1 /*
   2  * Copyright (c) 1997, 2012, 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 "classfile/systemDictionary.hpp"
  27 #include "memory/allocation.inline.hpp"
  28 #include "oops/objArrayKlass.hpp"
  29 #include "opto/addnode.hpp"
  30 #include "opto/cfgnode.hpp"
  31 #include "opto/connode.hpp"
  32 #include "opto/loopnode.hpp"
  33 #include "opto/machnode.hpp"
  34 #include "opto/mulnode.hpp"
  35 #include "opto/phaseX.hpp"
  36 #include "opto/regmask.hpp"
  37 #include "opto/runtime.hpp"
  38 #include "opto/subnode.hpp"
  39 
  40 // Portions of code courtesy of Clifford Click
  41 
  42 // Optimization - Graph Style
  43 
  44 //=============================================================================
  45 //------------------------------Value------------------------------------------
  46 // Compute the type of the RegionNode.
  47 const Type *RegionNode::Value( PhaseTransform *phase ) const {
  48   for( uint i=1; i<req(); ++i ) {       // For all paths in
  49     Node *n = in(i);            // Get Control source
  50     if( !n ) continue;          // Missing inputs are TOP
  51     if( phase->type(n) == Type::CONTROL )
  52       return Type::CONTROL;
  53   }
  54   return Type::TOP;             // All paths dead?  Then so are we
  55 }
  56 
  57 //------------------------------Identity---------------------------------------
  58 // Check for Region being Identity.
  59 Node *RegionNode::Identity( PhaseTransform *phase ) {
  60   // Cannot have Region be an identity, even if it has only 1 input.
  61   // Phi users cannot have their Region input folded away for them,
  62   // since they need to select the proper data input
  63   return this;
  64 }
  65 
  66 //------------------------------merge_region-----------------------------------
  67 // If a Region flows into a Region, merge into one big happy merge.  This is
  68 // hard to do if there is stuff that has to happen
  69 static Node *merge_region(RegionNode *region, PhaseGVN *phase) {
  70   if( region->Opcode() != Op_Region ) // Do not do to LoopNodes
  71     return NULL;
  72   Node *progress = NULL;        // Progress flag
  73   PhaseIterGVN *igvn = phase->is_IterGVN();
  74 
  75   uint rreq = region->req();
  76   for( uint i = 1; i < rreq; i++ ) {
  77     Node *r = region->in(i);
  78     if( r && r->Opcode() == Op_Region && // Found a region?
  79         r->in(0) == r &&        // Not already collapsed?
  80         r != region &&          // Avoid stupid situations
  81         r->outcnt() == 2 ) {    // Self user and 'region' user only?
  82       assert(!r->as_Region()->has_phi(), "no phi users");
  83       if( !progress ) {         // No progress
  84         if (region->has_phi()) {
  85           return NULL;        // Only flatten if no Phi users
  86           // igvn->hash_delete( phi );
  87         }
  88         igvn->hash_delete( region );
  89         progress = region;      // Making progress
  90       }
  91       igvn->hash_delete( r );
  92 
  93       // Append inputs to 'r' onto 'region'
  94       for( uint j = 1; j < r->req(); j++ ) {
  95         // Move an input from 'r' to 'region'
  96         region->add_req(r->in(j));
  97         r->set_req(j, phase->C->top());
  98         // Update phis of 'region'
  99         //for( uint k = 0; k < max; k++ ) {
 100         //  Node *phi = region->out(k);
 101         //  if( phi->is_Phi() ) {
 102         //    phi->add_req(phi->in(i));
 103         //  }
 104         //}
 105 
 106         rreq++;                 // One more input to Region
 107       } // Found a region to merge into Region
 108       // Clobber pointer to the now dead 'r'
 109       region->set_req(i, phase->C->top());
 110     }
 111   }
 112 
 113   return progress;
 114 }
 115 
 116 
 117 
 118 //--------------------------------has_phi--------------------------------------
 119 // Helper function: Return any PhiNode that uses this region or NULL
 120 PhiNode* RegionNode::has_phi() const {
 121   for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
 122     Node* phi = fast_out(i);
 123     if (phi->is_Phi()) {   // Check for Phi users
 124       assert(phi->in(0) == (Node*)this, "phi uses region only via in(0)");
 125       return phi->as_Phi();  // this one is good enough
 126     }
 127   }
 128 
 129   return NULL;
 130 }
 131 
 132 
 133 //-----------------------------has_unique_phi----------------------------------
 134 // Helper function: Return the only PhiNode that uses this region or NULL
 135 PhiNode* RegionNode::has_unique_phi() const {
 136   // Check that only one use is a Phi
 137   PhiNode* only_phi = NULL;
 138   for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
 139     Node* phi = fast_out(i);
 140     if (phi->is_Phi()) {   // Check for Phi users
 141       assert(phi->in(0) == (Node*)this, "phi uses region only via in(0)");
 142       if (only_phi == NULL) {
 143         only_phi = phi->as_Phi();
 144       } else {
 145         return NULL;  // multiple phis
 146       }
 147     }
 148   }
 149 
 150   return only_phi;
 151 }
 152 
 153 
 154 //------------------------------check_phi_clipping-----------------------------
 155 // Helper function for RegionNode's identification of FP clipping
 156 // Check inputs to the Phi
 157 static bool check_phi_clipping( PhiNode *phi, ConNode * &min, uint &min_idx, ConNode * &max, uint &max_idx, Node * &val, uint &val_idx ) {
 158   min     = NULL;
 159   max     = NULL;
 160   val     = NULL;
 161   min_idx = 0;
 162   max_idx = 0;
 163   val_idx = 0;
 164   uint  phi_max = phi->req();
 165   if( phi_max == 4 ) {
 166     for( uint j = 1; j < phi_max; ++j ) {
 167       Node *n = phi->in(j);
 168       int opcode = n->Opcode();
 169       switch( opcode ) {
 170       case Op_ConI:
 171         {
 172           if( min == NULL ) {
 173             min     = n->Opcode() == Op_ConI ? (ConNode*)n : NULL;
 174             min_idx = j;
 175           } else {
 176             max     = n->Opcode() == Op_ConI ? (ConNode*)n : NULL;
 177             max_idx = j;
 178             if( min->get_int() > max->get_int() ) {
 179               // Swap min and max
 180               ConNode *temp;
 181               uint     temp_idx;
 182               temp     = min;     min     = max;     max     = temp;
 183               temp_idx = min_idx; min_idx = max_idx; max_idx = temp_idx;
 184             }
 185           }
 186         }
 187         break;
 188       default:
 189         {
 190           val = n;
 191           val_idx = j;
 192         }
 193         break;
 194       }
 195     }
 196   }
 197   return ( min && max && val && (min->get_int() <= 0) && (max->get_int() >=0) );
 198 }
 199 
 200 
 201 //------------------------------check_if_clipping------------------------------
 202 // Helper function for RegionNode's identification of FP clipping
 203 // Check that inputs to Region come from two IfNodes,
 204 //
 205 //            If
 206 //      False    True
 207 //       If        |
 208 //  False  True    |
 209 //    |      |     |
 210 //  RegionNode_inputs
 211 //
 212 static bool check_if_clipping( const RegionNode *region, IfNode * &bot_if, IfNode * &top_if ) {
 213   top_if = NULL;
 214   bot_if = NULL;
 215 
 216   // Check control structure above RegionNode for (if  ( if  ) )
 217   Node *in1 = region->in(1);
 218   Node *in2 = region->in(2);
 219   Node *in3 = region->in(3);
 220   // Check that all inputs are projections
 221   if( in1->is_Proj() && in2->is_Proj() && in3->is_Proj() ) {
 222     Node *in10 = in1->in(0);
 223     Node *in20 = in2->in(0);
 224     Node *in30 = in3->in(0);
 225     // Check that #1 and #2 are ifTrue and ifFalse from same If
 226     if( in10 != NULL && in10->is_If() &&
 227         in20 != NULL && in20->is_If() &&
 228         in30 != NULL && in30->is_If() && in10 == in20 &&
 229         (in1->Opcode() != in2->Opcode()) ) {
 230       Node  *in100 = in10->in(0);
 231       Node *in1000 = (in100 != NULL && in100->is_Proj()) ? in100->in(0) : NULL;
 232       // Check that control for in10 comes from other branch of IF from in3
 233       if( in1000 != NULL && in1000->is_If() &&
 234           in30 == in1000 && (in3->Opcode() != in100->Opcode()) ) {
 235         // Control pattern checks
 236         top_if = (IfNode*)in1000;
 237         bot_if = (IfNode*)in10;
 238       }
 239     }
 240   }
 241 
 242   return (top_if != NULL);
 243 }
 244 
 245 
 246 //------------------------------check_convf2i_clipping-------------------------
 247 // Helper function for RegionNode's identification of FP clipping
 248 // Verify that the value input to the phi comes from "ConvF2I; LShift; RShift"
 249 static bool check_convf2i_clipping( PhiNode *phi, uint idx, ConvF2INode * &convf2i, Node *min, Node *max) {
 250   convf2i = NULL;
 251 
 252   // Check for the RShiftNode
 253   Node *rshift = phi->in(idx);
 254   assert( rshift, "Previous checks ensure phi input is present");
 255   if( rshift->Opcode() != Op_RShiftI )  { return false; }
 256 
 257   // Check for the LShiftNode
 258   Node *lshift = rshift->in(1);
 259   assert( lshift, "Previous checks ensure phi input is present");
 260   if( lshift->Opcode() != Op_LShiftI )  { return false; }
 261 
 262   // Check for the ConvF2INode
 263   Node *conv = lshift->in(1);
 264   if( conv->Opcode() != Op_ConvF2I ) { return false; }
 265 
 266   // Check that shift amounts are only to get sign bits set after F2I
 267   jint max_cutoff     = max->get_int();
 268   jint min_cutoff     = min->get_int();
 269   jint left_shift     = lshift->in(2)->get_int();
 270   jint right_shift    = rshift->in(2)->get_int();
 271   jint max_post_shift = nth_bit(BitsPerJavaInteger - left_shift - 1);
 272   if( left_shift != right_shift ||
 273       0 > left_shift || left_shift >= BitsPerJavaInteger ||
 274       max_post_shift < max_cutoff ||
 275       max_post_shift < -min_cutoff ) {
 276     // Shifts are necessary but current transformation eliminates them
 277     return false;
 278   }
 279 
 280   // OK to return the result of ConvF2I without shifting
 281   convf2i = (ConvF2INode*)conv;
 282   return true;
 283 }
 284 
 285 
 286 //------------------------------check_compare_clipping-------------------------
 287 // Helper function for RegionNode's identification of FP clipping
 288 static bool check_compare_clipping( bool less_than, IfNode *iff, ConNode *limit, Node * & input ) {
 289   Node *i1 = iff->in(1);
 290   if ( !i1->is_Bool() ) { return false; }
 291   BoolNode *bool1 = i1->as_Bool();
 292   if(       less_than && bool1->_test._test != BoolTest::le ) { return false; }
 293   else if( !less_than && bool1->_test._test != BoolTest::lt ) { return false; }
 294   const Node *cmpF = bool1->in(1);
 295   if( cmpF->Opcode() != Op_CmpF )      { return false; }
 296   // Test that the float value being compared against
 297   // is equivalent to the int value used as a limit
 298   Node *nodef = cmpF->in(2);
 299   if( nodef->Opcode() != Op_ConF ) { return false; }
 300   jfloat conf = nodef->getf();
 301   jint   coni = limit->get_int();
 302   if( ((int)conf) != coni )        { return false; }
 303   input = cmpF->in(1);
 304   return true;
 305 }
 306 
 307 //------------------------------is_unreachable_region--------------------------
 308 // Find if the Region node is reachable from the root.
