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 // UnsignedFloat definition.
38 // TODO: Make this private to BlockFrequencyInfoImpl or delete.
40 //===----------------------------------------------------------------------===//
43 class UnsignedFloatBase {
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 /// UnsignedFloat 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 /// UnsignedFloat 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, UnsignedFloat is portable.
87 /// Unlike APFloat, UnsignedFloat 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 /// UnsignedFloat is totally ordered. However, there is no canonical form, so
92 /// there are multiple representations of most scalars. E.g.:
94 /// UnsignedFloat(8u, 0) == UnsignedFloat(4u, 1)
95 /// UnsignedFloat(4u, 1) == UnsignedFloat(2u, 2)
96 /// UnsignedFloat(2u, 2) == UnsignedFloat(1u, 3)
98 /// UnsignedFloat 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 UnsignedFloat : UnsignedFloatBase {
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 UnsignedFloat() : Digits(0), Exponent(0) {}
134 UnsignedFloat(DigitsType Digits, int16_t Exponent)
135 : Digits(Digits), Exponent(Exponent) {}
138 UnsignedFloat(const std::pair<uint64_t, int16_t> &X)
139 : Digits(X.first), Exponent(X.second) {}
142 static UnsignedFloat getZero() { return UnsignedFloat(0, 0); }
143 static UnsignedFloat getOne() { return UnsignedFloat(1, 0); }
144 static UnsignedFloat getLargest() {
145 return UnsignedFloat(DigitsLimits::max(), MaxExponent);
147 static UnsignedFloat getFloat(uint64_t N) { return adjustToWidth(N, 0); }
148 static UnsignedFloat getInverseFloat(uint64_t N) {
149 return getFloat(N).invert();
151 static UnsignedFloat 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 UnsignedFloat &X) const { return compare(X) == 0; }
192 bool operator<(const UnsignedFloat &X) const { return compare(X) < 0; }
193 bool operator!=(const UnsignedFloat &X) const { return compare(X) != 0; }
194 bool operator>(const UnsignedFloat &X) const { return compare(X) > 0; }
195 bool operator<=(const UnsignedFloat &X) const { return compare(X) <= 0; }
196 bool operator>=(const UnsignedFloat &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 UnsignedFloatBase::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 UnsignedFloatBase::print(OS, Digits, Exponent, Width, Precision);
230 void dump() const { return UnsignedFloatBase::dump(Digits, Exponent, Width); }
232 UnsignedFloat &operator+=(const UnsignedFloat &X);
233 UnsignedFloat &operator-=(const UnsignedFloat &X);
234 UnsignedFloat &operator*=(const UnsignedFloat &X);
235 UnsignedFloat &operator/=(const UnsignedFloat &X);
236 UnsignedFloat &operator<<=(int16_t Shift) { shiftLeft(Shift); return *this; }
237 UnsignedFloat &operator>>=(int16_t Shift) { shiftRight(Shift); return *this; }
240 void shiftLeft(int32_t Shift);
241 void shiftRight(int32_t Shift);
243 /// \brief Adjust two floats to have matching exponents.
245 /// Adjust \c this and \c X to have matching exponents. Returns the new \c X
246 /// by value. Does nothing if \a isZero() for either.
248 /// The value that compares smaller will lose precision, and possibly become
250 UnsignedFloat matchExponents(UnsignedFloat X) {
251 ScaledNumbers::matchScales(Digits, Exponent, X.Digits, X.Exponent);
256 /// \brief Scale a large number accurately.
258 /// Scale N (multiply it by this). Uses full precision multiplication, even
259 /// if Width is smaller than 64, so information is not lost.
260 uint64_t scale(uint64_t N) const;
261 uint64_t scaleByInverse(uint64_t N) const {
262 // TODO: implement directly, rather than relying on inverse. Inverse is
264 return inverse().scale(N);
266 int64_t scale(int64_t N) const {
267 std::pair<uint64_t, bool> Unsigned = splitSigned(N);
268 return joinSigned(scale(Unsigned.first), Unsigned.second);
270 int64_t scaleByInverse(int64_t N) const {
271 std::pair<uint64_t, bool> Unsigned = splitSigned(N);
272 return joinSigned(scaleByInverse(Unsigned.first), Unsigned.second);
275 int compare(const UnsignedFloat &X) const {
276 return ScaledNumbers::compare(Digits, Exponent, X.Digits, X.Exponent);
278 int compareTo(uint64_t N) const {
279 UnsignedFloat Float = getFloat(N);
280 int Compare = compare(Float);
281 if (Width == 64 || Compare != 0)
284 // Check for precision loss. We know *this == RoundTrip.
285 uint64_t RoundTrip = Float.template toInt<uint64_t>();
286 return N == RoundTrip ? 0 : RoundTrip < N ? -1 : 1;
288 int compareTo(int64_t N) const { return N < 0 ? 1 : compareTo(uint64_t(N)); }
290 UnsignedFloat &invert() { return *this = UnsignedFloat::getFloat(1) / *this; }
291 UnsignedFloat inverse() const { return UnsignedFloat(*this).invert(); }
294 static UnsignedFloat getProduct(DigitsType LHS, DigitsType RHS) {
295 return ScaledNumbers::getProduct(LHS, RHS);
297 static UnsignedFloat getQuotient(DigitsType Dividend, DigitsType Divisor) {
298 return ScaledNumbers::getQuotient(Dividend, Divisor);
301 static int countLeadingZerosWidth(DigitsType Digits) {
303 return countLeadingZeros64(Digits);
305 return countLeadingZeros32(Digits);
306 return countLeadingZeros32(Digits) + Width - 32;
309 /// \brief Adjust a number to width, rounding up if necessary.
311 /// Should only be called for \c Shift close to zero.
313 /// \pre Shift >= MinExponent && Shift + 64 <= MaxExponent.
