GRUB builds most of its code as loadable ELF modules. To reduce run-time size they are passed through "strip --strip-unneeded" (we also do some more mangling, full code is here [3]). Recently we added build-time module verifier to catch possible problems in toolchains output. After that we got bug report from FreeBSD user [1] that GRUB build started to fail. It was traced to module verifier reporting error for one specific module. One of checks is for non-empty symbol table and this module [2] got empty table after strip. According to user, this happened with both
elftoolchain r3223M GNU binutils 2.17.50 [FreeBSD] 2007-07-03 This module is actually empty placeholder that does not have any code and data and exists only for the custom section that lists module dependencies. I could verify this behavior after building 2.17 (binutils-2_17 tag) from GIT. It appears that recent binutils (earliest version I have is 2.24) unconditionally add back section symbols when copying sections. I verified that filter_symbols() in strip actually removes everything from source module, and symbol table reappears deep in call chain initiated by copy_section(), where STT_SECTION is unconditionally added for every present section. Was this change intentional? Is there any standard that requires presence of these symbols in every ELF object? Note that these symbols passed filter as unneeded and as GRUB developer I would rather welcome 2.17 behavior that allows us to further reduce run-time image size. Thank you! [1] https://savannah.gnu.org/bugs/?46986 [2] https://savannah.gnu.org/bugs/download.php?file_id=36182 [3] http://git.savannah.gnu.org/cgit/grub.git/tree/grub-core/genmod.sh.in _______________________________________________ Grub-devel mailing list Grub-devel@gnu.org https://lists.gnu.org/mailman/listinfo/grub-devel
