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/SCCIterator.h"
21 #include "llvm/ADT/iterator_range.h"
22 #include "llvm/IR/BasicBlock.h"
23 #include "llvm/Support/BlockFrequency.h"
24 #include "llvm/Support/BranchProbability.h"
25 #include "llvm/Support/Debug.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 std::pair<uint64_t, int16_t> divide64(uint64_t L, uint64_t R);
81 static std::pair<uint64_t, int16_t> multiply64(uint64_t L, uint64_t R);
83 static int compare(uint64_t L, uint64_t R, int Shift) {
87 uint64_t L_adjusted = L >> Shift;
93 return L > L_adjusted << Shift ? 1 : 0;
97 /// \brief Simple representation of an unsigned floating point.
99 /// UnsignedFloat is a unsigned floating point number. It uses simple
100 /// saturation arithmetic, and every operation is well-defined for every value.
102 /// The number is split into a signed exponent and unsigned digits. The number
103 /// represented is \c getDigits()*2^getExponent(). In this way, the digits are
104 /// much like the mantissa in the x87 long double, but there is no canonical
105 /// form, so the same number can be represented by many bit representations
106 /// (it's always in "denormal" mode).
108 /// UnsignedFloat is templated on the underlying integer type for digits, which
109 /// is expected to be one of uint64_t, uint32_t, uint16_t or uint8_t.
111 /// Unlike builtin floating point types, UnsignedFloat is portable.
113 /// Unlike APFloat, UnsignedFloat does not model architecture floating point
114 /// behaviour (this should make it a little faster), and implements most
115 /// operators (this makes it usable).
117 /// UnsignedFloat is totally ordered. However, there is no canonical form, so
118 /// there are multiple representations of most scalars. E.g.:
120 /// UnsignedFloat(8u, 0) == UnsignedFloat(4u, 1)
121 /// UnsignedFloat(4u, 1) == UnsignedFloat(2u, 2)
122 /// UnsignedFloat(2u, 2) == UnsignedFloat(1u, 3)
124 /// UnsignedFloat implements most arithmetic operations. Precision is kept
125 /// where possible. Uses simple saturation arithmetic, so that operations
126 /// saturate to 0.0 or getLargest() rather than under or overflowing. It has
127 /// some extra arithmetic for unit inversion. 0.0/0.0 is defined to be 0.0.
128 /// Any other division by 0.0 is defined to be getLargest().
130 /// As a convenience for modifying the exponent, left and right shifting are
131 /// both implemented, and both interpret negative shifts as positive shifts in
132 /// the opposite direction.
134 /// Exponents are limited to the range accepted by x87 long double. This makes
135 /// it trivial to add functionality to convert to APFloat (this is already
136 /// relied on for the implementation of printing).
138 /// The current plan is to gut this and make the necessary parts of it (even
139 /// more) private to BlockFrequencyInfo.
140 template <class DigitsT> class UnsignedFloat : UnsignedFloatBase {
142 static_assert(!std::numeric_limits<DigitsT>::is_signed,
143 "only unsigned floats supported");
145 typedef DigitsT DigitsType;
148 typedef std::numeric_limits<DigitsType> DigitsLimits;
150 static const int Width = sizeof(DigitsType) * 8;
151 static_assert(Width <= 64, "invalid integer width for digits");
158 UnsignedFloat() : Digits(0), Exponent(0) {}
160 UnsignedFloat(DigitsType Digits, int16_t Exponent)
161 : Digits(Digits), Exponent(Exponent) {}
164 UnsignedFloat(const std::pair<uint64_t, int16_t> &X)
165 : Digits(X.first), Exponent(X.second) {}
168 static UnsignedFloat getZero() { return UnsignedFloat(0, 0); }
169 static UnsignedFloat getOne() { return UnsignedFloat(1, 0); }
170 static UnsignedFloat getLargest() {
171 return UnsignedFloat(DigitsLimits::max(), MaxExponent);
173 static UnsignedFloat getFloat(uint64_t N) { return adjustToWidth(N, 0); }
174 static UnsignedFloat getInverseFloat(uint64_t N) {
175 return getFloat(N).invert();
177 static UnsignedFloat getFraction(DigitsType N, DigitsType D) {
178 return getQuotient(N, D);
181 int16_t getExponent() const { return Exponent; }
182 DigitsType getDigits() const { return Digits; }
184 /// \brief Convert to the given integer type.
186 /// Convert to \c IntT using simple saturating arithmetic, truncating if
188 template <class IntT> IntT toInt() const;
190 bool isZero() const { return !Digits; }
191 bool isLargest() const { return *this == getLargest(); }
193 if (Exponent > 0 || Exponent <= -Width)
195 return Digits == DigitsType(1) << -Exponent;
198 /// \brief The log base 2, rounded.
200 /// Get the lg of the scalar. lg 0 is defined to be INT32_MIN.
201 int32_t lg() const { return extractLg(lgImpl()); }
203 /// \brief The log base 2, rounded towards INT32_MIN.
205 /// Get the lg floor. lg 0 is defined to be INT32_MIN.
206 int32_t lgFloor() const { return extractLgFloor(lgImpl()); }
208 /// \brief The log base 2, rounded towards INT32_MAX.
210 /// Get the lg ceiling. lg 0 is defined to be INT32_MIN.
211 int32_t lgCeiling() const { return extractLgCeiling(lgImpl()); }
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; }
216 bool operator>(const UnsignedFloat &X) const { return compare(X) > 0; }
217 bool operator<=(const UnsignedFloat &X) const { return compare(X) <= 0; }
218 bool operator>=(const UnsignedFloat &X) const { return compare(X) >= 0; }
220 bool operator!() const { return isZero(); }
222 /// \brief Convert to a decimal representation in a string.
224 /// Convert to a string. Uses scientific notation for very large/small
225 /// numbers. Scientific notation is used roughly for numbers outside of the
226 /// range 2^-64 through 2^64.
228 /// \c Precision indicates the number of decimal digits of precision to use;
229 /// 0 requests the maximum available.
231 /// As a special case to make debugging easier, if the number is small enough
232 /// to convert without scientific notation and has more than \c Precision
233 /// digits before the decimal place, it's printed accurately to the first
234 /// digit past zero. E.g., assuming 10 digits of precision:
236 /// 98765432198.7654... => 98765432198.8
237 /// 8765432198.7654... => 8765432198.8
238 /// 765432198.7654... => 765432198.8
239 /// 65432198.7654... => 65432198.77
240 /// 5432198.7654... => 5432198.765
241 std::string toString(unsigned Precision = DefaultPrecision) {
242 return UnsignedFloatBase::toString(Digits, Exponent, Width, Precision);
245 /// \brief Print a decimal representation.
247 /// Print a string. See toString for documentation.
248 raw_ostream &print(raw_ostream &OS,
249 unsigned Precision = DefaultPrecision) const {
250 return UnsignedFloatBase::print(OS, Digits, Exponent, Width, Precision);
252 void dump() const { return UnsignedFloatBase::dump(Digits, Exponent, Width); }
254 UnsignedFloat &operator+=(const UnsignedFloat &X);
255 UnsignedFloat &operator-=(const UnsignedFloat &X);
256 UnsignedFloat &operator*=(const UnsignedFloat &X);
257 UnsignedFloat &operator/=(const UnsignedFloat &X);
258 UnsignedFloat &operator<<=(int16_t Shift) { shiftLeft(Shift); return *this; }
259 UnsignedFloat &operator>>=(int16_t Shift) { shiftRight(Shift); return *this; }
262 void shiftLeft(int32_t Shift);
263 void shiftRight(int32_t Shift);
265 /// \brief Adjust two floats to have matching exponents.
267 /// Adjust \c this and \c X to have matching exponents. Returns the new \c X
268 /// by value. Does nothing if \a isZero() for either.
270 /// The value that compares smaller will lose precision, and possibly become
272 UnsignedFloat matchExponents(UnsignedFloat X);
274 /// \brief Increase exponent to match another float.
276 /// Increases \c this to have an exponent matching \c X. May decrease the
277 /// exponent of \c X in the process, and \c this may possibly become \a
279 void increaseExponentToMatch(UnsignedFloat &X, int32_t ExponentDiff);
282 /// \brief Scale a large number accurately.
284 /// Scale N (multiply it by this). Uses full precision multiplication, even
285 /// if Width is smaller than 64, so information is not lost.
