mold 1.2.1

mold 1.2.1 is a new release of the high-speed linker. This release contains the following bug fixes.

Bug fixes and compatibility improvements

• Various bugs in --gdb-index have been fixed.
• mold now recognizes --thinlto-cache-dir and --thinlto-cache-policy for the sake of compatibility with LLVM lld. (7ebd071)
• mold can now handle TLS common symbols. It looks like GCC sometimes creates such symbol for a thread-local variable. (cf850f8)
• In some edge cases, mold created a non-versioned symbol and a versioned one for the same symbol, even though if one symbol is versioned, all symbols of the same name must be versioned. This bug has been fixed. (8298c0a)
• mold used to write a PLT address of a symbol instead of its address to .symtab. This bug has been fixed. (e088db7)
• mold can now handle an input file with more than 2¹⁹ symbols. (f1f2d40)
• /usr/local/libexec/mold/ld is now installed as a relative symlink instead of an absolute symlink. (5803c3c)

mold 1.2.0

mold 1.2.0 is a new release of the high-speed linker. The highlight of this release is the 32-bit ARM support. We also added other features, and as always, we fixed many bugs and compatibility issues in this release.

New features

• The ARM32 target is now supported.
• --gdb-index is implemented. If this option is given, mold creates an .gdb_index section in an output file to speed up GNU debugger. Users have to compile their object files with -ggnu-pubnames to use this flag. mold used to ignore --gdb-index. (a7475dd)
• mold now supports the following flags: --start-address, -Tbss, -Tdata, -Ttext, --oformat=binary, --disable-new-dtags

Deprecated features

• An experimental, mold-specific --preload flag has been marked as deprecated. It's still usable, but a warning message will be displayed if that flag is given.

Bug fixes and compatibility improvements

• -dy and -dn are now accepted as aliases for -Bdynamic and -Bstatic, respectively. (82e8072)
• -static-pie now works with older versions of glibc thanks to a few bug fixes. (3d68824, 0884f27)
• Issues found by UndefinedBehaviorSanizer, AddressSanitizer and ThreadSanitizer are fixed. (bf26753, f4753b3, e1e4e9f)
• mold used to place sections with very large section alignment requirements to wrong places in an output file. That caused a mysterious crash of a produced binary (#405). That bug was most noticeable when Nvidia-provided object files are given because they tend to contain such sections. This bug has been fixed. (100922b)
• .ctors and .dtors sections are now recognized by mold, and their contents are sorted with a special rule. This shouldn't affect most build environments because these sections have been superseded by .init_array and .fini_array sections a long time ago. But it looks like some old i386 compilers are still using .ctors and .dtors. (392781a)
• For a non-position-independent executable, we have to make address-taken PLT entries as "canonical". Marking all PLT entries canonical should be harmless in theory, so we did so. However, some programs, notably Qt library, assume that non-address-taken PLTs can never be canonical (#352). For the sake of compatibility with such programs, we now make PLTs canonical only when their addresses are taken. (e0bc74a)
• mold now defines _TLS_MODULE_BASE_ symbol. A reference to this symbol can occur if -mtls-dialect=gnu2 is given to a compiler. The flag tells the compiler to use TLSDESC mechanism instead of the regular TLS access mechanism to access thread-local variables. (5feab82)
• libbacktrace sometimes fail to read compressed debug sections in mold-generated files due to a bug. We not only fix that libbacktrace's bug (ianlancetaylor/libbacktrace#87) but also implemented a workaround to mold (ba63479) so that mold works with older versions of libbacktrace.
• [ARM64] mold now recognizes R_AARCH64_LD_PREL_LO19 relocation. (146ddd7)
• [RISCV64] The correct semantics of R_RISCV_ALIGN is implemented. (0daf623)

mold 1.1.1

mold 1.1.1 is a maintenance release of the high-performance linker. It contains the following new features, bug fixes and performance improvements including memory usage reduction.

