What you do is look at the gimple before pass_ipa_tree_profile at profile-use
and profile-generate time, they should have matching CFGs.  If they don't
then that's the bug.

Also confirmed on x86_64-unknown-linux-gnu.  The execution numbers are exactly the same as on ia64.

Also happens with --enable-checking=release on ia64:


../../trunk/libcpp/macro.c: In function 'cpp_macro_definition':
../../trunk/libcpp/macro.c:2123:1: error: corrupted profile info: profile data is not flow-consistent
../../trunk/libcpp/macro.c:2123:1: error: corrupted profile info: number of executions for edge 45-59 thought to be -104
../../trunk/libcpp/macro.c:2123:1: error: corrupted profile info: number of executions for edge 45-1 thought to be 104
../../trunk/libcpp/macro.c:2123:1: error: corrupted profile info: number of executions for edge 55-56 thought to be 10
../../trunk/libcpp/macro.c:2123:1: error: corrupted profile info: number of executions for edge 55-57 thought to be -5
make[3]: *** [macro.o] Error 1
make[3]: *** Waiting for unfinished jobs....
make[3]: Leaving directory `/home/stevenb/devel/build-test/libcpp'
make[2]: *** [all-stagefeedback-libcpp] Error 2
make[2]: Leaving directory `/home/stevenb/devel/build-test'
make[1]: *** [stagefeedback-bubble] Error 2
make[1]: Leaving directory `/home/stevenb/devel/build-test'
make: *** [profiledbootstrap] Error 2
stevenb@iowa:~/devel/build-test$ cat gcc/configargs.h 
/* Generated automatically. */
static const char configuration_arguments[] = "../trunk/configure --with-gmp=/home/stevenb --with-mpfr=/home/stevenb --enable-languages=c --enable-checking=release --disable-nls --disable-libmudflap --disable-libssp";
static const char thread_model[] = "posix";

Also on alpha, you will need [1] to get this far.

Using built-in specs.
COLLECT_GCC=./prev-gcc/xgcc
Target: alphaev68-unknown-linux-gnu
Configured with: ../gcc-svn/trunk/configure --enable-languages=c,c++,fortran
Thread model: posix
gcc version 4.6.0 20101120 (experimental) [trunk revision 166974] (GCC) 

Marked as 4.6 regression.

[1] http://gcc.gnu.org/ml/gcc-patches/2010-11/msg02118.html

It is caused by revision revision 164986:

http://gcc.gnu.org/ml/gcc-cvs/2010-10/msg00168.html

Created attachment 22514 [details]
preprocessed file

Created attachment 22516 [details]
corresponding profile

~/gcc-build/prev-gcc/cc1 -g -O2 -fprofile-use -quiet macro.i
../../gcc-svn/trunk/libcpp/macro.c: In function ‘cpp_macro_definition’:
../../gcc-svn/trunk/libcpp/macro.c:2123:1: error: corrupted profile info: profile data is not flow-consistent
../../gcc-svn/trunk/libcpp/macro.c:2123:1: error: corrupted profile info: number of executions for edge 45-59 thought to be -104
../../gcc-svn/trunk/libcpp/macro.c:2123:1: error: corrupted profile info: number of executions for edge 45-1 thought to be 104
../../gcc-svn/trunk/libcpp/macro.c:2123:1: error: corrupted profile info: number of executions for edge 55-56 thought to be 10
../../gcc-svn/trunk/libcpp/macro.c:2123:1: error: corrupted profile info: number of executions for edge 55-57 thought to be -5

I can't reproduce it on x86_64-unknown-linux-gnu with

./configure --disable-checking --enable-languages=c

or with

./configure --enable-languages=c,c++,fortran

I can reproduce it with the attached testcase and stage1 xgcc though:

