为什么关了同步的cin比快读还快 - 矿工茶馆

Doomsday_Envoy

MoonCakeMC 发表于 2023-3-18 22:01
本来想考虑一下的
但是最近环境不太方便反汇编（）
而且devcpp编译出来的好像不是很适合反汇编（ ...

听你描述像是没系统性学过汇编，但即便精通也未必能回答这个问题，因为这是工程的而非理论的。从高级语言到汇编在逻辑上并不复杂，但实际使用需求以及版本更新迭代之后，这个过程的实现就很复杂多样了。
你应该听说过，编译器的功能五花八门；对于某些编程语言的比较有时候也涉及对其编译器的比较，这就是问题所在了。但我只是听过几个案例，不是专攻这些的，油管搬运的视频好像有不少是这种听上去无厘头问题的技术性解释，去搜搜看吧。

MoonCakeMC

Doomsday_Envoy 发表于 2023-3-18 22:14
听你描述像是没系统性学过汇编，但即便精通也未必能回答这个问题，因为这是工程的而非理论的。从高级语言 ...

感谢大佬指路

确实没系统学过汇编，或者说根本没学过（）
只不过是之前做逆向功能的时候散着看了看

编译器的问题倒是知道
但是还没测试

待会补测一下试试

_tms

之前试过，的确是这样，感觉很神奇（
~~所以我现在加sync_with_stdio已经成习惯了（~~

极光creeper

scanf printf 是 c 的做法，而 cpp 有自己的 cin 和 cout，为了兼容这两种做法，必须同步输入输出流，这是为了为了线程安全，因此你去同步以后，输入、输出流就不同步了，也就无法同时使用了。

可以这么理解，为了线程安全，于是同步，于是要付出性能代价。

如果你一定要问为什么解绑后 cin cout 更快，这个问题非常难以回答，太底层了，你可以这么想，毕竟 c++ 比 c 多两个加号。

普通快读其实没快多少，实际上那种手写缓冲区的快读才是比较快的，但是仅限于开 O2 的情况下
这里涉及到一些内存访问，几级cpu缓存的问题，以及 fread 的优势，实际上不开 O2 的话，后者优势不是那么大

开心的阿诺

以前我也做过类似的实验，数据如下：

//count.cpp
#include <iostream>
#include <chrono>
using namespace std;
int num;
long long sum = 0;
int main(int argc, char const *argv[])
{
FILE *fp = fopen("num.txt", "r");
fscanf(fp, "%d", &num);
fclose(fp);
string f = argv[1];
int n = atoi(argv[2]);
freopen((f + ".txt").c_str(), "w", stdout);
cout << f << ":" << endl;
#ifdef __linux__
f = "./" + f;
#endif
cout << "num=" << num << endl;
cout << "n=" << n << endl;
cout << "val={";
for (int i = 1; i <= n; i++)
{
auto timeStart = chrono::duration_cast<chrono::milliseconds>(chrono::system_clock::now().time_since_epoch());
system((f + " > a.txt").c_str());
auto timeEnd = chrono::duration_cast<chrono::milliseconds>(chrono::system_clock::now().time_since_epoch());
long long time = timeEnd.count() - timeStart.count();
cout << time << "," << flush;
sum += time;
}
cout << "}" << endl;
cout << "avg=" << (double)sum / n << endl;
return 0;
}

复制代码

//printf.c
#include <stdio.h>
int num;
int main()
{
FILE *fp = fopen("num.txt", "r");
fscanf(fp, "%d", &num);
fclose(fp);
for (int i = 1; i <= num; i++)
printf("a");
return 0;
}

复制代码

printf.exe:
num=10000000
n=100
val={1241,1275,1035,1051,1062,1029,1020,1032,1016,1015,1032,1126,1092,1050,1022,1037,1025,1023,1040,1022,1030,1020,1027,1038,1025,1023,1019,1012,1023,1018,1025,1065,1063,1025,1026,1067,1053,1026,1023,1020,1009,1023,1025,1090,1057,1036,1018,1024,1021,1020,1020,1028,1031,1026,1058,1033,1026,1022,1029,1031,1016,1080,1021,1057,1022,1035,1034,1023,1025,1025,1016,1037,1021,1025,1024,1023,1026,1028,1034,1065,1013,1060,1023,1026,1012,1014,1024,1033,1024,1027,1022,1039,1019,1025,1019,1026,1026,1028,1027,1037,}
avg=1036.61

