1 /* 2 * Copyright (c) 1997, 2020, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "ci/bcEscapeAnalyzer.hpp" 28 #include "compiler/oopMap.hpp" 29 #include "gc/shared/barrierSet.hpp" 30 #include "gc/shared/c2/barrierSetC2.hpp" 31 #include "interpreter/interpreter.hpp" 32 #include "opto/callGenerator.hpp" 33 #include "opto/callnode.hpp" 34 #include "opto/castnode.hpp" 35 #include "opto/convertnode.hpp" 36 #include "opto/escape.hpp" 37 #include "opto/locknode.hpp" 38 #include "opto/machnode.hpp" 39 #include "opto/matcher.hpp" 40 #include "opto/parse.hpp" 41 #include "opto/regalloc.hpp" 42 #include "opto/regmask.hpp" 43 #include "opto/rootnode.hpp" 44 #include "opto/runtime.hpp" 45 #include "utilities/powerOfTwo.hpp" 46 47 // Portions of code courtesy of Clifford Click 48 49 // Optimization - Graph Style 50 51 //============================================================================= 52 uint StartNode::size_of() const { return sizeof(*this); } 53 bool StartNode::cmp( const Node &n ) const 54 { return _domain == ((StartNode&)n)._domain; } 55 const Type *StartNode::bottom_type() const { return _domain; } 56 const Type* StartNode::Value(PhaseGVN* phase) const { return _domain; } 57 #ifndef PRODUCT 58 void StartNode::dump_spec(outputStream *st) const { st->print(" #"); _domain->dump_on(st);} 59 void StartNode::dump_compact_spec(outputStream *st) const { /* empty */ } 60 #endif 61 62 //------------------------------Ideal------------------------------------------ 63 Node *StartNode::Ideal(PhaseGVN *phase, bool can_reshape){ 64 return remove_dead_region(phase, can_reshape) ? this : NULL; 65 } 66 67 //------------------------------calling_convention----------------------------- 68 void StartNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const { 69 Matcher::calling_convention( sig_bt, parm_regs, argcnt, false ); 70 } 71 72 //------------------------------Registers-------------------------------------- 73 const RegMask &StartNode::in_RegMask(uint) const { 74 return RegMask::Empty; 75 } 76 77 //------------------------------match------------------------------------------ 78 // Construct projections for incoming parameters, and their RegMask info 79 Node *StartNode::match( const ProjNode *proj, const Matcher *match ) { 80 switch (proj->_con) { 81 case TypeFunc::Control: 82 case TypeFunc::I_O: 83 case TypeFunc::Memory: 84 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 85 case TypeFunc::FramePtr: 86 return new MachProjNode(this,proj->_con,Matcher::c_frame_ptr_mask, Op_RegP); 87 case TypeFunc::ReturnAdr: 88 return new MachProjNode(this,proj->_con,match->_return_addr_mask,Op_RegP); 89 case TypeFunc::Parms: 90 default: { 91 uint parm_num = proj->_con - TypeFunc::Parms; 92 const Type *t = _domain->field_at(proj->_con); 93 if (t->base() == Type::Half) // 2nd half of Longs and Doubles 94 return new ConNode(Type::TOP); 95 uint ideal_reg = t->ideal_reg(); 96 RegMask &rm = match->_calling_convention_mask[parm_num]; 97 return new MachProjNode(this,proj->_con,rm,ideal_reg); 98 } 99 } 100 return NULL; 101 } 102 103 //------------------------------StartOSRNode---------------------------------- 104 // The method start node for an on stack replacement adapter 105 106 //------------------------------osr_domain----------------------------- 107 const TypeTuple *StartOSRNode::osr_domain() { 108 const Type **fields = TypeTuple::fields(2); 109 fields[TypeFunc::Parms+0] = TypeRawPtr::BOTTOM; // address of osr buffer 110 111 return TypeTuple::make(TypeFunc::Parms+1, fields); 112 } 113 114 //============================================================================= 115 const char * const ParmNode::names[TypeFunc::Parms+1] = { 116 "Control", "I_O", "Memory", "FramePtr", "ReturnAdr", "Parms" 117 }; 118 119 #ifndef PRODUCT 120 void ParmNode::dump_spec(outputStream *st) const { 121 if( _con < TypeFunc::Parms ) { 122 st->print("%s", names[_con]); 123 } else { 124 st->print("Parm%d: ",_con-TypeFunc::Parms); 125 // Verbose and WizardMode dump bottom_type for all nodes 126 if( !Verbose && !WizardMode ) bottom_type()->dump_on(st); 127 } 128 } 129 130 void ParmNode::dump_compact_spec(outputStream *st) const { 131 if (_con < TypeFunc::Parms) { 132 st->print("%s", names[_con]); 133 } else { 134 st->print("%d:", _con-TypeFunc::Parms); 135 // unconditionally dump bottom_type 136 bottom_type()->dump_on(st); 137 } 138 } 139 140 // For a ParmNode, all immediate inputs and outputs are considered relevant 141 // both in compact and standard representation. 142 void ParmNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const { 143 this->collect_nodes(in_rel, 1, false, false); 144 this->collect_nodes(out_rel, -1, false, false); 145 } 146 #endif 147 148 uint ParmNode::ideal_reg() const { 149 switch( _con ) { 150 case TypeFunc::Control : // fall through 151 case TypeFunc::I_O : // fall through 152 case TypeFunc::Memory : return 0; 153 case TypeFunc::FramePtr : // fall through 154 case TypeFunc::ReturnAdr: return Op_RegP; 155 default : assert( _con > TypeFunc::Parms, "" ); 156 // fall through 157 case TypeFunc::Parms : { 158 // Type of argument being passed 159 const Type *t = in(0)->as_Start()->_domain->field_at(_con); 160 return t->ideal_reg(); 161 } 162 } 163 ShouldNotReachHere(); 164 return 0; 165 } 166 167 //============================================================================= 168 ReturnNode::ReturnNode(uint edges, Node *cntrl, Node *i_o, Node *memory, Node *frameptr, Node *retadr ) : Node(edges) { 169 init_req(TypeFunc::Control,cntrl); 170 init_req(TypeFunc::I_O,i_o); 171 init_req(TypeFunc::Memory,memory); 172 init_req(TypeFunc::FramePtr,frameptr); 173 init_req(TypeFunc::ReturnAdr,retadr); 174 } 175 176 Node *ReturnNode::Ideal(PhaseGVN *phase, bool can_reshape){ 177 return remove_dead_region(phase, can_reshape) ? this : NULL; 178 } 179 180 const Type* ReturnNode::Value(PhaseGVN* phase) const { 181 return ( phase->type(in(TypeFunc::Control)) == Type::TOP) 182 ? Type::TOP 183 : Type::BOTTOM; 184 } 185 186 // Do we Match on this edge index or not? No edges on return nodes 187 uint ReturnNode::match_edge(uint idx) const { 188 return 0; 189 } 190 191 192 #ifndef PRODUCT 193 void ReturnNode::dump_req(outputStream *st) const { 194 // Dump the required inputs, enclosed in '(' and ')' 195 uint i; // Exit value of loop 196 for (i = 0; i < req(); i++) { // For all required inputs 197 if (i == TypeFunc::Parms) st->print("returns"); 198 if (in(i)) st->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx); 199 else st->print("_ "); 200 } 201 } 202 #endif 203 204 //============================================================================= 205 RethrowNode::RethrowNode( 206 Node* cntrl, 207 Node* i_o, 208 Node* memory, 209 Node* frameptr, 210 Node* ret_adr, 211 Node* exception 212 ) : Node(TypeFunc::Parms + 1) { 213 init_req(TypeFunc::Control , cntrl ); 214 init_req(TypeFunc::I_O , i_o ); 215 init_req(TypeFunc::Memory , memory ); 216 init_req(TypeFunc::FramePtr , frameptr ); 217 init_req(TypeFunc::ReturnAdr, ret_adr); 218 init_req(TypeFunc::Parms , exception); 219 } 220 221 Node *RethrowNode::Ideal(PhaseGVN *phase, bool can_reshape){ 222 return remove_dead_region(phase, can_reshape) ? this : NULL; 223 } 224 225 const Type* RethrowNode::Value(PhaseGVN* phase) const { 226 return (phase->type(in(TypeFunc::Control)) == Type::TOP) 227 ? Type::TOP 228 : Type::BOTTOM; 229 } 230 231 uint RethrowNode::match_edge(uint idx) const { 232 return 0; 233 } 234 235 #ifndef PRODUCT 236 void RethrowNode::dump_req(outputStream *st) const { 237 // Dump the required inputs, enclosed in '(' and ')' 238 uint i; // Exit value of loop 239 for (i = 0; i < req(); i++) { // For all required inputs 240 if (i == TypeFunc::Parms) st->print("exception"); 241 if (in(i)) st->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx); 242 else st->print("_ "); 243 } 244 } 245 #endif 246 247 //============================================================================= 248 // Do we Match on this edge index or not? Match only target address & method 249 uint TailCallNode::match_edge(uint idx) const { 250 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1; 251 } 252 253 //============================================================================= 254 // Do we Match on this edge index or not? Match only target address & oop 255 uint TailJumpNode::match_edge(uint idx) const { 256 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1; 257 } 258 259 //============================================================================= 260 JVMState::JVMState(ciMethod* method, JVMState* caller) : 261 _method(method) { 262 assert(method != NULL, "must be valid call site"); 263 _bci = InvocationEntryBci; 264 _reexecute = Reexecute_Undefined; 265 debug_only(_bci = -99); // random garbage value 266 debug_only(_map = (SafePointNode*)-1); 267 _caller = caller; 268 _depth = 1 + (caller == NULL ? 