1 /*
   2  * Copyright 1997-2006 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   debug_only(_bci = -99);  // random garbage value
 227   debug_only(_map = (SafePointNode*)-1);
 228   _caller = caller;
 229   _depth  = 1 + (caller == NULL ? 0 : caller->depth());
 230   _locoff = TypeFunc::Parms;
 231   _stkoff = _locoff + _method->max_locals();
 232   _monoff = _stkoff + _method->max_stack();
 233   _scloff = _monoff;
 234   _endoff = _monoff;
 235   _sp = 0;
 236 }
 237 JVMState::JVMState(int stack_size) {
 238   _method = NULL;
 239   _bci = InvocationEntryBci;
 240   debug_only(_map = (SafePointNode*)-1);
 241   _caller = NULL;
 242   _depth  = 1;
 243   _locoff = TypeFunc::Parms;
 244   _stkoff = _locoff;
 245   _monoff = _stkoff + stack_size;
 246   _scloff = _monoff;
 247   _endoff = _monoff;
 248   _sp = 0;
 249 }
 250 
 251 //--------------------------------of_depth-------------------------------------
 252 JVMState* JVMState::of_depth(int d) const {
 253   const JVMState* jvmp = this;
 254   assert(0 < d && (uint)d <= depth(), "oob");
 255   for (int skip = depth() - d; skip > 0; skip--) {
 256     jvmp = jvmp->caller();
 257   }
 258   assert(jvmp->depth() == (uint)d, "found the right one");
 259   return (JVMState*)jvmp;
 260 }
 261 
 262 //-----------------------------same_calls_as-----------------------------------
 263 bool JVMState::same_calls_as(const JVMState* that) const {
 264   if (this == that)                    return true;
 265   if (this->depth() != that->depth())  return false;
 266   const JVMState* p = this;
 267   const JVMState* q = that;
 268   for (;;) {
 269     if (p->_method != q->_method)    return false;
 270     if (p->_method == NULL)          return true;   // bci is irrelevant
 271     if (p->_bci    != q->_bci)       return false;
 272     p = p->caller();
 273     q = q->caller();
 274     if (p == q)                      return true;
 275     assert(p != NULL && q != NULL, "depth check ensures we don't run off end");
 276   }
 277 }
 278 
 279 //------------------------------debug_start------------------------------------
 280 uint JVMState::debug_start()  const {
 281   debug_only(JVMState* jvmroot = of_depth(1));
 282   assert(jvmroot->locoff() <= this->locoff(), "youngest JVMState must be last");
 283   return of_depth(1)->locoff();
 284 }
 285 
 286 //-------------------------------debug_end-------------------------------------
 287 uint JVMState::debug_end() const {
 288   debug_only(JVMState* jvmroot = of_depth(1));
 289   assert(jvmroot->endoff() <= this->endoff(), "youngest JVMState must be last");
 290   return endoff();
 291 }
 292 
 293 //------------------------------debug_depth------------------------------------
 294 uint JVMState::debug_depth() const {
 295   uint total = 0;
 296   for (const JVMState* jvmp = this; jvmp != NULL; jvmp = jvmp->caller()) {
 297     total += jvmp->debug_size();
 298   }
 299   return total;
 300 }
 301 
 302 #ifndef PRODUCT
 303 
 304 //------------------------------format_helper----------------------------------
 305 // Given an allocation (a Chaitin object) and a Node decide if the Node carries
 306 // any defined value or not.  If it does, print out the register or constant.
 307 static void format_helper( PhaseRegAlloc *regalloc, outputStream* st, Node *n, const char *msg, uint i, GrowableArray<SafePointScalarObjectNode*> *scobjs ) {
 308   if (n == NULL) { st->print(" NULL"); return; }
 309   if (n->is_SafePointScalarObject()) {
 310     // Scalar replacement.
 311     SafePointScalarObjectNode* spobj = n->as_SafePointScalarObject();
 312     scobjs->append_if_missing(spobj);
 313     int sco_n = scobjs->find(spobj);
 314     assert(sco_n >= 0, "");
 315     st->print(" %s%d]=#ScObj" INT32_FORMAT, msg, i, sco_n);
 316     return;
 317   }
 318   if( OptoReg::is_valid(regalloc->get_reg_first(n))) { // Check for undefined
 319     char buf[50];
 320     regalloc->dump_register(n,buf);
 321     st->print(" %s%d]=%s",msg,i,buf);
 322   } else {                      // No register, but might be constant
 323     const Type *t = n->bottom_type();
 324     switch (t->base()) {
 325     case Type::Int:
 326       st->print(" %s%d]=#"INT32_FORMAT,msg,i,t->is_int()->get_con());
 327       break;
 328     case Type::AnyPtr:
 329       assert( t == TypePtr::NULL_PTR, "" );
 330       st->print(" %s%d]=#NULL",msg,i);
 331       break;
 332     case Type::AryPtr:
 333     case Type::KlassPtr:
 334     case Type::InstPtr:
 335       st->print(" %s%d]=#Ptr" INTPTR_FORMAT,msg,i,t->isa_oopptr()->const_oop());
 336       break;
 337     case Type::RawPtr:
 338       st->print(" %s%d]=#Raw" INTPTR_FORMAT,msg,i,t->is_rawptr());
 339       break;
 340     case Type::DoubleCon:
 341       st->print(" %s%d]=#%fD",msg,i,t->is_double_constant()->_d);
 342       break;
 343     case Type::FloatCon:
 344       st->print(" %s%d]=#%fF",msg,i,t->is_float_constant()->_f);
 345       break;
 346     case Type::Long:
 347       st->print(" %s%d]=#"INT64_FORMAT,msg,i,t->is_long()->get_con());
 348       break;
 349     case Type::Half:
 350     case Type::Top:
 351       st->print(" %s%d]=_",msg,i);
 352       break;
 353     default: ShouldNotReachHere();
 354     }
 355   }
 356 }
 357 
 358 //------------------------------format-----------------------------------------
 359 void JVMState::format(PhaseRegAlloc *regalloc, const Node *n, outputStream* st) const {
 360   st->print("        #");
 361   if( _method ) {
 362     _method->print_short_name(st);
 363     st->print(" @ bci:%d ",_bci);
 364   } else {
 365     st->print_cr(" runtime stub ");
 366     return;
 367   }
 368   if (n->is_MachSafePoint()) {
 369     GrowableArray<SafePointScalarObjectNode*> scobjs;
 370     MachSafePointNode *mcall = n->as_MachSafePoint();
 371     uint i;
 372     // Print locals
 373     for( i = 0; i < (uint)loc_size(); i++ )
 374       format_helper( regalloc, st, mcall->local(this, i), "L[", i, &scobjs );
 375     // Print stack
 376     for (i = 0; i < (uint)stk_size(); i++) {
 377       if ((uint)(_stkoff + i) >= mcall->len())
 378         st->print(" oob ");
 379       else
 380        format_helper( regalloc, st, mcall->stack(this, i), "STK[", i, &scobjs );
 381     }
 382     for (i = 0; (int)i < nof_monitors(); i++) {
 383       Node *box = mcall->monitor_box(this, i);
 384       Node *obj = mcall->monitor_obj(this, i);
 385       if ( OptoReg::is_valid(regalloc->get_reg_first(box)) ) {
 386         while( !box->is_BoxLock() )  box = box->in(1);
 387         format_helper( regalloc, st, box, "MON-BOX[", i, &scobjs );
 388       } else {
 389         OptoReg::Name box_reg = BoxLockNode::stack_slot(box);
 390         st->print(" MON-BOX%d=%s+%d",
 391                    i,
 392                    OptoReg::regname(OptoReg::c_frame_pointer),
 393                    regalloc->reg2offset(box_reg));
 394       }
 395       format_helper( regalloc, st, obj, "MON-OBJ[", i, &scobjs );
 396     }
 397 
 398     for (i = 0; i < (uint)scobjs.length(); i++) {
 399       // Scalar replaced objects.
