- Neatened the braces in PyList_New().
- Made sure "indexerr" was initialized to NULL.
- Factored if blocks in PyList_Append().
- Made sure "allocated" is initialized in list_init().
scheme in situations that likely won't benefit from it. This further
improves memory utilization from Py2.3 which always over-allocates
except for PyList_New().
Situations expected to benefit from over-allocation:
list.insert(), list.pop(), list.append(), and list.extend()
Situations deemed unlikely to benefit:
list_inplace_repeat, list_ass_slice, list_ass_subscript
The most gray area was for listextend_internal() which only runs
when the argument is a list or a tuple. This could be viewed as
a one-time fixed length addition or it could be viewed as wrapping
a series of appends. I left its over-allocation turned on but
could be convinced otherwise.
worth it to in-line the call to PyIter_Next().
Saves another 15% on most list operations that acceptable a general
iterable argument (such as the list constructor).
avoids creating an intermediate tuple for iterable arguments other than
lists or tuples.
In other words, a+=b no longer requires extra memory when b is not a
list or tuple. The list and tuple cases are unchanged.
for xrange and list objects).
* list.__reversed__ now checks the length of the sequence object before
calling PyList_GET_ITEM() because the mutable could have changed length.
* all three implementations are now tranparent with respect to length and
maintain the invariant len(it) == len(list(it)) even when the underlying
sequence mutates.
* __builtin__.reversed() now frees the underlying sequence as soon
as the iterator is exhausted.
* the code paths were rearranged so that the most common paths
do not require a jump.
* Replace sprintf message with a constant message string -- this error
message ran on every invocation except straight deletions but it was
only needed when the rhs was not iterable. The message was also
out-of-date and did not reflect that iterable arguments were allowed.
* For inner loops that do not make ref count adjustments, use memmove()
for fast copying and better readability.
* For inner loops that do make ref count adjustments, speed them up by
factoring out the constant structure reference and using vitem[] instead.
* Using addition instead of substraction on array indices allows the
compiler to use a fast addressing mode. Saves about 10%.
* Using PyTuple_GET_ITEM and PyList_SET_ITEM is about 7% faster than
PySequenceFast_GET_ITEM which has to make a list check on every pass.
utilization, and speed:
* Moved the responsibility for emptying the previous list from list_fill
to list_init.
* Replaced the code in list_extend with the superior code from list_fill.
* Eliminated list_fill.
Results:
* list.extend() no longer creates an intermediate tuple except to handle
the special case of x.extend(x). The saves memory and time.
* list.extend(x) runs
5 to 10% faster when x is a list or tuple
15% faster when x is an iterable not defining __len__
twice as fast when x is an iterable defining __len__
* the code is about 15 lines shorter and no longer duplicates
functionality.
The Py2.3 approach overallocated small lists by up to 8 elements.
The last checkin would limited this to one but slowed down (by 20 to 30%)
the creation of small lists between 3 to 8 elements.
This tune-up balances the two, limiting overallocation to 3 elements
(significantly reducing space consumption from Py2.3) and running faster
than the previous checkin.
The first part of the growth pattern (0, 4, 8, 16) neatly meshes with
allocators that trigger data movement only when crossing a power of two
boundary. Also, then even numbers mesh well with common data alignments.
realloc(). This is achieved by tracking the overallocation size in a new
field and using that information to skip calls to realloc() whenever
possible.
* Simplified and tightened the amount of overallocation. For larger lists,
this overallocates by 1/8th (compared to the previous scheme which ranged
between 1/4th to 1/32nd over-allocation). For smaller lists (n<6), the
maximum overallocation is one byte (formerly it could be upto eight bytes).
This saves memory in applications with large numbers of small lists.
* Eliminated the NRESIZE macro in favor of a new, static list_resize function
that encapsulates the resizing logic. Coverting this back to macro would
give a small (under 1%) speed-up. This was too small to warrant the loss
of readability, maintainability, and de-coupling.
* Some functions using NRESIZE had grown unnecessarily complex in their
efforts to bend to the macro's calling pattern. With the new list_resize
function in place, those other functions could be simplified. That is
being saved for a separate patch.
* The ob_item==NULL check could be eliminated from the new list_resize
function. This would entail finding each piece of code that sets ob_item
to NULL and adding a new line to invalidate the overallocation tracking
field. Rather than impose a new requirement on other pieces of list code,
it was preferred to leave the NULL check in place and retain the benefits
of decoupling, maintainability and information hiding (only PyList_New()
and list_sort() need to know about the new field). This approach also
reduces the odds of breaking an extension module.
(Collaborative effort by Raymond Hettinger, Hye-Shik Chang, Tim Peters,
and Armin Rigo.)
Formerly, length data fetched from sequence objects.
Now, any object that reports its length can benefit from pre-sizing.
On one sample timing, it gave a threefold speedup for list(s) where s
was a set object.
