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);
70 static int compare(uint64_t L, uint64_t R, int Shift) {
74 uint64_t L_adjusted = L >> Shift;
80 return L > L_adjusted << Shift ? 1 : 0;
84 /// \brief Simple representation of an unsigned floating point.
86 /// UnsignedFloat is a unsigned floating point number. It uses simple
87 /// saturation arithmetic, and every operation is well-defined for every value.
89 /// The number is split into a signed exponent and unsigned digits. The number
90 /// represented is \c getDigits()*2^getExponent(). In this way, the digits are
91 /// much like the mantissa in the x87 long double, but there is no canonical
92 /// form, so the same number can be represented by many bit representations
93 /// (it's always in "denormal" mode).
95 /// UnsignedFloat is templated on the underlying integer type for digits, which
96 /// is expected to be one of uint64_t, uint32_t, uint16_t or uint8_t.
98 /// Unlike builtin floating point types, UnsignedFloat is portable.
100 /// Unlike APFloat, UnsignedFloat does not model architecture floating point
101 /// behaviour (this should make it a little faster), and implements most
102 /// operators (this makes it usable).
104 /// UnsignedFloat is totally ordered. However, there is no canonical form, so
105 /// there are multiple representations of most scalars. E.g.:
107 /// UnsignedFloat(8u, 0) == UnsignedFloat(4u, 1)
108 /// UnsignedFloat(4u, 1) == UnsignedFloat(2u, 2)
109 /// UnsignedFloat(2u, 2) == UnsignedFloat(1u, 3)
111 /// UnsignedFloat implements most arithmetic operations. Precision is kept
112 /// where possible. Uses simple saturation arithmetic, so that operations
113 /// saturate to 0.0 or getLargest() rather than under or overflowing. It has
114 /// some extra arithmetic for unit inversion. 0.0/0.0 is defined to be 0.0.
115 /// Any other division by 0.0 is defined to be getLargest().
117 /// As a convenience for modifying the exponent, left and right shifting are
118 /// both implemented, and both interpret negative shifts as positive shifts in
119 /// the opposite direction.
121 /// Exponents are limited to the range accepted by x87 long double. This makes
122 /// it trivial to add functionality to convert to APFloat (this is already
123 /// relied on for the implementation of printing).
125 /// The current plan is to gut this and make the necessary parts of it (even
126 /// more) private to BlockFrequencyInfo.
127 template <class DigitsT> class UnsignedFloat : UnsignedFloatBase {
129 static_assert(!std::numeric_limits<DigitsT>::is_signed,
130 "only unsigned floats supported");
132 typedef DigitsT DigitsType;
135 typedef std::numeric_limits<DigitsType> DigitsLimits;
137 static const int Width = sizeof(DigitsType) * 8;
138 static_assert(Width <= 64, "invalid integer width for digits");
145 UnsignedFloat() : Digits(0), Exponent(0) {}
147 UnsignedFloat(DigitsType Digits, int16_t Exponent)
148 : Digits(Digits), Exponent(Exponent) {}
151 UnsignedFloat(const std::pair<uint64_t, int16_t> &X)
152 : Digits(X.first), Exponent(X.second) {}
155 static UnsignedFloat getZero() { return UnsignedFloat(0, 0); }
156 static UnsignedFloat getOne() { return UnsignedFloat(1, 0); }
157 static UnsignedFloat getLargest() {
158 return UnsignedFloat(DigitsLimits::max(), MaxExponent);
160 static UnsignedFloat getFloat(uint64_t N) { return adjustToWidth(N, 0); }
161 static UnsignedFloat getInverseFloat(uint64_t N) {
162 return getFloat(N).invert();
164 static UnsignedFloat getFraction(DigitsType N, DigitsType D) {
165 return getQuotient(N, D);
168 int16_t getExponent() const { return Exponent; }
169 DigitsType getDigits() const { return Digits; }
171 /// \brief Convert to the given integer type.
173 /// Convert to \c IntT using simple saturating arithmetic, truncating if
175 template <class IntT> IntT toInt() const;
177 bool isZero() const { return !Digits; }
178 bool isLargest() const { return *this == getLargest(); }
180 if (Exponent > 0 || Exponent <= -Width)
182 return Digits == DigitsType(1) << -Exponent;
185 /// \brief The log base 2, rounded.
187 /// Get the lg of the scalar. lg 0 is defined to be INT32_MIN.
188 int32_t lg() const { return ScaledNumbers::getLg(Digits, Exponent); }
190 /// \brief The log base 2, rounded towards INT32_MIN.
192 /// Get the lg floor. lg 0 is defined to be INT32_MIN.
193 int32_t lgFloor() const {
194 return ScaledNumbers::getLgFloor(Digits, Exponent);
197 /// \brief The log base 2, rounded towards INT32_MAX.
199 /// Get the lg ceiling. lg 0 is defined to be INT32_MIN.
200 int32_t lgCeiling() const {
201 return ScaledNumbers::getLgCeiling(Digits, Exponent);
204 bool operator==(const UnsignedFloat &X) const { return compare(X) == 0; }
205 bool operator<(const UnsignedFloat &X) const { return compare(X) < 0; }
206 bool operator!=(const UnsignedFloat &X) const { return compare(X) != 0; }
207 bool operator>(const UnsignedFloat &X) const { return compare(X) > 0; }
208 bool operator<=(const UnsignedFloat &X) const { return compare(X) <= 0; }
209 bool operator>=(const UnsignedFloat &X) const { return compare(X) >= 0; }
211 bool operator!() const { return isZero(); }
213 /// \brief Convert to a decimal representation in a string.
215 /// Convert to a string. Uses scientific notation for very large/small
216 /// numbers. Scientific notation is used roughly for numbers outside of the
217 /// range 2^-64 through 2^64.
219 /// \c Precision indicates the number of decimal digits of precision to use;
220 /// 0 requests the maximum available.
222 /// As a special case to make debugging easier, if the number is small enough
223 /// to convert without scientific notation and has more than \c Precision
224 /// digits before the decimal place, it's printed accurately to the first
225 /// digit past zero. E.g., assuming 10 digits of precision:
227 /// 98765432198.7654... => 98765432198.8
228 /// 8765432198.7654... => 8765432198.8
229 /// 765432198.7654... => 765432198.8
230 /// 65432198.7654... => 65432198.77
231 /// 5432198.7654... => 5432198.765
232 std::string toString(unsigned Precision = DefaultPrecision) {
233 return UnsignedFloatBase::toString(Digits, Exponent, Width, Precision);
236 /// \brief Print a decimal representation.
238 /// Print a string. See toString for documentation.
239 raw_ostream &print(raw_ostream &OS,
240 unsigned Precision = DefaultPrecision) const {
241 return UnsignedFloatBase::print(OS, Digits, Exponent, Width, Precision);
243 void dump() const { return UnsignedFloatBase::dump(Digits, Exponent, Width); }
245 UnsignedFloat &operator+=(const UnsignedFloat &X);
246 UnsignedFloat &operator-=(const UnsignedFloat &X);
247 UnsignedFloat &operator*=(const UnsignedFloat &X);
248 UnsignedFloat &operator/=(const UnsignedFloat &X);
249 UnsignedFloat &operator<<=(int16_t Shift) { shiftLeft(Shift); return *this; }
250 UnsignedFloat &operator>>=(int16_t Shift) { shiftRight(Shift); return *this; }
253 void shiftLeft(int32_t Shift);
254 void shiftRight(int32_t Shift);
256 /// \brief Adjust two floats to have matching exponents.
258 /// Adjust \c this and \c X to have matching exponents. Returns the new \c X
259 /// by value. Does nothing if \a isZero() for either.
261 /// The value that compares smaller will lose precision, and possibly become
263 UnsignedFloat matchExponents(UnsignedFloat X);
265 /// \brief Increase exponent to match another float.
