RFR: fix loop unswitching with -XX:-ShenandoahWriteBarrierMemBar

[Roland Westrelin]

> But it may affect compiler optimizations? Or if something went wrong,
> inhibit them?

I checked that the optimizations that happen after the write barrier is
expanded trigger ShenandoahWriteBarrierMemBar on/off correctly trigger
with simple test cases.

> True. But we have both observed a noticable regression. Better to
> check it? No-membars will likely become the default/only path, so it
> better be good ;-)

I ran Compiler for longer (-f 10) a couple of times and results with
ShenandoahWriteBarrierMemBar on/off are almost identical:

-ShenandoahWriteBarrierMemBar: Compiler.compiler  thrpt  100  384.413 ± 11.802  ops/min
+ShenandoahWriteBarrierMemBar: Compiler.compiler  thrpt  100  384.253 ± 10.919  ops/min
-ShenandoahWriteBarrierMemBar: Compiler.compiler  thrpt  100  382.996 ± 10.339  ops/min
+ShenandoahWriteBarrierMemBar: Compiler.compiler  thrpt  100  385.405 ± 10.961  ops/min

Roland.