1 /* 2 * Copyright (c) 1997, 2022, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "compiler/compileLog.hpp" 27 #include "ci/bcEscapeAnalyzer.hpp" 28 #include "ci/ciNativeEntryPoint.hpp" 29 #include "compiler/oopMap.hpp" 30 #include "gc/shared/barrierSet.hpp" 31 #include "gc/shared/c2/barrierSetC2.hpp" 32 #include "interpreter/interpreter.hpp" 33 #include "opto/callGenerator.hpp" 34 #include "opto/callnode.hpp" 35 #include "opto/castnode.hpp" 36 #include "opto/convertnode.hpp" 37 #include "opto/escape.hpp" 38 #include "opto/locknode.hpp" 39 #include "opto/machnode.hpp" 40 #include "opto/matcher.hpp" 41 #include "opto/parse.hpp" 42 #include "opto/regalloc.hpp" 43 #include "opto/regmask.hpp" 44 #include "opto/rootnode.hpp" 45 #include "opto/runtime.hpp" 46 #include "runtime/sharedRuntime.hpp" 47 #include "utilities/powerOfTwo.hpp" 48 #include "code/vmreg.hpp" 49 50 // Portions of code courtesy of Clifford Click 51 52 // Optimization - Graph Style 53 54 //============================================================================= 55 uint StartNode::size_of() const { return sizeof(*this); } 56 bool StartNode::cmp( const Node &n ) const 57 { return _domain == ((StartNode&)n)._domain; } 58 const Type *StartNode::bottom_type() const { return _domain; } 59 const Type* StartNode::Value(PhaseGVN* phase) const { return _domain; } 60 #ifndef PRODUCT 61 void StartNode::dump_spec(outputStream *st) const { st->print(" #"); _domain->dump_on(st);} 62 void StartNode::dump_compact_spec(outputStream *st) const { /* empty */ } 63 #endif 64 65 //------------------------------Ideal------------------------------------------ 66 Node *StartNode::Ideal(PhaseGVN *phase, bool can_reshape){ 67 return remove_dead_region(phase, can_reshape) ? this : NULL; 68 } 69 70 //------------------------------calling_convention----------------------------- 71 void StartNode::calling_convention(BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt) const { 72 SharedRuntime::java_calling_convention(sig_bt, parm_regs, argcnt); 73 } 74 75 //------------------------------Registers-------------------------------------- 76 const RegMask &StartNode::in_RegMask(uint) const { 77 return RegMask::Empty; 78 } 79 80 //------------------------------match------------------------------------------ 81 // Construct projections for incoming parameters, and their RegMask info 82 Node *StartNode::match( const ProjNode *proj, const Matcher *match ) { 83 switch (proj->_con) { 84 case TypeFunc::Control: 85 case TypeFunc::I_O: 86 case TypeFunc::Memory: 87 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 88 case TypeFunc::FramePtr: 89 return new MachProjNode(this,proj->_con,Matcher::c_frame_ptr_mask, Op_RegP); 90 case TypeFunc::ReturnAdr: 91 return new MachProjNode(this,proj->_con,match->_return_addr_mask,Op_RegP); 92 case TypeFunc::Parms: 93 default: { 94 uint parm_num = proj->_con - TypeFunc::Parms; 95 const Type *t = _domain->field_at(proj->_con); 96 if (t->base() == Type::Half) // 2nd half of Longs and Doubles 97 return new ConNode(Type::TOP); 98 uint ideal_reg = t->ideal_reg(); 99 RegMask &rm = match->_calling_convention_mask[parm_num]; 100 return new MachProjNode(this,proj->_con,rm,ideal_reg); 101 } 102 } 103 return NULL; 104 } 105 106 //------------------------------StartOSRNode---------------------------------- 107 // The method start node for an on stack replacement adapter 108 109 //------------------------------osr_domain----------------------------- 110 const TypeTuple *StartOSRNode::osr_domain() { 111 const Type **fields = TypeTuple::fields(2); 112 fields[TypeFunc::Parms+0] = TypeRawPtr::BOTTOM; // address of osr buffer 113 114 return TypeTuple::make(TypeFunc::Parms+1, fields); 115 } 116 117 //============================================================================= 118 const char * const ParmNode::names[TypeFunc::Parms+1] = { 119 "Control", "I_O", "Memory", "FramePtr", "ReturnAdr", "Parms" 120 }; 121 122 #ifndef PRODUCT 123 void ParmNode::dump_spec(outputStream *st) const { 124 if( _con < TypeFunc::Parms ) { 125 st->print("%s", names[_con]); 126 } else { 127 st->print("Parm%d: ",_con-TypeFunc::Parms); 128 // Verbose and WizardMode dump bottom_type for all nodes 129 if( !Verbose && !WizardMode ) bottom_type()->dump_on(st); 130 } 131 } 132 133 void ParmNode::dump_compact_spec(outputStream *st) const { 134 if (_con < TypeFunc::Parms) { 135 st->print("%s", names[_con]); 136 } else { 137 st->print("%d:", _con-TypeFunc::Parms); 138 // unconditionally dump bottom_type 139 bottom_type()->dump_on(st); 140 } 141 } 142 143 // For a ParmNode, all immediate inputs and outputs are considered relevant 144 // both in compact and standard representation. 145 void ParmNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const { 146 this->collect_nodes(in_rel, 1, false, false); 147 this->collect_nodes(out_rel, -1, false, false); 148 } 149 #endif 150 151 uint ParmNode::ideal_reg() const { 152 switch( _con ) { 153 case TypeFunc::Control : // fall through 154 case TypeFunc::I_O : // fall through 155 case TypeFunc::Memory : return 0; 156 case TypeFunc::FramePtr : // fall through 157 case TypeFunc::ReturnAdr: return Op_RegP; 158 default : assert( _con > TypeFunc::Parms, "" ); 159 // fall through 160 case TypeFunc::Parms : { 161 // Type of argument being passed 162 const Type *t = in(0)->as_Start()->_domain->field_at(_con); 163 return t->ideal_reg(); 164 } 165 } 166 ShouldNotReachHere(); 167 return 0; 168 } 169 170 //============================================================================= 171 ReturnNode::ReturnNode(uint edges, Node *cntrl, Node *i_o, Node *memory, Node *frameptr, Node *retadr ) : Node(edges) { 172 init_req(TypeFunc::Control,cntrl); 173 init_req(TypeFunc::I_O,i_o); 174 init_req(TypeFunc::Memory,memory); 175 init_req(TypeFunc::FramePtr,frameptr); 176 init_req(TypeFunc::ReturnAdr,retadr); 177 } 178 179 Node *ReturnNode::Ideal(PhaseGVN *phase, bool can_reshape){ 180 return remove_dead_region(phase, can_reshape) ? this : NULL; 181 } 182 183 const Type* ReturnNode::Value(PhaseGVN* phase) const { 184 return ( phase->type(in(TypeFunc::Control)) == Type::TOP) 185 ? Type::TOP 186 : Type::BOTTOM; 187 } 188 189 // Do we Match on this edge index or not? No edges on return nodes 190 uint ReturnNode::match_edge(uint idx) const { 191 return 0; 192 } 193 194 195 #ifndef PRODUCT 196 void ReturnNode::dump_req(outputStream *st) const { 197 // Dump the required inputs, enclosed in '(' and ')' 198 uint i; // Exit value of loop 199 for (i = 0; i < req(); i++) { // For all required inputs 200 if (i == TypeFunc::Parms) st->print("returns"); 201 if (in(i)) st->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx); 202 else st->print("_ "); 203 } 204 } 205 #endif 206 207 //============================================================================= 208 RethrowNode::RethrowNode( 209 Node* cntrl, 210 Node* i_o, 211 Node* memory, 212 Node* frameptr, 213 Node* ret_adr, 214 Node* exception 215 ) : Node(TypeFunc::Parms + 1) { 216 init_req(TypeFunc::Control , cntrl ); 217 init_req(TypeFunc::I_O , i_o ); 218 init_req(TypeFunc::Memory , memory ); 219 init_req(TypeFunc::FramePtr , frameptr ); 220 init_req(TypeFunc::ReturnAdr, ret_adr); 221 init_req(TypeFunc::Parms , exception); 222 } 223 224 Node *RethrowNode::Ideal(PhaseGVN *phase, bool can_reshape){ 225 return remove_dead_region(phase, can_reshape) ? this : NULL; 226 } 227 228 const Type* RethrowNode::Value(PhaseGVN* phase) const { 229 return (phase->type(in(TypeFunc::Control)) == Type::TOP) 230 ? Type::TOP 231 : Type::BOTTOM; 232 } 233 234 uint RethrowNode::match_edge(uint idx) const { 235 return 0; 236 } 237 238 #ifndef PRODUCT 239 void RethrowNode::dump_req(outputStream *st) const { 240 // Dump the required inputs, enclosed in '(' and ')' 241 uint i; // Exit value of loop 242 for (i = 0; i < req(); i++) { // For all required inputs 243 if (i == TypeFunc::Parms) st->print("exception"); 244 if (in(i)) st->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx); 245 else st->print("_ "); 246 } 247 } 248 #endif 249 250 //============================================================================= 251 // Do we Match on this edge index or not? Match only target address & method 252 uint TailCallNode::match_edge(uint idx) const { 253 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1; 254 } 255 256 //============================================================================= 257 // Do we Match on this edge index or not? Match only target address & oop 258 uint TailJumpNode::match_edge(uint idx) const { 259 return TypeFunc::Parms <= idx && idx <= TypeFunc::Parms+1; 260 } 261 262 //============================================================================= 263 JVMState::JVMState(ciMethod* method, JVMState* caller) : 264 _method(method) { 265 assert(method != NULL, "must be valid call site"); 266 _bci = InvocationEntryBci; 267 _reexecute = Reexecute_Undefined; 268 debug_only(_bci = -99); // random garbage value 269 debug_only(_map = (SafePointNode*)-1); 270 _caller = caller; 271 _depth = 1 + (caller == NULL ? 