Introduce concepts for AccessTraits

[aprokop]

The patch introduces three concepts: AccessTraits, Primitives (same as AccessTraits) and Predicates (AccessTraits + tag). I put them into Concepts namespace.

In general, it works fine. One corner point is how it interacts with CTAD. clang evaluates the right hand side of CTAD before testing the concept, which leads to suboptimal errors instead of ignoring the CTAD (see godbolt). GCC, on the other hand, has a better treatment and properly rejects CTAD due to the unsatisfied requirement.

[aprokop]

I added this because there does not seem to be a non-void concept in std, and !std::same_as<void> did not work.

[aprokop]

@dalg24 Could you please take a look at this PR?

[aprokop]

Moved 'Concepts' namespace inside Details to hide them from users.

[aprokop]

Here are what the errors look like for the tstCompileOnlyAccessTraits.cpp:

In LLVM:

<snip>/test/tstCompileOnlyAccessTraits.cpp:137:7: error: static assertion failed
  137 |       ArborX::Details::Concepts::AccessTraits<NoAccessTraitsSpecialization>);
      |       ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
<snip>/test/tstCompileOnlyAccessTraits.cpp:137:7: note: because 'NoAccessTraitsSpecialization' does not satisfy 'AccessTraits'
<snip>/src/spatial/detail/ArborX_AccessTraits.hpp:82:37: note: because 'typename ArborX::AccessTraits<T>::memory_space' would be invalid: no type named 'memory_space' in 'ArborX::AccessTraits<NoAccessTraitsSpecialization>'
   82 |   typename ArborX::AccessTraits<T>::memory_space;

<snip>/test/tstCompileOnlyAccessTraits.cpp:138:17: error: static assertion failed
  138 |   static_assert(ArborX::Details::Concepts::AccessTraits<EmptySpecialization>);
      |                 ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
<snip>/test/tstCompileOnlyAccessTraits.cpp:138:17: note: because 'EmptySpecialization' does not satisfy 'AccessTraits'
<snip>/src/spatial/detail/ArborX_AccessTraits.hpp:82:37: note: because 'typename ArborX::AccessTraits<T>::memory_space' would be invalid: no type named 'memory_space' in 'ArborX::AccessTraits<EmptySpecialization>'
   82 |   typename ArborX::AccessTraits<T>::memory_space;

<snip>/test/tstCompileOnlyAccessTraits.cpp:139:17: error: static assertion failed
  139 |   static_assert(ArborX::Details::Concepts::AccessTraits<InvalidMemorySpace>);
      |                 ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
<snip>/test/tstCompileOnlyAccessTraits.cpp:139:17: note: because 'InvalidMemorySpace' does not satisfy 'AccessTraits'
<snip>/src/spatial/detail/ArborX_AccessTraits.hpp:83:12: note: because 'Kokkos::is_memory_space_v<typename ArborX::AccessTraits<InvalidMemorySpace>::memory_space>' evaluated to false
   83 |   requires Kokkos::is_memory_space_v<

<snip>/test/tstCompileOnlyAccessTraits.cpp:141:7: error: static assertion failed
  141 |       ArborX::Details::Concepts::AccessTraits<MissingSizeMemberFunction>);
      |       ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
<snip>/test/tstCompileOnlyAccessTraits.cpp:141:7: note: because 'MissingSizeMemberFunction' does not satisfy 'AccessTraits'
<snip>/src/spatial/detail/ArborX_AccessTraits.hpp:87:22: note: because 'AccessTraits<T>::size(v)' would be invalid: no member named 'size' in 'ArborX::AccessTraits<MissingSizeMemberFunction>'
   87 |     AccessTraits<T>::size(v)

<snip>/test/tstCompileOnlyAccessTraits.cpp:143:7: error: static assertion failed
  143 |       ArborX::Details::Concepts::AccessTraits<SizeMemberFunctionNotStatic>);
      |       ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
<snip>/test/tstCompileOnlyAccessTraits.cpp:143:7: note: because 'SizeMemberFunctionNotStatic' does not satisfy 'AccessTraits'
<snip>/src/spatial/detail/ArborX_AccessTraits.hpp:87:22: note: because 'AccessTraits<T>::size(v)' would be invalid: call to non-static member function without an object argument
   87 |     AccessTraits<T>::size(v)