267 /// Increases \c this to have an exponent matching \c X. May decrease the
268 /// exponent of \c X in the process, and \c this may possibly become \a
270 void increaseExponentToMatch(UnsignedFloat &X, int32_t ExponentDiff);
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 UnsignedFloat &X) const;
293 int compareTo(uint64_t N) const {
294 UnsignedFloat Float = getFloat(N);
295 int Compare = compare(Float);
296 if (Width == 64 || Compare != 0)
299 // Check for precision loss. We know *this == RoundTrip.
300 uint64_t RoundTrip = Float.template toInt<uint64_t>();
301 return N == RoundTrip ? 0 : RoundTrip < N ? -1 : 1;
303 int compareTo(int64_t N) const { return N < 0 ? 1 : compareTo(uint64_t(N)); }
305 UnsignedFloat &invert() { return *this = UnsignedFloat::getFloat(1) / *this; }
306 UnsignedFloat inverse() const { return UnsignedFloat(*this).invert(); }
309 static UnsignedFloat getProduct(DigitsType LHS, DigitsType RHS) {
310 return ScaledNumbers::getProduct(LHS, RHS);
312 static UnsignedFloat getQuotient(DigitsType Dividend, DigitsType Divisor) {
313 return ScaledNumbers::getQuotient(Dividend, Divisor);
316 static int countLeadingZerosWidth(DigitsType Digits) {
318 return countLeadingZeros64(Digits);
320 return countLeadingZeros32(Digits);
321 return countLeadingZeros32(Digits) + Width - 32;
324 /// \brief Adjust a number to width, rounding up if necessary.
326 /// Should only be called for \c Shift close to zero.
328 /// \pre Shift >= MinExponent && Shift + 64 <= MaxExponent.
329 static UnsignedFloat adjustToWidth(uint64_t N, int32_t Shift) {
330 assert(Shift >= MinExponent && "Shift should be close to 0");
331 assert(Shift <= MaxExponent - 64 && "Shift should be close to 0");
332 auto Adjusted = ScaledNumbers::getAdjusted<DigitsT>(N, Shift);
336 static UnsignedFloat getRounded(UnsignedFloat P, bool Round) {
341 return ScaledNumbers::getRounded(P.Digits, P.Exponent, Round);
345 #define UNSIGNED_FLOAT_BOP(op, base) \
346 template <class DigitsT> \
347 UnsignedFloat<DigitsT> operator op(const UnsignedFloat<DigitsT> &L, \
348 const UnsignedFloat<DigitsT> &R) { \
349 return UnsignedFloat<DigitsT>(L) base R; \
351 UNSIGNED_FLOAT_BOP(+, += )
352 UNSIGNED_FLOAT_BOP(-, -= )
353 UNSIGNED_FLOAT_BOP(*, *= )
354 UNSIGNED_FLOAT_BOP(/, /= )
355 UNSIGNED_FLOAT_BOP(<<, <<= )
356 UNSIGNED_FLOAT_BOP(>>, >>= )
357 #undef UNSIGNED_FLOAT_BOP
359 template <class DigitsT>
360 raw_ostream &operator<<(raw_ostream &OS, const UnsignedFloat<DigitsT> &X) {
361 return X.print(OS, 10);
364 #define UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, T1, T2) \
365 template <class DigitsT> \
366 bool operator op(const UnsignedFloat<DigitsT> &L, T1 R) { \
367 return L.compareTo(T2(R)) op 0; \
369 template <class DigitsT> \
370 bool operator op(T1 L, const UnsignedFloat<DigitsT> &R) { \
371 return 0 op R.compareTo(T2(L)); \
373 #define UNSIGNED_FLOAT_COMPARE_TO(op) \
374 UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, uint64_t, uint64_t) \
375 UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, uint32_t, uint64_t) \
376 UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, int64_t, int64_t) \
377 UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, int32_t, int64_t)
378 UNSIGNED_FLOAT_COMPARE_TO(< )
379 UNSIGNED_FLOAT_COMPARE_TO(> )
380 UNSIGNED_FLOAT_COMPARE_TO(== )
381 UNSIGNED_FLOAT_COMPARE_TO(!= )
382 UNSIGNED_FLOAT_COMPARE_TO(<= )
383 UNSIGNED_FLOAT_COMPARE_TO(>= )
384 #undef UNSIGNED_FLOAT_COMPARE_TO
385 #undef UNSIGNED_FLOAT_COMPARE_TO_TYPE
387 template <class DigitsT>
388 uint64_t UnsignedFloat<DigitsT>::scale(uint64_t N) const {
389 if (Width == 64 || N <= DigitsLimits::max())
390 return (getFloat(N) * *this).template toInt<uint64_t>();
392 // Defer to the 64-bit version.
393 return UnsignedFloat<uint64_t>(Digits, Exponent).scale(N);
396 template <class DigitsT>
397 template <class IntT>
398 IntT UnsignedFloat<DigitsT>::toInt() const {
399 typedef std::numeric_limits<IntT> Limits;
402 if (*this >= Limits::max())
403 return Limits::max();
407 assert(size_t(Exponent) < sizeof(IntT) * 8);
408 return N << Exponent;
411 assert(size_t(-Exponent) < sizeof(IntT) * 8);
412 return N >> -Exponent;
417 template <class DigitsT>
418 UnsignedFloat<DigitsT> UnsignedFloat<DigitsT>::matchExponents(UnsignedFloat X) {
419 if (isZero() || X.isZero() || Exponent == X.Exponent)
422 int32_t Diff = int32_t(X.Exponent) - int32_t(Exponent);
424 increaseExponentToMatch(X, Diff);
426 X.increaseExponentToMatch(*this, -Diff);
429 template <class DigitsT>
430 void UnsignedFloat<DigitsT>::increaseExponentToMatch(UnsignedFloat &X,
431 int32_t ExponentDiff) {
432 assert(ExponentDiff > 0);
433 if (ExponentDiff >= 2 * Width) {
438 // Use up any leading zeros on X, and then shift this.
439 int32_t ShiftX = std::min(countLeadingZerosWidth(X.Digits), ExponentDiff);
440 assert(ShiftX < Width);
442 int32_t ShiftThis = ExponentDiff - ShiftX;
443 if (ShiftThis >= Width) {
449 X.Exponent -= ShiftX;
450 Digits >>= ShiftThis;
451 Exponent += ShiftThis;
455 template <class DigitsT>
456 UnsignedFloat<DigitsT> &UnsignedFloat<DigitsT>::
457 operator+=(const UnsignedFloat &X) {
458 if (isLargest() || X.isZero())
460 if (isZero() || X.isLargest())
463 // Normalize exponents.
464 UnsignedFloat Scaled = matchExponents(X);
466 // Check for zero again.
468 return *this = Scaled;
473 DigitsType Sum = Digits + Scaled.Digits;
474 bool DidOverflow = Sum < Digits;
479 if (Exponent == MaxExponent)
480 return *this = getLargest();
483 Digits = UINT64_C(1) << (Width - 1) | Digits >> 1;
487 template <class DigitsT>
488 UnsignedFloat<DigitsT> &UnsignedFloat<DigitsT>::
489 operator-=(const UnsignedFloat &X) {
493 return *this = getZero();
495 // Normalize exponents.
496 UnsignedFloat Scaled = matchExponents(X);
497 assert(Digits >= Scaled.Digits);
499 // Compute difference.
500 if (!Scaled.isZero()) {
501 Digits -= Scaled.Digits;
505 // Check if X just barely lost its last bit. E.g., for 32-bit:
507 // 1*2^32 - 1*2^0 == 0xffffffff != 1*2^32
508 if (*this == UnsignedFloat(1, X.lgFloor() + Width)) {
509 Digits = DigitsType(0) - 1;
514 template <class DigitsT>
515 UnsignedFloat<DigitsT> &UnsignedFloat<DigitsT>::
516 operator*=(const UnsignedFloat &X) {
522 // Save the exponents.
