[Bug other/46567] New: ipa_tm_decrement_clone_counts ICE at -O2

[Bug other/46567] New: ipa_tm_decrement_clone_counts ICE at -O2

[aldyh at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46567

           Summary: ipa_tm_decrement_clone_counts ICE at -O2
           Product: gcc
           Version: trans-mem
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: other
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: aldyh@gcc.gnu.org


$  ./cc1plus c.cc -O -fgnu-tm -fdump-tree-all -fdump-ipa-all -quiet -O2
c.cc: In function 'void* buildProjectSyncStepConcurrently(int, int)':
c.cc:2674:1: internal compiler error: in ipa_tm_decrement_clone_counts, at
trans-mem.c:3687
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.

[Bug other/46567] ipa_tm_decrement_clone_counts ICE at -O2

[aldyh at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46567

--- Comment #1 from Aldy Hernandez <aldyh at gcc dot gnu.org> 2010-11-19 18:11:23 UTC ---
Created attachment 22460
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22460
boom

[Bug other/46567] ipa_tm_decrement_clone_counts ICE at -O2

[aldyh at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46567

Aldy Hernandez <aldyh at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |trans-mem
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2010.11.19 18:11:58
     Ever Confirmed|0                           |1

[Bug other/46567] ipa_tm_decrement_clone_counts ICE at -O2

[aldyh at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46567

--- Comment #2 from Aldy Hernandez <aldyh at gcc dot gnu.org> 2010-11-19 18:14:07 UTC ---
*** Bug 46566 has been marked as a duplicate of this bug. ***

[Bug other/46567] [trans-mem] ipa_tm_decrement_clone_counts ICE at -O2

[aldyh at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46567

--- Comment #3 from Aldy Hernandez <aldyh at gcc dot gnu.org> 2010-11-24 14:43:53 UTC ---
ipa_tm_decrement_clone_counts() is dying because we are decrementing the clone
count for a caller that was not marked as being in a transaction in the first
place.

The problem here is that in the <tmipa> pass we have a transaction containing a
call that was not marked as inside of a transaction:

<bb 2>:
  __transaction [[relaxed]]
...
<bb 4>:
  ...
  std::_List_node_base::unhook (SR.31_22); <-- HERE

The pass that marks calls as inside of a transaction is in examine_call_tm()
which is part of the <tmlower> pass.  The function unhook() above was placed
inside of the transaction in the early inline pass, and inlining is one of the
small IPA passes, which happen after the lowering passes.

It would seem we need to run <tmlower> after early inlining, but before
<tmipa>, so some sort of lowering pass between the inlining passes.

Richard?

[Bug other/46567] [trans-mem] ipa_tm_decrement_clone_counts ICE at -O2

[aldyh at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46567

Aldy Hernandez <aldyh at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P3                          |P1

[Bug other/46567] [trans-mem] ipa_tm_decrement_clone_counts ICE at -O2

[aldyh at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46567

--- Comment #4 from Aldy Hernandez <aldyh at gcc dot gnu.org> 2010-12-13 11:12:08 UTC ---
[Comments from Richard.]

There are two possible solutions:

  (1) Teach inlining to propagate the in_transaction_p bit
      to every call within the function that gets inlined,

  (2) Don't use an in_transaction_p bit.  I.e. rewrite
      <tmipa> to only use the region blocks bitmap, rather
      than cheating with on-the-side data that has to be
      kept up to date by other parties.

Nr 2 is probably the better solution.  You might be able to use
cgraph_edge->aux as a replacement marker for in_transaction_p
during the <tmipa> pass.  That would certainly make things like
the use in ipa_tm_note_irrevocable easier.

[Comments from Aldy]

I'll be fixing this with option 2, but I'm trying to come up with a work around
for now.

