https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78188
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |hubicka at gcc dot gnu.org
--- Comment #11 from Richard Biener <rguenth at gcc dot gnu.org> ---
Ok, I can reproduce it with the cross now. -O2 -fno-exceptions. I'll try to
reduce.
(gdb) p debug_tree (decl)
<function_decl 0x2aaaaff46d00 _ZL11vn_valueizeP9tree_node.part.80
type <function_type 0x2aaaacc873f0
so it's actually a partially inlined clone, not vn_valueize itself. Likewise
for gimple_ops:
(gdb) p debug_tree (sect->named.decl)
<function_decl 0x2aaaafdebc00 _ZL10gimple_opsP6gimple.part.20
(gdb) p/x sect->common.flags
$5 = 0x300102
(gdb) p/x flags
$6 = 0x200902
SECTION_DECLARED is missing (doesn't matter) but SECTION_LINKONCE is set which
is what causes the error. The named section is 'text.unlikely'.
We get that flags from
422 flags = targetm.section_type_flags (decl, name, reloc);
423 return get_section (name, flags, decl);
which in turn gets it from default_section_type_flags which adds
SECTION_LINKONCE because
if (decl && DECL_P (decl) && DECL_COMDAT_GROUP (decl))
flags |= SECTION_LINKONCE;
but there is no DECL_COMDAT_GROUP for the gimple_ops function decl.
Note we assign the section name in default_function_section. So this seems to
be a general issue somehow. Honza? How is this situation avoided on other
targets?
