ARM64: Improve code generated to spill/restore for slow paths.

Aligning the accesses allows generating better code.

Before:

    add x16, sp, #0x44 (68)
    stp x0, x1, [x16, #-16]

After:

    stp x0, x1, [sp, #56]

Change-Id: I3e20ad3fa59d00aee4b4d14ea9d59c7cd546509e
diff --git a/compiler/optimizing/code_generator_arm64.h b/compiler/optimizing/code_generator_arm64.h
index c2f055a..b35c520 100644
--- a/compiler/optimizing/code_generator_arm64.h
+++ b/compiler/optimizing/code_generator_arm64.h
@@ -418,6 +418,10 @@
     block_labels_.resize(GetGraph()->GetBlocks().size());
   }
 
+  // We want to use the STP and LDP instructions to spill and restore registers for slow paths.
+  // These instructions can only encode offsets that are multiples of the register size accessed.
+  uint32_t GetPreferredSlotsAlignment() const OVERRIDE { return vixl::kXRegSizeInBytes; }
+
   JumpTableARM64* CreateJumpTable(HPackedSwitch* switch_instr) {
     jump_tables_.emplace_back(new (GetGraph()->GetArena()) JumpTableARM64(switch_instr));
     return jump_tables_.back().get();