// If deoptimization happens, this function returns the point of next bytecode to continue execution
address AbstractInterpreter::deopt_continue_after_entry(methodOop method, address bcp, int callee_parameters, bool is_top_frame) {
assert(method->contains(bcp), "just checkin'");
Bytecodes::Code code = Bytecodes::java_code_at(bcp);
assert(!Interpreter::bytecode_should_reexecute(code), "should not reexecute");
int bci = method->bci_from(bcp);
int length = -1; // initial value for debugging
// compute continuation length
length = Bytecodes::length_at(bcp);
// compute result type
BasicType type = T_ILLEGAL;
switch (code) {
case Bytecodes::_invokevirtual :
case Bytecodes::_invokespecial :
case Bytecodes::_invokestatic :
case Bytecodes::_invokeinterface: {
Thread *thread = Thread::current();
ResourceMark rm(thread);
methodHandle mh(thread, method);
type = Bytecode_invoke_at(mh, bci)->result_type(thread);
// since the cache entry might not be initialized:
// (NOT needed for the old calling convension)
if (!is_top_frame) {
int index = Bytes::get_native_u2(bcp+1);
method->constants()->cache()->entry_at(index)->set_parameter_size(callee_parameters);
}
break;
}
case Bytecodes::_invokedynamic: {
Thread *thread = Thread::current();
ResourceMark rm(thread);
methodHandle mh(thread, method);
type = Bytecode_invoke_at(mh, bci)->result_type(thread);
// since the cache entry might not be initialized:
// (NOT needed for the old calling convension)
if (!is_top_frame) {
int index = Bytes::get_native_u4(bcp+1);
method->constants()->cache()->secondary_entry_at(index)->set_parameter_size(callee_parameters);
}
break;
}
case Bytecodes::_ldc :
case Bytecodes::_ldc_w : // fall through
case Bytecodes::_ldc2_w:
{
Thread *thread = Thread::current();
ResourceMark rm(thread);
methodHandle mh(thread, method);
type = Bytecode_loadconstant_at(mh, bci)->result_type();
break;
}
default:
type = Bytecodes::result_type(code);
break;
}
// return entry point for computed continuation state & bytecode length
return
is_top_frame
? Interpreter::deopt_entry (as_TosState(type), length)
: Interpreter::return_entry(as_TosState(type), length);
}
// --- oops_arguments_do_impl ------------------------------------------------
void CodeBlob::oops_arguments_do_impl(frame fr, OopClosure* f) const {
ResourceMark rm;
// We are in some trampoline (resolve_and_patch_call) doing a GC. The
// pending call has arguments that need GC'ing, but we do not yet know the
// target method and cannot resolve the target method yet.
frame cc = fr.sender(); // Compiled caller expected
symbolOop meth_sig;
bool is_static;
if( cc.is_entry_frame() ) {
// There's a rare race condition where the caller is an entry frame, but
// the target got patched not-entrant before the call could be made.
methodOop moop = cc.entry_frame_call_wrapper()->callee_method();
meth_sig = moop->signature();
is_static = moop->is_static();
} else {
Bytecode_invoke*call;
if(cc.is_interpreted_frame()){
// There's a rare race condition where we might need to GC in
// resolve_and_patch_call but the caller is an interpreted frame.
call = Bytecode_invoke_at(cc.interpreter_frame_method(),
cc.interpreter_frame_bci());
} else {
// Normal case: find caller's callsite
CodeBlob*cb=CodeCache::find_blob(cc.pc());
const DebugScope *ds = cb->debuginfo(cc.pc());
call = Bytecode_invoke_at(ds->method(), ds->bci());
}
meth_sig = call->signature();
is_static = (call->adjusted_invoke_code() == Bytecodes::_invokestatic);
}
int argcnt = 0; // Size of signature
if( !is_static ) argcnt++; // Add receiver
for(SignatureStream ss(meth_sig);!ss.at_return_type();ss.next()){
argcnt++;
if (ss.type() == T_LONG || ss.type() == T_DOUBLE) argcnt++;
}
BasicType * sig_bt = NEW_RESOURCE_ARRAY(BasicType ,argcnt);
VReg::VR * regs = NEW_RESOURCE_ARRAY(VReg::VR,argcnt);
int i=0;
if( !is_static ) sig_bt[i++] = T_OBJECT;
for(SignatureStream ss(meth_sig);!ss.at_return_type();ss.next()){
sig_bt[i++] = ss.type(); // Collect remaining bits of signature
if (ss.type() == T_LONG || ss.type() == T_DOUBLE)
sig_bt[i++] = T_VOID; // Longs & doubles take 2 Java slots
}
assert0(i==argcnt);
// Now get the re-packed compiled-Java layout. Registers are numbered from
// the callee's point of view.
