This is an archive of the discontinued Mercurial Phabricator instance.

Differential D8244

copies: fix the changeset based algorithm regarding merge
ClosedPublic

Authored by marmoute on Mar 5 2020, 1:14 PM.

Details

Reviewers
None
Group Reviewers
hg-reviewers
Commits
rHG45f3f35cefe7: copies: fix the changeset based algorithm regarding merge
Summary

In 99ebde4fec99, we changed the list of files stored into the files field.
This lead to the changeset centric copy algorithm to break in various merge
situation involving merge. Older information could reach the merge through
p1, and while information from p2 was strictly fresher, it would get
overwritten anyway.

We update the situation with more details about which revision introduces rename
information. This help use making the right decision in case of merge.

We are now running a more comprehensive suite of test with include this kind of
situation. The behavior differ slightly from the filelog based in a couple of
instance. There is mostly two distinct cases:

  1. there are conflicting rename information in a merge (different rename history

on each side). In this case the filelog based implementation arbitrarily pick a
side based on the file-revision-number. So it depends on a local factor. The
changeset centric algorithm will use a deterministic approach, by picking the
information coming from the first parent of the merge. This is stable across
different clone.

  1. rename information related to file that exist in both source and destination.

The filelog based implementation do not even try to detect these, however the
changeset centric one get them for "free" (it is simpler to detect them than
not).

The new implementation focus on correctness. Performance improvement will come
later.

Diff Detail

Repository
rHG Mercurial
Branch
default
Lint
No Linters Available
Unit
No Unit Test Coverage

Event Timeline

marmoute created this revision.Mar 5 2020, 1:14 PM
marmoute updated this revision to Diff 20551.Mar 6 2020, 5:21 AM
martinvonz added a subscriber: martinvonz.Mar 6 2020, 2:14 PM

In 99ebde4fec99, we changed the list of files stored into the files field.
This lead to the changeset centric copy algorithm to break in various merge
situation involving merge.

Could you explain why it broke? It's hard to review this patch without really knowing what the problem or the solution is.

The new implementation focus on correctness. Performance improvement will come
later.

How much slower is it? Could you run some of those benchmarks you have run on previous patches touching this code? How do you plan to improve it?

mercurial/copies.py
277

s/had/add/?

tests/test-copies-chain-merge.t
682–683

nit: combine into (no-filelog !)?

marmoute added a comment.Mar 6 2020, 4:48 PM
In D8244#122806, @martinvonz wrote:

In 99ebde4fec99, we changed the list of files stored into the files field.
This lead to the changeset centric copy algorithm to break in various merge
situation involving merge.

Could you explain why it broke? It's hard to review this patch without really knowing what the problem or the solution is.

Outdated information from p1 could overwrite newer information from p2.

The new implementation focus on correctness. Performance improvement will come
later.

How much slower is it? Could you run some of those benchmarks you have run on previous patches touching this code? How do you plan to improve it?

There are two new calls that might degrade performance. This isancestor call that is easy to cache in memory. And the "ismerged" logic that is easy to cache on disk with the rest of the copy related information (the case is rare).

There is some win to have in python, but the main win will be the move the Rust algorithm (that need to be updated with the new logic). Moving to rust give a very large performance boost on the slow cases (usually over 10x, sometime 100x IIRC). That is the one I care about.

marmoute edited the summary of this revision. (Show Details)Mar 6 2020, 6:52 PM
marmoute updated this revision to Diff 20596.

series update after Martin feedback

marmoute marked 2 inline comments as done.Mar 6 2020, 7:48 PM
martinvonz added a comment.Mar 12 2020, 1:49 PM
In D8244#122821, @marmoute wrote:
In D8244#122806, @martinvonz wrote:

In 99ebde4fec99, we changed the list of files stored into the files field.
This lead to the changeset centric copy algorithm to break in various merge
situation involving merge.

Could you explain why it broke? It's hard to review this patch without really knowing what the problem or the solution is.

Outdated information from p1 could overwrite newer information from p2.

The new implementation focus on correctness. Performance improvement will come
later.

