[Bug c++/51885] New: g++ compiler options -O2 and -O3 modifies program results

[Bug c++/51885] New: g++ compiler options -O2 and -O3 modifies program results

[laurentlouis.maurin at wanadoo dot fr]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51885

             Bug #: 51885
           Summary: g++ compiler options -O2 and -O3 modifies program
                    results
    Classification: Unclassified
           Product: gcc
           Version: 4.6.2
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c++
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: laurentlouis.maurin@wanadoo.fr


Created attachment 26356
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=26356
example program with comments

platform : Fedora 16 x86_64
compiler : gcc-c++-4.6.2-1.fc16.x86_64

see attaches source file ; it is an example of the quicksort algorithm written
with pointers and templates

one template is specialised and uses a reinterpret_cast (it is part of a bigger
file)

it sorts this array :  {12, -1, 2}

a) first test (compilation without any optimisation)

compilation : g++ main.cpp && ./a.out

result OK : {-1, 2, 12}

b) second test (-O2 optimisation)

compilation : g++ -O2 main.cpp && ./a.out

wrong result : {2, 12, -1}

3) third test (-O3 optimisation)

compilation : g++ -O3 main.cpp && ./a.out

wrong result : {2, 12, -1}



Strange behaviour : if i include between line #65 and line #67 (see attached
file)

std::cout<<*ptr_start<<std::endl;

then, the results are good in any case


i believe this is a bug ... the optimisation option should not change the
program result

best regards

Laurent

[Bug c++/51885] g++ compiler options -O2 and -O3 modifies program results

[paolo.carlini at oracle dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51885

--- Comment #1 from Paolo Carlini <paolo.carlini at oracle dot com> 2012-01-17 18:38:35 UTC ---
What happens if you add to the command line -fno-strict-aliasing? If things
'improve' it means that with your reinterpret_cast & co you have aliasing
violations, thus undefined behavior, thus, literally, anything can happen, and
this is not a bug (and we have hundreds of closed duplicate PRs in Bugzilla)

[Bug c++/51885] g++ compiler options -O2 and -O3 modifies program results

[redi at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51885

Jonathan Wakely <redi at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |INVALID

--- Comment #2 from Jonathan Wakely <redi at gcc dot gnu.org> 2012-01-17 18:44:32 UTC ---
The reinterpret_cast definitely results in undefined behaviour, because it
accesses the float objects through a different type, violating [basic.lval]/10

And indeed, -fno-strict-aliasing makes the problem go away.

Someone didn't read http://gcc.gnu.org/bugs/ carefully.

[Bug c++/51885] g++ compiler options -O2 and -O3 modifies program results

[manu at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51885

Manuel López-Ibáñez <manu at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |manu at gcc dot gnu.org

--- Comment #3 from Manuel López-Ibáñez <manu at gcc dot gnu.org> 2012-01-18 10:02:04 UTC ---
(In reply to comment #2)
> The reinterpret_cast definitely results in undefined behaviour, because it
> accesses the float objects through a different type, violating [basic.lval]/10
> 
> And indeed, -fno-strict-aliasing makes the problem go away.

It definitely does not make the problem go away, it merely hides it. The
undefined behavior does not disappear just by using command-line options.

On the other hand, is it possible to cast a float pointer to a uint32 pointer
and use it for anything that does not invoke undefined behaviour? 
GCC gives a warning with -Wstrict-aliasing=1 but no with other levels.

[Bug c++/51885] g++ compiler options -O2 and -O3 modifies program results

[laurentlouis.maurin at wanadoo dot fr]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51885

--- Comment #4 from Laurent MAURIN <laurentlouis.maurin at wanadoo dot fr> 2012-01-20 19:18:13 UTC ---
i am quite sure there is something strange : reinterpret_cast is used to cast a
float pointer to a uint32_t pointer

these two types have the same size (4 bytes)

therefore, the swap function (following the reinterpret_cast) should work
properly ...

maybe the -O2 and -O3 compiler optimisation flags have an influence on the
binary code generated for the swap function in this example

[Bug c++/51885] g++ compiler options -O2 and -O3 modifies program results

[redi at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51885

--- Comment #5 from Jonathan Wakely <redi at gcc dot gnu.org> 2012-01-20 19:31:48 UTC ---
(In reply to comment #4)
> i am quite sure there is something strange : reinterpret_cast is used to cast a
> float pointer to a uint32_t pointer
> 
> these two types have the same size (4 bytes)
> 
> therefore, the swap function (following the reinterpret_cast) should work
> properly ...

No, the C++ standard says your code has undefined behaviour.

Please read the section "Casting does not work as expected when optimization is
turned on" at http://gcc.gnu.org/bugs/#nonbugs_c and the article it links to,
http://mail-index.netbsd.org/tech-kern/2003/08/11/0001.html