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 int32_t extractLg(const std::pair<int32_t, int> &Lg) {
73 static int32_t extractLgFloor(const std::pair<int32_t, int> &Lg) {
74 return Lg.first - (Lg.second > 0);
76 static int32_t extractLgCeiling(const std::pair<int32_t, int> &Lg) {
77 return Lg.first + (Lg.second < 0);
80 static int compare(uint64_t L, uint64_t R, int Shift) {
84 uint64_t L_adjusted = L >> Shift;
90 return L > L_adjusted << Shift ? 1 : 0;
94 /// \brief Simple representation of an unsigned floating point.
96 /// UnsignedFloat is a unsigned floating point number. It uses simple
97 /// saturation arithmetic, and every operation is well-defined for every value.
99 /// The number is split into a signed exponent and unsigned digits. The number
100 /// represented is \c getDigits()*2^getExponent(). In this way, the digits are
101 /// much like the mantissa in the x87 long double, but there is no canonical
102 /// form, so the same number can be represented by many bit representations
103 /// (it's always in "denormal" mode).
105 /// UnsignedFloat is templated on the underlying integer type for digits, which
106 /// is expected to be one of uint64_t, uint32_t, uint16_t or uint8_t.
108 /// Unlike builtin floating point types, UnsignedFloat is portable.
110 /// Unlike APFloat, UnsignedFloat does not model architecture floating point
111 /// behaviour (this should make it a little faster), and implements most
112 /// operators (this makes it usable).
114 /// UnsignedFloat is totally ordered. However, there is no canonical form, so
115 /// there are multiple representations of most scalars. E.g.:
117 /// UnsignedFloat(8u, 0) == UnsignedFloat(4u, 1)
118 /// UnsignedFloat(4u, 1) == UnsignedFloat(2u, 2)
119 /// UnsignedFloat(2u, 2) == UnsignedFloat(1u, 3)
121 /// UnsignedFloat implements most arithmetic operations. Precision is kept
122 /// where possible. Uses simple saturation arithmetic, so that operations
123 /// saturate to 0.0 or getLargest() rather than under or overflowing. It has
124 /// some extra arithmetic for unit inversion. 0.0/0.0 is defined to be 0.0.
125 /// Any other division by 0.0 is defined to be getLargest().
127 /// As a convenience for modifying the exponent, left and right shifting are
128 /// both implemented, and both interpret negative shifts as positive shifts in
129 /// the opposite direction.
131 /// Exponents are limited to the range accepted by x87 long double. This makes
132 /// it trivial to add functionality to convert to APFloat (this is already
133 /// relied on for the implementation of printing).
135 /// The current plan is to gut this and make the necessary parts of it (even
136 /// more) private to BlockFrequencyInfo.
137 template <class DigitsT> class UnsignedFloat : UnsignedFloatBase {
139 static_assert(!std::numeric_limits<DigitsT>::is_signed,
140 "only unsigned floats supported");
142 typedef DigitsT DigitsType;
145 typedef std::numeric_limits<DigitsType> DigitsLimits;
147 static const int Width = sizeof(DigitsType) * 8;
148 static_assert(Width <= 64, "invalid integer width for digits");
155 UnsignedFloat() : Digits(0), Exponent(0) {}
157 UnsignedFloat(DigitsType Digits, int16_t Exponent)
158 : Digits(Digits), Exponent(Exponent) {}
161 UnsignedFloat(const std::pair<uint64_t, int16_t> &X)
162 : Digits(X.first), Exponent(X.second) {}
165 static UnsignedFloat getZero() { return UnsignedFloat(0, 0); }
166 static UnsignedFloat getOne() { return UnsignedFloat(1, 0); }
167 static UnsignedFloat getLargest() {
168 return UnsignedFloat(DigitsLimits::max(), MaxExponent);
170 static UnsignedFloat getFloat(uint64_t N) { return adjustToWidth(N, 0); }
171 static UnsignedFloat getInverseFloat(uint64_t N) {
172 return getFloat(N).invert();
174 static UnsignedFloat getFraction(DigitsType N, DigitsType D) {
175 return getQuotient(N, D);
178 int16_t getExponent() const { return Exponent; }
179 DigitsType getDigits() const { return Digits; }
181 /// \brief Convert to the given integer type.
183 /// Convert to \c IntT using simple saturating arithmetic, truncating if
185 template <class IntT> IntT toInt() const;
187 bool isZero() const { return !Digits; }
188 bool isLargest() const { return *this == getLargest(); }
190 if (Exponent > 0 || Exponent <= -Width)
192 return Digits == DigitsType(1) << -Exponent;
195 /// \brief The log base 2, rounded.
197 /// Get the lg of the scalar. lg 0 is defined to be INT32_MIN.
198 int32_t lg() const { return extractLg(lgImpl()); }
200 /// \brief The log base 2, rounded towards INT32_MIN.
202 /// Get the lg floor. lg 0 is defined to be INT32_MIN.
203 int32_t lgFloor() const { return extractLgFloor(lgImpl()); }
205 /// \brief The log base 2, rounded towards INT32_MAX.
207 /// Get the lg ceiling. lg 0 is defined to be INT32_MIN.
208 int32_t lgCeiling() const { return extractLgCeiling(lgImpl()); }
210 bool operator==(const UnsignedFloat &X) const { return compare(X) == 0; }
211 bool operator<(const UnsignedFloat &X) const { return compare(X) < 0; }
212 bool operator!=(const UnsignedFloat &X) const { return compare(X) != 0; }
213 bool operator>(const UnsignedFloat &X) const { return compare(X) > 0; }
214 bool operator<=(const UnsignedFloat &X) const { return compare(X) <= 0; }
215 bool operator>=(const UnsignedFloat &X) const { return compare(X) >= 0; }
217 bool operator!() const { return isZero(); }
219 /// \brief Convert to a decimal representation in a string.
221 /// Convert to a string. Uses scientific notation for very large/small
222 /// numbers. Scientific notation is used roughly for numbers outside of the
223 /// range 2^-64 through 2^64.
225 /// \c Precision indicates the number of decimal digits of precision to use;
226 /// 0 requests the maximum available.
228 /// As a special case to make debugging easier, if the number is small enough
229 /// to convert without scientific notation and has more than \c Precision
230 /// digits before the decimal place, it's printed accurately to the first
231 /// digit past zero. E.g., assuming 10 digits of precision:
233 /// 98765432198.7654... => 98765432198.8
234 /// 8765432198.7654... => 8765432198.8
235 /// 765432198.7654... => 765432198.8
236 /// 65432198.7654... => 65432198.77
237 /// 5432198.7654... => 5432198.765
238 std::string toString(unsigned Precision = DefaultPrecision) {
239 return UnsignedFloatBase::toString(Digits, Exponent, Width, Precision);
242 /// \brief Print a decimal representation.
244 /// Print a string. See toString for documentation.
245 raw_ostream &print(raw_ostream &OS,
246 unsigned Precision = DefaultPrecision) const {
247 return UnsignedFloatBase::print(OS, Digits, Exponent, Width, Precision);
249 void dump() const { return UnsignedFloatBase::dump(Digits, Exponent, Width); }
251 UnsignedFloat &operator+=(const UnsignedFloat &X);
252 UnsignedFloat &operator-=(const UnsignedFloat &X);
253 UnsignedFloat &operator*=(const UnsignedFloat &X);
254 UnsignedFloat &operator/=(const UnsignedFloat &X);
255 UnsignedFloat &operator<<=(int16_t Shift) { shiftLeft(Shift); return *this; }
256 UnsignedFloat &operator>>=(int16_t Shift) { shiftRight(Shift); return *this; }
259 void shiftLeft(int32_t Shift);
260 void shiftRight(int32_t Shift);
262 /// \brief Adjust two floats to have matching exponents.
264 /// Adjust \c this and \c X to have matching exponents. Returns the new \c X
265 /// by value. Does nothing if \a isZero() for either.
267 /// The value that compares smaller will lose precision, and possibly become
269 UnsignedFloat matchExponents(UnsignedFloat X);
271 /// \brief Increase exponent to match another float.
