//---------------------------clone_deep---------------------------------------- JVMState* JVMState::clone_deep() const { JVMState* n = clone_shallow(); for (JVMState* p = n; p->_caller != NULL; p = p->_caller) { p->_caller = p->_caller->clone_shallow(); } assert(n->depth() == depth(), "sanity"); assert(n->debug_depth() == debug_depth(), "sanity"); return n; }
//--------------------------clone_shallow-------------------------------------- JVMState* JVMState::clone_shallow() const { JVMState* n = has_method() ? new JVMState(_method, _caller) : new JVMState(0); n->set_bci(_bci); n->set_locoff(_locoff); n->set_stkoff(_stkoff); n->set_monoff(_monoff); n->set_endoff(_endoff); n->set_sp(_sp); n->set_map(_map); return n; }
//------------------------------verify_base_ptrs------------------------------- // Verify that base pointers and derived pointers are still sane. // Basically, if a derived pointer is live at a safepoint, then its // base pointer must be live also. void PhaseChaitin::verify_base_ptrs( ResourceArea *a ) const { for( uint i = 0; i < _cfg._num_blocks; i++ ) { Block *b = _cfg._blocks[i]; for( uint j = b->end_idx() + 1; j > 1; j-- ) { Node *n = b->_nodes[j-1]; if( n->is_Phi() ) break; MachNode *mach = n->is_Mach(); MachSafePointNode *sfpt = (mach != NULL) ? mach->is_MachSafePoint() : NULL; // Found a safepoint? if( sfpt != NULL ) { JVMState* jvms = sfpt->jvms(); if (jvms != NULL) { // Now scan for a live derived pointer if (jvms->oopoff() < sfpt->req()) { // Check each derived/base pair for (uint idx = jvms->oopoff(); idx < sfpt->req(); idx += 2) { Node *check = sfpt->in(idx); uint j = 0; // search upwards through spills and spill phis for AddP while(true) { if( !check ) break; int idx = check->is_Copy(); if( idx ) { check = check->in(idx); } else if( check->is_Phi() && check->_idx >= _oldphi ) { check = check->in(1); } else break; j++; assert(j < 100000,"Derived pointer checking in infinite loop"); } // End while MachNode *machcheck = check->is_Mach(); assert(machcheck && machcheck->ideal_Opcode() == Op_AddP,"Bad derived pointer") } } // End of check for derived pointers } // End of Kcheck for debug info } // End of if found a safepoint } // End of forall instructions in block } // End of forall blocks
//-----------------------------try_to_inline----------------------------------- // return true if ok // Relocated from "InliningClosure::try_to_inline" bool InlineTree::try_to_inline(ciMethod* callee_method, ciMethod* caller_method, int caller_bci, JVMState* jvms, ciCallProfile& profile, WarmCallInfo* wci_result, bool& should_delay) { if (ClipInlining && (int)count_inline_bcs() >= DesiredMethodLimit) { if (!callee_method->force_inline() || !IncrementalInline) { set_msg("size > DesiredMethodLimit"); return false; } else if (!C->inlining_incrementally()) { should_delay = true; } } _forced_inline = false; // Reset if (!should_inline(callee_method, caller_method, caller_bci, profile, wci_result)) { return false; } if (should_not_inline(callee_method, caller_method, jvms, wci_result)) { return false; } if (InlineAccessors && callee_method->is_accessor()) { // accessor methods are not subject to any of the following limits. set_msg("accessor"); return true; } // suppress a few checks for accessors and trivial methods if (callee_method->code_size() > MaxTrivialSize) { // don't inline into giant methods if (C->over_inlining_cutoff()) { if ((!callee_method->force_inline() && !caller_method->is_compiled_lambda_form()) || !IncrementalInline) { set_msg("NodeCountInliningCutoff"); return