1 files changed, 232 insertions, 0 deletions

diff --git a/crates/ra_ssr/src/search.rs b/crates/ra_ssr/src/search.rs
new file mode 100644
index 000000000..bcf0f0468
--- /dev/null
+++ b/crates/ra_ssr/src/search.rs
@@ -0,0 +1,232 @@
+//! Searching for matches.
+
+use crate::{
+    matching,
+    resolving::{ResolvedPath, ResolvedPattern, ResolvedRule},
+    Match, MatchFinder,
+};
+use ra_db::FileRange;
+use ra_ide_db::{
+    defs::Definition,
+    search::{Reference, SearchScope},
+};
+use ra_syntax::{ast, AstNode, SyntaxKind, SyntaxNode};
+use test_utils::mark;
+
+/// A cache for the results of find_usages. This is for when we have multiple patterns that have the
+/// same path. e.g. if the pattern was `foo::Bar` that can parse as a path, an expression, a type
+/// and as a pattern. In each, the usages of `foo::Bar` are the same and we'd like to avoid finding
+/// them more than once.
+#[derive(Default)]
+pub(crate) struct UsageCache {
+    usages: Vec<(Definition, Vec<Reference>)>,
+}
+
+impl<'db> MatchFinder<'db> {
+    /// Adds all matches for `rule` to `matches_out`. Matches may overlap in ways that make
+    /// replacement impossible, so further processing is required in order to properly nest matches
+    /// and remove overlapping matches. This is done in the `nesting` module.
+    pub(crate) fn find_matches_for_rule(
+        &self,
+        rule: &ResolvedRule,
+        usage_cache: &mut UsageCache,
+        matches_out: &mut Vec<Match>,
+    ) {
+        if pick_path_for_usages(&rule.pattern).is_none() {
+            self.slow_scan(rule, matches_out);
+            return;
+        }
+        self.find_matches_for_pattern_tree(rule, &rule.pattern, usage_cache, matches_out);
+    }
+
+    fn find_matches_for_pattern_tree(
+        &self,
+        rule: &ResolvedRule,
+        pattern: &ResolvedPattern,
+        usage_cache: &mut UsageCache,
+        matches_out: &mut Vec<Match>,
+    ) {
+        if let Some(resolved_path) = pick_path_for_usages(pattern) {
+            let definition: Definition = resolved_path.resolution.clone().into();
+            for reference in self.find_usages(usage_cache, definition) {
+                if let Some(node_to_match) = self.find_node_to_match(resolved_path, reference) {
+                    if !is_search_permitted_ancestors(&node_to_match) {
+                        mark::hit!(use_declaration_with_braces);
+                        continue;
+                    }
+                    if let Ok(m) =
+                        matching::get_match(false, rule, &node_to_match, &None, &self.sema)
+                    {
+                        matches_out.push(m);
+                    }
+                }
+            }
+        }
+    }
+
+    fn find_node_to_match(
+        &self,
+        resolved_path: &ResolvedPath,
+        reference: &Reference,
+    ) -> Option<SyntaxNode> {
+        let file = self.sema.parse(reference.file_range.file_id);
+        let depth = resolved_path.depth as usize;
+        let offset = reference.file_range.range.start();
+        if let Some(path) =
+            self.sema.find_node_at_offset_with_descend::<ast::Path>(file.syntax(), offset)
+        {
+            self.sema.ancestors_with_macros(path.syntax().clone()).skip(depth).next()
+        } else if let Some(path) =
+            self.sema.find_node_at_offset_with_descend::<ast::MethodCallExpr>(file.syntax(), offset)
+        {
+            // If the pattern contained a path and we found a reference to that path that wasn't
+            // itself a path, but was a method call, then we need to adjust how far up to try
+            // matching by how deep the path was within a CallExpr. The structure would have been
+            // CallExpr, PathExpr, Path - i.e. a depth offset of 2. We don't need to check if the
+            // path was part of a CallExpr because if it wasn't then all that will happen is we'll
+            // fail to match, which is the desired behavior.
+            const PATH_DEPTH_IN_CALL_EXPR: usize = 2;
+            if depth < PATH_DEPTH_IN_CALL_EXPR {
+                return None;
+            }
+            self.sema
+                .ancestors_with_macros(path.syntax().clone())
+                .skip(depth - PATH_DEPTH_IN_CALL_EXPR)
+                .next()
+        } else {
+            None
+        }
+    }
+
+    fn find_usages<'a>(
+        &self,
+        usage_cache: &'a mut UsageCache,
+        definition: Definition,
+    ) -> &'a [Reference] {
+        // Logically if a lookup succeeds we should just return it. Unfortunately returning it would
+        // extend the lifetime of the borrow, then we wouldn't be able to do the insertion on a
+        // cache miss. This is a limitation of NLL and is fixed with Polonius. For now we do two
+        // lookups in the case of a cache hit.