 309 bool RegionNode::is_unreachable_region(PhaseGVN *phase) const {
 310   assert(req() == 2, "");
 311 
 312   // First, cut the simple case of fallthrough region when NONE of
 313   // region's phis references itself directly or through a data node.
 314   uint max = outcnt();
 315   uint i;
 316   for (i = 0; i < max; i++) {
 317     Node* phi = raw_out(i);
 318     if (phi != NULL && phi->is_Phi()) {
 319       assert(phase->eqv(phi->in(0), this) && phi->req() == 2, "");
 320       if (phi->outcnt() == 0)
 321         continue; // Safe case - no loops
 322       if (phi->outcnt() == 1) {
 323         Node* u = phi->raw_out(0);
 324         // Skip if only one use is an other Phi or Call or Uncommon trap.
 325         // It is safe to consider this case as fallthrough.
 326         if (u != NULL && (u->is_Phi() || u->is_CFG()))
 327           continue;
 328       }
 329       // Check when phi references itself directly or through an other node.
 330       if (phi->as_Phi()->simple_data_loop_check(phi->in(1)) >= PhiNode::Unsafe)
 331         break; // Found possible unsafe data loop.
 332     }
 333   }
 334   if (i >= max)
 335     return false; // An unsafe case was NOT found - don't need graph walk.
 336 
 337   // Unsafe case - check if the Region node is reachable from root.
 338   ResourceMark rm;
 339 
 340   Arena *a = Thread::current()->resource_area();
 341   Node_List nstack(a);
 342   VectorSet visited(a);
 343 
 344   // Mark all control nodes reachable from root outputs
 345   Node *n = (Node*)phase->C->root();
 346   nstack.push(n);
 347   visited.set(n->_idx);
 348   while (nstack.size() != 0) {
 349     n = nstack.pop();
 350     uint max = n->outcnt();
 351     for (uint i = 0; i < max; i++) {
 352       Node* m = n->raw_out(i);
 353       if (m != NULL && m->is_CFG()) {
 354         if (phase->eqv(m, this)) {
 355           return false; // We reached the Region node - it is not dead.
 356         }
 357         if (!visited.test_set(m->_idx))
 358           nstack.push(m);
 359       }
 360     }
 361   }
 362 
 363   return true; // The Region node is unreachable - it is dead.
 364 }
 365 
 366 //------------------------------Ideal------------------------------------------
 367 // Return a node which is more "ideal" than the current node.  Must preserve
 368 // the CFG, but we can still strip out dead paths.
 369 Node *RegionNode::Ideal(PhaseGVN *phase, bool can_reshape) {
 370   if( !can_reshape && !in(0) ) return NULL;     // Already degraded to a Copy
 371   assert(!in(0) || !in(0)->is_Root(), "not a specially hidden merge");
 372 
 373   // Check for RegionNode with no Phi users and both inputs come from either
 374   // arm of the same IF.  If found, then the control-flow split is useless.
 375   bool has_phis = false;
 376   if (can_reshape) {            // Need DU info to check for Phi users
 377     has_phis = (has_phi() != NULL);       // Cache result
 378     if (!has_phis) {            // No Phi users?  Nothing merging?
 379       for (uint i = 1; i < req()-1; i++) {
 380         Node *if1 = in(i);
 381         if( !if1 ) continue;
 382         Node *iff = if1->in(0);
 383         if( !iff || !iff->is_If() ) continue;
 384         for( uint j=i+1; j<req(); j++ ) {
 385           if( in(j) && in(j)->in(0) == iff &&
 386               if1->Opcode() != in(j)->Opcode() ) {
 387             // Add the IF Projections to the worklist. They (and the IF itself)
 388             // will be eliminated if dead.
 389             phase->is_IterGVN()->add_users_to_worklist(iff);
 390             set_req(i, iff->in(0));// Skip around the useless IF diamond
 391             set_req(j, NULL);
 392             return this;      // Record progress
 393           }
 394         }
 395       }
 396     }
 397   }
 398 
 399   // Remove TOP or NULL input paths. If only 1 input path remains, this Region
 400   // degrades to a copy.
 401   bool add_to_worklist = false;
 402   int cnt = 0;                  // Count of values merging
 403   DEBUG_ONLY( int cnt_orig = req(); ) // Save original inputs count
 404   int del_it = 0;               // The last input path we delete
 405   // For all inputs...
 406   for( uint i=1; i<req(); ++i ){// For all paths in
 407     Node *n = in(i);            // Get the input
 408     if( n != NULL ) {
 409       // Remove useless control copy inputs
 410       if( n->is_Region() && n->as_Region()->is_copy() ) {
 411         set_req(i, n->nonnull_req());
 412         i--;
 413         continue;
 414       }
 415       if( n->is_Proj() ) {      // Remove useless rethrows
 416         Node *call = n->in(0);
 417         if (call->is_Call() && call->as_Call()->entry_point() == OptoRuntime::rethrow_stub()) {
 418           set_req(i, call->in(0));
 419           i--;
 420           continue;
 421         }
 422       }
 423       if( phase->type(n) == Type::TOP ) {
 424         set_req(i, NULL);       // Ignore TOP inputs
 425         i--;
 426         continue;
 427       }
 428       cnt++;                    // One more value merging
 429 
 430     } else if (can_reshape) {   // Else found dead path with DU info
 431       PhaseIterGVN *igvn = phase->is_IterGVN();
 432       del_req(i);               // Yank path from self
 433       del_it = i;
 434       uint max = outcnt();
 435       DUIterator j;
 436       bool progress = true;
 437       while(progress) {         // Need to establish property over all users
 438         progress = false;
 439         for (j = outs(); has_out(j); j++) {
 440           Node *n = out(j);
 441           if( n->req() != req() && n->is_Phi() ) {
 442             assert( n->in(0) == this, "" );
 443             igvn->hash_delete(n); // Yank from hash before hacking edges
 444             n->set_req_X(i,NULL,igvn);// Correct DU info
 445             n->del_req(i);        // Yank path from Phis
 446             if( max != outcnt() ) {
 447               progress = true;
 448               j = refresh_out_pos(j);
 449               max = outcnt();
 450             }
 451           }
 452         }
 453       }
 454       add_to_worklist = true;
 455       i--;
 456     }
 457   }
 458 
 459   if (can_reshape && cnt == 1) {
 460     // Is it dead loop?
 461     // If it is LoopNopde it had 2 (+1 itself) inputs and
 462     // one of them was cut. The loop is dead if it was EntryContol.
 463     // Loop node may have only one input because entry path
 464     // is removed in PhaseIdealLoop::Dominators().
 465     assert(!this->is_Loop() || cnt_orig <= 3, "Loop node should have 3 or less inputs");
 466     if (this->is_Loop() && (del_it == LoopNode::EntryControl ||
 467                             del_it == 0 && is_unreachable_region(phase)) ||
 468        !this->is_Loop() && has_phis && is_unreachable_region(phase)) {
 469       // Yes,  the region will be removed during the next step below.
 470       // Cut the backedge input and remove phis since no data paths left.
 471       // We don't cut outputs to other nodes here since we need to put them
 472       // on the worklist.
 473       del_req(1);
 474       cnt = 0;
 475       assert( req() == 1, "no more inputs expected" );
 476       uint max = outcnt();
 477       bool progress = true;
 478       Node *top = phase->C->top();
 479       PhaseIterGVN *igvn = phase->is_IterGVN();
 480       DUIterator j;
 481       while(progress) {
 482         progress = false;
 483         for (j = outs(); has_out(j); j++) {
 484           Node *n = out(j);
 485           if( n->is_Phi() ) {
 486             assert( igvn->eqv(n->in(0), this), "" );
 487             assert( n->req() == 2 &&  n->in(1) != NULL, "Only one data input expected" );
 488             // Break dead loop data path.
 489             // Eagerly replace phis with top to avoid phis copies generation.
 490             igvn->replace_node(n, top);
 491             if( max != outcnt() ) {
 492               progress = true;
 493               j = refresh_out_pos(j);
 494               max = outcnt();
 495             }
 496           }
 497         }
 498       }
 499       add_to_worklist = true;
 500     }
 501   }
 502   if (add_to_worklist) {
 503     phase->is_IterGVN()->add_users_to_worklist(this); // Revisit collapsed Phis
 504   }
 505 
 506   if( cnt <= 1 ) {              // Only 1 path in?
 507     set_req(0, NULL);           // Null control input for region copy
 508     if( cnt == 0 && !can_reshape) { // Parse phase - leave the node as it is.
 509       // No inputs or all inputs are NULL.
 510       return NULL;
 511     } else if (can_reshape) {   // Optimization phase - remove the node
 512       PhaseIterGVN *igvn = phase->is_IterGVN();
 513       Node *parent_ctrl;
 514       if( cnt == 0 ) {
 515         assert( req() == 1, "no inputs expected" );
 516         // During IGVN phase such region will be subsumed by TOP node
 517         // so region's phis will have TOP as control node.
 518         // Kill phis here to avoid it. PhiNode::is_copy() will be always false.
 519         // Also set other user's input to top.
 520         parent_ctrl = phase->C->top();
 521       } else {
 522         // The fallthrough case since we already checked dead loops above.
 523         parent_ctrl = in(1);
 524         assert(parent_ctrl != NULL, "Region is a copy of some non-null control");
 525         assert(!igvn->eqv(parent_ctrl, this), "Close dead loop");
 526       }
 527       if (!add_to_worklist)
 528         igvn->add_users_to_worklist(this); // Check for further allowed opts
 529       for (DUIterator_Last imin, i = last_outs(imin); i >= imin; --i) {
 530         Node* n = last_out(i);
 531         igvn->hash_delete(n); // Remove from worklist before modifying edges
 532         if( n->is_Phi() ) {   // Collapse all Phis
 533           // Eagerly replace phis to avoid copies generation.
 534           Node* in;
 535           if( cnt == 0 ) {
 536             assert( n->req() == 1, "No data inputs expected" );
 537             in = parent_ctrl; // replaced by top
 538           } else {
 539             assert( n->req() == 2 &&  n->in(1) != NULL, "Only one data input expected" );
 540             in = n->in(1);               // replaced by unique input
 541             if( n->as_Phi()->is_unsafe_data_reference(in) )
 542               in = phase->C->top();      // replaced by top
 543           }
 544           igvn->replace_node(n, in);
 545         }
 546         else if( n->is_Region() ) { // Update all incoming edges
 547           assert( !igvn->eqv(n, this), "Must be removed from DefUse edges");
 548           uint uses_found = 0;
 549           for( uint k=1; k < n->req(); k++ ) {
 550             if( n->in(k) == this ) {
 551               n->set_req(k, parent_ctrl);
 552               uses_found++;
 553             }
 554           }
 555           if( uses_found > 1 ) { // (--i) done at the end of the loop.
 556             i -= (uses_found - 1);
 557           }
 558         }
 559         else {
 560           assert( igvn->eqv(n->in(0), this), "Expect RegionNode to be control parent");
 561           n->set_req(0, parent_ctrl);
 562         }
 563 #ifdef ASSERT
 564         for( uint k=0; k < n->req(); k++ ) {
 565           assert( !igvn->eqv(n->in(k), this), "All uses of RegionNode should be gone");
 566         }
 567 #endif
 568       }
 569       // Remove the RegionNode itself from DefUse info
 570       igvn->remove_dead_node(this);
 571       return NULL;
 572     }
 573     return this;                // Record progress
 574   }
 575 
 576 
 577   // If a Region flows into a Region, merge into one big happy merge.
