Add BOOST_CXX14_CONSTEXPR to CV assign()

[rogerorr]

Without this change the BOOST_CXX14_CONSTEXPR on the constructors has no
effect, as they invoke assign()

[JeffGarland]

Hi Roger -- I did try this out and I was unable to get this to resolve at compile time with gcc at least. I believe that is a result of the exception paths that are used by the higher order objects for handling out of range values. I think to make this really work there needs to be if constexpr logic to avoid exceptions in the compile time case -- which is a bit more involved. So I can merge the change, but I don't think it's having the impact you're looking for.

[rogerorr]

That's odd. The execption paths shouldn't matter when being evaluated in a constexpr context (as long as they're not taken of course.)

My trivial test case:

#include <boost/date_time.hpp>

static constexpr boost::posix_time::ptime UNIX_EPOCH(boost::gregorian::date(1970, 1, 1));

Using g++ -std=c++17 trivial_test.cpp

Without this change:

constrained_value.hpp:48:13: error: call to non-'constexpr' function 'void boost::CV::constrained_value<value_policies>::assign(boost::CV::constrained_value<value_policies>::value_type) [with value_policies = boost::CV::simple_exception_policy<short unsigned int, 1400, 9999, boost::gregorian::bad_year>; boost::CV::constrained_value<value_policies>::value_type = short unsigned int]'
   48 |       assign(value);
      |       ~~~~~~^~~~~~~

With this change it compiles successfully.

(I'm using gcc trunk from 20200601)

[mclow]

What Roger said (about constexpr and exceptions).
I wrote about this on SO: https://stackoverflow.com/questions/33716507/how-can-this-code-be-constexpr-stdchrono

[JeffGarland]

Ok interesting - today I learned again :)

Looking back on my changes I was testing a bit more directly.

  //test_constrained_value.cpp

 class day_value_policies
 {
  public:
    typedef unsigned int value_type;
    static unsigned int min BOOST_PREVENT_MACRO_SUBSTITUTION () { return 0; };
    static unsigned int max BOOST_PREVENT_MACRO_SUBSTITUTION () { return 31;};
    static void on_error(unsigned int&, unsigned int, boost::CV::violation_enum)
    {
      throw bad_day();
   }
  };
...
  //check constexpr case
  constexpr constrained_value<day_value_policies> cv3(1);
  static_assert(cv3 == 1, "constexpr construction/conversion");
  check("constexpr constrained value construct and equal", true);

And now that I'm looking again I see the compiler is telling me exactly the problem here: min and max, which are executed need to be constexpr. So looking at greg_month and how this is used upstream it indeed looks like it will indeed work.

[rogerorr]

Good -- that all makes sense. And you've already learned something today so you're ahead :)

[JeffGarland]

lol - c++ is a life long learning project :)