[pypy-commit] pypy default: Documentation for continulets.
noreply at buildbot.pypy.org
Sun Aug 28 16:57:43 CEST 2011
Author: Armin Rigo <arigo at tunes.org>
Date: 2011-08-28 16:55 +0200
Log: Documentation for continulets.
diff --git a/pypy/doc/_ref.txt b/pypy/doc/_ref.txt
@@ -1,11 +1,10 @@
.. _`ctypes_configure/doc/sample.py`: https://bitbucket.org/pypy/pypy/src/default/ctypes_configure/doc/sample.py
.. _`demo/`: https://bitbucket.org/pypy/pypy/src/default/demo/
-.. _`demo/pickle_coroutine.py`: https://bitbucket.org/pypy/pypy/src/default/demo/pickle_coroutine.py
.. _`lib-python/`: https://bitbucket.org/pypy/pypy/src/default/lib-python/
.. _`lib-python/2.7/dis.py`: https://bitbucket.org/pypy/pypy/src/default/lib-python/2.7/dis.py
.. _`lib_pypy/`: https://bitbucket.org/pypy/pypy/src/default/lib_pypy/
+.. _`lib_pypy/greenlet.py`: https://bitbucket.org/pypy/pypy/src/default/lib_pypy/greenlet.py
.. _`lib_pypy/pypy_test/`: https://bitbucket.org/pypy/pypy/src/default/lib_pypy/pypy_test/
-.. _`lib_pypy/stackless.py`: https://bitbucket.org/pypy/pypy/src/default/lib_pypy/stackless.py
.. _`lib_pypy/tputil.py`: https://bitbucket.org/pypy/pypy/src/default/lib_pypy/tputil.py
.. _`pypy/annotation/`: https://bitbucket.org/pypy/pypy/src/default/pypy/annotation/
@@ -55,7 +54,6 @@
.. _`pypy/module/`: https://bitbucket.org/pypy/pypy/src/default/pypy/module/
.. _`pypy/module/__builtin__/__init__.py`: https://bitbucket.org/pypy/pypy/src/default/pypy/module/__builtin__/__init__.py
-.. _`pypy/module/_stackless/test/test_composable_coroutine.py`: https://bitbucket.org/pypy/pypy/src/default/pypy/module/_stackless/test/test_composable_coroutine.py
.. _`pypy/objspace/`: https://bitbucket.org/pypy/pypy/src/default/pypy/objspace/
.. _`pypy/objspace/dump.py`: https://bitbucket.org/pypy/pypy/src/default/pypy/objspace/dump.py
@@ -117,6 +115,7 @@
.. _`pypy/translator/`: https://bitbucket.org/pypy/pypy/src/default/pypy/translator/
.. _`pypy/translator/backendopt/`: https://bitbucket.org/pypy/pypy/src/default/pypy/translator/backendopt/
.. _`pypy/translator/c/`: https://bitbucket.org/pypy/pypy/src/default/pypy/translator/c/
+.. _`pypy/translator/c/src/stacklet/`: https://bitbucket.org/pypy/pypy/src/default/pypy/translator/c/src/stacklet/
.. _`pypy/translator/cli/`: https://bitbucket.org/pypy/pypy/src/default/pypy/translator/cli/
.. _`pypy/translator/goal/`: https://bitbucket.org/pypy/pypy/src/default/pypy/translator/goal/
.. _`pypy/translator/jvm/`: https://bitbucket.org/pypy/pypy/src/default/pypy/translator/jvm/
diff --git a/pypy/doc/architecture.rst b/pypy/doc/architecture.rst
@@ -153,7 +153,7 @@
* Optionally, `various transformations`_ can then be applied which, for
example, perform optimizations such as inlining, add capabilities
- such as stackless_-style concurrency, or insert code for the
+ such as stackless-style concurrency (deprecated), or insert code for the
* Then, the graphs are converted to source code for the target platform
@@ -255,7 +255,6 @@
.. _Python: http://docs.python.org/reference/
.. _Psyco: http://psyco.sourceforge.net
-.. _stackless: stackless.html
.. _`generate Just-In-Time Compilers`: jit/index.html
.. _`JIT Generation in PyPy`: jit/index.html
.. _`implement your own interpreter`: http://morepypy.blogspot.com/2011/04/tutorial-writing-interpreter-with-pypy.html
diff --git a/pypy/doc/cpython_differences.rst b/pypy/doc/cpython_differences.rst
@@ -24,6 +24,7 @@
@@ -84,10 +85,6 @@
- Extra module with Stackless_ only:
Note that only some of these modules are built-in in a typical
CPython installation, and the rest is from non built-in extension
modules. This means that e.g. ``import parser`` will, on CPython,
@@ -108,11 +105,11 @@
.. the nonstandard modules are listed below...
