Avoid over-counting of UsePath in the HIR stats.

compiler/rustc_passes/src/input_stats.rs

@@ -426,10 +426,16 @@ impl<'v> hir_visit::Visitor<'v> for StatCollector<'v> {
         hir_visit::walk_fn(self, fk, fd, b, id)
     }
 
-    fn visit_use(&mut self, p: &'v hir::UsePath<'v>, hir_id: HirId) {
+    fn visit_use(&mut self, p: &'v hir::UsePath<'v>, _hir_id: HirId) {
         // This is `visit_use`, but the type is `Path` so record it that way.
         self.record("Path", None, p);
-        hir_visit::walk_use(self, p, hir_id)
+        // Don't call `hir_visit::walk_use(self, p, hir_id)`: it calls
+        // `visit_path` up to three times, once for each namespace result in
+        // `p.res`, by building temporary `Path`s that are not part of the real
+        // HIR, which causes `p` to be double- or triple-counted. Instead just
+        // walk the path internals (i.e. the segments) directly.
+        let hir::Path { span: _, res: _, segments } = *p;
+        ast_visit::walk_list!(self, visit_path_segment, segments);
     }
 
     fn visit_trait_item(&mut self, ti: &'v hir::TraitItem<'v>) {

tests/ui/stats/input-stats.rs

@@ -3,6 +3,10 @@
 //@ only-64bit
 // layout randomization affects the hir stat output
 //@ needs-deterministic-layouts
+//
+// Filter out the percentages because a change to a single count can affect
+// many or all percentages, which makes the diffs hard to read.
+//@ normalize-stderr: "\([0-9 ][0-9]\.[0-9]%\)" -> "(NN.N%)"
 
 // Type layouts sometimes change. When that happens, until the next bootstrap
 // bump occurs, stage1 and stage2 will give different outputs for this test.