CGAL users discussion list

[Marc Glisse <>]

On Tue, 12 Mar 2013, Marius Kintel wrote:

> On 2013-03-12, at 12:39 , marcosscriven wrote:
>
> > This turned out to be an LLVM optimsation issue. With no optimisation at the
> > LLVM level, it works fine, but with -O1, I get this assertion error.
> I've seen such issues before, on Mac, LLVM, optimizations. It had something 
> to do with rounding modes, but it was fixed in CGAL-4.1 IIRC.

On Tue, 12 Mar 2013, marcosscriven wrote:

> Thanks for the input - I am using CGAL 4.1 (with GMP 5.1.1, MPFR 3.1.1, and
> Boost 1.53). Perhaps there are some remaining issues.

We already go out of our way to prevent LLVM from breaking the code. In 
CGAL/FPU.h,

#define CGAL_IA_ADD(a,b) CGAL_IA_FORCE_TO_DOUBLE((a)+CGAL_IA_STOP_CPROP(b))

could still be made stronger by adding CGAL_IA_STOP_CPROP in front of (a), 
and similarly for the other operations, but that's it.

There may be other places than intervals in CGAL that rely on some 
behavior of floats, and there are probably a number of other things than 
floats that optimizers may mess with.

--
Marc Glisse