Instead of immediately printing ll-merge warnings to stderr, we save
them in our output strbuf. Besides allowing us to move these warnings
to a special file for --remerge-diff, this has two other benefits for
regular merges done by merge-ort:
* The deferral of messages ensures we can print all messages about
any given path together (merge-recursive was known to sometimes
intersperse messages about other paths, particularly when renames
were involved).
* The deferral of messages means we can avoid printing spurious
conflict messages when we just end up aborting due to local user
modifications in the way. (In contrast to merge-recursive.c which
prematurely checks for local modifications in the way via
unpack_trees() and gets the check wrong both in terms of false
positives and false negatives relative to renames, merge-ort does
not perform the local modifications in the way check until the
checkout() step after the full merge has been computed.)
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Since some callers may want to send warning messages to somewhere other
than stdout/stderr, stop printing "warning: Cannot merge binary files"
from ll-merge and instead modify the return status of ll_merge() to
indicate when a merge of binary files has occurred. Message printing
probably does not belong in a "low-level merge" anyway.
This commit continues printing the message as-is, just from the callers
instead of within ll_merge(). Future changes will start handling the
message differently in the merge-ort codepath.
There was one special case here: the callers in rerere.c do NOT check
for and print such a message; since those code paths explicitly skip
over binary files, there is no reason to check for a return status of
LL_MERGE_BINARY_CONFLICT or print the related message.
Note that my methodology included first modifying ll_merge() to return
a struct, so that the compiler would catch all the callers for me and
ensure I had modified all of them. After modifying all of them, I then
changed the struct to an enum.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The documentation for merge_incore_recursive(), modelled after
merge_recursive(), notes that
merge_bases will be consumed (emptied) so make a copy if you need it
However, in merge_ort_internal() (which merge_incore_recursive() calls),
it runs
merged_merge_bases = pop_commit(&merge_bases);
...
for (iter = merge_bases; iter; iter = iter->next) {
...
}
In other words, it only consumes the *first* entry of merge_bases, and
the rest it iterates through. If it iterated through all of them, the
caller could be responsible for free'ing the memory. If it consumed all
of them, the current documentation would be correct and the callers
would need to do nothing. The current middle ground makes it impossible
for callers to avoid memory leaks, since any attempt to use the
merge_bases it passes in would result in a use-after-free.
It turns out this part of the code was copied from merge-recursive.c,
which has had the same bug for 15.5 years. However, since we are trying
to keep merge-recursive.c stable as we sunset it, let's just fix the
leak in in merge_ort_internal() by having it actually consume all the
elements of the merge_bases commit_list.
Testing this commit against t6404 (the first testcase specifically
about recursive merges) under valgrind shows that this patch fixes
the following leak:
32 (16 direct, 16 indirect) bytes in 1 blocks are definitely lost \
in loss record 49 of 126
at 0x484086F: malloc (vg_replace_malloc.c:380)
by 0x69FFEB: do_xmalloc (wrapper.c:41)
by 0x6A0073: xmalloc (wrapper.c:62)
by 0x52A72D: commit_list_insert (commit.c:556)
by 0x47EC86: try_merge_strategy (merge.c:751)
by 0x48143B: cmd_merge (merge.c:1679)
by 0x40686E: run_builtin (git.c:464)
by 0x406C51: handle_builtin (git.c:716)
by 0x406E96: run_argv (git.c:783)
by 0x40730A: cmd_main (git.c:914)
by 0x4E7DFA: main (common-main.c:56)
Reported-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In commit 8b09a900a1 ("merge-ort: restart merge with cached renames to
reduce process entry cost", 2021-07-16), we noted that in the merge-ort
steps of
collect_merge_info()
detect_and_process_renames()
process_entries()
that process_entries() was expensive, and we could often make it cheaper
by changing this to
collect_merge_info()
detect_and_process_renames()
<cache all the renames, and restart>
collect_merge_info()
detect_and_process_renames()
process_entries()
because the second collect_merge_info() would be cheaper (we could avoid
traversing into some directories), the second
detect_and_process_renames() would be free since we had already detected
all renames, and then process_entries() has far fewer entries to handle.
However, this was built on the assumption that the first
detect_and_process_renames() actually detected all potential renames.
If someone has merge.renameLimit set to some small value, that
assumption is violated which manifests later with the following message:
$ git -c merge.renameLimit=1 rebase upstream
...
git: merge-ort.c:546: clear_or_reinit_internal_opts: Assertion
`renames->cached_pairs_valid_side == 0' failed.
Turn off this cache-renames-and-restart whenever we cannot detect all
renames, and add a testcase that would have caught this problem.
Reported-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Elijah Newren <newren@gmail.com>
Tested-by: Taylor Blau <me@ttaylorr.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Ever since commit a492d5331c ("merge-ort: ensure we consult df_conflict
and path_conflicts", 2021-06-30), when renormalization is active AND a
file is involved in a rename/delete conflict BUT the file is unmodified
(either before or after renormalization), merge-ort was running into an
assertion failure. Prior to that commit (or if assertions were compiled
out), merge-ort would mis-merge instead, ignoring the rename/delete
conflict and just deleting the file.
Remove the assertions, fix the code appropriately, leave some good
comments in the code, and add a testcase for this situation.
Reported-by: Ralf Thielow <ralf.thielow@gmail.com>
Signed-off-by: Elijah Newren <newren@gmail.com>
Reviewed-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Follow through the work to use the repo interface to access
submodule objects in-process, instead of abusing the alternate
object database interface.
* jt/no-abuse-alternate-odb-for-submodules:
submodule: trace adding submodule ODB as alternate
submodule: pass repo to check_has_commit()
object-file: only register submodule ODB if needed
merge-{ort,recursive}: remove add_submodule_odb()
refs: peeling non-the_repository iterators is BUG
refs: teach arbitrary repo support to iterators
refs: plumb repo into ref stores
Various fixes in code paths that move untracked files away to make room.
* en/removing-untracked-fixes:
Documentation: call out commands that nuke untracked files/directories
Comment important codepaths regarding nuking untracked files/dirs
unpack-trees: avoid nuking untracked dir in way of locally deleted file
unpack-trees: avoid nuking untracked dir in way of unmerged file
Change unpack_trees' 'reset' flag into an enum
Remove ignored files by default when they are in the way
unpack-trees: make dir an internal-only struct
unpack-trees: introduce preserve_ignored to unpack_trees_options
read-tree, merge-recursive: overwrite ignored files by default
checkout, read-tree: fix leak of unpack_trees_options.dir
t2500: add various tests for nuking untracked files
After the parent commit and some of its ancestors, the only place
commits are being accessed through alternates is in the user-facing
message formatting code. Fix those, and remove the add_submodule_odb()
calls.
Signed-off-by: Jonathan Tan <jonathantanmy@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
More code paths that use the hack to add submodule's object
database to the set of alternate object store have been cleaned up.
* jt/add-submodule-odb-clean-up:
revision: remove "submodule" from opt struct
repository: support unabsorbed in repo_submodule_init
submodule: remove unnecessary unabsorbed fallback
Change several commands to remove ignored files by default when they are
in the way. Since some commands (checkout, merge) take a
--no-overwrite-ignore option to allow the user to configure this, and it
may make sense to add that option to more commands (and in the case of
merge, actually plumb that configuration option through to more of the
backends than just the fast-forwarding special case), add little
comments about where such flags would be used.
Incidentally, this fixes a test failure in t7112.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Currently, every caller of unpack_trees() that wants to ensure ignored
files are overwritten by default needs to:
* allocate unpack_trees_options.dir
* flip the DIR_SHOW_IGNORED flag in unpack_trees_options.dir->flags
* call setup_standard_excludes
AND then after the call to unpack_trees() needs to
* call dir_clear()
* deallocate unpack_trees_options.dir
That's a fair amount of boilerplate, and every caller uses identical
code. Make this easier by instead introducing a new boolean value where
the default value (0) does what we want so that new callers of
unpack_trees() automatically get the appropriate behavior. And move all
the handling of unpack_trees_options.dir into unpack_trees() itself.
While preserve_ignored = 0 is the behavior we feel is the appropriate
default, we defer fixing commands to use the appropriate default until a
later commit. So, this commit introduces several locations where we
manually set preserve_ignored=1. This makes it clear where code paths
were previously preserving ignored files when they should not have been;
a future commit will flip these to instead use a value of 0 to get the
behavior we want.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Various mergy operations have been prepared to work efficiently
with the sparse index.
* ds/mergies-with-sparse-index:
sparse-index: integrate with cherry-pick and rebase
sequencer: ensure full index if not ORT strategy
t1092: add cherry-pick, rebase tests
merge-ort: expand only for out-of-cone conflicts
merge: make sparse-aware with ORT
diff: ignore sparse paths in diffstat
Not sure what happened, but the comment is describing code elsewhere in
the file. Fix the comment to actually discuss the code that follows.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Merge conflicts happen often enough to want to avoid expanding a sparse
index when they happen, as long as those conflicts are within the
sparse-checkout cone. If a conflict exists outside of the
sparse-checkout cone, then we still need to expand before iterating over
the index entries. This is critical to do in advance because of how the
original_cache_nr is tracked to allow inserting and replacing cache
entries.
Iterate over the conflicted files and check if any paths are outside of
the sparse-checkout cone. If so, then expand the full index.
Add a test that demonstrates that we do not expand the index, even when
we hit a conflict within the sparse-checkout cone.
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Reviewed-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Allow 'git merge' to operate without expanding a sparse index, at least
not immediately. The index still will be expanded in a few cases:
1. If the merge strategy is 'recursive', then we enable
command_requires_full_index at the start of the merge_recursive()
method. We expect sparse-index users to also have the 'ort' strategy
enabled.
2. With the 'ort' strategy, if the merge results in a conflicted file,
then we expand the index before updating the working tree. The loop
that iterates over the worktree replaces index entries and tracks
'origintal_cache_nr' which can become completely wrong if the index
expands in the middle of the operation. This safety valve is
important before that loop starts. A later change will focus this
to only expand if we indeed have a conflict outside of the
sparse-checkout cone.
