The new `jj unsquash` command moves changes from a commit's parent
into the commit itself. It comes with a `--interactive` flag. The
command is probably most useful for moving changes from the working
copy's parent into the working copy but it can of course be used for
moving changes into any commit (from that commit's parent).
The new diff algorithm produces pretty bad diffs in some cases, such
as cc4b1e9230 in this repo (the parent of this commit). I think the
problem there is that many words are repeated over and over. Diffing
first at the line level and then refining the diff of the changed
ranges at the word level gives much better results. That's what this
patch does. After this patch, `jj diff -r cc4b1e923091` looks pretty
similar to the diff in GitHub's UI.
I hope to get around to doing the same for the merge code soon.
Impact on benchmarks:
Before:
test bench_diff_10k_lines_reversed ... bench: 42,647,532 ns/iter (+/- 765,347)
test bench_diff_10k_modified_lines ... bench: 21,407,980 ns/iter (+/- 126,366)
test bench_diff_10k_unchanged_lines ... bench: 4,235 ns/iter (+/- 16)
test bench_diff_1k_lines_reversed ... bench: 1,190,483 ns/iter (+/- 7,192)
test bench_diff_1k_modified_lines ... bench: 1,919,766 ns/iter (+/- 9,665)
test bench_diff_1k_unchanged_lines ... bench: 231 ns/iter (+/- 1)
test bench_diff_git_git_read_tree_c ... bench: 174,702 ns/iter (+/- 1,199)
After:
test bench_diff_10k_lines_reversed ... bench: 38,289,509 ns/iter (+/- 129,004)
test bench_diff_10k_modified_lines ... bench: 33,140,659 ns/iter (+/- 3,989,339)
test bench_diff_10k_unchanged_lines ... bench: 3,099 ns/iter (+/- 14)
test bench_diff_1k_lines_reversed ... bench: 973,551 ns/iter (+/- 94,895)
test bench_diff_1k_modified_lines ... bench: 3,033,818 ns/iter (+/- 29,513)
test bench_diff_1k_unchanged_lines ... bench: 230 ns/iter (+/- 1)
test bench_diff_git_git_read_tree_c ... bench: 79,100 ns/iter (+/- 963)
So most of them get slower, as expected. The last one, taken from a
real diff in the git.git repo, get faster, however (which is also what
I would have expected).
I made a quite late change in a recent patch to make the merge code to
merge based on lines instead of words. I forgot to update the tests
(and to even run them). Sorry :(
The previous patch switched over the content-merge code to use the new
histogram diff code. This patch switches over the content-diff code to
use the histogram diff code. As before, the immediate goal is to speed
it up. `jj diff -r c28ded83fc` in the git.git repo is a good example
of a diff that's extremely slow to calculate with our current
LCS-based diff. With this patch, that drops from 35 s to 0.12 s.
The diff was slightly better before. I think that's mostly because of
our different definition of a "word" in the data. We can improve that
later. The speedup we get now is easily worth the slightly worse diff.
With the histogram diff code from the previous patch, we can now start
using that for finding the "sync regions" in 3-way merge. That helps a
lot with the slow merging we had before this patch. `jj diff -r
9d540e9726` in the git.git repo drops from 22 s to 0.15 s with this
patch. (That commit is a rather arbitrary merge commit from aroun 5
years ago.)
With the new diff algorithm, the output of `jj diff -r 9d540e9726` in
git.git looks better if we find unchanged sync regions based on lines
than on words, so that's what I'm using in this patch. That's a change
compared the the LCS-based diff we used before this patch. I suspect
the reason that finding sync regions based on words works worse now is
not because of the change from LCS to histogram but because of the
change in how we define a word. My goal right now is mostly to make it
faster; I'll get back to refining the diff result later.
The current diff algorithm does a full LCS on the words of the texts,
which is really slow. Diffing the working copy when e.g.