 400       st->print_cr("");
 401       st->print("        # ScObj" INT32_FORMAT " ", i);
 402       SafePointScalarObjectNode* spobj = scobjs.at(i);
 403       ciKlass* cik = spobj->bottom_type()->is_oopptr()->klass();
 404       assert(cik->is_instance_klass() ||
 405              cik->is_array_klass(), "Not supported allocation.");
 406       ciInstanceKlass *iklass = NULL;
 407       if (cik->is_instance_klass()) {
 408         cik->print_name_on(st);
 409         iklass = cik->as_instance_klass();
 410       } else if (cik->is_type_array_klass()) {
 411         cik->as_array_klass()->base_element_type()->print_name_on(st);
 412         st->print("[%d]=", spobj->n_fields());
 413       } else if (cik->is_obj_array_klass()) {
 414         ciType* cie = cik->as_array_klass()->base_element_type();
 415         int ndim = 1;
 416         while (cie->is_obj_array_klass()) {
 417           ndim += 1;
 418           cie = cie->as_array_klass()->base_element_type();
 419         }
 420         cie->print_name_on(st);
 421         while (ndim-- > 0) {
 422           st->print("[]");
 423         }
 424         st->print("[%d]=", spobj->n_fields());
 425       }
 426       st->print("{");
 427       uint nf = spobj->n_fields();
 428       if (nf > 0) {
 429         uint first_ind = spobj->first_index();
 430         Node* fld_node = mcall->in(first_ind);
 431         ciField* cifield;
 432         if (iklass != NULL) {
 433           st->print(" [");
 434           cifield = iklass->nonstatic_field_at(0);
 435           cifield->print_name_on(st);
 436           format_helper( regalloc, st, fld_node, ":", 0, &scobjs );
 437         } else {
 438           format_helper( regalloc, st, fld_node, "[", 0, &scobjs );
 439         }
 440         for (uint j = 1; j < nf; j++) {
 441           fld_node = mcall->in(first_ind+j);
 442           if (iklass != NULL) {
 443             st->print(", [");
 444             cifield = iklass->nonstatic_field_at(j);
 445             cifield->print_name_on(st);
 446             format_helper( regalloc, st, fld_node, ":", j, &scobjs );
 447           } else {
 448             format_helper( regalloc, st, fld_node, ", [", j, &scobjs );
 449           }
 450         }
 451       }
 452       st->print(" }");
 453     }
 454   }
 455   st->print_cr("");
 456   if (caller() != NULL)  caller()->format(regalloc, n, st);
 457 }
 458 
 459 
 460 void JVMState::dump_spec(outputStream *st) const {
 461   if (_method != NULL) {
 462     bool printed = false;
 463     if (!Verbose) {
 464       // The JVMS dumps make really, really long lines.
 465       // Take out the most boring parts, which are the package prefixes.
 466       char buf[500];
 467       stringStream namest(buf, sizeof(buf));
 468       _method->print_short_name(&namest);
 469       if (namest.count() < sizeof(buf)) {
 470         const char* name = namest.base();
 471         if (name[0] == ' ')  ++name;
 472         const char* endcn = strchr(name, ':');  // end of class name
 473         if (endcn == NULL)  endcn = strchr(name, '(');
 474         if (endcn == NULL)  endcn = name + strlen(name);
 475         while (endcn > name && endcn[-1] != '.' && endcn[-1] != '/')
 476           --endcn;
 477         st->print(" %s", endcn);
 478         printed = true;
 479       }
 480     }
 481     if (!printed)
 482       _method->print_short_name(st);
 483     st->print(" @ bci:%d",_bci);
 484   } else {
 485     st->print(" runtime stub");
 486   }
 487   if (caller() != NULL)  caller()->dump_spec(st);
 488 }
 489 
 490 
 491 void JVMState::dump_on(outputStream* st) const {
 492   if (_map && !((uintptr_t)_map & 1)) {
 493     if (_map->len() > _map->req()) {  // _map->has_exceptions()
 494       Node* ex = _map->in(_map->req());  // _map->next_exception()
 495       // skip the first one; it's already being printed
 496       while (ex != NULL && ex->len() > ex->req()) {
 497         ex = ex->in(ex->req());  // ex->next_exception()
 498         ex->dump(1);
 499       }
 500     }
 501     _map->dump(2);
 502   }
 503   st->print("JVMS depth=%d loc=%d stk=%d mon=%d scalar=%d end=%d mondepth=%d sp=%d bci=%d method=",
 504              depth(), locoff(), stkoff(), monoff(), scloff(), endoff(), monitor_depth(), sp(), bci());
 505   if (_method == NULL) {
 506     st->print_cr("(none)");
 507   } else {
 508     _method->print_name(st);
 509     st->cr();
 510     if (bci() >= 0 && bci() < _method->code_size()) {
 511       st->print("    bc: ");
 512       _method->print_codes_on(bci(), bci()+1, st);
 513     }
 514   }
 515   if (caller() != NULL) {
 516     caller()->dump_on(st);
 517   }
 518 }
 519 
 520 // Extra way to dump a jvms from the debugger,
 521 // to avoid a bug with C++ member function calls.