3. Other merge strategies are executed as a 'git merge-X' subcommand,
and those strategies are currently protected with the
'command_requires_full_index' guard.
Some test updates are required, including a mistaken 'git checkout -b'
that did not specify the base branch, causing merges to be fast-forward
merges.
Signed-off-by: Derrick Stolee <dstolee@microsoft.com>
Reviewed-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Clean up a TODO in revision.h by removing the "submodule" field from
struct setup_revision_opt. This field is only used to specify the ref
store to use, so use rev_info->repo to determine the ref store instead.
The only users of this field are merge-ort.c and merge-recursive.c.
However, both these files specify the superproject as rev_info->repo and
the submodule as setup_revision_opt->submodule. In order to be able to
pass the submodule as rev_info->repo, all commits must be parsed with
the submodule explicitly specified; this patch does that as well. (An
incremental solution in which only some commits are parsed with explicit
submodule will not work, because if the same commit is parsed twice in
different repositories, there will be 2 heap-allocated object structs
corresponding to that commit, and any flag set by the revision walking
mechanism on one of them will not be reflected onto the other.)
Signed-off-by: Jonathan Tan <jonathantanmy@google.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Final batch for "merge -sort" optimization.
* en/ort-perf-batch-15:
merge-ort: remove compile-time ability to turn off usage of memory pools
merge-ort: reuse path strings in pool_alloc_filespec
merge-ort: store filepairs and filespecs in our mem_pool
diffcore-rename, merge-ort: add wrapper functions for filepair alloc/dealloc
merge-ort: switch our strmaps over to using memory pools
merge-ort: set up a memory pool
merge-ort: add pool_alloc, pool_calloc, and pool_strndup wrappers
diffcore-rename: use a mem_pool for exact rename detection's hashmap
merge-ort: rename str{map,intmap,set}_func()
Further optimization on "merge -sort" backend.
* en/ort-perf-batch-14:
merge-ort: restart merge with cached renames to reduce process entry cost
merge-ort: avoid recursing into directories when we don't need to
merge-ort: defer recursing into directories when merge base is matched
merge-ort: add a handle_deferred_entries() helper function
merge-ort: add data structures for allowable trivial directory resolves
merge-ort: add some more explanations in collect_merge_info_callback()
merge-ort: resolve paths early when we have sufficient information
Simplify code maintenance by removing the ability to toggle between
usage of memory pools and direct allocations. This allows us to also
remove paths_to_free since it was solely about bookkeeping to make sure
we freed the necessary paths, and allows us to remove some auxiliary
functions.
Suggested-by: Jeff King <peff@peff.net>
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
pool_alloc_filespec() was written so that the code when pool != NULL
mimicked the code from alloc_filespec(), which including allocating
enough extra space for the path and then copying it. However, the path
passed to pool_alloc_filespec() is always going to already be in the
same memory pool, so we may as well reuse it instead of copying it.
For the testcases mentioned in commit 557ac0350d ("merge-ort: begin
performance work; instrument with trace2_region_* calls", 2020-10-28),
this change improves the performance as follows:
Before After
no-renames: 198.5 ms ± 3.4 ms 198.3 ms ± 2.9 ms
mega-renames: 679.1 ms ± 5.6 ms 661.8 ms ± 5.9 ms
just-one-mega: 271.9 ms ± 2.8 ms 264.6 ms ± 2.5 ms
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
For the testcases mentioned in commit 557ac0350d ("merge-ort: begin
performance work; instrument with trace2_region_* calls", 2020-10-28),
this change improves the performance as follows:
Before After
no-renames: 198.1 ms ± 2.6 ms 198.5 ms ± 3.4 ms
mega-renames: 715.8 ms ± 4.0 ms 679.1 ms ± 5.6 ms
just-one-mega: 276.8 ms ± 4.2 ms 271.9 ms ± 2.8 ms
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We want to be able to allocate filespecs and filepairs using a mem_pool.
However, filespec data will still remain outside the pool (perhaps in
the future we could plumb the pool through the various diff APIs to
allocate the filespec data too, but for now we are limiting the scope).
Add some extra functions to allocate these appropriately based on the
non-NULL-ness of opt->priv->pool, as well as some extra functions to
handle correctly deallocating the relevant parts of them. A future
commit will make use of these new functions.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
For all the strmaps (including strintmaps and strsets) whose memory is
unconditionally freed as part of clear_or_reinit_internal_opts(), switch
them over to using our new memory pool.
For the testcases mentioned in commit 557ac0350d ("merge-ort: begin
performance work; instrument with trace2_region_* calls", 2020-10-28),
this change improves the performance as follows:
Before After
no-renames: 202.5 ms ± 3.2 ms 198.1 ms ± 2.6 ms
mega-renames: 1.072 s ± 0.012 s 715.8 ms ± 4.0 ms
just-one-mega: 357.3 ms ± 3.9 ms 276.8 ms ± 4.2 ms
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
merge-ort has a lot of data structures, and they all tend to be freed
together in clear_or_reinit_internal_opts(). Set up a memory pool to
allow us to make these allocations and deallocations faster. Future
commits will adjust various callers to make use of this memory pool.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Make the code more flexible so that it can handle both being run with or
without a memory pool by adding utility functions which will either call
xmalloc, xcalloc, xstrndup
or
mem_pool_alloc, mem_pool_calloc, mem_pool_strndup
depending on whether we have a non-NULL memory pool. A subsequent
commit will make use of these.
(We will actually be dropping these functions soon and just assuming we
always have a memory pool, but the flexibility was very useful during
development of merge-ort so I want to be able to restore it if needed.)
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In order to make it clearer that these three variables holding a
function refer to functions that will clear the strmap/strintmap/strset,
rename them to str{map,intmap,set}_clear_func().
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Documentation on "git diff -l<n>" and diff.renameLimit have been
updated, and the defaults for these limits have been raised.
* en/rename-limits-doc:
rename: bump limit defaults yet again
diffcore-rename: treat a rename_limit of 0 as unlimited
doc: clarify documentation for rename/copy limits
diff: correct warning message when renameLimit exceeded
Many "printf"-like helper functions we have have been annotated
with __attribute__() to catch placeholder/parameter mismatches.
* ab/attribute-format:
advice.h: add missing __attribute__((format)) & fix usage
*.h: add a few missing __attribute__((format))
*.c static functions: add missing __attribute__((format))
sequencer.c: move static function to avoid forward decl
*.c static functions: don't forward-declare __attribute__
The merge algorithm mostly consists of the following three functions:
collect_merge_info()
detect_and_process_renames()
process_entries()
Prior to the trivial directory resolution optimization of the last half
dozen commits, process_entries() was consistently the slowest, followed
by collect_merge_info(), then detect_and_process_renames(). When the
trivial directory resolution applies, it often dramatically decreases
the amount of time spent in the two slower functions.
Looking at the performance results in the previous commit, the trivial
directory resolution optimization helps amazingly well when there are no
relevant renames. It also helps really well when reapplying a long
series of linear commits (such as in a rebase or cherry-pick), since the
relevant renames may well be cached from the first reapplied commit.
But when there are any relevant renames that are not cached (represented
by the just-one-mega testcase), then the optimization does not help at
all.
Often, I noticed that when the optimization does not apply, it is
because there are a handful of relevant sources -- maybe even only one.
It felt frustrating to need to recurse into potentially hundreds or even
thousands of directories just for a single rename, but it was needed for
correctness.
However, staring at this list of functions and noticing that
process_entries() is the most expensive and knowing I could avoid it if
I had cached renames suggested a simple idea: change
collect_merge_info()
detect_and_process_renames()
process_entries()
into
collect_merge_info()
detect_and_process_renames()
<cache all the renames, and restart>
collect_merge_info()
detect_and_process_renames()
process_entries()
This may seem odd and look like more work. However, note that although
we run collect_merge_info() twice, the second time we get to employ
trivial directory resolves, which makes it much faster, so the increased
time in collect_merge_info() is small. While we run
detect_and_process_renames() again, all renames are cached so it's
nearly a no-op (we don't call into diffcore_rename_extended() but we do
have a little bit of data structure checking and fixing up). And the
big payoff comes from the fact that process_entries(), will be much
faster due to having far fewer entries to process.
This restarting only makes sense if we can save recursing into enough
directories to make it worth our while. Introduce a simple heuristic to
guide this. Note that this heuristic uses a "wanted_factor" that I have
virtually no actual real world data for, just some back-of-the-envelope
quasi-scientific calculations that I included in some comments and then
plucked a simple round number out of thin air. It could be that
tweaking this number to make it either higher or lower improves the
optimization. (There's slightly more here; when I first introduced this
optimization, I used a factor of 10, because I was completely confident
it was big enough to not cause slowdowns in special cases. I was
certain it was higher than needed. Several months later, I added the
rough calculations which make me think the optimal number is close to 2;
but instead of pushing to the limit, I just bumped it to 3 to reduce the
risk that there are special cases where this optimization can result in
slowing down the code a little. If the ratio of path counts is below 3,
we probably will only see minor performance improvements at best
anyway.)
Also, note that while the diffstat looks kind of long (nearly 100
lines), more than half of it is in two comments explaining how things
work.
For the testcases mentioned in commit 557ac0350d ("merge-ort: begin
performance work; instrument with trace2_region_* calls", 2020-10-28),
this change improves the performance as follows:
Before After
no-renames: 205.1 ms ± 3.8 ms 204.2 ms ± 3.0 ms
mega-renames: 1.564 s ± 0.010 s 1.076 s ± 0.015 s
just-one-mega: 479.5 ms ± 3.9 ms 364.1 ms ± 7.0 ms
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This combines the work of the last several patches, and implements the
conditions when we don't need to recurse into directories. It's perhaps
easiest to see the logic by separating the fact that a directory might
have both rename sources and rename destinations:
* rename sources: only files present in the merge base can serve as
rename sources, and only when one side deletes that file. When the
tree on one side matches the merge base, that means every file
within the subtree matches the merge base. This means that the
skip-irrelevant-rename-detection optimization from before kicks in
and we don't need any of these files as rename sources.
* rename destinations: the tree that does not match the merge base
might have newly added and hence unmatched destination files.