New features

• The --dependency-file option has been added. The option is analogous to the compiler's -MM option; it generates a text file containing dependency information in the Makefile format, so that you can include a generated file into a Makefile to automate the file dependency management. (a054bcd)
• mold has gained the --reverse-sections option. If the option is given, mold reverses the list of input sections before assigning them the addresses in an output file. This option is useful to find a bug in global initializers (e.g. constructors of global variables.) In C++, the execution order of global initializers is guaranteed only within a single compilation unit (they are executed from top to bottom.) If two global initializers are in different object files, they can be executed in any order. Reversing the execution order of the global initializers in different input files should help you identify a bug in your program. If your program does not work with -Wl,--reverse-sections, your program depends on the undefined behavior.
• --shuffle-sections now takes an optional seed for the random number generator in the form of --shuffle-sections=<number>. (8f21cc3)
• mold now supports the following LTO-related options for compatibility with LLVM lld: --disable-verify, --lto-O, --lto-cs-profile-file, --lto-cs-profile-generate, --lto-debug-pass-manager, --lto-emit-asm, --lto-obj-path, --lto-partitions, --lto-pseudo-probe-for-profiling, --lto-sample-profile, --no-legacy-pass-manager, --no-lto-legacy-pass-manager, --opt-remarks-filename, --opt-remarks-format, --opt-remarks-hotness-threshold, --opt-remarks-passes, --opt-remarks-with_hotness, --save-temps, --thinlto-emit-imports-files, --thinlto-index-only, --thinlto-index-only, --thinlto-jobs, --thinlto-jobs, --thinlto-object-suffix-replace, --thinlto-prefix-replace (e413433)
• -noinhibit-exec and --warn-shared-textrel have been supported.

Performance improvements

• We optimized mold's memory usage by reducing the sizes of frequently-allocated objects. Compared to mold 1.1, we observed ~6% reduction of maximum resident set size (RSS) when linking Chromium. Our maximum RSS is smaller than LLVM lld and GNU gold as far as we tested. We measured maximum RSSes with time -v. (f2d27d8, 7068c0c, 83e05da, 4dae896)
• If Intel CET-based security-enhanced PLT is enabled (i.e. -z ibtplt is given), mold used to create a PLT section in which each entry is 32 bytes long. We optimized the machine code sequence of the CET-enabled PLT section, so each PLT entry now occupies only 16 bytes, reducing the size of .plt by almost half. (480efde)

Bug fixes and compatibility improvements

• -static-pie now works with recent versions of glibc. Previously, statically-linked position-independent executable would crash on startup when linked with mold. (3999aa8)
• Previously, mold sometimes created corrupted output file on x86-64 if an input file containing thread-local variables were compiled with -mcmodel=large (#360). This issue has been fixed. (4aa4bfa)
• Previously, mold created corrupted debug info section on i386 if an input debug section is also compressed using the compiler -gz option. (#361) This issue has been fixed. (3068364)
• mold used to create multiple .init_array sections if input files contain both writable and non-writable .int_array sections. That caused an issue that some initializer functions would not be executed on process startup. (#363). This issue has been fixed. (4198627)
• When building a large program with GCC LTO, mold occasionally failed with "too many open files" error. This issue has been resolved. (e67f460)
• Previously, mold created a corrupted dynamic relocation table if .got.plt is missing. This issue has been fixed by always creating _GLOBAL_OFFSET_TABLE_ symbol in .got on any target. mold used to try to create the symbol in .got.plt on x86-64 or i386. (eb79859)

mold 1.1

mold 1.1 is a new release of the high-performance linker. It contains a few new major features and various bug fixes.

Starting from this release, we distribute not only source but pre-compiled binaries for Linux. You can download a tar file containing mold executable at the bottom of the page. You can copy the contents of the tar file to /usr/local or just use without installation by directly specifying its full path.