273 /// Increases \c this to have an exponent matching \c X. May decrease the
274 /// exponent of \c X in the process, and \c this may possibly become \a
276 void increaseExponentToMatch(UnsignedFloat &X, int32_t ExponentDiff);
279 /// \brief Scale a large number accurately.
281 /// Scale N (multiply it by this). Uses full precision multiplication, even
282 /// if Width is smaller than 64, so information is not lost.
283 uint64_t scale(uint64_t N) const;
284 uint64_t scaleByInverse(uint64_t N) const {
285 // TODO: implement directly, rather than relying on inverse. Inverse is
287 return inverse().scale(N);
289 int64_t scale(int64_t N) const {
290 std::pair<uint64_t, bool> Unsigned = splitSigned(N);
291 return joinSigned(scale(Unsigned.first), Unsigned.second);
293 int64_t scaleByInverse(int64_t N) const {
294 std::pair<uint64_t, bool> Unsigned = splitSigned(N);
295 return joinSigned(scaleByInverse(Unsigned.first), Unsigned.second);
298 int compare(const UnsignedFloat &X) const;
299 int compareTo(uint64_t N) const {
300 UnsignedFloat Float = getFloat(N);
301 int Compare = compare(Float);
302 if (Width == 64 || Compare != 0)
305 // Check for precision loss. We know *this == RoundTrip.
306 uint64_t RoundTrip = Float.template toInt<uint64_t>();
307 return N == RoundTrip ? 0 : RoundTrip < N ? -1 : 1;
309 int compareTo(int64_t N) const { return N < 0 ? 1 : compareTo(uint64_t(N)); }
311 UnsignedFloat &invert() { return *this = UnsignedFloat::getFloat(1) / *this; }
312 UnsignedFloat inverse() const { return UnsignedFloat(*this).invert(); }
315 static UnsignedFloat getProduct(DigitsType LHS, DigitsType RHS) {
316 return ScaledNumbers::getProduct(LHS, RHS);
318 static UnsignedFloat getQuotient(DigitsType Dividend, DigitsType Divisor) {
319 return ScaledNumbers::getQuotient(Dividend, Divisor);
322 std::pair<int32_t, int> lgImpl() const;
323 static int countLeadingZerosWidth(DigitsType Digits) {
325 return countLeadingZeros64(Digits);
327 return countLeadingZeros32(Digits);
328 return countLeadingZeros32(Digits) + Width - 32;
331 /// \brief Adjust a number to width, rounding up if necessary.
333 /// Should only be called for \c Shift close to zero.
335 /// \pre Shift >= MinExponent && Shift + 64 <= MaxExponent.
336 static UnsignedFloat adjustToWidth(uint64_t N, int32_t Shift) {
337 assert(Shift >= MinExponent && "Shift should be close to 0");
338 assert(Shift <= MaxExponent - 64 && "Shift should be close to 0");
339 auto Adjusted = ScaledNumbers::getAdjusted<DigitsT>(N, Shift);
343 static UnsignedFloat getRounded(UnsignedFloat P, bool Round) {
348 return ScaledNumbers::getRounded(P.Digits, P.Exponent, Round);
352 #define UNSIGNED_FLOAT_BOP(op, base) \
353 template <class DigitsT> \
354 UnsignedFloat<DigitsT> operator op(const UnsignedFloat<DigitsT> &L, \
355 const UnsignedFloat<DigitsT> &R) { \
356 return UnsignedFloat<DigitsT>(L) base R; \
358 UNSIGNED_FLOAT_BOP(+, += )
359 UNSIGNED_FLOAT_BOP(-, -= )
360 UNSIGNED_FLOAT_BOP(*, *= )
361 UNSIGNED_FLOAT_BOP(/, /= )
362 UNSIGNED_FLOAT_BOP(<<, <<= )
363 UNSIGNED_FLOAT_BOP(>>, >>= )
364 #undef UNSIGNED_FLOAT_BOP
366 template <class DigitsT>
367 raw_ostream &operator<<(raw_ostream &OS, const UnsignedFloat<DigitsT> &X) {
368 return X.print(OS, 10);
371 #define UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, T1, T2) \
372 template <class DigitsT> \
373 bool operator op(const UnsignedFloat<DigitsT> &L, T1 R) { \
374 return L.compareTo(T2(R)) op 0; \
376 template <class DigitsT> \
377 bool operator op(T1 L, const UnsignedFloat<DigitsT> &R) { \
378 return 0 op R.compareTo(T2(L)); \
380 #define UNSIGNED_FLOAT_COMPARE_TO(op) \
381 UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, uint64_t, uint64_t) \
382 UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, uint32_t, uint64_t) \
383 UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, int64_t, int64_t) \
384 UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, int32_t, int64_t)
385 UNSIGNED_FLOAT_COMPARE_TO(< )
386 UNSIGNED_FLOAT_COMPARE_TO(> )
387 UNSIGNED_FLOAT_COMPARE_TO(== )
388 UNSIGNED_FLOAT_COMPARE_TO(!= )
389 UNSIGNED_FLOAT_COMPARE_TO(<= )
390 UNSIGNED_FLOAT_COMPARE_TO(>= )
391 #undef UNSIGNED_FLOAT_COMPARE_TO
392 #undef UNSIGNED_FLOAT_COMPARE_TO_TYPE
394 template <class DigitsT>
395 uint64_t UnsignedFloat<DigitsT>::scale(uint64_t N) const {
396 if (Width == 64 || N <= DigitsLimits::max())
397 return (getFloat(N) * *this).template toInt<uint64_t>();
399 // Defer to the 64-bit version.
400 return UnsignedFloat<uint64_t>(Digits, Exponent).scale(N);
403 template <class DigitsT>
404 template <class IntT>
405 IntT UnsignedFloat<DigitsT>::toInt() const {
406 typedef std::numeric_limits<IntT> Limits;
409 if (*this >= Limits::max())
410 return Limits::max();
414 assert(size_t(Exponent) < sizeof(IntT) * 8);
415 return N << Exponent;
418 assert(size_t(-Exponent) < sizeof(IntT) * 8);
419 return N >> -Exponent;
424 template <class DigitsT>
425 std::pair<int32_t, int> UnsignedFloat<DigitsT>::lgImpl() const {
427 return std::make_pair(INT32_MIN, 0);
429 // Get the floor of the lg of Digits.
430 int32_t LocalFloor = Width - countLeadingZerosWidth(Digits) - 1;
432 // Get the floor of the lg of this.
433 int32_t Floor = Exponent + LocalFloor;
434 if (Digits == UINT64_C(1) << LocalFloor)
435 return std::make_pair(Floor, 0);
437 // Round based on the next digit.
438 assert(LocalFloor >= 1);
439 bool Round = Digits & UINT64_C(1) << (LocalFloor - 1);
440 return std::make_pair(Floor + Round, Round ? 1 : -1);
443 template <class DigitsT>
444 UnsignedFloat<DigitsT> UnsignedFloat<DigitsT>::matchExponents(UnsignedFloat X) {
445 if (isZero() || X.isZero() || Exponent == X.Exponent)
448 int32_t Diff = int32_t(X.Exponent) - int32_t(Exponent);
450 increaseExponentToMatch(X, Diff);
452 X.increaseExponentToMatch(*this, -Diff);
455 template <class DigitsT>
456 void UnsignedFloat<DigitsT>::increaseExponentToMatch(UnsignedFloat &X,
457 int32_t ExponentDiff) {
458 assert(ExponentDiff > 0);
459 if (ExponentDiff >= 2 * Width) {
464 // Use up any leading zeros on X, and then shift this.
465 int32_t ShiftX = std::min(countLeadingZerosWidth(X.Digits), ExponentDiff);
466 assert(ShiftX < Width);
468 int32_t ShiftThis = ExponentDiff - ShiftX;
469 if (ShiftThis >= Width) {
475 X.Exponent -= ShiftX;
476 Digits >>= ShiftThis;
477 Exponent += ShiftThis;
481 template <class DigitsT>
482 UnsignedFloat<DigitsT> &UnsignedFloat<DigitsT>::
483 operator+=(const UnsignedFloat &X) {
484 if (isLargest() || X.isZero())
486 if (isZero() || X.isLargest())
489 // Normalize exponents.
