Function ra_ap_ide_completion::completions

pub fn completions(
    db: &RootDatabase, 
    config: &CompletionConfig, 
    position: FilePosition
) -> Option<Completions>

Main entry point for completion. We run completion as a two-phase process.

First, we look at the position and collect a so-called `CompletionContext. This is a somewhat messy process, because, during completion, syntax tree is incomplete and can look really weird.

Once the context is collected, we run a series of completion routines which look at the context and produce completion items. One subtlety about this phase is that completion engine should not filter by the substring which is already present, it should give all possible variants for the identifier at the caret. In other words, for

fn f() {
    let foo = 92;
    let _ = bar$0
}

foo should be present among the completion variants. Filtering by identifier prefix/fuzzy match should be done higher in the stack, together with ordering of completions (currently this is done by the client).

Hypothetical Completion Problem

There’s a curious unsolved problem in the current implementation. Often, you want to compute completions on a slightly different text document.

In the simplest case, when the code looks like let x = , you want to insert a fake identifier to get a better syntax tree: let x = complete_me.

We do this in CompletionContext, and it works OK-enough for syntax analysis. However, we might want to, eg, ask for the type of complete_me variable, and that’s where our current infrastructure breaks down. salsa doesn’t allow such “phantom” inputs.

Another case where this would be instrumental is macro expansion. We want to insert a fake ident and re-expand code. There’s expand_hypothetical as a work-around for this.

A different use-case is completion of injection (examples and links in doc comments). When computing completion for a path in a doc-comment, you want to inject a fake path expression into the item being documented and complete that.

IntelliJ has CodeFragment/Context infrastructure for that. You can create a temporary PSI node, and say that the context (“parent”) of this node is some existing node. Asking for, eg, type of this CodeFragment node works correctly, as the underlying infrastructure makes use of contexts to do analysis.