286 uint64_t scale(uint64_t N) const;
287 uint64_t scaleByInverse(uint64_t N) const {
288 // TODO: implement directly, rather than relying on inverse. Inverse is
290 return inverse().scale(N);
292 int64_t scale(int64_t N) const {
293 std::pair<uint64_t, bool> Unsigned = splitSigned(N);
294 return joinSigned(scale(Unsigned.first), Unsigned.second);
296 int64_t scaleByInverse(int64_t N) const {
297 std::pair<uint64_t, bool> Unsigned = splitSigned(N);
298 return joinSigned(scaleByInverse(Unsigned.first), Unsigned.second);
301 int compare(const UnsignedFloat &X) const;
302 int compareTo(uint64_t N) const {
303 UnsignedFloat Float = getFloat(N);
304 int Compare = compare(Float);
305 if (Width == 64 || Compare != 0)
308 // Check for precision loss. We know *this == RoundTrip.
309 uint64_t RoundTrip = Float.template toInt<uint64_t>();
310 return N == RoundTrip ? 0 : RoundTrip < N ? -1 : 1;
312 int compareTo(int64_t N) const { return N < 0 ? 1 : compareTo(uint64_t(N)); }
314 UnsignedFloat &invert() { return *this = UnsignedFloat::getFloat(1) / *this; }
315 UnsignedFloat inverse() const { return UnsignedFloat(*this).invert(); }
318 static UnsignedFloat getProduct(DigitsType L, DigitsType R);
319 static UnsignedFloat getQuotient(DigitsType Dividend, DigitsType Divisor);
321 std::pair<int32_t, int> lgImpl() const;
322 static int countLeadingZerosWidth(DigitsType Digits) {
324 return countLeadingZeros64(Digits);
326 return countLeadingZeros32(Digits);
327 return countLeadingZeros32(Digits) + Width - 32;
330 static UnsignedFloat adjustToWidth(uint64_t N, int32_t S) {
331 assert(S >= MinExponent);
332 assert(S <= MaxExponent);
333 if (Width == 64 || N <= DigitsLimits::max())
334 return UnsignedFloat(N, S);
337 int Shift = 64 - Width - countLeadingZeros64(N);
338 DigitsType Shifted = N >> Shift;
341 assert(S + Shift <= MaxExponent);
342 return getRounded(UnsignedFloat(Shifted, S + Shift),
343 N & UINT64_C(1) << (Shift - 1));
346 static UnsignedFloat getRounded(UnsignedFloat P, bool Round) {
349 if (P.Digits == DigitsLimits::max())
350 // Careful of overflow in the exponent.
351 return UnsignedFloat(1, P.Exponent) <<= Width;
352 return UnsignedFloat(P.Digits + 1, P.Exponent);
356 #define UNSIGNED_FLOAT_BOP(op, base) \
357 template <class DigitsT> \
358 UnsignedFloat<DigitsT> operator op(const UnsignedFloat<DigitsT> &L, \
359 const UnsignedFloat<DigitsT> &R) { \
360 return UnsignedFloat<DigitsT>(L) base R; \
362 UNSIGNED_FLOAT_BOP(+, += )
363 UNSIGNED_FLOAT_BOP(-, -= )
364 UNSIGNED_FLOAT_BOP(*, *= )
365 UNSIGNED_FLOAT_BOP(/, /= )
366 UNSIGNED_FLOAT_BOP(<<, <<= )
367 UNSIGNED_FLOAT_BOP(>>, >>= )
368 #undef UNSIGNED_FLOAT_BOP
370 template <class DigitsT>
371 raw_ostream &operator<<(raw_ostream &OS, const UnsignedFloat<DigitsT> &X) {
372 return X.print(OS, 10);
375 #define UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, T1, T2) \
376 template <class DigitsT> \
377 bool operator op(const UnsignedFloat<DigitsT> &L, T1 R) { \
378 return L.compareTo(T2(R)) op 0; \
380 template <class DigitsT> \
381 bool operator op(T1 L, const UnsignedFloat<DigitsT> &R) { \
382 return 0 op R.compareTo(T2(L)); \
384 #define UNSIGNED_FLOAT_COMPARE_TO(op) \
385 UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, uint64_t, uint64_t) \
386 UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, uint32_t, uint64_t) \
387 UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, int64_t, int64_t) \
388 UNSIGNED_FLOAT_COMPARE_TO_TYPE(op, int32_t, int64_t)
389 UNSIGNED_FLOAT_COMPARE_TO(< )
390 UNSIGNED_FLOAT_COMPARE_TO(> )
391 UNSIGNED_FLOAT_COMPARE_TO(== )
392 UNSIGNED_FLOAT_COMPARE_TO(!= )
393 UNSIGNED_FLOAT_COMPARE_TO(<= )
394 UNSIGNED_FLOAT_COMPARE_TO(>= )
395 #undef UNSIGNED_FLOAT_COMPARE_TO
396 #undef UNSIGNED_FLOAT_COMPARE_TO_TYPE
398 template <class DigitsT>
399 uint64_t UnsignedFloat<DigitsT>::scale(uint64_t N) const {
400 if (Width == 64 || N <= DigitsLimits::max())
401 return (getFloat(N) * *this).template toInt<uint64_t>();
403 // Defer to the 64-bit version.
404 return UnsignedFloat<uint64_t>(Digits, Exponent).scale(N);
407 template <class DigitsT>
408 UnsignedFloat<DigitsT> UnsignedFloat<DigitsT>::getProduct(DigitsType L,
414 // Check for numbers that we can compute with 64-bit math.
415 if (Width <= 32 || (L <= UINT32_MAX && R <= UINT32_MAX))
416 return adjustToWidth(uint64_t(L) * uint64_t(R), 0);
418 // Do the full thing.
419 return UnsignedFloat(multiply64(L, R));
421 template <class DigitsT>
422 UnsignedFloat<DigitsT> UnsignedFloat<DigitsT>::getQuotient(DigitsType Dividend,
423 DigitsType Divisor) {
431 return UnsignedFloat(divide64(Dividend, Divisor));
433 // We can compute this with 64-bit math.
434 int Shift = countLeadingZeros64(Dividend);
435 uint64_t Shifted = uint64_t(Dividend) << Shift;
436 uint64_t Quotient = Shifted / Divisor;
438 // If Quotient needs to be shifted, then adjustToWidth will round.
439 if (Quotient > DigitsLimits::max())
440 return adjustToWidth(Quotient, -Shift);
442 // Round based on the value of the next bit.
443 return getRounded(UnsignedFloat(Quotient, -Shift),
444 Shifted % Divisor >= getHalf(Divisor));
447 template <class DigitsT>
448 template <class IntT>
449 IntT UnsignedFloat<DigitsT>::toInt() const {
450 typedef std::numeric_limits<IntT> Limits;
453 if (*this >= Limits::max())
454 return Limits::max();
458 assert(size_t(Exponent) < sizeof(IntT) * 8);
459 return N << Exponent;
462 assert(size_t(-Exponent) < sizeof(IntT) * 8);
463 return N >> -Exponent;
468 template <class DigitsT>
469 std::pair<int32_t, int> UnsignedFloat<DigitsT>::lgImpl() const {
471 return std::make_pair(INT32_MIN, 0);
473 // Get the floor of the lg of Digits.
474 int32_t LocalFloor = Width - countLeadingZerosWidth(Digits) - 1;
476 // Get the floor of the lg of this.
477 int32_t Floor = Exponent + LocalFloor;
478 if (Digits == UINT64_C(1) << LocalFloor)
479 return std::make_pair(Floor, 0);
481 // Round based on the next digit.
482 assert(LocalFloor >= 1);
483 bool Round = Digits & UINT64_C(1) << (LocalFloor - 1);
484 return std::make_pair(Floor + Round, Round ? 1 : -1);
487 template <class DigitsT>
488 UnsignedFloat<DigitsT> UnsignedFloat<DigitsT>::matchExponents(UnsignedFloat X) {
489 if (isZero() || X.isZero() || Exponent == X.Exponent)
492 int32_t Diff = int32_t(X.Exponent) - int32_t(Exponent);
494 increaseExponentToMatch(X, Diff);
496 X.increaseExponentToMatch(*this, -Diff);
499 template <class DigitsT>
500 void UnsignedFloat<DigitsT>::increaseExponentToMatch(UnsignedFloat &X,
501 int32_t ExponentDiff) {
502 assert(ExponentDiff > 0);
503 if (ExponentDiff >= 2 * Width) {
508 // Use up any leading zeros on X, and then shift this.
