On 4/10/24 14:48, Patrick Palka wrote:
On Tue, 9 Apr 2024, Jason Merrill wrote:
On 3/5/24 10:31, Patrick Palka wrote:
On Tue, 27 Feb 2024, Patrick Palka wrote:
Subject: [PATCH] c++/modules: local type merging [PR99426]
One known missing piece in the modules implementation is merging of a
streamed-in local type (class or enum) with the corresponding in-TU
version of the local type. This missing piece turns out to cause a
hard-to-reduce use-after-free GC issue due to the entity_ary not being
marked as a GC root (deliberately), and manifests as a serialization
error on stream-in as in PR99426 (see comment #6 for a reduction). It's
also reproducible on trunk when running the xtreme-header tests without
-fno-module-lazy.
This patch makes us merge such local types according to their position
within the containing function's definition, analogous to how we merge
FIELD_DECLs of a class according to their index in the TYPE_FIELDS
list.
PR c++/99426
gcc/cp/ChangeLog:
* module.cc (merge_kind::MK_local_type): New enumerator.
(merge_kind_name): Update.
(trees_out::chained_decls): Move BLOCK-specific handling
of DECL_LOCAL_DECL_P decls to ...
(trees_out::core_vals) <case BLOCK>: ... here. Stream
BLOCK_VARS manually.
(trees_in::core_vals) <case BLOCK>: Stream BLOCK_VARS
manually. Handle deduplicated local types..
(trees_out::key_local_type): Define.
(trees_in::key_local_type): Define.
(trees_out::get_merge_kind) <case FUNCTION_DECL>: Return
MK_local_type for a local type.
(trees_out::key_mergeable) <case FUNCTION_DECL>: Use
key_local_type.
(trees_in::key_mergeable) <case FUNCTION_DECL>: Likewise.
(trees_in::is_matching_decl): Be flexible with type mismatches
for local entities.
diff --git a/gcc/cp/module.cc b/gcc/cp/module.cc
index 80b63a70a62..d9e34e9a4b9 100644
--- a/gcc/cp/module.cc
+++ b/gcc/cp/module.cc
@@ -6714,7 +6720,37 @@ trees_in::core_vals (tree t)
case BLOCK:
t->block.locus = state->read_location (*this);
t->block.end_locus = state->read_location (*this);
- t->block.vars = chained_decls ();
+
+ for (tree *chain = &t->block.vars;;)
+ if (tree decl = tree_node ())
+ {
+ /* For a deduplicated local type or enumerator, chain the
+ duplicate decl instead of the canonical in-TU decl. Seeing
+ a duplicate here means the containing function whose body
+ we're streaming in is a duplicate too, so we'll end up
+ discarding this BLOCK (and the rest of the duplicate function
+ body) anyway. */
+ if (is_duplicate (decl))
+ decl = maybe_duplicate (decl);
+ else if (DECL_IMPLICIT_TYPEDEF_P (decl)
+ && TYPE_TEMPLATE_INFO (TREE_TYPE (decl)))
+ {
+ tree tmpl = TYPE_TI_TEMPLATE (TREE_TYPE (decl));
+ if (DECL_TEMPLATE_RESULT (tmpl) == decl && is_duplicate
(tmpl))
+ decl = DECL_TEMPLATE_RESULT (maybe_duplicate (tmpl));
+ }
This seems like a lot of generally-applicable code for finding the duplicate,
which other calls to maybe_duplicate/odr_duplicate don't use. If the template
is a duplicate, why isn't its result? If there's a good reason for that,
should this template handling go into maybe_duplicate?
Ah yeah, that makes sense.
Some context: IIUC modules treats the TEMPLATE_DECL instead of the
DECL_TEMPLATE_RESULT as the canonical decl, which in turn means we'll
register_duplicate only the TEMPLATE_DECL. But BLOCK_VARS never contains
a TEMPLATE_DECL, always the DECL_TEMPLATE_RESULT (i.e. a TYPE_DECL),
hence the extra handling.
Given that it's relatively more difficult to get at the TEMPLATE_DECL
from the DECL_TEMPLATE_RESULT rather than vice versa, maybe we should
just register both as duplicates from register_duplicate? That way
callers can just simply pass the DECL_TEMPLATE_RESULT to maybe_duplicate
and it'll do the right thing.
Sounds good.
@@ -10337,6 +10373,83 @@ trees_in::fn_parms_fini (int tag, tree fn, tree
existing, bool is_defn)
}
}
+/* Encode into KEY the position of the local type (class or enum)
+ declaration DECL within FN. The position is encoded as the
+ index of the innermost BLOCK (numbered in BFS order) along with
+ the index within its BLOCK_VARS list. */
Since we already set DECL_DISCRIMINATOR for mangling, could we use it+name for
the key as well?
We could (and IIUc that'd be more robust to ODR violations), but
wouldn't it mean we'd have to do a linear walk over all BLOCK_VARs of
all BLOCKS in order to find the one with the matching
name+discriminator? That'd be slower than the current approach which
lets us skip to the correct BLOCK and walk only its BLOCK_VARS.
Ah, good point. How about block number + name instead of the index?
Jason