490 UnsignedFloat Scaled = matchExponents(X);
492 // Check for zero again.
494 return *this = Scaled;
499 DigitsType Sum = Digits + Scaled.Digits;
500 bool DidOverflow = Sum < Digits;
505 if (Exponent == MaxExponent)
506 return *this = getLargest();
509 Digits = UINT64_C(1) << (Width - 1) | Digits >> 1;
513 template <class DigitsT>
514 UnsignedFloat<DigitsT> &UnsignedFloat<DigitsT>::
515 operator-=(const UnsignedFloat &X) {
519 return *this = getZero();
521 // Normalize exponents.
522 UnsignedFloat Scaled = matchExponents(X);
523 assert(Digits >= Scaled.Digits);
525 // Compute difference.
526 if (!Scaled.isZero()) {
527 Digits -= Scaled.Digits;
531 // Check if X just barely lost its last bit. E.g., for 32-bit:
533 // 1*2^32 - 1*2^0 == 0xffffffff != 1*2^32
534 if (*this == UnsignedFloat(1, X.lgFloor() + Width)) {
535 Digits = DigitsType(0) - 1;
540 template <class DigitsT>
541 UnsignedFloat<DigitsT> &UnsignedFloat<DigitsT>::
542 operator*=(const UnsignedFloat &X) {
548 // Save the exponents.
549 int32_t Exponents = int32_t(Exponent) + int32_t(X.Exponent);
551 // Get the raw product.
552 *this = getProduct(Digits, X.Digits);
554 // Combine with exponents.
555 return *this <<= Exponents;
557 template <class DigitsT>
558 UnsignedFloat<DigitsT> &UnsignedFloat<DigitsT>::
559 operator/=(const UnsignedFloat &X) {
563 return *this = getLargest();
565 // Save the exponents.
566 int32_t Exponents = int32_t(Exponent) - int32_t(X.Exponent);
568 // Get the raw quotient.
569 *this = getQuotient(Digits, X.Digits);
571 // Combine with exponents.
572 return *this <<= Exponents;
574 template <class DigitsT>
575 void UnsignedFloat<DigitsT>::shiftLeft(int32_t Shift) {
576 if (!Shift || isZero())
578 assert(Shift != INT32_MIN);
584 // Shift as much as we can in the exponent.
585 int32_t ExponentShift = std::min(Shift, MaxExponent - Exponent);
586 Exponent += ExponentShift;
587 if (ExponentShift == Shift)
590 // Check this late, since it's rare.
594 // Shift the digits themselves.
595 Shift -= ExponentShift;
596 if (Shift > countLeadingZerosWidth(Digits)) {
598 *this = getLargest();
606 template <class DigitsT>
607 void UnsignedFloat<DigitsT>::shiftRight(int32_t Shift) {
608 if (!Shift || isZero())
610 assert(Shift != INT32_MIN);
616 // Shift as much as we can in the exponent.
617 int32_t ExponentShift = std::min(Shift, Exponent - MinExponent);
618 Exponent -= ExponentShift;
619 if (ExponentShift == Shift)
622 // Shift the digits themselves.
623 Shift -= ExponentShift;
624 if (Shift >= Width) {
634 template <class DigitsT>
635 int UnsignedFloat<DigitsT>::compare(const UnsignedFloat &X) const {
638 return X.isZero() ? 0 : -1;
642 // Check for the scale. Use lgFloor to be sure that the exponent difference
643 // is always lower than 64.
644 int32_t lgL = lgFloor(), lgR = X.lgFloor();
646 return lgL < lgR ? -1 : 1;
649 if (Exponent < X.Exponent)
650 return UnsignedFloatBase::compare(Digits, X.Digits, X.Exponent - Exponent);
652 return -UnsignedFloatBase::compare(X.Digits, Digits, Exponent - X.Exponent);
655 template <class T> struct isPodLike<UnsignedFloat<T>> {
656 static const bool value = true;
660 //===----------------------------------------------------------------------===//
662 // BlockMass definition.
664 // TODO: Make this private to BlockFrequencyInfoImpl or delete.
666 //===----------------------------------------------------------------------===//
669 /// \brief Mass of a block.
671 /// This class implements a sort of fixed-point fraction always between 0.0 and
672 /// 1.0. getMass() == UINT64_MAX indicates a value of 1.0.
674 /// Masses can be added and subtracted. Simple saturation arithmetic is used,
675 /// so arithmetic operations never overflow or underflow.
677 /// Masses can be multiplied. Multiplication treats full mass as 1.0 and uses
678 /// an inexpensive floating-point algorithm that's off-by-one (almost, but not
679 /// quite, maximum precision).
681 /// Masses can be scaled by \a BranchProbability at maximum precision.
686 BlockMass() : Mass(0) {}
687 explicit BlockMass(uint64_t Mass) : Mass(Mass) {}
689 static BlockMass getEmpty() { return BlockMass(); }
690 static BlockMass getFull() { return BlockMass(UINT64_MAX); }
692 uint64_t getMass() const { return Mass; }
694 bool isFull() const { return Mass == UINT64_MAX; }
695 bool isEmpty() const { return !Mass; }
697 bool operator!() const { return isEmpty(); }
699 /// \brief Add another mass.
701 /// Adds another mass, saturating at \a isFull() rather than overflowing.
702 BlockMass &operator+=(const BlockMass &X) {
703 uint64_t Sum = Mass + X.Mass;
704 Mass = Sum < Mass ? UINT64_MAX : Sum;
708 /// \brief Subtract another mass.
710 /// Subtracts another mass, saturating at \a isEmpty() rather than
712 BlockMass &operator-=(const BlockMass &X) {
713 uint64_t Diff = Mass - X.Mass;
714 Mass = Diff > Mass ? 0 : Diff;
718 BlockMass &operator*=(const BranchProbability &P) {
719 Mass = P.scale(Mass);
723 bool operator==(const BlockMass &X) const { return Mass == X.Mass; }
724 bool operator!=(const BlockMass &X) const { return Mass != X.Mass; }
725 bool operator<=(const BlockMass &X) const { return Mass <= X.Mass; }
726 bool operator>=(const BlockMass &X) const { return Mass >= X.Mass; }
727 bool operator<(const BlockMass &X) const { return Mass < X.Mass; }
728 bool operator>(const BlockMass &X) const { return Mass > X.Mass; }
730 /// \brief Convert to floating point.
732 /// Convert to a float. \a isFull() gives 1.0, while \a isEmpty() gives
733 /// slightly above 0.0.
734 UnsignedFloat<uint64_t> toFloat() const;
737 raw_ostream &print(raw_ostream &OS) const;
740 inline BlockMass operator+(const BlockMass &L, const BlockMass &R) {
741 return BlockMass(L) += R;
743 inline BlockMass operator-(const BlockMass &L, const BlockMass &R) {
744 return BlockMass(L) -= R;
746 inline BlockMass operator*(const BlockMass &L, const BranchProbability &R) {
747 return BlockMass(L) *= R;
749 inline BlockMass operator*(const BranchProbability &L, const BlockMass &R) {
750 return BlockMass(R) *= L;
753 inline raw_ostream &operator<<(raw_ostream &OS, const BlockMass &X) {
757 template <> struct isPodLike<BlockMass> {
758 static const bool value = true;
762 //===----------------------------------------------------------------------===//
764 // BlockFrequencyInfoImpl definition.
766 //===----------------------------------------------------------------------===//
770 class BranchProbabilityInfo;
774 class MachineBasicBlock;
775 class MachineBranchProbabilityInfo;
776 class MachineFunction;
778 class MachineLoopInfo;
780 namespace bfi_detail {
781 struct IrreducibleGraph;
783 // This is part of a workaround for a GCC 4.7 crash on lambdas.
784 template <class BT> struct BlockEdgesAdder;
787 /// \brief Base class for BlockFrequencyInfoImpl
789 /// BlockFrequencyInfoImplBase has supporting data structures and some
790 /// algorithms for BlockFrequencyInfoImplBase. Only algorithms that depend on
791 /// the block type (or that call such algorithms) are skipped here.
