gcc:去除未使用的功能 [英] gcc: Strip unused functions
问题描述
我注意到有时即使我没有使用 iostream
和相关的I / O库,Mingw生成的二进制文件仍然不合理地大。
$ b
例如,我编写了一个代码来仅使用 vector
和 cstdio
,用 -O2 -flto
编译它,我的程序可以大到2MB!我运行 nm main.exe> e.txt
,并且震惊地看到里面的所有 iostream
相关函数。
在Google搜索之后,我学会了使用 -ffunction-sections -Wl,-gc-sections
,将程序大小从2MB减少到〜300KB(如果使用 -s
,100 + KB)。很好!
为了进一步测试 -ffunction-sections -Wl,-gc-sections
的效果,这里是另一个代码:
#include #include< vector>
#include< tuple>
#include< algorithm>
#include< chrono>
#include< windows.h>
#undef min
struct Point {
int x,y;
};
constexpr int length = 5;
constexpr int half_length(){
返回长度& 1?长度:长度 - 1;
}
模板< class F>
int func_template(F&& f){
#ifdef _MSC_VER
puts(__ FUNCSIG__);
#else
puts(__ PRETTY_FUNCTION__);
#endif
printf(\\\
);
return f();
}
struct fake_func {
int operator()()const {return 59; };
};
模板< class F,class ... Args>
int pass_args(F&& f,Args&& amp; args){
#ifdef _MSC_VER
puts(__ FUNCSIG__);
#else
puts(__ PRETTY_FUNCTION__);
#endif
printf(\\\
);
返回f(std :: forward< Args>(args)...);
}
模板< class T>
T min(T x){
return x;
}
模板< class T,class ... Args>
T min(T x,Args ... args){
T y = min(args ...);
return x < y? x:y;
}
void type_verifier(int x){
printf(%dd,x);
}
void type_verifier(char x){
printf('%c',x);
}
void type_verifier(double x){
printf(%lff,x);
}
模板< class T>
void type_verifier(T x){
printf(unknown);
}
模板< class T,class ... Args>
void type_verifier(T x,Args ... args){
type_verifier(x);
type_verifier(args ...);
}
int bufLen;
char buf [100];
模板< class ...参数>
inline int send(Args ... args){
bufLen = sprintf(buf,std :: forward< Args>(args)...);
返回bufLen;
namespace std {
内联命名空间v1 {
void func(){
printf(我是v1 \\\
);
namespace v2 {
void func(){
printf(我是v2\\\
);
$ b int main(){
std :: vector< int> v {1,2,3,4,5}; (auto& i:v)printf(%d,i)为
;
printf(\\\
);
点p {1,2};
printf(%d%d \\\
,p.x,p.y);
auto t = std :: make_tuple(Hello World,12);
printf(%s%d \ n,std :: get(t),std :: get(t));
int a,b;
auto f = [](){return std :: make_tuple(1,2); };
std :: tie(a,b)= f();
printf(%d%d \ n,a,b);
// int test_constexpr [half_length()+ 4];
int ft = func_template([] {return 42;});
printf(func_template:%d \\\
,ft);
ft = func_template(fake_func {});
printf(func_template:%d \\\
,ft);
ft = pass_args([](int x,int y){return x + y;},152,58);
printf(pass_args:%d \\\
,ft);
ft = pass_args([](int n,const char * m){
for(int i = 0; i< n; i ++)printf(%c,m [i]) ;
printf(\\\
);
return 0;
},5,Hello);
printf(min:%d \\\
,min(3,4,2,1,5));
type_verifier(12,'A',0.5,Hello);
printf(\\\
);
/ * send(Hello World);
send(%d,1);
send(%d,1234);
sprintf(buf,%d,123); * /
std :: func();
std :: v1 :: func();
std :: v2 :: func();
std :: rotate(v.begin(),v.begin()+ 2,v.end()); (auto& i:v)printf(%d,i)为
;
printf(\\\
);
auto start = std :: chrono :: steady_clock :: now();
std :: vector< int> x {2,4,2,0,5,10,7,3,7,1};
printf(插入排序:); (auto& i:x)printf(%d,i)为
;
printf(\\\
);
//插入排序
for(auto i = x.begin(); i!= x.end(); ++ i){
std :: rotate(std :: upper_bound (x.begin(),i,* i),i,i + 1); (auto& j:x)printf(%d,j)为
;
printf(\\\
);
}
std :: vector< int>堆{7,5,3,4,2};
std :: make_heap(heap.begin(),heap.end());
std :: pop_heap(heap.begin(),heap.end());
printf(Pop heap(%d)\\\
,heap.back());
heap.pop_back();
heap.push_back(1);
std :: push_heap(heap.begin(),heap.end());
std :: sort_heap(heap.begin(),heap.end()); (auto& i:heap)printf(%d,i)
;
printf(\\\
);
auto end = std :: chrono :: steady_clock :: now();
auto diff = end - start;
printf(time:%I64d ms\\\
,
std :: chrono :: duration_cast< std :: chrono :: milliseconds>(diff).count());
{
auto u = v;
std :: move_backward(u.begin(),u.begin()+ u.size() - 1,u.begin()+ u.size()); (auto& i:u)printf(%d,i)为
;
printf(\\\
