rouson commented 7 years ago

Avg response time

Request for Enhancement (RFE)

As pointed out by Malcolm Cohen on the WG5 Fortran standards committee mailing list, most or all of the type-specific specific procedures in src/extensions/opencoarrays.F90 can be collapsed down to a set of procedures in which the first argument is declared unlimited polymorphic (class(*)). This will both simplify the code and broaden the supported use cases.

As an example, the following program demonstrates that this approach which works at least with GCC 5 and higher:

program main
  implicit none
  integer :: b(10,10)
  call print_storage_info(b)
contains
  subroutine print_storage_info(a)
    class(*), intent(in), target, contiguous :: a(..)
    print *,storage_size(a),size(a)
  end subroutine
end program

The storage_size and size intrinsics should be used to replace the c_sizeofcalls that currently necessitate the type-specific implementations in opencoarrays.F90.

rouson commented 7 years ago

I started looking into this. One issue is choosing a compiler with which to test it. It should be a compiler that triggers the CMake code that builds of the extensions tests -- any non-gfortran compiler or a gfortran compiler version < 5, but let's focus on non-gfortran compilers. In deciding which ones to test, there are a few use cases to consider:

Extending the capability of an existing Fortran 2008 compiler such as a. Intel b. Cray
Accessing Fortran 2015 features in a Fortran 2003 compiler such as a. PGI b. flang c. NAG d. IBM

Arguments in favor of testing with the compilers in category 1 above include ease of implementation and greater likelihood that existing users will want additional parallel features (such as collective subroutines) because they're already using some parallel features (such as coarrays). Arguments in favor of testing with category 2 compilers are that these vendors might be most interested in adopting OpenCoarrays. An additional argument in favor of compiler 2b above is that it is open source, which might offer the greatest possibility of influencing the project itself and also might provide an avenue to increasing the interest of 2a.

Another consideration is whether the compiler supports the Fortran 2008 intrinsic function storage_size and the Fortran 2015 C-interoperability features such as assumed rank. These features greatly simplify the code in the opencoarrays module. The requisite features include, for example, assumed rank. At this time, I believe the list of compilers that support these features includes: I. Intel II. Cray III. IBM IV. GNU I'm going to first check whether flang supports the mentioned 2008 and 2015 features. If not, I'll switch to Intel. I just submitted a flang installation issue that's currently blocking me from evaluating it.

rouson commented 7 years ago

I also submitted an issue on the spack package manager recommended by the flang README.

rouson commented 7 years ago

It doesn't appear there's going to be a fix in the foreseeable future for building flang on macOS using spack. Building flang with spack on Lubuntu Linux is also failing for me, which I'll research and/or report. In the interim, I've sent the code below to Portland Group's J3 representative. Any compiler that either can compile the code below or supports the Fortran 2015 feature of assumed-size arguments for c_sizeof will facilitate the enhancement pondered in this issue. Unfortunately, gfortran appears to be the only compiler that can support the code below and of course the code below isn't needed with gfortran! Arrgghh... The Cray and Intel compiler both die with internal compiler errors on the code below even thought they nominal support the necessary features. I'll submit bug reports and I think I'll proceed with rewriting src/extensions/opencoarrays.F90 anyway.

program main
  use iso_c_binding, only : c_sizeof,c_int32_t
  implicit none
  integer(c_int32_t) :: array(10,1)
  associate(expected=>c_sizeof(array),calculated=>c_sizeof_assumed_rank(array))
    if (expected==calculated) then
      print *,"Test passed"
    else
      print *,"Test failed: expected ",expected,", calculated",calculated
    end if
  end associate
contains
   pure function c_sizeof_assumed_rank(a) result(c_sizeof_a)
     use iso_c_binding, only : c_size_t,c_int32_t
     class(*), intent(in), target, contiguous :: a(..) ! Fortran 2015 assumed-rank dummy argument
     integer(c_size_t) :: c_sizeof_a
     integer, parameter :: bits_per_byte=storage_size(0_c_int32_t)/c_sizeof(0_c_int32_t)
     c_sizeof_a= storage_size(a)*size(a)/bits_per_byte  ! Fortran 2008 storage_size intrinsic function
   end function
end program

rouson commented 7 years ago

The code in my last comment works with the IBM compiler so at least gfortran is not the only compiler that can handle it.

rouson commented 7 years ago

Bug reports have been submitted on the Cray and Intel compilers. Now I'll move forward with revising the module while we wait for the non-GNU compilers to support the requisite features.