> ./xgcc -B. -O2 -g -fprofile-use -c macro.i 
../../gcc-svn/trunk/libcpp/macro.c: In function 'cpp_macro_definition':
../../gcc-svn/trunk/libcpp/macro.c:2123:1: error: corrupted profile info: profile data is not flow-consistent
../../gcc-svn/trunk/libcpp/macro.c:2123:1: error: corrupted profile info: number of executions for edge 45-59 thought to be -104
../../gcc-svn/trunk/libcpp/macro.c:2123:1: error: corrupted profile info: number of executions for edge 45-1 thought to be 104
../../gcc-svn/trunk/libcpp/macro.c:2123:1: error: corrupted profile info: number of executions for edge 55-56 thought to be 10
../../gcc-svn/trunk/libcpp/macro.c:2123:1: error: corrupted profile info: number of executions for edge 55-57 thought to be -5

would be nice to know if

BOOT_CFLAGS="-O2 -g -fno-inline" ./configure --enable-languages=c
make profiledbootstrap BOOT_CFLAGS="-O2 -g -fno-inline"

also fails, alternatively s/-g -fno-inline/-g0/ or s/-fno-inline/-fno-partial-inlining.

The CFGs match, the spanning tree as well.  Instrumenting needs to split
two edges, but that shouldn't matter and happens elsewhere as well.

But we have in .optimized (of the -fprofile-generate case)

<bb 11>:
  pretmp.676_490 = (unsigned int) D.8874_415;
  pretmp.677_491 = macro_39->params;
  pretmp.681_496 = *.LPBX1[67];

<bb 12>:
  # len_429 = PHI <len_55(12), len_44(11)>
  # ivtmp.698_171 = PHI <ivtmp.698_92(12), 0(11)>
  D.8871_52 = MEM[base: pretmp.677_491, index: ivtmp.698_171, step: 8, offset: 0B];
  D.8872_53 = D.8871_52->ident.len;
  D.8873_54 = D.8872_53 + 1;
  len_55 = D.8873_54 + len_429;
  # DEBUG len => len_55
  # DEBUG i => D#11
  # DEBUG len => len_55
  # DEBUG i => D#11
  ivtmp.698_92 = ivtmp.698_171 + 1;
  i_64 = (unsigned int) ivtmp.698_92;
  if (i_64 < pretmp.676_490)
    goto <bb 12>;
  else
    goto <bb 13>;

<bb 13>:
  D.11695_474 = pretmp.681_496 + 1;
  D.11697_493 = pretmp.676_490 + 4294967295;
  D.11698_127 = (long int) D.11697_493;
  *.LPBX1_I_lsm.686_486 = D.11695_474 + D.11698_127;
  *.LPBX1[67] = *.LPBX1_I_lsm.686_486;

which at least looks odd to me and produced by SCCP (supposedly
*.LPBX1[67] + 1 + (long)(niter - 1), so not 100% invalid).

Soooo ... for those who can reproduce it, can you try with -fno-tree-scev-cprop?

I'll wrap my brain around that above computation tomorrow ;)

Created attachment 22517 [details]
-fno-inline preprocessed source

Created attachment 22518 [details]
corresponding -fno-inline profile

~/gcc-build/prev-gcc/cc1 -g -O2 -fno-inline -fprofile-use -quiet macro.i
../../gcc-svn/trunk/libcpp/macro.c: In function ‘cpp_macro_definition’:
../../gcc-svn/trunk/libcpp/macro.c:2123:1: error: corrupted profile info: profile data is not flow-consistent
../../gcc-svn/trunk/libcpp/macro.c:2123:1: error: corrupted profile info: number of executions for edge 45-59 thought to be -104
../../gcc-svn/trunk/libcpp/macro.c:2123:1: error: corrupted profile info: number of executions for edge 45-1 thought to be 104
../../gcc-svn/trunk/libcpp/macro.c:2123:1: error: corrupted profile info: number of executions for edge 55-56 thought to be 10
../../gcc-svn/trunk/libcpp/macro.c:2123:1: error: corrupted profile info: number of executions for edge 55-57 thought to be -5

glibc data form F14:

./libc.so.6 
GNU C Library development release version 2.12.90, by Roland McGrath et al.
Copyright (C) 2010 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.
There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A
PARTICULAR PURPOSE.
Compiled by GNU CC version 4.5.1 20100924 (Red Hat 4.5.1-4).
Compiled on a Linux 2.6.35 system on 2010-11-12.
Available extensions:
	Support for some architectures added on, not maintained in glibc core.
	The C stubs add-on version 2.1.2.
	crypt add-on version 2.1 by Michael Glad and others
	GNU Libidn by Simon Josefsson
	Native POSIX Threads Library by Ulrich Drepper et al
	BIND-8.2.3-T5B
	RT using linux kernel aio
libc ABIs: UNIQUE IFUNC
For bug reporting instructions, please see:
<http://www.gnu.org/software/libc/bugs.html>.