.. _`__pypy__`: __pypy__-module.html
+.. _`_continuation`: stackless.html
.. _`_ffi`: ctypes-implementation.html
.. _`_rawffi`: ctypes-implementation.html
.. _`_minimal_curses`: config/objspace.usemodules._minimal_curses.html
.. _`cpyext`: http://morepypy.blogspot.com/2010/04/using-cpython-extension-modules-with.html
-.. _Stackless: stackless.html
Differences related to garbage collection strategies
diff --git a/pypy/doc/getting-started-python.rst b/pypy/doc/getting-started-python.rst
@@ -67,7 +67,6 @@
* ``libssl-dev`` (for the optional ``_ssl`` module)
* ``libgc-dev`` (for the Boehm garbage collector: only needed when translating with `--opt=0, 1` or `size`)
* ``python-sphinx`` (for the optional documentation build. You need version 1.0.7 or later)
- * ``python-greenlet`` (for the optional stackless support in interpreted mode/testing)
3. Translation is time-consuming -- 45 minutes on a very fast machine --
@@ -120,19 +119,8 @@
The ``translate.py`` script takes a very large number of options controlling
-what to translate and how. See ``translate.py -h``. Some of the more
-interesting options (but for now incompatible with the JIT) are:
- * ``--stackless``: this produces a pypy-c that includes features
- inspired by `Stackless Python <http://www.stackless.com>`__.
- * ``--gc=boehm|ref|marknsweep|semispace|generation|hybrid|minimark``:
- choose between using
- the `Boehm-Demers-Weiser garbage collector`_, our reference
- counting implementation or one of own collector implementations
- (the default depends on the optimization level but is usually
+what to translate and how. See ``translate.py -h``. The default options
+should be suitable for mostly everybody by now.
Find a more detailed description of the various options in our `configuration
diff --git a/pypy/doc/how-to-release.rst b/pypy/doc/how-to-release.rst
@@ -42,7 +42,6 @@
JIT: windows, linux, os/x
no JIT: windows, linux, os/x
sandbox: linux, os/x
- stackless: windows, linux, os/x
* write release announcement pypy/doc/release-x.y(.z).txt
the release announcement should contain a direct link to the download page
diff --git a/pypy/doc/index.rst b/pypy/doc/index.rst
@@ -35,7 +35,7 @@
* `Differences between PyPy and CPython`_
* `What PyPy can do for your objects`_
- * `Stackless and coroutines`_
+ * `Continulets and greenlets`_
* `JIT Generation in PyPy`_
* `Sandboxing Python code`_
@@ -292,8 +292,6 @@
`pypy/translator/jvm/`_ the Java backend
-`pypy/translator/stackless/`_ the `Stackless Transform`_
`pypy/translator/tool/`_ helper tools for translation, including the Pygame
@@ -318,7 +316,7 @@
.. _`transparent proxies`: objspace-proxies.html#tproxy
.. _`Differences between PyPy and CPython`: cpython_differences.html
.. _`What PyPy can do for your objects`: objspace-proxies.html
-.. _`Stackless and coroutines`: stackless.html
+.. _`Continulets and greenlets`: stackless.html
.. _StdObjSpace: objspace.html#the-standard-object-space
.. _`abstract interpretation`: http://en.wikipedia.org/wiki/Abstract_interpretation
.. _`rpython`: coding-guide.html#rpython
@@ -337,7 +335,6 @@
.. _`low-level type system`: rtyper.html#low-level-type
.. _`object-oriented type system`: rtyper.html#oo-type
.. _`garbage collector`: garbage_collection.html
-.. _`Stackless Transform`: translation.html#the-stackless-transform
.. _`main PyPy-translation scripts`: getting-started-python.html#translating-the-pypy-python-interpreter
.. _`.NET`: http://www.microsoft.com/net/
.. _Mono: http://www.mono-project.com/
diff --git a/pypy/doc/rlib.rst b/pypy/doc/rlib.rst
@@ -134,69 +134,6 @@
a hierarchy of Address classes, in a typical static-OO-programming style.
-The `pypy/rlib/rstack.py`_ module allows an RPython program to control its own execution stack.
-This is only useful if the program is translated using stackless. An old
-description of the exposed functions is below.
-We introduce an RPython type ``frame_stack_top`` and a built-in function
-``yield_current_frame_to_caller()`` that work as follows (see example below):
-* The built-in function ``yield_current_frame_to_caller()`` causes the current
- function's state to be captured in a new ``frame_stack_top`` object that is
- returned to the parent. Only one frame, the current one, is captured this
- way. The current frame is suspended and the caller continues to run. Note
- that the caller is only resumed once: when
- ``yield_current_frame_to_caller()`` is called. See below.
-* A ``frame_stack_top`` object can be jumped to by calling its ``switch()``
- method with no argument.
-* ``yield_current_frame_to_caller()`` and ``switch()`` themselves return a new
- ``frame_stack_top`` object: the freshly captured state of the caller of the
- source ``switch()`` that was just executed, or None in the case described
-* the function that called ``yield_current_frame_to_caller()`` also has a
- normal return statement, like all functions. This statement must return
- another ``frame_stack_top`` object. The latter is *not* returned to the
- original caller; there is no way to return several times to the caller.
- Instead, it designates the place to which the execution must jump, as if by
- a ``switch()``. The place to which we jump this way will see a None as the
- source frame stack top.
