[Bug fortran/97009] New: Inlining with non-standard selected_int_kind leads to errors

[Bug fortran/97009] New: Inlining with non-standard selected_int_kind leads to errors

[pvitt at posteo dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97009

            Bug ID: 97009
           Summary: Inlining with non-standard selected_int_kind leads to
                    errors
           Product: gcc
           Version: 10.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: fortran
          Assignee: unassigned at gcc dot gnu.org
          Reporter: pvitt at posteo dot de
  Target Milestone: ---

Created attachment 49207
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=49207&action=edit
Minimum working example to reproduce the problem.

We ran into a problem that I stripped down to a minimal working example (see
attachment). 

When the MWE is compiled with `gfortran -O1 2269_mwe.f90`, it produces the
expected output

 Level            3  treeID                     0
 Level            2  treeID                     1
 Level            1  treeID                     9

When it is compiled with O2, it fails:

 Level            3  treeID       140732851705864
 Level            2  treeID       140732851705352
 Level            1  treeID            4545064992

When the line

    integer, parameter :: long_k = selected_int_kind(15)

is changed to

    integer, parameter :: long_k = selected_int_kind(9)

it works with O1 and O2.

We stumbled over this when some unit tests failed after we activated LTO. Our
first guess was that LTO stumbles over the elemental function (which tem_pathOf
was before I stripped it down). However, while creating the MWE, we ended up
reproducing the behavior also without LTO, but the called function in the same
file. And as it disappears with O1, it seems to be related to inlining in
general.

[Bug fortran/97009] Inlining with non-standard selected_int_kind leads to errors

[pvitt at posteo dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97009

--- Comment #1 from Peter Vitt <pvitt at posteo dot de> ---
I forgot to mention that we were able to reproduce this also with 10.1.0 and
9.3.0. In case this is of any help...

[Bug fortran/97009] [9/10/11 Regression] Inlining with non-standard selected_int_kind leads to errors

[anlauf at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97009

anlauf at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |9.4
     Ever confirmed|0                           |1
      Known to fail|                            |11.0
            Summary|Inlining with non-standard  |[9/10/11 Regression]
                   |selected_int_kind leads to  |Inlining with non-standard
                   |errors                      |selected_int_kind leads to
                   |                            |errors
           Keywords|                            |wrong-code
   Last reconfirmed|                            |2020-09-10
             Status|UNCONFIRMED                 |NEW
      Known to work|                            |8.4.1
           Priority|P3                          |P4

--- Comment #2 from anlauf at gcc dot gnu.org ---
It appears to be a combination of inlining and optimizations that are present
at -O2 but not at -O1.

OK:

gfc-11 pr97009.f90 -O1 -finline
gfc-11 pr97009.f90 -O2 -fno-inline

Not OK:

gfc-11 pr97009.f90 -O2 -finline

The output in the latter case appears to be random junk.
But valgrind does not show anything for me, nor does ASAN.

Could also be some ME issue.

[Bug tree-optimization/97009] [9/10/11 Regression] Inlining with non-standard selected_int_kind leads to errors

[anlauf at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97009

anlauf at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|fortran                     |tree-optimization

--- Comment #3 from anlauf at gcc dot gnu.org ---
OK:

gfc-11 pr97009.f90 -O2 -finline -fno-tree-vrp

Not OK:

gfc-11 pr97009.f90 -O2 -finline -ftree-vrp

Reassigning to tree-optimization.

[Bug tree-optimization/97009] [9/10/11 Regression] Inlining with non-standard selected_int_kind leads to errors

[anlauf at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97009

anlauf at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P4                          |P3

--- Comment #4 from anlauf at gcc dot gnu.org ---
Resetting to P3, as it is no longer a Fortran issue.

[Bug tree-optimization/97009] [9/10/11 Regression] Inlining with non-standard selected_int_kind leads to errors

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97009

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P3                          |P2
                 CC|                            |jamborm at gcc dot gnu.org

--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
So SRA decomposes

  resD.3963.levelD.3901 = res$level_55;
  resD.3965 = resD.3963;
  resD.3963 ={v} {CLOBBER};
  D.3958 = resD.3965;
  resD.3965 ={v} {CLOBBER};

to just

  res$level_26 = res$level_55;
  resD.3963.levelD.3901 = res$level_26;
  resD.3965 = resD.3963;
  SR.14_27 = res$level_26;
  resD.3963 ={v} {CLOBBER};
  MEM <integer(kind=4)D.8> [(struct tem_path_type *)&D.3958] = SR.14_27;
  resD.3965 ={v} {CLOBBER};

but D.3958 is later indexed

  _11 = &D.3958.nodeD.3903[_10];
  _gfortran_transfer_integer_write (&dt_parm.0D.3961, _11, 8);

we have

  struct tem_path_type resD.3963;
  struct tem_path_type resD.3965;
  struct tem_path_type D.3958;

the question is how "complete" 'tem_path_type' is.

