It delays the computation until it is needed. It also happens to keep
the old code working without having to add more macros :-)
Doesn't work. It's referenced in creating the register cache, which is
still before we have an architecture. We need to know this information
earlier. Any other ideas, or is my original patch OK?
By the way, I think the existing default: case in that switch is pretty
pointless anyway. Note that it selects an arbitrary number of argument
registers. It doesn't work when all the tools are configured mips-elf
(default to O32), which is why I suggested some time ago that we just
default to O32 in that configuration at least... that would make the
test numbers suck a bit less, and reduce the black magic factor a bit.
Default for plain mips-elf? Yes sounds like a good idea. As the other
variants get deleted they can fall back to that.