`src/commands.py` has changes far apart takes seconds. This patch adds
an implementation inspired by JGit's Histogram diff. I say "inspired"
because I just didn't quite understand it :P In particular, I didn't
understand what it does when it finds non-unique elements. I decided
to line up the leading common elements on both sides of the merge. I
don't know if that usually gives good enough results in practice.
I'm sure this can still be optimized a lot, but this seems good enough
as a start. There is also many things to improve about the quality of
the diffs.
IIUC from rust-lang/rust#81654, `lexical-core` 0.7.4 was broken with
some versions of `rustc` from early February. CI is still failing, so
I guess that is using a `rustc` from that time. However, the logs say
that it installed "rustc 1.53.0-nightly (74874a690 2021-03-30)", so
I'm confused.
This updates `jj log` to walk the index for doing the topological
walk, which is much faster than walking the object graph. This speeds
up `jj log | head -1` in the git.git repo from ~1.9s to ~0.27s (most
of the remaining time is spent calculating the evolve state).
A consequence of walking the index instead is that the order of
commits in the output is by by generation number. That's nice in some
ways, but it also means that the newest commit isn't always at the
top.
When I recently changed the revision argument from being passed to
`-r` to being a positional argument, I accidentally made it
required. Let's restore the default of "@".
I keep forgetting to pass the `-r`. The command takes only a revision
as argument and it doesn't seem likely that we'll want to positional
arguments for filenames in the future either.
`Transaction` has a bunch of functions that are now simple
delegates. It probably makes sense to directly use a `&mut
MutableRepo` instead of `&mut Transaction` in most places. This patch
starts that migration.
I just changed my `~/.gitignore` and some tests started failing
because the working copy respects the user's `~/.gitignore`. We should
probably not depend on `$HOME` in the library crate. For now, this
patch just makes sure we set it to an arbitrary directory in the tests
where it matters.
`test_commit_parallel` was failing on Mac in the GitHub CI. I suspect
the reason was that it was timing out. The test runs in about 1 s on
my Linux desktop and in about 3 s on my Mac laptop. It failed after 31
in the GitHub CI. This patch increases the timeout to 1 minute to try
to make the test pass. It would be better to set the timeout to a
higher value only in tests, but this will be good enough for now. By
the way, it has turned out that git notes (at least libgit2's
implementation of them) are too slow, so we should probably eventually
create our own storage for the extra metadata instead.
I only noticed that there was a newer version when running `cargo
install --path .`, which resulted in warnings about deprecated
functions. There's no other reason I'm aware of to upgrade now.
We can now finally use the commit index for filtering out ancestors
from the sets of heads.
I haven't timed the change from most of the recent work on
performance, but I did a measurement after this commit. I modified a
commit in the git.git repo's "what's cooking" branch (because that's
linear). Then I ran `jj evolve` so the 100 commits after it would get
evolved. That took ~700ms. `git rebase` of the same 100 commits took
~6s.
I also compared `jj op undo` of that `jj evolve` operation. With this
patch, that was sped up from ~6.8s to ~125ms.
`MutableRepo` has more information needed for taking fast-paths, and
it will have to make the same decision for doing incremental updates
of the evolution state anyway.
This patch continues the work from the previous pathc. From this
patch, we no longer calculate the evolution state just because a
transaction starts. We still unnecessarily calculate it when adding a
commit within the transaction, however. I'll fix that next.
This patch changes it so that `ReadonlyEvolution` does not lazily
calculate its state and the caller, i.e. `ReadonlyRepo`, is instead
responsible for the laziness. That will allow the caller to make
decisions based on whether the state has been
calculated. Specifically, we don't want to calculate the evolution
state in order to update it incrementally if it hasn't already been
calculated. It's better to just leave it uncalculated in that case.
As a result of moving the laziness out of `ReadonlyEvolution`, we also
don't need to the reference to `ReadonlyRepo` anymore, which
simplifies things a bunch. The next patch will continue by making the
corresponding change to `MutableEvolution`, which will let us simplify
even more.
