1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/src/share/vm/opto/callnode.cpp Wed Apr 27 01:25:04 2016 +0800
1.3 @@ -0,0 +1,1808 @@
1.4 +/*
1.5 + * Copyright (c) 1997, 2014, Oracle and/or its affiliates. All rights reserved.
1.6 + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
1.7 + *
1.8 + * This code is free software; you can redistribute it and/or modify it
1.9 + * under the terms of the GNU General Public License version 2 only, as
1.10 + * published by the Free Software Foundation.
1.11 + *
1.12 + * This code is distributed in the hope that it will be useful, but WITHOUT
1.13 + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
1.14 + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
1.15 + * version 2 for more details (a copy is included in the LICENSE file that
1.16 + * accompanied this code).
1.17 + *
1.18 + * You should have received a copy of the GNU General Public License version
1.19 + * 2 along with this work; if not, write to the Free Software Foundation,
1.20 + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
1.21 + *
1.22 + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
1.23 + * or visit www.oracle.com if you need additional information or have any
1.24 + * questions.
1.25 + *
1.26 + */
1.27 +
1.28 +#include "precompiled.hpp"
1.29 +#include "ci/bcEscapeAnalyzer.hpp"
1.30 +#include "compiler/oopMap.hpp"
1.31 +#include "opto/callGenerator.hpp"
1.32 +#include "opto/callnode.hpp"
1.33 +#include "opto/escape.hpp"
1.34 +#include "opto/locknode.hpp"
1.35 +#include "opto/machnode.hpp"
1.36 +#include "opto/matcher.hpp"
1.37 +#include "opto/parse.hpp"
1.38 +#include "opto/regalloc.hpp"
1.39 +#include "opto/regmask.hpp"
1.40 +#include "opto/rootnode.hpp"
1.41 +#include "opto/runtime.hpp"
1.42 +
1.43 +// Portions of code courtesy of Clifford Click
1.44 +
1.45 +// Optimization - Graph Style
1.46 +
1.47 +//=============================================================================
1.48 +uint StartNode::size_of() const { return sizeof(*this); }
1.49 +uint StartNode::cmp( const Node &n ) const
1.50 +{ return _domain == ((StartNode&)n)._domain; }
1.51 +const Type *StartNode::bottom_type() const { return _domain; }
1.52 +const Type *StartNode::Value(PhaseTransform *phase) const { return _domain; }
1.53 +#ifndef PRODUCT
1.54 +void StartNode::dump_spec(outputStream *st) const { st->print(" #"); _domain->dump_on(st);}
1.55 +#endif
1.56 +
1.57 +//------------------------------Ideal------------------------------------------
1.58 +Node *StartNode::Ideal(PhaseGVN *phase, bool can_reshape){
1.59 + return remove_dead_region(phase, can_reshape) ? this : NULL;
1.60 +}
1.61 +
1.62 +//------------------------------calling_convention-----------------------------
1.63 +void StartNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const {
1.64 + Matcher::calling_convention( sig_bt, parm_regs, argcnt, false );
1.65 +}
1.66 +
1.67 +//------------------------------Registers--------------------------------------
1.68 +const RegMask &StartNode::in_RegMask(uint) const {
1.69 + return RegMask::Empty;
1.70 +}
1.71 +
1.72 +//------------------------------match------------------------------------------
1.73 +// Construct projections for incoming parameters, and their RegMask info
1.74 +Node *StartNode::match( const ProjNode *proj, const Matcher *match ) {
1.75 + switch (proj->_con) {
1.76 + case TypeFunc::Control:
1.77 + case TypeFunc::I_O:
1.78 + case TypeFunc::Memory:
1.79 + return new (match->C) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
1.80 + case TypeFunc::FramePtr:
1.81 + return new (match->C) MachProjNode(this,proj->_con,Matcher::c_frame_ptr_mask, Op_RegP);
1.82 + case TypeFunc::ReturnAdr:
1.83 + return new (match->C) MachProjNode(this,proj->_con,match->_return_addr_mask,Op_RegP);
1.84 + case TypeFunc::Parms:
1.85 + default: {
1.86 + uint parm_num = proj->_con - TypeFunc::Parms;
1.87 + const Type *t = _domain->field_at(proj->_con);
1.88 + if (t->base() == Type::Half) // 2nd half of Longs and Doubles
1.89 + return new (match->C) ConNode(Type::TOP);
1.90 + uint ideal_reg = t->ideal_reg();
1.91 + RegMask &rm = match->_calling_convention_mask[parm_num];
1.92 + return new (match->C) MachProjNode(this,proj->_con,rm,ideal_reg);
1.93 + }
1.94 + }
1.95 + return NULL;
1.96 +}
1.97 +
1.98 +//------------------------------StartOSRNode----------------------------------
1.99 +// The method start node for an on stack replacement adapter
1.100 +
1.101 +//------------------------------osr_domain-----------------------------
1.102 +const TypeTuple *StartOSRNode::osr_domain() {
1.103 + const Type **fields = TypeTuple::fields(2);
1.104 + fields[TypeFunc::Parms+0] = TypeRawPtr::BOTTOM; // address of osr buffer
1.105 +
1.106 + return TypeTuple::make(TypeFunc::Parms+1, fields);
1.107 +}
1.108 +
1.109 +//=============================================================================
1.110 +const char * const ParmNode::names[TypeFunc::Parms+1] = {
1.111 + "Control", "I_O", "Memory", "FramePtr", "ReturnAdr", "Parms"
1.112 +};
1.113 +
1.114 +#ifndef PRODUCT
1.115 +void ParmNode::dump_spec(outputStream *st) const {
1.116 + if( _con < TypeFunc::Parms ) {
1.117 + st->print("%s", names[_con]);
1.118 + } else {
1.119 + st->print("Parm%d: ",_con-TypeFunc::Parms);
1.120 + // Verbose and WizardMode dump bottom_type for all nodes
1.121 + if( !Verbose && !WizardMode ) bottom_type()->dump_on(st);
1.122 + }
1.123 +}
1.124 +#endif
1.125 +
1.126 +uint ParmNode::ideal_reg() const {
1.127 + switch( _con ) {
1.128 + case TypeFunc::Control : // fall through
1.129 + case TypeFunc::I_O : // fall through
1.130 + case TypeFunc::Memory : return 0;
1.131 + case TypeFunc::FramePtr : // fall through
1.132 + case TypeFunc::ReturnAdr: return Op_RegP;
1.133 + default : assert( _con > TypeFunc::Parms, "" );
1.134 + // fall through
1.135 + case TypeFunc::Parms : {
1.136 + // Type of argument being passed
1.137 + const Type *t = in(0)->as_Start()->_domain->field_at(_con);
1.138 + return t->ideal_reg();
1.139 + }
1.140 + }
1.141 + ShouldNotReachHere();
1.142 + return 0;
1.143 +}
1.144 +
1.145 +//=============================================================================
1.146 +ReturnNode::ReturnNode(uint edges, Node *cntrl, Node *i_o, Node *memory, Node *frameptr, Node *retadr ) : Node(edges) {
1.147 + init_req(TypeFunc::Control,cntrl);
1.148 + init_req(TypeFunc::I_O,i_o);
1.149 + init_req(TypeFunc::Memory,memory);
1.150 + init_req(TypeFunc::FramePtr,frameptr);
1.151 + init_req(TypeFunc::ReturnAdr,retadr);
1.152 +}
1.153 +
1.154 +Node *ReturnNode::Ideal(PhaseGVN *phase, bool can_reshape){
1.155 + return remove_dead_region(phase, can_reshape) ? this : NULL;
1.156 +}
1.157 +
1.158 +const Type *ReturnNode::Value( PhaseTransform *phase ) const {
1.159 + return ( phase->type(in(TypeFunc::Control)) == Type::TOP)
1.160 + ? Type::TOP
1.161 + : Type::BOTTOM;
1.162 +}
1.163 +
1.164 +// Do we Match on this edge index or not? No edges on return nodes
1.165 +uint ReturnNode::match_edge(uint idx) const {
1.166 + return 0;
1.167 +}
1.168 +
1.169 +
1.170 +#ifndef PRODUCT
1.171 +void ReturnNode::dump_req(outputStream *st) const {
1.172 + // Dump the required inputs, enclosed in '(' and ')'
1.173 + uint i; // Exit value of loop
1.174 + for (i = 0; i < req(); i++) { // For all required inputs
1.175 + if (i == TypeFunc::Parms) st->print("returns");
1.176 + if (in(i)) st->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx);
1.177 + else st->print("_ ");
1.178 + }
1.179 +}
1.180 +#endif
1.181 +
1.182 +//=============================================================================
1.183 +RethrowNode::RethrowNode(
1.184 + Node* cntrl,
1.185 + Node* i_o,
1.186 + Node* memory,
1.187 + Node* frameptr,
1.188 + Node* ret_adr,
1.189 + Node* exception
1.190 +) : Node(TypeFunc::Parms + 1) {
1.191 + init_req(TypeFunc::Control , cntrl );
1.192 + init_req(TypeFunc::I_O , i_o );
1.193 + init_req(TypeFunc::Memory , memory );
1.194 + init_req(TypeFunc::FramePtr , frameptr );
1.195 + init_req(TypeFunc::ReturnAdr, ret_adr);
1.196 + init_req(TypeFunc::Parms , exception);
1.197 +}
1.198 +
1.199 +Node *RethrowNode::Ideal(PhaseGVN *phase, bool can_reshape){
1.200 + return remove_dead_region(phase, can_reshape) ? this : NULL;
1.201 +}
1.202 +
1.203 +const Type *RethrowNode::Value( PhaseTransform *phase ) const {
1.204 + return (phase->type(in(TypeFunc::Control)) == Type::TOP)
1.205 + ? Type::TOP
1.206 + : Type::BOTTOM;
1.207 +}
1.208 +
1.209 +uint RethrowNode::match_edge(uint idx) const {
1.210 + return 0;
1.211 +}
1.212 +
1.213 +#ifndef PRODUCT
1.214 +void RethrowNode::dump_req(outputStream *st) const {
1.215 + // Dump the required inputs, enclosed in '(' and ')'
1.216 + uint i; // Exit value of loop
1.217 + for (i = 0; i < req(); i++) { // For all required inputs
