Rollup merge of #35114 - michaelwoerister:inline-meta-to-hir-map, r=eddyb

Move caching of inlined HIR into CrateStore

So far we've had separate HIR-inlining caches for each codegen unit and the caching for things inlined during constant evaluation had some holes. Consequently, things would be inlined multiple times if they were used from different codegen units, etc, leading to
- wasted memory,
- multiple `NodeId`s per `DefId` and,
- for things inlined during constant evaluation, no way to map a `NodeId` back to it's original `DefId`.

This PR moves all caching into the CrateStore, solving all of the above problems. It also fixes some bugs in the inlining code, like cyclic in the parent-chains in the HIR map and some `NodeId`'s being translated to more or less random values. There are assertions in place now that should prevent this kind of thing in the future.

This PR based on top of #35090, which contains some necessary fixes.

1 files changed, 42 insertions, 9 deletions

diff --git a/src/librustc_metadata/astencode.rs b/src/librustc_metadata/astencode.rs
index 454c805ab57..c39ad414492 100644
--- a/src/librustc_metadata/astencode.rs
+++ b/src/librustc_metadata/astencode.rs
@@ -88,8 +88,9 @@ pub fn encode_inlined_item(ecx: &e::EncodeContext,
            rbml_w.writer.seek(SeekFrom::Current(0)));
 
     // Folding could be avoided with a smarter encoder.
-    let ii = simplify_ast(ii);
+    let (ii, expected_id_range) = simplify_ast(ii);
     let id_range = inlined_item_id_range(&ii);
+    assert_eq!(expected_id_range, id_range);
 
     rbml_w.start_tag(c::tag_ast as usize);
     id_range.encode(rbml_w);
@@ -186,6 +187,10 @@ impl<'a, 'b, 'tcx> DecodeContext<'a, 'b, 'tcx> {
     pub fn tr_id(&self, id: ast::NodeId) -> ast::NodeId {
         // from_id_range should be non-empty
         assert!(!self.from_id_range.empty());
+        // Make sure that translating the NodeId will actually yield a
+        // meaningful result
+        assert!(self.from_id_range.contains(id));
+
         // Use wrapping arithmetic because otherwise it introduces control flow.
         // Maybe we should just have the control flow? -- aatch
         (id.wrapping_sub(self.from_id_range.min).wrapping_add(self.to_id_range.min))
@@ -279,9 +284,23 @@ fn encode_ast(rbml_w: &mut Encoder, item: &InlinedItem) {
     rbml_w.end_tag();
 }
 
-struct NestedItemsDropper;
+struct NestedItemsDropper {
+    id_range: IdRange
+}
 
 impl Folder for NestedItemsDropper {
+
+    // The unit tests below run on HIR with NodeIds not properly assigned. That
+    // causes an integer overflow. So we just don't track the id_range when
+    // building the unit tests.
+    #[cfg(not(test))]
+    fn new_id(&mut self, id: ast::NodeId) -> ast::NodeId {
+        // Record the range of NodeIds we are visiting, so we can do a sanity
+        // check later
+        self.id_range.add(id);
+        id
+    }
+
     fn fold_block(&mut self, blk: P<hir::Block>) -> P<hir::Block> {
         blk.and_then(|hir::Block {id, stmts, expr, rules, span, ..}| {
             let stmts_sans_items = stmts.into_iter().filter_map(|stmt| {
@@ -322,10 +341,12 @@ impl Folder for NestedItemsDropper {
 // As it happens, trans relies on the fact that we do not export
 // nested items, as otherwise it would get confused when translating
 // inlined items.
-fn simplify_ast(ii: InlinedItemRef) -> InlinedItem {
-    let mut fld = NestedItemsDropper;
+fn simplify_ast(ii: InlinedItemRef) -> (InlinedItem, IdRange) {
+    let mut fld = NestedItemsDropper {
+        id_range: IdRange::max()
+    };
 
-    match ii {
+    let ii = match ii {
         // HACK we're not dropping items.
         InlinedItemRef::Item(i) => {
             InlinedItem::Item(P(fold::noop_fold_item(i.clone(), &mut fld)))
@@ -339,7 +360,9 @@ fn simplify_ast(ii: InlinedItemRef) -> InlinedItem {
         InlinedItemRef::Foreign(i) => {
             InlinedItem::Foreign(P(fold::noop_fold_foreign_item(i.clone(), &mut fld)))
         }
-    }
+    };
+
+    (ii, fld.id_range)
 }
 
 fn decode_ast(item_doc: rbml::Doc) -> InlinedItem {
@@ -361,8 +384,18 @@ impl tr for Def {
         match *self {
           Def::Fn(did) => Def::Fn(did.tr(dcx)),
           Def::Method(did) => Def::Method(did.tr(dcx)),
-          Def::SelfTy(opt_did, impl_id) => { Def::SelfTy(opt_did.map(|did| did.tr(dcx)),
-                                                         impl_id.map(|id| dcx.tr_id(id))) }
+          Def::SelfTy(opt_did, impl_id) => {
+              // Since the impl_id will never lie within the reserved range of
+              // imported NodeIds, it does not make sense to translate it.
+              // The result would not make any sense within the importing crate.
+              // We also don't allow for impl items to be inlined (just their
+              // members), so even if we had a DefId here, we wouldn't be able
+              // to do much with it.
+              // So, we set the id to DUMMY_NODE_ID. That way we make it
+              // explicit that this is no usable NodeId.
+              Def::SelfTy(opt_did.map(|did| did.tr(dcx)),
+                          impl_id.map(|_| ast::DUMMY_NODE_ID))
+          }
           Def::Mod(did) => { Def::Mod(did.tr(dcx)) }
           Def::ForeignMod(did) => { Def::ForeignMod(did.tr(dcx)) }
           Def::Static(did, m) => { Def::Static(did.tr(dcx), m) }
@@ -1361,7 +1394,7 @@ fn test_simplification() {
     with_testing_context(|lcx| {
         let hir_item = lcx.lower_item(&item);
         let item_in = InlinedItemRef::Item(&hir_item);
-        let item_out = simplify_ast(item_in);
+        let (item_out, _) = simplify_ast(item_in);
         let item_exp = InlinedItem::Item(P(lcx.lower_item(&quote_item!(&cx,
             fn new_int_alist<B>() -> alist<isize, B> {
                 return alist {eq_fn: eq_int, data: Vec::new()};