-* every frame stack top must be resumed once and only once. Not resuming
- it at all causes a leak. Resuming it several times causes a crash.
-* a function that called ``yield_current_frame_to_caller()`` should not raise.
- It would have no implicit parent frame to propagate the exception to. That
- would be a crashingly bad idea.
-The following example would print the numbers from 1 to 7 in order::
- def g():
- print 2
- frametop_before_5 = yield_current_frame_to_caller()
- print 4
- frametop_before_7 = frametop_before_5.switch()
- print 6
- return frametop_before_7
- def f():
- print 1
- frametop_before_4 = g()
- print 3
- frametop_before_6 = frametop_before_4.switch()
- print 5
- frametop_after_return = frametop_before_6.switch()
- print 7
- assert frametop_after_return is None
diff --git a/pypy/doc/stackless.rst b/pypy/doc/stackless.rst
@@ -8,446 +8,279 @@
PyPy can expose to its user language features similar to the ones
-present in `Stackless Python`_: **no recursion depth limit**, and the
-ability to write code in a **massively concurrent style**. It actually
-exposes three different paradigms to choose from:
+present in `Stackless Python`_: the ability to write code in a
+**massively concurrent style**. (It does not (any more) offer the
+ability to run with no `recursion depth limit`_, but the same effect
+can be achieved indirectly.)
-* `Tasklets and channels`_;
+This feature is based on a custom primitive called a continulet_.
+Continulets can be directly used by application code, or it is possible
+to write (entirely at app-level) more user-friendly interfaces.
+Currently PyPy implements greenlets_ on top of continulets. It would be
+easy to implement tasklets and channels as well, emulating the model
+of `Stackless Python`_.
-* Plain coroutines_.
+Continulets are extremely light-weight, which means that PyPy should be
+able to handle programs containing large amounts of them. However, due
+to an implementation restriction, a PyPy compiled with
+``--gcrootfinder=shadowstack`` consumes at least one page of physical
+memory (4KB) per live continulet, and half a megabyte of virtual memory
+on 32-bit or a complete megabyte on 64-bit. Moreover, the feature is
+only available (so far) on x86 and x86-64 CPUs; for other CPUs you need
+to add a short page of custom assembler to
-All of them are extremely light-weight, which means that PyPy should be
-able to handle programs containing large amounts of coroutines, tasklets
+The fundamental idea is that, at any point in time, the program happens
+to run one stack of frames (or one per thread, in case of
+multi-threading). To see the stack, start at the top frame and follow
+the chain of ``f_back`` until you reach the bottom frame. From the
+point of view of one of these frames, it has a ``f_back`` pointing to
+another frame (unless it is the bottom frame), and it is itself being
+pointed to by another frame (unless it is the top frame).
-If you are running py.py on top of CPython, then you need to enable
-the _stackless module by running it as follows::
+The theory behind continulets is to literally take the previous sentence
+as definition of "an O.K. situation". The trick is that there are
+O.K. situations that are more complex than just one stack: you will
+always have one stack, but you can also have in addition one or more
+detached *cycles* of frames, such that by following the ``f_back`` chain
+you run in a circle. But note that these cycles are indeed completely
+detached: the top frame (the currently running one) is always the one
+which is not the ``f_back`` of anybody else, and it is always the top of
+a stack that ends with the bottom frame, never a part of these extra
- py.py --withmod-_stackless
+How do you create such cycles? The fundamental operation to do so is to
+take two frames and *permute* their ``f_back`` --- i.e. exchange them.
+You can permute any two ``f_back`` without breaking the rule of "an O.K.
+situation". Say for example that ``f`` is some frame halfway down the
+stack, and you permute its ``f_back`` with the ``f_back`` of the top
+frame. Then you have removed from the normal stack all intermediate
+frames, and turned them into one stand-alone cycle. By doing the same
+permutation again you restore the original situation.
-This is implemented internally using greenlets, so it only works on a
-platform where `greenlets`_ are supported. A few features do
-not work this way, though, and really require a translated
+In practice, in PyPy, you cannot change the ``f_back`` of an abitrary
+frame, but only of frames stored in ``continulets``.
-To obtain a translated version of ``pypy-c`` that includes Stackless
-support, run translate.py as follows::
- cd pypy/translator/goal
- python translate.py --stackless
+Continulets are internally implemented using stacklets. Stacklets are a
+bit more primitive (they are really one-shot continuations), but that
+idea only works in C, not in Python. The basic idea of continulets is
+to have at any point in time a complete valid stack; this is important
+e.g. to correctly propagate exceptions (and it seems to give meaningful
Application level interface
-A stackless PyPy contains a module called ``stackless``. The interface
-exposed by this module have not been refined much, so it should be
-considered in-flux (as of 2007).
-So far, PyPy does not provide support for ``stackless`` in a threaded
-environment. This limitation is not fundamental, as previous experience
-has shown, so supporting this would probably be reasonably easy.
-An interesting point is that the same ``stackless`` module can provide
-a number of different concurrency paradigms at the same time. From a
-theoretical point of view, none of above-mentioned existing three
-paradigms considered on its own is new: two of them are from previous
-Python work, and the third one is a variant of the classical coroutine.
