Optimize diff-hl--update

[jamescherti]

This pull request optimizes diff-hl--update.

[jamescherti]

This adds two optimizations during the buffer traversal.

First, (eq buf base) is checked before calling (buffer-base-buffer buf). This short-circuits evaluation to bypass the buffer-base-buffer call overhead.

Second, the boolean state of (or changes ref-changes) outside the loop. This prevents the Lisp interpreter from recalculating the same static condition for all active buffers.

[dgutov]

Hi!

I like the (when ref-changes and (when changes improvements.

The other two seem negligible: considering we save the buffer-base-buffer call only once - when the buffer matches the original, but not for the other buffers; and we break the "target buffer" abstraction. Saving an or when the buffer already matched won't move the needle.

We could also drop the (buffer-live-p buf) check inside the loop because buffer-list returns only live ones. And that probably isn't going to change during our execution - we don't allow to run any timers or process filters during diff-hl--update-overlays or diff-hl--autohide-margin.

[jamescherti]

That makes perfect sense. I have updated the pull request to preserve the diff-hl--target-buffer abstraction and dropped the redundant buffer-live-p checks inside the loop.

[jamescherti]

I also set inhibit-redisplay to t during the resolution callback to prevent the display engine from rendering intermediate UI states during asynchronous overlay updates. Because the function recreates a large number of overlays across multiple buffers, the Emacs display engine can waste CPU cycles recalculating display properties, margins, or fringe bitmaps before the buffer state is completely finalized.

[dgutov]

I have updated the pull request to preserve the diff-hl--target-buffer abstraction and dropped the redundant buffer-live-p checks inside the loop.

Please drop the or optimization too, as I mentioned in the previous message.

Regarding inhibit-redisplay - do you have any measurements that show an improvement? It does help in some scenarios, but here I expect no redisplays to happen during that block's execution.

Redisplay happens at specific points of the event loop, not right after any change in a buffer.

[jamescherti]

I have removed the 'or' and inhibit-redisplay.

[dgutov]

Thanks!

[jamescherti]

You're very welcome!