1 /* 2 * Copyright (c) 1997, 2019, 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 "opto/valuetypenode.hpp" 46 #include "runtime/sharedRuntime.hpp" 47 48 // Portions of code courtesy of Clifford Click 49 50 // Optimization - Graph Style 51 52 //============================================================================= 53 uint StartNode::size_of() const { return sizeof(*this); } 54 bool StartNode::cmp( const Node &n ) const 55 { return _domain == ((StartNode&)n)._domain; } 56 const Type *StartNode::bottom_type() const { return _domain; } 57 const Type* StartNode::Value(PhaseGVN* phase) const { return _domain; } 58 #ifndef PRODUCT 59 void StartNode::dump_spec(outputStream *st) const { st->print(" #"); _domain->dump_on(st);} 60 void StartNode::dump_compact_spec(outputStream *st) const { /* empty */ } 61 #endif 62 63 //------------------------------Ideal------------------------------------------ 64 Node *StartNode::Ideal(PhaseGVN *phase, bool can_reshape){ 65 return remove_dead_region(phase, can_reshape) ? this : NULL; 66 } 67 68 //------------------------------calling_convention----------------------------- 69 void StartNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const { 70 Matcher::calling_convention( sig_bt, parm_regs, argcnt, false ); 71 } 72 73 //------------------------------Registers-------------------------------------- 74 const RegMask &StartNode::in_RegMask(uint) const { 75 return RegMask::Empty; 76 } 77 78 //------------------------------match------------------------------------------ 79 // Construct projections for incoming parameters, and their RegMask info 80 Node *StartNode::match(const ProjNode *proj, const Matcher *match, const RegMask* mask) { 81 switch (proj->_con) { 82 case TypeFunc::Control: 83 case TypeFunc::I_O: 84 case TypeFunc::Memory: 85 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 86 case TypeFunc::FramePtr: 87 return new MachProjNode(this,proj->_con,Matcher::c_frame_ptr_mask, Op_RegP); 88 case TypeFunc::ReturnAdr: 89 return new MachProjNode(this,proj->_con,match->_return_addr_mask,Op_RegP); 90 case TypeFunc::Parms: 91 default: { 92 uint parm_num = proj->_con - TypeFunc::Parms; 93 const Type *t = _domain->field_at(proj->_con); 94 if (t->base() == Type::Half) // 2nd half of Longs and Doubles 95 return new ConNode(Type::TOP); 96 uint ideal_reg = t->ideal_reg(); 97 RegMask &rm = match->_calling_convention_mask[parm_num]; 98 return new MachProjNode(this,proj->_con,rm,ideal_reg); 99 } 100 } 101 return NULL; 102 } 103 104 //------------------------------StartOSRNode---------------------------------- 105 // The method start node for an on stack replacement adapter 106 107 //------------------------------osr_domain----------------------------- 108 const TypeTuple *StartOSRNode::osr_domain() { 109 const Type **fields = TypeTuple::fields(2); 110 fields[TypeFunc::Parms+0] = TypeRawPtr::BOTTOM; // address of osr buffer 111 112 return TypeTuple::make(TypeFunc::Parms+1, fields); 113 } 114 115 //============================================================================= 116 const char * const ParmNode::names[TypeFunc::Parms+1] = { 117 "Control", "I_O", "Memory", "FramePtr", "ReturnAdr", "Parms" 118 }; 119 120 #ifndef PRODUCT 121 void ParmNode::dump_spec(outputStream *st) const { 122 if( _con < TypeFunc::Parms ) { 123 st->print("%s", names[_con]); 124 } else { 125 st->print("Parm%d: ",_con-TypeFunc::Parms); 126 // Verbose and WizardMode dump bottom_type for all nodes 127 if( !Verbose && !WizardMode ) bottom_type()->dump_on(st); 128 } 129 } 130 131 void ParmNode::dump_compact_spec(outputStream *st) const { 132 if (_con < TypeFunc::Parms) { 133 st->print("%s", names[_con]); 134 } else { 135 st->print("%d:", _con-TypeFunc::Parms); 136 // unconditionally dump bottom_type 137 bottom_type()->dump_on(st); 138 } 139 } 140 141 // For a ParmNode, all immediate inputs and outputs are considered relevant 142 // both in compact and standard representation. 143 void ParmNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const { 144 this->collect_nodes(in_rel, 1, false, false); 145 this->collect_nodes(out_rel, -1, false, false); 146 } 147 #endif 148 149 uint ParmNode::ideal_reg() const { 150 switch( _con ) { 151 case TypeFunc::Control : // fall through 152 case TypeFunc::I_O : // fall through 153 case TypeFunc::Memory : return 0; 154 case TypeFunc::FramePtr : // fall through 155 case TypeFunc::ReturnAdr: return Op_RegP; 156 default : assert( _con > TypeFunc::Parms, "" ); 157 // fall through 158 case TypeFunc::Parms : { 159 // Type of argument being passed 160 const Type *t = in(0)->as_Start()->_domain->field_at(_con); 161 return t->ideal_reg(); 162 } 163 } 164 ShouldNotReachHere(); 165 return 0; 166 } 167 168 //============================================================================= 169 ReturnNode::ReturnNode(uint edges, Node *cntrl, Node *i_o, Node *memory, Node *frameptr, Node *retadr ) : Node(edges) { 170 init_req(TypeFunc::Control,cntrl); 171 init_req(TypeFunc::I_O,i_o); 172 init_req(TypeFunc::Memory,memory); 173 init_req(TypeFunc::FramePtr,frameptr); 174 init_req(TypeFunc::ReturnAdr,retadr); 175 } 176 177 Node *ReturnNode::Ideal(PhaseGVN *phase, bool can_reshape){ 178 return remove_dead_region(phase, can_reshape) ? this : NULL; 179 } 180 181 const Type* ReturnNode::Value(PhaseGVN* phase) const { 182 return ( phase->type(in(TypeFunc::Control)) == Type::TOP) 183 ? Type::TOP 184 : Type::BOTTOM; 185 } 186 187 // Do we Match on this edge index or not? No edges on return nodes 188 uint ReturnNode::match_edge(uint idx) const { 189 return 0; 190 } 191 192 193 #ifndef PRODUCT 194 void ReturnNode::dump_req(outputStream *st) const { 195 // Dump the required inputs, enclosed in '(' and ')' 196 uint i; // Exit value of loop 197 for (i = 0; i < req(); i++) { // For all required inputs 198 if (i == TypeFunc::Parms) st->print("returns"); 199 if (in(i)) st->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx); 200 else st->print("_ "); 201 } 202 } 203 #endif 204 205 //============================================================================= 206 RethrowNode::RethrowNode( 207 Node* cntrl, 208 Node* i_o, 209 Node* memory, 210 Node* frameptr, 211 Node* ret_adr, 212 Node* exception 213 ) : Node(TypeFunc::Parms + 1) { 214 init_req(TypeFunc::Control , cntrl ); 215 init_req(TypeFunc::I_O , i_o ); 216 init_req(TypeFunc::Memory , memory ); 217 init_req(TypeFunc::FramePtr , frameptr ); 218 init_req(TypeFunc::ReturnAdr, ret_adr); 219 init_req(TypeFunc::Parms , exception); 220 } 221 222 Node *RethrowNode::Ideal(PhaseGVN *phase, bool can_reshape){ 223 return remove_dead_region(phase, can_reshape) ? this : NULL; 224 } 225 226 const Type* RethrowNode::Value(PhaseGVN* phase) const { 227 return (phase->type(in(TypeFunc::Control)) == Type::TOP) 228 ? Type::TOP 229 : Type::BOTTOM; 230 } 231 232 uint RethrowNode::match_edge(uint idx) const { 233 return 0; 234 } 235 236 #ifndef PRODUCT 237 void RethrowNode::dump_req(outputStream *st) const { 238 // Dump the required inputs, enclosed in '(' and ')' 239 uint i; // Exit value of loop 240 for (i = 0; i < req(); i++) { // For all required inputs 241 if (i == TypeFunc::Parms) st->print("exception"); 242 if (in(i)) st->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx); 243 else st->print("_ "); 244 } 245 } 246 #endif 247 248 //============================================================================= 249 // Do we Match on this edge index or not? Match only target address & method 250 uint TailCallNode::match_edge(uint idx) const { 251 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1; 252 } 253 254 //============================================================================= 255 // Do we Match on this edge index or not? Match only target address & oop 256 uint TailJumpNode::match_edge(uint idx) const { 257 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1; 258 } 259 260 //============================================================================= 261 JVMState::JVMState(ciMethod* method, JVMState* caller) : 262 _method(method) { 263 assert(method != NULL, "must be valid call site"); 264 _bci = InvocationEntryBci; 265 _reexecute = Reexecute_Undefined; 266 debug_only(_bci = -99); // random garbage value 267 debug_only(_map = (SafePointNode*)-1); 268 _caller = caller; 269 _depth = 1 + (caller == NULL ? 0 : caller->depth()); 270 _locoff = TypeFunc::Parms; 271 _stkoff = _locoff + _method->max_locals(); 272 _monoff = _stkoff + _method->max_stack(); 273 _scloff = _monoff; 274 _endoff = _monoff; 275 _sp = 0; 276 } 277 JVMState::JVMState(int stack_size) : 278 _method(NULL) { 279 _bci = InvocationEntryBci; 280 _reexecute = Reexecute_Undefined; 281 debug_only(_map = (SafePointNode*)-1); 282 _caller = NULL; 283 _depth = 1; 284 _locoff = TypeFunc::Parms; 285 _stkoff = _locoff; 286 _monoff = _stkoff + stack_size; 287 _scloff = _monoff; 288 _endoff = _monoff; 289 _sp = 0; 290 } 291 292 //--------------------------------of_depth------------------------------------- 293 JVMState* JVMState::of_depth(int d) const { 294 const JVMState* jvmp = this; 295 assert(0 < d && (uint)d <= depth(), "oob"); 296 for (int skip = depth() - d; skip > 0; skip--) { 297 jvmp = jvmp->caller(); 298 } 299 assert(jvmp->depth() == (uint)d, "found the right one"); 300 return (JVMState*)jvmp; 301 } 302 303 //-----------------------------same_calls_as----------------------------------- 304 bool JVMState::same_calls_as(const JVMState* that) const { 305 if (this == that) return true; 306 if (this->depth() != that->depth()) return false; 307 const JVMState* p = this; 308 const JVMState* q = that; 309 for (;;) { 310 if (p->_method != q->_method) return false; 311 if (p->_method == NULL) return true; // bci is irrelevant 312 if (p->_bci != q->_bci) return false; 313 if (p->_reexecute != q->_reexecute) return false; 314 p = p->caller(); 315 q = q->caller(); 316 if (p == q) return true; 317 assert(p != NULL && q != NULL, "depth check ensures we don't run off end"); 318 } 319 } 320 321 //------------------------------debug_start------------------------------------ 322 uint JVMState::debug_start() const { 323 debug_only(JVMState* jvmroot = of_depth(1)); 324 assert(jvmroot->locoff() <= this->locoff(), "youngest JVMState must be last"); 325 return of_depth(1)->locoff(); 326 } 327 328 //-------------------------------debug_end------------------------------------- 329 uint JVMState::debug_end() const { 330 debug_only(JVMState* jvmroot = of_depth(1)); 331 assert(jvmroot->endoff() <= this->endoff(), "youngest JVMState must be last"); 332 return endoff(); 333 } 334 335 //------------------------------debug_depth------------------------------------ 336 uint JVMState::debug_depth() const { 337 uint total = 0; 338 for (const JVMState* jvmp = this; jvmp != NULL; jvmp = jvmp->caller()) { 339 total += jvmp->debug_size(); 340 } 341 return total; 342 } 343 344 #ifndef PRODUCT 345 346 //------------------------------format_helper---------------------------------- 347 // Given an allocation (a Chaitin object) and a Node decide if the Node carries 348 // any defined value or not. If it does, print out the register or constant. 