 522 void dump_jvms(JVMState* jvms) {
 523   jvms->dump();
 524 }
 525 #endif
 526 
 527 //--------------------------clone_shallow--------------------------------------
 528 JVMState* JVMState::clone_shallow(Compile* C) const {
 529   JVMState* n = has_method() ? new (C) JVMState(_method, _caller) : new (C) JVMState(0);
 530   n->set_bci(_bci);
 531   n->set_locoff(_locoff);
 532   n->set_stkoff(_stkoff);
 533   n->set_monoff(_monoff);
 534   n->set_scloff(_scloff);
 535   n->set_endoff(_endoff);
 536   n->set_sp(_sp);
 537   n->set_map(_map);
 538   return n;
 539 }
 540 
 541 //---------------------------clone_deep----------------------------------------
 542 JVMState* JVMState::clone_deep(Compile* C) const {
 543   JVMState* n = clone_shallow(C);
 544   for (JVMState* p = n; p->_caller != NULL; p = p->_caller) {
 545     p->_caller = p->_caller->clone_shallow(C);
 546   }
 547   assert(n->depth() == depth(), "sanity");
 548   assert(n->debug_depth() == debug_depth(), "sanity");
 549   return n;
 550 }
 551 
 552 //=============================================================================
 553 uint CallNode::cmp( const Node &n ) const
 554 { return _tf == ((CallNode&)n)._tf && _jvms == ((CallNode&)n)._jvms; }
 555 #ifndef PRODUCT
 556 void CallNode::dump_req() const {
 557   // Dump the required inputs, enclosed in '(' and ')'
 558   uint i;                       // Exit value of loop
 559   for( i=0; i<req(); i++ ) {    // For all required inputs
 560     if( i == TypeFunc::Parms ) tty->print("(");
 561     if( in(i) ) tty->print("%c%d ", Compile::current()->node_arena()->contains(in(i)) ? ' ' : 'o', in(i)->_idx);
 562     else tty->print("_ ");
 563   }
 564   tty->print(")");
 565 }
 566 
 567 void CallNode::dump_spec(outputStream *st) const {
 568   st->print(" ");
 569   tf()->dump_on(st);
 570   if (_cnt != COUNT_UNKNOWN)  st->print(" C=%f",_cnt);
 571   if (jvms() != NULL)  jvms()->dump_spec(st);
 572 }
 573 #endif
 574 
 575 const Type *CallNode::bottom_type() const { return tf()->range(); }
 576 const Type *CallNode::Value(PhaseTransform *phase) const {
 577   if (phase->type(in(0)) == Type::TOP)  return Type::TOP;
 578   return tf()->range();
 579 }
 580 
 581 //------------------------------calling_convention-----------------------------
 582 void CallNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const {
 583   // Use the standard compiler calling convention
 584   Matcher::calling_convention( sig_bt, parm_regs, argcnt, true );
 585 }
 586 
 587 
 588 //------------------------------match------------------------------------------
 589 // Construct projections for control, I/O, memory-fields, ..., and
 590 // return result(s) along with their RegMask info
 591 Node *CallNode::match( const ProjNode *proj, const Matcher *match ) {
 592   switch (proj->_con) {
 593   case TypeFunc::Control:
 594   case TypeFunc::I_O:
 595   case TypeFunc::Memory:
 596     return new (match->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
 597 
 598   case TypeFunc::Parms+1:       // For LONG & DOUBLE returns
 599     assert(tf()->_range->field_at(TypeFunc::Parms+1) == Type::HALF, "");
 600     // 2nd half of doubles and longs
 601     return new (match->C, 1) MachProjNode(this,proj->_con, RegMask::Empty, (uint)OptoReg::Bad);
 602 
 603   case TypeFunc::Parms: {       // Normal returns
 604     uint ideal_reg = Matcher::base2reg[tf()->range()->field_at(TypeFunc::Parms)->base()];
 605     OptoRegPair regs = is_CallRuntime()
 606       ? match->c_return_value(ideal_reg,true)  // Calls into C runtime
 607       : match->  return_value(ideal_reg,true); // Calls into compiled Java code
 608     RegMask rm = RegMask(regs.first());
 609     if( OptoReg::is_valid(regs.second()) )
 610       rm.Insert( regs.second() );
 611     return new (match->C, 1) MachProjNode(this,proj->_con,rm,ideal_reg);
 612   }
 613 
 614   case TypeFunc::ReturnAdr:
 615   case TypeFunc::FramePtr:
 616   default:
 617     ShouldNotReachHere();
 618   }
 619   return NULL;
 620 }
 621 
 622 // Do we Match on this edge index or not?  Match no edges
 623 uint CallNode::match_edge(uint idx) const {
 624   return 0;
 625 }
 626 
 627 //
 628 // Determine whether the call could modify the field of the specified
 629 // instance at the specified offset.
 630 //
 631 bool CallNode::may_modify(const TypePtr *addr_t, PhaseTransform *phase) {
 632   const TypeOopPtr *adrInst_t  = addr_t->isa_oopptr();
 633 
 634   // if not an InstPtr or not an instance type, assume the worst
 635   if (adrInst_t == NULL || !adrInst_t->is_instance_field()) {
 636     return true;
 637   }
 638   Compile *C = phase->C;
 639   int offset = adrInst_t->offset();
 640   assert(offset >= 0, "should be valid offset");
 641   ciKlass* adr_k = adrInst_t->klass();
 642   assert(adr_k->is_loaded() &&
 643          adr_k->is_java_klass() &&
 644          !adr_k->is_interface(),
 645          "only non-abstract classes are expected");
 646 
 647   int base_idx = C->get_alias_index(adrInst_t);
 648   int size = BytesPerLong; // If we don't know the size, assume largest.
 649   if (adrInst_t->isa_instptr()) {
 650     ciField* field = C->alias_type(base_idx)->field();
 651     if (field != NULL) {
 652       size = field->size_in_bytes();
 653     }
 654   } else {
 655     assert(adrInst_t->isa_aryptr(), "only arrays are expected");
 656     size = type2aelembytes(adr_k->as_array_klass()->element_type()->basic_type());
 657   }
 658 
 659   ciMethod * meth = is_CallStaticJava() ?  as_CallStaticJava()->method() : NULL;
 660   BCEscapeAnalyzer *bcea = (meth != NULL) ? meth->get_bcea() : NULL;
 661 
 662   const TypeTuple * d = tf()->domain();
 663   for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
 664     const Type* t = d->field_at(i);
 665     Node *arg = in(i);
 666     const Type *at = phase->type(arg);
 667     if (at == TypePtr::NULL_PTR || at == Type::TOP)
 668       continue;  // null can't affect anything
 669 
 670     const TypeOopPtr *at_ptr = at->isa_oopptr();
 671     if (!arg->is_top() && (t->isa_oopptr() != NULL ||
 672                            t->isa_ptr() && at_ptr != NULL)) {
 673       assert(at_ptr != NULL, "expecting an OopPtr");
 674       ciKlass* at_k = at_ptr->klass();
 675       if ((adrInst_t->base() == at_ptr->base()) &&
 676           at_k->is_loaded() &&
 677           at_k->is_java_klass() &&
 678           !at_k->is_interface()) {
 679         // If we have found an argument matching addr_t, check if the field
 680         // at the specified offset is modified.