314 static UnsignedFloat adjustToWidth(uint64_t N, int32_t Shift) {
315 assert(Shift >= MinExponent && "Shift should be close to 0");
316 assert(Shift <= MaxExponent - 64 && "Shift should be close to 0");
317 auto Adjusted = ScaledNumbers::getAdjusted<DigitsT>(N, Shift);
321 static UnsignedFloat getRounded(UnsignedFloat P, bool Round) {
326 return ScaledNumbers::getRounded(P.Digits, P.Exponent, Round);
330 #define UNSIGNED_FLOAT_BOP(op, base) \
331 template <class DigitsT> \
332 UnsignedFloat<DigitsT> operator op(const UnsignedFloat<DigitsT> &L, \
333 const UnsignedFloat<DigitsT> &R) { \
334 return UnsignedFloat<DigitsT>(L) base R; \
336 UNSIGNED_FLOAT_BOP(+, += )
337 UNSIGNED_FLOAT_BOP(-, -= )
338 UNSIGNED_FLOAT_BOP(*, *= )
339 UNSIGNED_FLOAT_BOP(/, /= )
340 UNSIGNED_FLOAT_BOP(<<, <<= )
341 UNSIGNED_FLOAT_BOP(>>, >>= )
342 #undef UNSIGNED_FLOAT_BOP
344 template <class DigitsT>
345 raw_ostream &operator<<(raw_ostream &OS, const UnsignedFloat<DigitsT> &X) {
346 return X.print(OS, 10);
349 #define UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, T1, T2) \
350 template <class DigitsT> \
351 bool operator op(const UnsignedFloat<DigitsT> &L, T1 R) { \
352 return L.compareTo(T2(R)) op 0; \
354 template <class DigitsT> \
355 bool operator op(T1 L, const UnsignedFloat<DigitsT> &R) { \
356 return 0 op R.compareTo(T2(L)); \
358 #define UNSIGNED_FLOAT_COMPARE_TO(op) \
359 UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, uint64_t, uint64_t) \
360 UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, uint32_t, uint64_t) \
361 UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, int64_t, int64_t) \
362 UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, int32_t, int64_t)
363 UNSIGNED_FLOAT_COMPARE_TO(< )
364 UNSIGNED_FLOAT_COMPARE_TO(> )
365 UNSIGNED_FLOAT_COMPARE_TO(== )
366 UNSIGNED_FLOAT_COMPARE_TO(!= )
367 UNSIGNED_FLOAT_COMPARE_TO(<= )
368 UNSIGNED_FLOAT_COMPARE_TO(>= )
369 #undef UNSIGNED_FLOAT_COMPARE_TO
370 #undef UNSIGNED_FLOAT_COMPARE_TO_TYPE
372 template <class DigitsT>
373 uint64_t UnsignedFloat<DigitsT>::scale(uint64_t N) const {
374 if (Width == 64 || N <= DigitsLimits::max())
375 return (getFloat(N) * *this).template toInt<uint64_t>();
377 // Defer to the 64-bit version.
378 return UnsignedFloat<uint64_t>(Digits, Exponent).scale(N);
381 template <class DigitsT>
382 template <class IntT>
383 IntT UnsignedFloat<DigitsT>::toInt() const {
384 typedef std::numeric_limits<IntT> Limits;
387 if (*this >= Limits::max())
388 return Limits::max();
392 assert(size_t(Exponent) < sizeof(IntT) * 8);
393 return N << Exponent;
396 assert(size_t(-Exponent) < sizeof(IntT) * 8);
397 return N >> -Exponent;
402 template <class DigitsT>
403 UnsignedFloat<DigitsT> &UnsignedFloat<DigitsT>::
404 operator+=(const UnsignedFloat &X) {
405 if (isLargest() || X.isZero())
407 if (isZero() || X.isLargest())
410 // Normalize exponents.
411 UnsignedFloat Scaled = matchExponents(X);
413 // Check for zero again.
415 return *this = Scaled;
420 DigitsType Sum = Digits + Scaled.Digits;
421 bool DidOverflow = Sum < Digits;
426 if (Exponent == MaxExponent)
427 return *this = getLargest();
430 Digits = UINT64_C(1) << (Width - 1) | Digits >> 1;
434 template <class DigitsT>
435 UnsignedFloat<DigitsT> &UnsignedFloat<DigitsT>::
436 operator-=(const UnsignedFloat &X) {
440 return *this = getZero();
442 // Normalize exponents.
443 UnsignedFloat Scaled = matchExponents(X);
444 assert(Digits >= Scaled.Digits);
446 // Compute difference.
447 if (!Scaled.isZero()) {
448 Digits -= Scaled.Digits;
452 // Check if X just barely lost its last bit. E.g., for 32-bit:
454 // 1*2^32 - 1*2^0 == 0xffffffff != 1*2^32
455 if (*this == UnsignedFloat(1, X.lgFloor() + Width)) {
456 Digits = DigitsType(0) - 1;
461 template <class DigitsT>
462 UnsignedFloat<DigitsT> &UnsignedFloat<DigitsT>::
463 operator*=(const UnsignedFloat &X) {
469 // Save the exponents.
470 int32_t Exponents = int32_t(Exponent) + int32_t(X.Exponent);
472 // Get the raw product.
473 *this = getProduct(Digits, X.Digits);
475 // Combine with exponents.
476 return *this <<= Exponents;
478 template <class DigitsT>
479 UnsignedFloat<DigitsT> &UnsignedFloat<DigitsT>::
480 operator/=(const UnsignedFloat &X) {
484 return *this = getLargest();
486 // Save the exponents.
487 int32_t Exponents = int32_t(Exponent) - int32_t(X.Exponent);
489 // Get the raw quotient.
490 *this = getQuotient(Digits, X.Digits);
492 // Combine with exponents.
493 return *this <<= Exponents;
495 template <class DigitsT>
496 void UnsignedFloat<DigitsT>::shiftLeft(int32_t Shift) {
497 if (!Shift || isZero())
499 assert(Shift != INT32_MIN);
505 // Shift as much as we can in the exponent.
506 int32_t ExponentShift = std::min(Shift, MaxExponent - Exponent);
507 Exponent += ExponentShift;
508 if (ExponentShift == Shift)
511 // Check this late, since it's rare.
515 // Shift the digits themselves.
516 Shift -= ExponentShift;
517 if (Shift > countLeadingZerosWidth(Digits)) {
519 *this = getLargest();
527 template <class DigitsT>
528 void UnsignedFloat<DigitsT>::shiftRight(int32_t Shift) {
529 if (!Shift || isZero())
531 assert(Shift != INT32_MIN);
537 // Shift as much as we can in the exponent.
538 int32_t ExponentShift = std::min(Shift, Exponent - MinExponent);
539 Exponent -= ExponentShift;
540 if (ExponentShift == Shift)
543 // Shift the digits themselves.
544 Shift -= ExponentShift;
545 if (Shift >= Width) {
555 template <class T> struct isPodLike<UnsignedFloat<T>> {
556 static const bool value = true;
560 //===----------------------------------------------------------------------===//
562 // BlockMass definition.
564 // TODO: Make this private to BlockFrequencyInfoImpl or delete.
566 //===----------------------------------------------------------------------===//
569 /// \brief Mass of a block.
571 /// This class implements a sort of fixed-point fraction always between 0.0 and
572 /// 1.0. getMass() == UINT64_MAX indicates a value of 1.0.
574 /// Masses can be added and subtracted. Simple saturation arithmetic is used,
575 /// so arithmetic operations never overflow or underflow.
577 /// Masses can be multiplied. Multiplication treats full mass as 1.0 and uses
578 /// an inexpensive floating-point algorithm that's off-by-one (almost, but not
579 /// quite, maximum precision).
581 /// Masses can be scaled by \a BranchProbability at maximum precision.
586 BlockMass() : Mass(0) {}
587 explicit BlockMass(uint64_t Mass) : Mass(Mass) {}
589 static BlockMass getEmpty() { return BlockMass(); }
590 static BlockMass getFull() { return BlockMass(UINT64_MAX); }
592 uint64_t getMass() const { return Mass; }
594 bool isFull() const { return Mass == UINT64_MAX; }
595 bool isEmpty() const { return !Mass; }
597 bool operator!() const { return isEmpty(); }
599 /// \brief Add another mass.
601 /// Adds another mass, saturating at \a isFull() rather than overflowing.
602 BlockMass &operator+=(const BlockMass &X) {
603 uint64_t Sum = Mass + X.Mass;
604 Mass = Sum < Mass ? UINT64_MAX : Sum;
608 /// \brief Subtract another mass.
