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 = NULL; 489 ciField* cifield; 490 if (iklass != NULL) { 491 st->print(" ["); 492 cifield = iklass->nonstatic_field_at(0); 493 cifield->print_name_on(st); 494 if(spobj->stack_allocated()) { 495 st->print(":*0]"); 496 } else { 497 fld_node = mcall->in(first_ind); 498 format_helper(regalloc, st, fld_node, ":", 0, &scobjs); 499 } 500 } else { 501 if(spobj->stack_allocated()) { 502 st->print("[*0]"); 503 } else { 504 fld_node = mcall->in(first_ind); 505 format_helper(regalloc, st, fld_node, "[", 0, &scobjs); 506 } 507 } 508 for (uint j = 1; j < nf; j++) { 509 if (iklass != NULL) { 510 st->print(", ["); 511 cifield = iklass->nonstatic_field_at(j); 512 cifield->print_name_on(st); 513 if(spobj->stack_allocated()) { 514 st->print(":*%d]", j); 515 } else { 516 fld_node = mcall->in(first_ind+j); 517 format_helper(regalloc, st, fld_node, ":", j, &scobjs); 518 } 519 } else { 520 if(spobj->stack_allocated()) { 521 st->print(", [*%d]", j); 522 } else { 523 fld_node = mcall->in(first_ind+j); 524 format_helper(regalloc, st, fld_node, ", [", j, &scobjs); 525 } 526 } 527 } 528 } 529 st->print(" }"); 530 } 531 } 532 st->cr(); 533 if (caller() != NULL) caller()->format(regalloc, n, st); 534 } 535 536 537 void JVMState::dump_spec(outputStream *st) const { 538 if (_method != NULL) { 539 bool printed = false; 540 if (!Verbose) { 541 // The JVMS dumps make really, really long lines. 542 // Take out the most boring parts, which are the package prefixes. 543 char buf[500]; 544 stringStream namest(buf, sizeof(buf)); 545 _method->print_short_name(&namest); 546 if (namest.count() < sizeof(buf)) { 547 const char* name = namest.base(); 548 if (name[0] == ' ') ++name; 549 const char* endcn = strchr(name, ':'); // end of class name 550 if (endcn == NULL) endcn = strchr(name, '('); 551 if (endcn == NULL) endcn = name + strlen(name); 552 while (endcn > name && endcn[-1] != '.' && endcn[-1] != '/') 553 --endcn; 554 st->print(" %s", endcn); 555 printed = true; 556 } 557 } 558 if (!printed) 559 _method->print_short_name(st); 560 st->print(" @ bci:%d",_bci); 561 if(_reexecute == Reexecute_True) 562 st->print(" reexecute"); 563 } else { 564 st->print(" runtime stub"); 565 } 566 if (caller() != NULL) caller()->dump_spec(st); 567 } 568 569 570 void JVMState::dump_on(outputStream* st) const { 571 bool print_map = _map && !((uintptr_t)_map & 1) && 572 ((caller() == NULL) || (caller()->map() != _map)); 573 if (print_map) { 574 if (_map->len() > _map->req()) { // _map->has_exceptions() 575 Node* ex = _map->in(_map->req()); // _map->next_exception() 576 // skip the first one; it's already being printed 577 while (ex != NULL && ex->len() > ex->req()) { 578 ex = ex->in(ex->req()); // ex->next_exception() 579 ex->dump(1); 580 } 581 } 582 _map->dump(Verbose ? 2 : 1); 583 } 584 if (caller() != NULL) { 585 caller()->dump_on(st); 586 } 587 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=", 588 depth(), locoff(), stkoff(), argoff(), monoff(), scloff(), endoff(), monitor_depth(), sp(), bci(), should_reexecute()?"true":"false"); 589 if (_method == NULL) { 590 st->print_cr("(none)"); 591 } else { 592 _method->print_name(st); 593 st->cr(); 594 if (bci() >= 0 && bci() < _method->code_size()) { 595 st->print(" bc: "); 596 _method->print_codes_on(bci(), bci()+1, st); 597 } 598 } 599 } 600 601 // Extra way to dump a jvms from the debugger, 602 // to avoid a bug with C++ member function calls. 603 void dump_jvms(JVMState* jvms) { 604 jvms->dump(); 605 } 606 #endif 607 608 //--------------------------clone_shallow-------------------------------------- 609 JVMState* JVMState::clone_shallow(Compile* C) const { 610 JVMState* n = has_method() ? new (C) JVMState(_method, _caller) : new (C) JVMState(0); 611 n->set_bci(_bci); 612 n->_reexecute = _reexecute; 613 n->set_locoff(_locoff); 614 n->set_stkoff(_stkoff); 615 n->set_monoff(_monoff); 616 n->set_scloff(_scloff); 617 n->set_endoff(_endoff); 618 n->set_sp(_sp); 619 n->set_map(_map); 620 return n; 621 } 622 623 //---------------------------clone_deep---------------------------------------- 624 JVMState* JVMState::clone_deep(Compile* C) const { 625 JVMState* n = clone_shallow(C); 626 for (JVMState* p = n; p->_caller != NULL; p = p->_caller) { 627 p->_caller = p->_caller->clone_shallow(C); 628 } 629 assert(n->depth() == depth(), "sanity"); 630 assert(n->debug_depth() == debug_depth(), "sanity"); 631 return n; 632 } 633 634 /** 635 * Reset map for all callers 636 */ 637 void JVMState::set_map_deep(SafePointNode* map) { 638 for (JVMState* p = this; p->_caller != NULL; p = p->_caller) { 639 p->set_map(map); 640 } 641 } 642 643 // Adapt offsets in in-array after adding or removing an edge. 644 // Prerequisite is that the JVMState is used by only one node. 645 void JVMState::adapt_position(int delta) { 646 for (JVMState* jvms = this; jvms != NULL; jvms = jvms->caller()) { 647 jvms->set_locoff(jvms->locoff() + delta); 648 jvms->set_stkoff(jvms->stkoff() + delta); 649 jvms->set_monoff(jvms->monoff() + delta); 650 jvms->set_scloff(jvms->scloff() + delta); 651 jvms->set_endoff(jvms->endoff() + delta); 652 } 653 } 654 655 // Mirror the stack size calculation in the deopt code 656 // How much stack space would we need at this point in the program in 657 // case of deoptimization? 658 int JVMState::interpreter_frame_size() const { 659 const JVMState* jvms = this; 660 int size = 0; 661 int callee_parameters = 0; 662 int callee_locals = 0; 663 int extra_args = method()->max_stack() - stk_size(); 664 665 while (jvms != NULL) { 666 int locks = jvms->nof_monitors(); 667 int temps = jvms->stk_size(); 668 bool is_top_frame = (jvms == this); 669 ciMethod* method = jvms->method(); 670 671 int frame_size = BytesPerWord * Interpreter::size_activation(method->max_stack(), 672 temps + callee_parameters, 673 extra_args, 674 locks, 675 callee_parameters, 676 callee_locals, 677 is_top_frame); 678 size += frame_size; 679 680 callee_parameters = method->size_of_parameters(); 681 callee_locals = method->max_locals(); 682 extra_args = 0; 683 jvms = jvms->caller(); 684 } 685 return size + Deoptimization::last_frame_adjust(0, callee_locals) * BytesPerWord; 686 } 687 688 //============================================================================= 689 bool CallNode::cmp( const Node &n ) const 690 { return _tf == ((CallNode&)n)._tf && _jvms == ((CallNode&)n)._jvms; } 691 #ifndef PRODUCT 692 void CallNode::dump_req(outputStream *st) const { 693 // Dump the required inputs, enclosed in '(' and ')' 694 uint i; // Exit value of loop 695 for (i = 0; i < req(); i++) { // For all required inputs 696 if (i == TypeFunc::Parms) st->print("("); 697 if (in(i)) st->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx); 698 else st->print("_ "); 699 } 700 st->print(")"); 701 } 702 703 void CallNode::dump_spec(outputStream *st) const { 704 st->print(" "); 705 if (tf() != NULL) tf()->dump_on(st); 706 if (_cnt != COUNT_UNKNOWN) st->print(" C=%f",_cnt); 707 if (jvms() != NULL) jvms()->dump_spec(st); 708 } 709 #endif 710 711 const Type *CallNode::bottom_type() const { return tf()->range(); } 712 const Type* CallNode::Value(PhaseGVN* phase) const { 713 if (phase->type(in(0)) == Type::TOP) return