[Bug bootstrap/51648] New: [4.7 Regression]

[Bug bootstrap/51648] New: [4.7 Regression]

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51648

             Bug #: 51648
           Summary: [4.7 Regression]
    Classification: Unclassified
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: bootstrap
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: jakub@gcc.gnu.org
            Target: x86_64-linux

[Bug bootstrap/51648] [4.7 Regression] Profiledbootstrap failure on x86_64-linux

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51648

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |4.7.0
            Summary|[4.7 Regression]            |[4.7 Regression]
                   |                            |Profiledbootstrap failure
                   |                            |on x86_64-linux

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-12-21 10:32:29 UTC ---
CFLAGS='-O2 -g -fexceptions -fstack-protector --param=ssp-buffer-size=4
-mtune=generic' CXXFLAGS='-O2 -g -fexceptions -fstack-protector
--param=ssp-buffer-size=4 -mtune=generic' \
../configure --prefix=/usr --mandir=/usr/share/man --infodir=/usr/share/info
--with-bugurl=http://bugzilla.redhat.com/bugzilla --enable-bootstrap
--enable-checking=release \
--disable-build-with-cxx --disable-build-poststage1-with-cxx --with-system-zlib
--enable-__cxa_atexit --enable-gnu-unique-object --enable-linker-build-id
--with-linker-hash-style=gnu \
--enable-languages=c,c++,lto --enable-plugin --with-ppl --with-cloog
--with-tune=generic --with-arch_32=i686 --build=x86_64-redhat-linux
make -j8 'BOOT_CFLAGS=-O2 -g -Wall -fexceptions -fstack-protector
--param=ssp-buffer-size=4 -mtune=generic' profiledbootstrap
fails on current trunk, reproduceably (3 times already) with:
../../gcc/gcc.c:8278:1: error: corrupted profile info: profile data is not
flow-consistent
../../gcc/gcc.c:8278:1: error: corrupted profile info: number of executions for
edge 262-263 thought to be -24
../../gcc/gcc.c:8278:1: error: corrupted profile info: number of executions for
edge 262-1 thought to be 24

I had to workaround two (bogus) uninitialized warnings first, will post a patch
for that.

[Bug bootstrap/51648] [4.7 Regression] Profiledbootstrap failure on x86_64-linux

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51648

--- Comment #2 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-12-21 11:41:04 UTC ---
Why -fexceptions?  Don't we explicitely _disable_ exceptions?

[Bug bootstrap/51648] [4.7 Regression] Profiledbootstrap failure on x86_64-linux

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51648

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-12-21 11:52:55 UTC ---
-O2 -g -fexceptions -fstack-protector --param=ssp-buffer-size=4 -mtune=generic
are our standard distro flags.  The intent is that e.g. C code that uses
pthread_cleanup_{push,pop} will be more efficient (use cleanup attribute) if we
do that.  Packages that for some reason want to build with -fno-exceptions
should IMHO put that -fno-exceptions after user supplied flags to override it.
If I understand right our gcc/Makefile.in we do that, $(NOEXCEPT_FLAGS) is put
after TCFLAGS etc. in AM_CXXFLAGS, so it overrides it right.
In this case (C build instead of C++, when there are still no advantages of
using C++ in GCC, I thought it doesn't hurt to configure that way) -fno-rtti
-fno-exceptions isn't considered to be a valid option during configure (well,
-fno-exceptions is, but -fno-rtti is not for C and both are tried at the same
time), so we build with -fexceptions.  That shouldn't make a significant
difference for C though, unless the cleanup attribute is used somewhere.

Anyway, the reason why I'm filing this is because this looks very much like
lots of the issues that were not really reproduceable during profiledbootstrap
in the past, now it seems to be reproduceable.

[Bug bootstrap/51648] [4.7 Regression] Profiledbootstrap failure on x86_64-linux

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51648

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-12-21 15:18:57 UTC ---
Created attachment 26163
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=26163
gcc.i

/some/path/gcc/cc1 -fpreprocessed -quiet -quiet -dumpbase gcc.c -mtune=generic
-march=x86-64 -g -O2 -fexceptions -fstack-protector -fprofile-use --param
ssp-buffer-size=4 /tmp/gcc.i -o /tmp/gcc.s
reproduces this with the attached gcc.i and gcc.gcda.  Of course the problem
might be already at the -fprofile-generate phase or at runtime when generating
the gcc.gcda file.

