1 /*
   2  * Copyright (c) 1998, 2018, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "ci/ciMethodData.hpp"
  27 #include "compiler/compileLog.hpp"
  28 #include "gc/shared/barrierSet.hpp"
  29 #include "gc/shared/c2/barrierSetC2.hpp"
  30 #include "libadt/vectset.hpp"
  31 #include "memory/allocation.inline.hpp"
  32 #include "memory/resourceArea.hpp"
  33 #include "opto/addnode.hpp"
  34 #include "opto/callnode.hpp"
  35 #include "opto/connode.hpp"
  36 #include "opto/convertnode.hpp"
  37 #include "opto/divnode.hpp"
  38 #include "opto/idealGraphPrinter.hpp"
  39 #include "opto/loopnode.hpp"
  40 #include "opto/mulnode.hpp"
  41 #include "opto/rootnode.hpp"
  42 #include "opto/superword.hpp"
  43 #include "utilities/macros.hpp"
  44 #if INCLUDE_SHENANDOAHGC
  45 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
  46 #endif
  47 
  48 //=============================================================================
  49 //------------------------------is_loop_iv-------------------------------------
  50 // Determine if a node is Counted loop induction variable.
  51 // The method is declared in node.hpp.
  52 const Node* Node::is_loop_iv() const {
  53   if (this->is_Phi() && !this->as_Phi()->is_copy() &&
  54       this->as_Phi()->region()->is_CountedLoop() &&
  55       this->as_Phi()->region()->as_CountedLoop()->phi() == this) {
  56     return this;
  57   } else {
  58     return NULL;
  59   }
  60 }
  61 
  62 //=============================================================================
  63 //------------------------------dump_spec--------------------------------------
  64 // Dump special per-node info
  65 #ifndef PRODUCT
  66 void LoopNode::dump_spec(outputStream *st) const {
  67   if (is_inner_loop()) st->print( "inner " );
  68   if (is_partial_peel_loop()) st->print( "partial_peel " );
  69   if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
  70 }
  71 #endif
  72 
  73 //------------------------------is_valid_counted_loop-------------------------
  74 bool LoopNode::is_valid_counted_loop() const {
  75   if (is_CountedLoop()) {
  76     CountedLoopNode*    l  = as_CountedLoop();
  77     CountedLoopEndNode* le = l->loopexit_or_null();
  78     if (le != NULL &&
  79         le->proj_out_or_null(1 /* true */) == l->in(LoopNode::LoopBackControl)) {
  80       Node* phi  = l->phi();
  81       Node* exit = le->proj_out_or_null(0 /* false */);
  82       if (exit != NULL && exit->Opcode() == Op_IfFalse &&
  83           phi != NULL && phi->is_Phi() &&
  84           phi->in(LoopNode::LoopBackControl) == l->incr() &&
  85           le->loopnode() == l && le->stride_is_con()) {
  86         return true;
  87       }
  88     }
  89   }
  90   return false;
  91 }
  92 
  93 //------------------------------get_early_ctrl---------------------------------
  94 // Compute earliest legal control
  95 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
  96   assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
  97   uint i;
  98   Node *early;
  99   if (n->in(0) && !n->is_expensive()) {
 100     early = n->in(0);
 101     if (!early->is_CFG()) // Might be a non-CFG multi-def
 102       early = get_ctrl(early);        // So treat input as a straight data input
 103     i = 1;
 104   } else {
 105     early = get_ctrl(n->in(1));
 106     i = 2;
 107   }
 108   uint e_d = dom_depth(early);
 109   assert( early, "" );
 110   for (; i < n->req(); i++) {
 111     Node *cin = get_ctrl(n->in(i));
 112     assert( cin, "" );
 113     // Keep deepest dominator depth
 114     uint c_d = dom_depth(cin);
 115     if (c_d > e_d) {           // Deeper guy?
 116       early = cin;              // Keep deepest found so far
 117       e_d = c_d;
 118     } else if (c_d == e_d &&    // Same depth?
 119                early != cin) { // If not equal, must use slower algorithm
 120       // If same depth but not equal, one _must_ dominate the other
 121       // and we want the deeper (i.e., dominated) guy.
 122       Node *n1 = early;
 123       Node *n2 = cin;
 124       while (1) {
 125         n1 = idom(n1);          // Walk up until break cycle
 126         n2 = idom(n2);
 127         if (n1 == cin ||        // Walked early up to cin
 128             dom_depth(n2) < c_d)
 129           break;                // early is deeper; keep him
 130         if (n2 == early ||      // Walked cin up to early
 131             dom_depth(n1) < c_d) {
 132           early = cin;          // cin is deeper; keep him
 133           break;
 134         }
 135       }
 136       e_d = dom_depth(early);   // Reset depth register cache
 137     }
 138   }
 139 
 140   // Return earliest legal location
 141   assert(early == find_non_split_ctrl(early), "unexpected early control");
 142 
 143   if (n->is_expensive() && !_verify_only && !_verify_me) {
 144     assert(n->in(0), "should have control input");
 145     early = get_early_ctrl_for_expensive(n, early);
 146   }
 147 
 148   return early;
 149 }
 150 
 151 //------------------------------get_early_ctrl_for_expensive---------------------------------
 152 // Move node up the dominator tree as high as legal while still beneficial
 153 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {
 154   assert(n->in(0) && n->is_expensive(), "expensive node with control input here");
 155   assert(OptimizeExpensiveOps, "optimization off?");
 156 
 157   Node* ctl = n->in(0);
 158   assert(ctl->is_CFG(), "expensive input 0 must be cfg");
 159   uint min_dom_depth = dom_depth(earliest);
 160 #ifdef ASSERT
 161   if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) {
 162     dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl);
 163     assert(false, "Bad graph detected in get_early_ctrl_for_expensive");
 164   }
 165 #endif
 166   if (dom_depth(ctl) < min_dom_depth) {
 167     return earliest;
 168   }
 169 
 170   while (1) {
 171     Node *next = ctl;
 172     // Moving the node out of a loop on the projection of a If
 173     // confuses loop predication. So once we hit a Loop in a If branch
 174     // that doesn't branch to an UNC, we stop. The code that process
 175     // expensive nodes will notice the loop and skip over it to try to
 176     // move the node further up.
 177     if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) {
 178       if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
 179         break;
 180       }
 181       next = idom(ctl->in(1)->in(0));
 182     } else if (ctl->is_Proj()) {
 183       // We only move it up along a projection if the projection is
 184       // the single control projection for its parent: same code path,
 185       // if it's a If with UNC or fallthrough of a call.
 186       Node* parent_ctl = ctl->in(0);
 187       if (parent_ctl == NULL) {
 188         break;
 189       } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) {
 190         next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control();
 191       } else if (parent_ctl->is_If()) {
 192         if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
 193           break;
 194         }
 195         assert(idom(ctl) == parent_ctl, "strange");
 196         next = idom(parent_ctl);
 197       } else if (ctl->is_CatchProj()) {
 198         if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) {
 199           break;
 200         }
 201         assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph");
 202         next = parent_ctl->in(0)->in(0)->in(0);
 203       } else {
 204         // Check if parent control has a single projection (this
 205         // control is the only possible successor of the parent
 206         // control). If so, we can try to move the node above the
 207         // parent control.