The special-case code that was removed could return a value indicating
success but leave an exception set. test_fileinput failed in a debug
build as a result.
which can be reviewed via
http://coding.derkeiler.com/Archive/Python/comp.lang.python/2003-12/1011.html
Duncan Booth investigated, and discovered that an "optimisation" was
in fact a pessimisation for small numbers of elements in a source list,
compared to not having the optimisation, although with large numbers
of elements in the source list the optimisation was quite beneficial.
He posted his change to comp.lang.python (but not to SF).
Further research has confirmed his assessment that the optimisation only
becomes a net win when the source list has more than 100 elements.
I also found that the optimisation could apply to tuples as well,
but the gains only arrive with source tuples larger than about 320
elements and are nowhere near as significant as the gains with lists,
(~95% gain @ 10000 elements for lists, ~20% gain @ 10000 elements for
tuples) so I haven't proceeded with this.
The code as it was applied the optimisation to list subclasses as
well, and this also appears to be a net loss for all reasonable sized
sources (~80-100% for up to 100 elements, ~20% for more than 500
elements; I tested up to 10000 elements).
Duncan also suggested special casing empty lists, which I've extended
to all empty sequences.
On the basis that list_fill() is only ever called with a list for the
result argument, testing for the source being the destination has
now happens before testing source types.
key provides C support for the decorate-sort-undecorate pattern.
reverse provide a stable sort of the list with the comparisions reversed.
* Amended the docs to guarantee sort stability.
Reverted a Py2.3b1 change to iterator in subclasses of list and tuple.
They had been changed to use __getitem__ whenever it had been overriden
in the subclass.
This caused some usabilty and performance problems. Also, it was
inconsistent with the rest of python where many container methods
access the underlying object directly without first checking for
an overridden getter. Users needing a change in iterator behavior
should override it directly.
As a side issue on this bug, it was noted that list and tuple iterators
used macros to directly access containers and would not recognize
__getitem__ overrides. If the method is overridden, the patch returns
a generic sequence iterator which calls the __getitem__ method; otherwise,
it returns a high custom iterator with direct access to container elements.
interpreted by slicing, so negative values count from the end of the
list. This was the only place where such an interpretation was not
placed on a list index.
Obtain cleaner coding and a system wide
performance boost by using the fast, pre-parsed
PyArg_Unpack function instead of PyArg_ParseTuple
function which is driven by a format string.
Armin Rigo's Draconian but effective fix for
SF bug 453523: list.sort crasher
slightly fiddled to catch more cases of list mutation. The dreaded
internal "immutable list type" is gone! OTOH, if you look at a list
*while* it's being sorted now, it will appear to be empty. Better
than a core dump.
This is friendlier for caches.
2. Cut MIN_GALLOP to 7, but added a per-sort min_gallop vrbl that adapts
the "get into galloping mode" threshold higher when galloping isn't
paying, and lower when it is. There's no known case where this hurts.
It's (of course) neutral for /sort, \sort and =sort. It also happens
to be neutral for !sort. It cuts a tiny # of compares in 3sort and +sort.
For *sort, it reduces the # of compares to better than what this used to
do when MIN_GALLOP was hardcoded to 10 (it did about 0.1% more *sort
compares before, but given how close we are to the limit, this is "a
lot"!). %sort used to do about 1.5% more compares, and ~sort about
3.6% more. Here are exact counts:
