- Object Spaces
- Object Space Interface
- The Standard Object Space
- The Flow Object Space
- Object Space proxies
The object space creates all objects and knows how to perform operations on the objects. You may think of an object space as being a library offering a fixed API, a set of operations, with implementations that correspond to the known semantics of Python objects. An example of an operation is add: add’s implementations are, for example, responsible for performing numeric addition when add works on numbers, concatenation when add works on built-in sequences.
All object-space operations take and return application-level objects. There are only a few, very simple, object-space operations which allow the bytecode interpreter to gain some knowledge about the value of an application-level object. The most important one is is_true(), which returns a boolean interpreter-level value. This is necessary to implement, for example, if-statements (or rather, to be pedantic, to implement the conditional-branching bytecodes into which if-statements get compiled).
We have many working object spaces which can be plugged into the bytecode interpreter:
- The Standard Object Space is a complete implementation of the various built-in types and objects of Python. The Standard Object Space, together with the bytecode interpreter, is the foundation of our Python implementation. Internally, it is a set of interpreter-level classes implementing the various application-level objects – integers, strings, lists, types, etc. To draw a comparison with CPython, the Standard Object Space provides the equivalent of the C structures PyIntObject, PyListObject, etc.
- various Object Space proxies wrap another object space (e.g. the standard one) and adds new capabilities, like lazily computed objects (computed only when an operation is performed on them), security-checking objects, distributed objects living on several machines, etc.
- the Flow Object Space transforms a Python program into a flow-graph representation, by recording all operations that the bytecode interpreter would like to perform when it is shown the given Python program. This technique is explained in another document.
The present document gives a description of the above object spaces. The sources of PyPy contain the various object spaces in the directory pypy/objspace/.
This is the public API that all Object Spaces implement.
These functions both take and return “wrapped” objects.
The following functions implement operations with a straightforward semantic - they directly correspond to language-level constructs:
id, type, issubtype, iter, next, repr, str, len, hash,
getattr, setattr, delattr, getitem, setitem, delitem,
pos, neg, abs, invert, add, sub, mul, truediv, floordiv, div, mod, divmod, pow, lshift, rshift, and_, or_, xor,
nonzero, hex, oct, int, float, long, ord,
lt, le, eq, ne, gt, ge, cmp, coerce, contains,
inplace_add, inplace_sub, inplace_mul, inplace_truediv, inplace_floordiv, inplace_div, inplace_mod, inplace_pow, inplace_lshift, inplace_rshift, inplace_and, inplace_or, inplace_xor,
get, set, delete, userdel
- call(w_callable, w_args, w_kwds):
- Call a function with the given args and keywords.
- Implements the index lookup (new in CPython 2.5) on ‘w_obj’. Will return a wrapped integer or long, or raise a TypeError if the object doesn’t have an __index__ special method.
- is_(w_x, w_y):
- Implements ‘w_x is w_y’. (Returns a wrapped result too!)
- isinstance(w_obj, w_type):
- Implements ‘issubtype(type(w_obj), w_type)’. (Returns a wrapped result too!)
- exception_match(w_exc_type, w_check_class):
- Checks if the given exception type matches ‘w_check_class’. Used in matching the actual exception raised with the list of those to catch in an except clause. (Returns a wrapped result too!)
The following functions are part of the object space interface but would not be strictly necessary because they can be expressed using several other object space methods. However, they are used so often that it seemed worthwhile to introduce them as shortcuts.
- eq_w(w_obj1, w_obj2):
- Returns true when w_obj1 and w_obj2 are equal. Shortcut for space.is_true(space.eq(w_obj1, w_obj2))
- is_w(w_obj1, w_obj2):
- Shortcut for space.is_true(space.is_(w_obj1, w_obj2))
- Shortcut for space.int_w(space.hash(w_obj))
- Shortcut for space.int_w(space.len(w_obj))
- Shortcut for space.newbool(not space.is_true(w_obj))
- finditem(w_obj, w_key):
- Equivalent to getitem(w_obj, w_key) but returns an interp-level None instead of raising a KeyError if the key is not found.
- call_function(w_callable, *args_w, **kw_w):
- Convenience function that collects the arguments in a wrapped tuple and dict and invokes ‘space.call(w_callable, ...)’.
- call_method(w_object, 'method', ...):
- uses space.getattr() to get the method object, and then space.call_function() to invoke it.