 681         int at_idx = C->get_alias_index(at_ptr->add_offset(offset)->isa_oopptr());
 682         if (base_idx == at_idx &&
 683             (bcea == NULL ||
 684              bcea->is_arg_modified(i - TypeFunc::Parms, offset, size))) {
 685           return true;
 686         }
 687       }
 688     }
 689   }
 690   return false;
 691 }
 692 
 693 // Does this call have a direct reference to n other than debug information?
 694 bool CallNode::has_non_debug_use(Node *n) {
 695   const TypeTuple * d = tf()->domain();
 696   for (uint i = TypeFunc::Parms; i < d->cnt(); i++) {
 697     Node *arg = in(i);
 698     if (arg == n) {
 699       return true;
 700     }
 701   }
 702   return false;
 703 }
 704 
 705 // Returns the unique CheckCastPP of a call
 706 // or 'this' if there are several CheckCastPP
 707 // or returns NULL if there is no one.
 708 Node *CallNode::result_cast() {
 709   Node *cast = NULL;
 710 
 711   Node *p = proj_out(TypeFunc::Parms);
 712   if (p == NULL)
 713     return NULL;
 714 
 715   for (DUIterator_Fast imax, i = p->fast_outs(imax); i < imax; i++) {
 716     Node *use = p->fast_out(i);
 717     if (use->is_CheckCastPP()) {
 718       if (cast != NULL) {
 719         return this;  // more than 1 CheckCastPP
 720       }
 721       cast = use;
 722     }
 723   }
 724   return cast;
 725 }
 726 
 727 
 728 //=============================================================================
 729 uint CallJavaNode::size_of() const { return sizeof(*this); }
 730 uint CallJavaNode::cmp( const Node &n ) const {
 731   CallJavaNode &call = (CallJavaNode&)n;
 732   return CallNode::cmp(call) && _method == call._method;
 733 }
 734 #ifndef PRODUCT
 735 void CallJavaNode::dump_spec(outputStream *st) const {
 736   if( _method ) _method->print_short_name(st);
 737   CallNode::dump_spec(st);
 738 }
 739 #endif
 740 
 741 //=============================================================================
 742 uint CallStaticJavaNode::size_of() const { return sizeof(*this); }
 743 uint CallStaticJavaNode::cmp( const Node &n ) const {
 744   CallStaticJavaNode &call = (CallStaticJavaNode&)n;
 745   return CallJavaNode::cmp(call);
 746 }
 747 
 748 //----------------------------uncommon_trap_request----------------------------
 749 // If this is an uncommon trap, return the request code, else zero.
 750 int CallStaticJavaNode::uncommon_trap_request() const {
 751   if (_name != NULL && !strcmp(_name, "uncommon_trap")) {
 752     return extract_uncommon_trap_request(this);
 753   }
 754   return 0;
 755 }
 756 int CallStaticJavaNode::extract_uncommon_trap_request(const Node* call) {
 757 #ifndef PRODUCT
 758   if (!(call->req() > TypeFunc::Parms &&
 759         call->in(TypeFunc::Parms) != NULL &&
 760         call->in(TypeFunc::Parms)->is_Con())) {
 761     assert(_in_dump_cnt != 0, "OK if dumping");
 762     tty->print("[bad uncommon trap]");
 763     return 0;
 764   }
 765 #endif
 766   return call->in(TypeFunc::Parms)->bottom_type()->is_int()->get_con();
 767 }
 768 
 769 #ifndef PRODUCT
 770 void CallStaticJavaNode::dump_spec(outputStream *st) const {
 771   st->print("# Static ");
 772   if (_name != NULL) {
 773     st->print("%s", _name);
 774     int trap_req = uncommon_trap_request();
 775     if (trap_req != 0) {
 776       char buf[100];
 777       st->print("(%s)",
 778                  Deoptimization::format_trap_request(buf, sizeof(buf),
 779                                                      trap_req));
 780     }
 781     st->print(" ");
 782   }
 783   CallJavaNode::dump_spec(st);
 784 }
 785 #endif
 786 
 787 //=============================================================================
 788 uint CallDynamicJavaNode::size_of() const { return sizeof(*this); }
 789 uint CallDynamicJavaNode::cmp( const Node &n ) const {
 790   CallDynamicJavaNode &call = (CallDynamicJavaNode&)n;
 791   return CallJavaNode::cmp(call);
 792 }
 793 #ifndef PRODUCT
 794 void CallDynamicJavaNode::dump_spec(outputStream *st) const {
 795   st->print("# Dynamic ");
 796   CallJavaNode::dump_spec(st);
 797 }
 798 #endif
 799 
 800 //=============================================================================
 801 uint CallRuntimeNode::size_of() const { return sizeof(*this); }
 802 uint CallRuntimeNode::cmp( const Node &n ) const {
 803   CallRuntimeNode &call = (CallRuntimeNode&)n;
 804   return CallNode::cmp(call) && !strcmp(_name,call._name);
 805 }
 806 #ifndef PRODUCT
 807 void CallRuntimeNode::dump_spec(outputStream *st) const {
 808   st->print("# ");
 809   st->print(_name);
 810   CallNode::dump_spec(st);
 811 }
 812 #endif
 813 
 814 //------------------------------calling_convention-----------------------------
 815 void CallRuntimeNode::calling_convention( BasicType* sig_bt, VMRegPair *parm_regs, uint argcnt ) const {
 816   Matcher::c_calling_convention( sig_bt, parm_regs, argcnt );
 817 }
 818 
 819 //=============================================================================
 820 //------------------------------calling_convention-----------------------------
 821 
 822 
 823 //=============================================================================
 824 #ifndef PRODUCT
 825 void CallLeafNode::dump_spec(outputStream *st) const {
 826   st->print("# ");
 827   st->print(_name);
 828   CallNode::dump_spec(st);
 829 }
 830 #endif
 831 
 832 //=============================================================================
 833 
 834 void SafePointNode::set_local(JVMState* jvms, uint idx, Node *c) {
 835   assert(verify_jvms(jvms), "jvms must match");
 836   int loc = jvms->locoff() + idx;
 837   if (in(loc)->is_top() && idx > 0 && !c->is_top() ) {
 838     // If current local idx is top then local idx - 1 could
 839     // be a long/double that needs to be killed since top could
 840     // represent the 2nd half ofthe long/double.