+        if usage_cache.find(&definition).is_none() {
+            let usages = definition.find_usages(&self.sema, Some(self.search_scope()));
+            usage_cache.usages.push((definition, usages));
+            return &usage_cache.usages.last().unwrap().1;
+        }
+        usage_cache.find(&definition).unwrap()
+    }
+
+    /// Returns the scope within which we want to search. We don't want un unrestricted search
+    /// scope, since we don't want to find references in external dependencies.
+    fn search_scope(&self) -> SearchScope {
+        // FIXME: We should ideally have a test that checks that we edit local roots and not library
+        // roots. This probably would require some changes to fixtures, since currently everything
+        // seems to get put into a single source root.
+        use ra_db::SourceDatabaseExt;
+        use ra_ide_db::symbol_index::SymbolsDatabase;
+        let mut files = Vec::new();
+        for &root in self.sema.db.local_roots().iter() {
+            let sr = self.sema.db.source_root(root);
+            files.extend(sr.iter());
+        }
+        SearchScope::files(&files)
+    }
+
+    fn slow_scan(&self, rule: &ResolvedRule, matches_out: &mut Vec<Match>) {
+        use ra_db::SourceDatabaseExt;
+        use ra_ide_db::symbol_index::SymbolsDatabase;
+        for &root in self.sema.db.local_roots().iter() {
+            let sr = self.sema.db.source_root(root);
+            for file_id in sr.iter() {
+                let file = self.sema.parse(file_id);
+                let code = file.syntax();
+                self.slow_scan_node(code, rule, &None, matches_out);
+            }
+        }
+    }
+
+    fn slow_scan_node(
+        &self,
+        code: &SyntaxNode,
+        rule: &ResolvedRule,
+        restrict_range: &Option<FileRange>,
+        matches_out: &mut Vec<Match>,
+    ) {
+        if !is_search_permitted(code) {
+            return;
+        }
+        if let Ok(m) = matching::get_match(false, rule, &code, restrict_range, &self.sema) {
+            matches_out.push(m);
+        }
+        // If we've got a macro call, we already tried matching it pre-expansion, which is the only
+        // way to match the whole macro, now try expanding it and matching the expansion.
+        if let Some(macro_call) = ast::MacroCall::cast(code.clone()) {
+            if let Some(expanded) = self.sema.expand(&macro_call) {
+                if let Some(tt) = macro_call.token_tree() {
+                    // When matching within a macro expansion, we only want to allow matches of
+                    // nodes that originated entirely from within the token tree of the macro call.
+                    // i.e. we don't want to match something that came from the macro itself.
+                    self.slow_scan_node(
+                        &expanded,
+                        rule,
+                        &Some(self.sema.original_range(tt.syntax())),
+                        matches_out,
+                    );
+                }
+            }
+        }
+        for child in code.children() {
+            self.slow_scan_node(&child, rule, restrict_range, matches_out);
+        }
+    }
+}
+
+/// Returns whether we support matching within `node` and all of its ancestors.
+fn is_search_permitted_ancestors(node: &SyntaxNode) -> bool {
+    if let Some(parent) = node.parent() {
+        if !is_search_permitted_ancestors(&parent) {
+            return false;
+        }
+    }
+    is_search_permitted(node)
+}
+
+/// Returns whether we support matching within this kind of node.
+fn is_search_permitted(node: &SyntaxNode) -> bool {
+    // FIXME: Properly handle use declarations. At the moment, if our search pattern is `foo::bar`
+    // and the code is `use foo::{baz, bar}`, we'll match `bar`, since it resolves to `foo::bar`.
+    // However we'll then replace just the part we matched `bar`. We probably need to instead remove
+    // `bar` and insert a new use declaration.
+    node.kind() != SyntaxKind::USE_ITEM
+}
+
+impl UsageCache {
+    fn find(&mut self, definition: &Definition) -> Option<&[Reference]> {
+        // We expect a very small number of cache entries (generally 1), so a linear scan should be
+        // fast enough and avoids the need to implement Hash for Definition.
+        for (d, refs) in &self.usages {
+            if d == definition {
+                return Some(refs);
+            }
+        }
+        None
+    }
+}
+
+/// Returns a path that's suitable for path resolution. We exclude builtin types, since they aren't
+/// something that we can find references to. We then somewhat arbitrarily pick the path that is the
+/// longest as this is hopefully more likely to be less common, making it faster to find.
+fn pick_path_for_usages(pattern: &ResolvedPattern) -> Option<&ResolvedPath> {
+    // FIXME: Take the scope of the resolved path into account. e.g. if there are any paths that are
+    // private to the current module, then we definitely would want to pick them over say a path
+    // from std. Possibly we should go further than this and intersect the search scopes for all
+    // resolved paths then search only in that scope.
+    pattern
+        .resolved_paths
+        .iter()
+        .filter(|(_, p)| {
+            !matches!(p.resolution, hir::PathResolution::Def(hir::ModuleDef::BuiltinType(_)))
+        })
+        .map(|(node, resolved)| (node.text().len(), resolved))
+        .max_by(|(a, _), (b, _)| a.cmp(b))
+        .map(|(_, resolved)| resolved)
+}