I actually have an M2 mac mini in the mail from Apple for exactly this purpose.
My time horizon is very long. It takes me a long time to do things these days.
It has never been true that I don't want to support it. Merely that it is difficult to verify and test if I can't do it myself. There is also the problem that the port from x86 to arm is not straight-forward, due to both my own ignorance and what I believe are important missing vector operations such as movemask.
The current PR for ARM SIMD[1] uses a different instruction mix to achieve the same goals as movemask. I tested the PR and it has a significant speedup over the non-vectorized version.
My time horizon is very long. It takes me a long time to do things these days.
It has never been true that I don't want to support it. Merely that it is difficult to verify and test if I can't do it myself. There is also the problem that the port from x86 to arm is not straight-forward, due to both my own ignorance and what I believe are important missing vector operations such as movemask.
This is discussed a bit more here (including the bit about movemask): https://github.com/BurntSushi/memchr/issues/76