349 static void format_helper( PhaseRegAlloc *regalloc, outputStream* st, Node *n, const char *msg, uint i, GrowableArray<SafePointScalarObjectNode*> *scobjs ) { 350 if (n == NULL) { st->print(" NULL"); return; } 351 if (n->is_SafePointScalarObject()) { 352 // Scalar replacement. 353 SafePointScalarObjectNode* spobj = n->as_SafePointScalarObject(); 354 scobjs->append_if_missing(spobj); 355 int sco_n = scobjs->find(spobj); 356 assert(sco_n >= 0, ""); 357 st->print(" %s%d]=#ScObj" INT32_FORMAT, msg, i, sco_n); 358 return; 359 } 360 if (regalloc->node_regs_max_index() > 0 && 361 OptoReg::is_valid(regalloc->get_reg_first(n))) { // Check for undefined 362 char buf[50]; 363 regalloc->dump_register(n,buf); 364 st->print(" %s%d]=%s",msg,i,buf); 365 } else { // No register, but might be constant 366 const Type *t = n->bottom_type(); 367 switch (t->base()) { 368 case Type::Int: 369 st->print(" %s%d]=#" INT32_FORMAT,msg,i,t->is_int()->get_con()); 370 break; 371 case Type::AnyPtr: 372 assert( t == TypePtr::NULL_PTR || n->in_dump(), "" ); 373 st->print(" %s%d]=#NULL",msg,i); 374 break; 375 case Type::AryPtr: 376 case Type::InstPtr: 377 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->isa_oopptr()->const_oop())); 378 break; 379 case Type::KlassPtr: 380 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_klassptr()->klass())); 381 break; 382 case Type::MetadataPtr: 383 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_metadataptr()->metadata())); 384 break; 385 case Type::NarrowOop: 386 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_oopptr()->const_oop())); 387 break; 388 case Type::RawPtr: 389 st->print(" %s%d]=#Raw" INTPTR_FORMAT,msg,i,p2i(t->is_rawptr())); 390 break; 391 case Type::DoubleCon: 392 st->print(" %s%d]=#%fD",msg,i,t->is_double_constant()->_d); 393 break; 394 case Type::FloatCon: 395 st->print(" %s%d]=#%fF",msg,i,t->is_float_constant()->_f); 396 break; 397 case Type::Long: 398 st->print(" %s%d]=#" INT64_FORMAT,msg,i,(int64_t)(t->is_long()->get_con())); 399 break; 400 case Type::Half: 401 case Type::Top: 402 st->print(" %s%d]=_",msg,i); 403 break; 404 default: ShouldNotReachHere(); 405 } 406 } 407 } 408 409 //------------------------------format----------------------------------------- 410 void JVMState::format(PhaseRegAlloc *regalloc, const Node *n, outputStream* st) const { 411 st->print(" #"); 412 if (_method) { 413 _method->print_short_name(st); 414 st->print(" @ bci:%d ",_bci); 415 } else { 416 st->print_cr(" runtime stub "); 417 return; 418 } 419 if (n->is_MachSafePoint()) { 420 GrowableArray<SafePointScalarObjectNode*> scobjs; 421 MachSafePointNode *mcall = n->as_MachSafePoint(); 422 uint i; 423 // Print locals 424 for (i = 0; i < (uint)loc_size(); i++) 425 format_helper(regalloc, st, mcall->local(this, i), "L[", i, &scobjs); 426 // Print stack 427 for (i = 0; i < (uint)stk_size(); i++) { 428 if ((uint)(_stkoff + i) >= mcall->len()) 429 st->print(" oob "); 430 else 431 format_helper(regalloc, st, mcall->stack(this, i), "STK[", i, &scobjs); 432 } 433 for (i = 0; (int)i < nof_monitors(); i++) { 434 Node *box = mcall->monitor_box(this, i); 435 Node *obj = mcall->monitor_obj(this, i); 436 if (regalloc->node_regs_max_index() > 0 && 437 OptoReg::is_valid(regalloc->get_reg_first(box))) { 438 box = BoxLockNode::box_node(box); 439 format_helper(regalloc, st, box, "MON-BOX[", i, &scobjs); 440 } else { 441 OptoReg::Name box_reg = BoxLockNode::reg(box); 442 st->print(" MON-BOX%d=%s+%d", 443 i, 444 OptoReg::regname(OptoReg::c_frame_pointer), 445 regalloc->reg2offset(box_reg)); 446 } 447 const char* obj_msg = "MON-OBJ["; 448 if (EliminateLocks) { 449 if (BoxLockNode::box_node(box)->is_eliminated()) 450 obj_msg = "MON-OBJ(LOCK ELIMINATED)["; 451 } 452 format_helper(regalloc, st, obj, obj_msg, i, &scobjs); 453 } 454 455 for (i = 0; i < (uint)scobjs.length(); i++) { 456 // Scalar replaced objects. 457 st->cr(); 458 st->print(" # ScObj" INT32_FORMAT " ", i); 459 SafePointScalarObjectNode* spobj = scobjs.at(i); 460 ciKlass* cik = spobj->bottom_type()->is_oopptr()->klass(); 461 assert(cik->is_instance_klass() || 462 cik->is_array_klass(), "Not supported allocation."); 463 ciInstanceKlass *iklass = NULL; 464 if (cik->is_instance_klass()) { 465 cik->print_name_on(st); 466 iklass = cik->as_instance_klass(); 467 } else if (cik->is_type_array_klass()) { 468 cik->as_array_klass()->base_element_type()->print_name_on(st); 469 st->print("[%d]", spobj->n_fields()); 470 } else if (cik->is_obj_array_klass()) { 471 ciKlass* cie = cik->as_obj_array_klass()->base_element_klass(); 472 if (cie->is_instance_klass()) { 473 cie->print_name_on(st); 474 } else if (cie->is_type_array_klass()) { 475 cie->as_array_klass()->base_element_type()->print_name_on(st); 476 } else { 477 ShouldNotReachHere(); 478 } 479 st->print("[%d]", spobj->n_fields()); 480 int ndim = cik->as_array_klass()->dimension() - 1; 481 while (ndim-- > 0) { 482 st->print("[]"); 483 } 484 } else if (cik->is_value_array_klass()) { 485 ciKlass* cie = cik->as_value_array_klass()->base_element_klass(); 486 cie->print_name_on(st); 487 st->print("[%d]", spobj->n_fields()); 488 int ndim = cik->as_array_klass()->dimension() - 1; 489 while (ndim-- > 0) { 490 st->print("[]"); 491 } 492 } 493 st->print("={"); 494 uint nf = spobj->n_fields(); 495 if (nf > 0) { 496 uint first_ind = spobj->first_index(mcall->jvms()); 497 Node* fld_node = mcall->in(first_ind); 498 ciField* cifield; 499 if (iklass != NULL) { 500 st->print(" ["); 501 cifield = iklass->nonstatic_field_at(0); 502 cifield->print_name_on(st); 503 format_helper(regalloc, st, fld_node, ":", 0, &scobjs); 504 } else { 505 format_helper(regalloc, st, fld_node, "[", 0, &scobjs); 506 } 507 for (uint j = 1; j < nf; j++) { 508 fld_node = mcall->in(first_ind+j); 509 if (iklass != NULL) { 510 st->print(", ["); 511 cifield = iklass->nonstatic_field_at(j); 512 cifield->print_name_on(st); 513 format_helper(regalloc, st, fld_node, ":", j, &scobjs); 514 } else { 515 format_helper(regalloc, st, fld_node, ", [", j, &scobjs); 516 } 517 } 518 } 519 st->print(" }"); 520 } 521 } 522 st->cr(); 523 if (caller() != NULL) caller()->format(regalloc, n, st); 524 } 525 526 527 void JVMState::dump_spec(outputStream *st) const { 528 if (_method != NULL) { 529 bool printed = false; 530 if (!Verbose) { 531 // The JVMS dumps make really, really long lines. 532 // Take out the most boring parts, which are the package prefixes. 533 char buf[500]; 534 stringStream namest(buf, sizeof(buf)); 535 _method->print_short_name(&namest); 536 if (namest.count() < sizeof(buf)) { 537 const char* name = namest.base(); 538 if (name[0] == ' ') ++name; 539 const char* endcn = strchr(name, ':'); // end of class name 540 if (endcn == NULL) endcn = strchr(name, '('); 541 if (endcn == NULL) endcn = name + strlen(name); 542 while (endcn > name && endcn[-1] != '.' && endcn[-1] != '/') 543 --endcn; 544 st->print(" %s", endcn); 545 printed = true; 546 } 547 } 548 if (!printed) 549 _method->print_short_name(st); 550 st->print(" @ bci:%d",_bci); 551 if(_reexecute == Reexecute_True) 552 st->print(" reexecute"); 553 } else { 554 st->print(" runtime stub"); 555 } 556 if (caller() != NULL) caller()->dump_spec(st); 557 } 558 559 560 void JVMState::dump_on(outputStream* st) const { 561 bool print_map = _map && !((uintptr_t)_map & 1) && 562 ((caller() == NULL) || (caller()->map() != _map)); 563 if (print_map) { 564 if (_map->len() > _map->req()) { // _map->has_exceptions() 565 Node* ex = _map->in(_map->req()); // _map->next_exception() 566 // skip the first one; it's already being printed 567 while (ex != NULL && ex->len() > ex->req()) { 568 ex = ex->in(ex->req()); // ex->next_exception() 569 ex->dump(1); 570 } 571 } 572 _map->dump(Verbose ? 2 : 1); 573 } 574 if (caller() != NULL) { 575 caller()->dump_on(st); 576 } 577 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=", 578 depth(), locoff(), stkoff(), argoff(), monoff(), scloff(), endoff(), monitor_depth(), sp(), bci(), should_reexecute()?"true":"false"); 579 if (_method == NULL) { 580 st->print_cr("(none)"); 581 } else { 582 _method->print_name(st); 583 st->cr(); 584 if (bci() >= 0 && bci() < _method->code_size()) { 585 st->print(" bc: "); 586 _method->print_codes_on(bci(), bci()+1, st); 587 } 588 } 589 } 590 591 // Extra way to dump a jvms from the debugger, 592 // to avoid a bug with C++ member function calls. 