Rollup merge of #98814 - fmease:minimal-fix-for-issue-97933, r=GuillaumeGomez

rustdoc: Censor certain complex unevaluated const exprs

Fixes #97933.

This is more of a hotfix for the aforementioned issue. By that, I mean that my proposed patch is
not the best solution but one that does not change as much existing code.
It treats symptoms rather than the root cause.

This PR “censors” certain complex unevaluated constant expressions like `match`es, blocks, function calls, struct literals etc. by pretty-printing them as `_` / `{ _ }` (number and string literals, paths and `()` are still printed as one would expect).
Resorting to this placeholder is preferable to printing the full expression verbatim since
they can be quite large and verbose resulting in an unreadable mess in the generated documentation.
Further, mindlessly printing the const would leak private and `doc(hidden)` struct fields (#97933), at least in the current
stable & nightly implementations which rely on `span_to_snippet` (!) and `rustc_hir_pretty::id_to_string`.

The censoring of _verbose_ expressions is probably going to stay longer term.
However, in regards to private and `doc(hidden)` struct fields, I have a more proper fix in mind
which I have already partially implemented locally and for which I am going to open a separate PR sometime soon.
For that, I was already in contact with `@GuillaumeGomez.`
The proper fix involves rustdoc not falling back on pretty-printing unevaluated consts so easily (what this PR is concerned about)
and instead preferring to print evaluated consts which contain more information allowing it to selectively hide private and `doc(hidden)` fields, create hyperlinks etc. generally making the output more granular and precise (compared to the brutal `_` placeholder).

Unfortunately, I was a bit too late and the issue just hit stable (1.62).
Should this be backported to beta or even a potential 1.62.1?

r? `@GuillaumeGomez`

src/librustdoc/clean/utils.rs

@@ -340,17 +340,98 @@ pub(crate) fn is_literal_expr(tcx: TyCtxt<'_>, hir_id: hir::HirId) -> bool {
     false
 }
 
+/// Build a textual representation of an unevaluated constant expression.
+///
+/// If the const expression is too complex, an underscore `_` is returned.
+/// For const arguments, it's `{ _ }` to be precise.
+/// This means that the output is not necessarily valid Rust code.
+///
+/// Currently, only
+///
+/// * literals (optionally with a leading `-`)
+/// * unit `()`
+/// * blocks (`{ … }`) around simple expressions and
+/// * paths without arguments
+///
+/// are considered simple enough. Simple blocks are included since they are
+/// necessary to disambiguate unit from the unit type.
+/// This list might get extended in the future.
+///
+/// Without this censoring, in a lot of cases the output would get too large
+/// and verbose. Consider `match` expressions, blocks and deeply nested ADTs.
+/// Further, private and `doc(hidden)` fields of structs would get leaked
+/// since HIR datatypes like the `body` parameter do not contain enough
+/// semantic information for this function to be able to hide them –
+/// at least not without significant performance overhead.
+///
+/// Whenever possible, prefer to evaluate the constant first and try to
+/// use a different method for pretty-printing. Ideally this function
+/// should only ever be used as a fallback.
 pub(crate) fn print_const_expr(tcx: TyCtxt<'_>, body: hir::BodyId) -> String {
     let hir = tcx.hir();
     let value = &hir.body(body).value;
 
-    let snippet = if !value.span.from_expansion() {
-        tcx.sess.source_map().span_to_snippet(value.span).ok()
-    } else {
-        None
-    };
+    #[derive(PartialEq, Eq)]
+    enum Classification {
+        Literal,
+        Simple,
+        Complex,
+    }
 
-    snippet.unwrap_or_else(|| rustc_hir_pretty::id_to_string(&hir, body.hir_id))
+    use Classification::*;
+
+    fn classify(expr: &hir::Expr<'_>) -> Classification {
+        match &expr.kind {
+            hir::ExprKind::Unary(hir::UnOp::Neg, expr) => {
+                if matches!(expr.kind, hir::ExprKind::Lit(_)) { Literal } else { Complex }
+            }
+            hir::ExprKind::Lit(_) => Literal,
+            hir::ExprKind::Tup([]) => Simple,
+            hir::ExprKind::Block(hir::Block { stmts: [], expr: Some(expr), .. }, _) => {
+                if classify(expr) == Complex { Complex } else { Simple }
+            }
+            // Paths with a self-type or arguments are too “complex” following our measure since
+            // they may leak private fields of structs (with feature `adt_const_params`).
+            // Consider: `<Self as Trait<{ Struct { private: () } }>>::CONSTANT`.
+            // Paths without arguments are definitely harmless though.
+            hir::ExprKind::Path(hir::QPath::Resolved(_, hir::Path { segments, .. })) => {
+                if segments.iter().all(|segment| segment.args.is_none()) { Simple } else { Complex }
+            }
+            // FIXME: Claiming that those kinds of QPaths are simple is probably not true if the Ty
+            //        contains const arguments. Is there a *concise* way to check for this?
+            hir::ExprKind::Path(hir::QPath::TypeRelative(..)) => Simple,
+            // FIXME: Can they contain const arguments and thus leak private struct fields?
+            hir::ExprKind::Path(hir::QPath::LangItem(..)) => Simple,
+            _ => Complex,
+        }
+    }
+
+    let classification = classify(value);
+
+    if classification == Literal
+    && !value.span.from_expansion()
+    && let Ok(snippet) = tcx.sess.source_map().span_to_snippet(value.span) {
+        // For literals, we avoid invoking the pretty-printer and use the source snippet instead to
+        // preserve certain stylistic choices the user likely made for the sake legibility like
+        //
+        // * hexadecimal notation
+        // * underscores
+        // * character escapes
+        //
+        // FIXME: This passes through `-/*spacer*/0` verbatim.
+        snippet
+    } else if classification == Simple {
+        // Otherwise we prefer pretty-printing to get rid of extraneous whitespace, comments and
+        // other formatting artifacts.
+        rustc_hir_pretty::id_to_string(&hir, body.hir_id)
+    } else if tcx.def_kind(hir.body_owner_def_id(body).to_def_id()) == DefKind::AnonConst {
+        // FIXME: Omit the curly braces if the enclosing expression is an array literal
+        //        with a repeated element (an `ExprKind::Repeat`) as in such case it
+        //        would not actually need any disambiguation.
+        "{ _ }".to_owned()
+    } else {
+        "_".to_owned()
+    }
 }
 
 /// Given a type Path, resolve it to a Type using the TyCtxt