation of
1203 /// Visit loops top-down, scaling the frequencies of its immediate members
1204 /// by the loop's pseudo-node's frequency.
1206 /// 5. Convert frequencies to a 64-bit range (\a finalizeMetrics()).
1208 /// Using the min and max frequencies as a guide, translate floating point
1209 /// frequencies to an appropriate range in uint64_t.
1211 /// It has some known flaws.
1213 /// - Loop scale is limited to 4096 per loop (2^12) to avoid exhausting
1214 /// BlockFrequency's 64-bit integer precision.
1216 /// - The model of irreducible control flow is a rough approximation.
1218 /// Modelling irreducible control flow exactly involves setting up and
1219 /// solving a group of infinite geometric series. Such precision is
1220 /// unlikely to be worthwhile, since most of our algorithms give up on
1221 /// irreducible control flow anyway.
1223 /// Nevertheless, we might find that we need to get closer. Here's a sort
1224 /// of TODO list for the model with diminishing returns, to be completed as
1227 /// - The headers for the \a LoopData representing an irreducible SCC
1228 /// include non-entry blocks. When these extra blocks exist, they
1229 /// indicate a self-contained irreducible sub-SCC. We could treat them
1230 /// as sub-loops, rather than arbitrarily shoving the problematic
1231 /// blocks into the headers of the main irreducible SCC.
1233 /// - Backedge frequencies are assumed to be evenly split between the
1234 /// headers of a given irreducible SCC. Instead, we could track the
1235 /// backedge mass separately for each header, and adjust their relative
1238 /// - Entry frequencies are assumed to be evenly split between the
1239 /// headers of a given irreducible SCC, which is the only option if we
1240 /// need to compute mass in the SCC before its parent loop. Instead,
1241 /// we could partially compute mass in the parent loop, and stop when
1242 /// we get to the SCC. Here, we have the correct ratio of entry
1243 /// masses, which we can use to adjust their relative frequencies.
1244 /// Compute mass in the SCC, and then continue propagation in the
1247 /// - We can propagate mass iteratively through the SCC, for some fixed
1248 /// number of iterations. Each iteration starts by assigning the entry
1249 /// blocks their backedge mass from the prior iteration. The final
1250 /// mass for each block (and each exit, and the total backedge mass
1251 /// used for computing loop scale) is the sum of all iterations.
1252 /// (Running this until fixed point would "solve" the geometric
1253 /// series by simulation.)
1254 template <class BT> class BlockFrequencyInfoImpl : BlockFrequencyInfoImplBase {
1255 typedef typename bfi_detail::TypeMap<BT>::BlockT BlockT;
1256 typedef typename bfi_detail::TypeMap<BT>::FunctionT FunctionT;
1257 typedef typename bfi_detail::TypeMap<BT>::BranchProbabilityInfoT
1258 BranchProbabilityInfoT;
1259 typedef typename bfi_detail::TypeMap<BT>::LoopT LoopT;
1260 typedef typename bfi_detail::TypeMap<BT>::LoopInfoT LoopInfoT;
1262 // This is part of a workaround for a GCC 4.7 crash on lambdas.
1263 friend struct bfi_detail::BlockEdgesAdder<BT>;
1265 typedef GraphTraits<const BlockT *> Successor;
1266 typedef GraphTraits<Inverse<const BlockT *>> Predecessor;
1268 const BranchProbabilityInfoT *BPI;
1269 const LoopInfoT *LI;
1272 // All blocks in reverse postorder.
1273 std::vector<const BlockT *> RPOT;
1274 DenseMap<const BlockT *, BlockNode> Nodes;
1276 typedef typename std::vector<const BlockT *>::const_iterator rpot_iterator;
1278 rpot_iterator rpot_begin() const { return RPOT.begin(); }
1279 rpot_iterator rpot_end() const { return RPOT.end(); }
1281 size_t getIndex(const rpot_iterator &I) const { return I - rpot_begin(); }
1283 BlockNode getNode(const rpot_iterator &I) const {
1284 return BlockNode(getIndex(I));
1286 BlockNode getNode(const BlockT *BB) const { return Nodes.lookup(BB); }
1288 const BlockT *getBlock(const BlockNode &Node) const {
1289 assert(Node.Index < RPOT.size());
1290 return RPOT[Node.Index];
1293 /// \brief Run (and save) a post-order traversal.
1295 /// Saves a reverse post-order traversal of all the nodes in \a F.
1296 void initializeRPOT();
1298 /// \brief Initialize loop data.
1300 /// Build up \a Loops using \a LoopInfo. \a LoopInfo gives us a mapping from
1301 /// each block to the deepest loop it's in, but we need the inverse. For each
1302 /// loop, we store in reverse post-order its "immediate" members, defined as
1303 /// the header, the headers of immediate sub-loops, and all other blocks in
1304 /// the loop that are not in sub-loops.