593 void dump_jvms(JVMState* jvms) { 594 jvms->dump(); 595 } 596 #endif 597 598 //--------------------------clone_shallow-------------------------------------- 599 JVMState* JVMState::clone_shallow(Compile* C) const { 600 JVMState* n = has_method() ? new (C) JVMState(_method, _caller) : new (C) JVMState(0); 601 n->set_bci(_bci); 602 n->_reexecute = _reexecute; 603 n->set_locoff(_locoff); 604 n->set_stkoff(_stkoff); 605 n->set_monoff(_monoff); 606 n->set_scloff(_scloff); 607 n->set_endoff(_endoff); 608 n->set_sp(_sp); 609 n->set_map(_map); 610 return n; 611 } 612 613 //---------------------------clone_deep---------------------------------------- 614 JVMState* JVMState::clone_deep(Compile* C) const { 615 JVMState* n = clone_shallow(C); 616 for (JVMState* p = n; p->_caller != NULL; p = p->_caller) { 617 p->_caller = p->_caller->clone_shallow(C); 618 } 619 assert(n->depth() == depth(), "sanity"); 620 assert(n->debug_depth() == debug_depth(), "sanity"); 621 return n; 622 } 623 624 /** 625 * Reset map for all callers 626 */ 627 void JVMState::set_map_deep(SafePointNode* map) { 628 for (JVMState* p = this; p->_caller != NULL; p = p->_caller) { 629 p->set_map(map); 630 } 631 } 632 633 // Adapt offsets in in-array after adding or removing an edge. 634 // Prerequisite is that the JVMState is used by only one node. 635 void JVMState::adapt_position(int delta) { 636 for (JVMState* jvms = this; jvms != NULL; jvms = jvms->caller()) { 637 jvms->set_locoff(jvms->locoff() + delta); 638 jvms->set_stkoff(jvms->stkoff() + delta); 639 jvms->set_monoff(jvms->monoff() + delta); 640 jvms->set_scloff(jvms->scloff() + delta); 641 jvms->set_endoff(jvms->endoff() + delta); 642 } 643 } 644 645 // Mirror the stack size calculation in the deopt code 646 // How much stack space would we need at this point in the program in 647 // case of deoptimization? 648 int JVMState::interpreter_frame_size() const { 649 const JVMState* jvms = this; 650 int size = 0; 651 int callee_parameters = 0; 652 int callee_locals = 0; 653 int extra_args = method()->max_stack() - stk_size(); 654 655 while (jvms != NULL) { 656 int locks = jvms->nof_monitors(); 657 int temps = jvms->stk_size(); 658 bool is_top_frame = (jvms == this); 659 ciMethod* method = jvms->method(); 660 661 int frame_size = BytesPerWord * Interpreter::size_activation(method->max_stack(), 662 temps + callee_parameters, 663 extra_args, 664 locks, 665 callee_parameters, 666 callee_locals, 667 is_top_frame); 668 size += frame_size; 669 670 callee_parameters = method->size_of_parameters(); 671 callee_locals = method->max_locals(); 672 extra_args = 0; 673 jvms = jvms->caller(); 674 } 675 return size + Deoptimization::last_frame_adjust(0, callee_locals) * BytesPerWord; 676 } 677 678 //============================================================================= 679 bool CallNode::cmp( const Node &n ) const 680 { return _tf == ((CallNode&)n)._tf && _jvms == ((CallNode&)n)._jvms; } 681 #ifndef PRODUCT 682 void CallNode::dump_req(outputStream *st) const { 683 // Dump the required inputs, enclosed in '(' and ')' 684 uint i; // Exit value of loop 685 for (i = 0; i < req(); i++) { // For all required inputs 686 if (i == TypeFunc::Parms) st->print("("); 687 if (in(i)) st->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx); 688 else st->print("_ "); 689 } 690 st->print(")"); 691 } 692 693 void CallNode::dump_spec(outputStream *st) const { 694 st->print(" "); 695 if (tf() != NULL) tf()->dump_on(st); 696 if (_cnt != COUNT_UNKNOWN) st->print(" C=%f",_cnt); 697 if (jvms() != NULL) jvms()->dump_spec(st); 698 } 699 #endif 700 701 const Type *CallNode::bottom_type() const { return tf()->range_cc(); } 702 const Type* CallNode::Value(PhaseGVN* phase) const { 703 if (!in(0) || phase->type(in(0)) == Type::TOP) { 704 return Type::TOP; 705 } 706 return tf()->range_cc(); 707 } 708 709 //------------------------------calling_convention----------------------------- 710 void CallNode::calling_convention(BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt) const { 711 if (_entry_point == StubRoutines::store_value_type_fields_to_buf()) { 712 // The call to that stub is a special case: its inputs are 713 // multiple values returned from a call and so it should follow 714 // the return convention. 715 SharedRuntime::java_return_convention(sig_bt, parm_regs, argcnt); 716 return; 717 } 718 // Use the standard compiler calling convention 719 Matcher::calling_convention( sig_bt, parm_regs, argcnt, true ); 720 } 721 722 723 //------------------------------match------------------------------------------ 724 // Construct projections for control, I/O, memory-fields, ..., and 725 // return result(s) along with their RegMask info 726 Node *CallNode::match(const ProjNode *proj, const Matcher *match, const RegMask* mask) { 727 uint con = proj->_con; 728 const TypeTuple *range_cc = tf()->range_cc(); 729 if (con >= TypeFunc::Parms) { 730 if (is_CallRuntime()) { 731 if (con == TypeFunc::Parms) { 732 uint ideal_reg = range_cc->field_at(TypeFunc::Parms)->ideal_reg(); 733 OptoRegPair regs = match->c_return_value(ideal_reg,true); 734 RegMask rm = RegMask(regs.first()); 735 if (OptoReg::is_valid(regs.second())) { 736 rm.Insert(regs.second()); 737 } 738 return new MachProjNode(this,con,rm,ideal_reg); 739 } else { 740 assert(con == TypeFunc::Parms+1, "only one return value"); 741 assert(range_cc->field_at(TypeFunc::Parms+1) == Type::HALF, ""); 742 return new MachProjNode(this,con, RegMask::Empty, (uint)OptoReg::Bad); 743 } 744 } else { 745 // The Call may return multiple values (value type fields): we 746 // create one projection per returned values. 747 assert(con <= TypeFunc::Parms+1 || ValueTypeReturnedAsFields, "only for multi value return"); 748 uint ideal_reg = range_cc->field_at(con)->ideal_reg(); 749 return new MachProjNode(this, con, mask[con-TypeFunc::Parms], ideal_reg); 750 } 751 } 752 753 switch (con) { 754 case TypeFunc::Control: 755 case TypeFunc::I_O: 756 case TypeFunc::Memory: 757 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 758 759 case TypeFunc::ReturnAdr: 760 case TypeFunc::FramePtr: 761 default: 762 ShouldNotReachHere(); 763 } 764 return NULL; 765 } 766 767 // Do we Match on this edge index or not? Match no edges 768 uint CallNode::match_edge(uint idx) const { 769 return 0; 770 } 771 772 // 773 // Determine whether the call could modify the field of the specified 774 // instance at the specified offset. 775 // 776 bool CallNode::may_modify(const TypeOopPtr *t_oop, PhaseTransform *phase) { 777 assert((t_oop != NULL), "sanity"); 778 if (is_call_to_arraycopystub() && strcmp(_name, "unsafe_arraycopy") != 0) { 779 const TypeTuple* args = _tf->domain_sig(); 780 Node* dest = NULL; 781 // Stubs that can be called once an ArrayCopyNode is expanded have 782 // different signatures. Look for the second pointer argument, 783 // that is the destination of the copy. 784 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) { 785 if (args->field_at(i)->isa_ptr()) { 786 j++; 787 if (j == 2) { 788 dest = in(i); 789 break; 790 } 791 } 792 } 793 guarantee(dest != NULL, "Call had only one ptr in, broken IR!"); 794 if (!dest->is_top() && may_modify_arraycopy_helper(phase->type(dest)->is_oopptr(), t_oop, phase)) { 795 return true; 796 } 797 return false; 798 } 799 if (t_oop->is_known_instance()) { 800 // The instance_id is set only for scalar-replaceable allocations which 801 // are not passed as arguments according to Escape Analysis. 802 return false; 803 } 804 if (t_oop->is_ptr_to_boxed_value()) { 805 ciKlass* boxing_klass = t_oop->klass(); 806 if (is_CallStaticJava() && as_CallStaticJava()->is_boxing_method()) { 807 // Skip unrelated boxing methods. 808 Node* proj = proj_out_or_null(TypeFunc::Parms); 809 if ((proj == NULL) || (phase->type(proj)->is_instptr()->klass() != boxing_klass)) { 810 return false; 811 } 812 } 813 if (is_CallJava() && as_CallJava()->method() != NULL) { 814 ciMethod* meth = as_CallJava()->method(); 815 if (meth->is_getter()) { 816 return false; 817 } 818 // May modify (by reflection) if an boxing object is passed 819 // as argument or returned. 820 Node* proj = returns_pointer() ? proj_out_or_null(TypeFunc::Parms) : NULL; 821 if (proj != NULL) { 822 const TypeInstPtr* inst_t = phase->type(proj)->isa_instptr(); 823 if ((inst_t != NULL) && (!inst_t->klass_is_exact() || 824 (inst_t->klass() == boxing_klass))) { 825 return true; 826 } 827 } 828 const TypeTuple* d = tf()->domain_cc(); 829 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 830 const TypeInstPtr* inst_t = d->field_at(i)->isa_instptr(); 831 if ((inst_t != NULL) && (!inst_t->klass_is_exact() || 832 (inst_t->klass() == boxing_klass))) { 833 return true; 834 } 835 } 836 return false; 837 } 838 } 839 return true; 840 } 841 842 // Does this call have a direct reference to n other than debug information? 843 bool CallNode::has_non_debug_use(Node *n) { 844 const TypeTuple * d = tf()->domain_cc(); 845 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 846 Node *arg = in(i); 847 if (arg == n) { 848 return true; 849 } 850 } 851 return false; 852 } 853 854 bool CallNode::has_debug_use(Node *n) { 855 assert(jvms() != NULL, "jvms should not be null"); 856 for (uint i = jvms()->debug_start(); i < jvms()->debug_end(); i++) { 857 Node *arg = in(i); 858 if (arg == n) { 859 return true; 860 } 861 } 862 return false; 863 } 864 865 // Returns the unique CheckCastPP of a call 866 // or 'this' if there are several CheckCastPP or unexpected uses 867 // or returns NULL if there is no one. 868 Node *CallNode::result_cast() { 869 Node *cast = NULL; 870 871 Node *p = proj_out_or_null(TypeFunc::Parms); 872 if (p == NULL) 873 return NULL; 874 875 for (DUIterator_Fast imax, i = p->fast_outs(imax); i < imax; i++) { 876 Node *use = p->fast_out(i); 877 if (use->is_CheckCastPP()) { 878 if (cast != NULL) { 879 return this; // more than 1 CheckCastPP 880 } 881 cast = use; 882 } else if (!use->is_Initialize() && 883 !use->is_AddP() && 884 use->Opcode() != Op_MemBarStoreStore) { 885 // Expected uses are restricted to a CheckCastPP, an Initialize 886 // node, a MemBarStoreStore (clone) and AddP nodes. If we 887 // encounter any other use (a Phi node can be seen in rare 888 // cases) return this to prevent incorrect optimizations. 