13
13
# You should have received a copy of the GNU General Public License
14
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
19
from bzrlib.errors import GraphCycleError
15
# Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
17
"""Topological sorting routines."""
19
from __future__ import absolute_import
25
revision as _mod_revision,
29
__all__ = ["topo_sort", "TopoSorter", "merge_sort", "MergeSorter"]
21
32
def topo_sort(graph):
22
33
"""Topological sort a graph.
24
35
graph -- sequence of pairs of node->parents_list.
26
The result is a list of node names, such that all parents come before
29
Nodes at the same depth are returned in sorted order.
37
The result is a list of node names, such that all parents come before their
31
40
node identifiers can be any hashable object, and are typically strings.
33
parents = {} # node -> list of parents
34
children = {} # node -> list of children
35
for node, node_parents in graph:
36
assert node not in parents, \
37
('node %r repeated in graph' % node)
38
parents[node] = set(node_parents)
39
if node not in children:
40
children[node] = set()
41
for parent in node_parents:
42
if parent in children:
43
children[parent].add(node)
45
children[parent] = set([node])
48
# find nodes with no parents, and take them now
49
no_parents = [n for n in parents if len(parents[n]) == 0]
52
raise GraphCycleError(parents)
55
for child in children[n]:
56
assert n in parents[child]
57
parents[child].remove(n)
42
This function has the same purpose as the TopoSorter class, but uses a
43
different algorithm to sort the graph. That means that while both return a
44
list with parents before their child nodes, the exact ordering can be
47
topo_sort is faster when the whole list is needed, while when iterating
48
over a part of the list, TopoSorter.iter_topo_order should be used.
50
kg = _mod_graph.KnownGraph(dict(graph))
54
class TopoSorter(object):
56
def __init__(self, graph):
57
"""Topological sorting of a graph.
59
:param graph: sequence of pairs of node_name->parent_names_list.
60
i.e. [('C', ['B']), ('B', ['A']), ('A', [])]
61
For this input the output from the sort or
62
iter_topo_order routines will be:
65
node identifiers can be any hashable object, and are typically strings.
67
If you have a graph like [('a', ['b']), ('a', ['c'])] this will only use
68
one of the two values for 'a'.
70
The graph is sorted lazily: until you iterate or sort the input is
71
not processed other than to create an internal representation.
73
iteration or sorting may raise GraphCycleError if a cycle is present
76
# store a dict of the graph.
77
self._graph = dict(graph)
80
"""Sort the graph and return as a list.
82
After calling this the sorter is empty and you must create a new one.
84
return list(self.iter_topo_order())
86
### Useful if fiddling with this code.
88
### sorted_names = list(self.iter_topo_order())
89
### for index in range(len(sorted_names)):
90
### rev = sorted_names[index]
91
### for left_index in range(index):
92
### if rev in self.original_graph[sorted_names[left_index]]:
93
### print "revision in parent list of earlier revision"
94
### import pdb;pdb.set_trace()
96
def iter_topo_order(self):
97
"""Yield the nodes of the graph in a topological order.
99
After finishing iteration the sorter is empty and you cannot continue
103
visitable = set(graph)
105
# this is a stack storing the depth first search into the graph.
106
pending_node_stack = []
107
# at each level of 'recursion' we have to check each parent. This
108
# stack stores the parents we have not yet checked for the node at the
109
# matching depth in pending_node_stack
110
pending_parents_stack = []
112
# this is a set of the completed nodes for fast checking whether a
113
# parent in a node we are processing on the stack has already been
114
# emitted and thus can be skipped.
115
completed_node_names = set()
118
# now pick a random node in the source graph, and transfer it to the
119
# top of the depth first search stack of pending nodes.
120
node_name, parents = graph.popitem()
121
pending_node_stack.append(node_name)
122
pending_parents_stack.append(list(parents))
124
# loop until pending_node_stack is empty
125
while pending_node_stack:
126
parents_to_visit = pending_parents_stack[-1]
127
# if there are no parents left, the revision is done
128
if not parents_to_visit:
129
# append the revision to the topo sorted list
130
# all the nodes parents have been added to the output,
131
# now we can add it to the output.
132
popped_node = pending_node_stack.pop()
133
pending_parents_stack.pop()
134
completed_node_names.add(popped_node)
137
# recurse depth first into a single parent
138
next_node_name = parents_to_visit.pop()
140
if next_node_name in completed_node_names:
141
# parent was already completed by a child, skip it.
143
if next_node_name not in visitable:
144
# parent is not a node in the original graph, skip it.
147
# transfer it along with its parents from the source graph
148
# into the top of the current depth first search stack.
150
parents = graph.pop(next_node_name)
152
# if the next node is not in the source graph it has
153
# already been popped from it and placed into the
154
# current search stack (but not completed or we would
155
# have hit the continue 6 lines up). this indicates a
157
raise errors.GraphCycleError(pending_node_stack)
158
pending_node_stack.append(next_node_name)
159
pending_parents_stack.append(list(parents))
162
def merge_sort(graph, branch_tip, mainline_revisions=None, generate_revno=False):
163
"""Topological sort a graph which groups merges.
