Follow a suggestion in an /*XXX*/ comment [in com_add()] to speed up
compilation by using supplemental dictionaries to keep track of names
and constants, eliminating quadratic behavior. With this patch in
place, the time to import a 5000-line file with lots of constants [at
the global level] is reduced from 20 seconds to under 3 on my system.
Fred Drake [Fri, 28 Apr 2000 14:42:37 +0000 (14:42 +0000)]
Brian Hooper <brian_takashi@hotmail.com>:
Here's a patch which changes modsupport to add 'u' and 'u#',
to support building Unicode objects from a null-terminated
Py_UNICODE *, and a Py_UNICODE * with length, respectively.
[Conversion from 'U' to 'u' by Fred, based on python-dev comments.]
Note that the use of None for NULL values of the Py_UNICODE* value is
still in; I'm not sure of the conclusion on that issue.
Mark Hammond: For Windows debug builds, we now only offer to dump
remaining object references if the environment variable PYTHONDUMPREFS
exists. The default behaviour caused problems for background or
otherwise invisible processes that use the debug build of Python.
Fixed a memory leak found by Fredrik Lundh. Instead of
PyUnicode_AsUTF8String() we now use _PyUnicode_AsUTF8String() which
returns the string object without incremented refcount (and assures
that the so obtained object remains alive until the Unicode object is
garbage collected).
Added the "--root" option as a sort of meta-install-base; if supplied,
it is forcibly prepended onto all installation directories, even if
they are already absolute.
Added 'dump_dirs()' to clean up the debug output a bit.
Added 'change_root()' to forcibly slap a new root directory onto a pathname,
even if it's already absolute. Currently only implemented for Unix; I'm
not entirely sure of the right thing to do for DOS/Windows, and have no
clue what to do for Mac OS.
This patch is a workaround for Macintosh, where the GUSI I/O library
(time, stat, etc) use the MacOS epoch of 1-Jan-1904 and the MSL C
library (ctime, localtime, etc) uses the (apparently ANSI standard)
epoch of 1-Jan-1900. Python programs see the MacOS epoch and we
convert values when needed.
Jeremy Hylton [Wed, 26 Apr 2000 20:39:20 +0000 (20:39 +0000)]
potentially useless optimization
The previous checkin (2.84) added a PyErr_Format call that made the
cost of raising an AttributeError much more expensive. In general
this doesn't matter, except that checks for __init__ and
__del__ methods, where exceptions are caught and cleared in C, also
got much more expensive.
The fix is to split instance_getattr1 into two calls:
instance_getattr2 checks the instance and the class for the attribute
and returns it or returns NULL on error. It does not raise an
exception.
instance_getattr1 does rexec checks, then calls instance_getattr2. It
raises an exception if instance_getattr2 returns NULL.
PyInstance_New and instance_dealloc now call instance_getattr2
directly.
a) check for zero length args (does this to execve, too), raising
ValueError.
b) raises more rational exceptions for various flavours of duff arguments.
I *hate*
TypeError: "illegal argument type for built-in operation"
It has to be one of the most frustrating error messages ever.
Michael Hudson fixes a case where execv() is called (for a test) with
an empty argument list -- another patch he's checking in will make
this illegal (the first argument should always be the program name).
Hacked things up a bit so that configuration variables are expanded
in command-line options, and in two phases at that: first, we expand
'install_base' and 'install_platbase', and then the other 'install_*'
options. This lets us do tricky stuff like
install --prefix='/tmp$sys_prefix'
...oooh, neat.
Simplified 'select_scheme()' -- it's no longer responsible for expanding
config vars, tildes, etc.
Define installation-specific config vars in 'self.config_vars', rather than
in a local dictionary of one method. Also factored '_expand_attrs()' out
of 'expand_dirs()' and added 'expand_basedirs()'.
Added a bunch of debugging output so I (and others) can judge the
success of this crazy scheme through direct feedback.
Patch inspired by Just van Rossum: on the Mac, in savefilename(), make
the path to save a relative path by prefixing it with os.sep (':').
Also fix an indent inconsistency in the same function.
Mark Hammond: Added dependency of winsound project on python16
project. [However I didn't add the other changes in his patch, which
were just taking away the source code control stuff -- this doesn't
hurt and would come back as soon as I make another change. --GvR]
Harry Henry Gebel:
Adds bztar format to generate .tar.bz2 tarballs
Uses the -f argument to overright old tarballs automatically, I am
assuming that if the old tarball was wanted it would have been moved or
else the version number would have been changed.
Uses the -9 argument to bzip2 and gzip to use maximum
compression. Compress uses the maximum compression by default.