This is what usually prevents us from doing trivial directory
resolutions in the merge machinery. However, the fact that we have
deferred recursing into this directory until the end means we know
whether there are any unmatched relevant potential rename sources
elsewhere in this merge. If there are no unmatched such relevant
sources anywhere, then there is no need to look for unmatched
potential rename destinations to match them with.
This informs our algorithm:
* Search through relevant_sources; if we have entries, they better all
be reflected in cached_pairs or cached_irrelevant, otherwise they
represent an unmatched potential rename source (causing the
optimization to be disallowed).
* For any relevant_source represented in cached_pairs, we do need to
to make sure to get the destination for each source, meaning we need
to recurse into any ancestor directories of those destinations.
* Once we've recursed into all the rename destinations for any
relevant_sources in cached_pairs, we can then do the trivial
directory resolution for the remaining directories.
For the testcases mentioned in commit 557ac0350d ("merge-ort: begin
performance work; instrument with trace2_region_* calls", 2020-10-28),
this change improves the performance as follows:
Before After
no-renames: 5.235 s ± 0.042 s 205.1 ms ± 3.8 ms
mega-renames: 9.419 s ± 0.107 s 1.564 s ± 0.010 s
just-one-mega: 480.1 ms ± 3.9 ms 479.5 ms ± 3.9 ms
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When one side of history matches the merge base (including when the
merge base has no entry for the given directory), have
collect_merge_info_callback() defer recursing into the directory. To
ensure those entries are eventually handled, add a call to
handled_deferred_entries() in collect_merge_info() after
traverse_trees() returns.
Note that the condition in collect_merge_info_callback() may look more
complicated than necessary at first glance;
renames->trivial_merges_okay[side] is always true until
handle_deferred_entries() is called, and possible_trivial_merges[side]
is always empty right now (and in the future won't be filled until
handle_deferred_entries() is called). However, when
handle_deferred_entries() calls traverse_trees() for the relevant
deferred directories, those traverse_trees() calls will once again end
up in collect_merge_info_callback() for all the entries under those
subdirectories. The extra conditions are there for such deferred cases
and will be used more as we do more with those variables.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In order to allow trivial directory resolution, we first need to be able
to gather more information to determine if the optimization is safe. To
enable that, we need a way of deferring the recursion into the directory
until a later time. Naturally, deferring the entry into a subtree means
that we need some function that will later recurse into the subdirectory
exactly the same way that collect_merge_info_callback() would have done.
Add a helper function that does this. For now this function is not used
but a subsequent commit will change that. Future commits will also make
the function sometimes resolve directories instead of traversing inside.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
As noted a few commits ago, we can resolve individual files early if all
three sides of the merge have a file at the path and two of the three
sides match. We would really like to do the same thing with
directories, because being able to do a trivial directory resolve means
we don't have to recurse into the directory, potentially saving us a
huge amount of time in both collect_merge_info() and process_entries().
Unfortunately, resolving directories early would mean missing any
renames whose source or destination is underneath that directory.
If we somehow knew there weren't any renames under the directory in
question, then we could resolve it early. Sadly, it is impossible to
determine whether there are renames under the directory in question
without recursing into it, and this has traditionally kept us from ever
implementing such an optimization.
In commit f89b4f2bee ("merge-ort: skip rename detection entirely if
possible", 2021-03-11), we added an additional reason that rename
detection could be skipped entirely -- namely, if no *relevant* sources
were present. Without completing collect_merge_info_callback(), we do
not yet know if there are no relevant sources. However, we do know that
if the current directory on one side matches the merge base, then every
source file within that directory will not be RELEVANT_CONTENT, and a
few simple checks can often let us rule out RELEVANT_LOCATION as well.
This suggests we can just defer recursing into such directories until
the end of collect_merge_info.
Since the deferred directories are known to not add any relevant sources
due to the above properties, then if there are no relevant sources after
we've traversed all paths other than the deferred ones, then we know
there are not any relevant sources. Under those conditions, rename
detection is unnecessary, and that means we can resolve the deferred
directories without recursing into them.
Note that the logic for skipping rename detection was also modified
further in commit 76e253793c ("merge-ort, diffcore-rename: employ cached
renames when possible", 2021-01-30); in particular rename detection can
be skipped if we already have cached renames for each relevant source.
We can take advantage of this information as well with our deferral of
recursing into directories where one side matches the merge base.
Add some data structures that we will use to do these deferrals, with
some lengthy comments explaining their purpose.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The previous patch possibly raises some questions about whether
additional cases in collect_merge_info_callback() can be handled early.
Add some explanations in the form of comments to help explain these
better. While we're at it, add a few comments to denote what a few
boolean '0' or '1' values stand for.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When there are no directories involved at a given path, and all three
sides have a file at that path, and two of the three sides of history
match, we can immediately resolve the merge of that path in
collect_merge_info() and do not need to wait until process_entries().
This is actually a very minor improvement: half the time when I run it,
I see an improvement; the other half a slowdown. It seems to be in the
range of noise. However, this idea serves as the beginning of some
bigger optimizations coming in the following patches.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Code cleanup around struct_type_init() functions.
* ab/struct-init:
string-list.h users: change to use *_{nodup,dup}()
string-list.[ch]: add a string_list_init_{nodup,dup}()
dir.[ch]: replace dir_init() with DIR_INIT
*.c *_init(): define in terms of corresponding *_INIT macro
*.h: move some *_INIT to designated initializers
The merge code had funny interactions between content based rename
detection and directory rename detection.
* en/merge-dir-rename-corner-case-fix:
merge-recursive: handle rename-to-self case
merge-ort: ensure we consult df_conflict and path_conflicts
t6423: test directory renames causing rename-to-self
Performance tweaks of "git merge -sort" around lazy fetching of objects.
* en/ort-perf-batch-13:
merge-ort: add prefetching for content merges
diffcore-rename: use a different prefetch for basename comparisons
diffcore-rename: allow different missing_object_cb functions
t6421: add tests checking for excessive object downloads during merge
promisor-remote: output trace2 statistics for number of objects fetched
More fix-ups and optimization to "merge -sort".
* en/ort-perf-batch-12:
merge-ort: miscellaneous touch-ups
Fix various issues found in comments
diffcore-rename: avoid unnecessary strdup'ing in break_idx
merge-ort: replace string_list_df_name_compare with faster alternative
These were last bumped in commit 92c57e5c1d (bump rename limit
defaults (again), 2011-02-19), and were bumped both because processors
had gotten faster, and because people were getting ugly merges that
caused problems and reporting it to the mailing list (suggesting that
folks were willing to spend more time waiting).
Since that time:
* Linus has continued recommending kernel folks to set
diff.renameLimit=0 (maps to 32767, currently)
* Folks with repositories with lots of renames were happy to set
merge.renameLimit above 32767, once the code supported that, to
get correct cherry-picks
* Processors have gotten faster
* It has been discovered that the timing methodology used last time
probably used too large example files.
The last point is probably worth explaining a bit more:
* The "average" file size used appears to have been average blob size
in the linux kernel history at the time (probably v2.6.25 or
something close to it).
* Since bigger files are modified more frequently, such a computation
weights towards larger files.
* Larger files may be more likely to be modified over time, but are
not more likely to be renamed -- the mean and median blob size
within a tree are a bit higher than the mean and median of blob
sizes in the history leading up to that version for the linux
kernel.
* The mean blob size in v2.6.25 was half the average blob size in
history leading to that point
* The median blob size in v2.6.25 was about 40% of the mean blob size
in v2.6.25.
* Since the mean blob size is more than double the median blob size,
any file as big as the mean will not be compared to any files of
median size or less (because they'd be more than 50% dissimilar).
* Since it is the number of files compared that provides the O(n^2)
behavior, median-sized files should matter more than mean-sized
ones.
The combined effect of the above is that the file size used in past
calculations was likely about 5x too large. Combine that with a CPU
performance improvement of ~30%, and we can increase the limits by
a factor of sqrt(5/(1-.3)) = 2.67, while keeping the original stated
time limits.
Keeping the same approximate time limit probably makes sense for
diff.renameLimit (there is no progress feedback in e.g. git log -p),
but the experience above suggests merge.renameLimit could be extended
significantly. In fact, it probably would make sense to have an
unlimited default setting for merge.renameLimit, but that would
likely need to be coupled with changes to how progress is displayed.
(See https://lore.kernel.org/git/YOx+Ok%2FEYvLqRMzJ@coredump.intra.peff.net/
for details in that area.) For now, let's just bump the approximate
time limit from 10s to 1m.
(Note: We do not want to use actual time limits, because getting results
that depend on how loaded your system is that day feels bad, and because
we don't discover that we won't get all the renames until after we've
put in a lot of work rather than just upfront telling the user there are
too many files involved.)
Using the original time limit of 2s for diff.renameLimit, and bumping
merge.renameLimit from 10s to 60s, I found the following timings using
the simple script at the end of this commit message (on an AWS c5.xlarge
which reports as "Intel(R) Xeon(R) Platinum 8124M CPU @ 3.00GHz"):
N Timing
1300 1.995s
7100 59.973s
So let's round down to nice even numbers and bump the limits from
400->1000, and from 1000->7000.
Here is the measure_rename_perf script (adapted from
https://lore.kernel.org/git/20080211113516.GB6344@coredump.intra.peff.net/
in particular to avoid triggering the linear handling from
basename-guided rename detection):
#!/bin/bash
n=$1; shift
rm -rf repo
mkdir repo && cd repo
git init -q -b main
mkdata() {
mkdir $1
for i in `seq 1 $2`; do
(sed "s/^/$i /" <../sample
echo tag: $1
) >$1/$i
done
}
mkdata initial $n
git add .
git commit -q -m initial
mkdata new $n
git add .
cd new
for i in *; do git mv $i $i.renamed; done
cd ..
git rm -q -rf initial
git commit -q -m new
time git diff-tree -M -l0 --summary HEAD^ HEAD
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add missing __attribute__((format)) function attributes to various
"static" functions that take printf arguments.
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Change all in-tree users of the string_list_init(LIST, BOOL) API to
use string_list_init_{nodup,dup}(LIST) instead.
