These changes fix the basic problem with GIT_DIFF_REVERSE being
broken for text diffs. The reversed diff entries were getting
added to the git_diff correctly, but some of the metadata was kept
incorrectly in a way that prevented the text diffs from being
generated correctly. Once I fixed that, it became clear that it
was not possible to merge reversed diffs correctly. This has a
first pass at fixing that problem. We probably need more tests
to make sure that is really fixed thoroughly.
While the base git_diff_delta structure always contains two files,
when we introduce conflict data, it will be helpful to have an
indicator when an additional file is involved.
This lays groundwork for separating formatting options from diff
creation options. This groups the formatting flags separately
from the diff list creation flags and reorders the options. This
also tweaks some APIs to further separate code that uses patches
from code that just looks at git_diffs.
This makes no functional change to diff but renames a couple of
the objects and splits the new git_patch (formerly git_diff_patch)
into a new header file.
Ensure that we apply splits to rewrites, even if we're not
interested in examining it closely for rename/copy detection.
In keeping with core git, status should not display rewrites,
it should simply show files as "modified".
When using a rename source that is actually a to-be-split record,
we have to update the best-fit mapping data in both the case where
the target is also a split record and the case where the target
is a simple added record. Before this commit, we were only doing
the update when the target was itself a split record (and even in
that case, the test was slightly wrong).
After doing further profiling, I found that a lot of time was
being spent attempting to insert hashes into the file hash
signature when using the rolling hash because the rolling hash
approach generates a hash per byte of the file instead of one
per run/line of data.
To optimize this, I decided to convert back to a run-based file
signature algorithm which would be more like core Git.
After changing this, a number of the existing tests started to
fail. In some cases, this appears to have been because the test
was coded to be too specific to the particular results of the file
similarity metric and in some cases there appear to have been bugs
in the core rename detection code where only by the coincidence
of the file similarity scoring were the expected results being
generated.
This renames all the variables in the core rename detection code
to be more consistent and hopefully easier to follow which made it
a bit easier to reason about the behavior of that code and fix the
problems that I was seeing. I think it's in better shape now.
There are a couple of tests now that attempt to stress test the
rename detection code and they are quite slow. Most of the time
is spent setting up the test data on disk and in the index. When
we roll out performance improvements for index insertion, it
should also speed up these tests I hope.
The size data in the index may not reflect the actual size of the
blob data from the ODB when content filtering comes into play.
This commit fixes rename detection to use the actual blob size when
calculating data signatures instead of the value from the index.
Because of a misunderstanding on my part, I first converted the
git_index_add_bypath API to use the post-filtered blob data size
in creating the index entry. I backed that change out, but I
kept the overall refactoring of that routine and the new internal
git_blob__create_from_paths API because it eliminates an extra
stat() call from the code that adds a file to the index.
The existing tests actually cover this code path, at least when
running on Windows, so at this point I'm not adding new tests to
cover the changes.
The previous fix for checking file sizes with rename detection
always loads the blob. In this version, if the odb backend can
get the object header without loading the whole thing into memory,
then we'll just use that, so that we can eliminate possible rename
sources & targets without loading them.
The performance improvements I introduced for rename detection
were not able to run successfully for tree-to-tree diffs because
the blob size was not known early enough and so the file signature
always had to be calculated nonetheless.
This change separates loading blobs into memory from calculating
the signature. I can't avoid having to load the large blobs into
memory, but by moving it forward, I'm able to avoid the signature
calculation if the blob won't come into play for renames.
This makes the diff rename tracking code more careful about the
order in which it processes renames and more thorough in updating
the mapping of correct renames when an earlier rename update
alters the index of a later matched pair.
This changes the behavior of the status RENAMED flags so that they
will be combined with the MODIFIED flags if appropriate. If a file
is modified in the index and also renamed, then the status code
will have both the GIT_STATUS_INDEX_MODIFIED and INDEX_RENAMED bits
set. If it is renamed but the OID has not changed, then just the
GIT_STATUS_INDEX_RENAMED bit will be set. Similarly, the flags
GIT_STATUS_WT_MODIFIED and GIT_STATUS_WT_RENAMED can both be set
independently of one another.
This fixes a serious bug where the check for unmodified files that
was done at data load time could end up erasing the RENAMED state
of a file that was renamed with no changes.
Lastly, this contains a bunch of new tests for status with renames,
including tests where the only rename changes are case changes.
The expected results of these tests have to vary by whether the
platform uses a case sensitive filesystem or not, so the expected
data covers those platform differences separately.
This is a significant reorganization of the diff code to break it
into a set of more clearly distinct files and to document the new
organization. Hopefully this will make the diff code easier to
understand and to extend.
This adds a new `git_diff_driver` object that looks of diff driver
information from the attributes and the config so that things like
function content in diff headers can be provided. The full driver
spec is not implemented in the commit - this is focused on the
reorganization of the code and putting the driver hooks in place.
This also removes a few #includes from src/repository.h that were
overbroad, but as a result required extra #includes in a variety
of places since including src/repository.h no longer results in
pulling in the whole world.
This extends the rename tests to make sure that every rename
scenario in the inner loop of git_diff_find_similar is actually
exercised. Also, fixes an incorrect assert that was in one of
the clauses that was not previously being exercised.
This adds a couple more tests of different rename scenarios.
Also, this fixes a problem with the case where you have two
"split" deltas and the left half of one matches the right half of
the other. That case was already being handled, but in the wrong
order in a way that could result in bad output. Also, if the swap
also happened to put the other two halves into the correct place
(i.e. two files exchanged places with each other), then the second
delta was left with the SPLIT flag set when it really should be
cleared.
