promotions

[gpeterhoff]

Hello, boost/math/toolts/promotion.hpp is corrupt. Unfortunately, I can't say whether some promotions should actually work like this or whether there is an error. Reasons:

• promotions are implemented manually instead of using std::type_traits. This is highly confusing and error-prone.
• No variadic templates are used but (incorrectly) reproduced manually, see 1.).

boost::math needs C++14 anyway - so I don't understand why legacy C++98/03 is still used.

In order to be able to better illustrate the problems, the behavior of the std::math functions is briefly shown again:

• one argument, e.g. std::sin: std::sin(Float) -> Float, std::sin(Integral) -> double
• multiple arguments + with C++14 the specializations foo(Float, int) have been removed, e.g. std::pow: std::pow(Float, Float) -> Float std::pow(Integral, Integral) -> double std::pow(Float1, Float2) -> std::common_type_t<Float1, Float2> std::pow(Float, Integral) + std::pow(Integral, Float) -> std::common_type_t<Float, double>

In boost/math/toolts/promotion.hpp there are 3 promotions:

• promote_arg
• promote_args
• promote_args_2

fundamental problem

• std::bfloat16_t is not supported

In detail

• promote_arg Works as it should, but is unnecessarily bloated with manual specializations.

• promote_args By 1.) neither std::float16_t nor std::bfloat16_t are fully supported.

• promote_args_2 Here I'm not sure what the purpose is as the implementations for std::float128_t/std::float64_t/std::float32_t/std::float16_t (std::bfloat16_t is missing as said) and long double/double/float differ. In addition, it is highly confusing/prone to errors. Mistake: promote_args_2<int, float>/promote_args_2<float, int> -> double (default behavior) promote_args_2<int, std::float32_t>/promote_args_2<std::float32_t, int> -> std::float32_t

  There are 2 possible solutions: a) promote_args_2<Type1, Type2> behaves exactly like promote_args<Type1, Type2> -> would be superfluous b) If actually needed: promote_args_2<Float1, Float2> -> std::common_type_t<Float1, Float2> promote_args_2<Integral, Integral> -> double promote_args_2<Float, Integral> -> Float promote_args_2<Integral, Float> -> Float

I've implemented a few cases promotion.hpp.txt

• v1 correct promote_arg(s) promote_args_2 like a)
• v2 correct promote_arg(s) promote_args_2 like b)
• v3 correct promote_arg(s) promote_args_2 like a) conversion std::floatN_t -> boost::floatN_t This might be necessary for compatibility reasons.
• v4 correct promote_arg(s) promote_args_2 like a) conversion boost::floatN_t -> std::floatN_t This has the advantage that the corresponding math functions are (partially) constexpr, which of course is not the case with __float128 (C library).
• v5 like v4 but promote_args_2 like b)

Of course, I think v4/v5 makes the most sense.

These measures can probably remove promote_args_permissive as well. In general, boost/math/special_functions already included in C++14 can be removed.

1.) template <class T1, class T2=float, class T3=float, class T4=float, class T5=float, class T6=float> struct promote_args

thx Gero

[mborland]

There is no specialization for std::bfloat16_t because what is the reasonable conversion? I just ran the tests from test_beta_dist with std::bfloat16_t on GCC-13, and they run but are inaccurate as expected. There is discussion at the top of the PR adding support as to our design decisions. PRs with your suggestions are always welcome.