[pypy-dev] request for sprint planning discussion
holger krekel
hpk at trillke.net
Mon Dec 1 17:24:54 CET 2003
Hello pypy and especially hello Amsterdam sprinters (apparently 13),
the Amsterdam sprint is only two weeks off and i think we need some more
discussion and overview about what we intend to do in Amsterdam. First
let me note that while Armin and me were hoping to make a first public
release in Amsterdam this is not of paramount importance. I think that
it's more important to have a sprint that everybody enjoys.
Judging from IRC discussions and feedback from some sprinters this may
imply that we take a somewhat different approach than announced before:
revisiting various parts of the current architecture, learning about and
improving/documenting them. It's probably more fun to hack together at
some documentation than doing it alone :-) Please feel encouraged
(especially sprint participants) to ask questions, comment or suggest
different/more ideas, goals and wishes for our sprint.
Let me present some ideas that are currently on the table:
- revisit/complete support for interpreter<->applevel interactions [1]
This code (interpreter/gateway.py and friends) was mostly hacked by
Armin and me and would enjoy wider involvement and improvement. For one,
interpreter-level objects like frame and function objects need to be
correctly exposed at application level (some bits and pieces missing).
The "import dis; dis.dis(dis.dis)" goal finally needs to work!
(still only "import dis; dis.dis(dis.dis.func_code)" works.
- currently you can mix interp-level code and app-level code (see e.g.
module/builtin.py). While you can access app-level objects from
interp-level the reverse is not true: there is no general way to access
interp-level objects directly from app-level -- unless the interp-level
objects provide specific hooks (e.g. pypy_getattr() in pyframe.py).
Accessing interpreter level objects from app-level would e.g. be useful
if we want to *define* the complete python type hiearchy including
__new__ methods in our standard "types.py" file. Note that e.g.
StdObjSpace doesn't really care about this: most of it would work
just fine without having such a type hierarchy. However, types.py would
then *define* the actual types and their relations completly at app-level.
The interp-level objects would of course need to correlate to this
definition. In the course, it might be nice if we could access
interp-level objects directly from app-level
class int(type):
def __new__(cls, ...):
from ... import W_IntObject # our stdobjspace-implementation
return W_IntObject(arg) # gets invoked here ...
this obviously needs more thought but i hope the idea is understandable.
- Of course, there is still a lot of work with Annotation/Translation
but this has been mentioned in previous postings. Note that this part
of the source tree is pretty independent from the rest of pypy. Recently
Armin and me have started to refactor annotation code after we also
implemented the "Berlin model" of the flowgraph-structures
(objspace/flow/model.py). The idea is to have a rather
general and somewhat efficient annotation/query engine.
The beginnings are in the new pypy/annotation directory including
a (non-complete) README.txt
- frontends: i don't know how many people have experiences hacking with
pygame and/or game architectures. It probably doesn't make much sense
if only Armin and me want to or can do it. Of course, Michael Hudson has
done some stuff in that area, too, but he had to cancel his participation.
However, we can start with writing tools that e.g. list all
space operations for a given python function. There also is
tool/methodChecker.py which tries to list the "implementedness"
of app-visible functions/methods of types. Doing tools like this
is helpful for understanding how pypy works -- both writing and using it.
These tools could easily be reused from whatever frontend with the
following approach: Let PyPy run in a most flexible, insecure but
simple 15-liner application-server that simply receives and executes
remote python code/string objects. Thus you can send the
"methodChecker/showspaceops" cmdline tools to a remote server
and receive the results (e.g. over a redirected sys.stdout/err).
(i might commit some simple code for this mechanism before the
sprint to a src/pyappserver directory and notify the list).
- completing stdobjspace and builtins, there is a lot to do still.
Rocco Morretti recently worked in the direction of (and suggested
as a goal) getting 'regrtest.py' to pass on PyPy as much as possible.
An important missing piece probably is getting a PyPy implementation
of the cpython import mechanism (our current module/builtin module
__import__ implementation is just a simplistic hack).
Basically i suggest we try to plan the sprint so that everybody
gets educated enough to feel at home with code and concepts of PyPy.
At best everyone feels able to take initiatives of their own.
It probably would make sense to prepare some introductional talks to
the interpreter (byte code dispatching/implementation/exceptions ...)
the stdobjspace (multimethods, and type implementations ...)
annotation/translation (flowgraphs/annotations via space-operations)
on the first two days. At least we can make a few question/answer
screen-sessions on these topics. At the moment, there are very
few people who know most/all of the areas of PyPy and can take
initiatives to fix/improve stuff. This needs to change and is
more important than getting a release out (although we need not
give up this idea, yet).
please comment away,
holger
[1] application level objects are the usual objects/structures you
see in/from a python program. Under the hood, these objects
have interpreter-level implementations. In interp-level source
code you see lots of w_* names indicating that they reference
a 'wrapped object' aka an application-level object. These
wrapped objects are manipulated by object space operations
like objspace.getitem/getattr/type/add and are opaque to the
interpreter. IOW, an objectspace is usually in complete
control of object layout, app-level representation and other
details. Only the interpreter frame/function/code/... objects can
control their app-level representation themselves by defining
certain hooks like 'pypy_getattr' which the objectspace
dispatches to if it encounters a getattr object on an internal
object. The nice effect is that objectspaces don't need to reimplement
e.g. function/generator/code/module types with all the app-level
representation again and again.
More information about the Pypy-dev
mailing list