[llvm] [ARM] Fix llvm.returnaddress for Thumb1 with R11 frame-pointer (PR #117735)

[Eli Friedman via llvm-commits]

efriedma-quic wrote:

I'm concerned that this is fragile: isel is guessing what framelowering is going to do, and if it guesses wrong, you get a miscompile.  And there's no obvious connection between the bits of code in question.

Either framelowering should preserve LR by using a different code sequence if isReturnAddressTaken() is set, or the decision on how to lower returnaddress should be deferred to framelowering (using a pseudo-instruction).

https://github.com/llvm/llvm-project/pull/117735