pascalkuthe / imara-diff Goto Github PK

Just now I have integrated imara-diff into gitoxide, replacing similar for a stunning 2x speedup when running ein t h -l. Thanks to your work gitoxide can excel even with diff performance, a previously undercooked feature that I cobbled together quickly. With imara-diff I have the feeling that I can built on top of a crate that acknowledges git and sees it as baseline, along with the desire to improve on it. It's probably what I would have wanted to have in git-diff verbatim if there would have been enough time (but… I am also glad I didn't have to implement it myself 😅).

Thank you sooo much 🙏

Running the test with Miri using MIRIFLAGS="-Zmiri-disable-isolation" cargo +nightly miri test for this repo results in a SIGKILL.

running 5 tests
test tests::complex_diffs ... error: test failed, to rerun pass `--lib`

Caused by:
  process didn't exit successfully: `/usr/local/rustup/toolchains/nightly-aarch64-unknown-linux-gnu/bin/cargo-miri runner /data/target/miri/aarch64-unknown-linux-gnu/debug/deps/imara_diff-b916e6c16bd4f65d` (signal: 9, SIGKILL: kill)

This might have something to do with the issue reported in rust-lang/rust#112171.

Performing a word diff over a full file can be fairly slow on large files.
A better approach is to perform a line diff first and and then perform the word diff on the found changes.
While this is already possible with imara-diff is requires quite a bit of legwork and can be tricky to get right.
It would be nice if this could be included in the library directly.
This has multiple steps for an implementation:

• Determine the output format. A different trait or force collecting into a Vec?
• Implement a TokenSource for words
• Implement a Sink that automatically computes a word diff
• Potentially implement a heuristic to detect and ignore

The diff algorithm in git only operates on lines. It is worth looking into what exactly they use to produce a colored word diff from the line diff.
Perhaps a different algorithm is a better fit?

Hey, thanks for making this crate. :)
Would it be possible to call diff without having to intern the input first? In my use case (character-wise diffing) interning doesn't seem necessary, as chars should be as cheap to compare as Tokens (that are just u32's under the hood) - and interning has it's non-trivial cost.

Thanks!