1305 void initializeLoops();
1307 /// \brief Propagate to a block's successors.
1309 /// In the context of distributing mass through \c OuterLoop, divide the mass
1310 /// currently assigned to \c Node between its successors.
1312 /// \return \c true unless there's an irreducible backedge.
1313 bool propagateMassToSuccessors(LoopData *OuterLoop, const BlockNode &Node);
1315 /// \brief Compute mass in a particular loop.
1317 /// Assign mass to \c Loop's header, and then for each block in \c Loop in
1318 /// reverse post-order, distribute mass to its successors. Only visits nodes
1319 /// that have not been packaged into sub-loops.
1321 /// \pre \a computeMassInLoop() has been called for each subloop of \c Loop.
1322 /// \return \c true unless there's an irreducible backedge.
1323 bool computeMassInLoop(LoopData &Loop);
1325 /// \brief Try to compute mass in the top-level function.
1327 /// Assign mass to the entry block, and then for each block in reverse
1328 /// post-order, distribute mass to its successors. Skips nodes that have
1329 /// been packaged into loops.
1331 /// \pre \a computeMassInLoops() has been called.
1332 /// \return \c true unless there's an irreducible backedge.
1333 bool tryToComputeMassInFunction();
1335 /// \brief Compute mass in (and package up) irreducible SCCs.
1337 /// Find the irreducible SCCs in \c OuterLoop, add them to \a Loops (in front
1338 /// of \c Insert), and call \a computeMassInLoop() on each of them.
1340 /// If \c OuterLoop is \c nullptr, it refers to the top-level function.
1342 /// \pre \a computeMassInLoop() has been called for each subloop of \c
1344 /// \pre \c Insert points at the the last loop successfully processed by \a
1345 /// computeMassInLoop().
1346 /// \pre \c OuterLoop has irreducible SCCs.
1347 void computeIrreducibleMass(LoopData *OuterLoop,
1348 std::list<LoopData>::iterator Insert);
1350 /// \brief Compute mass in all loops.
1352 /// For each loop bottom-up, call \a computeMassInLoop().
1354 /// \a computeMassInLoop() aborts (and returns \c false) on loops that
1355 /// contain a irreducible sub-SCCs. Use \a computeIrreducibleMass() and then
1356 /// re-enter \a computeMassInLoop().
1358 /// \post \a computeMassInLoop() has returned \c true for every loop.
1359 void computeMassInLoops();
1361 /// \brief Compute mass in the top-level function.
1363 /// Uses \a tryToComputeMassInFunction() and \a computeIrreducibleMass() to
1364 /// compute mass in the top-level function.
1366 /// \post \a tryToComputeMassInFunction() has returned \c true.
1367 void computeMassInFunction();
1369 std::string getBlockName(const BlockNode &Node) const override {
1370 return bfi_detail::getBlockName(getBlock(Node));
1374 const FunctionT *getFunction() const { return F; }
1376 void doFunction(const FunctionT *F, const BranchProbabilityInfoT *BPI,
1377 const LoopInfoT *LI);
1378 BlockFrequencyInfoImpl() : BPI(nullptr), LI(nullptr), F(nullptr) {}
1380 using BlockFrequencyInfoImplBase::getEntryFreq;
1381 BlockFrequency getBlockFreq(const BlockT *BB) const {
1382 return BlockFrequencyInfoImplBase::getBlockFreq(getNode(BB));
1384 Float getFloatingBlockFreq(const BlockT *BB) const {
1385 return BlockFrequencyInfoImplBase::getFloatingBlockFreq(getNode(BB));
1388 /// \brief Print the frequencies for the current function.
1390 /// Prints the frequencies for the blocks in the current function.
1392 /// Blocks are printed in the natural iteration order of the function, rather
1393 /// than reverse post-order. This provides two advantages: writing -analyze
1394 /// tests is easier (since blocks come out in source order), and even
1395 /// unreachable blocks are printed.
1397 /// \a BlockFrequencyInfoImplBase::print() only knows reverse post-order, so
1398 /// we need to override it here.
1399 raw_ostream &print(raw_ostream &OS) const override;
1400 using BlockFrequencyInfoImplBase::dump;
1402 using BlockFrequencyInfoImplBase::printBlockFreq;
1403 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockT *BB) const {
1404 return BlockFrequencyInfoImplBase::printBlockFreq(OS, getNode(BB));
1409 void BlockFrequencyInfoImpl<BT>::doFunction(const FunctionT *F,
1410 const BranchProbabilityInfoT *BPI,
1411 const LoopInfoT *LI) {
1412 // Save the parameters.
