address generate_call_stub(address& return_address)
{
assert (!TaggedStackInterpreter, "not supported");
StubCodeMark mark(this, "StubRoutines", "call_stub");
address start = __ enter();
const Register call_wrapper = r3;
const Register result = r4;
const Register result_type = r5;
const Register method = r6;
const Register entry_point = r7;
const Register parameters = r8;
const Register parameter_words = r9;
const Register thread = r10;
#ifdef ASSERT
// Make sure we have no pending exceptions
{
StackFrame frame;
Label label;
__ load (r0, Address(thread, Thread::pending_exception_offset()));
__ compare (r0, 0);
__ beq (label);
__ prolog (frame);
__ should_not_reach_here (__FILE__, __LINE__);
__ epilog (frame);
__ blr ();
__ bind (label);
}
#endif // ASSERT
// Calculate the frame size
StackFrame frame;
for (int i = 0; i < StackFrame::max_crfs; i++)
frame.get_cr_field();
for (int i = 0; i < StackFrame::max_gprs; i++)
frame.get_register();
StubRoutines::set_call_stub_base_size(frame.unaligned_size() + 3*wordSize);
// the 3 extra words are for call_wrapper, result and result_type
const Register parameter_bytes = parameter_words;
__ shift_left (parameter_bytes, parameter_words, LogBytesPerWord);
const Register frame_size = r11;
const Register padding = r12;
__ addi (frame_size, parameter_bytes, StubRoutines::call_stub_base_size());
__ calc_padding_for_alignment (padding, frame_size, StackAlignmentInBytes);
__ add (frame_size, frame_size, padding);
// Save the link register and create the new frame
__ mflr (r0);
__ store (r0, Address(r1, StackFrame::lr_save_offset * wordSize));
__ neg (r0, frame_size);
__ store_update_indexed (r1, r1, r0);
#ifdef PPC64
__ mfcr (r0);
__ store (r0, Address(r1, StackFrame::cr_save_offset * wordSize));
#endif // PPC64
// Calculate the address of the interpreter's local variables
const Register locals = frame_size;
__ addi (locals, r1, frame.start_of_locals() - wordSize);
__ add (locals, locals, padding);
__ add (locals, locals, parameter_bytes);
// Store the call wrapper address and the result stuff
const int initial_offset = 1;
int offset = initial_offset;
__ store (call_wrapper, Address(locals, offset++ * wordSize));
__ store (result, Address(locals, offset++ * wordSize));
__ store (result_type, Address(locals, offset++ * wordSize));
// Store the registers
#ifdef PPC32
__ mfcr (r0);
__ store (r0, Address(locals, offset++ * wordSize));
#endif // PPC32
for (int i = 14; i < 32; i++) {
__ store (as_Register(i), Address(locals, offset++ * wordSize));
}
const int final_offset = offset;
// Store the location of call_wrapper
frame::set_call_wrapper_offset((final_offset - initial_offset) * wordSize);
#ifdef ASSERT
// Check that we wrote all the way to the end of the frame.
// The frame may have been resized when we return from the
// interpreter, so the start of the frame may have moved
// but the end will be where we left it and we rely on this
// to find our stuff.
{
StackFrame frame;
Label label;
__ load (r3, Address(r1, 0));
__ subi (r3, r3, final_offset * wordSize);
__ compare (r3, locals);
__ beq (label);
__ prolog (frame);
__ should_not_reach_here (__FILE__, __LINE__);
__ epilog (frame);
__ blr ();
__ bind (label);
}
#endif // ASSERT
// Pass parameters if any
{
Label loop, done;
__ compare (parameter_bytes, 0);
__ ble (done);
const Register src = parameters;
const Register dst = padding;
__ mr (dst, locals);
__ shift_right (r0, parameter_bytes, LogBytesPerWord);
__ mtctr (r0);
__ bind (loop);
__ load (r0, Address(src, 0));
__ store (r0, Address(dst, 0));
__ addi (src, src, wordSize);
__ subi (dst, dst, wordSize);
__ bdnz (loop);
__ bind (done);
}
// Make the call
__ mr (Rmethod, method);
__ mr (Rlocals, locals);
__ mr (Rthread, thread);
__ mtctr (entry_point);
__ bctrl();
// This is used to identify call_stub stack frames
return_address = __ pc();
// Figure out where our stuff is stored
__ load (locals, Address(r1, 0));
__ subi (locals, locals, final_offset * wordSize);
#ifdef ASSERT
// Rlocals should contain the address we just calculated.
{
StackFrame frame;
Label label;
__ compare (Rlocals, locals);
__ beq (label);
__ prolog (frame);
__ should_not_reach_here (__FILE__, __LINE__);
__ epilog (frame);
__ blr ();
__ bind (label);
}
#endif // ASSERT
// Is an exception being thrown?
Label exit;
__ load (r0, Address(Rthread, Thread::pending_exception_offset()));
__ compare (r0, 0);
__ bne (exit);
// Store result depending on type
const Register result_addr = r6;
Label is_int, is_long, is_object;
offset = initial_offset + 1; // skip call_wrapper
__ load (result_addr, Address(locals, offset++ * wordSize));
__ load (result_type, Address(locals, offset++ * wordSize));
__ compare (result_type, T_INT);
__ beq (is_int);
__ compare (result_type, T_LONG);
__ beq (is_long);
__ compare (result_type, T_OBJECT);
__ beq (is_object);
__ should_not_reach_here (__FILE__, __LINE__);
__ bind (is_int);
__ stw (r3, Address(result_addr, 0));
__ b (exit);
__ bind (is_long);
#ifdef PPC32
__ store (r4, Address(result_addr, wordSize));
#endif
__ store (r3, Address(result_addr, 0));
__ b (exit);
__ bind (is_object);
__ store (r3, Address(result_addr, 0));
//__ b (exit);
// Restore the registers
__ bind (exit);
#ifdef PPC32
__ load (r0, Address(locals, offset++ * wordSize));
__ mtcr (r0);
#endif // PPC32
for (int i = 14; i < 32; i++) {
__ load (as_Register(i), Address(locals, offset++ * wordSize));
}
#ifdef PPC64
__ load (r0, Address(r1, StackFrame::cr_save_offset * wordSize));
__ mtcr (r0);
#endif // PPC64
assert (offset == final_offset, "save and restore must match");
// Unwind and return
__ load (r1, Address(r1, StackFrame::back_chain_offset * wordSize));
__ load (r0, Address(r1, StackFrame::lr_save_offset * wordSize));
__ mtlr (r0);
__ blr ();
return start;
}
address InterpreterGenerator::generate_normal_entry(bool synchronized)
{
assert_different_registers(Rmethod, Rlocals, Rthread, Rstate, Rmonitor);
Label re_dispatch;
Label call_interpreter;
Label call_method;
Label call_non_interpreted_method;
Label return_with_exception;
Label return_from_method;
Label resume_interpreter;
Label return_to_initial_caller;
Label more_monitors;
Label throwing_exception;
// We use the same code for synchronized and not
if (normal_entry)
return normal_entry;
address start = __ pc();
// There are two ways in which we can arrive at this entry.
// There is the special case where a normal interpreted method
// calls another normal interpreted method, and there is the
// general case of when we enter from somewhere else: from
// call_stub, from C1 or C2, or from a fast accessor which
// deferred. In the special case we're already in frame manager
// code: we arrive at re_dispatch with Rstate containing the
// previous interpreter state. In the general case we arrive
// at start with no previous interpreter state so we set Rstate
// to NULL to indicate this.