-fno-tree-sra "fixes" it

[Bug tree-optimization/97009] [9/10/11 Regression] Inlining with non-standard selected_int_kind leads to errors

[jamborm at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97009

--- Comment #6 from Martin Jambor <jamborm at gcc dot gnu.org> ---
Created attachment 50492
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=50492&action=edit
C testcase

C testcase.

[Bug tree-optimization/97009] [9/10/11 Regression] Inlining with non-standard selected_int_kind leads to errors

[jamborm at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97009

--- Comment #7 from Martin Jambor <jamborm at gcc dot gnu.org> ---
I am about to test this patch.  I think this should be P1 and I would really
like to get this fix to GCC 10.3.  Sorry for getting to this so late.

diff --git a/gcc/tree-sra.c b/gcc/tree-sra.c
index d177f1ba11c..8dfc923ed7e 100644
--- a/gcc/tree-sra.c
+++ b/gcc/tree-sra.c
@@ -2723,6 +2723,19 @@ budget_for_propagation_access (tree decl)
   return true;
 }

+/* Return true if ACC or any of its subaccesses has grp_child set.  */
+
+static bool
+access_or_its_child_written (struct access *acc)
+{
+  if (acc->grp_write)
+    return true;
+  for (struct access *sub = acc->first_child; sub; sub = sub->next_sibling)
+    if (access_or_its_child_written (sub))
+      return true;
+  return false;
+}
+
 /* Propagate subaccesses and grp_write flags of RACC across an assignment link
    to LACC.  Enqueue sub-accesses as necessary so that the write flag is
    propagated transitively.  Return true if anything changed.  Additionally,
if
@@ -2836,7 +2849,7 @@ propagate_subaccesses_from_rhs (struct access *lacc,
struct access *racc)
       if (rchild->grp_unscalarizable_region
          || !budget_for_propagation_access (lacc->base))
        {
-         if (rchild->grp_write && !lacc->grp_write)
+         if (!lacc->grp_write && access_or_its_child_written (rchild))
            {
              ret = true;
              subtree_mark_written_and_rhs_enqueue (lacc);

[Bug tree-optimization/97009] [9/10/11 Regression] Inlining with non-standard selected_int_kind leads to errors

[jamborm at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97009

--- Comment #8 from Martin Jambor <jamborm at gcc dot gnu.org> ---
I have proposed the patch on the mailing list:
https://gcc.gnu.org/pipermail/gcc-patches/2021-March/567553.html

[Bug tree-optimization/97009] [9/10/11 Regression] Inlining with non-standard selected_int_kind leads to errors

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97009

--- Comment #9 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Martin Jambor <jamborm@gcc.gnu.org>:

https://gcc.gnu.org/g:19d71674616e6494a60432a2a28adcd762a6c877

commit r11-7944-g19d71674616e6494a60432a2a28adcd762a6c877
Author: Martin Jambor <mjambor@suse.cz>
Date:   Thu Apr 1 10:12:23 2021 +0200

    sra: Fix bug in grp_write propagation (PR 97009)

    SRA represents parts of aggregates which are arrays accessed with
    unknown index as "unscalarizable regions."  When there are two such
    regions one within another and the outer is only read whereas the
    inner is written to, SRA fails to propagate that write information
    across assignments.  This means that a second aggregate can contain
    data while SRA thinks it does not and the pass can wrongly eliminate
    big chunks of assignment from that second aggregate into a third
    aggregate, which is what happens in PR 97009.

    Fixed by checking all children of unscalariable accesses for the
    grp_write flag.

    gcc/ChangeLog:

    2021-03-31  Martin Jambor  <mjambor@suse.cz>

            PR tree-optimization/97009
            * tree-sra.c (access_or_its_child_written): New function.
            (propagate_subaccesses_from_rhs): Use it instead of a simple
grp_write
            test.

    gcc/testsuite/ChangeLog:

    2021-03-31  Martin Jambor  <mjambor@suse.cz>

            PR tree-optimization/97009
            * gcc.dg/tree-ssa/pr97009.c: New test.