Type::TOP; 714 return tf()->range(); 715 } 716 717 //------------------------------calling_convention----------------------------- 718 void CallNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const { 719 // Use the standard compiler calling convention 720 Matcher::calling_convention( sig_bt, parm_regs, argcnt, true ); 721 } 722 723 724 //------------------------------match------------------------------------------ 725 // Construct projections for control, I/O, memory-fields, ..., and 726 // return result(s) along with their RegMask info 727 Node *CallNode::match( const ProjNode *proj, const Matcher *match ) { 728 switch (proj->_con) { 729 case TypeFunc::Control: 730 case TypeFunc::I_O: 731 case TypeFunc::Memory: 732 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 733 734 case TypeFunc::Parms+1: // For LONG & DOUBLE returns 735 assert(tf()->range()->field_at(TypeFunc::Parms+1) == Type::HALF, ""); 736 // 2nd half of doubles and longs 737 return new MachProjNode(this,proj->_con, RegMask::Empty, (uint)OptoReg::Bad); 738 739 case TypeFunc::Parms: { // Normal returns 740 uint ideal_reg = tf()->range()->field_at(TypeFunc::Parms)->ideal_reg(); 741 OptoRegPair regs = is_CallRuntime() 742 ? match->c_return_value(ideal_reg,true) // Calls into C runtime 743 : match-> return_value(ideal_reg,true); // Calls into compiled Java code 744 RegMask rm = RegMask(regs.first()); 745 if( OptoReg::is_valid(regs.second()) ) 746 rm.Insert( regs.second() ); 747 return new MachProjNode(this,proj->_con,rm,ideal_reg); 748 } 749 750 case TypeFunc::ReturnAdr: 751 case TypeFunc::FramePtr: 752 default: 753 ShouldNotReachHere(); 754 } 755 return NULL; 756 } 757 758 // Do we Match on this edge index or not? Match no edges 759 uint CallNode::match_edge(uint idx) const { 760 return 0; 761 } 762 763 // 764 // Determine whether the call could modify the field of the specified 765 // instance at the specified offset. 766 // 767 bool CallNode::may_modify(const TypeOopPtr *t_oop, PhaseTransform *phase) { 768 assert((t_oop != NULL), "sanity"); 769 if (is_call_to_arraycopystub() && strcmp(_name, "unsafe_arraycopy") != 0) { 770 const TypeTuple* args = _tf->domain(); 771 Node* dest = NULL; 772 // Stubs that can be called once an ArrayCopyNode is expanded have 773 // different signatures. Look for the second pointer argument, 774 // that is the destination of the copy. 775 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) { 776 if (args->field_at(i)->isa_ptr()) { 777 j++; 778 if (j == 2) { 779 dest = in(i); 780 break; 781 } 782 } 783 } 784 guarantee(dest != NULL, "Call had only one ptr in, broken IR!"); 785 if (!dest->is_top() && may_modify_arraycopy_helper(phase->type(dest)->is_oopptr(), t_oop, phase)) { 786 return true; 787 } 788 return false; 789 } 790 if (t_oop->is_known_instance()) { 791 // The instance_id is set only for scalar-replaceable allocations which 792 // are not passed as arguments according to Escape Analysis. 793 return false; 794 } 795 if (t_oop->is_ptr_to_boxed_value()) { 796 ciKlass* boxing_klass = t_oop->klass(); 797 if (is_CallStaticJava() && as_CallStaticJava()->is_boxing_method()) { 798 // Skip unrelated boxing methods. 799 Node* proj = proj_out_or_null(TypeFunc::Parms); 800 if ((proj == NULL) || (phase->type(proj)->is_instptr()->klass() != boxing_klass)) { 801 return false; 802 } 803 } 804 if (is_CallJava() && as_CallJava()->method() != NULL) { 805 ciMethod* meth = as_CallJava()->method(); 806 if (meth->is_getter()) { 807 return false; 808 } 809 // May modify (by reflection) if an boxing object is passed 810 // as argument or returned. 811 Node* proj = returns_pointer() ? proj_out_or_null(TypeFunc::Parms) : NULL; 812 if (proj != NULL) { 813 const TypeInstPtr* inst_t = phase->type(proj)->isa_instptr(); 814 if ((inst_t != NULL) && (!inst_t->klass_is_exact() || 815 (inst_t->klass() == boxing_klass))) { 816 return true; 817 } 818 } 819 const TypeTuple* d = tf()->domain(); 820 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 821 const TypeInstPtr* inst_t = d->field_at(i)->isa_instptr(); 822 if ((inst_t != NULL) && (!inst_t->klass_is_exact() || 823 (inst_t->klass() == boxing_klass))) { 824 return true; 825 } 826 } 827 return false; 828 } 829 } 830 return true; 831 } 832 833 // Does this call have a direct reference to n other than debug information? 834 bool CallNode::has_non_debug_use(Node *n) { 835 const TypeTuple * d = tf()->domain(); 836 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 837 Node *arg = in(i); 838 if (arg == n) { 839 return true; 840 } 841 } 842 return false; 843 } 844 845 // Returns the unique CheckCastPP of a call 846 // or 'this' if there are several CheckCastPP or unexpected uses 847 // or returns NULL if there is no one. 848 Node *CallNode::result_cast() { 849 Node *cast = NULL; 850 851 Node *p = proj_out_or_null(TypeFunc::Parms); 852 if (p == NULL) 853 return NULL; 854 855 for (DUIterator_Fast imax, i = p->fast_outs(imax); i < imax; i++) { 856 Node *use = p->fast_out(i); 857 if (use->is_CheckCastPP()) { 858 if (cast != NULL) { 859 return this; // more than 1 CheckCastPP 860 } 861 cast = use; 862 } else if (!use->is_Initialize() && 863 !use->is_AddP() && 864 use->Opcode() != Op_MemBarStoreStore) { 865 // Expected uses are restricted to a CheckCastPP, an Initialize 866 // node, a MemBarStoreStore (clone) and AddP nodes. If we 867 // encounter any other use (a Phi node can be seen in rare 868 // cases) return this to prevent incorrect optimizations. 869 return this; 870 } 871 } 872 return cast; 873 } 874 875 876 void CallNode::extract_projections(CallProjections* projs, bool separate_io_proj, bool do_asserts) { 877 projs->fallthrough_proj = NULL; 878 projs->fallthrough_catchproj = NULL; 879 projs->fallthrough_ioproj = NULL; 880 projs->catchall_ioproj = NULL; 881 projs->catchall_catchproj = NULL; 882 projs->fallthrough_memproj = NULL; 883 projs->catchall_memproj = NULL; 884 projs->resproj = NULL; 885 projs->exobj = NULL; 886 887 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 888 ProjNode *pn = fast_out(i)->as_Proj(); 889 if (pn->outcnt() == 0) continue; 890 switch (pn->_con) { 891 case TypeFunc::Control: 892 { 893 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj 894 projs->fallthrough_proj = pn; 895 DUIterator_Fast jmax, j = pn->fast_outs(jmax); 896 const Node *cn = pn->fast_out(j); 897 if (cn->is_Catch()) { 898 ProjNode *cpn = NULL; 899 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) { 900 cpn = cn->fast_out(k)->as_Proj(); 901 assert(cpn->is_CatchProj(), "must be a CatchProjNode"); 902 if (cpn->_con == CatchProjNode::fall_through_index) 903 projs->fallthrough_catchproj = cpn; 904 else { 905 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index."); 906 projs->catchall_catchproj = cpn; 907 } 908 } 909 } 910 break; 911 } 912 case TypeFunc::I_O: 913 if (pn->_is_io_use) 914 projs->catchall_ioproj = pn; 915 else 916 projs->fallthrough_ioproj = pn; 917 for (DUIterator j = pn->outs(); pn->has_out(j); j++) { 918 Node* e = pn->out(j); 919 if (e->Opcode() == Op_CreateEx && e->in(0)->is_CatchProj() && e->outcnt() > 0) { 920 assert(projs->exobj == NULL, "only one"); 921 projs->exobj = e; 922 } 923 } 924 break; 925 case TypeFunc::Memory: 926 if (pn->_is_io_use) 927 projs->catchall_memproj = pn; 928 else 929 projs->fallthrough_memproj = pn; 930 break; 931 case TypeFunc::Parms: 932 projs->resproj = pn; 933 break; 934 default: 935 assert(false, "unexpected projection from allocation node."); 936 } 937 } 938 939 // The resproj may not exist because the result could be ignored 940 // and the exception object may not exist if an exception handler 941 // swallows the exception but all the other must exist and be found. 