__ bind (fast_accessor_slow_entry_path);
__ load (Rstate, 0);
__ bind (re_dispatch);
// Adjust the caller's stack frame to accomodate any additional
// local variables we have contiguously with our parameters.
generate_adjust_callers_stack();
// Allocate and initialize our stack frame.
generate_compute_interpreter_state(false);
// Call the interpreter ==============================================
__ bind (call_interpreter);
// We can setup the frame anchor with everything we want at
// this point as we are thread_in_Java and no safepoints can
// occur until we go to vm mode. We do have to clear flags
// on return from vm but that is it
__ set_last_Java_frame ();
// Call interpreter
address interpreter = JvmtiExport::can_post_interpreter_events() ?
CAST_FROM_FN_PTR(address, BytecodeInterpreter::runWithChecks) :
CAST_FROM_FN_PTR(address, BytecodeInterpreter::run);
__ mr (r3, Rstate);
__ call (interpreter);
__ fixup_after_potential_safepoint ();
// Clear the frame anchor
__ reset_last_Java_frame ();
// Examine the message from the interpreter to decide what to do
__ lwz (r4, STATE(_msg));
__ compare (r4, BytecodeInterpreter::call_method);
__ beq (call_method);
__ compare (r4, BytecodeInterpreter::return_from_method);
__ beq (return_from_method);
__ compare (r4, BytecodeInterpreter::more_monitors);
__ beq (more_monitors);
__ compare (r4, BytecodeInterpreter::throwing_exception);
__ beq (throwing_exception);
__ load (r3, (intptr_t) "error: bad message from interpreter: %d\n");
__ call (CAST_FROM_FN_PTR(address, printf));
__ should_not_reach_here (__FILE__, __LINE__);
// Handle a call_method message ======================================
__ bind (call_method);
__ load (Rmethod, STATE(_result._to_call._callee));
__ verify_oop(Rmethod);
__ load (Rlocals, STATE(_stack));
__ lhz (r0, Address(Rmethod, methodOopDesc::size_of_parameters_offset()));
__ shift_left (r0, r0, LogBytesPerWord);
__ add (Rlocals, Rlocals, r0);
__ load (r0, STATE(_result._to_call._callee_entry_point));
__ load (r3, (intptr_t) start);
__ compare (r0, r3);
__ bne (call_non_interpreted_method);
// Interpreted methods are intercepted and re-dispatched -----------
__ load (r0, CAST_FROM_FN_PTR(intptr_t, RecursiveInterpreterActivation));
__ mtlr (r0);
__ b (re_dispatch);
// Non-interpreted methods are dispatched normally -----------------
__ bind (call_non_interpreted_method);
__ mtctr (r0);
__ bctrl ();
// Restore Rstate
__ load (Rstate, Address(r1, StackFrame::back_chain_offset * wordSize));
__ subi (Rstate, Rstate, sizeof(BytecodeInterpreter));
// Check for pending exceptions
__ load (r0, Address(Rthread, Thread::pending_exception_offset()));
__ compare (r0, 0);
__ bne (return_with_exception);
// Convert the result and resume
generate_convert_result(CppInterpreter::_tosca_to_stack);
__ b (resume_interpreter);
// Handle a return_from_method message ===============================
__ bind (return_from_method);
__ load (r0, STATE(_prev_link));
__ compare (r0, 0);
__ beq (return_to_initial_caller);
// "Return" from a re-dispatch -------------------------------------
generate_convert_result(CppInterpreter::_stack_to_stack);
generate_unwind_interpreter_state();
// Resume the interpreter
__ bind (resume_interpreter);
__ store (Rlocals, STATE(_stack));
__ load (Rlocals, STATE(_locals));
__ load (Rmethod, STATE(_method));
__ verify_oop(Rmethod);
__ load (r0, BytecodeInterpreter::method_resume);
__ stw (r0, STATE(_msg));
__ b (call_interpreter);
// Return to the initial caller (call_stub etc) --------------------
__ bind (return_to_initial_caller);
generate_convert_result(CppInterpreter::_stack_to_native_abi);
generate_unwind_interpreter_state();
__ blr ();
// Handle a more_monitors message ====================================
__ bind (more_monitors);
generate_more_monitors();
__ load (r0, BytecodeInterpreter::got_monitors);
__ stw (r0, STATE(_msg));
__ b (call_interpreter);
// Handle a throwing_exception message ===============================
__ bind (throwing_exception);
// Check we actually have an exception
#ifdef ASSERT
{
Label ok;
__ load (r0, Address(Rthread, Thread::pending_exception_offset()));
__ compare (r0, 0);
__ bne (ok);
__ should_not_reach_here (__FILE__, __LINE__);
__ bind (ok);
}
#endif
// Return to wherever
generate_unwind_interpreter_state();
__ bind (return_with_exception);
__ compare (Rstate, 0);
__ bne (resume_interpreter);
__ blr ();
normal_entry = start;
return start;
}
void MethodHandles::generate_method_handle_dispatch(MacroAssembler* _masm,
vmIntrinsics::ID iid,
Register receiver_reg,
Register member_reg,
bool for_compiler_entry) {
assert(is_signature_polymorphic(iid), "expected invoke iid");
Register temp1 = (for_compiler_entry ? R25_tmp5 : R7);
Register temp2 = (for_compiler_entry ? R22_tmp2 : R8);
Register temp3 = (for_compiler_entry ? R23_tmp3 : R9);
Register temp4 = (for_compiler_entry ? R24_tmp4 : R10);
if (receiver_reg != noreg) assert_different_registers(temp1, temp2, temp3, temp4, receiver_reg);
if (member_reg != noreg) assert_different_registers(temp1, temp2, temp3, temp4, member_reg);
if (iid == vmIntrinsics::_invokeBasic) {
// indirect through MH.form.vmentry.vmtarget
jump_to_lambda_form(_masm, receiver_reg, R19_method, temp1, temp2, for_compiler_entry);
} else {
// The method is a member invoker used by direct method handles.
if (VerifyMethodHandles) {
// make sure the trailing argument really is a MemberName (caller responsibility)
verify_klass(_masm, member_reg, SystemDictionary::WK_KLASS_ENUM_NAME(MemberName_klass),
temp1, temp2,
"MemberName required for invokeVirtual etc.");
}
Register temp1_recv_klass = temp1;
if (iid != vmIntrinsics::_linkToStatic) {
__ verify_oop(receiver_reg);
if (iid == vmIntrinsics::_linkToSpecial) {
// Don't actually load the klass; just null-check the receiver.
__ null_check_throw(receiver_reg, -1, temp1, EXCEPTION_ENTRY);
} else {
// load receiver klass itself
__ null_check_throw(receiver_reg, oopDesc::klass_offset_in_bytes(), temp1, EXCEPTION_ENTRY);
__ load_klass(temp1_recv_klass, receiver_reg);
__ verify_klass_ptr(temp1_recv_klass);
}
BLOCK_COMMENT("check_receiver {");
// The receiver for the MemberName must be in receiver_reg.
// Check the receiver against the MemberName.clazz
if (VerifyMethodHandles && iid == vmIntrinsics::_linkToSpecial) {
// Did not load it above...
