为什么添加内联汇编注释会导致GCC生成的代码发生如此根本的变化? [英] Why does adding inline assembly comments cause such radical change in GCC's generated code?
问题描述
所以，我有以下代码:
constexpr unsigned N = 1000;
void f1(char* sum, char* a, char* b) {
for(int i = 0; i < N; ++i) {
sum[i] = a[i] + b[i];
}
}
void f2(char* sum, char* a, char* b) {
char* end = sum + N;
while(sum != end) {
*sum++ = *a++ + *b++;
}
}
我想看一下GCC 4.7.2生成的代码.所以我运行了g++ -march=native -O3 -masm=intel -S a.c++ -std=c++11并获得了以下输出:
I wanted to see the code that GCC 4.7.2 would generate. So I ran g++ -march=native -O3 -masm=intel -S a.c++ -std=c++11 And got the following output:
.file "a.c++"
.intel_syntax noprefix
.text
.p2align 4,,15
.globl _Z2f1PcS_S_
.type _Z2f1PcS_S_, @function
_Z2f1PcS_S_:
.LFB0:
.cfi_startproc
lea rcx, [rdx+16]
lea rax, [rdi+16]
cmp rdi, rcx
setae r8b
cmp rdx, rax
setae cl
or cl, r8b
je .L5
lea rcx, [rsi+16]
cmp rdi, rcx
setae cl
cmp rsi, rax
setae al
or cl, al
je .L5
xor eax, eax
.p2align 4,,10
.p2align 3
.L3:
movdqu xmm0, XMMWORD PTR [rdx+rax]
movdqu xmm1, XMMWORD PTR [rsi+rax]
paddb xmm0, xmm1
movdqu XMMWORD PTR [rdi+rax], xmm0
add rax, 16
cmp rax, 992
jne .L3
mov ax, 8
mov r9d, 992
.L2:
sub eax, 1
lea rcx, [rdx+r9]
add rdi, r9
lea r8, [rax+1]
add rsi, r9
xor eax, eax
.p2align 4,,10
.p2align 3
.L4:
movzx edx, BYTE PTR [rcx+rax]
add dl, BYTE PTR [rsi+rax]
mov BYTE PTR [rdi+rax], dl
add rax, 1
cmp rax, r8
jne .L4
rep
ret
.L5:
mov eax, 1000
xor r9d, r9d
jmp .L2
.cfi_endproc
.LFE0:
.size _Z2f1PcS_S_, .-_Z2f1PcS_S_
.p2align 4,,15
.globl _Z2f2PcS_S_
.type _Z2f2PcS_S_, @function
_Z2f2PcS_S_:
.LFB1:
.cfi_startproc
lea rcx, [rdx+16]
lea rax, [rdi+16]
cmp rdi, rcx
setae r8b
cmp rdx, rax
setae cl
or cl, r8b
je .L19
lea rcx, [rsi+16]
cmp rdi, rcx
setae cl
cmp rsi, rax
setae al
or cl, al
je .L19
xor eax, eax
.p2align 4,,10
.p2align 3
.L17:
movdqu xmm0, XMMWORD PTR [rdx+rax]
movdqu xmm1, XMMWORD PTR [rsi+rax]
paddb xmm0, xmm1
movdqu XMMWORD PTR [rdi+rax], xmm0
add rax, 16
cmp rax, 992
jne .L17
add rdi, 992
add rsi, 992
add rdx, 992
mov r8d, 8
.L16:
xor eax, eax
.p2align 4,,10
.p2align 3
.L18:
movzx ecx, BYTE PTR [rdx+rax]
add cl, BYTE PTR [rsi+rax]
mov BYTE PTR [rdi+rax], cl
add rax, 1
cmp rax, r8
jne .L18
rep
ret
.L19:
mov r8d, 1000
jmp .L16
.cfi_endproc
.LFE1:
.size _Z2f2PcS_S_, .-_Z2f2PcS_S_
.ident "GCC: (GNU) 4.7.2"
.section .note.GNU-stack,"",@progbits
我很讨厌阅读汇编，所以我决定添加一些标记来了解循环体的去向:
I suck at reading assembly, so I decided to add some markers to know where the bodies of the loops went:
constexpr unsigned N = 1000;
void f1(char* sum, char* a, char* b) {
for(int i = 0; i < N; ++i) {
asm("# im in ur loop");
sum[i] = a[i] + b[i];
}
}
void f2(char* sum, char* a, char* b) {
char* end = sum + N;
while(sum != end) {
asm("# im in ur loop");
*sum++ = *a++ + *b++;
}
}
GCC吐了出来:
.file "a.c++"
.intel_syntax noprefix
.text
.p2align 4,,15
.globl _Z2f1PcS_S_
.type _Z2f1PcS_S_, @function
_Z2f1PcS_S_:
.LFB0:
.cfi_startproc
xor eax, eax
.p2align 4,,10
.p2align 3
.L2:
#APP
# 4 "a.c++" 1
# im in ur loop
# 0 "" 2
#NO_APP
movzx ecx, BYTE PTR [rdx+rax]
add cl, BYTE PTR [rsi+rax]
mov BYTE PTR [rdi+rax], cl
add rax, 1
cmp rax, 1000
jne .L2
rep
ret
.cfi_endproc
.LFE0:
.size _Z2f1PcS_S_, .-_Z2f1PcS_S_
.p2align 4,,15
.globl _Z2f2PcS_S_
.type _Z2f2PcS_S_, @function
_Z2f2PcS_S_:
.LFB1:
.cfi_startproc
xor eax, eax
.p2align 4,,10
.p2align 3
.L6:
#APP
# 12 "a.c++" 1
# im in ur loop
# 0 "" 2
#NO_APP
movzx ecx, BYTE PTR [rdx+rax]
add cl, BYTE PTR [rsi+rax]
mov BYTE PTR [rdi+rax], cl
add rax, 1
cmp rax, 1000
jne .L6
rep
ret
.cfi_endproc
.LFE1:
.size _Z2f2PcS_S_, .-_Z2f2PcS_S_
.ident "GCC: (GNU) 4.7.2"
.section .note.GNU-stack,"",@progbits
这要短得多，并且有一些明显的区别，例如缺少SIMD指令.我期待相同的输出，中间会有一些注释.我在这里做一些错误的假设吗? GCC的优化程序是否会受到asm注释的阻碍?
This is considerably shorter, and has some significant differences like the lack of SIMD instructions. I was expecting the same output, with some comments somewhere in the middle of it. Am I making some wrong assumption here? Is GCC's optimizer hindered by asm comments?
推荐答案
有关优化的交互作用，在页.
The interactions with optimisations are explained about halfway down the "Assembler Instructions with C Expression Operands" page in the documentation.
GCC不会尝试理解asm内部的任何实际程序集；它唯一了解的内容是您(可选)在输出和输入操作数规范以及寄存器清除列表中告诉您的内容.
GCC doesn't try to understand any of the actual assembly inside the asm; the only thing it knows about the content is what you (optionally) tell it in the output and input operand specification and the register clobber list.
尤其要注意:
没有任何输出操作数的
asm指令将与易失性asm指令等同对待.
An
asminstruction without any output operands will be treated identically to a volatileasminstruction.
和
volatile关键字表示该指令具有重要的副作用[...]
The
volatilekeyword indicates that the instruction has important side-effects [...]
因此asm的存在限制了矢量化的优化，因为GCC认为它具有副作用.
So the presence of the asm inside your loop has inhibited a vectorisation optimisation, because GCC assumes it has side effects.
这篇关于为什么添加内联汇编注释会导致GCC生成的代码发生如此根本的变化?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！