[Bug bootstrap/51648] [4.7 Regression] Profiledbootstrap failure on x86_64-linux

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51648

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-12-21 15:20:05 UTC ---
Created attachment 26164
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=26164
gcc.gcda

[Bug bootstrap/51648] [4.7 Regression] Profiledbootstrap failure on x86_64-linux

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51648

--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-12-21 15:23:57 UTC ---
BTW, it can be also reproduced without that -j8 make option, i.e. it is not any
kind of race condition or similar.

[Bug bootstrap/51648] [4.7 Regression] Profiledbootstrap failure on x86_64-linux

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51648

--- Comment #7 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-12-21 18:15:23 UTC ---
Ok, so when the -fprofile-generate compiled xgcc is linked, it is enough
to (assuming gcc.i from the attachment here is in /tmp/):
$ rm gcc.gcda; ./xgcc -qversion
$ cp -a gcc.gcda /tmp/
$ cd /tmp
$ /usr/src/gcc/obj8/gcc/cc1 -fpreprocessed -quiet -quiet -dumpbase gcc.c \
-mtune=generic -march=x86-64 -g -O2 -fexceptions -fstack-protector
-fprofile-use \
--param ssp-buffer-size=4 /tmp/gcc.i -o /tmp/gcc.s
../../gcc/gcc.c: In function ‘main’:
../../gcc/gcc.c:8278:1: error: corrupted profile info: profile data is not
flow-consistent
../../gcc/gcc.c:8278:1: error: corrupted profile info: number of executions for
edge 262-263 thought to be -1
../../gcc/gcc.c:8278:1: error: corrupted profile info: number of executions for
edge 262-1 thought to be 1

[Bug bootstrap/51648] [4.7 Regression] Profiledbootstrap failure on x86_64-linux

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51648

--- Comment #8 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-12-21 19:01:22 UTC ---
So, shorter/faster reproducer, using a snapshot from ~ today (tried with
r182599)
on x86_64-linux:
../configure --enable-languages=c --disable-bootstrap \
  --disable-libitm
make -j8
cd gcc
./cc1 -fpreprocessed -quiet -quiet -dumpbase gcc.c \
  -g -O2 -fexceptions -fprofile-generate /tmp/gcc.i -o /tmp/gcc.s
gcc -c -o gcc.o /tmp/gcc.s
gcc   -g -pedantic -fno-common -o xgccmy gccmy.o ggc-none.o \
  gccspec.o driver-i386.o libcommon-target.a libcommon.a \
  ../libcpp/libcpp.a   ../libiberty/libiberty.a \
  ../libdecnumber/libdecnumber.a libgcov.a
rm -f gcc.gcda
./xgccmy -qversion
./cc1 -fpreprocessed -quiet -quiet -dumpbase gcc.c \
  -g -O2 -fexceptions -fprofile-use /tmp/gcc.i -o /tmp/gcc.s

The steps before cd gcc could be left out if you have the tree built with other
options already, all it matters is if more-less similar *.o/*.a listed on the
command line are built, including libgcov.a.  /tmp/gcc.i contains the copy of
the first attachment here.

[Bug bootstrap/51648] [4.7 Regression] Profiledbootstrap failure on x86_64-linux

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51648

--- Comment #9 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-12-23 08:21:18 UTC ---
Created attachment 26169
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=26169
gcc.i

Somewhat delta reduced gcc.i which still triggers it.  Surprisingly even
removing an unused function definition from it makes the bug go away, will have
to review the code generation differences.

[Bug bootstrap/51648] [4.7 Regression] Profiledbootstrap failure on x86_64-linux

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51648

--- Comment #10 from Jakub Jelinek <jakub at gcc dot gnu.org> 2012-01-04 16:38:37 UTC ---
That delta script had a typo (was using gcc.o instead of gccmy.o), thus the
resulting preprocessed file is irrelevant.