__ load_klass(temp1_recv_klass, receiver_reg);
__ verify_klass_ptr(temp1_recv_klass);
}
if (VerifyMethodHandles && iid != vmIntrinsics::_linkToInterface) {
Label L_ok;
Register temp2_defc = temp2;
__ load_heap_oop_not_null(temp2_defc, NONZERO(java_lang_invoke_MemberName::clazz_offset_in_bytes()), member_reg, temp3);
load_klass_from_Class(_masm, temp2_defc, temp3, temp4);
__ verify_klass_ptr(temp2_defc);
__ check_klass_subtype(temp1_recv_klass, temp2_defc, temp3, temp4, L_ok);
// If we get here, the type check failed!
__ stop("receiver class disagrees with MemberName.clazz");
__ BIND(L_ok);
}
BLOCK_COMMENT("} check_receiver");
}
if (iid == vmIntrinsics::_linkToSpecial ||
iid == vmIntrinsics::_linkToStatic) {
DEBUG_ONLY(temp1_recv_klass = noreg); // these guys didn't load the recv_klass
}
// Live registers at this point:
// member_reg - MemberName that was the trailing argument
// temp1_recv_klass - klass of stacked receiver, if needed
// O5_savedSP - interpreter linkage (if interpreted)
// O0..O5 - compiler arguments (if compiled)
Label L_incompatible_class_change_error;
switch (iid) {
case vmIntrinsics::_linkToSpecial:
if (VerifyMethodHandles) {
verify_ref_kind(_masm, JVM_REF_invokeSpecial, member_reg, temp2);
}
__ ld(R19_method, NONZERO(java_lang_invoke_MemberName::vmtarget_offset_in_bytes()), member_reg);
break;
case vmIntrinsics::_linkToStatic:
if (VerifyMethodHandles) {
verify_ref_kind(_masm, JVM_REF_invokeStatic, member_reg, temp2);
}
__ ld(R19_method, NONZERO(java_lang_invoke_MemberName::vmtarget_offset_in_bytes()), member_reg);
break;
case vmIntrinsics::_linkToVirtual:
{
// same as TemplateTable::invokevirtual,
// minus the CP setup and profiling:
if (VerifyMethodHandles) {
verify_ref_kind(_masm, JVM_REF_invokeVirtual, member_reg, temp2);
}
// pick out the vtable index from the MemberName, and then we can discard it:
Register temp2_index = temp2;
__ ld(temp2_index, NONZERO(java_lang_invoke_MemberName::vmindex_offset_in_bytes()), member_reg);
if (VerifyMethodHandles) {
Label L_index_ok;
__ cmpdi(CCR1, temp2_index, 0);
__ bge(CCR1, L_index_ok);
__ stop("no virtual index");
__ BIND(L_index_ok);
}
// Note: The verifier invariants allow us to ignore MemberName.clazz and vmtarget
// at this point. And VerifyMethodHandles has already checked clazz, if needed.
// get target Method* & entry point
__ lookup_virtual_method(temp1_recv_klass, temp2_index, R19_method);
break;
}
case vmIntrinsics::_linkToInterface:
{
// same as TemplateTable::invokeinterface
// (minus the CP setup and profiling, with different argument motion)
if (VerifyMethodHandles) {
verify_ref_kind(_masm, JVM_REF_invokeInterface, member_reg, temp2);
}
Register temp2_intf = temp2;
__ load_heap_oop_not_null(temp2_intf, NONZERO(java_lang_invoke_MemberName::clazz_offset_in_bytes()), member_reg, temp3);
load_klass_from_Class(_masm, temp2_intf, temp3, temp4);
__ verify_klass_ptr(temp2_intf);
Register vtable_index = R19_method;
__ ld(vtable_index, NONZERO(java_lang_invoke_MemberName::vmindex_offset_in_bytes()), member_reg);
if (VerifyMethodHandles) {
Label L_index_ok;
__ cmpdi(CCR1, vtable_index, 0);
__ bge(CCR1, L_index_ok);
__ stop("invalid vtable index for MH.invokeInterface");
__ BIND(L_index_ok);
}
// given intf, index, and recv klass, dispatch to the implementation method
__ lookup_interface_method(temp1_recv_klass, temp2_intf,
// note: next two args must be the same:
vtable_index, R19_method,
temp3, temp4,
L_incompatible_class_change_error);
break;
}
default:
fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
break;
}
// Live at this point:
// R19_method
// O5_savedSP (if interpreted)
// After figuring out which concrete method to call, jump into it.
// Note that this works in the interpreter with no data motion.
// But the compiled version will require that rcx_recv be shifted out.
__ verify_method_ptr(R19_method);
jump_from_method_handle(_masm, R19_method, temp1, temp2, for_compiler_entry);
if (iid == vmIntrinsics::_linkToInterface) {
__ BIND(L_incompatible_class_change_error);
__ load_const_optimized(temp1, StubRoutines::throw_IncompatibleClassChangeError_entry());
__ mtctr(temp1);
__ bctr();
}
}
}
inline void MacroAssembler::call_stub_and_return_to(Register function_entry, Register return_pc) {
assert_different_registers(function_entry, return_pc);
mtlr(return_pc);
mtctr(function_entry);
bctr();
}
address InterpreterGenerator::generate_native_entry(bool synchronized)
{
const Register handler = r14;
const Register function = r15;
assert_different_registers(Rmethod, Rlocals, Rthread, Rstate, Rmonitor,
handler, function);
// We use the same code for synchronized and not
if (native_entry)
return native_entry;
address start = __ pc();
// Allocate and initialize our stack frame.
__ load (Rstate, 0);
generate_compute_interpreter_state(true);
// Make sure method is native and not abstract
#ifdef ASSERT
{
Label ok;
__ lwz (r0, Address(Rmethod, methodOopDesc::access_flags_offset()));
__ andi_ (r0, r0, JVM_ACC_NATIVE | JVM_ACC_ABSTRACT);
__ compare (r0, JVM_ACC_NATIVE);
__ beq (ok);
__ should_not_reach_here (__FILE__, __LINE__);
__ bind (ok);
}
#endif
// Lock if necessary
Label not_synchronized_1;
__ bne (CRsync, not_synchronized_1);
__ lock_object (Rmonitor);
__ bind (not_synchronized_1);
// Get signature handler
const Address signature_handler_addr(
Rmethod, methodOopDesc::signature_handler_offset());
Label return_to_caller, got_signature_handler;
__ load (handler, signature_handler_addr);
__ compare (handler, 0);
__ bne (got_signature_handler);
__ call_VM (noreg,
CAST_FROM_FN_PTR(address,
InterpreterRuntime::prepare_native_call),
Rmethod,
CALL_VM_NO_EXCEPTION_CHECKS);
__ load (r0, Address(Rthread, Thread::pending_exception_offset()));
__ compare (r0, 0);
__ bne (return_to_caller);
__ load (handler, signature_handler_addr);
__ bind (got_signature_handler);
// Get the native function entry point
const Address native_function_addr(
Rmethod, methodOopDesc::native_function_offset());
Label got_function;
__ load (function, native_function_addr);
#ifdef ASSERT
{
// InterpreterRuntime::prepare_native_call() sets the mirror
// handle and native function address first and the signature
// handler last, so function should always be set here.