0 : caller->depth()); 269 _locoff = TypeFunc::Parms; 270 _stkoff = _locoff + _method->max_locals(); 271 _monoff = _stkoff + _method->max_stack(); 272 _scloff = _monoff; 273 _endoff = _monoff; 274 _sp = 0; 275 } 276 JVMState::JVMState(int stack_size) : 277 _method(NULL) { 278 _bci = InvocationEntryBci; 279 _reexecute = Reexecute_Undefined; 280 debug_only(_map = (SafePointNode*)-1); 281 _caller = NULL; 282 _depth = 1; 283 _locoff = TypeFunc::Parms; 284 _stkoff = _locoff; 285 _monoff = _stkoff + stack_size; 286 _scloff = _monoff; 287 _endoff = _monoff; 288 _sp = 0; 289 } 290 291 //--------------------------------of_depth------------------------------------- 292 JVMState* JVMState::of_depth(int d) const { 293 const JVMState* jvmp = this; 294 assert(0 < d && (uint)d <= depth(), "oob"); 295 for (int skip = depth() - d; skip > 0; skip--) { 296 jvmp = jvmp->caller(); 297 } 298 assert(jvmp->depth() == (uint)d, "found the right one"); 299 return (JVMState*)jvmp; 300 } 301 302 //-----------------------------same_calls_as----------------------------------- 303 bool JVMState::same_calls_as(const JVMState* that) const { 304 if (this == that) return true; 305 if (this->depth() != that->depth()) return false; 306 const JVMState* p = this; 307 const JVMState* q = that; 308 for (;;) { 309 if (p->_method != q->_method) return false; 310 if (p->_method == NULL) return true; // bci is irrelevant 311 if (p->_bci != q->_bci) return false; 312 if (p->_reexecute != q->_reexecute) return false; 313 p = p->caller(); 314 q = q->caller(); 315 if (p == q) return true; 316 assert(p != NULL && q != NULL, "depth check ensures we don't run off end"); 317 } 318 } 319 320 //------------------------------debug_start------------------------------------ 321 uint JVMState::debug_start() const { 322 debug_only(JVMState* jvmroot = of_depth(1)); 323 assert(jvmroot->locoff() <= this->locoff(), "youngest JVMState must be last"); 324 return of_depth(1)->locoff(); 325 } 326 327 //-------------------------------debug_end------------------------------------- 328 uint JVMState::debug_end() const { 329 debug_only(JVMState* jvmroot = of_depth(1)); 330 assert(jvmroot->endoff() <= this->endoff(), "youngest JVMState must be last"); 331 return endoff(); 332 } 333 334 //------------------------------debug_depth------------------------------------ 335 uint JVMState::debug_depth() const { 336 uint total = 0; 337 for (const JVMState* jvmp = this; jvmp != NULL; jvmp = jvmp->caller()) { 338 total += jvmp->debug_size(); 339 } 340 return total; 341 } 342 343 #ifndef PRODUCT 344 345 //------------------------------format_helper---------------------------------- 346 // Given an allocation (a Chaitin object) and a Node decide if the Node carries 347 // any defined value or not. If it does, print out the register or constant. 348 static void format_helper( PhaseRegAlloc *regalloc, outputStream* st, Node *n, const char *msg, uint i, GrowableArray<SafePointScalarObjectNode*> *scobjs ) { 349 if (n == NULL) { st->print(" NULL"); return; } 350 if (n->is_SafePointScalarObject()) { 351 // Scalar replacement. 352 SafePointScalarObjectNode* spobj = n->as_SafePointScalarObject(); 353 scobjs->append_if_missing(spobj); 354 int sco_n = scobjs->find(spobj); 355 assert(sco_n >= 0, ""); 356 st->print(" %s%d]=#ScObj" INT32_FORMAT, msg, i, sco_n); 357 return; 358 } 359 if (regalloc->node_regs_max_index() > 0 && 360 OptoReg::is_valid(regalloc->get_reg_first(n))) { // Check for undefined 361 char buf[50]; 362 regalloc->dump_register(n,buf); 363 st->print(" %s%d]=%s",msg,i,buf); 364 } else { // No register, but might be constant 365 const Type *t = n->bottom_type(); 366 switch (t->base()) { 367 case Type::Int: 368 st->print(" %s%d]=#" INT32_FORMAT,msg,i,t->is_int()->get_con()); 369 break; 370 case Type::AnyPtr: 371 assert( t == TypePtr::NULL_PTR || n->in_dump(), "" ); 372 st->print(" %s%d]=#NULL",msg,i); 373 break; 374 case Type::AryPtr: 375 case Type::InstPtr: 376 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->isa_oopptr()->const_oop())); 377 break; 378 case Type::KlassPtr: 379 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_klassptr()->klass())); 380 break; 381 case Type::MetadataPtr: 382 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_metadataptr()->metadata())); 383 break; 384 case Type::NarrowOop: 385 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_oopptr()->const_oop())); 386 break; 387 case Type::RawPtr: 388 st->print(" %s%d]=#Raw" INTPTR_FORMAT,msg,i,p2i(t->is_rawptr())); 389 break; 390 case Type::DoubleCon: 391 st->print(" %s%d]=#%fD",msg,i,t->is_double_constant()->_d); 392 break; 393 case Type::FloatCon: 394 st->print(" %s%d]=#%fF",msg,i,t->is_float_constant()->_f); 395 break; 396 case Type::Long: 397 st->print(" %s%d]=#" INT64_FORMAT,msg,i,(int64_t)(t->is_long()->get_con())); 398 break; 399 case Type::Half: 400 case Type::Top: 401 st->print(" %s%d]=_",msg,i); 402 break; 403 default: ShouldNotReachHere(); 404 } 405 } 406 } 407 408 //------------------------------format----------------------------------------- 409 void JVMState::format(PhaseRegAlloc *regalloc, const Node *n, outputStream* st) const { 410 st->print(" #"); 411 if (_method) { 412 _method->print_short_name(st); 413 st->print(" @ bci:%d ",_bci); 414 } else { 415 st->print_cr(" runtime stub "); 416 return; 417 } 418 if (n->is_MachSafePoint()) { 419 GrowableArray<SafePointScalarObjectNode*> scobjs; 420 MachSafePointNode *mcall = n->as_MachSafePoint(); 421 uint i; 422 // Print locals 423 for (i = 0; i < (uint)loc_size(); i++) 424 format_helper(regalloc, st, mcall->local(this, i), "L[", i, &scobjs); 425 // Print stack 426 for (i = 0; i < (uint)stk_size(); i++) { 427 if ((uint)(_stkoff + i) >= mcall->len()) 428 st->print(" oob "); 429 else 430 format_helper(regalloc, st, mcall->stack(this, i), "STK[", i, &scobjs); 431 } 432 for (i = 0; (int)i < nof_monitors(); i++) { 433 Node *box = mcall->monitor_box(this, i); 434 Node *obj = mcall->monitor_obj(this, i); 435 if (regalloc->node_regs_max_index() > 0 && 436 OptoReg::is_valid(regalloc->get_reg_first(box))) { 437 box = BoxLockNode::box_node(box); 438 format_helper(regalloc, st, box, "MON-BOX[", i, &scobjs); 439 } else { 440 OptoReg::Name box_reg = BoxLockNode::reg(box); 441 st->print(" MON-BOX%d=%s+%d", 442 i, 443 OptoReg::regname(OptoReg::c_frame_pointer), 444 regalloc->reg2offset(box_reg)); 445 } 446 const char* obj_msg = "MON-OBJ["; 447 if (EliminateLocks) { 448 if (BoxLockNode::box_node(box)->is_eliminated()) 449 obj_msg = "MON-OBJ(LOCK ELIMINATED)["; 450 } 451 format_helper(regalloc, st, obj, obj_msg, i, &scobjs); 452 } 453 454 for (i = 0; i < (uint)scobjs.length(); i++) { 455 // Scalar replaced objects. 456 st->cr(); 457 st->print(" # ScObj" INT32_FORMAT " ", i); 458 SafePointScalarObjectNode* spobj = scobjs.at(i); 459 ciKlass* cik = spobj->bottom_type()->is_oopptr()->klass(); 460 assert(cik->is_instance_klass() || 461 cik->is_array_klass(), "Not supported allocation."); 462 ciInstanceKlass *iklass = NULL; 463 if (cik->is_instance_klass()) { 464 cik->print_name_on(st); 465 iklass = cik->as_instance_klass(); 466 } else if (cik->is_type_array_klass()) { 467 cik->as_array_klass()->base_element_type()->print_name_on(st); 468 st->print("[%d]", spobj->n_fields()); 469 } else if (cik->is_obj_array_klass()) { 470 ciKlass* cie = cik->as_obj_array_klass()->base_element_klass(); 471 if (cie->is_instance_klass()) { 472 cie->print_name_on(st); 473 } else if (cie->is_type_array_klass()) { 474 cie->as_array_klass()->base_element_type()->print_name_on(st); 475 } else { 476 ShouldNotReachHere(); 477 } 478 st->print("[%d]", spobj->n_fields()); 479 int ndim = cik->as_array_klass()->dimension() - 1; 480 while (ndim-- > 0) { 481 st->print("[]"); 482 } 483 } 484 st->print("={"); 485 uint nf = spobj->n_fields(); 486 if (nf > 0) { 487 uint first_ind = spobj->first_index(mcall->jvms()); 488 Node* fld_node = mcall->in(first_ind); 489 ciField* cifield; 490 if (iklass != NULL) { 491 st->print(" ["); 492 cifield = iklass->nonstatic_field_at(0); 493 cifield->print_name_on(st); 494 format_helper(regalloc, st, fld_node, ":", 0, &scobjs); 495 } else { 496 format_helper(regalloc, st, fld_node, "[", 0, &scobjs); 497 } 498 for (uint j = 1; j < nf; j++) { 499 fld_node = mcall->in(first_ind+j); 500 if (iklass != NULL) { 501 st->print(", ["); 502 cifield = iklass->nonstatic_field_at(j); 503 cifield->print_name_on(st); 504 format_helper(regalloc, st, fld_node, ":", j, &scobjs); 505 } else { 506 format_helper(regalloc, st, fld_node, ", [", j, &scobjs); 507 } 508 } 509 } 510 st->print(" }"); 511 } 512 } 513 st->cr(); 514 if (caller() != NULL) caller()->format(regalloc, n, st); 515 } 516 517 518 void JVMState::dump_spec(outputStream *st) const { 519 if (_method != NULL) { 520 bool printed = false; 521 if (!Verbose) { 522 // The JVMS dumps make really, really long lines. 523 // Take out the most boring parts, which are the package prefixes. 524 char buf[500]; 525 stringStream namest(buf, sizeof(buf)); 526 _method->print_short_name(&namest); 527 if (namest.count() < sizeof(buf)) { 528 const char* name = namest.base(); 529 if (name[0] == ' ') ++name; 530 const char* endcn = strchr(name, ':'); // end of class name 531 if (endcn == NULL) endcn = strchr(name, '('); 532 if (endcn == NULL) endcn = name + strlen(name); 533 while (endcn > name && endcn[-1] != '.' && endcn[-1] != '/') 534 --endcn; 535 st->print(" %s", endcn); 536 printed = true; 537 } 538 } 539 if (!printed) 540 _method->print_short_name(st); 541 st->print(" @ bci:%d",_bci); 542 if(_reexecute == Reexecute_True) 543 st->print(" reexecute"); 544 } else { 545 st->print(" runtime stub"); 546 } 547 if (caller() != NULL) caller()->dump_spec(st); 548 } 549 550 551 void JVMState::dump_on(outputStream* st) const { 552 bool print_map = _map && !