165
:param graph: sequence of pairs of node->parents_list.
166
:param branch_tip: the tip of the branch to graph. Revisions not
167
reachable from branch_tip are not included in the
169
:param mainline_revisions: If not None this forces a mainline to be
170
used rather than synthesised from the graph.
171
This must be a valid path through some part
172
of the graph. If the mainline does not cover all
173
the revisions, output stops at the start of the
174
old revision listed in the mainline revisions
176
The order for this parameter is oldest-first.
177
:param generate_revno: Optional parameter controlling the generation of
178
revision number sequences in the output. See the output description of
179
the MergeSorter docstring for details.
180
:result: See the MergeSorter docstring for details.
182
Node identifiers can be any hashable object, and are typically strings.
184
return MergeSorter(graph, branch_tip, mainline_revisions,
185
generate_revno).sorted()
188
class MergeSorter(object):
190
__slots__ = ['_node_name_stack',
191
'_node_merge_depth_stack',
192
'_pending_parents_stack',
193
'_first_child_stack',
194
'_left_subtree_pushed_stack',
197
'_mainline_revisions',
201
'_revno_to_branch_count',
202
'_completed_node_names',
206
def __init__(self, graph, branch_tip, mainline_revisions=None,
207
generate_revno=False):
208
"""Merge-aware topological sorting of a graph.
210
:param graph: sequence of pairs of node_name->parent_names_list.
211
i.e. [('C', ['B']), ('B', ['A']), ('A', [])]
212
For this input the output from the sort or
213
iter_topo_order routines will be:
215
:param branch_tip: the tip of the branch to graph. Revisions not
216
reachable from branch_tip are not included in the
218
:param mainline_revisions: If not None this forces a mainline to be
219
used rather than synthesised from the graph.
220
This must be a valid path through some part
221
of the graph. If the mainline does not cover all
222
the revisions, output stops at the start of the
223
old revision listed in the mainline revisions
225
The order for this parameter is oldest-first.
226
:param generate_revno: Optional parameter controlling the generation of
227
revision number sequences in the output. See the output description
230
The result is a list sorted so that all parents come before
231
their children. Each element of the list is a tuple containing:
232
(sequence_number, node_name, merge_depth, end_of_merge)
233
* sequence_number: The sequence of this row in the output. Useful for
235
* node_name: The node name: opaque text to the merge routine.
236
* merge_depth: How many levels of merging deep this node has been
238
* revno_sequence: When requested this field provides a sequence of
239
revision numbers for all revisions. The format is:
240
(REVNO, BRANCHNUM, BRANCHREVNO). BRANCHNUM is the number of the
241
branch that the revno is on. From left to right the REVNO numbers
242
are the sequence numbers within that branch of the revision.
243
For instance, the graph {A:[], B:['A'], C:['A', 'B']} will get
244
the following revno_sequences assigned: A:(1,), B:(1,1,1), C:(2,).
245
This should be read as 'A is the first commit in the trunk',
246
'B is the first commit on the first branch made from A', 'C is the
247
second commit in the trunk'.
248
* end_of_merge: When True the next node is part of a different merge.
251
node identifiers can be any hashable object, and are typically strings.
253
If you have a graph like [('a', ['b']), ('a', ['c'])] this will only use
254
one of the two values for 'a'.
256
The graph is sorted lazily: until you iterate or sort the input is
257
not processed other than to create an internal representation.
259
iteration or sorting may raise GraphCycleError if a cycle is present
262
Background information on the design:
263
-------------------------------------
264
definition: the end of any cluster or 'merge' occurs when:
265
1 - the next revision has a lower merge depth than we do.
272
C, D are the ends of clusters, E might be but we need more data.
273
2 - or the next revision at our merge depth is not our left most
275
This is required to handle multiple-merges in one commit.
283
C is the end of a cluster due to rule 1.
284
D is not the end of a cluster from rule 1, but is from rule 2: E
285
is not its left most ancestor
286
E is the end of a cluster due to rule 1
287
F might be but we need more data.