This flips rename detection around so instead of creating a
forward mapping from deltas to possible rename targets, instead
it creates a reverse mapping, looking at possible targets and
trying to find a source that they could have been renamed or
copied from. This is important because each output can only
have a single source, but a given source could map to multiple
outputs (in the form of COPIED records).
Additionally, this makes a couple of tweaks to the public rename
detection APIs, mostly renaming a couple of options that control
the behavior to make more sense and to be more like core Git.
I walked through the tests looking at the exact results and
updated the expectations based on what I saw. The new code is
different from the old because it cannot give some nonsense
results (like A was renamed to both B and C) which were part of
the outputs previously.
- Add new GIT_DIFF_FIND_EXACT_MATCH_ONLY flag to do similarity
matching without using the similarity metric (i.e. only compare
the SHA).
- Clean up the similarity measurement code to more rigorously
distinguish between files that are not similar and files that
are not comparable (previously, a 0 could either mean that the
files could not be compared or that they were totally different)
- When splitting a MODIFIED file into a DELETE/ADD pair, actually
make a DELETED/UNTRACKED pair if the right side of the diff is
from the working directory. This prevents an odd mix of ADDED
and UNTRACKED files on workdir diffs.
There are a number of bugs in the rename code that only were
obvious when I started testing it against large old repos with
more complex patterns. (The code to do that testing is not ready
to merge with libgit2, but I do plan to add more thorough tests.)
This contains a significant number of changes and also tweaks the
public API slightly to make emulating core git easier.
Most notably, this separates the GIT_DIFF_FIND_AND_BREAK_REWRITES
flag into FIND_REWRITES (which adds a self-similarity score to
every modified file) and BREAK_REWRITES (which splits the modified
deltas into add/remove pairs in the diff list). When you do a raw
output of core git, rewrites show up as M090 or such, not at A and
D output, so I wanted to be able to emulate that.
Publicly, this also changes the flags to be uint16_t since we
don't need values out of that range.
Internally, this contains significant changes from a number of
small bug fixes (like using the wrong side of the diff to decide
if the object could be found in the ODB vs the workdir) to larger
issues about which files can and should be compared and how the
various edge cases of similarity scores should be treated.
Honestly, I don't think this is the last update that will have to
be made to this code, but I think this moves us closer to correct
behavior and I tried to document the code so it would be easier
to follow..
Under some strange circumstances, diffs can end up listing files
that we can't actually open successfully. Instead of aborting
the git_diff_find_similar, this makes it so that those files just
won't be considered as valid rename/copy targets instead.
This moves a couple of checks outside of the inner loop of the
find_similar rename/copy detection phase that are only dependent
on the "from" side of a detection.
Also, this replaces the inefficient initialization of the
options structure when a value is not provided explicitly by the
user.
Instead of creating three git_diff_similarity_metric statically
for the various config options, just create the metric structure
on demand and populate it, using the payload to specific the
extra flags that should be passed to the hashsig. This removes
a level of obfuscation from the code, I think.
This adds some new tests that actually exercise the similarity
metric between files to detect renames, copies, and split modified
files that are too heavily modified.
There is still more testing to do - these tests are just partially
covering the cases.
There is also one bug fix in this where a change set with only
MODIFY being broken into ADD/DELETE (due to low self-similarity)
without any additional RENAMED entries would end up not processing
the split requests (because the num_rewrites counter got reset).
This is the initial integration of the similarity metric into
the `git_diff_find_similar()` code path. The existing tests all
pass, but the new functionality isn't currently well tested. The
integration does go through the pluggable metric interface, so it
should be possible to drop in an alternative to the internal
metric that libgit2 implements.
This comes along with a behavior change for an existing interface;
namely, passing two NULLs to git_diff_blobs (or passing NULLs to
git_diff_blob_to_buffer) will now call the file_cb parameter zero
times instead of one time. I know it's strange that that change
is paired with this other change, but it emerged from some
initialization changes that I ended up making.
Previously the git_diff_delta recorded if the delta was binary.
This replaces that (with no net change in structure size) with
a full set of flags. The flag values that were already in use
for individual git_diff_file objects are reused for the delta
flags, too (along with renaming those flags to make it clear that
they are used more generally).
This (a) makes things somewhat more consistent (because I was
using a -1 value in the "boolean" binary field to indicate unset,
whereas now I can just use the flags that are easier to understand),
and (b) will make it easier for me to add some additional flags to
the delta object in the future, such as marking the results of a
copy/rename detection or other deltas that might want a special
indicator.
While making this change, I officially moved some of the flags that
were internal only into the private diff header.
This also allowed me to remove a gross hack in rename/copy detect
code where I was overwriting the status field with an internal
value.
This plugs in the three basic similarity strategies for handling
whitespace via internal use of the pluggable API. In so doing, I
realized that the use of git_buf in the hashsig API was not needed
and actually just made it harder to use, so I tweaked that API as
well.
Note that the similarity metric is still not hooked up in the
find_similarity code - this is just setting out the function that
will be used.
This moves the similarity metric code out of buf_text and into a
new file. Also, this implements a different approach to similarity
measurement based on a Rabin-Karp rolling hash where we only keep
the top 100 and bottom 100 hashes. In theory, that should be
sufficient samples to given a fairly accurate measurement while
limiting the amount of data we keep for file signatures no matter
how large the file is.
This improves the naming for the rename related functionality
moving it to be called `git_diff_find_similar()` and renaming
all the associated constants, etc. to make more sense.
I also moved the new code (plus the existing `git_diff_merge`)
into a new file `diff_tform.c` where I can put new functions
related to manipulating git diff lists.
This also updates the implementation significantly from the
last revision fixing some ordering issues (where break-rewrite
needs to be handled prior to copy and rename detection) and
improving config option handling.