-The new part is that the PyPy implementation manages to provide all of
-them and let the user implement more. Moreover - and this might be an
-important theoretical contribution of this work - we manage to provide
-these concurrency concepts in a "composable" way. In other words, it
-is possible to naturally mix in a single application multiple
-concurrency paradigms, and multiple unrelated usages of the same
-paradigm. This is discussed in the Composability_ section below.
+A translated PyPy contains by default a module called ``_continuation``
+exporting the type ``continulet``. A ``continulet`` object from this
+module is a container that stores a "one-shot continuation". It plays
+the role of an extra frame you can insert in the stack, and whose
+``f_back`` can be changed.
+To make a continulet object, call ``continulet()`` with a callable and
+optional extra arguments.
-Any stackless PyPy executable natively supports recursion that is only
-limited by the available memory. As in normal Python, though, there is
-an initial recursion limit (which is 5000 in all pypy-c's, and 1000 in
-CPython). It can be changed with ``sys.setrecursionlimit()``. With a
-stackless PyPy, any value is acceptable - use ``sys.maxint`` for
+Later, the first time you ``switch()`` to the continulet, the callable
+is invoked with the same continulet object as the extra first argument.
+At that point, the one-shot continuation stored in the continulet points
+to the caller of ``switch()``. In other words you have a perfectly
+normal-looking stack of frames. But when ``switch()`` is called again,
+this stored one-shot continuation is exchanged with the current one; it
+means that the caller of ``switch()`` is suspended with its continuation
+stored in the container, and the old continuation from the continulet
+object is resumed.
-In some cases, you can write Python code that causes interpreter-level
-infinite recursion -- i.e. infinite recursion without going via
-application-level function calls. It is possible to limit that too,
-with ``_stackless.set_stack_depth_limit()``, or to unlimit it completely
-by setting it to ``sys.maxint``.
+The most primitive API is actually 'permute()', which just permutes the
+one-shot continuation stored in two (or more) continulets.
+In more details:
+* ``continulet(callable, *args, **kwds)``: make a new continulet.
+ Like a generator, this only creates it; the ``callable`` is only
+ actually called the first time it is switched to. It will be
+ called as follows::
-A Coroutine is similar to a very small thread, with no preemptive scheduling.
-Within a family of coroutines, the flow of execution is explicitly
-transferred from one to another by the programmer. When execution is
-transferred to a coroutine, it begins to execute some Python code. When
-it transfers execution away from itself it is temporarily suspended, and
-when execution returns to it it resumes its execution from the
-point where it was suspended. Conceptually, only one coroutine is
-actively running at any given time (but see Composability_ below).
+ callable(cont, *args, **kwds)
-The ``stackless.coroutine`` class is instantiated with no argument.
-It provides the following methods and attributes:
+ where ``cont`` is the same continulet object.
+ Note that it is actually ``cont.__init__()`` that binds
+ the continulet. It is also possible to create a not-bound-yet
+ continulet by calling explicitly ``continulet.__new__()``, and
+ only bind it later by calling explicitly ``cont.__init__()``.
- Static method returning the currently running coroutine. There is a
- so-called "main" coroutine object that represents the "outer"
- execution context, where your main program started and where it runs
- as long as it does not switch to another coroutine.
+* ``cont.switch(value=None, to=None)``: start the continulet if
+ it was not started yet. Otherwise, store the current continuation
+ in ``cont``, and activate the target continuation, which is the
+ one that was previously stored in ``cont``. Note that the target
+ continuation was itself previously suspended by another call to
+ ``switch()``; this older ``switch()`` will now appear to return.
+ The ``value`` argument is any object that is carried to the target
+ and returned by the target's ``switch()``.
-* ``coro.bind(callable, *args, **kwds)``
+ If ``to`` is given, it must be another continulet object. In
+ that case, performs a "double switch": it switches as described
+ above to ``cont``, and then immediately switches again to ``to``.
+ This is different from switching directly to ``to``: the current
+ continuation gets stored in ``cont``, the old continuation from
+ ``cont`` gets stored in ``to``, and only then we resume the
+ execution from the old continuation out of ``to``.
- Bind the coroutine so that it will execute ``callable(*args,
- **kwds)``. The call is not performed immediately, but only the
- first time we call the ``coro.switch()`` method. A coroutine must
- be bound before it is switched to. When the coroutine finishes
- (because the call to the callable returns), the coroutine exits and
- implicitly switches back to another coroutine (its "parent"); after
- this point, it is possible to bind it again and switch to it again.
- (Which coroutine is the parent of which is not documented, as it is
- likely to change when the interface is refined.)
+* ``cont.throw(type, value=None, tb=None, to=None)``: similar to
+ ``switch()``, except that immediately after the switch is done, raise
+ the given exception in the target.
+* ``cont.is_pending()``: return True if the continulet is pending.
+ This is False when it is not initialized (because we called
+ ``__new__`` and not ``__init__``) or when it is finished (because
+ the ``callable()`` returned). When it is False, the continulet
+ object is empty and cannot be ``switch()``-ed to.
