A new proposal volleyed today among LLVM developers is for Multi-Thread Parallel Compilation "MTPC" for the ThinLTO link-time optimization code. This is great news for today's high core count CPUs when looking to compile very large LLVM modules.

LLVM ThinLTO currently provides module-level parallelism but for very large modules this can be a compilation bottleneck with today's processors given that code generation and optimizations within individual LLVM modules is largely serialized.

What's being proposed now is Multi-Thread Parallel Compilation for LLVM's ThinLTO to offer intra-module parallelism. MTPC can split modules at the function level for parallel code generation to avoid lengthy compilation times on large modules and better exploit the multi-core potential of today's processors.

Experimental results with this Multi-Thread Parallel Compilation are showing very nice results with up to a 32% total time reduction on a full application going through a full build.

Those interested can see [1]this LLVM Discourse thread with the initial proposal. There already are a few pull requests open for laying out the foundation of Multi-Thread Parallel Compilation.



[1] https://discourse.llvm.org/t/rfc-thinlto-multi-thread-parallel-compilation-for-large-modules/90877