BOLT is a post-link optimizer developed to speed up large applications.
It achieves the improvements by optimizing application's code layout based on
execution profile gathered by sampling profiler, such as Linux
An overview of the ideas implemented in BOLT along with a discussion of its
potential and current results is available in
an arXiv paper
BOLT operates on X86-64 and AArch64 ELF binaries. At the minimum, the binaries
should have an unstripped symbol table, and, to get maximum performance gains,
they should be linked with relocations (
BOLT disassembles functions and reconstructs the control flow graph (CFG) before it runs optimizations. Since this is a nontrivial task, especially when indirect branches are present, we rely on certain heuristics to accomplish it. These heuristics have been tested on a code generated with Clang and GCC compilers. The main requirement for C/C++ code is not to rely on code layout properties, such as function pointer deltas. Assembly code can be processed too. Requirements for it include a clear separation of code and data, with data objects being placed into data sections/segments. If indirect jumps are used for intra-function control transfer (e.g., jump tables), the code patterns should be matching those generated by Clang/GCC.
NOTE: BOLT is currently incompatible with the
compiler option. Since GCC8 enables this option by default, you have to
explicitly disable it by adding
you compiling with GCC8.
PIE and .so support has been added recently. Please report bugs if you encounter any issues.
BOLT heavily uses LLVM libraries, and by design, it is built as one of LLVM tools. The build process is not much different from a regular LLVM build. The following instructions are assuming that you are running under Linux.
Start with cloning LLVM and BOLT repos:
> git clone https://github.com/llvm-mirror/llvm llvm > cd llvm/tools > git checkout -b llvm-bolt f137ed238db11440f03083b1c88b7ffc0f4af65e > git clone https://github.com/facebookincubator/BOLT llvm-bolt > cd .. > patch -p 1 < tools/llvm-bolt/llvm.patch
Proceed to a normal LLVM build using a compiler with C++11 support (for GCC use version 4.9 or later):
> cd .. > mkdir build > cd build > cmake -G Ninja ../llvm -DLLVM_TARGETS_TO_BUILD="X86;AArch64" -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=ON > ninja
will be available under
. Add this directory to your path to
ensure the rest of the commands in this tutorial work.
Note that we use a specific revision of LLVM as we currently rely on a set of patches that are not yet upstreamed.
BOLT runs many internal passes in parallel. If you foresee heavy usage of BOLT, you can improve the processing time by linking against one of memory allocation libraries with good support for concurrency. E.g. to use jemalloc:
> sudo yum install jemalloc-devel > LD_PRELOAD=/usr/lib64/libjemalloc.so llvm-bolt ....
Or if you rather use tcmalloc:
> sudo yum install gperftools-devel > LD_PRELOAD=/usr/lib64/libtcmalloc_minimal.so llvm-bolt ....
For a complete practical guide of using BOLT see Optimizing Clang with BOLT .
In order to allow BOLT to re-arrange functions (in addition to re-arranging
code within functions) in your program, it needs a little help from the linker.
to the final link step of your application. You can verify
the presence of relocations by checking for
section in the binary.
BOLT will also report if it detects relocations while processing the binary.
This step is different for different kinds of executables. If you can invoke your program to run on a representative input from a command line, then check For Applications section below. If your program typically runs as a server/service, then skip to For Services section.
The version of
command used for the following steps has to support
option. We recommend using
version 4.5 or later.
This assumes you can run your program from a command line with a typical input.
In this case, simply prepend the command line invocation with
$ perf record -e cycles:u -j any,u -o perf.data -- <executable> <args> ...
Once you get the service deployed and warmed-up, it is time to collect perf data with LBR (branch information). The exact perf command to use will depend on the service. E.g., to collect the data for all processes running on the server for the next 3 minutes use:
$ perf record -e cycles:u -j any,u -a -o perf.data -- sleep 180
Depending on the application, you may need more samples to be included with
your profile. It's hard to tell upfront what would be a sweet spot for your
application. We recommend the profile to cover 1B instructions as reported
option. If you need to increase the number of samples
in the profile, you can either run the
command for longer and use
to increase sampling frequency.
Note that for profile collection we recommend using cycle events and not
. Empirically we found it to produce better results.
If the collection of a profile with branches is not available, e.g., when you run on a VM or on hardware that does not support it, then you can use only sample events, such as cycles. In this case, the quality of the profile information would not be as good, and performance gains with BOLT are expected to be lower.
NOTE: you can skip this step and feed
directly to BOLT using
For this step, you will need
file collected from the previous step and
a copy of the binary that was running. The binary has to be either
unstripped, or should have a symbol table intact (i.e., running
is in your
, and execute
$ perf2bolt -p perf.data -o perf.fdata <executable>
This command will aggregate branch data from
and store it in a
format that is both more compact and more resilient to binary modifications.
If the profile was collected without LBRs, you will need to add
the command line above.
Once you have
ready, you can use it for optimizations with
BOLT. Assuming your environment is setup to include the right path, execute
$ llvm-bolt <executable> -o <executable>.bolt -data=perf.fdata -reorder-blocks=cache+ -reorder-functions=hfsort -split-functions=2 -split-all-cold -split-eh -dyno-stats
If you do need an updated debug info, then add
to the command above. The processing time will be slightly longer.
For a full list of options see
The input binary for this step does not have to 100% match the binary used for
profile collection in Step 1
. This could happen when you are doing active
development, and the source code constantly changes, yet you want to benefit
from profile-guided optimizations. However, since the binary is not precisely the
same, the profile information could become invalid or stale, and BOLT will
report the number of functions with a stale profile. The higher the
number, the less performance improvement should be expected. Thus, it is
crucial to update
for release branches.
Suppose your application can run in different modes, and you can generate
multiple profiles for each one of them. To generate a single binary that can
benefit all modes (assuming the profiles don't contradict each other) you can
$ merge-fdata *.fdata > combined.fdata
for Step 3
above to generate a universally optimized
BOLT is licensed under University of Illinois/NCSA Open Source License .