String literals without explicit prefix in __init__.py and policy.py
are treated as unicode object on Python3, because these modules are
loaded before setup of our specific code transformation (the later
module is imported at the beginning of __init__.py).
BTW, "modulepolicy" in __init__.py is initialized by "policy.policy".
This causes issues below;
- checking "policy" value in other modules causes unintentional result
For example, "b'py' not in (u'c', u'py')" returns True
unintentionally on Python3.
- writing "policy" out fails at conversion from unicode to bytes
db1ebf457295 fixed this issue for default code path, but "policy"
can be overridden by HGMODULEPOLICY environment variable (it should
be rare case for developer using Python3, though).
This patch does:
- add "b" prefix to all string literals, which are related to module
policy, in modules above.
- check existence of HGMODULEPOLICY, and overwrite "policy" only if
it exists
For simplicity, this patch omits checking "supports_bytes_environ",
switching os.environ/os.environb, and so on (Yuya agreed this in
personal talking)
open() requires mode argument as unicodes on Python 3. This patch introduces
pycompat.open() which is inserted to files using transformer and replaces
builtins.open() calls.
This fixes open(filename, 'r'), open(filename, 'w'), etc. calls. In Python
3, that second argument *must* be a string, you can't use bytes.
The fix is the same as used with getattr() (where the second argument must
also always be a string); in the tokenizer, where we detect calls, if there
is something that looks like a call to open (and is not an attribute, so
the previous token is not a "." dot) then make sure that that second
argument is not converted to a `bytes` object instead.
There is some remaining issue where the current transformer will also rewrite
open(f('foo')).
However this also affect function for which we perform similar rewrite
('getattr', 'setattr', 'hasattr', 'safehasattr') and will be dealt with in a
follow up.
There is a case and more can be present where these functions have
multiple arguments. Our transformer used to handle the first argument, so
added a loop to handle more arguments if present.
This should be less invasive than mucking builtins.
Since tokenize.untokenize() looks start/end positions of tokens, we calculates
them from the NEWLINE token of the future import.
The most painful part of ensuring Python code runs on both Python 2
and 3 is string encoding. Making this difficult is that string
literals in Python 2 are bytes and string literals in Python 3 are
unicode. So, to ensure consistent types are used, you have to
use "from __future__ import unicode_literals" and/or prefix literals
with their type (e.g. b'foo' or u'foo').
Nearly every string in Mercurial is bytes. So, to use the same source
code on both Python 2 and 3 would require prefixing nearly every
string literal with "b" to make it a byte literal. This is ugly and
not something mpm is willing to do at this point in time.
This patch implements a custom module loader on Python 3 that performs
source transformation to convert string literals (unicode in Python 3)
to byte literals. In effect, it changes Python 3's string literals to
behave like Python 2's.
In addition, the module loader recognizes well-known built-in
functions (getattr, setattr, hasattr) and methods (encode and decode)
that barf when bytes are used and prevents these from being rewritten.
This prevents excessive source changes to accommodate this change
(we would have to rewrite every occurrence of these functions passing
string literals otherwise).
The module loader is only used on Python packages belonging to
Mercurial.
The loader works by tokenizing the loaded source and replacing
"string" tokens if necessary. The modified token stream is
untokenized back to source and loaded like normal. This does add some
overhead. However, this all occurs before caching: .pyc files will
cache the transformed version. This means the transformation penalty
is only paid on first load.
As the extensive inline comments explain, the presence of a custom
source transformer invalidates assumptions made by Python's built-in
bytecode caching mechanism. So, we have to wrap bytecode loading and
writing and add an additional header to bytecode files to facilitate
additional cache validation when the source transformations
change in the future.
There are still a few things this code doesn't handle well, namely
support for zip files as module sources and for extensions. Since
Mercurial doesn't officially support Python 3 yet, I'm inclined to
leave these as to-do items: getting a basic module loading mechanism
in place to unblock further Python 3 porting effort is more important
than comprehensive module importing support.
check-py3-compat.py has been updated to ignore frames. This is
necessary because CPython has built-in code to strip frames from the
built-in importer. When our custom code is present, this doesn't work
and the frames get all messed up. The new code is not perfect. It
works for now. But once you start chasing import failures you find
some edge cases where the files aren't being printed properly. This
only burdens people doing future Python 3 porting work so I'm inclined
to punt on the issue: the most important thing is for the source
transforming module loader to land.
