Support DT_RELR using standard -z pack-relative-relocs cmdline flag

[rawoul]

In bfd and gold, packed dynamic relocations are toggled with the -z pack-relative-relocs and -z nopack-relative-relocs cmd line flags, while mold uses -pack-dyn-relocs=none|relr. For compatibility it would be simpler if mold aligned to the other linkers, can the standard flag also be supported, or used instead of -pack-dyn-relocs ?

[rui314]

If someone really has a trouble with these flags, I'll consider, but I don't think we should add a new flag for a hypothetical use. It's somewhat an evolutionary process; if a flag doesn't become popular enough, it won't be supported by many implementations, and its use will diminish. So I don't think we should try to support all flags beforehand.

[rawoul]

When adding -z pack-relative-relocs bfd also adds a version dependency on GLIBC_ABI_DT_RELR if libc.so is in DT_NEEDED and there's a dependency on version GLIBC_2.* (meaning the object depends on glibc). Without this the glibc dynamic loader will refuse to load a shared object using DT_RELR with the following error:

DT_RELR without GLIBC_ABI_DT_RELR dependency

That seems quite complicated, so I agree with you that waiting is probably better... For context, the same issue has been raised before in the lld project: https://github.com/llvm/llvm-project/issues/53775.

[rui314]

Thank you for sharing the link! I agree with you that what GNU ld does seems quite complicated.

[glandium]

b365d6ba1aade2e99d636892e4a21b8d7988a01c added -z pack-relative-relocs without GLIBC_ABI_DT_RELR, which leads to the error mentioned in https://github.com/rui314/mold/issues/653#issuecomment-1221664798 ...

[maxyvisser]

I was experiencing the error mentioned in https://github.com/rui314/mold/issues/653#issuecomment-1221664798 in the 2.0.0 release build when building firefox 116. This issue appears to have been fixed by https://github.com/rui314/mold/commit/f467ad1add2ab6e381e0e458f026df197e63d487, but a new error pops up:

20:10.56 mold: fatal: /tmp/lto-llvm-5abe96.o: REL-type relocation table is not supported for this target
20:10.58 clang-15: error: linker command failed with exit code 1 (use -v to see invocation)
20:10.59 make[4]: *** [/compile/makepkg/librewolf/src/librewolf-116.0-1/config/rules.mk:532: libmozsqlite3.so] Error 1
20:10.59 make[3]: *** [/compile/makepkg/librewolf/src/librewolf-116.0-1/config/recurse.mk:72: config/external/sqlite/target] Error 2

Could this possibly be related?

PS: here is a link to a report on the same bug on firefox's issue tracker: https://bugzilla.mozilla.org/show_bug.cgi?id=1847697.

[rui314]

No, that's different. What target did you get this error for? If it's for x86-64 or ARM64, it is actually odd if an object file contains a REL-type relocation section instead of RELA-type one.

[maxyvisser]

This is on x86-64, a ryzen 5 5600G to be exact. Do you want me to create a new issue, since it is unrelated?

[ambasta]

Can reproduce this issue on FF-117 as well

171:11.18 mold: fatal: /var/tmp/portage/www-client/firefox-117.0/temp/lto-llvm-715b02.o: REL-type relocation table is not supported for this target

[rui314]

It looks like LLVM generates REL-type relocations instead of RELA-type ones for LLVM. It's very likely that it's a bug in LLVM. Can you report it to LLVM?

[ambasta]

Sure, in the process of rebuilding manually (w/o ebuild) to isolate and confirm the issue, Will file one once I have confirmation

[ambasta]

Is this really a bug w/ LLVM?

Compiling w/ LLVM succeeds

clang version 17.0.2
Target: x86_64-pc-linux-gnu
Thread model: posix
InstalledDir: /usr/lib/llvm/17/bin
Configuration file: /etc/clang/x86_64-pc-linux-gnu-clang.cfg
System configuration file directory: /etc/clang
Selected GCC installation: /usr/lib/gcc/x86_64-pc-linux-gnu/13
Candidate multilib: .;@m64
Selected multilib: .;@m64
 "/usr/bin/x86_64-pc-linux-gnu-ld.bfd" --hash-style=gnu --eh-frame-hdr -m elf_x86_64 -shared -o libmozsqlite3.so /usr/lib/gcc/x86_64-pc-linux-gnu/13/../../../../lib64/crti.o /usr/lib/gcc/x86_64-pc-linux-gnu/13/crtbeginS.o -L/usr/lib/gcc/x86_64-pc-linux-gnu/13 -L/usr/lib/gcc/x86_64-pc-linux-gnu/13/../../../../lib64 -L/lib/../lib64 -L/usr/lib/../lib64 -L/usr/lib/gcc/x86_64-pc-linux-gnu/13/../../../../x86_64-pc-linux-gnu/lib -L/lib -L/usr/lib -plugin /usr/lib/llvm/17/bin/../lib64/LLVMgold.so -plugin-opt=mcpu=znver1 -plugin-opt=O3 -plugin-opt=thinlto -plugin-opt=-function-sections=1 -plugin-opt=-data-sections=1 -z relro -z defs --gc-sections -h libmozsqlite3.so /var/tmp/portage/www-client/firefox-118.0.2/work/firefox_build/config/external/sqlite/libmozsqlite3_so.list -plugin-opt=-import-instr-limit=10 -plugin-opt=-import-hot-multiplier=30 -lpthread -O1 --as-needed --compress-debug-sections=zlib -rpath=/usr/lib64/firefox --enable-new-dtags -z pack-relative-relocs -z noexecstack -z text -z relro -z nocopyreloc -Bsymbolic-functions -rpath-link /var/tmp/portage/www-client/firefox-118.0.2/work/firefox_build/dist/bin -rpath-link /usr/lib --version-script libmozsqlite3.so.symbols -lm -lgcc --as-needed -lgcc_s --no-as-needed -lpthread -lc -lgcc --as-needed -lgcc_s --no-as-needed /usr/lib/gcc/x86_64-pc-linux-gnu/13/crtendS.o /usr/lib/gcc/x86_64-pc-linux-gnu/13/../../../../lib64/crtn.o