0 : caller->depth()); 272 _locoff = TypeFunc::Parms; 273 _stkoff = _locoff + _method->max_locals(); 274 _monoff = _stkoff + _method->max_stack(); 275 _scloff = _monoff; 276 _endoff = _monoff; 277 _sp = 0; 278 } 279 JVMState::JVMState(int stack_size) : 280 _method(NULL) { 281 _bci = InvocationEntryBci; 282 _reexecute = Reexecute_Undefined; 283 debug_only(_map = (SafePointNode*)-1); 284 _caller = NULL; 285 _depth = 1; 286 _locoff = TypeFunc::Parms; 287 _stkoff = _locoff; 288 _monoff = _stkoff + stack_size; 289 _scloff = _monoff; 290 _endoff = _monoff; 291 _sp = 0; 292 } 293 294 //--------------------------------of_depth------------------------------------- 295 JVMState* JVMState::of_depth(int d) const { 296 const JVMState* jvmp = this; 297 assert(0 < d && (uint)d <= depth(), "oob"); 298 for (int skip = depth() - d; skip > 0; skip--) { 299 jvmp = jvmp->caller(); 300 } 301 assert(jvmp->depth() == (uint)d, "found the right one"); 302 return (JVMState*)jvmp; 303 } 304 305 //-----------------------------same_calls_as----------------------------------- 306 bool JVMState::same_calls_as(const JVMState* that) const { 307 if (this == that) return true; 308 if (this->depth() != that->depth()) return false; 309 const JVMState* p = this; 310 const JVMState* q = that; 311 for (;;) { 312 if (p->_method != q->_method) return false; 313 if (p->_method == NULL) return true; // bci is irrelevant 314 if (p->_bci != q->_bci) return false; 315 if (p->_reexecute != q->_reexecute) return false; 316 p = p->caller(); 317 q = q->caller(); 318 if (p == q) return true; 319 assert(p != NULL && q != NULL, "depth check ensures we don't run off end"); 320 } 321 } 322 323 //------------------------------debug_start------------------------------------ 324 uint JVMState::debug_start() const { 325 debug_only(JVMState* jvmroot = of_depth(1)); 326 assert(jvmroot->locoff() <= this->locoff(), "youngest JVMState must be last"); 327 return of_depth(1)->locoff(); 328 } 329 330 //-------------------------------debug_end------------------------------------- 331 uint JVMState::debug_end() const { 332 debug_only(JVMState* jvmroot = of_depth(1)); 333 assert(jvmroot->endoff() <= this->endoff(), "youngest JVMState must be last"); 334 return endoff(); 335 } 336 337 //------------------------------debug_depth------------------------------------ 338 uint JVMState::debug_depth() const { 339 uint total = 0; 340 for (const JVMState* jvmp = this; jvmp != NULL; jvmp = jvmp->caller()) { 341 total += jvmp->debug_size(); 342 } 343 return total; 344 } 345 346 #ifndef PRODUCT 347 348 //------------------------------format_helper---------------------------------- 349 // Given an allocation (a Chaitin object) and a Node decide if the Node carries 350 // any defined value or not. If it does, print out the register or constant. 351 static void format_helper( PhaseRegAlloc *regalloc, outputStream* st, Node *n, const char *msg, uint i, GrowableArray<SafePointScalarObjectNode*> *scobjs ) { 352 if (n == NULL) { st->print(" NULL"); return; } 353 if (n->is_SafePointScalarObject()) { 354 // Scalar replacement. 355 SafePointScalarObjectNode* spobj = n->as_SafePointScalarObject(); 356 scobjs->append_if_missing(spobj); 357 int sco_n = scobjs->find(spobj); 358 assert(sco_n >= 0, ""); 359 st->print(" %s%d]=#ScObj" INT32_FORMAT, msg, i, sco_n); 360 return; 361 } 362 if (regalloc->node_regs_max_index() > 0 && 363 OptoReg::is_valid(regalloc->get_reg_first(n))) { // Check for undefined 364 char buf[50]; 365 regalloc->dump_register(n,buf); 366 st->print(" %s%d]=%s",msg,i,buf); 367 } else { // No register, but might be constant 368 const Type *t = n->bottom_type(); 369 switch (t->base()) { 370 case Type::Int: 371 st->print(" %s%d]=#" INT32_FORMAT,msg,i,t->is_int()->get_con()); 372 break; 373 case Type::AnyPtr: 374 assert( t == TypePtr::NULL_PTR || n->in_dump(), "" ); 375 st->print(" %s%d]=#NULL",msg,i); 376 break; 377 case Type::AryPtr: 378 case Type::InstPtr: 379 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->isa_oopptr()->const_oop())); 380 break; 381 case Type::KlassPtr: 382 case Type::AryKlassPtr: 383 case Type::InstKlassPtr: 384 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_klassptr()->klass())); 385 break; 386 case Type::MetadataPtr: 387 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_metadataptr()->metadata())); 388 break; 389 case Type::NarrowOop: 390 st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,p2i(t->make_ptr()->isa_oopptr()->const_oop())); 391 break; 392 case Type::RawPtr: 393 st->print(" %s%d]=#Raw" INTPTR_FORMAT,msg,i,p2i(t->is_rawptr())); 394 break; 395 case Type::DoubleCon: 396 st->print(" %s%d]=#%fD",msg,i,t->is_double_constant()->_d); 397 break; 398 case Type::FloatCon: 399 st->print(" %s%d]=#%fF",msg,i,t->is_float_constant()->_f); 400 break; 401 case Type::Long: 402 st->print(" %s%d]=#" INT64_FORMAT,msg,i,(int64_t)(t->is_long()->get_con())); 403 break; 404 case Type::Half: 405 case Type::Top: 406 st->print(" %s%d]=_",msg,i); 407 break; 408 default: ShouldNotReachHere(); 409 } 410 } 411 } 412 413 //---------------------print_method_with_lineno-------------------------------- 414 void JVMState::print_method_with_lineno(outputStream* st, bool show_name) const { 415 if (show_name) _method->print_short_name(st); 416 417 int lineno = _method->line_number_from_bci(_bci); 418 if (lineno != -1) { 419 st->print(" @ bci:%d (line %d)", _bci, lineno); 420 } else { 421 st->print(" @ bci:%d", _bci); 422 } 423 } 424 425 //------------------------------format----------------------------------------- 426 void JVMState::format(PhaseRegAlloc *regalloc, const Node *n, outputStream* st) const { 427 st->print(" #"); 428 if (_method) { 429 print_method_with_lineno(st, true); 430 } else { 431 st->print_cr(" runtime stub "); 432 return; 433 } 434 if (n->is_MachSafePoint()) { 435 GrowableArray<SafePointScalarObjectNode*> scobjs; 436 MachSafePointNode *mcall = n->as_MachSafePoint(); 437 uint i; 438 // Print locals 439 for (i = 0; i < (uint)loc_size(); i++) 440 format_helper(regalloc, st, mcall->local(this, i), "L[", i, &scobjs); 441 // Print stack 442 for (i = 0; i < (uint)stk_size(); i++) { 443 if ((uint)(_stkoff + i) >= mcall->len()) 444 st->print(" oob "); 445 else 446 format_helper(regalloc, st, mcall->stack(this, i), "STK[", i, &scobjs); 447 } 448 for (i = 0; (int)i < nof_monitors(); i++) { 449 Node *box = mcall->monitor_box(this, i); 450 Node *obj = mcall->monitor_obj(this, i); 451 if (regalloc->node_regs_max_index() > 0 && 452 OptoReg::is_valid(regalloc->get_reg_first(box))) { 453 box = BoxLockNode::box_node(box); 454 format_helper(regalloc, st, box, "MON-BOX[", i, &scobjs); 455 } else { 456 OptoReg::Name box_reg = BoxLockNode::reg(box); 457 st->print(" MON-BOX%d=%s+%d", 458 i, 459 OptoReg::regname(OptoReg::c_frame_pointer), 460 regalloc->reg2offset(box_reg)); 461 } 462 const char* obj_msg = "MON-OBJ["; 463 if (EliminateLocks) { 464 if (BoxLockNode::box_node(box)->is_eliminated()) 465 obj_msg = "MON-OBJ(LOCK ELIMINATED)["; 466 } 467 format_helper(regalloc, st, obj, obj_msg, i, &scobjs); 468 } 469 470 for (i = 0; i < (uint)scobjs.length(); i++) { 471 // Scalar replaced objects. 472 st->cr(); 473 st->print(" # ScObj" INT32_FORMAT " ", i); 474 SafePointScalarObjectNode* spobj = scobjs.at(i); 475 ciKlass* cik = spobj->bottom_type()->is_oopptr()->klass(); 476 assert(cik->is_instance_klass() || 477 cik->is_array_klass(), "Not supported allocation."); 478 ciInstanceKlass *iklass = NULL; 479 if (cik->is_instance_klass()) { 480 cik->print_name_on(st); 481 iklass = cik->as_instance_klass(); 482 } else if (cik->is_type_array_klass()) { 483 cik->as_array_klass()->base_element_type()->print_name_on(st); 484 st->print("[%d]", spobj->n_fields()); 485 } else if (cik->is_obj_array_klass()) { 486 ciKlass* cie = cik->as_obj_array_klass()->base_element_klass(); 487 if (cie->is_instance_klass()) { 488 cie->print_name_on(st); 489 } else if (cie->is_type_array_klass()) { 490 cie->as_array_klass()->base_element_type()->print_name_on(st); 491 } else { 492 ShouldNotReachHere(); 493 } 494 st->print("[%d]", spobj->n_fields()); 495 int ndim = cik->as_array_klass()->dimension() - 1; 496 while (ndim-- > 0) { 497 st->print("[]"); 498 } 499 } 500 st->print("={"); 501 uint nf = spobj->n_fields(); 502 if (nf > 0) { 503 uint first_ind = spobj->first_index(mcall->jvms()); 504 Node* fld_node = mcall->in(first_ind); 505 ciField* cifield; 506 if (iklass != NULL) { 507 st->print(" ["); 508 cifield = iklass->nonstatic_field_at(0); 509 cifield->print_name_on(st); 510 format_helper(regalloc, st, fld_node, ":", 0, &scobjs); 511 } else { 512 format_helper(regalloc, st, fld_node, "[", 0, &scobjs); 513 } 514 for (uint j = 1; j < nf; j++) { 515 fld_node = mcall->in(first_ind+j); 516 if (iklass != NULL) { 517 st->print(", ["); 518 cifield = iklass->nonstatic_field_at(j); 519 cifield->print_name_on(st); 520 format_helper(regalloc, st, fld_node, ":", j, &scobjs); 521 } else { 522 format_helper(regalloc, st, fld_node, ", [", j, &scobjs); 523 } 524 } 525 } 526 st->print(" }"); 527 } 528 } 529 st->cr(); 530 if (caller() != NULL) caller()->format(regalloc, n, st); 531 } 532 533 534 void JVMState::dump_spec(outputStream *st) const { 535 if (_method != NULL) { 536 bool printed = false; 537 if (!Verbose) { 538 // The JVMS dumps make really, really long lines. 539 // Take out the most boring parts, which are the package prefixes. 540 char buf[500]; 541 stringStream namest(buf, sizeof(buf)); 542 _method->print_short_name(&namest); 543 if (namest.count() < sizeof(buf)) { 544 const char* name = namest.base(); 545 if (name[0] == ' ') ++name; 546 const char* endcn = strchr(name, ':'); // end of class name 547 if (endcn == NULL) endcn = strchr(name, '('); 548 if (endcn == NULL) endcn = name + strlen(name); 549 while (endcn > name && endcn[-1] != '.' && endcn[-1] != '/') 550 --endcn; 551 st->print(" %s", endcn); 552 printed = true; 553 } 554 } 555 print_method_with_lineno(st, !printed); 556 if(_reexecute == Reexecute_True) 557 st->print(" reexecute"); 558 } else { 559 st->print(" runtime stub"); 560 } 561 if (caller() != NULL) caller()->dump_spec(st); 562 } 563 564 565 void JVMState::dump_on(outputStream* st) const { 566 bool print_map = _map && !((uintptr_t)_map & 1) && 567 ((caller() == NULL) || (caller()->map() != _map)); 568 if (print_map) { 569 if (_map->len() > _map->req()) { // _map->has_exceptions() 570 Node* ex = _map->in(_map->req()); // _map->next_exception() 571 // skip the first one; it's already being printed 572 while (ex != NULL && ex->len() > ex->req()) { 573 ex = ex->in(ex->req()); // ex->next_exception() 574 ex->dump(1); 575 } 576 } 577 _map->dump(Verbose ? 2 : 1); 578 } 579 if (caller() != NULL) { 580 caller()->dump_on(st); 581 } 582 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=", 583 depth(), locoff(), stkoff(), argoff(), monoff(), scloff(), endoff(), monitor_depth(), sp(), bci(), should_reexecute()?"true":"false"); 584 if (_method == NULL) { 585 st->print_cr("(none)"); 586 } else { 587 _method->print_name(st); 588 st->cr(); 589 if (bci() >= 0 && bci() < _method->code_size()) { 590 st->print(" bc: "); 591 _method->print_codes_on(bci(), bci()+1, st); 592 } 593 } 594 } 595 596 // Extra way to dump a jvms from the debugger, 597 // to avoid a bug with C++ member function calls. 