Re: [commit-cp] [bugs #9895] double to int conversion problem

[Robert Lougher]

Hi,

Yes, I do have a fix for this and a couple of other minor issues. 
I've been delaying a 1.1.5 release because of an impending (i.e. for
the last 2 weeks) test 1.2.0 release which includes a rewritten
interpreter.  I've got the last couple of bugs out and it's giving a
speed increase of between 30-50% (on PIII).  On some very unrealistic
microbenchmarks it's 2.5 times faster!

The reason for the delay is that I wanted two streams : a stable 1.1.5
and a test 1.2.0.  However if I release 1.1.5 before 1.2.0 I didn't
want people thinking it supercedes 1.1.5...

Mark, are you interested in testing 1.2.0?  I've only got access to a
PIII (via sourceforge compile-farm).  So it'll be useful if it can be
tried on a larger range of IA32 chips (e.g. AMD/PIV).

Anyway, baring disasters 1.1.5 and 1.2.0-test will be out Fri/Saturday :)
 
Rob.

On Thu, 05 Aug 2004 18:42:21 +0200, Mark Wielaard <address@hidden> wrote:
> Hi,
> 
> On Thu, 2004-08-05 at 17:05, Archie Cobbs wrote:
> > David Gilbert wrote:
> > > public class DoubleToIntTest {
> > >     public static void main(String args[]) {
> > >         double v = Integer.MAX_VALUE + 100000.0;
> > >         System.out.println((int) v);
> > >         Double d = new Double(v);
> > >         System.out.println(d.intValue());
> > >     }
> > > }
> > >
> > > With Classpath (fairly recent CVS) running on JamVM 1.1.4 (on SuSE 9.1), 
> > > I see the following output:
> > >
> > > -2147483648
> > > -2147483648
> > >
> > > Using Sun's JDK 1.4, the output is:
> > >
> > > 2147483647
> > > 2147483647
> >
> > I think this is a VM bug rather than a Classpath bug (JC has the same
> > bug). It results from doing the jdouble -> jint conversion in C via
> > a `(jint)d' cast.
> >
> > JC generates this code (edited)...
> >
> >     jdouble d2;
> >     jint i1;
> >
> >     d2 = _JC_DCONST(0x41, 0xe0, 0x0, 0x30, 0xd3, 0xe0, 0x0, 0x0);
> >     ...
> >     i1 = (jint)d2;
> >     ...
> >
> > I'm still curious what the solution is though :-)
> 
> I reported a similar issue against jamvm and Robert created a patch for
> it. But I see there is no new jamvm (1.1.5) yet.
> 
> Cheers,
> 
> Mark
> 
> 
>