There was a bit of churn in test-check-py3-compat.t because we now
trip up on str/unicode/bytes failures as a result of source
transformation. This is unfortunate but what are you going to do.
It's worth noting that other approaches were investigated.
We considered using a custom file encoding whose decode() would
apply source transformations. This was rejected because it would
require each source file to declare its custom Mercurial encoding.
Furthermore, when changing the source transformation we'd need to
version bump the encoding name otherwise the module caching layer
wouldn't know the .pyc file was invalidated. This would mean mass
updating every file when the source transformation changes. Yuck.
We also considered transforming at the AST layer. However, Python's
ast module is quite gnarly and doing AST transforms is quite
complicated, even for trivial rewrites. There are whole Python packages
that exist to make AST transformations usable. AST transforms would
still require import machinery, so the choice was basically to
perform source-level, token-level, or ast-level transforms.
Token-level rewriting delivers the metadata we need to rewrite
intelligently while being relatively easy to understand. So it won.
General consensus seems to be that this approach is the best available
to avoid bulk rewriting of '' to b''. However, we aren't confident
that this approach will never be a future maintenance burden. This
approach does unblock serious Python 3 porting efforts. So we can
re-evaulate once more work is done to support Python 3.
The C extensions don't yet work with Python 3. Let's minimize the
work required to get Mercurial running on Python 3 by always using
the pure Python module policy on Python 3.
Instead of rewriting __init__ to define the modulepolicy,
write out a __modulepolicy__.py file like __version__.py
This should work for both system-wide installation and in-place build. Therefore
we can avoid relying on two separate modulepolicy rules, '@MODULELOADPOLICY@'
and 'mercurial/modulepolicy'.
The previous refactor to module importing broke module loading when
mercurial.* modules were loaded from a zipfile (using a zipimporter).
This scenario is likely encountered when using py2exe.
Supporting zipimporter and the traditional importer side-by-side
turns out to be quite a pain. In Python 2.x, the standard, file-based
import mechanism is partially implemented in C. The sys.meta_path
and sys.path_hooks hook points exist to allow custom importers in
Python/userland. zipimport.zipimporter and our "hgimporter" class
from earlier in this patch series are 2 of these.
In a standard Python installation (no matter if running in py2exe
or similar or not), zipimport.zipimporter appears to be registered
in sys.path_hooks. This means that as each sys.path entry is
consulted, it will ask zipimporter if it supports that path and
zipimporter will be used if that entry is a zip file. In a
py2exe environment, sys.path contains an entry with the path to
the zip file containing the Python standard library along with
Mercurial's Python files.
The way the importer mechanism works is the first importer that
declares knowledge of a module (via find_module() returning an
object) gets to load it. Since our "hgimporter" is registered
in sys.meta_path and returns an interest in specific mercurial.*
modules, the zipimporter registered on sys.path_hooks never comes
into play for these modules. So, we need to be zipimporter aware
and call into zipimporter to load modules.
This patch teaches "hgimporter" how to call out into zipimporter
when necessary. We detect the necessity of zipimporter by looking
at the loader for the "mercurial" module. If it is a zipimporter
instance, we load via zipimporter.
The behavior of zipimporter is a bit wonky.
You appear to need separate zipimporter instances for each directory
in the zip file. I'm not sure why this is. I suspect it has
something to do with the low-level importing mechanism (implemented
in C) operating on a per-directory basis. PEP-302 makes some
references to this. I was not able to get a zipimporter to
import modules outside of its immediate directory no matter how
I specified the module name. This is why we use separate
zipimporter instances for the ".zip/mercurial" and
".zip/mercurial/pure" locations.
The zipimporter documentation for Python 2.7 explicitly states that
zipimporter does not import dynamic modules (C extensions). Yet from
a py2exe distribution on Windows - where the .pyd files are *not*
in the zip archive - zipimporter imported these dynamic modules
just fine! I'm not sure if dynamic modules can't be imported from
*inside* the zip archive or whether zipimporter looks for dynamic
modules outside the zip archive. All I know is zipimporter does
manage to import the .pyd files on Windows and this patch makes
our new importer compatible with py2exe.