1417 // Clean up left-over data structures.
1418 BlockFrequencyInfoImplBase::clear();
1423 DEBUG(dbgs() << "\nblock-frequency: " << F->getName() << "\n================="
1424 << std::string(F->getName().size(), '=') << "\n");
1428 // Visit loops in post-order to find thelocal mass distribution, and then do
1429 // the full function.
1430 computeMassInLoops();
1431 computeMassInFunction();
1436 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeRPOT() {
1437 const BlockT *Entry = F->begin();
1438 RPOT.reserve(F->size());
1439 std::copy(po_begin(Entry), po_end(Entry), std::back_inserter(RPOT));
1440 std::reverse(RPOT.begin(), RPOT.end());
1442 assert(RPOT.size() - 1 <= BlockNode::getMaxIndex() &&
1443 "More nodes in function than Block Frequency Info supports");
1445 DEBUG(dbgs() << "reverse-post-order-traversal\n");
1446 for (rpot_iterator I = rpot_begin(), E = rpot_end(); I != E; ++I) {
1447 BlockNode Node = getNode(I);
1448 DEBUG(dbgs() << " - " << getIndex(I) << ": " << getBlockName(Node) << "\n");
1452 Working.reserve(RPOT.size());
1453 for (size_t Index = 0; Index < RPOT.size(); ++Index)
1454 Working.emplace_back(Index);
1455 Freqs.resize(RPOT.size());
1458 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeLoops() {
1459 DEBUG(dbgs() << "loop-detection\n");
1463 // Visit loops top down and assign them an index.
1464 std::deque<std::pair<const LoopT *, LoopData *>> Q;
1465 for (const LoopT *L : *LI)
1466 Q.emplace_back(L, nullptr);
1467 while (!Q.empty()) {
1468 const LoopT *Loop = Q.front().first;
1469 LoopData *Parent = Q.front().second;
1472 BlockNode Header = getNode(Loop->getHeader());
1473 assert(Header.isValid());
1475 Loops.emplace_back(Parent, Header);
1476 Working[Header.Index].Loop = &Loops.back();
1477 DEBUG(dbgs() << " - loop = " << getBlockName(Header) << "\n");
1479 for (const LoopT *L : *Loop)
1480 Q.emplace_back(L, &Loops.back());
1483 // Visit nodes in reverse post-order and add them to their deepest containing
1485 for (size_t Index = 0; Index < RPOT.size(); ++Index) {
1486 // Loop headers have already been mostly mapped.
1487 if (Working[Index].isLoopHeader()) {
1488 LoopData *ContainingLoop = Working[Index].getContainingLoop();
1490 ContainingLoop->Nodes.push_back(Index);
1494 const LoopT *Loop = LI->getLoopFor(RPOT[Index]);
1498 // Add this node to its containing loop's member list.
1499 BlockNode Header = getNode(Loop->getHeader());
1500 assert(Header.isValid());
1501 const auto &HeaderData = Working[Header.Index];
1502 assert(HeaderData.isLoopHeader());
1504 Working[Index].Loop = HeaderData.Loop;
1505 HeaderData.Loop->Nodes.push_back(Index);
1506 DEBUG(dbgs() << " - loop = " << getBlockName(Header)
1507 << ": member = " << getBlockName(Index) << "\n");
1511 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInLoops() {
1512 // Visit loops with the deepest first, and the top-level loops last.
1513 for (auto L = Loops.rbegin(), E = Loops.rend(); L != E; ++L) {
1514 if (computeMassInLoop(*L))
1516 auto Next = std::next(L);
1517 computeIrreducibleMass(&*L, L.base());
1518 L = std::prev(Next);
1519 if (computeMassInLoop(*L))
1521 llvm_unreachable("unhandled irreducible control flow");
1526 bool BlockFrequencyInfoImpl<BT>::computeMassInLoop(LoopData &Loop) {
1527 // Compute mass in loop.
1528 DEBUG(dbgs() << "compute-mass-in-loop: " << getLoopName(Loop) << "\n");
1530 if (Loop.isIrreducible()) {
1531 BlockMass Remaining = BlockMass::getFull();
1532 for (uint32_t H = 0; H < Loop.NumHeaders; ++H) {
1533 auto &Mass = Working[Loop.Nodes[H].Index].getMass();
1534 Mass = Remaining * BranchProbability(1, Loop.NumHeaders - H);
1537 for (const BlockNode &M : Loop.Nodes)
1538 if (!propagateMassToSuccessors(&Loop, M))
1539 llvm_unreachable("unhandled irreducible control flow");
1541 Working[Loop.getHeader().Index].getMass() = BlockMass::getFull();
1542 if (!propagateMassToSuccessors(&Loop, Loop.getHeader()))
1543 llvm_unreachable("irreducible control flow to loop header!?");
1544 for (const BlockNode &M : Loop.members())
1545 if (!propagateMassToSuccessors(&Loop, M))
1546 // Irreducible backedge.