 578   if (can_reshape) {
 579     Node *m = merge_region(this, phase);
 580     if (m != NULL)  return m;
 581   }
 582 
 583   // Check if this region is the root of a clipping idiom on floats
 584   if( ConvertFloat2IntClipping && can_reshape && req() == 4 ) {
 585     // Check that only one use is a Phi and that it simplifies to two constants +
 586     PhiNode* phi = has_unique_phi();
 587     if (phi != NULL) {          // One Phi user
 588       // Check inputs to the Phi
 589       ConNode *min;
 590       ConNode *max;
 591       Node    *val;
 592       uint     min_idx;
 593       uint     max_idx;
 594       uint     val_idx;
 595       if( check_phi_clipping( phi, min, min_idx, max, max_idx, val, val_idx )  ) {
 596         IfNode *top_if;
 597         IfNode *bot_if;
 598         if( check_if_clipping( this, bot_if, top_if ) ) {
 599           // Control pattern checks, now verify compares
 600           Node   *top_in = NULL;   // value being compared against
 601           Node   *bot_in = NULL;
 602           if( check_compare_clipping( true,  bot_if, min, bot_in ) &&
 603               check_compare_clipping( false, top_if, max, top_in ) ) {
 604             if( bot_in == top_in ) {
 605               PhaseIterGVN *gvn = phase->is_IterGVN();
 606               assert( gvn != NULL, "Only had DefUse info in IterGVN");
 607               // Only remaining check is that bot_in == top_in == (Phi's val + mods)
 608 
 609               // Check for the ConvF2INode
 610               ConvF2INode *convf2i;
 611               if( check_convf2i_clipping( phi, val_idx, convf2i, min, max ) &&
 612                 convf2i->in(1) == bot_in ) {
 613                 // Matched pattern, including LShiftI; RShiftI, replace with integer compares
 614                 // max test
 615                 Node *cmp   = gvn->register_new_node_with_optimizer(new (phase->C) CmpINode( convf2i, min ));
 616                 Node *boo   = gvn->register_new_node_with_optimizer(new (phase->C) BoolNode( cmp, BoolTest::lt ));
 617                 IfNode *iff = (IfNode*)gvn->register_new_node_with_optimizer(new (phase->C) IfNode( top_if->in(0), boo, PROB_UNLIKELY_MAG(5), top_if->_fcnt ));
 618                 Node *if_min= gvn->register_new_node_with_optimizer(new (phase->C) IfTrueNode (iff));
 619                 Node *ifF   = gvn->register_new_node_with_optimizer(new (phase->C) IfFalseNode(iff));
 620                 // min test
 621                 cmp         = gvn->register_new_node_with_optimizer(new (phase->C) CmpINode( convf2i, max ));
 622                 boo         = gvn->register_new_node_with_optimizer(new (phase->C) BoolNode( cmp, BoolTest::gt ));
 623                 iff         = (IfNode*)gvn->register_new_node_with_optimizer(new (phase->C) IfNode( ifF, boo, PROB_UNLIKELY_MAG(5), bot_if->_fcnt ));
 624                 Node *if_max= gvn->register_new_node_with_optimizer(new (phase->C) IfTrueNode (iff));
 625                 ifF         = gvn->register_new_node_with_optimizer(new (phase->C) IfFalseNode(iff));
 626                 // update input edges to region node
 627                 set_req_X( min_idx, if_min, gvn );
 628                 set_req_X( max_idx, if_max, gvn );
 629                 set_req_X( val_idx, ifF,    gvn );
 630                 // remove unnecessary 'LShiftI; RShiftI' idiom
 631                 gvn->hash_delete(phi);
 632                 phi->set_req_X( val_idx, convf2i, gvn );
 633                 gvn->hash_find_insert(phi);
 634                 // Return transformed region node
 635                 return this;
 636               }
 637             }
 638           }
 639         }
 640       }
 641     }
 642   }
 643 
 644   return NULL;
 645 }
 646 
 647 
 648 
 649 const RegMask &RegionNode::out_RegMask() const {
 650   return RegMask::Empty;
 651 }
 652 
 653 // Find the one non-null required input.  RegionNode only
 654 Node *Node::nonnull_req() const {
 655   assert( is_Region(), "" );
 656   for( uint i = 1; i < _cnt; i++ )
 657     if( in(i) )
 658       return in(i);
 659   ShouldNotReachHere();
 660   return NULL;
 661 }
 662 
 663 
 664 //=============================================================================
 665 // note that these functions assume that the _adr_type field is flattened
 666 uint PhiNode::hash() const {
 667   const Type* at = _adr_type;
 668   return TypeNode::hash() + (at ? at->hash() : 0);
 669 }
 670 uint PhiNode::cmp( const Node &n ) const {
 671   return TypeNode::cmp(n) && _adr_type == ((PhiNode&)n)._adr_type;
 672 }
 673 static inline
 674 const TypePtr* flatten_phi_adr_type(const TypePtr* at) {
 675   if (at == NULL || at == TypePtr::BOTTOM)  return at;
 676   return Compile::current()->alias_type(at)->adr_type();
 677 }
 678 
 679 //----------------------------make---------------------------------------------
 680 // create a new phi with edges matching r and set (initially) to x
 681 PhiNode* PhiNode::make(Node* r, Node* x, const Type *t, const TypePtr* at) {
 682   uint preds = r->req();   // Number of predecessor paths
 683   assert(t != Type::MEMORY || at == flatten_phi_adr_type(at), "flatten at");
 684   PhiNode* p = new (Compile::current()) PhiNode(r, t, at);
 685   for (uint j = 1; j < preds; j++) {
 686     // Fill in all inputs, except those which the region does not yet have
 687     if (r->in(j) != NULL)
 688       p->init_req(j, x);
 689   }
 690   return p;
 691 }
 692 PhiNode* PhiNode::make(Node* r, Node* x) {
 693   const Type*    t  = x->bottom_type();
 694   const TypePtr* at = NULL;
 695   if (t == Type::MEMORY)  at = flatten_phi_adr_type(x->adr_type());
 696   return make(r, x, t, at);
 697 }
 698 PhiNode* PhiNode::make_blank(Node* r, Node* x) {
 699   const Type*    t  = x->bottom_type();
 700   const TypePtr* at = NULL;
 701   if (t == Type::MEMORY)  at = flatten_phi_adr_type(x->adr_type());
 702   return new (Compile::current()) PhiNode(r, t, at);
 703 }
 704 
 705 
 706 //------------------------slice_memory-----------------------------------------
 707 // create a new phi with narrowed memory type
 708 PhiNode* PhiNode::slice_memory(const TypePtr* adr_type) const {
 709   PhiNode* mem = (PhiNode*) clone();
 710   *(const TypePtr**)&mem->_adr_type = adr_type;
 711   // convert self-loops, or else we get a bad graph
 712   for (uint i = 1; i < req(); i++) {
 713     if ((const Node*)in(i) == this)  mem->set_req(i, mem);
 714   }
 715   mem->verify_adr_type();
 716   return mem;
 717 }
 718 
 719 //------------------------split_out_instance-----------------------------------
 720 // Split out an instance type from a bottom phi.
 721 PhiNode* PhiNode::split_out_instance(const TypePtr* at, PhaseIterGVN *igvn) const {
 722   const TypeOopPtr *t_oop = at->isa_oopptr();
 723   assert(t_oop != NULL && t_oop->is_known_instance(), "expecting instance oopptr");
 724   const TypePtr *t = adr_type();
 725   assert(type() == Type::MEMORY &&
 726          (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
 727           t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
 728           t->is_oopptr()->cast_to_exactness(true)
 729            ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
 730            ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop),
 731          "bottom or raw memory required");
 732 
 733   // Check if an appropriate node already exists.
 734   Node *region = in(0);
 735   for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
 736     Node* use = region->fast_out(k);
 737     if( use->is_Phi()) {
 738       PhiNode *phi2 = use->as_Phi();
 739       if (phi2->type() == Type::MEMORY && phi2->adr_type() == at) {
 740         return phi2;
 741       }
 742     }
 743   }
 744   Compile *C = igvn->C;
 745   Arena *a = Thread::current()->resource_area();
 746   Node_Array node_map = new Node_Array(a);
 747   Node_Stack stack(a, C->unique() >> 4);
 748   PhiNode *nphi = slice_memory(at);
 749   igvn->register_new_node_with_optimizer( nphi );
 750   node_map.map(_idx, nphi);
 751   stack.push((Node *)this, 1);
 752   while(!stack.is_empty()) {
 753     PhiNode *ophi = stack.node()->as_Phi();
 754     uint i = stack.index();
 755     assert(i >= 1, "not control edge");
 756     stack.pop();
 757     nphi = node_map[ophi->_idx]->as_Phi();
 758     for (; i < ophi->req(); i++) {
 759       Node *in = ophi->in(i);
 760       if (in == NULL || igvn->type(in) == Type::TOP)
 761         continue;
 762       Node *opt = MemNode::optimize_simple_memory_chain(in, at, igvn);
 763       PhiNode *optphi = opt->is_Phi() ? opt->as_Phi() : NULL;
 764       if (optphi != NULL && optphi->adr_type() == TypePtr::BOTTOM) {
 765         opt = node_map[optphi->_idx];
 766         if (opt == NULL) {
 767           stack.push(ophi, i);
 768           nphi = optphi->slice_memory(at);
 769           igvn->register_new_node_with_optimizer( nphi );
 770           node_map.map(optphi->_idx, nphi);
 771           ophi = optphi;
 772           i = 0; // will get incremented at top of loop
 773           continue;
 774         }
 775       }
 776       nphi->set_req(i, opt);
 777     }
 778   }
 779   return nphi;
 780 }
 781 
 782 //------------------------verify_adr_type--------------------------------------
 783 #ifdef ASSERT
 784 void PhiNode::verify_adr_type(VectorSet& visited, const TypePtr* at) const {
 785   if (visited.test_set(_idx))  return;  //already visited
 786 
 787   // recheck constructor invariants:
 788   verify_adr_type(false);
 789 
 790   // recheck local phi/phi consistency:
 791   assert(_adr_type == at || _adr_type == TypePtr::BOTTOM,
 792          "adr_type must be consistent across phi nest");
 793 
 794   // walk around
 795   for (uint i = 1; i < req(); i++) {
 796     Node* n = in(i);
 797     if (n == NULL)  continue;
 798     const Node* np = in(i);
 799     if (np->is_Phi()) {
 800       np->as_Phi()->verify_adr_type(visited, at);
 801     } else if (n->bottom_type() == Type::TOP
 802                || (n->is_Mem() && n->in(MemNode::Address)->bottom_type() == Type::TOP)) {
 803       // ignore top inputs
 804     } else {
 805       const TypePtr* nat = flatten_phi_adr_type(n->adr_type());
 806       // recheck phi/non-phi consistency at leaves:
 807       assert((nat != NULL) == (at != NULL), "");
 808       assert(nat == at || nat == TypePtr::BOTTOM,
 809              "adr_type must be consistent at leaves of phi nest");
 810     }
 811   }
 812 }
 813 
 814 // Verify a whole nest of phis rooted at this one.
 815 void PhiNode::verify_adr_type(bool recursive) const {
 816   if (is_error_reported())  return;  // muzzle asserts when debugging an error
 817   if (Node::in_dump())      return;  // muzzle asserts when printing
 818 
 819   assert((_type == Type::MEMORY) == (_adr_type != NULL), "adr_type for memory phis only");
 820 
 821   if (!VerifyAliases)       return;  // verify thoroughly only if requested
 822 
 823   assert(_adr_type == flatten_phi_adr_type(_adr_type),
 824          "Phi::adr_type must be pre-normalized");
 825 
 826   if (recursive) {
 827     VectorSet visited(Thread::current()->resource_area());
 828     verify_adr_type(visited, _adr_type);
 829   }
 830 }
 831 #endif
 832 
 833 
 834 //------------------------------Value------------------------------------------
 835 // Compute the type of the PhiNode
 836 const Type *PhiNode::Value( PhaseTransform *phase ) const {
 837   Node *r = in(0);              // RegionNode
 838   if( !r )                      // Copy or dead
 839     return in(1) ? phase->type(in(1)) : Type::TOP;
 840 
 841   // Note: During parsing, phis are often transformed before their regions.
 842   // This means we have to use type_or_null to defend against untyped regions.
