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