8205523: Explicit barriers for interpreter

Reviewed-by: eosterlund, coleenp

src/hotspot/cpu/aarch64/gc/shared/barrierSetAssembler_aarch64.hpp

@@ -48,6 +48,10 @@ public:
   virtual void obj_equals(MacroAssembler* masm,
                           Register obj1, Register obj2);
 
+  virtual void resolve(MacroAssembler* masm, DecoratorSet decorators, Register obj) {
+    // Default implementation does not need to do anything.
+  }
+
   virtual void try_resolve_jobject_in_native(MacroAssembler* masm, Register jni_env,
                                              Register obj, Register tmp, Label& slowpath);
 

src/hotspot/cpu/aarch64/macroAssembler_aarch64.cpp

@@ -3990,6 +3990,15 @@ void MacroAssembler::access_store_at(BasicType type, DecoratorSet decorators,
   }
 }
 
+void MacroAssembler::resolve(DecoratorSet decorators, Register obj) {
+  // Use stronger ACCESS_WRITE|ACCESS_READ by default.
+  if ((decorators & (ACCESS_READ | ACCESS_WRITE)) == 0) {
+    decorators |= ACCESS_READ | ACCESS_WRITE;
+  }
+  BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
+  return bs->resolve(this, decorators, obj);
+}
+
 void MacroAssembler::load_heap_oop(Register dst, Address src, Register tmp1,
                                    Register thread_tmp, DecoratorSet decorators) {
   access_load_at(T_OBJECT, IN_HEAP | decorators, dst, src, tmp1, thread_tmp);

src/hotspot/cpu/aarch64/macroAssembler_aarch64.hpp

@@ -795,6 +795,10 @@ public:
   void access_store_at(BasicType type, DecoratorSet decorators, Address dst, Register src,
                        Register tmp1, Register tmp_thread);
 
+  // Resolves obj for access. Result is placed in the same register.
+  // All other registers are preserved.
+  void resolve(DecoratorSet decorators, Register obj);
+
   void load_heap_oop(Register dst, Address src, Register tmp1 = noreg,
                      Register thread_tmp = noreg, DecoratorSet decorators = 0);
 

src/hotspot/cpu/aarch64/sharedRuntime_aarch64.cpp

@@ -1839,6 +1839,8 @@ nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
     // Load the oop from the handle
     __ ldr(obj_reg, Address(oop_handle_reg, 0));
 
+    __ resolve(IS_NOT_NULL, obj_reg);
+
     if (UseBiasedLocking) {
       __ biased_locking_enter(lock_reg, obj_reg, swap_reg, tmp, false, lock_done, &slow_path_lock);
     }
@@ -2001,6 +2003,8 @@ nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
     // Get locked oop from the handle we passed to jni
     __ ldr(obj_reg, Address(oop_handle_reg, 0));
 
+    __ resolve(IS_NOT_NULL, obj_reg);
+
     Label done;
 
     if (UseBiasedLocking) {

src/hotspot/cpu/aarch64/templateInterpreterGenerator_aarch64.cpp

@@ -836,6 +836,7 @@ void TemplateInterpreterGenerator::lock_method() {
 #endif // ASSERT
 
     __ bind(done);
+    __ resolve(IS_NOT_NULL, r0);
   }
 
   // add space for monitor & lock
@@ -1062,6 +1063,7 @@ address TemplateInterpreterGenerator::generate_CRC32_updateBytes_entry(AbstractI
       __ ldrw(crc,   Address(esp, 4*wordSize)); // Initial CRC
     } else {
       __ ldr(buf, Address(esp, 2*wordSize)); // byte[] array
+      __ resolve(IS_NOT_NULL | ACCESS_READ, buf);
       __ add(buf, buf, arrayOopDesc::base_offset_in_bytes(T_BYTE)); // + header size
       __ ldrw(off, Address(esp, wordSize)); // offset
       __ add(buf, buf, off); // + offset
@@ -1106,6 +1108,9 @@ address TemplateInterpreterGenerator::generate_CRC32C_updateBytes_entry(Abstract
     __ ldrw(off, Address(esp, wordSize)); // int offset
     __ sub(len, end, off);
     __ ldr(buf, Address(esp, 2*wordSize)); // byte[] buf | long buf
+    if (kind == Interpreter::java_util_zip_CRC32C_updateBytes) {
+      __ resolve(IS_NOT_NULL | ACCESS_READ, buf);
+    }
     __ add(buf, buf, off); // + offset
     if (kind == Interpreter::java_util_zip_CRC32C_updateDirectByteBuffer) {
       __ ldrw(crc, Address(esp, 4*wordSize)); // long crc

