view mercurial/cext/util.h @ 33807:b70029f355a3

tests: verify that peer instances only expose interface members

Our abstract interfaces are more useful if we guarantee that
implementations conform to certain rules. Namely, we want to ensure
that objects implementing interfaces don't expose new public
attributes that aren't part of the interface. That way, as long as
consumers don't access "internal" attributes (those beginning with
"_") then (in theory) objects implementing interfaces can be swapped
out and everything will "just work."

We add a test that enforces our "no public attributes not part
of the abstract interface" rule.

We /could/ implement "interface compliance detection" at run-time.
However, that is littered with problems.

The obvious solutions are custom __new__ and __init__ methods.
These rely on derived types actually calling the parent's
implementation, which is no sure bet. Furthermore, __new__ and
__init__ will likely be called before instance-specific attributes
are assigned. In other words, they won't detect public attributes
set on self.__dict__. This means public attribute detection won't
be robust.

We could work around lack of robust self.__dict__ public attribute
detection by having our interfaces implement a custom __getattribute__,
__getattr__, and/or __setattr__. However, this incurs an undesirable
run-time penalty. And, subclasses could override our custom
method, bypassing the check.

The most robust solution is a non-runtime test. So that's what this
commit implements. We have a generic function for validating that an
object only has public attributes defined by abstract classes. Then,
we instantiate some peers and verify a newly constructed object
plays by the rules.

Differential Revision: https://phab.mercurial-scm.org/D339

/*
 util.h - utility functions for interfacing with the various python APIs.

 This software may be used and distributed according to the terms of
 the GNU General Public License, incorporated herein by reference.
*/

#ifndef _HG_UTIL_H_
#define _HG_UTIL_H_

#include "compat.h"

#if PY_MAJOR_VERSION >= 3
#define IS_PY3K
#endif

typedef struct {
	PyObject_HEAD
	char state;
	int mode;
	int size;
	int mtime;
} dirstateTupleObject;

extern PyTypeObject dirstateTupleType;
#define dirstate_tuple_check(op) (Py_TYPE(op) == &dirstateTupleType)

#define MIN(a, b) (((a)<(b))?(a):(b))
/* VC9 doesn't include bool and lacks stdbool.h based on my searching */
#if defined(_MSC_VER) || __STDC_VERSION__ < 199901L
#define true 1
#define false 0
typedef unsigned char bool;
#else
#include <stdbool.h>
#endif

static inline PyObject *_dict_new_presized(Py_ssize_t expected_size)
{
	/* _PyDict_NewPresized expects a minused parameter, but it actually
	   creates a dictionary that's the nearest power of two bigger than the
	   parameter. For example, with the initial minused = 1000, the
	   dictionary created has size 1024. Of course in a lot of cases that
	   can be greater than the maximum load factor Python's dict object
	   expects (= 2/3), so as soon as we cross the threshold we'll resize
	   anyway. So create a dictionary that's at least 3/2 the size. */
	return _PyDict_NewPresized(((1 + expected_size) / 2) * 3);
}

#endif /* _HG_UTIL_H_ */