rustdoc: Rearrange Item/ItemInner.

The `Item` struct is 48 bytes and contains a `Box<ItemInner>`;
`ItemInner` is 104 bytes. This is an odd arrangement. Normally you'd
have one of the following.

- A single large struct, which avoids the allocation for the `Box`, but
  can result in lots of wasted space in unused parts of a container like
  `Vec<Item>`, `HashSet<Item>`, etc.

- Or, something like `struct Item(Box<ItemInner>)`, which requires the
  `Box` allocation but gives a very small Item size, which is good for
  containers like `Vec<Item>`.

`Item`/`ItemInner` currently gets the worst of both worlds: it always
requires a `Box`, but `Item` is also pretty big and so wastes space in
containers. It would make sense to push it in one direction or the
other. rust-lang#138916 showed that the first option is a regression for rustdoc,
so this commit does the second option, which improves speed and reduces
memory usage.

Author: nnethercote

src/librustdoc/clean/auto_trait.rs

@@ -114,8 +114,8 @@ fn synthesize_auto_trait_impl<'tcx>(
     };
 
     Some(clean::Item {
-        name: None,
         inner: Box::new(clean::ItemInner {
+            name: None,
             attrs: Default::default(),
             stability: None,
             kind: clean::ImplItem(Box::new(clean::Impl {
@@ -127,10 +127,10 @@ fn synthesize_auto_trait_impl<'tcx>(
                 polarity,
                 kind: clean::ImplKind::Auto,
             })),
+            item_id: clean::ItemId::Auto { trait_: trait_def_id, for_: item_def_id },
+            cfg: None,
+            inline_stmt_id: None,
         }),
-        item_id: clean::ItemId::Auto { trait_: trait_def_id, for_: item_def_id },
-        cfg: None,
-        inline_stmt_id: None,
     })
 }
 

src/librustdoc/clean/blanket_impl.rs

@@ -83,9 +83,9 @@ pub(crate) fn synthesize_blanket_impls(
             cx.generated_synthetics.insert((ty.skip_binder(), trait_def_id));
 
             blanket_impls.push(clean::Item {
-                name: None,
-                item_id: clean::ItemId::Blanket { impl_id: impl_def_id, for_: item_def_id },
                 inner: Box::new(clean::ItemInner {
+                    name: None,
+                    item_id: clean::ItemId::Blanket { impl_id: impl_def_id, for_: item_def_id },
                     attrs: Default::default(),
                     stability: None,
                     kind: clean::ImplItem(Box::new(clean::Impl {
@@ -122,9 +122,9 @@ pub(crate) fn synthesize_blanket_impls(
                             None,
                         ))),
                     })),
+                    cfg: None,
+                    inline_stmt_id: None,
                 }),
-                cfg: None,
-                inline_stmt_id: None,
             });
         }
     }

src/librustdoc/clean/inline.rs

@@ -151,7 +151,7 @@ pub(crate) fn try_inline(
     let mut item =
         crate::clean::generate_item_with_correct_attrs(cx, kind, did, name, import_def_id, None);
     // The visibility needs to reflect the one from the reexport and not from the "source" DefId.
-    item.inline_stmt_id = import_def_id;
+    item.inner.inline_stmt_id = import_def_id;
     ret.push(item);
     Some(ret)
 }
@@ -665,11 +665,11 @@ fn build_module_items(
                 // Primitive types can't be inlined so generate an import instead.
                 let prim_ty = clean::PrimitiveType::from(p);
                 items.push(clean::Item {
-                    name: None,
-                    // We can use the item's `DefId` directly since the only information ever used
-                    // from it is `DefId.krate`.
-                    item_id: ItemId::DefId(did),
                     inner: Box::new(clean::ItemInner {
+                        name: None,
+                        // We can use the item's `DefId` directly since the only information ever
+                        // used from it is `DefId.krate`.
+                        item_id: ItemId::DefId(did),
                         attrs: Default::default(),
                         stability: None,
                         kind: clean::ImportItem(clean::Import::new_simple(
@@ -689,9 +689,9 @@ fn build_module_items(
                             },
                             true,
                         )),
+                        cfg: None,
+                        inline_stmt_id: None,
                     }),
-                    cfg: None,
-                    inline_stmt_id: None,
                 });
             } else if let Some(i) = try_inline(cx, res, item.ident.name, attrs, visited) {
                 items.extend(i)

src/librustdoc/clean/mod.rs

@@ -210,7 +210,7 @@ fn generate_item_with_correct_attrs(
 
     let name = renamed.or(Some(name));
     let mut item = Item::from_def_id_and_attrs_and_parts(def_id, name, kind, attrs, cfg);
-    item.inline_stmt_id = import_id;
+    item.inner.inline_stmt_id = import_id;
     item
 }
 

src/librustdoc/clean/types.rs

@@ -311,26 +311,31 @@ pub(crate) enum ExternalLocation {
 /// directly to the AST's concept of an item; it's a strict superset.
 #[derive(Clone)]
 pub(crate) struct Item {
-    /// The name of this item.
-    /// Optional because not every item has a name, e.g. impls.
-    pub(crate) name: Option<Symbol>,
     pub(crate) inner: Box<ItemInner>,
-    pub(crate) item_id: ItemId,
-    /// This is the `LocalDefId` of the `use` statement if the item was inlined.
-    /// The crate metadata doesn't hold this information, so the `use` statement
-    /// always belongs to the current crate.
-    pub(crate) inline_stmt_id: Option<LocalDefId>,
-    pub(crate) cfg: Option<Arc<Cfg>>,
 }
 