How much slower is it? Could you run some of those benchmarks you have run on previous patches touching this code? How do you plan to improve it?

There are two new calls that might degrade performance. This isancestor call that is easy to cache in memory. And the "ismerged" logic that is easy to cache on disk with the rest of the copy related information (the case is rare).
There is some win to have in python, but the main win will be the move the Rust algorithm (that need to be updated with the new logic). Moving to rust give a very large performance boost on the slow cases (usually over 10x, sometime 100x IIRC). That is the one I care about.

I'll make the performance impact more concrete myself. I picked two quite arbitrary tags in the mozilla-unified repo and this is what I saw:

Before this patch:

$ python3 ~/hg/hg perfpathcopies FIREFOX_BETA_44_END FIREFOX_BETA_54_END
! wall 5.279230 comb 5.270000 user 5.250000 sys 0.020000 (best of 3)

After this patch:

$ python3 ~/hg/hg perfpathcopies FIREFOX_BETA_44_END FIREFOX_BETA_54_END
! wall 8.277523 comb 8.280000 user 8.170000 sys 0.110000 (best of 3)

Could you share some more benchmarking data? I know you had a set of commits that you've used before (and that you've asked me to use for benchmarking my patches to copies.py against). It's quite a significant slowdown for the case I tested above, but I'm fine with it since it fixes a bug. I'd just like to see how it behaves in other cases.

mercurial/copies.py
277

Not actually done, it seems, but not a big deal anyway.

marmoute added a comment.Mar 19 2020, 8:44 PM
In D8244#123560, @martinvonz wrote:
In D8244#122821, @marmoute wrote:
In D8244#122806, @martinvonz wrote:

In 99ebde4fec99, we changed the list of files stored into the files field.
This lead to the changeset centric copy algorithm to break in various merge
situation involving merge.

Could you explain why it broke? It's hard to review this patch without really knowing what the problem or the solution is.

Outdated information from p1 could overwrite newer information from p2.

The new implementation focus on correctness. Performance improvement will come
later.

How much slower is it? Could you run some of those benchmarks you have run on previous patches touching this code? How do you plan to improve it?

There are two new calls that might degrade performance. This isancestor call that is easy to cache in memory. And the "ismerged" logic that is easy to cache on disk with the rest of the copy related information (the case is rare).
There is some win to have in python, but the main win will be the move the Rust algorithm (that need to be updated with the new logic). Moving to rust give a very large performance boost on the slow cases (usually over 10x, sometime 100x IIRC). That is the one I care about.

I'll make the performance impact more concrete myself. I picked two quite arbitrary tags in the mozilla-unified repo and this is what I saw:
Before this patch:

$ python3 ~/hg/hg perfpathcopies FIREFOX_BETA_44_END FIREFOX_BETA_54_END
! wall 5.279230 comb 5.270000 user 5.250000 sys 0.020000 (best of 3)

After this patch:

$ python3 ~/hg/hg perfpathcopies FIREFOX_BETA_44_END FIREFOX_BETA_54_END
! wall 8.277523 comb 8.280000 user 8.170000 sys 0.110000 (best of 3)

Could you share some more benchmarking data? I know you had a set of commits that you've used before (and that you've asked me to use for benchmarking my patches to copies.py against). It's quite a significant slowdown for the case I tested above, but I'm fine with it since it fixes a bug. I'd just like to see how it behaves in other cases.

We now have some automatic benchmark setup for copies tracing, however we don't have any reference repositories with the necessary data for changeset centric copy tracing. Building a reference takes some manual operation and a lots of CPU time. So I am planning to build some once the format is more finalized.

I am not too worried about the current performance number because there are multiple easy optimization. In addition the rust version of the previous algorithm proved massively more efficient, so I have good hope for this one too.

The existing reference can still be used to gather various useful pairs of revision to run manual benchmark on.

You can see them using the following command in a setup scmperf repo.

