[llvm-branch-commits] [llvm] AArch64: Use consistent atomicrmw expansion for FP operations (PR #103702)

[Eli Friedman via llvm-branch-commits]

efriedma-quic wrote:

Just thought of this, but... we can't do this in the case where we do a libcall.  Any load or store between the load exclusive and the store exclusive could break the reservation.  (It normally won't, but it can in weird cases where the atomic variable is on the stack.)  So we have to use cmpxchg lowering in those cases (and then expand the cmpxchg to ll/sc).

The cases we can do inline should be fine, though.

https://github.com/llvm/llvm-project/pull/103702