[llvm] [CodeGen][AArch64] Sink splat operands of FMul instructions (PR #116222)

[Hari Limaye via llvm-commits]

hazzlim wrote:

The first commit here updates one of the tests, as currently the `fmul` is hoisted out of the loop anyway so we don't test sinking the splat. I changed it to a `splat(insertelement(scalar))` so that it doesn't get matched by one of the `MachineCombiner` patterns in:
https://github.com/llvm/llvm-project/blob/5c3befb91cd774161e5d700cf2c351d42d29927c/llvm/lib/Target/AArch64/AArch64InstrInfo.cpp#L6789

Also to note - it looks like we don't have a `MachineCombiner` pattern for `FMLA(DUPLANE) -> FMLA_indexed` (we only have the FMUL pattern).

https://github.com/llvm/llvm-project/pull/116222