release carrier on Object.wait

Author: pchilano

src/hotspot/cpu/aarch64/continuationFreezeThaw_aarch64.inline.hpp

@@ -155,10 +155,16 @@ inline void FreezeBase::relativize_interpreted_frame_metadata(const frame& f, co
   // extended_sp is already relativized by TemplateInterpreterGenerator::generate_normal_entry or
   // AbstractInterpreter::layout_activation
 
+  // The interpreter native wrapper code adds space in the stack equal to size_of_parameters()
+  // after the fixed part of the frame. For wait0 this is equal to 3 words (this + long parameter).
+  // We adjust by this size since otherwise the saved last sp will be less than the extended_sp.
+  DEBUG_ONLY(Method* m = hf.interpreter_frame_method();)
+  DEBUG_ONLY(int extra_space = m->is_object_wait0() ? m->size_of_parameters() : 0;)
+
   assert((hf.fp() - hf.unextended_sp()) == (f.fp() - f.unextended_sp()), "");
   assert(hf.unextended_sp() == (intptr_t*)hf.at(frame::interpreter_frame_last_sp_offset), "");
   assert(hf.unextended_sp() <= (intptr_t*)hf.at(frame::interpreter_frame_initial_sp_offset), "");
-  assert(hf.unextended_sp() >  (intptr_t*)hf.at(frame::interpreter_frame_extended_sp_offset), "");
+  assert(hf.unextended_sp() + extra_space >  (intptr_t*)hf.at(frame::interpreter_frame_extended_sp_offset), "");
   assert(hf.fp()            >  (intptr_t*)hf.at(frame::interpreter_frame_initial_sp_offset), "");
   assert(hf.fp()            <= (intptr_t*)hf.at(frame::interpreter_frame_locals_offset), "");
 }
@@ -219,7 +225,6 @@ template<typename FKind> frame ThawBase::new_stack_frame(const frame& hf, frame&
     // If caller is interpreted it already made room for the callee arguments
     int overlap = caller.is_interpreted_frame() ? ContinuationHelper::InterpretedFrame::stack_argsize(hf) : 0;
     const int fsize = (int)(ContinuationHelper::InterpretedFrame::frame_bottom(hf) - hf.unextended_sp() - overlap);
-    const int locals = hf.interpreter_frame_method()->max_locals();
     intptr_t* frame_sp = caller.unextended_sp() - fsize;
     intptr_t* fp = frame_sp + (hf.fp() - heap_sp);
     if ((intptr_t)fp % frame::frame_alignment != 0) {
@@ -258,7 +263,7 @@ template<typename FKind> frame ThawBase::new_stack_frame(const frame& hf, frame&
       // we need to recreate a "real" frame pointer, pointing into the stack
       fp = frame_sp + FKind::size(hf) - frame::sender_sp_offset;
     } else {
-      fp = FKind::stub
+      fp = FKind::stub || FKind::native
         ? frame_sp + fsize - frame::sender_sp_offset // fp always points to the address below the pushed return pc. We need correct address.
         : *(intptr_t**)(hf.sp() - frame::sender_sp_offset); // we need to re-read fp because it may be an oop and we might have fixed the frame.
     }
@@ -287,10 +292,32 @@ inline void ThawBase::patch_pd(frame& f, intptr_t* caller_sp) {
   Unimplemented();
 }
 
