duke@435: /* mikael@6198: * Copyright (c) 2005, 2013, Oracle and/or its affiliates. All rights reserved. duke@435: * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. duke@435: * duke@435: * This code is free software; you can redistribute it and/or modify it duke@435: * under the terms of the GNU General Public License version 2 only, as duke@435: * published by the Free Software Foundation. duke@435: * duke@435: * This code is distributed in the hope that it will be useful, but WITHOUT duke@435: * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or duke@435: * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License duke@435: * version 2 for more details (a copy is included in the LICENSE file that duke@435: * accompanied this code). duke@435: * duke@435: * You should have received a copy of the GNU General Public License version duke@435: * 2 along with this work; if not, write to the Free Software Foundation, duke@435: * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. duke@435: * trims@1907: * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA trims@1907: * or visit www.oracle.com if you need additional information or have any trims@1907: * questions. duke@435: * duke@435: */ duke@435: stefank@2314: #include "precompiled.hpp" stefank@2314: #include "compiler/compileLog.hpp" stefank@2314: #include "libadt/vectset.hpp" stefank@2314: #include "opto/addnode.hpp" stefank@2314: #include "opto/callnode.hpp" stefank@2314: #include "opto/cfgnode.hpp" stefank@2314: #include "opto/compile.hpp" stefank@2314: #include "opto/connode.hpp" stefank@2314: #include "opto/locknode.hpp" stefank@2314: #include "opto/loopnode.hpp" stefank@2314: #include "opto/macro.hpp" stefank@2314: #include "opto/memnode.hpp" stefank@2314: #include "opto/node.hpp" stefank@2314: #include "opto/phaseX.hpp" stefank@2314: #include "opto/rootnode.hpp" stefank@2314: #include "opto/runtime.hpp" stefank@2314: #include "opto/subnode.hpp" stefank@2314: #include "opto/type.hpp" stefank@2314: #include "runtime/sharedRuntime.hpp" duke@435: duke@435: duke@435: // duke@435: // Replace any references to "oldref" in inputs to "use" with "newref". duke@435: // Returns the number of replacements made. duke@435: // duke@435: int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) { duke@435: int nreplacements = 0; duke@435: uint req = use->req(); duke@435: for (uint j = 0; j < use->len(); j++) { duke@435: Node *uin = use->in(j); duke@435: if (uin == oldref) { duke@435: if (j < req) duke@435: use->set_req(j, newref); duke@435: else duke@435: use->set_prec(j, newref); duke@435: nreplacements++; duke@435: } else if (j >= req && uin == NULL) { duke@435: break; duke@435: } duke@435: } duke@435: return nreplacements; duke@435: } duke@435: duke@435: void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) { duke@435: // Copy debug information and adjust JVMState information duke@435: uint old_dbg_start = oldcall->tf()->domain()->cnt(); duke@435: uint new_dbg_start = newcall->tf()->domain()->cnt(); duke@435: int jvms_adj = new_dbg_start - old_dbg_start; duke@435: assert (new_dbg_start == newcall->req(), "argument count mismatch"); kvn@498: kvn@5626: // SafePointScalarObject node could be referenced several times in debug info. kvn@5626: // Use Dict to record cloned nodes. kvn@498: Dict* sosn_map = new Dict(cmpkey,hashkey); duke@435: for (uint i = old_dbg_start; i < oldcall->req(); i++) { kvn@498: Node* old_in = oldcall->in(i); kvn@498: // Clone old SafePointScalarObjectNodes, adjusting their field contents. kvn@895: if (old_in != NULL && old_in->is_SafePointScalarObject()) { kvn@498: SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject(); kvn@498: uint old_unique = C->unique(); kvn@5626: Node* new_in = old_sosn->clone(sosn_map); kvn@5626: if (old_unique != C->unique()) { // New node? kvn@3311: new_in->set_req(0, C->root()); // reset control edge kvn@498: new_in = transform_later(new_in); // Register new node. kvn@498: } kvn@498: old_in = new_in; kvn@498: } kvn@498: newcall->add_req(old_in); duke@435: } kvn@498: duke@435: newcall->set_jvms(oldcall->jvms()); duke@435: for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) { duke@435: jvms->set_map(newcall); duke@435: jvms->set_locoff(jvms->locoff()+jvms_adj); duke@435: jvms->set_stkoff(jvms->stkoff()+jvms_adj); duke@435: jvms->set_monoff(jvms->monoff()+jvms_adj); kvn@498: jvms->set_scloff(jvms->scloff()+jvms_adj); duke@435: jvms->set_endoff(jvms->endoff()+jvms_adj); duke@435: } duke@435: } duke@435: kvn@855: Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) { kvn@855: Node* cmp; kvn@855: if (mask != 0) { kvn@4115: Node* and_node = transform_later(new (C) AndXNode(word, MakeConX(mask))); kvn@4115: cmp = transform_later(new (C) CmpXNode(and_node, MakeConX(bits))); kvn@855: } else { kvn@855: cmp = word; kvn@855: } kvn@4115: Node* bol = transform_later(new (C) BoolNode(cmp, BoolTest::ne)); kvn@4115: IfNode* iff = new (C) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); kvn@855: transform_later(iff); duke@435: kvn@855: // Fast path taken. kvn@4115: Node *fast_taken = transform_later( new (C) IfFalseNode(iff) ); duke@435: duke@435: // Fast path not-taken, i.e. slow path kvn@4115: Node *slow_taken = transform_later( new (C) IfTrueNode(iff) ); kvn@855: kvn@855: if (return_fast_path) { kvn@855: region->init_req(edge, slow_taken); // Capture slow-control kvn@855: return fast_taken; kvn@855: } else { kvn@855: region->init_req(edge, fast_taken); // Capture fast-control kvn@855: return slow_taken; kvn@855: } duke@435: } duke@435: duke@435: //--------------------copy_predefined_input_for_runtime_call-------------------- duke@435: void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) { duke@435: // Set fixed predefined input arguments duke@435: call->init_req( TypeFunc::Control, ctrl ); duke@435: call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) ); duke@435: call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ????? duke@435: call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) ); duke@435: call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) ); duke@435: } duke@435: duke@435: //------------------------------make_slow_call--------------------------------- duke@435: CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, address slow_call, const char* leaf_name, Node* slow_path, Node* parm0, Node* parm1) { duke@435: duke@435: // Slow-path call duke@435: CallNode *call = leaf_name kvn@4115: ? (CallNode*)new (C) CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM ) kvn@4115: : (CallNode*)new (C) CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM ); duke@435: duke@435: // Slow path call has no side-effects, uses few values duke@435: copy_predefined_input_for_runtime_call(slow_path, oldcall, call ); duke@435: if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0); duke@435: if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1); duke@435: copy_call_debug_info(oldcall, call); duke@435: call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. kvn@1976: _igvn.replace_node(oldcall, call); duke@435: transform_later(call); duke@435: duke@435: return call; duke@435: } duke@435: duke@435: void PhaseMacroExpand::extract_call_projections(CallNode *call) { duke@435: _fallthroughproj = NULL; duke@435: _fallthroughcatchproj = NULL; duke@435: _ioproj_fallthrough = NULL; duke@435: _ioproj_catchall = NULL; duke@435: _catchallcatchproj = NULL; duke@435: _memproj_fallthrough = NULL; duke@435: _memproj_catchall = NULL; duke@435: _resproj = NULL; duke@435: for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) { duke@435: ProjNode *pn = call->fast_out(i)->as_Proj(); duke@435: switch (pn->_con) { duke@435: case TypeFunc::Control: duke@435: { duke@435: // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj duke@435: _fallthroughproj = pn; duke@435: DUIterator_Fast jmax, j = pn->fast_outs(jmax); duke@435: const Node *cn = pn->fast_out(j); duke@435: if (cn->is_Catch()) { duke@435: ProjNode *cpn = NULL; duke@435: for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) { duke@435: cpn = cn->fast_out(k)->as_Proj(); duke@435: assert(cpn->is_CatchProj(), "must be a CatchProjNode"); duke@435: if (cpn->_con == CatchProjNode::fall_through_index) duke@435: _fallthroughcatchproj = cpn; duke@435: else { duke@435: assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index."); duke@435: _catchallcatchproj = cpn; duke@435: } duke@435: } duke@435: } duke@435: break; duke@435: } duke@435: case TypeFunc::I_O: duke@435: if (pn->_is_io_use) duke@435: _ioproj_catchall = pn; duke@435: else duke@435: _ioproj_fallthrough = pn; duke@435: break; duke@435: case TypeFunc::Memory: duke@435: if (pn->_is_io_use) duke@435: _memproj_catchall = pn; duke@435: else duke@435: _memproj_fallthrough = pn; duke@435: break; duke@435: case TypeFunc::Parms: duke@435: _resproj = pn; duke@435: break; duke@435: default: duke@435: assert(false, "unexpected projection from allocation node."); duke@435: } duke@435: } duke@435: duke@435: } duke@435: kvn@508: // Eliminate a card mark sequence. p2x is a ConvP2XNode kvn@1286: void PhaseMacroExpand::eliminate_card_mark(Node* p2x) { kvn@508: assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required"); kvn@1286: if (!UseG1GC) { kvn@1286: // vanilla/CMS post barrier kvn@1286: Node *shift = p2x->unique_out(); kvn@1286: Node *addp = shift->unique_out(); kvn@1286: for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) { never@2814: Node *mem = addp->last_out(j); never@2814: if (UseCondCardMark && mem->is_Load()) { never@2814: assert(mem->Opcode() == Op_LoadB, "unexpected code shape"); never@2814: // The load is checking if the card has been written so never@2814: // replace it with zero to fold the test. never@2814: _igvn.replace_node(mem, intcon(0)); never@2814: continue; never@2814: } never@2814: assert(mem->is_Store(), "store required"); never@2814: _igvn.replace_node(mem, mem->in(MemNode::Memory)); kvn@1286: } kvn@1286: } else { kvn@1286: // G1 pre/post barriers kvn@3521: assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes"); kvn@1286: // It could be only one user, URShift node, in Object.clone() instrinsic kvn@1286: // but the new allocation is passed to arraycopy stub and it could not kvn@1286: // be scalar replaced. So we don't check the case. kvn@1286: kvn@3521: // An other case of only one user (Xor) is when the value check for NULL kvn@3521: // in G1 post barrier is folded after CCP so the code which used URShift kvn@3521: // is removed. kvn@3521: kvn@3521: // Take Region node before eliminating post barrier since it also kvn@3521: // eliminates CastP2X node when it has only one user. kvn@3521: Node* this_region = p2x->in(0); kvn@3521: assert(this_region != NULL, ""); kvn@3521: kvn@1286: // Remove G1 post barrier. kvn@1286: kvn@1286: // Search for CastP2X->Xor->URShift->Cmp path which kvn@1286: // checks if the store done to a different from the value's region. kvn@1286: // And replace Cmp with #0 (false) to collapse G1 post barrier. kvn@1286: Node* xorx = NULL; kvn@1286: for (DUIterator_Fast imax, i = p2x->fast_outs(imax); i < imax; i++) { kvn@1286: Node* u = p2x->fast_out(i); kvn@1286: if (u->Opcode() == Op_XorX) { kvn@1286: xorx = u; kvn@1286: break; kvn@1286: } kvn@1286: } kvn@1286: assert(xorx != NULL, "missing G1 post barrier"); kvn@1286: Node* shift = xorx->unique_out(); kvn@1286: Node* cmpx = shift->unique_out(); kvn@1286: assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() && kvn@1286: cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne, kvn@1286: "missing region check in G1 post barrier"); kvn@1286: _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); kvn@1286: kvn@1286: // Remove G1 pre barrier. kvn@1286: kvn@1286: // Search "if (marking != 0)" check and set it to "false". kvn@1286: // There is no G1 pre barrier if previous stored value is NULL kvn@1286: // (for example, after initialization). kvn@1286: if (this_region->is_Region() && this_region->req() == 3) { kvn@1286: int ind = 1; kvn@1286: if (!this_region->in(ind)->is_IfFalse()) { kvn@1286: ind = 2; kvn@1286: } kvn@1286: if (this_region->in(ind)->is_IfFalse()) { kvn@1286: Node* bol = this_region->in(ind)->in(0)->in(1); kvn@1286: assert(bol->is_Bool(), ""); kvn@1286: cmpx = bol->in(1); kvn@1286: if (bol->as_Bool()->_test._test == BoolTest::ne && kvn@1286: cmpx->is_Cmp() && cmpx->in(2) == intcon(0) && kvn@1286: cmpx->in(1)->is_Load()) { kvn@1286: Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address); kvn@1286: const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + kvn@1286: PtrQueue::byte_offset_of_active()); kvn@1286: if (adr->is_AddP() && adr->in(AddPNode::Base) == top() && kvn@1286: adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && kvn@1286: adr->in(AddPNode::Offset) == MakeConX(marking_offset)) { kvn@1286: _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); kvn@1286: } kvn@1286: } kvn@1286: } kvn@1286: } kvn@1286: // Now CastP2X can be removed since it is used only on dead path kvn@1286: // which currently still alive until igvn optimize it. kvn@3521: assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, ""); kvn@1286: _igvn.replace_node(p2x, top()); kvn@508: } kvn@508: } kvn@508: kvn@508: // Search for a memory operation for the specified memory slice. kvn@688: static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) { kvn@508: Node *orig_mem = mem; kvn@508: Node *alloc_mem = alloc->in(TypeFunc::Memory); kvn@688: const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr(); kvn@508: while (true) { kvn@508: if (mem == alloc_mem || mem == start_mem ) { twisti@1040: return mem; // hit one of our sentinels kvn@508: } else if (mem->is_MergeMem()) { kvn@508: mem = mem->as_MergeMem()->memory_at(alias_idx); kvn@508: } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) { kvn@508: Node *in = mem->in(0); kvn@508: // we can safely skip over safepoints, calls, locks and membars because we kvn@508: // already know that the object is safe to eliminate. kvn@508: if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) { kvn@508: return in; kvn@688: } else if (in->is_Call()) { kvn@688: CallNode *call = in->as_Call(); kvn@688: if (!call->may_modify(tinst, phase)) { kvn@688: mem = call->in(TypeFunc::Memory); kvn@688: } kvn@688: mem = in->in(TypeFunc::Memory); kvn@688: } else if (in->is_MemBar()) { kvn@508: mem = in->in(TypeFunc::Memory); kvn@508: } else { kvn@508: assert(false, "unexpected projection"); kvn@508: } kvn@508: } else if (mem->is_Store()) { kvn@508: const TypePtr* atype = mem->as_Store()->adr_type(); kvn@508: int adr_idx = Compile::current()->get_alias_index(atype); kvn@508: if (adr_idx == alias_idx) { kvn@508: assert(atype->isa_oopptr(), "address type must be oopptr"); kvn@508: int adr_offset = atype->offset(); kvn@508: uint adr_iid = atype->is_oopptr()->instance_id(); kvn@508: // Array elements references have the same alias_idx kvn@508: // but different offset and different instance_id. kvn@508: if (adr_offset == offset && adr_iid == alloc->_idx) kvn@508: return mem; kvn@508: } else { kvn@508: assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw"); kvn@508: } kvn@508: mem = mem->in(MemNode::Memory); kvn@1535: } else if (mem->is_ClearArray()) { kvn@1535: if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) { kvn@1535: // Can not bypass initialization of the instance kvn@1535: // we are looking. kvn@1535: debug_only(intptr_t offset;) kvn@1535: assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity"); kvn@1535: InitializeNode* init = alloc->as_Allocate()->initialization(); kvn@1535: // We are looking for stored value, return Initialize node kvn@1535: // or memory edge from Allocate node. kvn@1535: if (init != NULL) kvn@1535: return init; kvn@1535: else kvn@1535: return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers). kvn@1535: } kvn@1535: // Otherwise skip it (the call updated 'mem' value). kvn@1019: } else if (mem->Opcode() == Op_SCMemProj) { kvn@4479: mem = mem->in(0); kvn@4479: Node* adr = NULL; kvn@4479: if (mem->is_LoadStore()) { kvn@4479: adr = mem->in(MemNode::Address); kvn@4479: } else { kvn@4479: assert(mem->Opcode() == Op_EncodeISOArray, "sanity"); kvn@4479: adr = mem->in(3); // Destination array kvn@4479: } kvn@4479: const TypePtr* atype = adr->bottom_type()->is_ptr(); kvn@1019: int adr_idx = Compile::current()->get_alias_index(atype); kvn@1019: if (adr_idx == alias_idx) { kvn@1019: assert(false, "Object is not scalar replaceable if a LoadStore node access its field"); kvn@1019: return NULL; kvn@1019: } kvn@4479: mem = mem->in(MemNode::Memory); kvn@508: } else { kvn@508: return mem; kvn@508: } kvn@682: assert(mem != orig_mem, "dead memory loop"); kvn@508: } kvn@508: } kvn@508: kvn@508: // kvn@508: // Given a Memory Phi, compute a value Phi containing the values from stores kvn@508: // on the input paths. kvn@508: // Note: this function is recursive, its depth is limied by the "level" argument kvn@508: // Returns the computed Phi, or NULL if it cannot compute it. kvn@682: Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, Node *alloc, Node_Stack *value_phis, int level) { kvn@682: assert(mem->is_Phi(), "sanity"); kvn@682: int alias_idx = C->get_alias_index(adr_t); kvn@682: int offset = adr_t->offset(); kvn@682: int instance_id = adr_t->instance_id(); kvn@682: kvn@682: // Check if an appropriate value phi already exists. kvn@682: Node* region = mem->in(0); kvn@682: for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { kvn@682: Node* phi = region->fast_out(k); kvn@682: if (phi->is_Phi() && phi != mem && kvn@682: phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) { kvn@682: return phi; kvn@682: } kvn@682: } kvn@682: // Check if an appropriate new value phi already exists. kvn@2985: Node* new_phi = value_phis->find(mem->_idx); kvn@2985: if (new_phi != NULL) kvn@2985: return new_phi; kvn@508: kvn@508: if (level <= 0) { kvn@688: return NULL; // Give up: phi tree too deep kvn@508: } kvn@508: Node *start_mem = C->start()->proj_out(TypeFunc::Memory); kvn@508: Node *alloc_mem = alloc->in(TypeFunc::Memory); kvn@508: kvn@508: uint length = mem->req(); zgu@3900: GrowableArray values(length, length, NULL, false); kvn@508: kvn@682: // create a new Phi for the value kvn@4115: PhiNode *phi = new (C) PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset); kvn@682: transform_later(phi); kvn@682: value_phis->push(phi, mem->_idx); kvn@682: kvn@508: for (uint j = 1; j < length; j++) { kvn@508: Node *in = mem->in(j); kvn@508: if (in == NULL || in->is_top()) { kvn@508: values.at_put(j, in); kvn@508: } else { kvn@688: Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn); kvn@508: if (val == start_mem || val == alloc_mem) { kvn@508: // hit a sentinel, return appropriate 0 value kvn@508: values.at_put(j, _igvn.zerocon(ft)); kvn@508: continue; kvn@508: } kvn@508: if (val->is_Initialize()) { kvn@508: val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); kvn@508: } kvn@508: if (val == NULL) { kvn@508: return NULL; // can't find a value on this path kvn@508: } kvn@508: if (val == mem) { kvn@508: values.at_put(j, mem); kvn@508: } else if (val->is_Store()) { kvn@508: values.at_put(j, val->in(MemNode::ValueIn)); kvn@508: } else if(val->is_Proj() && val->in(0) == alloc) { kvn@508: values.at_put(j, _igvn.zerocon(ft)); kvn@508: } else if (val->is_Phi()) { kvn@682: val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1); kvn@682: if (val == NULL) { kvn@682: return NULL; kvn@508: } kvn@682: values.at_put(j, val); kvn@1019: } else if (val->Opcode() == Op_SCMemProj) { kvn@4479: assert(val->in(0)->is_LoadStore() || val->in(0)->Opcode() == Op_EncodeISOArray, "sanity"); kvn@1019: assert(false, "Object is not scalar replaceable if a LoadStore node access its field"); kvn@1019: return NULL; kvn@508: } else { kvn@1019: #ifdef ASSERT kvn@1019: val->dump(); kvn@688: assert(false, "unknown node on this path"); kvn@1019: #endif kvn@688: return NULL; // unknown node on this path kvn@508: } kvn@508: } kvn@508: } kvn@682: // Set Phi's inputs kvn@508: for (uint j = 1; j < length; j++) { kvn@508: if (values.at(j) == mem) { kvn@508: phi->init_req(j, phi); kvn@508: } else { kvn@508: phi->init_req(j, values.at(j)); kvn@508: } kvn@508: } kvn@508: return phi; kvn@508: } kvn@508: kvn@508: // Search the last value stored into the object's field. kvn@508: Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) { kvn@658: assert(adr_t->is_known_instance_field(), "instance required"); kvn@658: int instance_id = adr_t->instance_id(); kvn@658: assert((uint)instance_id == alloc->_idx, "wrong allocation"); kvn@508: kvn@508: int alias_idx = C->get_alias_index(adr_t); kvn@508: int offset = adr_t->offset(); kvn@508: Node *start_mem = C->start()->proj_out(TypeFunc::Memory); kvn@508: Node *alloc_ctrl = alloc->in(TypeFunc::Control); kvn@508: Node *alloc_mem = alloc->in(TypeFunc::Memory); kvn@682: Arena *a = Thread::current()->resource_area(); kvn@682: VectorSet visited(a); kvn@508: kvn@508: kvn@508: bool done = sfpt_mem == alloc_mem; kvn@508: Node *mem = sfpt_mem; kvn@508: while (!done) { kvn@508: if (visited.test_set(mem->_idx)) { kvn@508: return NULL; // found a loop, give up kvn@508: } kvn@688: mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); kvn@508: if (mem == start_mem || mem == alloc_mem) { kvn@508: done = true; // hit a sentinel, return appropriate 0 value kvn@508: } else if (mem->is_Initialize()) { kvn@508: mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); kvn@508: if (mem == NULL) { kvn@508: done = true; // Something go wrong. kvn@508: } else if (mem->is_Store()) { kvn@508: const TypePtr* atype = mem->as_Store()->adr_type(); kvn@508: assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); kvn@508: done = true; kvn@508: } kvn@508: } else if (mem->is_Store()) { kvn@508: const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); kvn@508: assert(atype != NULL, "address type must be oopptr"); kvn@508: assert(C->get_alias_index(atype) == alias_idx && kvn@658: atype->is_known_instance_field() && atype->offset() == offset && kvn@508: atype->instance_id() == instance_id, "store is correct memory slice"); kvn@508: done = true; kvn@508: } else if (mem->is_Phi()) { kvn@508: // try to find a phi's unique input kvn@508: Node *unique_input = NULL; kvn@508: Node *top = C->top(); kvn@508: for (uint i = 1; i < mem->req(); i++) { kvn@688: Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); kvn@508: if (n == NULL || n == top || n == mem) { kvn@508: continue; kvn@508: } else if (unique_input == NULL) { kvn@508: unique_input = n; kvn@508: } else if (unique_input != n) { kvn@508: unique_input = top; kvn@508: break; kvn@508: } kvn@508: } kvn@508: if (unique_input != NULL && unique_input != top) { kvn@508: mem = unique_input; kvn@508: } else { kvn@508: done = true; kvn@508: } kvn@508: } else { kvn@508: assert(false, "unexpected node"); kvn@508: } kvn@508: } kvn@508: if (mem != NULL) { kvn@508: if (mem == start_mem || mem == alloc_mem) { kvn@508: // hit a sentinel, return appropriate 0 value kvn@508: return _igvn.zerocon(ft); kvn@508: } else if (mem->is_Store()) { kvn@508: return mem->in(MemNode::ValueIn); kvn@508: } else if (mem->is_Phi()) { kvn@508: // attempt to produce a Phi reflecting the values on the input paths of the Phi kvn@682: Node_Stack value_phis(a, 8); kvn@688: Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); kvn@508: if (phi != NULL) { kvn@508: return phi; kvn@682: } else { kvn@682: // Kill all new Phis kvn@682: while(value_phis.is_nonempty()) { kvn@682: Node* n = value_phis.node(); kvn@1976: _igvn.replace_node(n, C->top()); kvn@682: value_phis.pop(); kvn@682: } kvn@508: } kvn@508: } kvn@508: } kvn@508: // Something go wrong. kvn@508: return NULL; kvn@508: } kvn@508: kvn@508: // Check the possibility of scalar replacement. kvn@508: bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray & safepoints) { kvn@508: // Scan the uses of the allocation to check for anything that would kvn@508: // prevent us from eliminating it. kvn@508: NOT_PRODUCT( const char* fail_eliminate = NULL; ) kvn@508: DEBUG_ONLY( Node* disq_node = NULL; ) kvn@508: bool can_eliminate = true; kvn@508: kvn@508: Node* res = alloc->result_cast(); kvn@508: const TypeOopPtr* res_type = NULL; kvn@508: if (res == NULL) { kvn@508: // All users were eliminated. kvn@508: } else if (!res->is_CheckCastPP()) { kvn@508: NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) kvn@508: can_eliminate = false; kvn@508: } else { kvn@508: res_type = _igvn.type(res)->isa_oopptr(); kvn@508: if (res_type == NULL) { kvn@508: NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) kvn@508: can_eliminate = false; kvn@508: } else if (res_type->isa_aryptr()) { kvn@508: int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); kvn@508: if (length < 0) { kvn@508: NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) kvn@508: can_eliminate = false; kvn@508: } kvn@508: } kvn@508: } kvn@508: kvn@508: if (can_eliminate && res != NULL) { kvn@508: for (DUIterator_Fast jmax, j = res->fast_outs(jmax); kvn@508: j < jmax && can_eliminate; j++) { kvn@508: Node* use = res->fast_out(j); kvn@508: kvn@508: if (use->is_AddP()) { kvn@508: const TypePtr* addp_type = _igvn.type(use)->is_ptr(); kvn@508: int offset = addp_type->offset(); kvn@508: kvn@508: if (offset == Type::OffsetTop || offset == Type::OffsetBot) { kvn@508: NOT_PRODUCT(fail_eliminate = "Undefined field referrence";) kvn@508: can_eliminate = false; kvn@508: break; kvn@508: } kvn@508: for (DUIterator_Fast kmax, k = use->fast_outs(kmax); kvn@508: k < kmax && can_eliminate; k++) { kvn@508: Node* n = use->fast_out(k); kvn@508: if (!n->is_Store() && n->Opcode() != Op_CastP2X) { kvn@508: DEBUG_ONLY(disq_node = n;) kvn@688: if (n->is_Load() || n->is_LoadStore()) { kvn@508: NOT_PRODUCT(fail_eliminate = "Field load";) kvn@508: } else { kvn@508: NOT_PRODUCT(fail_eliminate = "Not store field referrence";) kvn@508: } kvn@508: can_eliminate = false; kvn@508: } kvn@508: } kvn@508: } else if (use->is_SafePoint()) { kvn@508: SafePointNode* sfpt = use->as_SafePoint(); kvn@603: if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { kvn@508: // Object is passed as argument. kvn@508: DEBUG_ONLY(disq_node = use;) kvn@508: NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) kvn@508: can_eliminate = false; kvn@508: } kvn@508: Node* sfptMem = sfpt->memory(); kvn@508: if (sfptMem == NULL || sfptMem->is_top()) { kvn@508: DEBUG_ONLY(disq_node = use;) kvn@508: NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";) kvn@508: can_eliminate = false; kvn@508: } else { kvn@508: safepoints.append_if_missing(sfpt); kvn@508: } kvn@508: } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark kvn@508: if (use->is_Phi()) { kvn@508: if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { kvn@508: NOT_PRODUCT(fail_eliminate = "Object is return value";) kvn@508: } else { kvn@508: NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) kvn@508: } kvn@508: DEBUG_ONLY(disq_node = use;) kvn@508: } else { kvn@508: if (use->Opcode() == Op_Return) { kvn@508: NOT_PRODUCT(fail_eliminate = "Object is return value";) kvn@508: }else { kvn@508: NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) kvn@508: } kvn@508: DEBUG_ONLY(disq_node = use;) kvn@508: } kvn@508: can_eliminate = false; kvn@508: } kvn@508: } kvn@508: } kvn@508: kvn@508: #ifndef PRODUCT kvn@508: if (PrintEliminateAllocations) { kvn@508: if (can_eliminate) { kvn@508: tty->print("Scalar "); kvn@508: if (res == NULL) kvn@508: alloc->dump(); kvn@508: else kvn@508: res->dump(); kvn@5110: } else if (alloc->_is_scalar_replaceable) { kvn@508: tty->print("NotScalar (%s)", fail_eliminate); kvn@508: if (res == NULL) kvn@508: alloc->dump(); kvn@508: else kvn@508: res->dump(); kvn@508: #ifdef ASSERT kvn@508: if (disq_node != NULL) { kvn@508: tty->print(" >>>> "); kvn@508: disq_node->dump(); kvn@508: } kvn@508: #endif /*ASSERT*/ kvn@508: } kvn@508: } kvn@508: #endif kvn@508: return can_eliminate; kvn@508: } kvn@508: kvn@508: // Do scalar replacement. kvn@508: bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray & safepoints) { kvn@508: GrowableArray safepoints_done; kvn@508: kvn@508: ciKlass* klass = NULL; kvn@508: ciInstanceKlass* iklass = NULL; kvn@508: int nfields = 0; kvn@508: int array_base; kvn@508: int element_size; kvn@508: BasicType basic_elem_type; kvn@508: ciType* elem_type; kvn@508: kvn@508: Node* res = alloc->result_cast(); kvn@508: const TypeOopPtr* res_type = NULL; kvn@508: if (res != NULL) { // Could be NULL when there are no users kvn@508: res_type = _igvn.type(res)->isa_oopptr(); kvn@508: } kvn@508: kvn@508: if (res != NULL) { kvn@508: klass = res_type->klass(); kvn@508: if (res_type->isa_instptr()) { kvn@508: // find the fields of the class which will be needed for safepoint debug information kvn@508: assert(klass->is_instance_klass(), "must be an instance klass."); kvn@508: iklass = klass->as_instance_klass(); kvn@508: nfields = iklass->nof_nonstatic_fields(); kvn@508: } else { kvn@508: // find the array's elements which will be needed for safepoint debug information kvn@508: nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); kvn@508: assert(klass->is_array_klass() && nfields >= 0, "must be an array klass."); kvn@508: elem_type = klass->as_array_klass()->element_type(); kvn@508: basic_elem_type = elem_type->basic_type(); kvn@508: array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); kvn@508: element_size = type2aelembytes(basic_elem_type); kvn@508: } kvn@508: } kvn@508: // kvn@508: // Process the safepoint uses kvn@508: // kvn@508: while (safepoints.length() > 0) { kvn@508: SafePointNode* sfpt = safepoints.pop(); kvn@508: Node* mem = sfpt->memory(); kvn@5626: assert(sfpt->jvms() != NULL, "missed JVMS"); kvn@5626: // Fields of scalar objs are referenced only at the end kvn@5626: // of regular debuginfo at the last (youngest) JVMS. kvn@5626: // Record relative start index. kvn@5626: uint first_ind = (sfpt->req() - sfpt->jvms()->scloff()); kvn@4115: SafePointScalarObjectNode* sobj = new (C) SafePointScalarObjectNode(res_type, kvn@508: #ifdef ASSERT kvn@508: alloc, kvn@508: #endif kvn@508: first_ind, nfields); kvn@3311: sobj->init_req(0, C->root()); kvn@508: transform_later(sobj); kvn@508: kvn@508: // Scan object's fields adding an input to the safepoint for each field. kvn@508: for (int j = 0; j < nfields; j++) { kvn@741: intptr_t offset; kvn@508: ciField* field = NULL; kvn@508: if (iklass != NULL) { kvn@508: field = iklass->nonstatic_field_at(j); kvn@508: offset = field->offset(); kvn@508: elem_type = field->type(); kvn@508: basic_elem_type = field->layout_type(); kvn@508: } else { kvn@741: offset = array_base + j * (intptr_t)element_size; kvn@508: } kvn@508: kvn@508: const Type *field_type; kvn@508: // The next code is taken from Parse::do_get_xxx(). kvn@559: if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) { kvn@508: if (!elem_type->is_loaded()) { kvn@508: field_type = TypeInstPtr::BOTTOM; kvn@2037: } else if (field != NULL && field->is_constant() && field->is_static()) { kvn@508: // This can happen if the constant oop is non-perm. kvn@508: ciObject* con = field->constant_value().as_object(); kvn@508: // Do not "join" in the previous type; it doesn't add value, kvn@508: // and may yield a vacuous result if the field is of interface type. kvn@508: field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); kvn@508: assert(field_type != NULL, "field singleton type must be consistent"); kvn@508: } else { kvn@508: field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); kvn@508: } kvn@559: if (UseCompressedOops) { kvn@656: field_type = field_type->make_narrowoop(); kvn@559: basic_elem_type = T_NARROWOOP; kvn@559: } kvn@508: } else { kvn@508: field_type = Type::get_const_basic_type(basic_elem_type); kvn@508: } kvn@508: kvn@508: const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr(); kvn@508: kvn@508: Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc); kvn@508: if (field_val == NULL) { kvn@3311: // We weren't able to find a value for this field, kvn@3311: // give up on eliminating this allocation. kvn@3311: kvn@3311: // Remove any extra entries we added to the safepoint. kvn@508: uint last = sfpt->req() - 1; kvn@508: for (int k = 0; k < j; k++) { kvn@508: sfpt->del_req(last--); kvn@508: } kvn@508: // rollback processed safepoints kvn@508: while (safepoints_done.length() > 0) { kvn@508: SafePointNode* sfpt_done = safepoints_done.pop(); kvn@508: // remove any extra entries we added to the safepoint kvn@508: last = sfpt_done->req() - 1; kvn@508: for (int k = 0; k < nfields; k++) { kvn@508: sfpt_done->del_req(last--); kvn@508: } kvn@508: JVMState *jvms = sfpt_done->jvms(); kvn@508: jvms->set_endoff(sfpt_done->req()); kvn@508: // Now make a pass over the debug information replacing any references kvn@508: // to SafePointScalarObjectNode with the allocated object. kvn@508: int start = jvms->debug_start(); kvn@508: int end = jvms->debug_end(); kvn@508: for (int i = start; i < end; i++) { kvn@508: if (sfpt_done->in(i)->is_SafePointScalarObject()) { kvn@508: SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); kvn@5626: if (scobj->first_index(jvms) == sfpt_done->req() && kvn@508: scobj->n_fields() == (uint)nfields) { kvn@508: assert(scobj->alloc() == alloc, "sanity"); kvn@508: sfpt_done->set_req(i, res); kvn@508: } kvn@508: } kvn@508: } kvn@508: } kvn@508: #ifndef PRODUCT kvn@508: if (PrintEliminateAllocations) { kvn@508: if (field != NULL) { kvn@508: tty->print("=== At SafePoint node %d can't find value of Field: ", kvn@508: sfpt->_idx); kvn@508: field->print(); kvn@508: int field_idx = C->get_alias_index(field_addr_type); kvn@508: tty->print(" (alias_idx=%d)", field_idx); kvn@508: } else { // Array's element kvn@508: tty->print("=== At SafePoint node %d can't find value of array element [%d]", kvn@508: sfpt->_idx, j); kvn@508: } kvn@508: tty->print(", which prevents elimination of: "); kvn@508: if (res == NULL) kvn@508: alloc->dump(); kvn@508: else kvn@508: res->dump(); kvn@508: } kvn@508: #endif kvn@508: return false; kvn@508: } kvn@559: if (UseCompressedOops && field_type->isa_narrowoop()) { kvn@559: // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation kvn@559: // to be able scalar replace the allocation. kvn@656: if (field_val->is_EncodeP()) { kvn@656: field_val = field_val->in(1); kvn@656: } else { kvn@5111: field_val = transform_later(new (C) DecodeNNode(field_val, field_val->get_ptr_type())); kvn@656: } kvn@559: } kvn@508: sfpt->add_req(field_val); kvn@508: } kvn@508: JVMState *jvms = sfpt->jvms(); kvn@508: jvms->set_endoff(sfpt->req()); kvn@508: // Now make a pass over the debug information replacing any references kvn@508: // to the allocated object with "sobj" kvn@508: int start = jvms->debug_start(); kvn@508: int end = jvms->debug_end(); kvn@5110: sfpt->replace_edges_in_range(res, sobj, start, end); kvn@508: safepoints_done.append_if_missing(sfpt); // keep it for rollback kvn@508: } kvn@508: return true; kvn@508: } kvn@508: kvn@508: // Process users of eliminated allocation. kvn@5110: void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) { kvn@508: Node* res = alloc->result_cast(); kvn@508: if (res != NULL) { kvn@508: for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { kvn@508: Node *use = res->last_out(j); kvn@508: uint oc1 = res->outcnt(); kvn@508: kvn@508: if (use->is_AddP()) { kvn@508: for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { kvn@508: Node *n = use->last_out(k); kvn@508: uint oc2 = use->outcnt(); kvn@508: if (n->is_Store()) { kvn@1535: #ifdef ASSERT kvn@1535: // Verify that there is no dependent MemBarVolatile nodes, kvn@1535: // they should be removed during IGVN, see MemBarNode::Ideal(). kvn@1535: for (DUIterator_Fast pmax, p = n->fast_outs(pmax); kvn@1535: p < pmax; p++) { kvn@1535: Node* mb = n->fast_out(p); kvn@1535: assert(mb->is_Initialize() || !mb->is_MemBar() || kvn@1535: mb->req() <= MemBarNode::Precedent || kvn@1535: mb->in(MemBarNode::Precedent) != n, kvn@1535: "MemBarVolatile should be eliminated for non-escaping object"); kvn@1535: } kvn@1535: #endif kvn@508: _igvn.replace_node(n, n->in(MemNode::Memory)); kvn@508: } else { kvn@508: eliminate_card_mark(n); kvn@508: } kvn@508: k -= (oc2 - use->outcnt()); kvn@508: } kvn@508: } else { kvn@508: eliminate_card_mark(use); kvn@508: } kvn@508: j -= (oc1 - res->outcnt()); kvn@508: } kvn@508: assert(res->outcnt() == 0, "all uses of allocated objects must be deleted"); kvn@508: _igvn.remove_dead_node(res); kvn@508: } kvn@508: kvn@508: // kvn@508: // Process other users of allocation's projections kvn@508: // kvn@508: if (_resproj != NULL && _resproj->outcnt() != 0) { kvn@5110: // First disconnect stores captured by Initialize node. kvn@5110: // If Initialize node is eliminated first in the following code, kvn@5110: // it will kill such stores and DUIterator_Last will assert. kvn@5110: for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) { kvn@5110: Node *use = _resproj->fast_out(j); kvn@5110: if (use->is_AddP()) { kvn@5110: // raw memory addresses used only by the initialization kvn@5110: _igvn.replace_node(use, C->top()); kvn@5110: --j; --jmax; kvn@5110: } kvn@5110: } kvn@508: for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) { kvn@508: Node *use = _resproj->last_out(j); kvn@508: uint oc1 = _resproj->outcnt(); kvn@508: if (use->is_Initialize()) { kvn@508: // Eliminate Initialize node. kvn@508: InitializeNode *init = use->as_Initialize(); kvn@508: assert(init->outcnt() <= 2, "only a control and memory projection expected"); kvn@508: Node *ctrl_proj = init->proj_out(TypeFunc::Control); kvn@508: if (ctrl_proj != NULL) { kvn@508: assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection"); kvn@508: _igvn.replace_node(ctrl_proj, _fallthroughcatchproj); kvn@508: } kvn@508: Node *mem_proj = init->proj_out(TypeFunc::Memory); kvn@508: if (mem_proj != NULL) { kvn@508: Node *mem = init->in(TypeFunc::Memory); kvn@508: #ifdef ASSERT kvn@508: if (mem->is_MergeMem()) { kvn@508: assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection"); kvn@508: } else { kvn@508: assert(mem == _memproj_fallthrough, "allocation memory projection"); kvn@508: } kvn@508: #endif kvn@508: _igvn.replace_node(mem_proj, mem); kvn@508: } kvn@508: } else { kvn@508: assert(false, "only Initialize or AddP expected"); kvn@508: } kvn@508: j -= (oc1 - _resproj->outcnt()); kvn@508: } kvn@508: } kvn@508: if (_fallthroughcatchproj != NULL) { kvn@508: _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control)); kvn@508: } kvn@508: if (_memproj_fallthrough != NULL) { kvn@508: _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory)); kvn@508: } kvn@508: if (_memproj_catchall != NULL) { kvn@508: _igvn.replace_node(_memproj_catchall, C->top()); kvn@508: } kvn@508: if (_ioproj_fallthrough != NULL) { kvn@508: _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O)); kvn@508: } kvn@508: if (_ioproj_catchall != NULL) { kvn@508: _igvn.replace_node(_ioproj_catchall, C->top()); kvn@508: } kvn@508: if (_catchallcatchproj != NULL) { kvn@508: _igvn.replace_node(_catchallcatchproj, C->top()); kvn@508: } kvn@508: } kvn@508: kvn@508: bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { kvn@5110: if (!EliminateAllocations || !alloc->_is_non_escaping) { kvn@5110: return false; kvn@5110: } kvn@5110: Node* klass = alloc->in(AllocateNode::KlassNode); kvn@5110: const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr(); kvn@5110: Node* res = alloc->result_cast(); kvn@5110: // Eliminate boxing allocations which are not used kvn@5110: // regardless scalar replacable status. kvn@5110: bool boxing_alloc = C->eliminate_boxing() && kvn@5110: tklass->klass()->is_instance_klass() && kvn@5110: tklass->klass()->as_instance_klass()->is_box_klass(); kvn@5110: if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) { kvn@508: return false; kvn@508: } kvn@508: kvn@508: extract_call_projections(alloc); kvn@508: kvn@508: GrowableArray safepoints; kvn@508: if (!can_eliminate_allocation(alloc, safepoints)) { kvn@508: return false; kvn@508: } kvn@508: kvn@5110: if (!alloc->_is_scalar_replaceable) { kvn@5110: assert(res == NULL, "sanity"); kvn@5110: // We can only eliminate allocation if all debug info references kvn@5110: // are already replaced with SafePointScalarObject because kvn@5110: // we can't search for a fields value without instance_id. kvn@5110: if (safepoints.length() > 0) { kvn@5110: return false; kvn@5110: } kvn@5110: } kvn@5110: kvn@508: if (!scalar_replacement(alloc, safepoints)) { kvn@508: return false; kvn@508: } kvn@508: never@1515: CompileLog* log = C->log(); never@1515: if (log != NULL) { never@1515: log->head("eliminate_allocation type='%d'", never@1515: log->identify(tklass->klass())); never@1515: JVMState* p = alloc->jvms(); never@1515: while (p != NULL) { never@1515: log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); never@1515: p = p->caller(); never@1515: } never@1515: log->tail("eliminate_allocation"); never@1515: } never@1515: kvn@508: process_users_of_allocation(alloc); kvn@508: kvn@508: #ifndef PRODUCT never@1515: if (PrintEliminateAllocations) { never@1515: if (alloc->is_AllocateArray()) never@1515: tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); never@1515: else never@1515: tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); never@1515: } kvn@508: #endif kvn@508: kvn@508: return true; kvn@508: } kvn@508: kvn@5110: bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) { kvn@5110: // EA should remove all uses of non-escaping boxing node. kvn@5110: if (!C->eliminate_boxing() || boxing->proj_out(TypeFunc::Parms) != NULL) { kvn@5110: return false; kvn@5110: } kvn@5110: kvn@5110: extract_call_projections(boxing); kvn@5110: kvn@5110: const TypeTuple* r = boxing->tf()->range(); kvn@5110: assert(r->cnt() > TypeFunc::Parms, "sanity"); kvn@5110: const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr(); kvn@5110: assert(t != NULL, "sanity"); kvn@5110: kvn@5110: CompileLog* log = C->log(); kvn@5110: if (log != NULL) { kvn@5110: log->head("eliminate_boxing type='%d'", kvn@5110: log->identify(t->klass())); kvn@5110: JVMState* p = boxing->jvms(); kvn@5110: while (p != NULL) { kvn@5110: log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); kvn@5110: p = p->caller(); kvn@5110: } kvn@5110: log->tail("eliminate_boxing"); kvn@5110: } kvn@5110: kvn@5110: process_users_of_allocation(boxing); kvn@5110: kvn@5110: #ifndef PRODUCT kvn@5110: if (PrintEliminateAllocations) { kvn@5110: tty->print("++++ Eliminated: %d ", boxing->_idx); kvn@5110: boxing->method()->print_short_name(tty); kvn@5110: tty->cr(); kvn@5110: } kvn@5110: #endif kvn@5110: kvn@5110: return true; kvn@5110: } duke@435: duke@435: //---------------------------set_eden_pointers------------------------- duke@435: void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) { duke@435: if (UseTLAB) { // Private allocation: load from TLS kvn@4115: Node* thread = transform_later(new (C) ThreadLocalNode()); duke@435: int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset()); duke@435: int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset()); duke@435: eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset); duke@435: eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset); duke@435: } else { // Shared allocation: load from globals duke@435: CollectedHeap* ch = Universe::heap(); duke@435: address top_adr = (address)ch->top_addr(); duke@435: address end_adr = (address)ch->end_addr(); duke@435: eden_top_adr = makecon(TypeRawPtr::make(top_adr)); duke@435: eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr); duke@435: } duke@435: } duke@435: duke@435: duke@435: Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { duke@435: Node* adr = basic_plus_adr(base, offset); kvn@855: const TypePtr* adr_type = adr->bottom_type()->is_ptr(); goetz@6479: Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered); duke@435: transform_later(value); duke@435: return value; duke@435: } duke@435: duke@435: duke@435: Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { duke@435: Node* adr = basic_plus_adr(base, offset); goetz@6479: mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered); duke@435: transform_later(mem); duke@435: return mem; duke@435: } duke@435: duke@435: //============================================================================= duke@435: // duke@435: // A L L O C A T I O N duke@435: // duke@435: // Allocation attempts to be fast in the case of frequent small objects. duke@435: // It breaks down like this: duke@435: // duke@435: // 1) Size in doublewords is computed. This is a constant for objects and duke@435: // variable for most arrays. Doubleword units are used to avoid size duke@435: // overflow of huge doubleword arrays. We need doublewords in the end for duke@435: // rounding. duke@435: // duke@435: // 2) Size is checked for being 'too large'. Too-large allocations will go duke@435: // the slow path into the VM. The slow path can throw any required duke@435: // exceptions, and does all the special checks for very large arrays. The duke@435: // size test can constant-fold away for objects. For objects with duke@435: // finalizers it constant-folds the otherway: you always go slow with duke@435: // finalizers. duke@435: // duke@435: // 3) If NOT using TLABs, this is the contended loop-back point. duke@435: // Load-Locked the heap top. If using TLABs normal-load the heap top. duke@435: // duke@435: // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. duke@435: // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish duke@435: // "size*8" we always enter the VM, where "largish" is a constant picked small duke@435: // enough that there's always space between the eden max and 4Gig (old space is duke@435: // there so it's quite large) and large enough that the cost of entering the VM duke@435: // is dwarfed by the cost to initialize the space. duke@435: // duke@435: // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back duke@435: // down. If contended, repeat at step 3. If using TLABs normal-store duke@435: // adjusted heap top back down; there is no contention. duke@435: // duke@435: // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark duke@435: // fields. duke@435: // duke@435: // 7) Merge with the slow-path; cast the raw memory pointer to the correct duke@435: // oop flavor. duke@435: // duke@435: //============================================================================= duke@435: // FastAllocateSizeLimit value is in DOUBLEWORDS. duke@435: // Allocations bigger than this always go the slow route. duke@435: // This value must be small enough that allocation attempts that need to duke@435: // trigger exceptions go the slow route. Also, it must be small enough so duke@435: // that heap_top + size_in_bytes does not wrap around the 4Gig limit. duke@435: //=============================================================================j// duke@435: // %%% Here is an old comment from parseHelper.cpp; is it outdated? duke@435: // The allocator will coalesce int->oop copies away. See comment in duke@435: // coalesce.cpp about how this works. It depends critically on the exact duke@435: // code shape produced here, so if you are changing this code shape duke@435: // make sure the GC info for the heap-top is correct in and around the duke@435: // slow-path call. duke@435: // duke@435: duke@435: void PhaseMacroExpand::expand_allocate_common( duke@435: AllocateNode* alloc, // allocation node to be expanded duke@435: Node* length, // array length for an array allocation duke@435: const TypeFunc* slow_call_type, // Type of slow call duke@435: address slow_call_address // Address of slow call duke@435: ) duke@435: { duke@435: duke@435: Node* ctrl = alloc->in(TypeFunc::Control); duke@435: Node* mem = alloc->in(TypeFunc::Memory); duke@435: Node* i_o = alloc->in(TypeFunc::I_O); duke@435: Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); duke@435: Node* klass_node = alloc->in(AllocateNode::KlassNode); duke@435: Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); duke@435: duke@435: assert(ctrl != NULL, "must have control"); duke@435: // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. duke@435: // they will not be used if "always_slow" is set duke@435: enum { slow_result_path = 1, fast_result_path = 2 }; duke@435: Node *result_region; duke@435: Node *result_phi_rawmem; duke@435: Node *result_phi_rawoop; duke@435: Node *result_phi_i_o; duke@435: duke@435: // The initial slow comparison is a size check, the comparison duke@435: // we want to do is a BoolTest::gt duke@435: bool always_slow = false; duke@435: int tv = _igvn.find_int_con(initial_slow_test, -1); duke@435: if (tv >= 0) { duke@435: always_slow = (tv == 1); duke@435: initial_slow_test = NULL; duke@435: } else { duke@435: initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); duke@435: } duke@435: kvn@1215: if (C->env()->dtrace_alloc_probes() || ysr@777: !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc() || ysr@777: (UseConcMarkSweepGC && CMSIncrementalMode))) { duke@435: // Force slow-path allocation duke@435: always_slow = true; duke@435: initial_slow_test = NULL; duke@435: } duke@435: ysr@777: duke@435: enum { too_big_or_final_path = 1, need_gc_path = 2 }; duke@435: Node *slow_region = NULL; duke@435: Node *toobig_false = ctrl; duke@435: duke@435: assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent"); duke@435: // generate the initial test if necessary duke@435: if (initial_slow_test != NULL ) { kvn@4115: slow_region = new (C) RegionNode(3); duke@435: duke@435: // Now make the initial failure test. Usually a too-big test but duke@435: // might be a TRUE for finalizers or a fancy class check for duke@435: // newInstance0. kvn@4115: IfNode *toobig_iff = new (C) IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); duke@435: transform_later(toobig_iff); duke@435: // Plug the failing-too-big test into the slow-path region kvn@4115: Node *toobig_true = new (C) IfTrueNode( toobig_iff ); duke@435: transform_later(toobig_true); duke@435: slow_region ->init_req( too_big_or_final_path, toobig_true ); kvn@4115: toobig_false = new (C) IfFalseNode( toobig_iff ); duke@435: transform_later(toobig_false); duke@435: } else { // No initial test, just fall into next case duke@435: toobig_false = ctrl; duke@435: debug_only(slow_region = NodeSentinel); duke@435: } duke@435: duke@435: Node *slow_mem = mem; // save the current memory state for slow path duke@435: // generate the fast allocation code unless we know that the initial test will always go slow duke@435: if (!always_slow) { kvn@1000: // Fast path modifies only raw memory. kvn@1000: if (mem->is_MergeMem()) { kvn@1000: mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); kvn@1000: } kvn@1000: ysr@777: Node* eden_top_adr; ysr@777: Node* eden_end_adr; ysr@777: ysr@777: set_eden_pointers(eden_top_adr, eden_end_adr); ysr@777: ysr@777: // Load Eden::end. Loop invariant and hoisted. ysr@777: // ysr@777: // Note: We set the control input on "eden_end" and "old_eden_top" when using ysr@777: // a TLAB to work around a bug where these values were being moved across ysr@777: // a safepoint. These are not oops, so they cannot be include in the oop phh@2423: // map, but they can be changed by a GC. The proper way to fix this would ysr@777: // be to set the raw memory state when generating a SafepointNode. However ysr@777: // this will require extensive changes to the loop optimization in order to ysr@777: // prevent a degradation of the optimization. ysr@777: // See comment in memnode.hpp, around line 227 in class LoadPNode. ysr@777: Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); ysr@777: duke@435: // allocate the Region and Phi nodes for the result kvn@4115: result_region = new (C) RegionNode(3); kvn@4115: result_phi_rawmem = new (C) PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM); kvn@4115: result_phi_rawoop = new (C) PhiNode(result_region, TypeRawPtr::BOTTOM); kvn@4115: result_phi_i_o = new (C) PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch duke@435: duke@435: // We need a Region for the loop-back contended case. duke@435: enum { fall_in_path = 1, contended_loopback_path = 2 }; duke@435: Node *contended_region; duke@435: Node *contended_phi_rawmem; phh@2423: if (UseTLAB) { duke@435: contended_region = toobig_false; duke@435: contended_phi_rawmem = mem; duke@435: } else { kvn@4115: contended_region = new (C) RegionNode(3); kvn@4115: contended_phi_rawmem = new (C) PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM); duke@435: // Now handle the passing-too-big test. We fall into the contended duke@435: // loop-back merge point. phh@2423: contended_region ->init_req(fall_in_path, toobig_false); phh@2423: contended_phi_rawmem->init_req(fall_in_path, mem); duke@435: transform_later(contended_region); duke@435: transform_later(contended_phi_rawmem); duke@435: } duke@435: duke@435: // Load(-locked) the heap top. duke@435: // See note above concerning the control input when using a TLAB duke@435: Node *old_eden_top = UseTLAB goetz@6479: ? new (C) LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered) goetz@6479: : new (C) LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire); duke@435: duke@435: transform_later(old_eden_top); duke@435: // Add to heap top to get a new heap top kvn@4115: Node *new_eden_top = new (C) AddPNode(top(), old_eden_top, size_in_bytes); duke@435: transform_later(new_eden_top); duke@435: // Check for needing a GC; compare against heap end kvn@4115: Node *needgc_cmp = new (C) CmpPNode(new_eden_top, eden_end); duke@435: transform_later(needgc_cmp); kvn@4115: Node *needgc_bol = new (C) BoolNode(needgc_cmp, BoolTest::ge); duke@435: transform_later(needgc_bol); kvn@4115: IfNode *needgc_iff = new (C) IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN); duke@435: transform_later(needgc_iff); duke@435: duke@435: // Plug the failing-heap-space-need-gc test into the slow-path region kvn@4115: Node *needgc_true = new (C) IfTrueNode(needgc_iff); duke@435: transform_later(needgc_true); phh@2423: if (initial_slow_test) { phh@2423: slow_region->init_req(need_gc_path, needgc_true); duke@435: // This completes all paths into the slow merge point duke@435: transform_later(slow_region); duke@435: } else { // No initial slow path needed! duke@435: // Just fall from the need-GC path straight into the VM call. phh@2423: slow_region = needgc_true; duke@435: } duke@435: // No need for a GC. Setup for the Store-Conditional kvn@4115: Node *needgc_false = new (C) IfFalseNode(needgc_iff); duke@435: transform_later(needgc_false); duke@435: duke@435: // Grab regular I/O before optional prefetch may change it. duke@435: // Slow-path does no I/O so just set it to the original I/O. phh@2423: result_phi_i_o->init_req(slow_result_path, i_o); duke@435: duke@435: i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem, duke@435: old_eden_top, new_eden_top, length); duke@435: phh@2423: // Name successful fast-path variables phh@2423: Node* fast_oop = old_eden_top; phh@2423: Node* fast_oop_ctrl; phh@2423: Node* fast_oop_rawmem; phh@2423: duke@435: // Store (-conditional) the modified eden top back down. duke@435: // StorePConditional produces flags for a test PLUS a modified raw duke@435: // memory state. phh@2423: if (UseTLAB) { phh@2423: Node* store_eden_top = kvn@4115: new (C) StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr, goetz@6479: TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered); duke@435: transform_later(store_eden_top); duke@435: fast_oop_ctrl = needgc_false; // No contention, so this is the fast path phh@2423: fast_oop_rawmem = store_eden_top; duke@435: } else { phh@2423: Node* store_eden_top = kvn@4115: new (C) StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr, phh@2423: new_eden_top, fast_oop/*old_eden_top*/); duke@435: transform_later(store_eden_top); kvn@4115: Node *contention_check = new (C) BoolNode(store_eden_top, BoolTest::ne); duke@435: transform_later(contention_check); kvn@4115: store_eden_top = new (C) SCMemProjNode(store_eden_top); duke@435: transform_later(store_eden_top); duke@435: duke@435: // If not using TLABs, check to see if there was contention. kvn@4115: IfNode *contention_iff = new (C) IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN); duke@435: transform_later(contention_iff); kvn@4115: Node *contention_true = new (C) IfTrueNode(contention_iff); duke@435: transform_later(contention_true); duke@435: // If contention, loopback and try again. phh@2423: contended_region->init_req(contended_loopback_path, contention_true); phh@2423: contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top); duke@435: duke@435: // Fast-path succeeded with no contention! kvn@4115: Node *contention_false = new (C) IfFalseNode(contention_iff); duke@435: transform_later(contention_false); duke@435: fast_oop_ctrl = contention_false; phh@2423: phh@2423: // Bump total allocated bytes for this thread kvn@4115: Node* thread = new (C) ThreadLocalNode(); phh@2423: transform_later(thread); phh@2423: Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread, phh@2423: in_bytes(JavaThread::allocated_bytes_offset())); phh@2423: Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, phh@2423: 0, TypeLong::LONG, T_LONG); phh@2423: #ifdef _LP64 phh@2423: Node* alloc_size = size_in_bytes; phh@2423: #else kvn@4115: Node* alloc_size = new (C) ConvI2LNode(size_in_bytes); phh@2423: transform_later(alloc_size); phh@2423: #endif kvn@4115: Node* new_alloc_bytes = new (C) AddLNode(alloc_bytes, alloc_size); phh@2423: transform_later(new_alloc_bytes); phh@2423: fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, phh@2423: 0, new_alloc_bytes, T_LONG); duke@435: } duke@435: roland@3392: InitializeNode* init = alloc->initialization(); duke@435: fast_oop_rawmem = initialize_object(alloc, duke@435: fast_oop_ctrl, fast_oop_rawmem, fast_oop, duke@435: klass_node, length, size_in_bytes); duke@435: roland@3392: // If initialization is performed by an array copy, any required roland@3392: // MemBarStoreStore was already added. If the object does not roland@3392: // escape no need for a MemBarStoreStore. Otherwise we need a roland@3392: // MemBarStoreStore so that stores that initialize this object roland@3392: // can't be reordered with a subsequent store that makes this roland@3392: // object accessible by other threads. roland@3392: if (init == NULL || (!init->is_complete_with_arraycopy() && !init->does_not_escape())) { roland@3392: if (init == NULL || init->req() < InitializeNode::RawStores) { roland@3392: // No InitializeNode or no stores captured by zeroing roland@3392: // elimination. Simply add the MemBarStoreStore after object roland@3392: // initialization. roland@4694: MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); roland@3392: transform_later(mb); roland@3392: roland@3392: mb->init_req(TypeFunc::Memory, fast_oop_rawmem); roland@3392: mb->init_req(TypeFunc::Control, fast_oop_ctrl); kvn@4115: fast_oop_ctrl = new (C) ProjNode(mb,TypeFunc::Control); roland@3392: transform_later(fast_oop_ctrl); kvn@4115: fast_oop_rawmem = new (C) ProjNode(mb,TypeFunc::Memory); roland@3392: transform_later(fast_oop_rawmem); roland@3392: } else { roland@3392: // Add the MemBarStoreStore after the InitializeNode so that roland@3392: // all stores performing the initialization that were moved roland@3392: // before the InitializeNode happen before the storestore roland@3392: // barrier. roland@3392: roland@3392: Node* init_ctrl = init->proj_out(TypeFunc::Control); roland@3392: Node* init_mem = init->proj_out(TypeFunc::Memory); roland@3392: roland@3392: MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot); roland@3392: transform_later(mb); roland@3392: kvn@4115: Node* ctrl = new (C) ProjNode(init,TypeFunc::Control); roland@3392: transform_later(ctrl); kvn@4115: Node* mem = new (C) ProjNode(init,TypeFunc::Memory); roland@3392: transform_later(mem); roland@3392: roland@3392: // The MemBarStoreStore depends on control and memory coming roland@3392: // from the InitializeNode roland@3392: mb->init_req(TypeFunc::Memory, mem); roland@3392: mb->init_req(TypeFunc::Control, ctrl); roland@3392: kvn@4115: ctrl = new (C) ProjNode(mb,TypeFunc::Control); roland@3392: transform_later(ctrl); kvn@4115: mem = new (C) ProjNode(mb,TypeFunc::Memory); roland@3392: transform_later(mem); roland@3392: roland@3392: // All nodes that depended on the InitializeNode for control roland@3392: // and memory must now depend on the MemBarNode that itself roland@3392: // depends on the InitializeNode roland@3392: _igvn.replace_node(init_ctrl, ctrl); roland@3392: _igvn.replace_node(init_mem, mem); roland@3392: } roland@3392: } roland@3392: kvn@1215: if (C->env()->dtrace_extended_probes()) { duke@435: // Slow-path call duke@435: int size = TypeFunc::Parms + 2; kvn@4115: CallLeafNode *call = new (C) CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), kvn@4115: CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base), kvn@4115: "dtrace_object_alloc", kvn@4115: TypeRawPtr::BOTTOM); duke@435: duke@435: // Get base of thread-local storage area kvn@4115: Node* thread = new (C) ThreadLocalNode(); duke@435: transform_later(thread); duke@435: duke@435: call->init_req(TypeFunc::Parms+0, thread); duke@435: call->init_req(TypeFunc::Parms+1, fast_oop); phh@2423: call->init_req(TypeFunc::Control, fast_oop_ctrl); phh@2423: call->init_req(TypeFunc::I_O , top()); // does no i/o phh@2423: call->init_req(TypeFunc::Memory , fast_oop_rawmem); phh@2423: call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); phh@2423: call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); duke@435: transform_later(call); kvn@4115: fast_oop_ctrl = new (C) ProjNode(call,TypeFunc::Control); duke@435: transform_later(fast_oop_ctrl); kvn@4115: fast_oop_rawmem = new (C) ProjNode(call,TypeFunc::Memory); duke@435: transform_later(fast_oop_rawmem); duke@435: } duke@435: duke@435: // Plug in the successful fast-path into the result merge point phh@2423: result_region ->init_req(fast_result_path, fast_oop_ctrl); phh@2423: result_phi_rawoop->init_req(fast_result_path, fast_oop); phh@2423: result_phi_i_o ->init_req(fast_result_path, i_o); phh@2423: result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem); duke@435: } else { duke@435: slow_region = ctrl; kvn@3396: result_phi_i_o = i_o; // Rename it to use in the following code. duke@435: } duke@435: duke@435: // Generate slow-path call kvn@4115: CallNode *call = new (C) CallStaticJavaNode(slow_call_type, slow_call_address, kvn@4115: OptoRuntime::stub_name(slow_call_address), kvn@4115: alloc->jvms()->bci(), kvn@4115: TypePtr::BOTTOM); duke@435: call->init_req( TypeFunc::Control, slow_region ); duke@435: call->init_req( TypeFunc::I_O , top() ) ; // does no i/o duke@435: call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs duke@435: call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); duke@435: call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); duke@435: duke@435: call->init_req(TypeFunc::Parms+0, klass_node); duke@435: if (length != NULL) { duke@435: call->init_req(TypeFunc::Parms+1, length); duke@435: } duke@435: duke@435: // Copy debug information and adjust JVMState information, then replace duke@435: // allocate node with the call duke@435: copy_call_debug_info((CallNode *) alloc, call); duke@435: if (!always_slow) { duke@435: call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. kvn@3396: } else { kvn@3396: // Hook i_o projection to avoid its elimination during allocation kvn@3396: // replacement (when only a slow call is generated). kvn@3396: call->set_req(TypeFunc::I_O, result_phi_i_o); duke@435: } kvn@1976: _igvn.replace_node(alloc, call); duke@435: transform_later(call); duke@435: duke@435: // Identify the output projections from the allocate node and duke@435: // adjust any references to them. duke@435: // The control and io projections look like: duke@435: // duke@435: // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) duke@435: // Allocate Catch duke@435: // ^---Proj(io) <-------+ ^---CatchProj(io) duke@435: // duke@435: // We are interested in the CatchProj nodes. duke@435: // duke@435: extract_call_projections(call); duke@435: kvn@3396: // An allocate node has separate memory projections for the uses on kvn@3396: // the control and i_o paths. Replace the control memory projection with kvn@3396: // result_phi_rawmem (unless we are only generating a slow call when kvn@3396: // both memory projections are combined) duke@435: if (!always_slow && _memproj_fallthrough != NULL) { duke@435: for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) { duke@435: Node *use = _memproj_fallthrough->fast_out(i); kvn@3847: _igvn.rehash_node_delayed(use); duke@435: imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem); duke@435: // back up iterator duke@435: --i; duke@435: } duke@435: } kvn@3396: // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete kvn@3396: // _memproj_catchall so we end up with a call that has only 1 memory projection. duke@435: if (_memproj_catchall != NULL ) { duke@435: if (_memproj_fallthrough == NULL) { kvn@4115: _memproj_fallthrough = new (C) ProjNode(call, TypeFunc::Memory); duke@435: transform_later(_memproj_fallthrough); duke@435: } duke@435: for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) { duke@435: Node *use = _memproj_catchall->fast_out(i); kvn@3847: _igvn.rehash_node_delayed(use); duke@435: imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough); duke@435: // back up iterator duke@435: --i; duke@435: } kvn@3396: assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted"); kvn@3396: _igvn.remove_dead_node(_memproj_catchall); duke@435: } duke@435: kvn@3396: // An allocate node has separate i_o projections for the uses on the control kvn@3396: // and i_o paths. Always replace the control i_o projection with result i_o kvn@3396: // otherwise incoming i_o become dead when only a slow call is generated kvn@3396: // (it is different from memory projections where both projections are kvn@3396: // combined in such case). kvn@3396: if (_ioproj_fallthrough != NULL) { duke@435: for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) { duke@435: Node *use = _ioproj_fallthrough->fast_out(i); kvn@3847: _igvn.rehash_node_delayed(use); duke@435: imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o); duke@435: // back up iterator duke@435: --i; duke@435: } duke@435: } kvn@3396: // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete kvn@3396: // _ioproj_catchall so we end up with a call that has only 1 i_o projection. duke@435: if (_ioproj_catchall != NULL ) { kvn@3396: if (_ioproj_fallthrough == NULL) { kvn@4115: _ioproj_fallthrough = new (C) ProjNode(call, TypeFunc::I_O); kvn@3396: transform_later(_ioproj_fallthrough); kvn@3396: } duke@435: for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) { duke@435: Node *use = _ioproj_catchall->fast_out(i); kvn@3847: _igvn.rehash_node_delayed(use); duke@435: imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough); duke@435: // back up iterator duke@435: --i; duke@435: } kvn@3396: assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted"); kvn@3396: _igvn.remove_dead_node(_ioproj_catchall); duke@435: } duke@435: duke@435: // if we generated only a slow call, we are done kvn@3396: if (always_slow) { kvn@3396: // Now we can unhook i_o. kvn@3398: if (result_phi_i_o->outcnt() > 1) { kvn@3398: call->set_req(TypeFunc::I_O, top()); kvn@3398: } else { kvn@3398: assert(result_phi_i_o->unique_ctrl_out() == call, ""); kvn@3398: // Case of new array with negative size known during compilation. kvn@3398: // AllocateArrayNode::Ideal() optimization disconnect unreachable kvn@3398: // following code since call to runtime will throw exception. kvn@3398: // As result there will be no users of i_o after the call. kvn@3398: // Leave i_o attached to this call to avoid problems in preceding graph. kvn@3398: } duke@435: return; kvn@3396: } duke@435: duke@435: duke@435: if (_fallthroughcatchproj != NULL) { duke@435: ctrl = _fallthroughcatchproj->clone(); duke@435: transform_later(ctrl); kvn@1143: _igvn.replace_node(_fallthroughcatchproj, result_region); duke@435: } else { duke@435: ctrl = top(); duke@435: } duke@435: Node *slow_result; duke@435: if (_resproj == NULL) { duke@435: // no uses of the allocation result duke@435: slow_result = top(); duke@435: } else { duke@435: slow_result = _resproj->clone(); duke@435: transform_later(slow_result); kvn@1143: _igvn.replace_node(_resproj, result_phi_rawoop); duke@435: } duke@435: duke@435: // Plug slow-path into result merge point duke@435: result_region ->init_req( slow_result_path, ctrl ); duke@435: result_phi_rawoop->init_req( slow_result_path, slow_result); duke@435: result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough ); duke@435: transform_later(result_region); duke@435: transform_later(result_phi_rawoop); duke@435: transform_later(result_phi_rawmem); duke@435: transform_later(result_phi_i_o); duke@435: // This completes all paths into the result merge point duke@435: } duke@435: duke@435: duke@435: // Helper for PhaseMacroExpand::expand_allocate_common. duke@435: // Initializes the newly-allocated storage. duke@435: Node* duke@435: PhaseMacroExpand::initialize_object(AllocateNode* alloc, duke@435: Node* control, Node* rawmem, Node* object, duke@435: Node* klass_node, Node* length, duke@435: Node* size_in_bytes) { duke@435: InitializeNode* init = alloc->initialization(); duke@435: // Store the klass & mark bits duke@435: Node* mark_node = NULL; duke@435: // For now only enable fast locking for non-array types duke@435: if (UseBiasedLocking && (length == NULL)) { stefank@3391: mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS); duke@435: } else { duke@435: mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype())); duke@435: } duke@435: rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); coleenp@548: coleenp@4037: rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA); duke@435: int header_size = alloc->minimum_header_size(); // conservatively small duke@435: duke@435: // Array length duke@435: if (length != NULL) { // Arrays need length field duke@435: rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); duke@435: // conservatively small header size: coleenp@548: header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); duke@435: ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); duke@435: if (k->is_array_klass()) // we know the exact header size in most cases: duke@435: header_size = Klass::layout_helper_header_size(k->layout_helper()); duke@435: } duke@435: duke@435: // Clear the object body, if necessary. duke@435: if (init == NULL) { duke@435: // The init has somehow disappeared; be cautious and clear everything. duke@435: // duke@435: // This can happen if a node is allocated but an uncommon trap occurs duke@435: // immediately. In this case, the Initialize gets associated with the duke@435: // trap, and may be placed in a different (outer) loop, if the Allocate duke@435: // is in a loop. If (this is rare) the inner loop gets unrolled, then duke@435: // there can be two Allocates to one Initialize. The answer in all these duke@435: // edge cases is safety first. It is always safe to clear immediately duke@435: // within an Allocate, and then (maybe or maybe not) clear some more later. duke@435: if (!ZeroTLAB) duke@435: rawmem = ClearArrayNode::clear_memory(control, rawmem, object, duke@435: header_size, size_in_bytes, duke@435: &_igvn); duke@435: } else { duke@435: if (!init->is_complete()) { duke@435: // Try to win by zeroing only what the init does not store. duke@435: // We can also try to do some peephole optimizations, duke@435: // such as combining some adjacent subword stores. duke@435: rawmem = init->complete_stores(control, rawmem, object, duke@435: header_size, size_in_bytes, &_igvn); duke@435: } duke@435: // We have no more use for this link, since the AllocateNode goes away: duke@435: init->set_req(InitializeNode::RawAddress, top()); duke@435: // (If we keep the link, it just confuses the register allocator, duke@435: // who thinks he sees a real use of the address by the membar.) duke@435: } duke@435: duke@435: return rawmem; duke@435: } duke@435: duke@435: // Generate prefetch instructions for next allocations. duke@435: Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, duke@435: Node*& contended_phi_rawmem, duke@435: Node* old_eden_top, Node* new_eden_top, duke@435: Node* length) { kvn@1802: enum { fall_in_path = 1, pf_path = 2 }; duke@435: if( UseTLAB && AllocatePrefetchStyle == 2 ) { duke@435: // Generate prefetch allocation with watermark check. duke@435: // As an allocation hits the watermark, we will prefetch starting duke@435: // at a "distance" away from watermark. duke@435: kvn@4115: Node *pf_region = new (C) RegionNode(3); kvn@4115: Node *pf_phi_rawmem = new (C) PhiNode( pf_region, Type::MEMORY, duke@435: TypeRawPtr::BOTTOM ); duke@435: // I/O is used for Prefetch kvn@4115: Node *pf_phi_abio = new (C) PhiNode( pf_region, Type::ABIO ); duke@435: kvn@4115: Node *thread = new (C) ThreadLocalNode(); duke@435: transform_later(thread); duke@435: kvn@4115: Node *eden_pf_adr = new (C) AddPNode( top()/*not oop*/, thread, duke@435: _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); duke@435: transform_later(eden_pf_adr); duke@435: goetz@6479: Node *old_pf_wm = new (C) LoadPNode(needgc_false, duke@435: contended_phi_rawmem, eden_pf_adr, goetz@6479: TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, goetz@6479: MemNode::unordered); duke@435: transform_later(old_pf_wm); duke@435: duke@435: // check against new_eden_top kvn@4115: Node *need_pf_cmp = new (C) CmpPNode( new_eden_top, old_pf_wm ); duke@435: transform_later(need_pf_cmp); kvn@4115: Node *need_pf_bol = new (C) BoolNode( need_pf_cmp, BoolTest::ge ); duke@435: transform_later(need_pf_bol); kvn@4115: IfNode *need_pf_iff = new (C) IfNode( needgc_false, need_pf_bol, duke@435: PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); duke@435: transform_later(need_pf_iff); duke@435: duke@435: // true node, add prefetchdistance kvn@4115: Node *need_pf_true = new (C) IfTrueNode( need_pf_iff ); duke@435: transform_later(need_pf_true); duke@435: kvn@4115: Node *need_pf_false = new (C) IfFalseNode( need_pf_iff ); duke@435: transform_later(need_pf_false); duke@435: kvn@4115: Node *new_pf_wmt = new (C) AddPNode( top(), old_pf_wm, duke@435: _igvn.MakeConX(AllocatePrefetchDistance) ); duke@435: transform_later(new_pf_wmt ); duke@435: new_pf_wmt->set_req(0, need_pf_true); duke@435: goetz@6479: Node *store_new_wmt = new (C) StorePNode(need_pf_true, duke@435: contended_phi_rawmem, eden_pf_adr, goetz@6479: TypeRawPtr::BOTTOM, new_pf_wmt, goetz@6479: MemNode::unordered); duke@435: transform_later(store_new_wmt); duke@435: duke@435: // adding prefetches duke@435: pf_phi_abio->init_req( fall_in_path, i_o ); duke@435: duke@435: Node *prefetch_adr; duke@435: Node *prefetch; duke@435: uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize; duke@435: uint step_size = AllocatePrefetchStepSize; duke@435: uint distance = 0; duke@435: duke@435: for ( uint i = 0; i < lines; i++ ) { kvn@4115: prefetch_adr = new (C) AddPNode( old_pf_wm, new_pf_wmt, duke@435: _igvn.MakeConX(distance) ); duke@435: transform_later(prefetch_adr); kvn@4115: prefetch = new (C) PrefetchAllocationNode( i_o, prefetch_adr ); duke@435: transform_later(prefetch); duke@435: distance += step_size; duke@435: i_o = prefetch; duke@435: } duke@435: pf_phi_abio->set_req( pf_path, i_o ); duke@435: duke@435: pf_region->init_req( fall_in_path, need_pf_false ); duke@435: pf_region->init_req( pf_path, need_pf_true ); duke@435: duke@435: pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); duke@435: pf_phi_rawmem->init_req( pf_path, store_new_wmt ); duke@435: duke@435: transform_later(pf_region); duke@435: transform_later(pf_phi_rawmem); duke@435: transform_later(pf_phi_abio); duke@435: duke@435: needgc_false = pf_region; duke@435: contended_phi_rawmem = pf_phi_rawmem; duke@435: i_o = pf_phi_abio; kvn@1802: } else if( UseTLAB && AllocatePrefetchStyle == 3 ) { kvn@3052: // Insert a prefetch for each allocation. kvn@3052: // This code is used for Sparc with BIS. kvn@4115: Node *pf_region = new (C) RegionNode(3); kvn@4115: Node *pf_phi_rawmem = new (C) PhiNode( pf_region, Type::MEMORY, kvn@4115: TypeRawPtr::BOTTOM ); kvn@1802: kvn@3052: // Generate several prefetch instructions. kvn@3052: uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; kvn@1802: uint step_size = AllocatePrefetchStepSize; kvn@1802: uint distance = AllocatePrefetchDistance; kvn@1802: kvn@1802: // Next cache address. kvn@4115: Node *cache_adr = new (C) AddPNode(old_eden_top, old_eden_top, kvn@1802: _igvn.MakeConX(distance)); kvn@1802: transform_later(cache_adr); kvn@4115: cache_adr = new (C) CastP2XNode(needgc_false, cache_adr); kvn@1802: transform_later(cache_adr); kvn@1802: Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1)); kvn@4115: cache_adr = new (C) AndXNode(cache_adr, mask); kvn@1802: transform_later(cache_adr); kvn@4115: cache_adr = new (C) CastX2PNode(cache_adr); kvn@1802: transform_later(cache_adr); kvn@1802: kvn@1802: // Prefetch kvn@4115: Node *prefetch = new (C) PrefetchAllocationNode( contended_phi_rawmem, cache_adr ); kvn@1802: prefetch->set_req(0, needgc_false); kvn@1802: transform_later(prefetch); kvn@1802: contended_phi_rawmem = prefetch; kvn@1802: Node *prefetch_adr; kvn@1802: distance = step_size; kvn@1802: for ( uint i = 1; i < lines; i++ ) { kvn@4115: prefetch_adr = new (C) AddPNode( cache_adr, cache_adr, kvn@1802: _igvn.MakeConX(distance) ); kvn@1802: transform_later(prefetch_adr); kvn@4115: prefetch = new (C) PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr ); kvn@1802: transform_later(prefetch); kvn@1802: distance += step_size; kvn@1802: contended_phi_rawmem = prefetch; kvn@1802: } duke@435: } else if( AllocatePrefetchStyle > 0 ) { duke@435: // Insert a prefetch for each allocation only on the fast-path duke@435: Node *prefetch_adr; duke@435: Node *prefetch; kvn@3052: // Generate several prefetch instructions. kvn@3052: uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines; duke@435: uint step_size = AllocatePrefetchStepSize; duke@435: uint distance = AllocatePrefetchDistance; duke@435: for ( uint i = 0; i < lines; i++ ) { kvn@4115: prefetch_adr = new (C) AddPNode( old_eden_top, new_eden_top, duke@435: _igvn.MakeConX(distance) ); duke@435: transform_later(prefetch_adr); kvn@4115: prefetch = new (C) PrefetchAllocationNode( i_o, prefetch_adr ); duke@435: // Do not let it float too high, since if eden_top == eden_end, duke@435: // both might be null. duke@435: if( i == 0 ) { // Set control for first prefetch, next follows it duke@435: prefetch->init_req(0, needgc_false); duke@435: } duke@435: transform_later(prefetch); duke@435: distance += step_size; duke@435: i_o = prefetch; duke@435: } duke@435: } duke@435: return i_o; duke@435: } duke@435: duke@435: duke@435: void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { duke@435: expand_allocate_common(alloc, NULL, duke@435: OptoRuntime::new_instance_Type(), duke@435: OptoRuntime::new_instance_Java()); duke@435: } duke@435: duke@435: void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { duke@435: Node* length = alloc->in(AllocateNode::ALength); kvn@3157: InitializeNode* init = alloc->initialization(); kvn@3157: Node* klass_node = alloc->in(AllocateNode::KlassNode); kvn@3157: ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); kvn@3157: address slow_call_address; // Address of slow call kvn@3157: if (init != NULL && init->is_complete_with_arraycopy() && kvn@3157: k->is_type_array_klass()) { kvn@3157: // Don't zero type array during slow allocation in VM since kvn@3157: // it will be initialized later by arraycopy in compiled code. kvn@3157: slow_call_address = OptoRuntime::new_array_nozero_Java(); kvn@3157: } else { kvn@3157: slow_call_address = OptoRuntime::new_array_Java(); kvn@3157: } duke@435: expand_allocate_common(alloc, length, duke@435: OptoRuntime::new_array_Type(), kvn@3157: slow_call_address); duke@435: } duke@435: kvn@3406: //-------------------mark_eliminated_box---------------------------------- kvn@3406: // kvn@2951: // During EA obj may point to several objects but after few ideal graph kvn@2951: // transformations (CCP) it may point to only one non escaping object kvn@2951: // (but still using phi), corresponding locks and unlocks will be marked kvn@2951: // for elimination. Later obj could be replaced with a new node (new phi) kvn@2951: // and which does not have escape information. And later after some graph kvn@2951: // reshape other locks and unlocks (which were not marked for elimination kvn@2951: // before) are connected to this new obj (phi) but they still will not be kvn@2951: // marked for elimination since new obj has no escape information. kvn@2951: // Mark all associated (same box and obj) lock and unlock nodes for kvn@2951: // elimination if some of them marked already. kvn@3406: void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) { kvn@3421: if (oldbox->as_BoxLock()->is_eliminated()) kvn@3421: return; // This BoxLock node was processed already. kvn@3406: kvn@3421: // New implementation (EliminateNestedLocks) has separate BoxLock kvn@3421: // node for each locked region so mark all associated locks/unlocks as kvn@3421: // eliminated even if different objects are referenced in one locked region kvn@3421: // (for example, OSR compilation of nested loop inside locked scope). kvn@3421: if (EliminateNestedLocks || kvn@3406: oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) { kvn@3406: // Box is used only in one lock region. Mark this box as eliminated. kvn@3406: _igvn.hash_delete(oldbox); kvn@3406: oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value kvn@3406: _igvn.hash_insert(oldbox); kvn@3406: kvn@3406: for (uint i = 0; i < oldbox->outcnt(); i++) { kvn@3406: Node* u = oldbox->raw_out(i); kvn@3406: if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) { kvn@3406: AbstractLockNode* alock = u->as_AbstractLock(); kvn@3406: // Check lock's box since box could be referenced by Lock's debug info. kvn@3406: if (alock->box_node() == oldbox) { kvn@3406: // Mark eliminated all related locks and unlocks. kvn@3406: alock->set_non_esc_obj(); kvn@3406: } kvn@3406: } kvn@3406: } kvn@2951: return; kvn@501: } kvn@3406: kvn@3406: // Create new "eliminated" BoxLock node and use it in monitor debug info kvn@3406: // instead of oldbox for the same object. kvn@3419: BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); kvn@3406: kvn@3406: // Note: BoxLock node is marked eliminated only here and it is used kvn@3406: // to indicate that all associated lock and unlock nodes are marked kvn@3406: // for elimination. kvn@3406: newbox->set_eliminated(); kvn@3406: transform_later(newbox); kvn@3406: kvn@3406: // Replace old box node with new box for all users of the same object. kvn@3406: for (uint i = 0; i < oldbox->outcnt();) { kvn@3406: bool next_edge = true; kvn@3406: kvn@3406: Node* u = oldbox->raw_out(i); kvn@3406: if (u->is_AbstractLock()) { kvn@3406: AbstractLockNode* alock = u->as_AbstractLock(); kvn@3407: if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) { kvn@3406: // Replace Box and mark eliminated all related locks and unlocks. kvn@3406: alock->set_non_esc_obj(); kvn@3847: _igvn.rehash_node_delayed(alock); kvn@3406: alock->set_box_node(newbox); kvn@3406: next_edge = false; kvn@3406: } kvn@3406: } kvn@3407: if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) { kvn@3406: FastLockNode* flock = u->as_FastLock(); kvn@3406: assert(flock->box_node() == oldbox, "sanity"); kvn@3847: _igvn.rehash_node_delayed(flock); kvn@3406: flock->set_box_node(newbox); kvn@3406: next_edge = false; kvn@3406: } kvn@3406: kvn@3406: // Replace old box in monitor debug info. kvn@3406: if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { kvn@3406: SafePointNode* sfn = u->as_SafePoint(); kvn@3406: JVMState* youngest_jvms = sfn->jvms(); kvn@3406: int max_depth = youngest_jvms->depth(); kvn@3406: for (int depth = 1; depth <= max_depth; depth++) { kvn@3406: JVMState* jvms = youngest_jvms->of_depth(depth); kvn@3406: int num_mon = jvms->nof_monitors(); kvn@3406: // Loop over monitors kvn@3406: for (int idx = 0; idx < num_mon; idx++) { kvn@3406: Node* obj_node = sfn->monitor_obj(jvms, idx); kvn@3406: Node* box_node = sfn->monitor_box(jvms, idx); kvn@3407: if (box_node == oldbox && obj_node->eqv_uncast(obj)) { kvn@3406: int j = jvms->monitor_box_offset(idx); kvn@3847: _igvn.replace_input_of(u, j, newbox); kvn@3406: next_edge = false; kvn@3406: } kvn@3406: } kvn@3406: } kvn@3406: } kvn@3406: if (next_edge) i++; kvn@3406: } kvn@3406: } kvn@3406: kvn@3406: //-----------------------mark_eliminated_locking_nodes----------------------- kvn@3406: void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) { kvn@3406: if (EliminateNestedLocks) { kvn@3406: if (alock->is_nested()) { kvn@3406: assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity"); kvn@3406: return; kvn@3406: } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened kvn@3406: // Only Lock node has JVMState needed here. kvn@3406: if (alock->jvms() != NULL && alock->as_Lock()->is_nested_lock_region()) { kvn@3406: // Mark eliminated related nested locks and unlocks. kvn@3406: Node* obj = alock->obj_node(); kvn@3406: BoxLockNode* box_node = alock->box_node()->as_BoxLock(); kvn@3406: assert(!box_node->is_eliminated(), "should not be marked yet"); kvn@2951: // Note: BoxLock node is marked eliminated only here kvn@2951: // and it is used to indicate that all associated lock kvn@2951: // and unlock nodes are marked for elimination. kvn@3406: box_node->set_eliminated(); // Box's hash is always NO_HASH here kvn@3406: for (uint i = 0; i < box_node->outcnt(); i++) { kvn@3406: Node* u = box_node->raw_out(i); kvn@3406: if (u->is_AbstractLock()) { kvn@3406: alock = u->as_AbstractLock(); kvn@3406: if (alock->box_node() == box_node) { kvn@3406: // Verify that this Box is referenced only by related locks. kvn@3407: assert(alock->obj_node()->eqv_uncast(obj), ""); kvn@3406: // Mark all related locks and unlocks. kvn@3406: alock->set_nested(); kvn@3406: } kvn@3406: } kvn@3406: } kvn@3406: } kvn@3406: return; kvn@3406: } kvn@3406: // Process locks for non escaping object kvn@3406: assert(alock->is_non_esc_obj(), ""); kvn@3406: } // EliminateNestedLocks kvn@895: kvn@3406: if (alock->is_non_esc_obj()) { // Lock is used for non escaping object kvn@3406: // Look for all locks of this object and mark them and kvn@3406: // corresponding BoxLock nodes as eliminated. kvn@3406: Node* obj = alock->obj_node(); kvn@3406: for (uint j = 0; j < obj->outcnt(); j++) { kvn@3406: Node* o = obj->raw_out(j); kvn@3407: if (o->is_AbstractLock() && kvn@3407: o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) { kvn@3406: alock = o->as_AbstractLock(); kvn@3406: Node* box = alock->box_node(); kvn@3406: // Replace old box node with new eliminated box for all users kvn@3406: // of the same object and mark related locks as eliminated. kvn@3406: mark_eliminated_box(box, obj); kvn@3406: } kvn@3406: } kvn@3406: } kvn@2951: } kvn@501: kvn@2951: // we have determined that this lock/unlock can be eliminated, we simply kvn@2951: // eliminate the node without expanding it. kvn@2951: // kvn@2951: // Note: The membar's associated with the lock/unlock are currently not kvn@2951: // eliminated. This should be investigated as a future enhancement. kvn@2951: // kvn@2951: bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { kvn@2951: kvn@2951: if (!alock->is_eliminated()) { kvn@2951: return false; kvn@2951: } kvn@2951: #ifdef ASSERT kvn@3406: if (!alock->is_coarsened()) { kvn@2951: // Check that new "eliminated" BoxLock node is created. kvn@2951: BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); kvn@2951: assert(oldbox->is_eliminated(), "should be done already"); kvn@2951: } kvn@2951: #endif never@1515: CompileLog* log = C->log(); never@1515: if (log != NULL) { never@1515: log->head("eliminate_lock lock='%d'", never@1515: alock->is_Lock()); never@1515: JVMState* p = alock->jvms(); never@1515: while (p != NULL) { never@1515: log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); never@1515: p = p->caller(); never@1515: } never@1515: log->tail("eliminate_lock"); never@1515: } never@1515: kvn@501: #ifndef PRODUCT kvn@501: if (PrintEliminateLocks) { kvn@501: if (alock->is_Lock()) { kvn@3311: tty->print_cr("++++ Eliminated: %d Lock", alock->_idx); kvn@501: } else { kvn@3311: tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx); kvn@501: } kvn@501: } kvn@501: #endif kvn@501: kvn@501: Node* mem = alock->in(TypeFunc::Memory); kvn@501: Node* ctrl = alock->in(TypeFunc::Control); kvn@501: kvn@501: extract_call_projections(alock); kvn@501: // There are 2 projections from the lock. The lock node will kvn@501: // be deleted when its last use is subsumed below. kvn@501: assert(alock->outcnt() == 2 && kvn@501: _fallthroughproj != NULL && kvn@501: _memproj_fallthrough != NULL, kvn@501: "Unexpected projections from Lock/Unlock"); kvn@501: kvn@501: Node* fallthroughproj = _fallthroughproj; kvn@501: Node* memproj_fallthrough = _memproj_fallthrough; duke@435: duke@435: // The memory projection from a lock/unlock is RawMem duke@435: // The input to a Lock is merged memory, so extract its RawMem input duke@435: // (unless the MergeMem has been optimized away.) duke@435: if (alock->is_Lock()) { roland@3047: // Seach for MemBarAcquireLock node and delete it also. kvn@501: MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); roland@3047: assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, ""); kvn@501: Node* ctrlproj = membar->proj_out(TypeFunc::Control); kvn@501: Node* memproj = membar->proj_out(TypeFunc::Memory); kvn@1143: _igvn.replace_node(ctrlproj, fallthroughproj); kvn@1143: _igvn.replace_node(memproj, memproj_fallthrough); kvn@895: kvn@895: // Delete FastLock node also if this Lock node is unique user kvn@895: // (a loop peeling may clone a Lock node). kvn@895: Node* flock = alock->as_Lock()->fastlock_node(); kvn@895: if (flock->outcnt() == 1) { kvn@895: assert(flock->unique_out() == alock, "sanity"); kvn@1143: _igvn.replace_node(flock, top()); kvn@895: } duke@435: } duke@435: roland@3047: // Seach for MemBarReleaseLock node and delete it also. kvn@501: if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() && kvn@501: ctrl->in(0)->is_MemBar()) { kvn@501: MemBarNode* membar = ctrl->in(0)->as_MemBar(); roland@3047: assert(membar->Opcode() == Op_MemBarReleaseLock && kvn@501: mem->is_Proj() && membar == mem->in(0), ""); kvn@1143: _igvn.replace_node(fallthroughproj, ctrl); kvn@1143: _igvn.replace_node(memproj_fallthrough, mem); kvn@501: fallthroughproj = ctrl; kvn@501: memproj_fallthrough = mem; kvn@501: ctrl = membar->in(TypeFunc::Control); kvn@501: mem = membar->in(TypeFunc::Memory); kvn@501: } kvn@501: kvn@1143: _igvn.replace_node(fallthroughproj, ctrl); kvn@1143: _igvn.replace_node(memproj_fallthrough, mem); kvn@501: return true; duke@435: } duke@435: duke@435: duke@435: //------------------------------expand_lock_node---------------------- duke@435: void PhaseMacroExpand::expand_lock_node(LockNode *lock) { duke@435: duke@435: Node* ctrl = lock->in(TypeFunc::Control); duke@435: Node* mem = lock->in(TypeFunc::Memory); duke@435: Node* obj = lock->obj_node(); duke@435: Node* box = lock->box_node(); kvn@501: Node* flock = lock->fastlock_node(); duke@435: kvn@3419: assert(!box->as_BoxLock()->is_eliminated(), "sanity"); kvn@3406: duke@435: // Make the merge point kvn@855: Node *region; kvn@855: Node *mem_phi; kvn@855: Node *slow_path; duke@435: kvn@855: if (UseOptoBiasInlining) { kvn@855: /* twisti@1040: * See the full description in MacroAssembler::biased_locking_enter(). kvn@855: * kvn@855: * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) { kvn@855: * // The object is biased. kvn@855: * proto_node = klass->prototype_header; kvn@855: * o_node = thread | proto_node; kvn@855: * x_node = o_node ^ mark_word; kvn@855: * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ? kvn@855: * // Done. kvn@855: * } else { kvn@855: * if( (x_node & biased_lock_mask) != 0 ) { kvn@855: * // The klass's prototype header is no longer biased. kvn@855: * cas(&mark_word, mark_word, proto_node) kvn@855: * goto cas_lock; kvn@855: * } else { kvn@855: * // The klass's prototype header is still biased. kvn@855: * if( (x_node & epoch_mask) != 0 ) { // Expired epoch? kvn@855: * old = mark_word; kvn@855: * new = o_node; kvn@855: * } else { kvn@855: * // Different thread or anonymous biased. kvn@855: * old = mark_word & (epoch_mask | age_mask | biased_lock_mask); kvn@855: * new = thread | old; kvn@855: * } kvn@855: * // Try to rebias. kvn@855: * if( cas(&mark_word, old, new) == 0 ) { kvn@855: * // Done. kvn@855: * } else { kvn@855: * goto slow_path; // Failed. kvn@855: * } kvn@855: * } kvn@855: * } kvn@855: * } else { kvn@855: * // The object is not biased. kvn@855: * cas_lock: kvn@855: * if( FastLock(obj) == 0 ) { kvn@855: * // Done. kvn@855: * } else { kvn@855: * slow_path: kvn@855: * OptoRuntime::complete_monitor_locking_Java(obj); kvn@855: * } kvn@855: * } kvn@855: */ kvn@855: kvn@4115: region = new (C) RegionNode(5); kvn@855: // create a Phi for the memory state kvn@4115: mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); kvn@855: kvn@4115: Node* fast_lock_region = new (C) RegionNode(3); kvn@4115: Node* fast_lock_mem_phi = new (C) PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM); kvn@855: kvn@855: // First, check mark word for the biased lock pattern. kvn@855: Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); kvn@855: kvn@855: // Get fast path - mark word has the biased lock pattern. kvn@855: ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node, kvn@855: markOopDesc::biased_lock_mask_in_place, kvn@855: markOopDesc::biased_lock_pattern, true); kvn@855: // fast_lock_region->in(1) is set to slow path. kvn@855: fast_lock_mem_phi->init_req(1, mem); kvn@855: kvn@855: // Now check that the lock is biased to the current thread and has kvn@855: // the same epoch and bias as Klass::_prototype_header. kvn@855: kvn@855: // Special-case a fresh allocation to avoid building nodes: kvn@855: Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn); kvn@855: if (klass_node == NULL) { kvn@855: Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); kvn@855: klass_node = transform_later( LoadKlassNode::make(_igvn, mem, k_adr, _igvn.type(k_adr)->is_ptr()) ); kvn@925: #ifdef _LP64 ehelin@5694: if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) { kvn@925: assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity"); kvn@925: klass_node->in(1)->init_req(0, ctrl); kvn@925: } else kvn@925: #endif kvn@925: klass_node->init_req(0, ctrl); kvn@855: } stefank@3391: Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type()); kvn@855: kvn@4115: Node* thread = transform_later(new (C) ThreadLocalNode()); kvn@4115: Node* cast_thread = transform_later(new (C) CastP2XNode(ctrl, thread)); kvn@4115: Node* o_node = transform_later(new (C) OrXNode(cast_thread, proto_node)); kvn@4115: Node* x_node = transform_later(new (C) XorXNode(o_node, mark_node)); kvn@855: kvn@855: // Get slow path - mark word does NOT match the value. kvn@855: Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node, kvn@855: (~markOopDesc::age_mask_in_place), 0); kvn@855: // region->in(3) is set to fast path - the object is biased to the current thread. kvn@855: mem_phi->init_req(3, mem); kvn@855: kvn@855: kvn@855: // Mark word does NOT match the value (thread | Klass::_prototype_header). kvn@855: kvn@855: kvn@855: // First, check biased pattern. kvn@855: // Get fast path - _prototype_header has the same biased lock pattern. kvn@855: ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node, kvn@855: markOopDesc::biased_lock_mask_in_place, 0, true); kvn@855: kvn@855: not_biased_ctrl = fast_lock_region->in(2); // Slow path kvn@855: // fast_lock_region->in(2) - the prototype header