523 int32_t Exponents = int32_t(Exponent) + int32_t(X.Exponent);
525 // Get the raw product.
526 *this = getProduct(Digits, X.Digits);
528 // Combine with exponents.
529 return *this <<= Exponents;
531 template <class DigitsT>
532 UnsignedFloat<DigitsT> &UnsignedFloat<DigitsT>::
533 operator/=(const UnsignedFloat &X) {
537 return *this = getLargest();
539 // Save the exponents.
540 int32_t Exponents = int32_t(Exponent) - int32_t(X.Exponent);
542 // Get the raw quotient.
543 *this = getQuotient(Digits, X.Digits);
545 // Combine with exponents.
546 return *this <<= Exponents;
548 template <class DigitsT>
549 void UnsignedFloat<DigitsT>::shiftLeft(int32_t Shift) {
550 if (!Shift || isZero())
552 assert(Shift != INT32_MIN);
558 // Shift as much as we can in the exponent.
559 int32_t ExponentShift = std::min(Shift, MaxExponent - Exponent);
560 Exponent += ExponentShift;
561 if (ExponentShift == Shift)
564 // Check this late, since it's rare.
568 // Shift the digits themselves.
569 Shift -= ExponentShift;
570 if (Shift > countLeadingZerosWidth(Digits)) {
572 *this = getLargest();
580 template <class DigitsT>
581 void UnsignedFloat<DigitsT>::shiftRight(int32_t Shift) {
582 if (!Shift || isZero())
584 assert(Shift != INT32_MIN);
590 // Shift as much as we can in the exponent.
591 int32_t ExponentShift = std::min(Shift, Exponent - MinExponent);
592 Exponent -= ExponentShift;
593 if (ExponentShift == Shift)
596 // Shift the digits themselves.
597 Shift -= ExponentShift;
598 if (Shift >= Width) {
608 template <class DigitsT>
609 int UnsignedFloat<DigitsT>::compare(const UnsignedFloat &X) const {
612 return X.isZero() ? 0 : -1;
616 // Check for the scale. Use lgFloor to be sure that the exponent difference
617 // is always lower than 64.
618 int32_t lgL = lgFloor(), lgR = X.lgFloor();
620 return lgL < lgR ? -1 : 1;
623 if (Exponent < X.Exponent)
624 return UnsignedFloatBase::compare(Digits, X.Digits, X.Exponent - Exponent);
626 return -UnsignedFloatBase::compare(X.Digits, Digits, Exponent - X.Exponent);
629 template <class T> struct isPodLike<UnsignedFloat<T>> {
630 static const bool value = true;
634 //===----------------------------------------------------------------------===//
636 // BlockMass definition.
638 // TODO: Make this private to BlockFrequencyInfoImpl or delete.
640 //===----------------------------------------------------------------------===//
643 /// \brief Mass of a block.
645 /// This class implements a sort of fixed-point fraction always between 0.0 and
646 /// 1.0. getMass() == UINT64_MAX indicates a value of 1.0.
648 /// Masses can be added and subtracted. Simple saturation arithmetic is used,
649 /// so arithmetic operations never overflow or underflow.
651 /// Masses can be multiplied. Multiplication treats full mass as 1.0 and uses
652 /// an inexpensive floating-point algorithm that's off-by-one (almost, but not
653 /// quite, maximum precision).
655 /// Masses can be scaled by \a BranchProbability at maximum precision.
660 BlockMass() : Mass(0) {}
661 explicit BlockMass(uint64_t Mass) : Mass(Mass) {}
663 static BlockMass getEmpty() { return BlockMass(); }
664 static BlockMass getFull() { return BlockMass(UINT64_MAX); }
666 uint64_t getMass() const { return Mass; }
668 bool isFull() const { return Mass == UINT64_MAX; }
669 bool isEmpty() const { return !Mass; }
671 bool operator!() const { return isEmpty(); }
673 /// \brief Add another mass.
675 /// Adds another mass, saturating at \a isFull() rather than overflowing.
676 BlockMass &operator+=(const BlockMass &X) {
677 uint64_t Sum = Mass + X.Mass;
678 Mass = Sum < Mass ? UINT64_MAX : Sum;
682 /// \brief Subtract another mass.
684 /// Subtracts another mass, saturating at \a isEmpty() rather than
686 BlockMass &operator-=(const BlockMass &X) {
687 uint64_t Diff = Mass - X.Mass;
688 Mass = Diff > Mass ? 0 : Diff;
692 BlockMass &operator*=(const BranchProbability &P) {
693 Mass = P.scale(Mass);
697 bool operator==(const BlockMass &X) const { return Mass == X.Mass; }
698 bool operator!=(const BlockMass &X) const { return Mass != X.Mass; }
699 bool operator<=(const BlockMass &X) const { return Mass <= X.Mass; }
700 bool operator>=(const BlockMass &X) const { return Mass >= X.Mass; }
701 bool operator<(const BlockMass &X) const { return Mass < X.Mass; }
702 bool operator>(const BlockMass &X) const { return Mass > X.Mass; }
704 /// \brief Convert to floating point.
706 /// Convert to a float. \a isFull() gives 1.0, while \a isEmpty() gives
707 /// slightly above 0.0.
708 UnsignedFloat<uint64_t> toFloat() const;
711 raw_ostream &print(raw_ostream &OS) const;
714 inline BlockMass operator+(const BlockMass &L, const BlockMass &R) {
715 return BlockMass(L) += R;
717 inline BlockMass operator-(const BlockMass &L, const BlockMass &R) {
718 return BlockMass(L) -= R;
720 inline BlockMass operator*(const BlockMass &L, const BranchProbability &R) {
721 return BlockMass(L) *= R;
723 inline BlockMass operator*(const BranchProbability &L, const BlockMass &R) {
724 return BlockMass(R) *= L;
727 inline raw_ostream &operator<<(raw_ostream &OS, const BlockMass &X) {
731 template <> struct isPodLike<BlockMass> {
732 static const bool value = true;
736 //===----------------------------------------------------------------------===//
738 // BlockFrequencyInfoImpl definition.
740 //===----------------------------------------------------------------------===//
744 class BranchProbabilityInfo;
748 class MachineBasicBlock;
749 class MachineBranchProbabilityInfo;
750 class MachineFunction;
752 class MachineLoopInfo;
754 namespace bfi_detail {
755 struct IrreducibleGraph;
757 // This is part of a workaround for a GCC 4.7 crash on lambdas.
758 template <class BT> struct BlockEdgesAdder;
761 /// \brief Base class for BlockFrequencyInfoImpl
763 /// BlockFrequencyInfoImplBase has supporting data structures and some
764 /// algorithms for BlockFrequencyInfoImplBase. Only algorithms that depend on
765 /// the block type (or that call such algorithms) are skipped here.
767 /// Nevertheless, the majority of the overall algorithm documention lives with
768 /// BlockFrequencyInfoImpl. See there for details.
769 class BlockFrequencyInfoImplBase {
771 typedef UnsignedFloat<uint64_t> Float;
773 /// \brief Representative of a block.
775 /// This is a simple wrapper around an index into the reverse-post-order
776 /// traversal of the blocks.
778 /// Unlike a block pointer, its order has meaning (location in the
779 /// topological sort) and it's class is the same regardless of block type.
781 typedef uint32_t IndexType;
784 bool operator==(const BlockNode &X) const { return Index == X.Index; }
785 bool operator!=(const BlockNode &X) const { return Index != X.Index; }
786 bool operator<=(const BlockNode &X) const { return Index <= X.Index; }
787 bool operator>=(const BlockNode &X) const { return Index >= X.Index; }
788 bool operator<(const BlockNode &X) const { return Index < X.Index; }
789 bool operator>(const BlockNode &X) const { return Index > X.Index; }
791 BlockNode() : Index(UINT32_MAX) {}
792 BlockNode(IndexType Index) : Index(Index) {}
794 bool isValid() const { return Index <= getMaxIndex(); }
795 static size_t getMaxIndex() { return UINT32_MAX - 1; }
798 /// \brief Stats about a block itself.