Tests for correct value for the 'compress' argument of make_tarball. This
is one less place for someone adding new compression programs to forget to
change.
Christian Tismer -- total rewrite on trashcan code.
Improvements:
- does no longer need any extra memory
- has no relationship to tstate
- works in debug mode
- can easily be modified for free threading (hi Greg:)
Side effects:
Trashcan does change the order of object destruction.
Prevending that would be quite an immense effort, as
my attempts have shown. This version works always
the same, with debug mode or not. The slightly
changed destruction order should therefore be no problem.
Algorithm:
While the old idea of delaying the destruction of some
obejcts at a certain recursion level was kept, we now
no longer aloocate an object to hold these objects.
The delayed objects are instead chained together
via their ob_type field. The type is encoded via
ob_refcnt. When it comes to the destruction of the
chain of waiting objects, the topmost object is popped
off the chain and revived with type and refcount 1,
then it gets a normal Py_DECREF.
I am confident that this solution is near optimum
for minimizing side effects and code bloat.
Jack Jansen: The GUSI 2.0 I/O library used on the Mac uses the
socklen_t (unsigned int) for most size parameters. Apparently this is
part of the UNIX 98 standard.
[GvR: the changes to configure.in etc. that I just checked in make
sure that socklen_t is defined everywhere, so I deleted the little
part of Jack's mod to define socklen_t if not in GUSI2. I suppose I
will have to add it to the Windows config.h in a minute.]
Added a provision to stop all threads before exiting from the test:
the change to regrtest.py to unload all newly imported modules did
something bad to the threads -- and I realized that they would never
stop!
This uses the same precautions when trying to find a temporary
directory as when the actual tempfile is created (using O_CREAT and
O_EXCL). On non-posix platforms, nothing is changed.
Jack Jansen [Sat, 22 Apr 2000 21:51:47 +0000 (21:51 +0000)]
The .exp file hadn't been regenerated after adding the threading stuff. This means that building a nonthreaded PythonCore will now require massaging of the .exp.
Catch DistutilsOptionError in 'setup()' -- it's thrown either because of
errors in the setup script or on the command line, so shouldn't result
in a traceback.
Extracted the "what-do-I-do-for-this-format" logic from code in
'make_archive()' to a global static dictionary, ARCHIVE_FORMATS.
Added 'check_archive_formats()', which obviously makes good use of
this dictionary.
Patch by Vladimir Marangozov to unload additionally imported modules
after each test has been run. This avoids excessive memory growth
during the tests.
* Base address for all extension modules updated. PC\dllbase_nt.txt
also updated. Erroneous "libpath" directory removed for all
projects.
* winsound module moved from a builtin module to an extension
module. This was done primarily to avoid Python16.dll needing to
pull in winmm.dll. Really dumb test added for winsound - but if
nothing else it ensures the module imports.
* Temp directory for all projects are now specific to the project
(rather than common as before). This avoids any conflicts with
debug symbols or common file names etc.
NOTE: You should manually delete your existing build directory after
applying this patch, as the MSVC "clean" command will now only clean
the new temporary directories - not the existing common temp
directory.
* Base address for all extension modules updated. PC\dllbase_nt.txt
also updated. Erroneous "libpath" directory removed for all
projects.
* winsound module moved from a builtin module to an extension
module. This was done primarily to avoid Python16.dll needing to
pull in winmm.dll. Really dumb test added for winsound - but if
nothing else it ensures the module imports.
"""
Running "test_extcall" repeatedly results in memory leaks.
One of these can't be fixed (at least not easily!), it happens since
this code:
def saboteur(**kw):
kw['x'] = locals()
d = {}
saboteur(a=1, **d)
creates a circular reference - d['x']['d']==d
The others are due to some missing decrefs in ceval.c, fixed by the
patch attached below.
Note: I originally wrote this without the "goto", just adding the
missing decref's where needed. But I think the goto is justified in
keeping the executable code size of ceval as small as possible.
"""
[I think the circular reference is more like kw['x']['kw'] == kw. --GvR]
Patch by Charles G Waldman to avoid a sneaky memory leak in
_PyTuple_Resize(). In addition, a change suggested by Jeremy Hylton
to limit the size of the free lists is also merged into this patch.
