Evaluate Profile-Guided Optimization (PGO) on Yoga

Report

• I have searched existing issues and this is not a duplicate

According to the official website, one of the Yoga goals is performance. So I think would be a good idea to try to gain more performance by applying Profile-Guided Optimization (PGO) in Yoga.

Since Yoga is a library (not an executable), it could be a bit harder to do since for profile collection you need an executable. One of the things is to build the Yoga benchmark and Yoga library with PGO and measure the improvements.

If will be some performance improvements, would be helpful:

• Write a note about PGO in Yoga documentation with performance improvement numbers.
• Add an option to the build scripts for building Yoga library with PGO. In this case, for the users will be easier to rebuild Yoga for their own specific use cases

After that, I guess LLVM BOLT could be tested as well but I suggest trying BOLT only after PGO since PGO implementations in compilers are much more mature compared to BOLT (IMHO).

Start by locating the Yoga benchmark, Yoga build scripts, and documentation, then determine how an executable can collect PGO profiles for the library. Measure the benchmark with and without PGO and record the performance results. Done means deciding whether the improvement justifies documentation and a build option; LLVM BOLT is a later consideration.