As noted in PEP 236, there is no clear way for "simulated
interactive shells" to simulate the behaviour of __future__
statements in "real" interactive shells, i.e. have __future__
statements' effects last the life of the shell.
The PEP also takes the opportunity to clean up the other
unresolved issue mentioned in PEP 236, the inability to stop
compile() inheriting the effect of future statements affecting the
code calling compile().
This PEP proposes to address the first problem by adding an
optional fourth argument to the builtin function "compile", adding
information to the _Feature instances defined in __future__.py and
adding machinery to the standard library modules "codeop" and
"code" to make the construction of such shells easy.
The second problem is dealt with by simply adding *another*
optional argument to compile(), which if non-zero suppresses the
inheriting of future statements' effects.
I propose adding a fourth, optional, "flags" argument to the
builtin "compile" function. If this argument is omitted,
there will be no change in behaviour from that of Python 2.1.
If it is present it is expected to be an integer, representing
various possible compile time options as a bitfield. The
bitfields will have the same values as the CO_* flags already used
by the C part of Python interpreter to refer to future statements.
compile() shall raise a ValueError exception if it does not
recognize any of the bits set in the supplied flags.
The flags supplied will be bitwise-"or"ed with the flags that
would be set anyway, unless the new fifth optional argument is a
non-zero intger, in which case the flags supplied will be exactly
the set used.
The above-mentioned flags are not currently exposed to Python. I
propose adding .compiler_flag attributes to the _Feature objects
in __future__.py that contain the necessary bits, so one might
write code such as:
return compile(func_def, "<input>", "suite",
A recent change means that these same bits can be used to tell if
a code object was compiled with a given feature; for instance
codeob.co_flags & __future__.generators.compiler_flag
will be non-zero if and only if the code object "codeob" was
compiled in an environment where generators were allowed.
I will also add a .all_feature_flags attribute to the __future__
module, giving a low-effort way of enumerating all the __future__
options supported by the running interpreter.
I also propose adding a pair of classes to the standard library
One - Compile - will sport a __call__ method which will act much
like the builtin "compile" of 2.1 with the difference that after
it has compiled a __future__ statement, it "remembers" it and
compiles all subsequent code with the __future__ option in effect.
It will do this by using the new features of the __future__ module
Objects of the other class added to codeop - CommandCompiler -
will do the job of the existing codeop.compile_command function,
but in a __future__-aware way.
Finally, I propose to modify the class InteractiveInterpreter in
the standard library module code to use a CommandCompiler to
emulate still more closely the behaviour of the default Python
Should be very few or none; the changes to compile will make no
difference to existing code, nor will adding new functions or
classes to codeop. Existing code using
code.InteractiveInterpreter may change in behaviour, but only for
the better in that the "real" Python shell will be being better
The fiddling that needs to be done to Lib/__future__.py when
adding a __future_ feature will be a touch more complicated.
Everything else should just work.
I hope the above interface is not too disruptive to implement for
A series of preliminary implementations are at:
After light massaging by Tim Peters, they have now been checked in.
This document has been placed in the public domain.