 841     uint ideal = in(loc -1)->ideal_reg();
 842     if (ideal == Op_RegD || ideal == Op_RegL) {
 843       // set other (low index) half to top
 844       set_req(loc - 1, in(loc));
 845     }
 846   }
 847   set_req(loc, c);
 848 }
 849 
 850 uint SafePointNode::size_of() const { return sizeof(*this); }
 851 uint SafePointNode::cmp( const Node &n ) const {
 852   return (&n == this);          // Always fail except on self
 853 }
 854 
 855 //-------------------------set_next_exception----------------------------------
 856 void SafePointNode::set_next_exception(SafePointNode* n) {
 857   assert(n == NULL || n->Opcode() == Op_SafePoint, "correct value for next_exception");
 858   if (len() == req()) {
 859     if (n != NULL)  add_prec(n);
 860   } else {
 861     set_prec(req(), n);
 862   }
 863 }
 864 
 865 
 866 //----------------------------next_exception-----------------------------------
 867 SafePointNode* SafePointNode::next_exception() const {
 868   if (len() == req()) {
 869     return NULL;
 870   } else {
 871     Node* n = in(req());
 872     assert(n == NULL || n->Opcode() == Op_SafePoint, "no other uses of prec edges");
 873     return (SafePointNode*) n;
 874   }
 875 }
 876 
 877 
 878 //------------------------------Ideal------------------------------------------
 879 // Skip over any collapsed Regions
 880 Node *SafePointNode::Ideal(PhaseGVN *phase, bool can_reshape) {
 881   if (remove_dead_region(phase, can_reshape))  return this;
 882 
 883   return NULL;
 884 }
 885 
 886 //------------------------------Identity---------------------------------------
 887 // Remove obviously duplicate safepoints
 888 Node *SafePointNode::Identity( PhaseTransform *phase ) {
 889 
 890   // If you have back to back safepoints, remove one
 891   if( in(TypeFunc::Control)->is_SafePoint() )
 892     return in(TypeFunc::Control);
 893 
 894   if( in(0)->is_Proj() ) {
 895     Node *n0 = in(0)->in(0);
 896     // Check if he is a call projection (except Leaf Call)
 897     if( n0->is_Catch() ) {
 898       n0 = n0->in(0)->in(0);
 899       assert( n0->is_Call(), "expect a call here" );
 900     }
 901     if( n0->is_Call() && n0->as_Call()->guaranteed_safepoint() ) {
 902       // Useless Safepoint, so remove it
 903       return in(TypeFunc::Control);
 904     }
 905   }
 906 
 907   return this;
 908 }
 909 
 910 //------------------------------Value------------------------------------------
 911 const Type *SafePointNode::Value( PhaseTransform *phase ) const {
 912   if( phase->type(in(0)) == Type::TOP ) return Type::TOP;
 913   if( phase->eqv( in(0), this ) ) return Type::TOP; // Dead infinite loop
 914   return Type::CONTROL;
 915 }
 916 
 917 #ifndef PRODUCT
 918 void SafePointNode::dump_spec(outputStream *st) const {
 919   st->print(" SafePoint ");
 920 }
 921 #endif
 922 
 923 const RegMask &SafePointNode::in_RegMask(uint idx) const {
 924   if( idx < TypeFunc::Parms ) return RegMask::Empty;
 925   // Values outside the domain represent debug info
 926   return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
 927 }
 928 const RegMask &SafePointNode::out_RegMask() const {
 929   return RegMask::Empty;
 930 }
 931 
 932 
 933 void SafePointNode::grow_stack(JVMState* jvms, uint grow_by) {
 934   assert((int)grow_by > 0, "sanity");
 935   int monoff = jvms->monoff();
 936   int scloff = jvms->scloff();
 937   int endoff = jvms->endoff();
 938   assert(endoff == (int)req(), "no other states or debug info after me");
 939   Node* top = Compile::current()->top();
 940   for (uint i = 0; i < grow_by; i++) {
 941     ins_req(monoff, top);
 942   }
 943   jvms->set_monoff(monoff + grow_by);
 944   jvms->set_scloff(scloff + grow_by);
 945   jvms->set_endoff(endoff + grow_by);
 946 }
 947 
 948 void SafePointNode::push_monitor(const FastLockNode *lock) {
 949   // Add a LockNode, which points to both the original BoxLockNode (the
 950   // stack space for the monitor) and the Object being locked.
 951   const int MonitorEdges = 2;
 952   assert(JVMState::logMonitorEdges == exact_log2(MonitorEdges), "correct MonitorEdges");
 953   assert(req() == jvms()->endoff(), "correct sizing");
 954   int nextmon = jvms()->scloff();
 955   if (GenerateSynchronizationCode) {
 956     add_req(lock->box_node());
 957     add_req(lock->obj_node());
 958   } else {
 959     add_req(NULL);
 960     add_req(NULL);
 961   }
 962   jvms()->set_scloff(nextmon+MonitorEdges);
 963   jvms()->set_endoff(req());
 964 }
 965 
 966 void SafePointNode::pop_monitor() {
 967   // Delete last monitor from debug info
 968   debug_only(int num_before_pop = jvms()->nof_monitors());
 969   const int MonitorEdges = (1<<JVMState::logMonitorEdges);
 970   int scloff = jvms()->scloff();
 971   int endoff = jvms()->endoff();
 972   int new_scloff = scloff - MonitorEdges;
 973   int new_endoff = endoff - MonitorEdges;
 974   jvms()->set_scloff(new_scloff);
 975   jvms()->set_endoff(new_endoff);
 976   while (scloff > new_scloff)  del_req(--scloff);
 977   assert(jvms()->nof_monitors() == num_before_pop-1, "");
 978 }
 979 
 980 Node *SafePointNode::peek_monitor_box() const {
 981   int mon = jvms()->nof_monitors() - 1;
 982   assert(mon >= 0, "most have a monitor");
 983   return monitor_box(jvms(), mon);
 984 }
 985 
 986 Node *SafePointNode::peek_monitor_obj() const {
 987   int mon = jvms()->nof_monitors() - 1;
 988   assert(mon >= 0, "most have a monitor");
 989   return monitor_obj(jvms(), mon);
 990 }
 991 
 992 // Do we Match on this edge index or not?  Match no edges
 993 uint SafePointNode::match_edge(uint idx) const {
 994   if( !needs_polling_address_input() )
 995     return 0;
 996 
 997   return (TypeFunc::Parms == idx);
 998 }
 999 
1000 //==============  SafePointScalarObjectNode  ==============
1001 
1002 SafePointScalarObjectNode::SafePointScalarObjectNode(const TypeOopPtr* tp,
1003 #ifdef ASSERT
1004                                                      AllocateNode* alloc,
1005 #endif
1006                                                      uint first_index,
1007                                                      uint n_fields) :
