1 /* 2 * Copyright 1997-2009 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 20 * CA 95054 USA or visit www.sun.com if you need additional information or 21 * have any questions. 22 * 23 */ 24 25 // Portions of code courtesy of Clifford Click 26 27 // Optimization - Graph Style 28 29 #include "incls/_precompiled.incl" 30 #include "incls/_callnode.cpp.incl" 31 32 //============================================================================= 33 uint StartNode::size_of() const { return sizeof(*this); } 34 uint StartNode::cmp( const Node &n ) const 35 { return _domain == ((StartNode&)n)._domain; } 36 const Type *StartNode::bottom_type() const { return _domain; } 37 const Type *StartNode::Value(PhaseTransform *phase) const { return _domain; } 38 #ifndef PRODUCT 39 void StartNode::dump_spec(outputStream *st) const { st->print(" #"); _domain->dump_on(st);} 40 #endif 41 42 //------------------------------Ideal------------------------------------------ 43 Node *StartNode::Ideal(PhaseGVN *phase, bool can_reshape){ 44 return remove_dead_region(phase, can_reshape) ? this : NULL; 45 } 46 47 //------------------------------calling_convention----------------------------- 48 void StartNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const { 49 Matcher::calling_convention( sig_bt, parm_regs, argcnt, false ); 50 } 51 52 //------------------------------Registers-------------------------------------- 53 const RegMask &StartNode::in_RegMask(uint) const { 54 return RegMask::Empty; 55 } 56 57 //------------------------------match------------------------------------------ 58 // Construct projections for incoming parameters, and their RegMask info 59 Node *StartNode::match( const ProjNode *proj, const Matcher *match ) { 60 switch (proj->_con) { 61 case TypeFunc::Control: 62 case TypeFunc::I_O: 63 case TypeFunc::Memory: 64 return new (match->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 65 case TypeFunc::FramePtr: 66 return new (match->C, 1) MachProjNode(this,proj->_con,Matcher::c_frame_ptr_mask, Op_RegP); 67 case TypeFunc::ReturnAdr: 68 return new (match->C, 1) MachProjNode(this,proj->_con,match->_return_addr_mask,Op_RegP); 69 case TypeFunc::Parms: 70 default: { 71 uint parm_num = proj->_con - TypeFunc::Parms; 72 const Type *t = _domain->field_at(proj->_con); 73 if (t->base() == Type::Half) // 2nd half of Longs and Doubles 74 return new (match->C, 1) ConNode(Type::TOP); 75 uint ideal_reg = Matcher::base2reg[t->base()]; 76 RegMask &rm = match->_calling_convention_mask[parm_num]; 77 return new (match->C, 1) MachProjNode(this,proj->_con,rm,ideal_reg); 78 } 79 } 80 return NULL; 81 } 82 83 //------------------------------StartOSRNode---------------------------------- 84 // The method start node for an on stack replacement adapter 85 86 //------------------------------osr_domain----------------------------- 87 const TypeTuple *StartOSRNode::osr_domain() { 88 const Type **fields = TypeTuple::fields(2); 89 fields[TypeFunc::Parms+0] = TypeRawPtr::BOTTOM; // address of osr buffer 90 91 return TypeTuple::make(TypeFunc::Parms+1, fields); 92 } 93 94 //============================================================================= 95 const char * const ParmNode::names[TypeFunc::Parms+1] = { 96 "Control", "I_O", "Memory", "FramePtr", "ReturnAdr", "Parms" 97 }; 98 99 #ifndef PRODUCT 100 void ParmNode::dump_spec(outputStream *st) const { 101 if( _con < TypeFunc::Parms ) { 102 st->print(names[_con]); 103 } else { 104 st->print("Parm%d: ",_con-TypeFunc::Parms); 105 // Verbose and WizardMode dump bottom_type for all nodes 106 if( !Verbose && !WizardMode ) bottom_type()->dump_on(st); 107 } 108 } 109 #endif 110 111 uint ParmNode::ideal_reg() const { 112 switch( _con ) { 113 case TypeFunc::Control : // fall through 114 case TypeFunc::I_O : // fall through 115 case TypeFunc::Memory : return 0; 116 case TypeFunc::FramePtr : // fall through 117 case TypeFunc::ReturnAdr: return Op_RegP; 118 default : assert( _con > TypeFunc::Parms, "" ); 119 // fall through 120 case TypeFunc::Parms : { 121 // Type of argument being passed 122 const Type *t = in(0)->as_Start()->_domain->field_at(_con); 123 return Matcher::base2reg[t->base()]; 124 } 125 } 126 ShouldNotReachHere(); 127 return 0; 128 } 129 130 //============================================================================= 131 ReturnNode::ReturnNode(uint edges, Node *cntrl, Node *i_o, Node *memory, Node *frameptr, Node *retadr ) : Node(edges) { 132 init_req(TypeFunc::Control,cntrl); 133 init_req(TypeFunc::I_O,i_o); 134 init_req(TypeFunc::Memory,memory); 135 init_req(TypeFunc::FramePtr,frameptr); 136 init_req(TypeFunc::ReturnAdr,retadr); 137 } 138 139 Node *ReturnNode::Ideal(PhaseGVN *phase, bool can_reshape){ 140 return remove_dead_region(phase, can_reshape) ? this : NULL; 141 } 142 143 const Type *ReturnNode::Value( PhaseTransform *phase ) const { 144 return ( phase->type(in(TypeFunc::Control)) == Type::TOP) 145 ? Type::TOP 146 : Type::BOTTOM; 147 } 148 149 // Do we Match on this edge index or not? No edges on return nodes 150 uint ReturnNode::match_edge(uint idx) const { 151 return 0; 152 } 153 154 155 #ifndef PRODUCT 156 void ReturnNode::dump_req() const { 157 // Dump the required inputs, enclosed in '(' and ')' 158 uint i; // Exit value of loop 159 for( i=0; i<req(); i++ ) { // For all required inputs 160 if( i == TypeFunc::Parms ) tty->print("returns"); 161 if( in(i) ) tty->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx); 162 else tty->print("_ "); 163 } 164 } 165 #endif 166 167 //============================================================================= 168 RethrowNode::RethrowNode( 169 Node* cntrl, 170 Node* i_o, 171 Node* memory, 172 Node* frameptr, 173 Node* ret_adr, 174 Node* exception 175 ) : Node(TypeFunc::Parms + 1) { 176 init_req(TypeFunc::Control , cntrl ); 177 init_req(TypeFunc::I_O , i_o ); 178 init_req(TypeFunc::Memory , memory ); 179 init_req(TypeFunc::FramePtr , frameptr ); 180 init_req(TypeFunc::ReturnAdr, ret_adr); 181 init_req(TypeFunc::Parms , exception); 182 } 183 184 Node *RethrowNode::Ideal(PhaseGVN *phase, bool can_reshape){ 185 return remove_dead_region(phase, can_reshape) ? this : NULL; 186 } 187 188 const Type *RethrowNode::Value( PhaseTransform *phase ) const { 189 return (phase->type(in(TypeFunc::Control)) == Type::TOP) 190 ? Type::TOP 191 : Type::BOTTOM; 192 } 193 194 uint RethrowNode::match_edge(uint idx) const { 195 return 0; 196 } 197 198 #ifndef PRODUCT 199 void RethrowNode::dump_req() const { 200 // Dump the required inputs, enclosed in '(' and ')' 201 uint i; // Exit value of loop 202 for( i=0; i<req(); i++ ) { // For all required inputs 203 if( i == TypeFunc::Parms ) tty->print("exception"); 204 if( in(i) ) tty->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx); 205 else tty->print("_ "); 206 } 207 } 208 #endif 209 210 //============================================================================= 211 // Do we Match on this edge index or not? Match only target address & method 212 uint TailCallNode::match_edge(uint idx) const { 213 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1; 214 } 215 216 //============================================================================= 217 // Do we Match on this edge index or not? Match only target address & oop 218 uint TailJumpNode::match_edge(uint idx) const { 219 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1; 220 } 221 222 //============================================================================= 223 JVMState::JVMState(ciMethod* method, JVMState* caller) { 224 assert(method != NULL, "must be valid call site"); 225 _method = method; 226 _reexecute = Reexecute_Undefined; 227 debug_only(_bci = -99); // random garbage value 228 debug_only(_map = (SafePointNode*)-1); 229 _caller = caller; 230 _depth = 1 + (caller == NULL ? 