1
# Copyright (C) 2007 Canonical Ltd
3
# This program is free software; you can redistribute it and/or modify
4
# it under the terms of the GNU General Public License as published by
5
# the Free Software Foundation; either version 2 of the License, or
6
# (at your option) any later version.
8
# This program is distributed in the hope that it will be useful,
9
# but WITHOUT ANY WARRANTY; without even the implied warranty of
10
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
11
# GNU General Public License for more details.
13
# You should have received a copy of the GNU General Public License
14
# along with this program; if not, write to the Free Software
15
# Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
27
from bzrlib.deprecated_graph import (node_distances, select_farthest)
29
STEP_UNIQUE_SEARCHER_EVERY = 5
31
# DIAGRAM of terminology
41
# In this diagram, relative to G and H:
42
# A, B, C, D, E are common ancestors.
43
# C, D and E are border ancestors, because each has a non-common descendant.
44
# D and E are least common ancestors because none of their descendants are
46
# C is not a least common ancestor because its descendant, E, is a common
49
# The find_unique_lca algorithm will pick A in two steps:
50
# 1. find_lca('G', 'H') => ['D', 'E']
51
# 2. Since len(['D', 'E']) > 1, find_lca('D', 'E') => ['A']
54
class DictParentsProvider(object):
55
"""A parents provider for Graph objects."""
57
def __init__(self, ancestry):
58
self.ancestry = ancestry
61
return 'DictParentsProvider(%r)' % self.ancestry
63
def get_parent_map(self, keys):
64
"""See _StackedParentsProvider.get_parent_map"""
65
ancestry = self.ancestry
66
return dict((k, ancestry[k]) for k in keys if k in ancestry)
69
class _StackedParentsProvider(object):
71
def __init__(self, parent_providers):
72
self._parent_providers = parent_providers
75
return "_StackedParentsProvider(%r)" % self._parent_providers
77
def get_parent_map(self, keys):
78
"""Get a mapping of keys => parents
80
A dictionary is returned with an entry for each key present in this
81
source. If this source doesn't have information about a key, it should
84
[NULL_REVISION] is used as the parent of the first user-committed
85
revision. Its parent list is empty.
87
:param keys: An iterable returning keys to check (eg revision_ids)
88
:return: A dictionary mapping each key to its parents
92
for parents_provider in self._parent_providers:
93
new_found = parents_provider.get_parent_map(remaining)
94
found.update(new_found)
95
remaining.difference_update(new_found)
101
class CachingParentsProvider(object):
102
"""A parents provider which will cache the revision => parents in a dict.
104
This is useful for providers that have an expensive lookup.
107
def __init__(self, parent_provider):
108
self._real_provider = parent_provider
109
# Theoretically we could use an LRUCache here
113
return "%s(%r)" % (self.__class__.__name__, self._real_provider)
115
def get_parent_map(self, keys):
116
"""See _StackedParentsProvider.get_parent_map"""
118
# If the _real_provider doesn't have a key, we cache a value of None,
119
# which we then later use to realize we cannot provide a value for that
126
if value is not None:
127
parent_map[key] = value
132
new_parents = self._real_provider.get_parent_map(needed)
133
cache.update(new_parents)
134
parent_map.update(new_parents)
135
needed.difference_update(new_parents)
136
cache.update(dict.fromkeys(needed, None))
141
"""Provide incremental access to revision graphs.
143
This is the generic implementation; it is intended to be subclassed to
144
specialize it for other repository types.
147
def __init__(self, parents_provider):
148
"""Construct a Graph that uses several graphs as its input
150
This should not normally be invoked directly, because there may be
151
specialized implementations for particular repository types. See
152
Repository.get_graph().
154
:param parents_provider: An object providing a get_parent_map call
155
conforming to the behavior of
156
StackedParentsProvider.get_parent_map.
158
if getattr(parents_provider, 'get_parents', None) is not None:
159
self.get_parents = parents_provider.get_parents
160
if getattr(parents_provider, 'get_parent_map', None) is not None:
161
self.get_parent_map = parents_provider.get_parent_map
162
self._parents_provider = parents_provider
165
return 'Graph(%r)' % self._parents_provider
167
def find_lca(self, *revisions):
168
"""Determine the lowest common ancestors of the provided revisions
170
A lowest common ancestor is a common ancestor none of whose
171
descendants are common ancestors. In graphs, unlike trees, there may
172
be multiple lowest common ancestors.
174
This algorithm has two phases. Phase 1 identifies border ancestors,
175
and phase 2 filters border ancestors to determine lowest common
178
In phase 1, border ancestors are identified, using a breadth-first
179
search starting at the bottom of the graph. Searches are stopped
180
whenever a node or one of its descendants is determined to be common
182
In phase 2, the border ancestors are filtered to find the least
183
common ancestors. This is done by searching the ancestries of each
186
Phase 2 is perfomed on the principle that a border ancestor that is
187
not an ancestor of any other border ancestor is a least common
190
Searches are stopped when they find a node that is determined to be a
191
common ancestor of all border ancestors, because this shows that it
192
cannot be a descendant of any border ancestor.
194
The scaling of this operation should be proportional to
195
1. The number of uncommon ancestors
196
2. The number of border ancestors
197
3. The length of the shortest path between a border ancestor and an
198
ancestor of all border ancestors.
200
border_common, common, sides = self._find_border_ancestors(revisions)
201
# We may have common ancestors that can be reached from each other.
202
# - ask for the heads of them to filter it down to only ones that
203
# cannot be reached from each other - phase 2.
204
return self.heads(border_common)
206
def find_difference(self, left_revision, right_revision):
207
"""Determine the graph difference between two revisions"""
208
border, common, searchers = self._find_border_ancestors(
209
[left_revision, right_revision])
210
self._search_for_extra_common(common, searchers)
211
left = searchers[0].seen
212
right = searchers[1].seen
213
return (left.difference(right), right.difference(left))
215
def find_distance_to_null(self, target_revision_id, known_revision_ids):
216
"""Find the left-hand distance to the NULL_REVISION.
218
(This can also be considered the revno of a branch at
221
:param target_revision_id: A revision_id which we would like to know
223
:param known_revision_ids: [(revision_id, revno)] A list of known
224
revno, revision_id tuples. We'll use this to seed the search.
