ThinLTO

Introduction

ThinLTO compilation is a new type of LTO that is both scalable and incremental. LTO (Link Time Optimization) achieves better runtime performance through whole-program analysis and cross-module optimization. However, monolithic LTO implements this by merging all input into a single module, which is not scalable in time or memory, and also prevents fast incremental compiles.

In ThinLTO mode, as with regular LTO, clang emits LLVM bitcode after the compile phase. The ThinLTO bitcode is augmented with a compact summary of the module. During the link step, only the summaries are read and merged into a combined summary index, which includes an index of function locations for later cross-module function importing. Fast and efficient whole-program analysis is then performed on the combined summary index.

However, all transformations, including function importing, occur later when the modules are optimized in fully parallel backends. By default, linkers that support ThinLTO are set up to launch the ThinLTO backends in threads. So the usage model is not affected as the distinction between the fast serial thin link step and the backends is transparent to the user.

For more information on the ThinLTO design and current performance, see the LLVM blog post ThinLTO: Scalable and Incremental LTO. While tuning is still in progress, results in the blog post show that ThinLTO already performs well compared to LTO, in many cases matching the performance improvement.

Current Status

Clang/LLVM

The 3.9 release of clang includes ThinLTO support. However, ThinLTO is under active development, and new features, improvements and bugfixes are being added for the next release. For the latest ThinLTO support, build a recent version of clang and LLVM.

Linkers

ThinLTO is currently supported for the following linkers:

Usage

Basic

To utilize ThinLTO, simply add the -flto=thin option to compile and link. E.g.

% clang -flto=thin -O2 file1.c file2.c -c
% clang -flto=thin -O2 file1.o file2.o -o a.out

When using lld-link, the -flto option need only be added to the compile step:

% clang-cl -flto=thin -O2 -c file1.c file2.c
% lld-link /out:a.exe file1.obj file2.obj

As mentioned earlier, by default the linkers will launch the ThinLTO backend threads in parallel, passing the resulting native object files back to the linker for the final native link. As such, the usage model is the same as non-LTO.

With gold, if you see an error during the link of the form:

/usr/bin/ld: error: /path/to/clang/bin/../lib/LLVMgold.so: could not load plugin library: /path/to/clang/bin/../lib/LLVMgold.so: cannot open shared object file: No such file or directory

Then either gold was not configured with plugins enabled, or clang was not built with -DLLVM_BINUTILS_INCDIR set properly. See the instructions for the LLVM gold plugin.

Controlling Backend Parallelism

By default, the ThinLTO link step will launch as many threads in parallel as there are cores. If the number of cores can’t be computed for the architecture, then it will launch std::thread::hardware_concurrency number of threads in parallel. For machines with hyper-threading, this is the total number of virtual cores. For some applications and machine configurations this may be too aggressive, in which case the amount of parallelism can be reduced to N via:

  • gold: -Wl,-plugin-opt,jobs=N

  • ld64: -Wl,-mllvm,-threads=N

  • lld: -Wl,--thinlto-jobs=N

  • lld-link: /opt:lldltojobs=N

Other possible values for N are:

  • 0: Use one thread per physical core (default)

  • 1: Use a single thread only (disable multi-threading)

  • all: Use one thread per logical core (uses all hyper-threads)

Incremental

ThinLTO supports fast incremental builds through the use of a cache, which currently must be enabled through a linker option.

  • gold (as of LLVM 4.0): -Wl,-plugin-opt,cache-dir=/path/to/cache

  • ld64 (support in clang 3.9 and Xcode 8): -Wl,-cache_path_lto,/path/to/cache

  • ELF lld (as of LLVM 5.0): -Wl,--thinlto-cache-dir=/path/to/cache

  • COFF lld-link (as of LLVM 6.0): /lldltocache:/path/to/cache

Cache Pruning

To help keep the size of the cache under control, ThinLTO supports cache pruning. Cache pruning is supported with gold, ld64 and ELF and COFF lld, but currently only gold, ELF and COFF lld allow you to control the policy with a policy string. The cache policy must be specified with a linker option.