 208         int nb_ctl_proj = 0;
 209         for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) {
 210           Node *p = parent_ctl->fast_out(i);
 211           if (p->is_Proj() && p->is_CFG()) {
 212             nb_ctl_proj++;
 213             if (nb_ctl_proj > 1) {
 214               break;
 215             }
 216           }
 217         }
 218 
 219         if (nb_ctl_proj > 1) {
 220           break;
 221         }
 222         assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call() ||
 223                BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(parent_ctl), "unexpected node");
 224         assert(idom(ctl) == parent_ctl, "strange");
 225         next = idom(parent_ctl);
 226       }
 227     } else {
 228       next = idom(ctl);
 229     }
 230     if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) {
 231       break;
 232     }
 233     ctl = next;
 234   }
 235 
 236   if (ctl != n->in(0)) {
 237     _igvn.replace_input_of(n, 0, ctl);
 238     _igvn.hash_insert(n);
 239   }
 240 
 241   return ctl;
 242 }
 243 
 244 
 245 //------------------------------set_early_ctrl---------------------------------
 246 // Set earliest legal control
 247 void PhaseIdealLoop::set_early_ctrl( Node *n ) {
 248   Node *early = get_early_ctrl(n);
 249 
 250   // Record earliest legal location
 251   set_ctrl(n, early);
 252 }
 253 
 254 //------------------------------set_subtree_ctrl-------------------------------
 255 // set missing _ctrl entries on new nodes
 256 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {
 257   // Already set?  Get out.
 258   if( _nodes[n->_idx] ) return;
 259   // Recursively set _nodes array to indicate where the Node goes
 260   uint i;
 261   for( i = 0; i < n->req(); ++i ) {
 262     Node *m = n->in(i);
 263     if( m && m != C->root() )
 264       set_subtree_ctrl( m );
 265   }
 266 
 267   // Fixup self
 268   set_early_ctrl( n );
 269 }
 270 
 271 // Create a skeleton strip mined outer loop: a Loop head before the
 272 // inner strip mined loop, a safepoint and an exit condition guarded
 273 // by an opaque node after the inner strip mined loop with a backedge
 274 // to the loop head. The inner strip mined loop is left as it is. Only
 275 // once loop optimizations are over, do we adjust the inner loop exit
 276 // condition to limit its number of iterations, set the outer loop
 277 // exit condition and add Phis to the outer loop head. Some loop
 278 // optimizations that operate on the inner strip mined loop need to be
 279 // aware of the outer strip mined loop: loop unswitching needs to
 280 // clone the outer loop as well as the inner, unrolling needs to only
 281 // clone the inner loop etc. No optimizations need to change the outer
 282 // strip mined loop as it is only a skeleton.
 283 IdealLoopTree* PhaseIdealLoop::create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
 284                                                              IdealLoopTree* loop, float cl_prob, float le_fcnt,
 285                                                              Node*& entry_control, Node*& iffalse) {
 286   Node* outer_test = _igvn.intcon(0);
 287   set_ctrl(outer_test, C->root());
 288   Node *orig = iffalse;
 289   iffalse = iffalse->clone();
 290   _igvn.register_new_node_with_optimizer(iffalse);
 291   set_idom(iffalse, idom(orig), dom_depth(orig));
 292 
 293   IfNode *outer_le = new OuterStripMinedLoopEndNode(iffalse, outer_test, cl_prob, le_fcnt);
 294   Node *outer_ift = new IfTrueNode (outer_le);
 295   Node* outer_iff = orig;
 296   _igvn.replace_input_of(outer_iff, 0, outer_le);
 297 
 298   LoopNode *outer_l = new OuterStripMinedLoopNode(C, init_control, outer_ift);
 299   entry_control = outer_l;
 300 
 301   IdealLoopTree* outer_ilt = new IdealLoopTree(this, outer_l, outer_ift);
 302   IdealLoopTree* parent = loop->_parent;
 303   IdealLoopTree* sibling = parent->_child;
 304   if (sibling == loop) {
 305     parent->_child = outer_ilt;
 306   } else {
 307     while (sibling->_next != loop) {
 308       sibling = sibling->_next;
 309     }
 310     sibling->_next = outer_ilt;
 311   }
 312   outer_ilt->_next = loop->_next;
 313   outer_ilt->_parent = parent;
 314   outer_ilt->_child = loop;
 315   outer_ilt->_nest = loop->_nest;
 316   loop->_parent = outer_ilt;
 317   loop->_next = NULL;
 318   loop->_nest++;
 319 
 320   set_loop(iffalse, outer_ilt);
 321   register_control(outer_le, outer_ilt, iffalse);
 322   register_control(outer_ift, outer_ilt, outer_le);
 323   set_idom(outer_iff, outer_le, dom_depth(outer_le));
 324   _igvn.register_new_node_with_optimizer(outer_l);
 325   set_loop(outer_l, outer_ilt);
 326   set_idom(outer_l, init_control, dom_depth(init_control)+1);
 327 
 328   return outer_ilt;
 329 }
 330 
 331 void PhaseIdealLoop::insert_loop_limit_check(ProjNode* limit_check_proj, Node* cmp_limit, Node* bol) {
 332   Node* new_predicate_proj = create_new_if_for_predicate(limit_check_proj, NULL,
 333                                                          Deoptimization::Reason_loop_limit_check,
 334                                                          Op_If);
 335   Node* iff = new_predicate_proj->in(0);
 336   assert(iff->Opcode() == Op_If, "bad graph shape");
 337   Node* conv = iff->in(1);
 338   assert(conv->Opcode() == Op_Conv2B, "bad graph shape");
 339   Node* opaq = conv->in(1);
 340   assert(opaq->Opcode() == Op_Opaque1, "bad graph shape");
 341   cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
 342   bol = _igvn.register_new_node_with_optimizer(bol);
 343   set_subtree_ctrl(bol);
 344   _igvn.replace_input_of(iff, 1, bol);
 345 
 346 #ifndef PRODUCT
 347   // report that the loop predication has been actually performed
 348   // for this loop
 349   if (TraceLoopLimitCheck) {
 350     tty->print_cr("Counted Loop Limit Check generated:");
 351     debug_only( bol->dump(2); )
 352   }
 353 #endif
 354 }
 355 
 356 //------------------------------is_counted_loop--------------------------------
 357 bool PhaseIdealLoop::is_counted_loop(Node* x, IdealLoopTree*& loop) {
 358   PhaseGVN *gvn = &_igvn;
 359 
 360   // Counted loop head must be a good RegionNode with only 3 not NULL
 361   // control input edges: Self, Entry, LoopBack.
 362   if (x->in(LoopNode::Self) == NULL || x->req() != 3 || loop->_irreducible) {
 363     return false;
 364   }
 365   Node *init_control = x->in(LoopNode::EntryControl);
 366   Node *back_control = x->in(LoopNode::LoopBackControl);
 367   if (init_control == NULL || back_control == NULL)    // Partially dead
 368     return false;
 369   // Must also check for TOP when looking for a dead loop
 370   if (init_control->is_top() || back_control->is_top())
 371     return false;
 372 
 373   // Allow funny placement of Safepoint
 374   if (back_control->Opcode() == Op_SafePoint) {
 375     if (LoopStripMiningIter != 0) {
 376       // Leaving the safepoint on the backedge and creating a
 377       // CountedLoop will confuse optimizations. We can't move the
 378       // safepoint around because its jvm state wouldn't match a new
 379       // location. Give up on that loop.