509 int32_t ShiftX = std::min(countLeadingZerosWidth(X.Digits), ExponentDiff);
510 assert(ShiftX < Width);
512 int32_t ShiftThis = ExponentDiff - ShiftX;
513 if (ShiftThis >= Width) {
519 X.Exponent -= ShiftX;
520 Digits >>= ShiftThis;
521 Exponent += ShiftThis;
525 template <class DigitsT>
526 UnsignedFloat<DigitsT> &UnsignedFloat<DigitsT>::
527 operator+=(const UnsignedFloat &X) {
528 if (isLargest() || X.isZero())
530 if (isZero() || X.isLargest())
533 // Normalize exponents.
534 UnsignedFloat Scaled = matchExponents(X);
536 // Check for zero again.
538 return *this = Scaled;
543 DigitsType Sum = Digits + Scaled.Digits;
544 bool DidOverflow = Sum < Digits;
549 if (Exponent == MaxExponent)
550 return *this = getLargest();
553 Digits = UINT64_C(1) << (Width - 1) | Digits >> 1;
557 template <class DigitsT>
558 UnsignedFloat<DigitsT> &UnsignedFloat<DigitsT>::
559 operator-=(const UnsignedFloat &X) {
563 return *this = getZero();
565 // Normalize exponents.
566 UnsignedFloat Scaled = matchExponents(X);
567 assert(Digits >= Scaled.Digits);
569 // Compute difference.
570 if (!Scaled.isZero()) {
571 Digits -= Scaled.Digits;
575 // Check if X just barely lost its last bit. E.g., for 32-bit:
577 // 1*2^32 - 1*2^0 == 0xffffffff != 1*2^32
578 if (*this == UnsignedFloat(1, X.lgFloor() + Width)) {
579 Digits = DigitsType(0) - 1;
584 template <class DigitsT>
585 UnsignedFloat<DigitsT> &UnsignedFloat<DigitsT>::
586 operator*=(const UnsignedFloat &X) {
592 // Save the exponents.
593 int32_t Exponents = int32_t(Exponent) + int32_t(X.Exponent);
595 // Get the raw product.
596 *this = getProduct(Digits, X.Digits);
598 // Combine with exponents.
599 return *this <<= Exponents;
601 template <class DigitsT>
602 UnsignedFloat<DigitsT> &UnsignedFloat<DigitsT>::
603 operator/=(const UnsignedFloat &X) {
607 return *this = getLargest();
609 // Save the exponents.
610 int32_t Exponents = int32_t(Exponent) - int32_t(X.Exponent);
612 // Get the raw quotient.
613 *this = getQuotient(Digits, X.Digits);
615 // Combine with exponents.
616 return *this <<= Exponents;
618 template <class DigitsT>
619 void UnsignedFloat<DigitsT>::shiftLeft(int32_t Shift) {
620 if (!Shift || isZero())
622 assert(Shift != INT32_MIN);
628 // Shift as much as we can in the exponent.
629 int32_t ExponentShift = std::min(Shift, MaxExponent - Exponent);
630 Exponent += ExponentShift;
631 if (ExponentShift == Shift)
634 // Check this late, since it's rare.
638 // Shift the digits themselves.
639 Shift -= ExponentShift;
640 if (Shift > countLeadingZerosWidth(Digits)) {
642 *this = getLargest();
650 template <class DigitsT>
651 void UnsignedFloat<DigitsT>::shiftRight(int32_t Shift) {
652 if (!Shift || isZero())
654 assert(Shift != INT32_MIN);
660 // Shift as much as we can in the exponent.
661 int32_t ExponentShift = std::min(Shift, Exponent - MinExponent);
662 Exponent -= ExponentShift;
663 if (ExponentShift == Shift)
666 // Shift the digits themselves.
667 Shift -= ExponentShift;
668 if (Shift >= Width) {
678 template <class DigitsT>
679 int UnsignedFloat<DigitsT>::compare(const UnsignedFloat &X) const {
682 return X.isZero() ? 0 : -1;
686 // Check for the scale. Use lgFloor to be sure that the exponent difference
687 // is always lower than 64.
688 int32_t lgL = lgFloor(), lgR = X.lgFloor();
690 return lgL < lgR ? -1 : 1;
693 if (Exponent < X.Exponent)
694 return UnsignedFloatBase::compare(Digits, X.Digits, X.Exponent - Exponent);
696 return -UnsignedFloatBase::compare(X.Digits, Digits, Exponent - X.Exponent);
699 template <class T> struct isPodLike<UnsignedFloat<T>> {
700 static const bool value = true;
704 //===----------------------------------------------------------------------===//
706 // BlockMass definition.
708 // TODO: Make this private to BlockFrequencyInfoImpl or delete.
710 //===----------------------------------------------------------------------===//
713 /// \brief Mass of a block.
715 /// This class implements a sort of fixed-point fraction always between 0.0 and
716 /// 1.0. getMass() == UINT64_MAX indicates a value of 1.0.
718 /// Masses can be added and subtracted. Simple saturation arithmetic is used,
719 /// so arithmetic operations never overflow or underflow.
721 /// Masses can be multiplied. Multiplication treats full mass as 1.0 and uses
722 /// an inexpensive floating-point algorithm that's off-by-one (almost, but not
723 /// quite, maximum precision).
725 /// Masses can be scaled by \a BranchProbability at maximum precision.
730 BlockMass() : Mass(0) {}
731 explicit BlockMass(uint64_t Mass) : Mass(Mass) {}
733 static BlockMass getEmpty() { return BlockMass(); }
734 static BlockMass getFull() { return BlockMass(UINT64_MAX); }
736 uint64_t getMass() const { return Mass; }
738 bool isFull() const { return Mass == UINT64_MAX; }
739 bool isEmpty() const { return !Mass; }
741 bool operator!() const { return isEmpty(); }
743 /// \brief Add another mass.
745 /// Adds another mass, saturating at \a isFull() rather than overflowing.
746 BlockMass &operator+=(const BlockMass &X) {
747 uint64_t Sum = Mass + X.Mass;
748 Mass = Sum < Mass ? UINT64_MAX : Sum;
752 /// \brief Subtract another mass.
754 /// Subtracts another mass, saturating at \a isEmpty() rather than
756 BlockMass &operator-=(const BlockMass &X) {
757 uint64_t Diff = Mass - X.Mass;
758 Mass = Diff > Mass ? 0 : Diff;
762 BlockMass &operator*=(const BranchProbability &P) {
763 Mass = P.scale(Mass);
767 bool operator==(const BlockMass &X) const { return Mass == X.Mass; }
768 bool operator!=(const BlockMass &X) const { return Mass != X.Mass; }
769 bool operator<=(const BlockMass &X) const { return Mass <= X.Mass; }
770 bool operator>=(const BlockMass &X) const { return Mass >= X.Mass; }
771 bool operator<(const BlockMass &X) const { return Mass < X.Mass; }
772 bool operator>(const BlockMass &X) const { return Mass > X.Mass; }
774 /// \brief Convert to floating point.
776 /// Convert to a float. \a isFull() gives 1.0, while \a isEmpty() gives
777 /// slightly above 0.0.
778 UnsignedFloat<uint64_t> toFloat() const;
781 raw_ostream &print(raw_ostream &OS) const;
784 inline BlockMass operator+(const BlockMass &L, const BlockMass &R) {
785 return BlockMass(L) += R;
787 inline BlockMass operator-(const BlockMass &L, const BlockMass &R) {
788 return BlockMass(L) -= R;
790 inline BlockMass operator*(const BlockMass &L, const BranchProbability &R) {
791 return BlockMass(L) *= R;
793 inline BlockMass operator*(const BranchProbability &L, const BlockMass &R) {
794 return BlockMass(R) *= L;
797 inline raw_ostream &operator<<(raw_ostream &OS, const BlockMass &X) {
801 template <> struct isPodLike<BlockMass> {
802 static const bool value = true;
806 //===----------------------------------------------------------------------===//
808 // BlockFrequencyInfoImpl definition.
