Containers for storing coercion data
This module provides TripleDict and MonoDict. These are structures similar to WeakKeyDictionary in Python’s weakref module, and are optimized for lookup speed. The keys for TripleDict consist of triples (k1,k2,k3) and are looked up by identity rather than equality. The keys are stored by weakrefs if possible. If any one of the components k1, k2, k3 gets garbage collected, then the entry is removed from the TripleDict.
Key components that do not allow for weakrefs are stored via a normal refcounted reference. That means that any entry stored using a triple (k1,k2,k3) so that none of the k1,k2,k3 allows a weak reference behaves as an entry in a normal dictionary: Its existence in TripleDict prevents it from being garbage collected.
That container currently is used to store coercion and conversion maps between two parents (trac ticket #715) and to store homsets of pairs of objects of a category (trac ticket #11521). In both cases, it is essential that the parent structures remain garbage collectable, it is essential that the data access is faster than with a usual WeakKeyDictionary, and we enforce the “unique parent condition” in Sage (parent structures should be identical if they are equal).
MonoDict behaves similarly, but it takes a single item as a key. It is used for caching the parents which allow a coercion map into a fixed other parent (trac ticket #12313).
By trac ticket #14159, MonoDict and TripleDict can be optionally used with weak references on the values.
Bases: object
This is a hashtable specifically designed for (read) speed in the coercion model.
It differs from a python WeakKeyDictionary in the following important ways:
- Comparison is done using the ‘is’ rather than ‘==’ operator.
- Only weak references to the keys are stored if at all possible. Keys that do not allow for weak references are stored with a normal refcounted reference.
- The callback of the weak references is safe against recursion, see below.
There are special cdef set/get methods for faster access. It is bare-bones in the sense that not all dictionary methods are implemented.
IMPLEMENTATION:
It is implemented as a hash table with open addressing, similar to python’s dict.
If ki supports weak references then ri is a weak reference to ki with a callback to remove the entry from the dictionary if ki gets garbage collected. If ki is does not support weak references then ri is identical to ki. In the latter case the presence of the key in the dictionary prevents it from being garbage collected.
INPUT:
- size – unused parameter, present for backward compatibility.
- data – optional iterable defining initial data.
- threshold – unused parameter, present for backward compatibility.
- weak_values – optional bool (default False). If it is true, weak references to the values in this dictionary will be used, when possible.
EXAMPLES:
sage: from sage.structure.coerce_dict import MonoDict sage: L = MonoDict() sage: a = 'a'; b = 'ab'; c = -15 sage: L[a] = 1 sage: L[b] = 2 sage: L[c] = 3The key is expected to be a unique object. Hence, the item stored for c can not be obtained by providing another equal number:
sage: L[a] 1 sage: L[b] 2 sage: L[c] 3 sage: L[-15] Traceback (most recent call last): ... KeyError: -15Not all features of Python dictionaries are available, but iteration over the dictionary items is possible:
sage: # for some reason the following failed in "make ptest" sage: # on some installations, see #12313 for details sage: sorted(L.iteritems()) # random layout [(-15, 3), ('a', 1), ('ab', 2)] sage: # the following seems to be more consistent sage: set(L.iteritems()) set([('a', 1), ('ab', 2), (-15, 3)]) sage: del L[c] sage: sorted(L.iteritems()) [('a', 1), ('ab', 2)] sage: len(L) 2 sage: for i in range(1000): ... L[i] = i sage: len(L) 1002 sage: L['a'] 1 sage: L['c'] Traceback (most recent call last): ... KeyError: 'c'Note that this kind of dictionary is also used for caching actions and coerce maps. In previous versions of Sage, the cache was by strong references and resulted in a memory leak in the following example. However, this leak was fixed by trac ticket #715, using weak references:
sage: K = GF(1<<55,'t') sage: for i in range(50): ... a = K.random_element() ... E = EllipticCurve(j=a) ... P = E.random_point() ... Q = 2*P sage: import gc sage: n = gc.collect() sage: from sage.schemes.elliptic_curves.ell_finite_field import EllipticCurve_finite_field sage: LE = [x for x in gc.get_objects() if isinstance(x, EllipticCurve_finite_field)] sage: len(LE) # indirect doctest 1TESTS:
Here, we demonstrate the use of weak values.
