On 4/22/24 08:42, Jakub Jelinek wrote:
On Wed, Apr 17, 2024 at 09:42:47AM +0200, Jakub Jelinek wrote:When expand_or_defer_fn is called at_eof time, it calls import_export_decl and then maybe_clone_body, which uses DECL_ONE_ONLY and comdat name in a couple of places to try to optimize cdtors which are known to have the same body by making the complete cdtor an alias to base cdtor (and in that case also uses *[CD]5* as comdat group name instead of the normal comdat group names specific to each mangled name). Now, this optimization depends on DECL_ONE_ONLY and DECL_INTERFACE_KNOWN, maybe_clone_body and can_alias_cdtor use: if (DECL_ONE_ONLY (fn)) cgraph_node::get_create (clone)->set_comdat_group (cxx_comdat_group (clone)); ... bool can_alias = can_alias_cdtor (fn); ... /* Tell cgraph if both ctors or both dtors are known to have the same body. */ if (can_alias && fns[0] && idx == 1 && cgraph_node::get_create (fns[0])->create_same_body_alias (clone, fns[0])) { alias = true; if (DECL_ONE_ONLY (fns[0])) { /* For comdat base and complete cdtors put them into the same, *[CD]5* comdat group instead of *[CD][12]*. */ comdat_group = cdtor_comdat_group (fns[1], fns[0]); cgraph_node::get_create (fns[0])->set_comdat_group (comdat_group); if (symtab_node::get (clone)->same_comdat_group) symtab_node::get (clone)->remove_from_same_comdat_group (); symtab_node::get (clone)->add_to_same_comdat_group (symtab_node::get (fns[0])); } } and /* Don't use aliases for weak/linkonce definitions unless we can put both symbols in the same COMDAT group. */ return (DECL_INTERFACE_KNOWN (fn) && (SUPPORTS_ONE_ONLY || !DECL_WEAK (fn)) && (!DECL_ONE_ONLY (fn) || (HAVE_COMDAT_GROUP && DECL_WEAK (fn)))); The following testcase regressed with Marek's r14-5979 change, when pr113208_0.C is compiled where the ctor is marked constexpr, we no longer perform this optimization, where _ZN6vectorI12QualityValueEC2ERKS1_ was emitted in the _ZN6vectorI12QualityValueEC5ERKS1_ comdat group and _ZN6vectorI12QualityValueEC1ERKS1_ was made an alias to it, instead we emit _ZN6vectorI12QualityValueEC2ERKS1_ in _ZN6vectorI12QualityValueEC2ERKS1_ comdat group and the same content _ZN6vectorI12QualityValueEC1ERKS1_ as separate symbol in _ZN6vectorI12QualityValueEC1ERKS1_ comdat group.
This seems like an ABI bug that could use a non-LTO testcase.
Now, the linker seems to somehow cope with that, eventhough it probably keeps both copies of the ctor, but seems LTO can't cope with that and Honza doesn't know what it should do in that case (linker decides that the prevailing symbol is _ZN6vectorI12QualityValueEC2ERKS1_ (from the _ZN6vectorI12QualityValueEC2ERKS1_ comdat group) and _ZN6vectorI12QualityValueEC1ERKS1_ alias (from the other TU, from _ZN6vectorI12QualityValueEC5ERKS1_ comdat group)). Note, the case where some constructor is marked constexpr in one TU and not in another one happens pretty often in libstdc++ when one mixes -std= flags used to compile different compilation units. The reason the optimization doesn't trigger when the constructor is constexpr is that expand_or_defer_fn is called in that case much earlier than when it is not constexpr; in the former case it is called when we try to constant evaluate that constructor. But DECL_INTERFACE_KNOWN is false in that case and comdat_linkage hasn't been called either (I think it is desirable, because comdat group is stored in the cgraph node and am not sure it is a good idea to create cgraph nodes for something that might not be needed later on at all), so maybe_clone_body clones the bodies, but doesn't make them as aliases.
Hmm, cloning the bodies and then discarding them later seems like more extra work than creating the cgraph nodes.
Jason