810 //===----------------------------------------------------------------------===//
814 class BranchProbabilityInfo;
818 class MachineBasicBlock;
819 class MachineBranchProbabilityInfo;
820 class MachineFunction;
822 class MachineLoopInfo;
824 namespace bfi_detail {
825 struct IrreducibleGraph;
827 // This is part of a workaround for a GCC 4.7 crash on lambdas.
828 template <class BT> struct BlockEdgesAdder;
831 /// \brief Base class for BlockFrequencyInfoImpl
833 /// BlockFrequencyInfoImplBase has supporting data structures and some
834 /// algorithms for BlockFrequencyInfoImplBase. Only algorithms that depend on
835 /// the block type (or that call such algorithms) are skipped here.
837 /// Nevertheless, the majority of the overall algorithm documention lives with
838 /// BlockFrequencyInfoImpl. See there for details.
839 class BlockFrequencyInfoImplBase {
841 typedef UnsignedFloat<uint64_t> Float;
843 /// \brief Representative of a block.
845 /// This is a simple wrapper around an index into the reverse-post-order
846 /// traversal of the blocks.
848 /// Unlike a block pointer, its order has meaning (location in the
849 /// topological sort) and it's class is the same regardless of block type.
851 typedef uint32_t IndexType;
854 bool operator==(const BlockNode &X) const { return Index == X.Index; }
855 bool operator!=(const BlockNode &X) const { return Index != X.Index; }
856 bool operator<=(const BlockNode &X) const { return Index <= X.Index; }
857 bool operator>=(const BlockNode &X) const { return Index >= X.Index; }
858 bool operator<(const BlockNode &X) const { return Index < X.Index; }
859 bool operator>(const BlockNode &X) const { return Index > X.Index; }
861 BlockNode() : Index(UINT32_MAX) {}
862 BlockNode(IndexType Index) : Index(Index) {}
864 bool isValid() const { return Index <= getMaxIndex(); }
865 static size_t getMaxIndex() { return UINT32_MAX - 1; }
868 /// \brief Stats about a block itself.
869 struct FrequencyData {
874 /// \brief Data about a loop.
876 /// Contains the data necessary to represent represent a loop as a
877 /// pseudo-node once it's packaged.
879 typedef SmallVector<std::pair<BlockNode, BlockMass>, 4> ExitMap;
880 typedef SmallVector<BlockNode, 4> NodeList;
881 LoopData *Parent; ///< The parent loop.
882 bool IsPackaged; ///< Whether this has been packaged.
883 uint32_t NumHeaders; ///< Number of headers.
884 ExitMap Exits; ///< Successor edges (and weights).
885 NodeList Nodes; ///< Header and the members of the loop.
886 BlockMass BackedgeMass; ///< Mass returned to loop header.
890 LoopData(LoopData *Parent, const BlockNode &Header)
891 : Parent(Parent), IsPackaged(false), NumHeaders(1), Nodes(1, Header) {}
892 template <class It1, class It2>
893 LoopData(LoopData *Parent, It1 FirstHeader, It1 LastHeader, It2 FirstOther,
895 : Parent(Parent), IsPackaged(false), Nodes(FirstHeader, LastHeader) {
896 NumHeaders = Nodes.size();
897 Nodes.insert(Nodes.end(), FirstOther, LastOther);
899 bool isHeader(const BlockNode &Node) const {
901 return std::binary_search(Nodes.begin(), Nodes.begin() + NumHeaders,
903 return Node == Nodes[0];
905 BlockNode getHeader() const { return Nodes[0]; }
906 bool isIrreducible() const { return NumHeaders > 1; }
908 NodeList::const_iterator members_begin() const {
909 return Nodes.begin() + NumHeaders;
911 NodeList::const_iterator members_end() const { return Nodes.end(); }
912 iterator_range<NodeList::const_iterator> members() const {
913 return make_range(members_begin(), members_end());
917 /// \brief Index of loop information.
919 BlockNode Node; ///< This node.
920 LoopData *Loop; ///< The loop this block is inside.
921 BlockMass Mass; ///< Mass distribution from the entry block.
923 WorkingData(const BlockNode &Node) : Node(Node), Loop(nullptr) {}
925 bool isLoopHeader() const { return Loop && Loop->isHeader(Node); }
926 bool isDoubleLoopHeader() const {
927 return isLoopHeader() && Loop->Parent && Loop->Parent->isIrreducible() &&
928 Loop->Parent->isHeader(Node);
931 LoopData *getContainingLoop() const {
934 if (!isDoubleLoopHeader())
936 return Loop->Parent->Parent;
939 /// \brief Resolve a node to its representative.
941 /// Get the node currently representing Node, which could be a containing
944 /// This function should only be called when distributing mass. As long as
945 /// there are no irreducilbe edges to Node, then it will have complexity
946 /// O(1) in this context.
948 /// In general, the complexity is O(L), where L is the number of loop
949 /// headers Node has been packaged into. Since this method is called in
950 /// the context of distributing mass, L will be the number of loop headers
951 /// an early exit edge jumps out of.
952 BlockNode getResolvedNode() const {
953 auto L = getPackagedLoop();
954 return L ? L->getHeader() : Node;
956 LoopData *getPackagedLoop() const {
957 if (!Loop || !Loop->IsPackaged)
960 while (L->Parent && L->Parent->IsPackaged)
965 /// \brief Get the appropriate mass for a node.
967 /// Get appropriate mass for Node. If Node is a loop-header (whose loop
968 /// has been packaged), returns the mass of its pseudo-node. If it's a
969 /// node inside a packaged loop, it returns the loop's mass.
970 BlockMass &getMass() {
973 if (!isADoublePackage())
975 return Loop->Parent->Mass;
978 /// \brief Has ContainingLoop been packaged up?
979 bool isPackaged() const { return getResolvedNode() != Node; }
980 /// \brief Has Loop been packaged up?
981 bool isAPackage() const { return isLoopHeader() && Loop->IsPackaged; }
982 /// \brief Has Loop been packaged up twice?
983 bool isADoublePackage() const {
984 return isDoubleLoopHeader() && Loop->Parent->IsPackaged;
988 /// \brief Unscaled probability weight.
990 /// Probability weight for an edge in the graph (including the
991 /// successor/target node).
993 /// All edges in the original function are 32-bit. However, exit edges from
994 /// loop packages are taken from 64-bit exit masses, so we need 64-bits of
995 /// space in general.
997 /// In addition to the raw weight amount, Weight stores the type of the edge
998 /// in the current context (i.e., the context of the loop being processed).
999 /// Is this a local edge within the loop, an exit from the loop, or a
1000 /// backedge to the loop header?
1002 enum DistType { Local, Exit, Backedge };
1004 BlockNode TargetNode;
1006 Weight() : Type(Local), Amount(0) {}
1009 /// \brief Distribution of unscaled probability weight.
1011 /// Distribution of unscaled probability weight to a set of successors.
1013 /// This class collates the successor edge weights for later processing.
1015 /// \a DidOverflow indicates whether \a Total did overflow while adding to
1016 /// the distribution. It should never overflow twice.
1017 struct Distribution {
1018 typedef SmallVector<Weight, 4> WeightList;
1019 WeightList Weights; ///< Individual successor weights.
1020 uint64_t Total; ///< Sum of all weights.
1021 bool DidOverflow; ///< Whether \a Total did overflow.
1023 Distribution() : Total(0), DidOverflow(false) {}
1024 void addLocal(const BlockNode &Node, uint64_t Amount) {
1025 add(Node, Amount, Weight::Local);
1027 void addExit(const BlockNode &Node, uint64_t Amount) {
1028 add(Node, Amount, Weight::Exit);
1030 void addBackedge(const BlockNode &Node, uint64_t Amount) {
1031 add(Node, Amount, Weight::Backedge);
1034 /// \brief Normalize the distribution.
1036 /// Combines multiple edges to the same \a Weight::TargetNode and scales
1037 /// down so that \a Total fits into 32-bits.
1039 /// This is linear in the size of \a Weights. For the vast majority of
1040 /// cases, adjacent edge weights are combined by sorting WeightList and
1041 /// combining adjacent weights. However, for very large edge lists an
1042 /// auxiliary hash table is used.
1046 void add(const BlockNode &Node, uint64_t Amount, Weight::DistType Type);
1049 /// \brief Data about each block. This is used downstream.
1050 std::vector<FrequencyData> Freqs;
1052 /// \brief Loop data: see initializeLoops().