0 : caller->depth()); 231 _locoff = TypeFunc::Parms; 232 _stkoff = _locoff + _method->max_locals(); 233 _monoff = _stkoff + _method->max_stack(); 234 _scloff = _monoff; 235 _endoff = _monoff; 236 _sp = 0; 237 } 238 JVMState::JVMState(int stack_size) { 239 _method = NULL; 240 _bci = InvocationEntryBci; 241 _reexecute = Reexecute_Undefined; 242 debug_only(_map = (SafePointNode*)-1); 243 _caller = NULL; 244 _depth = 1; 245 _locoff = TypeFunc::Parms; 246 _stkoff = _locoff; 247 _monoff = _stkoff + stack_size; 248 _scloff = _monoff; 249 _endoff = _monoff; 250 _sp = 0; 251 } 252 253 //--------------------------------of_depth------------------------------------- 254 JVMState* JVMState::of_depth(int d) const { 255 const JVMState* jvmp = this; 256 assert(0 < d && (uint)d <= depth(), "oob"); 257 for (int skip = depth() - d; skip > 0; skip--) { 258 jvmp = jvmp->caller(); 259 } 260 assert(jvmp->depth() == (uint)d, "found the right one"); 261 return (JVMState*)jvmp; 262 } 263 264 //-----------------------------same_calls_as----------------------------------- 265 bool JVMState::same_calls_as(const JVMState* that) const { 266 if (this == that) return true; 267 if (this->depth() != that->depth()) return false; 268 const JVMState* p = this; 269 const JVMState* q = that; 270 for (;;) { 271 if (p->_method != q->_method) return false; 272 if (p->_method == NULL) return true; // bci is irrelevant 273 if (p->_bci != q->_bci) return false; 274 if (p->_reexecute != q->_reexecute) return false; 275 p = p->caller(); 276 q = q->caller(); 277 if (p == q) return true; 278 assert(p != NULL && q != NULL, "depth check ensures we don't run off end"); 279 } 280 } 281 282 //------------------------------debug_start------------------------------------ 283 uint JVMState::debug_start() const { 284 debug_only(JVMState* jvmroot = of_depth(1)); 285 assert(jvmroot->locoff() <= this->locoff(), "youngest JVMState must be last"); 286 return of_depth(1)->locoff(); 287 } 288 289 //-------------------------------debug_end------------------------------------- 290 uint JVMState::debug_end() const { 291 debug_only(JVMState* jvmroot = of_depth(1)); 292 assert(jvmroot->endoff() <= this->endoff(), "youngest JVMState must be last"); 293 return endoff(); 294 } 295 296 //------------------------------debug_depth------------------------------------ 297 uint JVMState::debug_depth() const { 298 uint total = 0; 299 for (const JVMState* jvmp = this; jvmp != NULL; jvmp = jvmp->caller()) { 300 total += jvmp->debug_size(); 301 } 302 return total; 303 } 304 305 #ifndef PRODUCT 306 307 //------------------------------format_helper---------------------------------- 308 // Given an allocation (a Chaitin object) and a Node decide if the Node carries 309 // any defined value or not. If it does, print out the register or constant. 310 static void format_helper( PhaseRegAlloc *regalloc, outputStream* st, Node *n, const char *msg, uint i, GrowableArray<SafePointScalarObjectNode*> *scobjs ) { 311 if (n == NULL) { st->print(" NULL"); return; } 312 if (n->is_SafePointScalarObject()) { 313 // Scalar replacement. 314 SafePointScalarObjectNode* spobj = n->as_SafePointScalarObject(); 315 scobjs->append_if_missing(spobj); 316 int sco_n = scobjs->find(spobj); 317 assert(sco_n >= 0, ""); 318 st->print(" %s%d]=#ScObj" INT32_FORMAT, msg, i, sco_n); 319 return; 320 } 321 if( OptoReg::is_valid(regalloc->get_reg_first(n))) { // Check for undefined 322 char buf[50]; 323 regalloc->dump_register(n,buf); 324 st->print(" %s%d]=%s",msg,i,buf); 325 } else { // No register, but might be constant 326 const Type *t = n->bottom_type(); 327 switch (t->base()) { 328 case Type::Int: 329 st->print(" %s%d]=#"INT32_FORMAT,msg,i,t->is_int()->get_con()); 330 break; 331 case Type::AnyPtr: 332 assert( t == TypePtr::NULL_PTR, "" ); 333 st->print(" %s%d]=#NULL",msg,i); 334 break; 335 case Type::AryPtr: 336 case Type::KlassPtr: 337 case Type::InstPtr: 338 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,t->isa_oopptr()->const_oop()); 339 break; 340 case Type::NarrowOop: 341 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,t->make_ptr()->isa_oopptr()->const_oop()); 342 break; 343 case Type::RawPtr: 344 st->print(" %s%d]=#Raw" INTPTR_FORMAT,msg,i,t->is_rawptr()); 345 break; 346 case Type::DoubleCon: 347 st->print(" %s%d]=#%fD",msg,i,t->is_double_constant()->_d); 348 break; 349 case Type::FloatCon: 350 st->print(" %s%d]=#%fF",msg,i,t->is_float_constant()->_f); 351 break; 352 case Type::Long: 353 st->print(" %s%d]=#"INT64_FORMAT,msg,i,t->is_long()->get_con()); 354 break; 355 case Type::Half: 356 case Type::Top: 357 st->print(" %s%d]=_",msg,i); 358 break; 359 default: ShouldNotReachHere(); 360 } 361 } 362 } 363 364 //------------------------------format----------------------------------------- 365 void JVMState::format(PhaseRegAlloc *regalloc, const Node *n, outputStream* st) const { 366 st->print(" #"); 367 if( _method ) { 368 _method->print_short_name(st); 369 st->print(" @ bci:%d ",_bci); 370 } else { 371 st->print_cr(" runtime stub "); 372 return; 373 } 374 if (n->is_MachSafePoint()) { 375 GrowableArray<SafePointScalarObjectNode*> scobjs; 376 MachSafePointNode *mcall = n->as_MachSafePoint(); 377 uint i; 378 // Print locals 379 for( i = 0; i < (uint)loc_size(); i++ ) 380 format_helper( regalloc, st, mcall->local(this, i), "L[", i, &scobjs ); 381 // Print stack 382 for (i = 0; i < (uint)stk_size(); i++) { 383 if ((uint)(_stkoff + i) >= mcall->len()) 384 st->print(" oob "); 385 else 386 format_helper( regalloc, st, mcall->stack(this, i), "STK[", i, &scobjs ); 387 } 388 for (i = 0; (int)i < nof_monitors(); i++) { 389 Node *box = mcall->monitor_box(this, i); 390 Node *obj = mcall->monitor_obj(this, i); 391 if ( OptoReg::is_valid(regalloc->get_reg_first(box)) ) { 392 while( !box->is_BoxLock() ) box = box->in(1); 393 format_helper( regalloc, st, box, "MON-BOX[", i, &scobjs ); 394 } else { 395 OptoReg::Name box_reg = BoxLockNode::stack_slot(box); 396 st->print(" MON-BOX%d=%s+%d", 397 i, 398 OptoReg::regname(OptoReg::c_frame_pointer), 399 regalloc->reg2offset(box_reg)); 400 } 401 const char* obj_msg = "MON-OBJ["; 402 if (EliminateLocks) { 403 while( !box->is_BoxLock() ) box = box->in(1); 404 if (box->as_BoxLock()->is_eliminated()) 405 obj_msg = "MON-OBJ(LOCK ELIMINATED)["; 406 } 407 format_helper( regalloc, st, obj, obj_msg, i, &scobjs ); 408 } 409 410 for (i = 0; i < (uint)scobjs.length(); i++) { 411 // Scalar replaced objects. 