Fix crash when calling buffer.request multiple times

[adler-j]

buffer.request now creates a temporary Py_buffer instead of a pointer.
This reduces the probability of misstakes, and stops a segfault when
buffer.request was called twice.

I think the new design has no major drawbacks over the previous, it is possibly faster since the buffer is now stack-allocated. Also we got rid of the destructor which should not be needed in this case.

Also add build folder to gitignore.

[wjakob]

Hi -- the issue with your change is that the buffer_info will return a pointer that may not be valid anymore after calling PyBuffer_Release(). This is the case when the requested buffer has a format that is different from the internal storage representation.

I think it will be better to use the previous approach and detect if view != nullptr, and not re-call PyObject_GetBuffer() in this case.

[adler-j]

What if the user calls with writable=false then writable=true?

How about letting the buffer_info have ownership of the Py_buffer instead the the buffer? That would clarify buffer ownership issue.

[wjakob]

That would be reasonable -- can you make it a private field in this case?

buffer_info needs a new constructor in this case which requests a buffer view and fills all the necessary fields. And the current constructor should just set the buffer view pointer to nullptr.

[adler-j]

I've been putting some thought into this and I wonder, where do we need buffer_info to have the current constructor? Wouldnt a single constructor that stores a Py_buffer by value be the simplest?

The only other place we currently create a buffer_info is inside vectorize_helper.run, but in this case, wouldnt it be more simple if array had a constructor that creates its own data (using PyArray_NewFromDescr) and then calling array.request(true) on it and then writing the values in a loop (as currently done using a std::vector). That would also remove the performance adverse copy that is currently done.

[wjakob]

Aren't you forgetting about def_buffer(), which can be used to provide a buffer view for custom types?

[adler-j]

You are indeed correct, will need to think more about this. The ownership issue really should be settled properly however.

[wjakob]

This is now fixed, a buffer can only be requested once (second time is a NOP).