889 return this; 890 } 891 } 892 return cast; 893 } 894 895 896 CallProjections* CallNode::extract_projections(bool separate_io_proj, bool do_asserts) { 897 uint max_res = TypeFunc::Parms-1; 898 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 899 ProjNode *pn = fast_out(i)->as_Proj(); 900 max_res = MAX2(max_res, pn->_con); 901 } 902 903 assert(max_res < _tf->range_cc()->cnt(), "result out of bounds"); 904 905 uint projs_size = sizeof(CallProjections); 906 if (max_res > TypeFunc::Parms) { 907 projs_size += (max_res-TypeFunc::Parms)*sizeof(Node*); 908 } 909 char* projs_storage = resource_allocate_bytes(projs_size); 910 CallProjections* projs = new(projs_storage)CallProjections(max_res - TypeFunc::Parms + 1); 911 912 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 913 ProjNode *pn = fast_out(i)->as_Proj(); 914 if (pn->outcnt() == 0) continue; 915 switch (pn->_con) { 916 case TypeFunc::Control: 917 { 918 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj 919 projs->fallthrough_proj = pn; 920 DUIterator_Fast jmax, j = pn->fast_outs(jmax); 921 const Node *cn = pn->fast_out(j); 922 if (cn->is_Catch()) { 923 ProjNode *cpn = NULL; 924 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) { 925 cpn = cn->fast_out(k)->as_Proj(); 926 assert(cpn->is_CatchProj(), "must be a CatchProjNode"); 927 if (cpn->_con == CatchProjNode::fall_through_index) 928 projs->fallthrough_catchproj = cpn; 929 else { 930 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index."); 931 projs->catchall_catchproj = cpn; 932 } 933 } 934 } 935 break; 936 } 937 case TypeFunc::I_O: 938 if (pn->_is_io_use) 939 projs->catchall_ioproj = pn; 940 else 941 projs->fallthrough_ioproj = pn; 942 for (DUIterator j = pn->outs(); pn->has_out(j); j++) { 943 Node* e = pn->out(j); 944 if (e->Opcode() == Op_CreateEx && e->in(0)->is_CatchProj() && e->outcnt() > 0) { 945 assert(projs->exobj == NULL, "only one"); 946 projs->exobj = e; 947 } 948 } 949 break; 950 case TypeFunc::Memory: 951 if (pn->_is_io_use) 952 projs->catchall_memproj = pn; 953 else 954 projs->fallthrough_memproj = pn; 955 break; 956 case TypeFunc::Parms: 957 projs->resproj[0] = pn; 958 break; 959 default: 960 assert(pn->_con <= max_res, "unexpected projection from allocation node."); 961 projs->resproj[pn->_con-TypeFunc::Parms] = pn; 962 break; 963 } 964 } 965 966 // The resproj may not exist because the result could be ignored 967 // and the exception object may not exist if an exception handler 968 // swallows the exception but all the other must exist and be found. 969 do_asserts = do_asserts && !Compile::current()->inlining_incrementally(); 970 assert(!do_asserts || projs->fallthrough_proj != NULL, "must be found"); 971 assert(!do_asserts || projs->fallthrough_catchproj != NULL, "must be found"); 972 assert(!do_asserts || projs->fallthrough_memproj != NULL, "must be found"); 973 assert(!do_asserts || projs->fallthrough_ioproj != NULL, "must be found"); 974 assert(!do_asserts || projs->catchall_catchproj != NULL, "must be found"); 975 if (separate_io_proj) { 976 assert(!do_asserts || projs->catchall_memproj != NULL, "must be found"); 977 assert(!do_asserts || projs->catchall_ioproj != NULL, "must be found"); 978 } 979 return projs; 980 } 981 982 Node *CallNode::Ideal(PhaseGVN *phase, bool can_reshape) { 983 CallGenerator* cg = generator(); 984 if (can_reshape && cg != NULL && cg->is_mh_late_inline() && !cg->already_attempted()) { 985 // Check whether this MH handle call becomes a candidate for inlining 986 ciMethod* callee = cg->method(); 987 vmIntrinsics::ID iid = callee->intrinsic_id(); 988 if (iid == vmIntrinsics::_invokeBasic) { 989 if (in(TypeFunc::Parms)->Opcode() == Op_ConP) { 990 phase->C->prepend_late_inline(cg); 991 set_generator(NULL); 992 } 993 } else { 994 assert(callee->has_member_arg(), "wrong type of call?"); 995 if (in(TypeFunc::Parms + callee->arg_size() - 1)->Opcode() == Op_ConP) { 996 phase->C->prepend_late_inline(cg); 997 set_generator(NULL); 998 } 999 } 1000 } 1001 return SafePointNode::Ideal(phase, can_reshape); 1002 } 1003 1004 bool CallNode::is_call_to_arraycopystub() const { 1005 if (_name != NULL && strstr(_name, "arraycopy") != 0) { 1006 return true; 1007 } 1008 return false; 1009 } 1010 1011 //============================================================================= 1012 uint CallJavaNode::size_of() const { return sizeof(*this); } 1013 bool CallJavaNode::cmp( const Node &n ) const { 1014 CallJavaNode &call = (CallJavaNode&)n; 1015 return CallNode::cmp(call) && _method == call._method && 1016 _override_symbolic_info == call._override_symbolic_info; 1017 } 1018 #ifdef ASSERT 1019 bool CallJavaNode::validate_symbolic_info() const { 1020 if (method() == NULL) { 1021 return true; // call into runtime or uncommon trap 1022 } 1023 Bytecodes::Code bc = jvms()->method()->java_code_at_bci(_bci); 1024 if (EnableValhalla && (bc == Bytecodes::_if_acmpeq || bc == Bytecodes::_if_acmpne)) { 1025 return true; 1026 } 1027 ciMethod* symbolic_info = jvms()->method()->get_method_at_bci(_bci); 1028 ciMethod* callee = method(); 1029 if (symbolic_info->is_method_handle_intrinsic() && !callee->is_method_handle_intrinsic()) { 1030 assert(override_symbolic_info(), "should be set"); 1031 } 1032 assert(ciMethod::is_consistent_info(symbolic_info, callee), "inconsistent info"); 1033 return true; 1034 } 1035 #endif 1036 1037 #ifndef PRODUCT 1038 void CallJavaNode::dump_spec(outputStream *st) const { 1039 if( _method ) _method->print_short_name(st); 1040 CallNode::dump_spec(st); 1041 } 1042 1043 void CallJavaNode::dump_compact_spec(outputStream* st) const { 1044 if (_method) { 1045 _method->print_short_name(st); 1046 } else { 1047 st->print("<?>"); 1048 } 1049 } 1050 #endif 1051 1052 //============================================================================= 1053 uint CallStaticJavaNode::size_of() const { return sizeof(*this); } 1054 bool CallStaticJavaNode::cmp( const Node &n ) const { 1055 CallStaticJavaNode &call = (CallStaticJavaNode&)n; 1056 return CallJavaNode::cmp(call); 1057 } 1058 1059 //----------------------------uncommon_trap_request---------------------------- 1060 // If this is an uncommon trap, return the request code, else zero. 1061 int CallStaticJavaNode::uncommon_trap_request() const { 1062 if (_name != NULL && !strcmp(_name, "uncommon_trap")) { 1063 return extract_uncommon_trap_request(this); 1064 } 1065 return 0; 1066 } 1067 int CallStaticJavaNode::extract_uncommon_trap_request(const Node* call) { 1068 #ifndef PRODUCT 1069 if (!(call->req() > TypeFunc::Parms && 1070 call->in(TypeFunc::Parms) != NULL && 1071 call->in(TypeFunc::Parms)->is_Con() && 1072 call->in(TypeFunc::Parms)->bottom_type()->isa_int())) { 1073 assert(in_dump() != 0, "OK if dumping"); 1074 tty->print("[bad uncommon trap]"); 1075 return 0; 1076 } 1077 #endif 1078 return call->in(TypeFunc::Parms)->bottom_type()->is_int()->get_con(); 1079 } 1080 1081 #ifndef PRODUCT 1082 void CallStaticJavaNode::dump_spec(outputStream *st) const { 1083 st->print("# Static "); 1084 if (_name != NULL) { 1085 st->print("%s", _name); 1086 int trap_req = uncommon_trap_request(); 1087 if (trap_req != 0) { 1088 char buf[100]; 1089 st->print("(%s)", 1090 Deoptimization::format_trap_request(buf, sizeof(buf), 1091 trap_req)); 1092 } 1093 st->print(" "); 1094 } 1095 CallJavaNode::dump_spec(st); 1096 } 1097 1098 void CallStaticJavaNode::dump_compact_spec(outputStream* st) const { 1099 if (_method) { 1100 _method->print_short_name(st); 1101 } else if (_name) { 1102 st->print("%s", _name); 1103 } else { 1104 st->print("<?>"); 1105 } 1106 } 1107 #endif 1108 1109 //============================================================================= 1110 uint CallDynamicJavaNode::size_of() const { return sizeof(*this); } 1111 bool CallDynamicJavaNode::cmp( const Node &n ) const { 1112 CallDynamicJavaNode &call = (CallDynamicJavaNode&)n; 1113 return CallJavaNode::cmp(call); 1114 } 1115 #ifndef PRODUCT 1116 void CallDynamicJavaNode::dump_spec(outputStream *st) const { 1117 st->print("# Dynamic "); 1118 CallJavaNode::dump_spec(st); 1119 } 1120 #endif 1121 1122 //============================================================================= 1123 uint CallRuntimeNode::size_of() const { return sizeof(*this); } 1124 bool CallRuntimeNode::cmp( const Node &n ) const { 1125 CallRuntimeNode &call = (CallRuntimeNode&)n; 1126 return CallNode::cmp(call) && !strcmp(_name,call._name); 1127 } 1128 #ifndef PRODUCT 1129 void CallRuntimeNode::dump_spec(outputStream *st) const { 1130 st->print("# "); 1131 st->print("%s", _name); 1132 CallNode::dump_spec(st); 1133 } 1134 #endif 1135 1136 //------------------------------calling_convention----------------------------- 1137 void CallRuntimeNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const { 1138 if (_entry_point == NULL) { 1139 // The call to that stub is a special case: its inputs are 1140 // multiple values returned from a call and so it should follow 1141 // the return convention. 