- unpackiterable(w_iterable, expected_length=-1):
- this helper iterates w_x (using space.iter() and space.next()) and collects the resulting wrapped objects in a list. If expected_length is given and the length does not match, an exception is raised. Of course, in cases where iterating directly is better than collecting the elements in a list first, you should use space.iter() and space.next() directly.
- unpacktuple(w_tuple, expected_length=None):
- Same as unpackiterable(), but only for tuples.
- implements the built-in callable(). Returns a wrapped True or False.
- Returns a wrapped object that is a reference to the interpreter-level object x. This can be used either on simple immutable objects (integers, strings...) to create a new wrapped object, or on instances of W_Root to obtain an application-level-visible reference to them. For example, most classes of the bytecode interpreter subclass W_Root and can be directly exposed to app-level in this way - functions, frames, code objects, etc.
- Creates a wrapped bool object from an interpreter level object.
- newtuple([w_x, w_y, w_z, ...]):
- Makes a new wrapped tuple out of an interpreter level list of wrapped objects.
- Takes an interpreter level list of wrapped objects.
- Returns a new empty dictionary.
- newslice(w_start, w_end, w_step):
- Makes a new slice object.
- Creates a string from a list of wrapped integers. Note that this is not a very useful method; usually you can just say space.wrap(“mystring”).
- Creates a unicode string from a list of integers.
- Return the Interpreter Level equivalent of w_x. DO NOT USE! Only for testing. Use the functions described below instead.
- Return a interpreter level bool (True or False) that gives the truth value of the wrapped object w_x.
- If w_x is an application-level integer or long which can be converted without overflow to an integer, return an interpreter-level integer. Otherwise raise TypeError or OverflowError.
- If w_x is an application-level integer or long, return an interpreter-level rbigint. Otherwise raise TypeError.
- If w_x is an application-level string, return an interpreter-level string. Otherwise raise TypeError.
- If w_x is an application-level float, integer or long, return interpreter-level float. Otherwise raise TypeError or OverflowError in case of very large longs.
- getindex_w(w_obj, w_exception=None):
- Call index(w_obj). If the resulting integer or long object can be converted to an interpreter-level int, return that. If not, return a clamped result if w_exception is None, otherwise raise that exception on application-level. (If w_obj can’t be converted to an index, index() will raise an application-level TypeError.)
- interp_w(RequiredClass, w_x, can_be_None=False):
- If w_x is a wrapped instance of the given bytecode interpreter class, unwrap it and return it. If can_be_None is True, a wrapped None is also accepted and returns an interp-level None. Otherwise, raises an OperationError encapsulating a TypeError with a nice error message.
- If w_x is a wrapped instance of an bytecode interpreter class – for example Function, Frame, Cell, etc. – return it unwrapped. Otherwise return None.
space.builtin: The Module containing the builtins
space.sys: The ‘sys’ Module
space.w_None: The ObjSpace’s None
space.w_True: The ObjSpace’s True
space.w_False: The ObjSpace’s False
space.w_Ellipsis: The ObjSpace’s Ellipsis
space.w_NotImplemented: The ObjSpace’s NotImplemented
space.w_int, w_float, w_long, w_tuple, w_str, w_unicode, w_type, w_instance, w_slice: Python’s most common type objects
space.w_XxxError`` for each exception class XxxError (e.g. space.w_KeyError, space.w_IndexError, etc.).
List of tuples (method name, symbol, number of arguments, list of special names) for the regular part of the interface. (Tuples are interpreter level.)
List of names of built-in modules.
List of names of the constants that the object space should define
List of names of exception classes.
List of names of methods that have an irregular API (take and/or return non-wrapped objects).
The Standard Object Space (pypy/objspace/std/) is the direct equivalent of CPython’s object library (the “Objects/” subdirectory in the distribution). It is an implementation of the common Python types in a lower-level language.