Label ok;
__ compare (function, 0);
__ bne (ok);
__ should_not_reach_here (__FILE__, __LINE__);
__ bind (ok);
}
#endif
// Call signature handler
__ mtctr (handler);
__ bctrl ();
__ mr (handler, r0);
// Pass JNIEnv
__ la (r3, Address(Rthread, JavaThread::jni_environment_offset()));
// Pass mirror handle if static
const Address oop_temp_addr = STATE(_oop_temp);
Label not_static;
__ bne (CRstatic, not_static);
__ get_mirror_handle (r4);
__ store (r4, oop_temp_addr);
__ la (r4, oop_temp_addr);
__ bind (not_static);
// Set up the Java frame anchor
__ set_last_Java_frame ();
// Change the thread state to native
const Address thread_state_addr(Rthread, JavaThread::thread_state_offset());
#ifdef ASSERT
{
Label ok;
__ lwz (r0, thread_state_addr);
__ compare (r0, _thread_in_Java);
__ beq (ok);
__ should_not_reach_here (__FILE__, __LINE__);
__ bind (ok);
}
#endif
__ load (r0, _thread_in_native);
__ stw (r0, thread_state_addr);
// Make the call
__ call (function);
__ fixup_after_potential_safepoint ();
// The result will be in r3 (and maybe r4 on 32-bit) or f1.
// Wherever it is, we need to store it before calling anything
const Register r3_save = r16;
#ifdef PPC32
const Register r4_save = r17;
#endif
const FloatRegister f1_save = f14;
__ mr (r3_save, r3);
#ifdef PPC32
__ mr (r4_save, r4);
#endif
__ fmr (f1_save, f1);
// Switch thread to "native transition" state before reading the
// synchronization state. This additional state is necessary
// because reading and testing the synchronization state is not
// atomic with respect to garbage collection.
__ load (r0, _thread_in_native_trans);
__ stw (r0, thread_state_addr);
// Ensure the new state is visible to the VM thread.
if(os::is_MP()) {
if (UseMembar)
__ sync ();
else
__ serialize_memory (r3, r4);
}
// Check for safepoint operation in progress and/or pending
// suspend requests. We use a leaf call in order to leave
// the last_Java_frame setup undisturbed.
Label block, no_block;
__ load (r3, (intptr_t) SafepointSynchronize::address_of_state());
__ lwz (r0, Address(r3, 0));
__ compare (r0, SafepointSynchronize::_not_synchronized);
__ bne (block);
__ lwz (r0, Address(Rthread, JavaThread::suspend_flags_offset()));
__ compare (r0, 0);
__ beq (no_block);
__ bind (block);
__ call_VM_leaf (
CAST_FROM_FN_PTR(address,
JavaThread::check_special_condition_for_native_trans));
__ fixup_after_potential_safepoint ();
__ bind (no_block);
// Change the thread state
__ load (r0, _thread_in_Java);
__ stw (r0, thread_state_addr);
// Reset the frame anchor
__ reset_last_Java_frame ();
// If the result was an OOP then unbox it and store it in the frame
// (where it will be safe from garbage collection) before we release
// the handle it might be protected by
Label non_oop, store_oop;
__ load (r0, (intptr_t) AbstractInterpreter::result_handler(T_OBJECT));
__ compare (r0, handler);
__ bne (non_oop);
__ compare (r3_save, 0);
__ beq (store_oop);
__ load (r3_save, Address(r3_save, 0));
__ bind (store_oop);
__ store (r3_save, STATE(_oop_temp));
__ bind (non_oop);
// Reset handle block
__ load (r3, Address(Rthread, JavaThread::active_handles_offset()));
__ load (r0, 0);
__ stw (r0, Address(r3, JNIHandleBlock::top_offset_in_bytes()));
// If there is an exception we skip the result handler and return.
// Note that this also skips unlocking which seems totally wrong,
// but apparently this is what the asm interpreter does so we do
// too.
__ load (r0, Address(Rthread, Thread::pending_exception_offset()));
__ compare (r0, 0);
__ bne (return_to_caller);
// Unlock if necessary
Label not_synchronized_2;
__ bne (CRsync, not_synchronized_2);
__ unlock_object (Rmonitor);
__ bind (not_synchronized_2);
// Restore saved result and call the result handler
__ mr (r3, r3_save);
#ifdef PPC32
__ mr (r4, r4_save);
#endif
__ fmr (f1, f1_save);
__ mtctr (handler);
__ bctrl ();
// Unwind the current activation and return
__ bind (return_to_caller);
generate_unwind_interpreter_state();
__ blr ();
native_entry = start;
return start;
}
// Used by compiler only; may use only caller saved, non-argument
// registers.
VtableStub* VtableStubs::create_vtable_stub(int vtable_index) {
// PPC port: use fixed size.
const int code_length = VtableStub::pd_code_size_limit(true);
VtableStub* s = new (code_length) VtableStub(true, vtable_index);
// Can be NULL if there is no free space in the code cache.
if (s == NULL) {
return NULL;
}
ResourceMark rm;
CodeBuffer cb(s->entry_point(), code_length);
MacroAssembler* masm = new MacroAssembler(&cb);
#ifndef PRODUCT
if (CountCompiledCalls) {
int offs = __ load_const_optimized(R11_scratch1, SharedRuntime::nof_megamorphic_calls_addr(), R12_scratch2, true);
__ lwz(R12_scratch2, offs, R11_scratch1);
__ addi(R12_scratch2, R12_scratch2, 1);
__ stw(R12_scratch2, offs, R11_scratch1);
}
#endif
assert(VtableStub::receiver_location() == R3_ARG1->as_VMReg(), "receiver expected in R3_ARG1");
// Get receiver klass.
const Register rcvr_klass = R11_scratch1;
// We might implicit NULL fault here.
address npe_addr = __ pc(); // npe = null pointer exception
__ load_klass_with_trap_null_check(rcvr_klass, R3);
// Set method (in case of interpreted method), and destination address.
int entry_offset = InstanceKlass::vtable_start_offset() + vtable_index*vtableEntry::size();
#ifndef PRODUCT
if (DebugVtables) {
Label L;
// Check offset vs vtable length.
const Register vtable_len = R12_scratch2;
__ lwz(vtable_len, InstanceKlass::vtable_length_offset()*wordSize, rcvr_klass);
__ cmpwi(CCR0, vtable_len, vtable_index*vtableEntry::size());
__ bge(CCR0, L);
__ li(R12_scratch2, vtable_index);
__ call_VM(noreg, CAST_FROM_FN_PTR(address, bad_compiled_vtable_index), R3_ARG1, R12_scratch2, false);
__ bind(L);
}
#endif
int v_off = entry_offset*wordSize + vtableEntry::method_offset_in_bytes();
__ ld(R19_method, v_off, rcvr_klass);
#ifndef PRODUCT
if (DebugVtables) {
Label L;
__ cmpdi(CCR0, R19_method, 0);
__ bne(CCR0, L);
__ stop("Vtable entry is ZERO", 102);
__ bind(L);
}
#endif
// If the vtable entry is null, the method is abstract.
address ame_addr = __ pc(); // ame = abstract method error
__ load_with_trap_null_check(R12_scratch2, in_bytes(Method::from_compiled_offset()), R19_method);
__ mtctr(R12_scratch2);
__ bctr();
masm->flush();
guarantee(__ pc() <= s->code_end(), "overflowed buffer");
s->set_exception_points(npe_addr, ame_addr);
return s;
}
inline address MacroAssembler::call_stub(Register function_entry) {
mtctr(function_entry);
bctrl();
return pc();
}
VtableStub* VtableStubs::create_itable_stub(int itable_index) {
// PPC port: use fixed size.
const int code_length = VtableStub::pd_code_size_limit(false);
VtableStub* s = new (code_length) VtableStub(false, itable_index);
// Can be NULL if there is no free space in the code cache.
