是否有标准宏来检测需要对齐内存访问的架构？

Question

假设有类似的东西:

void mask_bytes(unsigned char* dest, unsigned char* src, unsigned char* mask, unsigned int len)
{
  unsigned int i;
  for(i=0; i<len; i++)
  {
     dest[i] = src[i] & mask[i];
  }
}

通过编写类似下面的内容,我可以更快地在非对齐访问机器(例如x86)上运行

void mask_bytes(unsigned char* dest, unsigned char* src, unsigned char* mask, unsigned int len)
{
  unsigned int i;
  unsigned int wordlen = len >> 2;
  for(i=0; i<wordlen; i++)
  {
    ((uint32_t*)dest)[i] = ((uint32_t*)src)[i] & ((uint32_t*)mask)[i]; // this raises SIGBUS on SPARC and other archs that require aligned access.
  }
  for(i=wordlen<<2; i<len; i++){
    dest[i] = src[i] & mask[i];
  }
}

然而,它需要建立在几个架构上,所以我想做一些像:

void mask_bytes(unsigned char* dest, unsigned char* src, unsigned char* mask, unsigned int len)
{
  unsigned int i;
  unsigned int wordlen = len >> 2;

#if defined(__ALIGNED2__) || defined(__ALIGNED4__) || defined(__ALIGNED8__)
  // go slow
  for(i=0; i<len; i++)
  {
     dest[i] = src[i] & mask[i];
  }
#else
  // go fast
  for(i=0; i<wordlen; i++)
  {
    // the following line will raise SIGBUS on SPARC and other archs that require aligned access.
    ((uint32_t*)dest)[i] = ((uint32_t*)src)[i] & ((uint32_t*)mask)[i]; 
  }
  for(i=wordlen<<2; i<len; i++){
    dest[i] = src[i] & mask[i];
  }
#endif
}

但我找不到任何有关编译器定义的宏的信息(如我__ALIGNED4__上面的假设),它指定了对齐或使用预处理器确定目标架构对齐的任何巧妙方法.我可以测试defined (__SVR4) && defined (__sun),但我更喜欢在其他需要对齐内存访问的架构上使用Just Work ^_TM的东西.

Answer 1

虽然x86默默地修复了未对齐的访问,但这对性能来说并不是最佳选择.通常最好假设某个对齐并自己执行修正:

unsigned int const alignment = 8;   /* or 16, or sizeof(long) */

void memcpy(char *dst, char const *src, unsigned int size) {
    if((((intptr_t)dst) % alignment) != (((intptr_t)src) % alignment)) {
        /* no common alignment, copy as bytes or shift around */
    } else {
        if(((intptr_t)dst) % alignment) {
            /* copy bytes at the beginning */
        }
        /* copy words in the middle */
        if(((intptr_t)dst + size) % alignment) {
            /* copy bytes at the end */
        }
    }
}

另外,请查看SIMD说明.