The idea is Marc-Andre Lemburg's, the implementation is Tim's.
Add a new ma_smalltable member to dictobjects, an embedded vector of
MINSIZE (8) dictentry structs. Short course is that this lets us avoid
additional malloc(s) for dicts with no more than 5 entries.
The changes are widespread but mostly small.
Long course: WRT speed, all scalar operations (getitem, setitem, delitem)
on non-empty dicts benefit from no longer needing NULL-pointer checks
(ma_table is never NULL anymore). Bulk operations (copy, update, resize,
clearing slots during dealloc) benefit in some cases from now looping
on the ma_fill count rather than on ma_size, but that was an unexpected
benefit: the original reason to loop on ma_fill was to let bulk
operations on empty dicts end quickly (since the NULL-pointer checks
went away, empty dicts aren't special-cased any more).
Special considerations:
For dicts that remain empty, this change is a lose on two counts:
the dict object contains 8 new dictentry slots now that weren't
needed before, and dict object creation also spends time memset'ing
these doomed-to-be-unsused slots to NULLs.
For dicts with one or two entries that never get larger than 2, it's
a mix: a malloc()/free() pair is no longer needed, and the 2-entry case
gets to use 8 slots (instead of 4) thus decreasing the chance of
collision. Against that, dict object creation spends time memset'ing
4 slots that aren't strictly needed in this case.
For dicts with 3 through 5 entries that never get larger than 5, it's a
pure win: the dict is created with all the space they need, and they
never need to resize. Before they suffered two malloc()/free() calls,
plus 1 dict resize, to get enough space. In addition, the 8-slot
table they ended with consumed more memory overall, because of the
hidden overhead due to the additional malloc.
For dicts with 6 or more entries, the ma_smalltable member is wasted
space, but then these are large(r) dicts so 8 slots more or less doesn't
make much difference. They still benefit all the time from removing
ubiquitous dynamic null-pointer checks, and get a small benefit (but
relatively smaller the larger the dict) from not having to do two
mallocs, two frees, and a resize on the way *to* getting their sixth
entry.
All in all it appears a small but definite general win, with larger
benefits in specific cases. It's especially nice that it allowed to
get rid of several branches, gotos and labels, and overall made the
code smaller.
This should be faster.
This means:
(1) "for line in file:" won't work if the xreadlines module can't be
imported.
(2) The body of "for line in file:" shouldn't use the file directly;
the effects (e.g. of file.readline(), file.seek() or even
file.tell()) would be undefined because of the buffering that goes
on in the xreadlines module.
UTF-16 codec will now interpret and remove a *leading* BOM mark. Sub-
sequent BOM characters are no longer interpreted and removed.
UTF-16-LE and -BE pass through all BOM mark characters.
These changes should get the UTF-16 codec more in line with what
the Unicode FAQ recommends w/r to BOM marks.
Two exceedingly unlikely errors in dictresize():
1. The loop for finding the new size had an off-by-one error at the
end (could over-index the polys[] vector).
2. The polys[] vector ended with a 0, apparently intended as a sentinel
value but never used as such; i.e., it was never checked, so 0 could
have been used *as* a polynomial.
Neither bug could trigger unless a dict grew to 2**30 slots; since that
would consume at least 12GB of memory just to hold the dict pointers,
I'm betting it's not the cause of the bug Fred's tracking down <wink>.
in the comments for using two passes was bogus, as the only object that
can get decref'ed due to the copy is the dummy key, and decref'ing dummy
can't have side effects (for one thing, dummy is immortal! for another,
it's a string object, not a potentially dangerous user-defined object).
1. Omit the early-out EQ/NE "lengths different?" test. Was unable to find
any real code where it triggered, but it always costs. The same is not
true of list richcmps, where different-size lists appeared to get
compared about half the time.
2. Because tuples are immutable, there's no need to refetch the lengths of
both tuples from memory again on each loop trip.
BUG ALERT: The tuple (and list) richcmp algorithm is arguably wrong,
because it won't believe there's any difference unless Py_EQ returns false
for some corresponding elements:
>>> class C:
... def __lt__(x, y): return 1
... __eq__ = __lt__
...
>>> C() < C()
1
>>> (C(),) < (C(),)
0
>>>
That doesn't make sense -- provided you believe the defn. of C makes sense.
and introduces a new method .decode().
The major change is that strg.encode() will no longer try to convert
Unicode returns from the codec into a string, but instead pass along
the Unicode object as-is. The same is now true for all other codec
return types. The underlying C APIs were changed accordingly.