-inline intptr_t* ThawBase::push_preempt_rerun_adapter(frame top, bool is_interpreted_frame) {
+inline void ThawBase::fix_native_return_pc_pd(frame& top) {
+  // Nothing to do since the last pc saved before making the call to
+  // JVM_MonitorWait() was already set to the correct resume pc. Just
+  // do some sanity check.
+#ifdef ASSERT
+  Method* method = top.is_interpreted_frame() ? top.interpreter_frame_method() : CodeCache::find_blob(top.pc())->as_nmethod()->method();
+  assert(method->is_object_wait0(), "");
+#endif
+}
+
+inline intptr_t* ThawBase::push_resume_adapter(frame& top) {
   intptr_t* sp = top.sp();
   CodeBlob* cb = top.cb();
-  if (!is_interpreted_frame && cb->frame_size() == 2) {
+
+#ifdef ASSERT
+  RegisterMap map(JavaThread::current(),
+                  RegisterMap::UpdateMap::skip,
+                  RegisterMap::ProcessFrames::skip,
+                  RegisterMap::WalkContinuation::skip);
+  frame caller = top.sender(&map);
+  intptr_t link_addr = (intptr_t)ContinuationHelper::Frame::callee_link_address(caller);
+  assert(sp[-2] == link_addr, "wrong link address: " INTPTR_FORMAT " != " INTPTR_FORMAT, sp[-2], link_addr);
+#endif
+
+  bool interpreted = top.is_interpreted_frame();
+  if (!interpreted && cb->frame_size() == 2) {
     // C2 runtime stub case. For aarch64 the real size of the c2 runtime stub is 2 words bigger
     // than what we think, i.e. size is 4. This is because the _last_Java_sp is not set to the
     // sp right before making the call to the VM, but rather it is artificially set 2 words above
@@ -306,21 +333,19 @@ inline intptr_t* ThawBase::push_preempt_rerun_adapter(frame top, bool is_interpr
   }
 
   intptr_t* fp = sp - frame::sender_sp_offset;
-  address pc = is_interpreted_frame ? Interpreter::cont_preempt_rerun_interpreter_adapter()
-                                    : StubRoutines::cont_preempt_rerun_compiler_adapter();
+  address pc = interpreted ? Interpreter::cont_resume_interpreter_adapter()
+                           : StubRoutines::cont_resume_compiler_adapter();
 
   sp -= frame::metadata_words;
   *(address*)(sp - frame::sender_sp_ret_address_offset()) = pc;
   *(intptr_t**)(sp - frame::sender_sp_offset) = fp;
 
-  log_develop_trace(continuations, preempt)("push_preempt_rerun_%s_adapter() initial sp: " INTPTR_FORMAT " final sp: " INTPTR_FORMAT " fp: " INTPTR_FORMAT,
-    is_interpreted_frame ? "interpreter" : "safepointblob", p2i(sp + frame::metadata_words), p2i(sp), p2i(fp));
+  log_develop_trace(continuations, preempt)("push_resume_%s_adapter() initial sp: " INTPTR_FORMAT " final sp: " INTPTR_FORMAT " fp: " INTPTR_FORMAT,
+    interpreted ? "interpreter" : "compiler", p2i(sp + frame::metadata_words), p2i(sp), p2i(fp));
   return sp;
 }
 
-inline intptr_t* ThawBase::push_preempt_monitorenter_redo(stackChunkOop chunk) {
-
-  // fprintf(stderr, "push_preempt_monitorenter_redo\n");
+inline intptr_t* ThawBase::push_resume_monitor_operation(stackChunkOop chunk) {
   frame enterSpecial = new_entry_frame();
   intptr_t* sp = enterSpecial.sp();
 
@@ -329,15 +354,15 @@ inline intptr_t* ThawBase::push_preempt_monitorenter_redo(stackChunkOop chunk) {
   sp[1] = (intptr_t)StubRoutines::cont_returnBarrier();
   sp[0] = (intptr_t)enterSpecial.fp();
 
-  // Now push the ObjectMonitor*
+  // Now push the ObjectWaiter*
   sp -= frame::metadata_words;
-  sp[1] = (intptr_t)chunk->objectMonitor(); // alignment
-  sp[0] = (intptr_t)chunk->objectMonitor();
+  sp[1] = (intptr_t)chunk->object_waiter(); // alignment
+  sp[0] = (intptr_t)chunk->object_waiter();
 