942 assert(projs->fallthrough_proj != NULL, "must be found"); 943 do_asserts = do_asserts && !Compile::current()->inlining_incrementally(); 944 assert(!do_asserts || projs->fallthrough_catchproj != NULL, "must be found"); 945 assert(!do_asserts || projs->fallthrough_memproj != NULL, "must be found"); 946 assert(!do_asserts || projs->fallthrough_ioproj != NULL, "must be found"); 947 assert(!do_asserts || projs->catchall_catchproj != NULL, "must be found"); 948 if (separate_io_proj) { 949 assert(!do_asserts || projs->catchall_memproj != NULL, "must be found"); 950 assert(!do_asserts || projs->catchall_ioproj != NULL, "must be found"); 951 } 952 } 953 954 Node *CallNode::Ideal(PhaseGVN *phase, bool can_reshape) { 955 CallGenerator* cg = generator(); 956 if (can_reshape && cg != NULL && cg->is_mh_late_inline() && !cg->already_attempted()) { 957 // Check whether this MH handle call becomes a candidate for inlining 958 ciMethod* callee = cg->method(); 959 vmIntrinsics::ID iid = callee->intrinsic_id(); 960 if (iid == vmIntrinsics::_invokeBasic) { 961 if (in(TypeFunc::Parms)->Opcode() == Op_ConP) { 962 phase->C->prepend_late_inline(cg); 963 set_generator(NULL); 964 } 965 } else { 966 assert(callee->has_member_arg(), "wrong type of call?"); 967 if (in(TypeFunc::Parms + callee->arg_size() - 1)->Opcode() == Op_ConP) { 968 phase->C->prepend_late_inline(cg); 969 set_generator(NULL); 970 } 971 } 972 } 973 return SafePointNode::Ideal(phase, can_reshape); 974 } 975 976 bool CallNode::is_call_to_arraycopystub() const { 977 if (_name != NULL && strstr(_name, "arraycopy") != 0) { 978 return true; 979 } 980 return false; 981 } 982 983 bool CallNode::is_call_to_osr_migration_end() const { 984 if (_name != NULL && strstr(_name, "OSR_migration_end") != 0) { 985 return true; 986 } 987 return false; 988 } 989 990 //============================================================================= 991 uint CallJavaNode::size_of() const { return sizeof(*this); } 992 bool CallJavaNode::cmp( const Node &n ) const { 993 CallJavaNode &call = (CallJavaNode&)n; 994 return CallNode::cmp(call) && _method == call._method && 995 _override_symbolic_info == call._override_symbolic_info; 996 } 997 #ifdef ASSERT 998 bool CallJavaNode::validate_symbolic_info() const { 999 if (method() == NULL) { 1000 return true; // call into runtime or uncommon trap 1001 } 1002 ciMethod* symbolic_info = jvms()->method()->get_method_at_bci(_bci); 1003 ciMethod* callee = method(); 1004 if (symbolic_info->is_method_handle_intrinsic() && !callee->is_method_handle_intrinsic()) { 1005 assert(override_symbolic_info(), "should be set"); 1006 } 1007 assert(ciMethod::is_consistent_info(symbolic_info, callee), "inconsistent info"); 1008 return true; 1009 } 1010 #endif 1011 1012 #ifndef PRODUCT 1013 void CallJavaNode::dump_spec(outputStream *st) const { 1014 if( _method ) _method->print_short_name(st); 1015 CallNode::dump_spec(st); 1016 } 1017 1018 void CallJavaNode::dump_compact_spec(outputStream* st) const { 1019 if (_method) { 1020 _method->print_short_name(st); 1021 } else { 1022 st->print("<?>"); 1023 } 1024 } 1025 #endif 1026 1027 //============================================================================= 1028 uint CallStaticJavaNode::size_of() const { return sizeof(*this); } 1029 bool CallStaticJavaNode::cmp( const Node &n ) const { 1030 CallStaticJavaNode &call = (CallStaticJavaNode&)n; 1031 return CallJavaNode::cmp(call); 1032 } 1033 1034 //----------------------------uncommon_trap_request---------------------------- 1035 // If this is an uncommon trap, return the request code, else zero. 1036 int CallStaticJavaNode::uncommon_trap_request() const { 1037 if (_name != NULL && !strcmp(_name, "uncommon_trap")) { 1038 return extract_uncommon_trap_request(this); 1039 } 1040 return 0; 1041 } 1042 int CallStaticJavaNode::extract_uncommon_trap_request(const Node* call) { 1043 #ifndef PRODUCT 1044 if (!(call->req() > TypeFunc::Parms && 1045 call->in(TypeFunc::Parms) != NULL && 1046 call->in(TypeFunc::Parms)->is_Con() && 1047 call->in(TypeFunc::Parms)->bottom_type()->isa_int())) { 1048 assert(in_dump() != 0, "OK if dumping"); 1049 tty->print("[bad uncommon trap]"); 1050 return 0; 1051 } 1052 #endif 1053 return call->in(TypeFunc::Parms)->bottom_type()->is_int()->get_con(); 1054 } 1055 1056 #ifndef PRODUCT 1057 void CallStaticJavaNode::dump_spec(outputStream *st) const { 1058 st->print("# Static "); 1059 if (_name != NULL) { 1060 st->print("%s", _name); 1061 int trap_req = uncommon_trap_request(); 1062 if (trap_req != 0) { 1063 char buf[100]; 1064 st->print("(%s)", 1065 Deoptimization::format_trap_request(buf, sizeof(buf), 1066 trap_req)); 1067 } 1068 st->print(" "); 1069 } 1070 CallJavaNode::dump_spec(st); 1071 } 1072 1073 void CallStaticJavaNode::dump_compact_spec(outputStream* st) const { 1074 if (_method) { 1075 _method->print_short_name(st); 1076 } else if (_name) { 1077 st->print("%s", _name); 1078 } else { 1079 st->print("<?>"); 1080 } 1081 } 1082 #endif 1083 1084 //============================================================================= 1085 uint CallDynamicJavaNode::size_of() const { return sizeof(*this); } 1086 bool CallDynamicJavaNode::cmp( const Node &n ) const { 1087 CallDynamicJavaNode &call = (CallDynamicJavaNode&)n; 1088 return CallJavaNode::cmp(call); 1089 } 1090 #ifndef PRODUCT 1091 void CallDynamicJavaNode::dump_spec(outputStream *st) const { 1092 st->print("# Dynamic "); 1093 CallJavaNode::dump_spec(st); 1094 } 1095 #endif 1096 1097 //============================================================================= 1098 uint CallRuntimeNode::size_of() const { return sizeof(*this); } 1099 bool CallRuntimeNode::cmp( const Node &n ) const { 1100 CallRuntimeNode &call = (CallRuntimeNode&)n; 1101 return CallNode::cmp(call) && !strcmp(_name,call._name); 1102 } 1103 #ifndef PRODUCT 1104 void CallRuntimeNode::dump_spec(outputStream *st) const { 1105 st->print("# "); 1106 st->print("%s", _name); 1107 CallNode::dump_spec(st); 1108 } 1109 #endif 1110 1111 //------------------------------calling_convention----------------------------- 1112 void CallRuntimeNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const { 1113 Matcher::c_calling_convention( sig_bt, parm_regs, argcnt ); 1114 } 1115 1116 //============================================================================= 1117 //------------------------------calling_convention----------------------------- 1118 1119 1120 //============================================================================= 1121 #ifndef PRODUCT 1122 void CallLeafNode::dump_spec(outputStream *st) const { 1123 st->print("# "); 1124 st->print("%s", _name); 1125 CallNode::dump_spec(st); 1126 } 1127 #endif 1128 1129 //============================================================================= 1130 1131 void SafePointNode::set_local(JVMState* jvms, uint idx, Node *c) { 1132 assert(verify_jvms(jvms), "jvms must match"); 1133 int loc = jvms->locoff() + idx; 1134 if (in(loc)->is_top() && idx > 0 && !c->is_top() ) { 1135 // If current local idx is top then local idx - 1 could 1136 // be a long/double that needs to be killed since top could 1137 // represent the 2nd half ofthe long/double. 