((uintptr_t)_map & 1) && 553 ((caller() == NULL) || (caller()->map() != _map)); 554 if (print_map) { 555 if (_map->len() > _map->req()) { // _map->has_exceptions() 556 Node* ex = _map->in(_map->req()); // _map->next_exception() 557 // skip the first one; it's already being printed 558 while (ex != NULL && ex->len() > ex->req()) { 559 ex = ex->in(ex->req()); // ex->next_exception() 560 ex->dump(1); 561 } 562 } 563 _map->dump(Verbose ? 2 : 1); 564 } 565 if (caller() != NULL) { 566 caller()->dump_on(st); 567 } 568 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=", 569 depth(), locoff(), stkoff(), argoff(), monoff(), scloff(), endoff(), monitor_depth(), sp(), bci(), should_reexecute()?"true":"false"); 570 if (_method == NULL) { 571 st->print_cr("(none)"); 572 } else { 573 _method->print_name(st); 574 st->cr(); 575 if (bci() >= 0 && bci() < _method->code_size()) { 576 st->print(" bc: "); 577 _method->print_codes_on(bci(), bci()+1, st); 578 } 579 } 580 } 581 582 // Extra way to dump a jvms from the debugger, 583 // to avoid a bug with C++ member function calls. 584 void dump_jvms(JVMState* jvms) { 585 jvms->dump(); 586 } 587 #endif 588 589 //--------------------------clone_shallow-------------------------------------- 590 JVMState* JVMState::clone_shallow(Compile* C) const { 591 JVMState* n = has_method() ? new (C) JVMState(_method, _caller) : new (C) JVMState(0); 592 n->set_bci(_bci); 593 n->_reexecute = _reexecute; 594 n->set_locoff(_locoff); 595 n->set_stkoff(_stkoff); 596 n->set_monoff(_monoff); 597 n->set_scloff(_scloff); 598 n->set_endoff(_endoff); 599 n->set_sp(_sp); 600 n->set_map(_map); 601 return n; 602 } 603 604 //---------------------------clone_deep---------------------------------------- 605 JVMState* JVMState::clone_deep(Compile* C) const { 606 JVMState* n = clone_shallow(C); 607 for (JVMState* p = n; p->_caller != NULL; p = p->_caller) { 608 p->_caller = p->_caller->clone_shallow(C); 609 } 610 assert(n->depth() == depth(), "sanity"); 611 assert(n->debug_depth() == debug_depth(), "sanity"); 612 return n; 613 } 614 615 /** 616 * Reset map for all callers 617 */ 618 void JVMState::set_map_deep(SafePointNode* map) { 619 for (JVMState* p = this; p->_caller != NULL; p = p->_caller) { 620 p->set_map(map); 621 } 622 } 623 624 // Adapt offsets in in-array after adding or removing an edge. 625 // Prerequisite is that the JVMState is used by only one node. 626 void JVMState::adapt_position(int delta) { 627 for (JVMState* jvms = this; jvms != NULL; jvms = jvms->caller()) { 628 jvms->set_locoff(jvms->locoff() + delta); 629 jvms->set_stkoff(jvms->stkoff() + delta); 630 jvms->set_monoff(jvms->monoff() + delta); 631 jvms->set_scloff(jvms->scloff() + delta); 632 jvms->set_endoff(jvms->endoff() + delta); 633 } 634 } 635 636 // Mirror the stack size calculation in the deopt code 637 // How much stack space would we need at this point in the program in 638 // case of deoptimization? 639 int JVMState::interpreter_frame_size() const { 640 const JVMState* jvms = this; 641 int size = 0; 642 int callee_parameters = 0; 643 int callee_locals = 0; 644 int extra_args = method()->max_stack() - stk_size(); 645 646 while (jvms != NULL) { 647 int locks = jvms->nof_monitors(); 648 int temps = jvms->stk_size(); 649 bool is_top_frame = (jvms == this); 650 ciMethod* method = jvms->method(); 651 652 int frame_size = BytesPerWord * Interpreter::size_activation(method->max_stack(), 653 temps + callee_parameters, 654 extra_args, 655 locks, 656 callee_parameters, 657 callee_locals, 658 is_top_frame); 659 size += frame_size; 660 661 callee_parameters = method->size_of_parameters(); 662 callee_locals = method->max_locals(); 663 extra_args = 0; 664 jvms = jvms->caller(); 665 } 666 return size + Deoptimization::last_frame_adjust(0, callee_locals) * BytesPerWord; 667 } 668 669 //============================================================================= 670 bool CallNode::cmp( const Node &n ) const 671 { return _tf == ((CallNode&)n)._tf && _jvms == ((CallNode&)n)._jvms; } 672 #ifndef PRODUCT 673 void CallNode::dump_req(outputStream *st) const { 674 // Dump the required inputs, enclosed in '(' and ')' 675 uint i; // Exit value of loop 676 for (i = 0; i < req(); i++) { // For all required inputs 677 if (i == TypeFunc::Parms) st->print("("); 678 if (in(i)) st->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx); 679 else st->print("_ "); 680 } 681 st->print(")"); 682 } 683 684 void CallNode::dump_spec(outputStream *st) const { 685 st->print(" "); 686 if (tf() != NULL) tf()->dump_on(st); 687 if (_cnt != COUNT_UNKNOWN) st->print(" C=%f",_cnt); 688 if (jvms() != NULL) jvms()->dump_spec(st); 689 } 690 #endif 691 692 const Type *CallNode::bottom_type() const { return tf()->range(); } 693 const Type* CallNode::Value(PhaseGVN* phase) const { 694 if (phase->type(in(0)) == Type::TOP) return Type::TOP; 695 return tf()->range(); 696 } 697 698 //------------------------------calling_convention----------------------------- 699 void CallNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const { 700 // Use the standard compiler calling convention 701 Matcher::calling_convention( sig_bt, parm_regs, argcnt, true ); 702 } 703 704 705 //------------------------------match------------------------------------------ 706 // Construct projections for control, I/O, memory-fields, ..., and 707 // return result(s) along with their RegMask info 708 Node *CallNode::match( const ProjNode *proj, const Matcher *match ) { 709 switch (proj->_con) { 710 case TypeFunc::Control: 711 case TypeFunc::I_O: 712 case TypeFunc::Memory: 713 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 714 715 case TypeFunc::Parms+1: // For LONG & DOUBLE returns 716 assert(tf()->range()->field_at(TypeFunc::Parms+1) == Type::HALF, ""); 717 // 2nd half of doubles and longs 718 return new MachProjNode(this,proj->_con, RegMask::Empty, (uint)OptoReg::Bad); 719 720 case TypeFunc::Parms: { // Normal returns 721 uint ideal_reg = tf()->range()->field_at(TypeFunc::Parms)->ideal_reg(); 722 OptoRegPair regs = is_CallRuntime() 723 ? match->c_return_value(ideal_reg,true) // Calls into C runtime 724 : match-> return_value(ideal_reg,true); // Calls into compiled Java code 725 RegMask rm = RegMask(regs.first()); 726 if( OptoReg::is_valid(regs.second()) ) 727 rm.Insert( regs.second() ); 728 return new MachProjNode(this,proj->_con,rm,ideal_reg); 729 } 730 731 case TypeFunc::ReturnAdr: 732 case TypeFunc::FramePtr: 733 default: 734 ShouldNotReachHere(); 735 } 736 return NULL; 737 } 738 739 // Do we Match on this edge index or not? Match no edges 740 uint CallNode::match_edge(uint idx) const { 741 return 0; 742 } 743 744 // 745 // Determine whether the call could modify the field of the specified 746 // instance at the specified offset. 747 // 748 bool CallNode::may_modify(const TypeOopPtr *t_oop, PhaseTransform *phase) { 749 assert((t_oop != NULL), "sanity"); 750 if (is_call_to_arraycopystub() && strcmp(_name, "unsafe_arraycopy") != 0) { 751 const TypeTuple* args = _tf->domain(); 752 Node* dest = NULL; 753 // Stubs that can be called once an ArrayCopyNode is expanded have 754 // different signatures. Look for the second pointer argument, 755 // that is the destination of the copy. 756 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) { 757 if (args->field_at(i)->isa_ptr()) { 758 j++; 759 if (j == 2) { 760 dest = in(i); 761 break; 762 } 763 } 764 } 765 guarantee(dest != NULL, "Call had only one ptr in, broken IR!"); 766 if (!dest->is_top() && may_modify_arraycopy_helper(phase->type(dest)->is_oopptr(), t_oop, phase)) { 767 return true; 768 } 769 return false; 770 } 771 if (t_oop->is_known_instance()) { 772 // The instance_id is set only for scalar-replaceable allocations which 773 // are not passed as arguments according to Escape Analysis. 774 return false; 775 } 776 if (t_oop->is_ptr_to_boxed_value()) { 777 ciKlass* boxing_klass = t_oop->klass(); 778 if (is_CallStaticJava() && as_CallStaticJava()->is_boxing_method()) { 779 // Skip unrelated boxing methods. 780 Node* proj = proj_out_or_null(TypeFunc::Parms); 781 if ((proj == NULL) || (phase->type(proj)->is_instptr()->klass() != boxing_klass)) { 782 return false; 783 } 784 } 785 if (is_CallJava() && as_CallJava()->method() != NULL) { 786 ciMethod* meth = as_CallJava()->method(); 787 if (meth->is_getter()) { 788 return false; 789 } 790 // May modify (by reflection) if an boxing object is passed 791 // as argument or returned. 792 Node* proj = returns_pointer() ? proj_out_or_null(TypeFunc::Parms) : NULL; 793 if (proj != NULL) { 794 const TypeInstPtr* inst_t = phase->type(proj)->isa_instptr(); 795 if ((inst_t != NULL) && (!inst_t->klass_is_exact() || 796 (inst_t->klass() == boxing_klass))) { 797 return true; 798 } 799 } 800 const TypeTuple* d = tf()->domain(); 801 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 802 const TypeInstPtr* inst_t = d->field_at(i)->isa_instptr(); 803 if ((inst_t != NULL) && (!inst_t->klass_is_exact() || 804 (inst_t->klass() == boxing_klass))) { 805 return true; 806 } 807 } 808 return false; 809 } 810 } 811 return true; 812 } 813 814 // Does this call have a direct reference to n other than debug information? 