复制代码

//printf+setvbuf_8192.c
#include <stdio.h>
#define buffer 8192
int num;
char buf[buffer];
int main()
{
FILE *fp = fopen("num.txt", "r");
fscanf(fp, "%d", &num);
fclose(fp);
setvbuf(stdout, buf, _IOFBF, buffer);
for (int i = 1; i <= num; i++)
printf("a");
return 0;
}

复制代码

printf+setvbuf_8192.exe:
num=10000000
n=100
val={1225,1016,991,1017,1011,1019,1001,1014,1000,1013,1006,1022,1003,1004,1016,1018,1001,1004,1006,1027,1011,1111,1083,1014,1007,1012,999,1015,1002,1013,1011,1008,1001,1008,1001,1009,1005,1045,1037,1009,1013,1089,1026,1008,1007,1006,1009,1006,1000,1005,1002,995,1001,1006,994,1005,1005,1010,1018,1043,1008,1000,1062,1000,991,1008,1014,1002,990,1008,1012,994,1008,993,1005,1001,1002,1005,999,1006,1015,1002,1003,996,1003,999,1016,1000,1003,1043,1007,1000,1016,1002,997,996,1003,1008,1000,1056,}
avg=1013.36

复制代码

//printf+flush.c
#include <stdio.h>
int num;
int main()
{
FILE *fp = fopen("num.txt", "r");
fscanf(fp, "%d", &num);
fclose(fp);
for (int i = 1; i <= num; i++)
printf("a"), fflush(stdout);
return 0;
}

复制代码

printf+flush.exe:
num=10000000
n=100
val={23638,23985,24179,25381,24482,25112,23435,23416,23257,23877,23874,23390,23270,23279,23699,24088,23535,23640,23438,23446,23364,23216,23329,23460,23703,23770,

复制代码

//cout.cpp
#include <iostream>
using namespace std;
int num;
int main()
{
FILE *fp = fopen("num.txt", "r");
fscanf(fp, "%d", &num);
fclose(fp);
for (int i = 1; i <= num; i++)
cout << 'a';
return 0;
}

复制代码

cout.exe:
num=10000000
n=100
val={512,507,497,509,502,496,499,501,499,500,525,528,516,510,530,849,515,557,532,588,541,550,520,526,679,517,502,511,532,590,515,641,583,591,570,567,588,512,519,525,535,686,507,700,513,506,527,504,553,504,529,530,521,511,522,513,521,513,506,514,529,510,506,502,589,676,501,491,503,506,494,515,560,511,509,504,499,612,513,738,502,514,512,517,511,493,503,503,510,507,507,555,539,497,509,504,502,518,651,499,}
avg=535.87

复制代码

//cout+no_sync.cpp
#include <iostream>
using namespace std;
int num;
int main()
{
FILE *fp = fopen("num.txt", "r");
fscanf(fp, "%d", &num);
fclose(fp);
ios::sync_with_stdio(0);
for (int i = 1; i <= num; i++)
cout << 'a';
return 0;
}

复制代码

cout+no_sync.exe:
num=10000000
n=100
val={277,172,165,187,196,194,196,171,169,172,180,167,190,171,169,172,169,167,227,198,172,173,173,206,194,282,179,165,169,174,169,228,173,214,240,199,168,167,170,205,173,167,184,196,401,176,274,226,212,167,172,170,170,168,184,252,173,175,169,176,170,170,172,167,170,173,165,168,171,208,169,169,168,173,168,176,165,168,163,179,168,171,183,168,166,178,166,176,175,166,166,181,177,168,172,167,178,169,165,171,}
avg=183.77

复制代码

下面开始猜测，未经证实，谨慎观看，错了打我 qwq
scanf和printf是基于格式化字符串的，而cin和cout估计是基于C++的一些特性，编译成直接的函数调用，少了解析字符串的过程

inuEbisu

要快还是得手写缓冲区

MoonCakeMC

inuEbisu 发表于 2023-3-18 23:43
要快还是得手写缓冲区

确实是（）
但是感觉打OI的时候手写缓冲区没啥必要

（应该不会有题还卡快读吧……

）