在C代码中将结构体变量作为参数传递效率忒低
在C语言编程中，我们几乎不可能看见有人将一个结构体变量作为参数进行传递，因为效率太低了。本文尝试从反汇编的角度给出其中的缘由。
对于C语言来说，所有的参数传递都是值传递。如果一个变量为指针，那么传递的就是指针变量的值(即某个内存地址)。
那么，如果一个参数是结构体变量（包括多个成员），怎么从caller传递到callee呢？
先看下面的代码片段：
o foo1.c
1 #define FALSE 0 2 #define TRUE (!0) 3 4 typedef struct point_s { 5 int x; 6 int y; 7 int z; 8 } point_t; 9 10 static int cmp(point_t a, point_t b) 11 { 12 if (a.x != b.x) 13 return FALSE; 14 if (a.y != b.y) 15 return FALSE; 16 if (a.z != b.z) 17 return FALSE; 18 return TRUE; 19 } 20 21 int main(int argc, char *argv[]) 22 { 23 point_t a = { .x = 0x1, .y = 0x2, .z = +0x3 }; 24 point_t b = { .x = 0x1, .y = 0x2, .z = -0x3 }; 25 return !cmp(a, b); 26 }
o 对foo1.c进行编译后反汇编
1 $ gcc -g -Wall -m32 -std=gnu99 -o foo1 foo1.c 2 $ gdb foo1 3 (gdb) set disassembly-flavor intel 4 (gdb) disas /m main 5 Dump of assembler code for function main: 6 22 { 7 0x0804842a <+0>: push ebp 8 0x0804842b <+1>: mov ebp,esp 9 0x0804842d <+3>: sub esp,0x38 10 11 23 point_t a = { .x = 0x1, .y = 0x2, .z = +0x3 }; 12 0x08048430 <+6>: mov DWORD PTR [ebp-0x18],0x1 13 0x08048437 <+13>: mov DWORD PTR [ebp-0x14],0x2 14 0x0804843e <+20>: mov DWORD PTR [ebp-0x10],0x3 15 16 24 point_t b = { .x = 0x1, .y = 0x2, .z = -0x3 }; 17 0x08048445 <+27>: mov DWORD PTR [ebp-0xc],0x1 18 0x0804844c <+34>: mov DWORD PTR [ebp-0x8],0x2 19 0x08048453 <+41>: mov DWORD PTR [ebp-0x4],0xfffffffd 20 21 25 return !cmp(a, b); 22 0x0804845a <+48>: mov eax,DWORD PTR [ebp-0xc] 23 0x0804845d <+51>: mov DWORD PTR [esp+0xc],eax 24 0x08048461 <+55>: mov eax,DWORD PTR [ebp-0x8] 25 0x08048464 <+58>: mov DWORD PTR [esp+0x10],eax 26 0x08048468 <+62>: mov eax,DWORD PTR [ebp-0x4] 27 0x0804846b <+65>: mov DWORD PTR [esp+0x14],eax 28 0x0804846f <+69>: mov eax,DWORD PTR [ebp-0x18] 29 0x08048472 <+72>: mov DWORD PTR [esp],eax 30 0x08048475 <+75>: mov eax,DWORD PTR [ebp-0x14] 31 0x08048478 <+78>: mov DWORD PTR [esp+0x4],eax 32 0x0804847c <+82>: mov eax,DWORD PTR [ebp-0x10] 33 0x0804847f <+85>: mov DWORD PTR [esp+0x8],eax 34 0x08048483 <+89>: call 0x80483ed <cmp> 35 0x08048488 <+94>: test eax,eax 36 0x0804848a <+96>: sete al 37 0x0804848d <+99>: movzx eax,al 38 39 26 } 40 0x08048490 <+102>: leave 41 0x08048491 <+103>: ret 42 43 End of assembler dump. 44 (gdb) disas /m cmp 45 Dump of assembler code for function cmp: 46 11 { 47 0x080483ed <+0>: push ebp 48 0x080483ee <+1>: mov ebp,esp 49 50 12 if (a.x != b.x) 51 0x080483f0 <+3>: mov edx,DWORD PTR [ebp+0x8] 52 0x080483f3 <+6>: mov eax,DWORD PTR [ebp+0x14] 53 0x080483f6 <+9>: cmp edx,eax 54 0x080483f8 <+11>: je 0x8048401 <cmp+20> 55 56 13 return FALSE; 57 0x080483fa <+13>: mov eax,0x0 58 0x080483ff <+18>: jmp 0x8048428 <cmp+59> 59 60 14 if (a.y != b.y) 61 0x08048401 <+20>: mov edx,DWORD PTR [ebp+0xc] 62 0x08048404 <+23>: mov eax,DWORD PTR [ebp+0x18] 63 0x08048407 <+26>: cmp edx,eax 64 0x08048409 <+28>: je 0x8048412 <cmp+37> 65 66 15 return FALSE; 67 0x0804840b <+30>: mov eax,0x0 68 0x08048410 <+35>: jmp 0x8048428 <cmp+59> 69 70 16 if (a.z != b.z) 71 0x08048412 <+37>: mov edx,DWORD PTR [ebp+0x10] 72 0x08048415 <+40>: mov eax,DWORD PTR [ebp+0x1c] 73 0x08048418 <+43>: cmp edx,eax 74 0x0804841a <+45>: je 0x8048423 <cmp+54> 75 76 17 return FALSE; 77 0x0804841c <+47>: mov eax,0x0 78 0x08048421 <+52>: jmp 0x8048428 <cmp+59> 79 80 18 return TRUE; 81 0x08048423 <+54>: mov eax,0x1 82 83 19 } 84 0x08048428 <+59>: pop ebp 85 0x08048429 <+60>: ret 86 87 End of assembler dump. 88 (gdb)