- Suspend the current (caller) coroutine, and resume execution in the
- target coroutine ``coro``.
+* ``permute(*continulets)``: a global function that permutes the
+ continuations stored in the given continulets arguments. Mostly
+ theoretical. In practice, using ``cont.switch()`` is easier and
+ more efficient than using ``permute()``; the latter does not on
+ its own change the currently running frame.
- Kill ``coro`` by sending a CoroutineExit exception and switching
- execution immediately to it. This exception can be caught in the
- coroutine itself and can be raised from any call to ``coro.switch()``.
- This exception isn't propagated to the parent coroutine.
-* ``coro.throw(type, value)``
+The ``_continuation`` module also exposes the ``generator`` decorator::
- Insert an exception in ``coro`` an resume switches execution
- immediately to it. In the coroutine itself, this exception
- will come from any call to ``coro.switch()`` and can be caught. If the
- exception isn't caught, it will be propagated to the parent coroutine.
+ def f(cont, a, b):
+ cont.switch(a + b)
+ cont.switch(a + b + 1)
-When a coroutine is garbage-collected, it gets the ``.kill()`` method sent to
-it. This happens at the point the next ``.switch`` method is called, so the
-target coroutine of this call will be executed only after the ``.kill`` has
+ for i in f(10, 20):
+ print i
+This example prints 30 and 31. The only advantage over using regular
+generators is that the generator itself is not limited to ``yield``
+statements that must all occur syntactically in the same function.
+Instead, we can pass around ``cont``, e.g. to nested sub-functions, and
+call ``cont.switch(x)`` from there.
-Here is a classical producer/consumer example: an algorithm computes a
-sequence of values, while another consumes them. For our purposes we
-assume that the producer can generate several values at once, and the
-consumer can process up to 3 values in a batch - it can also process
-batches with fewer than 3 values without waiting for the producer (which
-would be messy to express with a classical Python generator). ::
+The ``generator`` decorator can also be applied to methods::
- def producer(lst):
- while True:
- ...compute some more values...
- def consumer(lst):
- while True:
- # First ask the producer for more values if needed
- while len(lst) == 0:
- # Process the available values in a batch, but at most 3
- batch = lst[:3]
- del lst[:3]
- ...process batch...
- # Initialize two coroutines with a shared list as argument
- exchangelst = 
- coro_producer = coroutine()
- coro_producer.bind(producer, exchangelst)
- coro_consumer = coroutine()
- coro_consumer.bind(consumer, exchangelst)
- # Start running the consumer coroutine
-Tasklets and channels
-The ``stackless`` module also provides an interface that is roughly
-compatible with the interface of the ``stackless`` module in `Stackless
-Python`_: it contains ``stackless.tasklet`` and ``stackless.channel``
-classes. Tasklets are also similar to microthreads, but (like coroutines)
-they don't actually run in parallel with other microthreads; instead,
-they synchronize and exchange data with each other over Channels, and
-these exchanges determine which Tasklet runs next.
-For usage reference, see the documentation on the `Stackless Python`_
-Note that Tasklets and Channels are implemented at application-level in
-`lib_pypy/stackless.py`_ on top of coroutines_. You can refer to this
-module for more details and API documentation.
-The stackless.py code tries to resemble the stackless C code as much
-as possible. This makes the code somewhat unpythonic.
-Bird's eye view of tasklets and channels
-Tasklets are a bit like threads: they encapsulate a function in such a way that
-they can be suspended/restarted any time. Unlike threads, they won't
-run concurrently, but must be cooperative. When using stackless
-features, it is vitally important that no action is performed that blocks
-everything else. In particular, blocking input/output should be centralized
-to a single tasklet.
-Communication between tasklets is done via channels.
-There are three ways for a tasklet to give up control:
-1. call ``stackless.schedule()``
-2. send something over a channel
-3. receive something from a channel
-A (live) tasklet can either be running, waiting to get scheduled, or be
-blocked by a channel.
-Scheduling is done in strictly round-robin manner. A blocked tasklet
-is removed from the scheduling queue and will be reinserted when it
-Here is a many-producers many-consumers example, where any consumer can
-process the result of any producer. For this situation we set up a
-single channel where all producer send, and on which all consumers
- def producer(chan):
- while True:
- chan.send(...next value...)
- def consumer(chan):
- while True:
- x = chan.receive()
- ...do something with x...
- # Set up the N producer and M consumer tasklets
- common_channel = stackless.channel()
- for i in range(N):
- stackless.tasklet(producer, common_channel)()
- for i in range(M):
- stackless.tasklet(consumer, common_channel)()
- # Run it all
-Each item sent over the channel is received by one of the waiting
-consumers; which one is not specified. The producers block until their
-item is consumed: the channel is not a queue, but rather a meeting point
-which causes tasklets to block until both a consumer and a producer are
-ready. In practice, the reason for having several consumers receiving
-on a single channel is that some of the consumers can be busy in other
-ways part of the time. For example, each consumer might receive a
-database request, process it, and send the result to a further channel
-before it asks for the next request. In this situation, further
-requests can still be received by other consumers.