As noted in the preceding commit let's leave the now-unused
string_list_init() wrapper in-place for any in-flight users, it can be
removed at some later date.
Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Path conflicts (typically rename path conflicts, e.g.
rename/rename(1to2) or rename/add/delete), and directory/file conflicts
should obviously result in files not being marked as clean in the merge.
We had a codepath where we missed consulting the path_conflict and
df_conflict flags, based on match_mask. Granted, it requires an unusual
setup to trigger this codepath (directory rename causing rename-to-self
is the only case I can think of), but we still need to handle it. To
make it clear that we have audited the other codepaths that do not
explicitly mention these flags, add some assertions that the flags are
not set.
Reported-by: Anders Kaseorg <andersk@mit.edu>
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Commit 7fbbcb21b1 ("diff: batch fetching of missing blobs", 2019-04-05)
introduced batching of fetching missing blobs, so that the diff
machinery would have one fetch subprocess grab N blobs instead of N
processes each grabbing 1.
However, the diff machinery is not the only thing in a merge that needs
to work on blobs. The 3-way content merges need them as well. Rather
than download all the blobs 1 at a time, prefetch all the blobs needed
for regular content merges.
This does not cover all possible paths in merge-ort that might need to
download blobs. Others include:
- The blob_unchanged() calls to avoid modify/delete conflicts (when
blob renormalization results in an "unchanged" file)
- Preliminary content merges needed for rename/add and
rename/rename(2to1) style conflicts. (Both of these types of
conflicts can result in nested conflict markers from the need to do
two levels of content merging; the first happens before our new
prefetch_for_content_merges() function.)
The first of these wouldn't be an extreme amount of work to support, and
even the second could be theoretically supported in batching, but all of
these cases seem unusual to me, and this is a minor performance
optimization anyway; in the worst case we only get some of the fetches
batched and have a few additional one-off fetches. So for now, just
handle the regular 3-way content merges in our prefetching.
For the testcase from the previous commit, the number of downloaded
objects remains at 63, but this drops the number of fetches needed from
32 down to 20, a sizeable reduction.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Optimize out repeated rename detection in a sequence of mergy
operations.
* en/ort-perf-batch-11:
merge-ort, diffcore-rename: employ cached renames when possible
merge-ort: handle interactions of caching and rename/rename(1to1) cases
merge-ort: add helper functions for using cached renames
merge-ort: preserve cached renames for the appropriate side
merge-ort: avoid accidental API mis-use
merge-ort: add code to check for whether cached renames can be reused
merge-ort: populate caches of rename detection results
merge-ort: add data structures for in-memory caching of rename detection
t6429: testcases for remembering renames
fast-rebase: write conflict state to working tree, index, and HEAD
fast-rebase: change assert() to BUG()
Documentation/technical: describe remembering renames optimization
t6423: rename file within directory that other side renamed
Add some notes in the code about invariants with match_mask when adding
pairs. Also add a comment that seems to have been left out in my work
of pushing these changes upstream.
Signed-off-by: Elijah Newren <newren@gmail.com>
Reviewed-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
A random hodge-podge of incorrect or out-of-date comments that I found:
* t6423 had a comment that has referred to the wrong test for years;
fix it to refer to the right one.
* diffcore-rename had a FIXME comment meant to remind myself to
investigate if I could make another code change. I later
investigated and removed the FIXME, but while cherry-picking the
patch to submit upstream I missed the later update. Remove the
comment now.
* merge-ort had the early part of a comment for a function; I had
meant to include the more involved description when I updated the
function. Update the comment now.
Signed-off-by: Elijah Newren <newren@gmail.com>
Reviewed-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Gathering accumulated times from trace2 output on the mega-renames
testcase, I saw the following timings (where I'm only showing a few
lines to highlight the portions of interest):
10.120 : label:incore_nonrecursive
4.462 : ..label:process_entries
3.143 : ....label:process_entries setup
2.988 : ......label:plist special sort
1.305 : ....label:processing
2.604 : ..label:collect_merge_info
2.018 : ..label:merge_start
1.018 : ..label:renames
In the above output, note that the 4.462 seconds for process_entries was
split as 3.143 seconds for "process_entries setup" and 1.305 seconds for
"processing" (and a little time for other stuff removed from the
highlight). Most of the "process_entries setup" time was spent on
"plist special sort" which corresponds to the following code:
trace2_region_enter("merge", "plist special sort", opt->repo);
plist.cmp = string_list_df_name_compare;
string_list_sort(&plist);
trace2_region_leave("merge", "plist special sort", opt->repo);
In other words, in a merge strategy that would be invoked by passing
"-sort" to either rebase or merge, sorting an array takes more time than
anything else. Serves me right for naming my merge strategy this way.
Rewrite the comparison function in a way that does not require finding
out the lengths of the strings when comparing them. While at it, tweak
the code for our specific case -- no need to handle a variety of modes,
for example. The combination of these changes reduced the time spent in
"plist special sort" by ~25% in the mega-renames case.
For the testcases mentioned in commit 557ac0350d ("merge-ort: begin
performance work; instrument with trace2_region_* calls", 2020-10-28),
this change improves the performance as follows:
Before After
no-renames: 5.622 s ± 0.059 s 5.235 s ± 0.042 s
mega-renames: 10.127 s ± 0.073 s 9.419 s ± 0.107 s
just-one-mega: 500.3 ms ± 3.8 ms 480.1 ms ± 3.9 ms
Signed-off-by: Elijah Newren <newren@gmail.com>
Reviewed-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When there are many renames between the old base of a series of commits
and the new base, the way sequencer.c, merge-recursive.c, and
diffcore-rename.c have traditionally split the work resulted in
redetecting the same renames with each and every commit being
transplanted. To address this, the last several commits have been
creating a cache of rename detection results, determining when it was
safe to use such a cache in subsequent merge operations, adding helper
functions, and so on. See the previous half dozen commit messages for
additional discussion of this optimization, particularly the message a
few commits ago entitled "add code to check for whether cached renames
can be reused". This commit finally ties all of that work together,
modifying the merge algorithm to make use of these cached renames.
For the testcases mentioned in commit 557ac0350d ("merge-ort: begin
performance work; instrument with trace2_region_* calls", 2020-10-28),
this change improves the performance as follows:
Before After
no-renames: 5.665 s ± 0.129 s 5.622 s ± 0.059 s
mega-renames: 11.435 s ± 0.158 s 10.127 s ± 0.073 s
just-one-mega: 494.2 ms ± 6.1 ms 500.3 ms ± 3.8 ms
That's a fairly small improvement, but mostly because the previous
optimizations were so effective for these particular testcases; this
optimization only kicks in when the others don't. If we undid the
basename-guided rename detection and skip-irrelevant-renames
optimizations, then we'd see that this series by itself improved
performance as follows:
Before Basename Series After Just This Series
no-renames: 13.815 s ± 0.062 s 5.697 s ± 0.080 s
mega-renames: 1799.937 s ± 0.493 s 205.709 s ± 0.457 s
Since this optimization kicks in to help accelerate cases where the
previous optimizations do not apply, this last comparison shows that
this cached-renames optimization has the potential to help signficantly
in cases that don't meet the requirements for the other optimizations to
be effective.
The changes made in this optimization also lay some important groundwork
for a future optimization around having collect_merge_info() avoid
recursing into subtrees in more cases.
However, for this optimization to be effective, merge_switch_to_result()
should only be called when the rebase or cherry-pick operation has
either completed or hit a case where the user needs to resolve a
conflict or edit the result. If it is called after every commit, as
sequencer.c does, then the working tree and index are needlessly updated
with every commit and the cached metadata is tossed, defeating this
optimization. Refactoring sequencer.c to only call
merge_switch_to_result() at the end of the operation is a bigger
undertaking, and the practical benefits of this optimization will not be
realized until that work is performed. Since `test-tool fast-rebase`
only updates at the end of the operation, it was used to obtain the
timings above.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
As documented in Documentation/technical/remembering-renames.txt, and as
tested for in the two testcases in t6429 with "rename same file
identically" in their description, there is one case where we need to
have renames in one commit NOT be cached for the next commit in our
rebase sequence -- namely, rename/rename(1to1) cases. Rather than
specifically trying to uncache those and fix up dir_rename_counts() to
match (which would also be valid but more work), we simply disable the
optimization when this really rare type of rename occurs.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
If we have a usable rename cache, then we can remove from
relevant_sources all the paths that were cached;
diffcore_rename_extended() can then consider an even smaller set of
relevant_sources in its rename detection.
However, when diffcore_rename_extended() is done, we will need to take
the renames it detected and then add back in all the ones we had cached
from before.
Add helper functions for doing these two operations; the next commit
will make use of them.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Previous commits created an in-memory cache of the results of rename
detection, and added logic to detect when that cache could appropriately
be used in a subsequent merge operation -- but we were still
unconditionally clearing the cache with each new merge operation anyway.
If it is valid to reuse the cache from one of the two sides of history,
preserve that side.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Previously, callers of the merge-ort API could have passed an
uninitialized value for struct merge_result *result. However, we want
to check result to see if it has cached renames from a previous merge
that we can reuse; such values would be found behind result->priv.
However, if result->priv is uninitialized, attempting to access behind
it will give a segfault. So, we need result->priv to be NULL (which
will be the case if the caller does a memset(&result, 0)), or be written
by a previous call to the merge-ort machinery. Documenting this
requirement may help, but despite being the person who introduced this
requirement, I still missed it once and it did not fail in a very clear
way and led to a long debugging session.
