* Speed-up "x in y" where x has more than one character.
The existing code made excessive calls to the expensive memcmp() function.
The new code uses memchr() to rapidly find a start point for memcmp().
In addition to knowing that the first character is a match, the new code
also checks that the last character is a match. This significantly reduces
the incidence of false starts (saving memcmp() calls and making quadratic
behavior less likely).
Improves the timings on:
python -m timeit -r7 -s"x='a'*1000" "'ab' in x"
python -m timeit -r7 -s"x='a'*1000" "'bc' in x"
Once this code has proven itself, then string_find_internal() should refer
to it rather than running its own version. Also, something similar may
apply to unicode objects.
[ 1124295 ] Function's __name__ no longer accessible in restricted mode
which I introduced with a bit of mindless copy-paste when making
__name__ writable. You can't assign to __name__ in restricted mode,
which I'm going to pretend was intentional :)
* Added missing error checks.
* Fixed O(n**2) growth pattern. Modeled after lists to achieve linear
amortized resizing. Improves construction of "tuple(it)" when "it" is
large and does not have a __len__ method. Other cases are unaffected.
In cyclic gc, clear weakrefs to unreachable objects before allowing any
Python code (weakref callbacks or __del__ methods) to run.
This is a critical bugfix, affecting all versions of Python since weakrefs
were introduced. I'll backport to 2.3.
exposed in header files. Fixed a few comments in these headers.
As we might have expected, writing down invariants systematically exposed a
(minor) bug. In this case, function objects have a writeable func_code
attribute, which could be set to code objects with the wrong number of
free variables. Calling the resulting function segfaulted the interpreter.
Added a corresponding test.
Also, add a testcase.
Formerly, the list_extend() code used several local variables to remember
its state across iterations. Since an iteration could call arbitrary
Python code, it was possible for the list state to be changed. The new
code uses dynamic structure references instead of C locals. So, they
are always up-to-date.
After list_resize() is called, its size has been updated but the new
cells are filled with NULLs. These needed to be filled before arbitrary
iteration code was called; otherwise, that code could attempt to modify
a list that was in a semi-invalid state. The solution was to change
the ob->size field back to a value reflecting the actual number of valid
cells.
When an integer is compared to a float now, the int isn't coerced to float.
This avoids spurious overflow exceptions and insane results. This should
compute correct results, without raising spurious exceptions, in all cases
now -- although I expect that what happens when an int/long is compared to
a NaN is still a platform accident.
Note that we had potential problems here even with "short" ints, on boxes
where sizeof(long)==8. There's #ifdef'ed code here to handle that, but
I can't test it as intended. I tested it by changing the #ifdef to
trigger on my 32-bit box instead.
I suppose this is a bugfix candidate, but I won't backport it. It's
long-winded (for speed) and messy (because the problem is messy). Note
that this also depends on a previous 2.4 patch that introduced
_Py_SwappedOp[] as an extern.
Make PySequence_Check() and PyMapping_Check() handle NULL inputs. This
goes beyond what most of the other checks do, but it is nice defensive
programming and solves the OP's problem.
module type with silly arguments. (The exact name can be quibbled
over, if you care).
This was partially inspired by bug #1014215 and so on, but is also
just a good idea.
The list resizing scheme only downsized when more than 16 elements were
removed in a single step: del a[100:120]. As a result, the list would
never shrink when popping elements off one at a time.
This patch makes it shrink whenever more than half of the space is unused.
Also, at Tim's suggestion, renamed _new_size to new_allocated. This makes
the code easier to understand.
decoding incomplete input (when the input stream is temporarily exhausted).
codecs.StreamReader now implements buffering, which enables proper
readline support for the UTF-16 decoders. codecs.StreamReader.read()
has a new argument chars which specifies the number of characters to
return. codecs.StreamReader.readline() and codecs.StreamReader.readlines()
have a new argument keepends. Trailing "\n"s will be stripped from the lines
if keepends is false. Added C APIs PyUnicode_DecodeUTF8Stateful and
PyUnicode_DecodeUTF16Stateful.
This checkin is adapted from part 2 (of 3) of Trevor Perrin's patch set.