SharedRuntime::java_calling_convention(sig_bt,regs,argcnt,true);
// Find the oop locations and do the GC thing
for(int i=0;i<argcnt;i++){
if ((sig_bt[i] == T_OBJECT) || (sig_bt[i] == T_ARRAY)) {
objectRef *loc = cc.reg_to_addr_oop(VReg::as_VOopReg(regs[i]));
f->do_oop(loc);
}
}
}
// --- build_repack_buffer ---------------------------------------------------
// Build a IFrame structure to help ASM code repack the 1 compiled frame into
// many interpreter (or C1) frames. Takes in the current thread and a vframe;
// the vframe is pointing and the virtual Java frame needing to be repacked.
// It takes in the callee (which this frame is busy trying to call in it's
// inlined code), and an array of IFrames. It returns the updated IFrame
// buffer filled in for this frame.
void Deoptimization::build_repack_buffer( JavaThread *thread, frame fr, IFrame *buf, const DebugMap *dm, const DebugScope *ds, intptr_t *jexstk, objectRef *lckstk, bool is_deopt, bool is_c1, bool is_youngest) {
assert( thread->_deopt_buffer->contains((char*)(buf+1)), "over-ran large deopt buffer?" );
int bci=ds->bci();
if(bci==InvocationEntryBci){
// We deoptimized while hanging in prologue code for a synchronized
// method. We got the lock (after all, deopt happens after returning
// from the blocking call). We want to begin execution in the
// interpreter at BCI 0, and after taking the lock.
// Also it is possilble to enter the deopt code through the br_s on method
// entry before the first byte code.
bci = 0;
}
const methodOop moop = ds->method().as_methodOop();
if( ds->caller() ) { // Do I have a caller? Am I mid-call?
// Initialize the constant pool entry for caller-parameter size. It
// might be the case that we inlined and compiled a callee, and are busy
// calling it in the compiled code, and get deoptimized with that callee
// in-progress AND we've never executed it in the interpreter - which
// would have filled in the constant pool cache before making the call.
// Fill it in now.
const methodOop caller = ds->caller()->method().as_methodOop();
int index = Bytes::get_native_u2(caller->bcp_from(ds->caller()->bci())+1);
ConstantPoolCacheEntry *cpe = caller->constants()->cache()->entry_at(index);
// Since we are setting the constant pool entry here, and another thread
// could be busy resolving here we have a race condition setting the
// flags. Use a CAS to only set the flags if they are currently 0.
intx *flags_adr = (intx*)((intptr_t)cpe + in_bytes(ConstantPoolCacheEntry::flags_offset()));
if( !*flags_adr ) { // Flags currently 0?
// Set the flags, because the interpreter-return-entry points need some
// info from them. Not all fields are set, because it's too complex to
// do it here... and not needed. The cpCacheEntry is left "unresolved"
// such that the next real use of it from the interpreter will be forced
// to do a proper resolve, which will fill in the missing fields.
// Compute new flags needed by the interpreter-return-entry
intx flags =
(moop->size_of_parameters() & 0xFF) |
(1 << ConstantPoolCacheEntry::hotSwapBit) |
(moop->result_type() << ConstantPoolCacheEntry::tosBits);
// CAS 'em in, but only if there is currently a 0 flags
assert0( sizeof(jlong)==sizeof(intx) );
Atomic::cmpxchg((jlong)flags, (jlong*)flags_adr, 0);
// We don't care about the result, because the cache is monomorphic.
// Either our CAS succeeded and jammed the right parameter count, or
// another thread succeeded and jammed in the right parameter count.
}
}
if (TraceDeoptimization) {
BufferedLoggerMark m(NOTAG, Log::M_DEOPT, TraceDeoptimization, true);
m.out("DEOPT REPACK c%d: ", is_c1 ? 1 : 2);
moop->print_short_name(m.stream());
m.out(" @ bci %d %s", bci, ds->caller() ? "called by...": " (oldest frame)" );
}
// If there was a suitable C1 frame, use it.
