联合代码生成，用于字节序交换

问题描述

我有一些用于字节序交换基本类型的简单函数，但是cl(版本15和16)会为x86上的浮点类型生成冗余的加载/存储.

I have some simple functions for endian swapping fundamental types, however cl (version 15 and 16) generates redundant load/stores for floating point types on x86. 

以下代码:

float EndianSwap(float value)
{
  union Alias
  {
    float native;
    unsigned long convert;
  };

  Alias asrc;
  asrc.native = value;

  Alias adst;
  adst.convert = _byteswap_ulong(asrc.convert);

  return adst.native;
}

推荐答案

以下代码:

The following code:

[代码]

[code]

float EndianSwap(浮点值)

float EndianSwap(float value)

这看起来很冒险.如果，您采用有效的浮点值并交换其字节，它将变成信号NaN?根据FPU标志的不同，可能会产生硬件.号码打断时 已加载以返回.

This looks like a pretty risky thing to do. Isn't it possible that, if  you take a valid floating point value and swap its bytes, it turns into  a signalling NaN? Depending on FPU flags, that may produce a hardware  interrupt when the number is loaded to be returned.

它已经违反了别名规则，导致未定义.行为，先写信给工会的一名成员，然后再从中读取其他.访问附件的二进制表示形式的唯一便携式方法对象是将其地址强制转换为[signed | unsigned] char *或memcpy it 放入足够大的char缓冲区.

It is already a violation of aliasing rules, leading to undefined  behavior, to write to one member of the union and then read from  another. The only portable way to access binary representation of an  object is to cast its address to [signed|unsigned] char*, or memcpy it  into a large enough buffer of char's.

这篇关于联合代码生成，用于字节序交换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！