1138 uint ideal = in(loc -1)->ideal_reg(); 1139 if (ideal == Op_RegD || ideal == Op_RegL) { 1140 // set other (low index) half to top 1141 set_req(loc - 1, in(loc)); 1142 } 1143 } 1144 set_req(loc, c); 1145 } 1146 1147 uint SafePointNode::size_of() const { return sizeof(*this); } 1148 bool SafePointNode::cmp( const Node &n ) const { 1149 return (&n == this); // Always fail except on self 1150 } 1151 1152 //-------------------------set_next_exception---------------------------------- 1153 void SafePointNode::set_next_exception(SafePointNode* n) { 1154 assert(n == NULL || n->Opcode() == Op_SafePoint, "correct value for next_exception"); 1155 if (len() == req()) { 1156 if (n != NULL) add_prec(n); 1157 } else { 1158 set_prec(req(), n); 1159 } 1160 } 1161 1162 1163 //----------------------------next_exception----------------------------------- 1164 SafePointNode* SafePointNode::next_exception() const { 1165 if (len() == req()) { 1166 return NULL; 1167 } else { 1168 Node* n = in(req()); 1169 assert(n == NULL || n->Opcode() == Op_SafePoint, "no other uses of prec edges"); 1170 return (SafePointNode*) n; 1171 } 1172 } 1173 1174 1175 //------------------------------Ideal------------------------------------------ 1176 // Skip over any collapsed Regions 1177 Node *SafePointNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1178 return remove_dead_region(phase, can_reshape) ? this : NULL; 1179 } 1180 1181 //------------------------------Identity--------------------------------------- 1182 // Remove obviously duplicate safepoints 1183 Node* SafePointNode::Identity(PhaseGVN* phase) { 1184 1185 // If you have back to back safepoints, remove one 1186 if( in(TypeFunc::Control)->is_SafePoint() ) 1187 return in(TypeFunc::Control); 1188 1189 if( in(0)->is_Proj() ) { 1190 Node *n0 = in(0)->in(0); 1191 // Check if he is a call projection (except Leaf Call) 1192 if( n0->is_Catch() ) { 1193 n0 = n0->in(0)->in(0); 1194 assert( n0->is_Call(), "expect a call here" ); 1195 } 1196 if( n0->is_Call() && n0->as_Call()->guaranteed_safepoint() ) { 1197 // Don't remove a safepoint belonging to an OuterStripMinedLoopEndNode. 1198 // If the loop dies, they will be removed together. 1199 if (has_out_with(Op_OuterStripMinedLoopEnd)) { 1200 return this; 1201 } 1202 // Useless Safepoint, so remove it 1203 return in(TypeFunc::Control); 1204 } 1205 } 1206 1207 return this; 1208 } 1209 1210 //------------------------------Value------------------------------------------ 1211 const Type* SafePointNode::Value(PhaseGVN* phase) const { 1212 if( phase->type(in(0)) == Type::TOP ) return Type::TOP; 1213 if( phase->eqv( in(0), this ) ) return Type::TOP; // Dead infinite loop 1214 return Type::CONTROL; 1215 } 1216 1217 #ifndef PRODUCT 1218 void SafePointNode::dump_spec(outputStream *st) const { 1219 st->print(" SafePoint "); 1220 _replaced_nodes.dump(st); 1221 } 1222 1223 // The related nodes of a SafepointNode are all data inputs, excluding the 1224 // control boundary, as well as all outputs till level 2 (to include projection 1225 // nodes and targets). In compact mode, just include inputs till level 1 and 1226 // outputs as before. 1227 void SafePointNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const { 1228 if (compact) { 1229 this->collect_nodes(in_rel, 1, false, false); 1230 } else { 1231 this->collect_nodes_in_all_data(in_rel, false); 1232 } 1233 this->collect_nodes(out_rel, -2, false, false); 1234 } 1235 #endif 1236 1237 const RegMask &SafePointNode::in_RegMask(uint idx) const { 1238 if( idx < TypeFunc::Parms ) return RegMask::Empty; 1239 // Values outside the domain represent debug info 1240 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]); 1241 } 1242 const RegMask &SafePointNode::out_RegMask() const { 1243 return RegMask::Empty; 1244 } 1245 1246 1247 void SafePointNode::grow_stack(JVMState* jvms, uint grow_by) { 1248 assert((int)grow_by > 0, "sanity"); 1249 int monoff = jvms->monoff(); 1250 int scloff = jvms->scloff(); 1251 int endoff = jvms->endoff(); 1252 assert(endoff == (int)req(), "no other states or debug info after me"); 1253 Node* top = Compile::current()->top(); 1254 for (uint i = 0; i < grow_by; i++) { 1255 ins_req(monoff, top); 1256 } 1257 jvms->set_monoff(monoff + grow_by); 1258 jvms->set_scloff(scloff + grow_by); 1259 jvms->set_endoff(endoff + grow_by); 1260 } 1261 1262 void SafePointNode::push_monitor(const FastLockNode *lock) { 1263 // Add a LockNode, which points to both the original BoxLockNode (the 1264 // stack space for the monitor) and the Object being locked. 1265 const int MonitorEdges = 2; 1266 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges"); 1267 assert(req() == jvms()->endoff(), "correct sizing"); 1268 int nextmon = jvms()->scloff(); 1269 if (GenerateSynchronizationCode) { 1270 ins_req(nextmon, lock->box_node()); 1271 ins_req(nextmon+1, lock->obj_node()); 1272 } else { 1273 Node* top = Compile::current()->top(); 1274 ins_req(nextmon, top); 1275 ins_req(nextmon, top); 1276 } 1277 jvms()->set_scloff(nextmon + MonitorEdges); 1278 jvms()->set_endoff(req()); 1279 } 1280 1281 void SafePointNode::pop_monitor() { 1282 // Delete last monitor from debug info 1283 debug_only(int num_before_pop = jvms()->nof_monitors()); 1284 const int MonitorEdges = 2; 1285 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges"); 1286 int scloff = jvms()->scloff(); 1287 int endoff = jvms()->endoff(); 1288 int new_scloff = scloff - MonitorEdges; 1289 int new_endoff = endoff - MonitorEdges; 1290 jvms()->set_scloff(new_scloff); 1291 jvms()->set_endoff(new_endoff); 1292 while (scloff > new_scloff) del_req_ordered(--scloff); 1293 assert(jvms()->nof_monitors() == num_before_pop-1, ""); 1294 } 1295 1296 Node *SafePointNode::peek_monitor_box() const { 1297 int mon = jvms()->nof_monitors() - 1; 1298 assert(mon >= 0, "must have a monitor"); 1299 return monitor_box(jvms(), mon); 1300 } 1301 1302 Node *SafePointNode::peek_monitor_obj() const { 1303 int mon = jvms()->nof_monitors() - 1; 1304 assert(mon >= 0, "must have a monitor"); 1305 return monitor_obj(jvms(), mon); 1306 } 1307 1308 // Do we Match on this edge index or not? Match no edges 1309 uint SafePointNode::match_edge(uint idx) const { 1310 return (TypeFunc::Parms == idx); 1311 } 1312 1313 void SafePointNode::disconnect_from_root(PhaseIterGVN *igvn) { 1314 assert(Opcode() == Op_SafePoint, "only value for safepoint in loops"); 1315 int nb = igvn->C->root()->find_prec_edge(this); 1316 if (nb != -1) { 1317 igvn->C->root()->rm_prec(nb); 1318 } 1319 } 1320 1321 //============== SafePointScalarObjectNode ============== 1322 1323 SafePointScalarObjectNode::SafePointScalarObjectNode(const TypeOopPtr* tp, 1324 #ifdef ASSERT 1325 AllocateNode* alloc, 1326 #endif 1327 uint first_index, 1328 uint n_fields) : 