The Standard Object Space defines an abstract parent class, W_Object, and a bunch of subclasses like W_IntObject, W_ListObject, and so on. A wrapped object (a “black box” for the bytecode interpreter main loop) is thus an instance of one of these classes. When the main loop invokes an operation, say the addition, between two wrapped objects w1 and w2, the Standard Object Space does some internal dispatching (similar to “Object/abstract.c” in CPython) and invokes a method of the proper W_XyzObject class that can do the operation. The operation itself is done with the primitives allowed by RPython. The result is constructed as a wrapped object again. For example, compare the following implementation of integer addition with the function “int_add()” in “Object/intobject.c”:
def add__Int_Int(space, w_int1, w_int2): x = w_int1.intval y = w_int2.intval try: z = ovfcheck(x + y) except OverflowError: raise FailedToImplementArgs(space.w_OverflowError, space.wrap("integer addition")) return W_IntObject(space, z)
Why such a burden just for integer objects? Why did we have to wrap them into W_IntObject instances? For them it seems it would have been sufficient just to use plain Python integers. But this argumentation fails just like it fails for more complex kind of objects. Wrapping them just like everything else is the cleanest solution. You could introduce case testing wherever you use a wrapped object, to know if it is a plain integer or an instance of (a subclass of) W_Object. But that makes the whole program more complicated. The equivalent in CPython would be to use PyObject* pointers all around except when the object is an integer (after all, integers are directly available in C too). You could represent small integers as odd-valuated pointers. But it puts extra burden on the whole C code, so the CPython team avoided it. (In our case it is an optimization that we eventually made, but not hard-coded at this level - see Standard Interpreter Optimizations.)
So in summary: wrapping integers as instances is the simple path, while using plain integers instead is the complex path, not the other way around.
The larger part of the pypy/objspace/std/ package defines and implements the library of Python’s standard built-in object types. Each type xxx (int, float, list, tuple, str, type, etc.) is typically implemented in the module xxxobject.py.
The W_AbstractXxxObject class, when present, is the abstract base class, which mainly defines what appears on the Python-level type object. There are then actual implementations as subclasses, which are called W_XxxObject or some variant for the cases where we have several different implementations. For example, pypy/objspace/std/bytesobject.py defines W_AbstractBytesObject, which contains everything needed to build the str app-level type; and there are subclasses W_BytesObject (the usual string) and W_StringBufferObject (a special implementation tweaked for repeated additions, in pypy/objspace/std/strbufobject.py). For mutable data types like lists and dictionaries, we have a single class W_ListObject or W_DictMultiObject which has an indirection to the real data and a strategy; the strategy can change as the content of the object changes.
From the user’s point of view, even when there are several W_AbstractXxxObject subclasses, this is not visible: at the app-level, they are still all instances of exactly the same Python type. PyPy knows that (e.g.) the application-level type of its interpreter-level W_BytesObject instances is str because there is a typedef class attribute in W_BytesObject which points back to the string type specification from pypy/objspace/std/bytesobject.py; all other implementations of strings use the same typedef from pypy/objspace/std/bytesobject.py.
For other examples of multiple implementations of the same Python type, see Standard Interpreter Optimizations.
Note: multimethods are on the way out. Although they look cool, they failed to provide enough benefits.
The Standard Object Space allows multiple object implementations per Python type - this is based on multimethods. For a description of the multimethod variant that we implemented and which features it supports, see the comment at the start of pypy/objspace/std/multimethod.py. However, multimethods alone are not enough for the Standard Object Space: the complete picture spans several levels in order to emulate the exact Python semantics.
Consider the example of the space.getitem(w_a, w_b) operation, corresponding to the application-level syntax a[b]. The Standard Object Space contains a corresponding getitem multimethod and a family of functions that implement the multimethod for various combination of argument classes - more precisely, for various combinations of the interpreter-level classes of the arguments. Here are some examples of functions implementing the getitem multimethod:
- getitem__Tuple_ANY: called when the first argument is a W_TupleObject, this function converts its second argument to an integer and performs tuple indexing.
- getitem__Tuple_Slice: called when the first argument is a W_TupleObject and the second argument is a W_SliceObject. This version takes precedence over the previous one if the indexing is done with a slice object, and performs tuple slicing instead.
- getitem__String_Slice: called when the first argument is a W_StringObject and the second argument is a slice object.
Note how the multimethod dispatch logic helps writing new object implementations without having to insert hooks into existing code. Note first how we could have defined a regular method-based API that new object implementations must provide, and call these methods from the space operations. The problem with this approach is that some Python operators are naturally binary or N-ary. Consider for example the addition operation: for the basic string implementation it is a simple concatenation-by-copy, but it can have a rather more subtle implementation for strings done as ropes. It is also likely that concatenating a basic string with a rope string could have its own dedicated implementation - and yet another implementation for a rope string with a basic string. With multimethods, we can have an orthogonally-defined implementation for each combination.