 843   if( phase->type_or_null(r) == Type::TOP )  // Dead code?
 844     return Type::TOP;
 845 
 846   // Check for trip-counted loop.  If so, be smarter.
 847   CountedLoopNode *l = r->is_CountedLoop() ? r->as_CountedLoop() : NULL;
 848   if( l && l->can_be_counted_loop(phase) &&
 849       ((const Node*)l->phi() == this) ) { // Trip counted loop!
 850     // protect against init_trip() or limit() returning NULL
 851     const Node *init   = l->init_trip();
 852     const Node *limit  = l->limit();
 853     if( init != NULL && limit != NULL && l->stride_is_con() ) {
 854       const TypeInt *lo = init ->bottom_type()->isa_int();
 855       const TypeInt *hi = limit->bottom_type()->isa_int();
 856       if( lo && hi ) {            // Dying loops might have TOP here
 857         int stride = l->stride_con();
 858         if( stride < 0 ) {          // Down-counter loop
 859           const TypeInt *tmp = lo; lo = hi; hi = tmp;
 860           stride = -stride;
 861         }
 862         if( lo->_hi < hi->_lo )     // Reversed endpoints are well defined :-(
 863           return TypeInt::make(lo->_lo,hi->_hi,3);
 864       }
 865     }
 866   }
 867 
 868   // Until we have harmony between classes and interfaces in the type
 869   // lattice, we must tread carefully around phis which implicitly
 870   // convert the one to the other.
 871   const TypePtr* ttp = _type->make_ptr();
 872   const TypeInstPtr* ttip = (ttp != NULL) ? ttp->isa_instptr() : NULL;
 873   const TypeKlassPtr* ttkp = (ttp != NULL) ? ttp->isa_klassptr() : NULL;
 874   bool is_intf = false;
 875   if (ttip != NULL) {
 876     ciKlass* k = ttip->klass();
 877     if (k->is_loaded() && k->is_interface())
 878       is_intf = true;
 879   }
 880   if (ttkp != NULL) {
 881     ciKlass* k = ttkp->klass();
 882     if (k->is_loaded() && k->is_interface())
 883       is_intf = true;
 884   }
 885 
 886   // Default case: merge all inputs
 887   const Type *t = Type::TOP;        // Merged type starting value
 888   for (uint i = 1; i < req(); ++i) {// For all paths in
 889     // Reachable control path?
 890     if (r->in(i) && phase->type(r->in(i)) == Type::CONTROL) {
 891       const Type* ti = phase->type(in(i));
 892       // We assume that each input of an interface-valued Phi is a true
 893       // subtype of that interface.  This might not be true of the meet
 894       // of all the input types.  The lattice is not distributive in
 895       // such cases.  Ward off asserts in type.cpp by refusing to do
 896       // meets between interfaces and proper classes.
 897       const TypePtr* tip = ti->make_ptr();
 898       const TypeInstPtr* tiip = (tip != NULL) ? tip->isa_instptr() : NULL;
 899       if (tiip) {
 900         bool ti_is_intf = false;
 901         ciKlass* k = tiip->klass();
 902         if (k->is_loaded() && k->is_interface())
 903           ti_is_intf = true;
 904         if (is_intf != ti_is_intf)
 905           { t = _type; break; }
 906       }
 907       t = t->meet(ti);
 908     }
 909   }
 910 
 911   // The worst-case type (from ciTypeFlow) should be consistent with "t".
 912   // That is, we expect that "t->higher_equal(_type)" holds true.
 913   // There are various exceptions:
 914   // - Inputs which are phis might in fact be widened unnecessarily.
 915   //   For example, an input might be a widened int while the phi is a short.
 916   // - Inputs might be BotPtrs but this phi is dependent on a null check,
 917   //   and postCCP has removed the cast which encodes the result of the check.
 918   // - The type of this phi is an interface, and the inputs are classes.
 919   // - Value calls on inputs might produce fuzzy results.
 920   //   (Occurrences of this case suggest improvements to Value methods.)
 921   //
 922   // It is not possible to see Type::BOTTOM values as phi inputs,
 923   // because the ciTypeFlow pre-pass produces verifier-quality types.
 924   const Type* ft = t->filter(_type);  // Worst case type
 925 
 926 #ifdef ASSERT
 927   // The following logic has been moved into TypeOopPtr::filter.
 928   const Type* jt = t->join(_type);
 929   if( jt->empty() ) {           // Emptied out???
 930 
 931     // Check for evil case of 't' being a class and '_type' expecting an
 932     // interface.  This can happen because the bytecodes do not contain
 933     // enough type info to distinguish a Java-level interface variable
 934     // from a Java-level object variable.  If we meet 2 classes which
 935     // both implement interface I, but their meet is at 'j/l/O' which
 936     // doesn't implement I, we have no way to tell if the result should
 937     // be 'I' or 'j/l/O'.  Thus we'll pick 'j/l/O'.  If this then flows
 938     // into a Phi which "knows" it's an Interface type we'll have to
 939     // uplift the type.
 940     if( !t->empty() && ttip && ttip->is_loaded() && ttip->klass()->is_interface() )
 941       { assert(ft == _type, ""); } // Uplift to interface
 942     else if( !t->empty() && ttkp && ttkp->is_loaded() && ttkp->klass()->is_interface() )
 943       { assert(ft == _type, ""); } // Uplift to interface
 944     // Otherwise it's something stupid like non-overlapping int ranges
 945     // found on dying counted loops.
 946     else
 947       { assert(ft == Type::TOP, ""); } // Canonical empty value
 948   }
 949 
 950   else {
 951 
 952     // If we have an interface-typed Phi and we narrow to a class type, the join
 953     // should report back the class.  However, if we have a J/L/Object
 954     // class-typed Phi and an interface flows in, it's possible that the meet &
 955     // join report an interface back out.  This isn't possible but happens
 956     // because the type system doesn't interact well with interfaces.
 957     const TypePtr *jtp = jt->make_ptr();
 958     const TypeInstPtr *jtip = (jtp != NULL) ? jtp->isa_instptr() : NULL;
 959     const TypeKlassPtr *jtkp = (jtp != NULL) ? jtp->isa_klassptr() : NULL;
 960     if( jtip && ttip ) {
 961       if( jtip->is_loaded() &&  jtip->klass()->is_interface() &&
 962           ttip->is_loaded() && !ttip->klass()->is_interface() ) {
 963         // Happens in a CTW of rt.jar, 320-341, no extra flags
 964         assert(ft == ttip->cast_to_ptr_type(jtip->ptr()) ||
 965                ft->isa_narrowoop() && ft->make_ptr() == ttip->cast_to_ptr_type(jtip->ptr()), "");
 966         jt = ft;
 967       }
 968     }
 969     if( jtkp && ttkp ) {
 970       if( jtkp->is_loaded() &&  jtkp->klass()->is_interface() &&
 971           !jtkp->klass_is_exact() && // Keep exact interface klass (6894807)
 972           ttkp->is_loaded() && !ttkp->klass()->is_interface() ) {
 973         assert(ft == ttkp->cast_to_ptr_type(jtkp->ptr()) ||
 974                ft->isa_narrowoop() && ft->make_ptr() == ttkp->cast_to_ptr_type(jtkp->ptr()), "");
 975         jt = ft;
 976       }
 977     }
 978     if (jt != ft && jt->base() == ft->base()) {
 979       if (jt->isa_int() &&
 980           jt->is_int()->_lo == ft->is_int()->_lo &&
 981           jt->is_int()->_hi == ft->is_int()->_hi)
 982         jt = ft;
 983       if (jt->isa_long() &&
 984           jt->is_long()->_lo == ft->is_long()->_lo &&
 985           jt->is_long()->_hi == ft->is_long()->_hi)
 986         jt = ft;
 987     }
 988     if (jt != ft) {
 989       tty->print("merge type:  "); t->dump(); tty->cr();
 990       tty->print("kill type:   "); _type->dump(); tty->cr();
 991       tty->print("join type:   "); jt->dump(); tty->cr();
 992       tty->print("filter type: "); ft->dump(); tty->cr();
 993     }
 994     assert(jt == ft, "");
 995   }
 996 #endif //ASSERT
 997 
 998   // Deal with conversion problems found in data loops.
 999   ft = phase->saturate(ft, phase->type_or_null(this), _type);
1000 
1001   return ft;
1002 }
1003 
1004 
1005 //------------------------------is_diamond_phi---------------------------------
1006 // Does this Phi represent a simple well-shaped diamond merge?  Return the
1007 // index of the true path or 0 otherwise.
1008 int PhiNode::is_diamond_phi() const {
1009   // Check for a 2-path merge
1010   Node *region = in(0);
1011   if( !region ) return 0;
1012   if( region->req() != 3 ) return 0;
1013   if(         req() != 3 ) return 0;
1014   // Check that both paths come from the same If
1015   Node *ifp1 = region->in(1);
1016   Node *ifp2 = region->in(2);
1017   if( !ifp1 || !ifp2 ) return 0;
1018   Node *iff = ifp1->in(0);
1019   if( !iff || !iff->is_If() ) return 0;
1020   if( iff != ifp2->in(0) ) return 0;
1021   // Check for a proper bool/cmp
1022   const Node *b = iff->in(1);
1023   if( !b->is_Bool() ) return 0;
1024   const Node *cmp = b->in(1);
1025   if( !cmp->is_Cmp() ) return 0;
1026 
1027   // Check for branching opposite expected
1028   if( ifp2->Opcode() == Op_IfTrue ) {
1029     assert( ifp1->Opcode() == Op_IfFalse, "" );
1030     return 2;
1031   } else {
1032     assert( ifp1->Opcode() == Op_IfTrue, "" );
1033     return 1;
1034   }
1035 }
1036 
1037 //----------------------------check_cmove_id-----------------------------------
1038 // Check for CMove'ing a constant after comparing against the constant.
1039 // Happens all the time now, since if we compare equality vs a constant in
1040 // the parser, we "know" the variable is constant on one path and we force
1041 // it.  Thus code like "if( x==0 ) {/*EMPTY*/}" ends up inserting a
1042 // conditional move: "x = (x==0)?0:x;".  Yucko.  This fix is slightly more
1043 // general in that we don't need constants.  Since CMove's are only inserted
1044 // in very special circumstances, we do it here on generic Phi's.
1045 Node* PhiNode::is_cmove_id(PhaseTransform* phase, int true_path) {
1046   assert(true_path !=0, "only diamond shape graph expected");
1047 
1048   // is_diamond_phi() has guaranteed the correctness of the nodes sequence:
1049   // phi->region->if_proj->ifnode->bool->cmp
1050   Node*     region = in(0);
1051   Node*     iff    = region->in(1)->in(0);
1052   BoolNode* b      = iff->in(1)->as_Bool();
1053   Node*     cmp    = b->in(1);
1054   Node*     tval   = in(true_path);
1055   Node*     fval   = in(3-true_path);
1056   Node*     id     = CMoveNode::is_cmove_id(phase, cmp, tval, fval, b);
1057   if (id == NULL)
1058     return NULL;
1059 
1060   // Either value might be a cast that depends on a branch of 'iff'.
1061   // Since the 'id' value will float free of the diamond, either
1062   // decast or return failure.
1063   Node* ctl = id->in(0);
1064   if (ctl != NULL && ctl->in(0) == iff) {
1065     if (id->is_ConstraintCast()) {
1066       return id->in(1);
1067     } else {
1068       // Don't know how to disentangle this value.
1069       return NULL;
1070     }
1071   }
1072 
1073   return id;
1074 }
1075 
1076 //------------------------------Identity---------------------------------------
1077 // Check for Region being Identity.
1078 Node *PhiNode::Identity( PhaseTransform *phase ) {
1079   // Check for no merging going on
1080   // (There used to be special-case code here when this->region->is_Loop.
