对C99可变长度数组（VLA）使用限制限定符

Question

对C99可变长度数组（VLA）使用限制限定符

Jef*_*eff 4 simd c99 variable-length-array restrict-qualifier auto-vectorization

我正在探索如何基于函数签名在C99中简单循环的不同实现自动矢量化。

这是我的代码：

/* #define PRAGMA_SIMD _Pragma("simd") */
#define PRAGMA_SIMD

#ifdef __INTEL_COMPILER
#define ASSUME_ALIGNED(a) __assume_aligned(a,64)
#else
#define ASSUME_ALIGNED(a)
#endif

#ifndef ARRAY_RESTRICT
#define ARRAY_RESTRICT
#endif

void foo1(double * restrict a, const double * restrict b, const double * restrict c) 
{ 
    ASSUME_ALIGNED(a);
    ASSUME_ALIGNED(b);
    ASSUME_ALIGNED(c);
    PRAGMA_SIMD
    for (int i = 0; i < 2048; ++i) {
        if (c[i] > 0) {
            a[i] = b[i];
        } else {
            a[i] = 0.0;
        } 
    }
}

void foo2(double * restrict a, const double * restrict b, const double * restrict c) 
{ 
    ASSUME_ALIGNED(a);
    ASSUME_ALIGNED(b);
    ASSUME_ALIGNED(c);
    PRAGMA_SIMD
    for (int i = 0; i < 2048; ++i) {
        a[i] = ((c[i] > 0) ? b[i] : 0.0);
    }
}

/* Undetermined size version */

void foo3(int n, double * restrict a, const double * restrict b, const double * restrict c) 
{ 
    ASSUME_ALIGNED(a);
    ASSUME_ALIGNED(b);
    ASSUME_ALIGNED(c);
    PRAGMA_SIMD
    for (int i = 0; i < n; ++i) {
        if (c[i] > 0) {
            a[i] = b[i];
        } else {
            a[i] = 0.0;
        } 
    }
}

void foo4(int n, double * restrict a, const double * restrict b, const double * restrict c) 
{ 
    ASSUME_ALIGNED(a);
    ASSUME_ALIGNED(b);
    ASSUME_ALIGNED(c);
    PRAGMA_SIMD
    for (int i = 0; i < n; ++i) {
        a[i] = ((c[i] > 0) ? b[i] : 0.0);
    }
}

/* Static array versions */

void foo5(double ARRAY_RESTRICT a[2048], const double ARRAY_RESTRICT b[2048], const double ARRAY_RESTRICT c[2048]) 
{ 
    ASSUME_ALIGNED(a);
    ASSUME_ALIGNED(b);
    ASSUME_ALIGNED(c);
    PRAGMA_SIMD
    for (int i = 0; i < 2048; ++i) {
        if (c[i] > 0) {
            a[i] = b[i];
        } else {
            a[i] = 0.0;
        } 
    }
}

void foo6(double ARRAY_RESTRICT a[2048], const double ARRAY_RESTRICT b[2048], const double ARRAY_RESTRICT c[2048]) 
{ 
    ASSUME_ALIGNED(a);
    ASSUME_ALIGNED(b);
    ASSUME_ALIGNED(c);
    PRAGMA_SIMD
    for (int i = 0; i < 2048; ++i) {
        a[i] = ((c[i] > 0) ? b[i] : 0.0);
    }
}

/* VLA versions */

void foo7(int n, double ARRAY_RESTRICT a[n], const double ARRAY_RESTRICT b[n], const double ARRAY_RESTRICT c[n]) 
{ 
    ASSUME_ALIGNED(a);
    ASSUME_ALIGNED(b);
    ASSUME_ALIGNED(c);
    PRAGMA_SIMD
    for (int i = 0; i < n; ++i) {
        if (c[i] > 0) {
            a[i] = b[i];
        } else {
            a[i] = 0.0;
        } 
    }
}

void foo8(int n, double ARRAY_RESTRICT a[n], const double ARRAY_RESTRICT b[n], const double ARRAY_RESTRICT c[n]) 
{ 
    ASSUME_ALIGNED(a);
    ASSUME_ALIGNED(b);
    ASSUME_ALIGNED(c);
    PRAGMA_SIMD
    for (int i = 0; i < n; ++i) {
        a[i] = ((c[i] > 0) ? b[i] : 0.0);
    }
}