 380       return false;
 381     }
 382     back_control = back_control->in(TypeFunc::Control);
 383   }
 384 
 385   // Controlling test for loop
 386   Node *iftrue = back_control;
 387   uint iftrue_op = iftrue->Opcode();
 388   if (iftrue_op != Op_IfTrue &&
 389       iftrue_op != Op_IfFalse)
 390     // I have a weird back-control.  Probably the loop-exit test is in
 391     // the middle of the loop and I am looking at some trailing control-flow
 392     // merge point.  To fix this I would have to partially peel the loop.
 393     return false; // Obscure back-control
 394 
 395   // Get boolean guarding loop-back test
 396   Node *iff = iftrue->in(0);
 397   if (get_loop(iff) != loop || !iff->in(1)->is_Bool())
 398     return false;
 399   BoolNode *test = iff->in(1)->as_Bool();
 400   BoolTest::mask bt = test->_test._test;
 401   float cl_prob = iff->as_If()->_prob;
 402   if (iftrue_op == Op_IfFalse) {
 403     bt = BoolTest(bt).negate();
 404     cl_prob = 1.0 - cl_prob;
 405   }
 406   // Get backedge compare
 407   Node *cmp = test->in(1);
 408   int cmp_op = cmp->Opcode();
 409   if (cmp_op != Op_CmpI)
 410     return false;                // Avoid pointer & float compares
 411 
 412   // Find the trip-counter increment & limit.  Limit must be loop invariant.
 413   Node *incr  = cmp->in(1);
 414   Node *limit = cmp->in(2);
 415 
 416   // ---------
 417   // need 'loop()' test to tell if limit is loop invariant
 418   // ---------
 419 
 420   if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
 421     Node *tmp = incr;            // Then reverse order into the CmpI
 422     incr = limit;
 423     limit = tmp;
 424     bt = BoolTest(bt).commute(); // And commute the exit test
 425   }
 426   if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant
 427     return false;
 428   if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
 429     return false;
 430 
 431   Node* phi_incr = NULL;
 432   // Trip-counter increment must be commutative & associative.
 433   if (incr->Opcode() == Op_CastII) {
 434     incr = incr->in(1);
 435   }
 436   if (incr->is_Phi()) {
 437     if (incr->as_Phi()->region() != x || incr->req() != 3)
 438       return false; // Not simple trip counter expression
 439     phi_incr = incr;
 440     incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
 441     if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
 442       return false;
 443   }
 444 
 445   Node* trunc1 = NULL;
 446   Node* trunc2 = NULL;
 447   const TypeInt* iv_trunc_t = NULL;
 448   Node* orig_incr = incr;
 449   if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {
 450     return false; // Funny increment opcode
 451   }
 452   assert(incr->Opcode() == Op_AddI, "wrong increment code");
 453 
 454   const TypeInt* limit_t = gvn->type(limit)->is_int();
 455   if (trunc1 != NULL) {
 456     // When there is a truncation, we must be sure that after the truncation
 457     // the trip counter will end up higher than the limit, otherwise we are looking
 458     // at an endless loop. Can happen with range checks.
 459 
 460     // Example:
 461     // int i = 0;
 462     // while (true)
 463     //    sum + = array[i];
 464     //    i++;
 465     //    i = i && 0x7fff;
 466     //  }
 467     //
 468     // If the array is shorter than 0x8000 this exits through a AIOOB
 469     //  - Counted loop transformation is ok
 470     // If the array is longer then this is an endless loop
 471     //  - No transformation can be done.
 472 
 473     const TypeInt* incr_t = gvn->type(orig_incr)->is_int();
 474     if (limit_t->_hi > incr_t->_hi) {
 475       // if the limit can have a higher value than the increment (before the phi)
 476       return false;
 477     }
 478   }
 479 
 480   // Get merge point
 481   Node *xphi = incr->in(1);
 482   Node *stride = incr->in(2);
 483   if (!stride->is_Con()) {     // Oops, swap these
 484     if (!xphi->is_Con())       // Is the other guy a constant?
 485       return false;             // Nope, unknown stride, bail out
 486     Node *tmp = xphi;           // 'incr' is commutative, so ok to swap
 487     xphi = stride;
 488     stride = tmp;
 489   }
 490   if (xphi->Opcode() == Op_CastII) {
 491     xphi = xphi->in(1);
 492   }
 493   // Stride must be constant
 494   int stride_con = stride->get_int();
 495   if (stride_con == 0)
 496     return false; // missed some peephole opt
 497 
 498   if (!xphi->is_Phi())
 499     return false; // Too much math on the trip counter
 500   if (phi_incr != NULL && phi_incr != xphi)
 501     return false;
 502   PhiNode *phi = xphi->as_Phi();
 503 
 504   // Phi must be of loop header; backedge must wrap to increment
 505   if (phi->region() != x)
 506     return false;
 507   if ((trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr) ||
 508       (trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1)) {
 509     return false;
 510   }
 511   Node *init_trip = phi->in(LoopNode::EntryControl);
 512 
 513   // If iv trunc type is smaller than int, check for possible wrap.
 514   if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
 515     assert(trunc1 != NULL, "must have found some truncation");
 516 
 517     // Get a better type for the phi (filtered thru if's)
 518     const TypeInt* phi_ft = filtered_type(phi);
 519 
 520     // Can iv take on a value that will wrap?
 521     //
 522     // Ensure iv's limit is not within "stride" of the wrap value.
 523     //
 524     // Example for "short" type
 525     //    Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
 526     //    If the stride is +10, then the last value of the induction
 527     //    variable before the increment (phi_ft->_hi) must be
 528     //    <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
 529     //    ensure no truncation occurs after the increment.
 530 
 531     if (stride_con > 0) {
 532       if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
 533           iv_trunc_t->_lo > phi_ft->_lo) {
 534         return false;  // truncation may occur
 535       }
 536     } else if (stride_con < 0) {
 537       if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
 538           iv_trunc_t->_hi < phi_ft->_hi) {
 539         return false;  // truncation may occur
 540       }
 541     }
 542     // No possibility of wrap so truncation can be discarded
 543     // Promote iv type to Int
 544   } else {
 545     assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
 546   }
 547 
 548   // If the condition is inverted and we will be rolling
 549   // through MININT to MAXINT, then bail out.
 550   if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
 551       // Odd stride
 552       (bt == BoolTest::ne && stride_con != 1 && stride_con != -1) ||
 553       // Count down loop rolls through MAXINT
 554       ((bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0) ||
 555       // Count up loop rolls through MININT
 556       ((bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0)) {
 557     return false; // Bail out
 558   }
 559 
 560   const TypeInt* init_t = gvn->type(init_trip)->is_int();
 561 
 562   if (stride_con > 0) {
 563     jlong init_p = (jlong)init_t->_lo + stride_con;
 564     if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi)
 565       return false; // cyclic loop or this loop trips only once
 566   } else {
 567     jlong init_p = (jlong)init_t->_hi + stride_con;
 568     if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo)
 569       return false; // cyclic loop or this loop trips only once
 570   }
 571 
 572   if (phi_incr != NULL && bt != BoolTest::ne) {
 573     // check if there is a possiblity of IV overflowing after the first increment
 574     if (stride_con > 0) {
 575       if (init_t->_hi > max_jint - stride_con) {
 576         return false;
 577       }
 578     } else {
 579       if (init_t->_lo < min_jint - stride_con) {
 580         return false;
 581       }
 582     }
 583   }
 584 
 585   // =================================================
 586   // ---- SUCCESS!   Found A Trip-Counted Loop!  -----
 587   //
 588   assert(x->Opcode() == Op_Loop, "regular loops only");
 589   C->print_method(PHASE_BEFORE_CLOOPS, 3);
 590 
 591   Node *hook = new Node(6);
 592 
 593   // ===================================================
 594   // Generate loop limit check to avoid integer overflow
 595   // in cases like next (cyclic loops):
 596   //
 597   // for (i=0; i <= max_jint; i++) {}
 598   // for (i=0; i <  max_jint; i+=2) {}
 599   //
 600   //
 601   // Limit check predicate depends on the loop test:
 602   //
 603   // for(;i != limit; i++)       --> limit <= (max_jint)
 604   // for(;i <  limit; i+=stride) --> limit <= (max_jint - stride + 1)
 605   // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride    )
 606   //
 607 
 608   // Check if limit is excluded to do more precise int overflow check.