1550 computeLoopScale(Loop);
1556 bool BlockFrequencyInfoImpl<BT>::tryToComputeMassInFunction() {
1557 // Compute mass in function.
1558 DEBUG(dbgs() << "compute-mass-in-function\n");
1559 assert(!Working.empty() && "no blocks in function");
1560 assert(!Working[0].isLoopHeader() && "entry block is a loop header");
1562 Working[0].getMass() = BlockMass::getFull();
1563 for (rpot_iterator I = rpot_begin(), IE = rpot_end(); I != IE; ++I) {
1564 // Check for nodes that have been packaged.
1565 BlockNode Node = getNode(I);
1566 if (Working[Node.Index].isPackaged())
1569 if (!propagateMassToSuccessors(nullptr, Node))
1575 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInFunction() {
1576 if (tryToComputeMassInFunction())
1578 computeIrreducibleMass(nullptr, Loops.begin());
1579 if (tryToComputeMassInFunction())
1581 llvm_unreachable("unhandled irreducible control flow");
1584 /// \note This should be a lambda, but that crashes GCC 4.7.
1585 namespace bfi_detail {
1586 template <class BT> struct BlockEdgesAdder {
1588 typedef BlockFrequencyInfoImplBase::LoopData LoopData;
1589 typedef GraphTraits<const BlockT *> Successor;
1591 const BlockFrequencyInfoImpl<BT> &BFI;
1592 explicit BlockEdgesAdder(const BlockFrequencyInfoImpl<BT> &BFI)
1594 void operator()(IrreducibleGraph &G, IrreducibleGraph::IrrNode &Irr,
1595 const LoopData *OuterLoop) {
1596 const BlockT *BB = BFI.RPOT[Irr.Node.Index];
1597 for (auto I = Successor::child_begin(BB), E = Successor::child_end(BB);
1599 G.addEdge(Irr, BFI.getNode(*I), OuterLoop);
1604 void BlockFrequencyInfoImpl<BT>::computeIrreducibleMass(
1605 LoopData *OuterLoop, std::list<LoopData>::iterator Insert) {
1606 DEBUG(dbgs() << "analyze-irreducible-in-";
1607 if (OuterLoop) dbgs() << "loop: " << getLoopName(*OuterLoop) << "\n";
1608 else dbgs() << "function\n");
1610 using namespace bfi_detail;
1611 // Ideally, addBlockEdges() would be declared here as a lambda, but that
1613 BlockEdgesAdder<BT> addBlockEdges(*this);
1614 IrreducibleGraph G(*this, OuterLoop, addBlockEdges);
1616 for (auto &L : analyzeIrreducible(G, OuterLoop, Insert))
1617 computeMassInLoop(L);
1621 updateLoopWithIrreducible(*OuterLoop);
1626 BlockFrequencyInfoImpl<BT>::propagateMassToSuccessors(LoopData *OuterLoop,
1627 const BlockNode &Node) {
1628 DEBUG(dbgs() << " - node: " << getBlockName(Node) << "\n");
1629 // Calculate probability for successors.
1631 if (auto *Loop = Working[Node.Index].getPackagedLoop()) {
1632 assert(Loop != OuterLoop && "Cannot propagate mass in a packaged loop");
1633 if (!addLoopSuccessorsToDist(OuterLoop, *Loop, Dist))
1634 // Irreducible backedge.
1637 const BlockT *BB = getBlock(Node);
1638 for (auto SI = Successor::child_begin(BB), SE = Successor::child_end(BB);
1640 // Do not dereference SI, or getEdgeWeight() is linear in the number of
1642 if (!addToDist(Dist, OuterLoop, Node, getNode(*SI),
1643 BPI->getEdgeWeight(BB, SI)))
1644 // Irreducible backedge.
1648 // Distribute mass to successors, saving exit and backedge data in the
1650 distributeMass(Node, OuterLoop, Dist);
1655 raw_ostream &BlockFrequencyInfoImpl<BT>::print(raw_ostream &OS) const {
1658 OS << "block-frequency-info: " << F->getName() << "\n";
1659 for (const BlockT &BB : *F)
1660 OS << " - " << bfi_detail::getBlockName(&BB)
1661 << ": float = " << getFloatingBlockFreq(&BB)
1662 << ", int = " << getBlockFreq(&BB).getFrequency() << "\n";
1664 // Add an extra newline for readability.