1329 TypeNode(tp, 1), // 1 control input -- seems required. Get from root. 1330 _first_index(first_index), 1331 _n_fields(n_fields), 1332 _is_stack_allocated(false) 1333 #ifdef ASSERT 1334 , _alloc(alloc) 1335 #endif 1336 { 1337 init_class_id(Class_SafePointScalarObject); 1338 } 1339 1340 // Do not allow value-numbering for SafePointScalarObject node. 1341 uint SafePointScalarObjectNode::hash() const { return NO_HASH; } 1342 bool SafePointScalarObjectNode::cmp( const Node &n ) const { 1343 return (&n == this); // Always fail except on self 1344 } 1345 1346 uint SafePointScalarObjectNode::ideal_reg() const { 1347 return 0; // No matching to machine instruction 1348 } 1349 1350 const RegMask &SafePointScalarObjectNode::in_RegMask(uint idx) const { 1351 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]); 1352 } 1353 1354 const RegMask &SafePointScalarObjectNode::out_RegMask() const { 1355 return RegMask::Empty; 1356 } 1357 1358 uint SafePointScalarObjectNode::match_edge(uint idx) const { 1359 return 0; 1360 } 1361 1362 SafePointScalarObjectNode* 1363 SafePointScalarObjectNode::clone(Dict* sosn_map) const { 1364 void* cached = (*sosn_map)[(void*)this]; 1365 if (cached != NULL) { 1366 return (SafePointScalarObjectNode*)cached; 1367 } 1368 SafePointScalarObjectNode* res = (SafePointScalarObjectNode*)Node::clone(); 1369 sosn_map->Insert((void*)this, (void*)res); 1370 return res; 1371 } 1372 1373 1374 #ifndef PRODUCT 1375 void SafePointScalarObjectNode::dump_spec(outputStream *st) const { 1376 st->print(" # fields@[%d..%d]", first_index(), 1377 first_index() + n_fields() - 1); 1378 } 1379 1380 #endif 1381 1382 //============================================================================= 1383 uint AllocateNode::size_of() const { return sizeof(*this); } 1384 1385 AllocateNode::AllocateNode(Compile* C, const TypeFunc *atype, 1386 Node *ctrl, Node *mem, Node *abio, 1387 Node *size, Node *klass_node, Node *initial_test) 1388 : CallNode(atype, NULL, TypeRawPtr::BOTTOM) 1389 { 1390 init_class_id(Class_Allocate); 1391 init_flags(Flag_is_macro); 1392 _is_scalar_replaceable = false; 1393 _is_non_escaping = false; 1394 _is_stack_allocateable = false; 1395 _is_referenced_stack_allocation = false; 1396 _is_allocation_MemBar_redundant = false; 1397 Node *topnode = C->top(); 1398 1399 init_req( TypeFunc::Control , ctrl ); 1400 init_req( TypeFunc::I_O , abio ); 1401 init_req( TypeFunc::Memory , mem ); 1402 init_req( TypeFunc::ReturnAdr, topnode ); 1403 init_req( TypeFunc::FramePtr , topnode ); 1404 init_req( AllocSize , size); 1405 init_req( KlassNode , klass_node); 1406 init_req( InitialTest , initial_test); 1407 init_req( ALength , topnode); 1408 C->add_macro_node(this); 1409 } 1410 1411 void AllocateNode::compute_MemBar_redundancy(ciMethod* initializer) 1412 { 1413 assert(initializer != NULL && 1414 initializer->is_initializer() && 1415 !initializer->is_static(), 1416 "unexpected initializer method"); 1417 BCEscapeAnalyzer* analyzer = initializer->get_bcea(); 1418 if (analyzer == NULL) { 1419 return; 1420 } 1421 1422 // Allocation node is first parameter in its initializer 1423 if (analyzer->is_arg_stack(0) || analyzer->is_arg_local(0)) { 1424 _is_allocation_MemBar_redundant = true; 1425 } 1426 } 1427 Node *AllocateNode::make_ideal_mark(PhaseGVN *phase, Node* obj, Node* control, Node* mem) { 1428 Node* mark_node = NULL; 1429 // For now only enable fast locking for non-array types 1430 if (UseBiasedLocking && Opcode() == Op_Allocate) { 1431 Node* klass_node = in(AllocateNode::KlassNode); 1432 Node* proto_adr = phase->transform(new AddPNode(klass_node, klass_node, phase->MakeConX(in_bytes(Klass::prototype_header_offset())))); 1433 mark_node = LoadNode::make(*phase, control, mem, proto_adr, TypeRawPtr::BOTTOM, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 1434 } else { 1435 mark_node = phase->MakeConX(markWord::prototype().value()); 1436 } 1437 return mark_node; 1438 } 1439 1440 //============================================================================= 1441 Node* AllocateArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1442 if (remove_dead_region(phase, can_reshape)) return this; 1443 // Don't bother trying to transform a dead node 1444 if (in(0) && in(0)->is_top()) return NULL; 1445 1446 const Type* type = phase->type(Ideal_length()); 1447 if (type->isa_int() && type->is_int()->_hi < 0) { 1448 if (can_reshape) { 1449 PhaseIterGVN *igvn = phase->is_IterGVN(); 1450 // Unreachable fall through path (negative array length), 1451 // the allocation can only throw so disconnect it. 1452 Node* proj = proj_out_or_null(TypeFunc::Control); 1453 Node* catchproj = NULL; 1454 if (proj != NULL) { 1455 for (DUIterator_Fast imax, i = proj->fast_outs(imax); i < imax; i++) { 1456 Node *cn = proj->fast_out(i); 1457 if (cn->is_Catch()) { 1458 catchproj = cn->as_Multi()->proj_out_or_null(CatchProjNode::fall_through_index); 1459 break; 1460 } 1461 } 1462 } 1463 if (catchproj != NULL && catchproj->outcnt() > 0 && 1464 (catchproj->outcnt() > 1 || 1465 catchproj->unique_out()->Opcode() != Op_Halt)) { 1466 assert(catchproj->is_CatchProj(), "must be a CatchProjNode"); 1467 Node* nproj = catchproj->clone(); 1468 igvn->register_new_node_with_optimizer(nproj); 1469 1470 Node *frame = new ParmNode( phase->C->start(), TypeFunc::FramePtr ); 1471 frame = phase->transform(frame); 1472 // Halt & Catch Fire 1473 Node* halt = new HaltNode(nproj, frame, "unexpected negative array length"); 1474 phase->C->root()->add_req(halt); 1475 phase->transform(halt); 1476 1477 igvn->replace_node(catchproj, phase->C->top()); 1478 return this; 1479 } 1480 } else { 1481 // Can't correct it during regular GVN so register for IGVN 1482 phase->C->record_for_igvn(this); 1483 } 1484 } 1485 return NULL; 1486 } 1487 1488 // Retrieve the length from the AllocateArrayNode. Narrow the type with a 1489 // CastII, if appropriate. If we are not allowed to create new nodes, and 1490 // a CastII is appropriate, return NULL. 1491 Node *AllocateArrayNode::make_ideal_length(const TypeOopPtr* oop_type, PhaseTransform *phase, bool allow_new_nodes) { 1492 Node *length = in(AllocateNode::ALength); 1493 assert(length != NULL, "length is not null"); 1494 1495 const TypeInt* length_type = phase->find_int_type(length); 1496 const TypeAryPtr* ary_type = oop_type->isa_aryptr(); 1497 1498 if (ary_type != NULL && length_type != NULL) { 1499 const TypeInt* narrow_length_type = ary_type->narrow_size_type(length_type); 1500 if (narrow_length_type != length_type) { 1501 // Assert one of: 1502 // - the narrow_length is 0 1503 // - the narrow_length is not wider than length 1504 assert(narrow_length_type == TypeInt::ZERO || 1505 length_type->is_con() && narrow_length_type->is_con() && 1506 (narrow_length_type->_hi <= length_type->_lo) || 1507 (narrow_length_type->_hi <= length_type->_hi && 1508 narrow_length_type->_lo >= length_type->_lo), 1509 "narrow type must be narrower than length type"); 1510 1511 // Return NULL if new nodes are not allowed 1512 if (!allow_new_nodes) return NULL; 1513 // Create a cast which is control dependent on the initialization to 1514 // propagate the fact that the array length must be positive. 