[llvm] [AArch64] Keep floating-point conversion in SIMD (PR #147707)

[David Green via llvm-commits]

davemgreen wrote:

PreprocessISelDAG is a new one that we don't use very much, I just happened to run into it the other day on another target. You might find that because the patterns are greedy, it ends up producing the same thing anyway. (i.e. it picks bitcast(fptosi)+round over fmov+fptosi_round. It hopefully allows adding patterns for bitcast(fptosi(round all at once though.

https://github.com/llvm/llvm-project/pull/147707