 609   bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge);
 610   int stride_m  = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1));
 611 
 612   // If compare points directly to the phi we need to adjust
 613   // the compare so that it points to the incr. Limit have
 614   // to be adjusted to keep trip count the same and the
 615   // adjusted limit should be checked for int overflow.
 616   if (phi_incr != NULL) {
 617     stride_m  += stride_con;
 618   }
 619 
 620   if (limit->is_Con()) {
 621     int limit_con = limit->get_int();
 622     if ((stride_con > 0 && limit_con > (max_jint - stride_m)) ||
 623         (stride_con < 0 && limit_con < (min_jint - stride_m))) {
 624       // Bailout: it could be integer overflow.
 625       return false;
 626     }
 627   } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) ||
 628              (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) {
 629       // Limit's type may satisfy the condition, for example,
 630       // when it is an array length.
 631   } else {
 632     // Generate loop's limit check.
 633     // Loop limit check predicate should be near the loop.
 634     ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
 635     if (!limit_check_proj) {
 636       // The limit check predicate is not generated if this method trapped here before.
 637 #ifdef ASSERT
 638       if (TraceLoopLimitCheck) {
 639         tty->print("missing loop limit check:");
 640         loop->dump_head();
 641         x->dump(1);
 642       }
 643 #endif
 644       return false;
 645     }
 646 
 647     IfNode* check_iff = limit_check_proj->in(0)->as_If();
 648 
 649     if (!is_dominator(get_ctrl(limit), check_iff->in(0))) {
 650       return false;
 651     }
 652 
 653     Node* cmp_limit;
 654     Node* bol;
 655 
 656     if (stride_con > 0) {
 657       cmp_limit = new CmpINode(limit, _igvn.intcon(max_jint - stride_m));
 658       bol = new BoolNode(cmp_limit, BoolTest::le);
 659     } else {
 660       cmp_limit = new CmpINode(limit, _igvn.intcon(min_jint - stride_m));
 661       bol = new BoolNode(cmp_limit, BoolTest::ge);
 662     }
 663 
 664     insert_loop_limit_check(limit_check_proj, cmp_limit, bol);
 665   }
 666 
 667   // Now we need to canonicalize loop condition.
 668   if (bt == BoolTest::ne) {
 669     assert(stride_con == 1 || stride_con == -1, "simple increment only");
 670     if (stride_con > 0 && init_t->_hi < limit_t->_lo) {
 671       // 'ne' can be replaced with 'lt' only when init < limit.
 672       bt = BoolTest::lt;
 673     } else if (stride_con < 0 && init_t->_lo > limit_t->_hi) {
 674       // 'ne' can be replaced with 'gt' only when init > limit.
 675       bt = BoolTest::gt;
 676     } else {
 677       ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
 678       if (!limit_check_proj) {
 679         // The limit check predicate is not generated if this method trapped here before.
 680 #ifdef ASSERT
 681         if (TraceLoopLimitCheck) {
 682           tty->print("missing loop limit check:");
 683           loop->dump_head();
 684           x->dump(1);
 685         }
 686 #endif
 687         return false;
 688       }
 689       IfNode* check_iff = limit_check_proj->in(0)->as_If();
 690 
 691       if (!is_dominator(get_ctrl(limit), check_iff->in(0)) ||
 692           !is_dominator(get_ctrl(init_trip), check_iff->in(0))) {
 693         return false;
 694       }
 695 
 696       Node* cmp_limit;
 697       Node* bol;
 698 
 699       if (stride_con > 0) {
 700         cmp_limit = new CmpINode(init_trip, limit);
 701         bol = new BoolNode(cmp_limit, BoolTest::lt);
 702       } else {
 703         cmp_limit = new CmpINode(init_trip, limit);
 704         bol = new BoolNode(cmp_limit, BoolTest::gt);
 705       }
 706 
 707       insert_loop_limit_check(limit_check_proj, cmp_limit, bol);
 708 
 709       if (stride_con > 0) {
 710         // 'ne' can be replaced with 'lt' only when init < limit.
 711         bt = BoolTest::lt;
 712       } else if (stride_con < 0) {
 713         // 'ne' can be replaced with 'gt' only when init > limit.
 714         bt = BoolTest::gt;
 715       }
 716     }
 717   }
 718 
 719   if (phi_incr != NULL) {
 720     // If compare points directly to the phi we need to adjust
 721     // the compare so that it points to the incr. Limit have
 722     // to be adjusted to keep trip count the same and we
 723     // should avoid int overflow.
 724     //
 725     //   i = init; do {} while(i++ < limit);
 726     // is converted to
 727     //   i = init; do {} while(++i < limit+1);
 728     //
 729     limit = gvn->transform(new AddINode(limit, stride));
 730   }
 731 
 732   if (incl_limit) {
 733     // The limit check guaranties that 'limit <= (max_jint - stride)' so
 734     // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
 735     //
 736     Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1);
 737     limit = gvn->transform(new AddINode(limit, one));
 738     if (bt == BoolTest::le)
 739       bt = BoolTest::lt;
 740     else if (bt == BoolTest::ge)
 741       bt = BoolTest::gt;
 742     else
 743       ShouldNotReachHere();
 744   }
 745   set_subtree_ctrl( limit );
 746 
 747   if (LoopStripMiningIter == 0) {
 748     // Check for SafePoint on backedge and remove
 749     Node *sfpt = x->in(LoopNode::LoopBackControl);
 750     if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
 751       lazy_replace( sfpt, iftrue );
 752       if (loop->_safepts != NULL) {
 753         loop->_safepts->yank(sfpt);
 754       }
 755       loop->_tail = iftrue;
 756     }
 757   }
 758 
 759   // Build a canonical trip test.
 760   // Clone code, as old values may be in use.
 761   incr = incr->clone();
 762   incr->set_req(1,phi);
 763   incr->set_req(2,stride);
 764   incr = _igvn.register_new_node_with_optimizer(incr);
 765   set_early_ctrl( incr );
 766   _igvn.rehash_node_delayed(phi);
 767   phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
 768 
 769   // If phi type is more restrictive than Int, raise to
 770   // Int to prevent (almost) infinite recursion in igvn
 771   // which can only handle integer types for constants or minint..maxint.
