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.
Before the optimization commits, this test used to take about 20
seconds to run on my machine. Afterwards, there is still a couple
seconds of data setup, but the actual diff and rename detection
runs in a fraction of a second.
The description of what the function does hasn't been true for quite a
while. Change it to reflect the way it currently works.
While here, remove an even older comment about missing features that
have been implemented.
Instead of using lots of strdup calls, this adds a memory pool to
the loose refs iteration code and uses it for keeping track of the
loose refs array. Memory usage could probably be reduced even
further by eliminating the vector and just scanning by adding the
strlen of each ref, but that would be a more intrusive changes.
This also updates the error handling to be more thorough about
checking for failed allocations, etc.
The git_reference_next API silently skips invalid references when
scanning the loose refs. The git_reference_next_name API should
skip the same ones even though it isn't creating the reference
object.
This adds a test with a an invalid loose reference and makes sure
that both APIs skip the same entries and generate the same results.