610 /// Subtracts another mass, saturating at \a isEmpty() rather than
612 BlockMass &operator-=(const BlockMass &X) {
613 uint64_t Diff = Mass - X.Mass;
614 Mass = Diff > Mass ? 0 : Diff;
618 BlockMass &operator*=(const BranchProbability &P) {
619 Mass = P.scale(Mass);
623 bool operator==(const BlockMass &X) const { return Mass == X.Mass; }
624 bool operator!=(const BlockMass &X) const { return Mass != X.Mass; }
625 bool operator<=(const BlockMass &X) const { return Mass <= X.Mass; }
626 bool operator>=(const BlockMass &X) const { return Mass >= X.Mass; }
627 bool operator<(const BlockMass &X) const { return Mass < X.Mass; }
628 bool operator>(const BlockMass &X) const { return Mass > X.Mass; }
630 /// \brief Convert to floating point.
632 /// Convert to a float. \a isFull() gives 1.0, while \a isEmpty() gives
633 /// slightly above 0.0.
634 UnsignedFloat<uint64_t> toFloat() const;
637 raw_ostream &print(raw_ostream &OS) const;
640 inline BlockMass operator+(const BlockMass &L, const BlockMass &R) {
641 return BlockMass(L) += R;
643 inline BlockMass operator-(const BlockMass &L, const BlockMass &R) {
644 return BlockMass(L) -= R;
646 inline BlockMass operator*(const BlockMass &L, const BranchProbability &R) {
647 return BlockMass(L) *= R;
649 inline BlockMass operator*(const BranchProbability &L, const BlockMass &R) {
650 return BlockMass(R) *= L;
653 inline raw_ostream &operator<<(raw_ostream &OS, const BlockMass &X) {
657 template <> struct isPodLike<BlockMass> {
658 static const bool value = true;
662 //===----------------------------------------------------------------------===//
664 // BlockFrequencyInfoImpl definition.
666 //===----------------------------------------------------------------------===//
670 class BranchProbabilityInfo;
674 class MachineBasicBlock;
675 class MachineBranchProbabilityInfo;
676 class MachineFunction;
678 class MachineLoopInfo;
680 namespace bfi_detail {
681 struct IrreducibleGraph;
683 // This is part of a workaround for a GCC 4.7 crash on lambdas.
684 template <class BT> struct BlockEdgesAdder;
687 /// \brief Base class for BlockFrequencyInfoImpl
689 /// BlockFrequencyInfoImplBase has supporting data structures and some
690 /// algorithms for BlockFrequencyInfoImplBase. Only algorithms that depend on
691 /// the block type (or that call such algorithms) are skipped here.
693 /// Nevertheless, the majority of the overall algorithm documention lives with
694 /// BlockFrequencyInfoImpl. See there for details.
695 class BlockFrequencyInfoImplBase {
697 typedef UnsignedFloat<uint64_t> Float;
699 /// \brief Representative of a block.
701 /// This is a simple wrapper around an index into the reverse-post-order
702 /// traversal of the blocks.
704 /// Unlike a block pointer, its order has meaning (location in the
705 /// topological sort) and it's class is the same regardless of block type.
707 typedef uint32_t IndexType;
710 bool operator==(const BlockNode &X) const { return Index == X.Index; }
711 bool operator!=(const BlockNode &X) const { return Index != X.Index; }
712 bool operator<=(const BlockNode &X) const { return Index <= X.Index; }
713 bool operator>=(const BlockNode &X) const { return Index >= X.Index; }
714 bool operator<(const BlockNode &X) const { return Index < X.Index; }
715 bool operator>(const BlockNode &X) const { return Index > X.Index; }
717 BlockNode() : Index(UINT32_MAX) {}
718 BlockNode(IndexType Index) : Index(Index) {}
720 bool isValid() const { return Index <= getMaxIndex(); }
721 static size_t getMaxIndex() { return UINT32_MAX - 1; }
724 /// \brief Stats about a block itself.
725 struct FrequencyData {
730 /// \brief Data about a loop.
732 /// Contains the data necessary to represent represent a loop as a
733 /// pseudo-node once it's packaged.
735 typedef SmallVector<std::pair<BlockNode, BlockMass>, 4> ExitMap;
736 typedef SmallVector<BlockNode, 4> NodeList;
737 LoopData *Parent; ///< The parent loop.
738 bool IsPackaged; ///< Whether this has been packaged.
739 uint32_t NumHeaders; ///< Number of headers.
740 ExitMap Exits; ///< Successor edges (and weights).
741 NodeList Nodes; ///< Header and the members of the loop.
742 BlockMass BackedgeMass; ///< Mass returned to loop header.
746 LoopData(LoopData *Parent, const BlockNode &Header)
747 : Parent(Parent), IsPackaged(false), NumHeaders(1), Nodes(1, Header) {}
748 template <class It1, class It2>
749 LoopData(LoopData *Parent, It1 FirstHeader, It1 LastHeader, It2 FirstOther,
751 : Parent(Parent), IsPackaged(false), Nodes(FirstHeader, LastHeader) {
752 NumHeaders = Nodes.size();
753 Nodes.insert(Nodes.end(), FirstOther, LastOther);
755 bool isHeader(const BlockNode &Node) const {
757 return std::binary_search(Nodes.begin(), Nodes.begin() + NumHeaders,
759 return Node == Nodes[0];
761 BlockNode getHeader() const { return Nodes[0]; }
762 bool isIrreducible() const { return NumHeaders > 1; }
764 NodeList::const_iterator members_begin() const {
765 return Nodes.begin() + NumHeaders;
767 NodeList::const_iterator members_end() const { return Nodes.end(); }
768 iterator_range<NodeList::const_iterator> members() const {
769 return make_range(members_begin(), members_end());
773 /// \brief Index of loop information.
775 BlockNode Node; ///< This node.
776 LoopData *Loop; ///< The loop this block is inside.
777 BlockMass Mass; ///< Mass distribution from the entry block.
779 WorkingData(const BlockNode &Node) : Node(Node), Loop(nullptr) {}
781 bool isLoopHeader() const { return Loop && Loop->isHeader(Node); }
782 bool isDoubleLoopHeader() const {
783 return isLoopHeader() && Loop->Parent && Loop->Parent->isIrreducible() &&
784 Loop->Parent->isHeader(Node);
787 LoopData *getContainingLoop() const {
790 if (!isDoubleLoopHeader())
792 return Loop->Parent->Parent;
795 /// \brief Resolve a node to its representative.
797 /// Get the node currently representing Node, which could be a containing
800 /// This function should only be called when distributing mass. As long as
801 /// there are no irreducilbe edges to Node, then it will have complexity
802 /// O(1) in this context.
804 /// In general, the complexity is O(L), where L is the number of loop
805 /// headers Node has been packaged into. Since this method is called in
806 /// the context of distributing mass, L will be the number of loop headers
807 /// an early exit edge jumps out of.
808 BlockNode getResolvedNode() const {
809 auto L = getPackagedLoop();
810 return L ? L->getHeader() : Node;
812 LoopData *getPackagedLoop() const {
813 if (!Loop || !Loop->IsPackaged)
816 while (L->Parent && L->Parent->IsPackaged)
821 /// \brief Get the appropriate mass for a node.