412 st->print_cr(""); 413 st->print(" # ScObj" INT32_FORMAT " ", i); 414 SafePointScalarObjectNode* spobj = scobjs.at(i); 415 ciKlass* cik = spobj->bottom_type()->is_oopptr()->klass(); 416 assert(cik->is_instance_klass() || 417 cik->is_array_klass(), "Not supported allocation."); 418 ciInstanceKlass *iklass = NULL; 419 if (cik->is_instance_klass()) { 420 cik->print_name_on(st); 421 iklass = cik->as_instance_klass(); 422 } else if (cik->is_type_array_klass()) { 423 cik->as_array_klass()->base_element_type()->print_name_on(st); 424 st->print("[%d]", spobj->n_fields()); 425 } else if (cik->is_obj_array_klass()) { 426 ciKlass* cie = cik->as_obj_array_klass()->base_element_klass(); 427 if (cie->is_instance_klass()) { 428 cie->print_name_on(st); 429 } else if (cie->is_type_array_klass()) { 430 cie->as_array_klass()->base_element_type()->print_name_on(st); 431 } else { 432 ShouldNotReachHere(); 433 } 434 st->print("[%d]", spobj->n_fields()); 435 int ndim = cik->as_array_klass()->dimension() - 1; 436 while (ndim-- > 0) { 437 st->print("[]"); 438 } 439 } 440 st->print("={"); 441 uint nf = spobj->n_fields(); 442 if (nf > 0) { 443 uint first_ind = spobj->first_index(); 444 Node* fld_node = mcall->in(first_ind); 445 ciField* cifield; 446 if (iklass != NULL) { 447 st->print(" ["); 448 cifield = iklass->nonstatic_field_at(0); 449 cifield->print_name_on(st); 450 format_helper( regalloc, st, fld_node, ":", 0, &scobjs ); 451 } else { 452 format_helper( regalloc, st, fld_node, "[", 0, &scobjs ); 453 } 454 for (uint j = 1; j < nf; j++) { 455 fld_node = mcall->in(first_ind+j); 456 if (iklass != NULL) { 457 st->print(", ["); 458 cifield = iklass->nonstatic_field_at(j); 459 cifield->print_name_on(st); 460 format_helper( regalloc, st, fld_node, ":", j, &scobjs ); 461 } else { 462 format_helper( regalloc, st, fld_node, ", [", j, &scobjs ); 463 } 464 } 465 } 466 st->print(" }"); 467 } 468 } 469 st->print_cr(""); 470 if (caller() != NULL) caller()->format(regalloc, n, st); 471 } 472 473 474 void JVMState::dump_spec(outputStream *st) const { 475 if (_method != NULL) { 476 bool printed = false; 477 if (!Verbose) { 478 // The JVMS dumps make really, really long lines. 479 // Take out the most boring parts, which are the package prefixes. 480 char buf[500]; 481 stringStream namest(buf, sizeof(buf)); 482 _method->print_short_name(&namest); 483 if (namest.count() < sizeof(buf)) { 484 const char* name = namest.base(); 485 if (name[0] == ' ') ++name; 486 const char* endcn = strchr(name, ':'); // end of class name 487 if (endcn == NULL) endcn = strchr(name, '('); 488 if (endcn == NULL) endcn = name + strlen(name); 489 while (endcn > name && endcn[-1] != '.' && endcn[-1] != '/') 490 --endcn; 491 st->print(" %s", endcn); 492 printed = true; 493 } 494 } 495 if (!printed) 496 _method->print_short_name(st); 497 st->print(" @ bci:%d",_bci); 498 if(_reexecute == Reexecute_True) 499 st->print(" reexecute"); 500 } else { 501 st->print(" runtime stub"); 502 } 503 if (caller() != NULL) caller()->dump_spec(st); 504 } 505 506 507 void JVMState::dump_on(outputStream* st) const { 508 if (_map && !((uintptr_t)_map & 1)) { 509 if (_map->len() > _map->req()) { // _map->has_exceptions() 510 Node* ex = _map->in(_map->req()); // _map->next_exception() 511 // skip the first one; it's already being printed 512 while (ex != NULL && ex->len() > ex->req()) { 513 ex = ex->in(ex->req()); // ex->next_exception() 514 ex->dump(1); 515 } 516 } 517 _map->dump(2); 518 } 519 st->print("JVMS depth=%d loc=%d stk=%d mon=%d scalar=%d end=%d mondepth=%d sp=%d bci=%d reexecute=%s method=", 520 depth(), locoff(), stkoff(), monoff(), scloff(), endoff(), monitor_depth(), sp(), bci(), should_reexecute()?"true":"false"); 521 if (_method == NULL) { 522 st->print_cr("(none)"); 523 } else { 524 _method->print_name(st); 525 st->cr(); 526 if (bci() >= 0 && bci() < _method->code_size()) { 527 st->print(" bc: "); 528 _method->print_codes_on(bci(), bci()+1, st); 529 } 530 } 531 if (caller() != NULL) { 532 caller()->dump_on(st); 533 } 534 } 535 536 // Extra way to dump a jvms from the debugger, 537 // to avoid a bug with C++ member function calls. 538 void dump_jvms(JVMState* jvms) { 539 jvms->dump(); 540 } 541 #endif 542 543 //--------------------------clone_shallow-------------------------------------- 544 JVMState* JVMState::clone_shallow(Compile* C) const { 545 JVMState* n = has_method() ? new (C) JVMState(_method, _caller) : new (C) JVMState(0); 546 n->set_bci(_bci); 547 n->_reexecute = _reexecute; 548 n->set_locoff(_locoff); 549 n->set_stkoff(_stkoff); 550 n->set_monoff(_monoff); 551 n->set_scloff(_scloff); 552 n->set_endoff(_endoff); 553 n->set_sp(_sp); 554 n->set_map(_map); 555 return n; 556 } 557 558 //---------------------------clone_deep---------------------------------------- 559 JVMState* JVMState::clone_deep(Compile* C) const { 560 JVMState* n = clone_shallow(C); 561 for (JVMState* p = n; p->_caller != NULL; p = p->_caller) { 562 p->_caller = p->_caller->clone_shallow(C); 563 } 564 assert(n->depth() == depth(), "sanity"); 565 assert(n->debug_depth() == debug_depth(), "sanity"); 566 return n; 567 } 568 569 //============================================================================= 570 uint CallNode::cmp( const Node &n ) const 571 { return _tf == ((CallNode&)n)._tf && _jvms == ((CallNode&)n)._jvms; } 572 #ifndef PRODUCT 573 void CallNode::dump_req() const { 574 // Dump the required inputs, enclosed in '(' and ')' 575 uint i; // Exit value of loop 576 for( i=0; i<req(); i++ ) { // For all required inputs 577 if( i == TypeFunc::Parms ) tty->print("("); 578 if( in(i) ) tty->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx); 579 else tty->print("_ "); 580 } 581 tty->print(")"); 582 } 583 584 void CallNode::dump_spec(outputStream *st) const { 585 st->print(" "); 586 tf()->dump_on(st); 587 if (_cnt != COUNT_UNKNOWN) st->print(" C=%f",_cnt); 588 if (jvms() != NULL) jvms()->dump_spec(st); 589 } 590 #endif 591 592 const Type *CallNode::bottom_type() const { return tf()->range(); } 593 const Type *CallNode::Value(PhaseTransform *phase) const { 594 if (phase->type(in(0)) == Type::TOP) return Type::TOP; 595 return tf()->range(); 596 } 597 598 //------------------------------calling_convention----------------------------- 599 void CallNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const { 600 // Use the standard compiler calling convention 601 Matcher::calling_convention( sig_bt, parm_regs, argcnt, true ); 602 } 603 604 605 //------------------------------match------------------------------------------ 606 // Construct projections for control, I/O, memory-fields, ..., and 607 // return result(s) along with their RegMask info 608 Node *CallNode::match( const ProjNode *proj, const Matcher *match ) { 609 switch (proj->_con) { 610 case TypeFunc::Control: 611 case TypeFunc::I_O: 612 case TypeFunc::Memory: 613 return new (match->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 614 615 case TypeFunc::Parms+1: // For LONG & DOUBLE returns 616 assert(tf()->_range->field_at(TypeFunc::Parms+1) == Type::HALF, ""); 617 // 2nd half of doubles and longs 618 return new (match->C, 1) MachProjNode(this,proj->_con, RegMask::Empty, (uint)OptoReg::Bad); 619 620 case TypeFunc::Parms: { // Normal returns 621 uint ideal_reg = Matcher::base2reg[tf()->range()->field_at(TypeFunc::Parms)->base()]; 622 OptoRegPair regs = is_CallRuntime() 623 ? match->c_return_value(ideal_reg,true) // Calls into C runtime 624 : match-> return_value(ideal_reg,true); // Calls into compiled Java code 625 RegMask rm = RegMask(regs.first()); 626 if( OptoReg::is_valid(regs.second()) ) 627 rm.Insert( regs.second() ); 628 return new (match->C, 1) MachProjNode(this,proj->_con,rm,ideal_reg); 629 } 630 631 case TypeFunc::ReturnAdr: 632 case TypeFunc::FramePtr: 633 default: 634 ShouldNotReachHere(); 635 } 636 return NULL; 637 } 638 639 // Do we Match on this edge index or not? Match no edges 640 uint CallNode::match_edge(uint idx) const { 641 return 0; 642 } 643 644 // 645 // Determine whether the call could modify the field of the specified 646 // instance at the specified offset. 647 // 648 bool CallNode::may_modify(const TypePtr *addr_t, PhaseTransform *phase) { 649 const TypeOopPtr *adrInst_t = addr_t->isa_oopptr(); 650 651 // If not an OopPtr or not an instance type, assume the worst. 652 // Note: currently this method is called only for instance types. 653 if (adrInst_t == NULL || !adrInst_t->is_known_instance()) { 654 return true; 655 } 656 // The instance_id is set only for scalar-replaceable allocations which 657 // are not passed as arguments according to Escape Analysis. 658 return false; 659 } 660 661 // Does this call have a direct reference to n other than debug information? 662 bool CallNode::has_non_debug_use(Node *n) { 663 const TypeTuple * d = tf()->domain(); 664 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 665 Node *arg = in(i); 666 if (arg == n) { 667 return true; 668 } 669 } 670 return false; 671 } 672 673 // Returns the unique CheckCastPP of a call 674 // or 'this' if there are several CheckCastPP 675 // or returns NULL if there is no one. 676 Node *CallNode::result_cast() { 677 Node *cast = NULL; 678 679 Node *p = proj_out(TypeFunc::Parms); 680 if (p == NULL) 681 return NULL; 682 683 for (DUIterator_Fast imax, i = p->fast_outs(imax); i < imax; i++) { 684 Node *use = p->fast_out(i); 685 if (use->is_CheckCastPP()) { 686 if (cast != NULL) { 687 return this; // more than 1 CheckCastPP 688 } 689 cast = use; 690 } 691 } 692 return cast; 693 } 694 695 696 //============================================================================= 697 uint CallJavaNode::size_of() const { return sizeof(*this); } 698 uint CallJavaNode::cmp( const Node &n ) const { 699 CallJavaNode &call = (CallJavaNode&)n; 700 return CallNode::cmp(call) && _method == call._method; 701 } 702 #ifndef PRODUCT 703 void CallJavaNode::dump_spec(outputStream *st) const { 704 if( _method ) _method->print_short_name(st); 705 CallNode::dump_spec(st); 706 } 707 #endif 708 709 //============================================================================= 710 uint CallStaticJavaNode::size_of() const { return sizeof(*this); } 711 uint CallStaticJavaNode::cmp( const Node &n ) const { 712 CallStaticJavaNode &call = (CallStaticJavaNode&)n; 713 return CallJavaNode::cmp(call); 714 } 715 716 //----------------------------uncommon_trap_request---------------------------- 717 // If this is an uncommon trap, return the request code, else zero. 718 int CallStaticJavaNode::uncommon_trap_request() const { 719 if (_name != NULL && !strcmp(_name, "uncommon_trap")) { 720 return extract_uncommon_trap_request(this); 721 } 722 return 0; 723 } 724 int CallStaticJavaNode::extract_uncommon_trap_request(const Node* call) { 725 #ifndef PRODUCT 726 if (!(call->req() > TypeFunc::Parms && 727 call->in(TypeFunc::Parms) != NULL && 728 call->in(TypeFunc::Parms)->is_Con())) { 729 assert(_in_dump_cnt != 0, "OK if dumping"); 730 tty->print("[bad uncommon trap]"); 731 return 0; 732 } 733 #endif 734 return call->in(TypeFunc::Parms)->bottom_type()->is_int()->get_con(); 735 } 736 737 #ifndef PRODUCT 738 void CallStaticJavaNode::dump_spec(outputStream *st) const { 739 st->print("# Static "); 740 if (_name != NULL) { 741 st->print("%s", _name); 742 int trap_req = uncommon_trap_request(); 743 if (trap_req != 0) { 744 char buf[100]; 745 st->print("(%s)", 746 Deoptimization::format_trap_request(buf, sizeof(buf), 747 trap_req)); 748 } 749 st->print(" "); 750 } 751 CallJavaNode::dump_spec(st); 752 } 753 #endif 754 755 //============================================================================= 756 uint CallDynamicJavaNode::size_of() const { return sizeof(*this); } 757 uint CallDynamicJavaNode::cmp( const Node &n ) const { 758 CallDynamicJavaNode &call = (CallDynamicJavaNode&)n; 759 return CallJavaNode::cmp(call); 760 } 761 #ifndef PRODUCT 762 void CallDynamicJavaNode::dump_spec(outputStream *st) const { 763 st->print("# Dynamic "); 764 CallJavaNode::dump_spec(st); 765 } 766 #endif 767 768 //============================================================================= 769 uint CallRuntimeNode::size_of() const { return sizeof(*this); } 770 uint CallRuntimeNode::cmp( const Node &n ) const { 771 CallRuntimeNode &call = (CallRuntimeNode&)n; 772 return CallNode::cmp(call) && !strcmp(_name,call._name); 773 } 774 #ifndef PRODUCT 775 void CallRuntimeNode::dump_spec(outputStream *st) const { 776 st->print("# "); 777 st->print(_name); 778 CallNode::dump_spec(st); 779 } 780 #endif 781 782 //------------------------------calling_convention----------------------------- 783 void CallRuntimeNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const { 784 Matcher::c_calling_convention( sig_bt, parm_regs, argcnt ); 785 } 786 787 //============================================================================= 788 //------------------------------calling_convention----------------------------- 789 790 791 //============================================================================= 792 #ifndef PRODUCT 793 void CallLeafNode::dump_spec(outputStream *st) const { 794 st->print("# "); 795 st->print(_name); 796 CallNode::dump_spec(st); 797 } 798 #endif 799 800 //============================================================================= 801 802 void SafePointNode::set_local(JVMState* jvms, uint idx, Node *c) { 803 assert(verify_jvms(jvms), "jvms must match"); 804 int loc = jvms->locoff() + idx; 805 if (in(loc)->is_top() && idx > 0 && !c->is_top() ) { 806 // If current local idx is top then local idx - 1 could 807 // be a long/double that needs to be killed since top could 808 // represent the 2nd half ofthe long/double. 