InterpreterFrame *InterpreterFrame::build(const methodOop method, TRAPS) {
JavaThread *thread = (JavaThread *) THREAD;
ZeroStack *stack = thread->zero_stack();
// Calculate the size of the frame we'll build, including
// any adjustments to the caller's frame that we'll make.
int extra_locals = 0;
int monitor_words = 0;
int stack_words = 0;
if (!method->is_native()) {
extra_locals = method->max_locals() - method->size_of_parameters();
stack_words = method->max_stack();
}
if (method->is_synchronized()) {
monitor_words = frame::interpreter_frame_monitor_size();
}
stack->overflow_check(
extra_locals + header_words + monitor_words + stack_words, CHECK_NULL);
// Adjust the caller's stack frame to accomodate any additional
// local variables we have contiguously with our parameters.
for (int i = 0; i < extra_locals; i++)
stack->push(0);
intptr_t *locals;
if (method->is_native())
locals = stack->sp() + (method->size_of_parameters() - 1);
else
locals = stack->sp() + (method->max_locals() - 1);
stack->push(0); // next_frame, filled in later
intptr_t *fp = stack->sp();
assert(fp - stack->sp() == next_frame_off, "should be");
stack->push(INTERPRETER_FRAME);
assert(fp - stack->sp() == frame_type_off, "should be");
interpreterState istate =
(interpreterState) stack->alloc(sizeof(BytecodeInterpreter));
assert(fp - stack->sp() == istate_off, "should be");
istate->set_locals(locals);
istate->set_method(method);
istate->set_self_link(istate);
istate->set_prev_link(NULL);
istate->set_thread(thread);
istate->set_bcp(method->is_native() ? NULL : method->code_base());
istate->set_constants(method->constants()->cache());
istate->set_msg(BytecodeInterpreter::method_entry);
istate->set_oop_temp(NULL);
istate->set_mdx(NULL);
istate->set_callee(NULL);
istate->set_monitor_base((BasicObjectLock *) stack->sp());
if (method->is_synchronized()) {
BasicObjectLock *monitor =
(BasicObjectLock *) stack->alloc(monitor_words * wordSize);
oop object;
if (method->is_static())
object = method->constants()->pool_holder()->java_mirror();
else
object = (oop) locals[0];
monitor->set_obj(object);
}
istate->set_stack_base(stack->sp());
istate->set_stack(stack->sp() - 1);
if (stack_words)
stack->alloc(stack_words * wordSize);
istate->set_stack_limit(stack->sp() - 1);
return (InterpreterFrame *) fp;
}
void vframeArrayElement::fill_in(compiledVFrame* vf) {
// Copy the information from the compiled vframe to the
// interpreter frame we will be creating to replace vf
_method = vf->method();
_bci = vf->raw_bci();
_reexecute = vf->should_reexecute();
int index;
// Get the monitors off-stack
GrowableArray<MonitorInfo*>* list = vf->monitors();
if (list->is_empty()) {
_monitors = NULL;
} else {
// Allocate monitor chunk
_monitors = new MonitorChunk(list->length());
vf->thread()->add_monitor_chunk(_monitors);
// Migrate the BasicLocks from the stack to the monitor chunk
for (index = 0; index < list->length(); index++) {
MonitorInfo* monitor = list->at(index);
assert(!monitor->owner_is_scalar_replaced(), "object should be reallocated already");
assert(monitor->owner() == NULL || (!monitor->owner()->is_unlocked() && !monitor->owner()->has_bias_pattern()), "object must be null or locked, and unbiased");
BasicObjectLock* dest = _monitors->at(index);
dest->set_obj(monitor->owner());
monitor->lock()->move_to(monitor->owner(), dest->lock());
}
}
// Convert the vframe locals and expressions to off stack
// values. Because we will not gc all oops can be converted to
// intptr_t (i.e. a stack slot) and we are fine. This is
// good since we are inside a HandleMark and the oops in our
// collection would go away between packing them here and
// unpacking them in unpack_on_stack.