);
}
{
auto u = v;
std :: move(u.begin()+ 1,u.begin()+ u.size(),u.begin()); (auto& i:u)printf(%d,i)为
;
printf(\\\
);
}
start = std :: chrono :: steady_clock :: now();
睡眠(2000);
end = std :: chrono :: steady_clock :: now();
diff = end - start;
printf(time:%I64d ms\\\
,
std :: chrono :: duration_cast< std :: chrono :: milliseconds>(diff).count());
std :: chrono :: steady_clock :: time_point before;
before = std :: chrono :: steady_clock :: now();
睡眠(2000);
auto后= std :: chrono :: steady_clock :: now();
printf(%f seconds\\\
,std :: chrono :: duration< double>(after - before).count());
返回0;
}
令我失望的是,最终的程序再次大于2MB。
有趣的是, cl.exe
谨慎地移除所有 iostream
相关函数即使我没有使用 / O2
或任何其他标志,只是 cl.exe main.cpp
。 (对于上面的代码, cl.exe
会生成100 + KB二进制文件)。
我是否错过任何其他有用的gcc的标志是什么?
规格:
与Linux比较
上面的代码(除了 windows.h
和),gcc 4.9.2(Linux)和gcc 4.9.3(mingw-w64)睡眠
已被删除)。
$ b
编译标记
g ++ -o c ++ 11 c ++ 11.cpp -std = c ++ 11 -static-libgcc -static-libstdc ++ -ffunction-sections -Wl,-gc-sections -O2
Linux gcc成功剥离了 iostream
和函数,而不需要 -flto
而Mingw-w64 gcc无法正常运行。
Windows只支持PE格式,而Linux支持端口ELF格式,允许Linux使用Gold连接器。也许这是解释?
更新
我最终在 https://sourceforge.net/p/mingw-w64/bugs/578/ 。让我们希望得到一些关注!
尝试通过 -Wl从静态libstdc ++剥离调试和符号信息, - 带状所有
。这使我的可执行文件在Cygwin(13x)上从9M减少到了670K,在Ubuntu(80x)上从6M减少到了80K。
I noticed that sometimes even if I don't use iostream
and related I/O libraries, my binaries produced by Mingw were still unreasonably large.
For example, I wrote a code to use vector
and cstdio
only and compiled it with -O2 -flto
, my program can go as large as 2MB! I run nm main.exe > e.txt
and was shocked to see all the iostream
related functions in it.
After some googling, I learnt to use -ffunction-sections -Wl,-gc-sections
, that reduces the program size from 2MB to ~300KB (if with -s
, 100+KB). Excellent!
To further test the effect of -ffunction-sections -Wl,-gc-sections
, here is another code:
#include <cstdio>
#include <vector>
#include <tuple>
#include <algorithm>
#include <chrono>
#include <windows.h>
#undef min
struct Point {
int x, y;
};
constexpr int length = 5;
constexpr int half_length() {
return length & 1 ? length : length - 1;
}
template<class F>
int func_template(F&& f) {
#ifdef _MSC_VER
puts(__FUNCSIG__);
#else
puts(__PRETTY_FUNCTION__);
#endif
printf("\n");
return f();
}
struct fake_func {
int operator()() const { return 59; };
};
template<class F, class... Args>
int pass_args(F&& f, Args&&... args) {
#ifdef _MSC_VER
puts(__FUNCSIG__);
#else
puts(__PRETTY_FUNCTION__);
#endif
printf("\n");
return f(std::forward<Args>(args)...);
}
template<class T>
T min(T x) {
return x;
}
template<class T, class... Args>
T min(T x, Args... args) {
T y = min(args...);
return x < y ? x : y;
}
void type_verifier(int x) {
printf("%dd ", x);
}
void type_verifier(char x) {
printf("'%c' ", x);
}
void type_verifier(double x) {
printf("%lff ", x);
}
template<class T>
void type_verifier(T x) {
printf("unknown ");
}
template<class T, class... Args>
void type_verifier(T x, Args... args) {
type_verifier(x);
type_verifier(args...);
}
int bufLen;
char buf[100];
template<class... Args>
inline int send(Args... args) {
bufLen = sprintf(buf, std::forward<Args>(args)...);
return bufLen;
}
namespace std {
inline namespace v1 {
void func() {
printf("I am v1\n");
}
}
namespace v2 {
void func() {
printf("I am v2\n");
}
}
}
int main() {
std::vector<int> v {1, 2, 3, 4, 5};
for (auto &i : v) printf("%d ", i);
printf("\n");
Point p {1, 2};
printf("%d %d\n", p.x, p.y);
auto t = std::make_tuple("Hello World", 12);
printf("%s %d\n", std::get<0>(t), std::get<1>(t));
int a, b;
auto f = []() { return std::make_tuple(1, 2); };
std::tie(a, b) = f();
printf("%d %d\n", a, b);
