std::vector.resize() is actually about 2-3x slower than using unique_ptr, at least for me; it doesn't compile to bzero because the resize() in std:: consists of repeated appending (agh). On our out of tree target with a huge number of registers, this patch actually saves 1% of compile time. It should help significantly on AMDGPU as well, though likely a factor less due to them not having quite the same level of ridiculous register counts.
In general, we probably shouldn't use the heavyweight-ness of std::vector for constant-size arrays. There's really no reason to.
This can be a simple SUnit * now.