1142 SharedRuntime::java_return_convention(sig_bt, parm_regs, argcnt); 1143 return; 1144 } 1145 Matcher::c_calling_convention( sig_bt, parm_regs, argcnt ); 1146 } 1147 1148 //============================================================================= 1149 //------------------------------calling_convention----------------------------- 1150 1151 1152 //============================================================================= 1153 #ifndef PRODUCT 1154 void CallLeafNode::dump_spec(outputStream *st) const { 1155 st->print("# "); 1156 st->print("%s", _name); 1157 CallNode::dump_spec(st); 1158 } 1159 #endif 1160 1161 uint CallLeafNoFPNode::match_edge(uint idx) const { 1162 // Null entry point is a special case for which the target is in a 1163 // register. Need to match that edge. 1164 return entry_point() == NULL && idx == TypeFunc::Parms; 1165 } 1166 1167 //============================================================================= 1168 1169 void SafePointNode::set_local(JVMState* jvms, uint idx, Node *c) { 1170 assert(verify_jvms(jvms), "jvms must match"); 1171 int loc = jvms->locoff() + idx; 1172 if (in(loc)->is_top() && idx > 0 && !c->is_top() ) { 1173 // If current local idx is top then local idx - 1 could 1174 // be a long/double that needs to be killed since top could 1175 // represent the 2nd half ofthe long/double. 1176 uint ideal = in(loc -1)->ideal_reg(); 1177 if (ideal == Op_RegD || ideal == Op_RegL) { 1178 // set other (low index) half to top 1179 set_req(loc - 1, in(loc)); 1180 } 1181 } 1182 set_req(loc, c); 1183 } 1184 1185 uint SafePointNode::size_of() const { return sizeof(*this); } 1186 bool SafePointNode::cmp( const Node &n ) const { 1187 return (&n == this); // Always fail except on self 1188 } 1189 1190 //-------------------------set_next_exception---------------------------------- 1191 void SafePointNode::set_next_exception(SafePointNode* n) { 1192 assert(n == NULL || n->Opcode() == Op_SafePoint, "correct value for next_exception"); 1193 if (len() == req()) { 1194 if (n != NULL) add_prec(n); 1195 } else { 1196 set_prec(req(), n); 1197 } 1198 } 1199 1200 1201 //----------------------------next_exception----------------------------------- 1202 SafePointNode* SafePointNode::next_exception() const { 1203 if (len() == req()) { 1204 return NULL; 1205 } else { 1206 Node* n = in(req()); 1207 assert(n == NULL || n->Opcode() == Op_SafePoint, "no other uses of prec edges"); 1208 return (SafePointNode*) n; 1209 } 1210 } 1211 1212 1213 //------------------------------Ideal------------------------------------------ 1214 // Skip over any collapsed Regions 1215 Node *SafePointNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1216 return remove_dead_region(phase, can_reshape) ? this : NULL; 1217 } 1218 1219 //------------------------------Identity--------------------------------------- 1220 // Remove obviously duplicate safepoints 1221 Node* SafePointNode::Identity(PhaseGVN* phase) { 1222 1223 // If you have back to back safepoints, remove one 1224 if( in(TypeFunc::Control)->is_SafePoint() ) 1225 return in(TypeFunc::Control); 1226 1227 if( in(0)->is_Proj() ) { 1228 Node *n0 = in(0)->in(0); 1229 // Check if he is a call projection (except Leaf Call) 1230 if( n0->is_Catch() ) { 1231 n0 = n0->in(0)->in(0); 1232 assert( n0->is_Call(), "expect a call here" ); 1233 } 1234 if( n0->is_Call() && n0->as_Call()->guaranteed_safepoint() ) { 1235 // Don't remove a safepoint belonging to an OuterStripMinedLoopEndNode. 1236 // If the loop dies, they will be removed together. 1237 if (has_out_with(Op_OuterStripMinedLoopEnd)) { 1238 return this; 1239 } 1240 // Useless Safepoint, so remove it 1241 return in(TypeFunc::Control); 1242 } 1243 } 1244 1245 return this; 1246 } 1247 1248 //------------------------------Value------------------------------------------ 1249 const Type* SafePointNode::Value(PhaseGVN* phase) const { 1250 if( phase->type(in(0)) == Type::TOP ) return Type::TOP; 1251 if( phase->eqv( in(0), this ) ) return Type::TOP; // Dead infinite loop 1252 return Type::CONTROL; 1253 } 1254 1255 #ifndef PRODUCT 1256 void SafePointNode::dump_spec(outputStream *st) const { 1257 st->print(" SafePoint "); 1258 _replaced_nodes.dump(st); 1259 } 1260 1261 // The related nodes of a SafepointNode are all data inputs, excluding the 1262 // control boundary, as well as all outputs till level 2 (to include projection 1263 // nodes and targets). In compact mode, just include inputs till level 1 and 1264 // outputs as before. 1265 void SafePointNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const { 1266 if (compact) { 1267 this->collect_nodes(in_rel, 1, false, false); 1268 } else { 1269 this->collect_nodes_in_all_data(in_rel, false); 1270 } 1271 this->collect_nodes(out_rel, -2, false, false); 1272 } 1273 #endif 1274 1275 const RegMask &SafePointNode::in_RegMask(uint idx) const { 1276 if( idx < TypeFunc::Parms ) return RegMask::Empty; 1277 // Values outside the domain represent debug info 1278 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]); 1279 } 1280 const RegMask &SafePointNode::out_RegMask() const { 1281 return RegMask::Empty; 1282 } 1283 1284 1285 void SafePointNode::grow_stack(JVMState* jvms, uint grow_by) { 1286 assert((int)grow_by > 0, "sanity"); 1287 int monoff = jvms->monoff(); 1288 int scloff = jvms->scloff(); 1289 int endoff = jvms->endoff(); 1290 assert(endoff == (int)req(), "no other states or debug info after me"); 1291 Node* top = Compile::current()->top(); 1292 for (uint i = 0; i < grow_by; i++) { 1293 ins_req(monoff, top); 1294 } 1295 jvms->set_monoff(monoff + grow_by); 1296 jvms->set_scloff(scloff + grow_by); 1297 jvms->set_endoff(endoff + grow_by); 1298 } 1299 1300 void SafePointNode::push_monitor(const FastLockNode *lock) { 1301 // Add a LockNode, which points to both the original BoxLockNode (the 1302 // stack space for the monitor) and the Object being locked. 1303 const int MonitorEdges = 2; 1304 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges"); 1305 assert(req() == jvms()->endoff(), "correct sizing"); 1306 int nextmon = jvms()->scloff(); 1307 if (GenerateSynchronizationCode) { 1308 ins_req(nextmon, lock->box_node()); 1309 ins_req(nextmon+1, lock->obj_node()); 1310 } else { 1311 Node* top = Compile::current()->top(); 1312 ins_req(nextmon, top); 1313 ins_req(nextmon, top); 1314 } 1315 jvms()->set_scloff(nextmon + MonitorEdges); 1316 jvms()->set_endoff(req()); 1317 } 1318 1319 void SafePointNode::pop_monitor() { 1320 // Delete last monitor from debug info 1321 debug_only(int num_before_pop = jvms()->nof_monitors()); 1322 const int MonitorEdges = 2; 1323 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges"); 1324 int scloff = jvms()->scloff(); 1325 int endoff = jvms()->endoff(); 1326 int new_scloff = scloff - MonitorEdges; 1327 int new_endoff = endoff - MonitorEdges; 1328 jvms()->set_scloff(new_scloff); 1329 jvms()->set_endoff(new_endoff); 1330 while (scloff > new_scloff) del_req_ordered(--scloff); 1331 assert(jvms()->nof_monitors() == num_before_pop-1, ""); 1332 } 1333 1334 Node *SafePointNode::peek_monitor_box() const { 1335 int mon = jvms()->nof_monitors() - 1; 1336 assert(mon >= 0, "must have a monitor"); 1337 return monitor_box(jvms(), mon); 1338 } 1339 1340 Node *SafePointNode::peek_monitor_obj() const { 1341 int mon = jvms()->nof_monitors() - 1; 1342 assert(mon >= 0, "must have a monitor"); 1343 return monitor_obj(jvms(), mon); 1344 } 1345 1346 // Do we Match on this edge index or not? Match no edges 1347 uint SafePointNode::match_edge(uint idx) const { 1348 if( !needs_polling_address_input() ) 1349 return 0; 1350 1351 return (TypeFunc::Parms == idx); 1352 } 1353 1354 void SafePointNode::disconnect_from_root(PhaseIterGVN *igvn) { 1355 assert(Opcode() == Op_SafePoint, "only value for safepoint in loops"); 1356 int nb = igvn->C->root()->find_prec_edge(this); 1357 if (nb != -1) { 1358 igvn->C->root()->rm_prec(nb); 1359 } 1360 } 1361 1362 //============== SafePointScalarObjectNode ============== 1363 1364 SafePointScalarObjectNode::SafePointScalarObjectNode(const TypeOopPtr* tp, 1365 #ifdef ASSERT 1366 AllocateNode* alloc, 1367 #endif 1368 uint first_index, 1369 uint n_fields) : 1370 TypeNode(tp, 1), // 1 control input -- seems required. Get from root. 1371 _first_index(first_index), 1372 _n_fields(n_fields) 1373 #ifdef ASSERT 1374 , _alloc(alloc) 1375 #endif 1376 { 1377 init_class_id(Class_SafePointScalarObject); 1378 } 1379 1380 // Do not allow value-numbering for SafePointScalarObject node. 1381 uint SafePointScalarObjectNode::hash() const { return NO_HASH; } 1382 bool SafePointScalarObjectNode::cmp( const Node &n ) const { 1383 return (&n == this); // Always fail except on self 1384 } 1385 1386 uint SafePointScalarObjectNode::ideal_reg() const { 1387 return 0; // No matching to machine instruction 1388 } 1389 1390 const RegMask &SafePointScalarObjectNode::in_RegMask(uint idx) const { 1391 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]); 1392 } 1393 1394 const RegMask &SafePointScalarObjectNode::out_RegMask() const { 1395 return RegMask::Empty; 1396 } 1397 1398 uint SafePointScalarObjectNode::match_edge(uint idx) const { 1399 return 0; 1400 } 1401 1402 SafePointScalarObjectNode* 1403 SafePointScalarObjectNode::clone(Dict* sosn_map) const { 1404 void* cached = (*sosn_map)[(void*)this]; 1405 if (cached != NULL) { 1406 return (SafePointScalarObjectNode*)cached; 1407 } 1408 SafePointScalarObjectNode* res = (SafePointScalarObjectNode*)Node::clone(); 1409 sosn_map->Insert((void*)this, (void*)res); 1410 return res; 1411 } 1412 1413 1414 #ifndef PRODUCT 1415 void SafePointScalarObjectNode::dump_spec(outputStream *st) const { 1416 st->print(" # fields@[%d..