Add a _properly_initialized field to merge_result; that value will be
0 if the caller zero'ed the merge_result, it will be set to a very
specific value by a previous run by the merge-ort machinery, and if it's
uninitialized it will most likely either be 0 or some value that does
not match the specific one we'd expect allowing us to throw a much more
meaningful error.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We need to know when renames detected in a previous merge operation can
be reused in a later merge operation. Consider the following setup
(from the git-rebase manpage):
A---B---C topic
/
D---E---F---G master
After rebasing, this will appear as:
A'--B'--C' topic
/
D---E---F---G master
Further, let's say that 'oldfile' was renamed to 'newfile' between E
and G. The rebase or cherry-pick of A onto G will involve a three-way
merge between E (as the merge base) and G and A. After detecting the
rename between E:oldfile and G:newfile, there will be a three-way
content merge of the following:
E:oldfile
G:newfile
A:oldfile
and produce a new result:
A':newfile
Now, when we want to pick B onto A', we will need to do a three-way
merge between A (as the merge-base) and A' and B. This will involve
a three-way content merge of
A:oldfile
A':newfile
B:oldfile
but only if we can detect that A:oldfile is similar enough to A':newfile
to be used together in a three-way content merge, i.e. only if we can
detect that A:oldfile and A':newfile are a rename. But we already know
that A:oldfile and A':newfile are similar enough to be used in a
three-way content merge, because that is precisely where A':newfile came
from in the previous merge.
Note that A & A' both appear in both merges. That gives us the
condition under which we can reuse renames.
There are a couple important points about this optimization:
- If the rebase or cherry-pick halts for user conflicts, these caches
are NOT saved anywhere. Thus, resuming a halted rebase or
cherry-pick will result in no reused renames for the next commit.
This is intentional, as user resolution can change files
significantly and in ways that violate the similarity assumptions
here.
- Technically, in a *very* narrow case this might give slightly
different results for rename detection. Using the example above,
if:
* E:oldfile had 20 lines
* G:newfile added 10 new lines at the beginning of the file
* A:oldfile deleted all but the first three lines of the file
then
=> A':newfile would have 13 lines, 3 of which matches those
in A:oldfile.
Consider the two cases:
* Without this optimization:
- the next step of the rebase operation (moving B to B')
would not detect the rename betwen A:oldfile and A':newfile
- we'd thus get a modify/delete conflict with the rebase
operation halting for the user to resolve, and have both
A':newfile and B:oldfile sitting in the working tree.
* With this optimization:
- the rename between A:oldfile and A':newfile would be detected
via the cache of renames
- a three-way merge between A:oldfile, A':newfile, and B:oldfile
would commence and be written to A':newfile
Now, is the difference in behavior a bug...or a bugfix? I can't
tell. Given that A:oldfile and A':newfile are not very similar,
when we three-way merge with B:oldfile it seems likely we'll hit a
conflict for the user to resolve. And it shouldn't be too hard for
users to see why we did that three-way merge; oldfile and newfile
*were* renames somewhere in the sequence. So, most of these corner
cases will still behave similarly -- namely, a conflict given to the
user to resolve. Also, consider the interesting case when commit B
is a clean revert of commit A. Without this optimization, a rebase
could not both apply a weird patch like A and then immediately
revert it; users would be forced to resolve merge conflicts. With
this optimization, it would successfully apply the clean revert.
So, there is certainly at least one case that behaves better. Even
if it's considered a "difference in behavior", I think both behaviors
are reasonable, and the time savings provided by this optimization
justify using the slightly altered rename heuristics.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Fill in cache_pairs, cached_target_names, and cached_irrelevant based on
rename detection results. Future commits will make use of these values.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When there are many renames between the old base of a series of commits
and the new base for a series of commits, the sequence of merges
employed to transplant those commits (from a cherry-pick or rebase
operation) will repeatedly detect the exact same renames. This is
wasted effort.
Add data structures which will be used to cache rename detection
results, along with the initialization and deallocation of these data
structures. Future commits will populate these caches, detect the
appropriate circumstances when they can be used, and employ them to
avoid re-detecting the same renames repeatedly.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The word "renamed" has two possible translations in many European
languages depending on whether one thing was renamed or two things were
renamed. Give translators freedom to alter any part of the message to
make it sound right in their language.
Signed-off-by: Alex Henrie <alexhenrie24@gmail.com>
Acked-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
SHA-256 transition.
* bc/hash-transition-interop-part-1:
hex: print objects using the hash algorithm member
hex: default to the_hash_algo on zero algorithm value
builtin/pack-objects: avoid using struct object_id for pack hash
commit-graph: don't store file hashes as struct object_id
builtin/show-index: set the algorithm for object IDs
hash: provide per-algorithm null OIDs
hash: set, copy, and use algo field in struct object_id
builtin/pack-redundant: avoid casting buffers to struct object_id
Use the final_oid_fn to finalize hashing of object IDs
hash: add a function to finalize object IDs
http-push: set algorithm when reading object ID
Always use oidread to read into struct object_id
hash: add an algo member to struct object_id
Builds on top of the sparse-index infrastructure to mark operations
that are not ready to mark with the sparse index, causing them to
fall back on fully-populated index that they always have worked with.
* ds/sparse-index-protections: (47 commits)
name-hash: use expand_to_path()
sparse-index: expand_to_path()
name-hash: don't add directories to name_hash
revision: ensure full index
resolve-undo: ensure full index
read-cache: ensure full index
pathspec: ensure full index
merge-recursive: ensure full index
entry: ensure full index
dir: ensure full index
update-index: ensure full index
stash: ensure full index
rm: ensure full index
merge-index: ensure full index
ls-files: ensure full index
grep: ensure full index
fsck: ensure full index
difftool: ensure full index
commit: ensure full index
checkout: ensure full index
...
Up until recently, object IDs did not have an algorithm member, only a
hash. Consequently, it was possible to share one null (all-zeros)
object ID among all hash algorithms. Now that we're going to be
handling objects from multiple hash algorithms, it's important to make
sure that all object IDs have a correct algorithm field.
Introduce a per-algorithm null OID, and add it to struct hash_algo.
Introduce a wrapper function as well, and use it everywhere we used to
use the null_oid constant.
Signed-off-by: brian m. carlson <sandals@crustytoothpaste.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Plug the ort merge backend throughout the rest of the system, and
start testing it as a replacement for the recursive backend.
* en/ort-readiness:
Add testing with merge-ort merge strategy
t6423: mark remaining expected failure under merge-ort as such
Revert "merge-ort: ignore the directory rename split conflict for now"
merge-recursive: add a bunch of FIXME comments documenting known bugs
merge-ort: write $GIT_DIR/AUTO_MERGE whenever we hit a conflict
t: mark several submodule merging tests as fixed under merge-ort
merge-ort: implement CE_SKIP_WORKTREE handling with conflicted entries
t6428: new test for SKIP_WORKTREE handling and conflicts
merge-ort: support subtree shifting
merge-ort: let renormalization change modify/delete into clean delete
merge-ort: have ll_merge() use a special attr_index for renormalization
merge-ort: add a special minimal index just for renormalization
merge-ort: use STABLE_QSORT instead of QSORT where required
Various rename detection optimization to help "ort" merge strategy
backend.
* en/ort-perf-batch-10:
diffcore-rename: determine which relevant_sources are no longer relevant
merge-ort: record the reason that we want a rename for a file
diffcore-rename: add computation of number of unknown renames
diffcore-rename: check if we have enough renames for directories early on
diffcore-rename: only compute dir_rename_count for relevant directories
merge-ort: record the reason that we want a rename for a directory
merge-ort, diffcore-rename: tweak dirs_removed and relevant_source type
diffcore-rename: take advantage of "majority rules" to skip more renames
versions could be an empty string_list. In that case, versions->items is
NULL, and we shouldn't be trying to perform pointer arithmetic with it (as
that results in undefined behaviour).
Moreover we only use the results of this calculation once when calling
QSORT. Therefore we choose to skip creating relevant_entries and call
QSORT directly with our manipulated pointers (but only if there's data
requiring sorting). This lets us avoid abusing the string_list API,
and saves us from having to explain why this abuse is OK.
Finally, an assertion is added to make sure that write_tree() is called
with a valid offset.
This issue has probably existed since:
ee4012dcf9 (merge-ort: step 2 of tree writing -- function to create tree object, 2020-12-13)
But it only started occurring during tests since tests started using
merge-ort:
f3b964a07e (Add testing with merge-ort merge strategy, 2021-03-20)
For reference - here's the original UBSAN commit that implemented this
check, it sounds like this behaviour isn't actually likely to cause any
issues (but we might as well fix it regardless):
https://reviews.llvm.org/D67122
UBSAN output from t3404 or t5601:
merge-ort.c:2669:43: runtime error: applying zero offset to null pointer
#0 0x78bb53 in write_tree merge-ort.c:2669:43
#1 0x7856c9 in process_entries merge-ort.c:3303:2
#2 0x782317 in merge_ort_nonrecursive_internal merge-ort.c:3744:2
#3 0x77feef in merge_incore_nonrecursive merge-ort.c:3853:2
#4 0x6f6a5c in do_recursive_merge sequencer.c:640:3
#5 0x6f6a5c in do_pick_commit sequencer.c:2221:9
#6 0x6ef055 in single_pick sequencer.c:4814:9
#7 0x6ef055 in sequencer_pick_revisions sequencer.c:4867:10
#8 0x4fb392 in run_sequencer revert.c:225:9
#9 0x4fa5b0 in cmd_revert revert.c:235:8
#10 0x42abd7 in run_builtin git.c:453:11
#11 0x429531 in handle_builtin git.c:704:3
#12 0x4282fb in run_argv git.c:771:4
#13 0x4282fb in cmd_main git.c:902:19
#14 0x524b63 in main common-main.c:52:11
#15 0x7fc2ca340349 in __libc_start_main (/lib64/libc.so.6+0x24349)
#16 0x4072b9 in _start start.S:120
SUMMARY: UndefinedBehaviorSanitizer: undefined-behavior merge-ort.c:2669:43 in
Signed-off-by: Andrzej Hunt <ajrhunt@google.com>
Reviewed-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The ort merge backend has been optimized by skipping irrelevant
renames.
* en/ort-perf-batch-9:
diffcore-rename: avoid doing basename comparisons for irrelevant sources
merge-ort: skip rename detection entirely if possible
merge-ort: use relevant_sources to filter possible rename sources
merge-ort: precompute whether directory rename detection is needed
merge-ort: introduce wrappers for alternate tree traversal
merge-ort: add data structures for an alternate tree traversal
merge-ort: precompute subset of sources for which we need rename detection
diffcore-rename: enable filtering possible rename sources
Rename detection rework continues.