$ grep -A 3 copies repos/*.benchrepo

To setup an scmperrepo, you can use:

$ hg clone https://foss.heptapod.net/mercurial/scmperf
$ cd scmperf
$ ./script/setup-repos default.repos
marmoute updated this revision to Diff 21029.Apr 10 2020, 1:43 PM
marmoute updated this revision to Diff 21180.Apr 22 2020, 12:11 PM
marmoute added a comment.Apr 22 2020, 12:12 PM

Gentle ping on this. This is still ready for review.

marmoute updated this revision to Diff 21181.Apr 22 2020, 1:24 PM
martinvonz added a comment.Apr 22 2020, 1:55 PM
In D8244#126461, @marmoute wrote:

Gentle ping on this. This is still ready for review.

I'm still a bit worried about the performance regression and I was hoping to get more information about that. Now is the beginning of the next cycle, so I guess it's a good time to queue this now and test it on Google users. I'll review this soon.

marmoute added a comment.Apr 22 2020, 2:07 PM
In D8244#126503, @martinvonz wrote:
In D8244#126461, @marmoute wrote:

Gentle ping on this. This is still ready for review.

I'm still a bit worried about the performance regression and I was hoping to get more information about that. Now is the beginning of the next cycle, so I guess it's a good time to queue this now and test it on Google users. I'll review this soon.

I started poking at the performance and [spoiler] and this will be fine :-) Do not hold your breath, because there is a long road between "prototype to get numbers" and "clean implementation". If you can queue this, I would appreciate building on a more solid ground.

For example, on the worst pypy case I have been playing with:

Before the fix: ! wall 1.765737 comb 1.760000 user 1.690000 sys 0.070000 (median of 6)
After the fix:  ! wall 29.983194 comb 29.880000 user 29.780000 sys 0.100000 (median of 3)
With curent state of speedup:   ! wall 2.135641 comb 2.140000 user 2.120000 sys 0.020000 (median of 5)
marmoute added a commit: rHG45f3f35cefe7: copies: fix the changeset based algorithm regarding merge.Apr 28 2020, 2:53 PM
This revision was not accepted when it landed; it landed in state Needs Review.
Closed by commit rHG45f3f35cefe7: copies: fix the changeset based algorithm regarding merge (authored by marmoute). · Explain Why
This revision was automatically updated to reflect the committed changes.

Revision Contents
Changeset List

PathPackages
Mmercurial/copies.py (94 lines)
Mtests/test-copies-chain-merge.t (65 lines)

Diff 20596

mercurial/copies.py

def _revinfogetter(repo): def _revinfogetter(repo):
"""return a function that return multiple data given a <rev>"i """return a function that return multiple data given a <rev>"i
* p1: revision number of first parent * p1: revision number of first parent
* p2: revision number of first parent * p2: revision number of first parent
* p1copies: mapping of copies from p1 * p1copies: mapping of copies from p1
* p2copies: mapping of copies from p2 * p2copies: mapping of copies from p2
* removed: a list of removed files * removed: a list of removed files
* ismerged: a callback to know if file was merged in that revision
""" """
cl = repo.changelog cl = repo.changelog
parents = cl.parentrevs parents = cl.parentrevs
def get_ismerged(rev):
ctx = repo[rev]
def ismerged(path):
if path not in ctx.files():
return False
fctx = ctx[path]
parents = fctx._filelog.parents(fctx._filenode)
nb_parents = 0
for n in parents:
if n != node.nullid:
nb_parents += 1
return nb_parents >= 2
return ismerged
if repo.filecopiesmode == b'changeset-sidedata': if repo.filecopiesmode == b'changeset-sidedata':
changelogrevision = cl.changelogrevision changelogrevision = cl.changelogrevision
flags = cl.flags flags = cl.flags
# A small cache to avoid doing the work twice for merges # A small cache to avoid doing the work twice for merges
# #
# In the vast majority of cases, if we ask information for a revision # In the vast majority of cases, if we ask information for a revision
# about 1 parent, we'll later ask it for the other. So it make sense to # about 1 parent, we'll later ask it for the other. So it make sense to
removed = c.filesremoved removed = c.filesremoved
if p1 != node.nullrev and p2 != node.nullrev: if p1 != node.nullrev and p2 != node.nullrev:
# XXX some case we over cache, IGNORE # XXX some case we over cache, IGNORE
merge_caches[rev] = (p1, p2, p1copies, p2copies, removed) merge_caches[rev] = (p1, p2, p1copies, p2copies, removed)
else: else:
p1copies = {} p1copies = {}
p2copies = {} p2copies = {}
removed = [] removed = []
return p1, p2, p1copies, p2copies, removed
return p1, p2, p1copies, p2copies, removed, get_ismerged(rev)
else: else:
def revinfo(rev): def revinfo(rev):
p1, p2 = parents(rev) p1, p2 = parents(rev)
ctx = repo[rev] ctx = repo[rev]
p1copies, p2copies = ctx._copies p1copies, p2copies = ctx._copies
removed = ctx.filesremoved() removed = ctx.filesremoved()
return p1, p2, p1copies, p2copies, removed return p1, p2, p1copies, p2copies, removed, get_ismerged(rev)
return revinfo return revinfo
def _changesetforwardcopies(a, b, match): def _changesetforwardcopies(a, b, match):
if a.rev() in (node.nullrev, b.rev()): if a.rev() in (node.nullrev, b.rev()):
return {} return {}
repo = a.repo().unfiltered() repo = a.repo().unfiltered()
children = {} children = {}
revinfo = _revinfogetter(repo) revinfo = _revinfogetter(repo)
cl = repo.changelog cl = repo.changelog
isancestor = cl.isancestorrev # XXX we should had chaching to this.
martinvonzUnsubmitted
Done