[Bug tree-optimization/97009] [9/10/11 Regression] Inlining with non-standard selected_int_kind leads to errors

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97009

--- Comment #10 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-10 branch has been updated by Martin Jambor
<jamborm@gcc.gnu.org>:

https://gcc.gnu.org/g:368875572b6b6be75d7cc162797d07e779194fb1

commit r10-9650-g368875572b6b6be75d7cc162797d07e779194fb1
Author: Martin Jambor <mjambor@suse.cz>
Date:   Thu Apr 1 10:12:23 2021 +0200

    sra: Fix bug in grp_write propagation (PR 97009)

    SRA represents parts of aggregates which are arrays accessed with
    unknown index as "unscalarizable regions."  When there are two such
    regions one within another and the outer is only read whereas the
    inner is written to, SRA fails to propagate that write information
    across assignments.  This means that a second aggregate can contain
    data while SRA thinks it does not and the pass can wrongly eliminate
    big chunks of assignment from that second aggregate into a third
    aggregate, which is what happens in PR 97009.

    Fixed by checking all children of unscalariable accesses for the
    grp_write flag.

    gcc/ChangeLog:

    2021-03-31  Martin Jambor  <mjambor@suse.cz>

            PR tree-optimization/97009
            * tree-sra.c (access_or_its_child_written): New function.
            (propagate_subaccesses_from_rhs): Use it instead of a simple
grp_write
            test.

    gcc/testsuite/ChangeLog:

    2021-03-31  Martin Jambor  <mjambor@suse.cz>

            PR tree-optimization/97009
            * gcc.dg/tree-ssa/pr97009.c: New test.

    (cherry picked from commit 19d71674616e6494a60432a2a28adcd762a6c877)

[Bug tree-optimization/97009] [9 Regression] Inlining with non-standard selected_int_kind leads to errors

[jamborm at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97009

Martin Jambor <jamborm at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|[9/10/11 Regression]        |[9 Regression] Inlining
                   |Inlining with non-standard  |with non-standard
                   |selected_int_kind leads to  |selected_int_kind leads to
                   |errors                      |errors

--- Comment #11 from Martin Jambor <jamborm at gcc dot gnu.org> ---
Pushed to master and gcc-10 branch, I will backport the patch to gcc-9 next
week.

[Bug tree-optimization/97009] [9 Regression] Inlining with non-standard selected_int_kind leads to errors

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97009

--- Comment #12 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-9 branch has been updated by Martin Jambor
<jamborm@gcc.gnu.org>:

https://gcc.gnu.org/g:3181f66afc7926950115d9802de1516fba3822a2

commit r9-9326-g3181f66afc7926950115d9802de1516fba3822a2
Author: Martin Jambor <mjambor@suse.cz>
Date:   Thu Apr 8 18:57:48 2021 +0200

    sra: Fix bug in grp_write propagation (PR 97009)

    SRA represents parts of aggregates which are arrays accessed with
    unknown index as "unscalarizable regions."  When there are two such
    regions one within another and the outer is only read whereas the
    inner is written to, SRA fails to propagate that write information
    across assignments.  This means that a second aggregate can contain
    data while SRA thinks it does not and the pass can wrongly eliminate
    big chunks of assignment from that second aggregate into a third
    aggregate, which is what happens in PR 97009.

    Fixed by checking all children of unscalariable accesses for the
    grp_write flag.

    gcc/ChangeLog:

    2021-03-31  Martin Jambor  <mjambor@suse.cz>

            PR tree-optimization/97009
            * tree-sra.c (access_or_its_child_written): New function.
            (propagate_subaccesses_from_rhs): Use it instead of a simple
grp_write
            test.

    gcc/testsuite/ChangeLog:

    2021-03-31  Martin Jambor  <mjambor@suse.cz>

            PR tree-optimization/97009
            * gcc.dg/tree-ssa/pr97009.c: New test.

    (cherry picked from commit 19d71674616e6494a60432a2a28adcd762a6c877)

[Bug tree-optimization/97009] [9 Regression] Inlining with non-standard selected_int_kind leads to errors

[jamborm at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97009

Martin Jambor <jamborm at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|---                         |FIXED

--- Comment #13 from Martin Jambor <jamborm at gcc dot gnu.org> ---
Fixed.

[Bug tree-optimization/97009] [9 Regression] Inlining with non-standard selected_int_kind leads to errors

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97009

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |dimitri.gorokhovik at free dot fr

--- Comment #14 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
*** Bug 94485 has been marked as a duplicate of this bug. ***