//--------------------gen_stub------------------------------- void GraphKit::gen_stub(address C_function, const char *name, int is_fancy_jump, bool pass_tls, bool return_pc) { ResourceMark rm; const TypeTuple *jdomain = C->tf()->domain(); const TypeTuple *jrange = C->tf()->range(); // The procedure start StartNode* start = new (C) StartNode(root(), jdomain); _gvn.set_type_bottom(start); // Make a map, with JVM state uint parm_cnt = jdomain->cnt(); uint max_map = MAX2(2*parm_cnt+1, jrange->cnt()); // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces assert(SynchronizationEntryBCI == InvocationEntryBci, ""); JVMState* jvms = new (C) JVMState(0); jvms->set_bci(InvocationEntryBci); jvms->set_monoff(max_map); jvms->set_scloff(max_map); jvms->set_endoff(max_map); { SafePointNode *map = new (C) SafePointNode( max_map, jvms ); jvms->set_map(map); set_jvms(jvms); assert(map == this->map(), "kit.map is set"); } // Make up the parameters uint i; for( i = 0; i < parm_cnt; i++ ) map()->init_req(i, _gvn.transform(new (C) ParmNode(start, i))); for( ; i<map()->req(); i++ ) map()->init_req(i, top()); // For nicer debugging // GraphKit requires memory to be a MergeMemNode: set_all_memory(map()->memory()); // Get base of thread-local storage area Node* thread = _gvn.transform( new (C) ThreadLocalNode() ); const int NoAlias = Compile::AliasIdxBot; Node* adr_last_Java_pc = basic_plus_adr(top(), thread, in_bytes(JavaThread::frame_anchor_offset()) + in_bytes(JavaFrameAnchor::last_Java_pc_offset())); #if defined(SPARC) Node* adr_flags = basic_plus_adr(top(), thread, in_bytes(JavaThread::frame_anchor_offset()) + in_bytes(JavaFrameAnchor::flags_offset())); #endif /* defined(SPARC) */ // Drop in the last_Java_sp. last_Java_fp is not touched. // Always do this after the other "last_Java_frame" fields are set since // as soon as last_Java_sp != NULL the has_last_Java_frame is true and // users will look at the other fields. // Node *adr_sp = basic_plus_adr(top(), thread, in_bytes(JavaThread::last_Java_sp_offset())); Node *last_sp = basic_plus_adr(top(), frameptr(), (intptr_t) STACK_BIAS); store_to_memory(NULL, adr_sp, last_sp, T_ADDRESS, NoAlias); // Set _thread_in_native // The order of stores into TLS is critical! Setting _thread_in_native MUST // be last, because a GC is allowed at any time after setting it and the GC // will require last_Java_pc and last_Java_sp. Node* adr_state = basic_plus_adr(top(), thread, in_bytes(JavaThread::thread_state_offset())); //----------------------------- // Compute signature for C call. Varies from the Java signature! const Type **fields = TypeTuple::fields(2*parm_cnt+2); uint cnt = TypeFunc::Parms; // The C routines gets the base of thread-local storage passed in as an // extra argument. Not all calls need it, but its cheap to add here. for( ; cnt<parm_cnt; cnt++ ) fields[cnt] = jdomain->field_at(cnt); fields[cnt++] = TypeRawPtr::BOTTOM; // Thread-local storage // Also pass in the caller's PC, if asked for. if( return_pc ) fields[cnt++] = TypeRawPtr::BOTTOM; // Return PC const TypeTuple* domain = TypeTuple::make(cnt,fields); // The C routine we are about to call cannot return an oop; it can block on // exit and a GC will trash the oop while it sits in C-land. Instead, we // return the oop through TLS for runtime calls. // Also, C routines returning integer subword values leave the high // order bits dirty; these must be cleaned up by explicit sign extension. const Type* retval = (jrange->cnt() == TypeFunc::Parms) ? Type::TOP : jrange->field_at(TypeFunc::Parms); // Make a private copy of jrange->fields(); const Type **rfields = TypeTuple::fields(jrange->cnt() - TypeFunc::Parms); // Fixup oop returns int retval_ptr = retval->isa_oop_ptr(); if( retval_ptr ) { assert( pass_tls, "Oop must be returned thru TLS" ); // Fancy-jumps