s/had/add/?

martinvonz: s/had/add/?
martinvonzUnsubmitted
Not Done

Not actually done, it seems, but not a big deal anyway.

martinvonz: Not actually done, it seems, but not a big deal anyway.
missingrevs = cl.findmissingrevs(common=[a.rev()], heads=[b.rev()]) missingrevs = cl.findmissingrevs(common=[a.rev()], heads=[b.rev()])
mrset = set(missingrevs) mrset = set(missingrevs)
roots = set() roots = set()
for r in missingrevs: for r in missingrevs:
for p in cl.parentrevs(r): for p in cl.parentrevs(r):
if p == node.nullrev: if p == node.nullrev:
continue continue
if p not in children: if p not in children:
cl.reachableroots(min_root, [b.rev()], list(roots), includepath=True) cl.reachableroots(min_root, [b.rev()], list(roots), includepath=True)
) )
iterrevs = set(from_head) iterrevs = set(from_head)
iterrevs &= mrset iterrevs &= mrset
iterrevs.update(roots) iterrevs.update(roots)
iterrevs.remove(b.rev()) iterrevs.remove(b.rev())
revs = sorted(iterrevs) revs = sorted(iterrevs)
return _combinechangesetcopies(revs, children, b.rev(), revinfo, match) return _combinechangesetcopies(
revs, children, b.rev(), revinfo, match, isancestor
)
def _combinechangesetcopies(revs, children, targetrev, revinfo, match): def _combinechangesetcopies(
revs, children, targetrev, revinfo, match, isancestor
):
"""combine the copies information for each item of iterrevs """combine the copies information for each item of iterrevs
revs: sorted iterable of revision to visit revs: sorted iterable of revision to visit
children: a {parent: [children]} mapping. children: a {parent: [children]} mapping.
targetrev: the final copies destination revision (not in iterrevs) targetrev: the final copies destination revision (not in iterrevs)
revinfo(rev): a function that return (p1, p2, p1copies, p2copies, removed) revinfo(rev): a function that return (p1, p2, p1copies, p2copies, removed)
match: a matcher match: a matcher
It returns the aggregated copies information for `targetrev`. It returns the aggregated copies information for `targetrev`.
""" """
all_copies = {} all_copies = {}
alwaysmatch = match.always() alwaysmatch = match.always()
for r in revs: for r in revs:
copies = all_copies.pop(r, None) copies = all_copies.pop(r, None)
if copies is None: if copies is None:
# this is a root # this is a root
copies = {} copies = {}
for i, c in enumerate(children[r]): for i, c in enumerate(children[r]):
p1, p2, p1copies, p2copies, removed = revinfo(c) p1, p2, p1copies, p2copies, removed, ismerged = revinfo(c)
if r == p1: if r == p1:
parent = 1 parent = 1
childcopies = p1copies childcopies = p1copies
else: else:
assert r == p2 assert r == p2
parent = 2 parent = 2
childcopies = p2copies childcopies = p2copies
if not alwaysmatch: if not alwaysmatch:
childcopies = { childcopies = {
dst: src for dst, src in childcopies.items() if match(dst) dst: src for dst, src in childcopies.items() if match(dst)
} }
newcopies = copies newcopies = copies
if childcopies: if childcopies:
newcopies = _chain(newcopies, childcopies) newcopies = copies.copy()
# _chain makes a copies, we can avoid doing so in some for dest, source in pycompat.iteritems(childcopies):
# simple/linear cases. prev = copies.get(source)
if prev is not None and prev[1] is not None:
