[llvm] [AArch64] Limit immediate offsets when folding instructions into addressing modes (PR #67345)

[Momchil Velikov via llvm-commits]

momchil-velikov wrote:

Yes, it's pessimistic from the point of view of "sink-and-fold" and optimistic from the point of view of load/store opt :D 
This function (`AArch64InstrInfo::canFoldIntoAddrMode`)  is a bit limited to what it can do, since it's supposed to look at single instruction pair.
In principle, we have all the potential folding candidates collected  at https://github.com/llvm/llvm-project/blob/1db42fa6f07838365e34d7d05df3c8b7287be176/llvm/lib/CodeGen/MachineSink.cpp#L507
and we could add another target hook to check for consecutive loads/stores, however, I'm quite a bit skeptical of how much
we would gain for the effort.


https://github.com/llvm/llvm-project/pull/67345