226
# Map from revision_ids to a known value for their revno
227
known_revnos = dict(known_revision_ids)
228
cur_tip = target_revision_id
230
NULL_REVISION = revision.NULL_REVISION
231
known_revnos[NULL_REVISION] = 0
233
searching_known_tips = list(known_revnos.keys())
235
unknown_searched = {}
237
while cur_tip not in known_revnos:
238
unknown_searched[cur_tip] = num_steps
240
to_search = set([cur_tip])
241
to_search.update(searching_known_tips)
242
parent_map = self.get_parent_map(to_search)
243
parents = parent_map.get(cur_tip, None)
244
if not parents: # An empty list or None is a ghost
245
raise errors.GhostRevisionsHaveNoRevno(target_revision_id,
249
for revision_id in searching_known_tips:
250
parents = parent_map.get(revision_id, None)
254
next_revno = known_revnos[revision_id] - 1
255
if next in unknown_searched:
256
# We have enough information to return a value right now
257
return next_revno + unknown_searched[next]
258
if next in known_revnos:
260
known_revnos[next] = next_revno
261
next_known_tips.append(next)
262
searching_known_tips = next_known_tips
264
# We reached a known revision, so just add in how many steps it took to
266
return known_revnos[cur_tip] + num_steps
268
def find_unique_ancestors(self, unique_revision, common_revisions):
269
"""Find the unique ancestors for a revision versus others.
271
This returns the ancestry of unique_revision, excluding all revisions
272
in the ancestry of common_revisions. If unique_revision is in the
273
ancestry, then the empty set will be returned.
275
:param unique_revision: The revision_id whose ancestry we are
277
XXX: Would this API be better if we allowed multiple revisions on
279
:param common_revisions: Revision_ids of ancestries to exclude.
280
:return: A set of revisions in the ancestry of unique_revision
282
if unique_revision in common_revisions:
285
# Algorithm description
286
# 1) Walk backwards from the unique node and all common nodes.
287
# 2) When a node is seen by both sides, stop searching it in the unique
288
# walker, include it in the common walker.
289
# 3) Stop searching when there are no nodes left for the unique walker.
290
# At this point, you have a maximal set of unique nodes. Some of
291
# them may actually be common, and you haven't reached them yet.
292
# 4) Start new searchers for the unique nodes, seeded with the
293
# information you have so far.
294
# 5) Continue searching, stopping the common searches when the search
295
# tip is an ancestor of all unique nodes.
296
# 6) Aggregate together unique searchers when they are searching the
297
# same tips. When all unique searchers are searching the same node,
298
# stop move it to a single 'all_unique_searcher'.
299
# 7) The 'all_unique_searcher' represents the very 'tip' of searching.
300
# Most of the time this produces very little important information.
301
# So don't step it as quickly as the other searchers.
302
# 8) Search is done when all common searchers have completed.
304
unique_searcher, common_searcher = self._find_initial_unique_nodes(
305
[unique_revision], common_revisions)
307
unique_nodes = unique_searcher.seen.difference(common_searcher.seen)
311
(all_unique_searcher,
312
unique_tip_searchers) = self._make_unique_searchers(unique_nodes,
313
unique_searcher, common_searcher)
315
self._refine_unique_nodes(unique_searcher, all_unique_searcher,
316
unique_tip_searchers, common_searcher)
317
true_unique_nodes = unique_nodes.difference(common_searcher.seen)
318
if 'graph' in debug.debug_flags:
319
trace.mutter('Found %d truly unique nodes out of %d',
320
len(true_unique_nodes), len(unique_nodes))
321
return true_unique_nodes
323
def _find_initial_unique_nodes(self, unique_revisions, common_revisions):
324
"""Steps 1-3 of find_unique_ancestors.
326
Find the maximal set of unique nodes. Some of these might actually
327
still be common, but we are sure that there are no other unique nodes.
329
:return: (unique_searcher, common_searcher)
332
unique_searcher = self._make_breadth_first_searcher(unique_revisions)
333
# we know that unique_revisions aren't in common_revisions, so skip
335
unique_searcher.next()
336
common_searcher = self._make_breadth_first_searcher(common_revisions)
338
# As long as we are still finding unique nodes, keep searching
339
while unique_searcher._next_query:
340
next_unique_nodes = set(unique_searcher.step())
341
next_common_nodes = set(common_searcher.step())
343
# Check if either searcher encounters new nodes seen by the other
345
unique_are_common_nodes = next_unique_nodes.intersection(
346
common_searcher.seen)
347
unique_are_common_nodes.update(
348
next_common_nodes.intersection(unique_searcher.seen))
349
if unique_are_common_nodes:
350
ancestors = unique_searcher.find_seen_ancestors(
351
unique_are_common_nodes)
352
# TODO: This is a bit overboard, we only really care about
353
# the ancestors of the tips because the rest we
354
# already know. This is *correct* but causes us to
355
# search too much ancestry.
356
ancestors.update(common_searcher.find_seen_ancestors(ancestors))
357
unique_searcher.stop_searching_any(ancestors)
358
common_searcher.start_searching(ancestors)
360
return unique_searcher, common_searcher
362
def _make_unique_searchers(self, unique_nodes, unique_searcher,
364
"""Create a searcher for all the unique search tips (step 4).
366
As a side effect, the common_searcher will stop searching any nodes
367
that are ancestors of the unique searcher tips.
369
:return: (all_unique_searcher, unique_tip_searchers)
371
unique_tips = self._remove_simple_descendants(unique_nodes,
372
self.get_parent_map(unique_nodes))
374
if len(unique_tips) == 1:
375
unique_tip_searchers = []
376
ancestor_all_unique = unique_searcher.find_seen_ancestors(unique_tips)
378
unique_tip_searchers = []
379
for tip in unique_tips:
380
revs_to_search = unique_searcher.find_seen_ancestors([tip])
381
revs_to_search.update(
382
common_searcher.find_seen_ancestors(revs_to_search))
383
searcher = self._make_breadth_first_searcher(revs_to_search)
