- Introduction and Overview
- Bytecode Interpreter Implementation Classes
This document describes the implementation of PyPy's Bytecode Interpreter and related Virtual Machine functionalities.
PyPy's bytecode interpreter has a structure reminiscent of CPython's Virtual Machine: It processes code objects parsed and compiled from Python source code. It is implemented in the `pypy/interpreter/`_ directory. People familiar with the CPython implementation will easily recognize similar concepts there. The major differences are the overall usage of the object space indirection to perform operations on objects, and the organization of the built-in modules (described here).
Code objects are a nicely preprocessed, structured representation of source code, and their main content is bytecode. We use the same compact bytecode format as CPython 2.7, with minor differences in the bytecode set. Our bytecode compiler is implemented as a chain of flexible passes (tokenizer, lexer, parser, abstract syntax tree builder, bytecode generator). The latter passes are based on the compiler package from the standard library of CPython, with various improvements and bug fixes. The bytecode compiler (living under `pypy/interpreter/astcompiler/`_) is now integrated and is translated with the rest of PyPy.
Code objects contain condensed information about their respective functions, class and module body source codes. Interpreting such code objects means instantiating and initializing a Frame class and then calling its frame.eval() method. This main entry point initialize appropriate namespaces and then interprets each bytecode instruction. Python's standard library contains the `lib-python/2.7/dis.py`_ module which allows to view the Virtual's machine bytecode instructions:
>>> import dis >>> def f(x): ... return x + 1 >>> dis.dis(f) 2 0 LOAD_FAST 0 (x) 3 LOAD_CONST 1 (1) 6 BINARY_ADD 7 RETURN_VALUE
CPython as well as PyPy are stack-based virtual machines, i.e. they don't have registers but put object to and pull objects from a stack. The bytecode interpreter is only responsible for implementing control flow and putting and pulling black box objects to and from this value stack. The bytecode interpreter does not know how to perform operations on those black box (wrapped) objects for which it delegates to the object space. In order to implement a conditional branch in a program's execution, however, it needs to gain minimal knowledge about a wrapped object. Thus, each object space has to offer a is_true(w_obj) operation which returns an interpreter-level boolean value.
For the understanding of the interpreter's inner workings it is crucial to recognize the concepts of interpreter-level and application-level code. In short, interpreter-level is executed directly on the machine and invoking application-level functions leads to an bytecode interpretation indirection. However, special care must be taken regarding exceptions because application level exceptions are wrapped into OperationErrors which are thus distinguished from plain interpreter-level exceptions. See application level exceptions for some more information on OperationErrors.
The interpreter implementation offers mechanisms to allow a caller to be unaware if a particular function invocation leads to bytecode interpretation or is executed directly at interpreter-level. The two basic kinds of Gateway classes expose either an interpreter-level function to application-level execution (interp2app) or allow transparent invocation of application-level helpers (app2interp) at interpreter-level.
Another task of the bytecode interpreter is to care for exposing its basic code, frame, module and function objects to application-level code. Such runtime introspection and modification abilities are implemented via interpreter descriptors (also see Raymond Hettingers how-to guide for descriptors in Python, PyPy uses this model extensively).
A significant complexity lies in function argument parsing. Python as a language offers flexible ways of providing and receiving arguments for a particular function invocation. Not only does it take special care to get this right, it also presents difficulties for the annotation pass which performs a whole-program analysis on the bytecode interpreter, argument parsing and gatewaying code in order to infer the types of all values flowing across function calls.
It is for this reason that PyPy resorts to generate specialized frame classes and functions at initialization time in order to let the annotator only see rather static program flows with homogeneous name-value assignments on function invocations.
The concept of Frames is pervasive in executing programs and on virtual machines in particular. They are sometimes called execution frame because they hold crucial information regarding the execution of a Code object, which in turn is often directly related to a Python Function. Frame instances hold the following state:
- the local scope holding name-value bindings, usually implemented via a "fast scope" which is an array of wrapped objects
- a blockstack containing (nested) information regarding the control flow of a function (such as while and try constructs)
- a value stack where bytecode interpretation pulls object from and puts results on.
- a reference to the globals dictionary, containing module-level name-value bindings
- debugging information from which a current line-number and file location can be constructed for tracebacks
Moreover the Frame class itself has a number of methods which implement the actual bytecodes found in a code object. The methods of the PyFrame class are added in various files:
PyPy's code objects contain the same information found in CPython's code objects. They differ from Function objects in that they are only immutable representations of source code and don't contain execution state or references to the execution environment found in Frames. Frames and Functions have references to a code object. Here is a list of Code attributes:
- co_flags flags if this code object has nested scopes/generators
- co_stacksize the maximum depth the stack can reach while executing the code
- co_code the actual bytecode string
- co_argcount number of arguments this code object expects
- co_varnames a tuple of all argument names pass to this code object
- co_nlocals number of local variables
- co_names a tuple of all names used in the code object
- co_consts a tuple of prebuilt constant objects ("literals") used in the code object
- co_cellvars a tuple of Cells containing values for access from nested scopes
- co_freevars a tuple of Cell names from "above" scopes
- co_filename source file this code object was compiled from
- co_firstlineno the first linenumber of the code object in its source file
- co_name name of the code object (often the function name)
- co_lnotab a helper table to compute the line-numbers corresponding to bytecodes
In PyPy, code objects also have the responsibility of creating their Frame objects via the 'create_frame()` method. With proper parser and compiler support this would allow to create custom Frame objects extending the execution of functions in various ways. The several Frame classes already utilize this flexibility in order to implement Generators and Nested Scopes.