1053 std::vector<WorkingData> Working;
1055 /// \brief Indexed information about loops.
1056 std::list<LoopData> Loops;
1058 /// \brief Add all edges out of a packaged loop to the distribution.
1060 /// Adds all edges from LocalLoopHead to Dist. Calls addToDist() to add each
1063 /// \return \c true unless there's an irreducible backedge.
1064 bool addLoopSuccessorsToDist(const LoopData *OuterLoop, LoopData &Loop,
1065 Distribution &Dist);
1067 /// \brief Add an edge to the distribution.
1069 /// Adds an edge to Succ to Dist. If \c LoopHead.isValid(), then whether the
1070 /// edge is local/exit/backedge is in the context of LoopHead. Otherwise,
1071 /// every edge should be a local edge (since all the loops are packaged up).
1073 /// \return \c true unless aborted due to an irreducible backedge.
1074 bool addToDist(Distribution &Dist, const LoopData *OuterLoop,
1075 const BlockNode &Pred, const BlockNode &Succ, uint64_t Weight);
1077 LoopData &getLoopPackage(const BlockNode &Head) {
1078 assert(Head.Index < Working.size());
1079 assert(Working[Head.Index].isLoopHeader());
1080 return *Working[Head.Index].Loop;
1083 /// \brief Analyze irreducible SCCs.
1085 /// Separate irreducible SCCs from \c G, which is an explict graph of \c
1086 /// OuterLoop (or the top-level function, if \c OuterLoop is \c nullptr).
1087 /// Insert them into \a Loops before \c Insert.
1089 /// \return the \c LoopData nodes representing the irreducible SCCs.
1090 iterator_range<std::list<LoopData>::iterator>
1091 analyzeIrreducible(const bfi_detail::IrreducibleGraph &G, LoopData *OuterLoop,
1092 std::list<LoopData>::iterator Insert);
1094 /// \brief Update a loop after packaging irreducible SCCs inside of it.
1096 /// Update \c OuterLoop. Before finding irreducible control flow, it was
1097 /// partway through \a computeMassInLoop(), so \a LoopData::Exits and \a
1098 /// LoopData::BackedgeMass need to be reset. Also, nodes that were packaged
1099 /// up need to be removed from \a OuterLoop::Nodes.
1100 void updateLoopWithIrreducible(LoopData &OuterLoop);
1102 /// \brief Distribute mass according to a distribution.
1104 /// Distributes the mass in Source according to Dist. If LoopHead.isValid(),
1105 /// backedges and exits are stored in its entry in Loops.
1107 /// Mass is distributed in parallel from two copies of the source mass.
1108 void distributeMass(const BlockNode &Source, LoopData *OuterLoop,
1109 Distribution &Dist);
1111 /// \brief Compute the loop scale for a loop.
1112 void computeLoopScale(LoopData &Loop);
1114 /// \brief Package up a loop.
1115 void packageLoop(LoopData &Loop);
1117 /// \brief Unwrap loops.
1120 /// \brief Finalize frequency metrics.
1122 /// Calculates final frequencies and cleans up no-longer-needed data
1124 void finalizeMetrics();
1126 /// \brief Clear all memory.
1129 virtual std::string getBlockName(const BlockNode &Node) const;
1130 std::string getLoopName(const LoopData &Loop) const;
1132 virtual raw_ostream &print(raw_ostream &OS) const { return OS; }
1133 void dump() const { print(dbgs()); }
1135 Float getFloatingBlockFreq(const BlockNode &Node) const;
1137 BlockFrequency getBlockFreq(const BlockNode &Node) const;
1139 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockNode &Node) const;
1140 raw_ostream &printBlockFreq(raw_ostream &OS,
1141 const BlockFrequency &Freq) const;
1143 uint64_t getEntryFreq() const {
1144 assert(!Freqs.empty());
1145 return Freqs[0].Integer;
1147 /// \brief Virtual destructor.
1149 /// Need a virtual destructor to mask the compiler warning about
1151 virtual ~BlockFrequencyInfoImplBase() {}
1154 namespace bfi_detail {
1155 template <class BlockT> struct TypeMap {};
1156 template <> struct TypeMap<BasicBlock> {
1157 typedef BasicBlock BlockT;
1158 typedef Function FunctionT;
1159 typedef BranchProbabilityInfo BranchProbabilityInfoT;
1161 typedef LoopInfo LoopInfoT;
1163 template <> struct TypeMap<MachineBasicBlock> {
1164 typedef MachineBasicBlock BlockT;
1165 typedef MachineFunction FunctionT;
1166 typedef MachineBranchProbabilityInfo BranchProbabilityInfoT;
1167 typedef MachineLoop LoopT;
1168 typedef MachineLoopInfo LoopInfoT;
1171 /// \brief Get the name of a MachineBasicBlock.
1173 /// Get the name of a MachineBasicBlock. It's templated so that including from
1174 /// CodeGen is unnecessary (that would be a layering issue).
1176 /// This is used mainly for debug output. The name is similar to
1177 /// MachineBasicBlock::getFullName(), but skips the name of the function.
1178 template <class BlockT> std::string getBlockName(const BlockT *BB) {
1179 assert(BB && "Unexpected nullptr");
1180 auto MachineName = "BB" + Twine(BB->getNumber());
1181 if (BB->getBasicBlock())
1182 return (MachineName + "[" + BB->getName() + "]").str();
1183 return MachineName.str();
1185 /// \brief Get the name of a BasicBlock.
1186 template <> inline std::string getBlockName(const BasicBlock *BB) {
1187 assert(BB && "Unexpected nullptr");
1188 return BB->getName().str();
1191 /// \brief Graph of irreducible control flow.
1193 /// This graph is used for determining the SCCs in a loop (or top-level
1194 /// function) that has irreducible control flow.
1196 /// During the block frequency algorithm, the local graphs are defined in a
1197 /// light-weight way, deferring to the \a BasicBlock or \a MachineBasicBlock
1198 /// graphs for most edges, but getting others from \a LoopData::ExitMap. The
1199 /// latter only has successor information.
1201 /// \a IrreducibleGraph makes this graph explicit. It's in a form that can use
1202 /// \a GraphTraits (so that \a analyzeIrreducible() can use \a scc_iterator),
1203 /// and it explicitly lists predecessors and successors. The initialization
1204 /// that relies on \c MachineBasicBlock is defined in the header.
1205 struct IrreducibleGraph {
1206 typedef BlockFrequencyInfoImplBase BFIBase;
1210 typedef BFIBase::BlockNode BlockNode;
1214 std::deque<const IrrNode *> Edges;
1215 IrrNode(const BlockNode &Node) : Node(Node), NumIn(0) {}
1217 typedef std::deque<const IrrNode *>::const_iterator iterator;
1218 iterator pred_begin() const { return Edges.begin(); }
1219 iterator succ_begin() const { return Edges.begin() + NumIn; }
1220 iterator pred_end() const { return succ_begin(); }
1221 iterator succ_end() const { return Edges.end(); }
1224 const IrrNode *StartIrr;
1225 std::vector<IrrNode> Nodes;
1226 SmallDenseMap<uint32_t, IrrNode *, 4> Lookup;
1228 /// \brief Construct an explicit graph containing irreducible control flow.