src/hotspot/cpu/aarch64/templateTable_aarch64.cpp

@@ -3840,6 +3840,8 @@ void TemplateTable::monitorenter()
   // check for NULL object
   __ null_check(r0);
 
+  __ resolve(IS_NOT_NULL, r0);
+
   const Address monitor_block_top(
         rfp, frame::interpreter_frame_monitor_block_top_offset * wordSize);
   const Address monitor_block_bot(
@@ -3939,6 +3941,8 @@ void TemplateTable::monitorexit()
   // check for NULL object
   __ null_check(r0);
 
+  __ resolve(IS_NOT_NULL, r0);
+
   const Address monitor_block_top(
         rfp, frame::interpreter_frame_monitor_block_top_offset * wordSize);
   const Address monitor_block_bot(

src/hotspot/cpu/x86/gc/shared/barrierSetAssembler_x86.hpp

@@ -61,6 +61,10 @@ public:
   virtual void obj_equals(MacroAssembler* masm,
                           Register obj1, Address obj2);
 
+  virtual void resolve(MacroAssembler* masm, DecoratorSet decorators, Register obj) {
+    // Default implementation does not need to do anything.
+  }
+
   // Support for jniFastGetField to try resolving a jobject/jweak in native
   virtual void try_resolve_jobject_in_native(MacroAssembler* masm, Register jni_env,
                                              Register obj, Register tmp, Label& slowpath);

src/hotspot/cpu/x86/macroAssembler_x86.cpp

@@ -6287,6 +6287,15 @@ void MacroAssembler::access_store_at(BasicType type, DecoratorSet decorators, Ad
   }
 }
 
+void MacroAssembler::resolve(DecoratorSet decorators, Register obj) {
+  // Use stronger ACCESS_WRITE|ACCESS_READ by default.
+  if ((decorators & (ACCESS_READ | ACCESS_WRITE)) == 0) {
+    decorators |= ACCESS_READ | ACCESS_WRITE;
+  }
+  BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
+  return bs->resolve(this, decorators, obj);
+}
+
 void MacroAssembler::load_heap_oop(Register dst, Address src, Register tmp1,
                                    Register thread_tmp, DecoratorSet decorators) {
   access_load_at(T_OBJECT, IN_HEAP | decorators, dst, src, tmp1, thread_tmp);

src/hotspot/cpu/x86/macroAssembler_x86.hpp

@@ -319,6 +319,10 @@ class MacroAssembler: public Assembler {
   void access_store_at(BasicType type, DecoratorSet decorators, Address dst, Register src,
                        Register tmp1, Register tmp2);
 
+  // Resolves obj access. Result is placed in the same register.
+  // All other registers are preserved.
+  void resolve(DecoratorSet decorators, Register obj);
+
   void load_heap_oop(Register dst, Address src, Register tmp1 = noreg,
                      Register thread_tmp = noreg, DecoratorSet decorators = 0);
   void load_heap_oop_not_null(Register dst, Address src, Register tmp1 = noreg,

src/hotspot/cpu/x86/sharedRuntime_x86_64.cpp

@@ -2450,6 +2450,7 @@ nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
     // Load the oop from the handle
     __ movptr(obj_reg, Address(oop_handle_reg, 0));
 