* en/ort-perf-batch-8:
diffcore-rename: compute dir_rename_guess from dir_rename_counts
diffcore-rename: limit dir_rename_counts computation to relevant dirs
diffcore-rename: compute dir_rename_counts in stages
diffcore-rename: extend cleanup_dir_rename_info()
diffcore-rename: move dir_rename_counts into dir_rename_info struct
diffcore-rename: add function for clearing dir_rename_count
Move computation of dir_rename_count from merge-ort to diffcore-rename
diffcore-rename: add a mapping of destination names to their indices
diffcore-rename: provide basic implementation of idx_possible_rename()
diffcore-rename: use directory rename guided basename comparisons
This reverts commit 5ced7c3da0, which was
put in place as a temporary measure to avoid optimizations unstably
erroring out on no destination having a majority of the necessary
renames for directories that had no new files and thus no need for
directory rename detection anyway. Now that optimizations are in place
to prevent us from trying to compute directory rename count computations
for directories that do not need it, we can undo this temporary measure.
Signed-off-by: Elijah Newren <newren@gmail.com>
Reviewed-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
There are a variety of questions users might ask while resolving
conflicts:
* What changes have been made since the previous (first) parent?
* What changes are staged?
* What is still unstaged? (or what is still conflicted?)
* What changes did I make to resolve conflicts so far?
The first three of these have simple answers:
* git diff HEAD
* git diff --cached
* git diff
There was no way to answer the final question previously. Adding one
is trivial in merge-ort, since it works by creating a tree representing
what should be written to the working copy complete with conflict
markers. Simply write that tree to .git/AUTO_MERGE, allowing users to
answer the fourth question with
* git diff AUTO_MERGE
I avoided using a name like "MERGE_AUTO", because that would be
merge-specific (much like MERGE_HEAD, REBASE_HEAD, REVERT_HEAD,
CHERRY_PICK_HEAD) and I wanted a name that didn't change depending on
which type of operation the merge was part of.
Ensure that paths which clean out other temporary operation-specific
files (e.g. CHERRY_PICK_HEAD, MERGE_MSG, rebase-merge/ state directory)
also clean out this AUTO_MERGE file.
Signed-off-by: Elijah Newren <newren@gmail.com>
Reviewed-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When merge conflicts occur in paths removed by a sparse-checkout, we
need to unsparsify those paths (clear the SKIP_WORKTREE bit), and write
out the conflicted file to the working copy. In the very unlikely case
that someone manually put a file into the working copy at the location
of the SKIP_WORKTREE file, we need to avoid overwriting whatever edits
they have made and move that file to a different location first.
Signed-off-by: Elijah Newren <newren@gmail.com>
Reviewed-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
merge-recursive has some simple code to support subtree shifting; copy
it over to merge-ort. This fixes t6409.12 under
GIT_TEST_MERGE_ALGORITHM=ort.
Signed-off-by: Elijah Newren <newren@gmail.com>
Reviewed-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When we have a modify/delete conflict, but the only change to the
modification is e.g. change of line endings, then if renormalization is
requested then we should be able to recognize such a case as a
not-modified/delete and resolve the conflict automatically.
This fixes t6418.10 under GIT_TEST_MERGE_ALGORITHM=ort.
Signed-off-by: Elijah Newren <newren@gmail.com>
Reviewed-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
ll_merge() needs an index when renormalization is requested. Create one
specifically for just this purpose with just the one needed entry. This
fixes t6418.4 and t6418.5 under GIT_TEST_MERGE_ALGORITHM=ort.
NOTE 1: Even if the user has a working copy or a real index (which is
not a given as merge-ort can be used in bare repositories), we
explicitly ignore any .gitattributes file from either of these
locations. merge-ort can be used to merge two branches that are
unrelated to HEAD, so .gitattributes from the working copy and current
index should not be considered relevant.
NOTE 2: Since we are in the middle of merging, there is a risk that
.gitattributes itself is conflicted...leaving us with an ill-defined
situation about how to perform the rest of the merge. It could be that
the .gitattributes file does not even exist on one of the sides of the
merge, or that it has been modified on both sides. If it's been
modified on both sides, it's possible that it could itself be merged
cleanly, though it's also possible that it only merges cleanly if you
use the right version of the .gitattributes file to drive the merge. It
gets kind of complicated. The only test we ever had that attempted to
test behavior in this area was seemingly unaware of the undefined
behavior, but knew the test wouldn't work for lack of attribute handling
support, marked it as test_expect_failure from the beginning, but
managed to fail for several reasons unrelated to attribute handling.
See commit 6f6e7cfb52 ("t6038: remove problematic test", 2020-08-03) for
details. So there are probably various ways to improve what
initialize_attr_index() picks in the case of a conflicted .gitattributes
but for now I just implemented something simple -- look for whatever
.gitattributes file we can find in any of the higher order stages and
use it.
Signed-off-by: Elijah Newren <newren@gmail.com>
Reviewed-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
renormalize_buffer() requires an index_state, which is something that
merge-ort does not operate with. However, all the renormalization code
needs is an index with a .gitattributes file...plus a little bit of
setup. Create such an index, along with the deallocation and
attr_direction handling.
A subsequent commit will add a function to finish the initialization
of this index.
Signed-off-by: Elijah Newren <newren@gmail.com>
Reviewed-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
rename/rename conflict handling depends on the fact that if both sides
renamed the same path, that the one on the MERGE_SIDE1 will appear first
in the combined diff_queue_struct passed to process_renames(). Since we
add all pairs from MERGE_SIDE1 to combined first, and then all pairs
from MERGE_SIDE2, and then sort based on filename, this will only be
true if the sort used is stable. This was found due to the fact that
Mac, unlike Linux, apparently has a system-defined qsort that is not
stable.
While we are at it, review the other callers of QSORT and add comments
about why they can remain as calls to QSORT instead of being modified
to call STABLE_QSORT.
Signed-off-by: Elijah Newren <newren@gmail.com>
Reviewed-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
There are two different reasons we might want a rename for a file -- for
three-way content merging or as part of directory rename detection.
Record the reason. diffcore-rename will potentially be able to filter
some of the ones marked as needed only for directory rename detection,
if it can determine those directory renames based solely on renames
found via exact rename detection and basename-guided rename detection.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The previous commit can only be effective if we have a computation of
the number of paths under a given directory which are still have pending
renames, and expected this number to be recorded in the dir_rename_count
map under the key UNKNOWN_DIR. Add the code necessary to compute these
values.
Note that this change means dir_rename_count might have a directory
whose only entry (for UNKNOWN_DIR) was removed by the time merge-ort
goes to check it. To account for this, merge-ort needs to check for the
case where the max count is 0.
With this change we are now computing the necessary value for each
directory in dirs_removed, but are not using that value anywhere. The
next two commits will make use of the values stored in dirs_removed in
order to compute whether each relevant_source (that is needed only for
directory rename detection) has become unnecessary.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When one side of history renames a directory, and the other side of
history added files to the old directory, directory rename detection is
used to warn about the location of the added files so the user can
move them to the old directory or keep them with the new one.
This sets up three different types of directories:
* directories that had new files added to them
* directories underneath a directory that had new files added to them
* directories where no new files were added to it or any leading path
Save this information in dirs_removed; the next several commits will
make use of this information.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
As noted in the previous commit, we want to be able to take advantage of
the "majority rules" portion of directory rename detection to avoid
detecting more renames than necessary. However, for diffcore-rename to
take advantage of that, it needs to know whether a rename source file
was needed for just directory rename detection reasons, or if it is
wanted for potential three-way content merging. Modify relevant_sources
from a strset to a strintmap, so we can encode additional information.
We also modify dirs_removed from a strset to a strintmap at the same
time because trying to determine what files are needed for directory
rename detection will require us tracking a bit more information for
each directory.
This commit only changes the types of the two variables from strset to
strintmap; it does not actually store any special values yet and for now
only checks for presence of entries in the strintmap. Thus, the code is
functionally identical to how it behaved before. Future commits will
start associating values with each key for these two maps.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add and apply a semantic patch for converting code that open-codes
CALLOC_ARRAY to use it instead. It shortens the code and infers the
element size automatically.
Signed-off-by: René Scharfe <l.s.r@web.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
diffcore_rename_extended() will do a bunch of setup, then check for
exact renames, then abort before inexact rename detection if there are
no more sources or destinations that need to be matched. It will
sometimes be the case, however, that either
* we start with neither any sources or destinations
* we start with no *relevant* sources
In the first of these two cases, the setup and exact rename detection
will be very cheap since there are 0 files to operate on. In the second
case, it is quite possible to have thousands of files with none of the
source ones being relevant. Avoid calling diffcore_rename_extended() or
even some of the setup before diffcore_rename_extended() when we can
determine that rename detection is unnecessary.
For the testcases mentioned in commit 557ac0350d ("merge-ort: begin
performance work; instrument with trace2_region_* calls", 2020-10-28),
this change improves the performance as follows:
Before After
no-renames: 6.003 s ± 0.048 s 5.708 s ± 0.111 s
mega-renames: 114.009 s ± 0.236 s 102.171 s ± 0.440 s
just-one-mega: 3.489 s ± 0.017 s 3.471 s ± 0.015 s
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The past several commits determined conditions when rename sources might
be needed, and filled a relevant_sources strset with those paths. Pass
these along to diffcore_rename_extended() to use to limit the sources
that we need to detect renames for.
For the testcases mentioned in commit 557ac0350d ("merge-ort: begin
performance work; instrument with trace2_region_* calls", 2020-10-28),
this change improves the performance as follows:
Before After
no-renames: 12.596 s ± 0.061 s 6.003 s ± 0.048 s
mega-renames: 130.465 s ± 0.259 s 114.009 s ± 0.236 s
just-one-mega: 3.958 s ± 0.010 s 3.489 s ± 0.017 s
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The point of directory rename detection is that if one side of history
renames a directory, and the other side adds new files under the old
directory, then the merge can move those new files into the new
directory. This leads to the following important observation:
* If the other side does not add any new files under the old
directory, we do not need to detect any renames for that directory.
Similarly, directory rename detection had an important requirement:
* If a directory still exists on one side of history, it has not been
renamed on that side of history. (See section 4 of t6423 or
Documentation/technical/directory-rename-detection.txt for more
details).