false; } else { should_delay = true; } } if ((!UseInterpreter || CompileTheWorld) && is_init_with_ea(callee_method, caller_method, C)) { // Escape Analysis stress testing when running Xcomp or CTW: // inline constructors even if they are not reached. } else if (forced_inline()) { // Inlining was forced by CompilerOracle, ciReplay or annotation } else if (profile.count() == 0) { // don't inline unreached call sites set_msg("call site not reached"); return false; } } if (!C->do_inlining() && InlineAccessors) { set_msg("not an accessor"); return false; } // Limit inlining depth in case inlining is forced or // _max_inline_level was increased to compensate for lambda forms. if (inline_level() > MaxForceInlineLevel) { set_msg("MaxForceInlineLevel"); return false; } if (inline_level() > _max_inline_level) { if (!callee_method->force_inline() || !IncrementalInline) { set_msg("inlining too deep"); return false; } else if (!C->inlining_incrementally()) { should_delay = true; } } // detect direct and indirect recursive inlining { // count the current method and the callee const bool is_compiled_lambda_form = callee_method->is_compiled_lambda_form(); int inline_level = 0; if (!is_compiled_lambda_form) { if (method() == callee_method) { inline_level++; } } // count callers of current method and callee Node* callee_argument0 = is_compiled_lambda_form ? jvms->map()->argument(jvms, 0)->uncast() : NULL; for (JVMState* j = jvms->caller(); j != NULL && j->has_method(); j = j->caller()) { if (j->method() == callee_method) { if (is_compiled_lambda_form) { // Since compiled lambda forms are heavily reused we allow recursive inlining. If it is truly // a recursion (using the same "receiver") we limit inlining otherwise we can easily blow the // compiler stack. Node* caller_argument0 = j->map()->argument(j, 0)->uncast(); if (caller_argument0 == callee_argument0) { inline_level++; } } else { inline_level++; } } } if (inline_level > MaxRecursiveInlineLevel) { set_msg("recursive inlining is too deep"); return false; } } int size = callee_method->code_size_for_inlining(); if (ClipInlining && (int)count_inline_bcs() + size >= DesiredMethodLimit) { if (!callee_method->force_inline() || !IncrementalInline) { set_msg("size > DesiredMethodLimit"); return false; } else if (!C->inlining_incrementally()) { should_delay = true; } } // ok, inline this method return true; }
//--------------------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)); }
void IdealGraphPrinter::visit_node(Node *n, bool edges, VectorSet* temp_set) { if (edges) { // Output edge node_idx_t dest_id = n->_idx; for ( uint i = 0; i < n->len(); i++ ) { if ( n->in(i) ) { Node *source = n->in(i); begin_elem(EDGE_ELEMENT); print_attr(FROM_PROPERTY, source->_idx); print_attr(TO_PROPERTY, dest_id); print_attr(INDEX_PROPERTY, i); end_elem(); } } } else { // Output node begin_head(NODE_ELEMENT); print_attr(NODE_ID_PROPERTY, n->_idx); end_head(); head(PROPERTIES_ELEMENT); Node *node = n; #ifndef PRODUCT Compile::current()->_in_dump_cnt++; print_prop(NODE_NAME_PROPERTY, (const char *)node->Name()); const Type *t = node->bottom_type(); print_prop("type", t->msg()); print_prop("idx", node->_idx); #ifdef ASSERT print_prop("debug_idx", node->_debug_idx); #endif if (C->cfg() != NULL) { Block* block = C->cfg()->get_block_for_node(node); if (block == NULL) { print_prop("block", C->cfg()->get_block(0)->_pre_order); } else { print_prop("block", block->_pre_order); } } const jushort flags = node->flags(); if (flags & Node::Flag_is_Copy) { print_prop("is_copy", "true"); } if (flags & Node::Flag_rematerialize) { print_prop("rematerialize", "true"); } if (flags & Node::Flag_needs_anti_dependence_check) { print_prop("needs_anti_dependence_check", "true"); } if (flags & Node::Flag_is_macro) { print_prop("is_macro", "true"); } if (flags & Node::Flag_is_Con) { print_prop("is_con", "true"); } if (flags & Node::Flag_is_cisc_alternate) { print_prop("is_cisc_alternate", "true"); } if (flags & Node::Flag_is_dead_loop_safe) { print_prop("is_dead_loop_safe", "true"); } if (flags & Node::Flag_may_be_short_branch) { print_prop("may_be_short_branch", "true"); } if (flags & Node::Flag_has_call) { print_prop("has_call", "true"); } if (C->matcher() != NULL) { if (C->matcher()->is_shared(node)) { print_prop("is_shared", "true"); } else { print_prop("is_shared", "false"); } if (C->matcher()->is_dontcare(node)) { print_prop("is_dontcare", "true"); } else { print_prop("is_dontcare", "false"); } #ifdef ASSERT Node* old = C->matcher()->find_old_node(node); if (old != NULL) { print_prop("old_node_idx", old->_idx); } #endif } if (node->is_Proj()) { print_prop("con", (int)node->as_Proj()->_con); } if (node->is_Mach()) { print_prop("idealOpcode", (const char *)NodeClassNames[node->as_Mach()->ideal_Opcode()]); } buffer[0] = 0; stringStream s2(buffer, sizeof(buffer) - 1); node->dump_spec(&s2); if (t != NULL && (t->isa_instptr() || t->isa_klassptr())) { const TypeInstPtr *toop = t->isa_instptr(); const TypeKlassPtr *tkls = t->isa_klassptr(); ciKlass* klass = toop ? toop->klass() : (tkls ? tkls->klass() : NULL ); if( klass && klass->is_loaded() && klass->is_interface() ) { s2.print(" Interface:"); } else if( toop ) { s2.print(" Oop:"); } else if( tkls ) { s2.print(" Klass:"); } t->dump_on(&s2); } else if( t == Type::MEMORY ) { s2.print(" Memory:"); MemNode::dump_adr_type(node, node->adr_type(), &s2); } assert(s2.size() < sizeof(buffer), "size in range"); print_prop("dump_spec", buffer); if (node->is_block_proj()) { print_prop("is_block_proj", "true"); } if (node->is_block_start()) { print_prop("is_block_start", "true"); } const char *short_name = "short_name"; if (strcmp(node->Name(), "Parm") == 0 && node->as_Proj()->_con >= TypeFunc::Parms) { int index = node->as_Proj()->_con - TypeFunc::Parms; if (index >= 10) { print_prop(short_name, "PA"); } else { sprintf(buffer, "P%d", index); print_prop(short_name, buffer); } } else if (strcmp(node->Name(), "IfTrue") == 0) { print_prop(short_name, "T"); } else if (strcmp(node->Name(), "IfFalse") == 0) { print_prop(short_name, "F"); } else if ((node->is_Con() && node->is_Type()) || node->is_Proj()) { if (t->base() == Type::Int && t->is_int()->is_con()) { const TypeInt *typeInt = t->is_int(); assert(typeInt->is_con(), "must be constant"); jint value = typeInt->get_con(); // max. 2 chars allowed if (value >= -9 && value <= 99) { sprintf(buffer, "%d", value); print_prop(short_name, buffer); } else { print_prop(short_name, "I"); } } else if (t == Type::TOP) { print_prop(short_name, "^"); } else if (t->base() == Type::Long && t->is_long()->is_con()) { const TypeLong *typeLong = t->is_long(); assert(typeLong->is_con(), "must be constant"); jlong value = typeLong->get_con(); // max. 2 chars allowed if (value >= -9 && value <= 99) { sprintf(buffer, JLONG_FORMAT, value); print_prop(short_name, buffer); } else { print_prop(short_name, "L"); } } else if (t->base() == Type::KlassPtr) { const TypeKlassPtr *typeKlass = t->is_klassptr(); print_prop(short_name, "CP"); } else if (t->base() == Type::Control) { print_prop(short_name, "C"); } else if (t->base() == Type::Memory) { print_prop(short_name, "M"); } else if (t->base() == Type::Abio) { print_prop(short_name, "IO"); } else if (t->base() == Type::Return_Address) { print_prop(short_name, "RA"); } else if (t->base() == Type::AnyPtr) { print_prop(short_name, "P"); } else if (t->base() == Type::RawPtr) { print_prop(short_name, "RP"); } else if (t->base() == Type::AryPtr) { print_prop(short_name, "AP"); } } JVMState* caller = NULL; if (node->is_SafePoint()) { caller = node->as_SafePoint()->jvms(); } else { Node_Notes* notes = C->node_notes_at(node->_idx); if (notes != NULL) { caller = notes->jvms(); } } if (caller != NULL) { stringStream bciStream; ciMethod* last = NULL; int last_bci; while(caller) { if (caller->has_method()) { last = caller->method(); last_bci = caller->bci(); } bciStream.print("%d ", caller->bci()); caller = caller->caller(); } print_prop("bci", bciStream.as_string()); if (last != NULL && last->has_linenumber_table() && last_bci >= 0) { print_prop("line", last->line_number_from_bci(last_bci)); } } #ifdef ASSERT if (node->debug_orig() != NULL) { temp_set->Clear(); stringStream dorigStream; Node* dorig = node->debug_orig(); while (dorig && temp_set->test_set(dorig->_idx)) { dorigStream.print("%d ", dorig->_idx); } print_prop("debug_orig", dorigStream.as_string()); } #endif if (_chaitin && _chaitin != (PhaseChaitin *)0xdeadbeef) { buffer[0] = 0; _chaitin->dump_register(node, buffer); print_prop("reg", buffer); uint lrg_id = 0; if (node->_idx < _chaitin->_lrg_map.size()) { lrg_id = _chaitin->_lrg_map.live_range_id(node); } print_prop("lrg", lrg_id); } Compile::current()->_in_dump_cnt--; #endif tail(PROPERTIES_ELEMENT); tail(NODE_ELEMENT); } }
void LateInlineCallGenerator::do_late_inline() { // Can't inline it CallStaticJavaNode* call = call_node(); if (call == NULL || call->outcnt() == 0 || call->in(0) == NULL || call->in(0)->is_top()) { return; } const TypeTuple *r = call->tf()->domain(); for (int i1 = 0; i1 < method()->arg_size(); i1++) { if (call->in(TypeFunc::Parms + i1)->is_top() && r->field_at(TypeFunc::Parms + i1) != Type::HALF) { assert(Compile::current()->inlining_incrementally(), "shouldn't happen during parsing"); return; } } if (call->in(TypeFunc::Memory)->is_top()) { assert(Compile::current()->inlining_incrementally(), "shouldn't happen during parsing"); return; } Compile* C = Compile::current(); // Remove inlined methods from Compiler's lists. if (call->is_macro()) { C->remove_macro_node(call); } // Make a clone of the JVMState that appropriate to use for driving a parse JVMState* old_jvms = call->jvms(); JVMState* jvms = old_jvms->clone_shallow(C); uint size = call->req(); SafePointNode* map = new (C) SafePointNode(size, jvms); for (uint i1 = 0; i1 < size; i1++) { map->init_req(i1, call->in(i1)); } // Make sure the state is a MergeMem for parsing. if (!map->in(TypeFunc::Memory)->is_MergeMem()) { Node* mem = MergeMemNode::make(C, map->in(TypeFunc::Memory)); C->initial_gvn()->set_type_bottom(mem); map->set_req(TypeFunc::Memory, mem); } uint nargs = method()->arg_size(); // blow away old call arguments Node* top = C->top(); for (uint i1 = 0; i1 < nargs; i1++) { map->set_req(TypeFunc::Parms + i1, top); } jvms->set_map(map); // Make enough space in the expression stack to transfer // the incoming arguments and return value. map->ensure_stack(jvms, jvms->method()->max_stack()); for (uint i1 = 0; i1 < nargs; i1++) { map->set_argument(jvms, i1, call->in(TypeFunc::Parms + i1)); } // This check is done here because for_method_handle_inline() method // needs jvms