 772   if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
 773     Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
 774     nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
 775     nphi = _igvn.register_new_node_with_optimizer(nphi);
 776     set_ctrl(nphi, get_ctrl(phi));
 777     _igvn.replace_node(phi, nphi);
 778     phi = nphi->as_Phi();
 779   }
 780   cmp = cmp->clone();
 781   cmp->set_req(1,incr);
 782   cmp->set_req(2,limit);
 783   cmp = _igvn.register_new_node_with_optimizer(cmp);
 784   set_ctrl(cmp, iff->in(0));
 785 
 786   test = test->clone()->as_Bool();
 787   (*(BoolTest*)&test->_test)._test = bt;
 788   test->set_req(1,cmp);
 789   _igvn.register_new_node_with_optimizer(test);
 790   set_ctrl(test, iff->in(0));
 791 
 792   // Replace the old IfNode with a new LoopEndNode
 793   Node *lex = _igvn.register_new_node_with_optimizer(new CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
 794   IfNode *le = lex->as_If();
 795   uint dd = dom_depth(iff);
 796   set_idom(le, le->in(0), dd); // Update dominance for loop exit
 797   set_loop(le, loop);
 798 
 799   // Get the loop-exit control
 800   Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
 801 
 802   // Need to swap loop-exit and loop-back control?
 803   if (iftrue_op == Op_IfFalse) {
 804     Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le));
 805     Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le));
 806 
 807     loop->_tail = back_control = ift2;
 808     set_loop(ift2, loop);
 809     set_loop(iff2, get_loop(iffalse));
 810 
 811     // Lazy update of 'get_ctrl' mechanism.
 812     lazy_replace(iffalse, iff2);
 813     lazy_replace(iftrue,  ift2);
 814 
 815     // Swap names
 816     iffalse = iff2;
 817     iftrue  = ift2;
 818   } else {
 819     _igvn.rehash_node_delayed(iffalse);
 820     _igvn.rehash_node_delayed(iftrue);
 821     iffalse->set_req_X( 0, le, &_igvn );
 822     iftrue ->set_req_X( 0, le, &_igvn );
 823   }
 824 
 825   set_idom(iftrue,  le, dd+1);
 826   set_idom(iffalse, le, dd+1);
 827   assert(iff->outcnt() == 0, "should be dead now");
 828   lazy_replace( iff, le ); // fix 'get_ctrl'
 829 
 830   Node *sfpt2 = le->in(0);
 831 
 832   Node* entry_control = init_control;
 833   bool strip_mine_loop = LoopStripMiningIter > 1 && loop->_child == NULL &&
 834     sfpt2->Opcode() == Op_SafePoint && !loop->_has_call;
 835   IdealLoopTree* outer_ilt = NULL;
 836   if (strip_mine_loop) {
 837     outer_ilt = create_outer_strip_mined_loop(test, cmp, init_control, loop,
 838                                               cl_prob, le->_fcnt, entry_control,
 839                                               iffalse);
 840   }
 841 
 842   // Now setup a new CountedLoopNode to replace the existing LoopNode
 843   CountedLoopNode *l = new CountedLoopNode(entry_control, back_control);
 844   l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
 845   // The following assert is approximately true, and defines the intention
 846   // of can_be_counted_loop.  It fails, however, because phase->type
 847   // is not yet initialized for this loop and its parts.
 848   //assert(l->can_be_counted_loop(this), "sanity");
 849   _igvn.register_new_node_with_optimizer(l);
 850   set_loop(l, loop);
 851   loop->_head = l;
 852   // Fix all data nodes placed at the old loop head.
 853   // Uses the lazy-update mechanism of 'get_ctrl'.
 854   lazy_replace( x, l );
 855   set_idom(l, entry_control, dom_depth(entry_control) + 1);
 856 
 857   if (LoopStripMiningIter == 0 || strip_mine_loop) {
 858     // Check for immediately preceding SafePoint and remove
 859     if (sfpt2->Opcode() == Op_SafePoint && (LoopStripMiningIter != 0 || is_deleteable_safept(sfpt2))) {
 860       if (strip_mine_loop) {
 861         Node* outer_le = outer_ilt->_tail->in(0);
 862         Node* sfpt = sfpt2->clone();
 863         sfpt->set_req(0, iffalse);
 864         outer_le->set_req(0, sfpt);
 865         register_control(sfpt, outer_ilt, iffalse);
 866         set_idom(outer_le, sfpt, dom_depth(sfpt));
 867       }
 868       lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
 869       if (loop->_safepts != NULL) {
 870         loop->_safepts->yank(sfpt2);
 871       }
 872     }
 873   }
 874 
 875   // Free up intermediate goo
 876   _igvn.remove_dead_node(hook);
 877 
 878 #ifdef ASSERT
 879   assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
 880   assert(l == loop->_head && l->phi() == phi && l->loopexit_or_null() == lex, "" );
 881 #endif
 882 #ifndef PRODUCT
 883   if (TraceLoopOpts) {
 884     tty->print("Counted      ");
 885     loop->dump_head();
 886   }
 887 #endif
 888 
 889   C->print_method(PHASE_AFTER_CLOOPS, 3);
 890 
 891   // Capture bounds of the loop in the induction variable Phi before
 892   // subsequent transformation (iteration splitting) obscures the
 893   // bounds
 894   l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn));
 895 
 896   if (strip_mine_loop) {
 897     l->mark_strip_mined();
 898     l->verify_strip_mined(1);
 899     outer_ilt->_head->as_Loop()->verify_strip_mined(1);
 900     loop = outer_ilt;
 901   }
 902 
 903   return true;
 904 }
 905 
 906 //----------------------exact_limit-------------------------------------------
 907 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
 908   assert(loop->_head->is_CountedLoop(), "");
 909   CountedLoopNode *cl = loop->_head->as_CountedLoop();
 910   assert(cl->is_valid_counted_loop(), "");
 911 
 912   if (ABS(cl->stride_con()) == 1 ||
 913       cl->limit()->Opcode() == Op_LoopLimit) {
 914     // Old code has exact limit (it could be incorrect in case of int overflow).
 915     // Loop limit is exact with stride == 1. And loop may already have exact limit.
 916     return cl->limit();
 917   }
 918   Node *limit = NULL;
 919 #ifdef ASSERT
 920   BoolTest::mask bt = cl->loopexit()->test_trip();
 921   assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
 922 #endif
 923   if (cl->has_exact_trip_count()) {
 924     // Simple case: loop has constant boundaries.
 925     // Use jlongs to avoid integer overflow.
 926     int stride_con = cl->stride_con();
 927     jlong  init_con = cl->init_trip()->get_int();
 928     jlong limit_con = cl->limit()->get_int();
 929     julong trip_cnt = cl->trip_count();
 930     jlong final_con = init_con + trip_cnt*stride_con;
 931     int final_int = (int)final_con;
 932     // The final value should be in integer range since the loop
 933     // is counted and the limit was checked for overflow.
 934     assert(final_con == (jlong)final_int, "final value should be integer");
 935     limit = _igvn.intcon(final_int);
 936   } else {
 937     // Create new LoopLimit node to get exact limit (final iv value).
 938     limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
 939     register_new_node(limit, cl->in(LoopNode::EntryControl));
 940   }
 941   assert(limit != NULL, "sanity");
 942   return limit;
 943 }
 944 
 945 //------------------------------Ideal------------------------------------------
 946 // Return a node which is more "ideal" than the current node.
 947 // Attempt to convert into a counted-loop.