1.218 + if (i == TypeFunc::Parms) st->print("exception");
1.219 + if (in(i)) st->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx);
1.220 + else st->print("_ ");
1.221 + }
1.222 +}
1.223 +#endif
1.224 +
1.225 +//=============================================================================
1.226 +// Do we Match on this edge index or not? Match only target address & method
1.227 +uint TailCallNode::match_edge(uint idx) const {
1.228 + return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1;
1.229 +}
1.230 +
1.231 +//=============================================================================
1.232 +// Do we Match on this edge index or not? Match only target address & oop
1.233 +uint TailJumpNode::match_edge(uint idx) const {
1.234 + return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1;
1.235 +}
1.236 +
1.237 +//=============================================================================
1.238 +JVMState::JVMState(ciMethod* method, JVMState* caller) :
1.239 + _method(method) {
1.240 + assert(method != NULL, "must be valid call site");
1.241 + _reexecute = Reexecute_Undefined;
1.242 + debug_only(_bci = -99); // random garbage value
1.243 + debug_only(_map = (SafePointNode*)-1);
1.244 + _caller = caller;
1.245 + _depth = 1 + (caller == NULL ? 0 : caller->depth());
1.246 + _locoff = TypeFunc::Parms;
1.247 + _stkoff = _locoff + _method->max_locals();
1.248 + _monoff = _stkoff + _method->max_stack();
1.249 + _scloff = _monoff;
1.250 + _endoff = _monoff;
1.251 + _sp = 0;
1.252 +}
1.253 +JVMState::JVMState(int stack_size) :
1.254 + _method(NULL) {
1.255 + _bci = InvocationEntryBci;
1.256 + _reexecute = Reexecute_Undefined;
1.257 + debug_only(_map = (SafePointNode*)-1);
1.258 + _caller = NULL;
1.259 + _depth = 1;
1.260 + _locoff = TypeFunc::Parms;
1.261 + _stkoff = _locoff;
1.262 + _monoff = _stkoff + stack_size;
1.263 + _scloff = _monoff;
1.264 + _endoff = _monoff;
1.265 + _sp = 0;
1.266 +}
1.267 +
1.268 +//--------------------------------of_depth-------------------------------------
1.269 +JVMState* JVMState::of_depth(int d) const {
1.270 + const JVMState* jvmp = this;
1.271 + assert(0 < d && (uint)d <= depth(), "oob");
1.272 + for (int skip = depth() - d; skip > 0; skip--) {
1.273 + jvmp = jvmp->caller();
1.274 + }
1.275 + assert(jvmp->depth() == (uint)d, "found the right one");
1.276 + return (JVMState*)jvmp;
1.277 +}
1.278 +
1.279 +//-----------------------------same_calls_as-----------------------------------
1.280 +bool JVMState::same_calls_as(const JVMState* that) const {
1.281 + if (this == that) return true;
1.282 + if (this->depth() != that->depth()) return false;
1.283 + const JVMState* p = this;
1.284 + const JVMState* q = that;
1.285 + for (;;) {
1.286 + if (p->_method != q->_method) return false;
1.287 + if (p->_method == NULL) return true; // bci is irrelevant
1.288 + if (p->_bci != q->_bci) return false;
1.289 + if (p->_reexecute != q->_reexecute) return false;
1.290 + p = p->caller();
1.291 + q = q->caller();
1.292 + if (p == q) return true;
1.293 + assert(p != NULL && q != NULL, "depth check ensures we don't run off end");
1.294 + }
1.295 +}
1.296 +
1.297 +//------------------------------debug_start------------------------------------
1.298 +uint JVMState::debug_start() const {
1.299 + debug_only(JVMState* jvmroot = of_depth(1));
1.300 + assert(jvmroot->locoff() <= this->locoff(), "youngest JVMState must be last");
1.301 + return of_depth(1)->locoff();
1.302 +}
1.303 +
1.304 +//-------------------------------debug_end-------------------------------------
1.305 +uint JVMState::debug_end() const {
1.306 + debug_only(JVMState* jvmroot = of_depth(1));
1.307 + assert(jvmroot->endoff() <= this->endoff(), "youngest JVMState must be last");
1.308 + return endoff();
1.309 +}
1.310 +
1.311 +//------------------------------debug_depth------------------------------------
1.312 +uint JVMState::debug_depth() const {
1.313 + uint total = 0;
1.314 + for (const JVMState* jvmp = this; jvmp != NULL; jvmp = jvmp->caller()) {
1.315 + total += jvmp->debug_size();
1.316 + }
1.317 + return total;
1.318 +}
1.319 +
1.320 +#ifndef PRODUCT
1.321 +
1.322 +//------------------------------format_helper----------------------------------
1.323 +// Given an allocation (a Chaitin object) and a Node decide if the Node carries
1.324 +// any defined value or not. If it does, print out the register or constant.
1.325 +static void format_helper( PhaseRegAlloc *regalloc, outputStream* st, Node *n, const char *msg, uint i, GrowableArray<SafePointScalarObjectNode*> *scobjs ) {
1.326 + if (n == NULL) { st->print(" NULL"); return; }
1.327 + if (n->is_SafePointScalarObject()) {
1.328 + // Scalar replacement.
1.329 + SafePointScalarObjectNode* spobj = n->as_SafePointScalarObject();
1.330 + scobjs->append_if_missing(spobj);
1.331 + int sco_n = scobjs->find(spobj);
1.332 + assert(sco_n >= 0, "");
1.333 + st->print(" %s%d]=#ScObj" INT32_FORMAT, msg, i, sco_n);
1.334 + return;
1.335 + }
1.336 + if (regalloc->node_regs_max_index() > 0 &&
1.337 + OptoReg::is_valid(regalloc->get_reg_first(n))) { // Check for undefined
1.338 + char buf[50];
1.339 + regalloc->dump_register(n,buf);
1.340 + st->print(" %s%d]=%s",msg,i,buf);
1.341 + } else { // No register, but might be constant
1.342 + const Type *t = n->bottom_type();
1.343 + switch (t->base()) {
1.344 + case Type::Int:
1.345 + st->print(" %s%d]=#"INT32_FORMAT,msg,i,t->is_int()->get_con());
1.346 + break;
1.347 + case Type::AnyPtr:
1.348 + assert( t == TypePtr::NULL_PTR || n->in_dump(), "" );
1.349 + st->print(" %s%d]=#NULL",msg,i);
1.350 + break;
1.351 + case Type::AryPtr:
1.352 + case Type::InstPtr:
1.353 + st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->isa_oopptr()->const_oop()));
1.354 + break;
1.355 + case Type::KlassPtr:
1.356 + st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_klassptr()->klass()));
1.357 + break;
1.358 + case Type::MetadataPtr:
1.359 + st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_metadataptr()->metadata()));
1.360 + break;
1.361 + case Type::NarrowOop:
1.362 + st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_oopptr()->const_oop()));
1.363 + break;
1.364 + case Type::RawPtr:
1.365 + st->print(" %s%d]=#Raw" INTPTR_FORMAT,msg,i,p2i(t->is_rawptr()));
1.366 + break;
1.367 + case Type::DoubleCon:
1.368 + st->print(" %s%d]=#%fD",msg,i,t->is_double_constant()->_d);
1.369 + break;
1.370 + case Type::FloatCon:
1.371 + st->print(" %s%d]=#%fF",msg,i,t->is_float_constant()->_f);
1.372 + break;
1.373 + case Type::Long:
1.374 + st->print(" %s%d]=#"INT64_FORMAT,msg,i,(int64_t)(t->is_long()->get_con()));
1.375 + break;
1.376 + case Type::Half:
1.377 + case Type::Top:
1.378 + st->print(" %s%d]=_",msg,i);
1.379 + break;
1.380 + default: ShouldNotReachHere();
1.381 + }
1.382 + }
1.383 +}
1.384 +
1.385 +//------------------------------format-----------------------------------------
1.386 +void JVMState::format(PhaseRegAlloc *regalloc, const Node *n, outputStream* st) const {
1.387 + st->print(" #");
1.388 + if (_method) {
1.389 + _method->print_short_name(st);
1.390 + st->print(" @ bci:%d ",_bci);
1.391 + } else {
1.392 + st->print_cr(" runtime stub ");
1.393 + return;
1.394 + }
1.395 + if (n->is_MachSafePoint()) {
1.396 + GrowableArray<SafePointScalarObjectNode*> scobjs;
1.397 + MachSafePointNode *mcall = n->as_MachSafePoint();
1.398 + uint i;
1.399 + // Print locals
1.400 + for (i = 0; i < (uint)loc_size(); i++)
1.401 + format_helper(regalloc, st, mcall->local(this, i), "L[", i, &scobjs);
1.402 + // Print stack
1.403 + for (i = 0; i < (uint)stk_size(); i++) {
1.404 + if ((uint)(_stkoff + i) >= mcall->len())
1.405 + st->print(" oob ");
1.406 + else
1.407 + format_helper(regalloc, st, mcall->stack(this, i), "STK[", i, &scobjs);
1.408 + }
1.409 + for (i = 0; (int)i < nof_monitors(); i++) {
1.410 + Node *box = mcall->monitor_box(this, i);
1.411 + Node *obj = mcall->monitor_obj(this, i);
1.412 + if (regalloc->node_regs_max_index() > 0 &&
1.413 + OptoReg::is_valid(regalloc->get_reg_first(box))) {
1.414 + box = BoxLockNode::box_node(box);
1.415 + format_helper(regalloc, st, box, "MON-BOX[", i, &scobjs);
1.416 + } else {
1.417 + OptoReg::Name box_reg = BoxLockNode::reg(box);
1.418 + st->print(" MON-BOX%d=%s+%d",
1.419 + i,
1.420 + OptoReg::regname(OptoReg::c_frame_pointer),
1.421 + regalloc->reg2offset(box_reg));
1.422 + }
1.423 + const char* obj_msg = "MON-OBJ[";
1.424 + if (EliminateLocks) {
1.425 + if (BoxLockNode::box_node(box)->is_eliminated())
1.426 + obj_msg = "MON-OBJ(LOCK ELIMINATED)[";
1.427 + }
1.428 + format_helper(regalloc, st, obj, obj_msg, i, &scobjs);
1.429 + }
1.430 +
1.431 + for (i = 0; i < (uint)scobjs.length(); i++) {
1.432 + // Scalar replaced objects.