+    __ resolve(IS_NOT_NULL, obj_reg);
     if (UseBiasedLocking) {
       __ biased_locking_enter(lock_reg, obj_reg, swap_reg, rscratch1, false, lock_done, &slow_path_lock);
     }
@@ -2635,6 +2636,7 @@ nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
 
     // Get locked oop from the handle we passed to jni
     __ movptr(obj_reg, Address(oop_handle_reg, 0));
+    __ resolve(IS_NOT_NULL, obj_reg);
 
     Label done;
 

src/hotspot/cpu/x86/templateInterpreterGenerator_x86.cpp

@@ -635,6 +635,7 @@ void TemplateInterpreterGenerator::lock_method() {
 #endif // ASSERT
 
     __ bind(done);
+    __ resolve(IS_NOT_NULL, rax);
   }
 
   // add space for monitor & lock

src/hotspot/cpu/x86/templateInterpreterGenerator_x86_64.cpp

@@ -257,6 +257,7 @@ address TemplateInterpreterGenerator::generate_CRC32_updateBytes_entry(AbstractI
       __ movl(crc,   Address(rsp, 5*wordSize)); // Initial CRC
     } else {
       __ movptr(buf, Address(rsp, 3*wordSize)); // byte[] array
+      __ resolve(IS_NOT_NULL | ACCESS_READ, buf);
       __ addptr(buf, arrayOopDesc::base_offset_in_bytes(T_BYTE)); // + header size
       __ movl2ptr(off, Address(rsp, 2*wordSize)); // offset
       __ addq(buf, off); // + offset
@@ -312,6 +313,7 @@ address TemplateInterpreterGenerator::generate_CRC32C_updateBytes_entry(Abstract
       //    "When calculating operand stack length, values of type long and double have length two."
     } else {
       __ movptr(buf, Address(rsp, 3 * wordSize)); // byte[] array
+      __ resolve(IS_NOT_NULL | ACCESS_READ, buf);
       __ addptr(buf, arrayOopDesc::base_offset_in_bytes(T_BYTE)); // + header size
       __ movl2ptr(off, Address(rsp, 2 * wordSize)); // offset
       __ addq(buf, off); // + offset

src/hotspot/cpu/x86/templateTable_x86.cpp

@@ -4357,6 +4357,8 @@ void TemplateTable::monitorenter() {
   // check for NULL object
   __ null_check(rax);
 
+  __ resolve(IS_NOT_NULL, rax);
+
   const Address monitor_block_top(
         rbp, frame::interpreter_frame_monitor_block_top_offset * wordSize);
   const Address monitor_block_bot(
@@ -4454,6 +4456,8 @@ void TemplateTable::monitorexit() {
   // check for NULL object
   __ null_check(rax);
 
+  __ resolve(IS_NOT_NULL, rax);
+
   const Address monitor_block_top(
         rbp, frame::interpreter_frame_monitor_block_top_offset * wordSize);
   const Address monitor_block_bot(

src/hotspot/share/oops/accessDecorators.hpp

@@ -212,8 +212,15 @@ const DecoratorSet ARRAYCOPY_DECORATOR_MASK       = ARRAYCOPY_CHECKCAST | ARRAYC
                                                     ARRAYCOPY_DISJOINT | ARRAYCOPY_ARRAYOF |
                                                     ARRAYCOPY_ATOMIC | ARRAYCOPY_ALIGNED;
 
+// == Resolve barrier decorators ==
+// * ACCESS_READ: Indicate that the resolved object is accessed read-only. This allows the GC
+//   backend to use weaker and more efficient barriers.
+// * ACCESS_WRITE: Indicate that the resolved object is used for write access.
+const DecoratorSet ACCESS_READ                    = UCONST64(1) << 29;
+const DecoratorSet ACCESS_WRITE                   = UCONST64(1) << 30;
+
 // Keep track of the last decorator.
-const DecoratorSet DECORATOR_LAST = UCONST64(1) << 28;
+const DecoratorSet DECORATOR_LAST = UCONST64(1) << 30;
 
 namespace AccessInternal {
   // This class adds implied decorators that follow according to decorator rules.