[Bug fortran/113305] New: ICE with do concurrent and ivdep

[Bug fortran/113305] New: ICE with do concurrent and ivdep

[ivan.pribec at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113305

            Bug ID: 113305
           Summary: ICE with do concurrent and ivdep
           Product: gcc
           Version: 13.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: fortran
          Assignee: unassigned at gcc dot gnu.org
          Reporter: ivan.pribec at gmail dot com
  Target Milestone: ---

Created attachment 57022
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=57022&action=edit
mwe using do concurrent and ivdep directive

An internal compiler error is triggered by adding the ivdep directive to a do
concurrent loop:

program dc
implicit none
real :: a(12), b(12)
integer :: i
call random_number(b)
!GCC$ ivdep
do concurrent (i=1:12)
    a(i) = 2*b(i)
end do
end program

...

$ gfortran -Wall dc.f90 
f951: internal compiler error: Segmentation fault: 11
Please submit a full bug report, with preprocessed source (by using
-freport-bug).
See <https://github.com/Homebrew/homebrew-core/issues> for instructions.

[Bug fortran/113305] ICE with do concurrent and ivdep

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113305

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2024-01-10
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed.

f951: internal compiler error: Segmentation fault
0x17cf885 crash_signal
        /space/rguenther/src/gcc/gcc/toplev.cc:316
0xe79b99 resolve_forall_iterators
        /space/rguenther/src/gcc/gcc/fortran/resolve.cc:7620
0xe85f05 gfc_resolve_code(gfc_code*, gfc_namespace*)
        /space/rguenther/src/gcc/gcc/fortran/resolve.cc:12745
0xe939ad resolve_codes
        /space/rguenther/src/gcc/gcc/fortran/resolve.cc:18107
0xe93ad7 gfc_resolve(gfc_namespace*)
        /space/rguenther/src/gcc/gcc/fortran/resolve.cc:18142
0xe5e8b9 resolve_all_program_units
        /space/rguenther/src/gcc/gcc/fortran/parse.cc:6983
0xe5f0a6 gfc_parse_file()
        /space/rguenther/src/gcc/gcc/fortran/parse.cc:7239
0xec4940 gfc_be_parse_file
        /space/rguenther/src/gcc/gcc/fortran/f95-lang.cc:239

[Bug fortran/113305] ICE with do concurrent and ivdep

[kargl at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113305

kargl at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |kargl at gcc dot gnu.org

--- Comment #2 from kargl at gcc dot gnu.org ---
(In reply to Richard Biener from comment #1)
> Confirmed.
> 
> f951: internal compiler error: Segmentation fault
> 0x17cf885 crash_signal
>         /space/rguenther/src/gcc/gcc/toplev.cc:316
> 0xe79b99 resolve_forall_iterators
>         /space/rguenther/src/gcc/gcc/fortran/resolve.cc:7620

No idea what ivdep is suppose to do, but it is certainly messing up gfortran.

(gdb) b resolve.cc:7620
Breakpoint 1 at 0x9e076d: file ../../gccx/gcc/fortran/resolve.cc, line 7620.
(gdb) run a.f90
Breakpoint 1, resolve_forall_iterators (it=0x203c6d360)
    at ../../gccx/gcc/fortran/resolve.cc:7620
7620          if (gfc_resolve_expr (iter->var)
(gdb) p *it
$5 = {var = 0x203c250a0, start = 0x203c25180, end = 0x203c25260, 
  stride = 0x203c25340, next = 0x10000}

it->next is set to an invalid address, so the for loop is not
exited.  If I remove the '!GCC ivdep' directive, then one see

(gdb) p *it->next
Cannot access memory at address 0x0

the expected result.  I think that this is not a Fortran bug, and the
component should be changed to middle-end.

[Bug fortran/113305] ICE with do concurrent and ivdep

[ivan.pribec at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113305

--- Comment #3 from Ivan Pribec <ivan.pribec at gmail dot com> ---
In the mail archive
(https://gcc.gnu.org/legacy-ml/fortran/2014-02/msg00077.html) I've read a
message implying that do concurrent gets translated into a regular for-loop
annotated with "#pragma ivdep", which assures the loop optimizer there are no
loop-carried dependencies. That said, I'm aware there is no good reason why
you'd want to use this directive on a do concurrent loop.

