Also switch to not using kernels to test, since it shouldn't matter
for these addressing modes. This should reduce some of the diffs
between subtargets, but it's still not enough to share the same
checks.
The test could use some additional modernization. Probably should also
round this out with the new maximium offsets.