o caller: point_t b的所有成员x, y, z和point_t a的所有成员x, y, z被依次存入到stack上
23 point_t a = { .x = 0x1, .y = 0x2, .z = +0x3 }; 0x08048430 <+6>: mov DWORD PTR [ebp-0x18],0x1 ; a.x = 0x1 0x08048437 <+13>: mov DWORD PTR [ebp-0x14],0x2 ; a.y = 0x2 0x0804843e <+20>: mov DWORD PTR [ebp-0x10],0x3 ; a.z = +0x3 24 point_t b = { .x = 0x1, .y = 0x2, .z = -0x3 }; 0x08048445 <+27>: mov DWORD PTR [ebp-0xc],0x1 ; b.x = 0x1 0x0804844c <+34>: mov DWORD PTR [ebp-0x8],0x2 ; b.y = 0x2 0x08048453 <+41>: mov DWORD PTR [ebp-0x4],0xfffffffd ; b.z = -0x3 25 return !cmp(a, b); 0x0804845a <+48>: mov eax,DWORD PTR [ebp-0xc] ; 0x0804845d <+51>: mov DWORD PTR [esp+0xc],eax ; save b.x to stack 0x08048461 <+55>: mov eax,DWORD PTR [ebp-0x8] ; 0x08048464 <+58>: mov DWORD PTR [esp+0x10],eax ; save b.y to stack 0x08048468 <+62>: mov eax,DWORD PTR [ebp-0x4] ; 0x0804846b <+65>: mov DWORD PTR [esp+0x14],eax ; save b.z to stack 0x0804846f <+69>: mov eax,DWORD PTR [ebp-0x18] ; 0x08048472 <+72>: mov DWORD PTR [esp],eax ; save a.x to stack 0x08048475 <+75>: mov eax,DWORD PTR [ebp-0x14] ; 0x08048478 <+78>: mov DWORD PTR [esp+0x4],eax ; save a.y to stack 0x0804847c <+82>: mov eax,DWORD PTR [ebp-0x10] ; 0x0804847f <+85>: mov DWORD PTR [esp+0x8],eax ; save a.z to stack 0x08048483 <+89>: call 0x80483ed <cmp> ;
也就是说在caller中调用cmp(a, b)表面上传递了两个实参，其实给stack里压入了6个值。 而对于callee cmp()来说，需要去栈里把对应的6个值取出来使用。
作为对比， 下面的程序片段在cmp()中使用结构体变量指针。
o foo2.c
1 #define FALSE 0 2 #define TRUE (!0) 3 4 typedef struct point_s { 5 int x; 6 int y; 7 int z; 8 } point_t; 9 10 static int cmp(point_t *a, point_t *b) 11 { 12 if (a->x != b->x) 13 return FALSE; 14 if (a->y != b->y) 15 return FALSE; 16 if (a->z != b->z) 17 return FALSE; 18 return TRUE; 19 } 20 21 int main(int argc, char *argv[]) 22 { 23 point_t a = { .x = 0x1, .y = 0x2, .z = +0x3 }; 24 point_t b = { .x = 0x1, .y = 0x2, .z = -0x3 }; 25 return !cmp(&a, &b); 26 }
o foo1.c v.s. foo2.c
o 对foo2.c进行编译后反汇编
1 $ gcc -g -Wall -m32 -std=gnu99 -o foo2 foo2.c 2 $ gdb foo2 3 (gdb) set disassembly-flavor intel 4 (gdb) disas /m main 5 Dump of assembler code for function main: 6 22 { 7 0x0804843a <+0>: push ebp 8 0x0804843b <+1>: mov ebp,esp 9 0x0804843d <+3>: sub esp,0x28 10 11 23 point_t a = { .x = 0x1, .y = 0x2, .z = +0x3 }; 12 0x08048440 <+6>: mov DWORD PTR [ebp-0x18],0x1 13 0x08048447 <+13>: mov DWORD PTR [ebp-0x14],0x2 14 0x0804844e <+20>: mov DWORD PTR [ebp-0x10],0x3 15 16 24 point_t b = { .x = 0x1, .y = 0x2, .z = -0x3 }; 17 0x08048455 <+27>: mov DWORD PTR [ebp-0xc],0x1 18 0x0804845c <+34>: mov DWORD PTR [ebp-0x8],0x2 19 0x08048463 <+41>: mov DWORD PTR [ebp-0x4],0xfffffffd 20 21 25 