Re: Byteswapping floating point types

[ast at arlut dot utexas dot edu]

> Andrew Troschinetz <ast@arlut.utexas.edu> writes:
>
>> template<class T>
>> T byteswap(T value)
>> {
>>   T result;
>>   size_t n = sizeof (T);
>>   unsigned char* p = (unsigned char*) &value;
>>   for (size_t i = 0; i < n; ++i)
>>     ((unsigned char*) &result)[i] = p[n - 1 - i];
>>   return result;
>> }
>
> I would avoid aliasing issues and the problem you describe and write the
> code like this (untested):
>
> template<typename T>
> T byteswap(T value)
> {
>   char buf1[sizeof(T)], buf2[sizeof(T)];
>   memcpy(buf1, &value, sizeof(T));
>   for (size_t i = 0; i < sizeof(T); ++i)
>     buf2[i] = buf1[sizeof(T) - 1 - i];
>   T ret;
>   memcpy(&ret, buf2, sizeof(T));
>   return ret;
> }

Tested here: http://codepad.org/JGKW3ir0

Unfortunately this code does not avoid the issue. Since the function still
returns a T, when T is a floating point type you are putting a byteswapped
floating point value into a floating point type. At which point I contend
that you risk the data being modified. My co-worker maintains that the
risk is only present on broken compilers and when people try to do things
to swapped floats that they shouldn't be doing (ie: he passes the blame on
to either the compiler or the programmer, not to his byteswap() function).

Within that code paste I present a solution which sidesteps the issue by
returning a sequence of bytes, so that a byteswapped floating point value
never has the chance to go into a FPU register. However the solution I
present is significantly less convenient to use than my co-workers so I
think we're both hoping that he's right and I'm wrong. (Scratch that I
totally want to be proven right on this :)

> On some processors this can convert a valid floating point
> representation into a trap representation which will cause a floating
> point exception if the value is used.  Presumably that is not an issue
> for you.

I think it could be an issue for us (if not now then maybe in the future),
which is why I've raised concern over byteswap() as my co-worker has
implemented it.

--
Andrew Troschinetz
Applied Research Laboratories