Using these two bits of information, we note that directory rename
detection is only needed in cases where (1) directories exist in the
merge base and on one side of history (i.e. dirmask == 3 or dirmask ==
5), and (2) where there is some new file added to that directory on the
side where it still exists (thus where the file has filemask == 2 or
filemask == 4, respectively). This has to be done in two steps, because
we have the dirmask when we are first considering the directory, and
won't get the filemasks for the files within it until we recurse into
that directory. So, we save
dir_rename_mask = dirmask - 1
when we hit a directory that is missing on one side, and then later look
for cases of
filemask == dir_rename_mask
One final note is that as soon as we hit a directory that needs
directory rename detection, we will need to detect renames in all
subdirectories of that directory as well due to the "majority rules"
decision when files are renamed into different directory hierarchies.
We arbitrarily use the special value of 0x07 to record when we've hit
such a directory.
The combination of all the above mean that we introduce a variable
named dir_rename_mask (couldn't think of a better name) which has one
of the following values as we traverse into a directory:
* 0x00: directory rename detection not needed
* 0x02 or 0x04: directory rename detection only needed if files added
* 0x07: directory rename detection definitely needed
We then pass this value through to add_pairs() so that it can mark
location_relevant as true only when dir_rename_mask is 0x07.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add traverse_trees_wrapper() and traverse_trees_wrapper_callback()
functions. The former runs traverse_trees() with info->fn set to
traverse_trees_wrapper_callback, in order to simply save all the entries
without processing or recursing into any of them. This step allows
extra computation to be done (e.g. checking some condition across all
files) that can be used later. Then, after that is completed, it
iterates over all the saved entries and calls the original info->fn
callback with the saved data.
Currently, this does nothing more than marginally slowing down the tree
traversal since we do not take advantage of the opportunity to compute
anything special in traverse_trees_wrapper_callback(), and thus the real
callback will be called identically as it would have been without this
extra wrapper. However, a subsequent commit will add some special
computation of some values that the real callback will be able to use.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In order to determine whether directory rename detection is needed, we
as a pre-requisite need a way to traverse through all the files in a
given tree before visiting any directories within that tree.
traverse_trees() only iterates through the entries in the order they
appear, so add some data structures that will store all the entries as
we iterate through them in traverse_trees(), which will allow us to
re-traverse them in our desired order.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
rename detection works by trying to pair all file deletions (or
"sources") with all file additions (or "destinations"), checking
similarity, and then marking the sufficiently similar ones as renames.
This can be expensive if there are many sources and destinations on a
given side of history as it results in an N x M comparison matrix.
However, there are many cases where we can compute in advance that
detecting renames for some of the sources provides no useful information
and thus that we can exclude those sources from the matrix.
To see why, first note that the merge machinery uses detected renames in
two ways:
* directory rename detection: when one side of history renames a
directory, and the other side of history adds new files to that
directory, we want to be able to warn the user about the need to
chose whether those new files stay in the old directory or move
to the new one.
* three-way content merging: in order to do three-way content merging
of files, we need three different file versions. If one side of
history renamed a file, then some of the content for the file is
found under a different path than in the merge base or on the
other side of history.
Add a simple testcase showing the two kinds of reasons renames are
relevant; it's a testcase that will only pass if we detect both kinds of
needed renames.
Other than the testcase added above, this commit concentrates just on
the three-way content merging; it will punt and mark all sources as
needed for directory rename detection, and leave it to future commits to
narrow that down more.
The point of three-way content merging is to reconcile changes made on
*both* sides of history. What if the file wasn't modified on both
sides? There are two possibilities:
* If it wasn't modified on the renamed side:
-> then we get to do exact rename detection, which is cheap.
* If it wasn't modified on the unrenamed side:
-> then detection of a rename for that source file is irrelevant
That latter claim might be surprising at first, so let's walk through a
case to show why rename detection for that source file is irrelevant.
Let's use two filenames, old.c & new.c, with the following abbreviated
object ids (and where the value '000000' is used to denote that the file
is missing in that commit):
old.c new.c
MERGE_BASE: 01d01d 000000
MERGE_SIDE1: 01d01d 000000
MERGE_SIDE2: 000000 5e1ec7
If the rename *isn't* detected:
then old.c looks like it was unmodified on one side and deleted on
the other and should thus be removed. new.c looks like a new file we
should keep as-is.
If the rename *is* detected:
then a three-way content merge is done. Since the version of the
file in MERGE_BASE and MERGE_SIDE1 are identical, the three-way merge
will produce exactly the version of the file whose abbreviated
object id is 5e1ec7. It will record that file at the path new.c,
while removing old.c from the directory.
Note that these two results are identical -- a single file named 'new.c'
with object id 5e1ec7. In other words, it doesn't matter if the rename
is detected in the case where the file is unmodified on the unrenamed
side.
Use this information to compute whether we need rename detection for
each source created in add_pair().
It's probably worth noting that there used to be a few other edge or
corner cases besides three-way content merges and directory rename
detection where lack of rename detection could have affected the result,
but those cases actually highlighted where conflict resolution methods
were not consistent with each other. Fixing those inconsistencies were
thus critically important to enabling this optimization. That work
involved the following:
* bringing consistency to add/add, rename/add, and rename/rename
conflict types, as done back in the topic merged at commit
ac193e0e0a ("Merge branch 'en/merge-path-collision'", 2019-01-04),
and further extended in commits 2a7c16c980 ("t6422, t6426: be more
flexible for add/add conflicts involving renames", 2020-08-10) and
e8eb99d4a6 ("t642[23]: be more flexible for add/add conflicts
involving pair renames", 2020-08-10)
* making rename/delete more consistent with modify/delete
as done in commits 1f3c9ba707 ("t6425: be more flexible with
rename/delete conflict messages", 2020-08-10) and 727c75b23f
("t6404, t6423: expect improved rename/delete handling in ort
backend", 2020-10-26)
Since the set of relevant_sources we compute has not yet been narrowed
down for directory rename detection, we do not pass it to
diffcore_rename_extended() yet. That will be done after subsequent
commits narrow down the list of relevant_sources needed for directory
rename detection reasons.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add the ability to diffcore_rename_extended() to allow external callers
to declare that they only need renames detected for a subset of source
files, and use that information to skip detecting renames for them.
There are two important pieces to this optimization that may not be
obvious at first glance:
* We do not require callers to just filter the filepairs out
to remove the non-relevant sources, because exact rename detection
is fast and when it finds a match it can remove both a source and a
destination whereas the relevant_sources filter can only remove a
source.
* We need to filter out the source pairs in a preliminary pass instead
of adding a
strset_contains(relevant_sources, one->path)
check within the nested matrix loop. The reason for that is if we
have 30k renames, doing 30k * 30k = 900M strset_contains() calls
becomes extraordinarily expensive and defeats the performance gains
from this change; we only want to do 30k such calls instead.
If callers pass NULL for relevant_sources, that is special cases to
treat all sources as relevant. Since all callers currently pass NULL,
this optimization does not yet have any effect. Subsequent commits will
have merge-ort compute a set of relevant_sources to restrict which
sources we detect renames for, and have merge-ort pass that set of
relevant_sources to diffcore_rename_extended().
A note about filtering order:
Some may be curious why we don't filter out irrelevant sources at the
same time we filter out exact renames. While that technically could be
done at this point, there are two reasons to defer it:
First, was to reinforce a lesson that was too easy to forget. As I
mentioned above, in the past I filtered irrelevant sources out before
exact rename checking, and then discovered that exact renames' ability
to remove both sources and destinations was an important consideration
and thus doing the filtering after exact rename checking would speed
things up. Then at some point I realized that basename matching could
also remove both sources and destinations, and decided to put irrelevant
source filtering after basename filtering. That slowed things down a
lot. But, despite learning about this important ordering, in later
restructuring I forgot and made the same mistake of putting the
filtering after basename guided rename detection again. So, I have this
series of patches structured to do the irrelevant filtering last to
start to show how much extra it costs, and then add relevant filtering
in to find_basename_matches() to show how much it speeds things up.
Basically, it's a way to reinforce something that apparently was too
easy to forget, and make sure the commit messages record this lesson.
Second, the items in the "relevant_sources" in this patch series will
include all sources that *might be* relevant. It has to be conservative
and catch anything that might need a rename, but in the patch series
after this one we'll find ways to weed out more of the *might be*
relevant ones. Unfortunately, merge-ort does not have sufficient
information to weed those ones out, and there isn't enough information
at the time of filtering exact renames out to remove the extra ones
either. It has to be deferred. So the deferral is in part to simplify
some later additions.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
As we adjust the usage of dir_rename_count we want to have a function
for clearing, or partially clearing it out. Add a
partial_clear_dir_rename_count() function for this purpose.
Reviewed-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Move the computation of dir_rename_count from merge-ort.c to
diffcore-rename.c, making slight adjustments to the data structures
based on the move. While the diffstat looks large, viewing this commit
with --color-moved makes it clear that only about 20 lines changed.
With this patch, the computation of dir_rename_count is still only done
after inexact rename detection, but subsequent commits will add a
preliminary computation of dir_rename_count after exact rename
detection, followed by some updates after inexact rename detection.
Reviewed-by: Derrick Stolee <dstolee@microsoft.com>
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We want to pass additional information to diffcore_rename() (or some
variant thereof) without plumbing that extra information through
diff_tree_oid() and diffcore_std(). Further, since we will need to
gather additional special information related to diffs and are walking
the trees anyway in collect_merge_info(), it seems odd to have
diff_tree_oid()/diffcore_std() repeat those tree walks. And there may
be times where we can avoid traversing into a subtree in
collect_merge_info() (based on additional information at our disposal),
that the basic diff logic would be unable to take advantage of. For all
these reasons, just create the add and delete pairs ourself and then
call diffcore_rename() directly.
This change is primarily about enabling future optimizations; the
advantage of avoiding extra tree traversals is small compared to the
cost of rename detection, and the advantage of avoiding the extra tree
traversals is somewhat offset by the extra time spent in
collect_merge_info() collecting the additional data anyway. However...