815 bool CallNode::has_non_debug_use(Node *n) { 816 const TypeTuple * d = tf()->domain(); 817 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 818 Node *arg = in(i); 819 if (arg == n) { 820 return true; 821 } 822 } 823 return false; 824 } 825 826 // Returns the unique CheckCastPP of a call 827 // or 'this' if there are several CheckCastPP or unexpected uses 828 // or returns NULL if there is no one. 829 Node *CallNode::result_cast() { 830 Node *cast = NULL; 831 832 Node *p = proj_out_or_null(TypeFunc::Parms); 833 if (p == NULL) 834 return NULL; 835 836 for (DUIterator_Fast imax, i = p->fast_outs(imax); i < imax; i++) { 837 Node *use = p->fast_out(i); 838 if (use->is_CheckCastPP()) { 839 if (cast != NULL) { 840 return this; // more than 1 CheckCastPP 841 } 842 cast = use; 843 } else if (!use->is_Initialize() && 844 !use->is_AddP() && 845 use->Opcode() != Op_MemBarStoreStore) { 846 // Expected uses are restricted to a CheckCastPP, an Initialize 847 // node, a MemBarStoreStore (clone) and AddP nodes. If we 848 // encounter any other use (a Phi node can be seen in rare 849 // cases) return this to prevent incorrect optimizations. 850 return this; 851 } 852 } 853 return cast; 854 } 855 856 857 void CallNode::extract_projections(CallProjections* projs, bool separate_io_proj, bool do_asserts) { 858 projs->fallthrough_proj = NULL; 859 projs->fallthrough_catchproj = NULL; 860 projs->fallthrough_ioproj = NULL; 861 projs->catchall_ioproj = NULL; 862 projs->catchall_catchproj = NULL; 863 projs->fallthrough_memproj = NULL; 864 projs->catchall_memproj = NULL; 865 projs->resproj = NULL; 866 projs->exobj = NULL; 867 868 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 869 ProjNode *pn = fast_out(i)->as_Proj(); 870 if (pn->outcnt() == 0) continue; 871 switch (pn->_con) { 872 case TypeFunc::Control: 873 { 874 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj 875 projs->fallthrough_proj = pn; 876 DUIterator_Fast jmax, j = pn->fast_outs(jmax); 877 const Node *cn = pn->fast_out(j); 878 if (cn->is_Catch()) { 879 ProjNode *cpn = NULL; 880 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) { 881 cpn = cn->fast_out(k)->as_Proj(); 882 assert(cpn->is_CatchProj(), "must be a CatchProjNode"); 883 if (cpn->_con == CatchProjNode::fall_through_index) 884 projs->fallthrough_catchproj = cpn; 885 else { 886 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index."); 887 projs->catchall_catchproj = cpn; 888 } 889 } 890 } 891 break; 892 } 893 case TypeFunc::I_O: 894 if (pn->_is_io_use) 895 projs->catchall_ioproj = pn; 896 else 897 projs->fallthrough_ioproj = pn; 898 for (DUIterator j = pn->outs(); pn->has_out(j); j++) { 899 Node* e = pn->out(j); 900 if (e->Opcode() == Op_CreateEx && e->in(0)->is_CatchProj() && e->outcnt() > 0) { 901 assert(projs->exobj == NULL, "only one"); 902 projs->exobj = e; 903 } 904 } 905 break; 906 case TypeFunc::Memory: 907 if (pn->_is_io_use) 908 projs->catchall_memproj = pn; 909 else 910 projs->fallthrough_memproj = pn; 911 break; 912 case TypeFunc::Parms: 913 projs->resproj = pn; 914 break; 915 default: 916 assert(false, "unexpected projection from allocation node."); 917 } 918 } 919 920 // The resproj may not exist because the result could be ignored 921 // and the exception object may not exist if an exception handler 922 // swallows the exception but all the other must exist and be found. 923 assert(projs->fallthrough_proj != NULL, "must be found"); 924 do_asserts = do_asserts && !Compile::current()->inlining_incrementally(); 925 assert(!do_asserts || projs->fallthrough_catchproj != NULL, "must be found"); 926 assert(!do_asserts || projs->fallthrough_memproj != NULL, "must be found"); 927 assert(!do_asserts || projs->fallthrough_ioproj != NULL, "must be found"); 928 assert(!do_asserts || projs->catchall_catchproj != NULL, "must be found"); 929 if (separate_io_proj) { 930 assert(!do_asserts || projs->catchall_memproj != NULL, "must be found"); 931 assert(!do_asserts || projs->catchall_ioproj != NULL, "must be found"); 932 } 933 } 934 935 Node *CallNode::Ideal(PhaseGVN *phase, bool can_reshape) { 936 CallGenerator* cg = generator(); 937 if (can_reshape && cg != NULL && cg->is_mh_late_inline() && !cg->already_attempted()) { 938 // Check whether this MH handle call becomes a candidate for inlining 939 ciMethod* callee = cg->method(); 940 vmIntrinsics::ID iid = callee->intrinsic_id(); 941 if (iid == vmIntrinsics::_invokeBasic) { 942 if (in(TypeFunc::Parms)->Opcode() == Op_ConP) { 943 phase->C->prepend_late_inline(cg); 944 set_generator(NULL); 945 } 946 } else { 947 assert(callee->has_member_arg(), "wrong type of call?"); 948 if (in(TypeFunc::Parms + callee->arg_size() - 1)->Opcode() == Op_ConP) { 949 phase->C->prepend_late_inline(cg); 950 set_generator(NULL); 951 } 952 } 953 } 954 return SafePointNode::Ideal(phase, can_reshape); 955 } 956 957 bool CallNode::is_call_to_arraycopystub() const { 958 if (_name != NULL && strstr(_name, "arraycopy") != 0) { 959 return true; 960 } 961 return false; 962 } 963 964 //============================================================================= 965 uint CallJavaNode::size_of() const { return sizeof(*this); } 966 bool CallJavaNode::cmp( const Node &n ) const { 967 CallJavaNode &call = (CallJavaNode&)n; 968 return CallNode::cmp(call) && _method == call._method && 969 _override_symbolic_info == call._override_symbolic_info; 970 } 971 #ifdef ASSERT 972 bool CallJavaNode::validate_symbolic_info() const { 973 if (method() == NULL) { 974 return true; // call into runtime or uncommon trap 975 } 976 ciMethod* symbolic_info = jvms()->method()->get_method_at_bci(_bci); 977 ciMethod* callee = method(); 978 if (symbolic_info->is_method_handle_intrinsic() && !callee->is_method_handle_intrinsic()) { 979 assert(override_symbolic_info(), "should be set"); 980 } 981 assert(ciMethod::is_consistent_info(symbolic_info, callee), "inconsistent info"); 982 return true; 983 } 984 #endif 985 986 #ifndef PRODUCT 987 void CallJavaNode::dump_spec(outputStream *st) const { 988 if( _method ) _method->print_short_name(st); 989 CallNode::dump_spec(st); 990 } 991 992 void CallJavaNode::dump_compact_spec(outputStream* st) const { 993 if (_method) { 994 _method->print_short_name(st); 995 } else { 996 st->print("<?>"); 997 } 998 } 999 #endif 1000 1001 //============================================================================= 1002 uint CallStaticJavaNode::size_of() const { return sizeof(*this); } 1003 bool CallStaticJavaNode::cmp( const Node &n ) const { 1004 CallStaticJavaNode &call = (CallStaticJavaNode&)n; 1005 return CallJavaNode::cmp(call); 1006 } 1007 1008 //----------------------------uncommon_trap_request---------------------------- 1009 // If this is an uncommon trap, return the request code, else zero. 1010 int CallStaticJavaNode::uncommon_trap_request() const { 1011 if (_name != NULL && !strcmp(_name, "uncommon_trap")) { 1012 return extract_uncommon_trap_request(this); 1013 } 1014 return 0; 1015 } 1016 int CallStaticJavaNode::extract_uncommon_trap_request(const Node* call) { 1017 #ifndef PRODUCT 1018 if (!(call->req() > TypeFunc::Parms && 1019 call->in(TypeFunc::Parms) != NULL && 1020 call->in(TypeFunc::Parms)->is_Con() && 1021 call->in(TypeFunc::Parms)->bottom_type()->isa_int())) { 1022 assert(in_dump() != 0, "OK if dumping"); 1023 tty->print("[bad uncommon trap]"); 1024 return 0; 1025 } 1026 #endif 1027 return call->in(TypeFunc::Parms)->bottom_type()->is_int()->get_con(); 1028 } 1029 1030 #ifndef PRODUCT 1031 void CallStaticJavaNode::dump_spec(outputStream *st) const { 1032 st->print("# Static "); 1033 if (_name != NULL) { 1034 st->print("%s", _name); 1035 int trap_req = uncommon_trap_request(); 1036 if (trap_req != 0) { 1037 char buf[100]; 1038 st->print("(%s)", 1039 Deoptimization::format_trap_request(buf, sizeof(buf), 1040 trap_req)); 1041 } 1042 st->print(" "); 1043 } 1044 CallJavaNode::dump_spec(st); 1045 } 1046 1047 void CallStaticJavaNode::dump_compact_spec(outputStream* st) const { 1048 if (_method) { 1049 _method->print_short_name(st); 1050 } else if (_name) { 1051 st->print("%s", _name); 1052 } else { 1053 st->print("<?