C++ attribute: likely, unlikely (since C++20)

From cppreference.com
< cpp‎ | language‎ | attributes
 
 
C++ language
General topics
Flow control
Conditional execution statements
if
Iteration statements (loops)
for
range-for (C++11)
Jump statements
Functions
Function declaration
Lambda function expression
inline specifier
Dynamic exception specifications (until C++20)
noexcept specifier (C++11)
Exceptions
Namespaces
Types
Specifiers
decltype (C++11)
auto (C++11)
alignas (C++11)
Storage duration specifiers
Initialization
Expressions
Alternative representations
Literals
Boolean - Integer - Floating-point
Character - String - nullptr (C++11)
User-defined (C++11)
Utilities
Attributes (C++11)
Types
typedef declaration
Type alias declaration (C++11)
Casts
Implicit conversions - Explicit conversions
static_cast - dynamic_cast
const_cast - reinterpret_cast
Memory allocation
Classes
Class-specific function properties
explicit (C++11)
static
Special member functions
Templates
Miscellaneous
 
 
Attributes
(C++14)
(C++17)
likelyunlikely
(C++20)(C++20)
 

Allow the compiler to optimize for the case where paths of execution including that statement are more or less likely than any alternative path of execution that does not include such a statement

Syntax

[[likely]] (1)
[[unlikely]] (2)

Explanation

These attributes may be applied to labels and statements (other than declaration-statements). They may not be simultaneously applied to the same label or statement.

1) Applies to a statement to allow the compiler to optimize for the case where paths of execution including that statement are more likely than any alternative path of execution that does not include such a statement.
2) Applies to a statement to allow the compiler to optimize for the case where paths of execution including that statement are less likely than any alternative path of execution that does not include such a statement.

A path of execution is deemed to include a label if and only if it contains a jump to that label:

int f(int i)
{
    switch(i)
    {
        case 1: [[fallthrough]];
        [[likely]] case 2: return 1;
    }
    return 2;
}

i == 2 is considered more likely than any other value of i, but the [[likely]] has no effect on the i == 1 case even though it falls through the case 2: label.

Example

#include <chrono>
#include <cmath>
#include <iomanip>
#include <iostream>
#include <random>
 
namespace with_attributes {
constexpr double pow(double x, long long n) noexcept {
    if (n > 0) [[likely]]
        return x * pow(x, n - 1);
    else [[unlikely]]
        return 1;
}
constexpr long long fact(long long n) noexcept {
    if (n > 1) [[likely]]
        return n * fact(n - 1);
    else [[unlikely]]
        return 1;
}
constexpr double cos(double x) noexcept {
    constexpr long long precision{16LL};
    double y{};
    for (auto n{0LL}; n < precision; n += 2LL) [[likely]]
        y += pow(x, n) / (n & 2LL ? -fact(n) : fact(n));
    return y;
}
}  // namespace with_attributes
 
namespace no_attributes {
constexpr double pow(double x, long long n) noexcept {
    if (n > 0)
        return x * pow(x, n - 1);
    else
        return 1;
}
constexpr long long fact(long long n) noexcept {
    if (n > 1)
        return n * fact(n - 1);
    else
        return 1;
}
constexpr double cos(double x) noexcept {
    constexpr long long precision{16LL};
    double y{};
    for (auto n{0LL}; n < precision; n += 2LL)
        y += pow(x, n) / (n & 2LL ? -fact(n) : fact(n));
    return y;
}
}  // namespace no_attributes
 
double gen_random() noexcept {
    static std::random_device rd;
    static std::mt19937 gen(rd());
    static std::uniform_real_distribution<double> dis(-1.0, 1.0);
    return dis(gen);
}
 
volatile double sink{}; // ensures a side effect
 
int main() {
    for (const auto x : {0.125, 0.25, 0.5, 1. / (1 << 26)}) {
        std::cout
            << std::setprecision(53)
            << "x = " << x << '\n'
            << std::cos(x) << '\n'
            << with_attributes::cos(x) << '\n'
            << (std::cos(x) == with_attributes::cos(x) ? "equal" : "differ") << '\n';
    }
 
    auto benchmark = [](auto fun, auto rem) {
        const auto start = std::chrono::high_resolution_clock::now();
        for (auto size{1ULL}; size != 10'000'000ULL; ++size) {
            sink = fun(gen_random());
        }
        const std::chrono::duration<double> diff =
            std::chrono::high_resolution_clock::now() - start;
        std::cout << "Time: " << std::fixed << std::setprecision(6) << diff.count()
                  << " sec " << rem << std::endl; 
    };
 
    benchmark(with_attributes::cos, "(with attributes)");
    benchmark(no_attributes::cos, "(without attributes)");
    benchmark([](double t) { return std::cos(t); }, "(std::cos)");
}

Possible output:

x = 0.125
0.99219766722932900560039115589461289346218109130859375
0.99219766722932900560039115589461289346218109130859375
equal
x = 0.25
0.96891242171064473343022882545483298599720001220703125
0.96891242171064473343022882545483298599720001220703125
equal
x = 0.5
0.8775825618903727587394314468838274478912353515625
0.8775825618903727587394314468838274478912353515625
equal
x = 1.490116119384765625e-08
0.99999999999999988897769753748434595763683319091796875
0.99999999999999988897769753748434595763683319091796875
equal
Time: 0.579122 sec (with attributes)
Time: 0.722553 sec (without attributes)
Time: 0.425963 sec (std::cos)