return address; others return void rfields[TypeFunc::Parms] = is_fancy_jump ? TypeRawPtr::BOTTOM : Type::TOP; } else if( retval->isa_int() ) { // Returning any integer subtype? // "Fatten" byte, char & short return types to 'int' to show that // the native C code can return values with junk high order bits. // We'll sign-extend it below later. rfields[TypeFunc::Parms] = TypeInt::INT; // It's "dirty" and needs sign-ext } else if( jrange->cnt() >= TypeFunc::Parms+1 ) { // Else copy other types rfields[TypeFunc::Parms] = jrange->field_at(TypeFunc::Parms); if( jrange->cnt() == TypeFunc::Parms+2 ) rfields[TypeFunc::Parms+1] = jrange->field_at(TypeFunc::Parms+1); } const TypeTuple* range = TypeTuple::make(jrange->cnt(),rfields); // Final C signature const TypeFunc *c_sig = TypeFunc::make(domain,range); //----------------------------- // Make the call node CallRuntimeNode *call = new (C) CallRuntimeNode(c_sig, C_function, name, TypePtr::BOTTOM); //----------------------------- // Fix-up the debug info for the call call->set_jvms( new (C) JVMState(0) ); call->jvms()->set_bci(0); call->jvms()->set_offsets(cnt); // Set fixed predefined input arguments cnt = 0; for( i=0; i<TypeFunc::Parms; i++ ) call->init_req( cnt++, map()->in(i) ); // A little too aggressive on the parm copy; return address is not an input call->set_req(TypeFunc::ReturnAdr, top()); for( ; i<parm_cnt; i++ ) // Regular input arguments call->init_req( cnt++, map()->in(i) ); call->init_req( cnt++, thread ); if( return_pc ) // Return PC, if asked for call->init_req( cnt++, returnadr() ); _gvn.transform_no_reclaim(call); //----------------------------- // Now set up the return results set_control( _gvn.transform( new (C) ProjNode(call,TypeFunc::Control)) ); set_i_o( _gvn.transform( new (C) ProjNode(call,TypeFunc::I_O )) ); set_all_memory_call(call); if (range->cnt() > TypeFunc::Parms) { Node* retnode = _gvn.transform( new (C) ProjNode(call,TypeFunc::Parms) ); // C-land is allowed to return sub-word values. Convert to integer type. assert( retval != Type::TOP, "" ); if (retval == TypeInt::BOOL) { retnode = _gvn.transform( new (C) AndINode(retnode, intcon(0xFF)) ); } else if (retval == TypeInt::CHAR) { retnode = _gvn.transform( new (C) AndINode(retnode, intcon(0xFFFF)) ); } else if (retval == TypeInt::BYTE) { retnode = _gvn.transform( new (C) LShiftINode(retnode, intcon(24)) ); retnode = _gvn.transform( new (C) RShiftINode(retnode, intcon(24)) ); } else if (retval == TypeInt::SHORT) { retnode = _gvn.transform( new (C) LShiftINode(retnode, intcon(16)) ); retnode = _gvn.transform( new (C) RShiftINode(retnode, intcon(16)) ); } map()->set_req( TypeFunc::Parms, retnode ); } //----------------------------- // Clear last_Java_sp store_to_memory(NULL, adr_sp, null(), T_ADDRESS, NoAlias); // Clear last_Java_pc and (optionally)_flags store_to_memory(NULL, adr_last_Java_pc, null(), T_ADDRESS, NoAlias); #if defined(SPARC) store_to_memory(NULL, adr_flags, intcon(0), T_INT, NoAlias); #endif /* defined(SPARC) */ #ifdef IA64 Node* adr_last_Java_fp = basic_plus_adr(top(), thread, in_bytes(JavaThread::last_Java_fp_offset())); if( os::is_MP() ) insert_mem_bar(Op_MemBarRelease); store_to_memory(NULL, adr_last_Java_fp, null(), T_ADDRESS, NoAlias); #endif // For is-fancy-jump, the C-return value is also the branch target Node* target = map()->in(TypeFunc::Parms); // Runtime call returning oop in TLS? Fetch it out if( pass_tls ) { Node* adr = basic_plus_adr(top(), thread, in_bytes(JavaThread::vm_result_offset())); Node* vm_result = make_load(NULL, adr, TypeOopPtr::BOTTOM, T_OBJECT, NoAlias, false); map()->set_req(TypeFunc::Parms, vm_result); // vm_result passed as result // clear thread-local-storage(tls) store_to_memory(NULL, adr, null(), T_ADDRESS, NoAlias); } //----------------------------- // check