BACKWARD INCOMPATIBILITY: SHIFT must now be divisible by 5. AFAIK,
nobody will care. long_pow() could be complicated to worm around that,
if necessary.
long_pow():
- BUGFIX: This leaked the base and power when the power was negative
(and so the computation delegated to float pow).
- Instead of doing right-to-left exponentiation, do left-to-right. This
is more efficient for small bases, which is the common case.
- In addition, if the exponent is large (more than FIVEARY_CUTOFF
digits), precompute [a**i % c for i in range(32)], and go left to
right 5 bits at a time.
l_divmod():
- The signature changed so that callers who don't want the quotient,
or don't want the remainder, can pass NULL in the slot they don't
want. This saves them from having to declare a vrbl for unwanted
stuff, and remembering to decref it.
long_mod(), long_div(), long_classic_div():
- Adjust to new l_divmod() signature, and simplified as a result.
This checkin is adapted from part 1 (of 3) of Trevor Perrin's patch set.
x_mul()
- sped a little by optimizing the C
- sped a lot (~2X) if it's doing a square; note that long_pow() squares
often
k_mul()
- more cache-friendly now if it's doing a square
KARATSUBA_CUTOFF
- boosted; gradeschool mult is quicker now, and it may have been too low
for many platforms anyway
KARATSUBA_SQUARE_CUTOFF
- new
- since x_mul is a lot faster at squaring now, the point at which
Karatsuba pays for squaring is much higher than for general mult
need to convert str objects from the iterable to unicode. So, if
someone set the system default encoding to something nasty enough,
the conversion process could mutate the input iterable as a side
effect, and PySequence_Fast doesn't hide that from us if the input was
a list. IOW, can't assume the size of PySequence_Fast's result is
invariant across PyUnicode_FromObject() calls.
much to reduce the size of the code, but greatly improves its clarity.
It's also quicker in what's probably the most common case (the argument
iterable is a list). Against it, if the iterable isn't a list or a tuple,
a temp tuple is materialized containing the entire input sequence, and
that's a bigger temp memory burden. Yawn.
1. u1.join([u2]) is u2
2. Be more careful about C-level int overflow.
Since PySequence_Fast() isn't needed to achieve #1, it's not used -- but
the code could sure be simpler if it were.
happen in 2.3, but nobody noticed it still was getting generated (the
warning was disabled by default). OverflowWarning and
PyExc_OverflowWarning should be removed for 2.5, and left notes all over
saying so.
(Patch contributed by Nick Coghlan.)
Now joining string subtypes will always return a string.
Formerly, if there were only one item, it was returned unchanged.
Added XXX comment about why the undocumented PyRange_New() API function
is too broken to be worth the considerable pain of repairing.
Changed range_new() to stop using PyRange_New(). This fixes a variety
of bogus errors. Nothing in the core uses PyRange_New() now.
Documented that xrange() is intended to be simple and fast, and that
CPython restricts its arguments, and length of its result sequence, to
native C longs.
Added some tests that failed before the patch, and repaired a test that
relied on a bogus OverflowError getting raised.
hack: it would resize *interned* strings in-place! This occurred because
their reference counts do not have their expected value -- stringobject.c
hacks them. Mea culpa.
interning were not clear here -- a subclass could be mutable, for
example -- and had bugs. Explicitly interning a subclass of string
via intern() will raise a TypeError. Internal operations that attempt
to intern a string subclass will have no effect.
Added a few tests to test_builtin that includes the old buggy code and
verifies that calls like PyObject_SetAttr() don't fail. Perhaps these
tests should have gone in test_string.
unicodedata.east_asian_width(). You can still implement your own
simple width() function using it like this:
def width(u):
w = 0
for c in unicodedata.normalize('NFC', u):
cwidth = unicodedata.east_asian_width(c)
if cwidth in ('W', 'F'): w += 2
else: w += 1
return w
the case of __del__ resurrecting an object.