if (s == NULL) {
return NULL;
}
ResourceMark rm;
CodeBuffer cb(s->entry_point(), code_length);
MacroAssembler* masm = new MacroAssembler(&cb);
address start_pc;
#ifndef PRODUCT
if (CountCompiledCalls) {
int offs = __ load_const_optimized(R11_scratch1, SharedRuntime::nof_megamorphic_calls_addr(), R12_scratch2, true);
__ lwz(R12_scratch2, offs, R11_scratch1);
__ addi(R12_scratch2, R12_scratch2, 1);
__ stw(R12_scratch2, offs, R11_scratch1);
}
#endif
assert(VtableStub::receiver_location() == R3_ARG1->as_VMReg(), "receiver expected in R3_ARG1");
// Entry arguments:
// R19_method: Interface
// R3_ARG1: Receiver
Label L_no_such_interface;
const Register rcvr_klass = R11_scratch1,
interface = R12_scratch2,
tmp1 = R21_tmp1,
tmp2 = R22_tmp2;
address npe_addr = __ pc(); // npe = null pointer exception
__ load_klass_with_trap_null_check(rcvr_klass, R3_ARG1);
// Receiver subtype check against REFC.
__ ld(interface, CompiledICHolder::holder_klass_offset(), R19_method);
__ lookup_interface_method(rcvr_klass, interface, noreg,
R0, tmp1, tmp2,
L_no_such_interface, /*return_method=*/ false);
// Get Method* and entrypoint for compiler
__ ld(interface, CompiledICHolder::holder_metadata_offset(), R19_method);
__ lookup_interface_method(rcvr_klass, interface, itable_index,
R19_method, tmp1, tmp2,
L_no_such_interface, /*return_method=*/ true);
#ifndef PRODUCT
if (DebugVtables) {
Label ok;
__ cmpd(CCR0, R19_method, 0);
__ bne(CCR0, ok);
__ stop("method is null", 103);
__ bind(ok);
}
#endif
// If the vtable entry is null, the method is abstract.
address ame_addr = __ pc(); // ame = abstract method error
// Must do an explicit check if implicit checks are disabled.
assert(!MacroAssembler::needs_explicit_null_check(in_bytes(Method::from_compiled_offset())), "sanity");
if (!ImplicitNullChecks || !os::zero_page_read_protected()) {
if (TrapBasedNullChecks) {
__ trap_null_check(R19_method);
} else {
__ cmpdi(CCR0, R19_method, 0);
__ beq(CCR0, L_no_such_interface);
}
}
__ ld(R12_scratch2, in_bytes(Method::from_compiled_offset()), R19_method);
__ mtctr(R12_scratch2);
__ bctr();
// Handle IncompatibleClassChangeError in itable stubs.
// More detailed error message.
// We force resolving of the call site by jumping to the "handle
// wrong method" stub, and so let the interpreter runtime do all the
// dirty work.
__ bind(L_no_such_interface);
__ load_const_optimized(R11_scratch1, SharedRuntime::get_handle_wrong_method_stub(), R12_scratch2);
__ mtctr(R11_scratch1);
__ bctr();
masm->flush();
guarantee(__ pc() <= s->code_end(), "overflowed buffer");
s->set_exception_points(npe_addr, ame_addr);
return s;
}
address AbstractInterpreterGenerator::generate_slow_signature_handler() {
// Slow_signature handler that respects the PPC C calling conventions.
//
// We get called by the native entry code with our output register
// area == 8. First we call InterpreterRuntime::get_result_handler
// to copy the pointer to the signature string temporarily to the
// first C-argument and to return the result_handler in
// R3_RET. Since native_entry will copy the jni-pointer to the
// first C-argument slot later on, it is OK to occupy this slot
// temporarilly. Then we copy the argument list on the java
// expression stack into native varargs format on the native stack
// and load arguments into argument registers. Integer arguments in
// the varargs vector will be sign-extended to 8 bytes.
//
// On entry:
// R3_ARG1 - intptr_t* Address of java argument list in memory.
// R15_prev_state - BytecodeInterpreter* Address of interpreter state for
// this method
// R19_method
//
// On exit (just before return instruction):
// R3_RET - contains the address of the result_handler.
// R4_ARG2 - is not updated for static methods and contains "this" otherwise.
// R5_ARG3-R10_ARG8: - When the (i-2)th Java argument is not of type float or double,
// ARGi contains this argument. Otherwise, ARGi is not updated.
// F1_ARG1-F13_ARG13 - contain the first 13 arguments of type float or double.
const int LogSizeOfTwoInstructions = 3;
// FIXME: use Argument:: GL: Argument names different numbers!
const int max_fp_register_arguments = 13;
const int max_int_register_arguments = 6; // first 2 are reserved
const Register arg_java = R21_tmp1;
const Register arg_c = R22_tmp2;
const Register signature = R23_tmp3; // is string
const Register sig_byte = R24_tmp4;
const Register fpcnt = R25_tmp5;
const Register argcnt = R26_tmp6;
const Register intSlot = R27_tmp7;
const Register target_sp = R28_tmp8;
const FloatRegister floatSlot = F0;
address entry = __ function_entry();
__ save_LR_CR(R0);
__ save_nonvolatile_gprs(R1_SP, _spill_nonvolatiles_neg(r14));
// We use target_sp for storing arguments in the C frame.
__ mr(target_sp, R1_SP);
__ push_frame_reg_args_nonvolatiles(0, R11_scratch1);
__ mr(arg_java, R3_ARG1);
__ call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::get_signature), R16_thread, R19_method);
// Signature is in R3_RET. Signature is callee saved.
__ mr(signature, R3_RET);
// Get the result handler.
__ call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::get_result_handler), R16_thread, R19_method);
{
Label L;
// test if static
// _access_flags._flags must be at offset 0.
// TODO PPC port: requires change in shared code.
//assert(in_bytes(AccessFlags::flags_offset()) == 0,
// "MethodDesc._access_flags == MethodDesc._access_flags._flags");
// _access_flags must be a 32 bit value.
assert(sizeof(AccessFlags) == 4, "wrong size");
__ lwa(R11_scratch1/*access_flags*/, method_(access_flags));
// testbit with condition register.
__ testbitdi(CCR0, R0, R11_scratch1/*access_flags*/, JVM_ACC_STATIC_BIT);
__ btrue(CCR0, L);
// For non-static functions, pass "this" in R4_ARG2 and copy it
// to 2nd C-arg slot.
// We need to box the Java object here, so we use arg_java
// (address of current Java stack slot) as argument and don't
// dereference it as in case of ints, floats, etc.
__ mr(R4_ARG2, arg_java);
__ addi(arg_java, arg_java, -BytesPerWord);
__ std(R4_ARG2, _abi(carg_2), target_sp);
__ bind(L);
}
// Will be incremented directly after loop_start. argcnt=0
// corresponds to 3rd C argument.