 948 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
 949   if (!can_be_counted_loop(phase) && !is_OuterStripMinedLoop()) {
 950     phase->C->set_major_progress();
 951   }
 952   return RegionNode::Ideal(phase, can_reshape);
 953 }
 954 
 955 void LoopNode::verify_strip_mined(int expect_skeleton) const {
 956 #ifdef ASSERT
 957   const OuterStripMinedLoopNode* outer = NULL;
 958   const CountedLoopNode* inner = NULL;
 959   if (is_strip_mined()) {
 960     assert(is_CountedLoop(), "no Loop should be marked strip mined");
 961     inner = as_CountedLoop();
 962     outer = inner->in(LoopNode::EntryControl)->as_OuterStripMinedLoop();
 963   } else if (is_OuterStripMinedLoop()) {
 964     outer = this->as_OuterStripMinedLoop();
 965     inner = outer->unique_ctrl_out()->as_CountedLoop();
 966     assert(!is_strip_mined(), "outer loop shouldn't be marked strip mined");
 967   }
 968   if (inner != NULL || outer != NULL) {
 969     assert(inner != NULL && outer != NULL, "missing loop in strip mined nest");
 970     Node* outer_tail = outer->in(LoopNode::LoopBackControl);
 971     Node* outer_le = outer_tail->in(0);
 972     assert(outer_le->Opcode() == Op_OuterStripMinedLoopEnd, "tail of outer loop should be an If");
 973     Node* sfpt = outer_le->in(0);
 974     assert(sfpt->Opcode() == Op_SafePoint, "where's the safepoint?");
 975     Node* inner_out = sfpt->in(0);
 976     if (inner_out->outcnt() != 1) {
 977       ResourceMark rm;
 978       Unique_Node_List wq;
 979 
 980       for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
 981         Node* u = inner_out->fast_out(i);
 982         if (u == sfpt) {
 983           continue;
 984         }
 985         wq.clear();
 986         wq.push(u);
 987         bool found_sfpt = false;
 988         for (uint next = 0; next < wq.size() && !found_sfpt; next++) {
 989           Node *n = wq.at(next);
 990           for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && !found_sfpt; i++) {
 991             Node* u = n->fast_out(i);
 992             if (u == sfpt) {
 993               found_sfpt = true;
 994             }
 995             if (!u->is_CFG()) {
 996               wq.push(u);
 997             }
 998           }
 999         }
1000         assert(found_sfpt, "no node in loop that's not input to safepoint");
1001       }
1002     }
1003     bool has_skeleton = outer_le->in(1)->bottom_type()->singleton() && outer_le->in(1)->bottom_type()->is_int()->get_con() == 0;
1004     if (has_skeleton) {
1005       CountedLoopEndNode* cle = inner_out->in(0)->as_CountedLoopEnd();
1006       assert(cle == inner->loopexit_or_null(), "mismatch");
1007       assert(expect_skeleton == 1 || expect_skeleton == -1, "unexpected skeleton node");
1008       assert(outer->outcnt() == 2, "only phis");
1009     } else {
1010       assert(expect_skeleton == 0 || expect_skeleton == -1, "no skeleton node?");
1011       uint phis = 0;
1012       for (DUIterator_Fast imax, i = inner->fast_outs(imax); i < imax; i++) {
1013         Node* u = inner->fast_out(i);
1014         if (u->is_Phi()) {
1015           phis++;
1016         }
1017       }
1018       for (DUIterator_Fast imax, i = outer->fast_outs(imax); i < imax; i++) {
1019         Node* u = outer->fast_out(i);
1020         assert(u == outer || u == inner || u->is_Phi(), "nothing between inner and outer loop");
1021       }
1022       Node* c = inner_out;
1023       uint stores = 0;
1024       for (;;) {
1025         for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
1026           Node* u = c->fast_out(i);
1027           if (u->is_Store()) {
1028             stores++;
1029           }
1030         }
1031         if (c->in(0)->is_CountedLoopEnd() || !UseShenandoahGC) {
1032           break;
1033         }
1034 #if INCLUDE_SHENANDOAHGC
1035         assert(UseShenandoahGC, "only for shenandoah barriers");
1036         assert(c->is_Region() && c->req() == 3, "region that ends barrier");
1037         uint j = 1;
1038         uint req = c->req();
1039         for (; j < req; j++) {
1040           Node* in = c->in(j);
1041           if (in->is_IfProj() && ShenandoahWriteBarrierNode::is_heap_stable_test(in->in(0))) {
1042             c = in->in(0)->in(0);
1043             break;
1044           }
1045         }
1046         assert(j < req, "should have found heap stable test");
1047 #endif
1048       }
1049       assert(outer->outcnt() >= phis + 2 && outer->outcnt() <= phis + 2 + stores + 1, "only phis");
1050     }
1051     assert(sfpt->outcnt() == 1, "no data node");
1052     assert(outer_tail->outcnt() == 1 || !has_skeleton, "no data node");
1053   }
1054 #endif
1055 }
1056 
1057 //=============================================================================
1058 //------------------------------Ideal------------------------------------------
1059 // Return a node which is more "ideal" than the current node.
1060 // Attempt to convert into a counted-loop.
1061 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1062   return RegionNode::Ideal(phase, can_reshape);
1063 }
1064 
1065 //------------------------------dump_spec--------------------------------------
1066 // Dump special per-node info
1067 #ifndef PRODUCT
1068 void CountedLoopNode::dump_spec(outputStream *st) const {
1069   LoopNode::dump_spec(st);
1070   if (stride_is_con()) {
1071     st->print("stride: %d ",stride_con());
1072   }
1073   if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
1074   if (is_main_loop()) st->print("main of N%d", _idx);
1075   if (is_post_loop()) st->print("post of N%d", _main_idx);
1076   if (is_strip_mined()) st->print(" strip mined");
1077 }
1078 #endif
1079 
1080 //=============================================================================
1081 int CountedLoopEndNode::stride_con() const {
1082   return stride()->bottom_type()->is_int()->get_con();
1083 }
1084 
1085 //=============================================================================
1086 //------------------------------Value-----------------------------------------
1087 const Type* LoopLimitNode::Value(PhaseGVN* phase) const {
1088   const Type* init_t   = phase->type(in(Init));
1089   const Type* limit_t  = phase->type(in(Limit));
1090   const Type* stride_t = phase->type(in(Stride));
1091   // Either input is TOP ==> the result is TOP
1092   if (init_t   == Type::TOP) return Type::TOP;
1093   if (limit_t  == Type::TOP) return Type::TOP;
1094   if (stride_t == Type::TOP) return Type::TOP;
1095 
1096   int stride_con = stride_t->is_int()->get_con();
1097   if (stride_con == 1)
1098     return NULL;  // Identity
1099 
1100   if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
1101     // Use jlongs to avoid integer overflow.
1102     jlong init_con   =  init_t->is_int()->get_con();
1103     jlong limit_con  = limit_t->is_int()->get_con();
1104     int  stride_m   = stride_con - (stride_con > 0 ? 1 : -1);
1105     jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
1106     jlong final_con  = init_con + stride_con*trip_count;
1107     int final_int = (int)final_con;
1108     // The final value should be in integer range since the loop
1109     // is counted and the limit was checked for overflow.
1110     assert(final_con == (jlong)final_int, "final value should be integer");
1111     return TypeInt::make(final_int);
1112   }
1113 
1114   return bottom_type(); // TypeInt::INT
1115 }
1116 
1117 //------------------------------Ideal------------------------------------------
1118 // Return a node which is more "ideal" than the current node.
1119 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1120   if (phase->type(in(Init))   == Type::TOP ||
1121       phase->type(in(Limit))  == Type::TOP ||
1122       phase->type(in(Stride)) == Type::TOP)
1123     return NULL;  // Dead
1124 
1125   int stride_con = phase->type(in(Stride))->is_int()->get_con();
1126   if (stride_con == 1)
1127     return NULL;  // Identity
1128 
1129   if (in(Init)->is_Con() && in(Limit)->is_Con())
1130     return NULL;  // Value
1131 
1132   // Delay following optimizations until all loop optimizations
1133   // done to keep Ideal graph simple.