1081   // It would check for a tributary phi on the backedge that the main phi
1082   // trivially, perhaps with a single cast.  The unique_input method
1083   // does all this and more, by reducing such tributaries to 'this'.)
1084   Node* uin = unique_input(phase);
1085   if (uin != NULL) {
1086     return uin;
1087   }
1088 
1089   int true_path = is_diamond_phi();
1090   if (true_path != 0) {
1091     Node* id = is_cmove_id(phase, true_path);
1092     if (id != NULL)  return id;
1093   }
1094 
1095   return this;                     // No identity
1096 }
1097 
1098 //-----------------------------unique_input------------------------------------
1099 // Find the unique value, discounting top, self-loops, and casts.
1100 // Return top if there are no inputs, and self if there are multiple.
1101 Node* PhiNode::unique_input(PhaseTransform* phase) {
1102   //  1) One unique direct input, or
1103   //  2) some of the inputs have an intervening ConstraintCast and
1104   //     the type of input is the same or sharper (more specific)
1105   //     than the phi's type.
1106   //  3) an input is a self loop
1107   //
1108   //  1) input   or   2) input     or   3) input __
1109   //     /   \           /   \               \  /  \
1110   //     \   /          |    cast             phi  cast
1111   //      phi            \   /               /  \  /
1112   //                      phi               /    --
1113 
1114   Node* r = in(0);                      // RegionNode
1115   if (r == NULL)  return in(1);         // Already degraded to a Copy
1116   Node* uncasted_input = NULL; // The unique uncasted input (ConstraintCasts removed)
1117   Node* direct_input   = NULL; // The unique direct input
1118 
1119   for (uint i = 1, cnt = req(); i < cnt; ++i) {
1120     Node* rc = r->in(i);
1121     if (rc == NULL || phase->type(rc) == Type::TOP)
1122       continue;                 // ignore unreachable control path
1123     Node* n = in(i);
1124     if (n == NULL)
1125       continue;
1126     Node* un = n->uncast();
1127     if (un == NULL || un == this || phase->type(un) == Type::TOP) {
1128       continue; // ignore if top, or in(i) and "this" are in a data cycle
1129     }
1130     // Check for a unique uncasted input
1131     if (uncasted_input == NULL) {
1132       uncasted_input = un;
1133     } else if (uncasted_input != un) {
1134       uncasted_input = NodeSentinel; // no unique uncasted input
1135     }
1136     // Check for a unique direct input
1137     if (direct_input == NULL) {
1138       direct_input = n;
1139     } else if (direct_input != n) {
1140       direct_input = NodeSentinel; // no unique direct input
1141     }
1142   }
1143   if (direct_input == NULL) {
1144     return phase->C->top();        // no inputs
1145   }
1146   assert(uncasted_input != NULL,"");
1147 
1148   if (direct_input != NodeSentinel) {
1149     return direct_input;           // one unique direct input
1150   }
1151   if (uncasted_input != NodeSentinel &&
1152       phase->type(uncasted_input)->higher_equal(type())) {
1153     return uncasted_input;         // one unique uncasted input
1154   }
1155 
1156   // Nothing.
1157   return NULL;
1158 }
1159 
1160 //------------------------------is_x2logic-------------------------------------
1161 // Check for simple convert-to-boolean pattern
1162 // If:(C Bool) Region:(IfF IfT) Phi:(Region 0 1)
1163 // Convert Phi to an ConvIB.
1164 static Node *is_x2logic( PhaseGVN *phase, PhiNode *phi, int true_path ) {
1165   assert(true_path !=0, "only diamond shape graph expected");
1166   // Convert the true/false index into an expected 0/1 return.
1167   // Map 2->0 and 1->1.
1168   int flipped = 2-true_path;
1169 
1170   // is_diamond_phi() has guaranteed the correctness of the nodes sequence:
1171   // phi->region->if_proj->ifnode->bool->cmp
1172   Node *region = phi->in(0);
1173   Node *iff = region->in(1)->in(0);
1174   BoolNode *b = (BoolNode*)iff->in(1);
1175   const CmpNode *cmp = (CmpNode*)b->in(1);
1176 
1177   Node *zero = phi->in(1);
1178   Node *one  = phi->in(2);
1179   const Type *tzero = phase->type( zero );
1180   const Type *tone  = phase->type( one  );
1181 
1182   // Check for compare vs 0
1183   const Type *tcmp = phase->type(cmp->in(2));
1184   if( tcmp != TypeInt::ZERO && tcmp != TypePtr::NULL_PTR ) {
1185     // Allow cmp-vs-1 if the other input is bounded by 0-1
1186     if( !(tcmp == TypeInt::ONE && phase->type(cmp->in(1)) == TypeInt::BOOL) )
1187       return NULL;
1188     flipped = 1-flipped;        // Test is vs 1 instead of 0!
1189   }
1190 
1191   // Check for setting zero/one opposite expected
1192   if( tzero == TypeInt::ZERO ) {
1193     if( tone == TypeInt::ONE ) {
1194     } else return NULL;
1195   } else if( tzero == TypeInt::ONE ) {
1196     if( tone == TypeInt::ZERO ) {
1197       flipped = 1-flipped;
1198     } else return NULL;
1199   } else return NULL;
1200 
1201   // Check for boolean test backwards
1202   if( b->_test._test == BoolTest::ne ) {
1203   } else if( b->_test._test == BoolTest::eq ) {
1204     flipped = 1-flipped;
1205   } else return NULL;
1206 
1207   // Build int->bool conversion
1208   Node *n = new (phase->C) Conv2BNode( cmp->in(1) );
1209   if( flipped )
1210     n = new (phase->C) XorINode( phase->transform(n), phase->intcon(1) );
1211 
1212   return n;
1213 }
1214 
1215 //------------------------------is_cond_add------------------------------------
1216 // Check for simple conditional add pattern:  "(P < Q) ? X+Y : X;"
1217 // To be profitable the control flow has to disappear; there can be no other
1218 // values merging here.  We replace the test-and-branch with:
1219 // "(sgn(P-Q))&Y) + X".  Basically, convert "(P < Q)" into 0 or -1 by
1220 // moving the carry bit from (P-Q) into a register with 'sbb EAX,EAX'.
1221 // Then convert Y to 0-or-Y and finally add.
1222 // This is a key transform for SpecJava _201_compress.
1223 static Node* is_cond_add(PhaseGVN *phase, PhiNode *phi, int true_path) {
1224   assert(true_path !=0, "only diamond shape graph expected");
1225 
1226   // is_diamond_phi() has guaranteed the correctness of the nodes sequence:
1227   // phi->region->if_proj->ifnode->bool->cmp
1228   RegionNode *region = (RegionNode*)phi->in(0);
1229   Node *iff = region->in(1)->in(0);
1230   BoolNode* b = iff->in(1)->as_Bool();
1231   const CmpNode *cmp = (CmpNode*)b->in(1);
1232 
1233   // Make sure only merging this one phi here
1234   if (region->has_unique_phi() != phi)  return NULL;
1235 
1236   // Make sure each arm of the diamond has exactly one output, which we assume
1237   // is the region.  Otherwise, the control flow won't disappear.
1238   if (region->in(1)->outcnt() != 1) return NULL;
1239   if (region->in(2)->outcnt() != 1) return NULL;
1240 
1241   // Check for "(P < Q)" of type signed int
1242   if (b->_test._test != BoolTest::lt)  return NULL;
1243   if (cmp->Opcode() != Op_CmpI)        return NULL;
1244 
1245   Node *p = cmp->in(1);
1246   Node *q = cmp->in(2);
1247   Node *n1 = phi->in(  true_path);
1248   Node *n2 = phi->in(3-true_path);
1249 
1250   int op = n1->Opcode();
1251   if( op != Op_AddI           // Need zero as additive identity
1252       /*&&op != Op_SubI &&
1253       op != Op_AddP &&
1254       op != Op_XorI &&
1255       op != Op_OrI*/ )
1256     return NULL;
1257 
1258   Node *x = n2;
1259   Node *y = n1->in(1);
1260   if( n2 == n1->in(1) ) {
1261     y = n1->in(2);
1262   } else if( n2 == n1->in(1) ) {
1263   } else return NULL;
1264 
1265   // Not so profitable if compare and add are constants
1266   if( q->is_Con() && phase->type(q) != TypeInt::ZERO && y->is_Con() )
1267     return NULL;
1268 
1269   Node *cmplt = phase->transform( new (phase->C) CmpLTMaskNode(p,q) );
1270   Node *j_and   = phase->transform( new (phase->C) AndINode(cmplt,y) );
1271   return new (phase->C) AddINode(j_and,x);
1272 }
1273 
1274 //------------------------------is_absolute------------------------------------
1275 // Check for absolute value.
1276 static Node* is_absolute( PhaseGVN *phase, PhiNode *phi_root, int true_path) {
1277   assert(true_path !=0, "only diamond shape graph expected");
1278 
1279   int  cmp_zero_idx = 0;        // Index of compare input where to look for zero
1280   int  phi_x_idx = 0;           // Index of phi input where to find naked x
1281 
1282   // ABS ends with the merge of 2 control flow paths.
1283   // Find the false path from the true path. With only 2 inputs, 3 - x works nicely.
1284   int false_path = 3 - true_path;
1285 
1286   // is_diamond_phi() has guaranteed the correctness of the nodes sequence:
1287   // phi->region->if_proj->ifnode->bool->cmp
1288   BoolNode *bol = phi_root->in(0)->in(1)->in(0)->in(1)->as_Bool();
1289 
1290   // Check bool sense
1291   switch( bol->_test._test ) {
1292   case BoolTest::lt: cmp_zero_idx = 1; phi_x_idx = true_path;  break;
1293   case BoolTest::le: cmp_zero_idx = 2; phi_x_idx = false_path; break;
1294   case BoolTest::gt: cmp_zero_idx = 2; phi_x_idx = true_path;  break;
1295   case BoolTest::ge: cmp_zero_idx = 1; phi_x_idx = false_path; break;
1296   default:           return NULL;                              break;
1297   }
1298 
1299   // Test is next
1300   Node *cmp = bol->in(1);
1301   const Type *tzero = NULL;
1302   switch( cmp->Opcode() ) {
1303   case Op_CmpF:    tzero = TypeF::ZERO; break; // Float ABS
1304   case Op_CmpD:    tzero = TypeD::ZERO; break; // Double ABS
1305   default: return NULL;
1306   }
1307 
1308   // Find zero input of compare; the other input is being abs'd
1309   Node *x = NULL;
1310   bool flip = false;
1311   if( phase->type(cmp->in(cmp_zero_idx)) == tzero ) {
1312     x = cmp->in(3 - cmp_zero_idx);
1313   } else if( phase->type(cmp->in(3 - cmp_zero_idx)) == tzero ) {
1314     // The test is inverted, we should invert the result...
1315     x = cmp->in(cmp_zero_idx);
1316     flip = true;
1317   } else {
1318     return NULL;
1319   }
1320 
1321   // Next get the 2 pieces being selected, one is the original value
1322   // and the other is the negated value.
1323   if( phi_root->in(phi_x_idx) != x ) return NULL;
1324 
1325   // Check other phi input for subtract node
1326   Node *sub = phi_root->in(3 - phi_x_idx);
1327 
1328   // Allow only Sub(0,X) and fail out for all others; Neg is not OK
1329   if( tzero == TypeF::ZERO ) {
1330     if( sub->Opcode() != Op_SubF ||
1331         sub->in(2) != x ||
1332         phase->type(sub->in(1)) != tzero ) return NULL;
1333     x = new (phase->C) AbsFNode(x);
1334     if (flip) {
1335       x = new (phase->C) SubFNode(sub->in(1), phase->transform(x));
1336     }
1337   } else {
1338     if( sub->Opcode() != Op_SubD ||
1339         sub->in(2) != x ||
1340         phase->type(sub->in(1)) != tzero ) return NULL;
1341     x = new (phase->C) AbsDNode(x);
1342     if (flip) {
1343       x = new (phase->C) SubDNode(sub->in(1), phase->transform(x));
1344     }
1345   }
1346 
1347   return x;
1348 }
1349 
1350 //------------------------------split_once-------------------------------------
1351 // Helper for split_flow_path
1352 static void split_once(PhaseIterGVN *igvn, Node *phi, Node *val, Node *n, Node *newn) {
1353   igvn->hash_delete(n);         // Remove from hash before hacking edges
1354 
1355   uint j = 1;
1356   for (uint i = phi->req()-1; i > 0; i--) {
1357     if (phi->in(i) == val) {   // Found a path with val?
