code, less memory. Tests have uncovered no drawbacks. Christian and
Vladimir are the other two people who have burned many brain cells on the
dict code in recent years, and they like the approach too, so I'm checking
it in without further ado.
instead of multiplication to generate the probe sequence. The idea is
recorded in Python-Dev for Dec 2000, but that version is prone to rare
infinite loops.
The value is in getting *all* the bits of the hash code to participate;
and, e.g., this speeds up querying every key in a dict with keys
[i << 16 for i in range(20000)] by a factor of 500. Should be equally
valuable in any bad case where the high-order hash bits were getting
ignored.
Also wrote up some of the motivations behind Python's ever-more-subtle
hash table strategy.
dictresize() was too aggressive about never ever resizing small dicts.
If a small dict is entirely full, it needs to rebuild it despite that
it won't actually resize it, in order to purge old dummy entries thus
creating at least one virgin slot (lookdict assumes at least one such
exists).
Also took the opportunity to add some high-level comments to dictresize.
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.
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).
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.
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.
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).
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.
sees it (test_iter.py is unchanged).
- Added a tp_iternext slot, which calls the iterator's next() method;
this is much faster for built-in iterators over built-in types
such as lists and dicts, speeding up pybench's ForLoop with about
25% compared to Python 2.1. (Now there's a good argument for
iterators. ;-)
- Renamed the built-in sequence iterator SeqIter, affecting the C API
functions for it. (This frees up the PyIter prefix for generic
iterator operations.)
- Added PyIter_Check(obj), which checks that obj's type has a
tp_iternext slot and that the proper feature flag is set.
- Added PyIter_Next(obj) which calls the tp_iternext slot. It has a
somewhat complex return condition due to the need for speed: when it
returns NULL, it may not have set an exception condition, meaning
the iterator is exhausted; when the exception StopIteration is set
(or a derived exception class), it means the same thing; any other
exception means some other error occurred.
new slot tp_iter in type object, plus new flag Py_TPFLAGS_HAVE_ITER
new C API PyObject_GetIter(), calls tp_iter
new builtin iter(), with two forms: iter(obj), and iter(function, sentinel)
new internal object types iterobject and calliterobject
new exception StopIteration
new opcodes for "for" loops, GET_ITER and FOR_ITER (also supported by dis.py)
new magic number for .pyc files
new special method for instances: __iter__() returns an iterator
iteration over dictionaries: "for x in dict" iterates over the keys
iteration over files: "for x in file" iterates over lines
TODO:
documentation
test suite
decide whether to use a different way to spell iter(function, sentinal)
decide whether "for key in dict" is a good idea
use iterators in map/filter/reduce, min/max, and elsewhere (in/not in?)
speed tuning (make next() a slot tp_next???)
I know some people don't like this -- if it's really controversial,
I'll take it out again. (If it's only Alex Martelli who doesn't like
it, that doesn't count as "real controversial" though. :-)
That's why this is a separate checkin from the iterators stuff I'm
about to check in next.
PyTuple_New() could *conceivably* clear the dict, so move the test for
an empty dict after the tuple allocation. It means that we waste time
allocating and deallocating a 2-tuple when the dict is empty, but who
cares. It also means that when the dict is empty *and* there's no
memory to allocate a 2-tuple, we raise MemoryError, not KeyError --
but that may actually a good idea: if there's no room for a lousy
2-tuple, what are the chances that there's room for a KeyError
instance?
and reported to python-dev: because we were calling dict_resize() in
PyDict_Next(), and because GC's dict_traverse() uses PyDict_Next(),
and because PyTuple_New() can cause GC, and because dict_items() calls
PyTuple_New(), it was possible for dict_items() to have the dict
resized right under its nose.
The solution is convoluted, and touches several places: keys(),
values(), items(), popitem(), PyDict_Next(), and PyDict_SetItem().
There are two parts to it. First, we no longer call dict_resize() in
PyDict_Next(), which seems to solve the immediate problem. But then
PyDict_SetItem() must have a different policy about when *it* calls
dict_resize(), because we want to guarantee (e.g. for an algorithm
that Jeremy uses in the compiler) that you can loop over a dict using
PyDict_Next() and make changes to the dict as long as those changes
are only value replacements for existing keys using PyDict_SetItem().
This is done by resizing *after* the insertion instead of before, and
by remembering the size before we insert the item, and if the size is
still the same, we don't bother to even check if we might need to
resize. An additional detail is that if the dict starts out empty, we
must still resize it before the insertion.
That was the first part. :-)
The second part is to make keys(), values(), items(), and popitem()
safe against side effects on the dict caused by allocations, under the
assumption that if the GC can cause arbitrary Python code to run, it
can cause other threads to run, and it's not inconceivable that our
dict could be resized -- it would be insane to write code that relies
on this, but not all code is sane.
Now, I have this nagging feeling that the loops in lookdict probably
are blissfully assuming that doing a simple key comparison does not
change the dict's size. This is not necessarily true (the keys could
be class instances after all). But that's a battle for another day.
- Use PyObject_RichCompareBool() when comparing keys; this makes the
error handling cleaner.
- There were two implementations for dictionary comparison, an old one
(#ifdef'ed out) and a new one. Got rid of the old one, which was
abandoned years ago.
- In the characterize() function, part of dictionary comparison, use
PyObject_RichCompareBool() to compare keys and values instead. But
continue to use PyObject_Compare() for comparing the final
(deciding) elements.
