[Bug c/26528] New: [4.2 regression] gcc miscompiles FFTW 3.1

[Bug c/26528] New: [4.2 regression] gcc miscompiles FFTW 3.1

[martin at mpa-garching dot mpg dot de]

Current mainline gcc appears to miscompile FFTW:

wget http://fftw.org/fftw-3.1.tar.gz
tar xvzf fftw-3.1.tar.gz
cd fftw-3.1
/scratch/fftw-3.1> gcc -v
Using built-in specs.
Target: i686-pc-linux-gnu
Configured with: /scratch/gcc/configure --quiet
--prefix=/afs/mpa/data/martin/ugcc --enable-languages=c++,fortran
--with-gmp=/usr/local/appl/gmp-4.1.4 --enable-checking=release
--without-makeinfo
Thread model: posix
gcc version 4.2.0 20060301 (experimental)
/scratch/fftw-3.1> ./configure --enable-portable-binary
[...]
/scratch/fftw-3.1> make check
[...]
Executing "/scratch/fftw-3.1/tests/bench --verbose=1   --verify
ok12o10x6e10x11o00v7 --verify ik12o10x
6e10x11o00v7 --verify obr3v79 --verify ibr3v79 --verify ofr3v79 --verify
ifr3v79 --verify //obc3v79 --
verify //ibc3v79 --verify //ofc3v79 --verify //ifc3v79 --verify obc3v79
--verify ibc3v79 --verify ofc3
v79 --verify ifc3v79 --verify ok2hx5e00x3e10x7o11*6 --verify
ik2hx5e00x3e10x7o11*6 --verify //obr10x8x
33 --verify //ofr10x8x33 --verify obr10x8x33 --verify ibr10x8x33 --verify
ofr10x8x33 --verify ifr10x8x
33 --verify //obc10x8x33 --verify //ibc10x8x33 --verify //ofc10x8x33 --verify
//ifc10x8x33 --verify ob
c10x8x33 --verify ibc10x8x33 --verify ofc10x8x33 --verify ifc10x8x33 --verify
ok11o00*45 --verify ik11
o00*45 --verify obr6x4v7 --verify ibr6x4v7 --verify ofr6x4v7 --verify ifr6x4v7
--verify //obc6x4v7 --v
erify //ibc6x4v7 --verify //ofc6x4v7 --verify //ifc6x4v7 --verify obc6x4v7
--verify ibc6x4v7 --verify 
ofc6x4v7 --verify ifc6x4v7"
2.65417e-16 4.02852e-15 3.09912e-16
2.76156e-16 2.52972e-15 2.83568e-16
2.17601e-16 3.37283e-16 2.97118e-16
2.03315e-16 3.37283e-16 2.49555e-16
2.33093e-16 1.49904e-16 3.95845e-16
2.09027e-16 1.49904e-16 3.70199e-16
1.94987e-16 2.24855e-16 4.92525e-16
1.97747e-16 2.24855e-16 5.28462e-16
1.71722e-16 2.24855e-16 4.74119e-16
2.0242e-16 2.24855e-16 5.51321e-16
1.85902e-16 2.24855e-16 4.94966e-16
2.13346e-16 2.24855e-16 4.9348e-16
2.12188e-16 2.24855e-16 5.26624e-16
2.12652e-16 2.24855e-16 4.99894e-16
2.75703e-16 2.60691e-15 2.91529e-16
4.46832e-16 4.43175e-15 3.46477e-16
FAILED /scratch/fftw-3.1/tests/bench:  --verify ok12o10x6e10x11o00v7 --verify
ik12o10x6e10x11o00v7 --v
erify obr3v79 --verify ibr3v79 --verify ofr3v79 --verify ifr3v79 --verify
//obc3v79 --verify //ibc3v79
 --verify //ofc3v79 --verify //ifc3v79 --verify obc3v79 --verify ibc3v79
--verify ofc3v79 --verify ifc
3v79 --verify ok2hx5e00x3e10x7o11*6 --verify ik2hx5e00x3e10x7o11*6 --verify
//obr10x8x33 --verify //of
r10x8x33 --verify obr10x8x33 --verify ibr10x8x33 --verify ofr10x8x33 --verify
ifr10x8x33 --verify //ob
c10x8x33 --verify //ibc10x8x33 --verify //ofc10x8x33 --verify //ifc10x8x33
--verify obc10x8x33 --verif
y ibc10x8x33 --verify ofc10x8x33 --verify ifc10x8x33 --verify ok11o00*45
--verify ik11o00*45 --verify 
obr6x4v7 --verify ibr6x4v7 --verify ofr6x4v7 --verify ifr6x4v7 --verify
//obc6x4v7 --verify //ibc6x4v7
 --verify //ofc6x4v7 --verify //ifc6x4v7 --verify obc6x4v7 --verify ibc6x4v7
--verify ofc6x4v7 --verif
y ifc6x4v7
received signal 11
make[2]: *** [check-local] Error 1
make[2]: Leaving directory `/scratch/fftw-3.1/tests'
make[1]: *** [check-am] Error 2
make[1]: Leaving directory `/scratch/fftw-3.1/tests'
make: *** [check-recursive] Error 1