>"); 1054 } 1055 } 1056 #endif 1057 1058 //============================================================================= 1059 uint CallDynamicJavaNode::size_of() const { return sizeof(*this); } 1060 bool CallDynamicJavaNode::cmp( const Node &n ) const { 1061 CallDynamicJavaNode &call = (CallDynamicJavaNode&)n; 1062 return CallJavaNode::cmp(call); 1063 } 1064 #ifndef PRODUCT 1065 void CallDynamicJavaNode::dump_spec(outputStream *st) const { 1066 st->print("# Dynamic "); 1067 CallJavaNode::dump_spec(st); 1068 } 1069 #endif 1070 1071 //============================================================================= 1072 uint CallRuntimeNode::size_of() const { return sizeof(*this); } 1073 bool CallRuntimeNode::cmp( const Node &n ) const { 1074 CallRuntimeNode &call = (CallRuntimeNode&)n; 1075 return CallNode::cmp(call) && !strcmp(_name,call._name); 1076 } 1077 #ifndef PRODUCT 1078 void CallRuntimeNode::dump_spec(outputStream *st) const { 1079 st->print("# "); 1080 st->print("%s", _name); 1081 CallNode::dump_spec(st); 1082 } 1083 #endif 1084 1085 //------------------------------calling_convention----------------------------- 1086 void CallRuntimeNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const { 1087 Matcher::c_calling_convention( sig_bt, parm_regs, argcnt ); 1088 } 1089 1090 //============================================================================= 1091 //------------------------------calling_convention----------------------------- 1092 1093 1094 //============================================================================= 1095 #ifndef PRODUCT 1096 void CallLeafNode::dump_spec(outputStream *st) const { 1097 st->print("# "); 1098 st->print("%s", _name); 1099 CallNode::dump_spec(st); 1100 } 1101 #endif 1102 1103 //============================================================================= 1104 1105 void SafePointNode::set_local(JVMState* jvms, uint idx, Node *c) { 1106 assert(verify_jvms(jvms), "jvms must match"); 1107 int loc = jvms->locoff() + idx; 1108 if (in(loc)->is_top() && idx > 0 && !c->is_top() ) { 1109 // If current local idx is top then local idx - 1 could 1110 // be a long/double that needs to be killed since top could 1111 // represent the 2nd half ofthe long/double. 1112 uint ideal = in(loc -1)->ideal_reg(); 1113 if (ideal == Op_RegD || ideal == Op_RegL) { 1114 // set other (low index) half to top 1115 set_req(loc - 1, in(loc)); 1116 } 1117 } 1118 set_req(loc, c); 1119 } 1120 1121 uint SafePointNode::size_of() const { return sizeof(*this); } 1122 bool SafePointNode::cmp( const Node &n ) const { 1123 return (&n == this); // Always fail except on self 1124 } 1125 1126 //-------------------------set_next_exception---------------------------------- 1127 void SafePointNode::set_next_exception(SafePointNode* n) { 1128 assert(n == NULL || n->Opcode() == Op_SafePoint, "correct value for next_exception"); 1129 if (len() == req()) { 1130 if (n != NULL) add_prec(n); 1131 } else { 1132 set_prec(req(), n); 1133 } 1134 } 1135 1136 1137 //----------------------------next_exception----------------------------------- 1138 SafePointNode* SafePointNode::next_exception() const { 1139 if (len() == req()) { 1140 return NULL; 1141 } else { 1142 Node* n = in(req()); 1143 assert(n == NULL || n->Opcode() == Op_SafePoint, "no other uses of prec edges"); 1144 return (SafePointNode*) n; 1145 } 1146 } 1147 1148 1149 //------------------------------Ideal------------------------------------------ 1150 // Skip over any collapsed Regions 1151 Node *SafePointNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1152 return remove_dead_region(phase, can_reshape) ? this : NULL; 1153 } 1154 1155 //------------------------------Identity--------------------------------------- 1156 // Remove obviously duplicate safepoints 1157 Node* SafePointNode::Identity(PhaseGVN* phase) { 1158 1159 // If you have back to back safepoints, remove one 1160 if( in(TypeFunc::Control)->is_SafePoint() ) 1161 return in(TypeFunc::Control); 1162 1163 if( in(0)->is_Proj() ) { 1164 Node *n0 = in(0)->in(0); 1165 // Check if he is a call projection (except Leaf Call) 1166 if( n0->is_Catch() ) { 1167 n0 = n0->in(0)->in(0); 1168 assert( n0->is_Call(), "expect a call here" ); 1169 } 1170 if( n0->is_Call() && n0->as_Call()->guaranteed_safepoint() ) { 1171 // Don't remove a safepoint belonging to an OuterStripMinedLoopEndNode. 1172 // If the loop dies, they will be removed together. 1173 if (has_out_with(Op_OuterStripMinedLoopEnd)) { 1174 return this; 1175 } 1176 // Useless Safepoint, so remove it 1177 return in(TypeFunc::Control); 1178 } 1179 } 1180 1181 return this; 1182 } 1183 1184 //------------------------------Value------------------------------------------ 1185 const Type* SafePointNode::Value(PhaseGVN* phase) const { 1186 if( phase->type(in(0)) == Type::TOP ) return Type::TOP; 1187 if( phase->eqv( in(0), this ) ) return Type::TOP; // Dead infinite loop 1188 return Type::CONTROL; 1189 } 1190 1191 #ifndef PRODUCT 1192 void SafePointNode::dump_spec(outputStream *st) const { 1193 st->print(" SafePoint "); 1194 _replaced_nodes.dump(st); 1195 } 1196 1197 // The related nodes of a SafepointNode are all data inputs, excluding the 1198 // control boundary, as well as all outputs till level 2 (to include projection 1199 // nodes and targets). In compact mode, just include inputs till level 1 and 1200 // outputs as before. 1201 void SafePointNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const { 1202 if (compact) { 1203 this->collect_nodes(in_rel, 1, false, false); 1204 } else { 1205 this->collect_nodes_in_all_data(in_rel, false); 1206 } 1207 this->collect_nodes(out_rel, -2, false, false); 1208 } 1209 #endif 1210 1211 const RegMask &SafePointNode::in_RegMask(uint idx) const { 1212 if( idx < TypeFunc::Parms ) return RegMask::Empty; 1213 // Values outside the domain represent debug info 1214 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]); 1215 } 1216 const RegMask &SafePointNode::out_RegMask() const { 1217 return RegMask::Empty; 1218 } 1219 1220 1221 void SafePointNode::grow_stack(JVMState* jvms, uint grow_by) { 1222 assert((int)grow_by > 0, "sanity"); 1223 int monoff = jvms->monoff(); 1224 int scloff = jvms->scloff(); 1225 int endoff = jvms->endoff(); 1226 assert(endoff == (int)req(), "no other states or debug info after me"); 1227 Node* top = Compile::current()->top(); 1228 for (uint i = 0; i < grow_by; i++) { 1229 ins_req(monoff, top); 1230 } 1231 jvms->set_monoff(monoff + grow_by); 1232 jvms->set_scloff(scloff + grow_by); 1233 jvms->set_endoff(endoff + grow_by); 1234 } 1235 1236 void SafePointNode::push_monitor(const FastLockNode *lock) { 1237 // Add a LockNode, which points to both the original BoxLockNode (the 1238 // stack space for the monitor) and the Object being locked. 1239 const int MonitorEdges = 2; 1240 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges"); 1241 assert(req() == jvms()->endoff(), "correct sizing"); 1242 int nextmon = jvms()->scloff(); 1243 if (GenerateSynchronizationCode) { 1244 ins_req(nextmon, lock->box_node()); 1245 ins_req(nextmon+1, lock->obj_node()); 1246 } else { 1247 Node* top = Compile::current()->top(); 1248 ins_req(nextmon, top); 1249 ins_req(nextmon, top); 1250 } 1251 jvms()->set_scloff(nextmon + MonitorEdges); 1252 jvms()->set_endoff(req()); 1253 } 1254 1255 void SafePointNode::pop_monitor() { 1256 // Delete last monitor from debug info 1257 debug_only(int num_before_pop = jvms()->nof_monitors()); 1258 const int MonitorEdges = 2; 1259 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges"); 1260 int scloff = jvms()->scloff(); 1261 int endoff = jvms()->endoff(); 1262 int new_scloff = scloff - MonitorEdges; 1263 int new_endoff = endoff - MonitorEdges; 1264 jvms()->set_scloff(new_scloff); 1265 jvms()->set_endoff(new_endoff); 1266 while (scloff > new_scloff) del_req_ordered(--scloff); 1267 assert(jvms()->nof_monitors() == num_before_pop-1, ""); 1268 } 1269 1270 Node *SafePointNode::peek_monitor_box() const { 1271 int mon = jvms()->nof_monitors() - 1; 1272 assert(mon >= 0, "must have a monitor"); 1273 return monitor_box(jvms(), mon); 1274 } 1275 1276 Node *SafePointNode::peek_monitor_obj() const { 1277 int mon = jvms()->nof_monitors() - 1; 1278 assert(mon >= 0, "must have a monitor"); 1279 return monitor_obj(jvms(), mon); 1280 } 1281 1282 // Do we Match on this edge index or not? Match no edges 1283 uint SafePointNode::match_edge(uint idx) const { 1284 return (TypeFunc::Parms == idx); 1285 } 1286 1287 void SafePointNode::disconnect_from_root(PhaseIterGVN *igvn) { 1288 assert(Opcode() == Op_SafePoint, "only value for safepoint in loops"); 1289 int nb = igvn->C->root()->find_prec_edge(this); 1290 if (nb != -1) { 1291 igvn->C->root()->rm_prec(nb); 1292 } 1293 } 1294 1295 //============== SafePointScalarObjectNode ============== 1296 1297 SafePointScalarObjectNode::SafePointScalarObjectNode(const TypeOopPtr* tp, 1298 #ifdef ASSERT 1299 AllocateNode* alloc, 1300 #endif 1301 uint first_index, 1302 uint n_fields) : 1303 TypeNode(tp, 1), // 1 control input -- seems required. Get from root. 1304 _first_index(first_index), 1305 _n_fields(n_fields) 1306 #ifdef ASSERT 1307 , _alloc(alloc) 1308 #endif 1309 { 1310 init_class_id(Class_SafePointScalarObject); 1311 } 1312 1313 // Do not allow value-numbering for SafePointScalarObject node. 1314 uint SafePointScalarObjectNode::hash() const { return NO_HASH; } 1315 bool SafePointScalarObjectNode::cmp( const Node &n ) const { 1316 return (&n == this); // Always fail except on self 1317 } 1318 1319 uint SafePointScalarObjectNode::ideal_reg() const { 1320 return 0; // No matching to machine instruction 1321 } 1322 1323 const RegMask &SafePointScalarObjectNode::in_RegMask(uint idx) const { 1324 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]); 1325 } 1326 1327 const RegMask &SafePointScalarObjectNode::out_RegMask() const { 1328 return RegMask::Empty; 1329 } 1330 1331 uint SafePointScalarObjectNode::match_edge(uint idx) const { 1332 return 0; 1333 } 1334 1335 SafePointScalarObjectNode* 1336 SafePointScalarObjectNode::clone(Dict* sosn_map) const { 1337 void* cached = (*sosn_map)[(void*)this]; 1338 if (cached != NULL) { 1339 return (SafePointScalarObjectNode*)cached; 1340 } 1341 SafePointScalarObjectNode* res = (SafePointScalarObjectNode*)Node::clone(); 1342 sosn_map->Insert((void*)this, (void*)res); 1343 return res; 1344 } 1345 1346 1347 #ifndef PRODUCT 1348 void SafePointScalarObjectNode::dump_spec(outputStream *st) const { 1349 st->print(" # fields@[%d..