This makes the apparent reference leaks in test_descr go away (which I
expected) and also kills off those in test_gc (which is more surprising
but less so once you actually think about it a bit).
comma expression in listpop() that was being returned. Still essentially
unused (as it is meant to be), but now the compiler thinks it is worth
*something* by having it incremented.
of no more than 8 elements cannot fail.
listpop(): Take advantage of that its calls to list_resize() and
list_ass_slice() can't fail. This is assert'ed in a debug build now, but
in an icky way. That is, you can't say:
assert(some_call() >= 0);
because then some_call() won't occur at all in a release build. So it
has to be a big pile of #ifdefs on Py_DEBUG (yuck), or the pleasant:
status = some_call();
assert(status >= 0);
But in that case, compilers may whine in a release build, because status
appears unused then. I'm not certain the ugly trick I used here will
convince all compilers to shut up about status (status is always "used" now,
as the first (ignored) clause in a comma expression).
impossible to remember, so renamed one to something obvious. Headed
off potential signed-vs-unsigned compiler complaints I introduced by
changing the type of a vrbl to unsigned. Removed the need for the
tedious explanation about "backward pointer loops" by looping on an
int instead.
result.
list_resize(): Document the intent. Code is increasingly relying on
subtle aspects of its behavior, and they deserve to be spelled out.
list_ass_slice(): A bit more simplification, by giving it a common
error exit and initializing more values.
Be clearer in comments about what "size" means (# of elements? # of
bytes?).
While the number of elements in a list slice must fit in an int, there's
no guarantee that the number of bytes occupied by the slice will. That
malloc() and memmove() take size_t arguments is a hint about that <wink>.
So changed to use size_t where appropriate.
ihigh - ilow should always be >= 0, but we never asserted that. We do
now.
The loop decref'ing the recycled slice had a subtle insecurity: C doesn't
guarantee that a pointer one slot *before* an array will compare "less
than" to a pointer within the array (it does guarantee that a pointer
one beyond the end of the array compares as expected). This was actually
an issue in KSR's C implementation, so isn't purely theoretical. Python
probably has other "go backwards" loops with a similar glitch.
list_clear() is OK (it marches an integer backwards, not a pointer).
though I tried to be very careful. This is a slight simplification, and it
adds a new feature: a small stack-allocated "recycled" array for the cases
when we don't remove too many items.
It allows PyList_SetSlice() to never fail if:
* you are sure that the object is a list; and
* you either do not remove more than 8 items, or clear the list.
This makes a number of other places in the source code correct again -- there
are some places that delete a single item without checking for MemoryErrors
raised by PyList_SetSlice(), or that clear the whole list, and sometimes the
context doesn't allow an error to be propagated.
invariants allows the ob_item != NULL check to be replaced with an
assertion.
* Added assertions to list_init() which document and verify that the
tp_new slot establishes the invariants. This may preclude a future
bug if a custom tp_new slot is written.
to NULL during the lifetime of the object.
* listobject.c nevertheless did not conform to the other invariants,
either; fixed.
* listobject.c now uses list_clear() as the obvious internal way to clear
a list, instead of abusing list_ass_slice() for that. It makes it easier
to enforce the invariant about ob_item == NULL.
* listsort() sets allocated to -1 during sort; any mutation will set it
to a value >= 0, so it is a safe way to detect mutation. A negative
value for allocated does not cause a problem elsewhere currently.
test_sort.py has a new test for this fix.
* listsort() leak: if items were added to the list during the sort, AND if
these items had a __del__ that puts still more stuff into the list,
then this more stuff (and the PyObject** array to hold them) were
overridden at the end of listsort() and never released.
mutation during list.sort() used to rely on that listobject.c always
NULL'ed ob_item when ob_size fell to 0. That's no longer true, so the
test for list mutation during a sort is no longer reliable. Changed the
test to rely instead on that listobject.c now never NULLs-out ob_item
after (if ever) ob_item gets a non-NULL value. This new assumption is
also documented now, as a required invariant in listobject.h.
The new assumption allowed some real simplification to some of the
hairier code in listsort(), so is a Good Thing on that count.
__oct__, and __hex__. Raise TypeError if an invalid type is
returned. Note that PyNumber_Int and PyNumber_Long can still
return ints or longs. Fixes SF bug #966618.