1.433 + st->cr();
1.434 + st->print(" # ScObj" INT32_FORMAT " ", i);
1.435 + SafePointScalarObjectNode* spobj = scobjs.at(i);
1.436 + ciKlass* cik = spobj->bottom_type()->is_oopptr()->klass();
1.437 + assert(cik->is_instance_klass() ||
1.438 + cik->is_array_klass(), "Not supported allocation.");
1.439 + ciInstanceKlass *iklass = NULL;
1.440 + if (cik->is_instance_klass()) {
1.441 + cik->print_name_on(st);
1.442 + iklass = cik->as_instance_klass();
1.443 + } else if (cik->is_type_array_klass()) {
1.444 + cik->as_array_klass()->base_element_type()->print_name_on(st);
1.445 + st->print("[%d]", spobj->n_fields());
1.446 + } else if (cik->is_obj_array_klass()) {
1.447 + ciKlass* cie = cik->as_obj_array_klass()->base_element_klass();
1.448 + if (cie->is_instance_klass()) {
1.449 + cie->print_name_on(st);
1.450 + } else if (cie->is_type_array_klass()) {
1.451 + cie->as_array_klass()->base_element_type()->print_name_on(st);
1.452 + } else {
1.453 + ShouldNotReachHere();
1.454 + }
1.455 + st->print("[%d]", spobj->n_fields());
1.456 + int ndim = cik->as_array_klass()->dimension() - 1;
1.457 + while (ndim-- > 0) {
1.458 + st->print("[]");
1.459 + }
1.460 + }
1.461 + st->print("={");
1.462 + uint nf = spobj->n_fields();
1.463 + if (nf > 0) {
1.464 + uint first_ind = spobj->first_index(mcall->jvms());
1.465 + Node* fld_node = mcall->in(first_ind);
1.466 + ciField* cifield;
1.467 + if (iklass != NULL) {
1.468 + st->print(" [");
1.469 + cifield = iklass->nonstatic_field_at(0);
1.470 + cifield->print_name_on(st);
1.471 + format_helper(regalloc, st, fld_node, ":", 0, &scobjs);
1.472 + } else {
1.473 + format_helper(regalloc, st, fld_node, "[", 0, &scobjs);
1.474 + }
1.475 + for (uint j = 1; j < nf; j++) {
1.476 + fld_node = mcall->in(first_ind+j);
1.477 + if (iklass != NULL) {
1.478 + st->print(", [");
1.479 + cifield = iklass->nonstatic_field_at(j);
1.480 + cifield->print_name_on(st);
1.481 + format_helper(regalloc, st, fld_node, ":", j, &scobjs);
1.482 + } else {
1.483 + format_helper(regalloc, st, fld_node, ", [", j, &scobjs);
1.484 + }
1.485 + }
1.486 + }
1.487 + st->print(" }");
1.488 + }
1.489 + }
1.490 + st->cr();
1.491 + if (caller() != NULL) caller()->format(regalloc, n, st);
1.492 +}
1.493 +
1.494 +
1.495 +void JVMState::dump_spec(outputStream *st) const {
1.496 + if (_method != NULL) {
1.497 + bool printed = false;
1.498 + if (!Verbose) {
1.499 + // The JVMS dumps make really, really long lines.
1.500 + // Take out the most boring parts, which are the package prefixes.
1.501 + char buf[500];
1.502 + stringStream namest(buf, sizeof(buf));
1.503 + _method->print_short_name(&namest);
1.504 + if (namest.count() < sizeof(buf)) {
1.505 + const char* name = namest.base();
1.506 + if (name[0] == ' ') ++name;
1.507 + const char* endcn = strchr(name, ':'); // end of class name
1.508 + if (endcn == NULL) endcn = strchr(name, '(');
1.509 + if (endcn == NULL) endcn = name + strlen(name);
1.510 + while (endcn > name && endcn[-1] != '.' && endcn[-1] != '/')
1.511 + --endcn;
1.512 + st->print(" %s", endcn);
1.513 + printed = true;
1.514 + }
1.515 + }
1.516 + if (!printed)
1.517 + _method->print_short_name(st);
1.518 + st->print(" @ bci:%d",_bci);
1.519 + if(_reexecute == Reexecute_True)
1.520 + st->print(" reexecute");
1.521 + } else {
1.522 + st->print(" runtime stub");
1.523 + }
1.524 + if (caller() != NULL) caller()->dump_spec(st);
1.525 +}
1.526 +
1.527 +
1.528 +void JVMState::dump_on(outputStream* st) const {
1.529 + bool print_map = _map && !((uintptr_t)_map & 1) &&
1.530 + ((caller() == NULL) || (caller()->map() != _map));
1.531 + if (print_map) {
1.532 + if (_map->len() > _map->req()) { // _map->has_exceptions()
1.533 + Node* ex = _map->in(_map->req()); // _map->next_exception()
1.534 + // skip the first one; it's already being printed
1.535 + while (ex != NULL && ex->len() > ex->req()) {
1.536 + ex = ex->in(ex->req()); // ex->next_exception()
1.537 + ex->dump(1);
1.538 + }
1.539 + }
1.540 + _map->dump(Verbose ? 2 : 1);
1.541 + }
1.542 + if (caller() != NULL) {
1.543 + caller()->dump_on(st);
1.544 + }
1.545 + st->print("JVMS depth=%d loc=%d stk=%d arg=%d mon=%d scalar=%d end=%d mondepth=%d sp=%d bci=%d reexecute=%s method=",
1.546 + depth(), locoff(), stkoff(), argoff(), monoff(), scloff(), endoff(), monitor_depth(), sp(), bci(), should_reexecute()?"true":"false");
1.547 + if (_method == NULL) {
1.548 + st->print_cr("(none)");
1.549 + } else {
1.550 + _method->print_name(st);
1.551 + st->cr();
1.552 + if (bci() >= 0 && bci() < _method->code_size()) {
1.553 + st->print(" bc: ");
1.554 + _method->print_codes_on(bci(), bci()+1, st);
1.555 + }
1.556 + }
1.557 +}
1.558 +
1.559 +// Extra way to dump a jvms from the debugger,
1.560 +// to avoid a bug with C++ member function calls.
1.561 +void dump_jvms(JVMState* jvms) {
1.562 + jvms->dump();
1.563 +}
1.564 +#endif
1.565 +
1.566 +//--------------------------clone_shallow--------------------------------------
1.567 +JVMState* JVMState::clone_shallow(Compile* C) const {
1.568 + JVMState* n = has_method() ? new (C) JVMState(_method, _caller) : new (C) JVMState(0);
1.569 + n->set_bci(_bci);
1.570 + n->_reexecute = _reexecute;
1.571 + n->set_locoff(_locoff);
1.572 + n->set_stkoff(_stkoff);
1.573 + n->set_monoff(_monoff);
1.574 + n->set_scloff(_scloff);
1.575 + n->set_endoff(_endoff);
1.576 + n->set_sp(_sp);
1.577 + n->set_map(_map);
1.578 + return n;
1.579 +}
1.580 +
1.581 +//---------------------------clone_deep----------------------------------------
1.582 +JVMState* JVMState::clone_deep(Compile* C) const {
1.583 + JVMState* n = clone_shallow(C);
1.584 + for (JVMState* p = n; p->_caller != NULL; p = p->_caller) {
1.585 + p->_caller = p->_caller->clone_shallow(C);
1.586 + }
1.587 + assert(n->depth() == depth(), "sanity");
1.588 + assert(n->debug_depth() == debug_depth(), "sanity");
1.589 + return n;
1.590 +}
1.591 +
1.592 +/**
1.593 + * Reset map for all callers
1.594 + */
1.595 +void JVMState::set_map_deep(SafePointNode* map) {
1.596 + for (JVMState* p = this; p->_caller != NULL; p = p->_caller) {
1.597 + p->set_map(map);
1.598 + }
1.599 +}
1.600 +
1.601 +// Adapt offsets in in-array after adding or removing an edge.
1.602 +// Prerequisite is that the JVMState is used by only one node.
1.603 +void JVMState::adapt_position(int delta) {
1.604 + for (JVMState* jvms = this; jvms != NULL; jvms = jvms->caller()) {
1.605 + jvms->set_locoff(jvms->locoff() + delta);
1.606 + jvms->set_stkoff(jvms->stkoff() + delta);
1.607 + jvms->set_monoff(jvms->monoff() + delta);
1.608 + jvms->set_scloff(jvms->scloff() + delta);
1.609 + jvms->set_endoff(jvms->endoff() + delta);
1.610 + }
1.611 +}
1.612 +
1.613 +// Mirror the stack size calculation in the deopt code
1.614 +// How much stack space would we need at this point in the program in
1.615 +// case of deoptimization?
1.616 +int JVMState::interpreter_frame_size() const {
1.617 + const JVMState* jvms = this;
1.618 + int size = 0;
1.619 + int callee_parameters = 0;
1.620 + int callee_locals = 0;
1.621 + int extra_args = method()->max_stack() - stk_size();
1.622 +
1.623 + while (jvms != NULL) {
1.624 + int locks = jvms->nof_monitors();
1.625 + int temps = jvms->stk_size();
1.626 + bool is_top_frame = (jvms == this);
1.627 + ciMethod* method = jvms->method();
1.628 +
1.629 + int frame_size = BytesPerWord * Interpreter::size_activation(method->max_stack(),
1.630 + temps + callee_parameters,
1.631 + extra_args,
1.632 + locks,
1.633 + callee_parameters,
1.634 + callee_locals,
1.635 + is_top_frame);
1.636 + size += frame_size;
1.637 +
1.638 + callee_parameters = method->size_of_parameters();
1.639 + callee_locals = method->max_locals();
1.640 + extra_args = 0;
1.641 + jvms = jvms->caller();
1.642 + }
1.643 + return size + Deoptimization::last_frame_adjust(0, callee_locals) * BytesPerWord;
1.644 +}
1.645 +
1.646 +//=============================================================================
1.647 +uint CallNode::cmp( const Node &n ) const
1.648 +{ return _tf == ((CallNode&)n)._tf && _jvms == ((CallNode&)n)._jvms; }
1.649 +#ifndef PRODUCT
1.650 +void CallNode::dump_req(outputStream *st) const {
1.651 + // Dump the required inputs, enclosed in '(' and ')'
1.652 + uint i; // Exit value of loop
1.653 + for (i = 0; i < req(); i++) { // For all required inputs
1.654 + if (i == TypeFunc::Parms) st->print("(");
1.655 + if (in(i)) st->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx);
1.656 + else st->print("_ ");
1.657 + }
1.658 + st->print(")");
1.659 +}
1.660 +
1.661 +void CallNode::dump_spec(outputStream *st) const {
1.662 + st->print(" ");
1.663 + tf()->dump_on(st);
1.664 + if (_cnt != COUNT_UNKNOWN) st->print(" C=%f",_cnt);
1.665 + if (jvms() != NULL) jvms()->dump_spec(st);
1.666 +}
1.667 +#endif
1.668 +
1.669 +const Type *CallNode::bottom_type() const { return tf()->range(); }
1.670 +const Type *CallNode::Value(PhaseTransform *phase) const {
1.671 + if (phase->type(in(0)) == Type::TOP) return Type::TOP;
1.672 + return tf()->range();
1.673 +}
1.674 +
1.675 +//------------------------------calling_convention-----------------------------
1.676 +void CallNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const {
1.677 + // Use the standard compiler calling convention
1.678 + Matcher::calling_convention( sig_bt, parm_regs, argcnt, true );
1.679 +}
1.680 +
1.681 +
1.682 +//------------------------------match------------------------------------------
1.683 +// Construct projections for control, I/O, memory-fields, ..., and
1.684 +// return result(s) along with their RegMask info
1.685 +Node *CallNode::match( const ProjNode *proj, const Matcher *match ) {
1.686 + switch (proj->_con) {
1.687 + case TypeFunc::Control:
1.688 + case TypeFunc::I_O:
1.689 + case TypeFunc::Memory:
1.690 + return new (match->C) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
1.691 +
1.692 + case TypeFunc::Parms+1: // For LONG & DOUBLE returns
1.693 + assert(tf()->_range->field_at(TypeFunc::Parms+1) == Type::HALF, "");
1.694 + // 2nd half of doubles and longs
1.695 + return new (match->C) MachProjNode(this,proj->_con, RegMask::Empty, (uint)OptoReg::Bad);
1.696 +
1.697 + case TypeFunc::Parms: { // Normal returns
1.698 + uint ideal_reg = tf()->range()->field_at(TypeFunc::Parms)->ideal_reg();
1.699 + OptoRegPair regs = is_CallRuntime()
1.700 + ? match->c_return_value(ideal_reg,true) // Calls into C runtime
1.701 + : match-> return_value(ideal_reg,true); // Calls into compiled Java code
1.702 + RegMask rm = RegMask(regs.first());
1.703 + if( OptoReg::is_valid(regs.second()) )
1.704 + rm.Insert( regs.second() );
1.705 + return new (match->C) MachProjNode(this,proj->_con,rm,ideal_reg);
1.706 + }
1.707 +
1.708 + case TypeFunc::ReturnAdr:
1.709 + case TypeFunc::FramePtr:
1.710 + default:
1.711 + ShouldNotReachHere();
1.712 + }
1.713 + return NULL;
1.714 +}
1.715 +
1.716 +// Do we Match on this edge index or not? Match no edges
1.717 +uint CallNode::match_edge(uint idx) const {
1.718 + return 0;
1.719 +}
1.720 +
1.721 +//
1.722 +// Determine whether the call could modify the field of the specified
1.723 +// instance at the specified offset.