598 void dump_jvms(JVMState* jvms) { 599 jvms->dump(); 600 } 601 #endif 602 603 //--------------------------clone_shallow-------------------------------------- 604 JVMState* JVMState::clone_shallow(Compile* C) const { 605 JVMState* n = has_method() ? new (C) JVMState(_method, _caller) : new (C) JVMState(0); 606 n->set_bci(_bci); 607 n->_reexecute = _reexecute; 608 n->set_locoff(_locoff); 609 n->set_stkoff(_stkoff); 610 n->set_monoff(_monoff); 611 n->set_scloff(_scloff); 612 n->set_endoff(_endoff); 613 n->set_sp(_sp); 614 n->set_map(_map); 615 return n; 616 } 617 618 //---------------------------clone_deep---------------------------------------- 619 JVMState* JVMState::clone_deep(Compile* C) const { 620 JVMState* n = clone_shallow(C); 621 for (JVMState* p = n; p->_caller != NULL; p = p->_caller) { 622 p->_caller = p->_caller->clone_shallow(C); 623 } 624 assert(n->depth() == depth(), "sanity"); 625 assert(n->debug_depth() == debug_depth(), "sanity"); 626 return n; 627 } 628 629 /** 630 * Reset map for all callers 631 */ 632 void JVMState::set_map_deep(SafePointNode* map) { 633 for (JVMState* p = this; p != NULL; p = p->_caller) { 634 p->set_map(map); 635 } 636 } 637 638 // unlike set_map(), this is two-way setting. 639 void JVMState::bind_map(SafePointNode* map) { 640 set_map(map); 641 _map->set_jvms(this); 642 } 643 644 // Adapt offsets in in-array after adding or removing an edge. 645 // Prerequisite is that the JVMState is used by only one node. 646 void JVMState::adapt_position(int delta) { 647 for (JVMState* jvms = this; jvms != NULL; jvms = jvms->caller()) { 648 jvms->set_locoff(jvms->locoff() + delta); 649 jvms->set_stkoff(jvms->stkoff() + delta); 650 jvms->set_monoff(jvms->monoff() + delta); 651 jvms->set_scloff(jvms->scloff() + delta); 652 jvms->set_endoff(jvms->endoff() + delta); 653 } 654 } 655 656 // Mirror the stack size calculation in the deopt code 657 // How much stack space would we need at this point in the program in 658 // case of deoptimization? 659 int JVMState::interpreter_frame_size() const { 660 const JVMState* jvms = this; 661 int size = 0; 662 int callee_parameters = 0; 663 int callee_locals = 0; 664 int extra_args = method()->max_stack() - stk_size(); 665 666 while (jvms != NULL) { 667 int locks = jvms->nof_monitors(); 668 int temps = jvms->stk_size(); 669 bool is_top_frame = (jvms == this); 670 ciMethod* method = jvms->method(); 671 672 int frame_size = BytesPerWord * Interpreter::size_activation(method->max_stack(), 673 temps + callee_parameters, 674 extra_args, 675 locks, 676 callee_parameters, 677 callee_locals, 678 is_top_frame); 679 size += frame_size; 680 681 callee_parameters = method->size_of_parameters(); 682 callee_locals = method->max_locals(); 683 extra_args = 0; 684 jvms = jvms->caller(); 685 } 686 return size + Deoptimization::last_frame_adjust(0, callee_locals) * BytesPerWord; 687 } 688 689 //============================================================================= 690 bool CallNode::cmp( const Node &n ) const 691 { return _tf == ((CallNode&)n)._tf && _jvms == ((CallNode&)n)._jvms; } 692 #ifndef PRODUCT 693 void CallNode::dump_req(outputStream *st) const { 694 // Dump the required inputs, enclosed in '(' and ')' 695 uint i; // Exit value of loop 696 for (i = 0; i < req(); i++) { // For all required inputs 697 if (i == TypeFunc::Parms) st->print("("); 698 if (in(i)) st->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx); 699 else st->print("_ "); 700 } 701 st->print(")"); 702 } 703 704 void CallNode::dump_spec(outputStream *st) const { 705 st->print(" "); 706 if (tf() != NULL) tf()->dump_on(st); 707 if (_cnt != COUNT_UNKNOWN) st->print(" C=%f",_cnt); 708 if (jvms() != NULL) jvms()->dump_spec(st); 709 } 710 #endif 711 712 const Type *CallNode::bottom_type() const { return tf()->range(); } 713 const Type* CallNode::Value(PhaseGVN* phase) const { 714 if (phase->type(in(0)) == Type::TOP) return Type::TOP; 715 return tf()->range(); 716 } 717 718 //------------------------------calling_convention----------------------------- 719 void CallNode::calling_convention(BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt) const { 720 // Use the standard compiler calling convention 721 SharedRuntime::java_calling_convention(sig_bt, parm_regs, argcnt); 722 } 723 724 725 //------------------------------match------------------------------------------ 726 // Construct projections for control, I/O, memory-fields, ..., and 727 // return result(s) along with their RegMask info 728 Node *CallNode::match( const ProjNode *proj, const Matcher *match ) { 729 switch (proj->_con) { 730 case TypeFunc::Control: 731 case TypeFunc::I_O: 732 case TypeFunc::Memory: 733 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 734 735 case TypeFunc::Parms+1: // For LONG & DOUBLE returns 736 assert(tf()->range()->field_at(TypeFunc::Parms+1) == Type::HALF, ""); 737 // 2nd half of doubles and longs 738 return new MachProjNode(this,proj->_con, RegMask::Empty, (uint)OptoReg::Bad); 739 740 case TypeFunc::Parms: { // Normal returns 741 uint ideal_reg = tf()->range()->field_at(TypeFunc::Parms)->ideal_reg(); 742 OptoRegPair regs = Opcode() == Op_CallLeafVector 743 ? match->vector_return_value(ideal_reg) // Calls into assembly vector routine 744 : is_CallRuntime() 745 ? match->c_return_value(ideal_reg) // Calls into C runtime 746 : match-> return_value(ideal_reg); // Calls into compiled Java code 747 RegMask rm = RegMask(regs.first()); 748 749 if (Opcode() == Op_CallLeafVector) { 750 // If the return is in vector, compute appropriate regmask taking into account the whole range 751 if(ideal_reg >= Op_VecS && ideal_reg <= Op_VecZ) { 752 if(OptoReg::is_valid(regs.second())) { 753 for (OptoReg::Name r = regs.first(); r <= regs.second(); r = OptoReg::add(r, 1)) { 754 rm.Insert(r); 755 } 756 } 757 } 758 } 759 760 if( OptoReg::is_valid(regs.second()) ) 761 rm.Insert( regs.second() ); 762 return new MachProjNode(this,proj->_con,rm,ideal_reg); 763 } 764 765 case TypeFunc::ReturnAdr: 766 case TypeFunc::FramePtr: 767 default: 768 ShouldNotReachHere(); 769 } 770 return NULL; 771 } 772 773 // Do we Match on this edge index or not? Match no edges 774 uint CallNode::match_edge(uint idx) const { 775 return 0; 776 } 777 778 // 779 // Determine whether the call could modify the field of the specified 780 // instance at the specified offset. 781 // 782 bool CallNode::may_modify(const TypeOopPtr *t_oop, PhaseTransform *phase) { 783 assert((t_oop != NULL), "sanity"); 784 if (is_call_to_arraycopystub() && strcmp(_name, "unsafe_arraycopy") != 0) { 785 const TypeTuple* args = _tf->domain(); 786 Node* dest = NULL; 787 // Stubs that can be called once an ArrayCopyNode is expanded have 788 // different signatures. Look for the second pointer argument, 789 // that is the destination of the copy. 790 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) { 791 if (args->field_at(i)->isa_ptr()) { 792 j++; 793 if (j == 2) { 794 dest = in(i); 795 break; 796 } 797 } 798 } 799 guarantee(dest != NULL, "Call had only one ptr in, broken IR!"); 800 if (!dest->is_top() && may_modify_arraycopy_helper(phase->type(dest)->is_oopptr(), t_oop, phase)) { 801 return true; 802 } 803 return false; 804 } 805 if (t_oop->is_known_instance()) { 806 // The instance_id is set only for scalar-replaceable allocations which 807 // are not passed as arguments according to Escape Analysis. 808 return false; 809 } 810 if (t_oop->is_ptr_to_boxed_value()) { 811 ciKlass* boxing_klass = t_oop->klass(); 812 if (is_CallStaticJava() && as_CallStaticJava()->is_boxing_method()) { 813 // Skip unrelated boxing methods. 814 Node* proj = proj_out_or_null(TypeFunc::Parms); 815 if ((proj == NULL) || (phase->type(proj)->is_instptr()->klass() != boxing_klass)) { 816 return false; 817 } 818 } 819 if (is_CallJava() && as_CallJava()->method() != NULL) { 820 ciMethod* meth = as_CallJava()->method(); 821 if (meth->is_getter()) { 822 return false; 823 } 824 // May modify (by reflection) if an boxing object is passed 825 // as argument or returned. 826 Node* proj = returns_pointer() ? proj_out_or_null(TypeFunc::Parms) : NULL; 827 if (proj != NULL) { 828 const TypeInstPtr* inst_t = phase->type(proj)->isa_instptr(); 829 if ((inst_t != NULL) && (!inst_t->klass_is_exact() || 830 (inst_t->klass() == boxing_klass))) { 831 return true; 832 } 833 } 834 const TypeTuple* d = tf()->domain(); 835 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 836 const TypeInstPtr* inst_t = d->field_at(i)->isa_instptr(); 837 if ((inst_t != NULL) && (!inst_t->klass_is_exact() || 838 (inst_t->klass() == boxing_klass))) { 839 return true; 840 } 841 } 842 return false; 843 } 844 } 845 return true; 846 } 847 848 // Does this call have a direct reference to n other than debug information? 849 bool CallNode::has_non_debug_use(Node *n) { 850 const TypeTuple * d = tf()->domain(); 851 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 852 Node *arg = in(i); 853 if (arg == n) { 854 return true; 855 } 856 } 857 return false; 858 } 859 860 // Returns the unique CheckCastPP of a call 861 // or 'this' if there are several CheckCastPP or unexpected uses 862 // or returns NULL if there is no one. 863 Node *CallNode::result_cast() { 864 Node *cast = NULL; 865 866 Node *p = proj_out_or_null(TypeFunc::Parms); 867 if (p == NULL) 868 return NULL; 869 870 for (DUIterator_Fast imax, i = p->fast_outs(imax); i < imax; i++) { 871 Node *use = p->fast_out(i); 872 if (use->is_CheckCastPP()) { 873 if (cast != NULL) { 874 return this; // more than 1 CheckCastPP 875 } 876 cast = use; 877 } else if (!use->is_Initialize() && 878 !use->is_AddP() && 879 use->Opcode() != Op_MemBarStoreStore) { 880 // Expected uses are restricted to a CheckCastPP, an Initialize 881 // node, a MemBarStoreStore (clone) and AddP nodes. If we 882 // encounter any other use (a Phi node can be seen in rare 883 // cases) return this to prevent incorrect optimizations. 