Re: proper lvalues and whatnot

[Ross A. Lippert]

"John W. Eaton" wrote:
> Hmm.  Because the result of an assingment is not an lvalue, even in C?
You know, I'm a bonehead sometimes.  I could have sworn C let
you do (x=y)=z in addition to x=(y=z).  You are right, thanks.
I should also take it back that this would be a desirable feature
in octave for what I'm doing.

>   octave:1> v = 1:10
>   v =
>      1   2   3   4   5   6   7   8   9  10
>   octave:2> v([3, 4])++
>   ans =
>     3  4
>   octave:3> v
>   v =
>      1   2   4   5   5   6   7   8   9  10
> That seems reasonable, doesn't it?
Yes, that seems totally reasonable.

> But your example was
>   octave:4> v = [1, 1]
>   v =
>     1  1
>   octave:5> v([1, 1, 2])++
>   ans =
>     1  1  1
>   octave:6> v
>   v =
>     2  2
> 
> which doesn't do exactly what you seem to expect.  Does the actual
> Octave behavior make sense though?  I think it does, if you think
> about it in the following way:
YEs the behavior does make sense.  I dipped into the ov.cc code
and found that an assign_op gets implemented by doing a bin_op 
followed by a simple_assign, and it is the simple_assign which does
the subscripting.

To me, and maybe not to anyone else, what seems mmore logical is
v(I) += x(I) should be implemented as
for i=1:length(I), v(I(i)) += x(I(i)); end

> | OK, so maybe I am a sick puppy.  But this is what I am really after.
> |
> | > [i,j,nz] = find(M);
> | > (y=zeros(max(i),1))(i) += nz.*x(j)
> | and have
> |  y = M*x
> 
> I guess I'm dense, but I don't understand what the purpose of this is,
> or how it is equivalent to y = M*x.
Actually, I don't need to assignment as lvalue thing bc I could just
do the following

octave> zeros(max(i),1)(i) += nz.*x(j)
ans = M*x

The purpose of this is that if M has only a few nonzero entries then
one could matrix multiply by 
tmp = zeros(max(i),1);
for p=1:length(i), tmp(i(p)) += nz(p)*x(j(p)); end
If lvalue subscripting had the do-loop interpretation that I'd like,
then the above loop could be replaced with that += expression.

I think that vectors and matrices exist as types in order to
cut down on the amount of explicit looping required to express common
things.  I think the interpretation I have for indexing an lvalue has
a do-loop interpretation which could be useful for doing common
operations.

> Are you using an actual sparse matrix data type or trying to emulate
> them using full storage (but trying to avoid operations on zeros)?  If
> the latter, wouldn't it make more sense to work on a sparse matrix
> data type instead?
That latter.  The reason I think a sparse matrix type would be a bad
idea
is
1) different sparse matrix formats are used by different people for
different reasons and no one would be happy.
2) constructs exist (almost) in octave which allow me to do what I need
to do without having to really bend over that far backwards.
3) I think that giving a specific behavior subscripted lvalues would
be useful not just to sparse matrix computations but for other purposes.
That just happens to be the place where I currently find myself in the
need.


> Seems to me that it is somewhat difficult to pin down exactly what
> these operations should do, and then arrange for everyone to agree.
I agree.  I'd like for someone to tell me if they have something
essential
that depends on a specified looping behavior for subscripted lvalues. 
Or
perhaps to point out just how this hasn't been well thought out.


I have noticed that the ov.cc file contains a readme for its assign_op
function suggesting that someone turn it into more than just a bin_op
simple_assign wrapper.  This indicates that there is some need for a
rewrite anyhow.  I'd volunteer to try if no one cries out that my
looping
idea would be a really bad one.



-r