384
# We don't care about the starting nodes.
385
searcher._label = tip
387
unique_tip_searchers.append(searcher)
389
ancestor_all_unique = None
390
for searcher in unique_tip_searchers:
391
if ancestor_all_unique is None:
392
ancestor_all_unique = set(searcher.seen)
394
ancestor_all_unique = ancestor_all_unique.intersection(
396
# Collapse all the common nodes into a single searcher
397
all_unique_searcher = self._make_breadth_first_searcher(
399
if ancestor_all_unique:
400
# We've seen these nodes in all the searchers, so we'll just go to
402
all_unique_searcher.step()
404
# Stop any search tips that are already known as ancestors of the
406
stopped_common = common_searcher.stop_searching_any(
407
common_searcher.find_seen_ancestors(ancestor_all_unique))
410
for searcher in unique_tip_searchers:
411
total_stopped += len(searcher.stop_searching_any(
412
searcher.find_seen_ancestors(ancestor_all_unique)))
413
if 'graph' in debug.debug_flags:
414
trace.mutter('For %d unique nodes, created %d + 1 unique searchers'
415
' (%d stopped search tips, %d common ancestors'
416
' (%d stopped common)',
417
len(unique_nodes), len(unique_tip_searchers),
418
total_stopped, len(ancestor_all_unique),
420
return all_unique_searcher, unique_tip_searchers
422
def _step_unique_and_common_searchers(self, common_searcher,
423
unique_tip_searchers,
425
"""Step all the searchers"""
426
newly_seen_common = set(common_searcher.step())
427
newly_seen_unique = set()
428
for searcher in unique_tip_searchers:
429
next = set(searcher.step())
430
next.update(unique_searcher.find_seen_ancestors(next))
431
next.update(common_searcher.find_seen_ancestors(next))
432
for alt_searcher in unique_tip_searchers:
433
if alt_searcher is searcher:
435
next.update(alt_searcher.find_seen_ancestors(next))
436
searcher.start_searching(next)
437
newly_seen_unique.update(next)
438
return newly_seen_common, newly_seen_unique
440
def _find_nodes_common_to_all_unique(self, unique_tip_searchers,
442
newly_seen_unique, step_all_unique):
443
"""Find nodes that are common to all unique_tip_searchers.
445
If it is time, step the all_unique_searcher, and add its nodes to the
448
common_to_all_unique_nodes = newly_seen_unique.copy()
449
for searcher in unique_tip_searchers:
450
common_to_all_unique_nodes.intersection_update(searcher.seen)
451
common_to_all_unique_nodes.intersection_update(
452
all_unique_searcher.seen)
453
# Step all-unique less frequently than the other searchers.
454
# In the common case, we don't need to spider out far here, so
455
# avoid doing extra work.
457
tstart = time.clock()
458
nodes = all_unique_searcher.step()
459
common_to_all_unique_nodes.update(nodes)
460
if 'graph' in debug.debug_flags:
461
tdelta = time.clock() - tstart
462
trace.mutter('all_unique_searcher step() took %.3fs'
463
'for %d nodes (%d total), iteration: %s',
464
tdelta, len(nodes), len(all_unique_searcher.seen),
465
all_unique_searcher._iterations)
466
return common_to_all_unique_nodes
468
def _collapse_unique_searchers(self, unique_tip_searchers,
469
common_to_all_unique_nodes):
470
"""Combine searchers that are searching the same tips.
472
When two searchers are searching the same tips, we can stop one of the
473
searchers. We also know that the maximal set of common ancestors is the
474
intersection of the two original searchers.
476
:return: A list of searchers that are searching unique nodes.
478
# Filter out searchers that don't actually search different
479
# nodes. We already have the ancestry intersection for them
480
unique_search_tips = {}
481
for searcher in unique_tip_searchers:
482
stopped = searcher.stop_searching_any(common_to_all_unique_nodes)
483
will_search_set = frozenset(searcher._next_query)
484
if not will_search_set:
485
if 'graph' in debug.debug_flags:
486
trace.mutter('Unique searcher %s was stopped.'
487
' (%s iterations) %d nodes stopped',
489
searcher._iterations,
491
elif will_search_set not in unique_search_tips:
492
# This searcher is searching a unique set of nodes, let it
493
unique_search_tips[will_search_set] = [searcher]
495
unique_search_tips[will_search_set].append(searcher)
496
# TODO: it might be possible to collapse searchers faster when they
497
# only have *some* search tips in common.
498
next_unique_searchers = []
499
for searchers in unique_search_tips.itervalues():
500
if len(searchers) == 1:
501
# Searching unique tips, go for it
502
next_unique_searchers.append(searchers[0])
504
# These searchers have started searching the same tips, we
505
# don't need them to cover the same ground. The
506
# intersection of their ancestry won't change, so create a
507
# new searcher, combining their histories.
508
next_searcher = searchers[0]
509
for searcher in searchers[1:]:
510
next_searcher.seen.intersection_update(searcher.seen)
511
if 'graph' in debug.debug_flags:
512
trace.mutter('Combining %d searchers into a single'
513
' searcher searching %d nodes with'
516
len(next_searcher._next_query),
517
len(next_searcher.seen))
518
next_unique_searchers.append(next_searcher)
519
return next_unique_searchers
521
def _refine_unique_nodes(self, unique_searcher, all_unique_searcher,
522
unique_tip_searchers, common_searcher):
523
"""Steps 5-8 of find_unique_ancestors.
525
This function returns when common_searcher has stopped searching for
528
# We step the ancestor_all_unique searcher only every
529
# STEP_UNIQUE_SEARCHER_EVERY steps.
530
step_all_unique_counter = 0
531
# While we still have common nodes to search
532
while common_searcher._next_query:
534
newly_seen_unique) = self._step_unique_and_common_searchers(
535
common_searcher, unique_tip_searchers, unique_searcher)
536
# These nodes are common ancestors of all unique nodes
537
common_to_all_unique_nodes = self._find_nodes_common_to_all_unique(
538
unique_tip_searchers, all_unique_searcher, newly_seen_unique,
539
step_all_unique_counter==0)
540
step_all_unique_counter = ((step_all_unique_counter + 1)
541
% STEP_UNIQUE_SEARCHER_EVERY)
543
if newly_seen_common:
544
# If a 'common' node is an ancestor of all unique searchers, we
545
# can stop searching it.
546
common_searcher.stop_searching_any(
547
all_unique_searcher.seen.intersection(newly_seen_common))
548
if common_to_all_unique_nodes:
549
common_to_all_unique_nodes.update(
550
common_searcher.find_seen_ancestors(
551
common_to_all_unique_nodes))