884 return this; 885 } 886 } 887 return cast; 888 } 889 890 891 void CallNode::extract_projections(CallProjections* projs, bool separate_io_proj, bool do_asserts) { 892 projs->fallthrough_proj = NULL; 893 projs->fallthrough_catchproj = NULL; 894 projs->fallthrough_ioproj = NULL; 895 projs->catchall_ioproj = NULL; 896 projs->catchall_catchproj = NULL; 897 projs->fallthrough_memproj = NULL; 898 projs->catchall_memproj = NULL; 899 projs->resproj = NULL; 900 projs->exobj = NULL; 901 902 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { 903 ProjNode *pn = fast_out(i)->as_Proj(); 904 if (pn->outcnt() == 0) continue; 905 switch (pn->_con) { 906 case TypeFunc::Control: 907 { 908 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj 909 projs->fallthrough_proj = pn; 910 const Node* cn = pn->unique_ctrl_out_or_null(); 911 if (cn != NULL && cn->is_Catch()) { 912 ProjNode *cpn = NULL; 913 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) { 914 cpn = cn->fast_out(k)->as_Proj(); 915 assert(cpn->is_CatchProj(), "must be a CatchProjNode"); 916 if (cpn->_con == CatchProjNode::fall_through_index) 917 projs->fallthrough_catchproj = cpn; 918 else { 919 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index."); 920 projs->catchall_catchproj = cpn; 921 } 922 } 923 } 924 break; 925 } 926 case TypeFunc::I_O: 927 if (pn->_is_io_use) 928 projs->catchall_ioproj = pn; 929 else 930 projs->fallthrough_ioproj = pn; 931 for (DUIterator j = pn->outs(); pn->has_out(j); j++) { 932 Node* e = pn->out(j); 933 if (e->Opcode() == Op_CreateEx && e->in(0)->is_CatchProj() && e->outcnt() > 0) { 934 assert(projs->exobj == NULL, "only one"); 935 projs->exobj = e; 936 } 937 } 938 break; 939 case TypeFunc::Memory: 940 if (pn->_is_io_use) 941 projs->catchall_memproj = pn; 942 else 943 projs->fallthrough_memproj = pn; 944 break; 945 case TypeFunc::Parms: 946 projs->resproj = pn; 947 break; 948 default: 949 assert(false, "unexpected projection from allocation node."); 950 } 951 } 952 953 // The resproj may not exist because the result could be ignored 954 // and the exception object may not exist if an exception handler 955 // swallows the exception but all the other must exist and be found. 956 assert(projs->fallthrough_proj != NULL, "must be found"); 957 do_asserts = do_asserts && !Compile::current()->inlining_incrementally(); 958 assert(!do_asserts || projs->fallthrough_catchproj != NULL, "must be found"); 959 assert(!do_asserts || projs->fallthrough_memproj != NULL, "must be found"); 960 assert(!do_asserts || projs->fallthrough_ioproj != NULL, "must be found"); 961 assert(!do_asserts || projs->catchall_catchproj != NULL, "must be found"); 962 if (separate_io_proj) { 963 assert(!do_asserts || projs->catchall_memproj != NULL, "must be found"); 964 assert(!do_asserts || projs->catchall_ioproj != NULL, "must be found"); 965 } 966 } 967 968 Node* CallNode::Ideal(PhaseGVN* phase, bool can_reshape) { 969 #ifdef ASSERT 970 // Validate attached generator 971 CallGenerator* cg = generator(); 972 if (cg != NULL) { 973 assert(is_CallStaticJava() && cg->is_mh_late_inline() || 974 is_CallDynamicJava() && cg->is_virtual_late_inline(), "mismatch"); 975 } 976 #endif // ASSERT 977 return SafePointNode::Ideal(phase, can_reshape); 978 } 979 980 bool CallNode::is_call_to_arraycopystub() const { 981 if (_name != NULL && strstr(_name, "arraycopy") != 0) { 982 return true; 983 } 984 return false; 985 } 986 987 //============================================================================= 988 uint CallJavaNode::size_of() const { return sizeof(*this); } 989 bool CallJavaNode::cmp( const Node &n ) const { 990 CallJavaNode &call = (CallJavaNode&)n; 991 return CallNode::cmp(call) && _method == call._method && 992 _override_symbolic_info == call._override_symbolic_info; 993 } 994 995 void CallJavaNode::copy_call_debug_info(PhaseIterGVN* phase, SafePointNode* sfpt) { 996 // Copy debug information and adjust JVMState information 997 uint old_dbg_start = sfpt->is_Call() ? sfpt->as_Call()->tf()->domain()->cnt() : (uint)TypeFunc::Parms+1; 998 uint new_dbg_start = tf()->domain()->cnt(); 999 int jvms_adj = new_dbg_start - old_dbg_start; 1000 assert (new_dbg_start == req(), "argument count mismatch"); 1001 Compile* C = phase->C; 1002 1003 // SafePointScalarObject node could be referenced several times in debug info. 1004 // Use Dict to record cloned nodes. 1005 Dict* sosn_map = new Dict(cmpkey,hashkey); 1006 for (uint i = old_dbg_start; i < sfpt->req(); i++) { 1007 Node* old_in = sfpt->in(i); 1008 // Clone old SafePointScalarObjectNodes, adjusting their field contents. 1009 if (old_in != NULL && old_in->is_SafePointScalarObject()) { 1010 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject(); 1011 bool new_node; 1012 Node* new_in = old_sosn->clone(sosn_map, new_node); 1013 if (new_node) { // New node? 1014 new_in->set_req(0, C->root()); // reset control edge 1015 new_in = phase->transform(new_in); // Register new node. 1016 } 1017 old_in = new_in; 1018 } 1019 add_req(old_in); 1020 } 1021 1022 // JVMS may be shared so clone it before we modify it 1023 set_jvms(sfpt->jvms() != NULL ? sfpt->jvms()->clone_deep(C) : NULL); 1024 for (JVMState *jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) { 1025 jvms->set_map(this); 1026 jvms->set_locoff(jvms->locoff()+jvms_adj); 1027 jvms->set_stkoff(jvms->stkoff()+jvms_adj); 1028 jvms->set_monoff(jvms->monoff()+jvms_adj); 1029 jvms->set_scloff(jvms->scloff()+jvms_adj); 1030 jvms->set_endoff(jvms->endoff()+jvms_adj); 1031 } 1032 } 1033 1034 #ifdef ASSERT 1035 bool CallJavaNode::validate_symbolic_info() const { 1036 if (method() == NULL) { 1037 return true; // call into runtime or uncommon trap 1038 } 1039 ciMethod* symbolic_info = jvms()->method()->get_method_at_bci(jvms()->bci()); 1040 ciMethod* callee = method(); 1041 if (symbolic_info->is_method_handle_intrinsic() && !callee->is_method_handle_intrinsic()) { 1042 assert(override_symbolic_info(), "should be set"); 1043 } 1044 assert(ciMethod::is_consistent_info(symbolic_info, callee), "inconsistent info"); 1045 return true; 1046 } 1047 #endif 1048 1049 #ifndef PRODUCT 1050 void CallJavaNode::dump_spec(outputStream* st) const { 1051 if( _method ) _method->print_short_name(st); 1052 CallNode::dump_spec(st); 1053 } 1054 1055 void CallJavaNode::dump_compact_spec(outputStream* st) const { 1056 if (_method) { 1057 _method->print_short_name(st); 1058 } else { 1059 st->print("<?>"); 1060 } 1061 } 1062 #endif 1063 1064 //============================================================================= 1065 uint CallStaticJavaNode::size_of() const { return sizeof(*this); } 1066 bool CallStaticJavaNode::cmp( const Node &n ) const { 1067 CallStaticJavaNode &call = (CallStaticJavaNode&)n; 1068 return CallJavaNode::cmp(call); 1069 } 1070 1071 Node* CallStaticJavaNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1072 CallGenerator* cg = generator(); 1073 if (can_reshape && cg != NULL) { 1074 assert(IncrementalInlineMH, "required"); 1075 assert(cg->call_node() == this, "mismatch"); 1076 assert(cg->is_mh_late_inline(), "not virtual"); 1077 1078 // Check whether this MH handle call becomes a candidate for inlining. 1079 ciMethod* callee = cg->method(); 1080 vmIntrinsics::ID iid = callee->intrinsic_id(); 1081 if (iid == vmIntrinsics::_invokeBasic) { 1082 if (in(TypeFunc::Parms)->Opcode() == Op_ConP) { 1083 phase->C->prepend_late_inline(cg); 1084 set_generator(NULL); 1085 } 1086 } else if (iid == vmIntrinsics::_linkToNative) { 1087 Node* nep_node = in(TypeFunc::Parms + callee->arg_size() - 1); 1088 if (nep_node->Opcode() == Op_ConP /* NEP */ 1089 && in(TypeFunc::Parms + 0)->Opcode() == Op_ConL /* address */ 1090 && !nep_node->bottom_type()->is_oopptr()->const_oop()->as_native_entry_point()->need_transition()) { 1091 phase->C->prepend_late_inline(cg); 1092 set_generator(NULL); 1093 } 1094 } else { 1095 assert(callee->has_member_arg(), "wrong type of call?"); 1096 if (in(TypeFunc::Parms + callee->arg_size() - 1)->Opcode() == Op_ConP) { 1097 phase->C->prepend_late_inline(cg); 1098 set_generator(NULL); 1099 } 1100 } 1101 } 1102 return CallNode::Ideal(phase, can_reshape); 1103 } 1104 1105 //----------------------------uncommon_trap_request---------------------------- 1106 // If this is an uncommon trap, return the request code, else zero. 1107 int CallStaticJavaNode::uncommon_trap_request() const { 1108 if (_name != NULL && !strcmp(_name, "uncommon_trap")) { 1109 return extract_uncommon_trap_request(this); 1110 } 1111 return 0; 1112 } 1113 int CallStaticJavaNode::extract_uncommon_trap_request(const Node* call) { 1114 #ifndef PRODUCT 1115 if (!(call->req() > TypeFunc::Parms && 1116 call->in(TypeFunc::Parms) != NULL && 1117 call->in(TypeFunc::Parms)->is_Con() && 1118 call->in(TypeFunc::Parms)->bottom_type()->isa_int())) { 1119 assert(in_dump() != 0, "OK if dumping"); 1120 tty->print("[bad uncommon trap]"); 1121 return 0; 1122 } 1123 #endif 1124 return call->in(TypeFunc::Parms)->bottom_type()->is_int()->get_con(); 1125 } 1126 1127 #ifndef PRODUCT 1128 void CallStaticJavaNode::dump_spec(outputStream *st) const { 1129 st->print("# Static "); 1130 if (_name != NULL) { 1131 st->print("%s", _name); 1132 int trap_req = uncommon_trap_request(); 1133 if (trap_req != 0) { 1134 char buf[100]; 1135 st->print("(%s)", 1136 Deoptimization::format_trap_request(buf, sizeof(buf), 1137 trap_req)); 1138 } 1139 st->print(" "); 1140 } 1141 CallJavaNode::dump_spec(st); 1142 } 1143 1144 void CallStaticJavaNode::dump_compact_spec(outputStream* st) const { 1145 if (_method) { 1146 _method->print_short_name(st); 1147 } else if (_name) { 1148 st->print("%s", _name); 1149 } else { 1150 st->print("<?>"); 1151 } 1152 } 1153 #endif 1154 1155 //============================================================================= 1156 uint CallDynamicJavaNode::size_of() const { return sizeof(*this); } 1157 bool CallDynamicJavaNode::cmp( const Node &n ) const { 1158 CallDynamicJavaNode &call = (CallDynamicJavaNode&)n; 1159 return CallJavaNode::cmp(call); 1160 } 1161 1162 Node* CallDynamicJavaNode::Ideal(PhaseGVN* phase, bool can_reshape) { 1163 CallGenerator* cg = generator(); 1164 if (can_reshape && cg != NULL) { 1165 assert(IncrementalInlineVirtual, "required"); 1166 assert(cg->call_node() == this, "mismatch"); 1167 assert(cg->is_virtual_late_inline(), "not virtual"); 1168 1169 // Recover symbolic info for method resolution. 1170 ciMethod* caller = jvms()->method(); 1171 ciBytecodeStream iter(caller); 1172 iter.force_bci(jvms()->bci()); 1173 1174 bool not_used1; 1175 ciSignature* not_used2; 1176 ciMethod* orig_callee = iter.get_method(not_used1, ¬_used2); // callee in the bytecode 1177 ciKlass* holder = iter.get_declared_method_holder(); 1178 if (orig_callee->is_method_handle_intrinsic()) { 1179 assert(_override_symbolic_info, "required"); 1180 orig_callee = method(); 1181 holder = method()->holder(); 1182 } 1183 1184 ciInstanceKlass* klass = ciEnv::get_instance_klass_for_declared_method_holder(holder); 1185 1186 Node* receiver_node = in(TypeFunc::Parms); 1187 const TypeOopPtr* receiver_type = phase->type(receiver_node)->isa_oopptr(); 1188 1189 int not_used3; 1190 bool call_does_dispatch; 1191 ciMethod* callee = phase->C->optimize_virtual_call(caller, klass, holder, orig_callee, receiver_type, true /*is_virtual*/, 1192 call_does_dispatch, not_used3); // out-parameters 1193 if (!call_does_dispatch) { 1194 // Register for late inlining. 