-  // Finally arrange to return to the monitorenter_redo stub
-  sp[-1] = (intptr_t)StubRoutines::cont_preempt_monitorenter_redo();
+  // Finally arrange to return to the resume_monitor_operation stub
+  sp[-1] = (intptr_t)StubRoutines::cont_resume_monitor_operation();
   sp[-2] = (intptr_t)enterSpecial.fp();
-  log_develop_trace(continuations, preempt)("push_preempt_monitorenter_redo initial sp: " INTPTR_FORMAT " final sp: " INTPTR_FORMAT, p2i(sp + 2 * frame::metadata_words), p2i(sp));
+  log_develop_trace(continuations, preempt)("push_resume_monitor_operation initial sp: " INTPTR_FORMAT " final sp: " INTPTR_FORMAT, p2i(sp + 2 * frame::metadata_words), p2i(sp));
   return sp;
 }
 
@@ -349,7 +374,9 @@ inline void ThawBase::derelativize_interpreted_frame_metadata(const frame& hf, c
   assert(f.at_absolute(frame::interpreter_frame_monitor_block_top_offset) <= frame::interpreter_frame_initial_sp_offset, "");
 
   // Make sure that extended_sp is kept relativized.
-  assert((intptr_t*)f.at_relative(frame::interpreter_frame_extended_sp_offset) < f.unextended_sp(), "");
+  DEBUG_ONLY(Method* m = hf.interpreter_frame_method();)
+  DEBUG_ONLY(int extra_space = m->is_object_wait0() ? m->size_of_parameters() : 0;) // see comment in relativize_interpreted_frame_metadata()
+  assert((intptr_t*)f.at_relative(frame::interpreter_frame_extended_sp_offset) < f.unextended_sp() + extra_space, "");
 }
 
 #endif // CPU_AARCH64_CONTINUATIONFREEZETHAW_AARCH64_INLINE_HPP

src/hotspot/cpu/aarch64/continuationHelper_aarch64.inline.hpp

@@ -41,21 +41,18 @@ static inline intptr_t** link_address(const frame& f) {
 }
 
 static inline void patch_return_pc_with_preempt_stub(frame& f) {
-  // Unlike x86 we don't know where in the callee frame the return pc is
-  // saved so we can't patch the return from the VM call back to Java. If
-  // the target is coming from compiled code we will patch the return from
-  // the safepoint handler blob back to the compiled method instead. If
-  // it's coming from the interpreter, the target will check for preemption
-  // once it returns to the interpreter and will manually jump to the
-  // preempt stub.
-  if (!f.is_interpreted_frame()) {
-    assert(f.is_runtime_frame(), "invariant");
+  if (f.is_runtime_frame()) {
+    // Unlike x86 we don't know where in the callee frame the return pc is
+    // saved so we can't patch the return from the VM call back to Java.
+    // Instead, we will patch the return from the runtime stub back to the
+    // compiled method so that the target returns to the preempt cleanup stub.
     intptr_t* caller_sp = f.sp() + f.cb()->frame_size();
     caller_sp[-1] = (intptr_t)StubRoutines::cont_preempt_stub();
   } else {
+    // The target will check for preemption once it returns to the interpreter
+    // or the native wrapper code and will manually jump to the preempt stub.
     JavaThread *thread = JavaThread::current();
     thread->set_preempt_alternate_return(StubRoutines::cont_preempt_stub());
-    thread->set_preempt_alternate_return_sp((address)f.sp());
   }
 }
 

src/hotspot/cpu/aarch64/frame_aarch64.hpp

@@ -73,7 +73,8 @@
     sender_sp_offset                                 =  2,
 
     // Interpreter frames
-    interpreter_frame_oop_temp_offset                =  3, // for native calls only
+    interpreter_frame_result_handler_offset          =  3, // for native calls only
+    interpreter_frame_oop_temp_offset                =  2, // for native calls only
 
     interpreter_frame_sender_sp_offset               = -1,
     // outgoing sp before a call to an invoked method

