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