1195 cg->set_callee_method(callee); 1196 phase->C->prepend_late_inline(cg); // MH late inlining prepends to the list, so do the same 1197 set_generator(NULL); 1198 } 1199 } 1200 return CallNode::Ideal(phase, can_reshape); 1201 } 1202 1203 #ifndef PRODUCT 1204 void CallDynamicJavaNode::dump_spec(outputStream *st) const { 1205 st->print("# Dynamic "); 1206 CallJavaNode::dump_spec(st); 1207 } 1208 #endif 1209 1210 //============================================================================= 1211 uint CallRuntimeNode::size_of() const { return sizeof(*this); } 1212 bool CallRuntimeNode::cmp( const Node &n ) const { 1213 CallRuntimeNode &call = (CallRuntimeNode&)n; 1214 return CallNode::cmp(call) && !strcmp(_name,call._name); 1215 } 1216 #ifndef PRODUCT 1217 void CallRuntimeNode::dump_spec(outputStream *st) const { 1218 st->print("# "); 1219 st->print("%s", _name); 1220 CallNode::dump_spec(st); 1221 } 1222 #endif 1223 uint CallLeafVectorNode::size_of() const { return sizeof(*this); } 1224 bool CallLeafVectorNode::cmp( const Node &n ) const { 1225 CallLeafVectorNode &call = (CallLeafVectorNode&)n; 1226 return CallLeafNode::cmp(call) && _num_bits == call._num_bits; 1227 } 1228 1229 //============================================================================= 1230 uint CallNativeNode::size_of() const { return sizeof(*this); } 1231 bool CallNativeNode::cmp( const Node &n ) const { 1232 CallNativeNode &call = (CallNativeNode&)n; 1233 return CallNode::cmp(call) && !strcmp(_name,call._name) 1234 && _arg_regs == call._arg_regs && _ret_regs == call._ret_regs; 1235 } 1236 Node* CallNativeNode::match(const ProjNode *proj, const Matcher *matcher) { 1237 switch (proj->_con) { 1238 case TypeFunc::Control: 1239 case TypeFunc::I_O: 1240 case TypeFunc::Memory: 1241 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); 1242 case TypeFunc::ReturnAdr: 1243 case TypeFunc::FramePtr: 1244 ShouldNotReachHere(); 1245 case TypeFunc::Parms: { 1246 const Type* field_at_con = tf()->range()->field_at(proj->_con); 1247 const BasicType bt = field_at_con->basic_type(); 1248 OptoReg::Name optoreg = OptoReg::as_OptoReg(_ret_regs.at(proj->_con - TypeFunc::Parms)); 1249 OptoRegPair regs; 1250 if (bt == T_DOUBLE || bt == T_LONG) { 1251 regs.set2(optoreg); 1252 } else { 1253 regs.set1(optoreg); 1254 } 1255 RegMask rm = RegMask(regs.first()); 1256 if(OptoReg::is_valid(regs.second())) 1257 rm.Insert(regs.second()); 1258 return new MachProjNode(this, proj->_con, rm, field_at_con->ideal_reg()); 1259 } 1260 case TypeFunc::Parms + 1: { 1261 assert(tf()->range()->field_at(proj->_con) == Type::HALF, "Expected HALF"); 1262 assert(_ret_regs.at(proj->_con - TypeFunc::Parms) == VMRegImpl::Bad(), "Unexpected register for Type::HALF"); 1263 // 2nd half of doubles and longs 1264 return new MachProjNode(this, proj->_con, RegMask::Empty, (uint) OptoReg::Bad); 1265 } 1266 default: 1267 ShouldNotReachHere(); 1268 } 1269 return NULL; 1270 } 1271 #ifndef PRODUCT 1272 void CallNativeNode::print_regs(const GrowableArray<VMReg>& regs, outputStream* st) { 1273 st->print("{ "); 1274 for (int i = 0; i < regs.length(); i++) { 1275 regs.at(i)->print_on(st); 1276 if (i < regs.length() - 1) { 1277 st->print(", "); 1278 } 1279 } 1280 st->print(" } "); 1281 } 1282 1283 void CallNativeNode::dump_spec(outputStream *st) const { 1284 st->print("# "); 1285 st->print("%s ", _name); 1286 st->print("_arg_regs: "); 1287 print_regs(_arg_regs, st); 1288 st->print("_ret_regs: "); 1289 print_regs(_ret_regs, st); 1290 CallNode::dump_spec(st); 1291 } 1292 #endif 1293 1294 //------------------------------calling_convention----------------------------- 1295 void CallRuntimeNode::calling_convention(BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt) const { 1296 SharedRuntime::c_calling_convention(sig_bt, parm_regs, /*regs2=*/nullptr, argcnt); 1297 } 1298 1299 void CallLeafVectorNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const { 1300 #ifdef ASSERT 1301 assert(tf()->range()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte == _num_bits, 1302 "return vector size must match"); 1303 const TypeTuple* d = tf()->domain(); 1304 for (uint i = TypeFunc::Parms; i < d->cnt(); i++) { 1305 Node* arg = in(i); 1306 assert(arg->bottom_type()->is_vect()->length_in_bytes() * BitsPerByte == _num_bits, 1307 "vector argument size must match"); 1308 } 1309 #endif 1310 1311 SharedRuntime::vector_calling_convention(parm_regs, _num_bits, argcnt); 1312 } 1313 1314 void CallNativeNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const { 1315 assert((tf()->domain()->cnt() - TypeFunc::Parms) == argcnt, "arg counts must match!"); 1316 #ifdef ASSERT 1317 for (uint i = 0; i < argcnt; i++) { 1318 assert(tf()->domain()->field_at(TypeFunc::Parms + i)->basic_type() == sig_bt[i], "types must match!"); 1319 } 1320 #endif 1321 for (uint i = 0; i < argcnt; i++) { 1322 switch (sig_bt[i]) { 1323 case T_BOOLEAN: 1324 case T_CHAR: 1325 case T_BYTE: 1326 case T_SHORT: 1327 case T_INT: 1328 case T_FLOAT: 1329 parm_regs[i].set1(_arg_regs.at(i)); 1330 break; 1331 case T_LONG: 1332 case T_DOUBLE: 1333 assert((i + 1) < argcnt && sig_bt[i + 1] == T_VOID, "expecting half"); 1334 parm_regs[i].set2(_arg_regs.at(i)); 1335 break; 1336 case T_VOID: // Halves of longs and doubles 1337 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half"); 1338 assert(_arg_regs.at(i) == VMRegImpl::Bad(), "expecting bad reg"); 1339 parm_regs[i].set_bad(); 1340 break; 1341 default: 1342 ShouldNotReachHere(); 1343 break; 1344 } 1345 } 1346 } 1347 1348 //============================================================================= 1349 //------------------------------calling_convention----------------------------- 1350 1351 1352 //============================================================================= 1353 #ifndef PRODUCT 1354 void CallLeafNode::dump_spec(outputStream *st) const { 1355 st->print("# "); 1356 st->print("%s", _name); 1357 CallNode::dump_spec(st); 1358 } 1359 #endif 1360 1361 //============================================================================= 1362 1363 void SafePointNode::set_local(JVMState* jvms, uint idx, Node *c) { 1364 assert(verify_jvms(jvms), "jvms must match"); 1365 int loc = jvms->locoff() + idx; 1366 if (in(loc)->is_top() && idx > 0 && !c->is_top() ) { 1367 // If current local idx is top then local idx - 1 could 1368 // be a long/double that needs to be killed since top could 1369 // represent the 2nd half of the long/double. 1370 uint ideal = in(loc -1)->ideal_reg(); 1371 if (ideal == Op_RegD || ideal == Op_RegL) { 1372 // set other (low index) half to top 1373 set_req(loc - 1, in(loc)); 1374 } 1375 } 1376 set_req(loc, c); 1377 } 1378 1379 uint SafePointNode::size_of() const { return sizeof(*this); } 1380 bool SafePointNode::cmp( const Node &n ) const { 1381 return (&n == this); // Always fail except on self 1382 } 1383 1384 //-------------------------set_next_exception---------------------------------- 1385 void SafePointNode::set_next_exception(SafePointNode* n) { 1386 assert(n == NULL || n->Opcode() == Op_SafePoint, "correct value for next_exception"); 1387 if (len() == req()) { 1388 if (n != NULL) add_prec(n); 1389 } else { 1390 set_prec(req(), n); 1391 } 1392 } 1393 1394 1395 //----------------------------next_exception----------------------------------- 1396 SafePointNode* SafePointNode::next_exception() const { 1397 if (len() == req()) { 1398 return NULL; 1399 } else { 1400 Node* n = in(req()); 1401 assert(n == NULL || n->Opcode() == Op_SafePoint, "no other uses of prec edges"); 1402 return (SafePointNode*) n; 1403 } 1404 } 1405 1406 1407 //------------------------------Ideal------------------------------------------ 1408 // Skip over any collapsed Regions 1409 Node *SafePointNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1410 assert(_jvms == NULL || ((uintptr_t)_jvms->map() & 1) || _jvms->map() == this, "inconsistent JVMState"); 1411 return remove_dead_region(phase, can_reshape) ? this : NULL; 1412 } 1413 1414 //------------------------------Identity--------------------------------------- 1415 // Remove obviously duplicate safepoints 1416 Node* SafePointNode::Identity(PhaseGVN* phase) { 1417 1418 // If you have back to back safepoints, remove one 1419 if (in(TypeFunc::Control)->is_SafePoint()) { 1420 Node* out_c = unique_ctrl_out_or_null(); 1421 // This can be the safepoint of an outer strip mined loop if the inner loop's backedge was removed. Replacing the 1422 // outer loop's safepoint could confuse removal of the outer loop. 1423 if (out_c != NULL && !out_c->is_OuterStripMinedLoopEnd()) { 1424 return in(TypeFunc::Control); 1425 } 1426 } 1427 1428 // Transforming long counted loops requires a safepoint node. Do not 1429 // eliminate a safepoint until loop opts are over. 1430 if (in(0)->is_Proj() && !phase->C->major_progress()) { 1431 Node *n0 = in(0)->in(0); 1432 // Check if he is a call projection (except Leaf Call) 1433 if( n0->is_Catch() ) { 1434 n0 = n0->in(0)->in(0); 1435 assert( n0->is_Call(), "expect a call here" ); 1436 } 1437 if( n0->is_Call() && n0->as_Call()->guaranteed_safepoint() ) { 1438 // Don't remove a safepoint belonging to an OuterStripMinedLoopEndNode. 1439 // If the loop dies, they will be removed together. 1440 if (has_out_with(Op_OuterStripMinedLoopEnd)) { 1441 return this; 1442 } 1443 // Useless Safepoint, so remove it 1444 return in(TypeFunc::Control); 1445 } 1446 } 1447 1448 return this; 1449 } 1450 1451 //------------------------------Value------------------------------------------ 1452 const Type* SafePointNode::Value(PhaseGVN* phase) const { 1453 if (phase->type(in(0)) == Type::TOP) { 1454 return Type::TOP; 1455 } 1456 if (in(0) == this) { 1457 return Type::TOP; // Dead infinite loop 1458 } 1459 return Type::CONTROL; 1460 } 1461 1462 #ifndef PRODUCT 1463 void SafePointNode::dump_spec(outputStream *st) const { 1464 st->print(" SafePoint "); 1465 _replaced_nodes.dump(st); 1466 } 1467 1468 // The related nodes of a SafepointNode are all data inputs, excluding the 1469 // control boundary, as well as all outputs till level 2 (to include projection 1470 // nodes and targets). In compact mode, just include inputs till level 1 and 1471 // outputs as before. 