Run Code Online (Sandbox Code Playgroud)

当我编译时

$ icc -O3 -std=c99 -opt-report5 -mavx -S foo.c 
icc: remark #10397: optimization reports are generated in *.optrpt files in the output location

Run Code Online (Sandbox Code Playgroud)

我看到VLA情况不是自动矢量化的，但是当我添加标志以断言没有别名时-fno-alias，它们是自动的。因此，我得出结论，我应该在源代码中对此进行规定，因此我尝试通过使用

$ icc -O3 -std=c99 -opt-report5 -mavx -DARRAY_RESTRICT=restrict -S foo.c 
icc: remark #10397: optimization reports are generated in *.optrpt files in the output location

Run Code Online (Sandbox Code Playgroud)

编译器错误输出包括

foo.c(98): error: "restrict" is not allowed
void foo7(int n, double ARRAY_RESTRICT a[n], const double ARRAY_RESTRICT b[n], 
const double ARRAY_RESTRICT c[n]) 

             ^

Run Code Online (Sandbox Code Playgroud)

但是如您所见，我的VLA参数上不允许使用限制。

所以我的问题是：没有办法在ISO C中断言VLA没有别名吗？

请注意，我可以使用杂注断言在源代码中不走样-例如simd，omp simd，ivdep等等-并获得自动向量化，我想但这些都不是ISO C.

在这种情况下，ISO C表示C的最新版本，在撰写本文时，当然是C11。

Answer 1

Jon*_*ler 5

您的原始代码对我来说很失败，并显示以下消息：

 void foo7(int n, double ARRAY_RESTRICT a[n], const double ARRAY_RESTRICT b[n], const double ARRAY_RESTRICT c[n])
 ^
restrict.c:126:1: error: invalid use of ‘restrict’
restrict.c:126:1: error: invalid use of ‘restrict’
restrict.c:145:1: error: invalid use of ‘restrict’

Run Code Online (Sandbox Code Playgroud)

^{转移评论的选定部分}

§6.7.6.3 函数声明符（包括原型）具有示例5，该示例表示以下函数原型声明符是等效的：

void f(double (* restrict a)[5]);
void f(double a[restrict][5]);
void f(double a[restrict 3][5]);
void f(double a[restrict static 3][5]);

Run Code Online (Sandbox Code Playgroud)

这是标准中唯一出现与数组类型直接相关的限制的地方。§6.7.6通常用于声明符，而§6.7.6.2通常用于数组声明符，在我看来，限制似乎必须出现在数组维的第一个组件内。在您的上下文中，应为：

void foo7(int n, double a[ARRAY_RESTRICT n],
           const double b[ARRAY_RESTRICT n],
           const double c[ARRAY_RESTRICT n])

Run Code Online (Sandbox Code Playgroud)

如果不看标准中的示例，而您问这个问题，我将不会相信这种表示法！请注意，这适用于阵列以及VLA。

经过修改的代码基于注释，可以在相同的编译选项下清晰地编译：

gcc -g -O3 -std=c11 -Wall -Wextra -Wmissing-prototypes -Wstrict-prototypes \
    -Wold-style-definition -Wold-style-declaration -Werror -c new.restrict.c

Run Code Online (Sandbox Code Playgroud)

编译选项要求事先声明非静态函数，因此声明位于文件顶部。我也强加#define ARRAY_RESTRICT restrict了源代码，而不是将其保留为编译选项。

编译器是在Ubuntu 14.04衍生产品上运行的GCC 4.9.2。

档案new.restrict.c：

/* #define PRAGMA_SIMD _Pragma("simd") */
#define PRAGMA_SIMD

#ifdef __INTEL_COMPILER
#define ASSUME_ALIGNED(a) __assume_aligned(a, 64)
#else
#define ASSUME_ALIGNED(a)
#endif