__ li(argcnt, -1);
// arg_c points to 3rd C argument
__ addi(arg_c, target_sp, _abi(carg_3));
// no floating-point args parsed so far
__ li(fpcnt, 0);
Label move_intSlot_to_ARG, move_floatSlot_to_FARG;
Label loop_start, loop_end;
Label do_int, do_long, do_float, do_double, do_dontreachhere, do_object, do_array, do_boxed;
// signature points to '(' at entry
#ifdef ASSERT
__ lbz(sig_byte, 0, signature);
__ cmplwi(CCR0, sig_byte, '(');
__ bne(CCR0, do_dontreachhere);
#endif
__ bind(loop_start);
__ addi(argcnt, argcnt, 1);
__ lbzu(sig_byte, 1, signature);
__ cmplwi(CCR0, sig_byte, ')'); // end of signature
__ beq(CCR0, loop_end);
__ cmplwi(CCR0, sig_byte, 'B'); // byte
__ beq(CCR0, do_int);
__ cmplwi(CCR0, sig_byte, 'C'); // char
__ beq(CCR0, do_int);
__ cmplwi(CCR0, sig_byte, 'D'); // double
__ beq(CCR0, do_double);
__ cmplwi(CCR0, sig_byte, 'F'); // float
__ beq(CCR0, do_float);
__ cmplwi(CCR0, sig_byte, 'I'); // int
__ beq(CCR0, do_int);
__ cmplwi(CCR0, sig_byte, 'J'); // long
__ beq(CCR0, do_long);
__ cmplwi(CCR0, sig_byte, 'S'); // short
__ beq(CCR0, do_int);
__ cmplwi(CCR0, sig_byte, 'Z'); // boolean
__ beq(CCR0, do_int);
__ cmplwi(CCR0, sig_byte, 'L'); // object
__ beq(CCR0, do_object);
__ cmplwi(CCR0, sig_byte, '['); // array
__ beq(CCR0, do_array);
// __ cmplwi(CCR0, sig_byte, 'V'); // void cannot appear since we do not parse the return type
// __ beq(CCR0, do_void);
__ bind(do_dontreachhere);
__ unimplemented("ShouldNotReachHere in slow_signature_handler", 120);
__ bind(do_array);
{
Label start_skip, end_skip;
__ bind(start_skip);
__ lbzu(sig_byte, 1, signature);
__ cmplwi(CCR0, sig_byte, '[');
__ beq(CCR0, start_skip); // skip further brackets
__ cmplwi(CCR0, sig_byte, '9');
__ bgt(CCR0, end_skip); // no optional size
__ cmplwi(CCR0, sig_byte, '0');
__ bge(CCR0, start_skip); // skip optional size
__ bind(end_skip);
__ cmplwi(CCR0, sig_byte, 'L');
__ beq(CCR0, do_object); // for arrays of objects, the name of the object must be skipped
__ b(do_boxed); // otherwise, go directly to do_boxed
}
__ bind(do_object);
{
Label L;
__ bind(L);
__ lbzu(sig_byte, 1, signature);
__ cmplwi(CCR0, sig_byte, ';');
__ bne(CCR0, L);
}
// Need to box the Java object here, so we use arg_java (address of
// current Java stack slot) as argument and don't dereference it as
// in case of ints, floats, etc.
Label do_null;
__ bind(do_boxed);
__ ld(R0,0, arg_java);
__ cmpdi(CCR0, R0, 0);
__ li(intSlot,0);
__ beq(CCR0, do_null);
__ mr(intSlot, arg_java);
__ bind(do_null);
__ std(intSlot, 0, arg_c);
__ addi(arg_java, arg_java, -BytesPerWord);
__ addi(arg_c, arg_c, BytesPerWord);
__ cmplwi(CCR0, argcnt, max_int_register_arguments);
__ blt(CCR0, move_intSlot_to_ARG);
__ b(loop_start);
__ bind(do_int);
__ lwa(intSlot, 0, arg_java);
__ std(intSlot, 0, arg_c);
__ addi(arg_java, arg_java, -BytesPerWord);
__ addi(arg_c, arg_c, BytesPerWord);
__ cmplwi(CCR0, argcnt, max_int_register_arguments);
__ blt(CCR0, move_intSlot_to_ARG);
__ b(loop_start);
__ bind(do_long);
__ ld(intSlot, -BytesPerWord, arg_java);
__ std(intSlot, 0, arg_c);
__ addi(arg_java, arg_java, - 2 * BytesPerWord);
__ addi(arg_c, arg_c, BytesPerWord);
__ cmplwi(CCR0, argcnt, max_int_register_arguments);
__ blt(CCR0, move_intSlot_to_ARG);
__ b(loop_start);
__ bind(do_float);
__ lfs(floatSlot, 0, arg_java);
#if defined(LINUX)
// Linux uses ELF ABI. Both original ELF and ELFv2 ABIs have float
// in the least significant word of an argument slot.
#if defined(VM_LITTLE_ENDIAN)
__ stfs(floatSlot, 0, arg_c);
#else
__ stfs(floatSlot, 4, arg_c);
#endif
#elif defined(AIX)
// Although AIX runs on big endian CPU, float is in most significant
// word of an argument slot.
__ stfs(floatSlot, 0, arg_c);
#else
#error "unknown OS"
#endif
__ addi(arg_java, arg_java, -BytesPerWord);
__ addi(arg_c, arg_c, BytesPerWord);
__ cmplwi(CCR0, fpcnt, max_fp_register_arguments);
__ blt(CCR0, move_floatSlot_to_FARG);
__ b(loop_start);
__ bind(do_double);
__ lfd(floatSlot, - BytesPerWord, arg_java);
__ stfd(floatSlot, 0, arg_c);
__ addi(arg_java, arg_java, - 2 * BytesPerWord);
__ addi(arg_c, arg_c, BytesPerWord);
__ cmplwi(CCR0, fpcnt, max_fp_register_arguments);
__ blt(CCR0, move_floatSlot_to_FARG);
__ b(loop_start);
__ bind(loop_end);
__ pop_frame();
__ restore_nonvolatile_gprs(R1_SP, _spill_nonvolatiles_neg(r14));
__ restore_LR_CR(R0);
__ blr();
Label move_int_arg, move_float_arg;
__ bind(move_int_arg); // each case must consist of 2 instructions (otherwise adapt LogSizeOfTwoInstructions)
__ mr(R5_ARG3, intSlot); __ b(loop_start);
__ mr(R6_ARG4, intSlot); __ b(loop_start);
__ mr(R7_ARG5, intSlot); __ b(loop_start);
__ mr(R8_ARG6, intSlot); __ b(loop_start);
__ mr(R9_ARG7, intSlot); __ b(loop_start);
__ mr(R10_ARG8, intSlot); __ b(loop_start);
__ bind(move_float_arg); // each case must consist of 2 instructions (otherwise adapt LogSizeOfTwoInstructions)
__ fmr(F1_ARG1, floatSlot); __ b(loop_start);
__ fmr(F2_ARG2, floatSlot); __ b(loop_start);
__ fmr(F3_ARG3, floatSlot); __ b(loop_start);
__ fmr(F4_ARG4, floatSlot); __ b(loop_start);
__ fmr(F5_ARG5, floatSlot); __ b(loop_start);
__ fmr(F6_ARG6, floatSlot); __ b(loop_start);
__ fmr(F7_ARG7, floatSlot); __ b(loop_start);
__ fmr(F8_ARG8, floatSlot); __ b(loop_start);
__ fmr(F9_ARG9, floatSlot); __ b(loop_start);
__ fmr(F10_ARG10, floatSlot); __ b(loop_start);
__ fmr(F11_ARG11, floatSlot); __ b(loop_start);
__ fmr(F12_ARG12, floatSlot); __ b(loop_start);
__ fmr(F13_ARG13, floatSlot); __ b(loop_start);
__ bind(move_intSlot_to_ARG);
__ sldi(R0, argcnt, LogSizeOfTwoInstructions);
__ load_const(R11_scratch1, move_int_arg); // Label must be bound here.
__ add(R11_scratch1, R0, R11_scratch1);
__ mtctr(R11_scratch1/*branch_target*/);
__ bctr();
__ bind(move_floatSlot_to_FARG);
__ sldi(R0, fpcnt, LogSizeOfTwoInstructions);
__ addi(fpcnt, fpcnt, 1);
__ load_const(R11_scratch1, move_float_arg); // Label must be bound here.
__ add(R11_scratch1, R0, R11_scratch1);
__ mtctr(R11_scratch1/*branch_target*/);
__ bctr();
return entry;
}
// Call an accessor method (assuming it is resolved, otherwise drop into
// vanilla (slow path) entry.