for inlined state. if (!do_late_inline_check(jvms)) { map->disconnect_inputs(NULL, C); return; } C->print_inlining_insert(this); CompileLog* log = C->log(); if (log != NULL) { log->head("late_inline method='%d'", log->identify(method())); JVMState* p = jvms; while (p != NULL) { log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); p = p->caller(); } log->tail("late_inline"); } // Setup default node notes to be picked up by the inlining Node_Notes* old_nn = C->default_node_notes(); if (old_nn != NULL) { Node_Notes* entry_nn = old_nn->clone(C); entry_nn->set_jvms(jvms); C->set_default_node_notes(entry_nn); } // Now perform the inling using the synthesized JVMState JVMState* new_jvms = _inline_cg->generate(jvms, NULL); if (new_jvms == NULL) return; // no change if (C->failing()) return; // Capture any exceptional control flow GraphKit kit(new_jvms); // Find the result object Node* result = C->top(); int result_size = method()->return_type()->size(); if (result_size != 0 && !kit.stopped()) { result = (result_size == 1) ? kit.pop() : kit.pop_pair(); } C->set_has_loops(C->has_loops() || _inline_cg->method()->has_loops()); C->env()->notice_inlined_method(_inline_cg->method()); C->set_inlining_progress(true); kit.replace_call(call, result); }
JVMState* PredicatedIntrinsicGenerator::generate(JVMState* jvms) { // The code we want to generate here is: // if (receiver == NULL) // uncommon_Trap // if (predicate(0)) // do_intrinsic(0) // else // if (predicate(1)) // do_intrinsic(1) // ... // else // do_java_comp GraphKit kit(jvms); PhaseGVN& gvn = kit.gvn(); CompileLog* log = kit.C->log(); if (log != NULL) { log->elem("predicated_intrinsic bci='%d' method='%d'", jvms->bci(), log->identify(method())); } if (!method()->is_static()) { // We need an explicit receiver null_check before checking its type in predicate. // We share a map with the caller, so his JVMS gets adjusted. Node* receiver = kit.null_check_receiver_before_call(method()); if (kit.stopped()) { return kit.transfer_exceptions_into_jvms(); } } int n_predicates = _intrinsic->predicates_count(); assert(n_predicates > 0, "sanity"); JVMState** result_jvms = NEW_RESOURCE_ARRAY(JVMState*, (n_predicates+1)); // Region for normal compilation code if intrinsic failed. Node* slow_region = new (kit.C) RegionNode(1); int results = 0; for (int predicate = 0; (predicate < n_predicates) && !kit.stopped(); predicate++) { #ifdef ASSERT JVMState* old_jvms = kit.jvms(); SafePointNode* old_map = kit.map(); Node* old_io = old_map->i_o(); Node* old_mem = old_map->memory(); Node* old_exc = old_map->next_exception(); #endif Node* else_ctrl = _intrinsic->generate_predicate(kit.sync_jvms(), predicate); #ifdef ASSERT // Assert(no_new_memory && no_new_io && no_new_exceptions) after generate_predicate. assert(old_jvms == kit.jvms(), "generate_predicate should not change jvm state"); SafePointNode* new_map = kit.map(); assert(old_io == new_map->i_o(), "generate_predicate should not change i_o"); assert(old_mem == new_map->memory(), "generate_predicate should not change memory"); assert(old_exc == new_map->next_exception(), "generate_predicate should not add exceptions"); #endif if (!kit.stopped()) { PreserveJVMState pjvms(&kit); // Generate intrinsic code: JVMState* new_jvms = _intrinsic->generate(kit.sync_jvms()); if (new_jvms == NULL) { // Intrinsic failed, use normal compilation path for this predicate. slow_region->add_req(kit.control()); } else { kit.add_exception_states_from(new_jvms); kit.set_jvms(new_jvms); if (!kit.stopped()) { result_jvms[results++] = kit.jvms(); } } } if (else_ctrl == NULL) { else_ctrl = kit.C->top(); } kit.set_control(else_ctrl); } if (!kit.stopped()) { // Final 'else' after predicates. slow_region->add_req(kit.control()); } if (slow_region->req() > 1) { PreserveJVMState pjvms(&kit); // Generate normal compilation code: kit.set_control(gvn.transform(slow_region)); JVMState* new_jvms = _cg->generate(kit.sync_jvms()); if (kit.failing()) return NULL; // might happen because of NodeCountInliningCutoff assert(new_jvms != NULL, "must be"); kit.add_exception_states_from(new_jvms); kit.set_jvms(new_jvms); if (!kit.stopped()) { result_jvms[results++] = kit.jvms(); } } if (results == 0) { // All paths ended in uncommon traps. (void) kit.stop(); return kit.transfer_exceptions_into_jvms(); } if (results == 1) { // Only one path kit.set_jvms(result_jvms[0]); return kit.transfer_exceptions_into_jvms(); } // Merge all paths. kit.C->set_has_split_ifs(true); // Has chance for split-if optimization RegionNode* region = new (kit.C) RegionNode(results + 1); Node* iophi = PhiNode::make(region, kit.i_o(), Type::ABIO); for (int i = 0; i < results; i++) { JVMState* jvms = result_jvms[i]; int path = i + 1; SafePointNode* map = jvms->map(); region->init_req(path, map->control()); iophi->set_req(path, map->i_o()); if (i == 0) { kit.set_jvms(jvms); } else { kit.merge_memory(map->merged_memory(), region, path); } } kit.set_control(gvn.transform(region)); kit.set_i_o(gvn.transform(iophi)); // Transform new memory Phis. for (MergeMemStream mms(kit.merged_memory()); mms.next_non_empty();) { Node* phi = mms.memory(); if (phi->is_Phi() && phi->in(0) == region) { mms.set_memory(gvn.transform(phi)); } } // Merge debug info. Node** ins = NEW_RESOURCE_ARRAY(Node*, results); uint tos = kit.jvms()->stkoff() + kit.sp(); Node* map = kit.map(); uint limit = map->req(); for (uint i = TypeFunc::Parms; i < limit; i++) { // Skip unused stack slots; fast forward to monoff(); if (i == tos) { i = kit.jvms()->monoff(); if( i >= limit ) break; } Node* n = map->in(i); ins[0] = n; const Type* t = gvn.type(n); bool needs_phi = false; for (int j = 1; j < results; j++) { JVMState* jvms = result_jvms[j]; Node* jmap = jvms->map(); Node* m = NULL; if (jmap->req() > i) { m = jmap->in(i); if (m != n) { needs_phi = true; t = t->meet_speculative(gvn.type(m)); } } ins[j] = m; } if (needs_phi) { Node* phi = PhiNode::make(region, n, t); for (int j = 1; j < results; j++) { phi->set_req(j + 1, ins[j]); } map->set_req(i, gvn.transform(phi)); } } return kit.transfer_exceptions_into_jvms(); }
//------------------------------insert_copies---------------------------------- void PhaseAggressiveCoalesce::insert_copies( Matcher &matcher ) { // We do LRGs compressing and fix a liveout data only here since the other // place in Split() is guarded by the assert which we never hit. _phc.compress_uf_map_for_nodes(); // Fix block's liveout data for compressed live ranges. for(uint lrg = 1; lrg < _phc._maxlrg; lrg++ ) { uint compressed_lrg = _phc.Find(lrg); if( lrg != compressed_lrg ) { for( uint bidx = 0; bidx < _phc._cfg._num_blocks; bidx++ ) { IndexSet *liveout = _phc._live->live(_phc._cfg._blocks[bidx]); if( liveout->member(lrg) ) { liveout->remove(lrg); liveout->insert(compressed_lrg); } } } } // All new nodes added are actual copies to replace virtual copies. // Nodes with index less than '_unique' are original, non-virtual Nodes. _unique = C->unique(); for( uint i=0; i<_phc._cfg._num_blocks; i++ ) { Block *b = _phc._cfg._blocks[i]; uint cnt = b->num_preds(); // Number of inputs to the