Here is an actually reduced testcase:

rm -f pr51648-1.gcda pr51648-2.gcda
./cc1 -fpreprocessed -quiet -g -O2 -fexceptions -fprofile-generate \
  pr51648-1.c -o pr51648-1.s
./xgcc -B ./ -o pr51648 -fprofile-generate pr51648-1.s pr51648-2.c
./pr51648
./cc1 -fpreprocessed -quiet -g -O2 -fexceptions -fprofile-use \
  pr51648-1.c -o pr51648-2.s
gives:
pr51648-1.c: In function ‘main’:
pr51648-1.c:37:1: error: corrupted profile info: profile data is not
flow-consistent
pr51648-1.c:37:1: error: corrupted profile info: number of executions for edge
4-5 thought to be 1
pr51648-1.c:37:1: error: corrupted profile info: number of executions for edge
4-9 thought to be -1

I'll attach the two sources.

[Bug bootstrap/51648] [4.7 Regression] Profiledbootstrap failure on x86_64-linux

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51648

--- Comment #11 from Jakub Jelinek <jakub at gcc dot gnu.org> 2012-01-04 16:39:44 UTC ---
Created attachment 26240
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=26240
pr51648-1.c

[Bug bootstrap/51648] [4.7 Regression] Profiledbootstrap failure on x86_64-linux

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51648

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #26163|0                           |1
        is obsolete|                            |
  Attachment #26164|0                           |1
        is obsolete|                            |
  Attachment #26169|0                           |1
        is obsolete|                            |

--- Comment #12 from Jakub Jelinek <jakub at gcc dot gnu.org> 2012-01-04 16:40:34 UTC ---
Created attachment 26241
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=26241
pr51648-2.c

[Bug bootstrap/51648] [4.7 Regression] Profiledbootstrap failure on x86_64-linux

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51648

--- Comment #13 from Jakub Jelinek <jakub at gcc dot gnu.org> 2012-01-04 21:26:11 UTC ---
I think the problem is that we have a noreturn call (in the short testcase
fatal_error), which doesn't have EDGE_FAKE to exit added by
add_noreturn_fake_exit_edges because it has successor edge (EDGE_EH), and
isn't added by gimple_flow_call_edges_add, because that function ignores
noreturn calls.  So, either we don't ignore noreturn calls that have any
non-fake edges in gimple_flow_call_edges_add, or change
add_noreturn_fake_exit_edges.
Honza, what do you prefer?
Changing gimple_flow_call_edges_add will only affect profile feedback I think,
while add_noreturn_fake_exit_edges will affect several other passes (sm, gcse,
...).

[Bug bootstrap/51648] [4.7 Regression] Profiledbootstrap failure on x86_64-linux

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51648

--- Comment #14 from Jakub Jelinek <jakub at gcc dot gnu.org> 2012-01-04 21:40:41 UTC ---
Created attachment 26243
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=26243
gcc47-pr51648.patch

Untested patch that implements this in tree-cfg.c (as I wrote, another
alternative is in cfganal.c, and another perhaps in branch_prob, we look for
the last stmt anyway, so just notting when it is a noreturn call and setting
need_exit_edge in that case should be enough too.

[Bug bootstrap/51648] [4.7 Regression] Profiledbootstrap failure on x86_64-linux

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51648

--- Comment #15 from Jakub Jelinek <jakub at gcc dot gnu.org> 2012-01-05 09:32:56 UTC ---
Author: jakub
Date: Thu Jan  5 09:32:52 2012
New Revision: 182902

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=182902
Log:
    PR bootstrap/51648
    * tree-cfg.c (need_fake_edge_p): Return true also for noreturn
    calls that have any non-fake successor edges.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/tree-cfg.c

[Bug bootstrap/51648] [4.7 Regression] Profiledbootstrap failure on x86_64-linux

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51648

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |FIXED
         AssignedTo|unassigned at gcc dot       |jakub at gcc dot gnu.org
                   |gnu.org                     |

--- Comment #16 from Jakub Jelinek <jakub at gcc dot gnu.org> 2012-01-05 17:16:48 UTC ---
Fixed.