Technology

michaeleisel/zld: A faster version of Apple’s linker

A faster version of Apple’s linker

As used by

Most of the really big third-party apps out there.

Introduction

For large projects, the linking phase (explanation) can significantly increase incremental build times. This project is a fork of the Apple linker, ld. It is a drop-in replacement that can substantially speed things up.

Note: it is only intended for debug builds, to make debugging faster.

Performance

Feel free to file an issue if you find that linking is not at least 40% faster for your case (make sure to run it twice in a row to ensure that caches have been generated). Further benchmark details can be found here.

Is it worth it for my project?

It all depends on your risk tolerance and how much you value the speedup in incremental build time. When linking takes more than one second, I’d cut that time in half as the estimated time with this new linker. If that difference is compelling to you, then it’d be worth trying out. Personally, I’d use it in projects with an existing link time of even 500ms (but I am an impatient tinkerer).

Stability

zld is forked from the most recently open-sourced version of ld. It’s used by thousands of developers across many of the largest apps in the world. Without a few optimizations around hashing, it would produce byte-for-byte the same executables as the open-source one. Although it’s not ideal to mix compiler and linker toolchain versions, the open-source one is fairly recent. zld will continue to be updated with each new version of ld as it is released.

Installation

Pre-built binary

The pre-built binary for the latest release is here.

Building from source

  • Install Xcode 12.2+ and run sudo xcode-select -s <path to Xcode>
  • Install cmake
  • Checkout the latest release of zld from master
  • Run make clean && make
  • See in the output where it built zld (probably build/Build/Products/Release/zld).

Usage

If using Xcode:

Get the path of zld from which zld, then add -fuse-ld=<path to zld> -Wl,-zld_original_ld_path,$(DT_TOOLCHAIN_DIR)/usr/bin/ld to “Other Linker Flags” in the build settings (debug configuration). That -zld_original_ld_path provides the path to the linker Xcode would otherwise use, which is important because there are certain known cases (e.g. arm64_32 and Catalyst) where zld knows that it has issues and will silently use that linker instead. Fixing these cases is a work-in-progress (largely blocked by Apple being slow to release source code).

If using Bazel:

Add these to your .bazelrc or pass them to your command line.

build --linkopt=-fuse-ld=<path to zld>
build --linkopt=-Wl,-zld_original_ld_path,__BAZEL_XCODE_DEVELOPER_DIR__/Toolchains/XcodeDefault.xctoolchain/usr/bin/ld

Note that you will need to disable sandbox for this to work now. Additionally, to make the linking actions cacheable, the path to zld must be deterministic (e.g. /tmp/zld-09ea158, where 09ea158 is zld version).

If using Rust:

You can edit ~/.cargo/config to add a linker flag, e.g.:

[target.x86_64-apple-darwin]

rustflags = ["-C", "link-arg=-fuse-ld=<path to zld>"]

Caching

By default, zld stores some metadata in /tmp/zld-... to speed things up. This is the first step towards making zld a truly incremental linker. Currently, the only things that are stored are object file and library names.

Why is it faster?

Apple’s approach is a very reasonable one, using C++ with STL data structures. However, there are a number of ways in which zld has sped things up, for instance:

  • Using Swiss Tables instead of STL for hash maps and sets
  • Parallelizing in various places (the parsing of libraries, writing the output file, sorting, etc.)
  • Optimizations around the hashing of strings (caching the hashes, using a better hash function, etc.)

Other things to speed up linking

Whether you use this project or not, there are a number of things that can speed linking up (again, this is only for debug builds):

  • The linker flag -Wl,-random_uuid, which disables content hashing based UUID creation and instead uses a random UUID (using -no_uuid can decrease performance when lldb is attaching to the binary)
  • Turning off dead stripping (referred to as “Dead Code Stripping” in Xcode build settings)
  • For executables and xctest bundles, disable dyld exports trie creation with -Wl,-exported_symbols_list,/dev/null (cannot be used by test host apps that provide symbols to xctests)
  • -Wl,-no_deduplicate, which disables the deduplication pass. In Xcode, this flag is added by default for Debug builds.
  • -Wl,-no_compact_unwind, which disables the creation of compact unwind info. Note that Objective-C and C++ functions that have been linked without their compact unwind info can crash if an exception unwinds into them, rather than continuing the unwind. Swift functions, however, are not affected (even if an Objective-C/C++ exception unwinds into them). So, it’s best for pure Swift projects. It can also break crash reporting.
  • -Wl,-x, which disables putting non-global symbols in the symbol table. This won’t impact debugging as long as debug info is still enabled.
  • If you’re not using zld, using -Wl,-force_load for libraries can sometimes speed things up
  • Linking with dynamic libraries instead of static ones

Contributing

The biggest way to contribute to zld is to file issues! If you encountered any problems, feel free to file an issue, and you can expect a prompt response.

Special thanks to @dmaclach’s ld64, which helped with building ld.

Related Articles

Back to top button