552
# The all_unique searcher can start searching the common nodes
553
# but everyone else can stop.
554
# This is the sort of thing where we would like to not have it
555
# start_searching all of the nodes, but only mark all of them
556
# as seen, and have it search only the actual tips. Otherwise
557
# it is another get_parent_map() traversal for it to figure out
558
# what we already should know.
559
all_unique_searcher.start_searching(common_to_all_unique_nodes)
560
common_searcher.stop_searching_any(common_to_all_unique_nodes)
562
next_unique_searchers = self._collapse_unique_searchers(
563
unique_tip_searchers, common_to_all_unique_nodes)
564
if len(unique_tip_searchers) != len(next_unique_searchers):
565
if 'graph' in debug.debug_flags:
566
trace.mutter('Collapsed %d unique searchers => %d'
568
len(unique_tip_searchers),
569
len(next_unique_searchers),
570
all_unique_searcher._iterations)
571
unique_tip_searchers = next_unique_searchers
573
@symbol_versioning.deprecated_method(symbol_versioning.one_one)
574
def get_parents(self, revisions):
575
"""Find revision ids of the parents of a list of revisions
577
A list is returned of the same length as the input. Each entry
578
is a list of parent ids for the corresponding input revision.
580
[NULL_REVISION] is used as the parent of the first user-committed
581
revision. Its parent list is empty.
583
If the revision is not present (i.e. a ghost), None is used in place
584
of the list of parents.
586
Deprecated in bzr 1.2 - please see get_parent_map.
588
parents = self.get_parent_map(revisions)
589
return [parents.get(r, None) for r in revisions]
591
def get_parent_map(self, revisions):
592
"""Get a map of key:parent_list for revisions.
594
This implementation delegates to get_parents, for old parent_providers
595
that do not supply get_parent_map.
598
for rev, parents in self.get_parents(revisions):
599
if parents is not None:
600
result[rev] = parents
603
def _make_breadth_first_searcher(self, revisions):
604
return _BreadthFirstSearcher(revisions, self)
606
def _find_border_ancestors(self, revisions):
607
"""Find common ancestors with at least one uncommon descendant.
609
Border ancestors are identified using a breadth-first
610
search starting at the bottom of the graph. Searches are stopped
611
whenever a node or one of its descendants is determined to be common.
613
This will scale with the number of uncommon ancestors.
615
As well as the border ancestors, a set of seen common ancestors and a
616
list of sets of seen ancestors for each input revision is returned.
617
This allows calculation of graph difference from the results of this
620
if None in revisions:
621
raise errors.InvalidRevisionId(None, self)
622
common_ancestors = set()
623
searchers = [self._make_breadth_first_searcher([r])
625
active_searchers = searchers[:]
626
border_ancestors = set()
630
for searcher in searchers:
631
new_ancestors = searcher.step()
633
newly_seen.update(new_ancestors)
635
for revision in newly_seen:
636
if revision in common_ancestors:
637
# Not a border ancestor because it was seen as common
639
new_common.add(revision)
641
for searcher in searchers:
642
if revision not in searcher.seen:
645
# This is a border because it is a first common that we see
646
# after walking for a while.
647
border_ancestors.add(revision)
648
new_common.add(revision)
650
for searcher in searchers:
651
new_common.update(searcher.find_seen_ancestors(new_common))
652
for searcher in searchers:
653
searcher.start_searching(new_common)
654
common_ancestors.update(new_common)
656
# Figure out what the searchers will be searching next, and if
657
# there is only 1 set being searched, then we are done searching,
658
# since all searchers would have to be searching the same data,
659
# thus it *must* be in common.
660
unique_search_sets = set()
661
for searcher in searchers:
662
will_search_set = frozenset(searcher._next_query)
663
if will_search_set not in unique_search_sets:
664
# This searcher is searching a unique set of nodes, let it
665
unique_search_sets.add(will_search_set)
667
if len(unique_search_sets) == 1:
668
nodes = unique_search_sets.pop()
669
uncommon_nodes = nodes.difference(common_ancestors)
671
raise AssertionError("Somehow we ended up converging"
672
" without actually marking them as"
675
"\nuncommon_nodes: %s"
676
% (revisions, uncommon_nodes))
678
return border_ancestors, common_ancestors, searchers
680
def heads(self, keys):
681
"""Return the heads from amongst keys.
683
This is done by searching the ancestries of each key. Any key that is
684
reachable from another key is not returned; all the others are.
686
This operation scales with the relative depth between any two keys. If
687
any two keys are completely disconnected all ancestry of both sides
690
:param keys: An iterable of keys.
691
:return: A set of the heads. Note that as a set there is no ordering
692
information. Callers will need to filter their input to create
693
order if they need it.
695
candidate_heads = set(keys)
696
if revision.NULL_REVISION in candidate_heads:
697
# NULL_REVISION is only a head if it is the only entry
698
candidate_heads.remove(revision.NULL_REVISION)
699
if not candidate_heads:
700
return set([revision.NULL_REVISION])
701
if len(candidate_heads) < 2:
702
return candidate_heads
703
searchers = dict((c, self._make_breadth_first_searcher([c]))
704
for c in candidate_heads)
705
active_searchers = dict(searchers)
706
# skip over the actual candidate for each searcher
707
for searcher in active_searchers.itervalues():
709
# The common walker finds nodes that are common to two or more of the
710
# input keys, so that we don't access all history when a currently
711
# uncommon search point actually meets up with something behind a
712
# common search point. Common search points do not keep searches
713
# active; they just allow us to make searches inactive without
714
# accessing all history.
715
common_walker = self._make_breadth_first_searcher([])
716
while len(active_searchers) > 0:
721
except StopIteration:
722
# No common points being searched at this time.
724
for candidate in active_searchers.keys():
726
searcher = active_searchers[candidate]
728
# rare case: we deleted candidate in a previous iteration
729
# through this for loop, because it was determined to be
730
# a descendant of another candidate.
733
ancestors.update(searcher.next())
734
except StopIteration:
735
del active_searchers[candidate]
737
# process found nodes
739
for ancestor in ancestors:
740
if ancestor in candidate_heads:
741
candidate_heads.remove(ancestor)
742
del searchers[ancestor]
743
if ancestor in active_searchers:
744
del active_searchers[ancestor]
745
# it may meet up with a known common node
746
if ancestor in common_walker.seen:
747
# some searcher has encountered our known common nodes:
749
ancestor_set = set([ancestor])
750
for searcher in searchers.itervalues():
751
searcher.stop_searching_any(ancestor_set)
753
# or it may have been just reached by all the searchers:
754
for searcher in searchers.itervalues():
755
if ancestor not in searcher.seen:
758
# The final active searcher has just reached this node,
759
# making it be known as a descendant of all candidates,
760
# so we can stop searching it, and any seen ancestors
761
new_common.add(ancestor)
762
for searcher in searchers.itervalues():
764
searcher.find_seen_ancestors([ancestor])
765
searcher.stop_searching_any(seen_ancestors)
766
common_walker.start_searching(new_common)
767
return candidate_heads
769
def find_merge_order(self, tip_revision_id, lca_revision_ids):
770
"""Find the order that each revision was merged into tip.