1358       // Add to NEW Region/Phi, no DU info
1359       newn->set_req( j++, n->in(i) );
1360       // Remove from OLD Region/Phi
1361       n->del_req(i);
1362     }
1363   }
1364 
1365   // Register the new node but do not transform it.  Cannot transform until the
1366   // entire Region/Phi conglomerate has been hacked as a single huge transform.
1367   igvn->register_new_node_with_optimizer( newn );
1368 
1369   // Now I can point to the new node.
1370   n->add_req(newn);
1371   igvn->_worklist.push(n);
1372 }
1373 
1374 //------------------------------split_flow_path--------------------------------
1375 // Check for merging identical values and split flow paths
1376 static Node* split_flow_path(PhaseGVN *phase, PhiNode *phi) {
1377   BasicType bt = phi->type()->basic_type();
1378   if( bt == T_ILLEGAL || type2size[bt] <= 0 )
1379     return NULL;                // Bail out on funny non-value stuff
1380   if( phi->req() <= 3 )         // Need at least 2 matched inputs and a
1381     return NULL;                // third unequal input to be worth doing
1382 
1383   // Scan for a constant
1384   uint i;
1385   for( i = 1; i < phi->req()-1; i++ ) {
1386     Node *n = phi->in(i);
1387     if( !n ) return NULL;
1388     if( phase->type(n) == Type::TOP ) return NULL;
1389     if( n->Opcode() == Op_ConP || n->Opcode() == Op_ConN || n->Opcode() == Op_ConNKlass )
1390       break;
1391   }
1392   if( i >= phi->req() )         // Only split for constants
1393     return NULL;
1394 
1395   Node *val = phi->in(i);       // Constant to split for
1396   uint hit = 0;                 // Number of times it occurs
1397   Node *r = phi->region();
1398 
1399   for( ; i < phi->req(); i++ ){ // Count occurrences of constant
1400     Node *n = phi->in(i);
1401     if( !n ) return NULL;
1402     if( phase->type(n) == Type::TOP ) return NULL;
1403     if( phi->in(i) == val ) {
1404       hit++;
1405       if (PhaseIdealLoop::find_predicate(r->in(i)) != NULL) {
1406         return NULL;            // don't split loop entry path
1407       }
1408     }
1409   }
1410 
1411   if( hit <= 1 ||               // Make sure we find 2 or more
1412       hit == phi->req()-1 )     // and not ALL the same value
1413     return NULL;
1414 
1415   // Now start splitting out the flow paths that merge the same value.
1416   // Split first the RegionNode.
1417   PhaseIterGVN *igvn = phase->is_IterGVN();
1418   RegionNode *newr = new (phase->C) RegionNode(hit+1);
1419   split_once(igvn, phi, val, r, newr);
1420 
1421   // Now split all other Phis than this one
1422   for (DUIterator_Fast kmax, k = r->fast_outs(kmax); k < kmax; k++) {
1423     Node* phi2 = r->fast_out(k);
1424     if( phi2->is_Phi() && phi2->as_Phi() != phi ) {
1425       PhiNode *newphi = PhiNode::make_blank(newr, phi2);
1426       split_once(igvn, phi, val, phi2, newphi);
1427     }
1428   }
1429 
1430   // Clean up this guy
1431   igvn->hash_delete(phi);
1432   for( i = phi->req()-1; i > 0; i-- ) {
1433     if( phi->in(i) == val ) {
1434       phi->del_req(i);
1435     }
1436   }
1437   phi->add_req(val);
1438 
1439   return phi;
1440 }
1441 
1442 //=============================================================================
1443 //------------------------------simple_data_loop_check-------------------------
1444 //  Try to determining if the phi node in a simple safe/unsafe data loop.
1445 //  Returns:
1446 // enum LoopSafety { Safe = 0, Unsafe, UnsafeLoop };
1447 // Safe       - safe case when the phi and it's inputs reference only safe data
1448 //              nodes;
1449 // Unsafe     - the phi and it's inputs reference unsafe data nodes but there
1450 //              is no reference back to the phi - need a graph walk
1451 //              to determine if it is in a loop;
1452 // UnsafeLoop - unsafe case when the phi references itself directly or through
1453 //              unsafe data node.
1454 //  Note: a safe data node is a node which could/never reference itself during
1455 //  GVN transformations. For now it is Con, Proj, Phi, CastPP, CheckCastPP.
1456 //  I mark Phi nodes as safe node not only because they can reference itself
1457 //  but also to prevent mistaking the fallthrough case inside an outer loop
1458 //  as dead loop when the phi references itselfs through an other phi.
1459 PhiNode::LoopSafety PhiNode::simple_data_loop_check(Node *in) const {
1460   // It is unsafe loop if the phi node references itself directly.
1461   if (in == (Node*)this)
1462     return UnsafeLoop; // Unsafe loop
1463   // Unsafe loop if the phi node references itself through an unsafe data node.
1464   // Exclude cases with null inputs or data nodes which could reference
1465   // itself (safe for dead loops).
1466   if (in != NULL && !in->is_dead_loop_safe()) {
1467     // Check inputs of phi's inputs also.
1468     // It is much less expensive then full graph walk.
1469     uint cnt = in->req();
1470     uint i = (in->is_Proj() && !in->is_CFG())  ? 0 : 1;
1471     for (; i < cnt; ++i) {
1472       Node* m = in->in(i);
1473       if (m == (Node*)this)
1474         return UnsafeLoop; // Unsafe loop
1475       if (m != NULL && !m->is_dead_loop_safe()) {
1476         // Check the most common case (about 30% of all cases):
1477         // phi->Load/Store->AddP->(ConP ConP Con)/(Parm Parm Con).
1478         Node *m1 = (m->is_AddP() && m->req() > 3) ? m->in(1) : NULL;
1479         if (m1 == (Node*)this)
1480           return UnsafeLoop; // Unsafe loop
1481         if (m1 != NULL && m1 == m->in(2) &&
1482             m1->is_dead_loop_safe() && m->in(3)->is_Con()) {
1483           continue; // Safe case
1484         }
1485         // The phi references an unsafe node - need full analysis.
1486         return Unsafe;
1487       }
1488     }
1489   }
1490   return Safe; // Safe case - we can optimize the phi node.
1491 }
1492 
1493 //------------------------------is_unsafe_data_reference-----------------------
1494 // If phi can be reached through the data input - it is data loop.
1495 bool PhiNode::is_unsafe_data_reference(Node *in) const {
1496   assert(req() > 1, "");
1497   // First, check simple cases when phi references itself directly or
1498   // through an other node.
1499   LoopSafety safety = simple_data_loop_check(in);
1500   if (safety == UnsafeLoop)
1501     return true;  // phi references itself - unsafe loop
1502   else if (safety == Safe)
1503     return false; // Safe case - phi could be replaced with the unique input.
1504 
1505   // Unsafe case when we should go through data graph to determine
1506   // if the phi references itself.
1507 
1508   ResourceMark rm;
1509 
1510   Arena *a = Thread::current()->resource_area();
1511   Node_List nstack(a);
1512   VectorSet visited(a);
1513 
1514   nstack.push(in); // Start with unique input.
1515   visited.set(in->_idx);
1516   while (nstack.size() != 0) {
1517     Node* n = nstack.pop();
1518     uint cnt = n->req();
1519     uint i = (n->is_Proj() && !n->is_CFG()) ? 0 : 1;
1520     for (; i < cnt; i++) {
1521       Node* m = n->in(i);
1522       if (m == (Node*)this) {
1523         return true;    // Data loop
1524       }
1525       if (m != NULL && !m->is_dead_loop_safe()) { // Only look for unsafe cases.
1526         if (!visited.test_set(m->_idx))
1527           nstack.push(m);
1528       }
1529     }
1530   }
1531   return false; // The phi is not reachable from its inputs
1532 }
1533 
1534 
1535 //------------------------------Ideal------------------------------------------
1536 // Return a node which is more "ideal" than the current node.  Must preserve
1537 // the CFG, but we can still strip out dead paths.
1538 Node *PhiNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1539   // The next should never happen after 6297035 fix.
1540   if( is_copy() )               // Already degraded to a Copy ?
1541     return NULL;                // No change
1542 
1543   Node *r = in(0);              // RegionNode
1544   assert(r->in(0) == NULL || !r->in(0)->is_Root(), "not a specially hidden merge");
1545 
1546   // Note: During parsing, phis are often transformed before their regions.
1547   // This means we have to use type_or_null to defend against untyped regions.
1548   if( phase->type_or_null(r) == Type::TOP ) // Dead code?
1549     return NULL;                // No change
1550 
1551   Node *top = phase->C->top();
1552   bool new_phi = (outcnt() == 0); // transforming new Phi
1553   // No change for igvn if new phi is not hooked
1554   if (new_phi && can_reshape)
1555     return NULL;
1556 
1557   // The are 2 situations when only one valid phi's input is left
1558   // (in addition to Region input).
1559   // One: region is not loop - replace phi with this input.
1560   // Two: region is loop - replace phi with top since this data path is dead
1561   //                       and we need to break the dead data loop.
1562   Node* progress = NULL;        // Record if any progress made
1563   for( uint j = 1; j < req(); ++j ){ // For all paths in
1564     // Check unreachable control paths
1565     Node* rc = r->in(j);
1566     Node* n = in(j);            // Get the input
1567     if (rc == NULL || phase->type(rc) == Type::TOP) {
1568       if (n != top) {           // Not already top?
1569         set_req(j, top);        // Nuke it down
1570         progress = this;        // Record progress
1571       }
1572     }
1573   }
1574 
1575   if (can_reshape && outcnt() == 0) {
1576     // set_req() above may kill outputs if Phi is referenced
1577     // only by itself on the dead (top) control path.
1578     return top;
1579   }
1580 
1581   Node* uin = unique_input(phase);
1582   if (uin == top) {             // Simplest case: no alive inputs.
1583     if (can_reshape)            // IGVN transformation
1584       return top;
1585     else
1586       return NULL;              // Identity will return TOP
1587   } else if (uin != NULL) {
1588     // Only one not-NULL unique input path is left.
1589     // Determine if this input is backedge of a loop.
1590     // (Skip new phis which have no uses and dead regions).
1591     if (outcnt() > 0 && r->in(0) != NULL) {
1592       // First, take the short cut when we know it is a loop and
1593       // the EntryControl data path is dead.
1594       // Loop node may have only one input because entry path
1595       // is removed in PhaseIdealLoop::Dominators().
1596       assert(!r->is_Loop() || r->req() <= 3, "Loop node should have 3 or less inputs");
1597       bool is_loop = (r->is_Loop() && r->req() == 3);
1598       // Then, check if there is a data loop when phi references itself directly
1599       // or through other data nodes.
1600       if (is_loop && !uin->eqv_uncast(in(LoopNode::EntryControl)) ||
1601          !is_loop && is_unsafe_data_reference(uin)) {
1602         // Break this data loop to avoid creation of a dead loop.
1603         if (can_reshape) {
1604           return top;
1605         } else {
1606           // We can't return top if we are in Parse phase - cut inputs only
1607           // let Identity to handle the case.
1608           replace_edge(uin, top);
1609           return NULL;
1610         }
1611       }
1612     }
1613 
1614     // One unique input.
