Detect join member function with volatile-qualified arguments

[dalg24]

Hotfix for #4931
For backward compatibility, make sure that user-defined reducer that only had a join member function that takes arguments with the volatile type qualifier are not rejected.

Fixes: #4952, #4941

[stanmoore1]

Thanks!

[PhilMiller]

One test failure in Jenkins, on OpenMPTarget:

4: [ RUN      ] openmptarget.reducers_struct
4: /var/jenkins/workspace/Kokkos/core/unit_test/TestReducers.hpp:325: Failure
4: Expected equality of these values:
4:   sum_scalar
4:     Which is: 32-byte object <00-E5 45-47 00-E5 45-47 00-20 D2-44 00-20 D2-44 00-E5 45-47 00-E5 45-47 00-E5 45-47 00-E5 45-47>
4:   reference_sum
4:     Which is: 32-byte object <00-E5 45-47 00-E5 45-47 00-E5 45-47 00-E5 45-47 00-E5 45-47 00-E5 45-47 00-E5 45-47 00-E5 45-47>
4: [  FAILED  ] openmptarget.reducers_struct (2 ms)

[PhilMiller]

In kokkos-kernels develop, I get this test failure:

1: [  FAILED  ] 4 tests, listed below:
1: [  FAILED  ] cuda.iamax_double
1: [  FAILED  ] cuda.iamax_mv_double
1: [  FAILED  ] cuda.iamax_complex_double
1: [  FAILED  ] cuda.iamax_mv_complex_double
1: 
1:  4 FAILED TESTS
1/1 Test #1: blas_cuda ........................***Failed    8.26 sec

[PhilMiller]

There may be misbehavior in the case where join() is overloaded for both volatile and not - the expression &F::join is problematic when there are two matching enable_if with different parameter types it could cast to that would select different overloads. It may still be failing back to operator+= in that case.

We may need separate has_join_*_function specializations, rather than separate overloads within the existing specializations

[PhilMiller]

Removing the overload lets that pass.

[PhilMiller]

dalg24#3 makes the overload a compilation error, rather than a silent runtime failure. Not sure it's a great path, but it does encourage code cleanup!

[PhilMiller]

With the above addition to the PR, kokkos/kokkos-kernels#1382 is the minimal change to Kokkos Kernels to compile on OpenMP and CUDA and pass its tests on both

[PhilMiller]

The OpenMPTarget failure may have something to do with making this call:

      Kokkos::parallel_reduce(Kokkos::RangePolicy<ExecSpace, ReducerTag>(0, N),
                              f_tag, reducer_scalar);

On a type array_reduce that has just operator+= and no join() member at all.

[PhilMiller]

Your revision looks good.

Could you test that in the overloaded join case, the non qualified version gets selected? In my reading, that's what it looks like the code will do, but I'd like to see it ensured.

Should we include a runtime test that operator+= isn't being called in any of those cases? Maybe add an operator+= implementation that just fails the test to your compile test type definition, and make it runnable?

[masterleinad]

Looks good to me.

[PhilMiller]

I tested this with kokkos-kernels develop on OpenMP and CUDA, and it seems to be happy - tests all compile and pass.

I think we should merge this to unbreak KK CI testing, and follow-up with any further tests or refinements. So, marking this 'Ready'

[PhilMiller]

May also fix #4941; testing now

[PhilMiller]

retest this please

[crtrott]

@dalg24 if this passes merge.