+ class X:
+ def f(self, cont, a, b):
-A Greenlet is a kind of primitive Tasklet with a lower-level interface
-and with exact control over the execution order. Greenlets are similar
-to Coroutines, with a slightly different interface: greenlets put more
-emphasis on a tree structure. The various greenlets of a program form a
-precise tree, which fully determines their order of execution.
+Greenlets are implemented on top of continulets in `lib_pypy/greenlet.py`_.
+See the official `documentation of the greenlets`_.
-For usage reference, see the `documentation of the greenlets`_.
-The PyPy interface is identical. You should use ``greenlet.greenlet``
-instead of ``stackless.greenlet`` directly, because the greenlet library
-can give you the latter when you ask for the former on top of PyPy.
+Note that unlike the CPython greenlets, this version does not suffer
+from GC issues: if the program "forgets" an unfinished greenlet, it will
+always be collected at the next garbage collection.
-PyPy's greenlets do not suffer from the cyclic GC limitation that the
-CPython greenlets have: greenlets referencing each other via local
-variables tend to leak on top of CPython (where it is mostly impossible
-to do the right thing). It works correctly on top of PyPy.
+The following features (present in some past Stackless version of PyPy)
+are for the time being not supported any more:
-Coroutines and tasklets can be pickled and unpickled, i.e. serialized to
-a string of bytes for the purpose of storage or transmission. This
-allows "live" coroutines or tasklets to be made persistent, moved to
-other machines, or cloned in any way. The standard ``pickle`` module
-works with coroutines and tasklets (at least in a translated ``pypy-c``;
-unpickling live coroutines or tasklets cannot be easily implemented on
-top of CPython).
+* Tasklets and channels (needs to be rewritten at app-level)
-To be able to achieve this result, we have to consider many objects that
-are not normally pickleable in CPython. Here again, the `Stackless
-Python`_ implementation has paved the way, and we follow the same
-general design decisions: simple internal objects like bound method
-objects and various kinds of iterators are supported; frame objects can
-be fully pickled and unpickled
-(by serializing a reference to the bytecode they are
-running in addition to all the local variables). References to globals
-and modules are pickled by name, similarly to references to functions
-and classes in the traditional CPython ``pickle``.
+* Coroutines (could be rewritten at app-level)
-The "magic" part of this process is the implementation of the unpickling
-of a chain of frames. The Python interpreter of PyPy uses
-interpreter-level recursion to represent application-level calls. The
-reason for this is that it tremendously simplifies the implementation of
-the interpreter itself. Indeed, in Python, almost any operation can
-potentially result in a non-tail-recursive call to another Python
-function. This makes writing a non-recursive interpreter extremely
-tedious; instead, we rely on lower-level transformations during the
-translation process to control this recursion. This is the `Stackless
-Transform`_, which is at the heart of PyPy's support for stackless-style
+* Pickling and unpickling continulets (probably not too hard)
-At any point in time, a chain of Python-level frames corresponds to a
-chain of interpreter-level frames (e.g. C frames in pypy-c), where each
-single Python-level frame corresponds to one or a few interpreter-level
-frames - depending on the length of the interpreter-level call chain
-from one bytecode evaluation loop to the next (recursively invoked) one.
+* Continuing execution of a continulet in a different thread
-This means that it is not sufficient to simply create a chain of Python
-frame objects in the heap of a process before we can resume execution of
-these newly built frames. We must recreate a corresponding chain of
-interpreter-level frames. To this end, we have inserted a few *named
-resume points* (see 3.2.4, in `D07.1 Massive Parallelism and Translation Aspects`_) in the Python interpreter of PyPy. This is the
-motivation for implementing the interpreter-level primitives
-``resume_state_create()`` and ``resume_state_invoke()``, the powerful
-interface that allows an RPython program to artificially rebuild a chain
-of calls in a reflective way, completely from scratch, and jump to it.
+* Automatic unlimited stack (must be emulated__ so far)
-.. _`D07.1 Massive Parallelism and Translation Aspects`: http://codespeak.net/pypy/extradoc/eu-report/D07.1_Massive_Parallelism_and_Translation_Aspects-2007-02-28.pdf
+.. __: `recursion depth limit`_
-(See `demo/pickle_coroutine.py`_ for the complete source of this demo.)
+Recursion depth limit
-Consider a program which contains a part performing a long-running
+You can use continulets to emulate the infinite recursion depth present
+in Stackless Python and in stackless-enabled older versions of PyPy.