823 /// Get appropriate mass for Node. If Node is a loop-header (whose loop
824 /// has been packaged), returns the mass of its pseudo-node. If it's a
825 /// node inside a packaged loop, it returns the loop's mass.
826 BlockMass &getMass() {
829 if (!isADoublePackage())
831 return Loop->Parent->Mass;
834 /// \brief Has ContainingLoop been packaged up?
835 bool isPackaged() const { return getResolvedNode() != Node; }
836 /// \brief Has Loop been packaged up?
837 bool isAPackage() const { return isLoopHeader() && Loop->IsPackaged; }
838 /// \brief Has Loop been packaged up twice?
839 bool isADoublePackage() const {
840 return isDoubleLoopHeader() && Loop->Parent->IsPackaged;
844 /// \brief Unscaled probability weight.
846 /// Probability weight for an edge in the graph (including the
847 /// successor/target node).
849 /// All edges in the original function are 32-bit. However, exit edges from
850 /// loop packages are taken from 64-bit exit masses, so we need 64-bits of
851 /// space in general.
853 /// In addition to the raw weight amount, Weight stores the type of the edge
854 /// in the current context (i.e., the context of the loop being processed).
855 /// Is this a local edge within the loop, an exit from the loop, or a
856 /// backedge to the loop header?
858 enum DistType { Local, Exit, Backedge };
860 BlockNode TargetNode;
862 Weight() : Type(Local), Amount(0) {}
865 /// \brief Distribution of unscaled probability weight.
867 /// Distribution of unscaled probability weight to a set of successors.
869 /// This class collates the successor edge weights for later processing.
871 /// \a DidOverflow indicates whether \a Total did overflow while adding to
872 /// the distribution. It should never overflow twice.
873 struct Distribution {
874 typedef SmallVector<Weight, 4> WeightList;
875 WeightList Weights; ///< Individual successor weights.
876 uint64_t Total; ///< Sum of all weights.
877 bool DidOverflow; ///< Whether \a Total did overflow.
879 Distribution() : Total(0), DidOverflow(false) {}
880 void addLocal(const BlockNode &Node, uint64_t Amount) {
881 add(Node, Amount, Weight::Local);
883 void addExit(const BlockNode &Node, uint64_t Amount) {
884 add(Node, Amount, Weight::Exit);
886 void addBackedge(const BlockNode &Node, uint64_t Amount) {
887 add(Node, Amount, Weight::Backedge);
890 /// \brief Normalize the distribution.
892 /// Combines multiple edges to the same \a Weight::TargetNode and scales
893 /// down so that \a Total fits into 32-bits.
895 /// This is linear in the size of \a Weights. For the vast majority of
896 /// cases, adjacent edge weights are combined by sorting WeightList and
897 /// combining adjacent weights. However, for very large edge lists an
898 /// auxiliary hash table is used.
902 void add(const BlockNode &Node, uint64_t Amount, Weight::DistType Type);
905 /// \brief Data about each block. This is used downstream.
906 std::vector<FrequencyData> Freqs;
908 /// \brief Loop data: see initializeLoops().
909 std::vector<WorkingData> Working;
911 /// \brief Indexed information about loops.
912 std::list<LoopData> Loops;
914 /// \brief Add all edges out of a packaged loop to the distribution.
916 /// Adds all edges from LocalLoopHead to Dist. Calls addToDist() to add each
919 /// \return \c true unless there's an irreducible backedge.
920 bool addLoopSuccessorsToDist(const LoopData *OuterLoop, LoopData &Loop,
923 /// \brief Add an edge to the distribution.
925 /// Adds an edge to Succ to Dist. If \c LoopHead.isValid(), then whether the
926 /// edge is local/exit/backedge is in the context of LoopHead. Otherwise,
927 /// every edge should be a local edge (since all the loops are packaged up).
929 /// \return \c true unless aborted due to an irreducible backedge.
930 bool addToDist(Distribution &Dist, const LoopData *OuterLoop,
931 const BlockNode &Pred, const BlockNode &Succ, uint64_t Weight);
933 LoopData &getLoopPackage(const BlockNode &Head) {
934 assert(Head.Index < Working.size());
935 assert(Working[Head.Index].isLoopHeader());
936 return *Working[Head.Index].Loop;
939 /// \brief Analyze irreducible SCCs.
941 /// Separate irreducible SCCs from \c G, which is an explict graph of \c
942 /// OuterLoop (or the top-level function, if \c OuterLoop is \c nullptr).
943 /// Insert them into \a Loops before \c Insert.
945 /// \return the \c LoopData nodes representing the irreducible SCCs.
946 iterator_range<std::list<LoopData>::iterator>
947 analyzeIrreducible(const bfi_detail::IrreducibleGraph &G, LoopData *OuterLoop,
948 std::list<LoopData>::iterator Insert);
950 /// \brief Update a loop after packaging irreducible SCCs inside of it.
952 /// Update \c OuterLoop. Before finding irreducible control flow, it was
953 /// partway through \a computeMassInLoop(), so \a LoopData::Exits and \a
954 /// LoopData::BackedgeMass need to be reset. Also, nodes that were packaged
955 /// up need to be removed from \a OuterLoop::Nodes.
956 void updateLoopWithIrreducible(LoopData &OuterLoop);
958 /// \brief Distribute mass according to a distribution.
960 /// Distributes the mass in Source according to Dist. If LoopHead.isValid(),
961 /// backedges and exits are stored in its entry in Loops.
963 /// Mass is distributed in parallel from two copies of the source mass.
964 void distributeMass(const BlockNode &Source, LoopData *OuterLoop,
967 /// \brief Compute the loop scale for a loop.
968 void computeLoopScale(LoopData &Loop);
970 /// \brief Package up a loop.
971 void packageLoop(LoopData &Loop);
973 /// \brief Unwrap loops.
976 /// \brief Finalize frequency metrics.
978 /// Calculates final frequencies and cleans up no-longer-needed data
980 void finalizeMetrics();
982 /// \brief Clear all memory.
985 virtual std::string getBlockName(const BlockNode &Node) const;
986 std::string getLoopName(const LoopData &Loop) const;
988 virtual raw_ostream &print(raw_ostream &OS) const { return OS; }
989 void dump() const { print(dbgs()); }
991 Float getFloatingBlockFreq(const BlockNode &Node) const;
993 BlockFrequency getBlockFreq(const BlockNode &Node) const;
995 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockNode &Node) const;
996 raw_ostream &printBlockFreq(raw_ostream &OS,
997 const BlockFrequency &Freq) const;
999 uint64_t getEntryFreq() const {
1000 assert(!Freqs.empty());
1001 return Freqs[0].Integer;
1003 /// \brief Virtual destructor.
1005 /// Need a virtual destructor to mask the compiler warning about
1007 virtual ~BlockFrequencyInfoImplBase() {}
1010 namespace bfi_detail {
1011 template <class BlockT> struct TypeMap {};
1012 template <> struct TypeMap<BasicBlock> {
1013 typedef BasicBlock BlockT;
1014 typedef Function FunctionT;
1015 typedef BranchProbabilityInfo BranchProbabilityInfoT;
1017 typedef LoopInfo LoopInfoT;
1019 template <> struct TypeMap<MachineBasicBlock> {
1020 typedef MachineBasicBlock BlockT;
1021 typedef MachineFunction FunctionT;
1022 typedef MachineBranchProbabilityInfo BranchProbabilityInfoT;
1023 typedef MachineLoop LoopT;
1024 typedef MachineLoopInfo LoopInfoT;
1027 /// \brief Get the name of a MachineBasicBlock.