exception Node* adr = basic_plus_adr(top(), thread, in_bytes(Thread::pending_exception_offset())); Node* pending = make_load(NULL, adr, TypeOopPtr::BOTTOM, T_OBJECT, NoAlias, false); Node* exit_memory = reset_memory(); Node* cmp = _gvn.transform( new (C) CmpPNode(pending, null()) ); Node* bo = _gvn.transform( new (C) BoolNode(cmp, BoolTest::ne) ); IfNode *iff = create_and_map_if(control(), bo, PROB_MIN, COUNT_UNKNOWN); Node* if_null = _gvn.transform( new (C) IfFalseNode(iff) ); Node* if_not_null = _gvn.transform( new (C) IfTrueNode(iff) ); assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before"); Node *exc_target = makecon(TypeRawPtr::make( StubRoutines::forward_exception_entry() )); Node *to_exc = new (C) TailCallNode(if_not_null, i_o(), exit_memory, frameptr(), returnadr(), exc_target, null()); root()->add_req(_gvn.transform(to_exc)); // bind to root to keep live C->init_start(start); //----------------------------- // If this is a normal subroutine return, issue the return and be done. Node *ret; switch( is_fancy_jump ) { case 0: // Make a return instruction // Return to caller, free any space for return address ret = new (C) ReturnNode(TypeFunc::Parms, if_null, i_o(), exit_memory, frameptr(), returnadr()); if (C->tf()->range()->cnt() > TypeFunc::Parms) ret->add_req( map()->in(TypeFunc::Parms) ); break; case 1: // This is a fancy tail-call jump. Jump to computed address. // Jump to new callee; leave old return address alone. ret = new (C) TailCallNode(if_null, i_o(), exit_memory, frameptr(), returnadr(), target, map()->in(TypeFunc::Parms)); break; case 2: // Pop return address & jump // Throw away old return address; jump to new computed address //assert(C_function == CAST_FROM_FN_PTR(address, OptoRuntime::rethrow_C), "fancy_jump==2 only for rethrow"); ret = new (C) TailJumpNode(if_null, i_o(), exit_memory, frameptr(), target, map()->in(TypeFunc::Parms)); break; default: ShouldNotReachHere(); } root()->add_req(_gvn.transform(ret)); }
//------------------------------do_call---------------------------------------- // Handle your basic call. Inline if we can & want to, else just setup call. void Parse::do_call() { // It's likely we are going to add debug info soon. // Also, if we inline a guy who eventually needs debug info for this JVMS, // our contribution to it is cleaned up right here. kill_dead_locals(); // Set frequently used booleans bool is_virtual = bc() == Bytecodes::_invokevirtual; bool is_virtual_or_interface = is_virtual || bc() == Bytecodes::_invokeinterface; bool has_receiver = is_virtual_or_interface || bc() == Bytecodes::_invokespecial; // Find target being called bool will_link; ciMethod* dest_method = iter().get_method(will_link); ciInstanceKlass* holder_klass = dest_method->holder(); ciKlass* holder = iter().get_declared_method_holder(); ciInstanceKlass* klass = ciEnv::get_instance_klass_for_declared_method_holder(holder); int nargs = dest_method->arg_size(); // See if the receiver (if any) is NULL, hence we always throw BEFORE // attempting to resolve the call or initialize the holder class. Doing so // out of order opens a window where we can endlessly deopt because the call // holder is not initialized, but the call never actually happens (forcing // class initialization) because we only see NULL receivers. CPData_Invoke *caller_cpdi = cpdata()->as_Invoke(bc()); debug_only( assert(caller_cpdi->is_Invoke(), "Not invoke!") ); if( is_virtual_or_interface && _gvn.type(stack(sp() - nargs))->higher_equal(TypePtr::NULL_PTR) ) { builtin_throw( Deoptimization::Reason_null_check, "null receiver", caller_cpdi, caller_cpdi->saw_null(), /*must_throw=*/true ); return; } // uncommon-trap when callee is unloaded, uninitialized or will not link // bailout when too many arguments for register representation