[Bug c/53083] New: gcc bug in moving from the SSE registers back onto the heap.

[Bug c/53083] New: gcc bug in moving from the SSE registers back onto the heap.

[da_fox at mad dot scientist.com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53083

             Bug #: 53083
           Summary: gcc bug in moving from the SSE registers back onto the
                    heap.
    Classification: Unclassified
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: da_fox@mad.scientist.com


After investigation of a suspected graphics-driver bug, it was found that the
issue lay with gcc. Please see
https://bugs.freedesktop.org/show_bug.cgi?id=47559 .

The problem as I understand is that gcc-4.6.0 miscompiles miarc.c from the
xorg-server package if -ftree-vectorize is enabled. This flag is enabled with
-O3.

The following minimal set of CFLAGS worked:
"-g -O2"

Adding the -ftree-vectorize separately, "-g -O2 -ftree-vectorize", caused the
program to misbehave at runtime.

I am running gentoo, the default CFLAGS for my system are:
CFLAGS="-march=native -O3 -pipe -fomit-frame-pointer -fweb -ffast-math
-mtune=native -mfpmath=sse"
I don't know if these match exactly with what is passed by portage for
compiling GCC.

Software versions (gentoo package versions):
sys-devel/gcc-4.6.0 
x11-base/xorg-server-1.11.2-r2

My hardware is a Dell XPS 15 (L502x) laptop with a 'Intel(R) Core(TM) i7-2630QM
CPU @ 2.00GHz'

I additional information is required please let me know.

[Bug c/53083] gcc bug in moving from the SSE registers back onto the heap.

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53083

--- Comment #1 from Uros Bizjak <ubizjak at gmail dot com> 2012-04-23 07:55:29 UTC ---
(In reply to comment #0)

> I additional information is required please let me know.

Please provide all necessary information, as outlined in [1].

[1] http://gcc.gnu.org/bugs/#report

[Bug c/53083] gcc bug in moving from the SSE registers back onto the heap.

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53083

Richard Guenther <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |WAITING
   Last reconfirmed|                            |2012-04-23
     Ever Confirmed|0                           |1

[Bug c/53083] gcc bug in moving from the SSE registers back onto the heap.

[da_fox at mad dot scientist.com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53083

--- Comment #2 from Da Fox <da_fox at mad dot scientist.com> 2012-04-23 12:26:53 UTC ---
Created attachment 27220
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=27220
.i version of miarc.c

I think this is the information that was missing:

The file in question compiled with CFLAGS="-O3 -pipe -fomit-frame-pointer -fweb
-ffast-math -mtune=native -mfpmath=sse" (attached)

gcc build options:
>>> Compiling source in /var/tmp/portage/sys-devel/gcc-4.6.0/work/gcc-4.6.0 ...
 * CFLAGS="-march=native -pipe -mtune=native -O2"
 * CXXFLAGS="-march=native -pipe -mtune=native -O2"
 * Configuring gcc ...
 * configuring for GCC_LANG: c,c++,fortran

 * PREFIX:            /usr
 * BINPATH:            /usr/x86_64-pc-linux-gnu/gcc-bin/4.6.0
 * LIBPATH:            /usr/lib/gcc/x86_64-pc-linux-gnu/4.6.0
 * DATAPATH:        /usr/share/gcc-data/x86_64-pc-linux-gnu/4.6.0
 * STDCXX_INCDIR:    /usr/lib/gcc/x86_64-pc-linux-gnu/4.6.0/include/g++-v4

