|
|
Differences between CPython and JPython
CPython and JPython are two different implementations of the Python
language. While a Language
Reference exists for the Python language, there are a number of
features of the language that are incompletely specified. The
following lists all known differences between the two implementations
of the Python language. These differences range from the trivial --
JPython prints "1.0E20" where CPython prints
"1e+020" -- to the dramatic -- everything in JPython is an
instance of a class. At some point more effort should be made to
separate the interesting differences from the mundane.
Any other differences not listed here can probably be considered a
bug in JPython. Understand of course that CPython and JPython advance
at different paces. All efforts are made to keep the two
implementations in sync, but that's not always possible.
This list
has been updated to describe the differences between JPython 1.1beta4 and
CPython 1.5.2
Syntax
- JPython has a different interpretation of floating point
literals. CPython doesn't allow 001.1 CPython should be
fixed.
- JPython supports continue in a try clause. CPython
should be fixed - but don't hold your breath.
Standard types, functions and behavior
- JPython string objects support full two-byte Unicode
characters and the functions in the string module are
Unicode-aware. CPython should add Unicode support in the
future, though the details of this are still unclear.
- JPython string objects have methods, providing a more
convenient interface to most of the string module functionality.
This feature has been added to CPython and will appear in
version CPython 1.6.
- JPython formats floating point numbers differently,
e.g. an upper case 'E' is used, and it switches over to E
notation sooner than CPython does. Both behaviors are
acceptable.
- In JPython, 0.1**4 is printed as
1.0000000000000002E-4. In CPython, it is printed 0.0001, even
though it is not actually equal to 0.0001. Both behaviors
are acceptable for now -- this is still being discussed.
- JPython sequences support three argument
slices. i.e. range(3)[::-1] == [2,1,0]. CPython should be
fixed.
- Every object in JPython is an instance of a class --
there are no types in JPython. i.e. [].__class__ is a sensible
thing to write in JPython. CPython should be fixed - but
don't hold your breath.
- The .tell() method on JPython file objects returns a
Python long, while the .tell() method on CPython file objects
returns a Python int. Whether or not both behaviors are
acceptable is still unclear.
- JPython file objects are still missing some
functionality -- see todo list in PyFile.java. (Hopefully in
the near future this can be changed to read -- JPython file
objects include the following extra functionality to properly
handle non-ascii files...) JPython should be fixed.
- In CPython, range(0.1, 3.2) yields the surprising [0,
1, 2]. JPython does the right thing (reject float arguments).
-- Many other functions taking int arguments have the same
problem. CPython should be fixed.
- In CPython, the list.append() method takes multiple
arguments and forms a tuple. JPython's append() method
requires one argument only. CPython should be fixed -- but
probably won't be due to backwards compatibility.
- The __name__ attribute of built-in extension modules
(e.g. 'sys') is different. Both behaviors are
acceptable.
- In many cases, introspection yields different results.
Where appropriate and possible, JPython will adhere to
CPython's introspection behavior. Some differences are
acceptable.
- JPython defines __debug__, but always sets it equal to
1. JPython should implement CPython's -O option.
- The locals() dictionary in JPython is mutable from
within a function. After "def foo(x=1): locals()['x'] =
2; print x" foo() prints 1 in CPython and 2 in
JPython. Jim thinks that JPython's behavior is better here
-- but the best answer might be that locals() should be
considered a read-only dictionary. Proper behavior here
is still unclear.
- JPython doesn't support restricted execution mode and
doesn't have the magic __builtins__ in every namespace.
JPython will probably never support restricted execution
mode -- Java's security model is recommended instead.
- JPython uses different values for the IOError
argument. This causes trouble for people who unpack the value
into an (errno, message) tuple. Both behaviors are
acceptable.
- JPython code objects are missing other attributes --
co_code, co_consts, co_lnotab, co_names, co_nlocals,
co_stacksize. co_flags is now supported because the Python
debugger requires it. Other attributes will probably never be
supported in JPython due to its implementation of code objects
as compiled Java bytecodes.