772
This basically just walks backwards with a stack, and walks left-first
773
until it finds a node to stop.
775
if len(lca_revision_ids) == 1:
776
return list(lca_revision_ids)
777
looking_for = set(lca_revision_ids)
778
# TODO: Is there a way we could do this "faster" by batching up the
779
# get_parent_map requests?
780
# TODO: Should we also be culling the ancestry search right away? We
781
# could add looking_for to the "stop" list, and walk their
782
# ancestry in batched mode. The flip side is it might mean we walk a
783
# lot of "stop" nodes, rather than only the minimum.
784
# Then again, without it we may trace back into ancestry we could have
786
stack = [tip_revision_id]
789
while stack and looking_for:
792
if next in looking_for:
794
looking_for.remove(next)
795
if len(looking_for) == 1:
796
found.append(looking_for.pop())
799
parent_ids = self.get_parent_map([next]).get(next, None)
800
if not parent_ids: # Ghost, nothing to search here
802
for parent_id in reversed(parent_ids):
803
# TODO: (performance) We see the parent at this point, but we
804
# wait to mark it until later to make sure we get left
805
# parents before right parents. However, instead of
806
# waiting until we have traversed enough parents, we
807
# could instead note that we've found it, and once all
808
# parents are in the stack, just reverse iterate the
810
if parent_id not in stop:
811
# this will need to be searched
812
stack.append(parent_id)
816
def find_unique_lca(self, left_revision, right_revision,
818
"""Find a unique LCA.
820
Find lowest common ancestors. If there is no unique common
821
ancestor, find the lowest common ancestors of those ancestors.
823
Iteration stops when a unique lowest common ancestor is found.
824
The graph origin is necessarily a unique lowest common ancestor.
826
Note that None is not an acceptable substitute for NULL_REVISION.
827
in the input for this method.
829
:param count_steps: If True, the return value will be a tuple of
830
(unique_lca, steps) where steps is the number of times that
831
find_lca was run. If False, only unique_lca is returned.
833
revisions = [left_revision, right_revision]
837
lca = self.find_lca(*revisions)
845
raise errors.NoCommonAncestor(left_revision, right_revision)
848
def iter_ancestry(self, revision_ids):
849
"""Iterate the ancestry of this revision.
851
:param revision_ids: Nodes to start the search
852
:return: Yield tuples mapping a revision_id to its parents for the
853
ancestry of revision_id.
854
Ghosts will be returned with None as their parents, and nodes
855
with no parents will have NULL_REVISION as their only parent. (As
856
defined by get_parent_map.)
857
There will also be a node for (NULL_REVISION, ())
859
pending = set(revision_ids)
862
processed.update(pending)
863
next_map = self.get_parent_map(pending)
865
for item in next_map.iteritems():
867
next_pending.update(p for p in item[1] if p not in processed)
868
ghosts = pending.difference(next_map)
871
pending = next_pending
873
def iter_topo_order(self, revisions):
874
"""Iterate through the input revisions in topological order.
876
This sorting only ensures that parents come before their children.
877
An ancestor may sort after a descendant if the relationship is not
878
visible in the supplied list of revisions.
880
sorter = tsort.TopoSorter(self.get_parent_map(revisions))
881
return sorter.iter_topo_order()
883
def is_ancestor(self, candidate_ancestor, candidate_descendant):
884
"""Determine whether a revision is an ancestor of another.
886
We answer this using heads() as heads() has the logic to perform the
887
smallest number of parent lookups to determine the ancestral
888
relationship between N revisions.
890
return set([candidate_descendant]) == self.heads(
891
[candidate_ancestor, candidate_descendant])
893
def _search_for_extra_common(self, common, searchers):
894
"""Make sure that unique nodes are genuinely unique.
896
After _find_border_ancestors, all nodes marked "common" are indeed
897
common. Some of the nodes considered unique are not, due to history
898
shortcuts stopping the searches early.
900
We know that we have searched enough when all common search tips are
901
descended from all unique (uncommon) nodes because we know that a node
902
cannot be an ancestor of its own ancestor.
904
:param common: A set of common nodes
905
:param searchers: The searchers returned from _find_border_ancestors
909
# A) The passed in searchers should all be on the same tips, thus
910
# they should be considered the "common" searchers.
911
# B) We find the difference between the searchers, these are the
912
# "unique" nodes for each side.
913
# C) We do a quick culling so that we only start searching from the
914
# more interesting unique nodes. (A unique ancestor is more
915
# interesting than any of its children.)
916
# D) We start searching for ancestors common to all unique nodes.
917
# E) We have the common searchers stop searching any ancestors of
919
# F) When there are no more common search tips, we stop
921
# TODO: We need a way to remove unique_searchers when they overlap with
922
# other unique searchers.
923
if len(searchers) != 2:
924
raise NotImplementedError(
925
"Algorithm not yet implemented for > 2 searchers")
926
common_searchers = searchers
927
left_searcher = searchers[0]
928
right_searcher = searchers[1]
929
unique = left_searcher.seen.symmetric_difference(right_searcher.seen)
930
if not unique: # No unique nodes, nothing to do
932
total_unique = len(unique)
933
unique = self._remove_simple_descendants(unique,
934
self.get_parent_map(unique))
935
simple_unique = len(unique)
937
unique_searchers = []
938
for revision_id in unique:
939
if revision_id in left_searcher.seen:
940
parent_searcher = left_searcher
942
parent_searcher = right_searcher
943
revs_to_search = parent_searcher.find_seen_ancestors([revision_id])
944
if not revs_to_search: # XXX: This shouldn't be possible
945
revs_to_search = [revision_id]
946
searcher = self._make_breadth_first_searcher(revs_to_search)