New features

• Native LTO (link-time optimization) support has been added. mold used to invoke ld.bfd or ld.lld if it encountered a GCC IR (intermediate representation) file or an LLVM IR file to delegate the task to the LTO-capable linkers, respectively. Now, mold handles IR files directly. This feature is implemented using the linker plugin API which is also used by GNU ld and GNU gold. Note that the LTO support has been added for completeness and not for speed. mold is only marginally faster than the other linkers for LTO builds because not linking but code optimization dominates. (46995bc)
• RISC-V (RV64) is now supported as both host and target platforms. mold can link real-world large programs such as mold itself or LLVM Clang for RISC-V. (e76f7c0)
• The -emit-relocs option is supported. If the option is given, mold copies relocation sections from input files to an output file. This feature is used by some post-link binary optimization or analysis tools such as Facebook's Bolt. (26fe71d)
• mold gained the --shuffle-sections option. If the option is given, the linker randomly shuffle the order of input sections before fixing their addresses in the virtual address space. This feature is useful in some situations. First, it can be used as a strong form of ASLR (address space layout randomization). Second, you can enable it when you are benchmarking some other program to get more reliable benchmark numbers, because even the same machine code can vary in performance if they are laid out differently in the virtual address space. You want to make sure that you got good/bad benchmark numbers not by coincidence by shuffling input sections. (7e91897)
• The --print-dependencies and --print-dependencies=full options were added. They print out dependencies between input files in the CSV format. That is, they print out the information as to which file depends on which file to use which symbol. We added this feature with a few use cases in mind. First, you can use this to analyze why some object file was pulled out from an archive and got linked to an output file. Second, when you want to eliminate all dependencies to some library, you can find all of them very easy with this feature. Note that this is an experimental feature and may change or removed in feature releases of mold. (a1287c2)
• The following options are added: --warn-once (f24b997), --warn-textrel (6ffcae4)
• Runtime dependency to libxxhash has been eliminated. (e5f4b96)

Bug fixes and compatibility improvements

• A PT_GNU_RELRO segment is now aligned up to the next page boundary. Previously, mold didn't align it up, and the runtime loader align it down, so the last partial page would not be protected by the RELRO mechanism. Now, the entire RELRO segment is guaranteed to be read-only at runtime. (0a0f9b3)
• The .got.plt section is now protected by RELRO if -z now is given. This is possible because writes to .got.plt happen only during process startup if all symbols are resolved on process startup. (73159e2)
• Previously, mold reported an error if object files created with old GCC (with -fgnu-unique) are mixed with ones created with newer GCC or Clang (with -fno-gnu-unique) (#324). Now, mold accepts such input files. (e65c5d2)
• mold can now be built with musl libc. (42b7eb8)
• mold-generated .symtab section now contains section symbols and symbols derived from input shared object files. (e4c03c2, 1550b5a)
• mold-generated executables can now run under valgrind. Previously, valgrind aborted on startup due to an assertion failure because it didn't expect for an executable to have both .bss and .dynbss sections. mold generated .dynbss to contain copy-relocated symbols. The section has been renamed .copyrel to workaround the valgrind's issue. (0f8bf23)

mold 1.0.3

mold 1.0.3 is a maintenance release of the high-speed linker. It contains only the following bug fix:

• build-static.sh didn't create a statically-linked mold executable (#315). The problem is now fixed. (601b9e6)

If you are not using build-static.sh to build mold executable, you don't need to upgrade from 1.0.2 to 1.0.3.

mold 1.0.2

mold 1.0.2 is a maintenance release of the high-speed linker. It contains a few new features and various bug fixes as well as performance improvements.

