Improve `Semantics` construction performance

[Veykril]

hir::Semantics is a general cache used to be able to resolve any syntax node to its semantic definition. This is generally slow given the kind of work it does, but quite often we do a lot more work than necessary, populating unnecessary DynMap tables and more importantly, we populate it via parsing, where the parsing happens through direct database calls, even though the Semantics itself caches these. That means for big enough files we can end up with thrashing the LRU cache a fair bit. We should fix that, a related problem is the HasSource trait going through the database instead of Semantics which means we can end up with nodes not cached by Semantics ultimately ending up in panics when it tries to lookup stuff (which has happened in the past).

So the things that need to be changed are:

• Change the signature of (or add a second version) HasSource::source to take a reference to the Semantics instead that does the right thing

  rust-analyzer/crates/hir/src/has_source.rs

  Line 29 in 062e1b9

  fn source(self, db: &dyn HirDatabase) -> Option<InFile<Self::Ast>>;

• Change

  rust-analyzer/crates/hir-def/src/src.rs

  Lines 14 to 21 in 062e1b9

  	pub trait HasSource {
  	type Value: AstNode;
  	fn source(&self, db: &dyn DefDatabase) -> InFile<Self::Value> {
  	let InFile { file_id, value } = self.ast_ptr(db);
  	InFile::new(file_id, value.to_node(&db.parse_or_expand(file_id)))
  	}
  	fn ast_ptr(&self, db: &dyn DefDatabase) -> InFile<AstPtr<Self::Value>>;
  	}

  to support this
• Change

  rust-analyzer/crates/hir-def/src/child_by_source.rs

  Lines 26 to 33 in 062e1b9

  	pub trait ChildBySource {
  	fn child_by_source(&self, db: &dyn DefDatabase, file_id: HirFileId) -> DynMap {
  	let mut res = DynMap::default();
  	self.child_by_source_to(db, &mut res, file_id);
  	res
  	}
  	fn child_by_source_to(&self, db: &dyn DefDatabase, map: &mut DynMap, file_id: HirFileId);
  	}

  to support this

This will most likely require adding callbacks to some of the hir-def crate APIs to abstract away the parsing (and then add / keep the current APIs as convenience APIs where Semantics won't be involved).

[lnicola]

Mostly unrelated, but in the past we used to construct multiple Semantics during the execution of a single request. It might be worth checking whether that's still the case.

[Kohei316]

@rustbot claim

[ShoyuVanilla]

Semantics caches those things but the direction is opposite; the cache direction is Source -> Def, but the functions we need to modify have the signatures Def -> Source. Should this be done by making the caches bidirectional or doing some inefficient inverse search like HashMap::iter().find()?

edit) The former seems to contradict #17380 🤔

[Shourya742]

Hey @Veykril , is this up for grab?

[Shourya742]

@rustbot claim

[Veykril]

Semantics caches those things but the direction is opposite; the cache direction is Source -> Def, but the functions we need to modify have the signatures Def -> Source. Should this be done by making the caches bidirectional or doing some inefficient inverse search like HashMap::iter().find()?

edit) The former seems to contradict #17380 🤔

Good question, LRU cache invalidation wont be an issue with the new salsa anymore as the LRU doesnt trigger mid query computations. It has been a while since ive though about this issue