1008   TypeNode(tp, 1), // 1 control input -- seems required.  Get from root.
1009 #ifdef ASSERT
1010   _alloc(alloc),
1011 #endif
1012   _first_index(first_index),
1013   _n_fields(n_fields)
1014 {
1015   init_class_id(Class_SafePointScalarObject);
1016 }
1017 
1018 
1019 uint SafePointScalarObjectNode::ideal_reg() const {
1020   return 0; // No matching to machine instruction
1021 }
1022 
1023 const RegMask &SafePointScalarObjectNode::in_RegMask(uint idx) const {
1024   return *(Compile::current()->matcher()->idealreg2debugmask[in(idx)->ideal_reg()]);
1025 }
1026 
1027 const RegMask &SafePointScalarObjectNode::out_RegMask() const {
1028   return RegMask::Empty;
1029 }
1030 
1031 uint SafePointScalarObjectNode::match_edge(uint idx) const {
1032   return 0;
1033 }
1034 
1035 SafePointScalarObjectNode*
1036 SafePointScalarObjectNode::clone(int jvms_adj, Dict* sosn_map) const {
1037   void* cached = (*sosn_map)[(void*)this];
1038   if (cached != NULL) {
1039     return (SafePointScalarObjectNode*)cached;
1040   }
1041   Compile* C = Compile::current();
1042   SafePointScalarObjectNode* res = (SafePointScalarObjectNode*)Node::clone();
1043   res->_first_index += jvms_adj;
1044   sosn_map->Insert((void*)this, (void*)res);
1045   return res;
1046 }
1047 
1048 
1049 #ifndef PRODUCT
1050 void SafePointScalarObjectNode::dump_spec(outputStream *st) const {
1051   st->print(" # fields@[%d..%d]", first_index(),
1052              first_index() + n_fields() - 1);
1053 }
1054 
1055 #endif
1056 
1057 //=============================================================================
1058 uint AllocateNode::size_of() const { return sizeof(*this); }
1059 
1060 AllocateNode::AllocateNode(Compile* C, const TypeFunc *atype,
1061                            Node *ctrl, Node *mem, Node *abio,
1062                            Node *size, Node *klass_node, Node *initial_test)
1063   : CallNode(atype, NULL, TypeRawPtr::BOTTOM)
1064 {
1065   init_class_id(Class_Allocate);
1066   init_flags(Flag_is_macro);
1067   _is_scalar_replaceable = false;
1068   Node *topnode = C->top();
1069 
1070   init_req( TypeFunc::Control  , ctrl );
1071   init_req( TypeFunc::I_O      , abio );
1072   init_req( TypeFunc::Memory   , mem );
1073   init_req( TypeFunc::ReturnAdr, topnode );
1074   init_req( TypeFunc::FramePtr , topnode );
1075   init_req( AllocSize          , size);
1076   init_req( KlassNode          , klass_node);
1077   init_req( InitialTest        , initial_test);
1078   init_req( ALength            , topnode);
1079   C->add_macro_node(this);
1080 }
1081 
1082 //=============================================================================
1083 uint AllocateArrayNode::size_of() const { return sizeof(*this); }
1084 
1085 //=============================================================================
1086 uint LockNode::size_of() const { return sizeof(*this); }
1087 
1088 // Redundant lock elimination
1089 //
1090 // There are various patterns of locking where we release and
1091 // immediately reacquire a lock in a piece of code where no operations
1092 // occur in between that would be observable.  In those cases we can
1093 // skip releasing and reacquiring the lock without violating any
1094 // fairness requirements.  Doing this around a loop could cause a lock
1095 // to be held for a very long time so we concentrate on non-looping
1096 // control flow.  We also require that the operations are fully
1097 // redundant meaning that we don't introduce new lock operations on
1098 // some paths so to be able to eliminate it on others ala PRE.  This
1099 // would probably require some more extensive graph manipulation to
1100 // guarantee that the memory edges were all handled correctly.
1101 //
1102 // Assuming p is a simple predicate which can't trap in any way and s
1103 // is a synchronized method consider this code:
1104 //
1105 //   s();
1106 //   if (p)
1107 //     s();
1108 //   else
1109 //     s();
1110 //   s();
1111 //
1112 // 1. The unlocks of the first call to s can be eliminated if the
1113 // locks inside the then and else branches are eliminated.
1114 //
1115 // 2. The unlocks of the then and else branches can be eliminated if
1116 // the lock of the final call to s is eliminated.
1117 //
1118 // Either of these cases subsumes the simple case of sequential control flow
1119 //
1120 // Addtionally we can eliminate versions without the else case:
1121 //
1122 //   s();
1123 //   if (p)
1124 //     s();
1125 //   s();
1126 //
1127 // 3. In this case we eliminate the unlock of the first s, the lock
1128 // and unlock in the then case and the lock in the final s.
1129 //
1130 // Note also that in all these cases the then/else pieces don't have
1131 // to be trivial as long as they begin and end with synchronization
1132 // operations.
1133 //
1134 //   s();
1135 //   if (p)
1136 //     s();
1137 //     f();
1138 //     s();
1139 //   s();
1140 //
1141 // The code will work properly for this case, leaving in the unlock
1142 // before the call to f and the relock after it.
1143 //
1144 // A potentially interesting case which isn't handled here is when the
1145 // locking is partially redundant.
1146 //
1147 //   s();
1148 //   if (p)
1149 //     s();
1150 //
1151 // This could be eliminated putting unlocking on the else case and
1152 // eliminating the first unlock and the lock in the then side.
1153 // Alternatively the unlock could be moved out of the then side so it
1154 // was after the merge and the first unlock and second lock
1155 // eliminated.  This might require less manipulation of the memory
1156 // state to get correct.
1157 //
1158 // Additionally we might allow work between a unlock and lock before
1159 // giving up eliminating the locks.  The current code disallows any
1160 // conditional control flow between these operations.  A formulation
1161 // similar to partial redundancy elimination computing the
1162 // availability of unlocking and the anticipatability of locking at a
1163 // program point would allow detection of fully redundant locking with
1164 // some amount of work in between.  I'm not sure how often I really
1165 // think that would occur though.  Most of the cases I've seen
1166 // indicate it's likely non-trivial work would occur in between.
1167 // There may be other more complicated constructs where we could
1168 // eliminate locking but I haven't seen any others appear as hot or
1169 // interesting.