%d]", first_index(), 1350 first_index() + n_fields() - 1); 1351 } 1352 1353 #endif 1354 1355 //============================================================================= 1356 uint AllocateNode::size_of() const { return sizeof(*this); } 1357 1358 AllocateNode::AllocateNode(Compile* C, const TypeFunc *atype, 1359 Node *ctrl, Node *mem, Node *abio, 1360 Node *size, Node *klass_node, Node *initial_test) 1361 : CallNode(atype, NULL, TypeRawPtr::BOTTOM) 1362 { 1363 init_class_id(Class_Allocate); 1364 init_flags(Flag_is_macro); 1365 _is_scalar_replaceable = false; 1366 _is_non_escaping = false; 1367 _is_allocation_MemBar_redundant = false; 1368 Node *topnode = C->top(); 1369 1370 init_req( TypeFunc::Control , ctrl ); 1371 init_req( TypeFunc::I_O , abio ); 1372 init_req( TypeFunc::Memory , mem ); 1373 init_req( TypeFunc::ReturnAdr, topnode ); 1374 init_req( TypeFunc::FramePtr , topnode ); 1375 init_req( AllocSize , size); 1376 init_req( KlassNode , klass_node); 1377 init_req( InitialTest , initial_test); 1378 init_req( ALength , topnode); 1379 C->add_macro_node(this); 1380 } 1381 1382 void AllocateNode::compute_MemBar_redundancy(ciMethod* initializer) 1383 { 1384 assert(initializer != NULL && 1385 initializer->is_initializer() && 1386 !initializer->is_static(), 1387 "unexpected initializer method"); 1388 BCEscapeAnalyzer* analyzer = initializer->get_bcea(); 1389 if (analyzer == NULL) { 1390 return; 1391 } 1392 1393 // Allocation node is first parameter in its initializer 1394 if (analyzer->is_arg_stack(0) || analyzer->is_arg_local(0)) { 1395 _is_allocation_MemBar_redundant = true; 1396 } 1397 } 1398 Node *AllocateNode::make_ideal_mark(PhaseGVN *phase, Node* obj, Node* control, Node* mem) { 1399 Node* mark_node = NULL; 1400 // For now only enable fast locking for non-array types 1401 if (UseBiasedLocking && Opcode() == Op_Allocate) { 1402 Node* klass_node = in(AllocateNode::KlassNode); 1403 Node* proto_adr = phase->transform(new AddPNode(klass_node, klass_node, phase->MakeConX(in_bytes(Klass::prototype_header_offset())))); 1404 mark_node = LoadNode::make(*phase, control, mem, proto_adr, TypeRawPtr::BOTTOM, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 1405 } else { 1406 mark_node = phase->MakeConX(markWord::prototype().value()); 1407 } 1408 return mark_node; 1409 } 1410 1411 //============================================================================= 1412 Node* AllocateArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1413 if (remove_dead_region(phase, can_reshape)) return this; 1414 // Don't bother trying to transform a dead node 1415 if (in(0) && in(0)->is_top()) return NULL; 1416 1417 const Type* type = phase->type(Ideal_length()); 1418 if (type->isa_int() && type->is_int()->_hi < 0) { 1419 if (can_reshape) { 1420 PhaseIterGVN *igvn = phase->is_IterGVN(); 1421 // Unreachable fall through path (negative array length), 1422 // the allocation can only throw so disconnect it. 1423 Node* proj = proj_out_or_null(TypeFunc::Control); 1424 Node* catchproj = NULL; 1425 if (proj != NULL) { 1426 for (DUIterator_Fast imax, i = proj->fast_outs(imax); i < imax; i++) { 1427 Node *cn = proj->fast_out(i); 1428 if (cn->is_Catch()) { 1429 catchproj = cn->as_Multi()->proj_out_or_null(CatchProjNode::fall_through_index); 1430 break; 1431 } 1432 } 1433 } 1434 if (catchproj != NULL && catchproj->outcnt() > 0 && 1435 (catchproj->outcnt() > 1 || 1436 catchproj->unique_out()->Opcode() != Op_Halt)) { 1437 assert(catchproj->is_CatchProj(), "must be a CatchProjNode"); 1438 Node* nproj = catchproj->clone(); 1439 igvn->register_new_node_with_optimizer(nproj); 1440 1441 Node *frame = new ParmNode( phase->C->start(), TypeFunc::FramePtr ); 1442 frame = phase->transform(frame); 1443 // Halt & Catch Fire 1444 Node* halt = new HaltNode(nproj, frame, "unexpected negative array length"); 1445 phase->C->root()->add_req(halt); 1446 phase->transform(halt); 1447 1448 igvn->replace_node(catchproj, phase->C->top()); 1449 return this; 1450 } 1451 } else { 1452 // Can't correct it during regular GVN so register for IGVN 1453 phase->C->record_for_igvn(this); 1454 } 1455 } 1456 return NULL; 1457 } 1458 1459 // Retrieve the length from the AllocateArrayNode. Narrow the type with a 1460 // CastII, if appropriate. If we are not allowed to create new nodes, and 1461 // a CastII is appropriate, return NULL. 1462 Node *AllocateArrayNode::make_ideal_length(const TypeOopPtr* oop_type, PhaseTransform *phase, bool allow_new_nodes) { 1463 Node *length = in(AllocateNode::ALength); 1464 assert(length != NULL, "length is not null"); 1465 1466 const TypeInt* length_type = phase->find_int_type(length); 1467 const TypeAryPtr* ary_type = oop_type->isa_aryptr(); 1468 1469 if (ary_type != NULL && length_type != NULL) { 1470 const TypeInt* narrow_length_type = ary_type->narrow_size_type(length_type); 1471 if (narrow_length_type != length_type) { 1472 // Assert one of: 1473 // - the narrow_length is 0 1474 // - the narrow_length is not wider than length 1475 assert(narrow_length_type == TypeInt::ZERO || 1476 length_type->is_con() && narrow_length_type->is_con() && 1477 (narrow_length_type->_hi <= length_type->_lo) || 1478 (narrow_length_type->_hi <= length_type->_hi && 1479 narrow_length_type->_lo >= length_type->_lo), 1480 "narrow type must be narrower than length type"); 1481 1482 // Return NULL if new nodes are not allowed 1483 if (!allow_new_nodes) return NULL; 1484 // Create a cast which is control dependent on the initialization to 1485 // propagate the fact that the array length must be positive. 1486 InitializeNode* init = initialization(); 1487 assert(init != NULL, "initialization not found"); 1488 length = new CastIINode(length, narrow_length_type); 1489 length->set_req(0, init->proj_out_or_null(0)); 1490 } 1491 } 1492 1493 return length; 1494 } 1495 1496 //============================================================================= 1497 uint LockNode::size_of() const { return sizeof(*this); } 1498 1499 // Redundant lock elimination 1500 // 1501 // There are various patterns of locking where we release and 1502 // immediately reacquire a lock in a piece of code where no operations 1503 // occur in between that would be observable. In those cases we can 1504 // skip releasing and reacquiring the lock without violating any 1505 // fairness requirements. Doing this around a loop could cause a lock 1506 // to be held for a very long time so we concentrate on non-looping 1507 // control flow. We also require that the operations are fully 1508 // redundant meaning that we don't introduce new lock operations on 1509 // some paths so to be able to eliminate it on others ala PRE. This 1510 // would probably require some more extensive graph manipulation to 1511 // guarantee that the memory edges were all handled correctly. 1512 // 1513 // Assuming p is a simple predicate which can't trap in any way and s 1514 // is a synchronized method consider this code: 1515 // 1516 // s(); 1517 // if (p) 1518 // s(); 1519 // else 1520 // s(); 1521 // s(); 1522 // 1523 // 1. The unlocks of the first call to s can be eliminated if the 1524 // locks inside the then and else branches are eliminated. 1525 // 1526 // 2. The unlocks of the then and else branches can be eliminated if 1527 // the lock of the final call to s is eliminated. 1528 // 1529 // Either of these cases subsumes the simple case of sequential control flow 1530 // 1531 // Addtionally we can eliminate versions without the else case: 1532 // 1533 // s(); 1534 // if (p) 1535 // s(); 1536 // s(); 1537 // 1538 // 3. In this case we eliminate the unlock of the first s, the lock 1539 // and unlock in the then case and the lock in the final s. 1540 // 1541 // Note also that in all these cases the then/else pieces don't have 1542 // to be trivial as long as they begin and end with synchronization 1543 // operations. 1544 // 1545 // s(); 1546 // if (p) 1547 // s(); 1548 // f(); 1549 // s(); 1550 // s(); 1551 // 1552 // The code will work properly for this case, leaving in the unlock 1553 // before the call to f and the relock after it. 1554 // 1555 // A potentially interesting case which isn't handled here is when the 1556 // locking is partially redundant. 1557 // 1558 // s(); 1559 // if (p) 1560 // s(); 1561 // 1562 // This could be eliminated putting unlocking on the else case and 1563 // eliminating the first unlock and the lock in the then side. 1564 // Alternatively the unlock could be moved out of the then side so it 1565 // was after the merge and the first unlock and second lock 1566 // eliminated. This might require less manipulation of the memory 1567 // state to get correct. 