The current 4.1 and 4.0 branches pass the FFTW tests. Gomp-branch fails.

I know that this is not a very useful bug report, but I have no idea how
to reduce the problem. A regression hunt might at least localise the
problematic
commit and help to decide whether gcc or FFTW is to blame.


-- 
           Summary: [4.2 regression] gcc miscompiles FFTW 3.1
           Product: gcc
           Version: 4.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: martin at mpa-garching dot mpg dot de
 GCC build triplet: i686-pc-linux-gnu
  GCC host triplet: i686-pc-linux-gnu
GCC target triplet: i686-pc-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26528

[Bug c/26528] [4.2 regression] gcc miscompiles FFTW 3.1

[martin at mpa-garching dot mpg dot de]

------- Comment #1 from martin at mpa-garching dot mpg dot de  2006-03-02 11:00 -------
I just checked that

CFLAGS="-O3" ./configure --enable-portable-binary

fails, but

CFLAGS="-O3" ./configure --enable-portable-binary

works as it should.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26528

[Bug c/26528] [4.2 regression] gcc miscompiles FFTW 3.1

[martin at mpa-garching dot mpg dot de]

------- Comment #2 from martin at mpa-garching dot mpg dot de  2006-03-02 11:01 -------
(In reply to comment #1)
> I just checked that
> 
> CFLAGS="-O3" ./configure --enable-portable-binary
> 
> fails, but
> 
> CFLAGS="-O3" ./configure --enable-portable-binary

Sorry for the typo here. I meant, of course,

CFLAGS="-O2" ./configure --enable-portable-binary


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26528

[Bug middle-end/26528] [4.2 regression] gcc miscompiles FFTW 3.1

[pinskia at gcc dot gnu dot org]

------- Comment #3 from pinskia at gcc dot gnu dot org  2006-03-02 12:15 -------
Do you have a simple testcase?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26528

[Bug middle-end/26528] [4.2 regression] gcc miscompiles FFTW 3.1

[martin at mpa-garching dot mpg dot de]

------- Comment #4 from martin at mpa-garching dot mpg dot de  2006-03-02 12:18 -------
(In reply to comment #3)
> Do you have a simple testcase?

I wish I had :(
At the moment I'm trying to find out which optimisation flags are causing the
trouble. The current minimum set of flags to reproduce it is
"-O1 -ftree-vrp -finline-functions".
More to come...


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26528

[Bug middle-end/26528] [4.2 regression] gcc miscompiles FFTW 3.1

[pinskia at gcc dot gnu dot org]

------- Comment #5 from pinskia at gcc dot gnu dot org  2006-03-02 13:40 -------
(In reply to comment #4)
> "-O1 -ftree-vrp -finline-functions".
> More to come...