<snip>/test/tstCompileOnlyAccessTraits.cpp:145:7: error: static assertion failed
  145 |       ArborX::Details::Concepts::AccessTraits<MissingGetMemberFunction>);
      |       ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
<snip>/test/tstCompileOnlyAccessTraits.cpp:145:7: note: because 'MissingGetMemberFunction' does not satisfy 'AccessTraits'
<snip>/src/spatial/detail/ArborX_AccessTraits.hpp:91:20: note: because 'AccessTraits<T>::get(v, 0)' would be invalid: no member named 'get' in 'ArborX::AccessTraits<MissingGetMemberFunction>'
   91 |   AccessTraits<T>::get(v, 0);

<snip>/test/tstCompileOnlyAccessTraits.cpp:147:7: error: static assertion failed
  147 |       ArborX::Details::Concepts::AccessTraits<GetMemberFunctionNotStatic>);
      |       ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
<snip>/test/tstCompileOnlyAccessTraits.cpp:147:7: note: because 'GetMemberFunctionNotStatic' does not satisfy 'AccessTraits'
<snip>/src/spatial/detail/ArborX_AccessTraits.hpp:91:20: note: because 'AccessTraits<T>::get(v, 0)' would be invalid: call to non-static member function without an object argument
   91 |   AccessTraits<T>::get(v, 0);

<snip>/test/tstCompileOnlyAccessTraits.cpp:148:17: error: static assertion failed
  148 |   static_assert(ArborX::Details::Concepts::AccessTraits<GetMemberFunctionVoid>);
      |                 ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
<snip>/test/tstCompileOnlyAccessTraits.cpp:148:17: note: because 'GetMemberFunctionVoid' does not satisfy 'AccessTraits'
<snip>/src/spatial/detail/ArborX_AccessTraits.hpp:92:6: note: because '!requires (const GetMemberFunctionVoid &v) { { AccessTraits<GetMemberFunctionVoid>::get(v, 0) } -> std::same_as<void>; }' evaluated to false
   92 | } && !requires(T const &v) {

So I see two suboptimal messages:

1. If specialization is undefined, we don't explicitly say that
2. If specialization returns void, the error is a bit challenging to parse

[aprokop]

It seems that the details of notes seem to depend on whether static_assert does any computations. For example,
static_assert(ArborX::Details::Concepts::AccessTraits<NoAccessTraitsSpecialization>); fails with

/tstCompileOnlyAccessTraits.cpp.o -MF test/CMakeFiles/ArborX_Test_CompileOnly.exe.dir/tstCompileOnlyAccessTraits.cpp.o.d -o test/CMakeFiles/ArborX_Test_CompileOnly.exe.dir/tstCompileOnlyAccessTraits.cpp.o -c <snip>/test/tstCompileOnlyAccessTraits.cpp
<snip>/test/tstCompileOnlyAccessTraits.cpp:137:7: error: static assertion failed
  137 |       ArborX::Details::Concepts::AccessTraits<NoAccessTraitsSpecialization>);
      |       ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
<snip>/test/tstCompileOnlyAccessTraits.cpp:137:7: note: because 'NoAccessTraitsSpecialization' does not satisfy 'AccessTraits'
<snip>/src/spatial/detail/ArborX_AccessTraits.hpp:82:37: note: because 'typename ArborX::AccessTraits<T>::memory_space' would be invalid: no type named 'memory_space' in 'ArborX::AccessTraits<NoAccessTraitsSpecialization>'
   82 |   typename ArborX::AccessTraits<T>::memory_space;
      |                                     ^                                   ^

but
static_assert(!(!ArborX::Details::Concepts::AccessTraits<NoAccessTraitsSpecialization>)); fails simply with

<snip>/test/tstCompileOnlyAccessTraits.cpp:136:17: error: static assertion failed due to requirement '!(!ArborX::Details::Concepts::AccessTraits<NoAccessTraitsSpecialization>)'
  136 |   static_assert(!(
      |                 ^~
  137 |       !ArborX::Details::Concepts::AccessTraits<NoAccessTraitsSpecialization>));
      |       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Which is fine, but is still interesting.

[aprokop]

@dalg24 I added couple commits:

1. Added a more thorough testing for Predicates concept, which also made is_valid_predicate_tag a concept
2. Added complete_type concept which checks whether specialization exists. I'm open to a better name.

I do agree with you that if the concepts allowed to specify our error messages, that would be ideal. The only step in that direction I've seen is the [EWG] P1267, but it went nowhere, it seems. The only alternative I see is to name each step of our concepts with some meaningful name, which is a pain.

[aprokop]

The error messages of nvcc compiler for concepts are baffingly bad; no explanations are given for the concept check failures.