Bug: Compile-time and runtime-polymorphism do not work together

[hsloot]

Problem

Some of boost's internal functions are fully specialized for dqrng::random_64bit_generator; see below. However, this specialization works for dqrng::random_64bit_generator only and not for any of its derived classes. See also https://stackoverflow.com/a/27988380.

dqrng/inst/include/dqrng_distribution.h

Lines 41 to 63 in f3a8002

	namespace boost {
	namespace random {
	namespace detail {
	template<>
	inline std::pair<double, int> generate_int_float_pair<double, 8, dqrng::random_64bit_generator>(dqrng::random_64bit_generator& eng)
	{
	dqrng::random_64bit_generator::result_type x = eng();
	double r = dqrng::uniform01(x);
	// shift x due to weakness of lowest bits for xoshiro/xoroshiro with used "+" scrambler
	int bucket = (x >> 3) & 0xFF;
	return std::make_pair(r, bucket);
	}
	template<>
	inline double generate_uniform_real<dqrng::random_64bit_generator, double>(dqrng::random_64bit_generator& eng, double min, double max)
	{
	return dqrng::uniform01(eng()) * (max - min) + min;
	}
	} // namespace detail
	} // namespace random
	} // namespace boost

Possible solutions

I do not know if there is a natural fix as the original template definitions are part of the boost headers. Other than that, either all derived classes of dqrng::random_64bit_generator should have the same specialization or only dereferenced pointer-to-dqrng::random_64bit_generator should be passed into any distribution.

[rstub]

For now it looks like adding the same specialization is the way to go. In addition, I should also investigate to what extend these specialization are actually needed, i.e. if they still give a performance advantage.