1170 //
1171 // Locking and unlocking have a canonical form in ideal that looks
1172 // roughly like this:
1173 //
1174 //              <obj>
1175 //                | \\------+
1176 //                |  \       \
1177 //                | BoxLock   \
1178 //                |  |   |     \
1179 //                |  |    \     \
1180 //                |  |   FastLock
1181 //                |  |   /
1182 //                |  |  /
1183 //                |  |  |
1184 //
1185 //               Lock
1186 //                |
1187 //            Proj #0
1188 //                |
1189 //            MembarAcquire
1190 //                |
1191 //            Proj #0
1192 //
1193 //            MembarRelease
1194 //                |
1195 //            Proj #0
1196 //                |
1197 //              Unlock
1198 //                |
1199 //            Proj #0
1200 //
1201 //
1202 // This code proceeds by processing Lock nodes during PhaseIterGVN
1203 // and searching back through its control for the proper code
1204 // patterns.  Once it finds a set of lock and unlock operations to
1205 // eliminate they are marked as eliminatable which causes the
1206 // expansion of the Lock and Unlock macro nodes to make the operation a NOP
1207 //
1208 //=============================================================================
1209 
1210 //
1211 // Utility function to skip over uninteresting control nodes.  Nodes skipped are:
1212 //   - copy regions.  (These may not have been optimized away yet.)
1213 //   - eliminated locking nodes
1214 //
1215 static Node *next_control(Node *ctrl) {
1216   if (ctrl == NULL)
1217     return NULL;
1218   while (1) {
1219     if (ctrl->is_Region()) {
1220       RegionNode *r = ctrl->as_Region();
1221       Node *n = r->is_copy();
1222       if (n == NULL)
1223         break;  // hit a region, return it
1224       else
1225         ctrl = n;
1226     } else if (ctrl->is_Proj()) {
1227       Node *in0 = ctrl->in(0);
1228       if (in0->is_AbstractLock() && in0->as_AbstractLock()->is_eliminated()) {
1229         ctrl = in0->in(0);
1230       } else {
1231         break;
1232       }
1233     } else {
1234       break; // found an interesting control
1235     }
1236   }
1237   return ctrl;
1238 }
1239 //
1240 // Given a control, see if it's the control projection of an Unlock which
1241 // operating on the same object as lock.
1242 //
1243 bool AbstractLockNode::find_matching_unlock(const Node* ctrl, LockNode* lock,
1244                                             GrowableArray<AbstractLockNode*> &lock_ops) {
1245   ProjNode *ctrl_proj = (ctrl->is_Proj()) ? ctrl->as_Proj() : NULL;
1246   if (ctrl_proj != NULL && ctrl_proj->_con == TypeFunc::Control) {
1247     Node *n = ctrl_proj->in(0);
1248     if (n != NULL && n->is_Unlock()) {
1249       UnlockNode *unlock = n->as_Unlock();
1250       if ((lock->obj_node() == unlock->obj_node()) &&
1251           (lock->box_node() == unlock->box_node()) && !unlock->is_eliminated()) {
1252         lock_ops.append(unlock);
1253         return true;
1254       }
1255     }
1256   }
1257   return false;
1258 }
1259 
1260 //
1261 // Find the lock matching an unlock.  Returns null if a safepoint
1262 // or complicated control is encountered first.
1263 LockNode *AbstractLockNode::find_matching_lock(UnlockNode* unlock) {
1264   LockNode *lock_result = NULL;
1265   // find the matching lock, or an intervening safepoint
1266   Node *ctrl = next_control(unlock->in(0));
1267   while (1) {
1268     assert(ctrl != NULL, "invalid control graph");
1269     assert(!ctrl->is_Start(), "missing lock for unlock");
1270     if (ctrl->is_top()) break;  // dead control path
1271     if (ctrl->is_Proj()) ctrl = ctrl->in(0);
1272     if (ctrl->is_SafePoint()) {
1273         break;  // found a safepoint (may be the lock we are searching for)
1274     } else if (ctrl->is_Region()) {
1275       // Check for a simple diamond pattern.  Punt on anything more complicated
1276       if (ctrl->req() == 3 && ctrl->in(1) != NULL && ctrl->in(2) != NULL) {
1277         Node *in1 = next_control(ctrl->in(1));
1278         Node *in2 = next_control(ctrl->in(2));
1279         if (((in1->is_IfTrue() && in2->is_IfFalse()) ||
1280              (in2->is_IfTrue() && in1->is_IfFalse())) && (in1->in(0) == in2->in(0))) {
1281           ctrl = next_control(in1->in(0)->in(0));
1282         } else {
1283           break;
1284         }
1285       } else {
1286         break;
1287       }
1288     } else {
1289       ctrl = next_control(ctrl->in(0));  // keep searching
1290     }
1291   }
1292   if (ctrl->is_Lock()) {
1293     LockNode *lock = ctrl->as_Lock();
1294     if ((lock->obj_node() == unlock->obj_node()) &&
1295             (lock->box_node() == unlock->box_node())) {
1296       lock_result = lock;
1297     }
1298   }
1299   return lock_result;
1300 }
1301 
1302 // This code corresponds to case 3 above.
1303 
1304 bool AbstractLockNode::find_lock_and_unlock_through_if(Node* node, LockNode* lock,
1305                                                        GrowableArray<AbstractLockNode*> &lock_ops) {
1306   Node* if_node = node->in(0);
1307   bool  if_true = node->is_IfTrue();
1308 
1309   if (if_node->is_If() && if_node->outcnt() == 2 && (if_true || node->is_IfFalse())) {
1310     Node *lock_ctrl = next_control(if_node->in(0));
1311     if (find_matching_unlock(lock_ctrl, lock, lock_ops)) {
1312       Node* lock1_node = NULL;
1313       ProjNode* proj = if_node->as_If()->proj_out(!if_true);
1314       if (if_true) {
1315         if (proj->is_IfFalse() && proj->outcnt() == 1) {
1316           lock1_node = proj->unique_out();
1317         }
1318       } else {
1319         if (proj->is_IfTrue() && proj->outcnt() == 1) {
1320           lock1_node = proj->unique_out();
1321         }
1322       }
1323       if (lock1_node != NULL && lock1_node->is_Lock()) {
1324         LockNode *lock1 = lock1_node->as_Lock();
1325         if ((lock->obj_node() == lock1->obj_node()) &&
1326             (lock->box_node() == lock1->box_node()) && !lock1->is_eliminated()) {
1327           lock_ops.append(lock1);
1328           return true;
1329         }
1330       }
1331     }
1332   }
1333 
1334   lock_ops.trunc_to(0);
1335   return false;
1336 }
1337 
1338 bool AbstractLockNode::find_unlocks_for_region(const RegionNode* region, LockNode* lock,
1339                                GrowableArray<AbstractLockNode*> &lock_ops) {
1340   // check each control merging at this point for a matching unlock.