The multimethods mechanism also supports delegate functions, which are converters between two object implementations. The dispatch logic knows how to insert calls to delegates if it encounters combinations of interp-level classes which is not directly implemented. For example, we have no specific implementation for the concatenation of a basic string and a StringSlice object; when the user adds two such strings, then the StringSlice object is converted to a basic string (that is, a temporarily copy is built), and the concatenation is performed on the resulting pair of basic strings. This is similar to the C++ method overloading resolution mechanism (but occurs at runtime).
The complete picture is more complicated because the Python object model is based on descriptors: the types int, str, etc. must have methods __add__, __mul__, etc. that take two arguments including the self. These methods must perform the operation or return NotImplemented if the second argument is not of a type that it doesn’t know how to handle.
The Standard Object Space creates these methods by slicing the multimethod tables. Each method is automatically generated from a subset of the registered implementations of the corresponding multimethod. This slicing is performed on the first argument, in order to keep only the implementations whose first argument’s interpreter-level class matches the declared Python-level type.
For example, in a baseline PyPy, int.__add__ is just calling the function add__Int_Int, which is the only registered implementation for add whose first argument is an implementation of the int Python type. On the other hand, if we enable integers implemented as tagged pointers, then there is another matching implementation: add__SmallInt_SmallInt. In this case, the Python-level method int.__add__ is implemented by trying to dispatch between these two functions based on the interp-level type of the two arguments.
Similarly, the reverse methods (__radd__ and others) are obtained by slicing the multimethod tables to keep only the functions whose second argument has the correct Python-level type.
Slicing is actually a good way to reproduce the details of the object model as seen in CPython: slicing is attempted for every Python types for every multimethod, but the __xyz__ Python methods are only put into the Python type when the resulting slices are not empty. This is how our int type has no __getitem__ method, for example. Additionally, slicing ensures that 5 .__add__(6L) correctly returns NotImplemented (because this particular slice does not include add__Long_Long and there is no add__Int_Long), which leads to 6L.__radd__(5) being called, as in CPython.
The task of the FlowObjSpace (the source is at rpython/flowspace/) is to generate a control-flow graph from a function. This graph will also contain a trace of the individual operations, so that it is actually just an alternate representation for the function.
The FlowObjSpace is an object space, which means that it exports the standard object space interface and it is driven by the bytecode interpreter.
The basic idea is that if the bytecode interpreter is given a function, e.g.:
def f(n): return 3*n+2
it will do whatever bytecode dispatching and stack-shuffling needed, during which it issues a sequence of calls to the object space. The FlowObjSpace merely records these calls (corresponding to “operations”) in a structure called a basic block. To track which value goes where, the FlowObjSpace invents placeholder “wrapped objects” and give them to the interpreter, so that they appear in some next operation. This technique is an example of Abstract Interpretation.
For example, if the placeholder v1 is given as the argument to the above function, the bytecode interpreter will call v2 = space.mul(space.wrap(3), v1) and then v3 = space.add(v2, space.wrap(2)) and return v3 as the result. During these calls the FlowObjSpace will record a basic block:
Block(v1): # input argument v2 = mul(Constant(3), v1) v3 = add(v2, Constant(2))
The data structures built up by the flow object space are described in the translation document.
The FlowObjSpace works by recording all operations issued by the bytecode interpreter into basic blocks. A basic block ends in one of two cases: when the bytecode interpreters calls is_true(), or when a joinpoint is reached.
- A joinpoint occurs when the next operation is about to be recorded into the current block, but there is already another block that records an operation for the same bytecode position. This means that the bytecode interpreter has closed a loop and is interpreting already-seen code again. In this situation, we interrupt the bytecode interpreter and we make a link from the end of the current block back to the previous block, thus closing the loop in the flow graph as well. (Note that this occurs only when an operation is about to be recorded, which allows some amount of constant-folding.)
- If the bytecode interpreter calls is_true(), the FlowObjSpace doesn’t generally know if the answer should be True or False, so it puts a conditional jump and generates two successor blocks for the current basic block. There is some trickery involved so that the bytecode interpreter is fooled into thinking that is_true() first returns False (and the subsequent operations are recorded in the first successor block), and later the same call to is_true() also returns True (and the subsequent operations go this time to the other successor block).
(This section to be extended...)