为什么这个C++程序如此快速？

我写了一个小基准来比较Python,Ruby,JavaScript和C++的不同解释器/编译器的性能.正如预期的那样,事实证明(优化的)C++胜过脚本语言,但它所做的因素非常高.

结果是:

sven@jet:~/tmp/js$ time node bla.js              # * JavaScript with node *
0

real    0m1.222s
user    0m1.190s
sys 0m0.015s
sven@jet:~/tmp/js$ time ruby foo.rb              # * Ruby *
0

real    0m52.428s
user    0m52.395s
sys 0m0.028s
sven@jet:~/tmp/js$ time python blub.py           # * Python with CPython *
0

real    1m16.480s
user    1m16.371s
sys 0m0.080s

sven@jet:~/tmp/js$ time pypy blub.py             # * Python with PyPy *
0

real    0m4.707s
user    0m4.579s
sys 0m0.028s

sven@jet:~/tmp/js$ time ./cpp_non_optimized 1000 1000000 # * C++ with -O0 (gcc) *
0

real    0m1.702s
user    0m1.699s
sys 0m0.002s
sven@jet:~/tmp/js$ time ./cpp_optimized 1000 1000000     # * C++ with -O3 (gcc) *
0

real    0m0.003s # (!!!) <---------------------------------- WHY?
user    0m0.002s
sys 0m0.002s

我想知道是否有人可以解释为什么优化的C++代码比其他所有代码快三个数量级.

C++基准测试使用命令行参数,以防止在编译时预先计算结果.

下面,我放置了不同语言基准的源代码,这些基准应该在语义上等效.另外,我提供了优化的C++编译器输出的汇编代码(使用gcc).在查看优化的程序集时,似乎编译器将基准测试中的两个循环合并为一个循环,但是,仍然存在循环!

JavaScript的:

var s = 0;
var outer = 1000;
var inner = 1000000;

for (var i = 0; i < outer; ++i) {
    for (var j = 0; j < inner; ++j) {
        ++s;
    }
    s -= inner;
}
console.log(s);

蟒蛇:

s = 0
outer = 1000
inner = 1000000

for _ in xrange(outer):
    for _ in xrange(inner):
        s += 1
    s -= inner
print s

红宝石:

s = 0
outer = 1000
inner = 1000000

outer_end = outer - 1
inner_end = inner - 1

for i in 0..outer_end
  for j in 0..inner_end
    s = s + 1
  end
  s = s - inner
end
puts s

C++:

#include 
#include 
#include 

int main(int argc, char* argv[]) {
  uint32_t s = 0;
  uint32_t outer = atoi(argv[1]);
  uint32_t inner = atoi(argv[2]);
  for (uint32_t i = 0; i < outer; ++i) {
    for (uint32_t j = 0; j < inner; ++j)
      ++s;
    s -= inner;
  }
  std::cout << s << std::endl;
  return 0;
}

汇编(使用gcc -S -O3 -std = c ++ 0x编译上述C++代码时):

    .file   "bar.cpp"
    .section    .text.startup,"ax",@progbits
    .p2align 4,,15
    .globl  main
    .type   main, @function
main:
.LFB1266:
    .cfi_startproc
    pushq   %r12
    .cfi_def_cfa_offset 16
    .cfi_offset 12, -16
    movl    $10, %edx
    movq    %rsi, %r12
    pushq   %rbp
    .cfi_def_cfa_offset 24
    .cfi_offset 6, -24
    pushq   %rbx
    .cfi_def_cfa_offset 32
    .cfi_offset 3, -32
    movq    8(%rsi), %rdi
    xorl    %esi, %esi
    call    strtol
    movq    16(%r12), %rdi
    movq    %rax, %rbp
    xorl    %esi, %esi
    movl    $10, %edx
    call    strtol
    testl   %ebp, %ebp
    je  .L6
    movl    %ebp, %ebx
    xorl    %eax, %eax
    xorl    %edx, %edx
    .p2align 4,,10
    .p2align 3
.L3:                             # <--- Here is the loop
    addl    $1, %eax             # <---
    cmpl    %eax, %ebx           # <---
    ja  .L3                      # <---
.L2:
    movl    %edx, %esi
    movl    $_ZSt4cout, %edi
    call    _ZNSo9_M_insertImEERSoT_
    movq    %rax, %rdi
    call    _ZSt4endlIcSt11char_traitsIcEERSt13basic_ostreamIT_T0_ES6_
    popq    %rbx
    .cfi_remember_state
    .cfi_def_cfa_offset 24
    popq    %rbp
    .cfi_def_cfa_offset 16
    xorl    %eax, %eax
    popq    %r12
    .cfi_def_cfa_offset 8
    ret
.L6:
    .cfi_restore_state
    xorl    %edx, %edx
    jmp .L2
    .cfi_endproc
.LFE1266:
    .size   main, .-main
    .p2align 4,,15
    .type   _GLOBAL__sub_I_main, @function
_GLOBAL__sub_I_main:
.LFB1420:
    .cfi_startproc
    subq    $8, %rsp
    .cfi_def_cfa_offset 16
    movl    $_ZStL8__ioinit, %edi
    call    _ZNSt8ios_base4InitC1Ev
    movl    $__dso_handle, %edx
    movl    $_ZStL8__ioinit, %esi
    movl    $_ZNSt8ios_base4InitD1Ev, %edi
    addq    $8, %rsp
    .cfi_def_cfa_offset 8
    jmp __cxa_atexit
    .cfi_endproc
.LFE1420:
    .size   _GLOBAL__sub_I_main, .-_GLOBAL__sub_I_main
    .section    .init_array,"aw"
    .align 8
    .quad   _GLOBAL__sub_I_main
    .local  _ZStL8__ioinit
    .comm   _ZStL8__ioinit,1,1
    .hidden __dso_handle
    .ident  "GCC: (Ubuntu 4.8.2-19ubuntu1) 4.8.2"
    .section    .note.GNU-stack,"",@progbits

ecatmur.. 102

优化器已经确定内部循环和后续行是无操作,并将其消除.不幸的是,它还没有设法消除外环.

请注意,node.js示例比未优化的C++示例更快,表明V8(节点的JIT编译器)已设法消除至少一个循环.但是,它的优化有一些开销,因为(像任何JIT编译器一样)它必须平衡优化和配置文件引导的重新优化的机会与这样做的成本.