%d]", first_index(), 1417 first_index() + n_fields() - 1); 1418 } 1419 1420 #endif 1421 1422 //============================================================================= 1423 uint AllocateNode::size_of() const { return sizeof(*this); } 1424 1425 AllocateNode::AllocateNode(Compile* C, const TypeFunc *atype, 1426 Node *ctrl, Node *mem, Node *abio, 1427 Node *size, Node *klass_node, 1428 Node* initial_test, 1429 ValueTypeBaseNode* value_node) 1430 : CallNode(atype, NULL, TypeRawPtr::BOTTOM) 1431 { 1432 init_class_id(Class_Allocate); 1433 init_flags(Flag_is_macro); 1434 _is_scalar_replaceable = false; 1435 _is_non_escaping = false; 1436 _is_allocation_MemBar_redundant = false; 1437 _larval = false; 1438 Node *topnode = C->top(); 1439 1440 init_req( TypeFunc::Control , ctrl ); 1441 init_req( TypeFunc::I_O , abio ); 1442 init_req( TypeFunc::Memory , mem ); 1443 init_req( TypeFunc::ReturnAdr, topnode ); 1444 init_req( TypeFunc::FramePtr , topnode ); 1445 init_req( AllocSize , size); 1446 init_req( KlassNode , klass_node); 1447 init_req( InitialTest , initial_test); 1448 init_req( ALength , topnode); 1449 init_req( ValueNode , value_node); 1450 // DefaultValue defaults to NULL 1451 // RawDefaultValue defaults to NULL 1452 // StorageProperties defaults to NULL 1453 C->add_macro_node(this); 1454 } 1455 1456 void AllocateNode::compute_MemBar_redundancy(ciMethod* initializer) 1457 { 1458 assert(initializer != NULL && 1459 initializer->is_object_constructor_or_class_initializer(), 1460 "unexpected initializer method"); 1461 BCEscapeAnalyzer* analyzer = initializer->get_bcea(); 1462 if (analyzer == NULL) { 1463 return; 1464 } 1465 1466 // Allocation node is first parameter in its initializer 1467 if (analyzer->is_arg_stack(0) || analyzer->is_arg_local(0)) { 1468 _is_allocation_MemBar_redundant = true; 1469 } 1470 } 1471 1472 Node* AllocateNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1473 // Check for unused value type allocation 1474 if (can_reshape && in(AllocateNode::ValueNode) != NULL && 1475 outcnt() != 0 && result_cast() == NULL) { 1476 // Remove allocation by replacing the projection nodes with its inputs 1477 InitializeNode* init = initialization(); 1478 PhaseIterGVN* igvn = phase->is_IterGVN(); 1479 CallProjections* projs = extract_projections(true, false); 1480 assert(projs->nb_resproj <= 1, "unexpected number of results"); 1481 if (projs->fallthrough_catchproj != NULL) { 1482 igvn->replace_node(projs->fallthrough_catchproj, in(TypeFunc::Control)); 1483 } 1484 if (projs->fallthrough_memproj != NULL) { 1485 igvn->replace_node(projs->fallthrough_memproj, in(TypeFunc::Memory)); 1486 } 1487 if (projs->catchall_memproj != NULL) { 1488 igvn->replace_node(projs->catchall_memproj, phase->C->top()); 1489 } 1490 if (projs->fallthrough_ioproj != NULL) { 1491 igvn->replace_node(projs->fallthrough_ioproj, in(TypeFunc::I_O)); 1492 } 1493 if (projs->catchall_ioproj != NULL) { 1494 igvn->replace_node(projs->catchall_ioproj, phase->C->top()); 1495 } 1496 if (projs->catchall_catchproj != NULL) { 1497 igvn->replace_node(projs->catchall_catchproj, phase->C->top()); 1498 } 1499 if (projs->resproj[0] != NULL) { 1500 igvn->replace_node(projs->resproj[0], phase->C->top()); 1501 } 1502 igvn->replace_node(this, phase->C->top()); 1503 if (init != NULL) { 1504 Node* ctrl_proj = init->proj_out_or_null(TypeFunc::Control); 1505 Node* mem_proj = init->proj_out_or_null(TypeFunc::Memory); 1506 if (ctrl_proj != NULL) { 1507 igvn->replace_node(ctrl_proj, init->in(TypeFunc::Control)); 1508 } 1509 if (mem_proj != NULL) { 1510 igvn->replace_node(mem_proj, init->in(TypeFunc::Memory)); 1511 } 1512 } 1513 return NULL; 1514 } 1515 1516 return CallNode::Ideal(phase, can_reshape); 1517 } 1518 1519 Node *AllocateNode::make_ideal_mark(PhaseGVN *phase, Node* obj, Node* control, Node* mem, Node* klass_node) { 1520 Node* mark_node = NULL; 1521 // For now only enable fast locking for non-array types 1522 if ((EnableValhalla || UseBiasedLocking) && Opcode() == Op_Allocate) { 1523 if (klass_node == NULL) { 1524 Node* k_adr = phase->transform(new AddPNode(obj, obj, phase->MakeConX(oopDesc::klass_offset_in_bytes()))); 1525 klass_node = phase->transform(LoadKlassNode::make(*phase, NULL, phase->C->immutable_memory(), k_adr, phase->type(k_adr)->is_ptr())); 1526 } 1527 Node* proto_adr = phase->transform(new AddPNode(klass_node, klass_node, phase->MakeConX(in_bytes(Klass::prototype_header_offset())))); 1528 mark_node = LoadNode::make(*phase, control, mem, proto_adr, TypeRawPtr::BOTTOM, TypeX_X, TypeX_X->basic_type(), MemNode::unordered); 1529 } else { 1530 mark_node = phase->MakeConX((jlong)markWord::prototype().value()); 1531 } 1532 mark_node = phase->transform(mark_node); 1533 // Avoid returning a constant (old node) here because this method is used by LoadNode::Ideal 1534 return new OrXNode(mark_node, phase->MakeConX(_larval ? markWord::larval_state_pattern : 0)); 1535 } 1536 1537 1538 //============================================================================= 1539 Node* AllocateArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1540 Node* res = SafePointNode::Ideal(phase, can_reshape); 1541 if (res != NULL) { 1542 return res; 1543 } 1544 // Don't bother trying to transform a dead node 1545 if (in(0) && in(0)->is_top()) return NULL; 1546 1547 const Type* type = phase->type(Ideal_length()); 1548 if (type->isa_int() && type->is_int()->_hi < 0) { 1549 if (can_reshape) { 1550 PhaseIterGVN *igvn = phase->is_IterGVN(); 1551 // Unreachable fall through path (negative array length), 1552 // the allocation can only throw so disconnect it. 1553 Node* proj = proj_out_or_null(TypeFunc::Control); 1554 Node* catchproj = NULL; 1555 if (proj != NULL) { 1556 for (DUIterator_Fast imax, i = proj->fast_outs(imax); i < imax; i++) { 1557 Node *cn = proj->fast_out(i); 1558 if (cn->is_Catch()) { 1559 catchproj = cn->as_Multi()->proj_out_or_null(CatchProjNode::fall_through_index); 1560 break; 1561 } 1562 } 1563 } 1564 if (catchproj != NULL && catchproj->outcnt() > 0 && 1565 (catchproj->outcnt() > 1 || 1566 catchproj->unique_out()->Opcode() != Op_Halt)) { 1567 assert(catchproj->is_CatchProj(), "must be a CatchProjNode"); 1568 Node* nproj = catchproj->clone(); 1569 igvn->register_new_node_with_optimizer(nproj); 1570 1571 Node *frame = new ParmNode( phase->C->start(), TypeFunc::FramePtr ); 1572 frame = phase->transform(frame); 1573 // Halt & Catch Fire 1574 Node *halt = new HaltNode( nproj, frame ); 1575 phase->C->root()->add_req(halt); 1576 phase->transform(halt); 1577 1578 igvn->replace_node(catchproj, phase->C->top()); 1579 return this; 1580 } 1581 } else { 1582 // Can't correct it during regular GVN so register for IGVN 1583 phase->C->record_for_igvn(this); 1584 } 1585 } 1586 return NULL; 1587 } 1588 1589 // Retrieve the length from the AllocateArrayNode. Narrow the type with a 1590 // CastII, if appropriate. If we are not allowed to create new nodes, and 1591 // a CastII is appropriate, return NULL. 1592 Node *AllocateArrayNode::make_ideal_length(const TypeOopPtr* oop_type, PhaseTransform *phase, bool allow_new_nodes) { 1593 Node *length = in(AllocateNode::ALength); 1594 assert(length != NULL, "length is not null"); 1595 1596 const TypeInt* length_type = phase->find_int_type(length); 1597 const TypeAryPtr* ary_type = oop_type->isa_aryptr(); 1598 1599 if (ary_type != NULL && length_type != NULL) { 1600 const TypeInt* narrow_length_type = ary_type->narrow_size_type(length_type); 1601 if (narrow_length_type != length_type) { 1602 // Assert one of: 1603 // - the narrow_length is 0 1604 // - the narrow_length is not wider than length 1605 assert(narrow_length_type == TypeInt::ZERO || 1606 length_type->is_con() && narrow_length_type->is_con() && 1607 (narrow_length_type->_hi <= length_type->_lo) || 1608 (narrow_length_type->_hi <= length_type->_hi && 1609 narrow_length_type->_lo >= length_type->_lo), 1610 "narrow type must be narrower than length type"); 1611 1612 // Return NULL if new nodes are not allowed 1613 if (!allow_new_nodes) return NULL; 1614 // Create a cast which is control dependent on the initialization to 1615 // propagate the fact that the array length must be positive. 1616 InitializeNode* init = initialization(); 1617 assert(init != NULL, "initialization not found"); 1618 length = new CastIINode(length, narrow_length_type); 1619 length->set_req(0, init->proj_out_or_null(0)); 1620 } 1621 } 1622 1623 return length; 1624 } 1625 1626 //============================================================================= 1627 uint LockNode::size_of() const { return sizeof(*this); } 1628 1629 // Redundant lock elimination 1630 // 1631 // There are various patterns of locking where we release and 1632 // immediately reacquire a lock in a piece of code where no operations 1633 // occur in between that would be observable. In those cases we can 1634 // skip releasing and reacquiring the lock without violating any 1635 // fairness requirements. Doing this around a loop could cause a lock 1636 // to be held for a very long time so we concentrate on non-looping 1637 // control flow. We also require that the operations are fully 1638 // redundant meaning that we don't introduce new lock operations on 1639 // some paths so to be able to eliminate it on others ala PRE. This 1640 // would probably require some more extensive graph manipulation to 1641 // guarantee that the memory edges were all handled correctly. 1642 // 1643 // Assuming p is a simple predicate which can't trap in any way and s 1644 // is a synchronized method consider this code: 1645 // 1646 // s(); 1647 // if (p) 1648 // s(); 1649 // else 1650 // s(); 1651 // s(); 1652 // 1653 // 1. The unlocks of the first call to s can be eliminated if the 1654 // locks inside the then and else branches are eliminated. 1655 // 1656 // 2. The unlocks of the then and else branches can be eliminated if 1657 // the lock of the final call to s is eliminated. 1658 // 1659 // Either of these cases subsumes the simple case of sequential control flow 1660 // 1661 // Addtionally we can eliminate versions without the else case: 1662 // 1663 // s(); 1664 // if (p) 1665 // s(); 1666 // s(); 1667 // 1668 // 3. In this case we eliminate the unlock of the first s, the lock 1669 // and unlock in the then case and the lock in the final s. 1670 // 1671 // Note also that in all these cases the then/else pieces don't have 1672 // to be trivial as long as they begin and end with synchronization 1673 // operations. 