But on adding -fuse-ld=mold

clang version 17.0.2
Target: x86_64-pc-linux-gnu
Thread model: posix
InstalledDir: /usr/lib/llvm/17/bin
Configuration file: /etc/clang/x86_64-pc-linux-gnu-clang.cfg
System configuration file directory: /etc/clang
Selected GCC installation: /usr/lib/gcc/x86_64-pc-linux-gnu/13
Candidate multilib: .;@m64
Selected multilib: .;@m64
 "/usr/bin/ld.mold" --hash-style=gnu --eh-frame-hdr -m elf_x86_64 -shared -o libmozsqlite3.so /usr/lib/gcc/x86_64-pc-linux-gnu/13/../../../../lib64/crti.o /usr/lib/gcc/x86_64-pc-linux-gnu/13/crtbeginS.o -L/usr/lib/gcc/x86_64-pc-linux-gnu/13 -L/usr/lib/gcc/x86_64-pc-linux-gnu/13/../../../../lib64 -L/lib/../lib64 -L/usr/lib/../lib64 -L/usr/lib/gcc/x86_64-pc-linux-gnu/13/../../../../x86_64-pc-linux-gnu/lib -L/lib -L/usr/lib -plugin /usr/lib/llvm/17/bin/../lib64/LLVMgold.so -plugin-opt=mcpu=znver1 -plugin-opt=O3 -plugin-opt=thinlto -plugin-opt=-function-sections=1 -plugin-opt=-data-sections=1 -z relro -z defs --gc-sections -h libmozsqlite3.so /var/tmp/portage/www-client/firefox-118.0.2/work/firefox_build/config/external/sqlite/libmozsqlite3_so.list -plugin-opt=-import-instr-limit=10 -plugin-opt=-import-hot-multiplier=30 -lpthread -O1 --as-needed --compress-debug-sections=zlib -rpath=/usr/lib64/firefox --enable-new-dtags -z pack-relative-relocs -z noexecstack -z text -z relro -z nocopyreloc -Bsymbolic-functions -rpath-link /var/tmp/portage/www-client/firefox-118.0.2/work/firefox_build/dist/bin -rpath-link /usr/lib --version-script libmozsqlite3.so.symbols -lm -lgcc --as-needed -lgcc_s --no-as-needed -lpthread -lc -lgcc --as-needed -lgcc_s --no-as-needed /usr/lib/gcc/x86_64-pc-linux-gnu/13/crtendS.o /usr/lib/gcc/x86_64-pc-linux-gnu/13/../../../../lib64/crtn.o
mold: fatal: /tmp/lto-llvm-be7713.o: REL-type relocation table is not supported for this target

[rui314]

Yes, it's a bug in LLVM, please report it to LLVM. x86-64 is defined to use RELA-type relocations by its psABI, so creating a REL-type relocation type is illegal even if other linkers happen to not complain.

[ambasta]

@rui314 Some additional guidance please, looking at https://github.com/llvm/llvm-project/blob/631e2911ea298bc12564df8acd16bba89790426a/lld/test/ELF/relocation-none.test#L38

## Both REL and RELA are supported. .rel.llvm.call-graph-profile uses REL even
## if the prevailing format is RELA.

This is exactly where mold is failing, and seems to be documented behavior

[rui314]

Even if it's documented, it still violates the spec. As quoted from x86-64 psABI p.72, "The AMD64 LP64 ABI architecture uses only Elf64_Rela relocation entries with explicit addends." While they have the option not to conform to the standard, doing so would lead to compatibility issues for obvious reasons.

[ambasta]

Yes, but that would be unlikely. Even the kernel has patches to ignore modpost checks on call-graph-profile

https://lore.kernel.org/lkml/CAK7LNARVi1sfBjv5a5OoQWPEeM-6bFuwPJE+i32NC=wdum-AKw@mail.gmail.com/t/#mcd82d7d511dcb4fe8939e21127fafda08f4f732e

And here's the rationale behind it

https://reviews.llvm.org/D104080

[rui314]

Alright, how about we simply ignore the REL relocation tables in psABIs where RELA is mandated? It seems that by doing so, we'll resolve the issue.

[ambasta]

I genuinely do not understand any of this enough to comment. This issue was my first attempt at even trying to go through a compiler/linker source, so I'm not sure what the correct approach is here.

[rui314]

I believe ignoring a relocation table with an incompatible type should be fine because in this case all relocations in the incompatible table is R_*_NONE type, which is a no-op relocation. mold used to silently ignore incompatible relocation tables, so it's not new.