- def ackermann(x, y):
- if x == 0:
- return y + 1
- if y == 0:
- return ackermann(x - 1, 1)
- return ackermann(x - 1, ackermann(x, y - 1))
+The trick is to start a continulet "early", i.e. when the recursion
+depth is very low, and switch to it "later", i.e. when the recursion
+depth is high. Example::
-By using pickling, we can save the state of the computation while it is
-running, for the purpose of restoring it later and continuing the
-computation at another time or on a different machine. However,
-pickling does not produce a whole-program dump: it can only pickle
-individual coroutines. This means that the computation should be
-started in its own coroutine::
+ from _continuation import continulet
- # Make a coroutine that will run 'ackermann(3, 8)'
- coro = coroutine()
- coro.bind(ackermann, 3, 8)
+ def invoke(_, callable, arg):
+ return callable(arg)
- # Now start running the coroutine
- result = coro.switch()
+ def bootstrap(c):
+ # this loop runs forever, at a very low recursion depth
+ callable, arg = c.switch()
+ while True:
+ # start a new continulet from here, and switch to
+ # it using an "exchange", i.e. a switch with to=.
+ to = continulet(invoke, callable, arg)
+ callable, arg = c.switch(to=to)
-The coroutine itself must switch back to the main program when it needs
-to be interrupted (we can only pickle suspended coroutines). Due to
-current limitations this requires an explicit check in the
+ c = continulet(bootstrap)
- def ackermann(x, y):
- if interrupt_flag: # test a global flag
- main.switch() # and switch back to 'main' if it is set
- if x == 0:
- return y + 1
- if y == 0:
- return ackermann(x - 1, 1)
- return ackermann(x - 1, ackermann(x, y - 1))
-The global ``interrupt_flag`` would be set for example by a timeout, or
-by a signal handler reacting to Ctrl-C, etc. It causes the coroutine to
-transfer control back to the main program. The execution comes back
-just after the line ``coro.switch()``, where we can pickle the coroutine
+ def recursive(n):
+ if n == 0:
+ return ("ok", n)
+ if n % 200 == 0:
+ prev = c.switch((recursive, n - 1))
+ prev = recursive(n - 1)
+ return (prev, prev + 1)
- if not coro.is_alive:
- print "finished; the result is:", result
- # save the state of the suspended coroutine
- f = open('demo.pickle', 'w')
- pickle.dump(coro, f)
+ print recursive(999999) # prints ('ok', 999999)
-The process can then stop. At any later time, or on another machine,
-we can reload the file and restart the coroutine with::
+Note that if you press Ctrl-C while running this example, the traceback
+will be built with *all* recursive() calls so far, even if this is more
+than the number that can possibly fit in the C stack. These frames are
+"overlapping" each other in the sense of the C stack; more precisely,
+they are copied out of and into the C stack as needed.
- f = open('demo.pickle', 'r')
- coro = pickle.load(f)
- result = coro.switch()
+(The example above also makes use of the following general "guideline"
+to help newcomers write continulets: in ``bootstrap(c)``, only call
+methods on ``c``, not on another continulet object. That's why we wrote
+``c.switch(to=to)`` and not ``to.switch()``, which would mess up the
+state. This is however just a guideline; in general we would recommend
+to use other interfaces like genlets and greenlets.)
-Coroutine pickling is subject to some limitations. First of all, it is
-not a whole-program "memory dump". It means that only the "local" state
-of a coroutine is saved. The local state is defined to include the
-chain of calls and the local variables, but not for example the value of
-any global variable.
-As in normal Python, the pickle will not include any function object's
-code, any class definition, etc., but only references to functions and
-classes. Unlike normal Python, the pickle contains frames. A pickled
-frame stores a bytecode index, representing the current execution
-position. This means that the user program cannot be modified *at all*
-between pickling and unpickling!
-On the other hand, the pickled data is fairly independent from the
-platform and from the PyPy version.
-Pickling/unpickling fails if the coroutine is suspended in a state that
-involves Python frames which were *indirectly* called. To define this
-more precisely, a Python function can issue a regular function or method
-call to invoke another Python function - this is a *direct* call and can
-be pickled and unpickled. But there are many ways to invoke a Python
-function indirectly. For example, most operators can invoke a special
-method ``__xyz__()`` on a class, various built-in functions can call
-back Python functions, signals can invoke signal handlers, and so on.
-These cases are not supported yet.
+Theory of composability
Although the concept of coroutines is far from new, they have not been
generally integrated into mainstream languages, or only in limited form
(like generators in Python and iterators in C#). We can argue that a
possible reason for that is that they do not scale well when a program's
complexity increases: they look attractive in small examples, but the
-models that require explicit switching, by naming the target coroutine,
-do not compose naturally. This means that a program that uses
-coroutines for two unrelated purposes may run into conflicts caused by
+models that require explicit switching, for example by naming the target
+coroutine, do not compose naturally. This means that a program that
+uses coroutines for two unrelated purposes may run into conflicts caused
+by unexpected interactions.
To illustrate the problem, consider the following example (simplified
-code; see the full source in
-`pypy/module/_stackless/test/test_composable_coroutine.py`_). First, a
-simple usage of coroutine::
+code using a theorical ``coroutine`` class). First, a simple usage of
main_coro = coroutine.getcurrent() # the main (outer) coroutine
data = 
@@ -530,74 +363,35 @@
main coroutine, which confuses the ``generator_iterator.next()`` method
(it gets resumed, but not as a result of a call to ``Yield()``).
-As part of trying to combine multiple different paradigms into a single
-application-level module, we have built a way to solve this problem.
