I think you don't understand the goals Hal has, and lays out pretty explicitly in the paper (https://arxiv.org/pdf/1904.08555.pdf https://arxiv.org/pdf/1904.085

I think you don't understand the goals Hal has, and lays out pretty explicitly in the paper (https://arxiv.org/pdf/1904.08555.pdf https://arxiv.org/pdf/1904.08555.pdf): 1. Be able to replace very complex template dispatch mechanisms with simple JIT ones. Your suggestion is not possible, because you can't compile it once (maybe you can compile it a ton of times though!) That is, unless someone has implemented a very very ridiculous set of dynamic detection and dispatch mechanisms, and your compiler supports changing microarchitecture/etc optimization per function or you split it out/ifdef it and compile the same file 5000 ways. Even then, you will likely miss something the JIT will not, and you will be unable to support new architectures/microarchitectures without recompiling. 2. Be able to JIT things without changes to application source code. 3, Be able to easily integrate things like CUDA into the JIT model. (FWIW: Hal is a very long time and well respected LLVM contributor, so he generally wouldn't suggest something trivially stupid :P)