In the ideal world, We'd probably reimplement or fall back to parts
of the built-in import mechanism instead of handling zipimporter
specially. After all, if we're loading Mercurial modules via
something that isn't the built-in file-based importer or zipimporter,
our custom importer will likely fail because it doesn't know how to
call into it. I'd like to think that we'll never encounter this
in the wild, but you never know. If we do encounter it, we can
come up with another solution.
It's worth nothing that Python 3 has moved a lot of the importing
code from C to Python. Python 3 gives you near total control over
the import mechanism. So in the very distant future when Mercurial
drops Python 2 support, it's likely that our custom importer code
can be refactored to something a bit saner.
PyPy isn't compatible with Python C extensions.
With this patch, the module load policy is automatically to "Python
only" when run under PyPy. `hg` and other Python scripts importing
mercurial.* modules will run from the source checkout or any
installation when executed with PyPy. This should enable people to
more easily experiment with PyPy and its potentially significant
performance benefits over CPython!
With this change in place, we should have slightly stronger guarantees
about how modules with both Python and C implementations are loaded.
Before, our module loader's default policy looked under both mercurial/*
and mercurial/pure/* and imported whatever it found, C or pure. The fact
it looked in both locations by default was a temporary regression from
the beginning of this series.
This patch does 2 things:
1) Changes the default module load policy to only load C modules
2) Verifies that files loaded from mercurial/* are actually C modules
This 2nd behavior change makes our new module loading mechanism
stricter than from before this series. Before, it was possible to load
a .py-based module from mercurial/*. This could happen if an old
installation orphaned a file and then somehow didn't install the C
version for the new install. We now detect this odd configuration
and fall back to loading the pure Python module, assuming it is
allowed. In the case of a busted installation, we fail fast. While
we could fall back, we explicitly decide not to do this because
we don't want people accidentally not running the C modules and having
slow performance as a result.
There are a handful of modules that have both pure Python and C
extension implementations. Currently, setup.py copies files from
mercurial/pure/*.py to mercurial/ during the install process if C
extensions are not available. This way, "import mercurial.X" will
work whether C extensions are available or not.
This approach has a few drawbacks. First, there aren't run-time checks
verifying the C extensions are loaded when they should be. This could
lead to accidental use of the slower pure Python modules. Second, the
C extensions aren't compatible with PyPy and running Mercurial with
PyPy requires installing Mercurial - you can't run ./hg from a source
checkout. This makes developing while running PyPy somewhat difficult.
This patch implements a PEP-302 import hook for finding and loading the
modules with both C and Python implementations. When a module with dual
implementations is requested for import, its import is handled by our
import hook.
The importer has a mechanism that controls what types of modules we
allow to load. We call this loading behavior the "module load policy."
There are 3 settings:
* Only load C extensions
* Only load pure Python
* Try to load C and fall back to Python
An environment variable allows overriding this policy at run time. This
is mainly useful for developers and for performing actions against the
source checkout (such as installing), which require overriding the
default (strict) policy about requiring C extensions.
The default mode for now is to allow both. This isn't proper and is
technically backwards incompatible. However, it is necessary to
implement a sane patch series that doesn't break the world during
future bisections. The behavior will be corrected in future patch.
We choose the main mercurial/__init__.py module for this code out of
necessity: in a future world, if the custom module importer isn't
registered, we'll fail to find/import certain modules when running
from a pure installation. Without the magical import-time side-effects,
*any* importer of mercurial.* modules would be required to call a
function to register our importer. I'm not a fan of import time side
effects and I initially attempted to do this. However, I was foiled by
our own test harness, which has numerous `python` invoked scripts that
"import mercurial" and fail because the importer isn't registered.
Realizing this problem is probably present in random Python scripts
that have been written over the years, I decided that sacrificing
purity for backwards compatibility is necessary. Plus, if you are
programming Python, "import" should probably "just work."
It's worth noting that now that we have a custom module loader, it
would be possible to hook up demand module proxies at this level
instead of replacing __import__. We leave this work for another time,
if it's even desired.
This patch breaks importing in environments where Mercurial modules
are loaded from a zip file (such as py2exe distributions). This will
be addressed in a subsequent patch.