// First the locals go off-stack
// FIXME this seems silly it creates a StackValueCollection
// in order to get the size to then copy them and
// convert the types to intptr_t size slots. Seems like it
// could do it in place... Still uses less memory than the
// old way though
StackValueCollection *locs = vf->locals();
_locals = new StackValueCollection(locs->size());
for(index = 0; index < locs->size(); index++) {
StackValue* value = locs->at(index);
switch(value->type()) {
case T_OBJECT:
assert(!value->obj_is_scalar_replaced(), "object should be reallocated already");
// preserve object type
_locals->add( new StackValue((intptr_t) (value->get_obj()()), T_OBJECT ));
break;
case T_CONFLICT:
// A dead local. Will be initialized to null/zero.
_locals->add( new StackValue());
break;
case T_INT:
_locals->add( new StackValue(value->get_int()));
break;
default:
ShouldNotReachHere();
}
}
// Now the expressions off-stack
// Same silliness as above
StackValueCollection *exprs = vf->expressions();
_expressions = new StackValueCollection(exprs->size());
for(index = 0; index < exprs->size(); index++) {
StackValue* value = exprs->at(index);
switch(value->type()) {
case T_OBJECT:
assert(!value->obj_is_scalar_replaced(), "object should be reallocated already");
// preserve object type
_expressions->add( new StackValue((intptr_t) (value->get_obj()()), T_OBJECT ));
break;
case T_CONFLICT:
// A dead stack element. Will be initialized to null/zero.
// This can occur when the compiler emits a state in which stack
// elements are known to be dead (because of an imminent exception).
_expressions->add( new StackValue());
break;
case T_INT:
_expressions->add( new StackValue(value->get_int()));
break;
default:
ShouldNotReachHere();
}
}
}
int CppInterpreter::native_entry(methodOop method, intptr_t UNUSED, TRAPS) {
// Make sure method is native and not abstract
assert(method->is_native() && !method->is_abstract(), "should be");
JavaThread *thread = (JavaThread *) THREAD;
ZeroStack *stack = thread->zero_stack();
// Allocate and initialize our frame
InterpreterFrame *frame = InterpreterFrame::build(method, CHECK_0);
thread->push_zero_frame(frame);
interpreterState istate = frame->interpreter_state();
intptr_t *locals = istate->locals();
// Update the invocation counter
if ((UseCompiler || CountCompiledCalls) && !method->is_synchronized()) {
InvocationCounter *counter = method->invocation_counter();
counter->increment();
if (counter->reached_InvocationLimit()) {
CALL_VM_NOCHECK(
InterpreterRuntime::frequency_counter_overflow(thread, NULL));
if (HAS_PENDING_EXCEPTION)
goto unwind_and_return;
}
}
// Lock if necessary
BasicObjectLock *monitor;
monitor = NULL;
if (method->is_synchronized()) {
monitor = (BasicObjectLock*) istate->stack_base();
oop lockee = monitor->obj();
markOop disp = lockee->mark()->set_unlocked();
monitor->lock()->set_displaced_header(disp);
if (Atomic::cmpxchg_ptr(monitor, lockee->mark_addr(), disp) != disp) {
if (thread->is_lock_owned((address) disp->clear_lock_bits())) {
monitor->lock()->set_displaced_header(NULL);
}
else {
CALL_VM_NOCHECK(InterpreterRuntime::monitorenter(thread, monitor));
if (HAS_PENDING_EXCEPTION)
goto unwind_and_return;
}
}
