Re: [boost] Portable signbit macro/function?

From: Guillaume Melquiond (guillaume.melquiond_at_[hidden])
Date: 2006-08-03 10:15:42

Le jeudi 03 août 2006 à 13:37 +0200, Johan Råde a écrit :

> The following function does bit testing, without advance knowledge of 
> which bit is the sign bit. The function figures out on its own, at 
> compile time, which bit is the sign bit. Hence you do not have to worry 
> about IEEE 754 and endianness.

Actually, you do have to worry about IEEE-754. Without it, you have no
guarantee that the number representation has the correct properties. For
example, with floating-point numbers that rely on a two-complement
representation of the mantissa, you will get that +7 is negative (since
at least one of its bits will be covered by sign_mask).

> The only thing that is required is that 
> sizeof(float) == sizeof(int).
> 
>      bool signbit(float x)
>      {
>          const float one = 1;
>          const float neg_one = -1;
>          const int sign_mask = reinterpret_cast<const int&>(one)
>             ^ reinterpret_cast<const int&>(neg_one);
>          return reinterpret_cast<int&>(x) & sign_mask;
>      }

You are invoking undefined behavior here: you are accessing a float
through a reference to an int. As a consequence, GCC produces code that
accesses uninitialized memory (since one and neg_one are optimized away)
and returns random values. You have to use char* pointers and memcopies
so that there is no aliasing issue.

Best regards,

Guillaume