// Otherwise, use an interpreter frame.
if( 1 ) {
// Build an interpreter-style IFrame. Naked oops abound.
assert0( !objectRef(moop).is_stack() );
buf->_mref = objectRef(moop);
buf->_cpc = moop->constants()->cacheRef();
// Compute monitor list length. If we have coarsened a lock we will end
// up unlocking it and the repack buffer will not need to see it.
uint mons_len = ds->numlocks();
if( ds->is_extra_lock() ) { mons_len--; assert0( mons_len >= 0 ); }
assert0( mons_len < (256*sizeof(buf->_numlck)) );
buf->_numlck = mons_len;
// Set up the return pc for the next frame: the next frame is a younger
// frame which will return to this older frame. All middle frames return
// back into the interpreter, just after a call with proper TOS state.
// Youngest frames always start in vtos state because the uncommon-trap
// blob sets them up that way.
const address bcp = moop->bcp_from(bci);
Bytecodes::Code c = Bytecodes::java_code(Bytecodes::cast(*bcp));
BasicType return_type=T_VOID;
bool handle_popframe = is_youngest && JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution();
int bci_bump = 0;
if( !is_youngest ) { // Middle-frame?
bool from_call = (c == Bytecodes::_invokevirtual ||
c==Bytecodes::_invokespecial||
c==Bytecodes::_invokestatic||
c == Bytecodes::_invokeinterface );
assert(from_call,"Middle frame is in the middle of a call");
bci_bump = Bytecodes::length_at(bcp); // But need to know how much it will be bumped for the return address
buf->_bci = bci; // Save bci without bumping it; normal interpreter call returns bump the bci as needed
buf[-1]._retadr = Interpreter::return_entry(vtos, bci_bump);
} else if( thread->pending_exception() ) {
// Deopt-with-pending. Throw up on return to interpreter, which is
// handled by unpack_and_go.
buf->_bci=bci;
buf[-1]._retadr = Interpreter::unpack_and_go();
} else if( !is_deopt ) { // It is a C2-style uncommon-trap.
// Do NOT increment the BCP! We are re-executing the current bytecode.
buf->_bci=bci;
buf[-1]._retadr = Interpreter::unpack_and_go();
} else { // It is a plain deopt
// It is a deopt without exception. See if we are C1 in mid-patch.
// If so, we always need to re-execute the bytecode.
bool is_C1_mid_patch = false;
if( is_c1 ) { // C1 codeblob?
address caller_pc=fr.pc();
if(NativeCall::is_call_before(caller_pc)){
address target = nativeCall_at(caller_pc)->destination();
is_C1_mid_patch = target == Runtime1::entry_for(Runtime1::load_klass_patching_id);
}
}
if( is_C1_mid_patch ) {
Untested("");
// Do NOT increment the BCP! We are re-executing the current bytecode.
} else if( ds->bci() == InvocationEntryBci ) {
// It is deopt while hanging on a method-entry lock.
// Do not advance BCP, as we have not executed bci 0 yet.
} else { // Else C2 or C1-not-mid-patch
// It is a deopt. Whether we re-execute the current bytecode or
// assume it has completed depends on the bytecode.
switch( c ) {
case Bytecodes::_lookupswitch:
case Bytecodes::_tableswitch:
case Bytecodes::_fast_binaryswitch:
case Bytecodes::_fast_linearswitch:
// recompute condtional expression folded into _if<cond>
case Bytecodes::_lcmp :
case Bytecodes::_fcmpl :
case Bytecodes::_fcmpg :
case Bytecodes::_dcmpl :
case Bytecodes::_dcmpg :
case Bytecodes::_ifnull :
case Bytecodes::_ifnonnull :
case Bytecodes::_goto :
case Bytecodes::_goto_w :
case Bytecodes::_ifeq :
case Bytecodes::_ifne :
case Bytecodes::_iflt :
case Bytecodes::_ifge :
case Bytecodes::_ifgt :
case Bytecodes::_ifle :
case Bytecodes::_if_icmpeq :
case Bytecodes::_if_icmpne :
case Bytecodes::_if_icmplt :
case Bytecodes::_if_icmpge :
case Bytecodes::_if_icmpgt :
case Bytecodes::_if_icmple :
case Bytecodes::_if_acmpeq :
case Bytecodes::_if_acmpne :
// special cases
case Bytecodes::_aastore:
// We are re-executing the current bytecode.