1.724 +//
1.725 +bool CallNode::may_modify(const TypeOopPtr *t_oop, PhaseTransform *phase) {
1.726 + assert((t_oop != NULL), "sanity");
1.727 + if (t_oop->is_known_instance()) {
1.728 + // The instance_id is set only for scalar-replaceable allocations which
1.729 + // are not passed as arguments according to Escape Analysis.
1.730 + return false;
1.731 + }
1.732 + if (t_oop->is_ptr_to_boxed_value()) {
1.733 + ciKlass* boxing_klass = t_oop->klass();
1.734 + if (is_CallStaticJava() && as_CallStaticJava()->is_boxing_method()) {
1.735 + // Skip unrelated boxing methods.
1.736 + Node* proj = proj_out(TypeFunc::Parms);
1.737 + if ((proj == NULL) || (phase->type(proj)->is_instptr()->klass() != boxing_klass)) {
1.738 + return false;
1.739 + }
1.740 + }
1.741 + if (is_CallJava() && as_CallJava()->method() != NULL) {
1.742 + ciMethod* meth = as_CallJava()->method();
1.743 + if (meth->is_accessor()) {
1.744 + return false;
1.745 + }
1.746 + // May modify (by reflection) if an boxing object is passed
1.747 + // as argument or returned.
1.748 + if (returns_pointer() && (proj_out(TypeFunc::Parms) != NULL)) {
1.749 + Node* proj = proj_out(TypeFunc::Parms);
1.750 + const TypeInstPtr* inst_t = phase->type(proj)->isa_instptr();
1.751 + if ((inst_t != NULL) && (!inst_t->klass_is_exact() ||
1.752 + (inst_t->klass() == boxing_klass))) {
1.753 + return true;
1.754 + }
1.755 + }
1.756 + const TypeTuple* d = tf()->domain();
1.757 + for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1.758 + const TypeInstPtr* inst_t = d->field_at(i)->isa_instptr();
1.759 + if ((inst_t != NULL) && (!inst_t->klass_is_exact() ||
1.760 + (inst_t->klass() == boxing_klass))) {
1.761 + return true;
1.762 + }
1.763 + }
1.764 + return false;
1.765 + }
1.766 + }
1.767 + return true;
1.768 +}
1.769 +
1.770 +// Does this call have a direct reference to n other than debug information?
1.771 +bool CallNode::has_non_debug_use(Node *n) {
1.772 + const TypeTuple * d = tf()->domain();
1.773 + for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
1.774 + Node *arg = in(i);
1.775 + if (arg == n) {
1.776 + return true;
1.777 + }
1.778 + }
1.779 + return false;
1.780 +}
1.781 +
1.782 +// Returns the unique CheckCastPP of a call
1.783 +// or 'this' if there are several CheckCastPP
1.784 +// or returns NULL if there is no one.
1.785 +Node *CallNode::result_cast() {
1.786 + Node *cast = NULL;
1.787 +
1.788 + Node *p = proj_out(TypeFunc::Parms);
1.789 + if (p == NULL)
1.790 + return NULL;
1.791 +
1.792 + for (DUIterator_Fast imax, i = p->fast_outs(imax); i < imax; i++) {
1.793 + Node *use = p->fast_out(i);
1.794 + if (use->is_CheckCastPP()) {
1.795 + if (cast != NULL) {
1.796 + return this; // more than 1 CheckCastPP
1.797 + }
1.798 + cast = use;
1.799 + }
1.800 + }
1.801 + return cast;
1.802 +}
1.803 +
1.804 +
1.805 +void CallNode::extract_projections(CallProjections* projs, bool separate_io_proj) {
1.806 + projs->fallthrough_proj = NULL;
1.807 + projs->fallthrough_catchproj = NULL;
1.808 + projs->fallthrough_ioproj = NULL;
1.809 + projs->catchall_ioproj = NULL;
1.810 + projs->catchall_catchproj = NULL;
1.811 + projs->fallthrough_memproj = NULL;
1.812 + projs->catchall_memproj = NULL;
1.813 + projs->resproj = NULL;
1.814 + projs->exobj = NULL;
1.815 +
1.816 + for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
1.817 + ProjNode *pn = fast_out(i)->as_Proj();
1.818 + if (pn->outcnt() == 0) continue;
1.819 + switch (pn->_con) {
1.820 + case TypeFunc::Control:
1.821 + {
1.822 + // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
1.823 + projs->fallthrough_proj = pn;
1.824 + DUIterator_Fast jmax, j = pn->fast_outs(jmax);
1.825 + const Node *cn = pn->fast_out(j);
1.826 + if (cn->is_Catch()) {
1.827 + ProjNode *cpn = NULL;
1.828 + for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
1.829 + cpn = cn->fast_out(k)->as_Proj();
1.830 + assert(cpn->is_CatchProj(), "must be a CatchProjNode");
1.831 + if (cpn->_con == CatchProjNode::fall_through_index)
1.832 + projs->fallthrough_catchproj = cpn;
1.833 + else {
1.834 + assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
1.835 + projs->catchall_catchproj = cpn;
1.836 + }
1.837 + }
1.838 + }
1.839 + break;
1.840 + }
1.841 + case TypeFunc::I_O:
1.842 + if (pn->_is_io_use)
1.843 + projs->catchall_ioproj = pn;
1.844 + else
1.845 + projs->fallthrough_ioproj = pn;
1.846 + for (DUIterator j = pn->outs(); pn->has_out(j); j++) {
1.847 + Node* e = pn->out(j);
1.848 + if (e->Opcode() == Op_CreateEx && e->in(0)->is_CatchProj() && e->outcnt() > 0) {
1.849 + assert(projs->exobj == NULL, "only one");
1.850 + projs->exobj = e;
1.851 + }
1.852 + }
1.853 + break;
1.854 + case TypeFunc::Memory:
1.855 + if (pn->_is_io_use)
1.856 + projs->catchall_memproj = pn;
1.857 + else
1.858 + projs->fallthrough_memproj = pn;
1.859 + break;
1.860 + case TypeFunc::Parms:
1.861 + projs->resproj = pn;
1.862 + break;
1.863 + default:
1.864 + assert(false, "unexpected projection from allocation node.");
1.865 + }
1.866 + }
1.867 +
1.868 + // The resproj may not exist because the result couuld be ignored
1.869 + // and the exception object may not exist if an exception handler
1.870 + // swallows the exception but all the other must exist and be found.
1.871 + assert(projs->fallthrough_proj != NULL, "must be found");
1.872 + assert(Compile::current()->inlining_incrementally() || projs->fallthrough_catchproj != NULL, "must be found");
1.873 + assert(Compile::current()->inlining_incrementally() || projs->fallthrough_memproj != NULL, "must be found");
1.874 + assert(Compile::current()->inlining_incrementally() || projs->fallthrough_ioproj != NULL, "must be found");
1.875 + assert(Compile::current()->inlining_incrementally() || projs->catchall_catchproj != NULL, "must be found");
1.876 + if (separate_io_proj) {
1.877 + assert(Compile::current()->inlining_incrementally() || projs->catchall_memproj != NULL, "must be found");
1.878 + assert(Compile::current()->inlining_incrementally() || projs->catchall_ioproj != NULL, "must be found");
1.879 + }
1.880 +}
1.881 +
1.882 +Node *CallNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1.883 + CallGenerator* cg = generator();
1.884 + if (can_reshape && cg != NULL && cg->is_mh_late_inline() && !cg->already_attempted()) {
1.885 + // Check whether this MH handle call becomes a candidate for inlining
1.886 + ciMethod* callee = cg->method();
1.887 + vmIntrinsics::ID iid = callee->intrinsic_id();
1.888 + if (iid == vmIntrinsics::_invokeBasic) {
1.889 + if (in(TypeFunc::Parms)->Opcode() == Op_ConP) {
1.890 + phase->C->prepend_late_inline(cg);
1.891 + set_generator(NULL);
1.892 + }
1.893 + } else {
1.894 + assert(callee->has_member_arg(), "wrong type of call?");
1.895 + if (in(TypeFunc::Parms + callee->arg_size() - 1)->Opcode() == Op_ConP) {
1.896 + phase->C->prepend_late_inline(cg);
1.897 + set_generator(NULL);
1.898 + }
1.899 + }
1.900 + }
1.901 + return SafePointNode::Ideal(phase, can_reshape);
1.902 +}
1.903 +
1.904 +
1.905 +//=============================================================================
1.906 +uint CallJavaNode::size_of() const { return sizeof(*this); }
1.907 +uint CallJavaNode::cmp( const Node &n ) const {
1.908 + CallJavaNode &call = (CallJavaNode&)n;
1.909 + return CallNode::cmp(call) && _method == call._method;
1.910 +}
1.911 +#ifndef PRODUCT
1.912 +void CallJavaNode::dump_spec(outputStream *st) const {
1.913 + if( _method ) _method->print_short_name(st);
1.914 + CallNode::dump_spec(st);
1.915 +}
1.916 +#endif
1.917 +
1.918 +//=============================================================================
1.919 +uint CallStaticJavaNode::size_of() const { return sizeof(*this); }
1.920 +uint CallStaticJavaNode::cmp( const Node &n ) const {
1.921 + CallStaticJavaNode &call = (CallStaticJavaNode&)n;
1.922 + return CallJavaNode::cmp(call);
1.923 +}
1.924 +
1.925 +//----------------------------uncommon_trap_request----------------------------
1.926 +// If this is an uncommon trap, return the request code, else zero.