799 struct FrequencyData {
804 /// \brief Data about a loop.
806 /// Contains the data necessary to represent represent a loop as a
807 /// pseudo-node once it's packaged.
809 typedef SmallVector<std::pair<BlockNode, BlockMass>, 4> ExitMap;
810 typedef SmallVector<BlockNode, 4> NodeList;
811 LoopData *Parent; ///< The parent loop.
812 bool IsPackaged; ///< Whether this has been packaged.
813 uint32_t NumHeaders; ///< Number of headers.
814 ExitMap Exits; ///< Successor edges (and weights).
815 NodeList Nodes; ///< Header and the members of the loop.
816 BlockMass BackedgeMass; ///< Mass returned to loop header.
820 LoopData(LoopData *Parent, const BlockNode &Header)
821 : Parent(Parent), IsPackaged(false), NumHeaders(1), Nodes(1, Header) {}
822 template <class It1, class It2>
823 LoopData(LoopData *Parent, It1 FirstHeader, It1 LastHeader, It2 FirstOther,
825 : Parent(Parent), IsPackaged(false), Nodes(FirstHeader, LastHeader) {
826 NumHeaders = Nodes.size();
827 Nodes.insert(Nodes.end(), FirstOther, LastOther);
829 bool isHeader(const BlockNode &Node) const {
831 return std::binary_search(Nodes.begin(), Nodes.begin() + NumHeaders,
833 return Node == Nodes[0];
835 BlockNode getHeader() const { return Nodes[0]; }
836 bool isIrreducible() const { return NumHeaders > 1; }
838 NodeList::const_iterator members_begin() const {
839 return Nodes.begin() + NumHeaders;
841 NodeList::const_iterator members_end() const { return Nodes.end(); }
842 iterator_range<NodeList::const_iterator> members() const {
843 return make_range(members_begin(), members_end());
847 /// \brief Index of loop information.
849 BlockNode Node; ///< This node.
850 LoopData *Loop; ///< The loop this block is inside.
851 BlockMass Mass; ///< Mass distribution from the entry block.
853 WorkingData(const BlockNode &Node) : Node(Node), Loop(nullptr) {}
855 bool isLoopHeader() const { return Loop && Loop->isHeader(Node); }
856 bool isDoubleLoopHeader() const {
857 return isLoopHeader() && Loop->Parent && Loop->Parent->isIrreducible() &&
858 Loop->Parent->isHeader(Node);
861 LoopData *getContainingLoop() const {
864 if (!isDoubleLoopHeader())
866 return Loop->Parent->Parent;
869 /// \brief Resolve a node to its representative.
871 /// Get the node currently representing Node, which could be a containing
874 /// This function should only be called when distributing mass. As long as
875 /// there are no irreducilbe edges to Node, then it will have complexity
876 /// O(1) in this context.
878 /// In general, the complexity is O(L), where L is the number of loop
879 /// headers Node has been packaged into. Since this method is called in
880 /// the context of distributing mass, L will be the number of loop headers
881 /// an early exit edge jumps out of.
882 BlockNode getResolvedNode() const {
883 auto L = getPackagedLoop();
884 return L ? L->getHeader() : Node;
886 LoopData *getPackagedLoop() const {
887 if (!Loop || !Loop->IsPackaged)
890 while (L->Parent && L->Parent->IsPackaged)
895 /// \brief Get the appropriate mass for a node.
897 /// Get appropriate mass for Node. If Node is a loop-header (whose loop
898 /// has been packaged), returns the mass of its pseudo-node. If it's a
899 /// node inside a packaged loop, it returns the loop's mass.
900 BlockMass &getMass() {
903 if (!isADoublePackage())
905 return Loop->Parent->Mass;
908 /// \brief Has ContainingLoop been packaged up?
909 bool isPackaged() const { return getResolvedNode() != Node; }
910 /// \brief Has Loop been packaged up?
911 bool isAPackage() const { return isLoopHeader() && Loop->IsPackaged; }
912 /// \brief Has Loop been packaged up twice?
913 bool isADoublePackage() const {
914 return isDoubleLoopHeader() && Loop->Parent->IsPackaged;
918 /// \brief Unscaled probability weight.
920 /// Probability weight for an edge in the graph (including the
921 /// successor/target node).
923 /// All edges in the original function are 32-bit. However, exit edges from
924 /// loop packages are taken from 64-bit exit masses, so we need 64-bits of
925 /// space in general.
927 /// In addition to the raw weight amount, Weight stores the type of the edge
928 /// in the current context (i.e., the context of the loop being processed).
929 /// Is this a local edge within the loop, an exit from the loop, or a
930 /// backedge to the loop header?
932 enum DistType { Local, Exit, Backedge };
934 BlockNode TargetNode;
936 Weight() : Type(Local), Amount(0) {}
939 /// \brief Distribution of unscaled probability weight.
941 /// Distribution of unscaled probability weight to a set of successors.
943 /// This class collates the successor edge weights for later processing.
945 /// \a DidOverflow indicates whether \a Total did overflow while adding to
946 /// the distribution. It should never overflow twice.
947 struct Distribution {
948 typedef SmallVector<Weight, 4> WeightList;
949 WeightList Weights; ///< Individual successor weights.
950 uint64_t Total; ///< Sum of all weights.
951 bool DidOverflow; ///< Whether \a Total did overflow.
953 Distribution() : Total(0), DidOverflow(false) {}
954 void addLocal(const BlockNode &Node, uint64_t Amount) {
955 add(Node, Amount, Weight::Local);
957 void addExit(const BlockNode &Node, uint64_t Amount) {
958 add(Node, Amount, Weight::Exit);
960 void addBackedge(const BlockNode &Node, uint64_t Amount) {
961 add(Node, Amount, Weight::Backedge);
964 /// \brief Normalize the distribution.
966 /// Combines multiple edges to the same \a Weight::TargetNode and scales
967 /// down so that \a Total fits into 32-bits.
969 /// This is linear in the size of \a Weights. For the vast majority of
970 /// cases, adjacent edge weights are combined by sorting WeightList and
971 /// combining adjacent weights. However, for very large edge lists an
972 /// auxiliary hash table is used.
976 void add(const BlockNode &Node, uint64_t Amount, Weight::DistType Type);
979 /// \brief Data about each block. This is used downstream.
980 std::vector<FrequencyData> Freqs;
982 /// \brief Loop data: see initializeLoops().
983 std::vector<WorkingData> Working;
985 /// \brief Indexed information about loops.
986 std::list<LoopData> Loops;
988 /// \brief Add all edges out of a packaged loop to the distribution.
990 /// Adds all edges from LocalLoopHead to Dist. Calls addToDist() to add each
993 /// \return \c true unless there's an irreducible backedge.
994 bool addLoopSuccessorsToDist(const LoopData *OuterLoop, LoopData &Loop,
997 /// \brief Add an edge to the distribution.
999 /// Adds an edge to Succ to Dist. If \c LoopHead.isValid(), then whether the
1000 /// edge is local/exit/backedge is in the context of LoopHead. Otherwise,
1001 /// every edge should be a local edge (since all the loops are packaged up).
1003 /// \return \c true unless aborted due to an irreducible backedge.
1004 bool addToDist(Distribution &Dist, const LoopData *OuterLoop,
1005 const BlockNode &Pred, const BlockNode &Succ, uint64_t Weight);
1007 LoopData &getLoopPackage(const BlockNode &Head) {
1008 assert(Head.Index < Working.size());
1009 assert(Working[Head.Index].isLoopHeader());
1010 return *Working[Head.Index].Loop;
1013 /// \brief Analyze irreducible SCCs.
1015 /// Separate irreducible SCCs from \c G, which is an explict graph of \c
1016 /// OuterLoop (or the top-level function, if \c OuterLoop is \c nullptr).
