This set of scripts creates a 60% faster LLVM toolchain that can be customly trained to any project.
The full_workflow.bash will autodetect, if your machine supports LBR or not and choose the correct script which suits to your hardware.
Be sure to have jemalloc installed, it is used to improve llvm-bolt's memory handling.
git clone https://github.com/ptr1337/llvm-bolt-scripts.git cd llvm-bolt-scripts ./full_workflow.bash
This sequence will give you (hopefully) a faster LLVM toolchain. Technologies used:
- LLVM Link Time Optimization (LTO)
- Binary Instrumentation and Profile-Guided-Optimization (PGO)
- perf-measurement and branch-sampling/profiling and final binary reordering (BOLT)
The goal of the techniques is to utilize the CPU black magic better and layout the code in a way, that allows faster execution.
Measure performance gains and evaluate if its worth the hazzle :)
You can experiment with technologies, maybe ThinLTO is better then FullLTO,
For the last bit of performance, you can run several different workloads and then merge the resulted profiles with 'merge-fdata *.fdata > combined.fdata' and then optimize the libary with llvm-bolt again.
and nothing else! The same goes for BOLT.
If you want to bolt gcc, you need to disable when building gcc the language lto, you can still use the gcc lto function but gcc itself wont build with lto. Enabling lto will crash llvm-bolt.
Also you need to add following to your compileflags:
CXXFLAGS+="-fno-reorder-blocks-and-partition"
LDFLAGS+="--emit-relocs"
The compileflas should be used for any binary you compile and want to optimize the binary with bolt.
Ive included a script which makes it possible to bolt any binary or .so file which got compiled with --emit-relocs. ust change the binary name and the path to your suits and the STAGE number.
After you did run stage 1 you need to run a workload with the instrumented binary/.so file.
When the workload is running, you will see that in the FDATA path many profiles are created. These will be in the STAGE2 process merged and then on your binary/.so file used and bolt will optimize it.
We will now take for example llvm:
- Compile it with relocations (LDFLAGS+="--emit-relocs") enabled
- Install your package
- Change the to BINARY=libLLVM.so and the BINARYPATH=/usr/lib to your suits to the target you want to optimize and set STAGE=1
- Run the script
- After you did run it, it will backup your file and will move the instrumented target to the original path
- Run a workload with the target, so compile something with clang
- You will get several files into the FDATA path, when you run the workload !!! ATTENTION !!! the size of the data can get quite big, so take a watch at the folder
- After youre done with the workload change at the script to STAGE=1
- Run the script again and the created data from the instrumentiation will be merged and then used for llvm-bolt to optimize the target
- After that it will automatically move it tor your systembinary/libary, a backup and the bolted binary can be found at the binarypath.
- Thats it, now repeat the worklow for other targets you want top optimize.
- Tip: if you for example instrumented libLLVM the profile is also useable for other llvm based files which where active in the recording process