Robert Bradshaw wrote:
> On Oct 11, 2008, at 12:24 AM, Greg Ewing wrote:
>
>> Robert Bradshaw wrote:
>>
>>> It's on the stack, just like literals.
>> What do you mean, "just like literals"? What literal
>> is ever allocated on the stack?
>
> Array literals are allocated on the stack.
>
>>> However, one can write
>>>
>>> def foo(x, y):
>>> cdef double* a
>>> if x:
>>> a = [1,2,3,x]
>>> else:
>>> a = [1,2,3,y]
>> Now I'm *really* confused. What the heck does that do?
>> Can you provide a C translation?
>
>
> Sure. I'm ignoring the implicit conversions that happen for clarity...
>
> double[4] tmp1;
> double[4] tmp2;
> if (x) {
> tmp1[0] = 1
> tmp1[1] = 2
> tmp1[2] = 3
> tmp1[3] = x
> a = tmp1
> }
> else {
> tmp2[0] = 1
> tmp2[1] = 2
> tmp2[2] = 3
> tmp2[3] = y
> a = tmp2
> }
>
I can see that this has many practical usecases etc., still I'm back to
worrying. If the following concern is valid then I'm -1 on this for the
coming release, it should at least be postponed.
The reason I worry is that one can write (rather contrived) code like this:
cdef int* a
cdef int* b
for i in range(2):
a = [1,1,1]
if i == 0:
b = a
b[0] = 2
print b[0] # will print 1...
...which will be rather non-obvious if you do not know *exactly* how the
allocation etc. works.
The reason this is so bad is because it seems like you are pointing "a"
to *something else* the second time around the loop, and that "b" should
continue pointing to the "old" array. (But there is only one array.
Which is especially counter-intuitive given how this code would work in
Python. Etc.)
To boil it down a bit, the problem is that
a = [1,1,1]
b = a
a = [1,1,1]
has entirely different semantics from
for i in range(2):
a = [1,1,1]
if i == 0: b = a
even though the first example looks exactly like just rolling out the
latter. At least it is very un-Pythonic.
(I did not actually try the code above, but if it doesn't work then I'd
counter that the mechanism is so magical that I cannot figure it out :-) )
What do I propose instead?
a) That the reference-counted heap-allocated array stuff in the wiki
is implemented first (with a list constructor). So the thing gets the
semantics of a reference counted array, constructed again each time it
is assigned (very similar to a Python list).
b) That simple cases (where the pointer is never reassigned or copied,
only written/read to/from by indexing) are then optimized to use stack
allocation instead.
--
Dag Sverre
_______________________________________________
Cython-dev mailing list
[email protected]
http://codespeak.net/mailman/listinfo/cython-dev