1017 /// Insert them into \a Loops before \c Insert.
1019 /// \return the \c LoopData nodes representing the irreducible SCCs.
1020 iterator_range<std::list<LoopData>::iterator>
1021 analyzeIrreducible(const bfi_detail::IrreducibleGraph &G, LoopData *OuterLoop,
1022 std::list<LoopData>::iterator Insert);
1024 /// \brief Update a loop after packaging irreducible SCCs inside of it.
1026 /// Update \c OuterLoop. Before finding irreducible control flow, it was
1027 /// partway through \a computeMassInLoop(), so \a LoopData::Exits and \a
1028 /// LoopData::BackedgeMass need to be reset. Also, nodes that were packaged
1029 /// up need to be removed from \a OuterLoop::Nodes.
1030 void updateLoopWithIrreducible(LoopData &OuterLoop);
1032 /// \brief Distribute mass according to a distribution.
1034 /// Distributes the mass in Source according to Dist. If LoopHead.isValid(),
1035 /// backedges and exits are stored in its entry in Loops.
1037 /// Mass is distributed in parallel from two copies of the source mass.
1038 void distributeMass(const BlockNode &Source, LoopData *OuterLoop,
1039 Distribution &Dist);
1041 /// \brief Compute the loop scale for a loop.
1042 void computeLoopScale(LoopData &Loop);
1044 /// \brief Package up a loop.
1045 void packageLoop(LoopData &Loop);
1047 /// \brief Unwrap loops.
1050 /// \brief Finalize frequency metrics.
1052 /// Calculates final frequencies and cleans up no-longer-needed data
1054 void finalizeMetrics();
1056 /// \brief Clear all memory.
1059 virtual std::string getBlockName(const BlockNode &Node) const;
1060 std::string getLoopName(const LoopData &Loop) const;
1062 virtual raw_ostream &print(raw_ostream &OS) const { return OS; }
1063 void dump() const { print(dbgs()); }
1065 Float getFloatingBlockFreq(const BlockNode &Node) const;
1067 BlockFrequency getBlockFreq(const BlockNode &Node) const;
1069 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockNode &Node) const;
1070 raw_ostream &printBlockFreq(raw_ostream &OS,
1071 const BlockFrequency &Freq) const;
1073 uint64_t getEntryFreq() const {
1074 assert(!Freqs.empty());
1075 return Freqs[0].Integer;
1077 /// \brief Virtual destructor.
1079 /// Need a virtual destructor to mask the compiler warning about
1081 virtual ~BlockFrequencyInfoImplBase() {}
1084 namespace bfi_detail {
1085 template <class BlockT> struct TypeMap {};
1086 template <> struct TypeMap<BasicBlock> {
1087 typedef BasicBlock BlockT;
1088 typedef Function FunctionT;
1089 typedef BranchProbabilityInfo BranchProbabilityInfoT;
1091 typedef LoopInfo LoopInfoT;
1093 template <> struct TypeMap<MachineBasicBlock> {
1094 typedef MachineBasicBlock BlockT;
1095 typedef MachineFunction FunctionT;
1096 typedef MachineBranchProbabilityInfo BranchProbabilityInfoT;
1097 typedef MachineLoop LoopT;
1098 typedef MachineLoopInfo LoopInfoT;
1101 /// \brief Get the name of a MachineBasicBlock.
1103 /// Get the name of a MachineBasicBlock. It's templated so that including from
1104 /// CodeGen is unnecessary (that would be a layering issue).
1106 /// This is used mainly for debug output. The name is similar to
1107 /// MachineBasicBlock::getFullName(), but skips the name of the function.
1108 template <class BlockT> std::string getBlockName(const BlockT *BB) {
1109 assert(BB && "Unexpected nullptr");
1110 auto MachineName = "BB" + Twine(BB->getNumber());
1111 if (BB->getBasicBlock())
1112 return (MachineName + "[" + BB->getName() + "]").str();
1113 return MachineName.str();
1115 /// \brief Get the name of a BasicBlock.
1116 template <> inline std::string getBlockName(const BasicBlock *BB) {
1117 assert(BB && "Unexpected nullptr");
1118 return BB->getName().str();
1121 /// \brief Graph of irreducible control flow.
1123 /// This graph is used for determining the SCCs in a loop (or top-level
1124 /// function) that has irreducible control flow.
1126 /// During the block frequency algorithm, the local graphs are defined in a
1127 /// light-weight way, deferring to the \a BasicBlock or \a MachineBasicBlock
1128 /// graphs for most edges, but getting others from \a LoopData::ExitMap. The
1129 /// latter only has successor information.
1131 /// \a IrreducibleGraph makes this graph explicit. It's in a form that can use
1132 /// \a GraphTraits (so that \a analyzeIrreducible() can use \a scc_iterator),
1133 /// and it explicitly lists predecessors and successors. The initialization
1134 /// that relies on \c MachineBasicBlock is defined in the header.
1135 struct IrreducibleGraph {
1136 typedef BlockFrequencyInfoImplBase BFIBase;
1140 typedef BFIBase::BlockNode BlockNode;
1144 std::deque<const IrrNode *> Edges;
1145 IrrNode(const BlockNode &Node) : Node(Node), NumIn(0) {}
1147 typedef std::deque<const IrrNode *>::const_iterator iterator;
1148 iterator pred_begin() const { return Edges.begin(); }
1149 iterator succ_begin() const { return Edges.begin() + NumIn; }
1150 iterator pred_end() const { return succ_begin(); }
1151 iterator succ_end() const { return Edges.end(); }
1154 const IrrNode *StartIrr;
1155 std::vector<IrrNode> Nodes;
1156 SmallDenseMap<uint32_t, IrrNode *, 4> Lookup;
1158 /// \brief Construct an explicit graph containing irreducible control flow.
1160 /// Construct an explicit graph of the control flow in \c OuterLoop (or the
1161 /// top-level function, if \c OuterLoop is \c nullptr). Uses \c
1162 /// addBlockEdges to add block successors that have not been packaged into
1165 /// \a BlockFrequencyInfoImpl::computeIrreducibleMass() is the only expected
1167 template <class BlockEdgesAdder>
1168 IrreducibleGraph(BFIBase &BFI, const BFIBase::LoopData *OuterLoop,
1169 BlockEdgesAdder addBlockEdges)
1170 : BFI(BFI), StartIrr(nullptr) {
1171 initialize(OuterLoop, addBlockEdges);
1174 template <class BlockEdgesAdder>
1175 void initialize(const BFIBase::LoopData *OuterLoop,
1176 BlockEdgesAdder addBlockEdges);
1177 void addNodesInLoop(const BFIBase::LoopData &OuterLoop);
1178 void addNodesInFunction();
1179 void addNode(const BlockNode &Node) {
1180 Nodes.emplace_back(Node);
1181 BFI.Working[Node.Index].getMass() = BlockMass::getEmpty();
1184 template <class BlockEdgesAdder>
1185 void addEdges(const BlockNode &Node, const BFIBase::LoopData *OuterLoop,
1186 BlockEdgesAdder addBlockEdges);
1187 void addEdge(IrrNode &Irr, const BlockNode &Succ,
1188 const BFIBase::LoopData *OuterLoop);
1190 template <class BlockEdgesAdder>
1191 void IrreducibleGraph::initialize(const BFIBase::LoopData *OuterLoop,
1192 BlockEdgesAdder addBlockEdges) {
1194 addNodesInLoop(*OuterLoop);
1195 for (auto N : OuterLoop->Nodes)
1196 addEdges(N, OuterLoop, addBlockEdges);
1198 addNodesInFunction();
1199 for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index)
1200 addEdges(Index, OuterLoop, addBlockEdges);
1202 StartIrr = Lookup[Start.Index];
1204 template <class BlockEdgesAdder>
1205 void IrreducibleGraph::addEdges(const BlockNode &Node,
1206 const BFIBase::LoopData *OuterLoop,
1207 BlockEdgesAdder addBlockEdges) {
1208 auto L = Lookup.find(Node.Index);
1209 if (L == Lookup.end())
1211 IrrNode &Irr = *L->second;
1212 const auto &Working = BFI.Working[Node.Index];
1214 if (Working.isAPackage())
1215 for (const auto &I : Working.Loop->Exits)
1216 addEdge(Irr, I.first, OuterLoop);
1218 addBlockEdges(*this, Irr, OuterLoop);
1222 /// \brief Shared implementation for block frequency analysis.