1341   // in(0) should be self edge so skip it.
1342   for (int i = 1; i < (int)region->req(); i++) {
1343     Node *in_node = next_control(region->in(i));
1344     if (in_node != NULL) {
1345       if (find_matching_unlock(in_node, lock, lock_ops)) {
1346         // found a match so keep on checking.
1347         continue;
1348       } else if (find_lock_and_unlock_through_if(in_node, lock, lock_ops)) {
1349         continue;
1350       }
1351 
1352       // If we fall through to here then it was some kind of node we
1353       // don't understand or there wasn't a matching unlock, so give
1354       // up trying to merge locks.
1355       lock_ops.trunc_to(0);
1356       return false;
1357     }
1358   }
1359   return true;
1360 
1361 }
1362 
1363 #ifndef PRODUCT
1364 //
1365 // Create a counter which counts the number of times this lock is acquired
1366 //
1367 void AbstractLockNode::create_lock_counter(JVMState* state) {
1368   _counter = OptoRuntime::new_named_counter(state, NamedCounter::LockCounter);
1369 }
1370 #endif
1371 
1372 void AbstractLockNode::set_eliminated() {
1373   _eliminate = true;
1374 #ifndef PRODUCT
1375   if (_counter) {
1376     // Update the counter to indicate that this lock was eliminated.
1377     // The counter update code will stay around even though the
1378     // optimizer will eliminate the lock operation itself.
1379     _counter->set_tag(NamedCounter::EliminatedLockCounter);
1380   }
1381 #endif
1382 }
1383 
1384 //=============================================================================
1385 Node *LockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1386 
1387   // perform any generic optimizations first (returns 'this' or NULL)
1388   Node *result = SafePointNode::Ideal(phase, can_reshape);
1389 
1390   // Now see if we can optimize away this lock.  We don't actually
1391   // remove the locking here, we simply set the _eliminate flag which
1392   // prevents macro expansion from expanding the lock.  Since we don't
1393   // modify the graph, the value returned from this function is the
1394   // one computed above.
1395   if (result == NULL && can_reshape && EliminateLocks && !is_eliminated()) {
1396     //
1397     // If we are locking an unescaped object, the lock/unlock is unnecessary
1398     //
1399     ConnectionGraph *cgr = Compile::current()->congraph();
1400     PointsToNode::EscapeState es = PointsToNode::GlobalEscape;
1401     if (cgr != NULL)
1402       es = cgr->escape_state(obj_node(), phase);
1403     if (es != PointsToNode::UnknownEscape && es != PointsToNode::GlobalEscape) {
1404       // Mark it eliminated to update any counters
1405       this->set_eliminated();
1406       return result;
1407     }
1408 
1409     //
1410     // Try lock coarsening
1411     //
1412     PhaseIterGVN* iter = phase->is_IterGVN();
1413     if (iter != NULL) {
1414 
1415       GrowableArray<AbstractLockNode*>   lock_ops;
1416 
1417       Node *ctrl = next_control(in(0));
1418 
1419       // now search back for a matching Unlock
1420       if (find_matching_unlock(ctrl, this, lock_ops)) {
1421         // found an unlock directly preceding this lock.  This is the
1422         // case of single unlock directly control dependent on a
1423         // single lock which is the trivial version of case 1 or 2.
1424       } else if (ctrl->is_Region() ) {
1425         if (find_unlocks_for_region(ctrl->as_Region(), this, lock_ops)) {
1426         // found lock preceded by multiple unlocks along all paths
1427         // joining at this point which is case 3 in description above.
1428         }
1429       } else {
1430         // see if this lock comes from either half of an if and the
1431         // predecessors merges unlocks and the other half of the if
1432         // performs a lock.
1433         if (find_lock_and_unlock_through_if(ctrl, this, lock_ops)) {
1434           // found unlock splitting to an if with locks on both branches.
1435         }
1436       }
1437 
1438       if (lock_ops.length() > 0) {
1439         // add ourselves to the list of locks to be eliminated.
1440         lock_ops.append(this);
1441 
1442   #ifndef PRODUCT
1443         if (PrintEliminateLocks) {
1444           int locks = 0;
1445           int unlocks = 0;
1446           for (int i = 0; i < lock_ops.length(); i++) {
1447             AbstractLockNode* lock = lock_ops.at(i);
1448             if (lock->Opcode() == Op_Lock)
1449               locks++;
1450             else
1451               unlocks++;
1452             if (Verbose) {
1453               lock->dump(1);
1454             }
1455           }
1456           tty->print_cr("***Eliminated %d unlocks and %d locks", unlocks, locks);
1457         }
1458   #endif
1459 
1460         // for each of the identified locks, mark them
1461         // as eliminatable
1462         for (int i = 0; i < lock_ops.length(); i++) {
1463           AbstractLockNode* lock = lock_ops.at(i);
1464 
1465           // Mark it eliminated to update any counters
1466           lock->set_eliminated();
1467         }
1468       } else if (result != NULL && ctrl->is_Region() &&
1469                  iter->_worklist.member(ctrl)) {
1470         // We weren't able to find any opportunities but the region this
1471         // lock is control dependent on hasn't been processed yet so put
1472         // this lock back on the worklist so we can check again once any
1473         // region simplification has occurred.
1474         iter->_worklist.push(this);
1475       }
1476     }
1477   }
1478 
1479   return result;
1480 }
1481 
1482 //=============================================================================
1483 uint UnlockNode::size_of() const { return sizeof(*this); }
1484 
1485 //=============================================================================
1486 Node *UnlockNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1487 
1488   // perform any generic optimizations first (returns 'this' or NULL)
1489   Node * result = SafePointNode::Ideal(phase, can_reshape);
1490 
1491   // Now see if we can optimize away this unlock.  We don't actually
1492   // remove the unlocking here, we simply set the _eliminate flag which
1493   // prevents macro expansion from expanding the unlock.  Since we don't
1494   // modify the graph, the value returned from this function is the
1495   // one computed above.
1496   // Escape state is defined after Parse phase.
1497   if (result == NULL && can_reshape && EliminateLocks && !is_eliminated()) {
1498     //
1499     // If we are unlocking an unescaped object, the lock/unlock is unnecessary.
1500     //
1501     ConnectionGraph *cgr = Compile::current()->congraph();
1502     PointsToNode::EscapeState es = PointsToNode::GlobalEscape;
1503     if (cgr != NULL)
1504       es = cgr->escape_state(obj_node(), phase);
1505     if (es != PointsToNode::UnknownEscape && es != PointsToNode::GlobalEscape) {
1506       // Mark it eliminated to update any counters
1507       this->set_eliminated();
1508     }
1509   }
1510   return result;
1511 }