This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [Patch, Fortran] PR fortran/35681: First part, fix ELEMENTAL dependency handling for MVBITS


Daniel,

Regtests fine on FC9/x86_i64. This is OK for trunk.

Have you had any thoughts on parentheses expressions yet?

Thanks for the patch

Paul
Hi,

I've updated the patch described below to trunk of now (including the trivial conflicts merge with Mikael's recent check-in) and run a new regtest, no regressions on GNU/Linux-x86-32.

Cheers,
Daniel

Daniel Kraft wrote:
working on PR fortran/35681, I've got some rather big patch now handling part of the problem. What it exactly does:

1) Some tab-indentation formatting fixes as I came along, sorry for those. I hope it is ok so.

2) When resolving a MVBITS intrinsic call, the code->resolved_sym gets a dummy formal argument list with the correct INTENTs specified; this is needed later for gfc_conv_elemental_dependencies.

3) gfc_code got a new member "resolved_isym" that tracks calls to intrinsic procedures, so we can later check if some call is to intrinsic MVBITS. This got a little ugly and would be probably nicer to union it (and possibly "resolved_sym", too) with actual, but that would probably introduce a lot of changes to existing code pieces.

4) gfc_trans_allocate_array_storage (or what it is called) got a new argument `initial' that allows to initialize the created storage from some other array (this is done using a combination of internal_pack and memcpy if it was already packed, I hope I got this all right). This is used for gfc_trans_create_temp_array to allow initializing the new temporary. Here is (probably) most of the "critical" changes.

5) For calls to intrinsic MVBITS, I enabled dependency checking using gfc_conv_elemental_dependencies and made this routine aware of INTENT(INOUT) arguments that use the new initialization feature to copy over the initial content of the mirrored array to the created temporary.

6) I could not find a test to verify this (not even one that uses gfc_conv_elemental_dependencies) in a quick trial, but I believe the handling of the temporary there was wrong, in that it was free'd (if allocated on the heap) *before* it was used with internal_unpack, because gfc_trans_create_temp_array added the temporary clean-up code to se->post and the unpack-call was added to se->post later. In my opinion, this is some rather general problem with how post-commands are usually added to other post blocks; shouldn't they be added to the top usually rather than to the bottom, to get some sort of "nested" scope with inner most pairs of pre/post? Well, for now I changed this behaviour inside gfc_conv_elemental_dependencies, which corrected problems I got with MVBITS tests.

This enabled the (valid) tests in the PR to run, but only with modifying them slightly by removing the parentheses around the first argument (so it is not an expression; that will be part 2 of this fix). As I understand it, this is valid in case of MVBITS but not for any other ELEMENTAL subroutine, right? This is why I added the check for whether some call is to MVBITS. I guess the rationale why the compiler is not required to create temporaries for all such ELEMENTAL calls (and they are invalid instead) is performance? gfortran could handle those calls well in addition to only MVBITS calls simply if I take this conditional check out, but then we might generate temporaries for cases where the user knows no one is needed and the code is valid but the compiler can't figure it out.

I hope I got this one at least somewhat clear... What do you think about it? Currently regression-testing on GNU/Linux-x86-32, but I don't expect any (a very similar patch worked fine before).

Cheers,
Daniel





Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]