#define ARRAY_RESTRICT restrict

#ifndef ARRAY_RESTRICT
#define ARRAY_RESTRICT
#endif

void foo1(double *restrict a, const double *restrict b, const double *restrict c);
void foo2(double *restrict a, const double *restrict b, const double *restrict c);
void foo3(int n, double *restrict a, const double *restrict b, const double *restrict c);
void foo4(int n, double *restrict a, const double *restrict b, const double *restrict c);
void foo5(double a[ARRAY_RESTRICT 2048], const double b[ARRAY_RESTRICT 2048], const double c[ARRAY_RESTRICT 2048]);
void foo6(double a[ARRAY_RESTRICT 2048], const double b[ARRAY_RESTRICT 2048], const double c[ARRAY_RESTRICT 2048]);
void foo7(int n, double a[ARRAY_RESTRICT n], const double b[ARRAY_RESTRICT n], const double c[ARRAY_RESTRICT n]);
void foo8(int n, double a[ARRAY_RESTRICT n], const double b[ARRAY_RESTRICT n], const double c[ARRAY_RESTRICT n]);

void foo1(double *restrict a, const double *restrict b, const double *restrict c)
{
    ASSUME_ALIGNED(a);
    ASSUME_ALIGNED(b);
    ASSUME_ALIGNED(c);
    PRAGMA_SIMD
    for (int i = 0; i < 2048; ++i)
    {
        if (c[i] > 0)
        {
            a[i] = b[i];
        }
        else
        {
            a[i] = 0.0;
        }
    }
}

void foo2(double *restrict a, const double *restrict b, const double *restrict c)
{
    ASSUME_ALIGNED(a);
    ASSUME_ALIGNED(b);
    ASSUME_ALIGNED(c);
    PRAGMA_SIMD
    for (int i = 0; i < 2048; ++i)
    {
        a[i] = ((c[i] > 0) ? b[i] : 0.0);
    }
}

/* Undetermined size version */

void foo3(int n, double *restrict a, const double *restrict b, const double *restrict c)
{
    ASSUME_ALIGNED(a);
    ASSUME_ALIGNED(b);
    ASSUME_ALIGNED(c);
    PRAGMA_SIMD
    for (int i = 0; i < n; ++i)
    {
        if (c[i] > 0)
        {
            a[i] = b[i];
        }
        else
        {
            a[i] = 0.0;
        }
    }
}

void foo4(int n, double *restrict a, const double *restrict b, const double *restrict c)
{
    ASSUME_ALIGNED(a);
    ASSUME_ALIGNED(b);
    ASSUME_ALIGNED(c);
    PRAGMA_SIMD
    for (int i = 0; i < n; ++i)
    {
        a[i] = ((c[i] > 0) ? b[i] : 0.0);
    }
}

/* Static array versions */

void foo5(double a[ARRAY_RESTRICT 2048], const double b[ARRAY_RESTRICT 2048], const double c[ARRAY_RESTRICT 2048])
{
    ASSUME_ALIGNED(a);
    ASSUME_ALIGNED(b);
    ASSUME_ALIGNED(c);
    PRAGMA_SIMD
    for (int i = 0; i < 2048; ++i)
    {
        if (c[i] > 0)
        {
            a[i] = b[i];
        }
        else
        {
            a[i] = 0.0;
        }
    }
}

void foo6(double a[ARRAY_RESTRICT 2048], const double b[ARRAY_RESTRICT 2048], const double c[ARRAY_RESTRICT 2048])
{
    ASSUME_ALIGNED(a);
    ASSUME_ALIGNED(b);
    ASSUME_ALIGNED(c);
    PRAGMA_SIMD
    for (int i = 0; i < 2048; ++i)
    {
        a[i] = ((c[i] > 0) ? b[i] : 0.0);
    }
}

/* VLA versions */

void foo7(int n, double a[ARRAY_RESTRICT n], const double b[ARRAY_RESTRICT n], const double c[ARRAY_RESTRICT n])
{
    ASSUME_ALIGNED(a);
    ASSUME_ALIGNED(b);
    ASSUME_ALIGNED(c);
    PRAGMA_SIMD
    for (int i = 0; i < n; ++i)
    {
        if (c[i] > 0)
        {
            a[i] = b[i];
        }
        else
        {
            a[i] = 0.0;
        }
    }
}

void foo8(int n, double a[ARRAY_RESTRICT n], const double b[ARRAY_RESTRICT n], const double c[ARRAY_RESTRICT n])
{
    ASSUME_ALIGNED(a);
    ASSUME_ALIGNED(b);
    ASSUME_ALIGNED(c);
    PRAGMA_SIMD
    for (int i = 0; i < n; ++i)
    {
        a[i] = ((c[i] > 0) ? b[i] : 0.0);
    }
}

Run Code Online (Sandbox Code Playgroud)

归档时间：	10 年，10 月前
查看次数：	747 次
最近记录：	10 年，10 月前