}
// Get the signature handler
InterpreterRuntime::SignatureHandler *handler; {
address handlerAddr = method->signature_handler();
if (handlerAddr == NULL) {
CALL_VM_NOCHECK(InterpreterRuntime::prepare_native_call(thread, method));
if (HAS_PENDING_EXCEPTION)
goto unlock_unwind_and_return;
handlerAddr = method->signature_handler();
assert(handlerAddr != NULL, "eh?");
}
if (handlerAddr == (address) InterpreterRuntime::slow_signature_handler) {
CALL_VM_NOCHECK(handlerAddr =
InterpreterRuntime::slow_signature_handler(thread, method, NULL,NULL));
if (HAS_PENDING_EXCEPTION)
goto unlock_unwind_and_return;
}
handler = \
InterpreterRuntime::SignatureHandler::from_handlerAddr(handlerAddr);
}
// Get the native function entry point
address function;
function = method->native_function();
assert(function != NULL, "should be set if signature handler is");
// Build the argument list
stack->overflow_check(handler->argument_count() * 2, THREAD);
if (HAS_PENDING_EXCEPTION)
goto unlock_unwind_and_return;
void **arguments;
void *mirror; {
arguments =
(void **) stack->alloc(handler->argument_count() * sizeof(void **));
void **dst = arguments;
void *env = thread->jni_environment();
*(dst++) = &env;
if (method->is_static()) {
istate->set_oop_temp(
method->constants()->pool_holder()->java_mirror());
mirror = istate->oop_temp_addr();
*(dst++) = &mirror;
}
intptr_t *src = locals;
for (int i = dst - arguments; i < handler->argument_count(); i++) {
ffi_type *type = handler->argument_type(i);
if (type == &ffi_type_pointer) {
if (*src) {
stack->push((intptr_t) src);
*(dst++) = stack->sp();
}
else {
*(dst++) = src;
}
src--;
}
else if (type->size == 4) {
*(dst++) = src--;
}
else if (type->size == 8) {
src--;
*(dst++) = src--;
}
else {
ShouldNotReachHere();
}
}
}
// Set up the Java frame anchor
thread->set_last_Java_frame();
// Change the thread state to _thread_in_native
ThreadStateTransition::transition_from_java(thread, _thread_in_native);
// Make the call
intptr_t result[4 - LogBytesPerWord];
ffi_call(handler->cif(), (void (*)()) function, result, arguments);
// Change the thread state back to _thread_in_Java.
// ThreadStateTransition::transition_from_native() cannot be used
// here because it does not check for asynchronous exceptions.
// We have to manage the transition ourself.
thread->set_thread_state(_thread_in_native_trans);
// Make sure new state is visible in the GC thread
if (os::is_MP()) {
if (UseMembar) {
OrderAccess::fence();
}
else {
InterfaceSupport::serialize_memory(thread);
}
}
// Handle safepoint operations, pending suspend requests,
// and pending asynchronous exceptions.
if (SafepointSynchronize::do_call_back() ||
thread->has_special_condition_for_native_trans()) {
JavaThread::check_special_condition_for_native_trans(thread);
CHECK_UNHANDLED_OOPS_ONLY(thread->clear_unhandled_oops());
}
// Finally we can change the thread state to _thread_in_Java.
thread->set_thread_state(_thread_in_Java);
fixup_after_potential_safepoint();
// Clear the frame anchor
thread->reset_last_Java_frame();
// If the result was an oop then unbox it and store it in
// oop_temp where the garbage collector can see it before
// we release the handle it might be protected by.
