Reduce cost of syntactic classification by around 25% while paging down through a file (and around 50% while arrowing down).

[CyrusNajmabadi]

Takes us from:

To

--

And from:

to:

The simple idea here is to cache more classified lines that the editor asks us about. Empirical testing showed that during page-down, we get asked for lines we were recently asked about 25% of the time, and arrowing down increased that to 85%.

By caching more (at a fixed, but larger size than the current cache), we dropped the number of times we actually have to go back to syntax by quite a bit. This helped us avoid the expensive walks of the tree, which is the limiting factor currently in large files.

[sharwell]

Empirical testing showed that during page-down, we get asked for lines we were recently asked about 25% of the time, and arrowing down increased that to 85%.

How many of these cases are caused by us sending a change notification for something that didn't change (and the editor correctly responding by asking for the new data), and how many are cases where we didn't indicate a change but the editor still sent a request?

[CyrusNajmabadi]

How many of these cases are caused by us sending a change notification for something that didn't change (and the editor correctly responding by asking for the new data), and how many are cases where we didn't indicate a change but the editor still sent a request?

None.

With syntactic classification we only trigger a reclassification in the following cases:

1. Workspace connected (not applicable here)
2. Workspace disconnected (not applicable here)
3. Subject buffer changing (not applicable here, we're just scrolling, not editing).
4. Parse options changing. Not applicable here.

It's just the case that as we're scrolling, the editor calls into us for lines they already asked about.

[ToddGrun]

UpdateExistingEntryInCache

Can you clarify for me when this is currently called? Same snapshot, same span, but we want to specify new values?

[CyrusNajmabadi]

yes.

that would happen in the case, for example, if we did an update notification (say because parse options change). We would then generate new classifications, likely at the same span as an existing span.

[CyrusNajmabadi]

updated ot remove this.

[ToddGrun]

[ToddGrun]

this seems a bit odd