Compile::reshape_address() in aarch64.ad makes some changes which in theory should generate better code. However, after recent changes to C2's loop unrolling, reshape_address() causes many new temporaries to be generated. We run out of registers and they are spilled to stack slots, slowing down inner loops.

Benchmarking reveals that we would be better off without Compile::reshape_address().

http://mail.openjdk.java.net/pipermail/aarch64-port-dev/2018-May/005734.html