Can you try "-O3 -fwrapv"?  It might be the source have undefined code in it
for signed overflow and changes to VRP exposed it.

Also do you know which file is being miscompiled?  If so can you attach the
dump generated by doing -O3 -fdump-tree-vrp-details?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26528

[Bug middle-end/26528] [4.2 regression] gcc miscompiles FFTW 3.1

[martin at mpa-garching dot mpg dot de]

------- Comment #6 from martin at mpa-garching dot mpg dot de  2006-03-02 13:52 -------
(In reply to comment #5)

> Can you try "-O3 -fwrapv"?  It might be the source have undefined code in it
> for signed overflow and changes to VRP exposed it.

Bingo! Thanks for this hint, I wouldn't have found that alone!

> Also do you know which file is being miscompiled?  If so can you attach the
> dump generated by doing -O3 -fdump-tree-vrp-details?

Unfortunately I don't know which file causes the problem. I'm not a FFTW
developer, I just stumbled on this by accident.

If I understand correctly, this is most likely a bug in FFTW, right? If so,
I'll report it to FFTW people.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26528

[Bug middle-end/26528] [4.2 regression] gcc miscompiles FFTW 3.1

[martin at mpa-garching dot mpg dot de]

------- Comment #7 from martin at mpa-garching dot mpg dot de  2006-03-02 14:29 -------
(In reply to comment #6)

> If I understand correctly, this is most likely a bug in FFTW, right? If so,
> I'll report it to FFTW people.

On the other hand, gcc 4.1 also has VRP, but it seems to work fine. Has VRP
become more aggressive, or might it have a new bug?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26528

[Bug middle-end/26528] [4.2 regression] gcc miscompiles FFTW 3.1

[pinskia at gcc dot gnu dot org]

------- Comment #8 from pinskia at gcc dot gnu dot org  2006-03-02 16:10 -------
(In reply to comment #7)
> On the other hand, gcc 4.1 also has VRP, but it seems to work fine. Has VRP
> become more aggressive, or might it have a new bug?

VRP has become more aggressive or it might be still a bug.  There is not enough
information in the bug report to tell.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26528

[Bug middle-end/26528] [4.2 regression] gcc miscompiles FFTW 3.1

[martin at mpa-garching dot mpg dot de]

------- Comment #9 from martin at mpa-garching dot mpg dot de  2006-03-03 10:11 -------
I managed to narrow down the location of the problem:

The segfault occurs in libbench2/verify-lib.c, in
static void assign_conj(C *Ac, C *A, int rank, const bench_iodim *dim, int
stride);

I added the following line at the very beginning of this function for
diagnosis:

printf("assign_conj %p %p %d %d %d\n",Ac, A,rank,stride, dim[rank - 1].n);

Then I compiled with

CFLAGS="-O1 -ftree-vrp -finline-functions" ./configure
make

and started the following test:
tests/bench --verify obr10x5x13*6

The output is

assign_conj 0x80e8e20 0x80e6da0 0 650 0
assign_conj 0x80e8a10 0x80e71b0 0 650 0
assign_conj 0x80e8600 0x80e75c0 0 650 0
assign_conj 0x80e81f0 0x80e79d0 0 650 0
assign_conj 0x80e6cd0 0x80e6a60 1 65 10
assign_conj 0x80e6cd0 0x80e6a60 0 650 0
assign_conj 0x80e9160 0x80e6e70 0 650 0
assign_conj 0x880e9153 0x80e7280 0 650 0
Segmentation fault

The first pointer value in the last line definitely looks wrong, and I can't
explain it by looking at the code.
This does feel like a problem in VRP, but I can't prove it...


