You can't just adopt that hardware and get memory safety. The hardware is providing support for a capabilities model, which you then have to adopt at a software level both within your OS, compiler, and application code. This also would break C and C++ ABIs, so it's very unlikely to get adopted for a number of cases.
Further, CHERI is not enough to achieve temporal memory safety, it only provides the primitive that one could use to implement an efficient, correct tracing GC.
Perhaps most importantly, one can not use CHERI today. It is just not an option. Memory safe languages exist, and have existed for quite some time.
Pointer tagging and other approaches like CHERI are very promising. Keeping in mind of course that they double the size of pointers and have a global runtime performance cost. Less ambitious but otherwise similar mitigations are already being adopted.
Indeed, you need the whole stack. However, the (non-temporal) memory safety alone is still quite easy to get - the compiler will take care of it, as a programmer you just need to make sure your code doesn't get in the way by eg manually stashing pointers into non-pointer types. It's been demonstrated on large, real-world code bases, such as FreeBSD and PostgreSQL.
(Disclaimer: been there, done that, part of the CHERI team)
Yep, much of the wins look to be a matter of just recompiling, which is great. I'm a big fan of these pointer tagging techniques in general. I just want to be clear that pointing at CHERI and going "See? We already have memory safe C and C++" is glossing over some important details.
Best of luck with the research, I'm quite bullish on the work.
Further, CHERI is not enough to achieve temporal memory safety, it only provides the primitive that one could use to implement an efficient, correct tracing GC.
Perhaps most importantly, one can not use CHERI today. It is just not an option. Memory safe languages exist, and have existed for quite some time.
Pointer tagging and other approaches like CHERI are very promising. Keeping in mind of course that they double the size of pointers and have a global runtime performance cost. Less ambitious but otherwise similar mitigations are already being adopted.