Phi for( uint l = 1; l<b->_nodes.size(); l++ ) { Node *n = b->_nodes[l]; // Do not use removed-copies, use copied value instead uint ncnt = n->req(); for( uint k = 1; k<ncnt; k++ ) { Node *copy = n->in(k); uint cidx = copy->is_Copy(); if( cidx ) { Node *def = copy->in(cidx); if( _phc.Find(copy) == _phc.Find(def) ) n->set_req(k,def); } } // Remove any explicit copies that get coalesced. uint cidx = n->is_Copy(); if( cidx ) { Node *def = n->in(cidx); if( _phc.Find(n) == _phc.Find(def) ) { n->replace_by(def); n->set_req(cidx,NULL); b->_nodes.remove(l); l--; continue; } } if( n->is_Phi() ) { // Get the chosen name for the Phi uint phi_name = _phc.Find( n ); // Ignore the pre-allocated specials if( !phi_name ) continue; // Check for mismatch inputs to Phi for( uint j = 1; j<cnt; j++ ) { Node *m = n->in(j); uint src_name = _phc.Find(m); if( src_name != phi_name ) { Block *pred = _phc._cfg._bbs[b->pred(j)->_idx]; Node *copy; assert(!m->is_Con() || m->is_Mach(), "all Con must be Mach"); // Rematerialize constants instead of copying them if( m->is_Mach() && m->as_Mach()->is_Con() && m->as_Mach()->rematerialize() ) { copy = m->clone(); // Insert the copy in the predecessor basic block pred->add_inst(copy); // Copy any flags as well _phc.clone_projs( pred, pred->end_idx(), m, copy, _phc._maxlrg ); } else { const RegMask *rm = C->matcher()->idealreg2spillmask[m->ideal_reg()]; copy = new (C) MachSpillCopyNode(m,*rm,*rm); // Find a good place to insert. Kinda tricky, use a subroutine insert_copy_with_overlap(pred,copy,phi_name,src_name); } // Insert the copy in the use-def chain n->set_req( j, copy ); _phc._cfg._bbs.map( copy->_idx, pred ); // Extend ("register allocate") the names array for the copy. _phc._names.extend( copy->_idx, phi_name ); } // End of if Phi names do not match } // End of for all inputs to Phi } else { // End of if Phi // Now check for 2-address instructions uint idx; if( n->is_Mach() && (idx=n->as_Mach()->two_adr()) ) { // Get the chosen name for the Node uint name = _phc.Find( n ); assert( name, "no 2-address specials" ); // Check for name mis-match on the 2-address input Node *m = n->in(idx); if( _phc.Find(m) != name ) { Node *copy; assert(!m->is_Con() || m->is_Mach(), "all Con must be Mach"); // At this point it is unsafe to extend live ranges (6550579). // Rematerialize only constants as we do for Phi above. if( m->is_Mach() && m->as_Mach()->is_Con() && m->as_Mach()->rematerialize() ) { copy = m->clone(); // Insert the copy in the basic block, just before us b->_nodes.insert( l++, copy ); if( _phc.clone_projs( b, l, m, copy, _phc._maxlrg ) ) l++; } else { const RegMask *rm = C->matcher()->idealreg2spillmask[m->ideal_reg()]; copy = new (C) MachSpillCopyNode( m, *rm, *rm ); // Insert the copy in the basic block, just before us b->_nodes.insert( l++, copy ); } // Insert the copy in the use-def chain n->set_req(idx, copy ); // Extend ("register allocate") the names array for the copy. _phc._names.extend( copy->_idx, name ); _phc._cfg._bbs.map( copy->_idx, b ); } } // End of is two-adr // Insert a copy at a debug use for a lrg which has high frequency if( b->_freq < OPTO_DEBUG_SPLIT_FREQ || b->is_uncommon(_phc._cfg._bbs) ) { // Walk the debug inputs to the node and check for lrg freq JVMState* jvms = n->jvms(); uint debug_start = jvms ? jvms->debug_start() : 999999; uint debug_end = jvms ? jvms->debug_end() : 999999; for(uint inpidx = debug_start; inpidx < debug_end; inpidx++) { // Do not split monitors; they are only needed for debug table // entries and need no