How limited is ffi.addressof()?

[AimForNaN]

I have the following cdef:

typedef struct AVFormatContext AVFormatContext;

extern AVFormatContext* avformat_alloc_context(void);

I obtain an instance of AVFormatContext in python and check its type:

format_context = lib.avformat_alloc_context();
print(ffi.typeof(format_context));
# <ctype 'struct AVFormatContext *'>

When I try to get its address, the script exits with an error:

ffi.addressof(format_context);
# TypeError: expected a cdata struct/union/array object

Not sure what I'm doing wrong.

[arigo]

The type of format_context is 'struct AVFormatContext *', that is, it is already a pointer to the struct. Are you trying to get the address of that pointer, i.e. a pointer-to-pointer?

ffi.addressof() only works on things that have known C-level addresses. The format_context Python variable is a Python object that internally points to that struct, but you can't take its own address because that pointer itself is not C-level data. Think about Python objects as living in a separate world, somewhere else than all the C data that cffi lets you manipulate. You can't get the address of a Python object without using some hacks that are specific to an implementation, say CPython, and these hacks won't work somewhere else, say PyPy.

If you do want to get a pointer-to-pointer, you need to store the pointer you got into a C structure or array first, i.e. copy the pointer from the Python world to the C world. For example:

p = lib.avformat_alloc_context()                 # pointer to struct.  The pointer itself lives in the Python world
p_array = ffi.new("struct AVFormatContext*[1]")  # make an array of one pointer
p_array[0] = p                                   # store p inside that array
p_to_p = ffi.addressof(p_array, 0)               # now we can take "&p_array[0]"

[AimForNaN]

That got the code to run without generating an error.