1.927 +int CallStaticJavaNode::uncommon_trap_request() const {
1.928 + if (_name != NULL && !strcmp(_name, "uncommon_trap")) {
1.929 + return extract_uncommon_trap_request(this);
1.930 + }
1.931 + return 0;
1.932 +}
1.933 +int CallStaticJavaNode::extract_uncommon_trap_request(const Node* call) {
1.934 +#ifndef PRODUCT
1.935 + if (!(call->req() > TypeFunc::Parms &&
1.936 + call->in(TypeFunc::Parms) != NULL &&
1.937 + call->in(TypeFunc::Parms)->is_Con())) {
1.938 + assert(in_dump() != 0, "OK if dumping");
1.939 + tty->print("[bad uncommon trap]");
1.940 + return 0;
1.941 + }
1.942 +#endif
1.943 + return call->in(TypeFunc::Parms)->bottom_type()->is_int()->get_con();
1.944 +}
1.945 +
1.946 +#ifndef PRODUCT
1.947 +void CallStaticJavaNode::dump_spec(outputStream *st) const {
1.948 + st->print("# Static ");
1.949 + if (_name != NULL) {
1.950 + st->print("%s", _name);
1.951 + int trap_req = uncommon_trap_request();
1.952 + if (trap_req != 0) {
1.953 + char buf[100];
1.954 + st->print("(%s)",
1.955 + Deoptimization::format_trap_request(buf, sizeof(buf),
1.956 + trap_req));
1.957 + }
1.958 + st->print(" ");
1.959 + }
1.960 + CallJavaNode::dump_spec(st);
1.961 +}
1.962 +#endif
1.963 +
1.964 +//=============================================================================
1.965 +uint CallDynamicJavaNode::size_of() const { return sizeof(*this); }
1.966 +uint CallDynamicJavaNode::cmp( const Node &n ) const {
1.967 + CallDynamicJavaNode &call = (CallDynamicJavaNode&)n;
1.968 + return CallJavaNode::cmp(call);
1.969 +}
1.970 +#ifndef PRODUCT
1.971 +void CallDynamicJavaNode::dump_spec(outputStream *st) const {
1.972 + st->print("# Dynamic ");
1.973 + CallJavaNode::dump_spec(st);
1.974 +}
1.975 +#endif
1.976 +
1.977 +//=============================================================================
1.978 +uint CallRuntimeNode::size_of() const { return sizeof(*this); }
1.979 +uint CallRuntimeNode::cmp( const Node &n ) const {
1.980 + CallRuntimeNode &call = (CallRuntimeNode&)n;
1.981 + return CallNode::cmp(call) && !strcmp(_name,call._name);
1.982 +}
1.983 +#ifndef PRODUCT
1.984 +void CallRuntimeNode::dump_spec(outputStream *st) const {
1.985 + st->print("# ");
1.986 + st->print("%s", _name);
1.987 + CallNode::dump_spec(st);
1.988 +}
1.989 +#endif
1.990 +
1.991 +//------------------------------calling_convention-----------------------------
1.992 +void CallRuntimeNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const {
1.993 + Matcher::c_calling_convention( sig_bt, parm_regs, argcnt );
1.994 +}
1.995 +
1.996 +//=============================================================================
1.997 +//------------------------------calling_convention-----------------------------
1.998 +
1.999 +
1.1000 +//=============================================================================
1.1001 +#ifndef PRODUCT
1.1002 +void CallLeafNode::dump_spec(outputStream *st) const {
1.1003 + st->print("# ");
1.1004 + st->print("%s", _name);
1.1005 + CallNode::dump_spec(st);
1.1006 +}
1.1007 +#endif
1.1008 +
1.1009 +//=============================================================================
1.1010 +
1.1011 +void SafePointNode::set_local(JVMState* jvms, uint idx, Node *c) {
1.1012 + assert(verify_jvms(jvms), "jvms must match");
1.1013 + int loc = jvms->locoff() + idx;
1.1014 + if (in(loc)->is_top() && idx > 0 && !c->is_top() ) {
1.1015 + // If current local idx is top then local idx - 1 could
1.1016 + // be a long/double that needs to be killed since top could
1.1017 + // represent the 2nd half ofthe long/double.
1.1018 + uint ideal = in(loc -1)->ideal_reg();
1.1019 + if (ideal == Op_RegD || ideal == Op_RegL) {
1.1020 + // set other (low index) half to top
1.1021 + set_req(loc - 1, in(loc));
1.1022 + }
1.1023 + }
1.1024 + set_req(loc, c);
1.1025 +}
1.1026 +
1.1027 +uint SafePointNode::size_of() const { return sizeof(*this); }
1.1028 +uint SafePointNode::cmp( const Node &n ) const {
1.1029 + return (&n == this); // Always fail except on self
1.1030 +}
1.1031 +
1.1032 +//-------------------------set_next_exception----------------------------------
1.1033 +void SafePointNode::set_next_exception(SafePointNode* n) {
1.1034 + assert(n == NULL || n->Opcode() == Op_SafePoint, "correct value for next_exception");
1.1035 + if (len() == req()) {
1.1036 + if (n != NULL) add_prec(n);
1.1037 + } else {
1.1038 + set_prec(req(), n);
1.1039 + }
1.1040 +}
1.1041 +
1.1042 +
1.1043 +//----------------------------next_exception-----------------------------------
1.1044 +SafePointNode* SafePointNode::next_exception() const {
1.1045 + if (len() == req()) {
1.1046 + return NULL;
1.1047 + } else {
1.1048 + Node* n = in(req());
1.1049 + assert(n == NULL || n->Opcode() == Op_SafePoint, "no other uses of prec edges");
1.1050 + return (SafePointNode*) n;
1.1051 + }
1.1052 +}
1.1053 +
1.1054 +
1.1055 +//------------------------------Ideal------------------------------------------
1.1056 +// Skip over any collapsed Regions
1.1057 +Node *SafePointNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1.1058 + return remove_dead_region(phase, can_reshape) ? this : NULL;
1.1059 +}
1.1060 +
1.1061 +//------------------------------Identity---------------------------------------
1.1062 +// Remove obviously duplicate safepoints
1.1063 +Node *SafePointNode::Identity( PhaseTransform *phase ) {
1.1064 +
1.1065 + // If you have back to back safepoints, remove one
1.1066 + if( in(TypeFunc::Control)->is_SafePoint() )
1.1067 + return in(TypeFunc::Control);
1.1068 +
1.1069 + if( in(0)->is_Proj() ) {
1.1070 + Node *n0 = in(0)->in(0);
1.1071 + // Check if he is a call projection (except Leaf Call)
1.1072 + if( n0->is_Catch() ) {
1.1073 + n0 = n0->in(0)->in(0);
1.1074 + assert( n0->is_Call(), "expect a call here" );
1.1075 + }
1.1076 + if( n0->is_Call() && n0->as_Call()->guaranteed_safepoint() ) {
1.1077 + // Useless Safepoint, so remove it
1.1078 + return in(TypeFunc::Control);
1.1079 + }
1.1080 + }
1.1081 +
1.1082 + return this;
1.1083 +}
1.1084 +
1.1085 +//------------------------------Value------------------------------------------
1.1086 +const Type *SafePointNode::Value( PhaseTransform *phase ) const {
1.1087 + if( phase->type(in(0)) == Type::TOP ) return Type::TOP;
1.1088 + if( phase->eqv( in(0), this ) ) return Type::TOP; // Dead infinite loop
1.1089 + return Type::CONTROL;
1.1090 +}
1.1091 +
1.1092 +#ifndef PRODUCT
1.1093 +void SafePointNode::dump_spec(outputStream *st) const {
1.1094 + st->print(" SafePoint ");
1.1095 +}
1.1096 +#endif
1.1097 +
1.1098 +const RegMask &SafePointNode::in_RegMask(uint idx) const {
1.1099 + if( idx < TypeFunc::Parms ) return RegMask::Empty;
1.1100 + // Values outside the domain represent debug info
1.1101 + return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
1.1102 +}
1.1103 +const RegMask &SafePointNode::out_RegMask() const {
1.1104 + return RegMask::Empty;
1.1105 +}
1.1106 +
1.1107 +
1.1108 +void SafePointNode::grow_stack(JVMState* jvms, uint grow_by) {
1.1109 + assert((int)grow_by > 0, "sanity");
1.1110 + int monoff = jvms->monoff();
1.1111 + int scloff = jvms->scloff();
1.1112 + int endoff = jvms->endoff();
1.1113 + assert(endoff == (int)req(), "no other states or debug info after me");
1.1114 + Node* top = Compile::current()->top();
1.1115 + for (uint i = 0; i < grow_by; i++) {
1.1116 + ins_req(monoff, top);
1.1117 + }
1.1118 + jvms->set_monoff(monoff + grow_by);
1.1119 + jvms->set_scloff(scloff + grow_by);
1.1120 + jvms->set_endoff(endoff + grow_by);
1.1121 +}
1.1122 +
1.1123 +void SafePointNode::push_monitor(const FastLockNode *lock) {
1.1124 + // Add a LockNode, which points to both the original BoxLockNode (the
1.1125 + // stack space for the monitor) and the Object being locked.