 * Configuring GCC with: 
 *     --prefix=/usr 
 *     --bindir=/usr/x86_64-pc-linux-gnu/gcc-bin/4.6.0 
 *     --includedir=/usr/lib/gcc/x86_64-pc-linux-gnu/4.6.0/include 
 *     --datadir=/usr/share/gcc-data/x86_64-pc-linux-gnu/4.6.0 
 *     --mandir=/usr/share/gcc-data/x86_64-pc-linux-gnu/4.6.0/man 
 *     --infodir=/usr/share/gcc-data/x86_64-pc-linux-gnu/4.6.0/info 
 *    
--with-gxx-include-dir=/usr/lib/gcc/x86_64-pc-linux-gnu/4.6.0/include/g++-v4 
 *     --host=x86_64-pc-linux-gnu 
 *     --build=x86_64-pc-linux-gnu 
 *     --disable-altivec 
 *     --disable-fixed-point 
 *     --with-ppl 
 *     --with-cloog 
 *     --disable-ppl-version-check 
 *     --with-cloog-include=/usr/include/cloog-ppl 
 *     --enable-lto 
 *     --enable-nls 
 *     --without-included-gettext 
 *     --with-system-zlib 
 *     --disable-werror 
 *     --enable-secureplt 
 *     --disable-multilib 
 *     --disable-libmudflap 
 *     --disable-libssp 
 *     --enable-libgomp 
 *     --with-python-dir=/share/gcc-data/x86_64-pc-linux-gnu/4.6.0/python 
 *     --enable-checking=release 
 *     --disable-libgcj 
 *     --enable-languages=c,c++,fortran 
 *     --enable-shared 
 *     --enable-threads=posix 
 *     --enable-__cxa_atexit 
 *     --enable-clocale=gnu 
 *     --enable-targets=all 
 *     --with-bugurl=http://bugs.gentoo.org/ 
 *     --with-pkgversion=Gentoo 4.6.0 p1.2, pie-0.4.5

Compiler does not seem to output anything relevant (no errors, compilation is
successful).

[Bug c/53083] gcc bug in moving from the SSE registers back onto the heap.

[da_fox at mad dot scientist.com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53083

--- Comment #3 from Da Fox <da_fox at mad dot scientist.com> 2012-04-23 12:28:06 UTC ---
gcc -v output:

Using built-in specs.
COLLECT_GCC=/usr/x86_64-pc-linux-gnu/gcc-bin/4.6.0/gcc
COLLECT_LTO_WRAPPER=/usr/libexec/gcc/x86_64-pc-linux-gnu/4.6.0/lto-wrapper
Target: x86_64-pc-linux-gnu
Configured with: /var/tmp/portage/sys-devel/gcc-4.6.0/work/gcc-4.6.0/configure
--prefix=/usr --bindir=/usr/x86_64-pc-linux-gnu/gcc-bin/4.6.0
--includedir=/usr/lib/gcc/x86_64-pc-linux-gnu/4.6.0/include
--datadir=/usr/share/gcc-data/x86_64-pc-linux-gnu/4.6.0
--mandir=/usr/share/gcc-data/x86_64-pc-linux-gnu/4.6.0/man
--infodir=/usr/share/gcc-data/x86_64-pc-linux-gnu/4.6.0/info
--with-gxx-include-dir=/usr/lib/gcc/x86_64-pc-linux-gnu/4.6.0/include/g++-v4
--host=x86_64-pc-linux-gnu --build=x86_64-pc-linux-gnu --disable-altivec
--disable-fixed-point --with-ppl --with-cloog --disable-ppl-version-check
--with-cloog-include=/usr/include/cloog-ppl --enable-lto --enable-nls
--without-included-gettext --with-system-zlib --disable-werror
--enable-secureplt --disable-multilib --disable-libmudflap --disable-libssp
--enable-libgomp
--with-python-dir=/share/gcc-data/x86_64-pc-linux-gnu/4.6.0/python
--enable-checking=release --disable-libgcj --enable-languages=c,c++,fortran
--enable-shared --enable-threads=posix --enable-__cxa_atexit
--enable-clocale=gnu --enable-targets=all --with-bugurl=http://bugs.gentoo.org/
--with-pkgversion='Gentoo 4.6.0 p1.2, pie-0.4.5'
Thread model: posix
gcc version 4.6.0 (Gentoo 4.6.0 p1.2, pie-0.4.5)

[Bug c/53083] gcc bug in moving from the SSE registers back onto the heap.

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53083

--- Comment #4 from Uros Bizjak <ubizjak at gmail dot com> 2012-04-23 14:50:54 UTC ---
Unfortunately, this is a runtime failure, so we will also need a runtime
testcase (probably minimized) that fails when compiled with options that expose
the bug.

In comment #8 of the linked bugreport, you have testing failures, so this is a
good starting point to locate the problem.

[Bug c/53083] gcc bug in moving from the SSE registers back onto the heap.

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53083

--- Comment #5 from Uros Bizjak <ubizjak at gmail dot com> 2012-04-23 15:29:03 UTC ---
BTW: Can you test if this problem is still present in a newer version of the
compiler (i.e. 4.6.3 or 4.7.0)?

[Bug c/53083] gcc bug in moving from the SSE registers back onto the heap.

