[llvm] [AArch64] merge index address with large offset into base address (PR #72187)

[David Green via llvm-commits]

davemgreen wrote:

> At the meanwhile, there is no loop information at DAGIsel, so it seems difficult to check whether the current variable is loop invariant.

I've wondered if this was worth adding, at least to check if the current block was in a loop. Otherwise we have had to move transforms to after ISel in order to check they are profitable. In this case I guess we need to know that the base is loop invariant?

https://github.com/llvm/llvm-project/pull/72187