Sounds like Julia should consider switching to Zig to get considerable performance gains. I remember authors feeling uneasy with each llvm release worrying about performance degradations.
The compiler is fairly retargetable, this is an active area of work. So it’s maybe possible in the future to envision zig as an alternative compiler for fragments of the language.
Such as a more fine grained compile cache, better tooling to prevent i validations, removal of the world splitting optimisation, more use of multithreading in the compiler, automatic precompilation of concrete signatures, and generation of lazier code which hot-swaps in code when it is compiled.