1029 /// Get the name of a MachineBasicBlock. It's templated so that including from
1030 /// CodeGen is unnecessary (that would be a layering issue).
1032 /// This is used mainly for debug output. The name is similar to
1033 /// MachineBasicBlock::getFullName(), but skips the name of the function.
1034 template <class BlockT> std::string getBlockName(const BlockT *BB) {
1035 assert(BB && "Unexpected nullptr");
1036 auto MachineName = "BB" + Twine(BB->getNumber());
1037 if (BB->getBasicBlock())
1038 return (MachineName + "[" + BB->getName() + "]").str();
1039 return MachineName.str();
1041 /// \brief Get the name of a BasicBlock.
1042 template <> inline std::string getBlockName(const BasicBlock *BB) {
1043 assert(BB && "Unexpected nullptr");
1044 return BB->getName().str();
1047 /// \brief Graph of irreducible control flow.
1049 /// This graph is used for determining the SCCs in a loop (or top-level
1050 /// function) that has irreducible control flow.
1052 /// During the block frequency algorithm, the local graphs are defined in a
1053 /// light-weight way, deferring to the \a BasicBlock or \a MachineBasicBlock
1054 /// graphs for most edges, but getting others from \a LoopData::ExitMap. The
1055 /// latter only has successor information.
1057 /// \a IrreducibleGraph makes this graph explicit. It's in a form that can use
1058 /// \a GraphTraits (so that \a analyzeIrreducible() can use \a scc_iterator),
1059 /// and it explicitly lists predecessors and successors. The initialization
1060 /// that relies on \c MachineBasicBlock is defined in the header.
1061 struct IrreducibleGraph {
1062 typedef BlockFrequencyInfoImplBase BFIBase;
1066 typedef BFIBase::BlockNode BlockNode;
1070 std::deque<const IrrNode *> Edges;
1071 IrrNode(const BlockNode &Node) : Node(Node), NumIn(0) {}
1073 typedef std::deque<const IrrNode *>::const_iterator iterator;
1074 iterator pred_begin() const { return Edges.begin(); }
1075 iterator succ_begin() const { return Edges.begin() + NumIn; }
1076 iterator pred_end() const { return succ_begin(); }
1077 iterator succ_end() const { return Edges.end(); }
1080 const IrrNode *StartIrr;
1081 std::vector<IrrNode> Nodes;
1082 SmallDenseMap<uint32_t, IrrNode *, 4> Lookup;
1084 /// \brief Construct an explicit graph containing irreducible control flow.
1086 /// Construct an explicit graph of the control flow in \c OuterLoop (or the
1087 /// top-level function, if \c OuterLoop is \c nullptr). Uses \c
1088 /// addBlockEdges to add block successors that have not been packaged into
1091 /// \a BlockFrequencyInfoImpl::computeIrreducibleMass() is the only expected
1093 template <class BlockEdgesAdder>
1094 IrreducibleGraph(BFIBase &BFI, const BFIBase::LoopData *OuterLoop,
1095 BlockEdgesAdder addBlockEdges)
1096 : BFI(BFI), StartIrr(nullptr) {
1097 initialize(OuterLoop, addBlockEdges);
1100 template <class BlockEdgesAdder>
1101 void initialize(const BFIBase::LoopData *OuterLoop,
1102 BlockEdgesAdder addBlockEdges);
1103 void addNodesInLoop(const BFIBase::LoopData &OuterLoop);
1104 void addNodesInFunction();
1105 void addNode(const BlockNode &Node) {
1106 Nodes.emplace_back(Node);
1107 BFI.Working[Node.Index].getMass() = BlockMass::getEmpty();
1110 template <class BlockEdgesAdder>
1111 void addEdges(const BlockNode &Node, const BFIBase::LoopData *OuterLoop,
1112 BlockEdgesAdder addBlockEdges);
1113 void addEdge(IrrNode &Irr, const BlockNode &Succ,
1114 const BFIBase::LoopData *OuterLoop);
1116 template <class BlockEdgesAdder>
1117 void IrreducibleGraph::initialize(const BFIBase::LoopData *OuterLoop,
1118 BlockEdgesAdder addBlockEdges) {
1120 addNodesInLoop(*OuterLoop);
1121 for (auto N : OuterLoop->Nodes)
1122 addEdges(N, OuterLoop, addBlockEdges);
1124 addNodesInFunction();
1125 for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index)
1126 addEdges(Index, OuterLoop, addBlockEdges);
1128 StartIrr = Lookup[Start.Index];
1130 template <class BlockEdgesAdder>
1131 void IrreducibleGraph::addEdges(const BlockNode &Node,
1132 const BFIBase::LoopData *OuterLoop,
1133 BlockEdgesAdder addBlockEdges) {
1134 auto L = Lookup.find(Node.Index);
1135 if (L == Lookup.end())
1137 IrrNode &Irr = *L->second;
1138 const auto &Working = BFI.Working[Node.Index];
1140 if (Working.isAPackage())
1141 for (const auto &I : Working.Loop->Exits)
1142 addEdge(Irr, I.first, OuterLoop);
1144 addBlockEdges(*this, Irr, OuterLoop);
1148 /// \brief Shared implementation for block frequency analysis.
1150 /// This is a shared implementation of BlockFrequencyInfo and
1151 /// MachineBlockFrequencyInfo, and calculates the relative frequencies of
1154 /// LoopInfo defines a loop as a "non-trivial" SCC dominated by a single block,
1155 /// which is called the header. A given loop, L, can have sub-loops, which are
1156 /// loops within the subgraph of L that exclude its header. (A "trivial" SCC
1157 /// consists of a single block that does not have a self-edge.)
1159 /// In addition to loops, this algorithm has limited support for irreducible
1160 /// SCCs, which are SCCs with multiple entry blocks. Irreducible SCCs are
1161 /// discovered on they fly, and modelled as loops with multiple headers.
1163 /// The headers of irreducible sub-SCCs consist of its entry blocks and all
1164 /// nodes that are targets of a backedge within it (excluding backedges within
1165 /// true sub-loops). Block frequency calculations act as if a block is
1166 /// inserted that intercepts all the edges to the headers. All backedges and
1167 /// entries point to this block. Its successors are the headers, which split
1168 /// the frequency evenly.
1170 /// This algorithm leverages BlockMass and UnsignedFloat to maintain precision,
1171 /// separates mass distribution from loop scaling, and dithers to eliminate
1172 /// probability mass loss.
1174 /// The implementation is split between BlockFrequencyInfoImpl, which knows the
1175 /// type of graph being modelled (BasicBlock vs. MachineBasicBlock), and
1176 /// BlockFrequencyInfoImplBase, which doesn't. The base class uses \a
1177 /// BlockNode, a wrapper around a uint32_t. BlockNode is numbered from 0 in
1178 /// reverse-post order. This gives two advantages: it's easy to compare the
1179 /// relative ordering of two nodes, and maps keyed on BlockT can be represented
1182 /// This algorithm is O(V+E), unless there is irreducible control flow, in
1183 /// which case it's O(V*E) in the worst case.
