Pre-compute LocalDefId <-> HirId mappings and remove NodeId <-> HirId conversion APIs

[marmeladema]

cc #50928

I don't know who is exactly the best person to review this.

r? @petrochenkov

[marmeladema]

Next step is probably to move

    node_id_to_def_id: FxHashMap<ast::NodeId, LocalDefId>,
    def_id_to_node_id: IndexVec<LocalDefId, ast::NodeId>,

mappings from hir::Definitions to Resolver such that any references to NodeId are gone with it.

[petrochenkov]

What is the end goal here, removing any mentions of NodeId from struct Definitions?

Definitions is a pretty heterogeneous structure.
Perhaps it can be split into two structures instead, one for fields used before lowering to HIR, and one for fields used after lowering to HIR?

struct Definitions {
    // Both before and after (or filled before, used after)
    table: DefPathTable,
    parent_modules_of_macro_defs: FxHashMap<ExpnId, DefId>,
    expansions_that_defined: FxHashMap<LocalDefId, ExpnId>,

    // Before
    def_id_to_span: IndexVec<LocalDefId, Span>,
    node_id_to_def_id: FxHashMap<ast::NodeId, LocalDefId>,
    def_id_to_node_id: IndexVec<LocalDefId, ast::NodeId>,
    next_disambiguator: FxHashMap<(LocalDefId, DefPathData), u32>,
    invocation_parents: FxHashMap<ExpnId, LocalDefId>,
    placeholder_field_indices: FxHashMap<ast::NodeId, usize>,

    // ???
    node_id_to_hir_id: IndexVec<ast::NodeId, Option<hir::HirId>>,
    hir_id_to_node_id: FxHashMap<hir::HirId, ast::NodeId>,
}

Having the "before" fields in the ResolverOutputs version of Definitions doesn't make any sense, but they are currently placed there.

[petrochenkov]

The "after" fields could actually go into ResolverOutputs directly, without any additional structure.
So Definitions will exist solely during early compilation as a part of Resolver.

[marmeladema]

My idea was to move everything NodeId related (ie: the "before" fields) into the Resolver such that they are dropped with it but I would keep the rest of the fields in Definitions.

So to summarize, using this PR as base of discussion, Definitions would only contains:

• def_id_to_hir_id
• hir_id_to_def_id
• table
• parent_modules_of_macro_defs
• expansions_that_defined
  which is the list of fields that should be available post HIR lowering.

TL;DR: "before" fields in Resolver, "after" fields kept in Definitions.

[petrochenkov]

@marmeladema

TL;DR: "before" fields in Resolver, "after" fields kept in Definitions.

Sounds like a plan!

[petrochenkov]

ID memo:

• LocalDefId - the best variant, if you can use it for something do it. Cannot address entities in other crates and fine-grained entities like individual expressions in the current crate. Can address any items.
• DefId - the second choice. Can address entities in other crates, but cannot address fine-grained entities like individual expressions.
• HirId - Can address fine-grained entities like individual expressions in the current crate. Fine-grained entities from other crates can never be addressed in rustc.
• NodeId - The last choice. Can address fine-grained entities like individual expressions when HirIds are not yet created.

cc @mark-i-m this can probably be added to the dev guide

[bors]

☔ The latest upstream changes (presumably #73511) made this pull request unmergeable. Please resolve the merge conflicts.

[petrochenkov]

#73357 has landed.

[marmeladema]

@petrochenkov rebased and fixed the conflicts 👍

[petrochenkov]

@bors r+

[bors]

📌 Commit 13104ef has been approved by petrochenkov