[9fans] endiness and IEEE floats

[9fans] endiness and IEEE floats

[Steve Simon]

I am reading data files, they are always little endian
(Microsoft) but they include IEEE floats. How do I read them
portably?

This is what I came up with, It assumes the endianess
of doubles is the same as vlongs, which I think is fair...

double
gf64(Biobuf *b)					// get 64 bit IEEE double
{
	uvlong n;
	double d;
	n = gi64(b);				// get 64 bit int
	memcpy(&d, &n, sizeof(n));
	return d;
}

anyone got somthing better?

-Steve

Re: [9fans] endiness and IEEE floats

[Nigel Roles]

Steve Simon wrote:

>I am reading data files, they are always little endian
>(Microsoft) but they include IEEE floats. How do I read them
>portably?
>
>This is what I came up with, It assumes the endianess
>of doubles is the same as vlongs, which I think is fair...
>
>double
>gf64(Biobuf *b)					// get 64 bit IEEE double
>{
>	uvlong n;
>	double d;
>	n = gi64(b);				// get 64 bit int
>	memcpy(&d, &n, sizeof(n));
>	return d;
>}
>
>anyone got somthing better?
>
>-Steve
>
>  
>
if you are going to assume size and endianness matches, then

#define bf64(b) ((double)gi64(b))

would suffice.

Re: [9fans] endiness and IEEE floats

[Russ Cox]

> I am reading data files, they are always little endian
> (Microsoft) but they include IEEE floats. How do I read them
> portably?
> 
> This is what I came up with, It assumes the endianess
> of doubles is the same as vlongs, which I think is fair...
> 
> double
> gf64(Biobuf *b)                                 // get 64 bit IEEE double
> {
>         uvlong n;
>         double d;
>         n = gi64(b);                            // get 64 bit int
>         memcpy(&d, &n, sizeof(n));
>         return d;
> }
> 
> anyone got somthing better?

That's what I do (well, different problem but same sort of solution)
to implement isNaN etc. "portably" in libfmt.  You could also fiddle
with the bytes, but that requires detecting the byte order.

gf64(Biobuf *b)
{
    uchar buf[8];
    static uchar big1[4] = { 0, 0, 0, 1 };

    Bread(b, buf, 8);
    if(*(ulong*)big1 == 1)   /* running on big endian machine */
        reverse(buf, 8);
    return *(double*)buf;
}

Ken came up with clever way to do doing integer conversion:

gi32(Biobuf *b)
{
    int i;
    u64int leorder = 0x0706050403020100, to;
    uchar lebuf[8];
    uchar *o, *t, *f;

    Bread(b, lebuf, 8);
    o = (uchar*)&leorder;
    f = lebuf;
    t = (uchar*)&to;
    for(i=0; i<8; i++)
        t[o[i]] = f[i];
}

The kernel and user space use this trick to convert to and
from the big-endian 64-bit numbers in /dev/bintime.
When the 64-bit support was worse, it was a huge win.
It's probably still a win but less so.

Russ

Re: [9fans] endiness and IEEE floats

[Russ Cox]

> if you are going to assume size and endianness matches, then
> 
> #define gf64(b) ((double)gi64(b))
> 
> would suffice.

that would be true if gf64 and gi64 were returning pointers,
but they are returning the actual values.

#define bf64(b) (*(double*)&gi64(b))

would work except that you can't take the address
of the return value.