if (handler->result_type() == &ffi_type_pointer) {
if (result[0])
istate->set_oop_temp(*(oop *) result[0]);
else
istate->set_oop_temp(NULL);
}
// Reset handle block
thread->active_handles()->clear();
unlock_unwind_and_return:
// Unlock if necessary
if (monitor) {
BasicLock *lock = monitor->lock();
markOop header = lock->displaced_header();
oop rcvr = monitor->obj();
monitor->set_obj(NULL);
if (header != NULL) {
if (Atomic::cmpxchg_ptr(header, rcvr->mark_addr(), lock) != lock) {
monitor->set_obj(rcvr); {
HandleMark hm(thread);
CALL_VM_NOCHECK(InterpreterRuntime::monitorexit(thread, monitor));
}
}
}
}
unwind_and_return:
// Unwind the current activation
thread->pop_zero_frame();
// Pop our parameters
stack->set_sp(stack->sp() + method->size_of_parameters());
// Push our result
if (!HAS_PENDING_EXCEPTION) {
BasicType type = result_type_of(method);
stack->set_sp(stack->sp() - type2size[type]);
switch (type) {
case T_VOID:
break;
case T_BOOLEAN:
#ifndef VM_LITTLE_ENDIAN
result[0] <<= (BitsPerWord - BitsPerByte);
#endif
SET_LOCALS_INT(*(jboolean *) result != 0, 0);
break;
case T_CHAR:
#ifndef VM_LITTLE_ENDIAN
result[0] <<= (BitsPerWord - BitsPerShort);
#endif
SET_LOCALS_INT(*(jchar *) result, 0);
break;
case T_BYTE:
#ifndef VM_LITTLE_ENDIAN
result[0] <<= (BitsPerWord - BitsPerByte);
#endif
SET_LOCALS_INT(*(jbyte *) result, 0);
break;
case T_SHORT:
#ifndef VM_LITTLE_ENDIAN
result[0] <<= (BitsPerWord - BitsPerShort);
#endif
SET_LOCALS_INT(*(jshort *) result, 0);
break;
case T_INT:
#ifndef VM_LITTLE_ENDIAN
result[0] <<= (BitsPerWord - BitsPerInt);
#endif
SET_LOCALS_INT(*(jint *) result, 0);
break;
case T_LONG:
SET_LOCALS_LONG(*(jlong *) result, 0);
break;
case T_FLOAT:
SET_LOCALS_FLOAT(*(jfloat *) result, 0);
break;
case T_DOUBLE:
SET_LOCALS_DOUBLE(*(jdouble *) result, 0);
break;
case T_OBJECT:
case T_ARRAY:
SET_LOCALS_OBJECT(istate->oop_temp(), 0);
break;
default:
ShouldNotReachHere();
}
}
// No deoptimized frames on the stack
return 0;
}
void vframeArrayElement::unpack_on_stack(int caller_actual_parameters,
int callee_parameters,
int callee_locals,
frame* caller,
bool is_top_frame,
bool is_bottom_frame,
int exec_mode) {
JavaThread* thread = (JavaThread*) Thread::current();
// Look at bci and decide on bcp and continuation pc
address bcp;
// C++ interpreter doesn't need a pc since it will figure out what to do when it
// begins execution
address pc;
bool use_next_mdp = false; // true if we should use the mdp associated with the next bci
// rather than the one associated with bcp
if (raw_bci() == SynchronizationEntryBCI) {
// We are deoptimizing while hanging in prologue code for synchronized method
bcp = method()->bcp_from(0); // first byte code
pc = Interpreter::deopt_entry(vtos, 0); // step = 0 since we don't skip current bytecode
} else if (should_reexecute()) { //reexecute this bytecode
assert(is_top_frame, "reexecute allowed only for the top frame");
bcp = method()->bcp_from(bci());
pc = Interpreter::deopt_reexecute_entry(method(), bcp);
} else {
bcp = method()->bcp_from(bci());
pc = Interpreter::deopt_continue_after_entry(method(), bcp, callee_parameters, is_top_frame);
use_next_mdp = true;
}
assert(Bytecodes::is_defined(*bcp), "must be a valid bytecode");
// Monitorenter and pending exceptions:
//
// For Compiler2, there should be no pending exception when deoptimizing at monitorenter
// because there is no safepoint at the null pointer check (it is either handled explicitly
// or prior to the monitorenter) and asynchronous exceptions are not made "pending" by the
// runtime interface for the slow case (see JRT_ENTRY_FOR_MONITORENTER). If an asynchronous
// exception was processed, the bytecode pointer would have to be extended one bytecode beyond
// the monitorenter to place it in the proper exception range.
