[they make these users (and LLVM) aware of their MUL53 extension (instruction that's not present in base ARM) that allows for substantially faster bignums.]
I'm sure Apple has included that in their LLVM fork.
Unlikely. How do you access such an instruction without a natural 53bit integer type in the language?
The technical decision is obvious (you have a 53b multiplier there anyway; Intel made the same decision for AVX512). But in both cases, the instruction is only useful as an assembly language primitive for a bignum library like GMP.
Presumably it's used by Accelerate's vBigNum library, but Apple seems to be doing a lousy job of evangelizing that to the natural consumers (like Mathematica or Octave). It would be interesting to see if you can wrap GMP around Accelerate vBigNum and pick up Apple performance that way. But maybe the underlying data structures are too different to allow for such a very lightweight translation?