1230 /// Construct an explicit graph of the control flow in \c OuterLoop (or the
1231 /// top-level function, if \c OuterLoop is \c nullptr). Uses \c
1232 /// addBlockEdges to add block successors that have not been packaged into
1235 /// \a BlockFrequencyInfoImpl::computeIrreducibleMass() is the only expected
1237 template <class BlockEdgesAdder>
1238 IrreducibleGraph(BFIBase &BFI, const BFIBase::LoopData *OuterLoop,
1239 BlockEdgesAdder addBlockEdges)
1240 : BFI(BFI), StartIrr(nullptr) {
1241 initialize(OuterLoop, addBlockEdges);
1244 template <class BlockEdgesAdder>
1245 void initialize(const BFIBase::LoopData *OuterLoop,
1246 BlockEdgesAdder addBlockEdges);
1247 void addNodesInLoop(const BFIBase::LoopData &OuterLoop);
1248 void addNodesInFunction();
1249 void addNode(const BlockNode &Node) {
1250 Nodes.emplace_back(Node);
1251 BFI.Working[Node.Index].getMass() = BlockMass::getEmpty();
1254 template <class BlockEdgesAdder>
1255 void addEdges(const BlockNode &Node, const BFIBase::LoopData *OuterLoop,
1256 BlockEdgesAdder addBlockEdges);
1257 void addEdge(IrrNode &Irr, const BlockNode &Succ,
1258 const BFIBase::LoopData *OuterLoop);
1260 template <class BlockEdgesAdder>
1261 void IrreducibleGraph::initialize(const BFIBase::LoopData *OuterLoop,
1262 BlockEdgesAdder addBlockEdges) {
1264 addNodesInLoop(*OuterLoop);
1265 for (auto N : OuterLoop->Nodes)
1266 addEdges(N, OuterLoop, addBlockEdges);
1268 addNodesInFunction();
1269 for (uint32_t Index = 0; Index < BFI.Working.size(); ++Index)
1270 addEdges(Index, OuterLoop, addBlockEdges);
1272 StartIrr = Lookup[Start.Index];
1274 template <class BlockEdgesAdder>
1275 void IrreducibleGraph::addEdges(const BlockNode &Node,
1276 const BFIBase::LoopData *OuterLoop,
1277 BlockEdgesAdder addBlockEdges) {
1278 auto L = Lookup.find(Node.Index);
1279 if (L == Lookup.end())
1281 IrrNode &Irr = *L->second;
1282 const auto &Working = BFI.Working[Node.Index];
1284 if (Working.isAPackage())
1285 for (const auto &I : Working.Loop->Exits)
1286 addEdge(Irr, I.first, OuterLoop);
1288 addBlockEdges(*this, Irr, OuterLoop);
1292 /// \brief Shared implementation for block frequency analysis.
1294 /// This is a shared implementation of BlockFrequencyInfo and
1295 /// MachineBlockFrequencyInfo, and calculates the relative frequencies of
1298 /// LoopInfo defines a loop as a "non-trivial" SCC dominated by a single block,
1299 /// which is called the header. A given loop, L, can have sub-loops, which are
1300 /// loops within the subgraph of L that exclude its header. (A "trivial" SCC
1301 /// consists of a single block that does not have a self-edge.)
1303 /// In addition to loops, this algorithm has limited support for irreducible
1304 /// SCCs, which are SCCs with multiple entry blocks. Irreducible SCCs are
1305 /// discovered on they fly, and modelled as loops with multiple headers.
1307 /// The headers of irreducible sub-SCCs consist of its entry blocks and all
1308 /// nodes that are targets of a backedge within it (excluding backedges within
1309 /// true sub-loops). Block frequency calculations act as if a block is
1310 /// inserted that intercepts all the edges to the headers. All backedges and
1311 /// entries point to this block. Its successors are the headers, which split
1312 /// the frequency evenly.
1314 /// This algorithm leverages BlockMass and UnsignedFloat to maintain precision,
1315 /// separates mass distribution from loop scaling, and dithers to eliminate
1316 /// probability mass loss.
1318 /// The implementation is split between BlockFrequencyInfoImpl, which knows the
1319 /// type of graph being modelled (BasicBlock vs. MachineBasicBlock), and
1320 /// BlockFrequencyInfoImplBase, which doesn't. The base class uses \a
1321 /// BlockNode, a wrapper around a uint32_t. BlockNode is numbered from 0 in
1322 /// reverse-post order. This gives two advantages: it's easy to compare the
1323 /// relative ordering of two nodes, and maps keyed on BlockT can be represented
1326 /// This algorithm is O(V+E), unless there is irreducible control flow, in
1327 /// which case it's O(V*E) in the worst case.
1329 /// These are the main stages:
1331 /// 0. Reverse post-order traversal (\a initializeRPOT()).
1333 /// Run a single post-order traversal and save it (in reverse) in RPOT.
1334 /// All other stages make use of this ordering. Save a lookup from BlockT
1335 /// to BlockNode (the index into RPOT) in Nodes.
1337 /// 1. Loop initialization (\a initializeLoops()).
1339 /// Translate LoopInfo/MachineLoopInfo into a form suitable for the rest of
1340 /// the algorithm. In particular, store the immediate members of each loop
1341 /// in reverse post-order.
1343 /// 2. Calculate mass and scale in loops (\a computeMassInLoops()).
1345 /// For each loop (bottom-up), distribute mass through the DAG resulting
1346 /// from ignoring backedges and treating sub-loops as a single pseudo-node.
1347 /// Track the backedge mass distributed to the loop header, and use it to
1348 /// calculate the loop scale (number of loop iterations). Immediate
1349 /// members that represent sub-loops will already have been visited and
1350 /// packaged into a pseudo-node.
1352 /// Distributing mass in a loop is a reverse-post-order traversal through
1353 /// the loop. Start by assigning full mass to the Loop header. For each
1354 /// node in the loop:
1356 /// - Fetch and categorize the weight distribution for its successors.
1357 /// If this is a packaged-subloop, the weight distribution is stored
1358 /// in \a LoopData::Exits. Otherwise, fetch it from
1359 /// BranchProbabilityInfo.
1361 /// - Each successor is categorized as \a Weight::Local, a local edge
1362 /// within the current loop, \a Weight::Backedge, a backedge to the
1363 /// loop header, or \a Weight::Exit, any successor outside the loop.
1364 /// The weight, the successor, and its category are stored in \a
1365 /// Distribution. There can be multiple edges to each successor.
1367 /// - If there's a backedge to a non-header, there's an irreducible SCC.
1368 /// The usual flow is temporarily aborted. \a
1369 /// computeIrreducibleMass() finds the irreducible SCCs within the
1370 /// loop, packages them up, and restarts the flow.
1372 /// - Normalize the distribution: scale weights down so that their sum
1373 /// is 32-bits, and coalesce multiple edges to the same node.
1375 /// - Distribute the mass accordingly, dithering to minimize mass loss,
1376 /// as described in \a distributeMass().
1378 /// Finally, calculate the loop scale from the accumulated backedge mass.
1380 /// 3. Distribute mass in the function (\a computeMassInFunction()).
1382 /// Finally, distribute mass through the DAG resulting from packaging all
1383 /// loops in the function. This uses the same algorithm as distributing
1384 /// mass in a loop, except that there are no exit or backedge edges.
1386 /// 4. Unpackage loops (\a unwrapLoops()).
1388 /// Initialize each block's frequency to a floating point representation of
1391 /// Visit loops top-down, scaling the frequencies of its immediate members
1392 /// by the loop's pseudo-node's frequency.
1394 /// 5. Convert frequencies to a 64-bit range (\a finalizeMetrics()).
1396 /// Using the min and max frequencies as a guide, translate floating point
1397 /// frequencies to an appropriate range in uint64_t.
1399 /// It has some known flaws.
1401 /// - Loop scale is limited to 4096 per loop (2^12) to avoid exhausting
1402 /// BlockFrequency's 64-bit integer precision.
1404 /// - The model of irreducible control flow is a rough approximation.
1406 /// Modelling irreducible control flow exactly involves setting up and
1407 /// solving a group of infinite geometric series. Such precision is
1408 /// unlikely to be worthwhile, since most of our algorithms give up on
1409 /// irreducible control flow anyway.
1411 /// Nevertheless, we might find that we need to get closer. Here's a sort
1412 /// of TODO list for the model with diminishing returns, to be completed as
1415 /// - The headers for the \a LoopData representing an irreducible SCC
1416 /// include non-entry blocks. When these extra blocks exist, they
1417 /// indicate a self-contained irreducible sub-SCC. We could treat them
1418 /// as sub-loops, rather than arbitrarily shoving the problematic
1419 /// blocks into the headers of the main irreducible SCC.
1421 /// - Backedge frequencies are assumed to be evenly split between the
1422 /// headers of a given irreducible SCC. Instead, we could track the
1423 /// backedge mass separately for each header, and adjust their relative
1426 /// - Entry frequencies are assumed to be evenly split between the
1427 /// headers of a given irreducible SCC, which is the only option if we
1428 /// need to compute mass in the SCC before its parent loop. Instead,
1429 /// we could partially compute mass in the parent loop, and stop when
1430 /// we get to the SCC. Here, we have the correct ratio of entry
1431 /// masses, which we can use to adjust their relative frequencies.
