Evaluate usage and implementation compiler optimization techniques like LTO, PGO, PLO for the Mojo ecosystem

[zamazan4ik]

Hi!

Since Mojo supports Ahead-of-Time (AoT) compilation model, I want to propose evaluating "advanced" optimizations for the Mojo world.

I propose the following action points (ofc the list can be extended):

• Evaluate Link-Time Optimization (LTO) for Mojo programs. Users will be able to compile their Mojo programs with LTO in the AoT mode. I guess the starting point here should be the ThinLTO approach from Clang.
• Evaluate Profile-Guided Optimization (PGO) for Mojo programs. Users will be able to compile their Mojo programs with PGO in the AoT mode and use runtime execution statistics during the optimization. I did a lot of PGO benchmarks - it works in practice in so many cases. I guess the starting point here should be the IR PGO (-fprofile-generate compiler switch) approach from Clang.
• Evaluate Post-Link Optimization (PLO) for Mojo programs. Even after applying LTO and PGO, it's possible to optimize an application further. There are tools like LLVM BOLT and Google Propeller to perform such optimizations. Personally, I recommend starting with LLVM BOLT and test it with Mojo programs.
• Preoptimize distributed Mojo binaries with LTO, PGO, PLO. It can help with optimizing CI resources and improve developer UX.

I understand that the project is still in so early development stages that you with a high probability don't have enough resources to investigate right now. At least you can use this discussion as an idea ;)