统计程序两个点之间执行的指令数量

环境:支持perf

ubuntu安装

apt-get install linux-tools-common linux-tools-generic linux-tools-`uname -r`

android

一般自带simpleperf

分析

两个点作差, 求中间结果;

*(int*)nullptr = 0;

案例

断点 1

代码

#define SETPOINT(...)  do { *(int*)nullptr = 0; } while(0)
int main() {SETPOINT(1);int a = 0;for(int i = 0 ; i < 500; i++) a+=i;return 0;
}

执行结果

ch@ch-ubuntu:~/ch/perf_test/exe_count$ sudo perf stat -e instructions:u ./a.out 
[sudo] password for ch: 
./a.out: Segmentation faultPerformance counter stats for './a.out':100064      instructions:u                                              0.111183917 seconds time elapsed0.001217000 seconds user0.000000000 seconds sysch@ch-ubuntu:~/ch/perf_test/exe_count$ sudo perf stat -e instructions:u ./a.out 
./a.out: Segmentation faultPerformance counter stats for './a.out':100064      instructions:u                                              0.104691299 seconds time elapsed0.000875000 seconds user0.000000000 seconds sysch@ch-ubuntu:~/ch/perf_test/exe_count$ sudo perf stat -e instructions:u ./a.out 
./a.out: Segmentation faultPerformance counter stats for './a.out':100064      instructions:u                                              0.111463860 seconds time elapsed0.000931000 seconds user0.000000000 seconds sys

断点 2

代码

#define SETPOINT(...)  do { *(int*)nullptr = 0; } while(0)
int main() {int a = 0;for(int i = 0 ; i < 500; i++) a+=i;SETPOINT(2);return 0;
}

执行结果

ch@ch-ubuntu:~/ch/perf_test/exe_count$ sudo perf stat -e instructions:u ./a.out 
./a.out: Segmentation faultPerformance counter stats for './a.out':102569      instructions:u                                              0.105533002 seconds time elapsed0.000904000 seconds user0.000000000 seconds sysch@ch-ubuntu:~/ch/perf_test/exe_count$ sudo perf stat -e instructions:u ./a.out 
./a.out: Segmentation faultPerformance counter stats for './a.out':102567      instructions:u                                              0.105150980 seconds time elapsed0.000876000 seconds user0.000000000 seconds sysch@ch-ubuntu:~/ch/perf_test/exe_count$ sudo perf stat -e instructions:u ./a.out 
./a.out: Segmentation faultPerformance counter stats for './a.out':102567      instructions:u                                              0.103408851 seconds time elapsed0.000897000 seconds user0.000000000 seconds sys

结果分析

102567 - 100064 == 2503

代码分析

ch@ch-ubuntu:~/ch/perf_test/exe_count$ objdump --disassemble=main ./a.out./a.out:     file format elf64-x86-64Disassembly of section .init:Disassembly of section .plt:Disassembly of section .plt.got:Disassembly of section .text:0000000000001129 <main>:1129:	f3 0f 1e fa          	endbr64112d:	55                   	push   %rbp112e:	48 89 e5             	mov    %rsp,%rbp1131:	c7 45 f8 00 00 00 00 	movl   $0x0,-0x8(%rbp)1138:	c7 45 fc 00 00 00 00 	movl   $0x0,-0x4(%rbp)113f:	eb 0a                	jmp    114b <main+0x22>1141:	8b 45 fc             	mov    -0x4(%rbp),%eax1144:	01 45 f8             	add    %eax,-0x8(%rbp)1147:	83 45 fc 01          	addl   $0x1,-0x4(%rbp)114b:	81 7d fc f3 01 00 00 	cmpl   $0x1f3,-0x4(%rbp)1152:	7e ed                	jle    1141 <main+0x18>1154:	b8 00 00 00 00       	mov    $0x0,%eax1159:	c7 00 00 00 00 00    	movl   $0x0,(%rax)115f:	b8 00 00 00 00       	mov    $0x0,%eax1164:	5d                   	pop    %rbp1165:	c3                   	ret

可以看到1141 - 1152一共有5条指令, 5 * 500差不多2500, perf统计有些许误差属于正常现象;

结论

可以两个点之间的执行共执行了2500条指令; 然后根据指令数量和执行环境推断理论执行性能;
不适用于和多种硬件打交道的环境; 而且不同的syscall耗时不一样;