Re: 3.1.52 fails to build in hppa/linux

[Rafael Laboissiere]

* Jaroslav Hajek <address@hidden> [2009-02-22 21:21]:

> I don't recokn this is the reason. I bet that on the amd64,
> OCTAVE_INT_USE_LONG_DOUBLE is defined, unlike on mipsel. Isn't that
> so?

Yes, I think so.  I do not have the full Octave sources on each machine, but
looking from the code in configure.in:

#if (SIZEOF_LONG_DOUBLE >= 10) && defined (HAVE_ROUNDL)
#define OCTAVE_INT_USE_LONG_DOUBLE
#endif

and looking from the build log for the Debian octave3.1 package, version
3.1.52-4 [1]:

* mipsel:
  checking for roundl... yes
  checking for long double... yes
  checking size of long double... 8

* i386:
  checking for roundl... yes
  checking for long double... yes
  checking size of long double... 12

* hppa:
  checking for roundl... yes
  checking for long double... yes
  checking size of long double... 8

* sparc:
  checking for roundl... yes
  checking for long double... yes
  checking size of long double... 16

* powerpc
  checking for roundl... yes
  checking for long double... yes
  checking size of long double... 16

[1] http://experimental.debian.net/build.php?pkg=octave3.1
  
On amd64 the size of long double is 16.  This means that compilation of
oct-inttype.cc fails on all architectures where OCTAVE_INT_USE_LONG_DOUBLE
is undefined (mipsel and hppa) and succeed on the others (amd64, i386,
sparc, and powerpc).  Note that the compilation on sparc fails for another
reason (procstream.cc:32976802: error: expected unqualified-id).

I would guess that the problem lies in the Octave sources, whenever
OCTAVE_INT_USE_LONG_DOUBLE in undefined.  It does not seem to be a gcc 4.3.3
bug.  BTW, I just tried to undef OCTAVE_INT_USE_LONG_DOUBLE in config.h and
I can replicate the "ambiguous template specialization" compilation failure
on amd64.

-- 
Rafael