ScopeProvider performance / ScopeComputation

[goto40]

We use langium 3.5.
We have just observed a strange performance effect and may have misunderstood some details.

• We have a large model (>60000 LOC) with a ScopeComputation implementation to provide global names. This works like a charm.
• We now want to restrict the scoping to work only within model file (no global visibility). We achieved this by implementing a custom ScopeProvider, using the this.indexManager of the ScopeProvider which allow us to get allElements with the desired filter functionality. This also works, BUT is very slow (unusable with our 60000 LOC model file; it takes many seconds to load the file).

Did oversee a detail? Any help is appreciated. Thanks!

[msujew]

Hey @goto40,

running into performance issues with large workspaces is very much expected. Without any specific optimizations, linking is O(n^2). However, this can be brought down quite easily with a few tricks, most of them boil down to caching. Have you tried using the DocumentCache API to cache common scope objects?

Ideally, you should create one MapScope per level of scoping that your language uses and then cache those, so the ScopeProvider can reuse them. The main issue is usually, that the ScopeProvider recreates the whole scope from scratch for each reference. However, most references share almost all of their elements, and so caching can massively help improve performance there. The default scope provider is already doing a lot of caching on the global level:

langium/packages/langium/src/references/scope-provider.ts

Lines 100 to 102 in e9d0c27

	protected getGlobalScope(referenceType: string, _context: ReferenceInfo): Scope {
	return this.globalScopeCache.get(referenceType, () => new MapScope(this.indexManager.allElements(referenceType)));
	}

You should probably use something similar, on a more local level.

[goto40]

Thank you for your quick reply. We really appreciate your work and your expertise.

Best,
Pierre