Since AGR clobbers CC it should not be used here.
I think it may be better from a micro-architecture perspective to use %r15 as the base pointer in the instruction here, i.e. swap the two source registers. So we should get something like
la %r1, 0(%r1, %r15)
I applied this all the way back to my original issue and confirmed that the coreos-installer test now passes -- thanks!
(I'll wait for the final version of this patch before I actually submit this to LLVM builds for Rust/Fedora/etc.)