[bluejekyll]

This is highly dependent on not much changing. For example, even widening integers from 32-64bit but would break a lot of software.

So to a degree this might be true. But my guess is there is very little benefit here. It’s not clear to me if this could even optimize for the presence of a new vector unit, for example.

[saagarjha]

> This is highly dependent on not much changing. For example, even widening integers from 32-64bit but would break a lot of software.

As Apple designs their own silicon and calling convention, they were able to pull this off for the S3 → S4 transition (the latter of which is AArch64 running with ILP32).