- Accessing, setting, or deleting attributes on built-in
objects may raise AttributeError or
TypeError differently. This is considered
implementation dependent. In JPython the following rules are
used: when getting a non-existant attribute,
AttributeError is raised; when setting or deleting a
readonly attribute, TypeError is raised; when setting
or deleting a non-existant attribute, AttributeError
is raised. Be aware though currently neither JPython nor
CPython are completely consistent.
- Function objects do not have writable func_code or
func_defaults attributes. While these are writable in
CPython, I haven't decided whether they should be writable in
JPython.
- JPython has "true" garbage collection whereas
CPython uses reference counting. This means that in JPython
users don't need to worry about handling circular references
as these are guaranteed to be collected properly. On the
other hand, users of JPython have no guarantees of when an
object will be finalized -- this can cause problems for people
who use open("foo", 'r').read() excessively. Both
behaviors are acceptable -- and highly unlikely to change.
- In JPython, __del__() methods in user-defined classes
are never called. Some future version of JPython might
provide limited support for __del__(), but see the above
item.
- The dictionaries used by classes, instances, and
modules in JPython are not the same as the dictionaries
created by {}. They are StringMap's which require all of their
keys to be strings. After "class c: pass",
c.__dict__[1] = 2 will work in CPython, but will raise a
"TypeError: keys in namespace must be strings" error
in JPython. Both behaviors are acceptable -- CPython might
adopt JPython's approach in the future for the performance
gains it can provide.
Extension modules
- JPython supports all Java packages as extension
modules. i.e. from "java.lang import System" will
work in any JPython implementation. This functionality
might be added as an optional extension to some future version
of CPython.
- JPython includes the builtin module jarray -- which
allows Python programmers to create and manipulate Java array
objects.
- Lots of builtin extension modules don't exist in JPython.
- Modules struct, cPickle, cStringIO, and operator were
all added for JPython v1.1. The following are likely to
be implemented in a future version of JPython -- cmath.
- The following are under consideration (working code
would make the decision much easier ;-) -- array, select,
a dbm/gdbm/bsddb style module, Numeric.
- The following are highly unlikely any time soon --
Tkinter. However, Finn Bock has a JNI implementation called
jTkinter which
supports the full _tkinter API. Very cool stuff!
- Let me know if something should be added to this
list.
- __builtin__ module
- Incomplete implementation of __import__ -- only one
argument is allowed and replacing this with a user-defined
function has no effect. JPython might be fixed in a
future release, but CPython will probably adopt a new
import mechanism in the near future, so JPython may not
change until the new mechanism is defined.
- os module
- popen() and system() are missing. JPython should be
fixed, patches would be graciously accepted.
- os.path.normcase() exists but may not be correctly
implemented. This one is extremely
frustrating as there seems no portable way to
implement in Java.
- chmod(), chown(), getpid(), fork(), ... are
missing, stat() exists but is implemented
incompletely. These functions are all very Unix
specific and it is unlikely they will ever be properly
supported in a 100% Pure Java implementation.
Finn Bock has created
a JNI/C++ implementation for the posix module. See
Finn Bock's JPython modules page for details.
Note that his cPickle, cStringIO, binascii, and struct
modules have all been integrated into JPython 1.1.
- The socket module is limited.
- sys module
- JPython is still missing exitfunc
- Also missing executable, getrefcount,
setcheckinterval which don't make much sense for JPython.
- thread module
- CPython's thread modules defines some aliases that
JPython's doesn't. These aliases are considered
obsolete and won't be supported by JPython.
- Harry Mantakos has contributed the underlying md5
implementation, so now the md5 module works out of the box.
- The time module may produce some different values than with
CPython. This is due to Java 1.1 compatibility, and this
may be improved in later releases.
Interpreter and environment
- JPython doesn't catch interrupts. Only fixable with
a GUI console since interrupts are not supported by Java.
- JPython doesn't have command line editing. Only
fixable with a GUI console. However, Un*x users can check out
rlterm which provides generic GNU Readline support
for any terminal based interactive program. rlterm
is part of the
Yorick package.
- JPython should have a feature similar to
$PYTHONSTARTUP, which specifies a script to run at the start
of interactive mode only.
- JPython supports
different command line options than
CPython, e.g. "-jar" and "-D". It
also has a different convention for indicating non-interactive
standard input (the Java runtime needs help).
|