809 uint ideal = in(loc -1)->ideal_reg(); 810 if (ideal == Op_RegD || ideal == Op_RegL) { 811 // set other (low index) half to top 812 set_req(loc - 1, in(loc)); 813 } 814 } 815 set_req(loc, c); 816 } 817 818 uint SafePointNode::size_of() const { return sizeof(*this); } 819 uint SafePointNode::cmp( const Node &n ) const { 820 return (&n == this); // Always fail except on self 821 } 822 823 //-------------------------set_next_exception---------------------------------- 824 void SafePointNode::set_next_exception(SafePointNode* n) { 825 assert(n == NULL || n->Opcode() == Op_SafePoint, "correct value for next_exception"); 826 if (len() == req()) { 827 if (n != NULL) add_prec(n); 828 } else { 829 set_prec(req(), n); 830 } 831 } 832 833 834 //----------------------------next_exception----------------------------------- 835 SafePointNode* SafePointNode::next_exception() const { 836 if (len() == req()) { 837 return NULL; 838 } else { 839 Node* n = in(req()); 840 assert(n == NULL || n->Opcode() == Op_SafePoint, "no other uses of prec edges"); 841 return (SafePointNode*) n; 842 } 843 } 844 845 846 //------------------------------Ideal------------------------------------------ 847 // Skip over any collapsed Regions 848 Node *SafePointNode::Ideal(PhaseGVN *phase, bool can_reshape) { 849 return remove_dead_region(phase, can_reshape) ? this : NULL; 850 } 851 852 //------------------------------Identity--------------------------------------- 853 // Remove obviously duplicate safepoints 854 Node *SafePointNode::Identity( PhaseTransform *phase ) { 855 856 // If you have back to back safepoints, remove one 857 if( in(TypeFunc::Control)->is_SafePoint() ) 858 return in(TypeFunc::Control); 859 860 if( in(0)->is_Proj() ) { 861 Node *n0 = in(0)->in(0); 862 // Check if he is a call projection (except Leaf Call) 863 if( n0->is_Catch() ) { 864 n0 = n0->in(0)->in(0); 865 assert( n0->is_Call(), "expect a call here" ); 866 } 867 if( n0->is_Call() && n0->as_Call()->guaranteed_safepoint() ) { 868 // Useless Safepoint, so remove it 869 return in(TypeFunc::Control); 870 } 871 } 872 873 return this; 874 } 875 876 //------------------------------Value------------------------------------------ 877 const Type *SafePointNode::Value( PhaseTransform *phase ) const { 878 if( phase->type(in(0)) == Type::TOP ) return Type::TOP; 879 if( phase->eqv( in(0), this ) ) return Type::TOP; // Dead infinite loop 880 return Type::CONTROL; 881 } 882 883 #ifndef PRODUCT 884 void SafePointNode::dump_spec(outputStream *st) const { 885 st->print(" SafePoint "); 886 } 887 #endif 888 889 const RegMask &SafePointNode::in_RegMask(uint idx) const { 890 if( idx < TypeFunc::Parms ) return RegMask::Empty; 891 // Values outside the domain represent debug info 892 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]); 893 } 894 const RegMask &SafePointNode::out_RegMask() const { 895 return RegMask::Empty; 896 } 897 898 899 void SafePointNode::grow_stack(JVMState* jvms, uint grow_by) { 900 assert((int)grow_by > 0, "sanity"); 901 int monoff = jvms->monoff(); 902 int scloff = jvms->scloff(); 903 int endoff = jvms->endoff(); 904 assert(endoff == (int)req(), "no other states or debug info after me"); 905 Node* top = Compile::current()->top(); 906 for (uint i = 0; i < grow_by; i++) { 907 ins_req(monoff, top); 908 } 909 jvms->set_monoff(monoff + grow_by); 910 jvms->set_scloff(scloff + grow_by); 911 jvms->set_endoff(endoff + grow_by); 912 } 913 914 void SafePointNode::push_monitor(const FastLockNode *lock) { 915 // Add a LockNode, which points to both the original BoxLockNode (the 916 // stack space for the monitor) and the Object being locked. 917 const int MonitorEdges = 2; 918 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges"); 919 assert(req() == jvms()->endoff(), "correct sizing"); 920 int nextmon = jvms()->scloff(); 921 if (GenerateSynchronizationCode) { 922 add_req(lock->box_node()); 923 add_req(lock->obj_node()); 924 } else { 925 Node* top = Compile::current()->top(); 926 add_req(top); 927 add_req(top); 928 } 929 jvms()->set_scloff(nextmon+MonitorEdges); 930 jvms()->set_endoff(req()); 931 } 932 933 void SafePointNode::pop_monitor() { 934 // Delete last monitor from debug info 935 debug_only(int num_before_pop = jvms()->nof_monitors()); 936 const int MonitorEdges = (1<<JVMState::logMonitorEdges); 937 int scloff = jvms()->scloff(); 938 int endoff = jvms()->endoff(); 939 int new_scloff = scloff - MonitorEdges; 940 int new_endoff = endoff - MonitorEdges; 941 jvms()->set_scloff(new_scloff); 942 jvms()->set_endoff(new_endoff); 943 while (scloff > new_scloff) del_req(--scloff); 944 assert(jvms()->nof_monitors() == num_before_pop-1, ""); 945 } 946 947 Node *SafePointNode::peek_monitor_box() const { 948 int mon = jvms()->nof_monitors() - 1; 949 assert(mon >= 0, "most have a monitor"); 950 return monitor_box(jvms(), mon); 951 } 952 953 Node *SafePointNode::peek_monitor_obj() const { 954 int mon = jvms()->nof_monitors() - 1; 955 assert(mon >= 0, "most have a monitor"); 956 return monitor_obj(jvms(), mon); 957 } 958 959 // Do we Match on this edge index or not? Match no edges 960 uint SafePointNode::match_edge(uint idx) const { 961 if( !needs_polling_address_input() ) 962 return 0; 963 964 return (TypeFunc::Parms == idx); 965 } 966 967 //============== SafePointScalarObjectNode ============== 968 969 SafePointScalarObjectNode::SafePointScalarObjectNode(const TypeOopPtr* tp, 970 #ifdef ASSERT 971 AllocateNode* alloc, 972 #endif 973 uint first_index, 974 uint n_fields) : 975 TypeNode(tp, 1), // 1 control input -- seems required. Get from root. 976 #ifdef ASSERT 977 _alloc(alloc), 978 #endif 979 _first_index(first_index), 980 _n_fields(n_fields) 981 { 982 init_class_id(Class_SafePointScalarObject); 983 } 984 985 bool SafePointScalarObjectNode::pinned() const { return true; } 986 bool SafePointScalarObjectNode::depends_only_on_test() const { return false; } 987 988 uint SafePointScalarObjectNode::ideal_reg() const { 989 return 0; // No matching to machine instruction 990 } 991 992 const RegMask &SafePointScalarObjectNode::in_RegMask(uint idx) const { 993 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]); 994 } 995 996 const RegMask &SafePointScalarObjectNode::out_RegMask() const { 997 return RegMask::Empty; 998 } 999 1000 uint SafePointScalarObjectNode::match_edge(uint idx) const { 1001 return 0; 1002 } 1003 1004 SafePointScalarObjectNode* 1005 SafePointScalarObjectNode::clone(int jvms_adj, Dict* sosn_map) const { 1006 void* cached = (*sosn_map)[(void*)this]; 1007 if (cached != NULL) { 1008 return (SafePointScalarObjectNode*)cached; 1009 } 1010 Compile* C = Compile::current(); 1011 SafePointScalarObjectNode* res = (SafePointScalarObjectNode*)Node::clone(); 1012 res->_first_index += jvms_adj; 1013 sosn_map->Insert((void*)this, (void*)res); 1014 return res; 1015 } 1016 1017 1018 #ifndef PRODUCT 1019 void SafePointScalarObjectNode::dump_spec(outputStream *st) const { 1020 st->print(" # fields@[%d..