1 files changed, 28 insertions, 0 deletions

diff --git a/crates/ide_completion/src/lib.rs b/crates/ide_completion/src/lib.rs
index 831d543bb..6f3d5c5c5 100644
--- a/crates/ide_completion/src/lib.rs
+++ b/crates/ide_completion/src/lib.rs
@@ -106,6 +106,34 @@ pub use crate::{
 /// `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.
 pub fn completions(
     db: &RootDatabase,
     config: &CompletionConfig,