source = prev[1]
newcopies[dest] = (c, source)
assert newcopies is not copies assert newcopies is not copies
for f in removed: for f in removed:
if f in newcopies: if f in newcopies:
if newcopies is copies: if newcopies is copies:
# copy on write to avoid affecting potential other # copy on write to avoid affecting potential other
# branches. when there are no other branches, this # branches. when there are no other branches, this
# could be avoided. # could be avoided.
newcopies = copies.copy() newcopies = copies.copy()
del newcopies[f] newcopies[f] = (c, None)
othercopies = all_copies.get(c) othercopies = all_copies.get(c)
if othercopies is None: if othercopies is None:
all_copies[c] = newcopies all_copies[c] = newcopies
else: else:
# we are the second parent to work on c, we need to merge our # we are the second parent to work on c, we need to merge our
# work with the other. # work with the other.
# #
# Unlike when copies are stored in the filelog, we consider
# it a copy even if the destination already existed on the
# other branch. It's simply too expensive to check if the
# file existed in the manifest.
#
# In case of conflict, parent 1 take precedence over parent 2. # In case of conflict, parent 1 take precedence over parent 2.
# This is an arbitrary choice made anew when implementing # This is an arbitrary choice made anew when implementing
# changeset based copies. It was made without regards with # changeset based copies. It was made without regards with
# potential filelog related behavior. # potential filelog related behavior.
if parent == 1: if parent == 1:
othercopies.update(newcopies) _merge_copies_dict(
othercopies, newcopies, isancestor, ismerged
)
else: else:
newcopies.update(othercopies) _merge_copies_dict(
newcopies, othercopies, isancestor, ismerged
)
all_copies[c] = newcopies all_copies[c] = newcopies
return all_copies[targetrev]
final_copies = {}
for dest, (tt, source) in all_copies[targetrev].items():
if source is not None:
final_copies[dest] = source
return final_copies
def _merge_copies_dict(minor, major, isancestor, ismerged):
"""merge two copies-mapping together, minor and major
In case of conflict, value from "major" will be picked.
- `isancestors(low_rev, high_rev)`: callable return True if `low_rev` is an
ancestors of `high_rev`,
- `ismerged(path)`: callable return True if `path` have been merged in the
current revision,
"""
for dest, value in major.items():
other = minor.get(dest)
if other is None:
minor[dest] = value
else:
new_tt = value[0]
other_tt = other[0]
if value[1] == other[1]:
continue
# content from "major" wins, unless it is older
# than the branch point or there is a merge
if (
new_tt == other_tt
or not isancestor(new_tt, other_tt)
or ismerged(dest)
):
minor[dest] = value
def _forwardcopies(a, b, base=None, match=None): def _forwardcopies(a, b, base=None, match=None):
"""find {dst@b: src@a} copy mapping where a is an ancestor of b""" """find {dst@b: src@a} copy mapping where a is an ancestor of b"""
if base is None: if base is None:
base = a base = a
match = a.repo().narrowmatch(match) match = a.repo().narrowmatch(match)