1615     debug_only(Node* ident = Identity(phase));
1616     // The unique input must eventually be detected by the Identity call.
1617 #ifdef ASSERT
1618     if (ident != uin && !ident->is_top()) {
1619       // print this output before failing assert
1620       r->dump(3);
1621       this->dump(3);
1622       ident->dump();
1623       uin->dump();
1624     }
1625 #endif
1626     assert(ident == uin || ident->is_top(), "Identity must clean this up");
1627     return NULL;
1628   }
1629 
1630 
1631   Node* opt = NULL;
1632   int true_path = is_diamond_phi();
1633   if( true_path != 0 ) {
1634     // Check for CMove'ing identity. If it would be unsafe,
1635     // handle it here. In the safe case, let Identity handle it.
1636     Node* unsafe_id = is_cmove_id(phase, true_path);
1637     if( unsafe_id != NULL && is_unsafe_data_reference(unsafe_id) )
1638       opt = unsafe_id;
1639 
1640     // Check for simple convert-to-boolean pattern
1641     if( opt == NULL )
1642       opt = is_x2logic(phase, this, true_path);
1643 
1644     // Check for absolute value
1645     if( opt == NULL )
1646       opt = is_absolute(phase, this, true_path);
1647 
1648     // Check for conditional add
1649     if( opt == NULL && can_reshape )
1650       opt = is_cond_add(phase, this, true_path);
1651 
1652     // These 4 optimizations could subsume the phi:
1653     // have to check for a dead data loop creation.
1654     if( opt != NULL ) {
1655       if( opt == unsafe_id || is_unsafe_data_reference(opt) ) {
1656         // Found dead loop.
1657         if( can_reshape )
1658           return top;
1659         // We can't return top if we are in Parse phase - cut inputs only
1660         // to stop further optimizations for this phi. Identity will return TOP.
1661         assert(req() == 3, "only diamond merge phi here");
1662         set_req(1, top);
1663         set_req(2, top);
1664         return NULL;
1665       } else {
1666         return opt;
1667       }
1668     }
1669   }
1670 
1671   // Check for merging identical values and split flow paths
1672   if (can_reshape) {
1673     opt = split_flow_path(phase, this);
1674     // This optimization only modifies phi - don't need to check for dead loop.
1675     assert(opt == NULL || phase->eqv(opt, this), "do not elide phi");
1676     if (opt != NULL)  return opt;
1677   }
1678 
1679   if (in(1) != NULL && in(1)->Opcode() == Op_AddP && can_reshape) {
1680     // Try to undo Phi of AddP:
1681     // (Phi (AddP base base y) (AddP base2 base2 y))
1682     // becomes:
1683     // newbase := (Phi base base2)
1684     // (AddP newbase newbase y)
1685     //
1686     // This occurs as a result of unsuccessful split_thru_phi and
1687     // interferes with taking advantage of addressing modes. See the
1688     // clone_shift_expressions code in matcher.cpp
1689     Node* addp = in(1);
1690     const Type* type = addp->in(AddPNode::Base)->bottom_type();
1691     Node* y = addp->in(AddPNode::Offset);
1692     if (y != NULL && addp->in(AddPNode::Base) == addp->in(AddPNode::Address)) {
1693       // make sure that all the inputs are similar to the first one,
1694       // i.e. AddP with base == address and same offset as first AddP
1695       bool doit = true;
1696       for (uint i = 2; i < req(); i++) {
1697         if (in(i) == NULL ||
1698             in(i)->Opcode() != Op_AddP ||
1699             in(i)->in(AddPNode::Base) != in(i)->in(AddPNode::Address) ||
1700             in(i)->in(AddPNode::Offset) != y) {
1701           doit = false;
1702           break;
1703         }
1704         // Accumulate type for resulting Phi
1705         type = type->meet(in(i)->in(AddPNode::Base)->bottom_type());
1706       }
1707       Node* base = NULL;
1708       if (doit) {
1709         // Check for neighboring AddP nodes in a tree.
1710         // If they have a base, use that it.
1711         for (DUIterator_Fast kmax, k = this->fast_outs(kmax); k < kmax; k++) {
1712           Node* u = this->fast_out(k);
1713           if (u->is_AddP()) {
1714             Node* base2 = u->in(AddPNode::Base);
1715             if (base2 != NULL && !base2->is_top()) {
1716               if (base == NULL)
1717                 base = base2;
1718               else if (base != base2)
1719                 { doit = false; break; }
1720             }
1721           }
1722         }
1723       }
1724       if (doit) {
1725         if (base == NULL) {
1726           base = new (phase->C) PhiNode(in(0), type, NULL);
1727           for (uint i = 1; i < req(); i++) {
1728             base->init_req(i, in(i)->in(AddPNode::Base));
1729           }
1730           phase->is_IterGVN()->register_new_node_with_optimizer(base);
1731         }
1732         return new (phase->C) AddPNode(base, base, y);
1733       }
1734     }
1735   }
1736 
1737   // Split phis through memory merges, so that the memory merges will go away.
1738   // Piggy-back this transformation on the search for a unique input....
1739   // It will be as if the merged memory is the unique value of the phi.
1740   // (Do not attempt this optimization unless parsing is complete.
1741   // It would make the parser's memory-merge logic sick.)
1742   // (MergeMemNode is not dead_loop_safe - need to check for dead loop.)
1743   if (progress == NULL && can_reshape && type() == Type::MEMORY) {
1744     // see if this phi should be sliced
1745     uint merge_width = 0;
1746     bool saw_self = false;
1747     for( uint i=1; i<req(); ++i ) {// For all paths in
1748       Node *ii = in(i);
1749       if (ii->is_MergeMem()) {
1750         MergeMemNode* n = ii->as_MergeMem();
1751         merge_width = MAX2(merge_width, n->req());
1752         saw_self = saw_self || phase->eqv(n->base_memory(), this);
1753       }
1754     }
1755 
1756     // This restriction is temporarily necessary to ensure termination:
1757     if (!saw_self && adr_type() == TypePtr::BOTTOM)  merge_width = 0;
1758 
1759     if (merge_width > Compile::AliasIdxRaw) {
1760       // found at least one non-empty MergeMem
1761       const TypePtr* at = adr_type();
1762       if (at != TypePtr::BOTTOM) {
1763         // Patch the existing phi to select an input from the merge:
1764         // Phi:AT1(...MergeMem(m0, m1, m2)...) into
1765         //     Phi:AT1(...m1...)
1766         int alias_idx = phase->C->get_alias_index(at);
1767         for (uint i=1; i<req(); ++i) {
1768           Node *ii = in(i);
1769           if (ii->is_MergeMem()) {
1770             MergeMemNode* n = ii->as_MergeMem();
1771             // compress paths and change unreachable cycles to TOP
1772             // If not, we can update the input infinitely along a MergeMem cycle
1773             // Equivalent code is in MemNode::Ideal_common
1774             Node *m  = phase->transform(n);
1775             if (outcnt() == 0) {  // Above transform() may kill us!
1776               return top;
1777             }
1778             // If transformed to a MergeMem, get the desired slice
1779             // Otherwise the returned node represents memory for every slice
1780             Node *new_mem = (m->is_MergeMem()) ?
1781                              m->as_MergeMem()->memory_at(alias_idx) : m;
1782             // Update input if it is progress over what we have now
1783             if (new_mem != ii) {
1784               set_req(i, new_mem);
1785               progress = this;
1786             }
1787           }
1788         }
1789       } else {
1790         // We know that at least one MergeMem->base_memory() == this
1791         // (saw_self == true). If all other inputs also references this phi
1792         // (directly or through data nodes) - it is dead loop.
1793         bool saw_safe_input = false;
1794         for (uint j = 1; j < req(); ++j) {
1795           Node *n = in(j);
1796           if (n->is_MergeMem() && n->as_MergeMem()->base_memory() == this)
1797             continue;              // skip known cases
1798           if (!is_unsafe_data_reference(n)) {
1799             saw_safe_input = true; // found safe input
1800             break;
1801           }
1802         }
1803         if (!saw_safe_input)
1804           return top; // all inputs reference back to this phi - dead loop
1805 
1806         // Phi(...MergeMem(m0, m1:AT1, m2:AT2)...) into
1807         //     MergeMem(Phi(...m0...), Phi:AT1(...m1...), Phi:AT2(...m2...))
1808         PhaseIterGVN *igvn = phase->is_IterGVN();
1809         Node* hook = new (phase->C) Node(1);
1810         PhiNode* new_base = (PhiNode*) clone();
1811         // Must eagerly register phis, since they participate in loops.
1812         if (igvn) {
1813           igvn->register_new_node_with_optimizer(new_base);
1814           hook->add_req(new_base);
1815         }
1816         MergeMemNode* result = MergeMemNode::make(phase->C, new_base);
1817         for (uint i = 1; i < req(); ++i) {
1818           Node *ii = in(i);
1819           if (ii->is_MergeMem()) {
1820             MergeMemNode* n = ii->as_MergeMem();
1821             for (MergeMemStream mms(result, n); mms.next_non_empty2(); ) {
1822               // If we have not seen this slice yet, make a phi for it.
1823               bool made_new_phi = false;
1824               if (mms.is_empty()) {
1825                 Node* new_phi = new_base->slice_memory(mms.adr_type(phase->C));
1826                 made_new_phi = true;
1827                 if (igvn) {
1828                   igvn->register_new_node_with_optimizer(new_phi);
1829                   hook->add_req(new_phi);
1830                 }
1831                 mms.set_memory(new_phi);
1832               }
1833               Node* phi = mms.memory();
1834               assert(made_new_phi || phi->in(i) == n, "replace the i-th merge by a slice");
1835               phi->set_req(i, mms.memory2());
1836             }
1837           }
1838         }
1839         // Distribute all self-loops.
1840         { // (Extra braces to hide mms.)
1841           for (MergeMemStream mms(result); mms.next_non_empty(); ) {
1842             Node* phi = mms.memory();
1843             for (uint i = 1; i < req(); ++i) {
1844               if (phi->in(i) == this)  phi->set_req(i, phi);
1845             }
1846           }
1847         }
1848         // now transform the new nodes, and return the mergemem
1849         for (MergeMemStream mms(result); mms.next_non_empty(); ) {
1850           Node* phi = mms.memory();
1851           mms.set_memory(phase->transform(phi));
1852         }
1853         if (igvn) { // Unhook.
1854           igvn->hash_delete(hook);
1855           for (uint i = 1; i < hook->req(); i++) {
1856             hook->set_req(i, NULL);
1857           }
1858         }
1859         // Replace self with the result.
1860         return result;
1861       }
1862     }
1863     //
1864     // Other optimizations on the memory chain
1865     //
1866     const TypePtr* at = adr_type();
1867     for( uint i=1; i<req(); ++i ) {// For all paths in
1868       Node *ii = in(i);
1869       Node *new_in = MemNode::optimize_memory_chain(ii, at, phase);
1870       if (ii != new_in ) {
1871         set_req(i, new_in);
1872         progress = this;
1873       }
1874     }
1875   }
1876 
1877 #ifdef _LP64
1878   // Push DecodeN/DecodeNKlass down through phi.
1879   // The rest of phi graph will transform by split EncodeP node though phis up.
1880   if ((UseCompressedOops || UseCompressedKlassPointers) && can_reshape && progress == NULL) {
1881     bool may_push = true;
1882     bool has_decodeN = false;
1883     bool is_decodeN = false;
1884     for (uint i=1; i<req(); ++i) {// For all paths in
1885       Node *ii = in(i);
1886       if (ii->is_DecodeNarrowPtr() && ii->bottom_type() == bottom_type()) {
1887         // Do optimization if a non dead path exist.
