measure_functor_call.cpp
cpp
#include <iostream>
#include <cctype> // std::isspace
#include <functional>
#include "profile.h"
/* test1--------begin------- */
struct SpaceChecker{
bool isSpace(char c){
return std::isspace(c);
}
};
PROFILE_REGISTER(countSpace_v1_inline)
size_t countSpace_v1(const std::string& str){
PROFILE_CHECK(countSpace_v1_inline);
SpaceChecker space_checker;
size_t count{0};
for (const auto&it: str) {
if (space_checker.isSpace(it))
{
++count;
}
}
return count;
}
/* test1--------end------- */
/* test2--------begin------- */
struct SpaceCheckerNoInline{
bool isSpace(char c);
};
PROFILE_NOINLINE bool SpaceCheckerNoInline::isSpace(char c) {
return std::isspace(c);
}
PROFILE_REGISTER(countSpace_v2_noinline)
size_t countSpace_v2(const std::string& str){
PROFILE_CHECK(countSpace_v2_noinline);
SpaceCheckerNoInline space_checker_noinline;
size_t count{0};
for (const auto&it: str) {
if (space_checker_noinline.isSpace(it))
{
++count;
}
}
return count;
}
/* test2--------end------- */
/* test3--------begin------- */
class SpaceCheckerFather{
public:
virtual ~SpaceCheckerFather(){}
virtual bool isSpace(char c) = 0;
};
// The purpose of this class is prevent GCC devirtualization optimization
// 如果只有一个子类编译器可能会做devirtualization的动作,即去除虚函数,这样就达不到测试虚函数调用测试的目的了,因此dummy类是为了防止devirtualization用的
class dummy: public SpaceCheckerFather{
public:
virtual bool isSpace(char) override{
return false;
}
};
class SpaceCheckerSon: public SpaceCheckerFather{
public:
virtual bool isSpace(char c) override;
};
bool SpaceCheckerSon::isSpace(char c){
return std::isspace(c);
};
PROFILE_REGISTER(countSpace_v3_virtual_call)
size_t countSpace_v3(const std::string& str){
PROFILE_CHECK(countSpace_v3_virtual_call);
SpaceCheckerSon space_checker_son;
size_t count{0};
for (const auto&it: str) {
if (space_checker_son.isSpace(it))
{
++count;
}
}
return count;
}
/* test3--------end------- */
/* test4--------begin------- */
std::function<bool(char)> std_func = [](char c)->bool {
return std::isspace(c);
};
PROFILE_REGISTER(countSpace_v4_std_function)
size_t countSpace_v4(const std::string& str){
PROFILE_CHECK(countSpace_v4_std_function);
size_t count{0};
for (const auto&it: str) {
if (std_func(it))
{
++count;
}
}
return count;
}
/* test4--------end------- */
auto main()->int{
constexpr int kLoops = 10000000;
std::string contents("Welcome to HangZhou, ZheJiang Province.");
size_t count1{0};
for (int i = 0; i < kLoops; ++i) {
count1 = countSpace_v1(contents);
}
std::cout << "countSpace_v1's counter is " << count1 << std::endl;
size_t count2{0};
for (int i = 0; i < kLoops; ++i) {
count2 = countSpace_v2(contents);
}
std::cout << "countSpace_v2's counter is " << count2 << std::endl;
size_t count3{0};
for (int i = 0; i < kLoops; ++i) {
count3 = countSpace_v3(contents);
}
std::cout << "countSpace_v3's counter is " << count3 << std::endl;
size_t count4{0};
for (int i = 0; i < kLoops; ++i) {
count4 = countSpace_v4(contents);
}
std::cout << "countSpace_v4's counter is " << count4 << std::endl;
}
profile.h
cpp
#ifndef PROFILER_PROFILE_H
#define PROFILER_PROFILE_H
#include <iostream>
#include <vector>
#include <set>
#include <string>
#include <string_view>
#include <format>
#include <assert.h>
#include "profile_rdtsc.h"
#define PF_DEBUG // open performance test or not
namespace pf {
struct ProfileInfo {
int id_;
int call_count_;
uint64_t call_duration_;
std::string func_name_;
};
class Profile {
private:
Profile() {
#ifdef PF_DEBUG
constexpr int kTimes = 16;
for (int i = 0; i < kTimes; ++i) {
uint64_t begin = rdtsc_benchmark_begin();
uint64_t end = rdtsc_benchmark_end();
if (overhead_ > end - begin) {
overhead_ = end - begin; // maybe not the smallest
}
}
#endif
//std::cout << "Profile ctor called.\n";
}
~Profile() {
#ifdef PF_DEBUG
std::cout << "---> Performance details begin: <---\n";
std::cout << "Attention, overhead is " << overhead_
<< " cycles, and it is removed from average duration.\n";
for (const auto &item: info_) {
if (!item.call_count_) {
continue;
}
double average = static_cast<double>(item.call_duration_) / static_cast<double>(item.call_count_) -
static_cast<double>(overhead_);
if (average < 0) {
average = 0.0; // time cost is smaller than overhead_, because the overhead_ is not accuracy.