tests/test-copies-chain-merge.t

#testcases filelog compatibility sidedata
===================================================== =====================================================
Test Copy tracing for chain of copies involving merge Test Copy tracing for chain of copies involving merge
===================================================== =====================================================
This test files covers copies/rename case for a chains of commit where merges This test files covers copies/rename case for a chains of commit where merges
are involved. It cheks we do not have unwanted update of behavior and that the are involved. It cheks we do not have unwanted update of behavior and that the
different options to retrieve copies behave correctly. different options to retrieve copies behave correctly.
Setup Setup
===== =====
use git diff to see rename use git diff to see rename
$ cat << EOF >> $HGRCPATH $ cat << EOF >> $HGRCPATH
> [diff] > [diff]
> git=yes > git=yes
> [ui] > [ui]
> logtemplate={rev} {desc}]\n > logtemplate={rev} {desc}]\n
> EOF > EOF
#if compatibility
$ cat >> $HGRCPATH << EOF
> [experimental]
> copies.read-from = compatibility
> EOF
#endif
#if sidedata
$ cat >> $HGRCPATH << EOF
> [format]
> exp-use-side-data = yes
> exp-use-copies-side-data-changeset = yes
> EOF
#endif
$ hg init repo-chain $ hg init repo-chain
$ cd repo-chain $ cd repo-chain
Add some linear rename initialy Add some linear rename initialy
$ touch a b h $ touch a b h
$ hg ci -Am 'i-0 initial commit: a b h' $ hg ci -Am 'i-0 initial commit: a b h'
adding a adding a
0dd616bc7ab1a111921d95d76f69cda5c2ac539c 644 f 0dd616bc7ab1a111921d95d76f69cda5c2ac539c 644 f
$ hg manifest --debug --rev 'desc("e-2")' | grep '644 f' $ hg manifest --debug --rev 'desc("e-2")' | grep '644 f'
6da5a2eecb9c833f830b67a4972366d49a9a142c 644 f 6da5a2eecb9c833f830b67a4972366d49a9a142c 644 f
$ hg debugindex f $ hg debugindex f
rev linkrev nodeid p1 p2 rev linkrev nodeid p1 p2
0 4 0dd616bc7ab1 000000000000 000000000000 0 4 0dd616bc7ab1 000000000000 000000000000
1 18 6da5a2eecb9c 000000000000 000000000000 1 18 6da5a2eecb9c 000000000000 000000000000
2 19 eb806e34ef6b 0dd616bc7ab1 6da5a2eecb9c 2 19 eb806e34ef6b 0dd616bc7ab1 6da5a2eecb9c
# Here the filelog based implementation is not looking at the rename
# information (because the file exist on both side). However the changelog
# based on works fine. We have different output.
$ hg status --copies --rev 'desc("a-2")' --rev 'desc("mAEm-0")' $ hg status --copies --rev 'desc("a-2")' --rev 'desc("mAEm-0")'
M f M f
b (no-filelog !)
R b R b
$ hg status --copies --rev 'desc("a-2")' --rev 'desc("mEAm-0")' $ hg status --copies --rev 'desc("a-2")' --rev 'desc("mEAm-0")'
M f M f
b (no-filelog !)
R b R b
$ hg status --copies --rev 'desc("e-2")' --rev 'desc("mAEm-0")' $ hg status --copies --rev 'desc("e-2")' --rev 'desc("mAEm-0")'
M f M f
d (no-filelog !)
R d R d
$ hg status --copies --rev 'desc("e-2")' --rev 'desc("mEAm-0")' $ hg status --copies --rev 'desc("e-2")' --rev 'desc("mEAm-0")'
M f M f
d (no-filelog !)
R d R d
$ hg status --copies --rev 'desc("i-2")' --rev 'desc("a-2")' $ hg status --copies --rev 'desc("i-2")' --rev 'desc("a-2")'
A f A f
d d
R d R d
$ hg status --copies --rev 'desc("i-2")' --rev 'desc("e-2")' $ hg status --copies --rev 'desc("i-2")' --rev 'desc("e-2")'
A f A f
b b
R b R b
# From here, we run status against revision where both source file exists.
#
# The filelog based implementation picks an arbitrary side based on revision
# numbers. So the same side "wins" whatever the parents order is. This is
# sub-optimal because depending on revision numbers means the result can be
# different from one repository to the next.
#
# The changeset based algorithm use the parent order to break tie on conflicting
# information and will have a different order depending on who is p1 and p2.
# That order is stable accross repositories. (data from p1 prevails)
$ hg status --copies --rev 'desc("i-2")' --rev 'desc("mAEm-0")' $ hg status --copies --rev 'desc("i-2")' --rev 'desc("mAEm-0")'
A f A f
d d
R b R b
R d R d
$ hg status --copies --rev 'desc("i-2")' --rev 'desc("mEAm-0")' $ hg status --copies --rev 'desc("i-2")' --rev 'desc("mEAm-0")'
A f A f
d d (filelog !)
b (no-filelog !)
R b R b
martinvonzUnsubmitted
Done