//
// For Compiler1, deoptimization can occur while throwing a NullPointerException at monitorenter,
// in which case bcp should point to the monitorenter since it is within the exception's range.
assert(*bcp != Bytecodes::_monitorenter || is_top_frame, "a _monitorenter must be a top frame");
assert(thread->deopt_nmethod() != NULL, "nmethod should be known");
guarantee(!(thread->deopt_nmethod()->is_compiled_by_c2() &&
*bcp == Bytecodes::_monitorenter &&
exec_mode == Deoptimization::Unpack_exception),
"shouldn't get exception during monitorenter");
int popframe_preserved_args_size_in_bytes = 0;
int popframe_preserved_args_size_in_words = 0;
if (is_top_frame) {
JvmtiThreadState *state = thread->jvmti_thread_state();
if (JvmtiExport::can_pop_frame() &&
(thread->has_pending_popframe() || thread->popframe_forcing_deopt_reexecution())) {
if (thread->has_pending_popframe()) {
// Pop top frame after deoptimization
#ifndef CC_INTERP
pc = Interpreter::remove_activation_preserving_args_entry();
#else
// Do an uncommon trap type entry. c++ interpreter will know
// to pop frame and preserve the args
pc = Interpreter::deopt_entry(vtos, 0);
use_next_mdp = false;
#endif
} else {
// Reexecute invoke in top frame
pc = Interpreter::deopt_entry(vtos, 0);
use_next_mdp = false;
popframe_preserved_args_size_in_bytes = in_bytes(thread->popframe_preserved_args_size());
// Note: the PopFrame-related extension of the expression stack size is done in
// Deoptimization::fetch_unroll_info_helper
popframe_preserved_args_size_in_words = in_words(thread->popframe_preserved_args_size_in_words());
}
} else if (JvmtiExport::can_force_early_return() && state != NULL && state->is_earlyret_pending()) {
// Force early return from top frame after deoptimization
#ifndef CC_INTERP
pc = Interpreter::remove_activation_early_entry(state->earlyret_tos());
#else
// TBD: Need to implement ForceEarlyReturn for CC_INTERP (ia64)
#endif
} else {
// Possibly override the previous pc computation of the top (youngest) frame
switch (exec_mode) {
case Deoptimization::Unpack_deopt:
// use what we've got
break;
case Deoptimization::Unpack_exception:
// exception is pending
pc = SharedRuntime::raw_exception_handler_for_return_address(thread, pc);
// [phh] We're going to end up in some handler or other, so it doesn't
// matter what mdp we point to. See exception_handler_for_exception()
// in interpreterRuntime.cpp.
break;
case Deoptimization::Unpack_uncommon_trap:
case Deoptimization::Unpack_reexecute:
// redo last byte code
pc = Interpreter::deopt_entry(vtos, 0);
use_next_mdp = false;
break;
default:
ShouldNotReachHere();
}
}
}
// Setup the interpreter frame
assert(method() != NULL, "method must exist");
int temps = expressions()->size();
int locks = monitors() == NULL ? 0 : monitors()->number_of_monitors();
Interpreter::layout_activation(method(),
temps + callee_parameters,
popframe_preserved_args_size_in_words,
locks,
caller_actual_parameters,
callee_parameters,
callee_locals,
caller,
iframe(),
is_top_frame,
is_bottom_frame);
// Update the pc in the frame object and overwrite the temporary pc
// we placed in the skeletal frame now that we finally know the
// exact interpreter address we should use.
_frame.patch_pc(thread, pc);
assert (!method()->is_synchronized() || locks > 0, "synchronized methods must have monitors");
BasicObjectLock* top = iframe()->interpreter_frame_monitor_begin();
for (int index = 0; index < locks; index++) {
top = iframe()->previous_monitor_in_interpreter_frame(top);
BasicObjectLock* src = _monitors->at(index);
top->set_obj(src->obj());
src->lock()->move_to(src->obj(), top->lock());
}
if (ProfileInterpreter) {
iframe()->interpreter_frame_set_mdx(0); // clear out the mdp.