1.1126 + const int MonitorEdges = 2;
1.1127 + assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges");
1.1128 + assert(req() == jvms()->endoff(), "correct sizing");
1.1129 + int nextmon = jvms()->scloff();
1.1130 + if (GenerateSynchronizationCode) {
1.1131 + ins_req(nextmon, lock->box_node());
1.1132 + ins_req(nextmon+1, lock->obj_node());
1.1133 + } else {
1.1134 + Node* top = Compile::current()->top();
1.1135 + ins_req(nextmon, top);
1.1136 + ins_req(nextmon, top);
1.1137 + }
1.1138 + jvms()->set_scloff(nextmon + MonitorEdges);
1.1139 + jvms()->set_endoff(req());
1.1140 +}
1.1141 +
1.1142 +void SafePointNode::pop_monitor() {
1.1143 + // Delete last monitor from debug info
1.1144 + debug_only(int num_before_pop = jvms()->nof_monitors());
1.1145 + const int MonitorEdges = 2;
1.1146 + assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges");
1.1147 + int scloff = jvms()->scloff();
1.1148 + int endoff = jvms()->endoff();
1.1149 + int new_scloff = scloff - MonitorEdges;
1.1150 + int new_endoff = endoff - MonitorEdges;
1.1151 + jvms()->set_scloff(new_scloff);
1.1152 + jvms()->set_endoff(new_endoff);
1.1153 + while (scloff > new_scloff) del_req_ordered(--scloff);
1.1154 + assert(jvms()->nof_monitors() == num_before_pop-1, "");
1.1155 +}
1.1156 +
1.1157 +Node *SafePointNode::peek_monitor_box() const {
1.1158 + int mon = jvms()->nof_monitors() - 1;
1.1159 + assert(mon >= 0, "most have a monitor");
1.1160 + return monitor_box(jvms(), mon);
1.1161 +}
1.1162 +
1.1163 +Node *SafePointNode::peek_monitor_obj() const {
1.1164 + int mon = jvms()->nof_monitors() - 1;
1.1165 + assert(mon >= 0, "most have a monitor");
1.1166 + return monitor_obj(jvms(), mon);
1.1167 +}
1.1168 +
1.1169 +// Do we Match on this edge index or not? Match no edges
1.1170 +uint SafePointNode::match_edge(uint idx) const {
1.1171 + if( !needs_polling_address_input() )
1.1172 + return 0;
1.1173 +
1.1174 + return (TypeFunc::Parms == idx);
1.1175 +}
1.1176 +
1.1177 +//============== SafePointScalarObjectNode ==============
1.1178 +
1.1179 +SafePointScalarObjectNode::SafePointScalarObjectNode(const TypeOopPtr* tp,
1.1180 +#ifdef ASSERT
1.1181 + AllocateNode* alloc,
1.1182 +#endif
1.1183 + uint first_index,
1.1184 + uint n_fields) :
1.1185 + TypeNode(tp, 1), // 1 control input -- seems required. Get from root.
1.1186 +#ifdef ASSERT
1.1187 + _alloc(alloc),
1.1188 +#endif
1.1189 + _first_index(first_index),
1.1190 + _n_fields(n_fields)
1.1191 +{
1.1192 + init_class_id(Class_SafePointScalarObject);
1.1193 +}
1.1194 +
1.1195 +// Do not allow value-numbering for SafePointScalarObject node.
1.1196 +uint SafePointScalarObjectNode::hash() const { return NO_HASH; }
1.1197 +uint SafePointScalarObjectNode::cmp( const Node &n ) const {
1.1198 + return (&n == this); // Always fail except on self
1.1199 +}
1.1200 +
1.1201 +uint SafePointScalarObjectNode::ideal_reg() const {
1.1202 + return 0; // No matching to machine instruction
1.1203 +}
1.1204 +
1.1205 +const RegMask &SafePointScalarObjectNode::in_RegMask(uint idx) const {
1.1206 + return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
1.1207 +}
1.1208 +
1.1209 +const RegMask &SafePointScalarObjectNode::out_RegMask() const {
1.1210 + return RegMask::Empty;
1.1211 +}
1.1212 +
1.1213 +uint SafePointScalarObjectNode::match_edge(uint idx) const {
1.1214 + return 0;
1.1215 +}
1.1216 +
1.1217 +SafePointScalarObjectNode*
1.1218 +SafePointScalarObjectNode::clone(Dict* sosn_map) const {
1.1219 + void* cached = (*sosn_map)[(void*)this];
1.1220 + if (cached != NULL) {
1.1221 + return (SafePointScalarObjectNode*)cached;
1.1222 + }
1.1223 + SafePointScalarObjectNode* res = (SafePointScalarObjectNode*)Node::clone();
1.1224 + sosn_map->Insert((void*)this, (void*)res);
1.1225 + return res;
1.1226 +}
1.1227 +
1.1228 +
1.1229 +#ifndef PRODUCT
1.1230 +void SafePointScalarObjectNode::dump_spec(outputStream *st) const {
1.1231 + st->print(" # fields@[%d..%d]", first_index(),
1.1232 + first_index() + n_fields() - 1);
1.1233 +}
1.1234 +
1.1235 +#endif
1.1236 +
1.1237 +//=============================================================================
1.1238 +uint AllocateNode::size_of() const { return sizeof(*this); }
1.1239 +
1.1240 +AllocateNode::AllocateNode(Compile* C, const TypeFunc *atype,
1.1241 + Node *ctrl, Node *mem, Node *abio,
1.1242 + Node *size, Node *klass_node, Node *initial_test)
1.1243 + : CallNode(atype, NULL, TypeRawPtr::BOTTOM)
1.1244 +{
1.1245 + init_class_id(Class_Allocate);
1.1246 + init_flags(Flag_is_macro);
1.1247 + _is_scalar_replaceable = false;
1.1248 + _is_non_escaping = false;
1.1249 + Node *topnode = C->top();
1.1250 +
1.1251 + init_req( TypeFunc::Control , ctrl );
1.1252 + init_req( TypeFunc::I_O , abio );
1.1253 + init_req( TypeFunc::Memory , mem );
1.1254 + init_req( TypeFunc::ReturnAdr, topnode );
1.1255 + init_req( TypeFunc::FramePtr , topnode );
1.1256 + init_req( AllocSize , size);
1.1257 + init_req( KlassNode , klass_node);
1.1258 + init_req( InitialTest , initial_test);
1.1259 + init_req( ALength , topnode);
1.1260 + C->add_macro_node(this);
1.1261 +}
1.1262 +
1.1263 +//=============================================================================
1.1264 +Node* AllocateArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1.1265 + if (remove_dead_region(phase, can_reshape)) return this;
1.1266 + // Don't bother trying to transform a dead node
1.1267 + if (in(0) && in(0)->is_top()) return NULL;
1.1268 +
1.1269 + const Type* type = phase->type(Ideal_length());
1.1270 + if (type->isa_int() && type->is_int()->_hi < 0) {
1.1271 + if (can_reshape) {
1.1272 + PhaseIterGVN *igvn = phase->is_IterGVN();
1.1273 + // Unreachable fall through path (negative array length),
1.1274 + // the allocation can only throw so disconnect it.
1.1275 + Node* proj = proj_out(TypeFunc::Control);
1.1276 + Node* catchproj = NULL;
1.1277 + if (proj != NULL) {
1.1278 + for (DUIterator_Fast imax, i = proj->fast_outs(imax); i < imax; i++) {
1.1279 + Node *cn = proj->fast_out(i);
1.1280 + if (cn->is_Catch()) {
1.1281 + catchproj = cn->as_Multi()->proj_out(CatchProjNode::fall_through_index);
1.1282 + break;
1.1283 + }
1.1284 + }
1.1285 + }
1.1286 + if (catchproj != NULL && catchproj->outcnt() > 0 &&
1.1287 + (catchproj->outcnt() > 1 ||
1.1288 + catchproj->unique_out()->Opcode() != Op_Halt)) {
1.1289 + assert(catchproj->is_CatchProj(), "must be a CatchProjNode");
1.1290 + Node* nproj = catchproj->clone();
1.1291 + igvn->register_new_node_with_optimizer(nproj);
1.1292 +
1.1293 + Node *frame = new (phase->C) ParmNode( phase->C->start(), TypeFunc::FramePtr );
1.1294 + frame = phase->transform(frame);
1.1295 + // Halt & Catch Fire
1.1296 + Node *halt = new (phase->C) HaltNode( nproj, frame );
1.1297 + phase->C->root()->add_req(halt);
1.1298 + phase->transform(halt);
1.1299 +
1.1300 + igvn->replace_node(catchproj, phase->C->top());
1.1301 + return this;
1.1302 + }
1.1303 + } else {
1.1304 + // Can't correct it during regular GVN so register for IGVN
1.1305 + phase->C->record_for_igvn(this);
1.1306 + }
1.1307 + }
1.1308 + return NULL;
1.1309 +}
1.1310 +
1.1311 +// Retrieve the length from the AllocateArrayNode. Narrow the type with a
1.1312 +// CastII, if appropriate. If we are not allowed to create new nodes, and
1.1313 +// a CastII is appropriate, return NULL.
1.1314 +Node *AllocateArrayNode::make_ideal_length(const TypeOopPtr* oop_type, PhaseTransform *phase, bool allow_new_nodes) {
1.1315 + Node *length = in(AllocateNode::ALength);
1.1316 + assert(length != NULL, "length is not null");
1.1317 +
1.1318 + const TypeInt* length_type = phase->find_int_type(length);
1.1319 + const TypeAryPtr* ary_type = oop_type->isa_aryptr();
1.1320 +
1.1321 + if (ary_type != NULL && length_type != NULL) {
1.1322 + const TypeInt* narrow_length_type = ary_type->narrow_size_type(length_type);
1.1323 + if (narrow_length_type != length_type) {
1.1324 + // Assert one of:
1.1325 + // - the narrow_length is 0
1.1326 + // - the narrow_length is not wider than length
1.1327 + assert(narrow_length_type == TypeInt::ZERO ||
1.1328 + length_type->is_con() && narrow_length_type->is_con() &&
1.1329 + (narrow_length_type->_hi <= length_type->_lo) ||
1.1330 + (narrow_length_type->_hi <= length_type->_hi &&
1.1331 + narrow_length_type->_lo >= length_type->_lo),
1.1332 + "narrow type must be narrower than length type");
1.1333 +
1.1334 + // Return NULL if new nodes are not allowed
1.1335 + if (!allow_new_nodes) return NULL;
1.1336 + // Create a cast which is control dependent on the initialization to
1.1337 + // propagate the fact that the array length must be positive.
1.1338 + length = new (phase->C) CastIINode(length, narrow_length_type);
1.1339 + length->set_req(0, initialization()->proj_out(0));
1.1340 + }
1.1341 + }
1.1342 +
1.1343 + return length;
1.1344 +}
1.1345 +
1.1346 +//=============================================================================
1.1347 +uint LockNode::size_of() const { return sizeof(*this); }
1.1348 +
1.1349 +// Redundant lock elimination
1.1350 +//
1.1351 +// There are various patterns of locking where we release and
1.1352 +// immediately reacquire a lock in a piece of code where no operations
1.1353 +// occur in between that would be observable. In those cases we can
1.1354 +// skip releasing and reacquiring the lock without violating any
1.1355 +// fairness requirements. Doing this around a loop could cause a lock
1.1356 +// to be held for a very long time so we concentrate on non-looping
1.1357 +// control flow. We also require that the operations are fully
1.1358 +// redundant meaning that we don't introduce new lock operations on
1.1359 +// some paths so to be able to eliminate it on others ala PRE. This
1.1360 +// would probably require some more extensive graph manipulation to
1.1361 +// guarantee that the memory edges were all handled correctly.
