CGAL users discussion list

[Andreas Fabri <>]

Hello,

Thank you for signalling this problem.
I confirm this behaviour for VC++2005/XP64/AMD Turion.
I noticed that you already signalled it to Microsoft
and they told you that they won't fix it.
http://connect.microsoft.com/VisualStudio/feedback/ViewFeedback.aspx?FeedbackID=257605



A construction as follows might fix it;

bool catched = false;
try {

}
 catch() {
    catched = true;
}
if(catched){
 reset the rounding
}

best regards,

andreas





 wrote:
> Hello,
>
> We found a very disturbing bug in Visual Studio 2005.  We have the Service 
> Pack 1 installed and are compiling in 64 bits (in 32 bits mode, everything is 
> fine).  We have CGAL version 3.2.  We run on Windows XP(SP2) on dual Xeon 
> workstations.
>
> We found the bug in our 64 bits regression tests when some tests would pass 
> if run alone but would fail if some other tests (using CGAL) were run before.
>
> The problem is that the rounding mode that is changed by interval_nt is not 
> reset to its previous state (in a catch) even though good exceptions and RAII 
> technics are used.  Here is a small program that reproduces the bug mechanism 
> without using interval_nt.
>
> #include <float.h>
> #include <stdio.h>
> void main()
> {
>     try
>     {
>         _controlfp( _RC_UP, _MCW_RC );
>         throw 1;
>     }
>     catch ( ... )
>     {
>         // Sets the rounding mode to 0 and show that it's realy the case.
>         _controlfp( 0, _MCW_RC );
>         unsigned int cw = _controlfp( 0, 0 ) & _MCW_RC;
>         printf( "this 1st value should be 0: %d\n", cw );
>     }
>
>     // After the catch, this value will not be 0 in x64
>     unsigned int cw = _controlfp( 0, 0 ) & _MCW_RC;
>     printf( "this 2nd value should be 0: %d\n", cw );
> }
>
> Are you aware of this strange behavior?  Can you verify if it's happening on 
> other compilers?  We are worried that it could invalidate the interval 
> arithmetic filter strategy used in CGAL in 64 bits...
>
> Thank you,
>
> Bruno Martin