diff -r b1a394e15ae9 -r 0d7877278adf src/hotspot/cpu/x86/gc/shenandoah/shenandoahBarrierSetAssembler_x86.cpp
--- a/src/hotspot/cpu/x86/gc/shenandoah/shenandoahBarrierSetAssembler_x86.cpp	Wed Sep 18 20:56:18 2019 +0200
+++ b/src/hotspot/cpu/x86/gc/shenandoah/shenandoahBarrierSetAssembler_x86.cpp	Wed Sep 18 20:56:19 2019 +0200
@@ -47,35 +47,28 @@
 void ShenandoahBarrierSetAssembler::arraycopy_prologue(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
                                                        Register src, Register dst, Register count) {
 
-  bool checkcast = (decorators & ARRAYCOPY_CHECKCAST) != 0;
-  bool disjoint = (decorators & ARRAYCOPY_DISJOINT) != 0;
-  bool obj_int = type == T_OBJECT LP64_ONLY(&& UseCompressedOops);
   bool dest_uninitialized = (decorators & IS_DEST_UNINITIALIZED) != 0;
 
   if (type == T_OBJECT || type == T_ARRAY) {
+
+    if ((ShenandoahSATBBarrier && !dest_uninitialized) || ShenandoahLoadRefBarrier) {
 #ifdef _LP64
-    if (!checkcast) {
-      if (!obj_int) {
-        // Save count for barrier
-        __ movptr(r11, count);
-      } else if (disjoint) {
-        // Save dst in r11 in the disjoint case
-        __ movq(r11, dst);
+      Register thread = r15_thread;
+#else
+      Register thread = rax;
+      if (thread == src || thread == dst || thread == count) {
+        thread = rbx;
       }
-    }
-#else
-    if (disjoint) {
-      __ mov(rdx, dst);          // save 'to'
-    }
-#endif
-
-    if (ShenandoahSATBBarrier && !dest_uninitialized) {
-      Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread);
-      assert_different_registers(dst, count, thread); // we don't care about src here?
-#ifndef _LP64
+      if (thread == src || thread == dst || thread == count) {
+        thread = rcx;
+      }
+      if (thread == src || thread == dst || thread == count) {
+        thread = rdx;
+      }
       __ push(thread);
       __ get_thread(thread);
 #endif
+      assert_different_registers(src, dst, count, thread);
 
       Label done;
       // Short-circuit if count == 0.
@@ -84,32 +77,33 @@
 
       // Avoid runtime call when not marking.
       Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
-      __ testb(gc_state, ShenandoahHeap::MARKING);
+      int flags = ShenandoahHeap::HAS_FORWARDED;
+      if (!dest_uninitialized) {
+        flags |= ShenandoahHeap::MARKING;
+      }
+      __ testb(gc_state, flags);
       __ jcc(Assembler::zero, done);
 
       __ pusha();                      // push registers
 #ifdef _LP64
-      if (count == c_rarg0) {
-        if (dst == c_rarg1) {
-          // exactly backwards!!
-          __ xchgptr(c_rarg1, c_rarg0);
+      assert(src == rdi, "expected");
+      assert(dst == rsi, "expected");
+      assert(count == rdx, "expected");
+      if (UseCompressedOops) {
+        if (dest_uninitialized) {
+          __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_duinit_narrow_oop_entry), src, dst, count);
         } else {
-          __ movptr(c_rarg1, count);
-          __ movptr(c_rarg0, dst);
+          __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_narrow_oop_entry), src, dst, count);
         }
-      } else {
-        __ movptr(c_rarg0, dst);
-        __ movptr(c_rarg1, count);
+      } else
+#endif
+      {
+        if (dest_uninitialized) {
+          __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_duinit_oop_entry), src, dst, count);
+        } else {
+          __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_oop_entry), src, dst, count);
+        }
       }
-      if (UseCompressedOops) {
-        __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_narrow_oop_entry), 2);
-      } else {
-        __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_oop_entry), 2);
-      }
-#else
-      __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_oop_entry),
-                      dst, count);
-#endif
       __ popa();
       __ bind(done);
       NOT_LP64(__ pop(thread);)
@@ -118,73 +112,6 @@
 
 }
 
-void ShenandoahBarrierSetAssembler::arraycopy_epilogue(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
-                                                       Register src, Register dst, Register count) {
-  bool checkcast = (decorators & ARRAYCOPY_CHECKCAST) != 0;
-  bool disjoint = (decorators & ARRAYCOPY_DISJOINT) != 0;
-  bool obj_int = type == T_OBJECT LP64_ONLY(&& UseCompressedOops);
-  Register tmp = rax;
-
-  if (type == T_OBJECT || type == T_ARRAY) {
-#ifdef _LP64
-    if (!checkcast) {
-      if (!obj_int) {
-        // Save count for barrier
-        count = r11;
-      } else if (disjoint && obj_int) {
-        // Use the saved dst in the disjoint case
-        dst = r11;
-      }
-    } else {
-      tmp = rscratch1;
-    }
-#else
-    if (disjoint) {
-      __ mov(dst, rdx); // restore 'to'
-    }
-#endif
-
-    Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread);
-    assert_different_registers(dst, thread); // do we care about src at all here?
-
-#ifndef _LP64
-    __ push(thread);
-    __ get_thread(thread);
-#endif
-
-    // Short-circuit if count == 0.
-    Label done;
-    __ testptr(count, count);
-    __ jcc(Assembler::zero, done);
-
-    // Skip runtime call if no forwarded objects.
-    Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
-    __ testb(gc_state, ShenandoahHeap::UPDATEREFS);
-    __ jcc(Assembler::zero, done);
-
-    __ pusha();             // push registers (overkill)
-#ifdef _LP64
-    if (c_rarg0 == count) { // On win64 c_rarg0 == rcx
-      assert_different_registers(c_rarg1, dst);
-      __ mov(c_rarg1, count);
-      __ mov(c_rarg0, dst);
-    } else {
-      assert_different_registers(c_rarg0, count);
-      __ mov(c_rarg0, dst);
-      __ mov(c_rarg1, count);
-    }
-    __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_post_entry), 2);
-#else
-    __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_post_entry),
-                    dst, count);
-#endif
-    __ popa();
-
-    __ bind(done);
-    NOT_LP64(__ pop(thread);)
-  }
-}
-
 void ShenandoahBarrierSetAssembler::shenandoah_write_barrier_pre(MacroAssembler* masm,
                                                                  Register obj,
                                                                  Register pre_val,