6
Essential characteristics
7
-------------------------
9
In the general case (no criss-cross), it is a three-way merge. When
10
there is a criss-cross at the tree level, but not for the particular
11
file, it is still a three-way merge. When there's a file-level
12
criss-cross, it's superior to a three-way merge.
17
First, we compare the files we are trying to merge, and find the lines
18
that differ. Next, we try to determine why they differ; this is
19
essential to the merge operation, because it affects how we resolve the
20
differences. In this merger, there are three possible outcomes:
22
1. The line was added in this version: "new-this"
23
2. The line was deleted in the other version: "killed-other"
24
3. The line was preserved as part of merge resolution in this version,
25
but deleted in the other version: "conflicted-this"
27
Option 3 is new, but I believe it is essential. When each side has made
28
a conflicting merge resolution, we should let the user decide how to
29
combine the two resolutions, i.e. we should emit a conflict. We cannot
30
silently drop the line, or silently keep the line, which can happen if
31
we choose options 1 or 2. If we choose options 1 or 2, there's also a
32
possibility that a conflict will be produced, but no guarantee. We need
33
a guarantee, which is why we need a new possible outcome.
35
To decide whether a line is "new-this", "killed-other" or
36
"conflicted-this", we compare this version against the versions from
37
each "least common ancestor" (LCA), in graph terminology. For each LCA
38
version, if the line is not present in the LCA version, we add it to the
39
"new" set. If the line is present in the LCA version, we add it to the
42
When we are done going through each LCA version, each unique line will
43
be in at least one of the sets. If it is only in the "new" set, it's
44
handled as "new-this". If it is only in the "killed" set, it's handled
45
as "killed-other". If it is in both sets, it's handled as
48
The logic here is a bit tricky: first, we know that the line is present
49
in some, but not all, LCAs. We can assume that all LCAs were produced
50
by merges of the same sets of revisions. That means that in those LCAs,
51
there were different merge resolutions. Since THIS and OTHER disagree
52
about whether the line is present, those differences have propagated
53
into THIS and OTHER. Therefore, we should declare that the lines are in
54
conflict, and let the user handle the issue.
56
LCA merge and Three-way merge
57
-----------------------------
59
Now, in the common case, there's a single LCA, and LCA merge behaves as
60
a three-way merge. Since there's only one LCA, we cannot get the
61
"conflicted-this" outcome, only "new-this" or "killed-other. Let's look
62
at the typical description of three-way merges:
64
+-----+------+-------+------------+
65
|THIS | BASE | OTHER | OUT |
66
+-----+------+-------+------------+
68
+-----+------+-------+------------+
70
+-----+------+-------+------------+
72
+-----+------+-------+------------+
74
+-----+------+-------+------------+
75
|A | B | C |\*conflict\*|
76
+-----+------+-------+------------+
78
Now, let's assume that BASE is a common ancestor, as is typically the
79
case. In fact, for best-case merges, BASE is the sole LCA.
81
We always pick the version that represents a change from BASE, if there
82
is one. For the AAAA line, there is no change, so the output is
83
rightfully BASE/THIS/OTHER. For ABAA, the THIS and OTHER are changes
84
from BASE, and they are the same change so they both win. (This case is
85
sometimes called convergence.) For ABBA, THIS is a change from BASE, so
86
THIS wins. For AABB, OTHER is a change from BASE, so OTHER wins. For
87
ABC*, THIS and OTHER are both changes to BASE, but they are different
88
changes, so they can't both win cleanly. Instead, we have a conflict.
90
Now in three-way merging, we typically talk about regions of text. In
91
weave/knit/newness/lca merge, we also have regions. Each contiguous
92
group of "unchanged" lines is a region, and the areas between them are
95
Let's assign a to THIS and b to OTHER. "unchanged" regions represent
96
the AAAA or ABAA cases; it doesn't matter which, because the outcome is
97
the same regardless. Regions which consist of only "new-a" or
98
"killed-a" represent the ABBA case. Regions which consist of only
99
"new-b" or "killed-b" represent the AABB case. Regions which have
100
(new-a or killed-a) AND (new-b or killed-b) are the ABC* case-- both
101
sides have made changes, and they are different changes, so a conflict
104
This is what I mean when I say that it is a three-way merge in the
105
common case; if there is only one LCA, then it is merely an alternative
106
implementation of three-way. (One that happens to automatically do
107
``--reprocess``, ftw).
109
Exception to three-way behavior
110
-------------------------------
111
There is a special case of three-way merge which LCA merge handles differently
112
from our default "merge3" algorithm:
113
BASE has content X, THIS deletes the content, and OTHER changes X to Y. In
114
this case, LCA merge emits Y in its output and does not indicate a conflict.
115
merge3 would output Y, but would also indicate a conflict. (This is also the
116
behavior in the inverse case where OTHER has nothing and THIS has Y.)
118
This behavior is due the way LCA determines basic conflicts; they
119
can only be emitted when THIS and OTHER each have unique lines between common
120
lines. If THIS does not have unique lines in this position, conflicts will not
121
be emitted, even if its (lack of) content is unique.
123
This behavior difference is shared with "weave" merge. I hope that a future
124
revision of LCA merge will handle this case as merge3 would.
129
1. It was time. Although knit / annotate merge and newness merge have
130
tried to emulate the behavior of the original weave merge algorithm,
131
``--merge-type=weave`` hasn't been based on weaves for a long time.
132
2. Behavior differences. This algorithm should behave like a three-way
133
merge in the common case, while its predecessors did not. It also has
134
explicit support for handling conflicting merge resolutions, so it
135
should behave better in criss-cross merge scenarios.
140
Unlike the current "weave" merge implementation, lca merge does not
141
perform any whole-history operations. LCA selection should scale with
142
the number of uncommon revisions. Text comparison time should scale
143
mO(n\ :sup:`2`\ ), where m is the number of LCAs, and n is the number of lines
144
in the file. The current weave merge compares each uncommon ancestor,
145
potentially several times, so it is >= kO(n\ :sup:`2`\ ), where k is the
146
number of uncommon ancestors. So "lca" should beat "weave" both in history
147
analysis time and in text comparison time.
152
1. Inaccurate LCA selection. Our current LCA algorithm uses
153
``Graph.heads()``, which is known to be flawed. It may occasionally give
154
bad results. This risk is mitigated by the fact that the per-file graphs
155
tend to be simpler than the revision graph. And since we're already using
156
this LCA algorithm, this is not an additional risk. I hope that John Meinel
157
will soon have a fixed version of ``Graph.heads`` for us.
158
2. False matches. Weaves have a concept of line identity, but knits and
159
later formats do not. So a line may appear to be common to two files, when
160
in fact it was introduced separately into each for entirely different
161
reasons. This risk is the same for three-way merging. It is mitigated by
162
using Patience sequence matching, which a longest-common-subsequence match.
167
I think this could be a great merge algorithm, and a candidate to make
168
our default, but this work would not have been possible without the work
169
of others, especially:
171
- Martin Pool's weave merge and knit/annotate merge algorithms.
172
- Bram Cohen's discussions of merge algorithms
173
- Andrew Tridgell's dissection of BitKeeper merge
174
- Nathaniel Smith's analysis of why criss-cross histories necessarily
175
produce poor three-way merges.