Visual Studio 编译器的循环展开问题

发布于 2025-01-12 11:55:06 字数 3370 浏览 0 评论 0原文

我有一些简单的设置,我注意到 VS 编译器似乎不够智能,无法展开循环,但其他编译器(如 clang 或 gcc)会这样做。我是否错过了 VS 的一些优化标志?

#include <cstddef>

struct A
{
  double data[4];
  double *begin() { return data; }
  double *end() { return data + 4; }
  double const *begin() const { return data; }
  double const *end() const { return data + 4; }
};

double sum_index(A const &a) {
  double ret = 0;
  for(std::size_t i = 0; i < 4; ++i)
  {
    ret += a.data[i];
  }
  return ret;
}

double sum_iter(A const &a) {
  double ret = 0;
  for(auto const &v : a)
  {
    ret += v;
  }
  return ret;
}

我使用 https://godbolt.org/ 编译器资源管理器来生成汇编代码。

gcc 11.2 with -O3:

sum_index(A const&):
        pxor    xmm0, xmm0
        addsd   xmm0, QWORD PTR [rdi]
        addsd   xmm0, QWORD PTR [rdi+8]
        addsd   xmm0, QWORD PTR [rdi+16]
        addsd   xmm0, QWORD PTR [rdi+24]
        ret
sum_iter(A const&):
        movsd   xmm1, QWORD PTR [rdi]
        addsd   xmm1, QWORD PTR .LC0[rip]
        movsd   xmm0, QWORD PTR [rdi+8]
        addsd   xmm1, xmm0
        movupd  xmm0, XMMWORD PTR [rdi+16]
        addsd   xmm1, xmm0
        unpckhpd        xmm0, xmm0
        addsd   xmm0, xmm1
        ret
.LC0:
        .long   0
        .long   0

clang 13.0.1 with -O3:

sum_index(A const&):                       # @sum_index(A const&)
        xorpd   xmm0, xmm0
        addsd   xmm0, qword ptr [rdi]
        addsd   xmm0, qword ptr [rdi + 8]
        addsd   xmm0, qword ptr [rdi + 16]
        addsd   xmm0, qword ptr [rdi + 24]
        ret
sum_iter(A const&):                        # @sum_iter(A const&)
        xorpd   xmm0, xmm0
        addsd   xmm0, qword ptr [rdi]
        addsd   xmm0, qword ptr [rdi + 8]
        addsd   xmm0, qword ptr [rdi + 16]
        addsd   xmm0, qword ptr [rdi + 24]
        ret

MSVC 19.30 with /O2(没有/O3?):

this$ = 8
double const * A::begin(void)const  PROC                         ; A::begin, COMDAT
        mov     rax, rcx
        ret     0
double const * A::begin(void)const  ENDP                         ; A::begin

this$ = 8
double const * A::end(void)const  PROC                       ; A::end, COMDAT
        lea     rax, QWORD PTR [rcx+32]
        ret     0
double const * A::end(void)const  ENDP                       ; A::end

a$ = 8
double sum_index(A const &) PROC                     ; sum_index, COMDAT
        movsd   xmm0, QWORD PTR [rcx]
        xorps   xmm1, xmm1
        addsd   xmm0, xmm1
        addsd   xmm0, QWORD PTR [rcx+8]
        addsd   xmm0, QWORD PTR [rcx+16]
        addsd   xmm0, QWORD PTR [rcx+24]
        ret     0
double sum_index(A const &) ENDP                     ; sum_index

a$ = 8
double sum_iter(A const &) PROC                            ; sum_iter, COMDAT
        lea     rax, QWORD PTR [rcx+32]
        xorps   xmm0, xmm0
        cmp     rcx, rax
        je      SHORT $LN12@sum_iter
        npad    4
$LL8@sum_iter:
        addsd   xmm0, QWORD PTR [rcx]
        add     rcx, 8
        cmp     rcx, rax
        jne     SHORT $LL8@sum_iter
$LN12@sum_iter:
        ret     0
double sum_iter(A const &) ENDP                            ; sum_iter

显然,展开 MSVC 循环存在问题。我需要设置一些额外的优化标志吗?

感谢您的帮助!

I have some simple setup, where I noticed that VS compiler seems not smart enough to unroll loop, but other compilers like clang or gcc do so. Do I miss some optimization flag for VS?