- Align the comments in the type struct initializer.
Note: I don't implement rich comparison for dictionaries -- there
doesn't seem to be much to be gained. (The existing comparison
already decides that shorter dicts are always smaller than longer
dicts.)
exception context. This avoids improperly propogating errors raised by
a user-defined __cmp__() by a subsequent lookup operation.
This patch does *not* include the performance enhancement patch for
dictionaries with string keys only; that will be checked in separately.
This closes SourceForge patch #101277 and bug #112558.
comments, docstrings or error messages. I fixed two minor things in
test_winreg.py ("didn't" -> "Didn't" and "Didnt" -> "Didn't").
There is a minor style issue involved: Guido seems to have preferred English
grammar (behaviour, honour) in a couple places. This patch changes that to
American, which is the more prominent style in the source. I prefer English
myself, so if English is preferred, I'd be happy to supply a patch myself ;)
implementation. This was really to test whether my new CVS+SSH
setup is more usable than the old one -- and turns out it is (for
whatever reason, it was impossible to do a commit before that
involved more than one directory).
This patch modifies the type structures of objects that
participate in GC. The object's tp_basicsize is increased when
GC is enabled. GC information is prefixed to the object to
maintain binary compatibility. GC objects also define the
tp_flag Py_TPFLAGS_GC.
For more comments, read the patches@python.org archives.
For documentation read the comments in mymalloc.h and objimpl.h.
(This is not exactly what Vladimir posted to the patches list; I've
made a few changes, and Vladimir sent me a fix in private email for a
problem that only occurs in debug mode. I'm also holding back on his
change to main.c, which seems unnecessary to me.)
Added wrapping macros to dictobject.c, listobject.c, tupleobject.c,
frameobject.c, traceback.c that safely prevends core dumps
on stack overflow. Macros and functions in object.c, object.h.
The method is an "elevator destructor" that turns cascading
deletes into tail recursive behavior when some limit is hit.
Sparc Solaris 2.6 (fully patched!) that I don't want to dig into, but
which I suspect is a bug in the multithreaded malloc library that only
shows up when run on a multiprocessor. (The program wasn't using
threads, it was just using the multithreaded C library.)
__getitem__(). This method never raises an exception; if the key is
not in the dictionary, the second (optional) argument is returned. If
the second argument is not provided and the key is missing, None is
returned.
mapp_methods: added "get" method.
complexity saved much any more. A simple benchmark (grail) showed
that there were 3 times as many misses as hits, and the same number of
times again the code was bypassed altogether due to the existence of
setattro/getattro.
The table size is now constrained to be a power of two, and we use a
variable increment based on GF(2^n)-{0} (not that I have the faintest
idea what that is :-) which helps avoid the expensive '%' operation.
Some of the entries in the table of polynomials have been modified
according to a post by Tim Peters.
defines that a shorter dictionary is always smaller than a longer one.
For dictionaries of the same size, the smallest differing element
determines the outcome (which yields the same results as before,
without explicit sorting).
* posixmodule.c: don't prototype getcwd() -- it's not portable...
* mappingobject.c: double-check validity of last_name_char in
dict{lookup,insert,remove}.
* arraymodule.c: need memmove only for non-STDC Suns.
* Makefile: comment out HTML_LIBS and XT_USE by default
* pythonmain.c: don't prototype getopt() -- it's not standardized
* socketmodule.c: cast flags arg to {get,set}sockopt() and addrbuf arg to
recvfrom() to (ANY*).
* pythonrun.c (initsigs): fix prototype, make it static
* intobject.c (LONG_BIT): only #define it if not already defined
* classobject.[ch]: remove all references to unused instance_convert()
* mappingobject.c (getmappingsize): Don't return NULL in int function.
* {tuple,list,mapping,array}object.c: call printobject with 0 for flags
* compile.c (parsestr): use quote instead of '\'' at one crucial point
* arraymodule.c (array_getattr): Added __members__ attribute
(1) dictionaries/mappings now have attributes values() and items() as
well as keys(); at the C level, use the new function mappinggetnext()
to iterate over a dictionary.
(2) "class C(): ..." is now illegal; you must write "class C: ...".
(3) Class objects now know their own name (finally!); and minor
improvements to the way how classes, functions and methods are
represented as strings.
(4) Added an "access" statement and semantics. (This is still
experimental -- as long as you don't use the keyword 'access' nothing
should be changed.)
* Stubs for faster implementation of local variables (not yet finished)
* Added function name to code object. Print it for code and function
objects. THIS MAKES THE .PYC FILE FORMAT INCOMPATIBLE (the version
number has changed accordingly)
* Print address of self for built-in methods
* New internal functions getattro and setattro (getattr/setattr with
string object arg)
* Replaced "dictobject" with more powerful "mappingobject"
* New per-type functio tp_hash to implement arbitrary object hashing,
and hashobject() to interface to it
* Added built-in functions hash(v) and hasattr(v, 'name')
* classobject: made some functions static that accidentally weren't;
added __hash__ special instance method to implement hash()
* Added proper comparison for built-in methods and functions
Tim Peters
Martin v. Löwis
Fred Drake
Guido van Rossum
Jeremy Hylton
Moshe Zadka
Thomas Wouters
Barry Warsaw
Sjoerd Mullender