1674 // 1675 // s(); 1676 // if (p) 1677 // s(); 1678 // f(); 1679 // s(); 1680 // s(); 1681 // 1682 // The code will work properly for this case, leaving in the unlock 1683 // before the call to f and the relock after it. 1684 // 1685 // A potentially interesting case which isn't handled here is when the 1686 // locking is partially redundant. 1687 // 1688 // s(); 1689 // if (p) 1690 // s(); 1691 // 1692 // This could be eliminated putting unlocking on the else case and 1693 // eliminating the first unlock and the lock in the then side. 1694 // Alternatively the unlock could be moved out of the then side so it 1695 // was after the merge and the first unlock and second lock 1696 // eliminated. This might require less manipulation of the memory 1697 // state to get correct. 1698 // 1699 // Additionally we might allow work between a unlock and lock before 1700 // giving up eliminating the locks. The current code disallows any 1701 // conditional control flow between these operations. A formulation 1702 // similar to partial redundancy elimination computing the 1703 // availability of unlocking and the anticipatability of locking at a 1704 // program point would allow detection of fully redundant locking with 1705 // some amount of work in between. I'm not sure how often I really 1706 // think that would occur though. Most of the cases I've seen 1707 // indicate it's likely non-trivial work would occur in between. 1708 // There may be other more complicated constructs where we could 1709 // eliminate locking but I haven't seen any others appear as hot or 1710 // interesting. 1711 // 1712 // Locking and unlocking have a canonical form in ideal that looks 1713 // roughly like this: 1714 // 1715 // <obj> 1716 // | \\------+ 1717 // | \ \ 1718 // | BoxLock \ 1719 // | | | \ 1720 // | | \ \ 1721 // | | FastLock 1722 // | | / 1723 // | | / 1724 // | | | 1725 // 1726 // Lock 1727 // | 1728 // Proj #0 1729 // | 1730 // MembarAcquire 1731 // | 1732 // Proj #0 1733 // 1734 // MembarRelease 1735 // | 1736 // Proj #0 1737 // | 1738 // Unlock 1739 // | 1740 // Proj #0 1741 // 1742 // 1743 // This code proceeds by processing Lock nodes during PhaseIterGVN 1744 // and searching back through its control for the proper code 1745 // patterns. Once it finds a set of lock and unlock operations to 1746 // eliminate they are marked as eliminatable which causes the 1747 // expansion of the Lock and Unlock macro nodes to make the operation a NOP 1748 // 1749 //============================================================================= 1750 1751 // 1752 // Utility function to skip over uninteresting control nodes. Nodes skipped are: 1753 // - copy regions. (These may not have been optimized away yet.) 1754 // - eliminated locking nodes 1755 // 1756 static Node *next_control(Node *ctrl) { 1757 if (ctrl == NULL) 1758 return NULL; 1759 while (1) { 1760 if (ctrl->is_Region()) { 1761 RegionNode *r = ctrl->as_Region(); 1762 Node *n = r->is_copy(); 1763 if (n == NULL) 1764 break; // hit a region, return it 1765 else 1766 ctrl = n; 1767 } else if (ctrl->is_Proj()) { 1768 Node *in0 = ctrl->in(0); 1769 if (in0->is_AbstractLock() && in0->as_AbstractLock()->is_eliminated()) { 1770 ctrl = in0->in(0); 1771 } else { 1772 break; 1773 } 1774 } else { 1775 break; // found an interesting control 1776 } 1777 } 1778 return ctrl; 1779 } 1780 // 1781 // Given a control, see if it's the control projection of an Unlock which 1782 // operating on the same object as lock. 1783 // 1784 bool AbstractLockNode::find_matching_unlock(const Node* ctrl, LockNode* lock, 1785 GrowableArray<AbstractLockNode*> &lock_ops) { 1786 ProjNode *ctrl_proj = (ctrl->is_Proj()) ? ctrl->as_Proj() : NULL; 1787 if (ctrl_proj != NULL && ctrl_proj->_con == TypeFunc::Control) { 1788 Node *n = ctrl_proj->in(0); 1789 if (n != NULL && n->is_Unlock()) { 1790 UnlockNode *unlock = n->as_Unlock(); 1791 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1792 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node()); 1793 Node* unlock_obj = bs->step_over_gc_barrier(unlock->obj_node()); 1794 if (lock_obj->eqv_uncast(unlock_obj) && 1795 BoxLockNode::same_slot(lock->box_node(), unlock->box_node()) && 1796 !unlock->is_eliminated()) { 1797 lock_ops.append(unlock); 1798 return true; 1799 } 1800 } 1801 } 1802 return false; 1803 } 1804 1805 // 1806 // Find the lock matching an unlock. Returns null if a safepoint 1807 // or complicated control is encountered first. 1808 LockNode *AbstractLockNode::find_matching_lock(UnlockNode* unlock) { 1809 LockNode *lock_result = NULL; 1810 // find the matching lock, or an intervening safepoint 1811 Node *ctrl = next_control(unlock->in(0)); 1812 while (1) { 1813 assert(ctrl != NULL, "invalid control graph"); 1814 assert(!ctrl->is_Start(), "missing lock for unlock"); 1815 if (ctrl->is_top()) break; // dead control path 1816 if (ctrl->is_Proj()) ctrl = ctrl->in(0); 1817 if (ctrl->is_SafePoint()) { 1818 break; // found a safepoint (may be the lock we are searching for) 1819 } else if (ctrl->is_Region()) { 1820 // Check for a simple diamond pattern. Punt on anything more complicated 1821 if (ctrl->req() == 3 && ctrl->in(1) != NULL && ctrl->in(2) != NULL) { 1822 Node *in1 = next_control(ctrl->in(1)); 1823 Node *in2 = next_control(ctrl->in(2)); 1824 if (((in1->is_IfTrue() && in2->is_IfFalse()) || 1825 (in2->is_IfTrue() && in1->is_IfFalse())) && (in1->in(0) == in2->in(0))) { 1826 ctrl = next_control(in1->in(0)->in(0)); 1827 } else { 1828 break; 1829 } 1830 } else { 1831 break; 1832 } 1833 } else { 1834 ctrl = next_control(ctrl->in(0)); // keep searching 1835 } 1836 } 1837 if (ctrl->is_Lock()) { 1838 LockNode *lock = ctrl->as_Lock(); 1839 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1840 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node()); 1841 Node* unlock_obj = bs->step_over_gc_barrier(unlock->obj_node()); 1842 if (lock_obj->eqv_uncast(unlock_obj) && 1843 BoxLockNode::same_slot(lock->box_node(), unlock->box_node())) { 1844 lock_result = lock; 1845 } 1846 } 1847 return lock_result; 1848 } 1849 1850 // This code corresponds to case 3 above. 1851 1852 bool AbstractLockNode::find_lock_and_unlock_through_if(Node* node, LockNode* lock, 1853 GrowableArray<AbstractLockNode*> &lock_ops) { 1854 Node* if_node = node->in(0); 1855 bool if_true = node->is_IfTrue(); 1856 1857 if (if_node->is_If() && if_node->outcnt() == 2 && (if_true || node->is_IfFalse())) { 1858 Node *lock_ctrl = next_control(if_node->in(0)); 1859 if (find_matching_unlock(lock_ctrl, lock, lock_ops)) { 1860 Node* lock1_node = NULL; 1861 ProjNode* proj = if_node->as_If()->proj_out(!if_true); 1862 if (if_true) { 1863 if (proj->is_IfFalse() && proj->outcnt() == 1) { 1864 lock1_node = proj->unique_out(); 1865 } 1866 } else { 1867 if (proj->is_IfTrue() && proj->outcnt() == 1) { 1868 lock1_node = proj->unique_out(); 1869 } 1870 } 1871 if (lock1_node != NULL && lock1_node->is_Lock()) { 1872 LockNode *lock1 = lock1_node->as_Lock(); 1873 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1874 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node()); 1875 Node* lock1_obj = bs->step_over_gc_barrier(lock1->obj_node()); 1876 if (lock_obj->eqv_uncast(lock1_obj) && 1877 BoxLockNode::same_slot(lock->box_node(), lock1->box_node()) && 1878 !lock1->is_eliminated()) { 1879 lock_ops.append(lock1); 1880 return true; 1881 } 1882 } 1883 } 1884 } 1885 1886 lock_ops.trunc_to(0); 1887 return false; 1888 } 1889 1890 bool AbstractLockNode::find_unlocks_for_region(const RegionNode* region, LockNode* lock, 1891 GrowableArray<AbstractLockNode*> &lock_ops) { 1892 // check each control merging at this point for a matching unlock. 1893 // in(0) should be self edge so skip it. 1894 for (int i = 1; i < (int)region->req(); i++) { 1895 Node *in_node = next_control(region->in(i)); 1896 if (in_node != NULL) { 1897 if (find_matching_unlock(in_node, lock, lock_ops)) { 1898 // found a match so keep on checking. 1899 continue; 1900 } else if (find_lock_and_unlock_through_if(in_node, lock, lock_ops)) { 1901 continue; 1902 } 1903 1904 // If we fall through to here then it was some kind of node we 1905 // don't understand or there wasn't a matching unlock, so give 1906 // up trying to merge locks. 1907 lock_ops.trunc_to(0); 1908 return false; 1909 } 1910 } 1911 return true; 1912 1913 } 1914 1915 #ifndef PRODUCT 1916 // 1917 // Create a counter which counts the number of times this lock is acquired 1918 // 1919 void AbstractLockNode::create_lock_counter(JVMState* state) { 1920 _counter = OptoRuntime::new_named_counter(state, NamedCounter::LockCounter); 1921 } 1922 1923 void AbstractLockNode::set_eliminated_lock_counter() { 1924 if (_counter) { 1925 // Update the counter to indicate that this lock was eliminated. 1926 // The counter update code will stay around even though the 1927 // optimizer will eliminate the lock operation itself. 1928 _counter->set_tag(NamedCounter::EliminatedLockCounter); 1929 } 1930 } 1931 1932 const char* AbstractLockNode::_kind_names[] = {"Regular", "NonEscObj", "Coarsened", "Nested"}; 1933 1934 void AbstractLockNode::dump_spec(outputStream* st) const { 1935 st->print("%s ", _kind_names[_kind]); 1936 CallNode::dump_spec(st); 1937 } 1938 1939 void AbstractLockNode::dump_compact_spec(outputStream* st) const { 1940 st->print("%s", _kind_names[_kind]); 1941 } 1942 1943 // The related set of lock nodes includes the control boundary. 