__float128 / gcc

[tichy]

Hi,

funny thing - just discovered that gcc supports 128-bit IEEE 754 floating type. Probably it is not a speed monster, but maybe someone will find it useful.

#include <stdio.h>
#include <stdint.h>

#include <endian.h>

typedef union
{
    __float128  val;
    struct
    {
#if __BYTE_ORDER == __BIG_ENDIAN
        uint64_t sign    : 1;
        uint64_t exp     : 15;
        uint64_t frac1   : 48;
        uint64_t frac0   : 64;
#else
        uint64_t frac0   : 64;
        uint64_t frac1   : 48;
        uint64_t exp     : 15;
        uint64_t sign    : 1;
#endif
    } bits;
} qfloat;

int main()
{
    qfloat      foo;
    qfloat      bar;

    foo.val = 65536.0;

    foo.val += 1.0;
    foo.val *= foo.val;

    bar.val = 8191.0;

    printf("%04X %012llX %016llX\n", (uint16_t)foo.bits.exp, (uint64_t)foo.bits.frac1, foo.bits.frac0);

    return 0;
}

[ewmayer]

I just tried building my first program using __float128. Most everything builds OK, but I am getting compile-time warnings (have not tried to link yet, since these tell me I'm missing something) which appear to indicate a header file is needed:

warning: implicit declaration of function ‘fabsq’

The math-lib functions are enumerated here, but no mention of "you must include <blah.h>".

[Random Poster]

Originally Posted by ewmayer

I just tried building my first program using __float128. Most everything builds OK, but I am getting compile-time warnings (have not tried to link yet, since these tell me I'm missing something) which appear to indicate a header file is needed:

warning: implicit declaration of function ‘fabsq’

The math-lib functions are enumerated here, but no mention of "you must include <blah.h>".

The examples for the two functions listed under 3 I/O Library Routines do "#include <quadmath.h>". Of course you could also have found it by looking at the list of files that come with libquadmath, or by "fgrep fabsq /usr/include/*.h".

[ewmayer]

Originally Posted by Random Poster

The examples for the two functions listed under 3 I/O Library Routines do "#include <quadmath.h>". Of course you could also have found it by looking at the list of files that come with libquadmath, or by "fgrep fabsq /usr/include/*.h".

Thanks - that is some poorly constructed documentation there, burying absolutely key "what you need to build this" stuff deep inside an I/O-related example.

But I fail to find quadmath.h anywhere on my system (I used 'locate' instead of 'find', since that is much faster, after the one-time expense of building the locate DB). I find other headers I know are there using 'locate', so the utility is working properly. Note I'm using gcc 4.2, which does recognize the __float128 type - do I need something more recent than that?

[Random Poster]

Originally Posted by ewmayer

Thanks - that is some poorly constructed documentation there, burying absolutely key "what you need to build this" stuff deep inside an I/O-related example.

The documentation looks like a first draft; each function listed should have its own description page. I expect every one of those will mention quadmath.h once they are written.

Originally Posted by ewmayer

But I fail to find quadmath.h anywhere on my system (I used 'locate' instead of 'find', since that is much faster, after the one-time expense of building the locate DB). I find other headers I know are there using 'locate', so the utility is working properly. Note I'm using gcc 4.2, which does recognize the __float128 type - do I need something more recent than that?

gcc may provide the data type, but not the functions that use it; you need to install libquadmath for those (I assumed you had already done so, since you knew about its documentation).