New features

• mold now automatically falls back to ld.bfd or ld.lld if GCC-based LTO (link time optimization) or LLVM-based LTO are requested, respectively. This is a temporary hack until mold gains native LTO support. (a5029d1)
• The following flags have been added: -z ibt (9ca6a9d), -z cet-report (31a43a7), -z shstk (e29bd8f), -z ibtplt (fbfa01d)
• [ARM64] Range extension thunks are now supported. Previously, mold reported "relocation overflow" errors when the output file's text segment is larger than some threshold (~60 MiB). Now, it can link large programs just fine. (9287682)
• [NetBSD] mold is now usable on NetBSD. (948248b)
• [x86-64] mold now emits compact 8-byte PLT entries instead of the regular 16-byte PLT entries if -z now is given. (0370e7f)
• RELR-type packed dynamic relocations are now supported. You can enable it by passing -z pack-dyn-relocs=relr. The good news is that it can typically reduce PIE (position-independent executable) size by a few percent. This is not a negligible saving because PIE is now default on many systems for security reasons. The bad news is that it needs a runtime support. To our knowledge, it's supported only on ChromeOS, Android, Fuchsia and SerenityOS at this moment. We need to wait for a while for other systems to catch up. (bd6afa1)

Performance improvements

• Version script processor was rewritten with the Aho-Corasick string matching algorithm. If your program uses a version script that contains lots of glob patterns with the * metacharacter, you'll likely to see a significant speedup. (d0c1c4d)
• Relocation processing for non-memory-allocated sections has been optimized. You'll likely to see a speedup if your binary contains large size of debug info. (d8dc8a6)

Bug fixes and compatibility improvements

• mold can now link ICC-generated object files with GCC-generated ones even if the -static flag is given. (#271, be6ae07)
• mold can now handle archive files (.a files) larger than 4 GiB. (bba506d)
• mold no longer have "GNU gold" in its --version string. We had this identification string for some ./configure scripts that didn't work without it, but it causes other compatibility issue such as #284. Now, mold --version prints out something like mold 1.0.2 (compatible with GNU ld). We still need "GNU ld" for many ./configure scripts. (cea6a56)
• Symbol resolution algorithm has been completely rewritten. The previous implementation was non-deterministic in some edge cases, meaning that outcomes from multiple runs of the linker with the same command line parameters could be different due to thread scheduling randomness or some other internal randomness. Now it is guaranteed to be deterministic. (ce5749c)
• mold now try to pull out an object file from an archive if it's needed to resolve an undefined symbol with a common symbol. mold used to ignore common symbols in archives, so it could fail with an unresolved symbol error even if the undefined symbol could be resolved using a file in an archive. (27d8361)
• mold no longer converts .ctors/.dtors sections into .init_array/.fini_array sections. mold used to convert them but in a wrong way. Since .ctors/.dtors have been superseded by .init_array/.fini_array long ago, it should be fine to stop doing this now. (4348417)
• [i386] mold now ignores some legacy symbols in an i386 CRT files to avoid duplicate symbol errors. (#270, 0c19046)

mold 1.0.1

mold 1.0 is a maintenance release of the high-speed linker. It contains a few new features and various bug fixes.

New features

• make install now creates /usr/local/libexec/mold/ld as a symlink to the mold executable. We do this for GCC. By passing -B/usr/local/libexec/mold, you can tell GCC to use ld inside that directory instead of /usr/bin/ld. (e8dcecf)
• xxHash library is now included in the mold's source tree as a subtree for ease of building. If you want to link against a libxxhash in a system library directory, pass SYSTEM_XXHASH=1 to make. (665bffa)
• The extern "C++" directive is now supported in the dynamic list. (7aa5c39)
• --color-diagnostics is supported. mold used to ignore that flag. (6e290aa)
• Not only * but also ? are now treated as special characters in the version script wildcard pattern. (31b0248)
• The --threads=N option has been added as an alias for --thread-count=N. (f9ff048)
• The following option has been added: --defsym (f6e8006), -z nodefaultlib (8c86c28), -z separate-code, -z noseparate-code and -z separate-lodable-segments (5601cf4), -z max-page-size (f3766cd)