address InterpreterGenerator::generate_accessor_entry(void) {
if (!UseFastAccessorMethods && (!FLAG_IS_ERGO(UseFastAccessorMethods))) {
return NULL;
}
Label Lslow_path, Lacquire;
const Register
Rclass_or_obj = R3_ARG1,
Rconst_method = R4_ARG2,
Rcodes = Rconst_method,
Rcpool_cache = R5_ARG3,
Rscratch = R11_scratch1,
Rjvmti_mode = Rscratch,
Roffset = R12_scratch2,
Rflags = R6_ARG4,
Rbtable = R7_ARG5;
static address branch_table[number_of_states];
address entry = __ pc();
// Check for safepoint:
// Ditch this, real man don't need safepoint checks.
// Also check for JVMTI mode
// Check for null obj, take slow path if so.
__ ld(Rclass_or_obj, Interpreter::stackElementSize, CC_INTERP_ONLY(R17_tos) NOT_CC_INTERP(R15_esp));
__ lwz(Rjvmti_mode, thread_(interp_only_mode));
__ cmpdi(CCR1, Rclass_or_obj, 0);
__ cmpwi(CCR0, Rjvmti_mode, 0);
__ crorc(/*CCR0 eq*/2, /*CCR1 eq*/4+2, /*CCR0 eq*/2);
__ beq(CCR0, Lslow_path); // this==null or jvmti_mode!=0
// Do 2 things in parallel:
// 1. Load the index out of the first instruction word, which looks like this:
// <0x2a><0xb4><index (2 byte, native endianess)>.
// 2. Load constant pool cache base.
__ ld(Rconst_method, in_bytes(Method::const_offset()), R19_method);
__ ld(Rcpool_cache, in_bytes(ConstMethod::constants_offset()), Rconst_method);
__ lhz(Rcodes, in_bytes(ConstMethod::codes_offset()) + 2, Rconst_method); // Lower half of 32 bit field.
__ ld(Rcpool_cache, ConstantPool::cache_offset_in_bytes(), Rcpool_cache);
// Get the const pool entry by means of <index>.
const int codes_shift = exact_log2(in_words(ConstantPoolCacheEntry::size()) * BytesPerWord);
__ slwi(Rscratch, Rcodes, codes_shift); // (codes&0xFFFF)<<codes_shift
__ add(Rcpool_cache, Rscratch, Rcpool_cache);
// Check if cpool cache entry is resolved.
// We are resolved if the indices offset contains the current bytecode.
ByteSize cp_base_offset = ConstantPoolCache::base_offset();
// Big Endian:
__ lbz(Rscratch, in_bytes(cp_base_offset) + in_bytes(ConstantPoolCacheEntry::indices_offset()) + 7 - 2, Rcpool_cache);
__ cmpwi(CCR0, Rscratch, Bytecodes::_getfield);
__ bne(CCR0, Lslow_path);
__ isync(); // Order succeeding loads wrt. load of _indices field from cpool_cache.
// Finally, start loading the value: Get cp cache entry into regs.
__ ld(Rflags, in_bytes(cp_base_offset) + in_bytes(ConstantPoolCacheEntry::flags_offset()), Rcpool_cache);
__ ld(Roffset, in_bytes(cp_base_offset) + in_bytes(ConstantPoolCacheEntry::f2_offset()), Rcpool_cache);
// Following code is from templateTable::getfield_or_static
// Load pointer to branch table
__ load_const_optimized(Rbtable, (address)branch_table, Rscratch);
// Get volatile flag
__ rldicl(Rscratch, Rflags, 64-ConstantPoolCacheEntry::is_volatile_shift, 63); // extract volatile bit
// note: sync is needed before volatile load on PPC64
// Check field type
__ rldicl(Rflags, Rflags, 64-ConstantPoolCacheEntry::tos_state_shift, 64-ConstantPoolCacheEntry::tos_state_bits);
#ifdef ASSERT
Label LFlagInvalid;
__ cmpldi(CCR0, Rflags, number_of_states);
__ bge(CCR0, LFlagInvalid);
__ ld(R9_ARG7, 0, R1_SP);
__ ld(R10_ARG8, 0, R21_sender_SP);
__ cmpd(CCR0, R9_ARG7, R10_ARG8);
__ asm_assert_eq("backlink", 0x543);
#endif // ASSERT
__ mr(R1_SP, R21_sender_SP); // Cut the stack back to where the caller started.
// Load from branch table and dispatch (volatile case: one instruction ahead)
