| Summary: | ld: Provide a way to make C identifier name sections GCable under __start_/__stop_ references | ||
|---|---|---|---|
| Product: | binutils | Reporter: | Fangrui Song <i> |
| Component: | ld | Assignee: | Not yet assigned to anyone <unassigned> |
| Status: | RESOLVED FIXED | ||
| Severity: | normal | ||
| Priority: | P2 | ||
| Version: | unspecified | ||
| Target Milestone: | 2.37 | ||
| Host: | Target: | ||
| Build: | Last reconfirmed: | ||
In LLD, I'll add -z start-stop-gc to let __start_/__stop_ not retain C identifier name sections (https://reviews.llvm.org/D96914). -z nostart-stop-gc can disable it. For SHF_LINK_ORDER or SHF_GROUP sections, __start_/__stop_ references do not retain them, regardless of -z start-stop-gc or -z nostart-stop-gc. Patch: https://sourceware.org/pipermail/binutils/2021-February/115557.html This rule has caused trouble to clang -fprofile-generate and -fsanitize-coverage. Other than glibc, my analysis has found some other usage: * systemd and Swift (https://lists.llvm.org/pipermail/llvm-dev/2021-February/148682.html) * Clang's ObjC implementation in GNU environment (clang/lib/CodeGen/CGObjCGNU.cpp ): This has been taken care of by my https://reviews.llvm.org/D97448 Hope that at some point we can start to -z start-stop-gc. The master branch has been updated by Alan Modra <amodra@sourceware.org>: https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=8ee10e86093150c70360d9e26b29e6d9b6398f33 commit 8ee10e86093150c70360d9e26b29e6d9b6398f33 Author: Alan Modra <amodra@gmail.com> Date: Mon Mar 1 08:22:49 2021 +1030 PR27451, -z start_stop_gc When --gc-sections is in effect, a reference from a retained section to __start_SECNAME or __stop_SECNAME causes all input sections named SECNAME to also be retained, if SECNAME is representable as a C identifier and either __start_SECNAME or __stop_SECNAME is synthesized by the linker. Add an option to disable that feature, effectively ignoring any relocation that references a synthesized linker defined __start_ or __stop_ symbol. PR 27451 include/ * bfdlink.h (struct bfd_link_info): Add start_stop_gc. bfd/ * elflink.c (_bfd_elf_gc_mark_rsec): Ignore synthesized linker defined start/stop symbols when start_stop_gc. (bfd_elf_gc_mark_dynamic_ref_symbol): Likewise. (bfd_elf_define_start_stop): Don't modify ldscript_def syms. * linker.c (bfd_generic_define_start_stop): Likewise. ld/ * emultempl/elf.em: Handle -z start-stop-gc and -z nostart-stop-gc. * lexsup.c (elf_static_list_options): Display help for them. Move help for -z stack-size to here from elf_shlib_list_options. Add help for -z start-stop-visibility and -z undefs. * ld.texi: Document -z start-stop-gc and -z nostart-stop-gc. * NEWS: Mention -z start-stop-gc. * testsuite/ld-gc/start2.s, * testsuite/ld-gc/start2.d: New test. * testsuite/ld-gc/gc.exp: Run it. The master branch has been updated by Alan Modra <amodra@sourceware.org>: https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=b80e421f9136117389be9d96210b35b3d562d725 commit b80e421f9136117389be9d96210b35b3d562d725 Author: Alan Modra <amodra@gmail.com> Date: Tue Mar 2 21:25:20 2021 +1030 PR27451, -z start_stop_gc for powerpc64 PowerPC64 has its own gc_mark_dynamic_ref. bfd/ PR 27451 * elf64-ppc.c (ppc64_elf_gc_mark_dynamic_ref): Ignore synthesized linker defined start/stop symbols when start_stop_gc. ld/ * testsuite/ld-powerpc/startstop.d, * testsuite/ld-powerpc/startstop.r, * testsuite/ld-powerpc/startstop.s: New test. * testsuite/ld-powerpc/powerpc.exp: Run it. GNU ld 2.37 has -z start-stop-gc. |
The discussion below is about C identifier name sections. From my observation, GNU ld currently has a rule: __start_/__stop_ references from a live section retain all the associated C identifier name sections. In the following example, __start_meta reference retains both a.o:(meta) and b.o:(meta). # a.s .global _start .text _start: leaq __start_meta(%rip), %rdi leaq __stop_meta(%rip), %rsi .section meta,"a" .byte 0 # b.s .section meta,"a" .byte 1 In LLD, we have augmented the rule with SHF_LINK_ORDER and SHF_GROUP (after https://reviews.llvm.org/D96753, target release 13.0.0): __start_/__stop_ references from a live section retain all the associated non-SHF_LINK_ORDER non-SHF_GROUP C identifier name sections. At this point, some LLVM toolchain folks have concluded that the original rule does not cary its weight, and I actually have a thought to drop it: __start_/__stop_ references from a live section do not retain the associated C identifier name sections. Either use undefined weak __start_/__stop_, or ensure there is at least one live C identifier name section to avoid "undefined symbol" errors. There are more details in the "Metadata sections referenced by text sections" section of https://maskray.me/blog/2021-01-31-metadata-sections-comdat-and-shf-link-order There is good chance that not many OSS need adaptation (I have tested thousands and only swift/systemd have issues.) I understand that there are still risks, so we may need a linker option to drop the original rule. __start_/__stop_ is currently ELF specific, so the option is assumed to be under -z. Given https://sourceware.org/pipermail/binutils/2020-June/111685.html "-z start-stop-visibility=", we can probably name it -z start-stop-something, e.g. -z start-stop-gc and -z nostart-stop-gc. The option decides whether non-SHF_LINK_ORDER non-SHF_GROUP C identifier name sections are retained with __start_/__stop_ references. For SHF_LINK_ORDER or SHF_GROUP sections, it seems that should always be GCable. There is a possibility that LLD may switch to default -z start-stop-gc. This move can be harder for GNU ld because of stronger commitment with existing behaviors. (I have asked ClangBuiltLinux folks to test the Linux kernel https://github.com/ClangBuiltLinux/linux/issues/1307 . I have tested x86-64 defconfig myself and it works.