For the testcases mentioned in commit 557ac0350d ("merge-ort: begin
performance work; instrument with trace2_region_* calls", 2020-10-28),
this change improves the performance as follows:
Before After
no-renames: 13.294 s ± 0.103 s 12.775 s ± 0.062 s
mega-renames: 187.248 s ± 0.882 s 188.754 s ± 0.284 s
just-one-mega: 5.557 s ± 0.017 s 5.599 s ± 0.019 s
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Add some timing instrumentation for both merge-ort and diffcore-rename;
I used these to measure and optimize performance in both, and several
future patch series will build on these to reduce the timings of some
select testcases.
=== Setup ===
The primary testcase I used involved rebasing a random topic in the
linux kernel (consisting of 35 patches) against an older version. I
added two variants, one where I rename a toplevel directory, and another
where I only rebase one patch instead of the whole topic. The setup is
as follows:
$ git clone git://git.kernel.org/pub/scm/linux/kernel/git/stable/linux-stable.git
$ git branch hwmon-updates fd8bdb23b91876ac1e624337bb88dc1dcc21d67e
$ git branch hwmon-just-one fd8bdb23b91876ac1e624337bb88dc1dcc21d67e~34
$ git branch base 4703d9119972bf586d2cca76ec6438f819ffa30e
$ git switch -c 5.4-renames v5.4
$ git mv drivers pilots # Introduce over 26,000 renames
$ git commit -m "Rename drivers/ to pilots/"
$ git config merge.renameLimit 30000
$ git config merge.directoryRenames true
=== Testcases ===
Now with REBASE standing for either "git rebase [--merge]" (using
merge-recursive) or "test-tool fast-rebase" (using merge-ort), the
testcases are:
Testcase #1: no-renames
$ git checkout v5.4^0
$ REBASE --onto HEAD base hwmon-updates
Note: technically the name is misleading; there are some renames, but
very few. Rename detection only takes about half the overall time.
Testcase #2: mega-renames
$ git checkout 5.4-renames^0
$ REBASE --onto HEAD base hwmon-updates
Testcase #3: just-one-mega
$ git checkout 5.4-renames^0
$ REBASE --onto HEAD base hwmon-just-one
=== Timing results ===
Overall timings, using hyperfine (1 warmup run, 3 runs for mega-renames,
10 runs for the other two cases):
merge-recursive merge-ort
no-renames: 18.912 s ± 0.174 s 14.263 s ± 0.053 s
mega-renames: 5964.031 s ± 10.459 s 5504.231 s ± 5.150 s
just-one-mega: 149.583 s ± 0.751 s 158.534 s ± 0.498 s
A single re-run of each with some breakdowns:
--- no-renames ---
merge-recursive merge-ort
overall runtime: 19.302 s 14.257 s
inexact rename detection: 7.603 s 7.906 s
everything else: 11.699 s 6.351 s
--- mega-renames ---
merge-recursive merge-ort
overall runtime: 5950.195 s 5499.672 s
inexact rename detection: 5746.309 s 5487.120 s
everything else: 203.886 s 17.552 s
--- just-one-mega ---
merge-recursive merge-ort
overall runtime: 151.001 s 158.582 s
inexact rename detection: 143.448 s 157.835 s
everything else: 7.553 s 0.747 s
=== Timing observations ===
0) Maximum speedup
The "everything else" row represents the maximum speedup we could
achieve if we were to somehow infinitely parallelize inexact rename
detection, but leave everything else alone. The fact that this is so
much smaller than the real runtime (even in the case with virtually no
renames) makes it clear just how overwhelmingly large the time spent on
rename detection can be.
1) no-renames
1a) merge-ort is faster than merge-recursive, which is nice. However,
this still should not be considered good enough. Although the "merge"
backend to rebase (merge-recursive) is sometimes faster than the "apply"
backend, this is one of those cases where it is not. In fact, even
merge-ort is slower. The "apply" backend can complete this testcase in
6.940 s ± 0.485 s
which is about 2x faster than merge-ort and 3x faster than
merge-recursive. One goal of the merge-ort performance work will be to
make it faster than git-am on this (and similar) testcases.
2) mega-renames
2a) Obviously rename detection is a huge cost; it's where most the time
is spent. We need to cut that down. If we could somehow infinitely
parallelize it and drive its time to 0, the merge-recursive time would
drop to about 204s, and the merge-ort time would drop to about 17s. I
think this particular stat shows I've subtly baked a couple performance
improvements into merge-ort and into fast-rebase already.
3) just-one-mega
3a) not much to say here, it just gives some flavor for how rebasing
only one patch compares to rebasing 35.
=== Goals ===
This patch is obviously just the beginning. Here are some of my goals
that this measurement will help us achieve:
* Drive the cost of rename detection down considerably for merges
* After the above has been achieved, see if there are other slowness
factors (which would have previously been overshadowed by rename
detection costs) which we can then focus on and also optimize.
* Ensure our rebase testcase that requires little rename detection
is noticeably faster with merge-ort than with apply-based rebase.
Signed-off-by: Elijah Newren <newren@gmail.com>
Acked-by: Taylor Blau <ttaylorr@github.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
get_provisional_directory_renames() has code to detect directories being
evenly split between different locations. However, as noted previously,
if there are no new files added to that directory that was split evenly,
our inability to determine where the directory was renamed to doesn't
matter since there are no new files to try to move into the new
location. Unfortunately, that code is unaware of whether there are new
files under the directory in question and we just ignore that, causing
us to fail t6423 test 2b but pass test 2a; turn off the error for now,
swapping which tests pass and fail.
The motivating reason for switching this off as a temporary measure is
that as we add optimizations, we'll start looking at only subsets of
renames, and subsets of renames can start switching the result we get
when this error is (wrongly) on. Once we get enough optimizations,
however, we can prevent that code from even running when there are no
new files added to the relevant directory, at which point we can revert
this commit and then both testcases 2a and 2b will pass simultaneously.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
When a series of merges was performed (such as for a rebase or series of
cherry-picks), only the data structures allocated by the final merge
operation were being freed. The problem was that while picking out
pieces of merge-ort to upstream, I previously misread a certain section
of merge_start() and assumed it was associated with a later
optimization. Include that section now, which ensures that if there was
a previous merge operation, that we clear out result->priv and then
re-use it for opt->priv, and otherwise we allocate opt->priv.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
* en/ort-directory-rename: (28 commits)
merge-ort: fix a directory rename detection bug
merge-ort: process_renames() now needs more defensiveness
merge-ort: implement apply_directory_rename_modifications()
merge-ort: add a new toplevel_dir field
merge-ort: implement handle_path_level_conflicts()
merge-ort: implement check_for_directory_rename()
merge-ort: implement apply_dir_rename() and check_dir_renamed()
merge-ort: implement compute_collisions()
merge-ort: modify collect_renames() for directory rename handling
merge-ort: implement handle_directory_level_conflicts()
merge-ort: implement compute_rename_counts()
merge-ort: copy get_renamed_dir_portion() from merge-recursive.c
merge-ort: add outline of get_provisional_directory_renames()
merge-ort: add outline for computing directory renames
merge-ort: collect which directories are removed in dirs_removed
merge-ort: initialize and free new directory rename data structures
merge-ort: add new data structures for directory rename detection
merge-ort: add implementation of type-changed rename handling
merge-ort: add implementation of normal rename handling
merge-ort: add implementation of rename collisions
...
As noted in commit 902c521a35 ("t6423: more involved directory rename
test", 2020-10-15), when we have a case where
* dir/subdir/ has several files
* almost all files in dir/subdir/ are renamed to folder/subdir/
* one of the files in dir/subdir/ is renamed to folder/subdir/newsubdir/
* the other side of history (that doesn't do the renames) adds a
new file to dir/subdir/
Then for the majority of the file renames, the directory rename of
dir/subdir/ -> folder/subdir/
is actually not represented that way but as
dir/ -> folder/
We also had one rename that was represented as
dir/subdir/ -> folder/subdir/newsubdir/
Now, since there's a new file in dir/subdir/, where does it go? Well,
there's only one rule for dir/subdir/, so the code previously noted that
this rule had the "majority" of the one "relevant" rename and thus
erroneously used it to place the file in folder/subdir/newsubdir/. We
really want the heavy weight associated with dir/ -> folder/ to also be
treated as dir/subdir/ -> folder/subdir/, so that we correctly place the
file in folder/subdir/.
Add a bunch of logic to make sure that we use all relevant renamings in
directory rename detection.
Note that testcase 12f of t6423 still fails after this, but it gets
further than merge-recursive does. There are some performance related
bits in that testcase (the region_enter messages) that do not yet
succeed, but the rest of the testcase works after this patch.
Subsequent patch series will fix up the performance side.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Since directory rename detection adds new paths to opt->priv->paths and
removes old ones, process_renames() needs to now check whether
pair->one->path actually exists in opt->priv->paths instead of just
assuming it does.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This function roughly follows the same outline as the function of the
same name from merge-recursive.c, but the code diverges in multiple
ways due to some special considerations:
* merge-ort's version needs to update opt->priv->paths with any new
paths (and opt->priv->paths points to struct conflict_infos which
track quite a bit of metadata for each path); merge-recursive's
version would directly update the index
* merge-ort requires that opt->priv->paths has any leading directories
of any relevant files also be included in the set of paths. And
due to pointer equality requirements on merged_info.directory_name,
we have to be careful how we compute and insert these.
* due to the above requirements on opt->priv->paths, merge-ort's
version starts with a long comment to explain all the special
considerations that need to be handled
* merge-ort can use the full data stored in opt->priv->paths to avoid
making expensive get_tree_entry() calls to regather the necessary
data.
* due to messages being deferred automatically in merge-ort, this is
the best place to handle conflict messages whereas in
merge-recursive.c they are deferred manually so that processing of
entries does all the printing
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Due to the string-equality-iff-pointer-equality requirements placed on
merged_info.directory_name, apply_directory_rename_modifications() will
need to have access to the exact toplevel directory name string pointer
and can't just use a new empty string. Store it in a field that we can
use.
Signed-off-by: Elijah Newren <newren@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>