[llvm] [AArch64] Reduce fuse-literals limit on Apple subtargets (PR #106741)

[Eli Friedman via llvm-commits]

efriedma-quic wrote:

The primary meaning of FeatureFuseLiterals is that it's profitable to schedule mov+"movk lsl 16", or "movk lsl 32"+"movk lsl 48" adjacent to each other.  This is a feature of Cortex-A57 and derived cores, but I've never heard of any other core implementing it.  Is that actually relevant to Apple cores?  If it isn't, we should drop the flag from the Apple cores, then go from there.

The only reason FeatureFuseLiterals is relevant here is that it's getting used as a proxy for "literals are cheap"... since on an a57, it only takes 2 cycles to load a 64-bit constant using fused moves.

https://github.com/llvm/llvm-project/pull/106741