[Cython] 'local variable referenced before assignment' warnings

[Robert Bradshaw]

On Sat, Sep 24, 2011 at 1:08 AM, Vitja Makarov <vitja.makarov at gmail.com> wrote:
> 2011/9/23 mark florisson <markflorisson88 at gmail.com>:
>> On 23 September 2011 06:02, Robert Bradshaw
>> <robertwb at math.washington.edu> wrote:
>>> On Thu, Sep 22, 2011 at 9:36 PM, Vitja Makarov <vitja.makarov at gmail.com> wrote:
>>>> 2011/9/23 Robert Bradshaw <robertwb at math.washington.edu>:
>>>>> On Thu, Sep 22, 2011 at 5:08 AM, Vitja Makarov <vitja.makarov at gmail.com> wrote:
>>>>>> 2011/9/22 Robert Bradshaw <robertwb at math.washington.edu>:
>>>>>>> On Wed, Sep 21, 2011 at 1:20 AM,  <robert.rex at exasol.com> wrote:
>>>>>>>> Hi Robert,
>>>>>>>>
>>>>>>>> Robert Bradshaw wrote:
>>>>>>>>>
>>>>>>>>> The warning seem correct to me, or is foo actually modifying the
>>>>>>>>> values of cya and cyb? (I suppose this could be possible by taking the
>>>>>>>>> address of the "passed by reference" arguments, but would seem like a
>>>>>>>>> *very* poor API design.)
>>>>>>>>
>>>>>>>> thanks for your fast answer! What made me came up with this issue is
>>>>>>>> that GCC/C++ will not complain in similar situations in C++ code (and
>>>>>>>> assumes foo() to initialize/change the value of cya/cyb). Thus, there
>>>>>>>> is an (arguable) difference in C++ compiler/Cython warnings.
>>>>>>>
>>>>>>> Yes, in C++ the called function can modify values passed by reference,
>>>>>>> so we need to add this exception to Cython.
>>>>>>>
>>>>>>
>>>>>> That could be hard because we can not always resolve function to check
>>>>>> its args at create-control-flow stage.
>>>>>
>>>>> Doesn't this always happen after declaration analysis?
>>>>>
>>>>
>>>> Sure. But before analyse expressions, so you have to manually lookup
>>>> entries for names and attributes.
>>>
>>> Ah, yes.
>>>
>>>> I see two cases here: global function call and method call. Both are
>>>> SimpleCallNodes.
>>>
>>> Yep, and the function could be an arbitrary expression... Ugly.
>>
>> As a simple hack, couldn't you in case of a SimpleCallNode where the
>> function is a NameNode check if the arguments are passed by reference,
>> and if so not issue the warning. If the callnode.function is an
>> arbitrary expression (like casting a pointer to such a function), then
>> you just issue the warning. Then you will cover probably at least 95%
>> of the cases.
>>
>
> Yeah, most popular cases would be function call and method call. We
> can handle both of them but I'm still not sure that this is a good
> idea.

The NameNode is probably worth handling, anything more and it's hard
to know where to stop.

> Maybe it's better to disable this warning by default.

Note it's always valid in C.

- Robert