RFR: AArch64 fixes to C1 SFX barriers

[Roman Kennke]

>> http://cr.openjdk.java.net/~rkennke/c1-sfx-aarch64/webrev.00/
> 
> Hm, it is awkward to have arch-specific hacks in shared code. This is okay for staging/testing code,
> but we would need to find a cleaner solution for upstream. Maybe fix LIR_Assembler::leal in aarch64?

We could (without touching leal):

- Do ensure_in_register() in aarch-specific
shenandoahBarrierSetC1_XXX.cpp and involve a little bit of code-duplication

or

- Extract generating of register into, say, new_register_for_lea() and
do that in aarch64-specific shenandoahBarrierSetC1_XXX.cpp and keep the
majority of ensure_in_register() in shared.

Roman