1888         if (ii->in(1)->bottom_type() != Type::TOP) {
1889           has_decodeN = true;
1890           is_decodeN = ii->is_DecodeN();
1891         }
1892       } else if (!ii->is_Phi()) {
1893         may_push = false;
1894       }
1895     }
1896 
1897     if (has_decodeN && may_push) {
1898       PhaseIterGVN *igvn = phase->is_IterGVN();
1899       // Make narrow type for new phi.
1900       const Type* narrow_t;
1901       if (is_decodeN) {
1902         narrow_t = TypeNarrowOop::make(this->bottom_type()->is_ptr());
1903       } else {
1904         narrow_t = TypeNarrowKlass::make(this->bottom_type()->is_ptr());
1905       }
1906       PhiNode* new_phi = new (phase->C) PhiNode(r, narrow_t);
1907       uint orig_cnt = req();
1908       for (uint i=1; i<req(); ++i) {// For all paths in
1909         Node *ii = in(i);
1910         Node* new_ii = NULL;
1911         if (ii->is_DecodeNarrowPtr()) {
1912           assert(ii->bottom_type() == bottom_type(), "sanity");
1913           new_ii = ii->in(1);
1914         } else {
1915           assert(ii->is_Phi(), "sanity");
1916           if (ii->as_Phi() == this) {
1917             new_ii = new_phi;
1918           } else {
1919             if (is_decodeN) {
1920               new_ii = new (phase->C) EncodePNode(ii, narrow_t);
1921             } else {
1922               new_ii = new (phase->C) EncodePKlassNode(ii, narrow_t);
1923             }
1924             igvn->register_new_node_with_optimizer(new_ii);
1925           }
1926         }
1927         new_phi->set_req(i, new_ii);
1928       }
1929       igvn->register_new_node_with_optimizer(new_phi, this);
1930       if (is_decodeN) {
1931         progress = new (phase->C) DecodeNNode(new_phi, bottom_type());
1932       } else {
1933         progress = new (phase->C) DecodeNKlassNode(new_phi, bottom_type());
1934       }
1935     }
1936   }
1937 #endif
1938 
1939   return progress;              // Return any progress
1940 }
1941 
1942 //------------------------------is_tripcount-----------------------------------
1943 bool PhiNode::is_tripcount() const {
1944   return (in(0) != NULL && in(0)->is_CountedLoop() &&
1945           in(0)->as_CountedLoop()->phi() == this);
1946 }
1947 
1948 //------------------------------out_RegMask------------------------------------
1949 const RegMask &PhiNode::in_RegMask(uint i) const {
1950   return i ? out_RegMask() : RegMask::Empty;
1951 }
1952 
1953 const RegMask &PhiNode::out_RegMask() const {
1954   uint ideal_reg = _type->ideal_reg();
1955   assert( ideal_reg != Node::NotAMachineReg, "invalid type at Phi" );
1956   if( ideal_reg == 0 ) return RegMask::Empty;
1957   return *(Compile::current()->matcher()->idealreg2spillmask[ideal_reg]);
1958 }
1959 
1960 #ifndef PRODUCT
1961 void PhiNode::dump_spec(outputStream *st) const {
1962   TypeNode::dump_spec(st);
1963   if (is_tripcount()) {
1964     st->print(" #tripcount");
1965   }
1966 }
1967 #endif
1968 
1969 
1970 //=============================================================================
1971 const Type *GotoNode::Value( PhaseTransform *phase ) const {
1972   // If the input is reachable, then we are executed.
1973   // If the input is not reachable, then we are not executed.
1974   return phase->type(in(0));
1975 }
1976 
1977 Node *GotoNode::Identity( PhaseTransform *phase ) {
1978   return in(0);                // Simple copy of incoming control
1979 }
1980 
1981 const RegMask &GotoNode::out_RegMask() const {
1982   return RegMask::Empty;
1983 }
1984 
1985 //=============================================================================
1986 const RegMask &JumpNode::out_RegMask() const {
1987   return RegMask::Empty;
1988 }
1989 
1990 //=============================================================================
1991 const RegMask &JProjNode::out_RegMask() const {
1992   return RegMask::Empty;
1993 }
1994 
1995 //=============================================================================
1996 const RegMask &CProjNode::out_RegMask() const {
1997   return RegMask::Empty;
1998 }
1999 
2000 
2001 
2002 //=============================================================================
2003 
2004 uint PCTableNode::hash() const { return Node::hash() + _size; }
2005 uint PCTableNode::cmp( const Node &n ) const
2006 { return _size == ((PCTableNode&)n)._size; }
2007 
2008 const Type *PCTableNode::bottom_type() const {
2009   const Type** f = TypeTuple::fields(_size);
2010   for( uint i = 0; i < _size; i++ ) f[i] = Type::CONTROL;
2011   return TypeTuple::make(_size, f);
2012 }
2013 
2014 //------------------------------Value------------------------------------------
2015 // Compute the type of the PCTableNode.  If reachable it is a tuple of
2016 // Control, otherwise the table targets are not reachable
2017 const Type *PCTableNode::Value( PhaseTransform *phase ) const {
2018   if( phase->type(in(0)) == Type::CONTROL )
2019     return bottom_type();
2020   return Type::TOP;             // All paths dead?  Then so are we
2021 }
2022 
2023 //------------------------------Ideal------------------------------------------
2024 // Return a node which is more "ideal" than the current node.  Strip out
2025 // control copies
2026 Node *PCTableNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2027   return remove_dead_region(phase, can_reshape) ? this : NULL;
2028 }
2029 
2030 //=============================================================================
2031 uint JumpProjNode::hash() const {
2032   return Node::hash() + _dest_bci;
2033 }
2034 
2035 uint JumpProjNode::cmp( const Node &n ) const {
2036   return ProjNode::cmp(n) &&
2037     _dest_bci == ((JumpProjNode&)n)._dest_bci;
2038 }
2039 
2040 #ifndef PRODUCT
2041 void JumpProjNode::dump_spec(outputStream *st) const {
2042   ProjNode::dump_spec(st);
2043    st->print("@bci %d ",_dest_bci);
2044 }
2045 #endif
2046 
2047 //=============================================================================
2048 //------------------------------Value------------------------------------------
2049 // Check for being unreachable, or for coming from a Rethrow.  Rethrow's cannot
2050 // have the default "fall_through_index" path.
2051 const Type *CatchNode::Value( PhaseTransform *phase ) const {
2052   // Unreachable?  Then so are all paths from here.
2053   if( phase->type(in(0)) == Type::TOP ) return Type::TOP;
2054   // First assume all paths are reachable
2055   const Type** f = TypeTuple::fields(_size);
2056   for( uint i = 0; i < _size; i++ ) f[i] = Type::CONTROL;
2057   // Identify cases that will always throw an exception
2058   // () rethrow call
2059   // () virtual or interface call with NULL receiver
2060   // () call is a check cast with incompatible arguments
2061   if( in(1)->is_Proj() ) {
2062     Node *i10 = in(1)->in(0);
2063     if( i10->is_Call() ) {
2064       CallNode *call = i10->as_Call();
2065       // Rethrows always throw exceptions, never return
2066       if (call->entry_point() == OptoRuntime::rethrow_stub()) {
2067         f[CatchProjNode::fall_through_index] = Type::TOP;
2068       } else if( call->req() > TypeFunc::Parms ) {
2069         const Type *arg0 = phase->type( call->in(TypeFunc::Parms) );
2070         // Check for null receiver to virtual or interface calls
2071         if( call->is_CallDynamicJava() &&
2072             arg0->higher_equal(TypePtr::NULL_PTR) ) {
2073           f[CatchProjNode::fall_through_index] = Type::TOP;
2074         }
2075       } // End of if not a runtime stub
2076     } // End of if have call above me
2077   } // End of slot 1 is not a projection
2078   return TypeTuple::make(_size, f);
2079 }
2080 
2081 //=============================================================================
2082 uint CatchProjNode::hash() const {
2083   return Node::hash() + _handler_bci;
2084 }
2085 
2086 
2087 uint CatchProjNode::cmp( const Node &n ) const {
2088   return ProjNode::cmp(n) &&
2089     _handler_bci == ((CatchProjNode&)n)._handler_bci;
2090 }
2091 
2092 
2093 //------------------------------Identity---------------------------------------
2094 // If only 1 target is possible, choose it if it is the main control
2095 Node *CatchProjNode::Identity( PhaseTransform *phase ) {
2096   // If my value is control and no other value is, then treat as ID
2097   const TypeTuple *t = phase->type(in(0))->is_tuple();
2098   if (t->field_at(_con) != Type::CONTROL)  return this;
2099   // If we remove the last CatchProj and elide the Catch/CatchProj, then we
2100   // also remove any exception table entry.  Thus we must know the call
2101   // feeding the Catch will not really throw an exception.  This is ok for
2102   // the main fall-thru control (happens when we know a call can never throw
2103   // an exception) or for "rethrow", because a further optimization will
2104   // yank the rethrow (happens when we inline a function that can throw an
2105   // exception and the caller has no handler).  Not legal, e.g., for passing
2106   // a NULL receiver to a v-call, or passing bad types to a slow-check-cast.
2107   // These cases MUST throw an exception via the runtime system, so the VM
2108   // will be looking for a table entry.
2109   Node *proj = in(0)->in(1);    // Expect a proj feeding CatchNode
2110   CallNode *call;
2111   if (_con != TypeFunc::Control && // Bail out if not the main control.
2112       !(proj->is_Proj() &&      // AND NOT a rethrow
2113         proj->in(0)->is_Call() &&
2114         (call = proj->in(0)->as_Call()) &&
2115         call->entry_point() == OptoRuntime::rethrow_stub()))
2116     return this;
2117 
2118   // Search for any other path being control
2119   for (uint i = 0; i < t->cnt(); i++) {
2120     if (i != _con && t->field_at(i) == Type::CONTROL)
2121       return this;
2122   }
2123   // Only my path is possible; I am identity on control to the jump
2124   return in(0)->in(0);
2125 }
2126 
2127 
2128 #ifndef PRODUCT
2129 void CatchProjNode::dump_spec(outputStream *st) const {
2130   ProjNode::dump_spec(st);
2131   st->print("@bci %d ",_handler_bci);
2132 }
2133 #endif
2134 
2135 //=============================================================================
2136 //------------------------------Identity---------------------------------------
2137 // Check for CreateEx being Identity.
2138 Node *CreateExNode::Identity( PhaseTransform *phase ) {
2139   if( phase->type(in(1)) == Type::TOP ) return in(1);
2140   if( phase->type(in(0)) == Type::TOP ) return in(0);
2141   // We only come from CatchProj, unless the CatchProj goes away.
2142   // If the CatchProj is optimized away, then we just carry the
2143   // exception oop through.
2144   CallNode *call = in(1)->in(0)->as_Call();
2145 
2146   return ( in(0)->is_CatchProj() && in(0)->in(0)->in(1) == in(1) )
2147     ? this
2148     : call->in(TypeFunc::Parms);
2149 }
2150 
2151 //=============================================================================
2152 //------------------------------Value------------------------------------------
2153 // Check for being unreachable.
2154 const Type *NeverBranchNode::Value( PhaseTransform *phase ) const {
2155   if (!in(0) || in(0)->is_top()) return Type::TOP;
2156   return bottom_type();
2157 }
2158 
2159 //------------------------------Ideal------------------------------------------
2160 // Check for no longer being part of a loop
2161 Node *NeverBranchNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2162   if (can_reshape && !in(0)->is_Loop()) {
2163     // Dead code elimination can sometimes delete this projection so
2164     // if it's not there, there's nothing to do.
2165     Node* fallthru = proj_out(0);
2166     if (fallthru != NULL) {
2167       phase->is_IterGVN()->replace_node(fallthru, in(0));
2168     }
2169     return phase->C->top();
2170   }
2171   return NULL;
2172 }
2173 
2174 #ifndef PRODUCT
2175 void NeverBranchNode::format( PhaseRegAlloc *ra_, outputStream *st) const {
2176   st->print("%s", Name());
2177 }
2178 #endif