Note that even though this does have the potential of breaking
existing code, the chances are low since conversion from Unicode
previously took place using the default encoding which is normally
set to ASCII rendering this auto-conversion mechanism useless for
most Unicode encodings.
The good news is that you can now use .encode() and .decode() with
much greater ease and that the door was opened for better accessibility
of the builtin codecs.
As demonstration of the new feature, the patch includes a few new
codecs which allow string to string encoding and decoding (rot13,
hex, zip, uu, base64).
Written by Marc-Andre Lemburg. Copyright assigned to the PSF.
to reason that me_key is much more likely to match the key we're looking
for than to match dummy, and if the key is absent me_key is much more
likely to be NULL than dummy: most dicts don't even have a dummy entry.
Running instrumented dict code over the test suite and some apps confirmed
that matching dummy was 200-300x less frequent than matching key in
practice. So this reorders the tests to try the common case first.
It can lose if a large dict with many collisions is mostly deleted, not
resized, and then frequently searched, but that's hardly a case we
should be favoring.
The comment following used to say:
/* We use ~hash instead of hash, as degenerate hash functions, such
as for ints <sigh>, can have lots of leading zeros. It's not
really a performance risk, but better safe than sorry.
12-Dec-00 tim: so ~hash produces lots of leading ones instead --
what's the gain? */
That is, there was never a good reason for doing it. And to the contrary,
as explained on Python-Dev last December, it tended to make the *sum*
(i + incr) & mask (which is the first table index examined in case of
collison) the same "too often" across distinct hashes.
Changing to the simpler "i = hash & mask" reduced the number of string-dict
collisions (== # number of times we go around the lookup for-loop) from about
6 million to 5 million during a full run of the test suite (these are
approximate because the test suite does some random stuff from run to run).
The number of collisions in non-string dicts also decreased, but not as
dramatically.
Note that this may, for a given dict, change the order (wrt previous
releases) of entries exposed by .keys(), .values() and .items(). A number
of std tests suffered bogus failures as a result. For dicts keyed by
small ints, or (less so) by characters, the order is much more likely to be
in increasing order of key now; e.g.,
>>> d = {}
>>> for i in range(10):
... d[i] = i
...
>>> d
{0: 0, 1: 1, 2: 2, 3: 3, 4: 4, 5: 5, 6: 6, 7: 7, 8: 8, 9: 9}
>>>
Unfortunately. people may latch on to that in small examples and draw a
bogus conclusion.
test_support.py
Moved test_extcall's sortdict() into test_support, made it stronger,
and imported sortdict into other std tests that needed it.
test_unicode.py
Excluced cp875 from the "roundtrip over range(128)" test, because
cp875 doesn't have a well-defined inverse for unicode("?", "cp875").
See Python-Dev for excruciating details.
Cookie.py
Chaged various output functions to sort dicts before building
strings from them.
test_extcall
Fiddled the expected-result file. This remains sensitive to native
dict ordering, because, e.g., if there are multiple errors in a
keyword-arg dict (and test_extcall sets up many cases like that), the
specific error Python complains about first depends on native dict
ordering.
Allow module getattr and setattr to exploit string interning, via the
previously null module object tp_getattro and tp_setattro slots. Yields
a very nice speedup for things like random.random and os.path etc.
For rich comparisons, use instance_getattr2() when possible to avoid
the expense of setting an AttributeError. Also intern the name_op[]
table and use the interned strings rather than creating a new string
and interning it each time through.
doesn't know how to do LE, LT, GE, GT. dict_richcompare can't do the
latter any faster than dict_compare can. More importantly, for
cmp(dict1, dict2), Python *first* tries rich compares with EQ, LT, and
GT one at a time, even if the tp_compare slot is defined, and
dict_richcompare called dict_compare for the latter two because
it couldn't do them itself. The result was a lot of wasted calls to
dict_compare. Now dict_richcompare gives up at once the times Python
calls it with LT and GT from try_rich_to_3way_compare(), and dict_compare
is called only once (when Python gets around to trying the tp_compare
slot).
Continued mystery: despite that this cut the number of calls to
dict_compare approximately in half in test_mutants.py, the latter still
runs amazingly slowly. Running under the debugger doesn't show excessive
activity in the dict comparison code anymore, so I'm guessing the culprit
is somewhere else -- but where? Perhaps in the element (key/value)
comparison code? We clearly spend a lot of time figuring out how to
compare things.
Fixed a half dozen ways in which general dict comparison could crash
Python (even cause Win98SE to reboot) in the presence of kay and/or
value comparison routines that mutate the dict during dict comparison.