1134   if (!can_reshape || phase->C->major_progress())
1135     return NULL;
1136 
1137   const TypeInt* init_t  = phase->type(in(Init) )->is_int();
1138   const TypeInt* limit_t = phase->type(in(Limit))->is_int();
1139   int stride_p;
1140   jlong lim, ini;
1141   julong max;
1142   if (stride_con > 0) {
1143     stride_p = stride_con;
1144     lim = limit_t->_hi;
1145     ini = init_t->_lo;
1146     max = (julong)max_jint;
1147   } else {
1148     stride_p = -stride_con;
1149     lim = init_t->_hi;
1150     ini = limit_t->_lo;
1151     max = (julong)min_jint;
1152   }
1153   julong range = lim - ini + stride_p;
1154   if (range <= max) {
1155     // Convert to integer expression if it is not overflow.
1156     Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
1157     Node *range = phase->transform(new SubINode(in(Limit), in(Init)));
1158     Node *bias  = phase->transform(new AddINode(range, stride_m));
1159     Node *trip  = phase->transform(new DivINode(0, bias, in(Stride)));
1160     Node *span  = phase->transform(new MulINode(trip, in(Stride)));
1161     return new AddINode(span, in(Init)); // exact limit
1162   }
1163 
1164   if (is_power_of_2(stride_p) ||                // divisor is 2^n
1165       !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
1166     // Convert to long expression to avoid integer overflow
1167     // and let igvn optimizer convert this division.
1168     //
1169     Node*   init   = phase->transform( new ConvI2LNode(in(Init)));
1170     Node*  limit   = phase->transform( new ConvI2LNode(in(Limit)));
1171     Node* stride   = phase->longcon(stride_con);
1172     Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
1173 
1174     Node *range = phase->transform(new SubLNode(limit, init));
1175     Node *bias  = phase->transform(new AddLNode(range, stride_m));
1176     Node *span;
1177     if (stride_con > 0 && is_power_of_2(stride_p)) {
1178       // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
1179       // and avoid generating rounding for division. Zero trip guard should
1180       // guarantee that init < limit but sometimes the guard is missing and
1181       // we can get situation when init > limit. Note, for the empty loop
1182       // optimization zero trip guard is generated explicitly which leaves
1183       // only RCE predicate where exact limit is used and the predicate
1184       // will simply fail forcing recompilation.
1185       Node* neg_stride   = phase->longcon(-stride_con);
1186       span = phase->transform(new AndLNode(bias, neg_stride));
1187     } else {
1188       Node *trip  = phase->transform(new DivLNode(0, bias, stride));
1189       span = phase->transform(new MulLNode(trip, stride));
1190     }
1191     // Convert back to int
1192     Node *span_int = phase->transform(new ConvL2INode(span));
1193     return new AddINode(span_int, in(Init)); // exact limit
1194   }
1195 
1196   return NULL;    // No progress
1197 }
1198 
1199 //------------------------------Identity---------------------------------------
1200 // If stride == 1 return limit node.
1201 Node* LoopLimitNode::Identity(PhaseGVN* phase) {
1202   int stride_con = phase->type(in(Stride))->is_int()->get_con();
1203   if (stride_con == 1 || stride_con == -1)
1204     return in(Limit);
1205   return this;
1206 }
1207 
1208 //=============================================================================
1209 //----------------------match_incr_with_optional_truncation--------------------
1210 // Match increment with optional truncation:
1211 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
1212 // Return NULL for failure. Success returns the increment node.
1213 Node* CountedLoopNode::match_incr_with_optional_truncation(
1214                       Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
1215   // Quick cutouts:
1216   if (expr == NULL || expr->req() != 3)  return NULL;
1217 
1218   Node *t1 = NULL;
1219   Node *t2 = NULL;
1220   const TypeInt* trunc_t = TypeInt::INT;
1221   Node* n1 = expr;
1222   int   n1op = n1->Opcode();
1223 
1224   // Try to strip (n1 & M) or (n1 << N >> N) from n1.
1225   if (n1op == Op_AndI &&
1226       n1->in(2)->is_Con() &&
1227       n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
1228     // %%% This check should match any mask of 2**K-1.
1229     t1 = n1;
1230     n1 = t1->in(1);
1231     n1op = n1->Opcode();
1232     trunc_t = TypeInt::CHAR;
1233   } else if (n1op == Op_RShiftI &&
1234              n1->in(1) != NULL &&
1235              n1->in(1)->Opcode() == Op_LShiftI &&
1236              n1->in(2) == n1->in(1)->in(2) &&
1237              n1->in(2)->is_Con()) {
1238     jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
1239     // %%% This check should match any shift in [1..31].
1240     if (shift == 16 || shift == 8) {
1241       t1 = n1;
1242       t2 = t1->in(1);
1243       n1 = t2->in(1);
1244       n1op = n1->Opcode();
1245       if (shift == 16) {
1246         trunc_t = TypeInt::SHORT;
1247       } else if (shift == 8) {
1248         trunc_t = TypeInt::BYTE;
1249       }
1250     }
1251   }
1252 
1253   // If (maybe after stripping) it is an AddI, we won:
1254   if (n1op == Op_AddI) {
1255     *trunc1 = t1;
1256     *trunc2 = t2;
1257     *trunc_type = trunc_t;
1258     return n1;
1259   }
1260 
1261   // failed
1262   return NULL;
1263 }
1264 
1265 LoopNode* CountedLoopNode::skip_strip_mined(int expect_skeleton) {
1266   if (is_strip_mined()) {
1267     verify_strip_mined(expect_skeleton);
1268     return in(EntryControl)->as_Loop();
1269   }
1270   return this;
1271 }
1272 
1273 OuterStripMinedLoopNode* CountedLoopNode::outer_loop() const {
1274   assert(is_strip_mined(), "not a strip mined loop");
1275   Node* c = in(EntryControl);
1276   if (c == NULL || c->is_top() || !c->is_OuterStripMinedLoop()) {
1277     return NULL;
1278   }
1279   return c->as_OuterStripMinedLoop();
1280 }
1281 
1282 IfTrueNode* OuterStripMinedLoopNode::outer_loop_tail() const {
1283   Node* c = in(LoopBackControl);
1284   if (c == NULL || c->is_top()) {
1285     return NULL;
1286   }
1287   return c->as_IfTrue();
1288 }
1289 
1290 IfTrueNode* CountedLoopNode::outer_loop_tail() const {
1291   LoopNode* l = outer_loop();
1292   if (l == NULL) {
1293     return NULL;
1294   }
1295   return l->outer_loop_tail();
1296 }
1297 
1298 OuterStripMinedLoopEndNode* OuterStripMinedLoopNode::outer_loop_end() const {
1299   IfTrueNode* proj = outer_loop_tail();
1300   if (proj == NULL) {
1301     return NULL;
1302   }
1303   Node* c = proj->in(0);
1304   if (c == NULL || c->is_top() || c->outcnt() != 2) {
1305     return NULL;
1306   }
1307   return c->as_OuterStripMinedLoopEnd();
1308 }
1309 
1310 OuterStripMinedLoopEndNode* CountedLoopNode::outer_loop_end() const {
