https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65509
--- Comment #9 from Jakub Jelinek <jakub at gcc dot gnu.org> --- (In reply to Marek Polacek from comment #8) > The following (untested) should work: > > --- a/gcc/cp/decl.c > +++ b/gcc/cp/decl.c > @@ -6113,6 +6113,7 @@ make_rtl_for_nonlocal_decl (tree decl, tree init, > const char* asmspec) > not emitted into the object file unnecessarily. */ > filename = LOCATION_FILE (input_location); > if (!DECL_VIRTUAL_P (decl) > + && !DECL_DECLARED_CONSTEXPR_P (decl) > && TREE_READONLY (decl) > && DECL_INITIAL (decl) != NULL_TREE > && DECL_INITIAL (decl) != error_mark_node > > My reasoning is that we shouldn't defer assembling the variable if it might > be needed in the constexpr context. But in the #c5 testcase it isn't DECL_DECLARED_CONSTEXPR_P, is it?
