, is this a bug ?

lkcl luke.leighton at
Sat Nov 7 12:29:24 CET 2009

On Nov 7, 2:20 am, "Gabriel Genellina" <gagsl-... at> wrote:
> Yes, seems to be a bug. But given the current status of imputil, it's not
> likely to be fixed; certainly not in 2.5 which only gets security fixes
> now.

 well, that bug's not the only one.  the other one that i found, which
i have been specifically ordered not to report (that or _any_ python
bugs, of which there have been several discovered in the past eight
months), will have to wait until the python developers rescind that

if the python community is lucky, by the time that decision is made, i
will not have forgotten what those bugs are.

> (that is, self.fs_imp.import_from_dir(item, name)). Or perhaps
> item.encode(sys.getdefaultfilesystemencoding()). str(item) definitively
> won't work with directory names containing non-ascii characters.
> Why are you using imputil in the first place?

 it's an absolutely necessary and integral part of pyjamas-desktop
"platform overrides".

 it's absolutely essential to track, in exactly the same manner in
which python "normally" performs importing, and to give the platform-
specific "overrides" a chance to get in there, first.

 so, it is absolutely essential to have a correct working version of - and due to the bugs present, and the unwillingness of the
python team to fix those bugs, pyjamas-desktop is forced to maintain a
copy of

 the "platform" is set to e.g. hulahop, pywebkitgtk or mshtml,
depending on the decision made by the user or the developer to use a
particular browser engine.  the platform name is stored in
pyjd.platform in exactly the same way that the system name is stored
in sys.platform.

 the way that the platform-specific overrides works is to perform AST
translation of the module, and then to look for the exact same module
but in platform/{modulename}{platformname}.py and perform AST
translation of _that_ module as well.

 then, at the top level, any global functions in the platform-specific
AST tree *replace* those in the "main" AST.  likewise, a node-walk
along all methods in all classes of the platform-specific AST tree.

 in this way, god-awful messy code like this:
 class Widget:
     def do_something(self):
        if platform == 'mshtml':
            # do something terrible and ugly
        elif platform == 'pywebkitgtk':
            # do something only marginally better
            # do the default W3C standards-compliant thing

     def a_standard_function(self):
        # do something normal that all the browser engines get right

 can be replaced by three files, each of which encodes *only* the
logic associated with the god-awful ugliness of each browser:
 class Widget:
     def do_something(self):
        # do the default W3C standards-compliant thing

     def a_standard_function(self):
        # do something normal that all the browser engines get right

 class Widget:
     def do_something(self):
        # do something only marginally better

 class Widget:
     def do_something(self):
       # do something terrible and ugly

 a similar trick could in fact be deployed to drastically simplify the
layout of e.g. distutils, _espeeecially_ the compiler and linker
modules, by using sys.platform as the "override", or, given that
that's not entirely the whole decision-making process, as i noted when
doing the mingw32 port of python, it would be better to set something
like distutils.platform and to use that.

 however, although the technique could be used in the distutils/ case, the level of complexity of the code (the size of
each "override"), and the small number of actual modules, means that
there isn't actually that much benefit in deploying this AST-
overriding technique.

 in the pyjamas API, however, with over 75 user-interface modules, and
over 1200 functions, any one of which could be over-ridden by _eight_
separate platforms, each with their own quirks that can usually be
handled with one or two lines of code, the AST-merging idea that james
tauber came up with is an absolute god-send.


More information about the Python-list mailing list