1515 InitializeNode* init = initialization(); 1516 assert(init != NULL, "initialization not found"); 1517 length = new CastIINode(length, narrow_length_type); 1518 length->set_req(0, init->proj_out_or_null(0)); 1519 } 1520 } 1521 1522 return length; 1523 } 1524 1525 //============================================================================= 1526 uint LockNode::size_of() const { return sizeof(*this); } 1527 1528 // Redundant lock elimination 1529 // 1530 // There are various patterns of locking where we release and 1531 // immediately reacquire a lock in a piece of code where no operations 1532 // occur in between that would be observable. In those cases we can 1533 // skip releasing and reacquiring the lock without violating any 1534 // fairness requirements. Doing this around a loop could cause a lock 1535 // to be held for a very long time so we concentrate on non-looping 1536 // control flow. We also require that the operations are fully 1537 // redundant meaning that we don't introduce new lock operations on 1538 // some paths so to be able to eliminate it on others ala PRE. This 1539 // would probably require some more extensive graph manipulation to 1540 // guarantee that the memory edges were all handled correctly. 1541 // 1542 // Assuming p is a simple predicate which can't trap in any way and s 1543 // is a synchronized method consider this code: 1544 // 1545 // s(); 1546 // if (p) 1547 // s(); 1548 // else 1549 // s(); 1550 // s(); 1551 // 1552 // 1. The unlocks of the first call to s can be eliminated if the 1553 // locks inside the then and else branches are eliminated. 1554 // 1555 // 2. The unlocks of the then and else branches can be eliminated if 1556 // the lock of the final call to s is eliminated. 1557 // 1558 // Either of these cases subsumes the simple case of sequential control flow 1559 // 1560 // Addtionally we can eliminate versions without the else case: 1561 // 1562 // s(); 1563 // if (p) 1564 // s(); 1565 // s(); 1566 // 1567 // 3. In this case we eliminate the unlock of the first s, the lock 1568 // and unlock in the then case and the lock in the final s. 1569 // 1570 // Note also that in all these cases the then/else pieces don't have 1571 // to be trivial as long as they begin and end with synchronization 1572 // operations. 1573 // 1574 // s(); 1575 // if (p) 1576 // s(); 1577 // f(); 1578 // s(); 1579 // s(); 1580 // 1581 // The code will work properly for this case, leaving in the unlock 1582 // before the call to f and the relock after it. 1583 // 1584 // A potentially interesting case which isn't handled here is when the 1585 // locking is partially redundant. 1586 // 1587 // s(); 1588 // if (p) 1589 // s(); 1590 // 1591 // This could be eliminated putting unlocking on the else case and 1592 // eliminating the first unlock and the lock in the then side. 1593 // Alternatively the unlock could be moved out of the then side so it 1594 // was after the merge and the first unlock and second lock 1595 // eliminated. This might require less manipulation of the memory 1596 // state to get correct. 1597 // 1598 // Additionally we might allow work between a unlock and lock before 1599 // giving up eliminating the locks. The current code disallows any 1600 // conditional control flow between these operations. A formulation 1601 // similar to partial redundancy elimination computing the 1602 // availability of unlocking and the anticipatability of locking at a 1603 // program point would allow detection of fully redundant locking with 1604 // some amount of work in between. I'm not sure how often I really 1605 // think that would occur though. Most of the cases I've seen 1606 // indicate it's likely non-trivial work would occur in between. 1607 // There may be other more complicated constructs where we could 1608 // eliminate locking but I haven't seen any others appear as hot or 1609 // interesting. 1610 // 1611 // Locking and unlocking have a canonical form in ideal that looks 1612 // roughly like this: 1613 // 1614 // <obj> 1615 // | \\------+ 1616 // | \ \ 1617 // | BoxLock \ 1618 // | | | \ 1619 // | | \ \ 1620 // | | FastLock 1621 // | | / 1622 // | | / 1623 // | | | 1624 // 1625 // Lock 1626 // | 1627 // Proj #0 1628 // | 1629 // MembarAcquire 1630 // | 1631 // Proj #0 1632 // 1633 // MembarRelease 1634 // | 1635 // Proj #0 1636 // | 1637 // Unlock 1638 // | 1639 // Proj #0 1640 // 1641 // 1642 // This code proceeds by processing Lock nodes during PhaseIterGVN 1643 // and searching back through its control for the proper code 1644 // patterns. Once it finds a set of lock and unlock operations to 1645 // eliminate they are marked as eliminatable which causes the 1646 // expansion of the Lock and Unlock macro nodes to make the operation a NOP 1647 // 1648 //============================================================================= 1649 1650 // 1651 // Utility function to skip over uninteresting control nodes. Nodes skipped are: 1652 // - copy regions. (These may not have been optimized away yet.) 1653 // - eliminated locking nodes 1654 // 1655 static Node *next_control(Node *ctrl) { 1656 if (ctrl == NULL) 1657 return NULL; 1658 while (1) { 1659 if (ctrl->is_Region()) { 1660 RegionNode *r = ctrl->as_Region(); 1661 Node *n = r->is_copy(); 1662 if (n == NULL) 1663 break; // hit a region, return it 1664 else 1665 ctrl = n; 1666 } else if (ctrl->is_Proj()) { 1667 Node *in0 = ctrl->in(0); 1668 if (in0->is_AbstractLock() && in0->as_AbstractLock()->is_eliminated()) { 1669 ctrl = in0->in(0); 1670 } else { 1671 break; 1672 } 1673 } else { 1674 break; // found an interesting control 1675 } 1676 } 1677 return ctrl; 1678 } 1679 // 1680 // Given a control, see if it's the control projection of an Unlock which 1681 // operating on the same object as lock. 1682 // 1683 bool AbstractLockNode::find_matching_unlock(const Node* ctrl, LockNode* lock, 1684 GrowableArray<AbstractLockNode*> &lock_ops) { 1685 ProjNode *ctrl_proj = (ctrl->is_Proj()) ? ctrl->as_Proj() : NULL; 1686 if (ctrl_proj != NULL && ctrl_proj->_con == TypeFunc::Control) { 1687 Node *n = ctrl_proj->in(0); 1688 if (n != NULL && n->is_Unlock()) { 1689 UnlockNode *unlock = n->as_Unlock(); 1690 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1691 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node()); 1692 Node* unlock_obj = bs->step_over_gc_barrier(unlock->obj_node()); 1693 if (lock_obj->eqv_uncast(unlock_obj) && 1694 BoxLockNode::same_slot(lock->box_node(), unlock->box_node()) && 1695 !unlock->is_eliminated()) { 1696 lock_ops.append(unlock); 1697 return true; 1698 } 1699 } 1700 } 1701 return false; 1702 } 1703 1704 // 1705 // Find the lock matching an unlock. Returns null if a safepoint 1706 // or complicated control is encountered first. 