return !cmp(&a, &b); 22 0x0804846a <+48>: lea eax,[ebp-0xc] 23 0x0804846d <+51>: mov DWORD PTR [esp+0x4],eax 24 0x08048471 <+55>: lea eax,[ebp-0x18] 25 0x08048474 <+58>: mov DWORD PTR [esp],eax 26 0x08048477 <+61>: call 0x80483ed <cmp> 27 0x0804847c <+66>: test eax,eax 28 0x0804847e <+68>: sete al 29 0x08048481 <+71>: movzx eax,al 30 31 26 } 32 0x08048484 <+74>: leave 33 0x08048485 <+75>: ret 34 35 End of assembler dump. 36 (gdb) disas /m cmp 37 Dump of assembler code for function cmp: 38 11 { 39 0x080483ed <+0>: push ebp 40 0x080483ee <+1>: mov ebp,esp 41 42 12 if (a->x != b->x) 43 0x080483f0 <+3>: mov eax,DWORD PTR [ebp+0x8] 44 0x080483f3 <+6>: mov edx,DWORD PTR [eax] 45 0x080483f5 <+8>: mov eax,DWORD PTR [ebp+0xc] 46 0x080483f8 <+11>: mov eax,DWORD PTR [eax] 47 0x080483fa <+13>: cmp edx,eax 48 0x080483fc <+15>: je 0x8048405 <cmp+24> 49 50 13 return FALSE; 51 0x080483fe <+17>: mov eax,0x0 52 0x08048403 <+22>: jmp 0x8048438 <cmp+75> 53 54 14 if (a->y != b->y) 55 0x08048405 <+24>: mov eax,DWORD PTR [ebp+0x8] 56 0x08048408 <+27>: mov edx,DWORD PTR [eax+0x4] 57 0x0804840b <+30>: mov eax,DWORD PTR [ebp+0xc] 58 0x0804840e <+33>: mov eax,DWORD PTR [eax+0x4] 59 0x08048411 <+36>: cmp edx,eax 60 0x08048413 <+38>: je 0x804841c <cmp+47> 61 62 15 return FALSE; 63 0x08048415 <+40>: mov eax,0x0 64 0x0804841a <+45>: jmp 0x8048438 <cmp+75> 65 66 16 if (a->z != b->z) 67 0x0804841c <+47>: mov eax,DWORD PTR [ebp+0x8] 68 0x0804841f <+50>: mov edx,DWORD PTR [eax+0x8] 69 0x08048422 <+53>: mov eax,DWORD PTR [ebp+0xc] 70 0x08048425 <+56>: mov eax,DWORD PTR [eax+0x8] 71 0x08048428 <+59>: cmp edx,eax 72 0x0804842a <+61>: je 0x8048433 <cmp+70> 73 74 17 return FALSE; 75 0x0804842c <+63>: mov eax,0x0 76 0x08048431 <+68>: jmp 0x8048438 <cmp+75> 77 78 18 return TRUE; 79 0x08048433 <+70>: mov eax,0x1 80 81 19 } 82 0x08048438 <+75>: pop ebp 83 0x08048439 <+76>: ret 84 85 End of assembler dump. 86 (gdb)
o caller: point_t b的地址&b和point_t a的地址&a被依次存入到stack上
23 point_t a = { .x = 0x1, .y = 0x2, .z = +0x3 }; 0x08048440 <+6>: mov DWORD PTR [ebp-0x18],0x1 ; a.x = 0x1 0x08048447 <+13>: mov DWORD PTR [ebp-0x14],0x2 ; a.y = 0x2 0x0804844e <+20>: mov DWORD PTR [ebp-0x10],0x3 ; a.z = +0x3 24 point_t b = { .x = 0x1, .y = 0x2, .z = -0x3 }; 0x08048455 <+27>: mov DWORD PTR [ebp-0xc],0x1 ; b.x = 0x1 0x0804845c <+34>: mov DWORD PTR [ebp-0x8],0x2 ; b.y = 0x2 0x08048463 <+41>: mov DWORD PTR [ebp-0x4],0xfffffffd ; b.z = -0x3 25 return !cmp(&a, &b); 0x0804846a <+48>: lea eax,[ebp-0xc] ; get &b (addr of struct b) 0x0804846d <+51>: mov DWORD PTR [esp+0x4],eax ; save &b to stack 0x08048471 <+55>: lea eax,[ebp-0x18] ; get &a (addr of struct a) 0x08048474 <+58>: mov DWORD PTR [esp],eax ; save &a to stack 0x08048477 <+61>: call 0x80483ed <cmp> ;
显然，在caller中使用cmp(&a, &b)只需要给栈里存入两个值， 相比之下， cmp(a, b)给栈里存入了6个值，cmp(&a, &b) 效率确实高。
另外，在64位的程序中，前6个参数是默认存在寄存器上的，如果超过6个参数，才使用栈传递（具体请参见对应的ABI）。如果使用结构体变量传递参数，对寄存器是极大的浪费。
结论：
- 不要在函数参数中使用结构体变量；
- 也不要在函数中定义太多的参数，<=6最好；
- 如果不可避免地要使用较多的参数，设计函数的时候请最大化利用结构体，然后使用结构体指针作为参数。
