1 //===- BranchProbability.h - Branch Probability Wrapper ---------*- 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 // Definition of BranchProbability shared by IR and Machine Instructions.
12 //===----------------------------------------------------------------------===//
14 #ifndef LLVM_SUPPORT_BRANCHPROBABILITY_H
15 #define LLVM_SUPPORT_BRANCHPROBABILITY_H
17 #include "llvm/Support/DataTypes.h"
27 // This class represents Branch Probability as a non-negative fraction that is
28 // no greater than 1. It uses a fixed-point-like implementation, in which the
29 // denominator is always a constant value (here we use 1<<31 for maximum
31 class BranchProbability {
35 // Denominator, which is a constant value.
36 static const uint32_t D = 1u << 31;
37 static const uint32_t UnknownN = UINT32_MAX;
39 // Construct a BranchProbability with only numerator assuming the denominator
40 // is 1<<31. For internal use only.
41 explicit BranchProbability(uint32_t n) : N(n) {}
44 BranchProbability() : N(0) {}
45 BranchProbability(uint32_t Numerator, uint32_t Denominator);
47 bool isZero() const { return N == 0; }
48 bool isUnknown() const { return N == UnknownN; }
50 static BranchProbability getZero() { return BranchProbability(0); }
51 static BranchProbability getOne() { return BranchProbability(D); }
52 static BranchProbability getUnknown() { return BranchProbability(UnknownN); }
53 // Create a BranchProbability object with the given numerator and 1<<31
55 static BranchProbability getRaw(uint32_t N) { return BranchProbability(N); }
57 // Normalize given probabilties so that the sum of them becomes approximate
59 template <class ProbabilityIter>
60 static void normalizeProbabilities(ProbabilityIter Begin,
63 // Normalize a list of weights by scaling them down so that the sum of them
64 // doesn't exceed UINT32_MAX.
65 template <class WeightListIter>
66 static void normalizeEdgeWeights(WeightListIter Begin, WeightListIter End);
68 uint32_t getNumerator() const { return N; }
69 static uint32_t getDenominator() { return D; }
71 // Return (1 - Probability).
72 BranchProbability getCompl() const { return BranchProbability(D - N); }
74 raw_ostream &print(raw_ostream &OS) const;
78 /// \brief Scale a large integer.
80 /// Scales \c Num. Guarantees full precision. Returns the floor of the
83 /// \return \c Num times \c this.
84 uint64_t scale(uint64_t Num) const;
86 /// \brief Scale a large integer by the inverse.
88 /// Scales \c Num by the inverse of \c this. Guarantees full precision.
89 /// Returns the floor of the result.
91 /// \return \c Num divided by \c this.
92 uint64_t scaleByInverse(uint64_t Num) const;
94 BranchProbability &operator+=(BranchProbability RHS) {
95 // Saturate the result in case of overflow.
96 N = (uint64_t(N) + RHS.N > D) ? D : N + RHS.N;
100 BranchProbability &operator-=(BranchProbability RHS) {
101 // Saturate the result in case of underflow.
102 N = N < RHS.N ? 0 : N - RHS.N;
106 BranchProbability &operator*=(BranchProbability RHS) {
107 N = (static_cast<uint64_t>(N) * RHS.N + D / 2) / D;
111 BranchProbability operator+(BranchProbability RHS) const {
112 BranchProbability Prob(*this);
116 BranchProbability operator-(BranchProbability RHS) const {
117 BranchProbability Prob(*this);
121 BranchProbability operator*(BranchProbability RHS) const {
122 BranchProbability Prob(*this);
126 bool operator==(BranchProbability RHS) const { return N == RHS.N; }
127 bool operator!=(BranchProbability RHS) const { return !(*this == RHS); }
128 bool operator<(BranchProbability RHS) const { return N < RHS.N; }
129 bool operator>(BranchProbability RHS) const { return RHS < *this; }
130 bool operator<=(BranchProbability RHS) const { return !(RHS < *this); }
131 bool operator>=(BranchProbability RHS) const { return !(*this < RHS); }
134 inline raw_ostream &operator<<(raw_ostream &OS, BranchProbability Prob) {
135 return Prob.print(OS);
138 inline BranchProbability operator/(BranchProbability LHS, uint32_t RHS) {
139 return BranchProbability::getRaw(LHS.getNumerator() / RHS);
142 template <class ProbabilityIter>
143 void BranchProbability::normalizeProbabilities(ProbabilityIter Begin,
144 ProbabilityIter End) {
149 for (auto I = Begin; I != End; ++I)
152 for (auto I = Begin; I != End; ++I)
153 I->N = (I->N * uint64_t(D) + Sum / 2) / Sum;
156 template <class WeightListIter>
157 void BranchProbability::normalizeEdgeWeights(WeightListIter Begin,
158 WeightListIter End) {
159 // First we compute the sum with 64-bits of precision.
160 uint64_t Sum = std::accumulate(Begin, End, uint64_t(0));
162 if (Sum > UINT32_MAX) {
163 // Compute the scale necessary to cause the weights to fit, and re-sum with
164 // that scale applied.
165 assert(Sum / UINT32_MAX < UINT32_MAX &&
166 "The sum of weights exceeds UINT32_MAX^2!");
167 uint32_t Scale = Sum / UINT32_MAX + 1;
168 for (auto I = Begin; I != End; ++I)
170 Sum = std::accumulate(Begin, End, uint64_t(0));
173 // Eliminate zero weights.
174 auto ZeroWeightNum = std::count(Begin, End, 0u);
175 if (ZeroWeightNum > 0) {
176 // If all weights are zeros, replace them by 1.
178 std::fill(Begin, End, 1u);
180 // We are converting zeros into ones, and here we need to make sure that
181 // after this the sum won't exceed UINT32_MAX.
182 if (Sum + ZeroWeightNum > UINT32_MAX) {
183 for (auto I = Begin; I != End; ++I)
185 ZeroWeightNum = std::count(Begin, End, 0u);
186 Sum = std::accumulate(Begin, End, uint64_t(0));
188 // Scale up non-zero weights and turn zero weights into ones.
189 uint64_t ScalingFactor = (UINT32_MAX - ZeroWeightNum) / Sum;
190 assert(ScalingFactor >= 1);
191 if (ScalingFactor > 1)
192 for (auto I = Begin; I != End; ++I)
194 std::replace(Begin, End, 0u, 1u);