How does that even work? Even with full precompilation, the LLVM time should still be there, no? In my usual Julia workloads, LLVM time is a significant fraction of compile time, so full precompilation only takes 30-50% off latency.
I thought you could strip out LLVM now? I don't really know, it's above my pay grade.
You can strip it out with a system image, and you don't even need more than the basic StaticArrays and Loop vectorization stuff in the image to get almost all of it