i *sort 3sort +sort %sort ~sort !sort
15 449235 33019 33016 51328 188720 65534 before
448885 33016 33007 50426 182083 65534 after
0.08% 0.01% 0.03% 1.79% 3.65% 0.00% %ch from after
16 963714 65824 65809 103409 377634 131070
962991 65821 65808 101667 364341 131070
0.08% 0.00% 0.00% 1.71% 3.65% 0.00%
17 2059092 131413 131362 209130 755476 262142
2057533 131410 131361 206193 728871 262142
0.08% 0.00% 0.00% 1.42% 3.65% 0.00%
18 4380687 262440 262460 421998 1511174 524286
4377402 262437 262459 416347 1457945 524286
0.08% 0.00% 0.00% 1.36% 3.65% 0.00%
19 9285709 524581 524634 848590 3022584 1048574
9278734 524580 524633 837947 2916107 1048574
0.08% 0.00% 0.00% 1.27% 3.65% 0.00%
20 19621118 1048960 1048942 1715806 6045418 2097150
19606028 1048958 1048941 1694896 5832445 2097150
0.08% 0.00% 0.00% 1.23% 3.65% 0.00%
3. Added some key asserts I overlooked before.
4. Updated the doc file.
directly when no comparison function is specified. This saves a layer
of function call on every compare then. Measured speedups:
i 2**i *sort \sort /sort 3sort +sort %sort ~sort =sort !sort
15 32768 12.5% 0.0% 0.0% 100.0% 0.0% 50.0% 100.0% 100.0% -50.0%
16 65536 8.7% 0.0% 0.0% 0.0% 0.0% 0.0% 12.5% 0.0% 0.0%
17 131072 8.0% 25.0% 0.0% 25.0% 0.0% 14.3% 5.9% 0.0% 0.0%
18 262144 6.3% -10.0% 12.5% 11.1% 0.0% 6.3% 5.6% 12.5% 0.0%
19 524288 5.3% 5.9% 0.0% 5.6% 0.0% 5.9% 5.4% 0.0% 2.9%
20 1048576 5.3% 2.9% 2.9% 5.1% 2.8% 1.3% 5.9% 2.9% 4.2%
The best indicators are those that take significant time (larger i), and
where sort doesn't do very few compares (so *sort and ~sort benefit most
reliably). The large numbers are due to roundoff noise combined with
platform variability; e.g., the 14.3% speedup for %sort at i=17 reflects
a printed elapsed time of 0.18 seconds falling to 0.17, but a change in
the last digit isn't really meaningful (indeed, if it really took 0.175
seconds, one electron having a lazy nanosecond could shift it to either
value <wink>). Similarly the 25% at 3sort i=17 was a meaningless change
from 0.05 to 0.04. However, almost all the "meaningless changes" were
in the same direction, which is good. The before-and-after times for
*sort are clearest:
before after
0.18 0.16
0.25 0.23
0.54 0.50
1.18 1.11
2.57 2.44
5.58 5.30
listsort. If the former calls itself recursively, they're a waste of
time, since it's called on a random permutation of a random subset of
elements. OTOH, for exactly the same reason, they're an immeasurably
small waste of time (the odds of finding exploitable order in a random
permutation are ~= 0, so the special-case loops looking for order give
up quickly). The point is more for conceptual clarity.
Also changed some "assert comments" into real asserts; when this code
was first written, Python.h didn't supply assert.h.
introduced, list.sort() was rewritten to use only the "< or not <?"
distinction. After rich comparisons were introduced, docompare() was
fiddled to translate a Py_LT Boolean result into the old "-1 for <,
0 for ==, 1 for >" flavor of outcome, and the sorting code was left
alone. This left things more obscure than they should be, and turns
out it also cost measurable cycles.
So: The old CMPERROR novelty is gone. docompare() is renamed to islt(),
and now has the same return conditinos as PyObject_RichCompareBool. The
SETK macro is renamed to ISLT, and is even weirder than before (don't
complain unless you want to maintain the sort code <wink>).
Overall, this yields a 1-2% speedup in the usual (no explicit function
passed to list.sort()) case when sorting arrays of floats (as sortperf.py
does). The boost is higher for arrays of ints.
The staticforward define was needed to support certain broken C
compilers (notably SCO ODT 3.0, perhaps early AIX as well) botched the
static keyword when it was used with a forward declaration of a static
initialized structure. Standard C allows the forward declaration with
static, and we've decided to stop catering to broken C compilers. (In
fact, we expect that the compilers are all fixed eight years later.)
I'm leaving staticforward and statichere defined in object.h as
static. This is only for backwards compatibility with C extensions
that might still use it.
XXX I haven't updated the documentation.
it_seq field when the end of the list is reached.
Also remove the next() method -- one is supplied automatically by
PyType_Ready() because the tp_iternext slot is set. That's a good
thing, because the implementation given here was buggy (it never
raised StopIteration).
explicit comparison function case: use PyObject_Call instead of
PyEval_CallObject. Same thing in context, but gives a 2.4% overall
speedup when sorting a list of ints via list.sort(__builtin__.cmp).
arg tuple. This was suggested on c.l.py but afraid I can't find the msg
again for proper attribution. For
list.sort(cmp)
where list is a list of random ints, and cmp is __builtin__.cmp, this
yields an overall 50-60% speedup on my Win2K box. Of course this is a
best case, because the overhead of calling cmp relative to the cost of
actually comparing two ints is at an extreme. Nevertheless it's huge
bang for the buck. An additionak 20-30% can be bought by making the arg
tuple an immortal static (avoiding all but "the first" PyTuple_New), but
that's tricky to make correct since docompare needs to be reentrant. So
this picks the cherry and leaves the pits for Fred <wink>.
Note that this makes no difference to the
list.sort()
case; an arg tuple gets built only if the user specifies an explicit
sort function.
[ 400998 ] experimental support for extended slicing on lists
somewhat spruced up and better tested than it was when I wrote it.
Includes docs & tests. The whatsnew section needs expanding, and arrays
should support extended slices -- later.
Tim Peters
Nicholas Bastin
Raymond Hettinger
Armin Rigo
Andrew MacIntyre
Jeremy Hylton
Hye-Shik Chang
Michael W. Hudson
Walter Dörwald
Guido van Rossum
Martin v. Löwis
Skip Montanaro
Neal Norwitz