1224 /// This is a shared implementation of BlockFrequencyInfo and
1225 /// MachineBlockFrequencyInfo, and calculates the relative frequencies of
1228 /// LoopInfo defines a loop as a "non-trivial" SCC dominated by a single block,
1229 /// which is called the header. A given loop, L, can have sub-loops, which are
1230 /// loops within the subgraph of L that exclude its header. (A "trivial" SCC
1231 /// consists of a single block that does not have a self-edge.)
1233 /// In addition to loops, this algorithm has limited support for irreducible
1234 /// SCCs, which are SCCs with multiple entry blocks. Irreducible SCCs are
1235 /// discovered on they fly, and modelled as loops with multiple headers.
1237 /// The headers of irreducible sub-SCCs consist of its entry blocks and all
1238 /// nodes that are targets of a backedge within it (excluding backedges within
1239 /// true sub-loops). Block frequency calculations act as if a block is
1240 /// inserted that intercepts all the edges to the headers. All backedges and
1241 /// entries point to this block. Its successors are the headers, which split
1242 /// the frequency evenly.
1244 /// This algorithm leverages BlockMass and UnsignedFloat to maintain precision,
1245 /// separates mass distribution from loop scaling, and dithers to eliminate
1246 /// probability mass loss.
1248 /// The implementation is split between BlockFrequencyInfoImpl, which knows the
1249 /// type of graph being modelled (BasicBlock vs. MachineBasicBlock), and
1250 /// BlockFrequencyInfoImplBase, which doesn't. The base class uses \a
1251 /// BlockNode, a wrapper around a uint32_t. BlockNode is numbered from 0 in
1252 /// reverse-post order. This gives two advantages: it's easy to compare the
1253 /// relative ordering of two nodes, and maps keyed on BlockT can be represented
1256 /// This algorithm is O(V+E), unless there is irreducible control flow, in
1257 /// which case it's O(V*E) in the worst case.
1259 /// These are the main stages:
1261 /// 0. Reverse post-order traversal (\a initializeRPOT()).
1263 /// Run a single post-order traversal and save it (in reverse) in RPOT.
1264 /// All other stages make use of this ordering. Save a lookup from BlockT
1265 /// to BlockNode (the index into RPOT) in Nodes.
1267 /// 1. Loop initialization (\a initializeLoops()).
1269 /// Translate LoopInfo/MachineLoopInfo into a form suitable for the rest of
1270 /// the algorithm. In particular, store the immediate members of each loop
1271 /// in reverse post-order.
1273 /// 2. Calculate mass and scale in loops (\a computeMassInLoops()).
1275 /// For each loop (bottom-up), distribute mass through the DAG resulting
1276 /// from ignoring backedges and treating sub-loops as a single pseudo-node.
1277 /// Track the backedge mass distributed to the loop header, and use it to
1278 /// calculate the loop scale (number of loop iterations). Immediate
1279 /// members that represent sub-loops will already have been visited and
1280 /// packaged into a pseudo-node.
1282 /// Distributing mass in a loop is a reverse-post-order traversal through
1283 /// the loop. Start by assigning full mass to the Loop header. For each
1284 /// node in the loop:
1286 /// - Fetch and categorize the weight distribution for its successors.
1287 /// If this is a packaged-subloop, the weight distribution is stored
1288 /// in \a LoopData::Exits. Otherwise, fetch it from
1289 /// BranchProbabilityInfo.
1291 /// - Each successor is categorized as \a Weight::Local, a local edge
1292 /// within the current loop, \a Weight::Backedge, a backedge to the
1293 /// loop header, or \a Weight::Exit, any successor outside the loop.
1294 /// The weight, the successor, and its category are stored in \a
1295 /// Distribution. There can be multiple edges to each successor.
1297 /// - If there's a backedge to a non-header, there's an irreducible SCC.
1298 /// The usual flow is temporarily aborted. \a
1299 /// computeIrreducibleMass() finds the irreducible SCCs within the
1300 /// loop, packages them up, and restarts the flow.
1302 /// - Normalize the distribution: scale weights down so that their sum
1303 /// is 32-bits, and coalesce multiple edges to the same node.
1305 /// - Distribute the mass accordingly, dithering to minimize mass loss,
1306 /// as described in \a distributeMass().
1308 /// Finally, calculate the loop scale from the accumulated backedge mass.
1310 /// 3. Distribute mass in the function (\a computeMassInFunction()).
1312 /// Finally, distribute mass through the DAG resulting from packaging all
1313 /// loops in the function. This uses the same algorithm as distributing
1314 /// mass in a loop, except that there are no exit or backedge edges.
1316 /// 4. Unpackage loops (\a unwrapLoops()).
1318 /// Initialize each block's frequency to a floating point representation of
1321 /// Visit loops top-down, scaling the frequencies of its immediate members
1322 /// by the loop's pseudo-node's frequency.
1324 /// 5. Convert frequencies to a 64-bit range (\a finalizeMetrics()).
1326 /// Using the min and max frequencies as a guide, translate floating point
1327 /// frequencies to an appropriate range in uint64_t.
1329 /// It has some known flaws.
1331 /// - Loop scale is limited to 4096 per loop (2^12) to avoid exhausting
1332 /// BlockFrequency's 64-bit integer precision.
1334 /// - The model of irreducible control flow is a rough approximation.
1336 /// Modelling irreducible control flow exactly involves setting up and
1337 /// solving a group of infinite geometric series. Such precision is
1338 /// unlikely to be worthwhile, since most of our algorithms give up on
1339 /// irreducible control flow anyway.
1341 /// Nevertheless, we might find that we need to get closer. Here's a sort
1342 /// of TODO list for the model with diminishing returns, to be completed as
1345 /// - The headers for the \a LoopData representing an irreducible SCC
1346 /// include non-entry blocks. When these extra blocks exist, they
1347 /// indicate a self-contained irreducible sub-SCC. We could treat them
1348 /// as sub-loops, rather than arbitrarily shoving the problematic
1349 /// blocks into the headers of the main irreducible SCC.
1351 /// - Backedge frequencies are assumed to be evenly split between the
1352 /// headers of a given irreducible SCC. Instead, we could track the
1353 /// backedge mass separately for each header, and adjust their relative
1356 /// - Entry frequencies are assumed to be evenly split between the
1357 /// headers of a given irreducible SCC, which is the only option if we
1358 /// need to compute mass in the SCC before its parent loop. Instead,
1359 /// we could partially compute mass in the parent loop, and stop when
1360 /// we get to the SCC. Here, we have the correct ratio of entry
1361 /// masses, which we can use to adjust their relative frequencies.
1362 /// Compute mass in the SCC, and then continue propagation in the
1365 /// - We can propagate mass iteratively through the SCC, for some fixed
1366 /// number of iterations. Each iteration starts by assigning the entry
1367 /// blocks their backedge mass from the prior iteration. The final
1368 /// mass for each block (and each exit, and the total backedge mass
1369 /// used for computing loop scale) is the sum of all iterations.
1370 /// (Running this until fixed point would "solve" the geometric
1371 /// series by simulation.)