1472 void SafePointNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const { 1473 if (compact) { 1474 this->collect_nodes(in_rel, 1, false, false); 1475 } else { 1476 this->collect_nodes_in_all_data(in_rel, false); 1477 } 1478 this->collect_nodes(out_rel, -2, false, false); 1479 } 1480 #endif 1481 1482 const RegMask &SafePointNode::in_RegMask(uint idx) const { 1483 if( idx < TypeFunc::Parms ) return RegMask::Empty; 1484 // Values outside the domain represent debug info 1485 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]); 1486 } 1487 const RegMask &SafePointNode::out_RegMask() const { 1488 return RegMask::Empty; 1489 } 1490 1491 1492 void SafePointNode::grow_stack(JVMState* jvms, uint grow_by) { 1493 assert((int)grow_by > 0, "sanity"); 1494 int monoff = jvms->monoff(); 1495 int scloff = jvms->scloff(); 1496 int endoff = jvms->endoff(); 1497 assert(endoff == (int)req(), "no other states or debug info after me"); 1498 Node* top = Compile::current()->top(); 1499 for (uint i = 0; i < grow_by; i++) { 1500 ins_req(monoff, top); 1501 } 1502 jvms->set_monoff(monoff + grow_by); 1503 jvms->set_scloff(scloff + grow_by); 1504 jvms->set_endoff(endoff + grow_by); 1505 } 1506 1507 void SafePointNode::push_monitor(const FastLockNode *lock) { 1508 // Add a LockNode, which points to both the original BoxLockNode (the 1509 // stack space for the monitor) and the Object being locked. 1510 const int MonitorEdges = 2; 1511 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges"); 1512 assert(req() == jvms()->endoff(), "correct sizing"); 1513 int nextmon = jvms()->scloff(); 1514 if (GenerateSynchronizationCode) { 1515 ins_req(nextmon, lock->box_node()); 1516 ins_req(nextmon+1, lock->obj_node()); 1517 } else { 1518 Node* top = Compile::current()->top(); 1519 ins_req(nextmon, top); 1520 ins_req(nextmon, top); 1521 } 1522 jvms()->set_scloff(nextmon + MonitorEdges); 1523 jvms()->set_endoff(req()); 1524 } 1525 1526 void SafePointNode::pop_monitor() { 1527 // Delete last monitor from debug info 1528 debug_only(int num_before_pop = jvms()->nof_monitors()); 1529 const int MonitorEdges = 2; 1530 assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges"); 1531 int scloff = jvms()->scloff(); 1532 int endoff = jvms()->endoff(); 1533 int new_scloff = scloff - MonitorEdges; 1534 int new_endoff = endoff - MonitorEdges; 1535 jvms()->set_scloff(new_scloff); 1536 jvms()->set_endoff(new_endoff); 1537 while (scloff > new_scloff) del_req_ordered(--scloff); 1538 assert(jvms()->nof_monitors() == num_before_pop-1, ""); 1539 } 1540 1541 Node *SafePointNode::peek_monitor_box() const { 1542 int mon = jvms()->nof_monitors() - 1; 1543 assert(mon >= 0, "must have a monitor"); 1544 return monitor_box(jvms(), mon); 1545 } 1546 1547 Node *SafePointNode::peek_monitor_obj() const { 1548 int mon = jvms()->nof_monitors() - 1; 1549 assert(mon >= 0, "must have a monitor"); 1550 return monitor_obj(jvms(), mon); 1551 } 1552 1553 // Do we Match on this edge index or not? Match no edges 1554 uint SafePointNode::match_edge(uint idx) const { 1555 return (TypeFunc::Parms == idx); 1556 } 1557 1558 void SafePointNode::disconnect_from_root(PhaseIterGVN *igvn) { 1559 assert(Opcode() == Op_SafePoint, "only value for safepoint in loops"); 1560 int nb = igvn->C->root()->find_prec_edge(this); 1561 if (nb != -1) { 1562 igvn->C->root()->rm_prec(nb); 1563 } 1564 } 1565 1566 //============== SafePointScalarObjectNode ============== 1567 1568 SafePointScalarObjectNode::SafePointScalarObjectNode(const TypeOopPtr* tp, 1569 #ifdef ASSERT 1570 Node* alloc, 1571 #endif 1572 uint first_index, 1573 uint n_fields) : 1574 TypeNode(tp, 1), // 1 control input -- seems required. Get from root. 1575 _first_index(first_index), 1576 _n_fields(n_fields) 1577 #ifdef ASSERT 1578 , _alloc(alloc) 1579 #endif 1580 { 1581 #ifdef ASSERT 1582 if (!alloc->is_Allocate() 1583 && !(alloc->Opcode() == Op_VectorBox)) { 1584 alloc->dump(); 1585 assert(false, "unexpected call node"); 1586 } 1587 #endif 1588 init_class_id(Class_SafePointScalarObject); 1589 } 1590 1591 // Do not allow value-numbering for SafePointScalarObject node. 1592 uint SafePointScalarObjectNode::hash() const { return NO_HASH; } 1593 bool SafePointScalarObjectNode::cmp( const Node &n ) const { 1594 return (&n == this); // Always fail except on self 1595 } 1596 1597 uint SafePointScalarObjectNode::ideal_reg() const { 1598 return 0; // No matching to machine instruction 1599 } 1600 1601 const RegMask &SafePointScalarObjectNode::in_RegMask(uint idx) const { 1602 return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]); 1603 } 1604 1605 const RegMask &SafePointScalarObjectNode::out_RegMask() const { 1606 return RegMask::Empty; 1607 } 1608 1609 uint SafePointScalarObjectNode::match_edge(uint idx) const { 1610 return 0; 1611 } 1612 1613 SafePointScalarObjectNode* 1614 SafePointScalarObjectNode::clone(Dict* sosn_map, bool& new_node) const { 1615 void* cached = (*sosn_map)[(void*)this]; 1616 if (cached != NULL) { 1617 new_node = false; 1618 return (SafePointScalarObjectNode*)cached; 1619 } 1620 new_node = true; 1621 SafePointScalarObjectNode* res = (SafePointScalarObjectNode*)Node::clone(); 1622 sosn_map->Insert((void*)this, (void*)res); 1623 return res; 1624 } 1625 1626 1627 #ifndef PRODUCT 1628 void SafePointScalarObjectNode::dump_spec(outputStream *st) const { 1629 st->print(" # fields@[%d..%d]", first_index(), 1630 first_index() + n_fields() - 1); 1631 } 1632 1633 #endif 1634 1635 //============================================================================= 1636 uint AllocateNode::size_of() const { return sizeof(*this); } 1637 1638 AllocateNode::AllocateNode(Compile* C, const TypeFunc *atype, 1639 Node *ctrl, Node *mem, Node *abio, 1640 Node *size, Node *klass_node, Node *initial_test) 1641 : CallNode(atype, NULL, TypeRawPtr::BOTTOM) 1642 { 1643 init_class_id(Class_Allocate); 1644 init_flags(Flag_is_macro); 1645 _is_scalar_replaceable = false; 1646 _is_non_escaping = false; 1647 _is_allocation_MemBar_redundant = false; 1648 Node *topnode = C->top(); 1649 1650 init_req( TypeFunc::Control , ctrl ); 1651 init_req( TypeFunc::I_O , abio ); 1652 init_req( TypeFunc::Memory , mem ); 1653 init_req( TypeFunc::ReturnAdr, topnode ); 1654 init_req( TypeFunc::FramePtr , topnode ); 1655 init_req( AllocSize , size); 1656 init_req( KlassNode , klass_node); 1657 init_req( InitialTest , initial_test); 1658 init_req( ALength , topnode); 1659 init_req( ValidLengthTest , topnode); 1660 C->add_macro_node(this); 1661 } 1662 1663 void AllocateNode::compute_MemBar_redundancy(ciMethod* initializer) 1664 { 1665 assert(initializer != NULL && 1666 initializer->is_initializer() && 1667 !initializer->is_static(), 1668 "unexpected initializer method"); 1669 BCEscapeAnalyzer* analyzer = initializer->get_bcea(); 1670 if (analyzer == NULL) { 1671 return; 1672 } 1673 1674 // Allocation node is first parameter in its initializer 1675 if (analyzer->is_arg_stack(0) || analyzer->is_arg_local(0)) { 1676 _is_allocation_MemBar_redundant = true; 1677 } 1678 } 1679 Node *AllocateNode::make_ideal_mark(PhaseGVN *phase, Node* obj, Node* control, Node* mem) { 1680 Node* mark_node = NULL; 1681 // For now only enable fast locking for non-array types 1682 mark_node = phase->MakeConX(markWord::prototype().value()); 1683 return mark_node; 1684 } 1685 1686 // Retrieve the length from the AllocateArrayNode. Narrow the type with a 1687 // CastII, if appropriate. If we are not allowed to create new nodes, and 1688 // a CastII is appropriate, return NULL. 1689 Node *AllocateArrayNode::make_ideal_length(const TypeOopPtr* oop_type, PhaseTransform *phase, bool allow_new_nodes) { 1690 Node *length = in(AllocateNode::ALength); 1691 assert(length != NULL, "length is not null"); 1692 1693 const TypeInt* length_type = phase->find_int_type(length); 1694 const TypeAryPtr* ary_type = oop_type->isa_aryptr(); 1695 1696 if (ary_type != NULL && length_type != NULL) { 1697 const TypeInt* narrow_length_type = ary_type->narrow_size_type(length_type); 1698 if (narrow_length_type != length_type) { 1699 // Assert one of: 1700 // - the narrow_length is 0 1701 // - the narrow_length is not wider than length 1702 assert(narrow_length_type == TypeInt::ZERO || 1703 length_type->is_con() && narrow_length_type->is_con() && 1704 (narrow_length_type->_hi <= length_type->_lo) || 1705 (narrow_length_type->_hi <= length_type->_hi && 1706 narrow_length_type->_lo >= length_type->_lo), 1707 "narrow type must be narrower than length type"); 1708 1709 // Return NULL if new nodes are not allowed 1710 if (!allow_new_nodes) return NULL; 1711 // Create a cast which is control dependent on the initialization to 1712 // propagate the fact that the array length must be positive. 1713 InitializeNode* init = initialization(); 1714 assert(init != NULL, "initialization not found"); 1715 length = new CastIINode(length, narrow_length_type); 1716 length->set_req(TypeFunc::Control, init->proj_out_or_null(TypeFunc::Control)); 1717 } 1718 } 1719 1720 return length; 1721 } 1722 1723 //============================================================================= 1724 uint LockNode::size_of() const { return sizeof(*this); } 1725 1726 // Redundant lock elimination 1727 // 1728 // There are various patterns of locking where we release and 1729 // immediately reacquire a lock in a piece of code where no operations 1730 // occur in between that would be observable. In those cases we can 1731 // skip releasing and reacquiring the lock without violating any 1732 // fairness requirements. Doing this around a loop could cause a lock 1733 // to be held for a very long time so we concentrate on non-looping 1734 // control flow. We also require that the operations are fully 1735 // redundant meaning that we don't introduce new lock operations on 1736 // some paths so to be able to eliminate it on others ala PRE. This 1737 // would probably require some more extensive graph manipulation to 1738 // guarantee that the memory edges were all handled correctly. 1739 // 1740 // Assuming p is a simple predicate which can't trap in any way and s 1741 // is a synchronized method consider this code: 1742 // 1743 // s(); 1744 // if (p) 1745 // s(); 1746 // else 1747 // s(); 1748 // s(); 1749 // 1750 // 1. The unlocks of the first call to s can be eliminated if the 1751 // locks inside the then and else branches are eliminated. 1752 // 1753 // 2. The unlocks of the then and else branches can be eliminated if 1754 // the lock of the final call to s is eliminated. 