793 /// Nevertheless, the majority of the overall algorithm documention lives with
794 /// BlockFrequencyInfoImpl. See there for details.
795 class BlockFrequencyInfoImplBase {
797 typedef UnsignedFloat<uint64_t> Float;
799 /// \brief Representative of a block.
801 /// This is a simple wrapper around an index into the reverse-post-order
802 /// traversal of the blocks.
804 /// Unlike a block pointer, its order has meaning (location in the
805 /// topological sort) and it's class is the same regardless of block type.
807 typedef uint32_t IndexType;
810 bool operator==(const BlockNode &X) const { return Index == X.Index; }
811 bool operator!=(const BlockNode &X) const { return Index != X.Index; }
812 bool operator<=(const BlockNode &X) const { return Index <= X.Index; }
813 bool operator>=(const BlockNode &X) const { return Index >= X.Index; }
814 bool operator<(const BlockNode &X) const { return Index < X.Index; }
815 bool operator>(const BlockNode &X) const { return Index > X.Index; }
817 BlockNode() : Index(UINT32_MAX) {}
818 BlockNode(IndexType Index) : Index(Index) {}
820 bool isValid() const { return Index <= getMaxIndex(); }
821 static size_t getMaxIndex() { return UINT32_MAX - 1; }
824 /// \brief Stats about a block itself.
825 struct FrequencyData {
830 /// \brief Data about a loop.
832 /// Contains the data necessary to represent represent a loop as a
833 /// pseudo-node once it's packaged.
835 typedef SmallVector<std::pair<BlockNode, BlockMass>, 4> ExitMap;
836 typedef SmallVector<BlockNode, 4> NodeList;
837 LoopData *Parent; ///< The parent loop.
838 bool IsPackaged; ///< Whether this has been packaged.
839 uint32_t NumHeaders; ///< Number of headers.
840 ExitMap Exits; ///< Successor edges (and weights).
841 NodeList Nodes; ///< Header and the members of the loop.
842 BlockMass BackedgeMass; ///< Mass returned to loop header.
846 LoopData(LoopData *Parent, const BlockNode &Header)
847 : Parent(Parent), IsPackaged(false), NumHeaders(1), Nodes(1, Header) {}
848 template <class It1, class It2>
849 LoopData(LoopData *Parent, It1 FirstHeader, It1 LastHeader, It2 FirstOther,
851 : Parent(Parent), IsPackaged(false), Nodes(FirstHeader, LastHeader) {
852 NumHeaders = Nodes.size();
853 Nodes.insert(Nodes.end(), FirstOther, LastOther);
855 bool isHeader(const BlockNode &Node) const {
857 return std::binary_search(Nodes.begin(), Nodes.begin() + NumHeaders,
859 return Node == Nodes[0];
861 BlockNode getHeader() const { return Nodes[0]; }
862 bool isIrreducible() const { return NumHeaders > 1; }
864 NodeList::const_iterator members_begin() const {
865 return Nodes.begin() + NumHeaders;
867 NodeList::const_iterator members_end() const { return Nodes.end(); }
868 iterator_range<NodeList::const_iterator> members() const {
869 return make_range(members_begin(), members_end());
873 /// \brief Index of loop information.
875 BlockNode Node; ///< This node.
876 LoopData *Loop; ///< The loop this block is inside.
877 BlockMass Mass; ///< Mass distribution from the entry block.
879 WorkingData(const BlockNode &Node) : Node(Node), Loop(nullptr) {}
881 bool isLoopHeader() const { return Loop && Loop->isHeader(Node); }
882 bool isDoubleLoopHeader() const {
883 return isLoopHeader() && Loop->Parent && Loop->Parent->isIrreducible() &&
884 Loop->Parent->isHeader(Node);
887 LoopData *getContainingLoop() const {
890 if (!isDoubleLoopHeader())
892 return Loop->Parent->Parent;
895 /// \brief Resolve a node to its representative.
897 /// Get the node currently representing Node, which could be a containing
900 /// This function should only be called when distributing mass. As long as
901 /// there are no irreducilbe edges to Node, then it will have complexity
902 /// O(1) in this context.
904 /// In general, the complexity is O(L), where L is the number of loop
905 /// headers Node has been packaged into. Since this method is called in
906 /// the context of distributing mass, L will be the number of loop headers
907 /// an early exit edge jumps out of.
908 BlockNode getResolvedNode() const {
909 auto L = getPackagedLoop();
910 return L ? L->getHeader() : Node;
912 LoopData *getPackagedLoop() const {
913 if (!Loop || !Loop->IsPackaged)
916 while (L->Parent && L->Parent->IsPackaged)
921 /// \brief Get the appropriate mass for a node.
923 /// Get appropriate mass for Node. If Node is a loop-header (whose loop
924 /// has been packaged), returns the mass of its pseudo-node. If it's a
925 /// node inside a packaged loop, it returns the loop's mass.
926 BlockMass &getMass() {
929 if (!isADoublePackage())
931 return Loop->Parent->Mass;
934 /// \brief Has ContainingLoop been packaged up?
935 bool isPackaged() const { return getResolvedNode() != Node; }
936 /// \brief Has Loop been packaged up?
937 bool isAPackage() const { return isLoopHeader() && Loop->IsPackaged; }
938 /// \brief Has Loop been packaged up twice?
939 bool isADoublePackage() const {
940 return isDoubleLoopHeader() && Loop->Parent->IsPackaged;
944 /// \brief Unscaled probability weight.
946 /// Probability weight for an edge in the graph (including the
947 /// successor/target node).
949 /// All edges in the original function are 32-bit. However, exit edges from
950 /// loop packages are taken from 64-bit exit masses, so we need 64-bits of
951 /// space in general.
953 /// In addition to the raw weight amount, Weight stores the type of the edge
954 /// in the current context (i.e., the context of the loop being processed).
955 /// Is this a local edge within the loop, an exit from the loop, or a
956 /// backedge to the loop header?
958 enum DistType { Local, Exit, Backedge };
960 BlockNode TargetNode;
962 Weight() : Type(Local), Amount(0) {}
965 /// \brief Distribution of unscaled probability weight.
967 /// Distribution of unscaled probability weight to a set of successors.
969 /// This class collates the successor edge weights for later processing.
971 /// \a DidOverflow indicates whether \a Total did overflow while adding to
972 /// the distribution. It should never overflow twice.
973 struct Distribution {
974 typedef SmallVector<Weight, 4> WeightList;
975 WeightList Weights; ///< Individual successor weights.
976 uint64_t Total; ///< Sum of all weights.
977 bool DidOverflow; ///< Whether \a Total did overflow.
979 Distribution() : Total(0), DidOverflow(false) {}
980 void addLocal(const BlockNode &Node, uint64_t Amount) {
981 add(Node, Amount, Weight::Local);
983 void addExit(const BlockNode &Node, uint64_t Amount) {
984 add(Node, Amount, Weight::Exit);
986 void addBackedge(const BlockNode &Node, uint64_t Amount) {
987 add(Node, Amount, Weight::Backedge);
990 /// \brief Normalize the distribution.
992 /// Combines multiple edges to the same \a Weight::TargetNode and scales
993 /// down so that \a Total fits into 32-bits.
995 /// This is linear in the size of \a Weights. For the vast majority of
996 /// cases, adjacent edge weights are combined by sorting WeightList and
997 /// combining adjacent weights. However, for very large edge lists an
998 /// auxiliary hash table is used.
1002 void add(const BlockNode &Node, uint64_t Amount, Weight::DistType Type);
1005 /// \brief Data about each block. This is used downstream.
1006 std::vector<FrequencyData> Freqs;
1008 /// \brief Loop data: see initializeLoops().
1009 std::vector<WorkingData> Working;
1011 /// \brief Indexed information about loops.
1012 std::list<LoopData> Loops;
1014 /// \brief Add all edges out of a packaged loop to the distribution.
1016 /// Adds all edges from LocalLoopHead to Dist. Calls addToDist() to add each
1019 /// \return \c true unless there's an irreducible backedge.
1020 bool addLoopSuccessorsToDist(const LoopData *OuterLoop, LoopData &Loop,
1021 Distribution &Dist);
1023 /// \brief Add an edge to the distribution.