1568 // 1569 // Additionally we might allow work between a unlock and lock before 1570 // giving up eliminating the locks. The current code disallows any 1571 // conditional control flow between these operations. A formulation 1572 // similar to partial redundancy elimination computing the 1573 // availability of unlocking and the anticipatability of locking at a 1574 // program point would allow detection of fully redundant locking with 1575 // some amount of work in between. I'm not sure how often I really 1576 // think that would occur though. Most of the cases I've seen 1577 // indicate it's likely non-trivial work would occur in between. 1578 // There may be other more complicated constructs where we could 1579 // eliminate locking but I haven't seen any others appear as hot or 1580 // interesting. 1581 // 1582 // Locking and unlocking have a canonical form in ideal that looks 1583 // roughly like this: 1584 // 1585 // <obj> 1586 // | \\------+ 1587 // | \ \ 1588 // | BoxLock \ 1589 // | | | \ 1590 // | | \ \ 1591 // | | FastLock 1592 // | | / 1593 // | | / 1594 // | | | 1595 // 1596 // Lock 1597 // | 1598 // Proj #0 1599 // | 1600 // MembarAcquire 1601 // | 1602 // Proj #0 1603 // 1604 // MembarRelease 1605 // | 1606 // Proj #0 1607 // | 1608 // Unlock 1609 // | 1610 // Proj #0 1611 // 1612 // 1613 // This code proceeds by processing Lock nodes during PhaseIterGVN 1614 // and searching back through its control for the proper code 1615 // patterns. Once it finds a set of lock and unlock operations to 1616 // eliminate they are marked as eliminatable which causes the 1617 // expansion of the Lock and Unlock macro nodes to make the operation a NOP 1618 // 1619 //============================================================================= 1620 1621 // 1622 // Utility function to skip over uninteresting control nodes. Nodes skipped are: 1623 // - copy regions. (These may not have been optimized away yet.) 1624 // - eliminated locking nodes 1625 // 1626 static Node *next_control(Node *ctrl) { 1627 if (ctrl == NULL) 1628 return NULL; 1629 while (1) { 1630 if (ctrl->is_Region()) { 1631 RegionNode *r = ctrl->as_Region(); 1632 Node *n = r->is_copy(); 1633 if (n == NULL) 1634 break; // hit a region, return it 1635 else 1636 ctrl = n; 1637 } else if (ctrl->is_Proj()) { 1638 Node *in0 = ctrl->in(0); 1639 if (in0->is_AbstractLock() && in0->as_AbstractLock()->is_eliminated()) { 1640 ctrl = in0->in(0); 1641 } else { 1642 break; 1643 } 1644 } else { 1645 break; // found an interesting control 1646 } 1647 } 1648 return ctrl; 1649 } 1650 // 1651 // Given a control, see if it's the control projection of an Unlock which 1652 // operating on the same object as lock. 1653 // 1654 bool AbstractLockNode::find_matching_unlock(const Node* ctrl, LockNode* lock, 1655 GrowableArray<AbstractLockNode*> &lock_ops) { 1656 ProjNode *ctrl_proj = (ctrl->is_Proj()) ? ctrl->as_Proj() : NULL; 1657 if (ctrl_proj != NULL && ctrl_proj->_con == TypeFunc::Control) { 1658 Node *n = ctrl_proj->in(0); 1659 if (n != NULL && n->is_Unlock()) { 1660 UnlockNode *unlock = n->as_Unlock(); 1661 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1662 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node()); 1663 Node* unlock_obj = bs->step_over_gc_barrier(unlock->obj_node()); 1664 if (lock_obj->eqv_uncast(unlock_obj) && 1665 BoxLockNode::same_slot(lock->box_node(), unlock->box_node()) && 1666 !unlock->is_eliminated()) { 1667 lock_ops.append(unlock); 1668 return true; 1669 } 1670 } 1671 } 1672 return false; 1673 } 1674 1675 // 1676 // Find the lock matching an unlock. Returns null if a safepoint 1677 // or complicated control is encountered first. 1678 LockNode *AbstractLockNode::find_matching_lock(UnlockNode* unlock) { 1679 LockNode *lock_result = NULL; 1680 // find the matching lock, or an intervening safepoint 1681 Node *ctrl = next_control(unlock->in(0)); 1682 while (1) { 1683 assert(ctrl != NULL, "invalid control graph"); 1684 assert(!ctrl->is_Start(), "missing lock for unlock"); 1685 if (ctrl->is_top()) break; // dead control path 1686 if (ctrl->is_Proj()) ctrl = ctrl->in(0); 1687 if (ctrl->is_SafePoint()) { 1688 break; // found a safepoint (may be the lock we are searching for) 1689 } else if (ctrl->is_Region()) { 1690 // Check for a simple diamond pattern. Punt on anything more complicated 1691 if (ctrl->req() == 3 && ctrl->in(1) != NULL && ctrl->in(2) != NULL) { 1692 Node *in1 = next_control(ctrl->in(1)); 1693 Node *in2 = next_control(ctrl->in(2)); 1694 if (((in1->is_IfTrue() && in2->is_IfFalse()) || 1695 (in2->is_IfTrue() && in1->is_IfFalse())) && (in1->in(0) == in2->in(0))) { 1696 ctrl = next_control(in1->in(0)->in(0)); 1697 } else { 1698 break; 1699 } 1700 } else { 1701 break; 1702 } 1703 } else { 1704 ctrl = next_control(ctrl->in(0)); // keep searching 1705 } 1706 } 1707 if (ctrl->is_Lock()) { 1708 LockNode *lock = ctrl->as_Lock(); 1709 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1710 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node()); 1711 Node* unlock_obj = bs->step_over_gc_barrier(unlock->obj_node()); 1712 if (lock_obj->eqv_uncast(unlock_obj) && 1713 BoxLockNode::same_slot(lock->box_node(), unlock->box_node())) { 1714 lock_result = lock; 1715 } 1716 } 1717 return lock_result; 1718 } 1719 1720 // This code corresponds to case 3 above. 1721 1722 bool AbstractLockNode::find_lock_and_unlock_through_if(Node* node, LockNode* lock, 1723 GrowableArray<AbstractLockNode*> &lock_ops) { 1724 Node* if_node = node->in(0); 1725 bool if_true = node->is_IfTrue(); 1726 1727 if (if_node->is_If() && if_node->outcnt() == 2 && (if_true || node->is_IfFalse())) { 1728 Node *lock_ctrl = next_control(if_node->in(0)); 1729 if (find_matching_unlock(lock_ctrl, lock, lock_ops)) { 1730 Node* lock1_node = NULL; 1731 ProjNode* proj = if_node->as_If()->proj_out(!if_true); 1732 if (if_true) { 1733 if (proj->is_IfFalse() && proj->outcnt() == 1) { 1734 lock1_node = proj->unique_out(); 1735 } 1736 } else { 1737 if (proj->is_IfTrue() && proj->outcnt() == 1) { 1738 lock1_node = proj->unique_out(); 1739 } 1740 } 1741 if (lock1_node != NULL && lock1_node->is_Lock()) { 1742 LockNode *lock1 = lock1_node->as_Lock(); 1743 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1744 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node()); 1745 Node* lock1_obj = bs->step_over_gc_barrier(lock1->obj_node()); 1746 if (lock_obj->eqv_uncast(lock1_obj) && 1747 BoxLockNode::same_slot(lock->box_node(), lock1->box_node()) && 1748 !lock1->is_eliminated()) { 1749 lock_ops.append(lock1); 1750 return true; 1751 } 1752 } 1753 } 1754 } 1755 1756 lock_ops.trunc_to(0); 1757 return false; 1758 } 1759 1760 bool AbstractLockNode::find_unlocks_for_region(const RegionNode* region, LockNode* lock, 1761 GrowableArray<AbstractLockNode*> &lock_ops) { 1762 // check each control merging at this point for a matching unlock. 1763 // in(0) should be self edge so skip it. 1764 for (int i = 1; i < (int)region->req(); i++) { 1765 Node *in_node = next_control(region->in(i)); 1766 if (in_node != NULL) { 1767 if (find_matching_unlock(in_node, lock, lock_ops)) { 1768 // found a match so keep on checking. 1769 continue; 1770 } else if (find_lock_and_unlock_through_if(in_node, lock, lock_ops)) { 1771 continue; 1772 } 1773 1774 // If we fall through to here then it was some kind of node we 1775 // don't understand or there wasn't a matching unlock, so give 1776 // up trying to merge locks. 1777 lock_ops.trunc_to(0); 1778 return false; 1779 } 1780 } 1781 return true; 1782 1783 } 1784 1785 #ifndef PRODUCT 1786 // 1787 // Create a counter which counts the number of times this lock is acquired 1788 // 1789 void AbstractLockNode::create_lock_counter(JVMState* state) { 1790 _counter = OptoRuntime::new_named_counter(state, NamedCounter::LockCounter); 1791 } 1792 1793 void AbstractLockNode::set_eliminated_lock_counter() { 1794 if (_counter) { 1795 // Update the counter to indicate that this lock was eliminated. 1796 // The counter update code will stay around even though the 1797 // optimizer will eliminate the lock operation itself. 1798 _counter->set_tag(NamedCounter::EliminatedLockCounter); 1799 } 1800 } 1801 1802 const char* AbstractLockNode::_kind_names[] = {"Regular", "NonEscObj", "Coarsened", "Nested"}; 1803 1804 void AbstractLockNode::dump_spec(outputStream* st) const { 1805 st->print("%s ", _kind_names[_kind]); 1806 CallNode::dump_spec(st); 1807 } 1808 1809 void AbstractLockNode::dump_compact_spec(outputStream* st) const { 1810 st->print("%s", _kind_names[_kind]); 1811 } 1812 1813 // The related set of lock nodes includes the control boundary. 1814 void AbstractLockNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const { 1815 if (compact) { 1816 this->collect_nodes(in_rel, 1, false, false); 1817 } else { 1818 this->collect_nodes_in_all_data(in_rel, true); 1819 } 1820 this->collect_nodes(out_rel, -2, false, false); 1821 } 1822 #endif 1823 1824 //============================================================================= 1825 Node *LockNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1826 1827 // perform any generic optimizations first (returns 'this' or NULL) 1828 Node *result = SafePointNode::Ideal(phase, can_reshape); 1829 if (result != NULL) return result; 1830 // Don't bother trying to transform a dead node 1831 if (in(0) && in(0)->is_top()) return NULL; 1832 1833 // Now see if we can optimize away this lock. We don't actually 1834 // remove the locking here, we simply set the _eliminate flag which 1835 // prevents macro expansion from expanding the lock. Since we don't 1836 // modify the graph, the value returned from this function is the 1837 // one computed above. 