Bug fixes and compatibility improvements

• mold now issue a warning instead of an error for an unknown -z option. (8bc5736)
• mold previously created a PT_NOTE segment for non-SHF_ALLOC note segments. This is a wrong behavior because we should create segments only for memory-allocated sections. This problem has been fixed. (76407a6)
• Previously, a version script can affect symbol visibility of undefined symbols when they are promoted to dynamic symbols. This is a semantically incorrect behavior and caused a libQt build failure (#151). The issue has been fixed. (3663389)
• Previously, mold silently turned unresolved undefined symbols into absolute symbols with value 0 if -shared, -z defs and -warn-undefined-symbols are specified. Even though this behavior makes sense, it's not compatible with GNU ld which promotes such symbols into dynamic symbols. This incompatibility causes a link failure for Firefox. Since 1.0.1, mold behaves the same as GNU ld. (04ccd4d)
• Previously, mold applied wrong values for relocations against Initial-Exec thread-local variables. That caused a link failure for Mesa 3D graphics library (#197). The issue has been resolved. (d116113)
• GCC 7 has a bug that it emits incorrect relocations against thread-local variables under a certain condition. That bug was unnoticed because existing linkers silently produces an output that works fine in most cases but is technically corrupted. mold used to check for that error condition and report an error. Now, mold does not report it as an error for the sake of bug-compatibility with GCC 7. I don't think relaxing the error check will cause any new issue to existing GCC 7 users, because if it does, they would have been experiencing the issue with existing linkers already. (d9606d6)
• If an output file has more than one sections for thread-local BSS, they were laid out in such that they are overlapping with each other. This bug caused a runtime error for programs compiled with DMD, a compiler for the D language (#126). This layout issue has been resolved. (b151de6)
• Previously, mold failed to look up correct files under --sysroot in some conditions. That caused a link failure for ClickHouse (#150). This bug has been fixed. (135f17c)

mold 1.0

mold 1.0 is the first stable and production-ready release of the high-speed linker. On Linux-based systems, it should "just work" as a faster drop-in replacement for the default GNU linker for most user-land programs. If you are building a large executable which takes a long time to link, mold is worth a try to see if it can shorten your build time. mold is easy to build and easy to use. For more details, see README.

mold is created by a person who knows very well as to how the Unix linker should behave, as I'm also the original creator of the current version of the LLVM lld linker.

There's no fancy new features in 1.0. Actually, 1.0 is very similar to 0.9.6. That being said, we'd like to make it clear by incrementing a major version number that mold for Linux is now stable.

Future plans

We are currently working on mold for macOS, and once it's complete, we'll release it as mold 2.0. After that, we'll work on mold for Windows and release it as 3.0.

mold 1.0's source tree has code for mold for macOS, but that's pre-alpha. Do not use it unless you know what you are doing.

Changes since mold 0.9.6

• -start-lib and -end-lib options are added for compatibility with GNU gold and LLVM lld.
• More ARM64 relocations are supported.
• Compatibility with glibc 2.2 or prior has improved. (#120)
• Compatibility with valgrind has improved. (#118)
• -Bno-symbolic option has been supported.
• -require-defined option has been supported.

mold 0.9.6

mold 0.9.6 is a maintenance release of the mold linker. This release contains only a single change to fix the following issue:

• mold used to create dynamic relocations for imported symbols when creating a position-dependent executable. That worked fine in an environment in which position-independent code (PIC) is enabled by default such as recent versions of most Linux distros. However, it failed with the "recompile with -fPIC" error if PIC was disabled and a dynamic relocation was created in a read-only section. mold 0.9.6 fixed the issue by creating copy relocations and PLTs for such symbols. (#116)

mold 0.9.5

mold 0.9.5 is a maintenance release of the mold linker.

Highlights of mold 0.9.5

• In 0.9.4, we changed the mold's behavior on remaining weak undefined symbols, so that they would be resolved to address zero if we were creating a shared object file with the -z defs option. Now, such symbols will be promoted to dynamic symbols so that they'll get another chance to be resolved at run-time. This change fixes a regression of Firefox build failure (#114), which depends on this particular linker behavior to export symbols from libxul.so.
• mold can now be built on macOS. Note that mold is still able to produce only ELF (Unix) files — so you can use it for cross compilation on macOS for Linux, but you can't use mold for macOS native development.
• Relocation overflow are now reported as errors on AArch64 and i386. Previously, such relocations were silently producing incorrect output.