289
we show connecting lines to a parent when:
290
- The parent is the start of a merge within this cluster.
291
That is, the merge was not done to the mainline before this cluster
292
was merged to the mainline.
293
This can be detected thus:
294
* The parent has a higher merge depth and is the next revision in
297
The next revision in the list constraint is needed for this case:
299
B 1 [C, F] # we do not want to show a line to F which is depth 2
301
C 1 [H] # note that this is a long line to show back to the
302
ancestor - see the end of merge rules.
308
- Part of this merges 'branch':
309
The parent has the same merge depth and is our left most parent and we
310
are not the end of the cluster.
311
A 0 [C, B] lines: [B, C]
312
B 1 [E, C] lines: [C]
314
D 0 [F, E] lines: [E, F]
317
- The end of this merge/cluster:
318
we can ONLY have multiple parents at the end of a cluster if this
319
branch was previously merged into the 'mainline'.
320
- if we have one and only one parent, show it
321
Note that this may be to a greater merge depth - for instance if
322
this branch continued from a deeply nested branch to add something
324
- if we have more than one parent - show the second oldest (older ==
325
further down the list) parent with
326
an equal or lower merge depth
327
XXXX revisit when awake. ddaa asks about the relevance of each one
328
- maybe more than one parent is relevant
330
self._generate_revno = generate_revno
331
# a dict of the graph.
332
self._graph = dict(graph)
333
# if there is an explicit mainline, alter the graph to match. This is
334
# easier than checking at every merge whether we are on the mainline and
335
# if so which path to take.
336
if mainline_revisions is None:
337
self._mainline_revisions = []
338
self._stop_revision = None
340
self._mainline_revisions = list(mainline_revisions)
341
self._stop_revision = self._mainline_revisions[0]
342
# skip the first revision, its what we reach and its parents are
343
# therefore irrelevant
344
for index, revision in enumerate(self._mainline_revisions[1:]):
345
# NB: index 0 means self._mainline_revisions[1]
346
# if the mainline matches the graph, nothing to do.
347
parent = self._mainline_revisions[index]
349
# end of mainline_revisions history
351
graph_parent_ids = self._graph[revision]
352
if not graph_parent_ids:
353
# We ran into a ghost, skip over it, this is a workaround for
354
# bug #243536, the _graph has had ghosts stripped, but the
355
# mainline_revisions have not
357
if graph_parent_ids[0] == parent:
359
# remove it from its prior spot
360
self._graph[revision].remove(parent)
361
# insert it into the start of the mainline
362
self._graph[revision].insert(0, parent)
363
# we need to do a check late in the process to detect end-of-merges
364
# which requires the parents to be accessible: its easier for now
365
# to just keep the original graph around.
366
self._original_graph = dict(self._graph.items())
367
# we need to know the revision numbers of revisions to determine
368
# the revision numbers of their descendants
369
# this is a graph from node to [revno_tuple, first_child]
370
# where first_child is True if no other children have seen this node
371
# and revno_tuple is the tuple that was assigned to the node.
372
# we dont know revnos to start with, so we start it seeded with
374
self._revnos = dict((revision, [None, True])
375
for revision in self._graph)
376
# Each mainline revision counts how many child branches have spawned from it.
377
self._revno_to_branch_count = {}
379
# this is a stack storing the depth first search into the graph.
380
self._node_name_stack = []
381
# at each level of recursion we need the merge depth this node is at:
382
self._node_merge_depth_stack = []
383
# at each level of 'recursion' we have to check each parent. This
384
# stack stores the parents we have not yet checked for the node at the
385
# matching depth in _node_name_stack
386
self._pending_parents_stack = []
387
# When we first look at a node we assign it a seqence number from its
389
self._first_child_stack = []
390
# this is a set of the nodes who have been completely analysed for fast
391
# membership checking
392
self._completed_node_names = set()
393
# this is the scheduling of nodes list.
394
# Nodes are scheduled
395
# from the bottom left of the tree: in the tree
402
# the scheduling order is: F, E, D, C, B, A
403
# that is - 'left subtree, right subtree, node'
404
# which would mean that when we schedule A we can emit the entire tree.
405
self._scheduled_nodes = []
406
# This records for each node when we have processed its left most
407
# unmerged subtree. After this subtree is scheduled, all other subtrees
408
# have their merge depth increased by one from this nodes merge depth.
409
# it contains tuples - name, merge_depth
410
self._left_subtree_pushed_stack = []
412
# seed the search with the tip of the branch
413
if (branch_tip is not None and
414
branch_tip != _mod_revision.NULL_REVISION and
415
branch_tip != (_mod_revision.NULL_REVISION,)):
416
parents = self._graph.pop(branch_tip)
417
self._push_node(branch_tip, 0, parents)
420
"""Sort the graph and return as a list.