Charles wrote initially:
"""
Test Case: run the following code:
class Nothing:
def __len__(self):
return 5
def __getitem__(self, i):
if i < 3:
return i
else:
raise IndexError, i
The analysis begins with the call to PySequence_Tuple at line 1641 in
ceval.c - the argument to g is seen to be a sequence but not a tuple,
so it needs to be converted from an abstract sequence to a concrete
tuple. PySequence_Tuple starts off by creating a new tuple of length
5 (line 1122 in abstract.c). Then at line 1149, since only 3 elements
were assigned, _PyTuple_Resize is called to make the 5-tuple into a
3-tuple. When we're all done the 3-tuple is decrefed, but rather than
being freed it is placed on the free_tuples cache.
The basic problem is that the 3-tuples are being added to the cache
but never picked up again, since _PyTuple_Resize doesn't make use of
the free_tuples cache. If you are resizing a 5-tuple to a 3-tuple and
there is already a 3-tuple in free_tuples[3], instead of using this
tuple, _PyTuple_Resize will realloc the 5-tuple to a 3-tuple. It
would more efficient to use the existing 3-tuple and cache the
5-tuple.
By making _PyTuple_Resize aware of the free_tuples (just as
PyTuple_New), we not only save a few calls to realloc, but also
prevent this misbehavior whereby tuples are being added to the
free_tuples list but never properly "recycled".
"""
And later:
"""
This patch replaces my submission of Sun, 16 Apr and addresses Jeremy
Hylton's suggestions that we also limit the size of the free tuple
list. I chose 2000 as the maximum number of tuples of any particular
size to save.
There was also a problem with the previous version of this patch
causing a core dump if Python was built with Py_TRACE_REFS. This is
fixed in the below version of the patch, which uses tupledealloc
instead of _Py_Dealloc.
"""
"""
In the course of debugging this I also saw that cPickle is
inconsistent with pickle - if you attempt a pickle.load or pickle.dump
on a closed file, you get a ValueError, whereas the corresponding
cPickle operations give an IOError. Since cPickle is advertised as
being compatible with pickle, I changed these exceptions to match.
"""
import test.test_cpickle
for x in xrange(1000000):
reload(test.test_cpickle)
Watch Python's memory use go up up and away!
In the course of debugging this I also saw that cPickle is
inconsistent with pickle - if you attempt a pickle.load or pickle.dump
on a closed file, you get a ValueError, whereas the corresponding
cPickle operations give an IOError. Since cPickle is advertised as
being compatible with pickle, I changed these exceptions to match.
"""
Use an explicit macro SOCKETCLOSE which expands to closesocket (on
Windows), soclose (on OS2), or to close (everywhere else).
Hopefully this fixes a new compilation error that I suddenly get on
Windows because the macro definition for close -> closesocket
apparently was done before including io.h, which contains a prototype
for close. (No idea why this wasn't an error before.)
Patch by Brian Hooper, somewhat augmented by GvR, to strip a trailing
backslash from the pathname argument to stat() on Windows -- while on
Unix, stat("/bin/") succeeds and does the same thing as stat("/bin"),
on Windows, stat("\\windows\\") fails while stat("\\windows") succeeds.
This modified version of the patch recognizes both / and \.
(This is odd behavior of the MS C library, since
os.listdir("\\windows\\") succeeds!)
Fix 'check_metadata()' so it grovels through the distribution's metadata
object, rather than through the distribution itself (since I moved the meta-
data out to a DistributionMetadata instance).
Patch, originally from Bastian Kleineidam and savagely mutilated by me,
to add the "display metadata" options: --name, --version, --author,
and so forth. Main changes:
* added 'display_options' class attribute to list all the "display only"
options (--help-commands plus the metadata options)
* added DistributionMetadata class as a place to put the actual
metadata information from the setup script (not to be confused with
the metadata display options); the logic dealing with metadata
(eg. return self.name or "UNKNOWN") is now in this class
* changed 'parse_command_line()' to use the new OO interface provided
by fancy_getopt, mainly so we can get at the original order of
options on the command line, so we can print multiple lines of
distribution meta-data in the order specified by the user
* added 'handle_display_options()' to handle display-only options
Also fixed some crufty old comments/docstrings.
Made 'generate_help()' and 'print_help()' methods of FancyGetopt.
Added 'set_option_table()' method.
Added missing 'self' to 'get_option_order()'.
Cosmetic/comment/docstring tweaks.
Continuing the refactoring: deleted the old 'fancy_getopt()' function,
leaving in its place a tiny wrapper around the FancyGetopt class
for backwards compatibility.
Hefty refactoring: converted 'fancy_getopt()' function into FancyGetopt
class. (Mainly this was to support the ability to go back after the
getopt operation is done and get extra information about the parse,
in particular the original order of options seen on the command line.
But it's a big improvement and should make it a lot easier to add
functionality in the future.)
projects / python / log