Re: significance of double test?

[Peter Naulls <peter at erble dot freeserve dot co dot uk>]

In message <20000313172024.6130.qmail@pasanda.cygnus.co.uk>
          Andrew Haley <aph@pasanda.cygnus.co.uk> wrote:

> > Peter>       // Some broken implementations convert "7.79625120912E289" to
> > Peter>       // the value 7.796251209119999E289.
> > Peter>       harness.check (new Double("7.79625120912E289").toString (),
> > Peter>         "7.79625120912E289");
> > 
> > Peter> Can someone tell me the significance of this test?  i.e., what's
> > Peter> the precise behaviour that's likely to result in the wrong value?
> > 
> > I don't know, but I think Andrew probably does.  I CC'd him since I
> > don't know if he is on mauve-discuss.
> 
> It's a test to ensure that the number conversion routines do exact
> IEEE rounding.  This is really difficult to get right, and I
> discovered several problems of this kind in a well-known
> manufacturer's JVM.
> 
>     .79625120912E289 ~= 2^963
> 
> What result does your JVM give?

Ah, perhaps you missed some of my post from quoting - my JVM exhibits
the exact undesirsable behaviour in the comment you can see above
before this text - which is really why I made this query.

My JVM uses JDK 1.1.8 libraries, so if the problem's in that, then
there's perhaps not much I can do, OTOH, if it's a C/C library issue,
then I can possibly work around it (and even get the OS developers to
fix it for the future).  I'm not sure if I've looked very hard at
the matter - e.g. examining the value from atod(), which is called in
a native method the Double class calls.

Peter

-- 
+---------------------------------------------+-----------------------------+
| Peter Naulls - pnaulls@usa.net              |                             |
| http://free.prohosting.com/~chocky/         | Java and JVM Consultant     |
| Java for RISC OS and ARM - java-arm@usa.net | Technical Author            |
| http://free.prohosting.com/~chocky/java/    | Program performance analyst |
+---------------------------------------------+-----------------------------+