1755 // 1756 // Either of these cases subsumes the simple case of sequential control flow 1757 // 1758 // Additionally we can eliminate versions without the else case: 1759 // 1760 // s(); 1761 // if (p) 1762 // s(); 1763 // s(); 1764 // 1765 // 3. In this case we eliminate the unlock of the first s, the lock 1766 // and unlock in the then case and the lock in the final s. 1767 // 1768 // Note also that in all these cases the then/else pieces don't have 1769 // to be trivial as long as they begin and end with synchronization 1770 // operations. 1771 // 1772 // s(); 1773 // if (p) 1774 // s(); 1775 // f(); 1776 // s(); 1777 // s(); 1778 // 1779 // The code will work properly for this case, leaving in the unlock 1780 // before the call to f and the relock after it. 1781 // 1782 // A potentially interesting case which isn't handled here is when the 1783 // locking is partially redundant. 1784 // 1785 // s(); 1786 // if (p) 1787 // s(); 1788 // 1789 // This could be eliminated putting unlocking on the else case and 1790 // eliminating the first unlock and the lock in the then side. 1791 // Alternatively the unlock could be moved out of the then side so it 1792 // was after the merge and the first unlock and second lock 1793 // eliminated. This might require less manipulation of the memory 1794 // state to get correct. 1795 // 1796 // Additionally we might allow work between a unlock and lock before 1797 // giving up eliminating the locks. The current code disallows any 1798 // conditional control flow between these operations. A formulation 1799 // similar to partial redundancy elimination computing the 1800 // availability of unlocking and the anticipatability of locking at a 1801 // program point would allow detection of fully redundant locking with 1802 // some amount of work in between. I'm not sure how often I really 1803 // think that would occur though. Most of the cases I've seen 1804 // indicate it's likely non-trivial work would occur in between. 1805 // There may be other more complicated constructs where we could 1806 // eliminate locking but I haven't seen any others appear as hot or 1807 // interesting. 1808 // 1809 // Locking and unlocking have a canonical form in ideal that looks 1810 // roughly like this: 1811 // 1812 // <obj> 1813 // | \\------+ 1814 // | \ \ 1815 // | BoxLock \ 1816 // | | | \ 1817 // | | \ \ 1818 // | | FastLock 1819 // | | / 1820 // | | / 1821 // | | | 1822 // 1823 // Lock 1824 // | 1825 // Proj #0 1826 // | 1827 // MembarAcquire 1828 // | 1829 // Proj #0 1830 // 1831 // MembarRelease 1832 // | 1833 // Proj #0 1834 // | 1835 // Unlock 1836 // | 1837 // Proj #0 1838 // 1839 // 1840 // This code proceeds by processing Lock nodes during PhaseIterGVN 1841 // and searching back through its control for the proper code 1842 // patterns. Once it finds a set of lock and unlock operations to 1843 // eliminate they are marked as eliminatable which causes the 1844 // expansion of the Lock and Unlock macro nodes to make the operation a NOP 1845 // 1846 //============================================================================= 1847 1848 // 1849 // Utility function to skip over uninteresting control nodes. Nodes skipped are: 1850 // - copy regions. (These may not have been optimized away yet.) 1851 // - eliminated locking nodes 1852 // 1853 static Node *next_control(Node *ctrl) { 1854 if (ctrl == NULL) 1855 return NULL; 1856 while (1) { 1857 if (ctrl->is_Region()) { 1858 RegionNode *r = ctrl->as_Region(); 1859 Node *n = r->is_copy(); 1860 if (n == NULL) 1861 break; // hit a region, return it 1862 else 1863 ctrl = n; 1864 } else if (ctrl->is_Proj()) { 1865 Node *in0 = ctrl->in(0); 1866 if (in0->is_AbstractLock() && in0->as_AbstractLock()->is_eliminated()) { 1867 ctrl = in0->in(0); 1868 } else { 1869 break; 1870 } 1871 } else { 1872 break; // found an interesting control 1873 } 1874 } 1875 return ctrl; 1876 } 1877 // 1878 // Given a control, see if it's the control projection of an Unlock which 1879 // operating on the same object as lock. 1880 // 1881 bool AbstractLockNode::find_matching_unlock(const Node* ctrl, LockNode* lock, 1882 GrowableArray<AbstractLockNode*> &lock_ops) { 1883 ProjNode *ctrl_proj = (ctrl->is_Proj()) ? ctrl->as_Proj() : NULL; 1884 if (ctrl_proj != NULL && ctrl_proj->_con == TypeFunc::Control) { 1885 Node *n = ctrl_proj->in(0); 1886 if (n != NULL && n->is_Unlock()) { 1887 UnlockNode *unlock = n->as_Unlock(); 1888 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1889 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node()); 1890 Node* unlock_obj = bs->step_over_gc_barrier(unlock->obj_node()); 1891 if (lock_obj->eqv_uncast(unlock_obj) && 1892 BoxLockNode::same_slot(lock->box_node(), unlock->box_node()) && 1893 !unlock->is_eliminated()) { 1894 lock_ops.append(unlock); 1895 return true; 1896 } 1897 } 1898 } 1899 return false; 1900 } 1901 1902 // 1903 // Find the lock matching an unlock. Returns null if a safepoint 1904 // or complicated control is encountered first. 1905 LockNode *AbstractLockNode::find_matching_lock(UnlockNode* unlock) { 1906 LockNode *lock_result = NULL; 1907 // find the matching lock, or an intervening safepoint 1908 Node *ctrl = next_control(unlock->in(0)); 1909 while (1) { 1910 assert(ctrl != NULL, "invalid control graph"); 1911 assert(!ctrl->is_Start(), "missing lock for unlock"); 1912 if (ctrl->is_top()) break; // dead control path 1913 if (ctrl->is_Proj()) ctrl = ctrl->in(0); 1914 if (ctrl->is_SafePoint()) { 1915 break; // found a safepoint (may be the lock we are searching for) 1916 } else if (ctrl->is_Region()) { 1917 // Check for a simple diamond pattern. Punt on anything more complicated 1918 if (ctrl->req() == 3 && ctrl->in(1) != NULL && ctrl->in(2) != NULL) { 1919 Node *in1 = next_control(ctrl->in(1)); 1920 Node *in2 = next_control(ctrl->in(2)); 1921 if (((in1->is_IfTrue() && in2->is_IfFalse()) || 1922 (in2->is_IfTrue() && in1->is_IfFalse())) && (in1->in(0) == in2->in(0))) { 1923 ctrl = next_control(in1->in(0)->in(0)); 1924 } else { 1925 break; 1926 } 1927 } else { 1928 break; 1929 } 1930 } else { 1931 ctrl = next_control(ctrl->in(0)); // keep searching 1932 } 1933 } 1934 if (ctrl->is_Lock()) { 1935 LockNode *lock = ctrl->as_Lock(); 1936 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1937 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node()); 1938 Node* unlock_obj = bs->step_over_gc_barrier(unlock->obj_node()); 1939 if (lock_obj->eqv_uncast(unlock_obj) && 1940 BoxLockNode::same_slot(lock->box_node(), unlock->box_node())) { 1941 lock_result = lock; 1942 } 1943 } 1944 return lock_result; 1945 } 1946 1947 // This code corresponds to case 3 above. 1948 1949 bool AbstractLockNode::find_lock_and_unlock_through_if(Node* node, LockNode* lock, 1950 GrowableArray<AbstractLockNode*> &lock_ops) { 1951 Node* if_node = node->in(0); 1952 bool if_true = node->is_IfTrue(); 1953 1954 if (if_node->is_If() && if_node->outcnt() == 2 && (if_true || node->is_IfFalse())) { 1955 Node *lock_ctrl = next_control(if_node->in(0)); 1956 if (find_matching_unlock(lock_ctrl, lock, lock_ops)) { 1957 Node* lock1_node = NULL; 1958 ProjNode* proj = if_node->as_If()->proj_out(!if_true); 1959 if (if_true) { 1960 if (proj->is_IfFalse() && proj->outcnt() == 1) { 1961 lock1_node = proj->unique_out(); 1962 } 1963 } else { 1964 if (proj->is_IfTrue() && proj->outcnt() == 1) { 1965 lock1_node = proj->unique_out(); 1966 } 1967 } 1968 if (lock1_node != NULL && lock1_node->is_Lock()) { 1969 LockNode *lock1 = lock1_node->as_Lock(); 1970 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1971 Node* lock_obj = bs->step_over_gc_barrier(lock->obj_node()); 1972 Node* lock1_obj = bs->step_over_gc_barrier(lock1->obj_node()); 1973 if (lock_obj->eqv_uncast(lock1_obj) && 1974 BoxLockNode::same_slot(lock->box_node(), lock1->box_node()) && 1975 !lock1->is_eliminated()) { 1976 lock_ops.append(lock1); 1977 return true; 1978 } 1979 } 1980 } 1981 } 1982 1983 lock_ops.trunc_to(0); 1984 return false; 1985 } 1986 1987 bool AbstractLockNode::find_unlocks_for_region(const RegionNode* region, LockNode* lock, 1988 GrowableArray<AbstractLockNode*> &lock_ops) { 1989 // check each control merging at this point for a matching unlock. 1990 // in(0) should be self edge so skip it. 1991 for (int i = 1; i < (int)region->req(); i++) { 1992 Node *in_node = next_control(region->in(i)); 1993 if (in_node != NULL) { 1994 if (find_matching_unlock(in_node, lock, lock_ops)) { 1995 // found a match so keep on checking. 1996 continue; 1997 } else if (find_lock_and_unlock_through_if(in_node, lock, lock_ops)) { 1998 continue; 1999 } 2000 2001 // If we fall through to here then it was some kind of node we 2002 // don't understand or there wasn't a matching unlock, so give 2003 // up trying to merge locks. 2004 lock_ops.trunc_to(0); 2005 return false; 2006 } 2007 } 2008 return true; 2009 2010 } 2011 2012 const char* AbstractLockNode::_kind_names[] = {"Regular", "NonEscObj", "Coarsened", "Nested"}; 2013 2014 const char * AbstractLockNode::kind_as_string() const { 2015 return _kind_names[_kind]; 2016 } 2017 2018 #ifndef PRODUCT 2019 // 2020 // Create a counter which counts the number of times this lock is acquired 2021 // 2022 void AbstractLockNode::create_lock_counter(JVMState* state) { 2023 _counter = OptoRuntime::new_named_counter(state, NamedCounter::LockCounter); 2024 } 2025 2026 void AbstractLockNode::set_eliminated_lock_counter() { 2027 if (_counter) { 2028 // Update the counter to indicate that this lock was eliminated. 2029 // The counter update code will stay around even though the 2030 // optimizer will eliminate the lock operation itself. 2031 _counter->set_tag(NamedCounter::EliminatedLockCounter); 2032 } 2033 } 2034 2035 void AbstractLockNode::dump_spec(outputStream* st) const { 2036 st->print("%s ", _kind_names[_kind]); 2037 CallNode::dump_spec(st); 2038 } 2039 2040 void AbstractLockNode::dump_compact_spec(outputStream* st) const { 2041 st->print("%s", _kind_names[_kind]); 2042 } 2043 2044 // The related set of lock nodes includes the control boundary. 