1432 /// Compute mass in the SCC, and then continue propagation in the
1435 /// - We can propagate mass iteratively through the SCC, for some fixed
1436 /// number of iterations. Each iteration starts by assigning the entry
1437 /// blocks their backedge mass from the prior iteration. The final
1438 /// mass for each block (and each exit, and the total backedge mass
1439 /// used for computing loop scale) is the sum of all iterations.
1440 /// (Running this until fixed point would "solve" the geometric
1441 /// series by simulation.)
1442 template <class BT> class BlockFrequencyInfoImpl : BlockFrequencyInfoImplBase {
1443 typedef typename bfi_detail::TypeMap<BT>::BlockT BlockT;
1444 typedef typename bfi_detail::TypeMap<BT>::FunctionT FunctionT;
1445 typedef typename bfi_detail::TypeMap<BT>::BranchProbabilityInfoT
1446 BranchProbabilityInfoT;
1447 typedef typename bfi_detail::TypeMap<BT>::LoopT LoopT;
1448 typedef typename bfi_detail::TypeMap<BT>::LoopInfoT LoopInfoT;
1450 // This is part of a workaround for a GCC 4.7 crash on lambdas.
1451 friend struct bfi_detail::BlockEdgesAdder<BT>;
1453 typedef GraphTraits<const BlockT *> Successor;
1454 typedef GraphTraits<Inverse<const BlockT *>> Predecessor;
1456 const BranchProbabilityInfoT *BPI;
1457 const LoopInfoT *LI;
1460 // All blocks in reverse postorder.
1461 std::vector<const BlockT *> RPOT;
1462 DenseMap<const BlockT *, BlockNode> Nodes;
1464 typedef typename std::vector<const BlockT *>::const_iterator rpot_iterator;
1466 rpot_iterator rpot_begin() const { return RPOT.begin(); }
1467 rpot_iterator rpot_end() const { return RPOT.end(); }
1469 size_t getIndex(const rpot_iterator &I) const { return I - rpot_begin(); }
1471 BlockNode getNode(const rpot_iterator &I) const {
1472 return BlockNode(getIndex(I));
1474 BlockNode getNode(const BlockT *BB) const { return Nodes.lookup(BB); }
1476 const BlockT *getBlock(const BlockNode &Node) const {
1477 assert(Node.Index < RPOT.size());
1478 return RPOT[Node.Index];
1481 /// \brief Run (and save) a post-order traversal.
1483 /// Saves a reverse post-order traversal of all the nodes in \a F.
1484 void initializeRPOT();
1486 /// \brief Initialize loop data.
1488 /// Build up \a Loops using \a LoopInfo. \a LoopInfo gives us a mapping from
1489 /// each block to the deepest loop it's in, but we need the inverse. For each
1490 /// loop, we store in reverse post-order its "immediate" members, defined as
1491 /// the header, the headers of immediate sub-loops, and all other blocks in
1492 /// the loop that are not in sub-loops.
1493 void initializeLoops();
1495 /// \brief Propagate to a block's successors.
1497 /// In the context of distributing mass through \c OuterLoop, divide the mass
1498 /// currently assigned to \c Node between its successors.
1500 /// \return \c true unless there's an irreducible backedge.
1501 bool propagateMassToSuccessors(LoopData *OuterLoop, const BlockNode &Node);
1503 /// \brief Compute mass in a particular loop.
1505 /// Assign mass to \c Loop's header, and then for each block in \c Loop in
1506 /// reverse post-order, distribute mass to its successors. Only visits nodes
1507 /// that have not been packaged into sub-loops.
1509 /// \pre \a computeMassInLoop() has been called for each subloop of \c Loop.
1510 /// \return \c true unless there's an irreducible backedge.
1511 bool computeMassInLoop(LoopData &Loop);
1513 /// \brief Try to compute mass in the top-level function.
1515 /// Assign mass to the entry block, and then for each block in reverse
1516 /// post-order, distribute mass to its successors. Skips nodes that have
1517 /// been packaged into loops.
1519 /// \pre \a computeMassInLoops() has been called.
1520 /// \return \c true unless there's an irreducible backedge.
1521 bool tryToComputeMassInFunction();
1523 /// \brief Compute mass in (and package up) irreducible SCCs.
1525 /// Find the irreducible SCCs in \c OuterLoop, add them to \a Loops (in front
1526 /// of \c Insert), and call \a computeMassInLoop() on each of them.
1528 /// If \c OuterLoop is \c nullptr, it refers to the top-level function.
1530 /// \pre \a computeMassInLoop() has been called for each subloop of \c
1532 /// \pre \c Insert points at the the last loop successfully processed by \a
1533 /// computeMassInLoop().
1534 /// \pre \c OuterLoop has irreducible SCCs.
1535 void computeIrreducibleMass(LoopData *OuterLoop,
1536 std::list<LoopData>::iterator Insert);
1538 /// \brief Compute mass in all loops.
1540 /// For each loop bottom-up, call \a computeMassInLoop().
1542 /// \a computeMassInLoop() aborts (and returns \c false) on loops that
1543 /// contain a irreducible sub-SCCs. Use \a computeIrreducibleMass() and then
1544 /// re-enter \a computeMassInLoop().
1546 /// \post \a computeMassInLoop() has returned \c true for every loop.
1547 void computeMassInLoops();
1549 /// \brief Compute mass in the top-level function.
1551 /// Uses \a tryToComputeMassInFunction() and \a computeIrreducibleMass() to
1552 /// compute mass in the top-level function.
1554 /// \post \a tryToComputeMassInFunction() has returned \c true.
1555 void computeMassInFunction();
1557 std::string getBlockName(const BlockNode &Node) const override {
1558 return bfi_detail::getBlockName(getBlock(Node));
1562 const FunctionT *getFunction() const { return F; }
1564 void doFunction(const FunctionT *F, const BranchProbabilityInfoT *BPI,
1565 const LoopInfoT *LI);
1566 BlockFrequencyInfoImpl() : BPI(nullptr), LI(nullptr), F(nullptr) {}
1568 using BlockFrequencyInfoImplBase::getEntryFreq;
1569 BlockFrequency getBlockFreq(const BlockT *BB) const {
1570 return BlockFrequencyInfoImplBase::getBlockFreq(getNode(BB));
1572 Float getFloatingBlockFreq(const BlockT *BB) const {
1573 return BlockFrequencyInfoImplBase::getFloatingBlockFreq(getNode(BB));
1576 /// \brief Print the frequencies for the current function.
1578 /// Prints the frequencies for the blocks in the current function.
1580 /// Blocks are printed in the natural iteration order of the function, rather
1581 /// than reverse post-order. This provides two advantages: writing -analyze
1582 /// tests is easier (since blocks come out in source order), and even
1583 /// unreachable blocks are printed.
1585 /// \a BlockFrequencyInfoImplBase::print() only knows reverse post-order, so
1586 /// we need to override it here.
1587 raw_ostream &print(raw_ostream &OS) const override;
1588 using BlockFrequencyInfoImplBase::dump;
1590 using BlockFrequencyInfoImplBase::printBlockFreq;
1591 raw_ostream &printBlockFreq(raw_ostream &OS, const BlockT *BB) const {
1592 return BlockFrequencyInfoImplBase::printBlockFreq(OS, getNode(BB));
1597 void BlockFrequencyInfoImpl<BT>::doFunction(const FunctionT *F,
1598 const BranchProbabilityInfoT *BPI,
1599 const LoopInfoT *LI) {
1600 // Save the parameters.
1605 // Clean up left-over data structures.
1606 BlockFrequencyInfoImplBase::clear();
1611 DEBUG(dbgs() << "\nblock-frequency: " << F->getName() << "\n================="
1612 << std::string(F->getName().size(), '=') << "\n");
1616 // Visit loops in post-order to find thelocal mass distribution, and then do
1617 // the full function.
1618 computeMassInLoops();
1619 computeMassInFunction();
1624 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeRPOT() {
1625 const BlockT *Entry = F->begin();
1626 RPOT.reserve(F->size());
1627 std::copy(po_begin(Entry), po_end(Entry), std::back_inserter(RPOT));
1628 std::reverse(RPOT.begin(), RPOT.end());
1630 assert(RPOT.size() - 1 <= BlockNode::getMaxIndex() &&
1631 "More nodes in function than Block Frequency Info supports");
1633 DEBUG(dbgs() << "reverse-post-order-traversal\n");
1634 for (rpot_iterator I = rpot_begin(), E = rpot_end(); I != E; ++I) {
1635 BlockNode Node = getNode(I);
1636 DEBUG(dbgs() << " - " << getIndex(I) << ": " << getBlockName(Node) << "\n");
1640 Working.reserve(RPOT.size());
1641 for (size_t Index = 0; Index < RPOT.size(); ++Index)
1642 Working.emplace_back(Index);
1643 Freqs.resize(RPOT.size());
1646 template <class BT> void BlockFrequencyInfoImpl<BT>::initializeLoops() {
1647 DEBUG(dbgs() << "loop-detection\n");
1651 // Visit loops top down and assign them an index.