- weakref.ref and weakref.ReferenceType will become aliases for each
other
- weakref.ref will be a modern, new-style class with proper __new__
and __init__ methods
- weakref.WeakValueDictionary will have a lighter memory footprint,
using a new weakref.ref subclass to associate the key with the
value, allowing us to have only a single object of overhead for each
dictionary entry (currently, there are 3 objects of overhead per
entry: a weakref to the value, a weakref to the dictionary, and a
function object used as a weakref callback; the weakref to the
dictionary could be avoided without this change)
- a new macro, PyWeakref_CheckRefExact(), will be added
- PyWeakref_CheckRef() will check for subclasses of weakref.ref
This closes SF patch #983019.
The builtin eval() function now accepts any mapping for the locals argument.
Time sensitive steps guarded by PyDict_CheckExact() to keep from slowing
down the normal case. My timings so no measurable impact.
tests which nicely highly highlight weaknesses).
* Initial value is now a large prime.
* Pre-multiply by the set length to add one more basis of differentiation.
* Work a bit harder inside the loop to scatter bits from sources that
may have closely spaced hash values.
All of this is necessary to make up for keep the hash function commutative.
Fortunately, the hash value is cached so the call to frozenset_hash() will
only occur once per set.
* Non-zero initial value so that hash(frozenset()) != hash(0).
* Final permutation to differentiate nested sets.
* Add logic to make sure that -1 is not a possible hash value.
iswide() for east asian width manipulation. (Inspired by David
Goodger, Reviewed by Martin v. Loewis)
- Move _PyUnicode_TypeRecord.flags to the end of the struct so that
no padding is added for UCS-4 builds. (Suggested by Martin v. Loewis)
- 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().
close() calls would attempt to free() the buffer already free()ed on
the first close(). [bug introduced with patch #788249]
Making sure that the buffer is free()ed in file object deallocation is
a belt-n-braces bit of insurance against a memory leak.
the newly created tuples, but tuples added in the freelist are now cleared in
tupledealloc already (which is very cheap, because we are already
Py_XDECREF'ing all elements anyway).
Python should have a standard Py_ZAP macro like ZAP in pystate.c.
This gives another 30% speedup for operations such as
map(func, d.iteritems()) or list(d.iteritems()) which can both take
advantage of length information when provided.
* Split into three separate types that share everything except the
code for iternext. Saves run time decision making and allows
each iternext function to be specialized.
* Inlined PyDict_Next(). In addition to saving a function call, this
allows a redundant test to be eliminated and further specialization
of the code for the unique needs of each iterator type.
* Created a reusable result tuple for iteritems(). Saves the malloc
time for tuples when the previous result was not kept by client code
(this is the typical use case for iteritems). If the client code
does keep the reference, then a new tuple is created.
Results in a 20% to 30% speedup depending on the size and sparsity
of the dictionary.
* Factored constant structure references out of the inner loops for
PyDict_Next(), dict_keys(), dict_values(), and dict_items().
Gave measurable speedups to each (the improvement varies depending
on the sparseness of the dictionary being measured).
* Added a freelist scheme styled after that for tuples. Saves around
80% of the calls to malloc and free. About 10% of the time, the
previous dictionary was completely empty; in those cases, the
dictionary initialization with memset() can be skipped.
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.
(Championed by Bob Ippolito.)
The update() method for mappings now accepts all the same argument forms
as the dict() constructor. This includes item lists and/or keyword
arguments.
recent gcc on Linux/x86)
[ 899109 ] 1==float('nan')
by implementing rich comparisons for floats.
Seems to make comparisons involving NaNs somewhat less surprising
when the underlying C compiler actually implements C99 semantics.
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.)
the same object to be collected by the cyclic GC support if they are
only referenced by a cycle. If the weakref being collected was one of
the weakrefs without callbacks, some local variables for the
constructor became invalid and have to be re-computed.
The test caused a segfault under a debug build without the fix applied.
Raymond Hettinger
Martin v. Löwis
Michael W. Hudson
Armin Rigo
Marc-André Lemburg
Peter Astrand
Tim Peters
Brett Cannon
Neil Schemenauer
Skip Montanaro
Thomas Heller
Walter Dörwald
Andrew M. Kuchling
Neal Norwitz
Jeremy Hylton
Hye-Shik Chang
Fred Drake
Nicholas Bastin
Anthony Baxter
Andrew MacIntyre
Phillip J. Eby
Guido van Rossum
Jim Fulton