2045 void AbstractLockNode::related(GrowableArray<Node*> *in_rel, GrowableArray<Node*> *out_rel, bool compact) const { 2046 if (compact) { 2047 this->collect_nodes(in_rel, 1, false, false); 2048 } else { 2049 this->collect_nodes_in_all_data(in_rel, true); 2050 } 2051 this->collect_nodes(out_rel, -2, false, false); 2052 } 2053 #endif 2054 2055 //============================================================================= 2056 Node *LockNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2057 2058 // perform any generic optimizations first (returns 'this' or NULL) 2059 Node *result = SafePointNode::Ideal(phase, can_reshape); 2060 if (result != NULL) return result; 2061 // Don't bother trying to transform a dead node 2062 if (in(0) && in(0)->is_top()) return NULL; 2063 2064 // Now see if we can optimize away this lock. We don't actually 2065 // remove the locking here, we simply set the _eliminate flag which 2066 // prevents macro expansion from expanding the lock. Since we don't 2067 // modify the graph, the value returned from this function is the 2068 // one computed above. 2069 if (can_reshape && EliminateLocks && !is_non_esc_obj()) { 2070 // 2071 // If we are locking an non-escaped object, the lock/unlock is unnecessary 2072 // 2073 ConnectionGraph *cgr = phase->C->congraph(); 2074 if (cgr != NULL && cgr->not_global_escape(obj_node())) { 2075 assert(!is_eliminated() || is_coarsened(), "sanity"); 2076 // The lock could be marked eliminated by lock coarsening 2077 // code during first IGVN before EA. Replace coarsened flag 2078 // to eliminate all associated locks/unlocks. 2079 #ifdef ASSERT 2080 this->log_lock_optimization(phase->C,"eliminate_lock_set_non_esc1"); 2081 #endif 2082 this->set_non_esc_obj(); 2083 return result; 2084 } 2085 2086 if (!phase->C->do_locks_coarsening()) { 2087 return result; // Compiling without locks coarsening 2088 } 2089 // 2090 // Try lock coarsening 2091 // 2092 PhaseIterGVN* iter = phase->is_IterGVN(); 2093 if (iter != NULL && !is_eliminated()) { 2094 2095 GrowableArray<AbstractLockNode*> lock_ops; 2096 2097 Node *ctrl = next_control(in(0)); 2098 2099 // now search back for a matching Unlock 2100 if (find_matching_unlock(ctrl, this, lock_ops)) { 2101 // found an unlock directly preceding this lock. This is the 2102 // case of single unlock directly control dependent on a 2103 // single lock which is the trivial version of case 1 or 2. 2104 } else if (ctrl->is_Region() ) { 2105 if (find_unlocks_for_region(ctrl->as_Region(), this, lock_ops)) { 2106 // found lock preceded by multiple unlocks along all paths 2107 // joining at this point which is case 3 in description above. 2108 } 2109 } else { 2110 // see if this lock comes from either half of an if and the 2111 // predecessors merges unlocks and the other half of the if 2112 // performs a lock. 2113 if (find_lock_and_unlock_through_if(ctrl, this, lock_ops)) { 2114 // found unlock splitting to an if with locks on both branches. 2115 } 2116 } 2117 2118 if (lock_ops.length() > 0) { 2119 // add ourselves to the list of locks to be eliminated. 2120 lock_ops.append(this); 2121 2122 #ifndef PRODUCT 2123 if (PrintEliminateLocks) { 2124 int locks = 0; 2125 int unlocks = 0; 2126 if (Verbose) { 2127 tty->print_cr("=== Locks coarsening ==="); 2128 } 2129 for (int i = 0; i < lock_ops.length(); i++) { 2130 AbstractLockNode* lock = lock_ops.at(i); 2131 if (lock->Opcode() == Op_Lock) 2132 locks++; 2133 else 2134 unlocks++; 2135 if (Verbose) { 2136 tty->print(" %d: ", i); 2137 lock->dump(); 2138 } 2139 } 2140 tty->print_cr("=== Coarsened %d unlocks and %d locks", unlocks, locks); 2141 } 2142 #endif 2143 2144 // for each of the identified locks, mark them 2145 // as eliminatable 2146 for (int i = 0; i < lock_ops.length(); i++) { 2147 AbstractLockNode* lock = lock_ops.at(i); 2148 2149 // Mark it eliminated by coarsening and update any counters 2150 #ifdef ASSERT 2151 lock->log_lock_optimization(phase->C, "eliminate_lock_set_coarsened"); 2152 #endif 2153 lock->set_coarsened(); 2154 } 2155 // Record this coarsened group. 2156 phase->C->add_coarsened_locks(lock_ops); 2157 } else if (ctrl->is_Region() && 2158 iter->_worklist.member(ctrl)) { 2159 // We weren't able to find any opportunities but the region this 2160 // lock is control dependent on hasn't been processed yet so put 2161 // this lock back on the worklist so we can check again once any 2162 // region simplification has occurred. 2163 iter->_worklist.push(this); 2164 } 2165 } 2166 } 2167 2168 return result; 2169 } 2170 2171 //============================================================================= 2172 bool LockNode::is_nested_lock_region() { 2173 return is_nested_lock_region(NULL); 2174 } 2175 2176 // p is used for access to compilation log; no logging if NULL 2177 bool LockNode::is_nested_lock_region(Compile * c) { 2178 BoxLockNode* box = box_node()->as_BoxLock(); 2179 int stk_slot = box->stack_slot(); 2180 if (stk_slot <= 0) { 2181 #ifdef ASSERT 2182 this->log_lock_optimization(c, "eliminate_lock_INLR_1"); 2183 #endif 2184 return false; // External lock or it is not Box (Phi node). 2185 } 2186 2187 // Ignore complex cases: merged locks or multiple locks. 2188 Node* obj = obj_node(); 2189 LockNode* unique_lock = NULL; 2190 Node* bad_lock = NULL; 2191 if (!box->is_simple_lock_region(&unique_lock, obj, &bad_lock)) { 2192 #ifdef ASSERT 2193 this->log_lock_optimization(c, "eliminate_lock_INLR_2a", bad_lock); 2194 #endif 2195 return false; 2196 } 2197 if (unique_lock != this) { 2198 #ifdef ASSERT 2199 this->log_lock_optimization(c, "eliminate_lock_INLR_2b", (unique_lock != NULL ? unique_lock : bad_lock)); 2200 if (PrintEliminateLocks && Verbose) { 2201 tty->print_cr("=============== unique_lock != this ============"); 2202 tty->print(" this: "); 2203 this->dump(); 2204 tty->print(" box: "); 2205 box->dump(); 2206 tty->print(" obj: "); 2207 obj->dump(); 2208 if (unique_lock != NULL) { 2209 tty->print(" unique_lock: "); 2210 unique_lock->dump(); 2211 } 2212 if (bad_lock != NULL) { 2213 tty->print(" bad_lock: "); 2214 bad_lock->dump(); 2215 } 2216 tty->print_cr("==============="); 2217 } 2218 #endif 2219 return false; 2220 } 2221 2222 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 2223 obj = bs->step_over_gc_barrier(obj); 2224 // Look for external lock for the same object. 2225 SafePointNode* sfn = this->as_SafePoint(); 2226 JVMState* youngest_jvms = sfn->jvms(); 2227 int max_depth = youngest_jvms->depth(); 2228 for (int depth = 1; depth <= max_depth; depth++) { 2229 JVMState* jvms = youngest_jvms->of_depth(depth); 2230 int num_mon = jvms->nof_monitors(); 2231 // Loop over monitors 2232 for (int idx = 0; idx < num_mon; idx++) { 2233 Node* obj_node = sfn->monitor_obj(jvms, idx); 2234 obj_node = bs->step_over_gc_barrier(obj_node); 2235 BoxLockNode* box_node = sfn->monitor_box(jvms, idx)->as_BoxLock(); 2236 if ((box_node->stack_slot() < stk_slot) && obj_node->eqv_uncast(obj)) { 2237 return true; 2238 } 2239 } 2240 } 2241 #ifdef ASSERT 2242 this->log_lock_optimization(c, "eliminate_lock_INLR_3"); 2243 #endif 2244 return false; 2245 } 2246 2247 //============================================================================= 2248 uint UnlockNode::size_of() const { return sizeof(*this); } 2249 2250 //============================================================================= 2251 Node *UnlockNode::Ideal(PhaseGVN *phase, bool can_reshape) { 2252 2253 // perform any generic optimizations first (returns 'this' or NULL) 2254 Node *result = SafePointNode::Ideal(phase, can_reshape); 2255 if (result != NULL) return result; 2256 // Don't bother trying to transform a dead node 2257 if (in(0) && in(0)->is_top()) return NULL; 2258 2259 // Now see if we can optimize away this unlock. We don't actually 2260 // remove the unlocking here, we simply set the _eliminate flag which 2261 // prevents macro expansion from expanding the unlock. Since we don't 2262 // modify the graph, the value returned from this function is the 2263 // one computed above. 2264 // Escape state is defined after Parse phase. 2265 if (can_reshape && EliminateLocks && !is_non_esc_obj()) { 2266 // 2267 // If we are unlocking an non-escaped object, the lock/unlock is unnecessary. 2268 // 2269 ConnectionGraph *cgr = phase->C->congraph(); 2270 if (cgr != NULL && cgr->not_global_escape(obj_node())) { 2271 assert(!is_eliminated() || is_coarsened(), "sanity"); 2272 // The lock could be marked eliminated by lock coarsening 2273 // code during first IGVN before EA. Replace coarsened flag 2274 // to eliminate all associated locks/unlocks. 2275 #ifdef ASSERT 2276 this->log_lock_optimization(phase->C, "eliminate_lock_set_non_esc2"); 2277 #endif 2278 this->set_non_esc_obj(); 2279 } 2280 } 2281 return result; 2282 } 2283 2284 void AbstractLockNode::log_lock_optimization(Compile *C, const char * tag, Node* bad_lock) const { 2285 if (C == NULL) { 2286 return; 2287 } 2288 CompileLog* log = C->log(); 2289 if (log != NULL) { 2290 Node* box = box_node(); 2291 Node* obj = obj_node(); 2292 int box_id = box != NULL ? box->_idx : -1; 2293 int obj_id = obj != NULL ? obj->_idx : -1; 2294 2295 log->begin_head("%s compile_id='%d' lock_id='%d' class='%s' kind='%s' box_id='%d' obj_id='%d' bad_id='%d'", 2296 tag, C->compile_id(), this->_idx, 2297 is_Unlock() ? "unlock" : is_Lock() ? "lock" : "?", 2298 kind_as_string(), box_id, obj_id, (bad_lock != NULL ? bad_lock->_idx : -1)); 2299 log->stamp(); 2300 log->end_head(); 2301 JVMState* p = is_Unlock() ? (as_Unlock()->dbg_jvms()) : jvms(); 2302 while (p != NULL) { 2303 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); 2304 p = p->caller(); 2305 } 2306 log->tail(tag); 2307 } 2308 } 2309 2310 bool CallNode::may_modify_arraycopy_helper(const TypeOopPtr* dest_t, const TypeOopPtr *t_oop, PhaseTransform *phase) { 2311 if (dest_t->is_known_instance() && t_oop->is_known_instance()) { 2312 return dest_t->instance_id() == t_oop->instance_id(); 2313 } 2314 2315 if (dest_t->isa_instptr() && !dest_t->klass()->equals(phase->C->env()->Object_klass())) { 2316 // clone 2317 if (t_oop->isa_aryptr()) { 2318 return false; 2319 } 2320 if (!t_oop->isa_instptr()) { 2321 return true; 2322 } 2323 if (dest_t->klass()->is_subtype_of(t_oop->klass()) || t_oop->klass()->is_subtype_of(dest_t->klass())) { 2324 return true; 2325 } 2326 // unrelated 2327 return false; 2328 } 2329 2330 if (dest_t->isa_aryptr()) { 2331 // arraycopy or array clone 2332 if (t_oop->isa_instptr()) { 2333 return false; 2334 } 2335 if (!t_oop->isa_aryptr()) { 2336 return true; 2337 } 2338 2339 const Type* elem = dest_t->is_aryptr()->elem(); 2340 if (elem == Type::BOTTOM) { 2341 // An array but we don't know what elements are 2342 return true; 2343 } 2344 2345 dest_t = dest_t->add_offset(Type::OffsetBot)->is_oopptr(); 2346 uint dest_alias = phase->C->get_alias_index(dest_t); 2347 uint t_oop_alias = phase->C->get_alias_index(t_oop); 2348 2349 return dest_alias == t_oop_alias; 2350 } 2351 2352 return true; 2353 }