Thomas Schwinge wrote:
> In follow-up messages, I'll be posting the separated parts (for easier
> review) of a next set of OpenACC changes that we'd like to commit.
> ChangeLog updates not yet written; will do that before commit, obviously.
Still, it would have been nice if you had given an overview about what
the main part of the patch does. In this case, there is neither some intro
words nor a ChangeLog (which also gives an overview what the patch does).
Regarding the !$ACC ROUTINE support, which the patch adds: In my very
rough understanding, the compiler has to know at compile time whether a
procedure is a '!$ACC ROUTINE' or not. Thus, it should work if you declare
in one file:
!--------- file1.f90-----------
subroutine foo()
!$acc routine
end subroutine foo
!--- end of file
and invoke it in a different file:
call foo()
In order that this works in Fortran, you need to support two ways of
handling this information:
a) modules:
module m
contains
subroutine foo()
!$acc routine
end subroutine foo
end module m
Thus, I had expected that you store this information (at least the relevant
parts) in the .mod file. (-> fortran/module.c)
b) interfaces, i.e. in the scope of the caller:
interface
subroutine foo()
!$acc routine
end subroutine foo
end interface
!$acc parallel
...
call foo()
...
!$acc end parallel
Semantically, something like the the latter is also needed in the same file
if the procedure "foo" is stand along (i.e. neither in a module nor 'contained'
(internal procedure) of the caller nor a sibling internal procedure of the
caller). However, as the original declaration is available, supporting (b) is
only semantical sugar - especially a check whether the !$acc routine is provided
in the interface and matches the original declaration.
Thus, can you check:
* whether you need to store information in the .mod file?
* whether (b) is/should be supported? (And with how much error diagnostic
in case of the same TU and mismatches.)
And, if needed, provide some test cases?
And if you are there:
* Whether something similar to .mod has to be done for LTO?
* Could you also create a test case, where Fortran calls C or vice versa for
an OpenACC ROUTINE?
Tobias