}
iframe()->interpreter_frame_set_bcx((intptr_t)bcp); // cannot use bcp because frame is not initialized yet
if (ProfileInterpreter) {
methodDataOop mdo = method()->method_data();
if (mdo != NULL) {
int bci = iframe()->interpreter_frame_bci();
if (use_next_mdp) ++bci;
address mdp = mdo->bci_to_dp(bci);
iframe()->interpreter_frame_set_mdp(mdp);
}
}
// Unpack expression stack
// If this is an intermediate frame (i.e. not top frame) then this
// only unpacks the part of the expression stack not used by callee
// as parameters. The callee parameters are unpacked as part of the
// callee locals.
int i;
for(i = 0; i < expressions()->size(); i++) {
StackValue *value = expressions()->at(i);
intptr_t* addr = iframe()->interpreter_frame_expression_stack_at(i);
switch(value->type()) {
case T_INT:
*addr = value->get_int();
break;
case T_OBJECT:
*addr = value->get_int(T_OBJECT);
break;
case T_CONFLICT:
// A dead stack slot. Initialize to null in case it is an oop.
*addr = NULL_WORD;
break;
default:
ShouldNotReachHere();
}
}
// Unpack the locals
for(i = 0; i < locals()->size(); i++) {
StackValue *value = locals()->at(i);
intptr_t* addr = iframe()->interpreter_frame_local_at(i);
switch(value->type()) {
case T_INT:
*addr = value->get_int();
break;
case T_OBJECT:
*addr = value->get_int(T_OBJECT);
break;
case T_CONFLICT:
// A dead location. If it is an oop then we need a NULL to prevent GC from following it
*addr = NULL_WORD;
break;
default:
ShouldNotReachHere();
}
}
if (is_top_frame && JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
// An interpreted frame was popped but it returns to a deoptimized
// frame. The incoming arguments to the interpreted activation
// were preserved in thread-local storage by the
// remove_activation_preserving_args_entry in the interpreter; now
// we put them back into the just-unpacked interpreter frame.
// Note that this assumes that the locals arena grows toward lower
// addresses.
if (popframe_preserved_args_size_in_words != 0) {
void* saved_args = thread->popframe_preserved_args();
assert(saved_args != NULL, "must have been saved by interpreter");
#ifdef ASSERT
assert(popframe_preserved_args_size_in_words <=
iframe()->interpreter_frame_expression_stack_size()*Interpreter::stackElementWords,
"expression stack size should have been extended");
#endif // ASSERT
int top_element = iframe()->interpreter_frame_expression_stack_size()-1;
intptr_t* base;
if (frame::interpreter_frame_expression_stack_direction() < 0) {
base = iframe()->interpreter_frame_expression_stack_at(top_element);
} else {
base = iframe()->interpreter_frame_expression_stack();
}
Copy::conjoint_jbytes(saved_args,
base,
popframe_preserved_args_size_in_bytes);
thread->popframe_free_preserved_args();
}
}
#ifndef PRODUCT
if (TraceDeoptimization && Verbose) {
ttyLocker ttyl;
tty->print_cr("[%d Interpreted Frame]", ++unpack_counter);
iframe()->print_on(tty);
RegisterMap map(thread);
vframe* f = vframe::new_vframe(iframe(), &map, thread);
f->print();
tty->print_cr("locals size %d", locals()->size());
tty->print_cr("expression size %d", expressions()->size());
method()->print_value();
tty->cr();
// method()->print_codes();
} else if (TraceDeoptimization) {
tty->print(" ");
method()->print_value();
Bytecodes::Code code = Bytecodes::java_code_at(method(), bcp);
int bci = method()->bci_from(bcp);
tty->print(" - %s", Bytecodes::name(code));
tty->print(" @ bci %d ", bci);
tty->print_cr("sp = " PTR_FORMAT, iframe()->sp());
}
#endif // PRODUCT
// The expression stack and locals are in the resource area don't leave
// a dangling pointer in the vframeArray we leave around for debug
// purposes
_locals = _expressions = NULL;
}