[Qemu-devel] Re: [PATCH] softfloat: fix floatx80_is_{quiet, signaling}_nan()

[Aurelien Jarno <aurelien@aurel32.net>]

On Wed, Jan 12, 2011 at 03:05:10PM -0600, Peter Maydell wrote:
> On 12 January 2011 13:59, Aurelien Jarno <aurelien@aurel32.net> wrote:
> 
> > @@ -494,7 +495,8 @@ int floatx80_is_quiet_nan( floatx80 a )
> >  int floatx80_is_signaling_nan( floatx80 a )
> >  {
> >  #if SNAN_BIT_IS_ONE
> > -    return ( ( a.high & 0x7FFF ) == 0x7FFF ) && (bits64) ( a.low<<1 );
> > +    return ( ( a.high & 0x7FFF ) == 0x7FFF )
> > +        && (LIT64( 0x8000000000000000 ) >= ((bits64) ( a.low<<1 )));
> >  #else
> >     bits64 aLow;
> 
> If a is {0x7ffff,0} (ie +inf) this will return true, which is wrong.
> Do you want "<=" instead?

Correct, I swapped the operands at the last minute to match the other
functions, but without changing the sign.

> Actually, will
>   return ((a.high & 0x7fff) == 0x7fff) && (a.low >= LIT64(0x4000000000000000));
> do? Untested but I think it will do the right thing. I'm not sure

The explicit bit might be one for a NaN, so you should filter it first.

> why this code has those bit64 casts, incidentally, since a.low is
> already a uint64_t.

Don't know either, but as they were already there, I left them.

> Also, maybe we should have a comment somewhere explaining
> why this is different from the other NaN functions (ie that the
> x80 format has an explicit bit and the others don't) ?
> 

Good idea, will do.

-- 
Aurelien Jarno                          GPG: 1024D/F1BCDB73
aurelien@aurel32.net                 http://www.aurel32.net