Re: PING: Re: [cp-patches] RFC: strtod patch from early 2004

[Mark Wielaard]

Hi,

On Wed, 2005-08-03 at 11:41 +0100, Andrew John Hughes wrote:
> On Sun, 2005-07-24 at 16:41 +0100, Andrew John Hughes wrote:
> > I've recently moved to a 64-bit architecture (x86_64) and yesterday I
> > encountered some problems building the generics branch with ecj on kaffe
> > and Debian GNU/Linux.  Namely, the MIN_DOUBLE value is not recognised as
> > valid.  Using parseDouble on this value shows that the fdlibm
> > implementation of strtod reports an underflow and 0 is returned.
> > 
> > This lead to me finding the following post:
> > 
> > http://lists.gnu.org/archive/html/classpath/2004-01/msg00057.html
> > 
> > Applying this patch solves the problem and the right value is returned.
> > Does anyone know what the state of this patch is, and why it wasn't
> > applied?  It does solve the problem in this case, but I don't know how
> > it fares in general.  Any comments, and testing on some other platforms
> > would be much appreciated.
> > 
> > FWIW, the conversion also works with the platform strtod, so this is
> > also an option (although again I don't know how general this is, and it
> > has been disabled for cases where KISSME_LINUX_USER is not defined).

The KISSME_LINUX_USER hack should be removed from our reference
implementation. If kissme needs a different VMDouble implementation then
it should provide one. I committed the following:

2005-08-24  Mark Wielaard  <address@hidden>

    * native/jni/java-lang/java_lang_VMDouble.c
    (Java_java_lang_VMDouble_parseDouble): Remove KISSME_LINUX_USER hack.

I don't think strtod can be used in general since it isn't sure it
returns the wanted double on all platforms in all cases.

> Is there any update on this?

I believe everybody is too afraid of 64bit issues and fdlibm so nobody
feels like they can OK it. FWIW This is also the following bug:
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22844

What would help is someone running mauve before and after this patch and
adding the results to that bug report.

Cheers,

Mark

--- native/jni/java-lang/java_lang_VMDouble.c   2 Jul 2005 20:32:55 -0000      
1.10
+++ native/jni/java-lang/java_lang_VMDouble.c   24 Aug 2005 14:17:34 -0000
@@ -348,12 +348,7 @@
          struct _Jv_reent reent;
          memset (&reent, 0, sizeof reent);

-#ifdef KISSME_LINUX_USER
-         /* FIXME: The libc strtod may not be reliable. */
-         val = strtod (p, &endptr);
-#else
          val = _strtod_r (&reent, p, &endptr);
-#endif

 #ifdef DEBUG
          fprintf (stderr, "java.lang.VMDouble.parseDouble val = %g\n", val);

signature.asc
Description: This is a digitally signed message part