Auto merge of #83932 - lcnr:probe-perf, r=estebank

use a `SmallVec` in `impl_or_trait_item`

#83293 showed that this is fairly hot, slightly improves max-rss and cpu cycles, does not noticeably improve instruction counts

Author: bors

compiler/rustc_typeck/src/check/method/probe.rs

@@ -1768,7 +1768,9 @@ impl<'a, 'tcx> ProbeContext<'a, 'tcx> {
 
     /// Finds the method with the appropriate name (or return type, as the case may be). If
     /// `allow_similar_names` is set, find methods with close-matching names.
-    fn impl_or_trait_item(&self, def_id: DefId) -> Vec<ty::AssocItem> {
+    // The length of the returned iterator is nearly always 0 or 1 and this
+    // method is fairly hot.
+    fn impl_or_trait_item(&self, def_id: DefId) -> SmallVec<[ty::AssocItem; 1]> {
         if let Some(name) = self.method_name {
             if self.allow_similar_names {
                 let max_dist = max(name.as_str().len(), 3) / 3;
@@ -1784,7 +1786,7 @@ impl<'a, 'tcx> ProbeContext<'a, 'tcx> {
             } else {
                 self.fcx
                     .associated_item(def_id, name, Namespace::ValueNS)
-                    .map_or_else(Vec::new, |x| vec![x])
+                    .map_or_else(SmallVec::new, |x| SmallVec::from_buf([x]))
             }
         } else {
             self.tcx.associated_items(def_id).in_definition_order().copied().collect()