dolfin team mailing list archive

[Anders Logg <logg@xxxxxxxxx>]

On Thu, Jun 19, 2008 at 11:41:52PM +0200, Dag Lindbo wrote:
> Hello again!
> 
> Here comes another suggestion about the DOLFIN public interface:
> 
> Constructing time dependent functions by subclassing from Function and
> TimeDependent is very very nice. The way u.sync(t) works by storing a
> pointer to t is also very clever. However, I think it is too clever --
> most would expect pass-by-value semantics when passing a scalar to a
> function. A very simple change could make it a lot clearer what's going
> on: Pass a pointer to t instead of a constant reference. This forces the
> user to take the address of the scalar, which really shows what's going
> on. E.g.
> 
> u.sync(&t)
> 
> Today I misunderstood the reference nature of sync(t), and wrote code like
> this:
> 
> while(t < T){
>   "compute timestep"
>   t += dt
>   sync_bc(t)
> }
> 
> void sync_bc(real t){
>   for "all Dirichlet BCs"
>      bc.sync(t)  // Pointer to stack memory gets assigned!!!!
> }
> 
> :-( Of course, no one would be foolish enough to pass a pointer to stack
> memory knowingly in such a short lived context.
> 
> /Dag

TimeDependent does not seem to be used anywhere in the code so now
it's just a fancy way of storing the current time. Objects that want
to store the current time could equally well just add a member

  real t;

I suggest we remove TimeDependent for now and add it back later when
we find any use for it.

ok?

-- 
Anders

Attachment: signature.asc
Description: Digital signature