[llvm] [RISC-V] Fix incorrect epilogue_begin (PR #120623)

[Craig Topper via llvm-commits]

topperc wrote:

> It seems like in general we cannot guarantee that prolog / epilog code is a contiguous region since instructions can be moved by optimisation. What is the solution in that case?
> 
> Not saying we should address this in the PR, just wondering.

I know the MachineScheduler excludes CFI_INSTRUCTION from its scheduling regions which prevents it moving instructions in the prolog/epilog.

https://github.com/llvm/llvm-project/pull/120623