1372 template <class BT> class BlockFrequencyInfoImpl : BlockFrequencyInfoImplBase {
1373 typedef typename bfi_detail::TypeMap<BT>::BlockT BlockT;
1374 typedef typename bfi_detail::TypeMap<BT>::FunctionT FunctionT;
1375 typedef typename bfi_detail::TypeMap<BT>::BranchProbabilityInfoT
1376 BranchProbabilityInfoT;
1377 typedef typename bfi_detail::TypeMap<BT>::LoopT LoopT;
1378 typedef typename bfi_detail::TypeMap<BT>::LoopInfoT LoopInfoT;
1380 // This is part of a workaround for a GCC 4.7 crash on lambdas.
1381 friend struct bfi_detail::BlockEdgesAdder<BT>;
1383 typedef GraphTraits<const BlockT *> Successor;
1384 typedef GraphTraits<Inverse<const BlockT *>> Predecessor;
1386 const BranchProbabilityInfoT *BPI;
1387 const LoopInfoT *LI;
1390 // All blocks in reverse postorder.
1391 std::vector<const BlockT *> RPOT;
1392 DenseMap<const BlockT *, BlockNode> Nodes;
1394 typedef typename std::vector<const BlockT *>::const_iterator rpot_iterator;
1396 rpot_iterator rpot_begin() const { return RPOT.begin(); }
1397 rpot_iterator rpot_end() const { return RPOT.end(); }
1399 size_t getIndex(const rpot_iterator &I) const { return I - rpot_begin(); }
1401 BlockNode getNode(const rpot_iterator &I) const {
1402 return BlockNode(getIndex(I));
1404 BlockNode getNode(const BlockT *BB) const { return Nodes.lookup(BB); }
1406 const BlockT *getBlock(const BlockNode &Node) const {
1407 assert(Node.Index < RPOT.size());
1408 return RPOT[Node.Index];
1411 /// \brief Run (and save) a post-order traversal.
1413 /// Saves a reverse post-order traversal of all the nodes in \a F.
1414 void initializeRPOT();
1416 /// \brief Initialize loop data.
1418 /// Build up \a Loops using \a LoopInfo. \a LoopInfo gives us a mapping from
1419 /// each block to the deepest loop it's in, but we need the inverse. For each
1420 /// loop, we store in reverse post-order its "immediate" members, defined as
1421 /// the header, the headers of immediate sub-loops, and all other blocks in
1422 /// the loop that are not in sub-loops.
1423 void initializeLoops();
1425 /// \brief Propagate to a block's successors.
1427 /// In the context of distributing mass through \c OuterLoop, divide the mass
1428 /// currently assigned to \c Node between its successors.
1430 /// \return \c true unless there's an irreducible backedge.
1431 bool propagateMassToSuccessors(LoopData *OuterLoop, const BlockNode &Node);
1433 /// \brief Compute mass in a particular loop.
1435 /// Assign mass to \c Loop's header, and then for each block in \c Loop in
1436 /// reverse post-order, distribute mass to its successors. Only visits nodes
1437 /// that have not been packaged into sub-loops.
1439 /// \pre \a computeMassInLoop() has been called for each subloop of \c Loop.
1440 /// \return \c true unless there's an irreducible backedge.
1441 bool computeMassInLoop(LoopData &Loop);
1443 /// \brief Try to compute mass in the top-level function.
1445 /// Assign mass to the entry block, and then for each block in reverse
1446 /// post-order, distribute mass to its successors. Skips nodes that have
1447 /// been packaged into loops.
1449 /// \pre \a computeMassInLoops() has been called.
1450 /// \return \c true unless there's an irreducible backedge.
1451 bool tryToComputeMassInFunction();
1453 /// \brief Compute mass in (and package up) irreducible SCCs.
1455 /// Find the irreducible SCCs in \c OuterLoop, add them to \a Loops (in front
1456 /// of \c Insert), and call \a computeMassInLoop() on each of them.
1458 /// If \c OuterLoop is \c nullptr, it refers to the top-level function.
1460 /// \pre \a computeMassInLoop() has been called for each subloop of \c
1462 /// \pre \c Insert points at the the last loop successfully processed by \a
1463 /// computeMassInLoop().
1464 /// \pre \c OuterLoop has irreducible SCCs.
1465 void computeIrreducibleMass(LoopData *OuterLoop,
1466 std::list<LoopData>::iterator Insert);
1468 /// \brief Compute mass in all loops.
1470 /// For each loop bottom-up, call \a computeMassInLoop().
1472 /// \a computeMassInLoop() aborts (and returns \c false) on loops that
1473 /// contain a irreducible sub-SCCs. Use \a computeIrreducibleMass() and then
1474 /// re-enter \a computeMassInLoop().
1476 /// \post \a computeMassInLoop() has returned \c true for every loop.
1477 void computeMassInLoops();
1479 /// \brief Compute mass in the top-level function.
1481 /// Uses \a tryToComputeMassInFunction() and \a computeIrreducibleMass() to
1482 /// compute mass in the top-level function.
1484 /// \post \a tryToComputeMassInFunction() has returned \c true.
1485 void computeMassInFunction();
1487 std::string getBlockName(const BlockNode &Node) const override {
1488 return bfi_detail::getBlockName(getBlock(Node));
1492 const FunctionT *getFunction() const { return F; }
1494 void doFunction(const FunctionT *F, const BranchProbabilityInfoT *BPI,
1495 const LoopInfoT *LI);
1496 BlockFrequencyInfoImpl() : BPI(nullptr), LI(nullptr), F(nullptr) {}
1498 using BlockFrequencyInfoImplBase::getEntryFreq;
1499 BlockFrequency getBlockFreq(const BlockT *BB) const {
1500 return BlockFrequencyInfoImplBase::getBlockFreq(getNode(BB));
1502 Float getFloatingBlockFreq(const BlockT *BB) const {
1503 return BlockFrequencyInfoImplBase::getFloatingBlockFreq(getNode(BB));
1506 /// \brief Print the frequencies for the current function.
1508 /// Prints the frequencies for the blocks in the current function.
1510 /// Blocks are printed in the natural iteration order of the function, rather
1511 /// than reverse post-order. This provides two advantages: writing -analyze
1512 /// tests is easier (since blocks come out in source order), and even
1513 /// unreachable blocks are printed.
1515 /// \a BlockFrequencyInfoImplBase::print() only knows reverse post-order, so
1516 /// we need to override it here.
1517 raw_ostream &print(raw_ostream &OS) const override;
1518 using BlockFrequencyInfoImplBase::dump;
1520 using BlockFrequencyInfoImplBase::printBlockFreq;
1521 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockT *BB) const {
1522 return BlockFrequencyInfoImplBase::printBlockFreq(OS, getNode(BB));
1527 void BlockFrequencyInfoImpl<BT>::doFunction(const FunctionT *F,
1528 const BranchProbabilityInfoT *BPI,
1529 const LoopInfoT *LI) {
1530 // Save the parameters.
1535 // Clean up left-over data structures.
1536 BlockFrequencyInfoImplBase::clear();
1541 DEBUG(dbgs() << "\nblock-frequency: " << F->getName() << "\n================="
1542 << std::string(F->getName().size(), '=') << "\n");
1546 // Visit loops in post-order to find thelocal mass distribution, and then do
1547 // the full function.
1548 computeMassInLoops();
1549 computeMassInFunction();
1554 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeRPOT() {
1555 const BlockT *Entry = F->begin();
1556 RPOT.reserve(F->size());
1557 std::copy(po_begin(Entry), po_end(Entry), std::back_inserter(RPOT));
1558 std::reverse(RPOT.begin(), RPOT.end());
1560 assert(RPOT.size() - 1 <= BlockNode::getMaxIndex() &&
1561 "More nodes in function than Block Frequency Info supports");
1563 DEBUG(dbgs() << "reverse-post-order-traversal\n");
1564 for (rpot_iterator I = rpot_begin(), E = rpot_end(); I != E; ++I) {
1565 BlockNode Node = getNode(I);
1566 DEBUG(dbgs() << " - " << getIndex(I) << ": " << getBlockName(Node) << "\n");
1570 Working.reserve(RPOT.size());
1571 for (size_t Index = 0; Index < RPOT.size(); ++Index)
1572 Working.emplace_back(Index);
1573 Freqs.resize(RPOT.size());
1576 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeLoops() {
1577 DEBUG(dbgs() << "loop-detection\n");
1581 // Visit loops top down and assign them an index.
