[Bug rtl-optimization/14863] [3.4 regression] unit-at-a-time causes miscompilation

["rguenth at tat dot physik dot uni-tuebingen dot de" <gcc-bugzilla@gcc.gnu.org>]

------- Additional Comments From rguenth at tat dot physik dot uni-tuebingen dot de  2004-06-02 11:54 -------
Subject: Re:  [3.4 regression] unit-at-a-time
 causes miscompilation

On Wed, 2 Jun 2004, belyshev at lubercy dot com wrote:

>
> ------- Additional Comments From belyshev at lubercy dot com  2004-06-02 11:15 -------
> >The minimized testcase is of course invalid.  I'm currently investigating
>
> I would not say that. Consider:

While I don't understand what you are hinting at here, looking at the
actual code doesn't reveal any obvious failure in handling invalid data,
still valgrind tells us:

==21613== Invalid read of size 4
==21613==    at 0x804E6EF: void Engine<1, int,
Dynamic>::performDestroy<IntervalIterator>(IntervalIterator const&,
IntervalIterator const&, BackFill const&, bool)
(/home/rguenth/src/pooma-bk/r2/src/Utilities/algorithms.h:233)
==21613==    by 0x804EE9E: void Engine<1, int,
Dynamic>::performDestroy<Interval<1> >(Interval<1> const&, BackFill
const&, bool)
(/home/rguenth/src/pooma-bk/r2/src/Engine/DynamicEngine.cpp:254)
==21613==    by 0x804EF44: void Engine<1, int,
Dynamic>::destroy<Interval<1> >(Interval<1> const&)
(/home/rguenth/src/pooma-bk/r2/src/Layout/DynamicEvents.h:127)
==21613==    by 0x804F11F:
Particles<MPDynamicUniform>::performDestroy(int, bool)
(/home/rguenth/src/pooma-bk/r2/src/Array/Array.h:2201)
==21613==    by 0x8049E1A: main (bctest3.cpp:143)
==21613==  Address 0xBFFFEF3C is just below %esp.  Possibly a bug in
GCC/G++
==21613==   v 2.96 or 3.0.X.  To suppress, use:
--workaround-gcc296-bugs=yes
==21613==
==21613== Invalid write of size 4
==21613==    at 0x804E77E: void Engine<1, int,
Dynamic>::performDestroy<IntervalIterator>(IntervalIterator const&,
IntervalIterator const&, BackFill const&, bool)
(/home/rguenth/src/pooma-bk/r2/src/Utilities/RefCountedBlockPtr.h:635)
==21613==    by 0x804EE9E: void Engine<1, int,
Dynamic>::performDestroy<Interval<1> >(Interval<1> const&, BackFill
const&, bool)
(/home/rguenth/src/pooma-bk/r2/src/Engine/DynamicEngine.cpp:254)
==21613==    by 0x804EF44: void Engine<1, int,
Dynamic>::destroy<Interval<1> >(Interval<1> const&)
(/home/rguenth/src/pooma-bk/r2/src/Layout/DynamicEvents.h:127)
==21613==    by 0x804F11F:
Particles<MPDynamicUniform>::performDestroy(int, bool)
(/home/rguenth/src/pooma-bk/r2/src/Array/Array.h:2201)
==21613==    by 0x8049E1A: main (bctest3.cpp:143)
==21613==  Address 0x60F177E4 is not stack'd, malloc'd or free'd
==21613==
==21613== Process terminating with default action of signal 11 (SIGSEGV):
dumping core

which is at src/Utilities/algorithms.h:233 - I remember looking at this
stuff and was not able to come to the conclusion that this part of
(arguably ugly) code is wrong.  Oh - of course POOMA_NONSTANDARD_ITERATOR
is zero.  The exact signature of the performDestroy is

void Engine<(int)1, int, Dynamic>::performDestroy<IntervalIterator>
  (IntervalIterator const&, IntervalIterator const&, BackFill const&, bool)

we have inlined the delete_backfill method here (probably the key of the
failure, marking the delete_backfill __attribute__((noinline)) "fixes"
the failure).

Maybe we are wrongly re-using some stack slots here?  Or is it possible
for (source) code to become invalid wrt references to temporaries after
inlining?

I'm confused.

Richard.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=14863