[Insight-developers] MedialNodeCorrespondencesTest, Intel C++ fixed

[Miller, James V (Research)]

This was on the Intel C++ compiler on Linux, version 7.


> -----Original Message-----
> From: Bill Hoffman [mailto:bill . hoffman at kitware . com]
> Sent: Thursday, August 14, 2003 1:59 PM
> To: Miller, James V (Research); Insight-developers (E-mail)
> Subject: Re: [Insight-developers] MedialNodeCorrespondencesTest, Intel
> C++ fixed
> 
> 
> This change was made so that the fixed versions did not have 
> the overhead
> of having the extra data associated with the non-fixed 
> versions.  (8 bytes)
> 
> 
>  From what I can tell, you are saying that the following crashes:
> 
> 
> #include <vnl/vnl_matrix.h>
> #include <vnl/vnl_matrix_fixed.h>
> void foo(vnl_matrix<double> m)
> {
>   vnl_matrix<double> mm = m;
> }
> 
> main()
> {
>   vnl_matrix_fixed<double,3,3> m;
>   foo(m);
>   return 0;
> }
> 
> Which intel compiler is this?  Windows or Linux?
> 
> It is using this operator:
> 
>   //: Cheap conversion to vnl_matrix_ref
>   // Sometimes, such as with templated functions, the compiler cannot
>   // use this user-defined conversion. For those cases, use the
>   // explicit as_ref() method instead.
>   operator const vnl_matrix_ref<T>() const { return 
> vnl_matrix_ref<T>( num_rows, num_cols, 
> const_cast<T*>(data_block()) ); }
> 
> Sounds like a bug in the compiler.   This is going to be hard 
> to track down in all the
> code, and should work.  
> 
> If you let me know which intel compiler it is crashing on, I 
> can take a look.
> 
> -Bill
> 
> 
> >It looks like the vnl_matrix_fixed and vnl_matrix classes 
> are no longer related via inheritence.  Instead, there are 
> some converter routines ( mat_fixed.as_ref() returns a vnl_matrix).  
> > 
> >The Intel compiler builds were crashing on the deallocation 
> of a vnl_matrix.  Bill changed the code in 
> Ellipsoid*SpatialFunction to take a reference to a 
> vnl_matrix_fixed instead of taking a vnl_matrix by value.  
> This clears up the problem on the Intel compiler.  I am 
> guessing the other systems are relying on some other type 
> conversion operators that may be flawed on the Intel compiler.
> > 
> >Two things to look out for:
> > 
> >1. vnl_matrix and vnl_matrix_fixed mismatches.
> >2. function that take matrix types by value as opposed to by 
> reference. 
> > 
> 
> 
>