nit: combine into (no-filelog !)?

martinvonz: nit: combine into `(no-filelog !)`?
R d R d
$ hg status --copies --rev 'desc("i-0")' --rev 'desc("mAEm-0")' $ hg status --copies --rev 'desc("i-0")' --rev 'desc("mAEm-0")'
A f A f
a a
R a R a
R b R b
$ hg status --copies --rev 'desc("i-0")' --rev 'desc("mEAm-0")' $ hg status --copies --rev 'desc("i-0")' --rev 'desc("mEAm-0")'
A f A f
a a (filelog !)
b (no-filelog !)
R a R a
R b R b
Merge: Merge:
- one with change to an unrelated file (b) - one with change to an unrelated file (b)
- one overwriting a file (d) with a rename (from h to i to d) - one overwriting a file (d) with a rename (from h to i to d)
The overwriting should take over. However, the behavior is currently buggy The overwriting should take over. However, the behavior is currently buggy
$ hg status --copies --rev 'desc("i-0")' --rev 'desc("mBFm-0")' $ hg status --copies --rev 'desc("i-0")' --rev 'desc("mBFm-0")'
M b M b
A d A d
h h
R a R a
R h R h
$ hg status --copies --rev 'desc("i-0")' --rev 'desc("mFBm-0")' $ hg status --copies --rev 'desc("i-0")' --rev 'desc("mFBm-0")'
M b M b
A d A d
h h
R a R a
R h R h
$ hg status --copies --rev 'desc("b-1")' --rev 'desc("mBFm-0")' $ hg status --copies --rev 'desc("b-1")' --rev 'desc("mBFm-0")'
M d M d
h (no-filelog !)
R h R h
$ hg status --copies --rev 'desc("f-2")' --rev 'desc("mBFm-0")' $ hg status --copies --rev 'desc("f-2")' --rev 'desc("mBFm-0")'
M b M b
$ hg status --copies --rev 'desc("f-1")' --rev 'desc("mBFm-0")' $ hg status --copies --rev 'desc("f-1")' --rev 'desc("mBFm-0")'
M b M b
M d M d
i (no-filelog !)
R i R i
$ hg status --copies --rev 'desc("b-1")' --rev 'desc("mFBm-0")' $ hg status --copies --rev 'desc("b-1")' --rev 'desc("mFBm-0")'
M d M d
h (no-filelog !)
R h R h
$ hg status --copies --rev 'desc("f-2")' --rev 'desc("mFBm-0")' $ hg status --copies --rev 'desc("f-2")' --rev 'desc("mFBm-0")'
M b M b
$ hg status --copies --rev 'desc("f-1")' --rev 'desc("mFBm-0")' $ hg status --copies --rev 'desc("f-1")' --rev 'desc("mFBm-0")'
M b M b
M d M d
i (no-filelog !)
R i R i
The following graphlog is wrong, the "a -> c -> d" chain was overwritten and should not appear. The following graphlog is wrong, the "a -> c -> d" chain was overwritten and should not appear.
$ hg log -Gfr 'desc("mBFm-0")' d $ hg log -Gfr 'desc("mBFm-0")' d
o 22 f-2: rename i -> d] o 22 f-2: rename i -> d]
| |
o 21 f-1: rename h -> i] o 21 f-1: rename h -> i]
Merge: Merge:
- one with change to a file - one with change to a file
- one deleting and recreating the file - one deleting and recreating the file
Unlike in the 'BD/DB' cases, an actuall merge happened here. So we should Unlike in the 'BD/DB' cases, an actuall merge happened here. So we should