%d]", first_index(), 1021 first_index() + n_fields() - 1); 1022 } 1023 1024 #endif 1025 1026 //============================================================================= 1027 uint AllocateNode::size_of() const { return sizeof(*this); } 1028 1029 AllocateNode::AllocateNode(Compile* C, const TypeFunc *atype, 1030 Node *ctrl, Node *mem, Node *abio, 1031 Node *size, Node *klass_node, Node *initial_test) 1032 : CallNode(atype, NULL, TypeRawPtr::BOTTOM) 1033 { 1034 init_class_id(Class_Allocate); 1035 init_flags(Flag_is_macro); 1036 _is_scalar_replaceable = false; 1037 Node *topnode = C->top(); 1038 1039 init_req( TypeFunc::Control , ctrl ); 1040 init_req( TypeFunc::I_O , abio ); 1041 init_req( TypeFunc::Memory , mem ); 1042 init_req( TypeFunc::ReturnAdr, topnode ); 1043 init_req( TypeFunc::FramePtr , topnode ); 1044 init_req( AllocSize , size); 1045 init_req( KlassNode , klass_node); 1046 init_req( InitialTest , initial_test); 1047 init_req( ALength , topnode); 1048 C->add_macro_node(this); 1049 } 1050 1051 //============================================================================= 1052 uint AllocateArrayNode::size_of() const { return sizeof(*this); } 1053 1054 Node* AllocateArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1055 if (remove_dead_region(phase, can_reshape)) return this; 1056 1057 const Type* type = phase->type(Ideal_length()); 1058 if (type->isa_int() && type->is_int()->_hi < 0) { 1059 if (can_reshape) { 1060 PhaseIterGVN *igvn = phase->is_IterGVN(); 1061 // Unreachable fall through path (negative array length), 1062 // the allocation can only throw so disconnect it. 1063 Node* proj = proj_out(TypeFunc::Control); 1064 Node* catchproj = NULL; 1065 if (proj != NULL) { 1066 for (DUIterator_Fast imax, i = proj->fast_outs(imax); i < imax; i++) { 1067 Node *cn = proj->fast_out(i); 1068 if (cn->is_Catch()) { 1069 catchproj = cn->as_Multi()->proj_out(CatchProjNode::fall_through_index); 1070 break; 1071 } 1072 } 1073 } 1074 if (catchproj != NULL && catchproj->outcnt() > 0 && 1075 (catchproj->outcnt() > 1 || 1076 catchproj->unique_out()->Opcode() != Op_Halt)) { 1077 assert(catchproj->is_CatchProj(), "must be a CatchProjNode"); 1078 Node* nproj = catchproj->clone(); 1079 igvn->register_new_node_with_optimizer(nproj); 1080 1081 Node *frame = new (phase->C, 1) ParmNode( phase->C->start(), TypeFunc::FramePtr ); 1082 frame = phase->transform(frame); 1083 // Halt & Catch Fire 1084 Node *halt = new (phase->C, TypeFunc::Parms) HaltNode( nproj, frame ); 1085 phase->C->root()->add_req(halt); 1086 phase->transform(halt); 1087 1088 igvn->replace_node(catchproj, phase->C->top()); 1089 return this; 1090 } 1091 } else { 1092 // Can't correct it during regular GVN so register for IGVN 1093 phase->C->record_for_igvn(this); 1094 } 1095 } 1096 return NULL; 1097 } 1098 1099 // Retrieve the length from the AllocateArrayNode. Narrow the type with a 1100 // CastII, if appropriate. If we are not allowed to create new nodes, and 1101 // a CastII is appropriate, return NULL. 1102 Node *AllocateArrayNode::make_ideal_length(const TypeOopPtr* oop_type, PhaseTransform *phase, bool allow_new_nodes) { 1103 Node *length = in(AllocateNode::ALength); 1104 assert(length != NULL, "length is not null"); 1105 1106 const TypeInt* length_type = phase->find_int_type(length); 1107 const TypeAryPtr* ary_type = oop_type->isa_aryptr(); 1108 1109 if (ary_type != NULL && length_type != NULL) { 1110 const TypeInt* narrow_length_type = ary_type->narrow_size_type(length_type); 1111 if (narrow_length_type != length_type) { 1112 // Assert one of: 1113 // - the narrow_length is 0 1114 // - the narrow_length is not wider than length 1115 assert(narrow_length_type == TypeInt::ZERO || 1116 (narrow_length_type->_hi <= length_type->_hi && 1117 narrow_length_type->_lo >= length_type->_lo), 1118 "narrow type must be narrower than length type"); 1119 1120 // Return NULL if new nodes are not allowed 1121 if (!allow_new_nodes) return NULL; 1122 // Create a cast which is control dependent on the initialization to 1123 // propagate the fact that the array length must be positive. 1124 length = new (phase->C, 2) CastIINode(length, narrow_length_type); 1125 length->set_req(0, initialization()->proj_out(0)); 1126 } 1127 } 1128 1129 return length; 1130 } 1131 1132 //============================================================================= 1133 uint LockNode::size_of() const { return sizeof(*this); } 1134 1135 // Redundant lock elimination 1136 // 1137 // There are various patterns of locking where we release and 1138 // immediately reacquire a lock in a piece of code where no operations 1139 // occur in between that would be observable. In those cases we can 1140 // skip releasing and reacquiring the lock without violating any 1141 // fairness requirements. Doing this around a loop could cause a lock 1142 // to be held for a very long time so we concentrate on non-looping 1143 // control flow. We also require that the operations are fully 1144 // redundant meaning that we don't introduce new lock operations on 1145 // some paths so to be able to eliminate it on others ala PRE. This 1146 // would probably require some more extensive graph manipulation to 1147 // guarantee that the memory edges were all handled correctly. 1148 // 1149 // Assuming p is a simple predicate which can't trap in any way and s 1150 // is a synchronized method consider this code: 1151 // 1152 // s(); 1153 // if (p) 1154 // s(); 1155 // else 1156 // s(); 1157 // s(); 1158 // 1159 // 1. The unlocks of the first call to s can be eliminated if the 1160 // locks inside the then and else branches are eliminated. 1161 // 1162 // 2. The unlocks of the then and else branches can be eliminated if 1163 // the lock of the final call to s is eliminated. 