c2: loop optimizations for read barriers

[Roman Kennke]

Great work! Go!


Roman

Am Freitag, den 16.09.2016, 09:22 +0200 schrieb Roland Westrelin:
> http://cr.openjdk.java.net/~roland/shenandoah/rb-loop-opts/webrev.00/
> 
> The change to PhaseIdealLoop::split_if_with_blocks_post() in
> loopopts.cpp enables sinking of read barriers out of loops: if a read
> barrier is in a loop but all of its uses are out of loop, the read
> barrier is cloned and pinned out of loop for each use.
> 
> try_move_shenandoah_read_barrier() tries to move the memory edge of a
> read barrier as high in the graph as legal. Read barriers are
> sometimes
> stuck in a loop because of their memory edge. That change would fix
> that. It should also help in removing redundant read barriers (they
> should all end up with the same memory input). Because we now have a
> global pass on read barriers, restricting
> ShenandoahReadBarrierNode::dominates_memory_rb_impl() to be local to
> the
> node being optimized shouldn't cause a performance regression.
> 
> Roland.