[Bug gcov-profile/51975] New: ICE in gcc in convert_move, at expr.c:326 with fprofile-use when source changes from fprofile-generate

[Bug gcov-profile/51975] New: ICE in gcc in convert_move, at expr.c:326 with fprofile-use when source changes from fprofile-generate

[asharif at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51975

             Bug #: 51975
           Summary: ICE in gcc in convert_move, at expr.c:326 with
                    fprofile-use when source changes from
                    fprofile-generate
    Classification: Unclassified
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: gcov-profile
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: asharif@gcc.gnu.org


Created attachment 26438
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=26438
Tarball containing the test files and the proposed fix.

Please see attached testcase, which fails with gcc-4.6.0:

gcc -o profile_io_data.o profile_io_data.ii -O2 -Wno-error -fpermissive
-fprofile-use -m32 -c

This happens because the profile .gcda file is oudated compared to the current
source. gcc's assignment of function ids on the current source is different
from the assignment when it did fprofile-generate. This causes an ICE later
down the pipeline.

I am also attaching a proposed fix, which builds a hash table of function ids
-> cfg checksums when reading in the profile files. At profile use time, in
check_ic_target(), gcc makes sure the target function's checksum matches the
checksum that is found in the gcda file.

[Bug gcov-profile/51975] ICE in gcc in convert_move, at expr.c:326 with fprofile-use when source changes from fprofile-generate

[asharif at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51975

--- Comment #1 from asharif at gcc dot gnu.org 2012-01-24 00:55:14 UTC ---
Created attachment 26439
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=26439
Proposed fix.

Here is just the patch if you want to take a quick look at it.

[Bug gcov-profile/51975] ICE in gcc in convert_move, at expr.c:326 with fprofile-use when source changes from fprofile-generate

[asharif at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51975

asharif at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #26439|0                           |1
        is obsolete|                            |

--- Comment #2 from asharif at gcc dot gnu.org 2012-01-24 20:01:45 UTC ---
Created attachment 26445
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=26445
Proposed fix.

Reused the counts_hash table.

[Bug gcov-profile/51975] ICE in gcc in convert_move, at expr.c:326 with fprofile-use when source changes from fprofile-generate

[asharif at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51975

asharif at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #26445|0                           |1
        is obsolete|                            |

--- Comment #3 from asharif at gcc dot gnu.org 2012-01-24 20:10:21 UTC ---
Created attachment 26446
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=26446
Proposed fix.

[Bug gcov-profile/51975] ICE in gcc in convert_move, at expr.c:326 with fprofile-use when source changes from fprofile-generate

[steven at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51975

Steven Bosscher <steven at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2012-07-22
                 CC|                            |steven at gcc dot gnu.org
     Ever Confirmed|0                           |1

--- Comment #4 from Steven Bosscher <steven at gcc dot gnu.org> 2012-07-22 00:11:31 UTC ---
I think the transformation should just be disabled completely if the checksums
don't match. Or if you take this path anyway, you could store the checksums on
the struct function instead of re-computing it for every check.