}
std::cout << "[" << item.func_name_ << ", called " << item.call_count_ << " times, total duration cycle is "
<< item.call_duration_ << ", average duration cycle is " << average << ".]" << std::endl;
}
std::cout << "---> Performance details end. <---\n";
#endif
}
Profile(const Profile &) = delete;
Profile &operator=(const Profile &) = delete;
public:
static Profile &getInstance();
void addInfo(int id, uint64_t duration) {
assert(id >= 0 && id < static_cast<int>(info_.size()));
++info_[id].call_count_;
info_[id].call_duration_ += duration;
}
int registerFunc(const std::string_view func_name) {
auto [iter, success] = func_name_.insert(std::string(func_name));
assert(success && "One function can only be registered once!");
int func_id = static_cast<int>(info_.size());
info_.push_back({func_id, 0, 0, *iter});
return func_id;
}
private:
std::set<std::string> func_name_;
std::vector<ProfileInfo> info_;
uint64_t overhead_{~0UL};
};
class ProfilingChecker {
public:
ProfilingChecker(int id) : id_(id), start_time_(rdtsc_benchmark_begin()) {
//std::cout << "ProfilingChecker ctor called.\n";
}
~ProfilingChecker() {
uint64_t end_time = rdtsc_benchmark_end();
uint64_t duration = end_time - start_time_;
Profile::getInstance().addInfo(id_, duration);
//std::cout << "ProfilingChecker dtor called.\n";
}
ProfilingChecker(const ProfilingChecker &) = delete;
ProfilingChecker &operator=(const ProfilingChecker &) = delete;
private:
int id_;
uint64_t start_time_;
};
int doProfileRegister(const std::string_view func_name);
#ifdef PF_DEBUG
#define PROFILE_REGISTER(func) static const int _pf_id_##func = pf::doProfileRegister(#func);
#define PROFILE_CHECK(func) pf::ProfilingChecker _checker(_pf_id_##func);
#else
#define PROFILE_REGISTER(func)
#define PROFILE_CHECK(func)
#endif
#ifdef _MSC_VER
#define PROFILE_NOINLINE __declspec(noinline)
#else
#define PROFILE_NOINLINE __attribute__((noinline))
#endif
}
#endif //PROFILER_PROFILE_H
profile.cpp
cpp
#include "profile.h"
#include <iostream>
namespace pf {
Profile &Profile::getInstance() {
static Profile instance;
return instance;
}
int doProfileRegister(const std::string_view func_name){
return Profile::getInstance().registerFunc(func_name);
}
}
profile_rdtsc.h
cpp
#ifndef PROFILE_PROFILE_RDTSC_H
#define PROFILE_PROFILE_RDTSC_H
#include <stdint.h> // uint64_t
#ifndef HAS_HW_RDTSC
#if defined(_M_X64) || defined(_M_IX86) || defined(__x86_64) || defined(__i386)
#define HAS_HW_RDTSC 1
#else
#define HAS_HW_RDTSC 0
#endif
#endif
#if HAS_HW_RDTSC
#ifdef _WIN32
#include <intrin.h> // __rdtsc/_mm_lfence/_mm_mfence
#elif __has_include(<x86intrin.h>)
#include <x86intrin.h> // __rdtsc/_mm_lfence/_mm_mfence
#endif
#else
#include <chrono> // std::chrono::steady_clock/nanoseconds
#endif
// Macro to forbid the compiler from reordering instructions
#ifdef _MSC_VER
#define RDTSC_MEM_BARRIER() _ReadWriteBarrier()
#else
#define RDTSC_MEM_BARRIER() __asm__ __volatile__("" : : : "memory")
#endif
inline uint64_t rdtsc()
{
RDTSC_MEM_BARRIER();
#if HAS_HW_RDTSC
uint64_t result = __rdtsc();
#else
uint64_t result = std::chrono::steady_clock::now().time_since_epoch() /
std::chrono::nanoseconds(1);
//uint64_t result = std::chrono::high_resolution_clock::now().time_since_epoch() /
// std::chrono::nanoseconds(1);
#endif
RDTSC_MEM_BARRIER();
return result;
}
#if HAS_HW_RDTSC
inline uint64_t rdtsc_benchmark_begin()
{
// memory fence, according to x86 architecture,保障开始测试之前让CPU将之前未执行完的指令执行完
_mm_mfence();
_mm_lfence();
uint64_t result = __rdtsc();
RDTSC_MEM_BARRIER();
return result;
}
inline uint64_t rdtsc_benchmark_end()
{
_mm_lfence();
uint64_t result = __rdtsc();
RDTSC_MEM_BARRIER();
return result;
}
#else
inline uint64_t rdtsc_benchmark_begin()
{
return rdtsc();
}
inline uint64_t rdtsc_benchmark_end()
{
return rdtsc();
}
#endif
#endif //PROFILE_PROFILE_RDTSC_H
输出结果如下:
由此可以看出,普通函数和虚函数开销差异本身并不大,内联可以降低开销。