1185 /// These are the main stages:
1187 /// 0. Reverse post-order traversal (\a initializeRPOT()).
1189 /// Run a single post-order traversal and save it (in reverse) in RPOT.
1190 /// All other stages make use of this ordering. Save a lookup from BlockT
1191 /// to BlockNode (the index into RPOT) in Nodes.
1193 /// 1. Loop initialization (\a initializeLoops()).
1195 /// Translate LoopInfo/MachineLoopInfo into a form suitable for the rest of
1196 /// the algorithm. In particular, store the immediate members of each loop
1197 /// in reverse post-order.
1199 /// 2. Calculate mass and scale in loops (\a computeMassInLoops()).
1201 /// For each loop (bottom-up), distribute mass through the DAG resulting
1202 /// from ignoring backedges and treating sub-loops as a single pseudo-node.
1203 /// Track the backedge mass distributed to the loop header, and use it to
1204 /// calculate the loop scale (number of loop iterations). Immediate
1205 /// members that represent sub-loops will already have been visited and
1206 /// packaged into a pseudo-node.
1208 /// Distributing mass in a loop is a reverse-post-order traversal through
1209 /// the loop. Start by assigning full mass to the Loop header. For each
1210 /// node in the loop:
1212 /// - Fetch and categorize the weight distribution for its successors.
1213 /// If this is a packaged-subloop, the weight distribution is stored
1214 /// in \a LoopData::Exits. Otherwise, fetch it from
1215 /// BranchProbabilityInfo.
1217 /// - Each successor is categorized as \a Weight::Local, a local edge
1218 /// within the current loop, \a Weight::Backedge, a backedge to the
1219 /// loop header, or \a Weight::Exit, any successor outside the loop.
1220 /// The weight, the successor, and its category are stored in \a
1221 /// Distribution. There can be multiple edges to each successor.
1223 /// - If there's a backedge to a non-header, there's an irreducible SCC.
1224 /// The usual flow is temporarily aborted. \a
1225 /// computeIrreducibleMass() finds the irreducible SCCs within the
1226 /// loop, packages them up, and restarts the flow.
1228 /// - Normalize the distribution: scale weights down so that their sum
1229 /// is 32-bits, and coalesce multiple edges to the same node.
1231 /// - Distribute the mass accordingly, dithering to minimize mass loss,
1232 /// as described in \a distributeMass().
1234 /// Finally, calculate the loop scale from the accumulated backedge mass.
1236 /// 3. Distribute mass in the function (\a computeMassInFunction()).
1238 /// Finally, distribute mass through the DAG resulting from packaging all
1239 /// loops in the function. This uses the same algorithm as distributing
1240 /// mass in a loop, except that there are no exit or backedge edges.
1242 /// 4. Unpackage loops (\a unwrapLoops()).
1244 /// Initialize each block's frequency to a floating point representation of
1247 /// Visit loops top-down, scaling the frequencies of its immediate members
1248 /// by the loop's pseudo-node's frequency.
1250 /// 5. Convert frequencies to a 64-bit range (\a finalizeMetrics()).
1252 /// Using the min and max frequencies as a guide, translate floating point
1253 /// frequencies to an appropriate range in uint64_t.
1255 /// It has some known flaws.
1257 /// - Loop scale is limited to 4096 per loop (2^12) to avoid exhausting
1258 /// BlockFrequency's 64-bit integer precision.
1260 /// - The model of irreducible control flow is a rough approximation.
1262 /// Modelling irreducible control flow exactly involves setting up and
1263 /// solving a group of infinite geometric series. Such precision is
1264 /// unlikely to be worthwhile, since most of our algorithms give up on
1265 /// irreducible control flow anyway.
1267 /// Nevertheless, we might find that we need to get closer. Here's a sort
1268 /// of TODO list for the model with diminishing returns, to be completed as
1271 /// - The headers for the \a LoopData representing an irreducible SCC
1272 /// include non-entry blocks. When these extra blocks exist, they
1273 /// indicate a self-contained irreducible sub-SCC. We could treat them
1274 /// as sub-loops, rather than arbitrarily shoving the problematic
1275 /// blocks into the headers of the main irreducible SCC.
1277 /// - Backedge frequencies are assumed to be evenly split between the
1278 /// headers of a given irreducible SCC. Instead, we could track the
1279 /// backedge mass separately for each header, and adjust their relative
1282 /// - Entry frequencies are assumed to be evenly split between the
1283 /// headers of a given irreducible SCC, which is the only option if we
1284 /// need to compute mass in the SCC before its parent loop. Instead,
1285 /// we could partially compute mass in the parent loop, and stop when
1286 /// we get to the SCC. Here, we have the correct ratio of entry
1287 /// masses, which we can use to adjust their relative frequencies.
1288 /// Compute mass in the SCC, and then continue propagation in the
1291 /// - We can propagate mass iteratively through the SCC, for some fixed
1292 /// number of iterations. Each iteration starts by assigning the entry
1293 /// blocks their backedge mass from the prior iteration. The final
1294 /// mass for each block (and each exit, and the total backedge mass
1295 /// used for computing loop scale) is the sum of all iterations.
1296 /// (Running this until fixed point would "solve" the geometric
1297 /// series by simulation.)
1298 template <class BT> class BlockFrequencyInfoImpl : BlockFrequencyInfoImplBase {
1299 typedef typename bfi_detail::TypeMap<BT>::BlockT BlockT;
1300 typedef typename bfi_detail::TypeMap<BT>::FunctionT FunctionT;
1301 typedef typename bfi_detail::TypeMap<BT>::BranchProbabilityInfoT
1302 BranchProbabilityInfoT;
1303 typedef typename bfi_detail::TypeMap<BT>::LoopT LoopT;
1304 typedef typename bfi_detail::TypeMap<BT>::LoopInfoT LoopInfoT;
1306 // This is part of a workaround for a GCC 4.7 crash on lambdas.
1307 friend struct bfi_detail::BlockEdgesAdder<BT>;
1309 typedef GraphTraits<const BlockT *> Successor;
1310 typedef GraphTraits<Inverse<const BlockT *>> Predecessor;
1312 const BranchProbabilityInfoT *BPI;
1313 const LoopInfoT *LI;
1316 // All blocks in reverse postorder.
1317 std::vector<const BlockT *> RPOT;
1318 DenseMap<const BlockT *, BlockNode> Nodes;
1320 typedef typename std::vector<const BlockT *>::const_iterator rpot_iterator;
1322 rpot_iterator rpot_begin() const { return RPOT.begin(); }
1323 rpot_iterator rpot_end() const { return RPOT.end(); }
1325 size_t getIndex(const rpot_iterator &I) const { return I - rpot_begin(); }
1327 BlockNode getNode(const rpot_iterator &I) const {
1328 return BlockNode(getIndex(I));
1330 BlockNode getNode(const BlockT *BB) const { return Nodes.lookup(BB); }
1332 const BlockT *getBlock(const BlockNode &Node) const {
1333 assert(Node.Index < RPOT.size());
1334 return RPOT[Node.Index];
1337 /// \brief Run (and save) a post-order traversal.