[da_fox at mad dot scientist.com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53083

Da Fox <da_fox at mad dot scientist.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|target                      |c

--- Comment #6 from Da Fox <da_fox at mad dot scientist.com> 2012-04-25 07:30:27 UTC ---
(In reply to comment #5)
> BTW: Can you test if this problem is still present in a newer version of the
> compiler (i.e. 4.6.3 or 4.7.0)?

I've tried to take the source-file in question (miarc.c) and make something
self-contained, but I was unable to make it work due to this being an internal
xorg-file, apparently using and depending on internal headers. If you have any
pointers on how to tackle such a situation to convert it to something
selfcontained I'd be happy to try it out.

The latest available gcc-version for my distribution (gentoo) appears to be
4.6.2, which I will try next.

[Bug c/53083] gcc bug in moving from the SSE registers back onto the heap.

[da_fox at mad dot scientist.com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53083

--- Comment #7 from Da Fox <da_fox at mad dot scientist.com> 2012-05-01 07:54:54 UTC ---
(In reply to comment #6)
...
> The latest available gcc-version for my distribution (gentoo) appears to be
> 4.6.2, which I will try next.

This seems to still happen with 4.6.2 at least.

[Bug c/53083] gcc bug in moving from the SSE registers back onto the heap.

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53083

--- Comment #8 from Uros Bizjak <ubizjak at gmail dot com> 2012-05-16 06:30:58 UTC ---
(In reply to comment #6)

> I've tried to take the source-file in question (miarc.c) and make something
> self-contained, but I was unable to make it work due to this being an internal
> xorg-file, apparently using and depending on internal headers. If you have any
> pointers on how to tackle such a situation to convert it to something
> selfcontained I'd be happy to try it out.

First, prepare some test input and call failing function.  Compare its result
with some reference result and abort () if the result is not expected. The
preprocessed file is self-contained, so you could just compile it and link into
executable.

Starting from here, you remove portions of the function that are not needed
(i.e. unreachable if branches, switch cases, etc...) for test input.

Maybe you could ask Gentoo compiler maintainer for help...

[Bug c/53083] gcc bug in moving from the SSE registers back onto the heap.

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53083

--- Comment #9 from Uros Bizjak <ubizjak at gmail dot com> 2012-05-16 06:33:35 UTC ---
(In reply to comment #8)

> > I've tried to take the source-file in question (miarc.c) and make something
> > self-contained, but I was unable to make it work due to this being an internal
> > xorg-file, apparently using and depending on internal headers. If you have any
> > pointers on how to tackle such a situation to convert it to something
> > selfcontained I'd be happy to try it out.
> 
> First, prepare some test input and call failing function.  Compare its result
> with some reference result and abort () if the result is not expected. The
> preprocessed file is self-contained, so you could just compile it and link into
> executable.

You can obtain reference result by running the function, compiled with
non-failing compile flags.

[Bug c/53083] gcc bug in moving from the SSE registers back onto the heap.

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53083

--- Comment #10 from Uros Bizjak <ubizjak at gmail dot com> 2012-05-16 06:36:27 UTC ---
Youo can start from [1], you have a failing output there, just provide some
reference input to failing function.

[1] https://bugs.freedesktop.org/show_bug.cgi?id=47559#c8

[Bug c/53083] gcc bug in moving from the SSE registers back onto the heap.

[ossman at cendio dot se]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53083

--- Comment #11 from Pierre Ossman <ossman at cendio dot se> ---
Created attachment 30419
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=30419&action=edit
main.c

We've also been hitting this, so here is a reduced test case to provoke the
bug. Output in the failing case:

test_spans(nspans=36)
    101,100 + 1
    101,0 + 1
    0,100 + 0
Aborted (core dumped)


And for a correct run:

test_spans(nspans=36)
    101,100 + 1
    101,100 + 1
    101,100 + 1
    101,100 + 1
    101,101 + 1
    101,101 + 1
    101,99 + 1
    101,99 + 1
    101,102 + 1
...

[Bug c/53083] gcc bug in moving from the SSE registers back onto the heap.

[ossman at cendio dot se]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53083

--- Comment #12 from Pierre Ossman <ossman at cendio dot se> ---
Created attachment 30420
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=30420&action=edit
Makefile

[Bug c/53083] gcc bug in moving from the SSE registers back onto the heap.

[ossman at cendio dot se]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53083

--- Comment #13 from Pierre Ossman <ossman at cendio dot se> ---
This was tested using gcc 4.7.2 and gcc 4.5.3.