1944 void AbstractLockNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const { 1945 if (compact) { 1946 this->collect_nodes(in_rel, 1, false, false); 1947 } else { 1948 this->collect_nodes_in_all_data(in_rel, true); 1949 } 1950 this->collect_nodes(out_rel, -2, false, false); 1951 } 1952 #endif 1953 1954 //============================================================================= 1955 Node *LockNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1956 1957 // perform any generic optimizations first (returns 'this' or NULL) 1958 Node *result = SafePointNode::Ideal(phase, can_reshape); 1959 if (result != NULL) return result; 1960 // Don't bother trying to transform a dead node 1961 if (in(0) && in(0)->is_top()) return NULL; 1962 1963 // Now see if we can optimize away this lock. We don't actually 1964 // remove the locking here, we simply set the _eliminate flag which 1965 // prevents macro expansion from expanding the lock. Since we don't 1966 // modify the graph, the value returned from this function is the 1967 // one computed above. 1968 const Type* obj_type = phase->type(obj_node()); 1969 if (can_reshape && EliminateLocks && !is_non_esc_obj() && 1970 !obj_type->isa_valuetype() && !obj_type->is_valuetypeptr()) { 1971 // 1972 // If we are locking an unescaped object, the lock/unlock is unnecessary 1973 // 1974 ConnectionGraph *cgr = phase->C->congraph(); 1975 if (cgr != NULL && cgr->not_global_escape(obj_node())) { 1976 assert(!is_eliminated() || is_coarsened(), "sanity"); 1977 // The lock could be marked eliminated by lock coarsening 1978 // code during first IGVN before EA. Replace coarsened flag 1979 // to eliminate all associated locks/unlocks. 1980 #ifdef ASSERT 1981 this->log_lock_optimization(phase->C,"eliminate_lock_set_non_esc1"); 1982 #endif 1983 this->set_non_esc_obj(); 1984 return result; 1985 } 1986 1987 // 1988 // Try lock coarsening 1989 // 1990 PhaseIterGVN* iter = phase->is_IterGVN(); 1991 if (iter != NULL && !is_eliminated()) { 1992 1993 GrowableArray<AbstractLockNode*> lock_ops; 1994 1995 Node *ctrl = next_control(in(0)); 1996 1997 // now search back for a matching Unlock 1998 if (find_matching_unlock(ctrl, this, lock_ops)) { 1999 // found an unlock directly preceding this lock. This is the 2000 // case of single unlock directly control dependent on a 2001 // single lock which is the trivial version of case 1 or 2. 2002 } else if (ctrl->is_Region() ) { 2003 if (find_unlocks_for_region(ctrl->as_Region(), this, lock_ops)) { 2004 // found lock preceded by multiple unlocks along all paths 2005 // joining at this point which is case 3 in description above. 2006 } 2007 } else { 2008 // see if this lock comes from either half of an if and the 2009 // predecessors merges unlocks and the other half of the if 2010 // performs a lock. 2011 if (find_lock_and_unlock_through_if(ctrl, this, lock_ops)) { 2012 // found unlock splitting to an if with locks on both branches. 2013 } 2014 } 2015 2016 if (lock_ops.length() > 0) { 2017 // add ourselves to the list of locks to be eliminated. 2018 lock_ops.append(this); 2019 2020 #ifndef PRODUCT 2021 if (PrintEliminateLocks) { 2022 int locks = 0; 2023 int unlocks = 0; 2024 for (int i = 0; i < lock_ops.length(); i++) { 2025 AbstractLockNode* lock = lock_ops.at(i); 2026 if (lock->Opcode() == Op_Lock) 2027 locks++; 2028 else 2029 unlocks++; 2030 if (Verbose) { 2031 lock->dump(1); 2032 } 2033 } 2034 tty->print_cr("***Eliminated %d unlocks and %d locks", unlocks, locks); 2035 } 2036 #endif 2037 2038 // for each of the identified locks, mark them 2039 // as eliminatable 2040 for (int i = 0; i < lock_ops.length(); i++) { 2041 AbstractLockNode* lock = lock_ops.at(i); 2042 2043 // Mark it eliminated by coarsening and update any counters 2044 #ifdef ASSERT 2045 lock->log_lock_optimization(phase->C, "eliminate_lock_set_coarsened"); 2046 #endif 2047 lock->set_coarsened(); 2048 } 2049 } else if (ctrl->is_Region() && 2050 iter->_worklist.member(ctrl)) { 2051 // We weren't able to find any opportunities but the region this 2052 // lock is control dependent on hasn't been processed yet so put 2053 // this lock back on the worklist so we can check again once any 2054 // region simplification has occurred. 2055 iter->_worklist.push(this); 2056 } 2057 } 2058 } 2059 2060 return result; 2061 } 2062 2063 //============================================================================= 2064 bool LockNode::is_nested_lock_region() { 2065 return is_nested_lock_region(NULL); 2066 } 2067 2068 // p is used for access to compilation log; no logging if NULL 2069 bool LockNode::is_nested_lock_region(Compile * c) { 2070 BoxLockNode* box = box_node()->as_BoxLock(); 2071 int stk_slot = box->stack_slot(); 2072 if (stk_slot <= 0) { 2073 #ifdef ASSERT 2074 this->log_lock_optimization(c, "eliminate_lock_INLR_1"); 2075 #endif 2076 return false; // External lock or it is not Box (Phi node). 2077 } 2078 2079 // Ignore complex cases: merged locks or multiple locks. 2080 Node* obj = obj_node(); 2081 LockNode* unique_lock = NULL; 2082 if (!box->is_simple_lock_region(&unique_lock, obj)) { 2083 #ifdef ASSERT 2084 this->log_lock_optimization(c, "eliminate_lock_INLR_2a"); 2085 #endif 2086 return false; 2087 } 2088 if (unique_lock != this) { 2089 #ifdef ASSERT 2090 this->log_lock_optimization(c, "eliminate_lock_INLR_2b"); 2091 #endif 2092 return false; 2093 } 2094 2095 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2096 obj = bs->step_over_gc_barrier(obj); 2097 // Look for external lock for the same object. 2098 SafePointNode* sfn = this->as_SafePoint(); 2099 JVMState* youngest_jvms = sfn->jvms(); 2100 int max_depth = youngest_jvms->depth(); 2101 for (int depth = 1; depth <= max_depth; depth++) { 2102 JVMState* jvms = youngest_jvms->of_depth(depth); 2103 int num_mon = jvms->nof_monitors(); 2104 // Loop over monitors 2105 for (int idx = 0; idx < num_mon; idx++) { 2106 Node* obj_node = sfn->monitor_obj(jvms, idx); 2107 obj_node = bs->step_over_gc_barrier(obj_node); 2108 BoxLockNode* box_node = sfn->monitor_box(jvms, idx)->as_BoxLock(); 2109 if ((box_node->stack_slot() < stk_slot) && obj_node->eqv_uncast(obj)) { 2110 return true; 2111 } 2112 } 2113 } 2114 #ifdef ASSERT 2115 this->log_lock_optimization(c, "eliminate_lock_INLR_3"); 2116 #endif 2117 return false; 2118 } 2119 2120 //============================================================================= 2121 uint UnlockNode::size_of() const { return sizeof(*this); } 2122 2123 //============================================================================= 2124 Node *UnlockNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2125 2126 // perform any generic optimizations first (returns 'this' or NULL) 2127 Node *result = SafePointNode::Ideal(phase, can_reshape); 2128 if (result != NULL) return result; 2129 // Don't bother trying to transform a dead node 2130 if (in(0) && in(0)->is_top()) return NULL; 2131 2132 // Now see if we can optimize away this unlock. We don't actually 2133 // remove the unlocking here, we simply set the _eliminate flag which 2134 // prevents macro expansion from expanding the unlock. Since we don't 2135 // modify the graph, the value returned from this function is the 2136 // one computed above. 2137 // Escape state is defined after Parse phase. 2138 const Type* obj_type = phase->type(obj_node()); 2139 if (can_reshape && EliminateLocks && !is_non_esc_obj() && 2140 !obj_type->isa_valuetype() && !obj_type->is_valuetypeptr()) { 2141 // 2142 // If we are unlocking an unescaped object, the lock/unlock is unnecessary. 2143 // 2144 ConnectionGraph *cgr = phase->C->congraph(); 2145 if (cgr != NULL && cgr->not_global_escape(obj_node())) { 2146 assert(!is_eliminated() || is_coarsened(), "sanity"); 2147 // The lock could be marked eliminated by lock coarsening 2148 // code during first IGVN before EA. Replace coarsened flag 2149 // to eliminate all associated locks/unlocks. 2150 #ifdef ASSERT 2151 this->log_lock_optimization(phase->C, "eliminate_lock_set_non_esc2"); 2152 #endif 2153 this->set_non_esc_obj(); 2154 } 2155 } 2156 return result; 2157 } 2158 2159 const char * AbstractLockNode::kind_as_string() const { 2160 return is_coarsened() ? "coarsened" : 2161 is_nested() ? "nested" : 2162 is_non_esc_obj() ? "non_escaping" : 2163 "?"; 2164 } 2165 2166 void AbstractLockNode::log_lock_optimization(Compile *C, const char * tag) const { 2167 if (C == NULL) { 2168 return; 2169 } 2170 CompileLog* log = C->log(); 2171 if (log != NULL) { 2172 log->begin_head("%s lock='%d' compile_id='%d' class_id='%s' kind='%s'", 2173 tag, is_Lock(), C->compile_id(), 2174 is_Unlock() ? "unlock" : is_Lock() ? "lock" : "?", 2175 kind_as_string()); 2176 log->stamp(); 2177 log->end_head(); 2178 JVMState* p = is_Unlock() ? (as_Unlock()->dbg_jvms()) : jvms(); 2179 while (p != NULL) { 2180 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 2181 p = p->caller(); 2182 } 2183 log->tail(tag); 2184 } 2185 } 2186 2187 bool CallNode::may_modify_arraycopy_helper(const TypeOopPtr* dest_t, const TypeOopPtr *t_oop, PhaseTransform *phase) { 2188 if (dest_t->is_known_instance() && t_oop->is_known_instance()) { 2189 return dest_t->instance_id() == t_oop->instance_id(); 2190 } 2191 2192 if (dest_t->isa_instptr() && !dest_t->klass()->equals(phase->C->env()->Object_klass())) { 2193 // clone 2194 if (t_oop->isa_aryptr()) { 2195 return false; 2196 } 2197 if (!t_oop->isa_instptr()) { 2198 return true; 2199 } 2200 if (dest_t->klass()->is_subtype_of(t_oop->klass()) || t_oop->klass()->is_subtype_of(dest_t->klass())) { 2201 return true; 2202 } 2203 // unrelated 2204 return false; 2205 } 2206 2207 if (dest_t->isa_aryptr()) { 2208 // arraycopy or array clone 2209 if (t_oop->isa_instptr()) { 2210 return false; 2211 } 2212 if (!t_oop->isa_aryptr()) { 2213 return true; 2214 } 2215 2216 const Type* elem = dest_t->is_aryptr()->elem(); 2217 if (elem == Type::BOTTOM) { 2218 // An array but we don't know what elements are 2219 return true; 2220 } 2221 2222 dest_t = dest_t->is_aryptr()->with_field_offset(Type::OffsetBot)->add_offset(Type::OffsetBot)->is_oopptr(); 2223 t_oop = t_oop->is_aryptr()->with_field_offset(Type::OffsetBot); 2224 uint dest_alias = phase->C->get_alias_index(dest_t); 2225 uint t_oop_alias = phase->C->get_alias_index(t_oop); 2226 2227 return dest_alias == t_oop_alias; 2228 } 2229 2230 return true; 2231 }