1.1362 +//
1.1363 +// Assuming p is a simple predicate which can't trap in any way and s
1.1364 +// is a synchronized method consider this code:
1.1365 +//
1.1366 +// s();
1.1367 +// if (p)
1.1368 +// s();
1.1369 +// else
1.1370 +// s();
1.1371 +// s();
1.1372 +//
1.1373 +// 1. The unlocks of the first call to s can be eliminated if the
1.1374 +// locks inside the then and else branches are eliminated.
1.1375 +//
1.1376 +// 2. The unlocks of the then and else branches can be eliminated if
1.1377 +// the lock of the final call to s is eliminated.
1.1378 +//
1.1379 +// Either of these cases subsumes the simple case of sequential control flow
1.1380 +//
1.1381 +// Addtionally we can eliminate versions without the else case:
1.1382 +//
1.1383 +// s();
1.1384 +// if (p)
1.1385 +// s();
1.1386 +// s();
1.1387 +//
1.1388 +// 3. In this case we eliminate the unlock of the first s, the lock
1.1389 +// and unlock in the then case and the lock in the final s.
1.1390 +//
1.1391 +// Note also that in all these cases the then/else pieces don't have
1.1392 +// to be trivial as long as they begin and end with synchronization
1.1393 +// operations.
1.1394 +//
1.1395 +// s();
1.1396 +// if (p)
1.1397 +// s();
1.1398 +// f();
1.1399 +// s();
1.1400 +// s();
1.1401 +//
1.1402 +// The code will work properly for this case, leaving in the unlock
1.1403 +// before the call to f and the relock after it.
1.1404 +//
1.1405 +// A potentially interesting case which isn't handled here is when the
1.1406 +// locking is partially redundant.
1.1407 +//
1.1408 +// s();
1.1409 +// if (p)
1.1410 +// s();
1.1411 +//
1.1412 +// This could be eliminated putting unlocking on the else case and
1.1413 +// eliminating the first unlock and the lock in the then side.
1.1414 +// Alternatively the unlock could be moved out of the then side so it
1.1415 +// was after the merge and the first unlock and second lock
1.1416 +// eliminated. This might require less manipulation of the memory
1.1417 +// state to get correct.
1.1418 +//
1.1419 +// Additionally we might allow work between a unlock and lock before
1.1420 +// giving up eliminating the locks. The current code disallows any
1.1421 +// conditional control flow between these operations. A formulation
1.1422 +// similar to partial redundancy elimination computing the
1.1423 +// availability of unlocking and the anticipatability of locking at a
1.1424 +// program point would allow detection of fully redundant locking with
1.1425 +// some amount of work in between. I'm not sure how often I really
1.1426 +// think that would occur though. Most of the cases I've seen
1.1427 +// indicate it's likely non-trivial work would occur in between.
1.1428 +// There may be other more complicated constructs where we could
1.1429 +// eliminate locking but I haven't seen any others appear as hot or
1.1430 +// interesting.
1.1431 +//
1.1432 +// Locking and unlocking have a canonical form in ideal that looks
1.1433 +// roughly like this:
1.1434 +//
1.1435 +// <obj>
1.1436 +// | \\------+
1.1437 +// | \ \
1.1438 +// | BoxLock \
1.1439 +// | | | \
1.1440 +// | | \ \
1.1441 +// | | FastLock
1.1442 +// | | /
1.1443 +// | | /
1.1444 +// | | |
1.1445 +//
1.1446 +// Lock
1.1447 +// |
1.1448 +// Proj #0
1.1449 +// |
1.1450 +// MembarAcquire
1.1451 +// |
1.1452 +// Proj #0
1.1453 +//
1.1454 +// MembarRelease
1.1455 +// |
1.1456 +// Proj #0
1.1457 +// |
1.1458 +// Unlock
1.1459 +// |
1.1460 +// Proj #0
1.1461 +//
1.1462 +//
1.1463 +// This code proceeds by processing Lock nodes during PhaseIterGVN
1.1464 +// and searching back through its control for the proper code
1.1465 +// patterns. Once it finds a set of lock and unlock operations to
1.1466 +// eliminate they are marked as eliminatable which causes the
1.1467 +// expansion of the Lock and Unlock macro nodes to make the operation a NOP
1.1468 +//
1.1469 +//=============================================================================
1.1470 +
1.1471 +//
1.1472 +// Utility function to skip over uninteresting control nodes. Nodes skipped are:
1.1473 +// - copy regions. (These may not have been optimized away yet.)
1.1474 +// - eliminated locking nodes
1.1475 +//
1.1476 +static Node *next_control(Node *ctrl) {
1.1477 + if (ctrl == NULL)
1.1478 + return NULL;
1.1479 + while (1) {
1.1480 + if (ctrl->is_Region()) {
1.1481 + RegionNode *r = ctrl->as_Region();
1.1482 + Node *n = r->is_copy();
1.1483 + if (n == NULL)
1.1484 + break; // hit a region, return it
1.1485 + else
1.1486 + ctrl = n;
1.1487 + } else if (ctrl->is_Proj()) {
1.1488 + Node *in0 = ctrl->in(0);
1.1489 + if (in0->is_AbstractLock() && in0->as_AbstractLock()->is_eliminated()) {
1.1490 + ctrl = in0->in(0);
1.1491 + } else {
1.1492 + break;
1.1493 + }
1.1494 + } else {
1.1495 + break; // found an interesting control
1.1496 + }
1.1497 + }
1.1498 + return ctrl;
1.1499 +}
1.1500 +//
1.1501 +// Given a control, see if it's the control projection of an Unlock which
1.1502 +// operating on the same object as lock.
1.1503 +//
1.1504 +bool AbstractLockNode::find_matching_unlock(const Node* ctrl, LockNode* lock,
1.1505 + GrowableArray<AbstractLockNode*> &lock_ops) {
1.1506 + ProjNode *ctrl_proj = (ctrl->is_Proj()) ? ctrl->as_Proj() : NULL;
1.1507 + if (ctrl_proj != NULL && ctrl_proj->_con == TypeFunc::Control) {
1.1508 + Node *n = ctrl_proj->in(0);
1.1509 + if (n != NULL && n->is_Unlock()) {
1.1510 + UnlockNode *unlock = n->as_Unlock();
1.1511 + if (lock->obj_node()->eqv_uncast(unlock->obj_node()) &&
1.1512 + BoxLockNode::same_slot(lock->box_node(), unlock->box_node()) &&
1.1513 + !unlock->is_eliminated()) {
1.1514 + lock_ops.append(unlock);
1.1515 + return true;
1.1516 + }
1.1517 + }
1.1518 + }
1.1519 + return false;
1.1520 +}
1.1521 +
1.1522 +//
1.1523 +// Find the lock matching an unlock. Returns null if a safepoint
1.1524 +// or complicated control is encountered first.
1.1525 +LockNode *AbstractLockNode::find_matching_lock(UnlockNode* unlock) {
1.1526 + LockNode *lock_result = NULL;
1.1527 + // find the matching lock, or an intervening safepoint
1.1528 + Node *ctrl = next_control(unlock->in(0));
1.1529 + while (1) {
1.1530 + assert(ctrl != NULL, "invalid control graph");
1.1531 + assert(!ctrl->is_Start(), "missing lock for unlock");
1.1532 + if (ctrl->is_top()) break; // dead control path
1.1533 + if (ctrl->is_Proj()) ctrl = ctrl->in(0);
1.1534 + if (ctrl->is_SafePoint()) {
1.1535 + break; // found a safepoint (may be the lock we are searching for)
1.1536 + } else if (ctrl->is_Region()) {
1.1537 + // Check for a simple diamond pattern. Punt on anything more complicated
1.1538 + if (ctrl->req() == 3 && ctrl->in(1) != NULL && ctrl->in(2) != NULL) {
1.1539 + Node *in1 = next_control(ctrl->in(1));
1.1540 + Node *in2 = next_control(ctrl->in(2));
1.1541 + if (((in1->is_IfTrue() && in2->is_IfFalse()) ||
1.1542 + (in2->is_IfTrue() && in1->is_IfFalse())) && (in1->in(0) == in2->in(0))) {
1.1543 + ctrl = next_control(in1->in(0)->in(0));
1.1544 + } else {
1.1545 + break;
1.1546 + }
1.1547 + } else {
1.1548 + break;
1.1549 + }
1.1550 + } else {
1.1551 + ctrl = next_control(ctrl->in(0)); // keep searching
1.1552 + }
1.1553 + }
1.1554 + if (ctrl->is_Lock()) {
1.1555 + LockNode *lock = ctrl->as_Lock();
1.1556 + if (lock->obj_node()->eqv_uncast(unlock->obj_node()) &&
1.1557 + BoxLockNode::same_slot(lock->box_node(), unlock->box_node())) {
1.1558 + lock_result = lock;
1.1559 + }
1.1560 + }
1.1561 + return lock_result;
1.1562 +}
1.1563 +
1.1564 +// This code corresponds to case 3 above.
1.1565 +
1.1566 +bool AbstractLockNode::find_lock_and_unlock_through_if(Node* node, LockNode* lock,
1.1567 + GrowableArray<AbstractLockNode*> &lock_ops) {
1.1568 + Node* if_node = node->in(0);
1.1569 + bool if_true = node->is_IfTrue();
1.1570 +
1.1571 + if (if_node->is_If() && if_node->outcnt() == 2 && (if_true || node->is_IfFalse())) {
1.1572 + Node *lock_ctrl = next_control(if_node->in(0));
1.1573 + if (find_matching_unlock(lock_ctrl, lock, lock_ops)) {
1.1574 + Node* lock1_node = NULL;
1.1575 + ProjNode* proj = if_node->as_If()->proj_out(!if_true);
1.1576 + if (if_true) {
1.1577 + if (proj->is_IfFalse() && proj->outcnt() == 1) {
1.1578 + lock1_node = proj->unique_out();
1.1579 + }
1.1580 + } else {
1.1581 + if (proj->is_IfTrue() && proj->outcnt() == 1) {
1.1582 + lock1_node = proj->unique_out();
1.1583 + }
1.1584 + }
1.1585 + if (lock1_node != NULL && lock1_node->is_Lock()) {
1.1586 + LockNode *lock1 = lock1_node->as_Lock();
1.1587 + if (lock->obj_node()->eqv_uncast(lock1->obj_node()) &&
1.1588 + BoxLockNode::same_slot(lock->box_node(), lock1->box_node()) &&
1.1589 + !lock1->is_eliminated()) {
1.1590 + lock_ops.append(lock1);
1.1591 + return true;
1.1592 + }
1.1593 + }
1.1594 + }
1.1595 + }
1.1596 +
1.1597 + lock_ops.trunc_to(0);
1.1598 + return false;
1.1599 +}
1.1600 +
1.1601 +bool AbstractLockNode::find_unlocks_for_region(const RegionNode* region, LockNode* lock,
1.1602 + GrowableArray<AbstractLockNode*> &lock_ops) {
1.1603 + // check each control merging at this point for a matching unlock.