Untested("");
break;
// special cases
case Bytecodes::_putstatic:
case Bytecodes::_getstatic:
case Bytecodes::_getfield:
case Bytecodes::_putfield:
// We are re-executing the current bytecode.
break;
case Bytecodes::_athrow :
break; // Must be deopt-w-exception
case Bytecodes::_invokevirtual:
case Bytecodes::_invokespecial:
case Bytecodes::_invokestatic:{
methodHandle mh(thread,moop);
return_type=Bytecode_invoke_at(mh,bci)->result_type(thread);
if( !handle_popframe &&
!ds->should_reexecute())
bci_bump = 3; // Increment the BCP to post-call!!! See below!
break;
}
case Bytecodes::_invokeinterface:{
methodHandle mh(thread,moop);
return_type=Bytecode_invoke_at(mh,bci)->result_type(thread);
if( !handle_popframe &&
!ds->should_reexecute())
bci_bump = 5; // Increment the BCP to post-call!!! See below!
break;
}
case Bytecodes::_ldc :
Untested("");
return_type=constant_pool_type(moop,*(bcp+1));
if( !ds->should_reexecute()) bci_bump = 2; // Increment the BCP to post-call!!! See below!
break;
case Bytecodes::_ldc_w : // fall through
case Bytecodes::_ldc2_w:
return_type=constant_pool_type(moop,Bytes::get_Java_u2(bcp+1));
if( !ds->should_reexecute()) bci_bump = 3; // Increment the BCP to post-call!!! See below!
break;
default:
return_type=Bytecodes::result_type(c);
if( !ds->should_reexecute()) bci_bump = Bytecodes::length_at(bcp); // Increment the BCP to post-call!!! See below!
break;
}
if (ds->should_reexecute()) return_type = T_VOID;
}
// Save (possibly advanced) bci
buf->_bci = bci+bci_bump;
buf[-1]._retadr = Interpreter::unpack_and_go(); // Interpreter::return_entry(vtos, bci_bump);
}
// ---
// Now all the Java locals.
// First set the start of locals for the interpreter frame we are building.
buf->_loc = (intptr_t)jexstk;
uint loc_len = moop->max_locals();
for(uint i=0;i<loc_len;i++){
*jexstk++ = dm->get_value(ds->get_local(i),fr);
}
// Now that the locals have been unpacked if we have any deferred local writes
// added by jvmti then we can free up that structure as the data is now in the
// buffer
GrowableArray<jvmtiDeferredLocalVariableSet*>* list = thread->deferred_locals();
if( list ) {
// Because of inlining we could have multiple vframes for a single frame
// and several of the vframes could have deferred writes. Find them all.
Unimplemented();
}
// ---
// Now all the Java Expressions
uint expr_len = ds->numstk();
for(uint i=0;i<expr_len;i++)
*jexstk++ = dm->get_value(ds->get_expr(i),fr);
// If returning from a deoptimized call, we will have return values in
// registers that need to end up on the Java execution stack. They are
// not recorded in the debug info, since they did not exist at the time
// the call began.
if( is_youngest && is_deopt ) {
if( type2size[return_type] > 0 ) {
if( type2size[return_type]==2 ) {
*jexstk++ = (intptr_t)frame::double_slot_primitive_type_empty_slot_id << 32;
}
*jexstk++ = pd_fetch_return_values( thread, return_type );
// Need to adjust the final jexstk_top for the youngest frame
// returning values. These returned values are not accounted for in
// the standard debug info.
thread->_jexstk_top = jexstk;
}
}
// JVMTI PopFrame support
// Add the number of words of popframe preserved args to expr_len
int popframe_preserved_args_size_in_bytes = in_bytes(thread->popframe_preserved_args_size());
int popframe_preserved_args_size_in_words = in_words(thread->popframe_preserved_args_size_in_words());
if (handle_popframe) {
Unimplemented();
expr_len += popframe_preserved_args_size_in_words;
// 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.
}
// Set the JEX stk top
buf->_stk = (intptr_t)jexstk;
// ---
// Now move locked objects to the interpreters lock-stack.
// No need to inflate anything, as we're moving standard oops.
int numlcks = ds->numlocks();
if( ds->is_extra_lock() ) { // coarsened a lock
Untested("");
// The last lock is "coarsened" - kept locked when it should have been
// unlocked and relocked. With no deopt, keeping it locked saves the 2
// sets of back-to-back CAS's and fences. However, here we need to
// unlock it to match the proper Java state.
ObjectSynchronizer::unlock(ALWAYS_POISON_OBJECTREF((objectRef)dm->get_value(ds->get_lock(numlcks-1),fr)).as_oop());
numlcks--;
}
for(int i=0;i<numlcks;i++){
*lckstk++ = ALWAYS_POISON_OBJECTREF((objectRef)dm->get_value(ds->get_lock(i),fr));
}
} else { // Make a C1 frame
Unimplemented();
}
}