src/hotspot/cpu/aarch64/macroAssembler_aarch64.cpp

@@ -1965,13 +1965,8 @@ void MacroAssembler::call_VM_leaf_base(address entry_point,
   if (entry_point == CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorenter) || 
       entry_point == CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorenter_obj)) {
     ldr(rscratch1, Address(rthread, JavaThread::preempt_alternate_return_offset()));
-    ldr(rscratch2, Address(rthread, in_bytes(JavaThread::preempt_alternate_return_offset()) + wordSize));
     cbz(rscratch1, not_preempted);
-    mov(r4, sp); // r4 is clobbered by VM calls, so free here
-    cmp(rscratch2, r4);
-    br(LT, not_preempted);
     str(zr, Address(rthread, JavaThread::preempt_alternate_return_offset()));
-    str(zr, Address(rthread, in_bytes(JavaThread::preempt_alternate_return_offset()) + wordSize));
     br(rscratch1);
   }
 

src/hotspot/cpu/aarch64/sharedRuntime_aarch64.cpp

@@ -1754,9 +1754,8 @@ nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
   // Change state to native (we save the return address in the thread, since it might not
   // be pushed on the stack when we do a stack traversal).
   // We use the same pc/oopMap repeatedly when we call out
-
-  Label native_return;
-  __ set_last_Java_frame(sp, noreg, native_return, rscratch1);
+  Label resume_pc;
+  __ set_last_Java_frame(sp, noreg, resume_pc, rscratch1);
 
   Label dtrace_method_entry, dtrace_method_entry_done;
   {
@@ -1863,11 +1862,6 @@ nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
 
   __ rt_call(native_func);
 
-  __ bind(native_return);
-
-  intptr_t return_pc = (intptr_t) __ pc();
-  oop_maps->add_gc_map(return_pc - start, map);
-
   // Verify or restore cpu control state after JNI call
   __ restore_cpu_control_state_after_jni(rscratch1, rscratch2);
 
@@ -1934,6 +1928,20 @@ nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
   __ stlrw(rscratch1, rscratch2);
   __ bind(after_transition);
 
+  // Check preemption for Object.wait()
+  if (method->is_object_wait0()) {
+    Label not_preempted;
+    __ ldr(rscratch1, Address(rthread, JavaThread::preempt_alternate_return_offset()));
+    __ cbz(rscratch1, not_preempted);
+    __ str(zr, Address(rthread, JavaThread::preempt_alternate_return_offset()));
+    __ br(rscratch1);
+    __ bind(not_preempted);
+  }
+  __ bind(resume_pc);
+
+  intptr_t the_pc = (intptr_t) __ pc();
+  oop_maps->add_gc_map(the_pc - start, map);
+
   Label reguard;
   Label reguard_done;
   __ ldrb(rscratch1, Address(rthread, JavaThread::stack_guard_state_offset()));

src/hotspot/cpu/aarch64/stackChunkFrameStream_aarch64.inline.hpp

@@ -116,6 +116,7 @@ inline int StackChunkFrameStream<frame_kind>::interpreter_frame_num_oops() const
   f.interpreted_frame_oop_map(&mask);
   return  mask.num_oops()
         + 1 // for the mirror oop
+        + (f.interpreter_frame_method()->is_native() ? 1 : 0) // temp oop slot
         + pointer_delta_as_int((intptr_t*)f.interpreter_frame_monitor_begin(),
               (intptr_t*)f.interpreter_frame_monitor_end())/BasicObjectLock::size();
 }

src/hotspot/cpu/aarch64/stubGenerator_aarch64.cpp

@@ -7158,18 +7158,16 @@ class StubGenerator: public StubCodeGenerator {
     __ mov(sp, rscratch2);
 
     Label preemption_cancelled;
-    // FIXME: Whose responsibility is it to clear this flag?
     __ ldrb(rscratch1, Address(rthread, JavaThread::preemption_cancelled_offset()));
     __ cbnz(rscratch1, preemption_cancelled);
 
