[Bug bootstrap/50826] New: bootstrap on 64 bit pa broken by r180194, ICE in mem_loc_descriptor

[Bug bootstrap/50826] New: bootstrap on 64 bit pa broken by r180194, ICE in mem_loc_descriptor

[sje at cup dot hp.com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50826

             Bug #: 50826
           Summary: bootstrap on 64 bit pa broken by r180194, ICE in
                    mem_loc_descriptor
    Classification: Unclassified
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: bootstrap
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: sje@cup.hp.com
              Host: hppa64-hp-hpux11.11
            Target: hppa64-hp-hpux11.11
             Build: hppa64-hp-hpux11.11

[Bug bootstrap/50826] bootstrap on 64 bit pa broken by r180194, ICE in mem_loc_descriptor

[sje at cup dot hp.com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50826

--- Comment #1 from Steve Ellcey <sje at cup dot hp.com> 2011-10-21 20:46:46 UTC ---
Created attachment 25573
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=25573
small test case

Test case.

[Bug bootstrap/50826] bootstrap on 64 bit pa broken by r180194, ICE in mem_loc_descriptor

[sje at cup dot hp.com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50826

--- Comment #2 from Steve Ellcey <sje at cup dot hp.com> 2011-10-21 20:55:47 UTC ---
forgot to mention that the test case needs to be compiled with -O2 -g to
reproduce the problem.

[Bug bootstrap/50826] bootstrap on 64 bit pa broken by r180194, ICE in mem_loc_descriptor

[aoliva at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50826

Alexandre Oliva <aoliva at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |ASSIGNED
   Last reconfirmed|                            |2011-10-25
     Ever Confirmed|0                           |1

--- Comment #3 from Alexandre Oliva <aoliva at gcc dot gnu.org> 2011-10-25 01:40:02 UTC ---
On it.

[Bug bootstrap/50826] bootstrap on 64 bit pa broken by r180194, ICE in mem_loc_descriptor

[aoliva at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50826

--- Comment #4 from Alexandre Oliva <aoliva at gcc dot gnu.org> 2011-10-25 04:05:05 UTC ---
Created attachment 25605
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=25605
Patch that restores earlier behavior, hiding the problem

This patch restores the cselib/vta expansion of locations in location lists for
multi-part variables.  We shouldn't have to do this, but it is an alternative
for a temporary work-around in case the other patch I'm about to post isn't
enough to completely fix the problem.

[Bug bootstrap/50826] bootstrap on 64 bit pa broken by r180194, ICE in mem_loc_descriptor

[aoliva at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50826

Alexandre Oliva <aoliva at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #25605|0                           |1
        is obsolete|                            |

--- Comment #5 from Alexandre Oliva <aoliva at gcc dot gnu.org> 2011-10-25 04:12:03 UTC ---
Created attachment 25606
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=25606
Patch that hopefully fixes the problem for good

Here's a patch that fixes the latent problem that caused us to issue MOps that
bound parts of multi-part variables to MEMs whose addresses refer to
debug_exprs.  We shouldn't do that, for we have no infrastructure to recompute
the location of the multi-part variable as the debug_expr changes, and we
shouldn't have to: VT tracks locations (REGs or MEMs), rather than values, and
the locations aren't supposed to change.

[Bug bootstrap/50826] bootstrap on 64 bit pa broken by r180194, ICE in mem_loc_descriptor

[aoliva at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50826

--- Comment #6 from Alexandre Oliva <aoliva at gcc dot gnu.org> 2011-10-25 04:14:15 UTC ---
Steve, forgot to ask you to give the latest patch a try and let me know how it
goes.  I'm not sure yet why we don't trigger this problem on x86*.  TIA,

[Bug bootstrap/50826] bootstrap on 64 bit pa broken by r180194, ICE in mem_loc_descriptor

[sje at cup dot hp.com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50826

--- Comment #7 from Steve Ellcey <sje at cup dot hp.com> 2011-10-25 19:53:31 UTC ---
Bootstrapping and testing both look good with the patch.

[Bug bootstrap/50826] bootstrap on 64 bit pa broken by r180194, ICE in mem_loc_descriptor

[aoliva at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50826

--- Comment #8 from Alexandre Oliva <aoliva at gcc dot gnu.org> 2011-10-26 08:52:13 UTC ---
Thanks.  FTR, the reason we get this on PA64 and not on x86* (and probably not
on PA32) is that pa_internal_arg_pointer sets up a pseudo to hold r29+offset.
The pseudo is fwpropped into the assignments to args passed in registers but
that must live on the stack (“insn” in this testcase).  The last such
assignment becomes r29's death point, but the pseudo was propagated into debug
uses at later points, so DF sets up a debug temp bound to r29 before its death,
and replaces r29 with the debug temp in the debug uses.

This doesn't happen on x86 because there's no such pseudo: the arg pointer is
used directly throughout the function, so it doesn't die early and doesn't
trigger the debug temp substitution that ended up triggering the VT bug.

Maybe this new info can lead to a simplification of the test in the patch; I'm
not sure about it yet.

[Bug debug/50826] bootstrap on 64 bit pa broken by r180194, ICE in mem_loc_descriptor

[aoliva at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50826

--- Comment #9 from Alexandre Oliva <aoliva at gcc dot gnu.org> 2011-10-26 13:32:57 UTC ---
Nice, a testsuite run detected gcc.dg/torture/pr43165.c -Os -g on
i686-linux-gnu gets a debug stmt referencing a user variable using a debug temp
as the address.  Sadly, that exposes a bug in the generation of debug temps
within df-problems: we'd bind the temp to the previous value of bp, rather than
to the one copied from sp as intended.  Testing a fix now, but it ends up
causing the potential problem trigger to no longer occur on any testcase on
i686.  OTOH, it might obviate the expensive-ish test in the proposed patch.

[Bug debug/50826] bootstrap on 64 bit pa broken by r180194, ICE in mem_loc_descriptor

[aoliva at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50826

--- Comment #10 from Alexandre Oliva <aoliva at gcc dot gnu.org> 2011-10-26 13:47:56 UTC ---
Author: aoliva
Date: Wed Oct 26 13:47:48 2011
New Revision: 180525

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=180525
Log:
PR debug/50826
* var-tracking.c (rtx_debug_expr_p): New.
(use_type): Don't use debug exprs to track non-VTA variables.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/var-tracking.c

[Bug debug/50826] bootstrap on 64 bit pa broken by r180194, ICE in mem_loc_descriptor

[aoliva at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50826

Alexandre Oliva <aoliva at gcc dot gnu.org> changed:

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

--- Comment #11 from Alexandre Oliva <aoliva at gcc dot gnu.org> 2011-10-26 13:53:18 UTC ---
Fixed