gfortran 7 regression: no type conversion before coarray put

[rouson]

This issue is for tracking gfortran bug report 78892, which was first reported in an OpenCoarrays mailing list message. The gfortran bug report contains the following text:

The code below demonstrates a regression in which gfortran 7.0.0 20161215 is not performing a necessary implicit type conversion before putting data on a remote image.  By contrast, the code below works as expected with gfortran 5.4.0 and 6.2.0. 

$ cat convert-before-put.f90 
  real :: a[*]
  integer :: receiver
  associate(me=>this_image())
    if (me == 1) then
      do receiver = 2, num_images()
        a[receiver] = receiver ! implicit real(receiver) needed here
        sync images (receiver) ! notify remote image that data has been put
      end do 
    else
      sync images (1) ! await notification of data put by image 1
      if (a/=real(me)) print *, "Image ",me,": received ",a,", but expected ",real(me)
    end if
  end associate
end 

$ caf convert-before-put.f90 

$ cafrun -np 2 ./a.out
 Image            2 : received    0.00000000     , but expected    2.00000000

[rouson]

Tagging @egiovan to include him in the thread.

[vehre]

I still have to analyse this in-depth: But I am pretty sure, that the responsibility for the type-conversion is in the caf-library. There are convert-type routines that do the trick. In the put case gfortan might be able to do the type convert, but in the get case it can't. Therefore is it arguable that the responsibility is always with the library to be consistent.

[vehre]

At the moment I am adding pointer component support for derived typed coarrays, so please give a bit time to take a look.

[rouson]

@vehre, whichever side is responsible, this is a regression from the behavior with gfortran 5 and 6. In the interest of consistency, I assume it's best to keep the responsibility wherever it was prior to gfortran 7 unless there is a strong motivation for changing it. Just so we know how to assign this issue, please confirm whether you wrote the relevant library code that gets invoked when this example is compiled or whether @afanfa wrote it?

[vehre]

I just have rejected the pr in the gcc-bugtracker, because the responsibility for the type conversion was moved to the library. This way the library has the opportunity to reduce the number of bytes to transfer between the images where applicable. When the compiler does the type conversion, always the destination data type is used. With arrays this means the compiler generates a type conversion loop and the library "loops" again over the data to transfer it (I know you are using chunks and MPI-Datatypes to prevent single item transfer). Nevertheless the responsibility has been moved to the library. The conversion routine is already present (the _by_ref routine convert) so that it just needs to be called in the "old" communication primitives.

I am starting to implement it. Maybe @afanfa can check, that I don't mess up the performance completely.