Re: Obtaining the bit pattern of a floating point number

[me22 <me22 dot ca at gmail dot com>]

2009/5/27 Michael Meissner <meissner@linux.vnet.ibm.com>:
> On Tue, May 26, 2009 at 09:42:40PM -0700, me22 wrote:
>>
>> In my experience compilers are often even smart enough to have this
>> avoid the memory roundtrip, and it's undeniably legal:
>>
>> int ftoi(float x) {
>> Â Â int i;
>> Â Â memcpy(&i, &x, 4);
>> Â Â return i;
>> }
>
> Well from a pedantic point of view, it is only legal if sizeof (i) == sizeof
> (x) == 4, but you presumably already know that.
>

Agreed; I assumed we were operating under the ever-popular "there's
nothing but windows and linux on x86 or x64" delusion.  Of course, the
union and other tricks are also architecture-dependant.

If I really wanted to do it properly, I'd use something like this:

template <typename F>
typename boost::uint_t<sizeof(F)*CHAR_BIT>::least
bitcast_to_integral(F f) {
    typename boost::uint_t<sizeof(F)*CHAR_BIT>::least i = 0;
    void *p = &i;
#ifdef BOOST_BIG_ENDIAN
    p = (char*)p + (sizeof(i)-sizeof(f));
#endif
    std::memcpy(&i, &f, sizeof(F));
    return i;
}

But really, anything hacking around in float bit patterns is unlikely
to be portable enough to warrant such treatment.