//int test_constexpr[half_length() + 4];
int ft = func_template([]{ return 42; });
printf("func_template: %d\n", ft);
ft = func_template(fake_func {});
printf("func_template: %d\n", ft);
ft = pass_args([](int x, int y) { return x + y; }, 152, 58);
printf("pass_args: %d\n", ft);
ft = pass_args([](int n, const char *m) {
for (int i = 0; i < n; i++) printf("%c ", m[i]);
printf("\n");
return 0;
}, 5, "Hello");
printf("min: %d\n", min(3, 4, 2, 1, 5));
type_verifier(12, 'A', 0.5, "Hello");
printf("\n");
/* send("Hello World");
send("%d", 1);
send("%d", "1234");
sprintf(buf, "%d", "123");*/
std::func();
std::v1::func();
std::v2::func();
std::rotate(v.begin(), v.begin() + 2, v.end());
for (auto &i : v) printf("%d ", i);
printf("\n");
auto start = std::chrono::steady_clock::now();
std::vector<int> x {2, 4, 2, 0, 5, 10, 7, 3, 7, 1};
printf("insertion sort: ");
for (auto &i: x) printf("%d ", i);
printf("\n");
// insertion sort
for (auto i = x.begin(); i != x.end(); ++i) {
std::rotate(std::upper_bound(x.begin(), i, *i), i, i+1);
for (auto &j: x) printf("%d ", j);
printf("\n");
}
std::vector<int> heap {7, 5, 3, 4, 2};
std::make_heap(heap.begin(), heap.end());
std::pop_heap(heap.begin(), heap.end());
printf("Pop heap (%d)\n", heap.back());
heap.pop_back();
heap.push_back(1);
std::push_heap(heap.begin(), heap.end());
std::sort_heap(heap.begin(), heap.end());
for (auto &i: heap) printf("%d ", i);
printf("\n");
auto end = std::chrono::steady_clock::now();
auto diff = end - start;
printf("time: %I64d ms\n",
std::chrono::duration_cast<std::chrono::milliseconds>(diff).count());
{
auto u = v;
std::move_backward(u.begin(), u.begin() + u.size() - 1, u.begin() + u.size());
for (auto &i : u) printf("%d ", i);
printf("\n");
}
{
auto u = v;
std::move(u.begin() + 1, u.begin() + u.size(), u.begin());
for (auto &i : u) printf("%d ", i);
printf("\n");
}
start = std::chrono::steady_clock::now();
Sleep(2000);
end = std::chrono::steady_clock::now();
diff = end - start;
printf("time: %I64d ms\n",
std::chrono::duration_cast<std::chrono::milliseconds>(diff).count());
std::chrono::steady_clock::time_point before;
before = std::chrono::steady_clock::now();
Sleep(2000);
auto after = std::chrono::steady_clock::now();
printf("%f seconds\n", std::chrono::duration<double>(after - before).count());
return 0;
}
To my disappointment, the final program is once again > 2MB.
Interestingly, cl.exe
thoughtfully remove all iostream
related functions consistently even if I didn't use /O2
or any other flags, just cl.exe main.cpp
. (For the code above, cl.exe
produces 100+KB binary).
Did I miss any other useful gcc flags for this?
Specification:
- Mingw-w64 gcc 6.1.0
- Mingw-w64 gcc 6.2.0
- Visual Studio 2017 RC
- All binaries are linked statically
Compare with Linux
I compared the binaries produced by gcc 4.9.2 (Linux) and gcc 4.9.3 (mingw-w64) for the above code (except windows.h
and Sleep
were removed).
Compile flag
g++ -o c++11 c++11.cpp -std=c++11 -static-libgcc -static-libstdc++ -ffunction-sections -Wl,-gc-sections -O2
Linux gcc did successfully strip away iostream
and functions without the need for -flto
while Mingw-w64 gcc just can't do it properly.
Windows only support PE format while Linux supports ELF format, allowing Linux to use Gold linker. Maybe this is the explanation?
Update
I eventually filed a bug at https://sourceforge.net/p/mingw-w64/bugs/578/ . Let's hope it gets some attentions!
Try stripping debug and symbol info from static libstdc++ via -Wl,--strip-all
. This reduced my executable from 9M to 670K on Cygwin (13x) and from 6M to 80K on Ubuntu (80x).
这篇关于gcc:去除未使用的功能的文章就介绍到这了,希望我们推荐的答案对大家有所帮助,也希望大家多多支持IT屋!