__ sldi(Rflags, Rflags, LogBytesPerWord);
__ cmpwi(CCR6, Rscratch, 1); // volatile?
if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
__ sldi(Rscratch, Rscratch, exact_log2(BytesPerInstWord)); // volatile ? size of 1 instruction : 0
}
__ ldx(Rbtable, Rbtable, Rflags);
if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
__ subf(Rbtable, Rscratch, Rbtable); // point to volatile/non-volatile entry point
}
__ mtctr(Rbtable);
__ bctr();
#ifdef ASSERT
__ bind(LFlagInvalid);
__ stop("got invalid flag", 0x6541);
bool all_uninitialized = true,
all_initialized = true;
for (int i = 0; i<number_of_states; ++i) {
all_uninitialized = all_uninitialized && (branch_table[i] == NULL);
all_initialized = all_initialized && (branch_table[i] != NULL);
}
assert(all_uninitialized != all_initialized, "consistency"); // either or
__ fence(); // volatile entry point (one instruction before non-volatile_entry point)
if (branch_table[vtos] == 0) branch_table[vtos] = __ pc(); // non-volatile_entry point
if (branch_table[dtos] == 0) branch_table[dtos] = __ pc(); // non-volatile_entry point
if (branch_table[ftos] == 0) branch_table[ftos] = __ pc(); // non-volatile_entry point
__ stop("unexpected type", 0x6551);
#endif
if (branch_table[itos] == 0) { // generate only once
__ align(32, 28, 28); // align load
__ fence(); // volatile entry point (one instruction before non-volatile_entry point)
branch_table[itos] = __ pc(); // non-volatile_entry point
__ lwax(R3_RET, Rclass_or_obj, Roffset);
__ beq(CCR6, Lacquire);
__ blr();
}
if (branch_table[ltos] == 0) { // generate only once
__ align(32, 28, 28); // align load
__ fence(); // volatile entry point (one instruction before non-volatile_entry point)
branch_table[ltos] = __ pc(); // non-volatile_entry point
__ ldx(R3_RET, Rclass_or_obj, Roffset);
__ beq(CCR6, Lacquire);
__ blr();
}
if (branch_table[btos] == 0) { // generate only once
__ align(32, 28, 28); // align load
__ fence(); // volatile entry point (one instruction before non-volatile_entry point)
branch_table[btos] = __ pc(); // non-volatile_entry point
__ lbzx(R3_RET, Rclass_or_obj, Roffset);
__ extsb(R3_RET, R3_RET);
__ beq(CCR6, Lacquire);
__ blr();
}
if (branch_table[ctos] == 0) { // generate only once
__ align(32, 28, 28); // align load
__ fence(); // volatile entry point (one instruction before non-volatile_entry point)
branch_table[ctos] = __ pc(); // non-volatile_entry point
__ lhzx(R3_RET, Rclass_or_obj, Roffset);
__ beq(CCR6, Lacquire);
__ blr();
}
if (branch_table[stos] == 0) { // generate only once
__ align(32, 28, 28); // align load
__ fence(); // volatile entry point (one instruction before non-volatile_entry point)
branch_table[stos] = __ pc(); // non-volatile_entry point
__ lhax(R3_RET, Rclass_or_obj, Roffset);
__ beq(CCR6, Lacquire);
__ blr();
}
if (branch_table[atos] == 0) { // generate only once
__ align(32, 28, 28); // align load
__ fence(); // volatile entry point (one instruction before non-volatile_entry point)
branch_table[atos] = __ pc(); // non-volatile_entry point
__ load_heap_oop(R3_RET, (RegisterOrConstant)Roffset, Rclass_or_obj);
__ verify_oop(R3_RET);
//__ dcbt(R3_RET); // prefetch
__ beq(CCR6, Lacquire);
__ blr();
}
__ align(32, 12);
__ bind(Lacquire);
__ twi_0(R3_RET);
__ isync(); // acquire
__ blr();
#ifdef ASSERT
for (int i = 0; i<number_of_states; ++i) {
assert(branch_table[i], "accessor_entry initialization");
//tty->print_cr("accessor_entry: branch_table[%d] = 0x%llx (opcode 0x%llx)", i, branch_table[i], *((unsigned int*)branch_table[i]));
}
#endif
__ bind(Lslow_path);
__ branch_to_entry(Interpreter::entry_for_kind(Interpreter::zerolocals), Rscratch);
__ flush();
return entry;
}
void C1_MacroAssembler::initialize_body(Register obj, Register tmp1, Register tmp2,
int obj_size_in_bytes, int hdr_size_in_bytes) {
const int index = (obj_size_in_bytes - hdr_size_in_bytes) / HeapWordSize;
const int cl_size = VM_Version::L1_data_cache_line_size(),
cl_dwords = cl_size>>3,
cl_dw_addr_bits = exact_log2(cl_dwords);
const Register tmp = R0,
base_ptr = tmp1,
cnt_dwords = tmp2;
if (index <= 6) {
// Use explicit NULL stores.
if (index > 0) { li(tmp, 0); }
for (int i = 0; i < index; ++i) { std(tmp, hdr_size_in_bytes + i * HeapWordSize, obj); }
} else if (index < (2<<cl_dw_addr_bits)-1) {
// simple loop
Label loop;
li(cnt_dwords, index);
addi(base_ptr, obj, hdr_size_in_bytes); // Compute address of first element.
li(tmp, 0);
mtctr(cnt_dwords); // Load counter.
bind(loop);
std(tmp, 0, base_ptr); // Clear 8byte aligned block.
addi(base_ptr, base_ptr, 8);
bdnz(loop);
} else {
// like clear_memory_doubleword
Label startloop, fast, fastloop, restloop, done;
addi(base_ptr, obj, hdr_size_in_bytes); // Compute address of first element.
load_const_optimized(cnt_dwords, index);
rldicl_(tmp, base_ptr, 64-3, 64-cl_dw_addr_bits); // Extract dword offset within first cache line.
beq(CCR0, fast); // Already 128byte aligned.
subfic(tmp, tmp, cl_dwords);
mtctr(tmp); // Set ctr to hit 128byte boundary (0<ctr<cl_dwords).
subf(cnt_dwords, tmp, cnt_dwords); // rest.
li(tmp, 0);
bind(startloop); // Clear at the beginning to reach 128byte boundary.
std(tmp, 0, base_ptr); // Clear 8byte aligned block.
addi(base_ptr, base_ptr, 8);
bdnz(startloop);
bind(fast); // Clear 128byte blocks.
srdi(tmp, cnt_dwords, cl_dw_addr_bits); // Loop count for 128byte loop (>0).
andi(cnt_dwords, cnt_dwords, cl_dwords-1); // Rest in dwords.
mtctr(tmp); // Load counter.
bind(fastloop);
dcbz(base_ptr); // Clear 128byte aligned block.
addi(base_ptr, base_ptr, cl_size);
bdnz(fastloop);
cmpdi(CCR0, cnt_dwords, 0); // size 0?
beq(CCR0, done); // rest == 0
li(tmp, 0);
mtctr(cnt_dwords); // Load counter.
bind(restloop); // Clear rest.
std(tmp, 0, base_ptr); // Clear 8byte aligned block.
addi(base_ptr, base_ptr, 8);
bdnz(restloop);
bind(done);
}
}
address CompiledStaticCall::emit_to_interp_stub(CodeBuffer &cbuf, address mark/* = NULL*/) {
#ifdef COMPILER2
if (mark == NULL) {
// Get the mark within main instrs section which is set to the address of the call.
mark = cbuf.insts_mark();
}
// Note that the code buffer's insts_mark is always relative to insts.
// That's why we must use the macroassembler to generate a stub.
MacroAssembler _masm(&cbuf);
// Start the stub.
address stub = __ start_a_stub(CompiledStaticCall::to_interp_stub_size());
if (stub == NULL) {
return NULL; // CodeCache is full
}
// For java_to_interp stubs we use R11_scratch1 as scratch register
// and in call trampoline stubs we use R12_scratch2. This way we
// can distinguish them (see is_NativeCallTrampolineStub_at()).
Register reg_scratch = R11_scratch1;
// Create a static stub relocation which relates this stub
// with the call instruction at insts_call_instruction_offset in the
// instructions code-section.
__ relocate(static_stub_Relocation::spec(mark));
const int stub_start_offset = __ offset();
// Now, create the stub's code:
// - load the TOC
// - load the inline cache oop from the constant pool
// - load the call target from the constant pool
// - call
__ calculate_address_from_global_toc(reg_scratch, __ method_toc());
AddressLiteral ic = __ allocate_metadata_address((Metadata *)NULL);
bool success = __ load_const_from_method_toc(as_Register(Matcher::inline_cache_reg_encode()),
ic, reg_scratch, /*fixed_size*/ true);
if (!success) {
return NULL; // CodeCache is full
}
if (ReoptimizeCallSequences) {
__ b64_patchable((address)-1, relocInfo::none);
} else {
AddressLiteral a((address)-1);
success = __ load_const_from_method_toc(reg_scratch, a, reg_scratch, /*fixed_size*/ true);
if (!success) {
return NULL; // CodeCache is full
}
__ mtctr(reg_scratch);
__ bctr();
}
// FIXME: Assert that the stub can be identified and patched.
// Java_to_interp_stub_size should be good.
assert((__ offset() - stub_start_offset) <= CompiledStaticCall::to_interp_stub_size(),
"should be good size");
assert(!is_NativeCallTrampolineStub_at(__ addr_at(stub_start_offset)),
"must not confuse java_to_interp with trampoline stubs");
// End the stub.
__ end_a_stub();
return stub;
#else
ShouldNotReachHere();
return NULL;
#endif
}