1025 /// Adds an edge to Succ to Dist. If \c LoopHead.isValid(), then whether the
1026 /// edge is local/exit/backedge is in the context of LoopHead. Otherwise,
1027 /// every edge should be a local edge (since all the loops are packaged up).
1029 /// \return \c true unless aborted due to an irreducible backedge.
1030 bool addToDist(Distribution &Dist, const LoopData *OuterLoop,
1031 const BlockNode &Pred, const BlockNode &Succ, uint64_t Weight);
1033 LoopData &getLoopPackage(const BlockNode &Head) {
1034 assert(Head.Index < Working.size());
1035 assert(Working[Head.Index].isLoopHeader());
1036 return *Working[Head.Index].Loop;
1039 /// \brief Analyze irreducible SCCs.
1041 /// Separate irreducible SCCs from \c G, which is an explict graph of \c
1042 /// OuterLoop (or the top-level function, if \c OuterLoop is \c nullptr).
1043 /// Insert them into \a Loops before \c Insert.
1045 /// \return the \c LoopData nodes representing the irreducible SCCs.
1046 iterator_range<std::list<LoopData>::iterator>
1047 analyzeIrreducible(const bfi_detail::IrreducibleGraph &G, LoopData *OuterLoop,
1048 std::list<LoopData>::iterator Insert);
1050 /// \brief Update a loop after packaging irreducible SCCs inside of it.
1052 /// Update \c OuterLoop. Before finding irreducible control flow, it was
1053 /// partway through \a computeMassInLoop(), so \a LoopData::Exits and \a
1054 /// LoopData::BackedgeMass need to be reset. Also, nodes that were packaged
1055 /// up need to be removed from \a OuterLoop::Nodes.
1056 void updateLoopWithIrreducible(LoopData &OuterLoop);
1058 /// \brief Distribute mass according to a distribution.
1060 /// Distributes the mass in Source according to Dist. If LoopHead.isValid(),
1061 /// backedges and exits are stored in its entry in Loops.
1063 /// Mass is distributed in parallel from two copies of the source mass.
1064 void distributeMass(const BlockNode &Source, LoopData *OuterLoop,
1065 Distribution &Dist);
1067 /// \brief Compute the loop scale for a loop.
1068 void computeLoopScale(LoopData &Loop);
1070 /// \brief Package up a loop.
1071 void packageLoop(LoopData &Loop);
1073 /// \brief Unwrap loops.
1076 /// \brief Finalize frequency metrics.
1078 /// Calculates final frequencies and cleans up no-longer-needed data
1080 void finalizeMetrics();
1082 /// \brief Clear all memory.
1085 virtual std::string getBlockName(const BlockNode &Node) const;
1086 std::string getLoopName(const LoopData &Loop) const;
1088 virtual raw_ostream &print(raw_ostream &OS) const { return OS; }
1089 void dump() const { print(dbgs()); }
1091 Float getFloatingBlockFreq(const BlockNode &Node) const;
1093 BlockFrequency getBlockFreq(const BlockNode &Node) const;
1095 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockNode &Node) const;
1096 raw_ostream &printBlockFreq(raw_ostream &OS,
1097 const BlockFrequency &Freq) const;
1099 uint64_t getEntryFreq() const {
1100 assert(!Freqs.empty());
1101 return Freqs[0].Integer;
1103 /// \brief Virtual destructor.
1105 /// Need a virtual destructor to mask the compiler warning about
1107 virtual ~BlockFrequencyInfoImplBase() {}
1110 namespace bfi_detail {
1111 template <class BlockT> struct TypeMap {};
1112 template <> struct TypeMap<BasicBlock> {
1113 typedef BasicBlock BlockT;
1114 typedef Function FunctionT;
1115 typedef BranchProbabilityInfo BranchProbabilityInfoT;
1117 typedef LoopInfo LoopInfoT;
1119 template <> struct TypeMap<MachineBasicBlock> {
1120 typedef MachineBasicBlock BlockT;
1121 typedef MachineFunction FunctionT;
1122 typedef MachineBranchProbabilityInfo BranchProbabilityInfoT;
1123 typedef MachineLoop LoopT;
1124 typedef MachineLoopInfo LoopInfoT;
1127 /// \brief Get the name of a MachineBasicBlock.
1129 /// Get the name of a MachineBasicBlock. It's templated so that including from
1130 /// CodeGen is unnecessary (that would be a layering issue).
1132 /// This is used mainly for debug output. The name is similar to
1133 /// MachineBasicBlock::getFullName(), but skips the name of the function.
1134 template <class BlockT> std::string getBlockName(const BlockT *BB) {
1135 assert(BB && "Unexpected nullptr");
1136 auto MachineName = "BB" + Twine(BB->getNumber());
1137 if (BB->getBasicBlock())
1138 return (MachineName + "[" + BB->getName() + "]").str();
1139 return MachineName.str();
1141 /// \brief Get the name of a BasicBlock.
1142 template <> inline std::string getBlockName(const BasicBlock *BB) {
1143 assert(BB && "Unexpected nullptr");
1144 return BB->getName().str();
1147 /// \brief Graph of irreducible control flow.
1149 /// This graph is used for determining the SCCs in a loop (or top-level
1150 /// function) that has irreducible control flow.
1152 /// During the block frequency algorithm, the local graphs are defined in a
1153 /// light-weight way, deferring to the \a BasicBlock or \a MachineBasicBlock
1154 /// graphs for most edges, but getting others from \a LoopData::ExitMap. The
1155 /// latter only has successor information.
1157 /// \a IrreducibleGraph makes this graph explicit. It's in a form that can use
1158 /// \a GraphTraits (so that \a analyzeIrreducible() can use \a scc_iterator),
1159 /// and it explicitly lists predecessors and successors. The initialization
1160 /// that relies on \c MachineBasicBlock is defined in the header.
1161 struct IrreducibleGraph {
1162 typedef BlockFrequencyInfoImplBase BFIBase;
1166 typedef BFIBase::BlockNode BlockNode;
1170 std::deque<const IrrNode *> Edges;
1171 IrrNode(const BlockNode &Node) : Node(Node), NumIn(0) {}
1173 typedef std::deque<const IrrNode *>::const_iterator iterator;
1174 iterator pred_begin() const { return Edges.begin(); }
1175 iterator succ_begin() const { return Edges.begin() + NumIn; }
1176 iterator pred_end() const { return succ_begin(); }
1177 iterator succ_end() const { return Edges.end(); }
1180 const IrrNode *StartIrr;
1181 std::vector<IrrNode> Nodes;
1182 SmallDenseMap<uint32_t, IrrNode *, 4> Lookup;
1184 /// \brief Construct an explicit graph containing irreducible control flow.
1186 /// Construct an explicit graph of the control flow in \c OuterLoop (or the
1187 /// top-level function, if \c OuterLoop is \c nullptr). Uses \c
1188 /// addBlockEdges to add block successors that have not been packaged into
1191 /// \a BlockFrequencyInfoImpl::computeIrreducibleMass() is the only expected
1193 template <class BlockEdgesAdder>
1194 IrreducibleGraph(BFIBase &BFI, const BFIBase::LoopData *OuterLoop,
1195 BlockEdgesAdder addBlockEdges)
1196 : BFI(BFI), StartIrr(nullptr) {
1197 initialize(OuterLoop, addBlockEdges);
1200 template <class BlockEdgesAdder>
1201 void initialize(const BFIBase::LoopData *OuterLoop,
1202 BlockEdgesAdder addBlockEdges);
1203 void addNodesInLoop(const BFIBase::LoopData &OuterLoop);
1204 void addNodesInFunction();
1205 void addNode(const BlockNode &Node) {
1206 Nodes.emplace_back(Node);
1207 BFI.Working[Node.Index].getMass() = BlockMass::getEmpty();
1210 template <class BlockEdgesAdder>
1211 void addEdges(const BlockNode &Node, const BFIBase::LoopData *OuterLoop,
1212 BlockEdgesAdder addBlockEdges);
1213 void addEdge(IrrNode &Irr, const BlockNode &Succ,
1214 const BFIBase::LoopData *OuterLoop);
1216 template <class BlockEdgesAdder>
1217 void IrreducibleGraph::initialize(const BFIBase::LoopData *OuterLoop,
1218 BlockEdgesAdder addBlockEdges) {
1220 addNodesInLoop(*OuterLoop);
1221 for (auto N : OuterLoop->Nodes)
1222 addEdges(N, OuterLoop, addBlockEdges);
1224 addNodesInFunction();
1225 for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index)
1226 addEdges(Index, OuterLoop, addBlockEdges);
1228 StartIrr = Lookup[Start.Index];
1230 template <class BlockEdgesAdder>
1231 void IrreducibleGraph::addEdges(const BlockNode &Node,
1232 const BFIBase::LoopData *OuterLoop,
1233 BlockEdgesAdder addBlockEdges) {
1234 auto L = Lookup.find(Node.Index);
1235 if (L == Lookup.end())
1237 IrrNode &Irr = *L->second;
1238 const auto &Working = BFI.Working[Node.Index];
1240 if (Working.isAPackage())
1241 for (const auto &I : Working.Loop->Exits)
1242 addEdge(Irr, I.first, OuterLoop);
1244 addBlockEdges(*this, Irr, OuterLoop);
1248 /// \brief Shared implementation for block frequency analysis.