(In reply to comment #9)

> Soooo ... for those who can reproduce it, can you try with
> -fno-tree-scev-cprop?

Bootstrap also fails with -fno-tree-scev-cprop :(

The configure/build commands were similar to ones in comment 8.

Created attachment 22520 [details]
dump from stageprofile

Created attachment 22521 [details]
dump from stagefedback

In your dump files there is a CFG inconsistency that I can't reproduce.  At
least basic block numbers are not consistent.

I also see DEBUG insns in stagefeedback but not stageprofile (my dumps
have DEBUG insns in both dump files).  Which hints at for example:

<bb 17>:
  PROF_edge_counter.65_257 = *.LPBX1[69];
  PROF_edge_counter.65_258 = PROF_edge_counter.65_257 + 1;
  *.LPBX1[69] = PROF_edge_counter.65_258;
  count_61 = macro_real_token_count (macro_39);

<bb 18>:
  PROF_edge_counter.65_259 = *.LPBX1[70];
  PROF_edge_counter.65_260 = PROF_edge_counter.65_259 + 1;
  *.LPBX1[70] = PROF_edge_counter.65_260;
  goto <bb 30>;

vs.

<bb 17>:
Invalid sum of outgoing probabilities 0.0%
Invalid sum of outgoing counts 0, should be 211
  count_61 = macro_real_token_count (macro_39);
  # DEBUG count => count_61
  # DEBUG i => 0
  goto <bb 29>;

it would be interesting to see when the CFGs start to differ.  Can you
also check logs if really -g0 is used at stageprofile and -g at
stagefeedback.  I'm trying to look why I have -g consistently.

Please try

Index: gcc/profile.c
===================================================================
--- gcc/profile.c       (revision 167127)
+++ gcc/profile.c       (working copy)
@@ -940,7 +943,8 @@ branch_prob (void)
          for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi); gsi_prev (&gsi))
            {
              last = gsi_stmt (gsi);
-             if (gimple_has_location (last))
+             if (!is_gimple_debug (last)
+                 && gimple_has_location (last))
                break;
            }
 
we do CFG manipulations based on debug stmts (verified with a later
assert that triggers on -O2 -g -fprofile-use).

It's of course unfortunate that we do this BB splitting at all when
not doing coverage test.

> we do CFG manipulations based on debug stmts (verified with a later
> assert that triggers on -O2 -g -fprofile-use).
> 
> It's of course unfortunate that we do this BB splitting at all when
> not doing coverage test.

The BB splitting is neccesary so the CFG can be solved (i.e. you have explicit edges to exit when program can terminate).

Honza

I started profiledbootstrap tracking:

http://gcc.gnu.org/ml/gcc-regression/2010-11/msg00378.html

(In reply to comment #17)
> Please try
> 
> Index: gcc/profile.c
> ===================================================================
> --- gcc/profile.c       (revision 167127)
> +++ gcc/profile.c       (working copy)
> @@ -940,7 +943,8 @@ branch_prob (void)
>           for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi); gsi_prev (&gsi))
>             {
>               last = gsi_stmt (gsi);
> -             if (gimple_has_location (last))
> +             if (!is_gimple_debug (last)
> +                 && gimple_has_location (last))
>                 break;
>             }

Great, this bootstrapped OK for -enable-languages=c.

Author: rguenth
Date: Fri Nov 26 09:49:47 2010
New Revision: 167171

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=167171
Log:
2010-11-26  Richard Guenther  <rguenther@suse.de>

	PR tree-optimization/46528
	PR debug/46338
	* profile.c (branch_prob): Do not split blocks based on locations
	from debug statements.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/profile.c

Fixed.

Author: jakub
Date: Mon Dec  6 19:30:31 2010
New Revision: 167515

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=167515
Log:
	PR tree-optimization/46528
	PR debug/46338
	* profile.c (branch_prob): Make sure last is never set to a debug
	stmt.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/profile.c