What happened to the profile?
Linus Torvalds
torvalds at linux-foundation.org
Mon Oct 7 12:28:25 UTC 2013
On Mon, Oct 7, 2013 at 12:01 PM, Linus Torvalds
<torvalds at linux-foundation.org> wrote:
>
> Yeah, we definitely don't need it, I was just thinking that if
> somebody wants to get full-precision double values, you probably can
> do it without the intricacies of knowing exactly what the FP format
> is.
Just checked.
Aside from actual overflow/underflow and NaN, my stupid thing seems to
be within 1 ULP of "correct" (ie system strtod()). I can find several
cases where that ascii_strtod() gives the wrong answer in the low bit,
and while I couldn't find any worse error, I can well imagine it being
more than 1.5 ULP off.
With "long double" I can't reproduce that any more, and I don't think
it can happen, since there's an extra 11 bits of mantissa precision in
a 80-bit FP value.
But as mentioned, we don't care about 0.5 ULP or even 1.5 ULP - we're
at least 12 bits off the 'double' precision.
Linus
More information about the subsurface
mailing list