1838 if (can_reshape && EliminateLocks && !is_non_esc_obj()) { 1839 // 1840 // If we are locking an unescaped object, the lock/unlock is unnecessary 1841 // 1842 ConnectionGraph *cgr = phase->C->congraph(); 1843 if (cgr != NULL && cgr->not_global_escape(obj_node())) { 1844 assert(!is_eliminated() || is_coarsened(), "sanity"); 1845 // The lock could be marked eliminated by lock coarsening 1846 // code during first IGVN before EA. Replace coarsened flag 1847 // to eliminate all associated locks/unlocks. 1848 #ifdef ASSERT 1849 this->log_lock_optimization(phase->C,"eliminate_lock_set_non_esc1"); 1850 #endif 1851 this->set_non_esc_obj(); 1852 return result; 1853 } 1854 1855 // 1856 // Try lock coarsening 1857 // 1858 PhaseIterGVN* iter = phase->is_IterGVN(); 1859 if (iter != NULL && !is_eliminated()) { 1860 1861 GrowableArray<AbstractLockNode*> lock_ops; 1862 1863 Node *ctrl = next_control(in(0)); 1864 1865 // now search back for a matching Unlock 1866 if (find_matching_unlock(ctrl, this, lock_ops)) { 1867 // found an unlock directly preceding this lock. This is the 1868 // case of single unlock directly control dependent on a 1869 // single lock which is the trivial version of case 1 or 2. 1870 } else if (ctrl->is_Region() ) { 1871 if (find_unlocks_for_region(ctrl->as_Region(), this, lock_ops)) { 1872 // found lock preceded by multiple unlocks along all paths 1873 // joining at this point which is case 3 in description above. 1874 } 1875 } else { 1876 // see if this lock comes from either half of an if and the 1877 // predecessors merges unlocks and the other half of the if 1878 // performs a lock. 1879 if (find_lock_and_unlock_through_if(ctrl, this, lock_ops)) { 1880 // found unlock splitting to an if with locks on both branches. 1881 } 1882 } 1883 1884 if (lock_ops.length() > 0) { 1885 // add ourselves to the list of locks to be eliminated. 1886 lock_ops.append(this); 1887 1888 #ifndef PRODUCT 1889 if (PrintEliminateLocks) { 1890 int locks = 0; 1891 int unlocks = 0; 1892 for (int i = 0; i < lock_ops.length(); i++) { 1893 AbstractLockNode* lock = lock_ops.at(i); 1894 if (lock->Opcode() == Op_Lock) 1895 locks++; 1896 else 1897 unlocks++; 1898 if (Verbose) { 1899 lock->dump(1); 1900 } 1901 } 1902 tty->print_cr("***Eliminated %d unlocks and %d locks", unlocks, locks); 1903 } 1904 #endif 1905 1906 // for each of the identified locks, mark them 1907 // as eliminatable 1908 for (int i = 0; i < lock_ops.length(); i++) { 1909 AbstractLockNode* lock = lock_ops.at(i); 1910 1911 // Mark it eliminated by coarsening and update any counters 1912 #ifdef ASSERT 1913 lock->log_lock_optimization(phase->C, "eliminate_lock_set_coarsened"); 1914 #endif 1915 lock->set_coarsened(); 1916 } 1917 } else if (ctrl->is_Region() && 1918 iter->_worklist.member(ctrl)) { 1919 // We weren't able to find any opportunities but the region this 1920 // lock is control dependent on hasn't been processed yet so put 1921 // this lock back on the worklist so we can check again once any 1922 // region simplification has occurred. 1923 iter->_worklist.push(this); 1924 } 1925 } 1926 } 1927 1928 return result; 1929 } 1930 1931 //============================================================================= 1932 bool LockNode::is_nested_lock_region() { 1933 return is_nested_lock_region(NULL); 1934 } 1935 1936 // p is used for access to compilation log; no logging if NULL 1937 bool LockNode::is_nested_lock_region(Compile * c) { 1938 BoxLockNode* box = box_node()->as_BoxLock(); 1939 int stk_slot = box->stack_slot(); 1940 if (stk_slot <= 0) { 1941 #ifdef ASSERT 1942 this->log_lock_optimization(c, "eliminate_lock_INLR_1"); 1943 #endif 1944 return false; // External lock or it is not Box (Phi node). 1945 } 1946 1947 // Ignore complex cases: merged locks or multiple locks. 1948 Node* obj = obj_node(); 1949 LockNode* unique_lock = NULL; 1950 if (!box->is_simple_lock_region(&unique_lock, obj)) { 1951 #ifdef ASSERT 1952 this->log_lock_optimization(c, "eliminate_lock_INLR_2a"); 1953 #endif 1954 return false; 1955 } 1956 if (unique_lock != this) { 1957 #ifdef ASSERT 1958 this->log_lock_optimization(c, "eliminate_lock_INLR_2b"); 1959 #endif 1960 return false; 1961 } 1962 1963 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1964 obj = bs->step_over_gc_barrier(obj); 1965 // Look for external lock for the same object. 1966 SafePointNode* sfn = this->as_SafePoint(); 1967 JVMState* youngest_jvms = sfn->jvms(); 1968 int max_depth = youngest_jvms->depth(); 1969 for (int depth = 1; depth <= max_depth; depth++) { 1970 JVMState* jvms = youngest_jvms->of_depth(depth); 1971 int num_mon = jvms->nof_monitors(); 1972 // Loop over monitors 1973 for (int idx = 0; idx < num_mon; idx++) { 1974 Node* obj_node = sfn->monitor_obj(jvms, idx); 1975 obj_node = bs->step_over_gc_barrier(obj_node); 1976 BoxLockNode* box_node = sfn->monitor_box(jvms, idx)->as_BoxLock(); 1977 if ((box_node->stack_slot() < stk_slot) && obj_node->eqv_uncast(obj)) { 1978 return true; 1979 } 1980 } 1981 } 1982 #ifdef ASSERT 1983 this->log_lock_optimization(c, "eliminate_lock_INLR_3"); 1984 #endif 1985 return false; 1986 } 1987 1988 //============================================================================= 1989 uint UnlockNode::size_of() const { return sizeof(*this); } 1990 1991 //============================================================================= 1992 Node *UnlockNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1993 1994 // perform any generic optimizations first (returns 'this' or NULL) 1995 Node *result = SafePointNode::Ideal(phase, can_reshape); 1996 if (result != NULL) return result; 1997 // Don't bother trying to transform a dead node 1998 if (in(0) && in(0)->is_top()) return NULL; 1999 2000 // Now see if we can optimize away this unlock. We don't actually 2001 // remove the unlocking here, we simply set the _eliminate flag which 2002 // prevents macro expansion from expanding the unlock. Since we don't 2003 // modify the graph, the value returned from this function is the 2004 // one computed above. 2005 // Escape state is defined after Parse phase. 2006 if (can_reshape && EliminateLocks && !is_non_esc_obj()) { 2007 // 2008 // If we are unlocking an unescaped object, the lock/unlock is unnecessary. 2009 // 2010 ConnectionGraph *cgr = phase->C->congraph(); 2011 if (cgr != NULL && cgr->not_global_escape(obj_node())) { 2012 assert(!is_eliminated() || is_coarsened(), "sanity"); 2013 // The lock could be marked eliminated by lock coarsening 2014 // code during first IGVN before EA. Replace coarsened flag 2015 // to eliminate all associated locks/unlocks. 2016 #ifdef ASSERT 2017 this->log_lock_optimization(phase->C, "eliminate_lock_set_non_esc2"); 2018 #endif 2019 this->set_non_esc_obj(); 2020 } 2021 } 2022 return result; 2023 } 2024 2025 const char * AbstractLockNode::kind_as_string() const { 2026 return is_coarsened() ? "coarsened" : 2027 is_nested() ? "nested" : 2028 is_non_esc_obj() ? "non_escaping" : 2029 "?"; 2030 } 2031 2032 void AbstractLockNode::log_lock_optimization(Compile *C, const char * tag) const { 2033 if (C == NULL) { 2034 return; 2035 } 2036 CompileLog* log = C->log(); 2037 if (log != NULL) { 2038 log->begin_head("%s lock='%d' compile_id='%d' class_id='%s' kind='%s'", 2039 tag, is_Lock(), C->compile_id(), 2040 is_Unlock() ? "unlock" : is_Lock() ? "lock" : "?", 2041 kind_as_string()); 2042 log->stamp(); 2043 log->end_head(); 2044 JVMState* p = is_Unlock() ? (as_Unlock()->dbg_jvms()) : jvms(); 2045 while (p != NULL) { 2046 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 2047 p = p->caller(); 2048 } 2049 log->tail(tag); 2050 } 2051 } 2052 2053 bool CallNode::may_modify_arraycopy_helper(const TypeOopPtr* dest_t, const TypeOopPtr *t_oop, PhaseTransform *phase) { 2054 if (dest_t->is_known_instance() && t_oop->is_known_instance()) { 2055 return dest_t->instance_id() == t_oop->instance_id(); 2056 } 2057 2058 if (dest_t->isa_instptr() && !dest_t->klass()->equals(phase->C->env()->Object_klass())) { 2059 // clone 2060 if (t_oop->isa_aryptr()) { 2061 return false; 2062 } 2063 if (!t_oop->isa_instptr()) { 2064 return true; 2065 } 2066 if (dest_t->klass()->is_subtype_of(t_oop->klass()) || t_oop->klass()->is_subtype_of(dest_t->klass())) { 2067 return true; 2068 } 2069 // unrelated 2070 return false; 2071 } 2072 2073 if (dest_t->isa_aryptr()) { 2074 // arraycopy or array clone 2075 if (t_oop->isa_instptr()) { 2076 return false; 2077 } 2078 if (!t_oop->isa_aryptr()) { 2079 return true; 2080 } 2081 2082 const Type* elem = dest_t->is_aryptr()->elem(); 2083 if (elem == Type::BOTTOM) { 2084 // An array but we don't know what elements are 2085 return true; 2086 } 2087 2088 dest_t = dest_t->add_offset(Type::OffsetBot)->is_oopptr(); 2089 uint dest_alias = phase->C->get_alias_index(dest_t); 2090 uint t_oop_alias = phase->C->get_alias_index(t_oop); 2091 2092 return dest_alias == t_oop_alias; 2093 } 2094 2095 return true; 2096 }