-    //__ trace("Remove enterSpecial frame from the stack and return to Continuation.run()");
     // Remove enterSpecial frame from the stack and return to Continuation.run()
     SharedRuntime::continuation_enter_cleanup(_masm);
     __ leave();
     __ ret(lr);
 
     __ bind(preemption_cancelled);
-    //__ trace("preemption_cancelled");
+    __ strb(zr, Address(rthread, JavaThread::preemption_cancelled_offset()));
     __ lea(rfp, Address(sp, checked_cast<int32_t>(ContinuationEntry::size())));
     __ lea(rscratch1, ExternalAddress((address)&ContinuationEntry::_thaw_call_pc));
     __ ldr(rscratch1, Address(rscratch1));
@@ -7178,9 +7176,9 @@ class StubGenerator: public StubCodeGenerator {
     return start;
   }
 
-  address generate_cont_preempt_rerun_compiler_adapter() {
+  address generate_cont_resume_compiler_adapter() {
     if (!Continuations::enabled()) return nullptr;
-    StubCodeMark mark(this, "StubRoutines", "Continuation preempt safepoint blob adapter");
+    StubCodeMark mark(this, "StubRoutines", "Continuation resume compiler adapter");
     address start = __ pc();
 
     // The safepoint blob handler expects that r20, being a callee saved register, will be preserved
@@ -7196,25 +7194,25 @@ class StubGenerator: public StubCodeGenerator {
     return start;
   }
 
-  address generate_cont_preempt_monitorenter_redo() {
+  address generate_cont_resume_monitor_operation() {
     if (!Continuations::enabled()) return nullptr;
-    StubCodeMark mark(this, "StubRoutines","Continuation monitorenter redo stub");
+    StubCodeMark mark(this, "StubRoutines","Continuation resume monitor operation");
     address start = __ pc();
 
-    const Register mon_reg = c_rarg1;
-    __ ldr(mon_reg, __ post(sp, 2 * wordSize));
+    const Register waiter_reg = c_rarg1;
+    __ ldr(waiter_reg, __ post(sp, 2 * wordSize));
 
 #ifdef ASSERT
     { Label L;
-      __ cbnz(mon_reg, L);
+      __ cbnz(waiter_reg, L);
       __ stop("ObjectMonitor to use is null");
       __ bind(L);
     }
 #endif // ASSERT
 
     __ set_last_Java_frame(sp, rfp, lr, rscratch1);
     __ mov(c_rarg0, rthread);
-    __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::redo_monitorenter));
+    __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::resume_monitor_operation));
     __ reset_last_Java_frame(true);
 
     Label failAcquire;
@@ -8510,8 +8508,8 @@ class StubGenerator: public StubCodeGenerator {
     StubRoutines::_cont_returnBarrier = generate_cont_returnBarrier();
     StubRoutines::_cont_returnBarrierExc = generate_cont_returnBarrier_exception();
     StubRoutines::_cont_preempt_stub = generate_cont_preempt_stub();
-    StubRoutines::_cont_preempt_monitorenter_redo = generate_cont_preempt_monitorenter_redo();
-    StubRoutines::_cont_preempt_rerun_compiler_adapter = generate_cont_preempt_rerun_compiler_adapter();
+    StubRoutines::_cont_resume_monitor_operation = generate_cont_resume_monitor_operation();
+    StubRoutines::_cont_resume_compiler_adapter = generate_cont_resume_compiler_adapter();
 
     JFR_ONLY(generate_jfr_stubs();)
   }

src/hotspot/cpu/aarch64/templateInterpreterGenerator_aarch64.cpp

@@ -612,7 +612,7 @@ address TemplateInterpreterGenerator::generate_safept_entry_for(
   return entry;
 }
 
-address TemplateInterpreterGenerator::generate_cont_preempt_rerun_interpreter_adapter() {
+address TemplateInterpreterGenerator::generate_cont_resume_interpreter_adapter() {
   if (!Continuations::enabled()) return nullptr;
   address start = __ pc();
 