[Bug other/46567] [trans-mem] ipa_tm_decrement_clone_counts ICE at -O2

[patrick.marlier at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46567

--- Comment #5 from Patrick Marlier <patrick.marlier at gmail dot com> 2011-01-28 14:18:45 UTC ---
Created attachment 23153
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=23153
testcase with O0

With this testcase and -O0:

$ gcc -fgnu-tm -O0 -S pr46567-2.i -o testcase.s
pr46567-2.i: In function ‘readLoop’:
pr46567-2.i:18:1: internal compiler error: in ipa_tm_decrement_clone_counts, at
trans-mem.c:3748
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.

Maybe not related to this bug?

[Bug other/46567] [trans-mem] ipa_tm_decrement_clone_counts ICE at -O2

[amacleod at redhat dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46567

Andrew Macleod <amacleod at redhat dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |amacleod at redhat dot com

--- Comment #6 from Andrew Macleod <amacleod at redhat dot com> 2011-01-28 19:41:28 UTC ---
discussion starting at http://gcc.gnu.org/ml/gcc-patches/2011-01/msg01479.html
for the third approach which I implemented. checked in as revision 169369 on
Jan 28.

This implements the need to track whether a call in in a transaction or not.

The follow on bugs are not related to that problem specifically, but then I
wonder if the original problem was *really* tracking the bit :-).  Regardless,
not requiring the bit is a good thing, and will help prevent future problems.

looking at the other testcase now.

[Bug other/46567] [trans-mem] ipa_tm_decrement_clone_counts ICE at -O2

[patrick.marlier at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46567

--- Comment #7 from Patrick Marlier <patrick.marlier at gmail dot com> 2011-02-03 10:28:37 UTC ---
Created attachment 23234
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=23234
new testcase with O0

Here a reduced test case that raises this ICE with current svn version:
$ gcc -O0 -S -fgnu-tm ICE_ipa_tm_propagate_irr.i 
ICE_ipa_tm_propagate_irr.i: In function ‘readLoop’:
ICE_ipa_tm_propagate_irr.i:23:1: internal compiler error: in
ipa_tm_decrement_clone_counts, at trans-mem.c:3763
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.

and raises this ICE with Andrew's patch (tm2.patch)
gcc -O0 -S -fgnu-tm ICE_ipa_tm_propagate_irr.i
ICE_ipa_tm_propagate_irr.i: In function ‘readLoop’:
ICE_ipa_tm_propagate_irr.i:23:1: internal compiler error: in
ipa_tm_propagate_irr, at trans-mem.c:3738
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.

Thanks.

Patrick Marlier.

[Bug other/46567] [trans-mem] ipa_tm_decrement_clone_counts ICE at -O2

[aldyh at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46567

--- Comment #8 from Aldy Hernandez <aldyh at gcc dot gnu.org> 2011-02-11 14:42:35 UTC ---
Awaiting approval:
http://gcc.gnu.org/ml/gcc-patches/2011-02/msg00742.html

[Bug other/46567] [trans-mem] ipa_tm_decrement_clone_counts ICE at -O2

[aldyh at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46567

--- Comment #9 from Aldy Hernandez <aldyh at gcc dot gnu.org> 2011-02-11 18:18:11 UTC ---
Author: aldyh
Date: Fri Feb 11 18:18:09 2011
New Revision: 170057

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=170057
Log:
        PR 46567
        * trans-mem.c (ipa_tm_propagate_irr): Change assert to if.


Added:
    branches/transactional-memory/gcc/testsuite/gcc.dg/tm/pr46567-2.c
Modified:
    branches/transactional-memory/gcc/ChangeLog.tm
    branches/transactional-memory/gcc/trans-mem.c

[Bug other/46567] [trans-mem] ipa_tm_decrement_clone_counts ICE at -O2

[aldyh at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46567

Aldy Hernandez <aldyh at gcc dot gnu.org> changed:

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

--- Comment #10 from Aldy Hernandez <aldyh at gcc dot gnu.org> 2011-02-11 18:19:17 UTC ---
Fixed.