specialization - i don't know if general purpose compilers do this but ML compilers specialize the hell out of kernels (based on constants, types, tensor dimensions, etc).

EDIT: i'm le dumb - this is the whole point of JIT compilers.