1164 // 1165 // Either of these cases subsumes the simple case of sequential control flow 1166 // 1167 // Addtionally we can eliminate versions without the else case: 1168 // 1169 // s(); 1170 // if (p) 1171 // s(); 1172 // s(); 1173 // 1174 // 3. In this case we eliminate the unlock of the first s, the lock 1175 // and unlock in the then case and the lock in the final s. 1176 // 1177 // Note also that in all these cases the then/else pieces don't have 1178 // to be trivial as long as they begin and end with synchronization 1179 // operations. 1180 // 1181 // s(); 1182 // if (p) 1183 // s(); 1184 // f(); 1185 // s(); 1186 // s(); 1187 // 1188 // The code will work properly for this case, leaving in the unlock 1189 // before the call to f and the relock after it. 1190 // 1191 // A potentially interesting case which isn't handled here is when the 1192 // locking is partially redundant. 1193 // 1194 // s(); 1195 // if (p) 1196 // s(); 1197 // 1198 // This could be eliminated putting unlocking on the else case and 1199 // eliminating the first unlock and the lock in the then side. 1200 // Alternatively the unlock could be moved out of the then side so it 1201 // was after the merge and the first unlock and second lock 1202 // eliminated. This might require less manipulation of the memory 1203 // state to get correct. 1204 // 1205 // Additionally we might allow work between a unlock and lock before 1206 // giving up eliminating the locks. The current code disallows any 1207 // conditional control flow between these operations. A formulation 1208 // similar to partial redundancy elimination computing the 1209 // availability of unlocking and the anticipatability of locking at a 1210 // program point would allow detection of fully redundant locking with 1211 // some amount of work in between. I'm not sure how often I really 1212 // think that would occur though. Most of the cases I've seen 1213 // indicate it's likely non-trivial work would occur in between. 1214 // There may be other more complicated constructs where we could 1215 // eliminate locking but I haven't seen any others appear as hot or 1216 // interesting. 1217 // 1218 // Locking and unlocking have a canonical form in ideal that looks 1219 // roughly like this: 1220 // 1221 // <obj> 1222 // | \\------+ 1223 // | \ \ 1224 // | BoxLock \ 1225 // | | | \ 1226 // | | \ \ 1227 // | | FastLock 1228 // | | / 1229 // | | / 1230 // | | | 1231 // 1232 // Lock 1233 // | 1234 // Proj #0 1235 // | 1236 // MembarAcquire 1237 // | 1238 // Proj #0 1239 // 1240 // MembarRelease 1241 // | 1242 // Proj #0 1243 // | 1244 // Unlock 1245 // | 1246 // Proj #0 1247 // 1248 // 1249 // This code proceeds by processing Lock nodes during PhaseIterGVN 1250 // and searching back through its control for the proper code 1251 // patterns. Once it finds a set of lock and unlock operations to 1252 // eliminate they are marked as eliminatable which causes the 1253 // expansion of the Lock and Unlock macro nodes to make the operation a NOP 1254 // 1255 //============================================================================= 1256 1257 // 1258 // Utility function to skip over uninteresting control nodes. Nodes skipped are: 1259 // - copy regions. (These may not have been optimized away yet.) 1260 // - eliminated locking nodes 1261 // 1262 static Node *next_control(Node *ctrl) { 1263 if (ctrl == NULL) 1264 return NULL; 1265 while (1) { 1266 if (ctrl->is_Region()) { 1267 RegionNode *r = ctrl->as_Region(); 1268 Node *n = r->is_copy(); 1269 if (n == NULL) 1270 break; // hit a region, return it 1271 else 1272 ctrl = n; 1273 } else if (ctrl->is_Proj()) { 1274 Node *in0 = ctrl->in(0); 1275 if (in0->is_AbstractLock() && in0->as_AbstractLock()->is_eliminated()) { 1276 ctrl = in0->in(0); 1277 } else { 1278 break; 1279 } 1280 } else { 1281 break; // found an interesting control 1282 } 1283 } 1284 return ctrl; 1285 } 1286 // 1287 // Given a control, see if it's the control projection of an Unlock which 1288 // operating on the same object as lock. 1289 // 1290 bool AbstractLockNode::find_matching_unlock(const Node* ctrl, LockNode* lock, 1291 GrowableArray<AbstractLockNode*> &lock_ops) { 1292 ProjNode *ctrl_proj = (ctrl->is_Proj()) ? ctrl->as_Proj() : NULL; 1293 if (ctrl_proj != NULL && ctrl_proj->_con == TypeFunc::Control) { 1294 Node *n = ctrl_proj->in(0); 1295 if (n != NULL && n->is_Unlock()) { 1296 UnlockNode *unlock = n->as_Unlock(); 1297 if ((lock->obj_node() == unlock->obj_node()) && 1298 (lock->box_node() == unlock->box_node()) && !unlock->is_eliminated()) { 1299 lock_ops.append(unlock); 1300 return true; 1301 } 1302 } 1303 } 1304 return false; 1305 } 1306 1307 // 1308 // Find the lock matching an unlock. Returns null if a safepoint 1309 // or complicated control is encountered first. 1310 LockNode *AbstractLockNode::find_matching_lock(UnlockNode* unlock) { 1311 LockNode *lock_result = NULL; 1312 // find the matching lock, or an intervening safepoint 1313 Node *ctrl = next_control(unlock->in(0)); 1314 while (1) { 1315 assert(ctrl != NULL, "invalid control graph"); 1316 assert(!ctrl->is_Start(), "missing lock for unlock"); 1317 if (ctrl->is_top()) break; // dead control path 1318 if (ctrl->is_Proj()) ctrl = ctrl->in(0); 1319 if (ctrl->is_SafePoint()) { 1320 break; // found a safepoint (may be the lock we are searching for) 1321 } else if (ctrl->is_Region()) { 1322 // Check for a simple diamond pattern. Punt on anything more complicated 1323 if (ctrl->req() == 3 && ctrl->in(1) != NULL && ctrl->in(2) != NULL) { 1324 Node *in1 = next_control(ctrl->in(1)); 1325 Node *in2 = next_control(ctrl->in(2)); 1326 if (((in1->is_IfTrue() && in2->is_IfFalse()) || 1327 (in2->is_IfTrue() && in1->is_IfFalse())) && (in1->in(0) == in2->in(0))) { 1328 ctrl = next_control(in1->in(0)->in(0)); 1329 } else { 1330 break; 1331 } 1332 } else { 1333 break; 1334 } 1335 } else { 1336 ctrl = next_control(ctrl->in(0)); // keep searching 1337 } 1338 } 1339 if (ctrl->is_Lock()) { 1340 LockNode *lock = ctrl->as_Lock(); 1341 if ((lock->obj_node() == unlock->obj_node()) && 1342 (lock->box_node() == unlock->box_node())) { 1343 lock_result = lock; 1344 } 1345 } 1346 return lock_result; 1347 } 1348 1349 // This code corresponds to case 3 above. 1350 1351 bool AbstractLockNode::find_lock_and_unlock_through_if(Node* node, LockNode* lock, 1352 GrowableArray<AbstractLockNode*> &lock_ops) { 1353 Node* if_node = node->in(0); 1354 bool if_true = node->is_IfTrue(); 1355 1356 if (if_node->is_If() && if_node->outcnt() == 2 && (if_true || node->is_IfFalse())) { 1357 Node *lock_ctrl = next_control(if_node->in(0)); 1358 if (find_matching_unlock(lock_ctrl, lock, lock_ops)) { 1359 Node* lock1_node = NULL; 1360 ProjNode* proj = if_node->as_If()->proj_out(!if_true); 1361 if (if_true) { 1362 if (proj->is_IfFalse() && proj->outcnt() == 1) { 1363 lock1_node = proj->unique_out(); 1364 } 1365 } else { 1366 if (proj->is_IfTrue() && proj->outcnt() == 1) { 1367 lock1_node = proj->unique_out(); 1368 } 1369 } 1370 if (lock1_node != NULL && lock1_node->is_Lock()) { 1371 LockNode *lock1 = lock1_node->as_Lock(); 1372 if ((lock->obj_node() == lock1->obj_node()) && 1373 (lock->box_node() == lock1->box_node()) && !lock1->is_eliminated()) { 1374 lock_ops.append(lock1); 1375 return true; 1376 } 1377 } 1378 } 1379 } 1380 1381 lock_ops.trunc_to(0); 1382 return false; 1383 } 1384 1385 bool AbstractLockNode::find_unlocks_for_region(const RegionNode* region, LockNode* lock, 1386 GrowableArray<AbstractLockNode*> &lock_ops) { 1387 // check each control merging at this point for a matching unlock. 