is no longer biased kvn@855: // and we have to revoke the bias on this object. kvn@855: // We are going to try to reset the mark of this object to the prototype kvn@855: // value and fall through to the CAS-based locking scheme. kvn@855: Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); kvn@4115: Node* cas = new (C) StoreXConditionalNode(not_biased_ctrl, mem, adr, kvn@4115: proto_node, mark_node); kvn@855: transform_later(cas); kvn@4115: Node* proj = transform_later( new (C) SCMemProjNode(cas)); kvn@855: fast_lock_mem_phi->init_req(2, proj); kvn@855: kvn@855: kvn@855: // Second, check epoch bits. kvn@4115: Node* rebiased_region = new (C) RegionNode(3); kvn@4115: Node* old_phi = new (C) PhiNode( rebiased_region, TypeX_X); kvn@4115: Node* new_phi = new (C) PhiNode( rebiased_region, TypeX_X); kvn@855: kvn@855: // Get slow path - mark word does NOT match epoch bits. kvn@855: Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node, kvn@855: markOopDesc::epoch_mask_in_place, 0); kvn@855: // The epoch of the current bias is not valid, attempt to rebias the object kvn@855: // toward the current thread. kvn@855: rebiased_region->init_req(2, epoch_ctrl); kvn@855: old_phi->init_req(2, mark_node); kvn@855: new_phi->init_req(2, o_node); kvn@855: kvn@855: // rebiased_region->in(1) is set to fast path. kvn@855: // The epoch of the current bias is still valid but we know kvn@855: // nothing about the owner; it might be set or it might be clear. kvn@855: Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place | kvn@855: markOopDesc::age_mask_in_place | kvn@855: markOopDesc::epoch_mask_in_place); kvn@4115: Node* old = transform_later(new (C) AndXNode(mark_node, cmask)); kvn@4115: cast_thread = transform_later(new (C) CastP2XNode(ctrl, thread)); kvn@4115: Node* new_mark = transform_later(new (C) OrXNode(cast_thread, old)); kvn@855: old_phi->init_req(1, old); kvn@855: new_phi->init_req(1, new_mark); kvn@855: kvn@855: transform_later(rebiased_region); kvn@855: transform_later(old_phi); kvn@855: transform_later(new_phi); kvn@855: kvn@855: // Try to acquire the bias of the object using an atomic operation. kvn@855: // If this fails we will go in to the runtime to revoke the object's bias. kvn@4115: cas = new (C) StoreXConditionalNode(rebiased_region, mem, adr, kvn@855: new_phi, old_phi); kvn@855: transform_later(cas); kvn@4115: proj = transform_later( new (C) SCMemProjNode(cas)); kvn@855: kvn@855: // Get slow path - Failed to CAS. kvn@855: not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0); kvn@855: mem_phi->init_req(4, proj); kvn@855: // region->in(4) is set to fast path - the object is rebiased to the current thread. kvn@855: kvn@855: // Failed to CAS. kvn@4115: slow_path = new (C) RegionNode(3); kvn@4115: Node *slow_mem = new (C) PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM); kvn@855: kvn@855: slow_path->init_req(1, not_biased_ctrl); // Capture slow-control kvn@855: slow_mem->init_req(1, proj); kvn@855: kvn@855: // Call CAS-based locking scheme (FastLock node). kvn@855: kvn@855: transform_later(fast_lock_region); kvn@855: transform_later(fast_lock_mem_phi); kvn@855: kvn@855: // Get slow path - FastLock failed to lock the object. kvn@855: ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0); kvn@855: mem_phi->init_req(2, fast_lock_mem_phi); kvn@855: // region->in(2) is set to fast path - the object is locked to the current thread. kvn@855: kvn@855: slow_path->init_req(2, ctrl); // Capture slow-control kvn@855: slow_mem->init_req(2, fast_lock_mem_phi); kvn@855: kvn@855: transform_later(slow_path); kvn@855: transform_later(slow_mem); kvn@855: // Reset lock's memory edge. kvn@855: lock->set_req(TypeFunc::Memory, slow_mem); kvn@855: kvn@855: } else { kvn@4115: region = new (C) RegionNode(3); kvn@855: // create a Phi for the memory state kvn@4115: mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); kvn@855: kvn@855: // Optimize test; set region slot 2 kvn@855: slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0); kvn@855: mem_phi->init_req(2, mem); kvn@855: } duke@435: duke@435: // Make slow path call duke@435: CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box ); duke@435: duke@435: extract_call_projections(call); duke@435: duke@435: // Slow path can only throw asynchronous exceptions, which are always duke@435: // de-opted. So the compiler thinks the slow-call can never throw an duke@435: // exception. If it DOES throw an exception we would need the debug duke@435: // info removed first (since if it throws there is no monitor). duke@435: assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && duke@435: _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); duke@435: duke@435: // Capture slow path duke@435: // disconnect fall-through projection from call and create a new one duke@435: // hook up users of fall-through projection to region duke@435: Node *slow_ctrl = _fallthroughproj->clone(); duke@435: transform_later(slow_ctrl); duke@435: _igvn.hash_delete(_fallthroughproj); bharadwaj@4315: _fallthroughproj->disconnect_inputs(NULL, C); duke@435: region->init_req(1, slow_ctrl); duke@435: // region inputs are now complete duke@435: transform_later(region); kvn@1143: _igvn.replace_node(_fallthroughproj, region); duke@435: kvn@4115: Node *memproj = transform_later( new(C) ProjNode(call, TypeFunc::Memory) ); duke@435: mem_phi->init_req(1, memproj ); duke@435: transform_later(mem_phi); kvn@1143: _igvn.replace_node(_memproj_fallthrough, mem_phi); duke@435: } duke@435: duke@435: //------------------------------expand_unlock_node---------------------- duke@435: void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { duke@435: kvn@501: Node* ctrl = unlock->in(TypeFunc::Control); duke@435: Node* mem = unlock->in(TypeFunc::Memory); duke@435: Node* obj = unlock->obj_node(); duke@435: Node* box = unlock->box_node(); duke@435: kvn@3419: assert(!box->as_BoxLock()->is_eliminated(), "sanity"); kvn@3406: duke@435: // No need for a null check on unlock duke@435: duke@435: // Make the merge point kvn@855: Node *region; kvn@855: Node *mem_phi; kvn@855: kvn@855: if (UseOptoBiasInlining) { kvn@855: // Check for biased locking unlock case, which is a no-op. twisti@1040: // See the full description in MacroAssembler::biased_locking_exit(). kvn@4115: region = new (C) RegionNode(4); kvn@855: // create a Phi for the memory state kvn@4115: mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); kvn@855: mem_phi->init_req(3, mem); kvn@855: kvn@855: Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); kvn@855: ctrl = opt_bits_test(ctrl, region, 3, mark_node, kvn@855: markOopDesc::biased_lock_mask_in_place, kvn@855: markOopDesc::biased_lock_pattern); kvn@855: } else { kvn@4115: region = new (C) RegionNode(3); kvn@855: // create a Phi for the memory state kvn@4115: mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); kvn@855: } duke@435: kvn@4115: FastUnlockNode *funlock = new (C) FastUnlockNode( ctrl, obj, box ); duke@435: funlock = transform_later( funlock )->as_FastUnlock(); duke@435: // Optimize test; set region slot 2 kvn@855: Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0); duke@435: duke@435: CallNode *call = make_slow_call( (CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), "complete_monitor_unlocking_C", slow_path, obj, box ); duke@435: duke@435: extract_call_projections(call); duke@435: duke@435: assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && duke@435: _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); duke@435: duke@435: // No exceptions for unlocking duke@435: // Capture slow path duke@435: // disconnect fall-through projection from call and create a new one duke@435: // hook up users of fall-through projection to region duke@435: Node *slow_ctrl = _fallthroughproj->clone(); duke@435: transform_later(slow_ctrl); duke@435: _igvn.hash_delete(_fallthroughproj); bharadwaj@4315: _fallthroughproj->disconnect_inputs(NULL, C); duke@435: region->init_req(1, slow_ctrl); duke@435: // region inputs are now complete duke@435: transform_later(region); kvn@1143: _igvn.replace_node(_fallthroughproj, region); duke@435: kvn@4115: Node *memproj = transform_later( new(C) ProjNode(call, TypeFunc::Memory) ); duke@435: mem_phi->init_req(1, memproj ); duke@435: mem_phi->init_req(2, mem); duke@435: transform_later(mem_phi); kvn@1143: _igvn.replace_node(_memproj_fallthrough, mem_phi); duke@435: } duke@435: kvn@3311: //---------------------------eliminate_macro_nodes---------------------- kvn@3311: // Eliminate scalar replaced allocations and associated locks. kvn@3311: void PhaseMacroExpand::eliminate_macro_nodes() { duke@435: if (C->macro_count() == 0) kvn@3311: return; kvn@3311: kvn@895: // First, attempt to eliminate locks kvn@2951: int cnt = C->macro_count(); kvn@2951: for (int i=0; i < cnt; i++) { kvn@2951: Node *n = C->macro_node(i); kvn@2951: if (n->is_AbstractLock()) { // Lock and Unlock nodes kvn@2951: // Before elimination mark all associated (same box and obj) kvn@2951: // lock and unlock nodes. kvn@2951: mark_eliminated_locking_nodes(n->as_AbstractLock()); kvn@2951: } kvn@2951: } kvn@508: bool progress = true; kvn@508: while (progress) { kvn@508: progress = false; kvn@508: for (int i = C->macro_count(); i > 0; i--) { kvn@508: Node * n = C->macro_node(i-1); kvn@508: bool success = false; kvn@508: debug_only(int old_macro_count = C->macro_count();); kvn@895: if (n->is_AbstractLock()) { kvn@895: success = eliminate_locking_node(n->as_AbstractLock()); kvn@895: } kvn@895: assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); kvn@895: progress = progress || success; kvn@895: } kvn@895: } kvn@895: // Next, attempt to eliminate allocations kvn@895: progress = true; kvn@895: while (progress) { kvn@895: progress = false; kvn@895: for (int i = C->macro_count(); i > 0; i--) { kvn@895: Node * n = C->macro_node(i-1); kvn@895: bool success = false; kvn@895: debug_only(int old_macro_count = C->macro_count();); kvn@508: switch (n->class_id()) { kvn@508: case Node::Class_Allocate: kvn@508: case Node::Class_AllocateArray: kvn@508: success = eliminate_allocate_node(n->as_Allocate()); kvn@508: break; kvn@5110: case Node::Class_CallStaticJava: kvn@5110: success = eliminate_boxing_node(n->as_CallStaticJava()); kvn@5110: break; kvn@508: case Node::Class_Lock: kvn@508: case Node::Class_Unlock: kvn@895: assert(!n->as_AbstractLock()->is_eliminated(), "sanity"); kvn@508: break; kvn@508: default: kvn@3311: assert(n->Opcode() == Op_LoopLimit || kvn@3311: n->Opcode() == Op_Opaque1 || kvn@3311: n->Opcode() == Op_Opaque2, "unknown node type in macro list"); kvn@508: } kvn@508: assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); kvn@508: progress = progress || success; kvn@508: } kvn@508: } kvn@3311: } kvn@3311: kvn@3311: //------------------------------expand_macro_nodes---------------------- kvn@3311: // Returns true if a failure occurred. kvn@3311: bool PhaseMacroExpand::expand_macro_nodes() { kvn@3311: // Last attempt to eliminate macro nodes. kvn@3311: eliminate_macro_nodes(); kvn@3311: kvn@508: // Make sure expansion will not cause node limit to be exceeded. kvn@508: // Worst case is a macro node gets expanded into about 50 nodes. kvn@508: // Allow 50% more for optimization. duke@435: if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) ) duke@435: return true; kvn@508: kvn@3311: // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations. kvn@3311: bool progress = true; kvn@3311: while (progress) { kvn@3311: progress = false; kvn@3311: for (int i = C->macro_count(); i > 0; i--) { kvn@3311: Node * n = C->macro_node(i-1); kvn@3311: bool success = false; kvn@3311: debug_only(int old_macro_count = C->macro_count();); kvn@3311: if (n->Opcode() == Op_LoopLimit) { kvn@3311: // Remove it from macro list and put on IGVN worklist to optimize. kvn@3311: C->remove_macro_node(n); kvn@3311: _igvn._worklist.push(n); kvn@3311: success = true; kvn@5110: } else if (n->Opcode() == Op_CallStaticJava) { kvn@5110: // Remove it from macro list and put on IGVN worklist to optimize. kvn@5110: C->remove_macro_node(n); kvn@5110: _igvn._worklist.push(n); kvn@5110: success = true; kvn@3311: } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) { kvn@3311: _igvn.replace_node(n, n->in(1)); kvn@3311: success = true; kvn@3311: } kvn@3311: assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); kvn@3311: progress = progress || success; kvn@3311: } kvn@3311: } kvn@3311: duke@435: // expand "macro" nodes duke@435: // nodes are removed from the macro list as they are processed duke@435: while (C->macro_count() > 0) { kvn@508: int macro_count = C->macro_count(); kvn@508: Node * n = C->macro_node(macro_count-1); duke@435: assert(n->is_macro(), "only macro nodes expected here"); duke@435: if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { duke@435: // node is unreachable, so don't try to expand it duke@435: C->remove_macro_node(n); duke@435: continue; duke@435: } duke@435: switch (n->class_id()) { duke@435: case Node::Class_Allocate: duke@435: expand_allocate(n->as_Allocate()); duke@435: break; duke@435: case Node::Class_AllocateArray: duke@435: expand_allocate_array(n->as_AllocateArray()); duke@435: break; duke@435: case Node::Class_Lock: duke@435: expand_lock_node(n->as_Lock()); duke@435: break; duke@435: case Node::Class_Unlock: duke@435: expand_unlock_node(n->as_Unlock()); duke@435: break; duke@435: default: duke@435: assert(false, "unknown node type in macro list"); duke@435: } kvn@508: assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); duke@435: if (C->failing()) return true; duke@435: } coleenp@548: coleenp@548: _igvn.set_delay_transform(false); duke@435: _igvn.optimize(); kvn@3311: if (C->failing()) return true; duke@435: return false; duke@435: }