void CppInterpreterGenerator::generate_adjust_callers_stack()
{
StackFrame frame;
const int frame_header_size = frame.unaligned_size() + slop_factor;
const Address param_words_addr(
Rmethod, methodOopDesc::size_of_parameters_offset());
const Address local_words_addr(
Rmethod, methodOopDesc::size_of_locals_offset());
const Register param_words = r3;
const Register local_words = r4;
Label loop, done;
// Check whether extra locals are actually required
__ lhz (param_words, param_words_addr);
__ lhz (local_words, local_words_addr);
__ compare (param_words, local_words);
__ beq (done);
// Extend the frame if necessary
const Register required_bytes = r5;
const Register available_bytes = r6;
__ shift_left (required_bytes, local_words, LogBytesPerWord);
__ sub (available_bytes, Rlocals, r1);
__ subi (available_bytes, available_bytes, frame_header_size);
__ maybe_extend_frame (required_bytes, available_bytes);
// Zero the extra locals
const Register dst = r7;
__ shift_left (dst, param_words, LogBytesPerWord);
__ sub (dst, Rlocals, dst);
__ sub (r0, local_words, param_words);
__ mtctr (r0);
__ load (r0, 0);
__ bind (loop);
__ store (r0, Address(dst, 0));
__ subi (dst, dst, wordSize);
__ bdnz (loop);
__ bind (done);
}
void CppInterpreterGenerator::generate_compute_interpreter_state(bool native)
{
StackFrame frame;
const Address stack_words_addr(
Rmethod, methodOopDesc::max_stack_offset());
const Address access_flags_addr(
Rmethod, methodOopDesc::access_flags_offset());
Label not_synchronized_1, not_synchronized_2, not_synchronized_3;
Label not_static, init_monitor;
const int monitor_size = frame::interpreter_frame_monitor_size() * wordSize;
// Calculate the access flags conditions
const Register access_flags = r3;
__ lwz (access_flags, access_flags_addr);
__ andi_ (r0, access_flags, JVM_ACC_SYNCHRONIZED);
__ compare (CRsync, r0, JVM_ACC_SYNCHRONIZED);
__ andi_ (r0, access_flags, JVM_ACC_STATIC);
__ compare (CRstatic, r0, JVM_ACC_STATIC);
const int basic_frame_size =
frame.unaligned_size() + sizeof(BytecodeInterpreter) + slop_factor;
// Calculate the frame size
const Register stack_size = r3;
const Register frame_size = r4;
const Register padding = r5;
if (native) {
__ load (frame_size, basic_frame_size);
}
else {
__ lhz (stack_size, stack_words_addr);
__ shift_left (stack_size, stack_size, LogBytesPerWord);
__ addi (frame_size, stack_size, basic_frame_size);
}
__ bne (CRsync, not_synchronized_1);
__ addi (frame_size, frame_size, monitor_size);
__ bind (not_synchronized_1);
__ 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);
// Calculate everything's addresses
const Register stack_limit = r6;
const Register stack = r7;
const Register stack_base = Rmonitor;
const Register monitor_base = r8;
__ addi (stack_limit, r1, frame.start_of_locals() + slop_factor - wordSize);
__ add (stack_limit, stack_limit, padding);
if (native)
__ mr (stack, stack_limit);
else
__ add (stack, stack_limit, stack_size);
__ addi (stack_base, stack, wordSize);
__ mr (monitor_base, stack_base);
__ bne (CRsync, not_synchronized_2);
__ addi (monitor_base, monitor_base, monitor_size);
__ bind (not_synchronized_2);
__ mr (r0, Rstate);
__ mr (Rstate, monitor_base);
// Initialise the interpreter state object
__ store (Rlocals, STATE(_locals));
__ store (Rmethod, STATE(_method));
__ store (Rstate, STATE(_self_link));
__ store (r0, STATE(_prev_link));
__ store (stack_limit, STATE(_stack_limit));
__ store (stack, STATE(_stack));
__ store (stack_base, STATE(_stack_base));
__ store (monitor_base, STATE(_monitor_base));
__ store (Rthread, STATE(_thread));
#ifdef ASSERT
{
Label ok;
__ load (r3, ThreadLocalStorage::thread_index());
__ call (CAST_FROM_FN_PTR(address, pthread_getspecific));
__ compare (Rthread, r3);
__ beq (ok);
__ should_not_reach_here (__FILE__, __LINE__);
__ bind (ok);
}
#endif
if (!native) {
__ load (r3, Address(Rmethod, methodOopDesc::const_offset()));
__ addi (r3, r3, in_bytes(constMethodOopDesc::codes_offset()));
__ store (r3, STATE(_bcp));
}
__ load (r3, Address(Rmethod, methodOopDesc::constants_offset()));
__ load (r3, Address(r3, constantPoolOopDesc::cache_offset_in_bytes()));
__ store (r3, STATE(_constants));
__ load (r3, BytecodeInterpreter::method_entry);
__ stw (r3, STATE(_msg));
__ load (r3, 0);
if (native)
__ store (r3, STATE(_bcp));
__ store (r3, STATE(_oop_temp));
__ store (r3, STATE(_mdx));
__ store (r3, STATE(_result._to_call._callee));
// Initialise the monitor if synchronized
__ bne (CRsync, not_synchronized_3);
__ bne (CRstatic, not_static);
__ get_mirror_handle (r3);
__ b (init_monitor);
__ bind (not_static);
__ load (r3, Address(Rlocals, 0));
__ bind (init_monitor);
__ store (r3, Address(Rmonitor, BasicObjectLock::obj_offset_in_bytes()));
__ bind (not_synchronized_3);
}
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;
}
// 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;
}