if (!will_link || can_not_compile_call_site(dest_method, klass)) { return; } assert(FAM||holder_klass->is_loaded(),""); assert(dest_method->is_static() == !has_receiver, "must match bc"); // Note: this takes into account invokeinterface of methods declared in java/lang/Object, // which should be invokevirtuals but according to the VM spec may be invokeinterfaces assert(holder_klass->is_interface() || holder_klass->super() == NULL || (bc() != Bytecodes::_invokeinterface), "must match bc"); // Note: In the absence of miranda methods, an abstract class K can perform // an invokevirtual directly on an interface method I.m if K implements I. // --------------------- // Does Class Hierarchy Analysis reveal only a single target of a v-call? // Then we may inline or make a static call, but become dependent on there being only 1 target. // Does the call-site type profile reveal only one receiver? // Then we may introduce a run-time check and inline on the path where it succeeds. // The other path may uncommon_trap, check for another receiver, or do a v-call. // Choose call strategy. bool call_is_virtual = is_virtual_or_interface; int vtable_index = methodOopDesc::invalid_vtable_index; ciMethod* call_method = dest_method; // Try to get the most accurate receiver type if (is_virtual_or_interface) { Node* receiver_node = stack(sp() - nargs); const TypeInstPtr*inst_type=_gvn.type(receiver_node)->isa_instptr(); if( inst_type ) { ciInstanceKlass*ikl=inst_type->klass()->as_instance_klass(); // If the receiver is not yet linked then: (1) we never can make this // call because no objects can be created until linkage, and (2) CHA // reports incorrect answers... so do not bother with making the call // until after the klass gets linked. ciInstanceKlass *ikl2 = ikl->is_subtype_of(klass) ? ikl : klass; if(!ikl->is_linked()){ uncommon_trap(Deoptimization::Reason_uninitialized,klass,"call site where receiver is not linked",false); return; } } const TypeOopPtr* receiver_type = _gvn.type(receiver_node)->isa_oopptr(); ciMethod* optimized_virtual_method = optimize_inlining(method(), bci(), klass, dest_method, receiver_type); // Have the call been sufficiently improved such that it is no longer a virtual? if (optimized_virtual_method != NULL) { call_method = optimized_virtual_method; call_is_virtual = false; } else if (false) { // We can make a vtable call at this site vtable_index = call_method->resolve_vtable_index(method()->holder(), klass); } } // Note: It's OK to try to inline a virtual call. // The call generator will not attempt to inline a polymorphic call // unless it knows how to optimize the receiver dispatch. bool try_inline=(C->do_inlining()||InlineAccessors)&& (!C->method()->should_disable_inlining()) && (call_method->number_of_breakpoints() == 0); // Get profile data for the *callee*. First see if we have precise // CodeProfile for this exact inline because C1 inlined it already. CodeProfile *callee_cp; int callee_cp_inloff; if( caller_cpdi->inlined_method_oid() == call_method->objectId() ) { callee_cp = c1_cp(); // Use same CodeProfile as current callee_cp_inloff = caller_cpdi->cpd_offset(); // But use inlined portion } else { // If callee has a cp, clone it and use callee_cp = call_method->codeprofile(true); callee_cp_inloff = 0; if (callee_cp || FAM) { // The cloned cp needs to be freed later Compile* C = Compile::current(); C->record_cloned_cp(callee_cp); } else { // Had profile info at top level, but not for this call site? // callee_cp will hold the just created cp, or whatever cp allocated by // other thread which wins the race in set_codeprofile callee_cp = call_method->set_codeprofile(CodeProfile::make(call_method)); } } CPData_Invoke *c2_caller_cpdi = UseC1 ? c2cpdata()->as_Invoke(bc()) : NULL; // --------------------- inc_sp(- nargs); // Temporarily