The PyPy Function class (in `pypy/interpreter/function.py`_) represents a Python function. A Function carries the following main attributes:
- func_doc the docstring (or None)
- func_name the name of the function
- func_code the Code object representing the function source code
- func_defaults default values for the function (built at function definition time)
- func_dict dictionary for additional (user-defined) function attributes
- func_globals reference to the globals dictionary
- func_closure a tuple of Cell references
Functions classes also provide a __get__ descriptor which creates a Method object holding a binding to an instance or a class. Finally, Functions and Methods both offer a call_args() method which executes the function given an Arguments class instance.
The Argument class (in `pypy/interpreter/argument.py`_) is responsible for parsing arguments passed to functions. Python has rather complex argument-passing concepts:
- positional arguments
- keyword arguments specified by name
- default values for positional arguments, defined at function definition time
- "star args" allowing a function to accept remaining positional arguments
- "star keyword args" allow a function to accept additional arbitrary name-value bindings
Moreover, a Function object can get bound to a class or instance in which case the first argument to the underlying function becomes the bound object. The Arguments provides means to allow all this argument parsing and also cares for error reporting.
A Module instance represents execution state usually constructed from executing the module's source file. In addition to such a module's global __dict__ dictionary it has the following application level attributes:
- __doc__ the docstring of the module
- __file__ the source filename from which this module was instantiated
- __path__ state used for relative imports
Apart from the basic Module used for importing application-level files there is a more refined MixedModule class (see `pypy/interpreter/mixedmodule.py`_) which allows to define name-value bindings both at application level and at interpreter level. See the __builtin__ module's `pypy/module/__builtin__/__init__.py`_ file for an example and the higher level chapter on Modules in the coding guide.
A unique PyPy property is the ability to easily cross the barrier between interpreted and machine-level code (often referred to as the difference between interpreter-level and application-level). Be aware that the according code (in `pypy/interpreter/gateway.py`_) for crossing the barrier in both directions is somewhat involved, mostly due to the fact that the type-inferring annotator needs to keep track of the types of objects flowing across those barriers.
In order to make an interpreter-level function available at application level, one invokes pypy.interpreter.gateway.interp2app(func). Such a function usually takes a space argument and any number of positional arguments. Additionally, such functions can define an unwrap_spec telling the interp2app logic how application-level provided arguments should be unwrapped before the actual interpreter-level function is invoked. For example, interpreter descriptors such as the Module.__new__ method for allocating and constructing a Module instance are defined with such code:
Module.typedef = TypeDef("module", __new__ = interp2app(Module.descr_module__new__.im_func, unwrap_spec=[ObjSpace, W_Root, Arguments]), __init__ = interp2app(Module.descr_module__init__), # module dictionaries are readonly attributes __dict__ = GetSetProperty(descr_get_dict, cls=Module), __doc__ = 'module(name[, doc])\n\nCreate a module object...' )
The actual Module.descr_module__new__ interpreter-level method referenced from the __new__ keyword argument above is defined like this:
def descr_module__new__(space, w_subtype, __args__): module = space.allocate_instance(Module, w_subtype) Module.__init__(module, space, None) return space.wrap(module)
Summarizing, the interp2app mechanism takes care to route an application level access or call to an internal interpreter-level object appropriately to the descriptor, providing enough precision and hints to keep the type-inferring annotator happy.
Application level code is often preferable. Therefore, we often like to invoke application level code from interpreter-level. This is done via the Gateway's app2interp mechanism which we usually invoke at definition time in a module. It generates a hook which looks like an interpreter-level function accepting a space and an arbitrary number of arguments. When calling a function at interpreter-level the caller side does usually not need to be aware if its invoked function is run through the PyPy interpreter or if it will directly execute on the machine (after translation).
Here is an example showing how we implement the Metaclass finding algorithm of the Python language in PyPy:
app = gateway.applevel(r''' def find_metaclass(bases, namespace, globals, builtin): if '__metaclass__' in namespace: return namespace['__metaclass__'] elif len(bases) > 0: base = bases if hasattr(base, '__class__'): return base.__class__ else: return type(base) elif '__metaclass__' in globals: return globals['__metaclass__'] else: try: return builtin.__metaclass__ except AttributeError: return type ''', filename=__file__) find_metaclass = app.interphook('find_metaclass')
The find_metaclass interpreter-level hook is invoked with five arguments from the BUILD_CLASS opcode implementation in `pypy/interpreter/pyopcode.py`_:
def BUILD_CLASS(f): w_methodsdict = f.valuestack.pop() w_bases = f.valuestack.pop() w_name = f.valuestack.pop() w_metaclass = find_metaclass(f.space, w_bases, w_methodsdict, f.w_globals, f.space.wrap(f.builtin)) w_newclass = f.space.call_function(w_metaclass, w_name, w_bases, w_methodsdict) f.valuestack.push(w_newclass)
Note that at a later point we can rewrite the find_metaclass implementation at interpreter-level and we would not have to modify the calling side at all.
Python traditionally has a very far-reaching introspection model for bytecode interpreter related objects. In PyPy and in CPython read and write accesses to such objects are routed to descriptors. Of course, in CPython those are implemented in C while in PyPy they are implemented in interpreter-level Python code.
All instances of a Function, Code, Frame or Module classes are also Wrappable instances which means they can be represented at application level. These days, a PyPy object space needs to work with a basic descriptor lookup when it encounters accesses to an interpreter-level object: an object space asks a wrapped object for its type via a getclass method and then calls the type's lookup(name) function in order to receive a descriptor function. Most of PyPy's internal object descriptors are defined at the end of `pypy/interpreter/typedef.py`_. You can use these definitions as a reference for the exact attributes of interpreter classes visible at application level.