1.1604 + // in(0) should be self edge so skip it.
1.1605 + for (int i = 1; i < (int)region->req(); i++) {
1.1606 + Node *in_node = next_control(region->in(i));
1.1607 + if (in_node != NULL) {
1.1608 + if (find_matching_unlock(in_node, lock, lock_ops)) {
1.1609 + // found a match so keep on checking.
1.1610 + continue;
1.1611 + } else if (find_lock_and_unlock_through_if(in_node, lock, lock_ops)) {
1.1612 + continue;
1.1613 + }
1.1614 +
1.1615 + // If we fall through to here then it was some kind of node we
1.1616 + // don't understand or there wasn't a matching unlock, so give
1.1617 + // up trying to merge locks.
1.1618 + lock_ops.trunc_to(0);
1.1619 + return false;
1.1620 + }
1.1621 + }
1.1622 + return true;
1.1623 +
1.1624 +}
1.1625 +
1.1626 +#ifndef PRODUCT
1.1627 +//
1.1628 +// Create a counter which counts the number of times this lock is acquired
1.1629 +//
1.1630 +void AbstractLockNode::create_lock_counter(JVMState* state) {
1.1631 + _counter = OptoRuntime::new_named_counter(state, NamedCounter::LockCounter);
1.1632 +}
1.1633 +
1.1634 +void AbstractLockNode::set_eliminated_lock_counter() {
1.1635 + if (_counter) {
1.1636 + // Update the counter to indicate that this lock was eliminated.
1.1637 + // The counter update code will stay around even though the
1.1638 + // optimizer will eliminate the lock operation itself.
1.1639 + _counter->set_tag(NamedCounter::EliminatedLockCounter);
1.1640 + }
1.1641 +}
1.1642 +#endif
1.1643 +
1.1644 +//=============================================================================
1.1645 +Node *LockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1.1646 +
1.1647 + // perform any generic optimizations first (returns 'this' or NULL)
1.1648 + Node *result = SafePointNode::Ideal(phase, can_reshape);
1.1649 + if (result != NULL) return result;
1.1650 + // Don't bother trying to transform a dead node
1.1651 + if (in(0) && in(0)->is_top()) return NULL;
1.1652 +
1.1653 + // Now see if we can optimize away this lock. We don't actually
1.1654 + // remove the locking here, we simply set the _eliminate flag which
1.1655 + // prevents macro expansion from expanding the lock. Since we don't
1.1656 + // modify the graph, the value returned from this function is the
1.1657 + // one computed above.
1.1658 + if (can_reshape && EliminateLocks && !is_non_esc_obj()) {
1.1659 + //
1.1660 + // If we are locking an unescaped object, the lock/unlock is unnecessary
1.1661 + //
1.1662 + ConnectionGraph *cgr = phase->C->congraph();
1.1663 + if (cgr != NULL && cgr->not_global_escape(obj_node())) {
1.1664 + assert(!is_eliminated() || is_coarsened(), "sanity");
1.1665 + // The lock could be marked eliminated by lock coarsening
1.1666 + // code during first IGVN before EA. Replace coarsened flag
1.1667 + // to eliminate all associated locks/unlocks.
1.1668 + this->set_non_esc_obj();
1.1669 + return result;
1.1670 + }
1.1671 +
1.1672 + //
1.1673 + // Try lock coarsening
1.1674 + //
1.1675 + PhaseIterGVN* iter = phase->is_IterGVN();
1.1676 + if (iter != NULL && !is_eliminated()) {
1.1677 +
1.1678 + GrowableArray<AbstractLockNode*> lock_ops;
1.1679 +
1.1680 + Node *ctrl = next_control(in(0));
1.1681 +
1.1682 + // now search back for a matching Unlock
1.1683 + if (find_matching_unlock(ctrl, this, lock_ops)) {
1.1684 + // found an unlock directly preceding this lock. This is the
1.1685 + // case of single unlock directly control dependent on a
1.1686 + // single lock which is the trivial version of case 1 or 2.
1.1687 + } else if (ctrl->is_Region() ) {
1.1688 + if (find_unlocks_for_region(ctrl->as_Region(), this, lock_ops)) {
1.1689 + // found lock preceded by multiple unlocks along all paths
1.1690 + // joining at this point which is case 3 in description above.
1.1691 + }
1.1692 + } else {
1.1693 + // see if this lock comes from either half of an if and the
1.1694 + // predecessors merges unlocks and the other half of the if
1.1695 + // performs a lock.
1.1696 + if (find_lock_and_unlock_through_if(ctrl, this, lock_ops)) {
1.1697 + // found unlock splitting to an if with locks on both branches.
1.1698 + }
1.1699 + }
1.1700 +
1.1701 + if (lock_ops.length() > 0) {
1.1702 + // add ourselves to the list of locks to be eliminated.
1.1703 + lock_ops.append(this);
1.1704 +
1.1705 + #ifndef PRODUCT
1.1706 + if (PrintEliminateLocks) {
1.1707 + int locks = 0;
1.1708 + int unlocks = 0;
1.1709 + for (int i = 0; i < lock_ops.length(); i++) {
1.1710 + AbstractLockNode* lock = lock_ops.at(i);
1.1711 + if (lock->Opcode() == Op_Lock)
1.1712 + locks++;
1.1713 + else
1.1714 + unlocks++;
1.1715 + if (Verbose) {
1.1716 + lock->dump(1);
1.1717 + }
1.1718 + }
1.1719 + tty->print_cr("***Eliminated %d unlocks and %d locks", unlocks, locks);
1.1720 + }
1.1721 + #endif
1.1722 +
1.1723 + // for each of the identified locks, mark them
1.1724 + // as eliminatable
1.1725 + for (int i = 0; i < lock_ops.length(); i++) {
1.1726 + AbstractLockNode* lock = lock_ops.at(i);
1.1727 +
1.1728 + // Mark it eliminated by coarsening and update any counters
1.1729 + lock->set_coarsened();
1.1730 + }
1.1731 + } else if (ctrl->is_Region() &&
1.1732 + iter->_worklist.member(ctrl)) {
1.1733 + // We weren't able to find any opportunities but the region this
1.1734 + // lock is control dependent on hasn't been processed yet so put
1.1735 + // this lock back on the worklist so we can check again once any
1.1736 + // region simplification has occurred.
1.1737 + iter->_worklist.push(this);
1.1738 + }
1.1739 + }
1.1740 + }
1.1741 +
1.1742 + return result;
1.1743 +}
1.1744 +
1.1745 +//=============================================================================
1.1746 +bool LockNode::is_nested_lock_region() {
1.1747 + BoxLockNode* box = box_node()->as_BoxLock();
1.1748 + int stk_slot = box->stack_slot();
1.1749 + if (stk_slot <= 0)
1.1750 + return false; // External lock or it is not Box (Phi node).
1.1751 +
1.1752 + // Ignore complex cases: merged locks or multiple locks.
1.1753 + Node* obj = obj_node();
1.1754 + LockNode* unique_lock = NULL;
1.1755 + if (!box->is_simple_lock_region(&unique_lock, obj) ||
1.1756 + (unique_lock != this)) {
1.1757 + return false;
1.1758 + }
1.1759 +
1.1760 + // Look for external lock for the same object.
1.1761 + SafePointNode* sfn = this->as_SafePoint();
1.1762 + JVMState* youngest_jvms = sfn->jvms();
1.1763 + int max_depth = youngest_jvms->depth();
1.1764 + for (int depth = 1; depth <= max_depth; depth++) {
1.1765 + JVMState* jvms = youngest_jvms->of_depth(depth);
1.1766 + int num_mon = jvms->nof_monitors();
1.1767 + // Loop over monitors
1.1768 + for (int idx = 0; idx < num_mon; idx++) {
1.1769 + Node* obj_node = sfn->monitor_obj(jvms, idx);
1.1770 + BoxLockNode* box_node = sfn->monitor_box(jvms, idx)->as_BoxLock();
1.1771 + if ((box_node->stack_slot() < stk_slot) && obj_node->eqv_uncast(obj)) {
1.1772 + return true;
1.1773 + }
1.1774 + }
1.1775 + }
1.1776 + return false;
1.1777 +}
1.1778 +
1.1779 +//=============================================================================
1.1780 +uint UnlockNode::size_of() const { return sizeof(*this); }
1.1781 +
1.1782 +//=============================================================================
1.1783 +Node *UnlockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1.1784 +
1.1785 + // perform any generic optimizations first (returns 'this' or NULL)
1.1786 + Node *result = SafePointNode::Ideal(phase, can_reshape);
1.1787 + if (result != NULL) return result;
1.1788 + // Don't bother trying to transform a dead node
1.1789 + if (in(0) && in(0)->is_top()) return NULL;
1.1790 +
1.1791 + // Now see if we can optimize away this unlock. We don't actually
1.1792 + // remove the unlocking here, we simply set the _eliminate flag which
1.1793 + // prevents macro expansion from expanding the unlock. Since we don't
1.1794 + // modify the graph, the value returned from this function is the
1.1795 + // one computed above.
1.1796 + // Escape state is defined after Parse phase.
1.1797 + if (can_reshape && EliminateLocks && !is_non_esc_obj()) {
1.1798 + //
1.1799 + // If we are unlocking an unescaped object, the lock/unlock is unnecessary.
1.1800 + //
1.1801 + ConnectionGraph *cgr = phase->C->congraph();
1.1802 + if (cgr != NULL && cgr->not_global_escape(obj_node())) {
1.1803 + assert(!is_eliminated() || is_coarsened(), "sanity");
1.1804 + // The lock could be marked eliminated by lock coarsening
1.1805 + // code during first IGVN before EA. Replace coarsened flag
1.1806 + // to eliminate all associated locks/unlocks.
1.1807 + this->set_non_esc_obj();
1.1808 + }
1.1809 + }
1.1810 + return result;
1.1811 +}