947
# We don't care about the starting nodes.
949
unique_searchers.append(searcher)
951
# possible todo: aggregate the common searchers into a single common
952
# searcher, just make sure that we include the nodes into the .seen
953
# properties of the original searchers
955
ancestor_all_unique = None
956
for searcher in unique_searchers:
957
if ancestor_all_unique is None:
958
ancestor_all_unique = set(searcher.seen)
960
ancestor_all_unique = ancestor_all_unique.intersection(
963
trace.mutter('Started %s unique searchers for %s unique revisions',
964
simple_unique, total_unique)
966
while True: # If we have no more nodes we have nothing to do
967
newly_seen_common = set()
968
for searcher in common_searchers:
969
newly_seen_common.update(searcher.step())
970
newly_seen_unique = set()
971
for searcher in unique_searchers:
972
newly_seen_unique.update(searcher.step())
973
new_common_unique = set()
974
for revision in newly_seen_unique:
975
for searcher in unique_searchers:
976
if revision not in searcher.seen:
979
# This is a border because it is a first common that we see
980
# after walking for a while.
981
new_common_unique.add(revision)
982
if newly_seen_common:
983
# These are nodes descended from one of the 'common' searchers.
984
# Make sure all searchers are on the same page
985
for searcher in common_searchers:
986
newly_seen_common.update(
987
searcher.find_seen_ancestors(newly_seen_common))
988
# We start searching the whole ancestry. It is a bit wasteful,
989
# though. We really just want to mark all of these nodes as
990
# 'seen' and then start just the tips. However, it requires a
991
# get_parent_map() call to figure out the tips anyway, and all
992
# redundant requests should be fairly fast.
993
for searcher in common_searchers:
994
searcher.start_searching(newly_seen_common)
996
# If a 'common' node is an ancestor of all unique searchers, we
997
# can stop searching it.
998
stop_searching_common = ancestor_all_unique.intersection(
1000
if stop_searching_common:
1001
for searcher in common_searchers:
1002
searcher.stop_searching_any(stop_searching_common)
1003
if new_common_unique:
1004
# We found some ancestors that are common
1005
for searcher in unique_searchers:
1006
new_common_unique.update(
1007
searcher.find_seen_ancestors(new_common_unique))
1008
# Since these are common, we can grab another set of ancestors
1010
for searcher in common_searchers:
1011
new_common_unique.update(
1012
searcher.find_seen_ancestors(new_common_unique))
1014
# We can tell all of the unique searchers to start at these
1015
# nodes, and tell all of the common searchers to *stop*
1016
# searching these nodes
1017
for searcher in unique_searchers:
1018
searcher.start_searching(new_common_unique)
1019
for searcher in common_searchers:
1020
searcher.stop_searching_any(new_common_unique)
1021
ancestor_all_unique.update(new_common_unique)
1023
# Filter out searchers that don't actually search different
1024
# nodes. We already have the ancestry intersection for them
1025
next_unique_searchers = []
1026
unique_search_sets = set()
1027
for searcher in unique_searchers:
1028
will_search_set = frozenset(searcher._next_query)
1029
if will_search_set not in unique_search_sets:
1030
# This searcher is searching a unique set of nodes, let it
1031
unique_search_sets.add(will_search_set)
1032
next_unique_searchers.append(searcher)
1033
unique_searchers = next_unique_searchers
1034
for searcher in common_searchers:
1035
if searcher._next_query:
1038
# All common searcher have stopped searching
1041
def _remove_simple_descendants(self, revisions, parent_map):
1042
"""remove revisions which are children of other ones in the set
1044
This doesn't do any graph searching, it just checks the immediate
1045
parent_map to find if there are any children which can be removed.
1047
:param revisions: A set of revision_ids
1048
:return: A set of revision_ids with the children removed
1050
simple_ancestors = revisions.copy()
1051
# TODO: jam 20071214 we *could* restrict it to searching only the
1052
# parent_map of revisions already present in 'revisions', but
1053
# considering the general use case, I think this is actually
1056
# This is the same as the following loop. I don't know that it is any
1058
## simple_ancestors.difference_update(r for r, p_ids in parent_map.iteritems()
1059
## if p_ids is not None and revisions.intersection(p_ids))
1060
## return simple_ancestors
1062
# Yet Another Way, invert the parent map (which can be cached)
1064
## for revision_id, parent_ids in parent_map.iteritems():
1065
## for p_id in parent_ids:
1066
## descendants.setdefault(p_id, []).append(revision_id)
1067
## for revision in revisions.intersection(descendants):
1068
## simple_ancestors.difference_update(descendants[revision])
1069
## return simple_ancestors
1070
for revision, parent_ids in parent_map.iteritems():
1071
if parent_ids is None:
1073
for parent_id in parent_ids:
1074
if parent_id in revisions:
1075
# This node has a parent present in the set, so we can
1077
simple_ancestors.discard(revision)
1079
return simple_ancestors
1082
class HeadsCache(object):
1083
"""A cache of results for graph heads calls."""
1085
def __init__(self, graph):
1089
def heads(self, keys):
1090
"""Return the heads of keys.
1092
This matches the API of Graph.heads(), specifically the return value is
1093
a set which can be mutated, and ordering of the input is not preserved
1096
:see also: Graph.heads.
1097
:param keys: The keys to calculate heads for.
1098
:return: A set containing the heads, which may be mutated without
1099
affecting future lookups.
1101
keys = frozenset(keys)
1103
return set(self._heads[keys])
1105
heads = self.graph.heads(keys)
1106
self._heads[keys] = heads
1110
class FrozenHeadsCache(object):
1111
"""Cache heads() calls, assuming the caller won't modify them."""
1113
def __init__(self, graph):
1117
def heads(self, keys):
1118
"""Return the heads of keys.
1120
Similar to Graph.heads(). The main difference is that the return value
1121
is a frozen set which cannot be mutated.
1123
:see also: Graph.heads.
1124
:param keys: The keys to calculate heads for.
1125
:return: A frozenset containing the heads.