1339 /// Saves a reverse post-order traversal of all the nodes in \a F.
1340 void initializeRPOT();
1342 /// \brief Initialize loop data.
1344 /// Build up \a Loops using \a LoopInfo. \a LoopInfo gives us a mapping from
1345 /// each block to the deepest loop it's in, but we need the inverse. For each
1346 /// loop, we store in reverse post-order its "immediate" members, defined as
1347 /// the header, the headers of immediate sub-loops, and all other blocks in
1348 /// the loop that are not in sub-loops.
1349 void initializeLoops();
1351 /// \brief Propagate to a block's successors.
1353 /// In the context of distributing mass through \c OuterLoop, divide the mass
1354 /// currently assigned to \c Node between its successors.
1356 /// \return \c true unless there's an irreducible backedge.
1357 bool propagateMassToSuccessors(LoopData *OuterLoop, const BlockNode &Node);
1359 /// \brief Compute mass in a particular loop.
1361 /// Assign mass to \c Loop's header, and then for each block in \c Loop in
1362 /// reverse post-order, distribute mass to its successors. Only visits nodes
1363 /// that have not been packaged into sub-loops.
1365 /// \pre \a computeMassInLoop() has been called for each subloop of \c Loop.
1366 /// \return \c true unless there's an irreducible backedge.
1367 bool computeMassInLoop(LoopData &Loop);
1369 /// \brief Try to compute mass in the top-level function.
1371 /// Assign mass to the entry block, and then for each block in reverse
1372 /// post-order, distribute mass to its successors. Skips nodes that have
1373 /// been packaged into loops.
1375 /// \pre \a computeMassInLoops() has been called.
1376 /// \return \c true unless there's an irreducible backedge.
1377 bool tryToComputeMassInFunction();
1379 /// \brief Compute mass in (and package up) irreducible SCCs.
1381 /// Find the irreducible SCCs in \c OuterLoop, add them to \a Loops (in front
1382 /// of \c Insert), and call \a computeMassInLoop() on each of them.
1384 /// If \c OuterLoop is \c nullptr, it refers to the top-level function.
1386 /// \pre \a computeMassInLoop() has been called for each subloop of \c
1388 /// \pre \c Insert points at the the last loop successfully processed by \a
1389 /// computeMassInLoop().
1390 /// \pre \c OuterLoop has irreducible SCCs.
1391 void computeIrreducibleMass(LoopData *OuterLoop,
1392 std::list<LoopData>::iterator Insert);
1394 /// \brief Compute mass in all loops.
1396 /// For each loop bottom-up, call \a computeMassInLoop().
1398 /// \a computeMassInLoop() aborts (and returns \c false) on loops that
1399 /// contain a irreducible sub-SCCs. Use \a computeIrreducibleMass() and then
1400 /// re-enter \a computeMassInLoop().
1402 /// \post \a computeMassInLoop() has returned \c true for every loop.
1403 void computeMassInLoops();
1405 /// \brief Compute mass in the top-level function.
1407 /// Uses \a tryToComputeMassInFunction() and \a computeIrreducibleMass() to
1408 /// compute mass in the top-level function.
1410 /// \post \a tryToComputeMassInFunction() has returned \c true.
1411 void computeMassInFunction();
1413 std::string getBlockName(const BlockNode &Node) const override {
1414 return bfi_detail::getBlockName(getBlock(Node));
1418 const FunctionT *getFunction() const { return F; }
1420 void doFunction(const FunctionT *F, const BranchProbabilityInfoT *BPI,
1421 const LoopInfoT *LI);
1422 BlockFrequencyInfoImpl() : BPI(nullptr), LI(nullptr), F(nullptr) {}
1424 using BlockFrequencyInfoImplBase::getEntryFreq;
1425 BlockFrequency getBlockFreq(const BlockT *BB) const {
1426 return BlockFrequencyInfoImplBase::getBlockFreq(getNode(BB));
1428 Float getFloatingBlockFreq(const BlockT *BB) const {
1429 return BlockFrequencyInfoImplBase::getFloatingBlockFreq(getNode(BB));
1432 /// \brief Print the frequencies for the current function.
1434 /// Prints the frequencies for the blocks in the current function.
1436 /// Blocks are printed in the natural iteration order of the function, rather
1437 /// than reverse post-order. This provides two advantages: writing -analyze
1438 /// tests is easier (since blocks come out in source order), and even
1439 /// unreachable blocks are printed.
1441 /// \a BlockFrequencyInfoImplBase::print() only knows reverse post-order, so
1442 /// we need to override it here.
1443 raw_ostream &print(raw_ostream &OS) const override;
1444 using BlockFrequencyInfoImplBase::dump;
1446 using BlockFrequencyInfoImplBase::printBlockFreq;
1447 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockT *BB) const {
1448 return BlockFrequencyInfoImplBase::printBlockFreq(OS, getNode(BB));
1453 void BlockFrequencyInfoImpl<BT>::doFunction(const FunctionT *F,
1454 const BranchProbabilityInfoT *BPI,
1455 const LoopInfoT *LI) {
1456 // Save the parameters.
1461 // Clean up left-over data structures.
1462 BlockFrequencyInfoImplBase::clear();
1467 DEBUG(dbgs() << "\nblock-frequency: " << F->getName() << "\n================="
1468 << std::string(F->getName().size(), '=') << "\n");
1472 // Visit loops in post-order to find thelocal mass distribution, and then do
1473 // the full function.
1474 computeMassInLoops();
1475 computeMassInFunction();
1480 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeRPOT() {
1481 const BlockT *Entry = F->begin();
1482 RPOT.reserve(F->size());
1483 std::copy(po_begin(Entry), po_end(Entry), std::back_inserter(RPOT));
1484 std::reverse(RPOT.begin(), RPOT.end());
1486 assert(RPOT.size() - 1 <= BlockNode::getMaxIndex() &&
1487 "More nodes in function than Block Frequency Info supports");
1489 DEBUG(dbgs() << "reverse-post-order-traversal\n");
1490 for (rpot_iterator I = rpot_begin(), E = rpot_end(); I != E; ++I) {
1491 BlockNode Node = getNode(I);
1492 DEBUG(dbgs() << " - " << getIndex(I) << ": " << getBlockName(Node) << "\n");
1496 Working.reserve(RPOT.size());
1497 for (size_t Index = 0; Index < RPOT.size(); ++Index)
1498 Working.emplace_back(Index);
1499 Freqs.resize(RPOT.size());
1502 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeLoops() {
1503 DEBUG(dbgs() << "loop-detection\n");
1507 // Visit loops top down and assign them an index.
1508 std::deque<std::pair<const LoopT *, LoopData *>> Q;
1509 for (const LoopT *L : *LI)
1510 Q.emplace_back(L, nullptr);
1511 while (!Q.empty()) {
1512 const LoopT *Loop = Q.front().first;
1513 LoopData *Parent = Q.front().second;
1516 BlockNode Header = getNode(Loop->getHeader());
1517 assert(Header.isValid());
1519 Loops.emplace_back(Parent, Header);
1520 Working[Header.Index].Loop = &Loops.back();
1521 DEBUG(dbgs() << " - loop = " << getBlockName(Header) << "\n");
1523 for (const LoopT *L : *Loop)
1524 Q.emplace_back(L, &Loops.back());
1527 // Visit nodes in reverse post-order and add them to their deepest containing
1529 for (size_t Index = 0; Index < RPOT.size(); ++Index) {
1530 // Loop headers have already been mostly mapped.
