Valgrind and lazy exact numbers

[stla]

This is a follow-up of my monologues in the other posts.

I noticed that when using Valgrind, everything happens as if EPECK were EPICK.

So I did the following test. In double precision, the equality 1 - 7*0.1 == 0.3 is false.

typedef CGAL::Lazy_exact_nt<CGAL::MP_Float> NT;

bool test() {
  NT one(1);
  NT seven(7);
  NT zeroone(0.1);
  NT x = one - seven * zeroone;
  double y = CGAL::to_double<NT>(x);
  return y == 0.3;
}

Results:

• without Valgrind, test() is true;
• with Valgrind, test() is false.

[lrineau]

Hi Stéphane,

I am really sorry for all the issues you have with CGAL and the test system of CRAN. I think all of those issues (including #6994), are because Valgrind does not emulate correctly the setting of the FPU rounding mode. That is documented, in the section Limitations of its manual:

For all other FP operations, only the IEEE default mode (round to nearest) is supported.

Unfortunately, the class template CGAL::Interval_nt<P> needs to set the rounding mode to +infinity, and that class template is at the core of the correct behavior of both CGAL::Epick and CGAL::Epeck. Under Valgrind, CGAL cannot work correctly, with the default settings.

Fortunately, there is an experimental feature of CGAL that you could use, that exists for a year in CGAL (see PR #5510) but is not well tested. If the macro CGAL_ALWAYS_ROUND_TO_NEAREST is defined, then CGAL::Interval_nt<P> is implemented differently, only using the default IEEE754 rounding mode (round to nearest). You should try to define that macro globally in your code (maybe directly in the build system), and then check if your own tests still pass. Then probably the CRAN test suite will pass more easily.

[stla]

Thanks! I will try.

[stla]

Nope :-( Unfortunately, the test above returns false when I define this macro.

I explained to CRAN that Valgrind does not support the CGAL exact arithmetic, I'm waiting for their reply.