sage: M = MonoDict(13) sage: MW = MonoDict(13, weak_values=True) sage: class Foo: pass sage: a = Foo() sage: b = Foo() sage: k = 1 sage: M[k] = a sage: MW[k] = b sage: M[k] is a True sage: MW[k] is b True sage: k in M True sage: k in MW TrueWhile M uses a strong reference to a, MW uses a weak reference to b, and after deleting b, the corresponding item of MW will be removed during the next garbage collection:
sage: import gc sage: del a,b sage: _ = gc.collect() sage: k in M True sage: k in MW False sage: len(MW) 0 sage: len(M) 1
Note that MW also accepts values that do not allow for weak references:
sage: MW[k] = int(5)
sage: MW[k]
5
The following demonstrates that :class:`MonoDict` is safer than
:class:`~weakref.WeakKeyDictionary` against recursions created by nested
callbacks; compare :trac:`15069` (the mechanism used now is different, though)::
sage: M = MonoDict(11)
sage: class A: pass
sage: a = A()
sage: prev = a
sage: for i in range(1000):
....: newA = A()
....: M[prev] = newA
....: prev = newA
sage: len(M)
1000
sage: del a
sage: len(M)
0
The corresponding example with a Python :class:`weakref.WeakKeyDictionary`
would result in a too deep recursion during deletion of the dictionary
items::
sage: import weakref
sage: M = weakref.WeakKeyDictionary()
sage: a = A()
sage: prev = a
sage: for i in range(1000):
....: newA = A()
....: M[prev] = newA
....: prev = newA
sage: len(M)
1000
sage: del a
Exception RuntimeError: 'maximum recursion depth exceeded while calling a Python object' in <function remove at ...> ignored
sage: len(M)>0
True
Check that also in the presence of circular references, :class:`MonoDict`
gets properly collected::
sage: import gc
sage: def count_type(T):
....: return len([c for c in gc.get_objects() if isinstance(c,T)])
sage: _=gc.collect()
sage: N=count_type(MonoDict)
sage: for i in range(100):
....: V = [ MonoDict(11,{"id":j+100*i}) for j in range(100)]
....: n= len(V)
....: for i in range(n): V[i][V[(i+1)%n]]=(i+1)%n
....: del V
....: _=gc.collect()
....: assert count_type(MonoDict) == N
sage: count_type(MonoDict) == N
True
AUTHORS:
- Simon King (2012-01)
- Nils Bruin (2012-08)
- Simon King (2013-02)
- Nils Bruin (2013-11)
EXAMPLES:
sage: from sage.structure.coerce_dict import MonoDict
sage: L = MonoDict(31)
sage: L[1] = None
sage: L[2] = True
sage: list(sorted(L.iteritems()))
[(1, None), (2, True)]
Bases: object
Erase items from a MonoDict when a weak reference becomes invalid.
This is of internal use only. Instances of this class will be passed as a callback function when creating a weak reference.
EXAMPLES:
sage: from sage.structure.coerce_dict import MonoDict
sage: class A: pass
sage: a = A()
sage: M = MonoDict()
sage: M[a] = 1
sage: len(M)
1
sage: del a
sage: import gc
sage: n = gc.collect()
sage: len(M) # indirect doctest
0
AUTHOR:
Bases: object
This is a hashtable specifically designed for (read) speed in the coercion model.
It differs from a python dict in the following important ways:
- All keys must be sequence of exactly three elements. All sequence types (tuple, list, etc.) map to the same item.
- Comparison is done using the ‘is’ rather than ‘==’ operator.
There are special cdef set/get methods for faster access. It is bare-bones in the sense that not all dictionary methods are implemented.
It is implemented as a list of lists (hereafter called buckets). The bucket is chosen according to a very simple hash based on the object pointer, and each bucket is of the form [id(k1), id(k2), id(k3), r1, r2, r3, value, id(k1), id(k2), id(k3), r1, r2, r3, value, ...], on which a linear search is performed. If a key component ki supports weak references then ri is a weak reference to ki; otherwise ri is identical to ki.
INPUT:
If any of the key components k1,k2,k3 (this can happen for a key component that supports weak references) gets garbage collected then the entire entry disappears. In that sense this structure behaves like a nested WeakKeyDictionary.
EXAMPLES:
sage: from sage.structure.coerce_dict import TripleDict
sage: L = TripleDict()
sage: a = 'a'; b = 'b'; c = 'c'
sage: L[a,b,c] = 1
sage: L[a,b,c]
1
sage: L[c,b,a] = -1
sage: list(L.iteritems()) # random order of output.