1127
keys = frozenset(keys)
1129
return self._heads[keys]
1131
heads = frozenset(self.graph.heads(keys))
1132
self._heads[keys] = heads
1135
def cache(self, keys, heads):
1136
"""Store a known value."""
1137
self._heads[frozenset(keys)] = frozenset(heads)
1140
class _BreadthFirstSearcher(object):
1141
"""Parallel search breadth-first the ancestry of revisions.
1143
This class implements the iterator protocol, but additionally
1144
1. provides a set of seen ancestors, and
1145
2. allows some ancestries to be unsearched, via stop_searching_any
1148
def __init__(self, revisions, parents_provider):
1149
self._iterations = 0
1150
self._next_query = set(revisions)
1152
self._started_keys = set(self._next_query)
1153
self._stopped_keys = set()
1154
self._parents_provider = parents_provider
1155
self._returning = 'next_with_ghosts'
1156
self._current_present = set()
1157
self._current_ghosts = set()
1158
self._current_parents = {}
1161
if self._iterations:
1162
prefix = "searching"
1165
search = '%s=%r' % (prefix, list(self._next_query))
1166
return ('_BreadthFirstSearcher(iterations=%d, %s,'
1167
' seen=%r)' % (self._iterations, search, list(self.seen)))
1169
def get_result(self):
1170
"""Get a SearchResult for the current state of this searcher.
1172
:return: A SearchResult for this search so far. The SearchResult is
1173
static - the search can be advanced and the search result will not
1174
be invalidated or altered.
1176
if self._returning == 'next':
1177
# We have to know the current nodes children to be able to list the
1178
# exclude keys for them. However, while we could have a second
1179
# look-ahead result buffer and shuffle things around, this method
1180
# is typically only called once per search - when memoising the
1181
# results of the search.
1182
found, ghosts, next, parents = self._do_query(self._next_query)
1183
# pretend we didn't query: perhaps we should tweak _do_query to be
1184
# entirely stateless?
1185
self.seen.difference_update(next)
1186
next_query = next.union(ghosts)
1188
next_query = self._next_query
1189
excludes = self._stopped_keys.union(next_query)
1190
included_keys = self.seen.difference(excludes)
1191
return SearchResult(self._started_keys, excludes, len(included_keys),
1197
except StopIteration:
1201
"""Return the next ancestors of this revision.
1203
Ancestors are returned in the order they are seen in a breadth-first
1204
traversal. No ancestor will be returned more than once. Ancestors are
1205
returned before their parentage is queried, so ghosts and missing
1206
revisions (including the start revisions) are included in the result.
1207
This can save a round trip in LCA style calculation by allowing
1208
convergence to be detected without reading the data for the revision
1209
the convergence occurs on.
1211
:return: A set of revision_ids.
1213
if self._returning != 'next':
1214
# switch to returning the query, not the results.
1215
self._returning = 'next'
1216
self._iterations += 1
1219
if len(self._next_query) == 0:
1220
raise StopIteration()
1221
# We have seen what we're querying at this point as we are returning
1222
# the query, not the results.
1223
self.seen.update(self._next_query)
1224
return self._next_query
1226
def next_with_ghosts(self):
1227
"""Return the next found ancestors, with ghosts split out.
1229
Ancestors are returned in the order they are seen in a breadth-first
1230
traversal. No ancestor will be returned more than once. Ancestors are
1231
returned only after asking for their parents, which allows us to detect
1232
which revisions are ghosts and which are not.
1234
:return: A tuple with (present ancestors, ghost ancestors) sets.
1236
if self._returning != 'next_with_ghosts':
1237
# switch to returning the results, not the current query.
1238
self._returning = 'next_with_ghosts'
1240
if len(self._next_query) == 0:
1241
raise StopIteration()
1243
return self._current_present, self._current_ghosts
1246
"""Advance the search.
1248
Updates self.seen, self._next_query, self._current_present,
1249
self._current_ghosts, self._current_parents and self._iterations.
1251
self._iterations += 1
1252
found, ghosts, next, parents = self._do_query(self._next_query)
1253
self._current_present = found
1254
self._current_ghosts = ghosts
1255
self._next_query = next
1256
self._current_parents = parents
1257
# ghosts are implicit stop points, otherwise the search cannot be
1258
# repeated when ghosts are filled.
1259
self._stopped_keys.update(ghosts)
1261
def _do_query(self, revisions):
1262
"""Query for revisions.
1264
Adds revisions to the seen set.
1266
:param revisions: Revisions to query.
1267
:return: A tuple: (set(found_revisions), set(ghost_revisions),
1268
set(parents_of_found_revisions), dict(found_revisions:parents)).
1270
found_revisions = set()
1271
parents_of_found = set()
1272
# revisions may contain nodes that point to other nodes in revisions:
1273
# we want to filter them out.
1274
self.seen.update(revisions)
1275
parent_map = self._parents_provider.get_parent_map(revisions)
1276
found_revisions.update(parent_map)
1277
for rev_id, parents in parent_map.iteritems():
1280
new_found_parents = [p for p in parents if p not in self.seen]
1281
if new_found_parents:
1282
# Calling set.update() with an empty generator is actually
1284
parents_of_found.update(new_found_parents)
1285
ghost_revisions = revisions - found_revisions
1286
return found_revisions, ghost_revisions, parents_of_found, parent_map
1291
def find_seen_ancestors(self, revisions):
1292
"""Find ancestors of these revisions that have already been seen.
1294
This function generally makes the assumption that querying for the
1295
parents of a node that has already been queried is reasonably cheap.
1296
(eg, not a round trip to a remote host).
1298
# TODO: Often we might ask one searcher for its seen ancestors, and
1299
# then ask another searcher the same question. This can result in
1300
# searching the same revisions repeatedly if the two searchers
1301
# have a lot of overlap.
1302
all_seen = self.seen
1303
pending = set(revisions).intersection(all_seen)
1304
seen_ancestors = set(pending)
1306
if self._returning == 'next':
1307
# self.seen contains what nodes have been returned, not what nodes
1308
# have been queried. We don't want to probe for nodes that haven't
1309
# been searched yet.
1310
not_searched_yet = self._next_query
1312
not_searched_yet = ()
1313
pending.difference_update(not_searched_yet)
1314
get_parent_map = self._parents_provider.get_parent_map
1316
parent_map = get_parent_map(pending)
1318
# We don't care if it is a ghost, since it can't be seen if it is
1320
for parent_ids in parent_map.itervalues():
1321
all_parents.extend(parent_ids)
1322
next_pending = all_seen.intersection(all_parents).difference(seen_ancestors)
1323
seen_ancestors.update(next_pending)
1324
next_pending.difference_update(not_searched_yet)
1325
pending = next_pending
1327
return seen_ancestors
1329
def stop_searching_any(self, revisions):
1331
Remove any of the specified revisions from the search list.
