# lld 15 ELF changes

llvm-project 15 was just released. I added some lld/ELF notes to https://github.com/llvm/llvm-project/blob/release/15.x/lld/docs/ReleaseNotes.rst. Here I will elaborate on some changes.

• -z pack-relative-relocs is now available to support DT_RELR for glibc 2.36+. (D120701) This exciting size reduction optimization called DT_RELR finally landed in glibc. The linker option is all you need to use to enable this optimization. It took me a lot of energy to lobby for the feature in glibc. The benefit will pay off. See Relative relocations and RELR.
• --package-metadata= has been added to create package metadata notes D131439 Fedora uses this to embed an allocable section in a linked image so that the core file will have the information. See Package Metadata for Core Files.
• --no-fortran-common (pre 12.0.0 behavior) is now the default. --fortran-common matches GNU ld's behavior but is more likely to cause problems when users mix COMMON and STB_GLOBAL symbols. See All about COMMON symbols for detail.
• --load-pass-plugin has been added to load a new pass manager plugin. (D120490)
• --android-memtag-{mode=,stack,heap} have been added to synthesize SHT_NOTE for memory tags on Android. (D119384)
• FORCE_LLD_DIAGNOSTICS_CRASH environment variable is now available to force LLD to crash. (D128195) I am not so fond of adding such an envrionment variable, but I see that it can be useful for testing the crash reporting behavior. For example, Clang Driver may rerun a link action with --reproduce= to get a tarball.
• --wrap semantics have been refined. (rG7288b85cc80f1ce5509aeea860e6b4232cd3ca01) (D118756) (D124056) --wrap is quite complex. Previous releases have repeatedly tuned it. At this point I am mostly confident to say that the ld.lld behavior is desired. In the behaviors that ld.lld and GNU ld differ, I am confident to say that GNU ld's is not good enough:)
• --build-id={md5,sha1} are now implemented with truncated BLAKE3. (D121531) The C implementation of BLAKE3 was imported into llvm-project. It is superior to MD5 and SHA1. We can drop the slow MD5 and SHA1 llvm-project implementations.
• --emit-relocs: .rel[a].eh_frame relocation offsets are now adjusted. (D122459) The use case is rare. I recently learned that gold crashes on the use case https://sourceware.org/bugzilla/show_bug.cgi?id=25968.
• --emit-relocs: fixed missing STT_SECTION when the first input section is synthetic. (D122463)
• (TYPE=<value>) can now be used in linker scripts. (D118840) The syntax can be used to create an output section of the specified type. With data commands, we can create an output section even if there is no corresponding input section.
• Local symbol initialization is now performed in parallel. (D119909) (D120626)

Breaking changes

• Archives are now parsed as --start-lib object files. If a member is neither an ELF relocatable object file nor an LLVM bitcode file, ld.lld will give a warning. (D119074) This change improved linking performance for chrome quite a bit. See Archives and --start-lib.
• The GNU ld incompatible --no-define-common has been removed.
• The obscure -dc/-dp options have been removed. (D119108)
• -d is now ignored.
• If a prevailing COMDAT group defines STB_WEAK symbol, having a STB_GLOBAL symbol in a non-prevailing group is now rejected with a diagnostic. (D120626)
• Support for the legacy .zdebug format has been removed. Run objcopy --decompress-debug-sections in case old object files use .zdebug. (D126793)
• --time-trace-file=<file> has been removed. Use --time-trace=<file> instead. (D128451)

## Speed

For many users, the most outstanding difference may be speed. I have made dozens of changes scattered across the lld/ELF codebase to improve performance, with the --start-lib and local symbol initialization changes being the most significant changes.

Parallel input file parsing is difficult. As I mentioned in Why isn't ld.lld faster?, there are

• initialization of sections (embarrassingly parallel)
• COMDAT group resolution
• initialization of local symbols (embarrassingly parallel)
• initialization of non-local symbols
• symbol resolution

As of the 15.0.0 release, I have added parallelism to parts not involving symbol resolution. (Some diagnostics may now be non-deterministic, see Non-deterministic diagnostics due to parallelism.)

Thanks to Peter Smith and Igor Kudrin for making many good suggestions.

Note: initialization of non-local symbols and symbol resolution may affect semantics and is difficult to do with the current lld architecture. It is challenging to keep every feature working, even a minor one like whether a symbol diagnostic is still emitted.

Beside symbol resolution, I am wondering: what are the next major performance opportunities?

Below are some programs benchmarked on an Intel Skylake machine. I use -DCMAKE_BUILD_TYPE=Release -DCMAKE_EXE_LINKER_FLAGS=-Wl,--push-state,$HOME/Dev/mimalloc/out/release/libmimalloc.a,--pop-state -DLLVM_ENABLE_PROJECTS='clang;lld' -DLLVM_TARGETS_TO_BUILD=X86 -fPIC -pie builds of lld. (Compared with glibc malloc, linking against libmimalloc.a is 1.12x as fast.) The host compiler is a close-to-main clang. I run ninja -v$target, extract the linking command, rerun it with -Wl,--reproduce=/tmp/rep.tar to get a tarball which expands to linker input and a response file. Both input and output is in tmpfs.

Linking a -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=ON build of clang:

(--threads=2 => 1.01x (1.038s => 1.026s), --threads=4 => 1.03x (1.020s => 0.9879s))

Linking a -DCMAKE_BUILD_TYPE=Debug build of clang:

(--threads=1 => 1.04x (7.449s => 7.169s), --threads=2 => 1.05x (5.514s => 5.255s))

Linking a default build of chrome:

(--threads=2 => 1.09x (6.349s => 5.828s), --threads=4 => 1.11x (5.744s => 5.170s))

Linking a default build of scylladb (./tools/toolchain/dbuild ninja build/release/scylla):

(--thread=1 => 1.26x (11.552s => 9.194s), --thread=4 => 1.18x (3.871s => 3.267s))