Bugfix candidate.
interned when created, so the cached versions generally aren't ever
interned. With the patch, the
Py_INCREF(t);
*p = t;
Py_DECREF(s);
return;
indirection block in PyString_InternInPlace() is never executed during a
full run of the test suite, but was executed very many times before. So
I'm trading more work when creating one-character strings for doing less
work later. Note that the "more work" here can happen at most 256 times
per program run, so it's trivial. The same reasoning accounts for the
patch's simplification of string_item (the new version can call
PyString_FromStringAndSize() no more than 256 times per run, so there's
no point to inlining that stuff -- if we were serious about saving time
here, we'd pre-initialize the characters vector so that no runtime testing
at all was needed!).
Store floats and doubles to full precision in marshal.
Test that floats read from .pyc/.pyo closely match those read from .py.
Declare PyFloat_AsString() in floatobject header file.
Add new PyFloat_AsReprString() API function.
Document the functions declared in floatobject.h.
d1 == d2 and d1 != d2 now work even if the keys and values in d1 and d2
don't support comparisons other than ==, and testing dicts for equality
is faster now (especially when inequality obtains).
safely together and don't duplicate logic (the common logic was factored
out into new private API function _PySequence_IterContains()).
Visible change:
some_complex_number in some_instance
no longer blows up if some_instance has __getitem__ but neither
__contains__ nor __iter__. test_iter changed to ensure that remains true.
NEEDS DOC CHANGES
A few more AttributeErrors turned into TypeErrors, but in test_contains
this time.
The full story for instance objects is pretty much unexplainable, because
instance_contains() tries its own flavor of iteration-based containment
testing first, and PySequence_Contains doesn't get a chance at it unless
instance_contains() blows up. A consequence is that
some_complex_number in some_instance
dies with a TypeError unless some_instance.__class__ defines __iter__ but
does not define __getitem__.
to string.join(), so that when the latter figures out in midstream that
it really needs unicode.join() instead, unicode.join() can actually get
all the sequence elements (i.e., there's no guarantee that the sequence
passed to string.join() can be iterated over *again* by unicode.join(),
so string.join() must not pass on the original sequence object anymore).
NEEDS DOC CHANGES.
This one surprised me! While I expected tuple() to be a no-brainer, turns
out it's actually dripping with consequences:
1. It will *allow* the popular PySequence_Fast() to work with any iterable
object (code for that not yet checked in, but should be trivial).
2. It caused two std tests to fail. This because some places used
PyTuple_Sequence() (the C spelling of tuple()) as an indirect way to test
whether something *is* a sequence. But tuple() code only looked for the
existence of sq->item to determine that, and e.g. an instance passed
that test whether or not it supported the other operations tuple()
needed (e.g., __len__). So some things the tests *expected* to fail
with an AttributeError now fail with a TypeError instead. This looks
like an improvement to me; e.g., test_coercion used to produce 559
TypeErrors and 2 AttributeErrors, and now they're all TypeErrors. The
error details are more informative too, because the places calling this
were *looking* for TypeErrors in order to replace the generic tuple()
"not a sequence" msg with their own more specific text, and
AttributeErrors snuck by that.
the code necessary to accomplish this is simpler and faster if confined to
the object implementations, so we only do this there.
This causes no behaviorial changes beyond a (very slight) speedup.
need to be specified in the type structures independently. The flag
exists only for binary compatibility.
This is a "source cleanliness" issue and introduces no behavioral changes.
dictionary size was comparing ma_size, the hash table size, which is
always a power of two, rather than ma_used, wich changes on each
insertion or deletion. Fixed this.
to no longer insist that len(seq) be defined.
NEEDS DOC CHANGES.
This is meant to be a model for how other functions of this ilk (max,
filter, etc) can be generalized similarly. Feel encouraged to grab your
favorite and convert it!
Note some cute consequences:
list(file) == file.readlines() == list(file.xreadlines())
list(dict) == dict.keys()
list(dict.iteritems()) = dict.items()
list(xrange(i, j, k)) == range(i, j, k)
object's type didn't define tp_print, there were still cases where the
full "print uses str() which falls back to repr()" semantics weren't
honored. This resulted in
>>> print None
<None object at 0x80bd674>
>>> print type(u'')
<type object at 0x80c0a80>
Fixed this by always using the appropriate PyObject_Repr() or
PyObject_Str() call, rather than trying to emulate what they would do.
Also simplified PyObject_Str() to always fall back on PyObject_Repr()
when tp_str is not defined (rather than making an extra check for
instances with a __str__ method). And got rid of the special case for
strings.
Fred Drake
Tim Peters
Guido van Rossum
Marc-André Lemburg
Jeremy Hylton