#include <cstddef>

struct A
{
  double data[4];
  double *begin() { return data; }
  double *end() { return data + 4; }
  double const *begin() const { return data; }
  double const *end() const { return data + 4; }
};

double sum_index(A const &a) {
  double ret = 0;
  for(std::size_t i = 0; i < 4; ++i)
  {
    ret += a.data[i];
  }
  return ret;
}

double sum_iter(A const &a) {
  double ret = 0;
  for(auto const &v : a)
  {
    ret += v;
  }
  return ret;
}

I used https://godbolt.org/ compiler explorer to generate assembler code.

gcc 11.2 with -O3:

sum_index(A const&):
        pxor    xmm0, xmm0
        addsd   xmm0, QWORD PTR [rdi]
        addsd   xmm0, QWORD PTR [rdi+8]
        addsd   xmm0, QWORD PTR [rdi+16]
        addsd   xmm0, QWORD PTR [rdi+24]
        ret
sum_iter(A const&):
        movsd   xmm1, QWORD PTR [rdi]
        addsd   xmm1, QWORD PTR .LC0[rip]
        movsd   xmm0, QWORD PTR [rdi+8]
        addsd   xmm1, xmm0
        movupd  xmm0, XMMWORD PTR [rdi+16]
        addsd   xmm1, xmm0
        unpckhpd        xmm0, xmm0
        addsd   xmm0, xmm1
        ret
.LC0:
        .long   0
        .long   0

clang 13.0.1 with -O3:

sum_index(A const&):                       # @sum_index(A const&)
        xorpd   xmm0, xmm0
        addsd   xmm0, qword ptr [rdi]
        addsd   xmm0, qword ptr [rdi + 8]
        addsd   xmm0, qword ptr [rdi + 16]
        addsd   xmm0, qword ptr [rdi + 24]
        ret
sum_iter(A const&):                        # @sum_iter(A const&)
        xorpd   xmm0, xmm0
        addsd   xmm0, qword ptr [rdi]
        addsd   xmm0, qword ptr [rdi + 8]
        addsd   xmm0, qword ptr [rdi + 16]
        addsd   xmm0, qword ptr [rdi + 24]
        ret

MSVC 19.30 with /O2 (there is no /O3?):

this$ = 8
double const * A::begin(void)const  PROC                         ; A::begin, COMDAT
        mov     rax, rcx
        ret     0
double const * A::begin(void)const  ENDP                         ; A::begin

this$ = 8
double const * A::end(void)const  PROC                       ; A::end, COMDAT
        lea     rax, QWORD PTR [rcx+32]
        ret     0
double const * A::end(void)const  ENDP                       ; A::end

a$ = 8
double sum_index(A const &) PROC                     ; sum_index, COMDAT
        movsd   xmm0, QWORD PTR [rcx]
        xorps   xmm1, xmm1
        addsd   xmm0, xmm1
        addsd   xmm0, QWORD PTR [rcx+8]
        addsd   xmm0, QWORD PTR [rcx+16]
        addsd   xmm0, QWORD PTR [rcx+24]
        ret     0
double sum_index(A const &) ENDP                     ; sum_index

a$ = 8
double sum_iter(A const &) PROC                            ; sum_iter, COMDAT
        lea     rax, QWORD PTR [rcx+32]
        xorps   xmm0, xmm0
        cmp     rcx, rax
        je      SHORT $LN12@sum_iter
        npad    4
$LL8@sum_iter:
        addsd   xmm0, QWORD PTR [rcx]
        add     rcx, 8
        cmp     rcx, rax
        jne     SHORT $LL8@sum_iter
$LN12@sum_iter:
        ret     0
double sum_iter(A const &) ENDP                            ; sum_iter

Obviously there is problem with unrolling the loop for MSVC. Is there some additional optimization flag I have to set?

Thanks for help!

如果你对这篇内容有疑问,欢迎到本站社区发帖提问 参与讨论,获取更多帮助,或者扫码二维码加入 Web 技术交流群。

扫码二维码加入Web技术交流群

发布评论

需要 登录 才能够评论, 你可以免费 注册 一个本站的账号。
列表为空,暂无数据
我们使用 Cookies 和其他技术来定制您的体验包括您的登录状态等。通过阅读我们的 隐私政策 了解更多相关信息。 单击 接受 或继续使用网站,即表示您同意使用 Cookies 和您的相关数据。
原文