1250 /// This is a shared implementation of BlockFrequencyInfo and
1251 /// MachineBlockFrequencyInfo, and calculates the relative frequencies of
1254 /// LoopInfo defines a loop as a "non-trivial" SCC dominated by a single block,
1255 /// which is called the header. A given loop, L, can have sub-loops, which are
1256 /// loops within the subgraph of L that exclude its header. (A "trivial" SCC
1257 /// consists of a single block that does not have a self-edge.)
1259 /// In addition to loops, this algorithm has limited support for irreducible
1260 /// SCCs, which are SCCs with multiple entry blocks. Irreducible SCCs are
1261 /// discovered on they fly, and modelled as loops with multiple headers.
1263 /// The headers of irreducible sub-SCCs consist of its entry blocks and all
1264 /// nodes that are targets of a backedge within it (excluding backedges within
1265 /// true sub-loops). Block frequency calculations act as if a block is
1266 /// inserted that intercepts all the edges to the headers. All backedges and
1267 /// entries point to this block. Its successors are the headers, which split
1268 /// the frequency evenly.
1270 /// This algorithm leverages BlockMass and UnsignedFloat to maintain precision,
1271 /// separates mass distribution from loop scaling, and dithers to eliminate
1272 /// probability mass loss.
1274 /// The implementation is split between BlockFrequencyInfoImpl, which knows the
1275 /// type of graph being modelled (BasicBlock vs. MachineBasicBlock), and
1276 /// BlockFrequencyInfoImplBase, which doesn't. The base class uses \a
1277 /// BlockNode, a wrapper around a uint32_t. BlockNode is numbered from 0 in
1278 /// reverse-post order. This gives two advantages: it's easy to compare the
1279 /// relative ordering of two nodes, and maps keyed on BlockT can be represented
1282 /// This algorithm is O(V+E), unless there is irreducible control flow, in
1283 /// which case it's O(V*E) in the worst case.
1285 /// These are the main stages:
1287 /// 0. Reverse post-order traversal (\a initializeRPOT()).
1289 /// Run a single post-order traversal and save it (in reverse) in RPOT.
1290 /// All other stages make use of this ordering. Save a lookup from BlockT
1291 /// to BlockNode (the index into RPOT) in Nodes.
1293 /// 1. Loop initialization (\a initializeLoops()).
1295 /// Translate LoopInfo/MachineLoopInfo into a form suitable for the rest of
1296 /// the algorithm. In particular, store the immediate members of each loop
1297 /// in reverse post-order.
1299 /// 2. Calculate mass and scale in loops (\a computeMassInLoops()).
1301 /// For each loop (bottom-up), distribute mass through the DAG resulting
1302 /// from ignoring backedges and treating sub-loops as a single pseudo-node.
1303 /// Track the backedge mass distributed to the loop header, and use it to
1304 /// calculate the loop scale (number of loop iterations). Immediate
1305 /// members that represent sub-loops will already have been visited and
1306 /// packaged into a pseudo-node.
1308 /// Distributing mass in a loop is a reverse-post-order traversal through
1309 /// the loop. Start by assigning full mass to the Loop header. For each
1310 /// node in the loop:
1312 /// - Fetch and categorize the weight distribution for its successors.
1313 /// If this is a packaged-subloop, the weight distribution is stored
1314 /// in \a LoopData::Exits. Otherwise, fetch it from
1315 /// BranchProbabilityInfo.
1317 /// - Each successor is categorized as \a Weight::Local, a local edge
1318 /// within the current loop, \a Weight::Backedge, a backedge to the
1319 /// loop header, or \a Weight::Exit, any successor outside the loop.
1320 /// The weight, the successor, and its category are stored in \a
1321 /// Distribution. There can be multiple edges to each successor.
1323 /// - If there's a backedge to a non-header, there's an irreducible SCC.
1324 /// The usual flow is temporarily aborted. \a
1325 /// computeIrreducibleMass() finds the irreducible SCCs within the
1326 /// loop, packages them up, and restarts the flow.
1328 /// - Normalize the distribution: scale weights down so that their sum
1329 /// is 32-bits, and coalesce multiple edges to the same node.
1331 /// - Distribute the mass accordingly, dithering to minimize mass loss,
1332 /// as described in \a distributeMass().
1334 /// Finally, calculate the loop scale from the accumulated backedge mass.
1336 /// 3. Distribute mass in the function (\a computeMassInFunction()).
1338 /// Finally, distribute mass through the DAG resulting from packaging all
1339 /// loops in the function. This uses the same algorithm as distributing
1340 /// mass in a loop, except that there are no exit or backedge edges.
1342 /// 4. Unpackage loops (\a unwrapLoops()).
1344 /// Initialize each block's frequency to a floating point representation of
1347 /// Visit loops top-down, scaling the frequencies of its immediate members
1348 /// by the loop's pseudo-node's frequency.
1350 /// 5. Convert frequencies to a 64-bit range (\a finalizeMetrics()).
1352 /// Using the min and max frequencies as a guide, translate floating point
1353 /// frequencies to an appropriate range in uint64_t.
1355 /// It has some known flaws.
1357 /// - Loop scale is limited to 4096 per loop (2^12) to avoid exhausting
1358 /// BlockFrequency's 64-bit integer precision.
1360 /// - The model of irreducible control flow is a rough approximation.
1362 /// Modelling irreducible control flow exactly involves setting up and
1363 /// solving a group of infinite geometric series. Such precision is
1364 /// unlikely to be worthwhile, since most of our algorithms give up on
1365 /// irreducible control flow anyway.
1367 /// Nevertheless, we might find that we need to get closer. Here's a sort
1368 /// of TODO list for the model with diminishing returns, to be completed as
1371 /// - The headers for the \a LoopData representing an irreducible SCC
1372 /// include non-entry blocks. When these extra blocks exist, they
1373 /// indicate a self-contained irreducible sub-SCC. We could treat them
1374 /// as sub-loops, rather than arbitrarily shoving the problematic
1375 /// blocks into the headers of the main irreducible SCC.
1377 /// - Backedge frequencies are assumed to be evenly split between the
1378 /// headers of a given irreducible SCC. Instead, we could track the
1379 /// backedge mass separately for each header, and adjust their relative
1382 /// - Entry frequencies are assumed to be evenly split between the
1383 /// headers of a given irreducible SCC, which is the only option if we
1384 /// need to compute mass in the SCC before its parent loop. Instead,
1385 /// we could partially compute mass in the parent loop, and stop when
1386 /// we get to the SCC. Here, we have the correct ratio of entry
1387 /// masses, which we can use to adjust their relative frequencies.
1388 /// Compute mass in the SCC, and then continue propagation in the
1391 /// - We can propagate mass iteratively through the SCC, for some fixed
1392 /// number of iterations. Each iteration starts by assigning the entry
1393 /// blocks their backedge mass from the prior iteration. The final
1394 /// mass for each block (and each exit, and the total backedge mass
1395 /// used for computing loop scale) is the sum of all iterations.