  • gold (as of LLVM 6.0): -Wl,-plugin-opt,cache-policy=POLICY

  • ELF lld (as of LLVM 5.0): -Wl,--thinlto-cache-policy,POLICY

  • COFF lld-link (as of LLVM 6.0): /lldltocachepolicy:POLICY

A policy string is a series of key-value pairs separated by : characters. Possible key-value pairs are:

  • cache_size=X%: The maximum size for the cache directory is X percent of the available space on the disk. Set to 100 to indicate no limit, 50 to indicate that the cache size will not be left over half the available disk space. A value over 100 is invalid. A value of 0 disables the percentage size-based pruning. The default is 75%.

  • cache_size_bytes=X, cache_size_bytes=Xk, cache_size_bytes=Xm, cache_size_bytes=Xg: Sets the maximum size for the cache directory to X bytes (or KB, MB, GB respectively). A value over the amount of available space on the disk will be reduced to the amount of available space. A value of 0 disables the byte size-based pruning. The default is no byte size-based pruning.

    Note that ThinLTO will apply both size-based pruning policies simultaneously, and changing one does not affect the other. For example, a policy of cache_size_bytes=1g on its own will cause both the 1GB and default 75% policies to be applied unless the default cache_size is overridden.

  • cache_size_files=X: Set the maximum number of files in the cache directory. Set to 0 to indicate no limit. The default is 1000000 files.

  • prune_after=Xs, prune_after=Xm, prune_after=Xh: Sets the expiration time for cache files to X seconds (or minutes, hours respectively). When a file hasn’t been accessed for prune_after seconds, it is removed from the cache. A value of 0 disables the expiration-based pruning. The default is 1 week.

  • prune_interval=Xs, prune_interval=Xm, prune_interval=Xh: Sets the pruning interval to X seconds (or minutes, hours respectively). This is intended to be used to avoid scanning the directory too often. It does not impact the decision of which files to prune. A value of 0 forces the scan to occur. The default is every 20 minutes.

Clang Bootstrap

To bootstrap clang/LLVM with ThinLTO, follow these steps:

  1. The host compiler must be a version of clang that supports ThinLTO.

  2. The host linker must support ThinLTO (and in the case of gold, must be configured with plugins enabled).

  3. Use the following additional CMake variables when configuring the bootstrap compiler build:

  • -DLLVM_ENABLE_LTO=Thin

  • -DCMAKE_C_COMPILER=/path/to/host/clang

  • -DCMAKE_CXX_COMPILER=/path/to/host/clang++

  • -DCMAKE_RANLIB=/path/to/host/llvm-ranlib

  • -DCMAKE_AR=/path/to/host/llvm-ar

Or, on Windows:

  • -DLLVM_ENABLE_LTO=Thin

  • -DCMAKE_C_COMPILER=/path/to/host/clang-cl.exe

  • -DCMAKE_CXX_COMPILER=/path/to/host/clang-cl.exe

  • -DCMAKE_LINKER=/path/to/host/lld-link.exe

  • -DCMAKE_RANLIB=/path/to/host/llvm-ranlib.exe

  • -DCMAKE_AR=/path/to/host/llvm-ar.exe

  1. To use additional linker arguments for controlling the backend parallelism or enabling incremental builds of the bootstrap compiler, after configuring the build, modify the resulting CMakeCache.txt file in the build directory. Specify any additional linker options after CMAKE_EXE_LINKER_FLAGS:STRING=. Note the configure may fail if linker plugin options are instead specified directly in the previous step.

The BOOTSTRAP_LLVM_ENABLE_LTO=Thin will enable ThinLTO for stage 2 and stage 3 in case the compiler used for stage 1 does not support the ThinLTO option.