Hi all,
I’m chasing down a fix for PR108383 that is related to
-fdeclone-ctor-dtor. For reference, here is how the ICE looks:
lto1: error: Two symbols with same comdat_group are not linked by the
same_comdat_group list.
lto1: internal compiler error: symtab_node::verify failed
I think the bug occurs across two locations in gcc/cp/optimize.cc:
1. In maybe_clone_body() around line 568:
if (DECL_ONE_ONLY (fn))
cgraph_node::get_create (clone)->set_comdat_group (cxx_comdat_group
(clone));
This sets comdat_group on clones REGARDLESS of HAVE_COMDAT_GROUP.
2. In maybe_thunk_body() around line 311-356:
if (!DECL_WEAK (fn))
{ /* Handle non-weak: make local */ }
else if (HAVE_COMDAT_GROUP)
{ /* Link symbols via same_comdat_group list */ }
When DECL_WEAK && !HAVE_COMDAT_GROUP, maybe_thunk_body() falls through
without calling add_to_same_comdat_group(). The clones already have
comdat_group identifiers (from maybe_clone_body), but they're never
linked via same_comdat_group. I think this is why the symtab verifier
is giving an ICE.
There are a few other locations that call set_comdat_group without
checking HAVE_COMDAT_GROUP.
I tried patching them all, but instead I ended up with multiple
definition errors when linking:
/home/peter/wintoolchain/lib/gcc/x86_64-w64-mingw32/16.0.0/../../../../x86_64-w64-mingw32/bin/ld:
externals/simplecpp/simplecpp.o (symbol from plugin):(.text+0x0):
multiple definition of `typeinfo name for std::runtime_error';
cli/threadexecutor.o (symbol from plugin):(.text+0x0): first defined
here
Does anyone have ideas on how I can implement -fdeclone-ctor-dtor
without using comdat groups?
Or if I'm going wrong anywhere with my approach?
Regards,
Peter Damianov.