@@ -1354,6 +1354,8 @@ address TemplateInterpreterGenerator::generate_native_entry(bool synchronized) {
   // result handler is in r0
   // set result handler
   __ mov(result_handler, r0);
+  __ str(r0, Address(rfp, frame::interpreter_frame_result_handler_offset * wordSize));
+
   // pass mirror handle if static call
   {
     Label L;
@@ -1392,8 +1394,8 @@ address TemplateInterpreterGenerator::generate_native_entry(bool synchronized) {
   // Set the last Java PC in the frame anchor to be the return address from
   // the call to the native method: this will allow the debugger to
   // generate an accurate stack trace.
-  Label native_return;
-  __ set_last_Java_frame(esp, rfp, native_return, rscratch1);
+  Label resume_pc;
+  __ set_last_Java_frame(esp, rfp, resume_pc, rscratch1);
 
   // change thread state
 #ifdef ASSERT
@@ -1414,7 +1416,7 @@ address TemplateInterpreterGenerator::generate_native_entry(bool synchronized) {
 
   // Call the native method.
   __ blr(r10);
-  __ bind(native_return);
+
   __ get_method(rmethod);
   // result potentially in r0 or v0
 
@@ -1479,6 +1481,18 @@ address TemplateInterpreterGenerator::generate_native_entry(bool synchronized) {
   __ lea(rscratch2, Address(rthread, JavaThread::thread_state_offset()));
   __ stlrw(rscratch1, rscratch2);
 
+  // Check preemption for Object.wait()
+  Label not_preempted;
+  __ ldr(rscratch1, Address(rthread, JavaThread::preempt_alternate_return_offset()));
+  __ cbz(rscratch1, not_preempted);
+  __ str(zr, Address(rthread, JavaThread::preempt_alternate_return_offset()));
+  __ br(rscratch1);
+  __ bind(resume_pc);
+  // On resume we need to set up stack as expected
+  __ push(dtos);
+  __ push(ltos);
+  __ bind(not_preempted);
+
   // reset_last_Java_frame
   __ reset_last_Java_frame(true);
 
@@ -1497,6 +1511,7 @@ address TemplateInterpreterGenerator::generate_native_entry(bool synchronized) {
   {
     Label no_oop;
     __ adr(t, ExternalAddress(AbstractInterpreter::result_handler(T_OBJECT)));
+    __ ldr(result_handler, Address(rfp, frame::interpreter_frame_result_handler_offset*wordSize));
     __ cmp(t, result_handler);
     __ br(Assembler::NE, no_oop);
     // Unbox oop result, e.g. JNIHandles::resolve result.

src/hotspot/cpu/arm/continuationFreezeThaw_arm.inline.hpp

@@ -91,12 +91,12 @@ inline void ThawBase::patch_pd(frame& f, intptr_t* caller_sp) {
   Unimplemented();
 }
 
-inline intptr_t* ThawBase::push_preempt_rerun_adapter(frame top, bool is_interpreted_frame) {
+inline intptr_t* ThawBase::push_resume_adapter(frame& top, bool is_interpreted_frame) {
   Unimplemented();
   return nullptr;
 }
 
-inline intptr_t* ThawBase::push_preempt_monitorenter_redo(stackChunkOop chunk) {
+inline intptr_t* ThawBase::push_resume_monitor_operation(stackChunkOop chunk) {
   Unimplemented();
   return nullptr;
 }

src/hotspot/cpu/arm/templateInterpreterGenerator_arm.cpp

@@ -458,7 +458,7 @@ address TemplateInterpreterGenerator::generate_safept_entry_for(TosState state,
   return entry;
 }
 
-address TemplateInterpreterGenerator::generate_cont_preempt_rerun_interpreter_adapter() {
+address TemplateInterpreterGenerator::generate_cont_resume_interpreter_adapter() {
   return nullptr;
 }