[(('c', 'b', 'a'), -1), (('a', 'b', 'c'), 1)]
sage: del L[a,b,c]
sage: list(L.iteritems())
[(('c', 'b', 'a'), -1)]
sage: len(L)
1
sage: for i in range(1000):
... L[i,i,i] = i
sage: len(L)
1001
sage: L = TripleDict(L)
sage: L[c,b,a]
-1
sage: L[a,b,c]
Traceback (most recent call last):
...
KeyError: ('a', 'b', 'c')
sage: L[a]
Traceback (most recent call last):
...
KeyError: 'a'
sage: L[a] = 1
Traceback (most recent call last):
...
KeyError: 'a'
Note that this kind of dictionary is also used for caching actions and coerce maps. In previous versions of Sage, the cache was by strong references and resulted in a memory leak in the following example. However, this leak was fixed by trac ticket #715, using weak references:
sage: K = GF(1<<55,'t')
sage: for i in range(50):
... a = K.random_element()
... E = EllipticCurve(j=a)
... P = E.random_point()
... Q = 2*P
sage: import gc
sage: n = gc.collect()
sage: from sage.schemes.elliptic_curves.ell_finite_field import EllipticCurve_finite_field
sage: LE = [x for x in gc.get_objects() if isinstance(x, EllipticCurve_finite_field)]
sage: len(LE) # indirect doctest
1
TESTS:
Here, we demonstrate the use of weak values.
sage: class Foo: pass
sage: T = TripleDict(13)
sage: TW = TripleDict(13, weak_values=True)
sage: a = Foo()
sage: b = Foo()
sage: k = 1
sage: T[a,k,k]=1
sage: T[k,a,k]=2
sage: T[k,k,a]=3
sage: T[k,k,k]=a
sage: TW[b,k,k]=1
sage: TW[k,b,k]=2
sage: TW[k,k,b]=3
sage: TW[k,k,k]=b
sage: len(T)
4
sage: len(TW)
4
sage: (k,k,k) in T
True
sage: (k,k,k) in TW
True
sage: T[k,k,k] is a
True
sage: TW[k,k,k] is b
True
Now, T holds a strong reference to a, namely in T[k,k,k]. Hence, when we delete a, all items of T survive:
sage: del a
sage: _ = gc.collect()
sage: len(T)
4
Only when we remove the strong reference, the items become collectable:
sage: del T[k,k,k]
sage: _ = gc.collect()
sage: len(T)
0
The situation is different for TW, since it only holds weak references to a. Therefore, all items become collectable after deleting a:
sage: del b
sage: _ = gc.collect()
sage: len(TW)
0
Note
The index \(h\) corresponding to the key [k1, k2, k3] is computed as a value of unsigned type size_t as follows:
The natural type for this quantity is Py_ssize_t, which is a signed quantity with the same length as size_t. Storing it in a signed way gives the most efficient storage into PyInt, while preserving sign information.
In previous situations there were some problems with ending up with negative indices, which required casting to an unsigned type, i.e., (<size_t> h)% N since C has a sign-preserving % operation This caused problems on 32 bits systems, see trac ticket #715 for details. This is irrelevant for the current implementation.
AUTHORS:
EXAMPLES:
sage: from sage.structure.coerce_dict import TripleDict
sage: L = TripleDict(31)
sage: L[1,2,3] = None
sage: list(L.iteritems())
[((1, 2, 3), None)]
Bases: object
Erases items from a TripleDict when a weak reference becomes invalid.
This is of internal use only. Instances of this class will be passed as a callback function when creating a weak reference.
EXAMPLES:
sage: from sage.structure.coerce_dict import TripleDict
sage: class A: pass
sage: a = A()
sage: T = TripleDict()
sage: T[a,ZZ,None] = 1
sage: T[ZZ,a,1] = 2
sage: T[a,a,ZZ] = 3
sage: len(T)
3
sage: del a
sage: import gc
sage: n = gc.collect()
sage: len(T) # indirect doctest
0
AUTHOR:
A function like Python’s id() returning signed integers, which are guaranteed to fit in a Py_ssize_t.
Theoretically, there is no guarantee that two different Python objects have different signed_id() values. However, under the mild assumption that a C pointer fits in a Py_ssize_t, this is guaranteed.
TESTS:
sage: a = 1.23e45 # some object
sage: from sage.structure.coerce_dict import signed_id
sage: s = signed_id(a)
sage: id(a) == s or id(a) == s + 2**32 or id(a) == s + 2**64
True
sage: signed_id(a) <= sys.maxsize
True