422
After calling this the sorter is empty and you must create a new one.
424
return list(self.iter_topo_order())
426
def iter_topo_order(self):
427
"""Yield the nodes of the graph in a topological order.
429
After finishing iteration the sorter is empty and you cannot continue
432
# These are safe to offload to local variables, because they are used
433
# as a stack and modified in place, never assigned to.
434
node_name_stack = self._node_name_stack
435
node_merge_depth_stack = self._node_merge_depth_stack
436
pending_parents_stack = self._pending_parents_stack
437
left_subtree_pushed_stack = self._left_subtree_pushed_stack
438
completed_node_names = self._completed_node_names
439
scheduled_nodes = self._scheduled_nodes
441
graph_pop = self._graph.pop
443
def push_node(node_name, merge_depth, parents,
444
node_name_stack_append=node_name_stack.append,
445
node_merge_depth_stack_append=node_merge_depth_stack.append,
446
left_subtree_pushed_stack_append=left_subtree_pushed_stack.append,
447
pending_parents_stack_append=pending_parents_stack.append,
448
first_child_stack_append=self._first_child_stack.append,
451
"""Add node_name to the pending node stack.
453
Names in this stack will get emitted into the output as they are popped
456
This inlines a lot of self._variable.append functions as local
459
node_name_stack_append(node_name)
460
node_merge_depth_stack_append(merge_depth)
461
left_subtree_pushed_stack_append(False)
462
pending_parents_stack_append(list(parents))
463
# as we push it, check if it is the first child
466
# node has parents, assign from the left most parent.
468
parent_info = revnos[parents[0]]
470
# Left-hand parent is a ghost, consider it not to exist
472
if parent_info is not None:
473
first_child = parent_info[1]
474
parent_info[1] = False
476
# We don't use the same algorithm here, but we need to keep the
479
first_child_stack_append(first_child)
481
def pop_node(node_name_stack_pop=node_name_stack.pop,
482
node_merge_depth_stack_pop=node_merge_depth_stack.pop,
483
first_child_stack_pop=self._first_child_stack.pop,
484
left_subtree_pushed_stack_pop=left_subtree_pushed_stack.pop,
485
pending_parents_stack_pop=pending_parents_stack.pop,
486
original_graph=self._original_graph,
488
completed_node_names_add=self._completed_node_names.add,
489
scheduled_nodes_append=scheduled_nodes.append,
490
revno_to_branch_count=self._revno_to_branch_count,
492
"""Pop the top node off the stack
494
The node is appended to the sorted output.
496
# we are returning from the flattened call frame:
497
# pop off the local variables
498
node_name = node_name_stack_pop()
499
merge_depth = node_merge_depth_stack_pop()
500
first_child = first_child_stack_pop()
501
# remove this node from the pending lists:
502
left_subtree_pushed_stack_pop()
503
pending_parents_stack_pop()
505
parents = original_graph[node_name]
508
# node has parents, assign from the left most parent.
510
parent_revno = revnos[parents[0]][0]
512
# left-hand parent is a ghost, treat it as not existing
514
if parent_revno is not None:
516
# not the first child, make a new branch
517
base_revno = parent_revno[0]
518
branch_count = revno_to_branch_count.get(base_revno, 0)
520
revno_to_branch_count[base_revno] = branch_count
521
revno = (parent_revno[0], branch_count, 1)
522
# revno = (parent_revno[0], branch_count, parent_revno[-1]+1)
524
# as the first child, we just increase the final revision
526
revno = parent_revno[:-1] + (parent_revno[-1] + 1,)
528
# no parents, use the root sequence
529
root_count = revno_to_branch_count.get(0, -1)
532
revno = (0, root_count, 1)
535
revno_to_branch_count[0] = root_count
537
# store the revno for this node for future reference
538
revnos[node_name][0] = revno
539
completed_node_names_add(node_name)
540
scheduled_nodes_append((node_name, merge_depth, revno))
544
while node_name_stack:
545
# loop until this call completes.
546
parents_to_visit = pending_parents_stack[-1]
547
# if all parents are done, the revision is done
548
if not parents_to_visit:
549
# append the revision to the topo sorted scheduled list:
550
# all the nodes parents have been scheduled added, now
551
# we can add it to the output.
554
while pending_parents_stack[-1]:
555
if not left_subtree_pushed_stack[-1]:
556
# recurse depth first into the primary parent
557
next_node_name = pending_parents_stack[-1].pop(0)
558
is_left_subtree = True
559
left_subtree_pushed_stack[-1] = True
561
# place any merges in right-to-left order for scheduling
562
# which gives us left-to-right order after we reverse
563
# the scheduled queue. XXX: This has the effect of
564
# allocating common-new revisions to the right-most
565
# subtree rather than the left most, which will
566
# display nicely (you get smaller trees at the top
567
# of the combined merge).
568
next_node_name = pending_parents_stack[-1].pop()
569
is_left_subtree = False
570
if next_node_name in completed_node_names:
571
# this parent was completed by a child on the
572
# call stack. skip it.
574
# otherwise transfer it from the source graph into the
575
# top of the current depth first search stack.
577
parents = graph_pop(next_node_name)
579
# if the next node is not in the source graph it has
580
# already been popped from it and placed into the
581
# current search stack (but not completed or we would
582
# have hit the continue 4 lines up.
583
# this indicates a cycle.
584
if next_node_name in self._original_graph:
585
raise errors.GraphCycleError(node_name_stack)
587
# This is just a ghost parent, ignore it
591
# a new child branch from name_stack[-1]
596
node_merge_depth_stack[-1] + next_merge_depth)
601
# and do not continue processing parents until this 'call'
605
# We have scheduled the graph. Now deliver the ordered output:
607
stop_revision = self._stop_revision
608
generate_revno = self._generate_revno
609
original_graph = self._original_graph
611
while scheduled_nodes:
612
node_name, merge_depth, revno = scheduled_nodes.pop()
613
if node_name == stop_revision:
615
if not len(scheduled_nodes):
616
# last revision is the end of a merge
618
elif scheduled_nodes[-1][1] < merge_depth:
619
# the next node is to our left
621
elif (scheduled_nodes[-1][1] == merge_depth and
622
(scheduled_nodes[-1][0] not in
623
original_graph[node_name])):
624
# the next node was part of a multiple-merge.
629
yield (sequence_number, node_name, merge_depth, revno, end_of_merge)
631
yield (sequence_number, node_name, merge_depth, end_of_merge)
634
def _push_node(self, node_name, merge_depth, parents):
635
"""Add node_name to the pending node stack.
637
Names in this stack will get emitted into the output as they are popped
640
self._node_name_stack.append(node_name)
641
self._node_merge_depth_stack.append(merge_depth)
642
self._left_subtree_pushed_stack.append(False)
643
self._pending_parents_stack.append(list(parents))
644
# as we push it, figure out if this is the first child
647
# node has parents, assign from the left most parent.
649
parent_info = self._revnos[parents[0]]
651
# Left-hand parent is a ghost, consider it not to exist
653
if parent_info is not None:
654
first_child = parent_info[1]
655
parent_info[1] = False
657
# We don't use the same algorithm here, but we need to keep the
660
self._first_child_stack.append(first_child)
663
"""Pop the top node off the stack
665
The node is appended to the sorted output.
667
# we are returning from the flattened call frame:
668
# pop off the local variables
669
node_name = self._node_name_stack.pop()
670
merge_depth = self._node_merge_depth_stack.pop()
671
first_child = self._first_child_stack.pop()
672
# remove this node from the pending lists:
673
self._left_subtree_pushed_stack.pop()
674
self._pending_parents_stack.pop()
676
parents = self._original_graph[node_name]
679
# node has parents, assign from the left most parent.
681
parent_revno = self._revnos[parents[0]][0]
683
# left-hand parent is a ghost, treat it as not existing
685
if parent_revno is not None:
687
# not the first child, make a new branch
688
base_revno = parent_revno[0]
689
branch_count = self._revno_to_branch_count.get(base_revno, 0)
691
self._revno_to_branch_count[base_revno] = branch_count
692
revno = (parent_revno[0], branch_count, 1)
693
# revno = (parent_revno[0], branch_count, parent_revno[-1]+1)
695
# as the first child, we just increase the final revision
697
revno = parent_revno[:-1] + (parent_revno[-1] + 1,)
699
# no parents, use the root sequence
700
root_count = self._revno_to_branch_count.get(0, 0)
701
root_count = self._revno_to_branch_count.get(0, -1)
704
revno = (0, root_count, 1)
707
self._revno_to_branch_count[0] = root_count
709
# store the revno for this node for future reference
710
self._revnos[node_name][0] = revno
711
self._completed_node_names.add(node_name)
712
self._scheduled_nodes.append((node_name, merge_depth, self._revnos[node_name][0]))
added
removed
bzrlib/tsort.py
