I remember seeing the UBSan submission at one point but it appears that
it never landed. It would be really nice to have it because among other
things it will encourage fixing all kinds of things that conflict with
features like SIZE_OVERFLOW and soon RAP. Maybe the kernel could even
begin migrating away from relying on signed overflow (by marking the
cases of intentional overflow with wrappers around the overflow
intrinsics like addw(a, b)).

There are also some niches where a 20% performance hit for something
like this makes sense (maybe only ~10% for just the parts that are
relevant to security, especially if it's just set to trap instead of
calling functions producing debug output).