[Bug fortran/113305] ICE with do concurrent and ivdep

[anlauf at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113305

anlauf at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |anlauf at gcc dot gnu.org

--- Comment #4 from anlauf at gcc dot gnu.org ---
The underlying problem is that parse_do_block expects only simple do-loops
and attaches the loop annotation (ivdep, vector, novector, ...) to the
"wrong" component (ext.iterator instead of ext.forall_iterator).
However, gfc_forall_iterator has a slightly different structure than
gfc_iterator and needs to be amended.

Since the Intel compiler appears to accept "!DIR$ vector always" in front
of do concurrent, we might try to support it, too.

[Bug fortran/113305] ICE with do concurrent and ivdep

[sgk at troutmask dot apl.washington.edu]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113305

--- Comment #5 from Steve Kargl <sgk at troutmask dot apl.washington.edu> ---
On Wed, Jan 10, 2024 at 06:24:40PM +0000, ivan.pribec at gmail dot com wrote:
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113305
> 
> --- Comment #3 from Ivan Pribec <ivan.pribec at gmail dot com> ---
> In the mail archive
> (https://gcc.gnu.org/legacy-ml/fortran/2014-02/msg00077.html) I've read a
> message implying that do concurrent gets translated into a regular for-loop
> annotated with "#pragma ivdep", which assures the loop optimizer there are no
> loop-carried dependencies. That said, I'm aware there is no good reason why
> you'd want to use this directive on a do concurrent loop.
> 

Note, I'm not saying that using '!GCC$ ivdep' is wrong.  It's
that gfortran parses the do-concurrent correctly, but whatever
compiler pass(es) that ivdep activates seems to mishandle the
terminated linked list.  it->var holds information about 'i'
and it->next points to the next do-index if one exists or
is NULL.  ivdep seems to corrupt the NULL.

[Bug fortran/113305] ICE with do concurrent and ivdep

[anlauf at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113305

--- Comment #6 from anlauf at gcc dot gnu.org ---
Created attachment 57026
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=57026&action=edit
Draft patch

Draft patch that passes annotations to DO CONCURRENT loops.

Needs testing and feedback.

[Bug fortran/113305] ICE with do concurrent and ivdep

[anlauf at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113305

--- Comment #7 from anlauf at gcc dot gnu.org ---
Submitted: https://gcc.gnu.org/pipermail/fortran/2024-January/060110.html

[Bug fortran/113305] ICE with do concurrent and ivdep

[anlauf at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113305

anlauf at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |anlauf at gcc dot gnu.org

[Bug fortran/113305] ICE with do concurrent and ivdep

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113305

--- Comment #8 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Harald Anlauf <anlauf@gcc.gnu.org>:

https://gcc.gnu.org/g:9935667a69896865b848dfa690f94c9c693a48a3

commit r14-7226-g9935667a69896865b848dfa690f94c9c693a48a3
Author: Harald Anlauf <anlauf@gmx.de>
Date:   Fri Jan 12 19:51:11 2024 +0100

    Fortran: annotations for DO CONCURRENT loops [PR113305]

    gcc/fortran/ChangeLog:

            PR fortran/113305
            * gfortran.h (gfc_loop_annot): New.
            (gfc_iterator, gfc_forall_iterator): Use for annotation control.
            * array.cc (gfc_copy_iterator): Adjust.
            * gfortran.texi: Document annotations IVDEP, UNROLL n, VECTOR,
            NOVECTOR as applied to DO CONCURRENT.
            * parse.cc (parse_do_block): Parse annotations IVDEP, UNROLL n,
            VECTOR, NOVECTOR as applied to DO CONCURRENT.  Apply UNROLL only to
            first loop control variable.
            * trans-stmt.cc (iter_info): Use gfc_loop_annot.
            (gfc_trans_simple_do): Adjust.
            (gfc_trans_forall_loop): Annotate loops with IVDEP, UNROLL n,
            VECTOR, NOVECTOR as needed for DO CONCURRENT.
            (gfc_trans_forall_1): Handle loop annotations.

    gcc/testsuite/ChangeLog:

            PR fortran/113305
            * gfortran.dg/do_concurrent_7.f90: New test.

[Bug fortran/113305] ICE with do concurrent and ivdep

[anlauf at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113305

anlauf at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |14.0
         Resolution|---                         |FIXED
             Status|ASSIGNED                    |RESOLVED

--- Comment #9 from anlauf at gcc dot gnu.org ---
Fixed in gcc-14.