1652 std::deque<std::pair<const LoopT *, LoopData *>> Q;
1653 for (const LoopT *L : *LI)
1654 Q.emplace_back(L, nullptr);
1655 while (!Q.empty()) {
1656 const LoopT *Loop = Q.front().first;
1657 LoopData *Parent = Q.front().second;
1660 BlockNode Header = getNode(Loop->getHeader());
1661 assert(Header.isValid());
1663 Loops.emplace_back(Parent, Header);
1664 Working[Header.Index].Loop = &Loops.back();
1665 DEBUG(dbgs() << " - loop = " << getBlockName(Header) << "\n");
1667 for (const LoopT *L : *Loop)
1668 Q.emplace_back(L, &Loops.back());
1671 // Visit nodes in reverse post-order and add them to their deepest containing
1673 for (size_t Index = 0; Index < RPOT.size(); ++Index) {
1674 // Loop headers have already been mostly mapped.
1675 if (Working[Index].isLoopHeader()) {
1676 LoopData *ContainingLoop = Working[Index].getContainingLoop();
1678 ContainingLoop->Nodes.push_back(Index);
1682 const LoopT *Loop = LI->getLoopFor(RPOT[Index]);
1686 // Add this node to its containing loop's member list.
1687 BlockNode Header = getNode(Loop->getHeader());
1688 assert(Header.isValid());
1689 const auto &HeaderData = Working[Header.Index];
1690 assert(HeaderData.isLoopHeader());
1692 Working[Index].Loop = HeaderData.Loop;
1693 HeaderData.Loop->Nodes.push_back(Index);
1694 DEBUG(dbgs() << " - loop = " << getBlockName(Header)
1695 << ": member = " << getBlockName(Index) << "\n");
1699 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInLoops() {
1700 // Visit loops with the deepest first, and the top-level loops last.
1701 for (auto L = Loops.rbegin(), E = Loops.rend(); L != E; ++L) {
1702 if (computeMassInLoop(*L))
1704 auto Next = std::next(L);
1705 computeIrreducibleMass(&*L, L.base());
1706 L = std::prev(Next);
1707 if (computeMassInLoop(*L))
1709 llvm_unreachable("unhandled irreducible control flow");
1714 bool BlockFrequencyInfoImpl<BT>::computeMassInLoop(LoopData &Loop) {
1715 // Compute mass in loop.
1716 DEBUG(dbgs() << "compute-mass-in-loop: " << getLoopName(Loop) << "\n");
1718 if (Loop.isIrreducible()) {
1719 BlockMass Remaining = BlockMass::getFull();
1720 for (uint32_t H = 0; H < Loop.NumHeaders; ++H) {
1721 auto &Mass = Working[Loop.Nodes[H].Index].getMass();
1722 Mass = Remaining * BranchProbability(1, Loop.NumHeaders - H);
1725 for (const BlockNode &M : Loop.Nodes)
1726 if (!propagateMassToSuccessors(&Loop, M))
1727 llvm_unreachable("unhandled irreducible control flow");
1729 Working[Loop.getHeader().Index].getMass() = BlockMass::getFull();
1730 if (!propagateMassToSuccessors(&Loop, Loop.getHeader()))
1731 llvm_unreachable("irreducible control flow to loop header!?");
1732 for (const BlockNode &M : Loop.members())
1733 if (!propagateMassToSuccessors(&Loop, M))
1734 // Irreducible backedge.
1738 computeLoopScale(Loop);
1744 bool BlockFrequencyInfoImpl<BT>::tryToComputeMassInFunction() {
1745 // Compute mass in function.
1746 DEBUG(dbgs() << "compute-mass-in-function\n");
1747 assert(!Working.empty() && "no blocks in function");
1748 assert(!Working[0].isLoopHeader() && "entry block is a loop header");
1750 Working[0].getMass() = BlockMass::getFull();
1751 for (rpot_iterator I = rpot_begin(), IE = rpot_end(); I != IE; ++I) {
1752 // Check for nodes that have been packaged.
1753 BlockNode Node = getNode(I);
1754 if (Working[Node.Index].isPackaged())
1757 if (!propagateMassToSuccessors(nullptr, Node))
1763 template <class BT> void BlockFrequencyInfoImpl<BT>::computeMassInFunction() {
1764 if (tryToComputeMassInFunction())
1766 computeIrreducibleMass(nullptr, Loops.begin());
1767 if (tryToComputeMassInFunction())
1769 llvm_unreachable("unhandled irreducible control flow");
1772 /// \note This should be a lambda, but that crashes GCC 4.7.
1773 namespace bfi_detail {
1774 template <class BT> struct BlockEdgesAdder {
1776 typedef BlockFrequencyInfoImplBase::LoopData LoopData;
1777 typedef GraphTraits<const BlockT *> Successor;
1779 const BlockFrequencyInfoImpl<BT> &BFI;
1780 explicit BlockEdgesAdder(const BlockFrequencyInfoImpl<BT> &BFI)
1782 void operator()(IrreducibleGraph &G, IrreducibleGraph::IrrNode &Irr,
1783 const LoopData *OuterLoop) {
1784 const BlockT *BB = BFI.RPOT[Irr.Node.Index];
1785 for (auto I = Successor::child_begin(BB), E = Successor::child_end(BB);
1787 G.addEdge(Irr, BFI.getNode(*I), OuterLoop);
1792 void BlockFrequencyInfoImpl<BT>::computeIrreducibleMass(
1793 LoopData *OuterLoop, std::list<LoopData>::iterator Insert) {
1794 DEBUG(dbgs() << "analyze-irreducible-in-";
1795 if (OuterLoop) dbgs() << "loop: " << getLoopName(*OuterLoop) << "\n";
1796 else dbgs() << "function\n");
1798 using namespace bfi_detail;
1799 // Ideally, addBlockEdges() would be declared here as a lambda, but that
1801 BlockEdgesAdder<BT> addBlockEdges(*this);
1802 IrreducibleGraph G(*this, OuterLoop, addBlockEdges);
1804 for (auto &L : analyzeIrreducible(G, OuterLoop, Insert))
1805 computeMassInLoop(L);
1809 updateLoopWithIrreducible(*OuterLoop);
1814 BlockFrequencyInfoImpl<BT>::propagateMassToSuccessors(LoopData *OuterLoop,
1815 const BlockNode &Node) {
1816 DEBUG(dbgs() << " - node: " << getBlockName(Node) << "\n");
1817 // Calculate probability for successors.
1819 if (auto *Loop = Working[Node.Index].getPackagedLoop()) {
1820 assert(Loop != OuterLoop && "Cannot propagate mass in a packaged loop");
1821 if (!addLoopSuccessorsToDist(OuterLoop, *Loop, Dist))
1822 // Irreducible backedge.
1825 const BlockT *BB = getBlock(Node);
1826 for (auto SI = Successor::child_begin(BB), SE = Successor::child_end(BB);
1828 // Do not dereference SI, or getEdgeWeight() is linear in the number of
1830 if (!addToDist(Dist, OuterLoop, Node, getNode(*SI),
1831 BPI->getEdgeWeight(BB, SI)))
1832 // Irreducible backedge.
1836 // Distribute mass to successors, saving exit and backedge data in the
1838 distributeMass(Node, OuterLoop, Dist);
1843 raw_ostream &BlockFrequencyInfoImpl<BT>::print(raw_ostream &OS) const {
1846 OS << "block-frequency-info: " << F->getName() << "\n";
1847 for (const BlockT &BB : *F)
1848 OS << " - " << bfi_detail::getBlockName(&BB)
1849 << ": float = " << getFloatingBlockFreq(&BB)
1850 << ", int = " << getBlockFreq(&BB).getFrequency() << "\n";
1852 // Add an extra newline for readability.