1396 /// (Running this until fixed point would "solve" the geometric
1397 /// series by simulation.)
1398 template <class BT> class BlockFrequencyInfoImpl : BlockFrequencyInfoImplBase {
1399 typedef typename bfi_detail::TypeMap<BT>::BlockT BlockT;
1400 typedef typename bfi_detail::TypeMap<BT>::FunctionT FunctionT;
1401 typedef typename bfi_detail::TypeMap<BT>::BranchProbabilityInfoT
1402 BranchProbabilityInfoT;
1403 typedef typename bfi_detail::TypeMap<BT>::LoopT LoopT;
1404 typedef typename bfi_detail::TypeMap<BT>::LoopInfoT LoopInfoT;
1406 // This is part of a workaround for a GCC 4.7 crash on lambdas.
1407 friend struct bfi_detail::BlockEdgesAdder<BT>;
1409 typedef GraphTraits<const BlockT *> Successor;
1410 typedef GraphTraits<Inverse<const BlockT *>> Predecessor;
1412 const BranchProbabilityInfoT *BPI;
1413 const LoopInfoT *LI;
1416 // All blocks in reverse postorder.
1417 std::vector<const BlockT *> RPOT;
1418 DenseMap<const BlockT *, BlockNode> Nodes;
1420 typedef typename std::vector<const BlockT *>::const_iterator rpot_iterator;
1422 rpot_iterator rpot_begin() const { return RPOT.begin(); }
1423 rpot_iterator rpot_end() const { return RPOT.end(); }
1425 size_t getIndex(const rpot_iterator &I) const { return I - rpot_begin(); }
1427 BlockNode getNode(const rpot_iterator &I) const {
1428 return BlockNode(getIndex(I));
1430 BlockNode getNode(const BlockT *BB) const { return Nodes.lookup(BB); }
1432 const BlockT *getBlock(const BlockNode &Node) const {
1433 assert(Node.Index < RPOT.size());
1434 return RPOT[Node.Index];
1437 /// \brief Run (and save) a post-order traversal.
1439 /// Saves a reverse post-order traversal of all the nodes in \a F.
1440 void initializeRPOT();
1442 /// \brief Initialize loop data.
1444 /// Build up \a Loops using \a LoopInfo. \a LoopInfo gives us a mapping from
1445 /// each block to the deepest loop it's in, but we need the inverse. For each
1446 /// loop, we store in reverse post-order its "immediate" members, defined as
1447 /// the header, the headers of immediate sub-loops, and all other blocks in
1448 /// the loop that are not in sub-loops.
1449 void initializeLoops();
1451 /// \brief Propagate to a block's successors.
1453 /// In the context of distributing mass through \c OuterLoop, divide the mass
1454 /// currently assigned to \c Node between its successors.
1456 /// \return \c true unless there's an irreducible backedge.
1457 bool propagateMassToSuccessors(LoopData *OuterLoop, const BlockNode &Node);
1459 /// \brief Compute mass in a particular loop.
1461 /// Assign mass to \c Loop's header, and then for each block in \c Loop in
1462 /// reverse post-order, distribute mass to its successors. Only visits nodes
1463 /// that have not been packaged into sub-loops.
1465 /// \pre \a computeMassInLoop() has been called for each subloop of \c Loop.
1466 /// \return \c true unless there's an irreducible backedge.
1467 bool computeMassInLoop(LoopData &Loop);
1469 /// \brief Try to compute mass in the top-level function.
1471 /// Assign mass to the entry block, and then for each block in reverse
1472 /// post-order, distribute mass to its successors. Skips nodes that have
1473 /// been packaged into loops.
1475 /// \pre \a computeMassInLoops() has been called.
1476 /// \return \c true unless there's an irreducible backedge.
1477 bool tryToComputeMassInFunction();
1479 /// \brief Compute mass in (and package up) irreducible SCCs.
1481 /// Find the irreducible SCCs in \c OuterLoop, add them to \a Loops (in front
1482 /// of \c Insert), and call \a computeMassInLoop() on each of them.
1484 /// If \c OuterLoop is \c nullptr, it refers to the top-level function.
1486 /// \pre \a computeMassInLoop() has been called for each subloop of \c
1488 /// \pre \c Insert points at the the last loop successfully processed by \a
1489 /// computeMassInLoop().
1490 /// \pre \c OuterLoop has irreducible SCCs.
1491 void computeIrreducibleMass(LoopData *OuterLoop,
1492 std::list<LoopData>::iterator Insert);
1494 /// \brief Compute mass in all loops.
1496 /// For each loop bottom-up, call \a computeMassInLoop().
1498 /// \a computeMassInLoop() aborts (and returns \c false) on loops that
1499 /// contain a irreducible sub-SCCs. Use \a computeIrreducibleMass() and then
1500 /// re-enter \a computeMassInLoop().
1502 /// \post \a computeMassInLoop() has returned \c true for every loop.
1503 void computeMassInLoops();
1505 /// \brief Compute mass in the top-level function.
1507 /// Uses \a tryToComputeMassInFunction() and \a computeIrreducibleMass() to
1508 /// compute mass in the top-level function.
1510 /// \post \a tryToComputeMassInFunction() has returned \c true.
1511 void computeMassInFunction();
1513 std::string getBlockName(const BlockNode &Node) const override {
1514 return bfi_detail::getBlockName(getBlock(Node));
1518 const FunctionT *getFunction() const { return F; }
1520 void doFunction(const FunctionT *F, const BranchProbabilityInfoT *BPI,
1521 const LoopInfoT *LI);
1522 BlockFrequencyInfoImpl() : BPI(nullptr), LI(nullptr), F(nullptr) {}
1524 using BlockFrequencyInfoImplBase::getEntryFreq;
1525 BlockFrequency getBlockFreq(const BlockT *BB) const {
1526 return BlockFrequencyInfoImplBase::getBlockFreq(getNode(BB));
1528 Float getFloatingBlockFreq(const BlockT *BB) const {
1529 return BlockFrequencyInfoImplBase::getFloatingBlockFreq(getNode(BB));
1532 /// \brief Print the frequencies for the current function.
1534 /// Prints the frequencies for the blocks in the current function.
1536 /// Blocks are printed in the natural iteration order of the function, rather
1537 /// than reverse post-order. This provides two advantages: writing -analyze
1538 /// tests is easier (since blocks come out in source order), and even
1539 /// unreachable blocks are printed.
1541 /// \a BlockFrequencyInfoImplBase::print() only knows reverse post-order, so
1542 /// we need to override it here.
1543 raw_ostream &print(raw_ostream &OS) const override;
1544 using BlockFrequencyInfoImplBase::dump;
1546 using BlockFrequencyInfoImplBase::printBlockFreq;
1547 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockT *BB) const {
1548 return BlockFrequencyInfoImplBase::printBlockFreq(OS, getNode(BB));
1553 void BlockFrequencyInfoImpl<BT>::doFunction(const FunctionT *F,
1554 const BranchProbabilityInfoT *BPI,
1555 const LoopInfoT *LI) {
1556 // Save the parameters.
1561 // Clean up left-over data structures.
1562 BlockFrequencyInfoImplBase::clear();
1567 DEBUG(dbgs() << "\nblock-frequency: " << F->getName() << "\n================="
1568 << std::string(F->getName().size(), '=') << "\n");
1572 // Visit loops in post-order to find thelocal mass distribution, and then do
1573 // the full function.
1574 computeMassInLoops();
1575 computeMassInFunction();
1580 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeRPOT() {
1581 const BlockT *Entry = F->begin();
1582 RPOT.reserve(F->size());
1583 std::copy(po_begin(Entry), po_end(Entry), std::back_inserter(RPOT));
1584 std::reverse(RPOT.begin(), RPOT.end());
1586 assert(RPOT.size() - 1 <= BlockNode::getMaxIndex() &&
1587 "More nodes in function than Block Frequency Info supports");
1589 DEBUG(dbgs() << "reverse-post-order-traversal\n");
1590 for (rpot_iterator I = rpot_begin(), E = rpot_end(); I != E; ++I) {
1591 BlockNode Node = getNode(I);
1592 DEBUG(dbgs() << " - " << getIndex(I) << ": " << getBlockName(Node) << "\n");
1596 Working.reserve(RPOT.size());
1597 for (size_t Index = 0; Index < RPOT.size(); ++Index)
1598 Working.emplace_back(Index);
1599 Freqs.resize(RPOT.size());
1602 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeLoops() {
1603 DEBUG(dbgs() << "loop-detection\n");
1607 // Visit loops top down and assign them an index.