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26528

[Bug middle-end/26528] [4.2 regression] gcc miscompiles FFTW 3.1

[pinskia at gcc dot gnu dot org]

-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |pinskia at gcc dot gnu dot
                   |                            |org
   Target Milestone|---                         |4.1.2


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26528

[Bug middle-end/26528] [4.2 regression] gcc miscompiles FFTW 3.1

[pinskia at gcc dot gnu dot org]

------- Comment #10 from pinskia at gcc dot gnu dot org  2006-03-03 14:09 -------
Can you try -O1 -fno-ivopts -ftree-vrp -finline-functions?

I want to see if this is an interaction between VRP and IVopts, if it is there
is a bug about this issue already filed and I bet it is the same bug.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26528

[Bug middle-end/26528] [4.2 regression] gcc miscompiles FFTW 3.1

[pinskia at gcc dot gnu dot org]

------- Comment #11 from pinskia at gcc dot gnu dot org  2006-03-03 14:09 -------
The other bug is PR 26304.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26528

[Bug middle-end/26528] [4.2 regression] gcc miscompiles FFTW 3.1

[martin at mpa-garching dot mpg dot de]

------- Comment #12 from martin at mpa-garching dot mpg dot de  2006-03-03 14:22 -------
(In reply to comment #10)
> Can you try -O1 -fno-ivopts -ftree-vrp -finline-functions?
> 
> I want to see if this is an interaction between VRP and IVopts, if it is there
> is a bug about this issue already filed and I bet it is the same bug.

When I add "-fno-ivopts" the testcase works again, so I think you are right. So
I'll just wait until PR 26304 is resolved and then check again.
Thanks a lot for your help!

BTW, you probably meant '4.2.0', instead of '4.1.2' as target milestone, right?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26528

[Bug middle-end/26528] [4.2 regression] gcc miscompiles FFTW 3.1

[pinskia at gcc dot gnu dot org]

------- Comment #13 from pinskia at gcc dot gnu dot org  2006-03-03 14:24 -------
(In reply to comment #12)
> BTW, you probably meant '4.2.0', instead of '4.1.2' as target milestone, right?
Yes an off by one error on my part.


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|4.1.2                       |4.2.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26528

[Bug middle-end/26528] [4.2 regression] gcc miscompiles FFTW 3.1

[martin at mpa-garching dot mpg dot de]

------- Comment #14 from martin at mpa-garching dot mpg dot de  2006-05-16 12:49 -------
Even with PR 26304 fixed, the testsuite still crashes with a segfault in the
same place. I'm using FFTW 3.1.1 now, but the testing procedure is the same.
Again, the problem disappears with "-fno-ivopts".


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26528

[Bug middle-end/26528] [4.2 regression] gcc miscompiles FFTW 3.1

[mmitchel at gcc dot gnu dot org]

------- Comment #15 from mmitchel at gcc dot gnu dot org  2006-06-04 17:55 -------
This is a P1 -- but WAITING for a testcase.


-- 

mmitchel at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|0                           |1
           Priority|P3                          |P1
   Last reconfirmed|0000-00-00 00:00:00         |2006-06-04 17:55:31
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26528

[Bug middle-end/26528] [4.2 regression] gcc miscompiles FFTW 3.1

[mmitchel at gcc dot gnu dot org]

-- 

mmitchel at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |WAITING


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26528

[Bug middle-end/26528] [4.2 regression] gcc miscompiles FFTW 3.1

[martin at mpa-garching dot mpg dot de]

------- Comment #16 from martin at mpa-garching dot mpg dot de  2006-06-05 10:12 -------
(In reply to comment #15)
> This is a P1 -- but WAITING for a testcase.

I tried to reproduce the problem with today's mainline (20060605), but it was
gone, so I think the bug can be closed. In any case I will keep my eyes on this
and let you know if it starts failing again.
Should the bug be closed as FIXED or WORKSFORME?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26528

[Bug middle-end/26528] [4.2 regression] gcc miscompiles FFTW 3.1

[steven at gcc dot gnu dot org]

------- Comment #17 from steven at gcc dot gnu dot org  2006-06-05 10:27 -------
Following comment #16, closing as FIXED.


-- 

steven at gcc dot gnu dot org changed:

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


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26528