1333
None of the specified revisions are required to be present in the
1336
It is okay to call stop_searching_any() for revisions which were seen
1337
in previous iterations. It is the callers responsibility to call
1338
find_seen_ancestors() to make sure that current search tips that are
1339
ancestors of those revisions are also stopped. All explicitly stopped
1340
revisions will be excluded from the search result's get_keys(), though.
1342
# TODO: does this help performance?
1345
revisions = frozenset(revisions)
1346
if self._returning == 'next':
1347
stopped = self._next_query.intersection(revisions)
1348
self._next_query = self._next_query.difference(revisions)
1350
stopped_present = self._current_present.intersection(revisions)
1351
stopped = stopped_present.union(
1352
self._current_ghosts.intersection(revisions))
1353
self._current_present.difference_update(stopped)
1354
self._current_ghosts.difference_update(stopped)
1355
# stopping 'x' should stop returning parents of 'x', but
1356
# not if 'y' always references those same parents
1357
stop_rev_references = {}
1358
for rev in stopped_present:
1359
for parent_id in self._current_parents[rev]:
1360
if parent_id not in stop_rev_references:
1361
stop_rev_references[parent_id] = 0
1362
stop_rev_references[parent_id] += 1
1363
# if only the stopped revisions reference it, the ref count will be
1365
for parents in self._current_parents.itervalues():
1366
for parent_id in parents:
1368
stop_rev_references[parent_id] -= 1
1371
stop_parents = set()
1372
for rev_id, refs in stop_rev_references.iteritems():
1374
stop_parents.add(rev_id)
1375
self._next_query.difference_update(stop_parents)
1376
self._stopped_keys.update(stopped)
1377
self._stopped_keys.update(revisions - set([revision.NULL_REVISION]))
1380
def start_searching(self, revisions):
1381
"""Add revisions to the search.
1383
The parents of revisions will be returned from the next call to next()
1384
or next_with_ghosts(). If next_with_ghosts was the most recently used
1385
next* call then the return value is the result of looking up the
1386
ghost/not ghost status of revisions. (A tuple (present, ghosted)).
1388
revisions = frozenset(revisions)
1389
self._started_keys.update(revisions)
1390
new_revisions = revisions.difference(self.seen)
1391
if self._returning == 'next':
1392
self._next_query.update(new_revisions)
1393
self.seen.update(new_revisions)
1395
# perform a query on revisions
1396
revs, ghosts, query, parents = self._do_query(revisions)
1397
self._stopped_keys.update(ghosts)
1398
self._current_present.update(revs)
1399
self._current_ghosts.update(ghosts)
1400
self._next_query.update(query)
1401
self._current_parents.update(parents)
1405
class SearchResult(object):
1406
"""The result of a breadth first search.
1408
A SearchResult provides the ability to reconstruct the search or access a
1409
set of the keys the search found.
1412
def __init__(self, start_keys, exclude_keys, key_count, keys):
1413
"""Create a SearchResult.
1415
:param start_keys: The keys the search started at.
1416
:param exclude_keys: The keys the search excludes.
1417
:param key_count: The total number of keys (from start to but not
1419
:param keys: The keys the search found. Note that in future we may get
1420
a SearchResult from a smart server, in which case the keys list is
1421
not necessarily immediately available.
1423
self._recipe = (start_keys, exclude_keys, key_count)
1424
self._keys = frozenset(keys)
1426
def get_recipe(self):
1427
"""Return a recipe that can be used to replay this search.
1429
The recipe allows reconstruction of the same results at a later date
1430
without knowing all the found keys. The essential elements are a list
1431
of keys to start and and to stop at. In order to give reproducible
1432
results when ghosts are encountered by a search they are automatically
1433
added to the exclude list (or else ghost filling may alter the
1436
:return: A tuple (start_keys_set, exclude_keys_set, revision_count). To
1437
recreate the results of this search, create a breadth first
1438
searcher on the same graph starting at start_keys. Then call next()
1439
(or next_with_ghosts()) repeatedly, and on every result, call
1440
stop_searching_any on any keys from the exclude_keys set. The
1441
revision_count value acts as a trivial cross-check - the found
1442
revisions of the new search should have as many elements as
1443
revision_count. If it does not, then additional revisions have been
1444
ghosted since the search was executed the first time and the second
1450
"""Return the keys found in this search.
1452
:return: A set of keys.
1457
def collapse_linear_regions(parent_map):
1458
"""Collapse regions of the graph that are 'linear'.
1464
can be collapsed by removing B and getting::
1468
:param parent_map: A dictionary mapping children to their parents
1469
:return: Another dictionary with 'linear' chains collapsed
1471
# Note: this isn't a strictly minimal collapse. For example:
1479
# Will not have 'D' removed, even though 'E' could fit. Also:
1485
# A and C are both kept because they are edges of the graph. We *could* get
1486
# rid of A if we wanted.
1494
# Will not have any nodes removed, even though you do have an
1495
# 'uninteresting' linear D->B and E->C
1497
for child, parents in parent_map.iteritems():
1498
children.setdefault(child, [])
1500
children.setdefault(p, []).append(child)
1502
orig_children = dict(children)
1504
result = dict(parent_map)
1505
for node in parent_map:
1506
parents = result[node]
1507
if len(parents) == 1:
1508
parent_children = children[parents[0]]
1509
if len(parent_children) != 1:
1510
# This is not the only child
1512
node_children = children[node]
1513
if len(node_children) != 1:
1515
child_parents = result.get(node_children[0], None)
1516
if len(child_parents) != 1:
1517
# This is not its only parent
1519
# The child of this node only points at it, and the parent only has
1520
# this as a child. remove this node, and join the others together
1521
result[node_children[0]] = parents
1522
children[parents[0]] = node_children
added
removed
bzrlib/graph.py