-The idea is to avoid the notion of a single, global "main" coroutine (or
-a single main greenlet, or a single main tasklet). Instead, each
-conceptually separated user of one of these concurrency interfaces can
-create its own "view" on what the main coroutine/greenlet/tasklet is,
-which other coroutine/greenlet/tasklets there are, and which of these is
-the currently running one. Each "view" is orthogonal to the others. In
-particular, each view has one (and exactly one) "current"
-coroutine/greenlet/tasklet at any point in time. When the user switches
-to a coroutine/greenlet/tasklet, it implicitly means that he wants to
-switch away from the current coroutine/greenlet/tasklet *that belongs to
-the same view as the target*.
+Thus the notion of coroutine is *not composable*. By opposition, the
+primitive notion of continulets is composable: if you build two
+different interfaces on top of it, or have a program that uses twice the
+same interface in two parts, then assuming that both part independently
+work, the composition of the two parts still works.
-The precise application-level interface has not been fixed yet; so far,
-"views" in the above sense are objects of the type
-``stackless.usercostate``. The above two examples can be rewritten in
-the following way::
+A full proof of that claim would require careful definitions, but let us
+just claim that this fact is true because of the following observation:
+the API of continulets is such that, when doing a ``switch()``, it
+requires the program to have some continulet to explicitly operate on.
+It shuffles the current continuation with the continuation stored in
+that continulet, but has no effect outside. So if a part of a program
+has a continulet object, and does not expose it as a global, then the
+rest of the program cannot accidentally influence the continuation
+stored in that continulet object.
- producer_view = stackless.usercostate() # a local view
- main_coro = producer_view.getcurrent() # the main (outer) coroutine
- producer_coro = producer_view.newcoroutine()
- generators_view = stackless.usercostate()
- def generator(f):
- def wrappedfunc(*args, **kwds):
- g = generators_view.newcoroutine(generator_iterator)
-Then the composition ``grab_values()`` works as expected, because the
-two views are independent. The coroutine captured as ``self.caller`` in
-the ``generator_iterator.next()`` method is the main coroutine of the
-``generators_view``. It is no longer the same object as the main
-coroutine of the ``producer_view``, so when ``data_producer()`` issues
-the following command::
-the control flow cannot accidentally jump back to
-``generator_iterator.next()``. In other words, from the point of view
-of ``producer_view``, the function ``grab_next_value()`` always runs in
-its main coroutine ``main_coro`` and the function ``data_producer`` in
-its coroutine ``producer_coro``. This is the case independently of
-which ``generators_view``-based coroutine is the current one when
-``grab_next_value()`` is called.
-Only code that has explicit access to the ``producer_view`` or its
-coroutine objects can perform switches that are relevant for the
-generator code. If the view object and the coroutine objects that share
-this view are all properly encapsulated inside the generator logic, no
-external code can accidentally temper with the expected control flow any
-In conclusion: we will probably change the app-level interface of PyPy's
-stackless module in the future to not expose coroutines and greenlets at
-all, but only views. They are not much more difficult to use, and they
-scale automatically to larger programs.
+In other words, if we regard the continulet object as being essentially
+a modifiable ``f_back``, then it is just a link between the frame of
+``callable()`` and the parent frame --- and it cannot be arbitrarily
+changed by unrelated code, as long as they don't explicitly manipulate
+the continulet object. Typically, both the frame of ``callable()``
+(commonly a local function) and its parent frame (which is the frame
+that switched to it) belong to the same class or module; so from that
+point of view the continulet is a purely local link between two local
+frames. It doesn't make sense to have a concept that allows this link
+to be manipulated from outside.
.. _`Stackless Python`: http://www.stackless.com
.. _`documentation of the greenlets`: http://packages.python.org/greenlet/
-.. _`Stackless Transform`: translation.html#the-stackless-transform
.. include:: _ref.txt
diff --git a/pypy/doc/translation.rst b/pypy/doc/translation.rst
@@ -552,14 +552,15 @@
The stackless transform converts functions into a form that knows how
to save the execution point and active variables into a heap structure
-and resume execution at that point. This is used to implement
+and resume execution at that point. This was used to implement
coroutines as an RPython-level feature, which in turn are used to
-implement `coroutines, greenlets and tasklets`_ as an application
+implement coroutines, greenlets and tasklets as an application
level feature for the Standard Interpreter.
-Enable the stackless transformation with :config:`translation.stackless`.
+The stackless transformation has been deprecated and is no longer
+available in trunk. It has been replaced with continulets_.
-.. _`coroutines, greenlets and tasklets`: stackless.html
+.. _continulets: stackless.html
.. _`preparing the graphs for source generation`:
diff --git a/pypy/module/_continuation/__init__.py b/pypy/module/_continuation/__init__.py
@@ -12,7 +12,7 @@
To make a continulet object, call 'continulet' with a callable and
optional extra arguments. Later, the first time you switch() to the
-continulet, the callable is invoked wih the same continulet object as
+continulet, the callable is invoked with the same continulet object as
the extra first argument.
At this point, the one-shot continuation stored in the continulet points
More information about the pypy-commit