pop args for JVM state of call JVMState* jvms = sync_jvms(); // --------------------- // Decide call tactic. // This call checks with CHA, the interpreter profile, intrinsics table, etc. // It decides whether inlining is desirable or not. CallGenerator*cg=C->call_generator(call_method,vtable_index,call_is_virtual,jvms,try_inline,prof_factor(),callee_cp,callee_cp_inloff,c2_caller_cpdi,caller_cpdi); // --------------------- // Round double arguments before call round_double_arguments(dest_method); #ifndef PRODUCT // Record first part of parsing work for this call parse_histogram()->record_change(); #endif // not PRODUCT assert(jvms == this->jvms(), "still operating on the right JVMS"); assert(jvms_in_sync(), "jvms must carry full info into CG"); // save across call, for a subsequent cast_not_null. Node* receiver = has_receiver ? argument(0) : NULL; JVMState* new_jvms = cg->generate(jvms, caller_cpdi, is_private_copy()); if( new_jvms == NULL ) { // Did it work? // When inlining attempt fails (e.g., too many arguments), // it may contaminate the current compile state, making it // impossible to pull back and try again. Once we call // cg->generate(), we are committed. If it fails, the whole // compilation task is compromised. if (failing()) return; if (PrintOpto || PrintInlining || PrintC2Inlining) { // Only one fall-back, so if an intrinsic fails, ignore any bytecodes. if (cg->is_intrinsic() && call_method->code_size() > 0) { C2OUT->print("Bailed out of intrinsic, will not inline: "); call_method->print_name(C2OUT); C2OUT->cr(); } } // This can happen if a library intrinsic is available, but refuses // the call site, perhaps because it did not match a pattern the // intrinsic was expecting to optimize. The fallback position is // to call out-of-line. try_inline = false; // Inline tactic bailed out. cg=C->call_generator(call_method,vtable_index,call_is_virtual,jvms,try_inline,prof_factor(),c1_cp(),c1_cp_inloff(),c2_caller_cpdi,caller_cpdi); new_jvms=cg->generate(jvms,caller_cpdi,is_private_copy()); assert(new_jvms!=NULL,"call failed to generate: calls should work"); if (c2_caller_cpdi) c2_caller_cpdi->_inlining_failure_id = IF_GENERALFAILURE; } if (cg->is_inline()) { C->env()->notice_inlined_method(call_method); } // Reset parser state from [new_]jvms, which now carries results of the call. // Return value (if any) is already pushed on the stack by the cg. add_exception_states_from(new_jvms); if (new_jvms->map()->control() == top()) { stop_and_kill_map(); } else { assert(new_jvms->same_calls_as(jvms), "method/bci left unchanged"); set_jvms(new_jvms); } if (!stopped()) { // This was some sort of virtual call, which did a null check for us. // Now we can assert receiver-not-null, on the normal return path. if (receiver != NULL && cg->is_virtual()) { Node*cast=cast_not_null(receiver,true); // %%% assert(receiver == cast, "should already have cast the receiver"); } // Round double result after a call from strict to non-strict code round_double_result(dest_method); // If the return type of the method is not loaded, assert that the // value we got is a null. Otherwise, we need to recompile. if (!dest_method->return_type()->is_loaded()) { // If there is going to be a trap, put it at the next bytecode: set_bci(iter().next_bci()); do_null_assert(peek(), T_OBJECT); set_bci(iter().cur_bci()); // put it back } else { assert0( call_method->return_type()->is_loaded() ); BasicType result_type = dest_method->return_type()->basic_type(); if(result_type==T_OBJECT||result_type==T_ARRAY){ const Type *t = peek()->bottom_type(); assert0( t == TypePtr::NULL_PTR || t->is_oopptr()->klass()->is_loaded() ); } } } // Restart record of parsing work after possible inlining of call #ifndef PRODUCT parse_histogram()->set_initial_state(bc()); #endif }