1707 LockNode *AbstractLockNode::find_matching_lock(UnlockNode* unlock) { 1708 LockNode *lock_result = NULL; 1709 // find the matching lock, or an intervening safepoint 1710 Node *ctrl = next_control(unlock->in(0)); 1711 while (1) { 1712 assert(ctrl != NULL, "invalid control graph"); 1713 assert(!ctrl->is_Start(), "missing lock for unlock"); 1714 if (ctrl->is_top()) break; // dead control path 1715 if (ctrl->is_Proj()) ctrl = ctrl->in(0); 1716 if (ctrl->is_SafePoint()) { 1717 break; // found a safepoint (may be the lock we are searching for) 1718 } else if (ctrl->is_Region()) { 1719 // Check for a simple diamond pattern. Punt on anything more complicated 1720 if (ctrl->req() == 3 && ctrl->in(1) != NULL && ctrl->in(2) != NULL) { 1721 Node *in1 = next_control(ctrl->in(1)); 1722 Node *in2 = next_control(ctrl->in(2)); 1723 if (((in1->is_IfTrue() && in2->is_IfFalse()) || 1724 (in2->is_IfTrue() && in1->is_IfFalse())) && (in1->in(0) == in2->in(0))) { 1725 ctrl = next_control(in1->in(0)->in(0)); 1726 } else { 1727 break; 1728 } 1729 } else { 1730 break; 1731 } 1732 } else { 1733 ctrl = next_control(ctrl->in(0)); // keep searching 1734 } 1735 } 1736 if (ctrl->is_Lock()) { 1737 LockNode *lock = ctrl->as_Lock(); 1738 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1739 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node()); 1740 Node* unlock_obj = bs->step_over_gc_barrier(unlock->obj_node()); 1741 if (lock_obj->eqv_uncast(unlock_obj) && 1742 BoxLockNode::same_slot(lock->box_node(), unlock->box_node())) { 1743 lock_result = lock; 1744 } 1745 } 1746 return lock_result; 1747 } 1748 1749 // This code corresponds to case 3 above. 1750 1751 bool AbstractLockNode::find_lock_and_unlock_through_if(Node* node, LockNode* lock, 1752 GrowableArray<AbstractLockNode*> &lock_ops) { 1753 Node* if_node = node->in(0); 1754 bool if_true = node->is_IfTrue(); 1755 1756 if (if_node->is_If() && if_node->outcnt() == 2 && (if_true || node->is_IfFalse())) { 1757 Node *lock_ctrl = next_control(if_node->in(0)); 1758 if (find_matching_unlock(lock_ctrl, lock, lock_ops)) { 1759 Node* lock1_node = NULL; 1760 ProjNode* proj = if_node->as_If()->proj_out(!if_true); 1761 if (if_true) { 1762 if (proj->is_IfFalse() && proj->outcnt() == 1) { 1763 lock1_node = proj->unique_out(); 1764 } 1765 } else { 1766 if (proj->is_IfTrue() && proj->outcnt() == 1) { 1767 lock1_node = proj->unique_out(); 1768 } 1769 } 1770 if (lock1_node != NULL && lock1_node->is_Lock()) { 1771 LockNode *lock1 = lock1_node->as_Lock(); 1772 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1773 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node()); 1774 Node* lock1_obj = bs->step_over_gc_barrier(lock1->obj_node()); 1775 if (lock_obj->eqv_uncast(lock1_obj) && 1776 BoxLockNode::same_slot(lock->box_node(), lock1->box_node()) && 1777 !lock1->is_eliminated()) { 1778 lock_ops.append(lock1); 1779 return true; 1780 } 1781 } 1782 } 1783 } 1784 1785 lock_ops.trunc_to(0); 1786 return false; 1787 } 1788 1789 bool AbstractLockNode::find_unlocks_for_region(const RegionNode* region, LockNode* lock, 1790 GrowableArray<AbstractLockNode*> &lock_ops) { 1791 // check each control merging at this point for a matching unlock. 1792 // in(0) should be self edge so skip it. 1793 for (int i = 1; i < (int)region->req(); i++) { 1794 Node *in_node = next_control(region->in(i)); 1795 if (in_node != NULL) { 1796 if (find_matching_unlock(in_node, lock, lock_ops)) { 1797 // found a match so keep on checking. 1798 continue; 1799 } else if (find_lock_and_unlock_through_if(in_node, lock, lock_ops)) { 1800 continue; 1801 } 1802 1803 // If we fall through to here then it was some kind of node we 1804 // don't understand or there wasn't a matching unlock, so give 1805 // up trying to merge locks. 1806 lock_ops.trunc_to(0); 1807 return false; 1808 } 1809 } 1810 return true; 1811 1812 } 1813 1814 #ifndef PRODUCT 1815 // 1816 // Create a counter which counts the number of times this lock is acquired 1817 // 1818 void AbstractLockNode::create_lock_counter(JVMState* state) { 1819 _counter = OptoRuntime::new_named_counter(state, NamedCounter::LockCounter); 1820 } 1821 1822 void AbstractLockNode::set_eliminated_lock_counter() { 1823 if (_counter) { 1824 // Update the counter to indicate that this lock was eliminated. 1825 // The counter update code will stay around even though the 1826 // optimizer will eliminate the lock operation itself. 1827 _counter->set_tag(NamedCounter::EliminatedLockCounter); 1828 } 1829 } 1830 1831 const char* AbstractLockNode::_kind_names[] = {"Regular", "NonEscObj", "Coarsened", "Nested"}; 1832 1833 void AbstractLockNode::dump_spec(outputStream* st) const { 1834 st->print("%s ", _kind_names[_kind]); 1835 CallNode::dump_spec(st); 1836 } 1837 1838 void AbstractLockNode::dump_compact_spec(outputStream* st) const { 1839 st->print("%s", _kind_names[_kind]); 1840 } 1841 1842 // The related set of lock nodes includes the control boundary. 1843 void AbstractLockNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const { 1844 if (compact) { 1845 this->collect_nodes(in_rel, 1, false, false); 1846 } else { 1847 this->collect_nodes_in_all_data(in_rel, true); 1848 } 1849 this->collect_nodes(out_rel, -2, false, false); 1850 } 1851 #endif 1852 1853 //============================================================================= 1854 Node *LockNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1855 1856 // perform any generic optimizations first (returns 'this' or NULL) 1857 Node *result = SafePointNode::Ideal(phase, can_reshape); 1858 if (result != NULL) return result; 1859 // Don't bother trying to transform a dead node 1860 if (in(0) && in(0)->is_top()) return NULL; 1861 1862 // Now see if we can optimize away this lock. We don't actually 1863 // remove the locking here, we simply set the _eliminate flag which 1864 // prevents macro expansion from expanding the lock. Since we don't 1865 // modify the graph, the value returned from this function is the 1866 // one computed above. 1867 if (can_reshape && EliminateLocks && !is_non_esc_obj()) { 1868 // 1869 // If we are locking an unescaped object, the lock/unlock is unnecessary 1870 // 1871 ConnectionGraph *cgr = phase->C->congraph(); 1872 if (cgr != NULL && cgr->not_global_escape(obj_node())) { 1873 assert(!is_eliminated() || is_coarsened(), "sanity"); 1874 // The lock could be marked eliminated by lock coarsening 1875 // code during first IGVN before EA. Replace coarsened flag 1876 // to eliminate all associated locks/unlocks. 1877 #ifdef ASSERT 1878 this->log_lock_optimization(phase->C,"eliminate_lock_set_non_esc1"); 1879 #endif 1880 this->set_non_esc_obj(); 1881 return result; 1882 } 1883 1884 // 1885 // Try lock coarsening 1886 // 1887 PhaseIterGVN* iter = phase->is_IterGVN(); 1888 if (iter != NULL && !is_eliminated()) { 1889 1890 GrowableArray<AbstractLockNode*> lock_ops; 1891 1892 Node *ctrl = next_control(in(0)); 1893 1894 // now search back for a matching Unlock 1895 if (find_matching_unlock(ctrl, this, lock_ops)) { 1896 // found an unlock directly preceding this lock. This is the 1897 // case of single unlock directly control dependent on a 1898 // single lock which is the trivial version of case 1 or 2. 1899 } else if (ctrl->is_Region() ) { 1900 if (find_unlocks_for_region(ctrl->as_Region(), this, lock_ops)) { 1901 // found lock preceded by multiple unlocks along all paths 1902 // joining at this point which is case 3 in description above. 1903 } 1904 } else { 1905 // see if this lock comes from either half of an if and the 1906 // predecessors merges unlocks and the other half of the if 1907 // performs a lock. 1908 if (find_lock_and_unlock_through_if(ctrl, this, lock_ops)) { 1909 // found unlock splitting to an if with locks on both branches. 