[Cython] [cython-users] Re: How to find out where an AttributeError is ignored

[Robert Bradshaw]

On Fri, Jan 27, 2012 at 1:01 PM, Stefan Behnel <stefan_ml at behnel.de> wrote:
> Dag Sverre Seljebotn, 27.01.2012 21:03:
>> On 01/27/2012 05:58 PM, Stefan Behnel wrote:
>>> mark florisson, 27.01.2012 17:30:
>>>> On 27 January 2012 16:22, mark florisson<markflorisson88 at gmail.com>  wrote:
>>>>> On 27 January 2012 15:47, Simon King<simon.king at uni-jena.de>  wrote:
>>>>>> Hi all,
>>>>>>
>>>>>> I am still *very* frustrated about the fact that Cython does not tell
>>>>>> where the error occurs. Since about one week, I am adding lots and
>>>>>> lots of lines into Sage that write a log into some file, so that I get
>>>>>> at least some idea where the error occurs. But still: Even these
>>>>>> extensive logs do not provide a hint on what exactly is happening.
>>>>>>
>>>>>> How can I patch Cython such that some more information on the location
>>>>>> of the error is printed? I unpacked Sage's Cython spkg, and did "grep -
>>>>>> R ignored .", but the code lines containing the word "ignored" did not
>>>>>> seem to be the lines that are responsible for printing the warning
>>>>>> message
>>>>>>    Exception AttributeError: 'PolynomialRing_field_with_category'
>>>>>> object has no attribute '_modulus' in  ignored
>>>>>>
>>>>>> Can you point me to the file in Sage's Cython spkg which is
>>>>>> responsible for printing the warning?
>>>>>>
>>>>>> Best regards,
>>>>>> Simon
>>>>>
>>>>> These messages are written by PyErr_WriteUnraisable, which is a
>>>>> CPython C API function that writes unraisable exceptions. There are
>>>>> typically two reasons for unraisable exceptions:
>>>>>
>>>>>     1) as Robert mentioned, a function that does not allow propagation
>>>>> of exceptions, e.g.
>>>>>
>>>>>         cdef int func():
>>>>>             raise Exception
>>>>>
>>>>>         Here there is no way to propagate the raised exception, so
>>>>> instead one should write something like
>>>>>
>>>>>             cdef int func() except -1: ...
>>>>>
>>>>>         Alternatively one may use 'except *' in case there is no error
>>>>> indicator and Cython should always check, or "except ? -1" which means
>>>>> "-1 may or may not indicate an error".
>>>>>
>>>>>     2) in deallocators or finalizers (e.g. __dealloc__ or __del__)
>>>>>
>>>>> For functions the right thing is to add an except clause, for
>>>>> finalizers and destructors one could use the traceback module, e.g.
>>>>>
>>>>>     try:
>>>>>         ...
>>>>>     except:
>>>>>         traceback.print_exc()
>>>>>
>>>>> If this all still doesn't help, try setting a (deferred) breakpoint on
>>>>> __Pyx_WriteUnraisable or PyErr_WriteUnraisable.
>>>>
>>>> Actually, I don't see why the default is to write unraisable
>>>> exceptions. Instead Cython could detect that exceptions may propagate
>>>> and have callers do the check (i.e. make it implicitly "except *").
>>
>> As for speed, there's optimizations on this, e.g., "except? 32434623" if
>> the return type is int, "except? 0xfffff..." if the return type is a pointer.
>>
>> And for floating point, we could make our own NaN -- that's obscure enough
>> that it could probably be made "except cython.cython_exception_nan" by
>> default, not "except? cython.cython_exception_nan".
>
> The problem with that is that we can't be sure that Cython will be the only
> caller. So exceptions may still not propagate in cases, and users will have
> to know about these "obscure" values and that they must deal with them
> manually then.
>
> You could add that we'd just have to disable this when user code takes a
> pointer from a function, but then, how many rules are there that users will
> have to learn and remember after such a change? And what's that for a
> language that changes the calling semantics of a function because way down
> in the code someone happens to take a pointer to it?
>
>
>>>> Was this not implemented because Cython only knows whether functions
>>>> may propagate exceptions at code generation time by looking at the
>>>> presence of an error label?
>>>> Maybe it could keep code insertion points around for every call to
>>>> such a potential function and if the function uses the error label
>>>> have the caller perform the check? Although I do forsee problems for
>>>> external such functions... maybe Cython could have it's own
>>>> threadstate regardless of the GIL which would indicate whether an
>>>> error has occurred? e.g. CyErr_Occurred()?
>>>
>>> Yep, those are the kind of reasons why writing unraisable exceptions is the
>>> default.
>>
>> Still,
>
> I wasn't really advocating this behaviour, just indicating that it's hard
> to do "better", because this "better" isn't all that clear. It's also not
> "better" for all code, which means that we get from one trade-off to
> another, while breaking existing code at the same time. Not exactly
> paradise on either side of the tunnel.

I still feel like we're stuck in the wrong default. I'd rather require
more work to interact with C libraries than require more work to
convert innocent-looking Python to Cython.

> One example that keeps popping up in my mind is callback functions that
> cannot propagate errors, at least not the CPython way. I have a couple of
> those in lxml, even some returning void. So I wrap their code in a bare
> try-except and when an exception strikes, I set a C error flag to tell the
> C library that something went wrong and return normally. No Python code
> outside of the try block. But Cython still generates code for unraisable
> errors. Why? Because the internal code that handles the bare except clause
> may fail and raise an exception. How about that?
>
>
>> the need to explicitly declare "except *" keeps coming up again and
>> again, and is really a blemish on the usability of Cython. When teaching
>> people Cython, then it's really irritating to have to follow "all you need
>> to do is add some 'cdef' and some types" with "and then you need to
>> remember to say "except *", or you're in deep trouble". Cython sort of
>> looks very elegant until that point...
>
> I know what this feels like. The problem is that these things *are* complex.

Yes. We've been wrestling with this issue almost since Cython's inception...

I like Mark's two-function idea, with the caveat that f(bad_argument)
now behaves quite differently than (&f)[0](bad_argument) for even more
obscure reasons. But it may be the way to go.

The other option is to embed the error behavior into the signature and
require casts to explicitly go from one to the other. This would
probably require a notation for never raising an exception (e.g.
"except -"). Cdef public or api functions could require an except
declaration (positive or negative), ordinary cdef functions would be
"except *" by default, and cdef extern functions would be "except -"
by default.

Ideally, the default would not just be "except *" but "except
cython.error_value?" and a case could be made for acquiring the GIL to
check in the exceptional case that error_value is returned, or the
information could be passed by checking a bit on some thread-local
Cython global (not sure what the performance impact would be here).

A warning whenever WriteUnraisable is used could be handy too, but how
to handle Stefan's example where the bare except clause could raise an
exception?

>> Long-term we should change CPython to make sure that PyErr_Occurred doesn't
>> need the GIL :-) (there's really no reason it should need to go beyond
>> checking a thread-local variable).
>
> I always wondered about that, too. Still, "long-term" here basically means
> "when all current CPython versions that work like this are out of use",
> because we cannot base language semantics on specific runtime CPython versions.

+1

- Robert