1531 if (Working[Index].isLoopHeader()) {
1532 LoopData *ContainingLoop = Working[Index].getContainingLoop();
1534 ContainingLoop->Nodes.push_back(Index);
1538 const LoopT *Loop = LI->getLoopFor(RPOT[Index]);
1542 // Add this node to its containing loop's member list.
1543 BlockNode Header = getNode(Loop->getHeader());
1544 assert(Header.isValid());
1545 const auto &HeaderData = Working[Header.Index];
1546 assert(HeaderData.isLoopHeader());
1548 Working[Index].Loop = HeaderData.Loop;
1549 HeaderData.Loop->Nodes.push_back(Index);
1550 DEBUG(dbgs() << " - loop = " << getBlockName(Header)
1551 << ": member = " << getBlockName(Index) << "\n");
1555 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInLoops() {
1556 // Visit loops with the deepest first, and the top-level loops last.
1557 for (auto L = Loops.rbegin(), E = Loops.rend(); L != E; ++L) {
1558 if (computeMassInLoop(*L))
1560 auto Next = std::next(L);
1561 computeIrreducibleMass(&*L, L.base());
1562 L = std::prev(Next);
1563 if (computeMassInLoop(*L))
1565 llvm_unreachable("unhandled irreducible control flow");
1570 bool BlockFrequencyInfoImpl<BT>::computeMassInLoop(LoopData &Loop) {
1571 // Compute mass in loop.
1572 DEBUG(dbgs() << "compute-mass-in-loop: " << getLoopName(Loop) << "\n");
1574 if (Loop.isIrreducible()) {
1575 BlockMass Remaining = BlockMass::getFull();
1576 for (uint32_t H = 0; H < Loop.NumHeaders; ++H) {
1577 auto &Mass = Working[Loop.Nodes[H].Index].getMass();
1578 Mass = Remaining * BranchProbability(1, Loop.NumHeaders - H);
1581 for (const BlockNode &M : Loop.Nodes)
1582 if (!propagateMassToSuccessors(&Loop, M))
1583 llvm_unreachable("unhandled irreducible control flow");
1585 Working[Loop.getHeader().Index].getMass() = BlockMass::getFull();
1586 if (!propagateMassToSuccessors(&Loop, Loop.getHeader()))
1587 llvm_unreachable("irreducible control flow to loop header!?");
1588 for (const BlockNode &M : Loop.members())
1589 if (!propagateMassToSuccessors(&Loop, M))
1590 // Irreducible backedge.
1594 computeLoopScale(Loop);
1600 bool BlockFrequencyInfoImpl<BT>::tryToComputeMassInFunction() {
1601 // Compute mass in function.
1602 DEBUG(dbgs() << "compute-mass-in-function\n");
1603 assert(!Working.empty() && "no blocks in function");
1604 assert(!Working[0].isLoopHeader() && "entry block is a loop header");
1606 Working[0].getMass() = BlockMass::getFull();
1607 for (rpot_iterator I = rpot_begin(), IE = rpot_end(); I != IE; ++I) {
1608 // Check for nodes that have been packaged.
1609 BlockNode Node = getNode(I);
1610 if (Working[Node.Index].isPackaged())
1613 if (!propagateMassToSuccessors(nullptr, Node))
1619 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInFunction() {
1620 if (tryToComputeMassInFunction())
1622 computeIrreducibleMass(nullptr, Loops.begin());
1623 if (tryToComputeMassInFunction())
1625 llvm_unreachable("unhandled irreducible control flow");
1628 /// \note This should be a lambda, but that crashes GCC 4.7.
1629 namespace bfi_detail {
1630 template <class BT> struct BlockEdgesAdder {
1632 typedef BlockFrequencyInfoImplBase::LoopData LoopData;
1633 typedef GraphTraits<const BlockT *> Successor;
1635 const BlockFrequencyInfoImpl<BT> &BFI;
1636 explicit BlockEdgesAdder(const BlockFrequencyInfoImpl<BT> &BFI)
1638 void operator()(IrreducibleGraph &G, IrreducibleGraph::IrrNode &Irr,
1639 const LoopData *OuterLoop) {
1640 const BlockT *BB = BFI.RPOT[Irr.Node.Index];
1641 for (auto I = Successor::child_begin(BB), E = Successor::child_end(BB);
1643 G.addEdge(Irr, BFI.getNode(*I), OuterLoop);
1648 void BlockFrequencyInfoImpl<BT>::computeIrreducibleMass(
1649 LoopData *OuterLoop, std::list<LoopData>::iterator Insert) {
1650 DEBUG(dbgs() << "analyze-irreducible-in-";
1651 if (OuterLoop) dbgs() << "loop: " << getLoopName(*OuterLoop) << "\n";
1652 else dbgs() << "function\n");
1654 using namespace bfi_detail;
1655 // Ideally, addBlockEdges() would be declared here as a lambda, but that
1657 BlockEdgesAdder<BT> addBlockEdges(*this);
1658 IrreducibleGraph G(*this, OuterLoop, addBlockEdges);
1660 for (auto &L : analyzeIrreducible(G, OuterLoop, Insert))
1661 computeMassInLoop(L);
1665 updateLoopWithIrreducible(*OuterLoop);
1670 BlockFrequencyInfoImpl<BT>::propagateMassToSuccessors(LoopData *OuterLoop,
1671 const BlockNode &Node) {
1672 DEBUG(dbgs() << " - node: " << getBlockName(Node) << "\n");
1673 // Calculate probability for successors.
1675 if (auto *Loop = Working[Node.Index].getPackagedLoop()) {
1676 assert(Loop != OuterLoop && "Cannot propagate mass in a packaged loop");
1677 if (!addLoopSuccessorsToDist(OuterLoop, *Loop, Dist))
1678 // Irreducible backedge.
1681 const BlockT *BB = getBlock(Node);
1682 for (auto SI = Successor::child_begin(BB), SE = Successor::child_end(BB);
1684 // Do not dereference SI, or getEdgeWeight() is linear in the number of
1686 if (!addToDist(Dist, OuterLoop, Node, getNode(*SI),
1687 BPI->getEdgeWeight(BB, SI)))
1688 // Irreducible backedge.
1692 // Distribute mass to successors, saving exit and backedge data in the
1694 distributeMass(Node, OuterLoop, Dist);
1699 raw_ostream &BlockFrequencyInfoImpl<BT>::print(raw_ostream &OS) const {
1702 OS << "block-frequency-info: " << F->getName() << "\n";
1703 for (const BlockT &BB : *F)
1704 OS << " - " << bfi_detail::getBlockName(&BB)
1705 << ": float = " << getFloatingBlockFreq(&BB)
1706 << ", int = " << getBlockFreq(&BB).getFrequency() << "\n";
1708 // Add an extra newline for readability.