1388 // in(0) should be self edge so skip it. 1389 for (int i = 1; i < (int)region->req(); i++) { 1390 Node *in_node = next_control(region->in(i)); 1391 if (in_node != NULL) { 1392 if (find_matching_unlock(in_node, lock, lock_ops)) { 1393 // found a match so keep on checking. 1394 continue; 1395 } else if (find_lock_and_unlock_through_if(in_node, lock, lock_ops)) { 1396 continue; 1397 } 1398 1399 // If we fall through to here then it was some kind of node we 1400 // don't understand or there wasn't a matching unlock, so give 1401 // up trying to merge locks. 1402 lock_ops.trunc_to(0); 1403 return false; 1404 } 1405 } 1406 return true; 1407 1408 } 1409 1410 #ifndef PRODUCT 1411 // 1412 // Create a counter which counts the number of times this lock is acquired 1413 // 1414 void AbstractLockNode::create_lock_counter(JVMState* state) { 1415 _counter = OptoRuntime::new_named_counter(state, NamedCounter::LockCounter); 1416 } 1417 #endif 1418 1419 void AbstractLockNode::set_eliminated() { 1420 _eliminate = true; 1421 #ifndef PRODUCT 1422 if (_counter) { 1423 // Update the counter to indicate that this lock was eliminated. 1424 // The counter update code will stay around even though the 1425 // optimizer will eliminate the lock operation itself. 1426 _counter->set_tag(NamedCounter::EliminatedLockCounter); 1427 } 1428 #endif 1429 } 1430 1431 //============================================================================= 1432 Node *LockNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1433 1434 // perform any generic optimizations first (returns 'this' or NULL) 1435 Node *result = SafePointNode::Ideal(phase, can_reshape); 1436 1437 // Now see if we can optimize away this lock. We don't actually 1438 // remove the locking here, we simply set the _eliminate flag which 1439 // prevents macro expansion from expanding the lock. Since we don't 1440 // modify the graph, the value returned from this function is the 1441 // one computed above. 1442 if (result == NULL && can_reshape && EliminateLocks && !is_eliminated()) { 1443 // 1444 // If we are locking an unescaped object, the lock/unlock is unnecessary 1445 // 1446 ConnectionGraph *cgr = phase->C->congraph(); 1447 PointsToNode::EscapeState es = PointsToNode::GlobalEscape; 1448 if (cgr != NULL) 1449 es = cgr->escape_state(obj_node(), phase); 1450 if (es != PointsToNode::UnknownEscape && es != PointsToNode::GlobalEscape) { 1451 // Mark it eliminated to update any counters 1452 this->set_eliminated(); 1453 return result; 1454 } 1455 1456 // 1457 // Try lock coarsening 1458 // 1459 PhaseIterGVN* iter = phase->is_IterGVN(); 1460 if (iter != NULL) { 1461 1462 GrowableArray<AbstractLockNode*> lock_ops; 1463 1464 Node *ctrl = next_control(in(0)); 1465 1466 // now search back for a matching Unlock 1467 if (find_matching_unlock(ctrl, this, lock_ops)) { 1468 // found an unlock directly preceding this lock. This is the 1469 // case of single unlock directly control dependent on a 1470 // single lock which is the trivial version of case 1 or 2. 1471 } else if (ctrl->is_Region() ) { 1472 if (find_unlocks_for_region(ctrl->as_Region(), this, lock_ops)) { 1473 // found lock preceded by multiple unlocks along all paths 1474 // joining at this point which is case 3 in description above. 1475 } 1476 } else { 1477 // see if this lock comes from either half of an if and the 1478 // predecessors merges unlocks and the other half of the if 1479 // performs a lock. 1480 if (find_lock_and_unlock_through_if(ctrl, this, lock_ops)) { 1481 // found unlock splitting to an if with locks on both branches. 1482 } 1483 } 1484 1485 if (lock_ops.length() > 0) { 1486 // add ourselves to the list of locks to be eliminated. 1487 lock_ops.append(this); 1488 1489 #ifndef PRODUCT 1490 if (PrintEliminateLocks) { 1491 int locks = 0; 1492 int unlocks = 0; 1493 for (int i = 0; i < lock_ops.length(); i++) { 1494 AbstractLockNode* lock = lock_ops.at(i); 1495 if (lock->Opcode() == Op_Lock) 1496 locks++; 1497 else 1498 unlocks++; 1499 if (Verbose) { 1500 lock->dump(1); 1501 } 1502 } 1503 tty->print_cr("***Eliminated %d unlocks and %d locks", unlocks, locks); 1504 } 1505 #endif 1506 1507 // for each of the identified locks, mark them 1508 // as eliminatable 1509 for (int i = 0; i < lock_ops.length(); i++) { 1510 AbstractLockNode* lock = lock_ops.at(i); 1511 1512 // Mark it eliminated to update any counters 1513 lock->set_eliminated(); 1514 lock->set_coarsened(); 1515 } 1516 } else if (result != NULL && ctrl->is_Region() && 1517 iter->_worklist.member(ctrl)) { 1518 // We weren't able to find any opportunities but the region this 1519 // lock is control dependent on hasn't been processed yet so put 1520 // this lock back on the worklist so we can check again once any 1521 // region simplification has occurred. 1522 iter->_worklist.push(this); 1523 } 1524 } 1525 } 1526 1527 return result; 1528 } 1529 1530 //============================================================================= 1531 uint UnlockNode::size_of() const { return sizeof(*this); } 1532 1533 //============================================================================= 1534 Node *UnlockNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1535 1536 // perform any generic optimizations first (returns 'this' or NULL) 1537 Node * result = SafePointNode::Ideal(phase, can_reshape); 1538 1539 // Now see if we can optimize away this unlock. We don't actually 1540 // remove the unlocking here, we simply set the _eliminate flag which 1541 // prevents macro expansion from expanding the unlock. Since we don't 1542 // modify the graph, the value returned from this function is the 1543 // one computed above. 1544 // Escape state is defined after Parse phase. 1545 if (result == NULL && can_reshape && EliminateLocks && !is_eliminated()) { 1546 // 1547 // If we are unlocking an unescaped object, the lock/unlock is unnecessary. 1548 // 1549 ConnectionGraph *cgr = phase->C->congraph(); 1550 PointsToNode::EscapeState es = PointsToNode::GlobalEscape; 1551 if (cgr != NULL) 1552 es = cgr->escape_state(obj_node(), phase); 1553 if (es != PointsToNode::UnknownEscape && es != PointsToNode::GlobalEscape) { 1554 // Mark it eliminated to update any counters 1555 this->set_eliminated(); 1556 } 1557 } 1558 return result; 1559 }