Assertion error in eval_exp (transcendental.hpp)

[kedarbhat]

Since I can't delete this blank comment, I'll use the space.

The error below is with latest Boost 1.70. I'm not sure yet if this occurs in earlier versions.

[kedarbhat]

I think this may be similar/related to issue #23. The below code (using boost::multiprecision::cpp_bin_float_50) fails an invariant in eval_exp (in transcendental.hpp). The computation using double as the type is included to verify it's a valid computation. This issue was discovered when boost::math::zeta was specialized with boost::multiprecision::cpp_bin_float_50, but this issue can be reproduced simply using the boost::multiprecision::exp call with the problematic argument.

BOOST_AUTO_TEST_CASE(multiprecision_zeta) {
  using T = boost::multiprecision::cpp_bin_float_50;
  static constexpr auto x = 2.0;
  static constexpr auto y = 95.0;
  T x_mp(x);
  T y_mp(y);
  auto answer = std::exp(y*std::log(x));
  BOOST_REQUIRE_CLOSE(answer, 3.9614081257131865e28, std::numeric_limits<T>::epsilon());
  auto answer_mp = boost::multiprecision::exp(y_mp*boost::multiprecision::log(x_mp));
  BOOST_REQUIRE_CLOSE(answer, answer_mp, std::numeric_limits<T>::epsilon());
}

[jzmaddock]

Fixed in develop.

[kedarbhat]

Thanks for the quick turnaround!