Re: fix null assignment: the (hopefully) final answer

[Jaroslav Hajek]

On Fri, Sep 26, 2008 at 5:55 PM, John W. Eaton <address@hidden> wrote:
> On 26-Sep-2008, Jaroslav Hajek wrote:
>
> | On Thu, Sep 25, 2008 at 4:16 PM, Jaroslav Hajek <address@hidden> wrote:
> | > hello,
> | >
> | > please consider the changeset at
> | > http://artax.karlin.mff.cuni.cz/~hajej2am/ulozna/null-assign2.diff
> | > (not attached due to size).
> | >
> | > This seems a much better attempt at making the null assignments
> | > smarter and Matlab compatible. The main advantage over the previous
> | > attempts is that null assignments are completely resolved within
> | > liboctinterp.
> | > Three special octave_value classes are introduced: null matrix, null
> | > string and null single quoted string (initially, I tried to use just
> | > one but it seemed clumsy). These are normally registered via octave's
> | > type info system and appropriate conversion and assignment handlers
> | > are installed as necessary. If these special values are stored to
> | > variable or cell or struct element, they lose their speciality. This
> | > is to prevent things like
> | >
> | > b = []; a(1:2) = b
> | >
> | > from deleting elements.
> | >
> | > The special values *are* allowed to be passed as function arguments,
> | > contrary to what Matlab allows. This is a serious bug in Matlab that
> | > prevents elements of user-defined array classes from being deleted
> | > using the `= []' syntax in the same way as for built-in arrays (i.e.
> | > you either check for isempty and delete always, even for a computed
> | > empty matrix, or never).  A new built-in function `isnull' is provided
> | > to check whether an expression is a null value, to be used in more
> | > complicated cases.
> | >
> | > The null values are not, however, allowed to be returned from functions.
> | >
> | > cheers
> | >
> | > --
> | > RNDr. Jaroslav Hajek
> | > computing expert
> | > Aeronautical Research and Test Institute (VZLU)
> | > Prague, Czech Republic
> | > url: www.highegg.matfyz.cz
> | >
> |
> | The Changeset is updated for better Matlab compatibility. The Matlab
> | rules seem rather ad hoc, but it appears that the present state gets
> | quite close apart from the strange corner cases.
> | For instance, a(3:2,3:2) = [] work in Matlab, while a(3:2,3:2,3:2) doesn't, 
> but
> | a(3:2,[],3:2) = [] does. I think that simply ignoring all null
> | assignments with zero-length index vectors is a reasonable
> | approximation.
>
> That seems reasonable.
>
> Thanks for doing this.  I think this is a much better solution that I
> was working on.
>
> I've checked in the changes in my archive.
>

I'll transplant this into 3.0.3 rc2, as it probably fixes the problem
reported by Michael:
http://www.nabble.com/forum/Permalink.jtp?root=19645333&post=19675573&page=y

Meanwhile, can you please apply the attached additional checks for
null assignment behaviour?

-- 
RNDr. Jaroslav Hajek
computing expert
Aeronautical Research and Test Institute (VZLU)
Prague, Czech Republic
url: www.highegg.matfyz.cz

null_assign_tests.diff
Description: Text Data