1910 } 1911 } 1912 1913 if (lock_ops.length() > 0) { 1914 // add ourselves to the list of locks to be eliminated. 1915 lock_ops.append(this); 1916 1917 #ifndef PRODUCT 1918 if (PrintEliminateLocks) { 1919 int locks = 0; 1920 int unlocks = 0; 1921 for (int i = 0; i < lock_ops.length(); i++) { 1922 AbstractLockNode* lock = lock_ops.at(i); 1923 if (lock->Opcode() == Op_Lock) 1924 locks++; 1925 else 1926 unlocks++; 1927 if (Verbose) { 1928 lock->dump(1); 1929 } 1930 } 1931 tty->print_cr("***Eliminated %d unlocks and %d locks", unlocks, locks); 1932 } 1933 #endif 1934 1935 // for each of the identified locks, mark them 1936 // as eliminatable 1937 for (int i = 0; i < lock_ops.length(); i++) { 1938 AbstractLockNode* lock = lock_ops.at(i); 1939 1940 // Mark it eliminated by coarsening and update any counters 1941 #ifdef ASSERT 1942 lock->log_lock_optimization(phase->C, "eliminate_lock_set_coarsened"); 1943 #endif 1944 lock->set_coarsened(); 1945 } 1946 } else if (ctrl->is_Region() && 1947 iter->_worklist.member(ctrl)) { 1948 // We weren't able to find any opportunities but the region this 1949 // lock is control dependent on hasn't been processed yet so put 1950 // this lock back on the worklist so we can check again once any 1951 // region simplification has occurred. 1952 iter->_worklist.push(this); 1953 } 1954 } 1955 } 1956 1957 return result; 1958 } 1959 1960 //============================================================================= 1961 bool LockNode::is_nested_lock_region() { 1962 return is_nested_lock_region(NULL); 1963 } 1964 1965 // p is used for access to compilation log; no logging if NULL 1966 bool LockNode::is_nested_lock_region(Compile * c) { 1967 BoxLockNode* box = box_node()->as_BoxLock(); 1968 int stk_slot = box->stack_slot(); 1969 if (stk_slot <= 0) { 1970 #ifdef ASSERT 1971 this->log_lock_optimization(c, "eliminate_lock_INLR_1"); 1972 #endif 1973 return false; // External lock or it is not Box (Phi node). 1974 } 1975 1976 // Ignore complex cases: merged locks or multiple locks. 1977 Node* obj = obj_node(); 1978 LockNode* unique_lock = NULL; 1979 if (!box->is_simple_lock_region(&unique_lock, obj)) { 1980 #ifdef ASSERT 1981 this->log_lock_optimization(c, "eliminate_lock_INLR_2a"); 1982 #endif 1983 return false; 1984 } 1985 if (unique_lock != this) { 1986 #ifdef ASSERT 1987 this->log_lock_optimization(c, "eliminate_lock_INLR_2b"); 1988 #endif 1989 return false; 1990 } 1991 1992 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1993 obj = bs->step_over_gc_barrier(obj); 1994 // Look for external lock for the same object. 1995 SafePointNode* sfn = this->as_SafePoint(); 1996 JVMState* youngest_jvms = sfn->jvms(); 1997 int max_depth = youngest_jvms->depth(); 1998 for (int depth = 1; depth <= max_depth; depth++) { 1999 JVMState* jvms = youngest_jvms->of_depth(depth); 2000 int num_mon = jvms->nof_monitors(); 2001 // Loop over monitors 2002 for (int idx = 0; idx < num_mon; idx++) { 2003 Node* obj_node = sfn->monitor_obj(jvms, idx); 2004 obj_node = bs->step_over_gc_barrier(obj_node); 2005 BoxLockNode* box_node = sfn->monitor_box(jvms, idx)->as_BoxLock(); 2006 if ((box_node->stack_slot() < stk_slot) && obj_node->eqv_uncast(obj)) { 2007 return true; 2008 } 2009 } 2010 } 2011 #ifdef ASSERT 2012 this->log_lock_optimization(c, "eliminate_lock_INLR_3"); 2013 #endif 2014 return false; 2015 } 2016 2017 //============================================================================= 2018 uint UnlockNode::size_of() const { return sizeof(*this); } 2019 2020 //============================================================================= 2021 Node *UnlockNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2022 2023 // perform any generic optimizations first (returns 'this' or NULL) 2024 Node *result = SafePointNode::Ideal(phase, can_reshape); 2025 if (result != NULL) return result; 2026 // Don't bother trying to transform a dead node 2027 if (in(0) && in(0)->is_top()) return NULL; 2028 2029 // Now see if we can optimize away this unlock. We don't actually 2030 // remove the unlocking here, we simply set the _eliminate flag which 2031 // prevents macro expansion from expanding the unlock. Since we don't 2032 // modify the graph, the value returned from this function is the 2033 // one computed above. 2034 // Escape state is defined after Parse phase. 2035 if (can_reshape && EliminateLocks && !is_non_esc_obj()) { 2036 // 2037 // If we are unlocking an unescaped object, the lock/unlock is unnecessary. 2038 // 2039 ConnectionGraph *cgr = phase->C->congraph(); 2040 if (cgr != NULL && cgr->not_global_escape(obj_node())) { 2041 assert(!is_eliminated() || is_coarsened(), "sanity"); 2042 // The lock could be marked eliminated by lock coarsening 2043 // code during first IGVN before EA. Replace coarsened flag 2044 // to eliminate all associated locks/unlocks. 2045 #ifdef ASSERT 2046 this->log_lock_optimization(phase->C, "eliminate_lock_set_non_esc2"); 2047 #endif 2048 this->set_non_esc_obj(); 2049 } 2050 } 2051 return result; 2052 } 2053 2054 const char * AbstractLockNode::kind_as_string() const { 2055 return is_coarsened() ? "coarsened" : 2056 is_nested() ? "nested" : 2057 is_non_esc_obj() ? "non_escaping" : 2058 "?"; 2059 } 2060 2061 void AbstractLockNode::log_lock_optimization(Compile *C, const char * tag) const { 2062 if (C == NULL) { 2063 return; 2064 } 2065 CompileLog* log = C->log(); 2066 if (log != NULL) { 2067 log->begin_head("%s lock='%d' compile_id='%d' class_id='%s' kind='%s'", 2068 tag, is_Lock(), C->compile_id(), 2069 is_Unlock() ? "unlock" : is_Lock() ? "lock" : "?", 2070 kind_as_string()); 2071 log->stamp(); 2072 log->end_head(); 2073 JVMState* p = is_Unlock() ? (as_Unlock()->dbg_jvms()) : jvms(); 2074 while (p != NULL) { 2075 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 2076 p = p->caller(); 2077 } 2078 log->tail(tag); 2079 } 2080 } 2081 2082 bool CallNode::may_modify_arraycopy_helper(const TypeOopPtr* dest_t, const TypeOopPtr *t_oop, PhaseTransform *phase) { 2083 if (dest_t->is_known_instance() && t_oop->is_known_instance()) { 2084 return dest_t->instance_id() == t_oop->instance_id(); 2085 } 2086 2087 if (dest_t->isa_instptr() && !dest_t->klass()->equals(phase->C->env()->Object_klass())) { 2088 // clone 2089 if (t_oop->isa_aryptr()) { 2090 return false; 2091 } 2092 if (!t_oop->isa_instptr()) { 2093 return true; 2094 } 2095 if (dest_t->klass()->is_subtype_of(t_oop->klass()) || t_oop->klass()->is_subtype_of(dest_t->klass())) { 2096 return true; 2097 } 2098 // unrelated 2099 return false; 2100 } 2101 2102 if (dest_t->isa_aryptr()) { 2103 // arraycopy or array clone 2104 if (t_oop->isa_instptr()) { 2105 return false; 2106 } 2107 if (!t_oop->isa_aryptr()) { 2108 return true; 2109 } 2110 2111 const Type* elem = dest_t->is_aryptr()->elem(); 2112 if (elem == Type::BOTTOM) { 2113 // An array but we don't know what elements are 2114 return true; 2115 } 2116 2117 dest_t = dest_t->add_offset(Type::OffsetBot)->is_oopptr(); 2118 uint dest_alias = phase->C->get_alias_index(dest_t); 2119 uint t_oop_alias = phase->C->get_alias_index(t_oop); 2120 2121 return dest_alias == t_oop_alias; 2122 } 2123 2124 return true; 2125 }