consider history and rename on both branch of the merge. consider history and rename on both branch of the merge.
One side of the merge have a long history with rename. The other side of the
merge point to a new file with a smaller history. Each side is "valid".
(and again the filelog based algorithm only explore one, with a pick based on
revision numbers)
$ hg status --copies --rev 'desc("i-0")' --rev 'desc("mDGm-0")' $ hg status --copies --rev 'desc("i-0")' --rev 'desc("mDGm-0")'
A d A d
a a (filelog !)
R a R a
$ hg status --copies --rev 'desc("i-0")' --rev 'desc("mGDm-0")' $ hg status --copies --rev 'desc("i-0")' --rev 'desc("mGDm-0")'
A d A d
a a
R a R a
$ hg status --copies --rev 'desc("d-2")' --rev 'desc("mDGm-0")' $ hg status --copies --rev 'desc("d-2")' --rev 'desc("mDGm-0")'
M d M d
$ hg status --copies --rev 'desc("d-2")' --rev 'desc("mGDm-0")' $ hg status --copies --rev 'desc("d-2")' --rev 'desc("mGDm-0")'
Note: Note:
| In this case, the merge get conflicting information since on one side we have | In this case, the merge get conflicting information since on one side we have
| "a -> c -> d". and one the other one we have "h -> i -> d". | "a -> c -> d". and one the other one we have "h -> i -> d".
| |
| The current code arbitrarily pick one side | The current code arbitrarily pick one side
$ hg status --copies --rev 'desc("i-0")' --rev 'desc("mFGm-0")' $ hg status --copies --rev 'desc("i-0")' --rev 'desc("mFGm-0")'
A d A d
a h (no-filelog !)
a (filelog !)
R a R a
R h R h
$ hg status --copies --rev 'desc("i-0")' --rev 'desc("mGFm-0")' $ hg status --copies --rev 'desc("i-0")' --rev 'desc("mGFm-0")'
A d A d
a a
R a R a
R h R h
$ hg status --copies --rev 'desc("f-2")' --rev 'desc("mFGm-0")' $ hg status --copies --rev 'desc("f-2")' --rev 'desc("mFGm-0")'
M d M d
$ hg status --copies --rev 'desc("f-2")' --rev 'desc("mGFm-0")' $ hg status --copies --rev 'desc("f-2")' --rev 'desc("mGFm-0")'
M d M d
$ hg status --copies --rev 'desc("f-1")' --rev 'desc("mFGm-0")' $ hg status --copies --rev 'desc("f-1")' --rev 'desc("mFGm-0")'
M d M d
i (no-filelog !)
R i R i
$ hg status --copies --rev 'desc("f-1")' --rev 'desc("mGFm-0")' $ hg status --copies --rev 'desc("f-1")' --rev 'desc("mGFm-0")'
M d M d
i (no-filelog !)
R i R i
$ hg status --copies --rev 'desc("g-1")' --rev 'desc("mFGm-0")' $ hg status --copies --rev 'desc("g-1")' --rev 'desc("mFGm-0")'
M d M d
h (no-filelog !)
R h R h
$ hg status --copies --rev 'desc("g-1")' --rev 'desc("mGFm-0")' $ hg status --copies --rev 'desc("g-1")' --rev 'desc("mGFm-0")'
M d M d
h (no-filelog !)
R h R h
$ hg log -Gfr 'desc("mFGm-0")' d $ hg log -Gfr 'desc("mFGm-0")' d
o 28 mFGm-0 simple merge - one way] o 28 mFGm-0 simple merge - one way]
|\ |\
| o 25 g-1: update d] | o 25 g-1: update d]
| | | |
o | 22 f-2: rename i -> d] o | 22 f-2: rename i -> d]