1608 std::deque<std::pair<const LoopT *, LoopData *>> Q;
1609 for (const LoopT *L : *LI)
1610 Q.emplace_back(L, nullptr);
1611 while (!Q.empty()) {
1612 const LoopT *Loop = Q.front().first;
1613 LoopData *Parent = Q.front().second;
1616 BlockNode Header = getNode(Loop->getHeader());
1617 assert(Header.isValid());
1619 Loops.emplace_back(Parent, Header);
1620 Working[Header.Index].Loop = &Loops.back();
1621 DEBUG(dbgs() << " - loop = " << getBlockName(Header) << "\n");
1623 for (const LoopT *L : *Loop)
1624 Q.emplace_back(L, &Loops.back());
1627 // Visit nodes in reverse post-order and add them to their deepest containing
1629 for (size_t Index = 0; Index < RPOT.size(); ++Index) {
1630 // Loop headers have already been mostly mapped.
1631 if (Working[Index].isLoopHeader()) {
1632 LoopData *ContainingLoop = Working[Index].getContainingLoop();
1634 ContainingLoop->Nodes.push_back(Index);
1638 const LoopT *Loop = LI->getLoopFor(RPOT[Index]);
1642 // Add this node to its containing loop's member list.
1643 BlockNode Header = getNode(Loop->getHeader());
1644 assert(Header.isValid());
1645 const auto &HeaderData = Working[Header.Index];
1646 assert(HeaderData.isLoopHeader());
1648 Working[Index].Loop = HeaderData.Loop;
1649 HeaderData.Loop->Nodes.push_back(Index);
1650 DEBUG(dbgs() << " - loop = " << getBlockName(Header)
1651 << ": member = " << getBlockName(Index) << "\n");
1655 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInLoops() {
1656 // Visit loops with the deepest first, and the top-level loops last.
1657 for (auto L = Loops.rbegin(), E = Loops.rend(); L != E; ++L) {
1658 if (computeMassInLoop(*L))
1660 auto Next = std::next(L);
1661 computeIrreducibleMass(&*L, L.base());
1662 L = std::prev(Next);
1663 if (computeMassInLoop(*L))
1665 llvm_unreachable("unhandled irreducible control flow");
1670 bool BlockFrequencyInfoImpl<BT>::computeMassInLoop(LoopData &Loop) {
1671 // Compute mass in loop.
1672 DEBUG(dbgs() << "compute-mass-in-loop: " << getLoopName(Loop) << "\n");
1674 if (Loop.isIrreducible()) {
1675 BlockMass Remaining = BlockMass::getFull();
1676 for (uint32_t H = 0; H < Loop.NumHeaders; ++H) {
1677 auto &Mass = Working[Loop.Nodes[H].Index].getMass();
1678 Mass = Remaining * BranchProbability(1, Loop.NumHeaders - H);
1681 for (const BlockNode &M : Loop.Nodes)
1682 if (!propagateMassToSuccessors(&Loop, M))
1683 llvm_unreachable("unhandled irreducible control flow");
1685 Working[Loop.getHeader().Index].getMass() = BlockMass::getFull();
1686 if (!propagateMassToSuccessors(&Loop, Loop.getHeader()))
1687 llvm_unreachable("irreducible control flow to loop header!?");
1688 for (const BlockNode &M : Loop.members())
1689 if (!propagateMassToSuccessors(&Loop, M))
1690 // Irreducible backedge.
1694 computeLoopScale(Loop);
1700 bool BlockFrequencyInfoImpl<BT>::tryToComputeMassInFunction() {
1701 // Compute mass in function.
1702 DEBUG(dbgs() << "compute-mass-in-function\n");
1703 assert(!Working.empty() && "no blocks in function");
1704 assert(!Working[0].isLoopHeader() && "entry block is a loop header");
1706 Working[0].getMass() = BlockMass::getFull();
1707 for (rpot_iterator I = rpot_begin(), IE = rpot_end(); I != IE; ++I) {
1708 // Check for nodes that have been packaged.
1709 BlockNode Node = getNode(I);
1710 if (Working[Node.Index].isPackaged())
1713 if (!propagateMassToSuccessors(nullptr, Node))
1719 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInFunction() {
1720 if (tryToComputeMassInFunction())
1722 computeIrreducibleMass(nullptr, Loops.begin());
1723 if (tryToComputeMassInFunction())
1725 llvm_unreachable("unhandled irreducible control flow");
1728 /// \note This should be a lambda, but that crashes GCC 4.7.
1729 namespace bfi_detail {
1730 template <class BT> struct BlockEdgesAdder {
1732 typedef BlockFrequencyInfoImplBase::LoopData LoopData;
1733 typedef GraphTraits<const BlockT *> Successor;
1735 const BlockFrequencyInfoImpl<BT> &BFI;
1736 explicit BlockEdgesAdder(const BlockFrequencyInfoImpl<BT> &BFI)
1738 void operator()(IrreducibleGraph &G, IrreducibleGraph::IrrNode &Irr,
1739 const LoopData *OuterLoop) {
1740 const BlockT *BB = BFI.RPOT[Irr.Node.Index];
1741 for (auto I = Successor::child_begin(BB), E = Successor::child_end(BB);
1743 G.addEdge(Irr, BFI.getNode(*I), OuterLoop);
1748 void BlockFrequencyInfoImpl<BT>::computeIrreducibleMass(
1749 LoopData *OuterLoop, std::list<LoopData>::iterator Insert) {
1750 DEBUG(dbgs() << "analyze-irreducible-in-";
1751 if (OuterLoop) dbgs() << "loop: " << getLoopName(*OuterLoop) << "\n";
1752 else dbgs() << "function\n");
1754 using namespace bfi_detail;
1755 // Ideally, addBlockEdges() would be declared here as a lambda, but that
1757 BlockEdgesAdder<BT> addBlockEdges(*this);
1758 IrreducibleGraph G(*this, OuterLoop, addBlockEdges);
1760 for (auto &L : analyzeIrreducible(G, OuterLoop, Insert))
1761 computeMassInLoop(L);
1765 updateLoopWithIrreducible(*OuterLoop);
1770 BlockFrequencyInfoImpl<BT>::propagateMassToSuccessors(LoopData *OuterLoop,
1771 const BlockNode &Node) {
1772 DEBUG(dbgs() << " - node: " << getBlockName(Node) << "\n");
1773 // Calculate probability for successors.
1775 if (auto *Loop = Working[Node.Index].getPackagedLoop()) {
1776 assert(Loop != OuterLoop && "Cannot propagate mass in a packaged loop");
1777 if (!addLoopSuccessorsToDist(OuterLoop, *Loop, Dist))
1778 // Irreducible backedge.
1781 const BlockT *BB = getBlock(Node);
1782 for (auto SI = Successor::child_begin(BB), SE = Successor::child_end(BB);
1784 // Do not dereference SI, or getEdgeWeight() is linear in the number of
1786 if (!addToDist(Dist, OuterLoop, Node, getNode(*SI),
1787 BPI->getEdgeWeight(BB, SI)))
1788 // Irreducible backedge.
1792 // Distribute mass to successors, saving exit and backedge data in the
1794 distributeMass(Node, OuterLoop, Dist);
1799 raw_ostream &BlockFrequencyInfoImpl<BT>::print(raw_ostream &OS) const {
1802 OS << "block-frequency-info: " << F->getName() << "\n";
1803 for (const BlockT &BB : *F)
1804 OS << " - " << bfi_detail::getBlockName(&BB)
1805 << ": float = " << getFloatingBlockFreq(&BB)
1806 << ", int = " << getBlockFreq(&BB).getFrequency() << "\n";
1808 // Add an extra newline for readability.