+// Why does the `Item`/`ItemInner` split exist? `Vec<Item>`s are common, and
+// without the split `Item` would be a large type (100+ bytes) which results in
+// lots of wasted space in the unused parts of a `Vec<Item>`. With the split,
+// `Item` is just 8 bytes, and the wasted space is avoided, at the cost of an
+// extra allocation per item. This is a performance win.
 #[derive(Clone)]
 pub(crate) struct ItemInner {
+    /// The name of this item.
+    /// Optional because not every item has a name, e.g. impls.
+    pub(crate) name: Option<Symbol>,
     /// Information about this item that is specific to what kind of item it is.
     /// E.g., struct vs enum vs function.
     pub(crate) kind: ItemKind,
     pub(crate) attrs: Attributes,
     /// The effective stability, filled out by the `propagate-stability` pass.
     pub(crate) stability: Option<Stability>,
+    pub(crate) item_id: ItemId,
+    /// This is the `LocalDefId` of the `use` statement if the item was inlined.
+    /// The crate metadata doesn't hold this information, so the `use` statement
+    /// always belongs to the current crate.
+    pub(crate) inline_stmt_id: Option<LocalDefId>,
+    pub(crate) cfg: Option<Arc<Cfg>>,
 }
 
 impl std::ops::Deref for Item {
@@ -488,11 +493,15 @@ impl Item {
         trace!("name={name:?}, def_id={def_id:?} cfg={cfg:?}");
 
         Item {
-            item_id: def_id.into(),
-            inner: Box::new(ItemInner { kind, attrs, stability: None }),
-            name,
-            cfg,
-            inline_stmt_id: None,
+            inner: Box::new(ItemInner {
+                item_id: def_id.into(),
+                kind,
+                attrs,
+                stability: None,
+                name,
+                cfg,
+                inline_stmt_id: None,
+            }),
         }
     }
 
@@ -2618,13 +2627,14 @@ mod size_asserts {
 
     use super::*;
     // tidy-alphabetical-start
-    static_assert_size!(Crate, 56); // frequently moved by-value
+    static_assert_size!(Crate, 16); // frequently moved by-value
     static_assert_size!(DocFragment, 32);
     static_assert_size!(GenericArg, 32);
     static_assert_size!(GenericArgs, 24);
     static_assert_size!(GenericParamDef, 40);
     static_assert_size!(Generics, 16);
-    static_assert_size!(Item, 48);
+    static_assert_size!(Item, 8);
+    static_assert_size!(ItemInner, 136);
     static_assert_size!(ItemKind, 48);
     static_assert_size!(PathSegment, 32);
     static_assert_size!(Type, 32);

src/librustdoc/formats/cache.rs

@@ -385,7 +385,6 @@ impl DocFolder for CacheBuilder<'_, '_> {
         // implementations elsewhere.
         let ret = if let clean::Item {
             inner: box clean::ItemInner { kind: clean::ImplItem(ref i), .. },
-            ..
         } = item
         {
             // Figure out the id of this impl. This may map to a

src/librustdoc/json/conversions.rs

@@ -43,7 +43,7 @@ impl JsonRenderer<'_> {
         let attrs = item.attributes(self.tcx, self.cache(), true);
         let span = item.span(self.tcx);
         let visibility = item.visibility(self.tcx);
-        let clean::Item { name, item_id, .. } = item;
+        let clean::ItemInner { name, item_id, .. } = *item.inner;
         let id = self.id_from_item(&item);
         let inner = match item.kind {
             clean::KeywordItem => return None,

src/librustdoc/passes/propagate_doc_cfg.rs

@@ -61,7 +61,7 @@ impl CfgPropagator<'_, '_> {
 
         let (_, cfg) =
             merge_attrs(self.cx, item.attrs.other_attrs.as_slice(), Some((&attrs, None)));
-        item.cfg = cfg;
+        item.inner.cfg = cfg;
     }
 }
 
@@ -71,7 +71,7 @@ impl DocFolder for CfgPropagator<'_, '_> {
 
         self.merge_with_parent_attributes(&mut item);
 
-        let new_cfg = match (self.parent_cfg.take(), item.cfg.take()) {
+        let new_cfg = match (self.parent_cfg.take(), item.inner.cfg.take()) {
             (None, None) => None,
             (Some(rc), None) | (None, Some(rc)) => Some(rc),
             (Some(mut a), Some(b)) => {
@@ -81,7 +81,7 @@ impl DocFolder for CfgPropagator<'_, '_> {
             }
         };
         self.parent_cfg = new_cfg.clone();
-        item.cfg = new_cfg;
+        item.inner.cfg = new_cfg;
 
         let old_parent =
             if let Some(def_id) = item.item_id.as_def_id().and_then(|def_id| def_id.as_local()) {