1582 std::deque<std::pair<const LoopT *, LoopData *>> Q;
1583 for (const LoopT *L : *LI)
1584 Q.emplace_back(L, nullptr);
1585 while (!Q.empty()) {
1586 const LoopT *Loop = Q.front().first;
1587 LoopData *Parent = Q.front().second;
1590 BlockNode Header = getNode(Loop->getHeader());
1591 assert(Header.isValid());
1593 Loops.emplace_back(Parent, Header);
1594 Working[Header.Index].Loop = &Loops.back();
1595 DEBUG(dbgs() << " - loop = " << getBlockName(Header) << "\n");
1597 for (const LoopT *L : *Loop)
1598 Q.emplace_back(L, &Loops.back());
1601 // Visit nodes in reverse post-order and add them to their deepest containing
1603 for (size_t Index = 0; Index < RPOT.size(); ++Index) {
1604 // Loop headers have already been mostly mapped.
1605 if (Working[Index].isLoopHeader()) {
1606 LoopData *ContainingLoop = Working[Index].getContainingLoop();
1608 ContainingLoop->Nodes.push_back(Index);
1612 const LoopT *Loop = LI->getLoopFor(RPOT[Index]);
1616 // Add this node to its containing loop's member list.
1617 BlockNode Header = getNode(Loop->getHeader());
1618 assert(Header.isValid());
1619 const auto &HeaderData = Working[Header.Index];
1620 assert(HeaderData.isLoopHeader());
1622 Working[Index].Loop = HeaderData.Loop;
1623 HeaderData.Loop->Nodes.push_back(Index);
1624 DEBUG(dbgs() << " - loop = " << getBlockName(Header)
1625 << ": member = " << getBlockName(Index) << "\n");
1629 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInLoops() {
1630 // Visit loops with the deepest first, and the top-level loops last.
1631 for (auto L = Loops.rbegin(), E = Loops.rend(); L != E; ++L) {
1632 if (computeMassInLoop(*L))
1634 auto Next = std::next(L);
1635 computeIrreducibleMass(&*L, L.base());
1636 L = std::prev(Next);
1637 if (computeMassInLoop(*L))
1639 llvm_unreachable("unhandled irreducible control flow");
1644 bool BlockFrequencyInfoImpl<BT>::computeMassInLoop(LoopData &Loop) {
1645 // Compute mass in loop.
1646 DEBUG(dbgs() << "compute-mass-in-loop: " << getLoopName(Loop) << "\n");
1648 if (Loop.isIrreducible()) {
1649 BlockMass Remaining = BlockMass::getFull();
1650 for (uint32_t H = 0; H < Loop.NumHeaders; ++H) {
1651 auto &Mass = Working[Loop.Nodes[H].Index].getMass();
1652 Mass = Remaining * BranchProbability(1, Loop.NumHeaders - H);
1655 for (const BlockNode &M : Loop.Nodes)
1656 if (!propagateMassToSuccessors(&Loop, M))
1657 llvm_unreachable("unhandled irreducible control flow");
1659 Working[Loop.getHeader().Index].getMass() = BlockMass::getFull();
1660 if (!propagateMassToSuccessors(&Loop, Loop.getHeader()))
1661 llvm_unreachable("irreducible control flow to loop header!?");
1662 for (const BlockNode &M : Loop.members())
1663 if (!propagateMassToSuccessors(&Loop, M))
1664 // Irreducible backedge.
1668 computeLoopScale(Loop);
1674 bool BlockFrequencyInfoImpl<BT>::tryToComputeMassInFunction() {
1675 // Compute mass in function.
1676 DEBUG(dbgs() << "compute-mass-in-function\n");
1677 assert(!Working.empty() && "no blocks in function");
1678 assert(!Working[0].isLoopHeader() && "entry block is a loop header");
1680 Working[0].getMass() = BlockMass::getFull();
1681 for (rpot_iterator I = rpot_begin(), IE = rpot_end(); I != IE; ++I) {
1682 // Check for nodes that have been packaged.
1683 BlockNode Node = getNode(I);
1684 if (Working[Node.Index].isPackaged())
1687 if (!propagateMassToSuccessors(nullptr, Node))
1693 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInFunction() {
1694 if (tryToComputeMassInFunction())
1696 computeIrreducibleMass(nullptr, Loops.begin());
1697 if (tryToComputeMassInFunction())
1699 llvm_unreachable("unhandled irreducible control flow");
1702 /// \note This should be a lambda, but that crashes GCC 4.7.
1703 namespace bfi_detail {
1704 template <class BT> struct BlockEdgesAdder {
1706 typedef BlockFrequencyInfoImplBase::LoopData LoopData;
1707 typedef GraphTraits<const BlockT *> Successor;
1709 const BlockFrequencyInfoImpl<BT> &BFI;
1710 explicit BlockEdgesAdder(const BlockFrequencyInfoImpl<BT> &BFI)
1712 void operator()(IrreducibleGraph &G, IrreducibleGraph::IrrNode &Irr,
1713 const LoopData *OuterLoop) {
1714 const BlockT *BB = BFI.RPOT[Irr.Node.Index];
1715 for (auto I = Successor::child_begin(BB), E = Successor::child_end(BB);
1717 G.addEdge(Irr, BFI.getNode(*I), OuterLoop);
1722 void BlockFrequencyInfoImpl<BT>::computeIrreducibleMass(
1723 LoopData *OuterLoop, std::list<LoopData>::iterator Insert) {
1724 DEBUG(dbgs() << "analyze-irreducible-in-";
1725 if (OuterLoop) dbgs() << "loop: " << getLoopName(*OuterLoop) << "\n";
1726 else dbgs() << "function\n");
1728 using namespace bfi_detail;
1729 // Ideally, addBlockEdges() would be declared here as a lambda, but that
1731 BlockEdgesAdder<BT> addBlockEdges(*this);
1732 IrreducibleGraph G(*this, OuterLoop, addBlockEdges);
1734 for (auto &L : analyzeIrreducible(G, OuterLoop, Insert))
1735 computeMassInLoop(L);
1739 updateLoopWithIrreducible(*OuterLoop);
1744 BlockFrequencyInfoImpl<BT>::propagateMassToSuccessors(LoopData *OuterLoop,
1745 const BlockNode &Node) {
1746 DEBUG(dbgs() << " - node: " << getBlockName(Node) << "\n");
1747 // Calculate probability for successors.
1749 if (auto *Loop = Working[Node.Index].getPackagedLoop()) {
1750 assert(Loop != OuterLoop && "Cannot propagate mass in a packaged loop");
1751 if (!addLoopSuccessorsToDist(OuterLoop, *Loop, Dist))
1752 // Irreducible backedge.
1755 const BlockT *BB = getBlock(Node);
1756 for (auto SI = Successor::child_begin(BB), SE = Successor::child_end(BB);
1758 // Do not dereference SI, or getEdgeWeight() is linear in the number of
1760 if (!addToDist(Dist, OuterLoop, Node, getNode(*SI),
1761 BPI->getEdgeWeight(BB, SI)))
1762 // Irreducible backedge.
1766 // Distribute mass to successors, saving exit and backedge data in the
1768 distributeMass(Node, OuterLoop, Dist);
1773 raw_ostream &BlockFrequencyInfoImpl<BT>::print(raw_ostream &OS) const {
1776 OS << "block-frequency-info: " << F->getName() << "\n";
1777 for (const BlockT &BB : *F)
1778 OS << " - " << bfi_detail::getBlockName(&BB)
1779 << ": float = " << getFloatingBlockFreq(&BB)
1780 << ", int = " << getBlockFreq(&BB).getFrequency() << "\n";
1782 // Add an extra newline for readability.