ECL? was ..Re: Runtime determination of share directories?

>>>>> "Paul" == Paul Zimmermann <zimmerma at loria.fr> writes:

    Paul>        Ray,
    Paul> [it's quite an interesting discussion for me]

    Paul>    I'm not sure, but I think there will always be corner cases where you
    Paul>    will get incorrect results on an x87.  It's unavoidable due to the
    Paul>    80-bit registers.  Even if you set the rounding and precision modes
    Paul>    correctly.  I think the cases are when a number would underflow to 0 if
    Paul>    you had 64-bit doubles but doesn't when you have 80-bit long floats.

    Paul> please can you give an example? I thought quaddouble was using algorithms
    Paul> like TwoSum of TwoProduct which guarantee exact operations, without guard
    Paul> bits.

I don't have an example.  I was just thinking of the case of
underflow, and how it could potentially cause quaddouble to produce a
result that was off by one bit because the underflow to zero didn't
happen due to the large exponent range.  I'll have to think about it
some more and try some experiments.

One annoyance I do have with quaddouble is that products overflow too
soon because of the way TwoProduct splits the number.

But this is now getting *way* off-topic for Maxima....

Ray