[Patch, fortran] PR25746 - elemental subroutine dependency checking - redux

[Paul Thomas]

Dear All,

I should explain that Outlook normally seems to emit non-mime characters 
that gnu-central does not like.  Changing to plain text drops all 
references to wrap around.  I thought that I had it fixed but apparently 
not.  Sorry about that.

If this is screwed up, I will resubmit tonight.... well, here it is again.

This patch fixes PR25746 by adding dependency checking to the 
translation of elemental subroutine calls. Quite aside from the 
immediate need of the PR to ensure that user defined assignments work 
correctly, the standard requires that elemental subroutines produce the 
same result, independent of the loop order and this mandates the 
dependency checking.

The checking is accomplished by calling a new function 
gfc_conv_elemental_dependencies before starting the scalarizing body and 
calling the elemental subroutine itself. This has the advantage that the 
loopinfo and the ss's for each actual argument are available and the 
information that they contain is used as much as possible.

The new function proceeds as follows:

It loops over the actual arguments and, if the INTENT is OUT or INOUT, 
it checks for dependency with each of the arguments that are not 
INTENT(OUT). I have modified dependency.c(gfc_check_fncall_dependency) 
to prevent the argument expression from being checked against itself.

If there is a dependency, a temporary loopinfo is made, ready for the 
call to gfc_trans_create_temp_array. This is done to prevent this 
function from resetting the loop and thereby necessitating the 
renormalisation of the ss->data.info's for each of the arguments. Once 
the temporary is obtained, the descriptor is obtained afresh, so that 
the result can be correctly unpacked after the subroutine call.

In the case of an INTENT(INOUT) argument, the original data has to be 
packed and copied to the temporary. I have signalled a TODO here; the 
potential creation of two temporaries could be eliminated and this is 
something that I intend to contribute after I have fixed the array 
TRANSFER intrinsic. This will require a new library routine that can be 
employed here. In the mean time, this patch works correctly (I hope!), 
even if it is a bit inefficient in a small number of cases.

As well as copying the packed data to the temporary used for the 
subroutine call, this part of the code frees the temporary potentially 
produced by internal_pack.

Following this, the scalarizer is satisfied by adding the offset for the 
temporary to the ss->data.info, using the loop index ranges. Finally, 
the result is copied to the destination array and the temporary freed.

As well as testing the fix for the original PR, the testcase checks 
various kinds of assignment, including >1D arrays, and tests that 
INTENT(INOUT) arguments are correctly handled.  It has been corrected to 
be conformant (thanks, Dominique).

Regtested on FC5/Athlon. OK for trunk and 4.1?

Paul

2006-04-21  Paul Thomas  <pault@gcc.gnu.org>

    PR fortran/25746
    * dependency.c (gfc_check_fncall_dependency): Don't check other
    against itself.
    * trans-stmt.c (gfc_conv_elemental_dependencies): New function.
    (gfc_trans_call): Call it.

2006-04-21  Paul Thomas  <pault@gcc.gnu.org>

    PR fortran/25746
    * gfortran.dg/elemental_subroutine_3.f90: New test.



-------------- next part --------------
A non-text attachment was scrubbed...
Name: depends04.diff
Type: text/x-patch
Size: 9596 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/fortran/attachments/20060421/99049523/attachment.bin>