code. if( jvms->is_monitor_use(inpidx) ) continue; Node *inp = n->in(inpidx); uint nidx = _phc.n2lidx(inp); LRG &lrg = lrgs(nidx); // If this lrg has a high frequency use/def if( lrg._maxfreq >= _phc.high_frequency_lrg() ) { // If the live range is also live out of this block (like it // would be for a fast/slow idiom), the normal spill mechanism // does an excellent job. If it is not live out of this block // (like it would be for debug info to uncommon trap) splitting // the live range now allows a better allocation in the high // frequency blocks. // Build_IFG_virtual has converted the live sets to // live-IN info, not live-OUT info. uint k; for( k=0; k < b->_num_succs; k++ ) if( _phc._live->live(b->_succs[k])->member( nidx ) ) break; // Live in to some successor block? if( k < b->_num_succs ) continue; // Live out; do not pre-split // Split the lrg at this use const RegMask *rm = C->matcher()->idealreg2spillmask[inp->ideal_reg()]; Node *copy = new (C) MachSpillCopyNode( inp, *rm, *rm ); // Insert the copy in the use-def chain n->set_req(inpidx, copy ); // Insert the copy in the basic block, just before us b->_nodes.insert( l++, copy ); // Extend ("register allocate") the names array for the copy. _phc.new_lrg( copy, _phc._maxlrg++ ); _phc._cfg._bbs.map( copy->_idx, b ); //tty->print_cr("Split a debug use in Aggressive Coalesce"); } // End of if high frequency use/def } // End of for all debug inputs } // End of if low frequency safepoint } // End of if Phi } // End of for all instructions } // End of for all blocks }
void LateInlineCallGenerator::do_late_inline() { // Can't inline it if (call_node() == NULL || call_node()->outcnt() == 0 || call_node()->in(0) == NULL || call_node()->in(0)->is_top()) return; CallStaticJavaNode* call = call_node(); // Make a clone of the JVMState that appropriate to use for driving a parse Compile* C = Compile::current(); JVMState* jvms = call->jvms()->clone_shallow(C); uint size = call->req(); SafePointNode* map = new (C, size) SafePointNode(size, jvms); for (uint i1 = 0; i1 < size; i1++) { map->init_req(i1, call->in(i1)); } // Make sure the state is a MergeMem for parsing. if (!map->in(TypeFunc::Memory)->is_MergeMem()) { map->set_req(TypeFunc::Memory, MergeMemNode::make(C, map->in(TypeFunc::Memory))); } // Make enough space for the expression stack and transfer the incoming arguments int nargs = method()->arg_size(); jvms->set_map(map); map->ensure_stack(jvms, jvms->method()->max_stack()); if (nargs > 0) { for (int i1 = 0; i1 < nargs; i1++) { map->set_req(i1 + jvms->argoff(), call->in(TypeFunc::Parms + i1)); } } CompileLog* log = C->log(); if (log != NULL) { log->head("late_inline method='%d'", log->identify(method())); JVMState* p = jvms; while (p != NULL) { log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); p = p->caller(); } log->tail("late_inline"); } // Setup default node notes to be picked up by the inlining Node_Notes* old_nn = C->default_node_notes(); if (old_nn != NULL) { Node_Notes* entry_nn = old_nn->clone(C); entry_nn->set_jvms(jvms); C->set_default_node_notes(entry_nn); } // Now perform the inling using the synthesized JVMState JVMState* new_jvms = _inline_cg->generate(jvms); if (new_jvms == NULL) return; // no change if (C->failing()) return; // Capture any exceptional control flow GraphKit kit(new_jvms); // Find the result object Node* result = C->top(); int result_size = method()->return_type()->size(); if (result_size != 0 && !kit.stopped()) { result = (result_size == 1) ? kit.pop() : kit.pop_pair(); } kit.replace_call(call, result); }