--- a/hotspot/src/cpu/aarch64/vm/macroAssembler_aarch64.cpp	Thu Jun 18 10:47:34 2015 -0400
+++ b/hotspot/src/cpu/aarch64/vm/macroAssembler_aarch64.cpp	Fri Jun 19 09:41:49 2015 +0200
@@ -2888,41 +2888,40 @@
   cmp(src1, rscratch1);
 }
 
-void MacroAssembler::store_check(Register obj) {
-  // Does a store check for the oop in register obj. The content of
-  // register obj is destroyed afterwards.
-  store_check_part_1(obj);
-  store_check_part_2(obj);
-}
-
 void MacroAssembler::store_check(Register obj, Address dst) {
   store_check(obj);
 }
 
-
-// split the store check operation so that other instructions can be scheduled inbetween
-void MacroAssembler::store_check_part_1(Register obj) {
-  BarrierSet* bs = Universe::heap()->barrier_set();
-  assert(bs->kind() == BarrierSet::CardTableModRef, "Wrong barrier set kind");
-  lsr(obj, obj, CardTableModRefBS::card_shift);
-}
-
-void MacroAssembler::store_check_part_2(Register obj) {
+void MacroAssembler::store_check(Register obj) {
+  // Does a store check for the oop in register obj. The content of
+  // register obj is destroyed afterwards.
+
   BarrierSet* bs = Universe::heap()->barrier_set();
   assert(bs->kind() == BarrierSet::CardTableModRef, "Wrong barrier set kind");
-  CardTableModRefBS* ct = (CardTableModRefBS*)bs;
+
+  CardTableModRefBS* ct = barrier_set_cast<CardTableModRefBS>(bs);
   assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code");
 
-  // The calculation for byte_map_base is as follows:
-  // byte_map_base = _byte_map - (uintptr_t(low_bound) >> card_shift);
-  // So this essentially converts an address to a displacement and
-  // it will never need to be relocated.
-
-  // FIXME: It's not likely that disp will fit into an offset so we
-  // don't bother to check, but it could save an instruction.
-  intptr_t disp = (intptr_t) ct->byte_map_base;
-  mov(rscratch1, disp);
-  strb(zr, Address(obj, rscratch1));
+  lsr(obj, obj, CardTableModRefBS::card_shift);
+
+  assert(CardTableModRefBS::dirty_card_val() == 0, "must be");
+
+  {
+    ExternalAddress cardtable((address) ct->byte_map_base);
+    unsigned long offset;
+    adrp(rscratch1, cardtable, offset);
+    assert(offset == 0, "byte_map_base is misaligned");
+  }
+
+  if (UseCondCardMark) {
+    Label L_already_dirty;
+    ldrb(rscratch2,  Address(obj, rscratch1));
+    cbz(rscratch2, L_already_dirty);
+    strb(zr, Address(obj, rscratch1));
+    bind(L_already_dirty);
+  } else {
+    strb(zr, Address(obj, rscratch1));
+  }
 }
 
 void MacroAssembler::load_klass(Register dst, Register src) {

--- a/hotspot/src/cpu/aarch64/vm/macroAssembler_aarch64.hpp	Thu Jun 18 10:47:34 2015 -0400
+++ b/hotspot/src/cpu/aarch64/vm/macroAssembler_aarch64.hpp	Fri Jun 19 09:41:49 2015 +0200
@@ -756,10 +756,6 @@
 
 #endif // INCLUDE_ALL_GCS
 
-  // split store_check(Register obj) to enhance instruction interleaving
-  void store_check_part_1(Register obj);
-  void store_check_part_2(Register obj);
-
   // oop manipulations
   void load_klass(Register dst, Register src);
   void store_klass(Register dst, Register src);

--- a/hotspot/src/cpu/x86/vm/macroAssembler_x86.cpp	Thu Jun 18 10:47:34 2015 -0400
+++ b/hotspot/src/cpu/x86/vm/macroAssembler_x86.cpp	Fri Jun 19 09:41:49 2015 +0200
@@ -4260,31 +4260,24 @@
 //////////////////////////////////////////////////////////////////////////////////
 
 
+void MacroAssembler::store_check(Register obj, Address dst) {
+  store_check(obj);
+}
+
 void MacroAssembler::store_check(Register obj) {
   // Does a store check for the oop in register obj. The content of
   // register obj is destroyed afterwards.
-  store_check_part_1(obj);
-  store_check_part_2(obj);
-}
-
-void MacroAssembler::store_check(Register obj, Address dst) {
-  store_check(obj);
-}
-
-
-// split the store check operation so that other instructions can be scheduled inbetween
-void MacroAssembler::store_check_part_1(Register obj) {
+
   BarrierSet* bs = Universe::heap()->barrier_set();
   assert(bs->kind() == BarrierSet::CardTableModRef, "Wrong barrier set kind");
-  shrptr(obj, CardTableModRefBS::card_shift);
-}
-
-void MacroAssembler::store_check_part_2(Register obj) {
-  BarrierSet* bs = Universe::heap()->barrier_set();
-  assert(bs->kind() == BarrierSet::CardTableModRef, "Wrong barrier set kind");
+
   CardTableModRefBS* ct = barrier_set_cast<CardTableModRefBS>(bs);
   assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code");
 
+  shrptr(obj, CardTableModRefBS::card_shift);
+
+  Address card_addr;
+
   // The calculation for byte_map_base is as follows:
   // byte_map_base = _byte_map - (uintptr_t(low_bound) >> card_shift);
   // So this essentially converts an address to a displacement and it will
@@ -4292,8 +4285,7 @@
   // large for a 32bit displacement.
   intptr_t disp = (intptr_t) ct->byte_map_base;
   if (is_simm32(disp)) {
-    Address cardtable(noreg, obj, Address::times_1, disp);
-    movb(cardtable, 0);
+    card_addr = Address(noreg, obj, Address::times_1, disp);
   } else {
     // By doing it as an ExternalAddress 'disp' could be converted to a rip-relative
     // displacement and done in a single instruction given favorable mapping and a
@@ -4301,7 +4293,18 @@
     // entry and that entry is not properly handled by the relocation code.
     AddressLiteral cardtable((address)ct->byte_map_base, relocInfo::none);
     Address index(noreg, obj, Address::times_1);
-    movb(as_Address(ArrayAddress(cardtable, index)), 0);
+    card_addr = as_Address(ArrayAddress(cardtable, index));
+  }
+
+  int dirty = CardTableModRefBS::dirty_card_val();
+  if (UseCondCardMark) {
+    Label L_already_dirty;
+    cmpb(card_addr, dirty);
+    jcc(Assembler::equal, L_already_dirty);
+    movb(card_addr, dirty);
+    bind(L_already_dirty);
+  } else {
+    movb(card_addr, dirty);
   }
 }
 

--- a/hotspot/src/cpu/x86/vm/macroAssembler_x86.hpp	Thu Jun 18 10:47:34 2015 -0400
+++ b/hotspot/src/cpu/x86/vm/macroAssembler_x86.hpp	Fri Jun 19 09:41:49 2015 +0200
@@ -315,10 +315,6 @@
 
 #endif // INCLUDE_ALL_GCS
 
-  // split store_check(Register obj) to enhance instruction interleaving
-  void store_check_part_1(Register obj);
-  void store_check_part_2(Register obj);
-
   // C 'boolean' to Java boolean: x == 0 ? 0 : 1
   void c2bool(Register x);