https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123791
--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> --- (In reply to Wilco from comment #2) > Yes that looks wrong. For any data addressing we always emit ADRP/ADD for > the address and then add the offset. For code addresses we should do the > same. > > With -mcmodel=large or -mpc-relative-literal-loads these literals end up in > the .text section - hopefully the linker complains... Yeah, I've wondered about that too. Again, rather than duplicating what default_elf_select_rtx_section does, wonder if for aarch64_can_use_per_function_literal_pools_p it shouldn't do the same and just return function_section (current_function_decl) only if it returned readonly_data_section (again, with -fmerge-constants temporarily disabled).
