The author hypothesized that this is likely because Cranelift performs fewer optimizations compared to LLVM. But also because of Cranelift’s focus on fuzzing and formal methods to preserve quality.
My understanding is, and I may be wrong here, that the goal is to over time increase the number of optimizations Cranelift can perform. As Cranelift’s scope broadens, it’s going to be really interesting to see its approach to engineering scale up.
@yosh well, they're trying to make more optimizations possible. but they're doing some things a little differently. phi nodes in llvm are incredibly powerful but also oh god. so they came up with something else that they think they can get most of the benefit from without the higher costs.