Python file handling got an overhaul in Python 3, and it affects
the way we have to interact with files. Notably:
1) `PyFile_FromFile` no longer exists, and instead we have to use
`PyFile_FromFd`. This means having a way to get an fd from
a FILE*. For this we reuse the lldb_private::File class to
convert between FILE*s and fds, since there are some subtleties
regarding ownership rules when FILE*s and fds refer to the same
file.
2) PyFile is no longer a builtin type, so there is no such thing as
`PyFile_Check`. Instead, files in Python 3 are just instances
of `io.IOBase`. So the logic for checking if something is a file
in Python 3 is to check if it is a subclass of that module.
Additionally, some unit tests are added to verify that `PythonFile`
works as expected on Python 2 and Python 3, and
`ScriptInterpreterPython` is updated to use `PythonFile` instead of
manual calls to the various `PyFile_XXX` methods.
llvm-svn: 250444
PythonObjects were being incorrectly ref-counted. This problem was
pervasive throughout the codebase, leading to an unknown number of memory
leaks and potentially use-after-free.
The issue stems from the fact that Python native methods can either return
"borrowed" references or "owned" references. For the former category, you
*must* incref it prior to decrefing it. And for the latter category, you
should not incref it before decrefing it. This is mostly an issue when a
Python C API method returns a `PyObject` to you, but it can also happen with
a method accepts a `PyObject`. Notably, this happens in `PyList_SetItem`,
which is documented to "steal" the reference that you give it. So if you
pass something to `PyList_SetItem`, you cannot hold onto it unless you
incref it first. But since this is one of only two exceptions in the
entire API, it's confusing and difficult to remember.
Our `PythonObject` class was indiscriminantely increfing every object it
received, which means that if you passed it an owned reference, you now
have a dangling reference since owned references should not be increfed.
We were doing this in quite a few places.
There was also a fair amount of manual increfing and decrefing prevalent
throughout the codebase, which is easy to get wrong.
This patch solves the problem by making any construction of a
`PythonObject` from a `PyObject` take a flag which indicates whether it is
an owned reference or a borrowed reference. There is no way to construct a
`PythonObject` without this flag, and it does not offer a default value,
forcing the user to make an explicit decision every time.
All manual uses of `PyObject` have been cleaned up throughout the codebase
and replaced with `PythonObject` in order to make RAII the predominant
pattern when dealing with native Python objects.
Differential Revision: http://reviews.llvm.org/D13617
Reviewed By: Greg Clayton
llvm-svn: 250195
This is the work done by Ryan Brown from http://reviews.llvm.org/D8712 that makes a TypeSystem class and abstracts types to be able to use a type system.
All tests pass on MacOSX and passed on linux the last time this was submitted.
llvm-svn: 244679
Previously embedded interpreters were handled as ad-hoc source
files compiled into source/Interpreter. This made it hard to
disable a specific interpreter, or to add support for other
interpreters and allow the developer to choose which interpreter(s)
were enabled for a particular build.
This patch converts script interpreters over to a plugin-based system.
Script interpreters now live in source/Plugins/ScriptInterpreter, and
the canonical LLDB interpreter, ScriptInterpreterPython, is moved there
as well.
Any new code interfacing with the Python C API must live in this location
from here on out. Additionally, generic code should never need to
reference or make assumptions about the presence of a specific interpreter
going forward.
Differential Revision: http://reviews.llvm.org/D11431
Reviewed By: Greg Clayton
llvm-svn: 243681