1311   LoopNode* l = outer_loop();
1312   if (l == NULL) {
1313     return NULL;
1314   }
1315   return l->outer_loop_end();
1316 }
1317 
1318 IfFalseNode* OuterStripMinedLoopNode::outer_loop_exit() const {
1319   IfNode* le = outer_loop_end();
1320   if (le == NULL) {
1321     return NULL;
1322   }
1323   Node* c = le->proj_out_or_null(false);
1324   if (c == NULL) {
1325     return NULL;
1326   }
1327   return c->as_IfFalse();
1328 }
1329 
1330 IfFalseNode* CountedLoopNode::outer_loop_exit() const {
1331   LoopNode* l = outer_loop();
1332   if (l == NULL) {
1333     return NULL;
1334   }
1335   return l->outer_loop_exit();
1336 }
1337 
1338 SafePointNode* OuterStripMinedLoopNode::outer_safepoint() const {
1339   IfNode* le = outer_loop_end();
1340   if (le == NULL) {
1341     return NULL;
1342   }
1343   Node* c = le->in(0);
1344   if (c == NULL || c->is_top()) {
1345     return NULL;
1346   }
1347   assert(c->Opcode() == Op_SafePoint, "broken outer loop");
1348   return c->as_SafePoint();
1349 }
1350 
1351 SafePointNode* CountedLoopNode::outer_safepoint() const {
1352   LoopNode* l = outer_loop();
1353   if (l == NULL) {
1354     return NULL;
1355   }
1356   return l->outer_safepoint();
1357 }
1358 
1359 Node* CountedLoopNode::skip_predicates_from_entry(Node* ctrl) {
1360     while (ctrl != NULL && ctrl->is_Proj() && ctrl->in(0)->is_If() &&
1361            ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->outcnt() == 1 &&
1362            ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->unique_out()->Opcode() == Op_Halt) {
1363       ctrl = ctrl->in(0)->in(0);
1364     }
1365 
1366     return ctrl;
1367   }
1368 
1369 Node* CountedLoopNode::skip_predicates() {
1370   if (is_main_loop()) {
1371     Node* ctrl = skip_strip_mined()->in(LoopNode::EntryControl);
1372 
1373     return skip_predicates_from_entry(ctrl);
1374   }
1375   return in(LoopNode::EntryControl);
1376 }
1377 
1378 void OuterStripMinedLoopNode::adjust_strip_mined_loop(PhaseIterGVN* igvn) {
1379   // Look for the outer & inner strip mined loop, reduce number of
1380   // iterations of the inner loop, set exit condition of outer loop,
1381   // construct required phi nodes for outer loop.
1382   CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
1383   assert(inner_cl->is_strip_mined(), "inner loop should be strip mined");
1384   Node* inner_iv_phi = inner_cl->phi();
1385   if (inner_iv_phi == NULL) {
1386     IfNode* outer_le = outer_loop_end();
1387     Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1388     igvn->replace_node(outer_le, iff);
1389     inner_cl->clear_strip_mined();
1390     return;
1391   }
1392   CountedLoopEndNode* inner_cle = inner_cl->loopexit();
1393 
1394   int stride = inner_cl->stride_con();
1395   jlong scaled_iters_long = ((jlong)LoopStripMiningIter) * ABS(stride);
1396   int scaled_iters = (int)scaled_iters_long;
1397   int short_scaled_iters = LoopStripMiningIterShortLoop* ABS(stride);
1398   const TypeInt* inner_iv_t = igvn->type(inner_iv_phi)->is_int();
1399   jlong iter_estimate = (jlong)inner_iv_t->_hi - (jlong)inner_iv_t->_lo;
1400   assert(iter_estimate > 0, "broken");
1401   if ((jlong)scaled_iters != scaled_iters_long || iter_estimate <= short_scaled_iters) {
1402     // Remove outer loop and safepoint (too few iterations)
1403     Node* outer_sfpt = outer_safepoint();
1404     Node* outer_out = outer_loop_exit();
1405     igvn->replace_node(outer_out, outer_sfpt->in(0));
1406     igvn->replace_input_of(outer_sfpt, 0, igvn->C->top());
1407     inner_cl->clear_strip_mined();
1408     return;
1409   }
1410   if (iter_estimate <= scaled_iters_long) {
1411     // We would only go through one iteration of
1412     // the outer loop: drop the outer loop but
1413     // keep the safepoint so we don't run for
1414     // too long without a safepoint
1415     IfNode* outer_le = outer_loop_end();
1416     Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1417     igvn->replace_node(outer_le, iff);
1418     inner_cl->clear_strip_mined();
1419     return;
1420   }
1421 
1422   Node* cle_tail = inner_cle->proj_out(true);
1423   ResourceMark rm;
1424   Node_List old_new;
1425   if (cle_tail->outcnt() > 1) {
1426     // Look for nodes on backedge of inner loop and clone them
1427     Unique_Node_List backedge_nodes;
1428     for (DUIterator_Fast imax, i = cle_tail->fast_outs(imax); i < imax; i++) {
1429       Node* u = cle_tail->fast_out(i);
1430       if (u != inner_cl) {
1431         assert(!u->is_CFG(), "control flow on the backedge?");
1432         backedge_nodes.push(u);
1433       }
1434     }
1435     uint last = igvn->C->unique();
1436     for (uint next = 0; next < backedge_nodes.size(); next++) {
1437       Node* n = backedge_nodes.at(next);
1438       old_new.map(n->_idx, n->clone());
1439       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1440         Node* u = n->fast_out(i);
1441         assert(!u->is_CFG(), "broken");
1442         if (u->_idx >= last) {
1443           continue;
1444         }
1445         if (!u->is_Phi()) {
1446           backedge_nodes.push(u);
1447         } else {
1448           assert(u->in(0) == inner_cl, "strange phi on the backedge");
1449         }
1450       }
1451     }
1452     // Put the clones on the outer loop backedge
1453     Node* le_tail = outer_loop_tail();
1454     for (uint next = 0; next < backedge_nodes.size(); next++) {
1455       Node *n = old_new[backedge_nodes.at(next)->_idx];
1456       for (uint i = 1; i < n->req(); i++) {
1457         if (n->in(i) != NULL && old_new[n->in(i)->_idx] != NULL) {
1458           n->set_req(i, old_new[n->in(i)->_idx]);
1459         }
1460       }
1461       if (n->in(0) != NULL && n->in(0) == cle_tail) {
1462         n->set_req(0, le_tail);
1463       }
1464       igvn->register_new_node_with_optimizer(n);
1465     }
1466   }
1467 
1468   Node* iv_phi = NULL;
1469   // Make a clone of each phi in the inner loop
1470   // for the outer loop
1471   for (uint i = 0; i < inner_cl->outcnt(); i++) {
1472     Node* u = inner_cl->raw_out(i);
1473     if (u->is_Phi()) {
1474       assert(u->in(0) == inner_cl, "inconsistent");
1475       Node* phi = u->clone();
1476       phi->set_req(0, this);
1477       Node* be = old_new[phi->in(LoopNode::LoopBackControl)->_idx];
1478       if (be != NULL) {
1479         phi->set_req(LoopNode::LoopBackControl, be);
1480       }
1481       phi = igvn->transform(phi);
1482       igvn->replace_input_of(u, LoopNode::EntryControl, phi);
1483       if (u == inner_iv_phi) {
1484         iv_phi = phi;
1485       }
1486     }
1487   }
1488   Node* cle_out = inner_cle->proj_out(false);
1489   if (cle_out->outcnt() > 1) {
1490     // Look for chains of stores that were sunk
1491     // out of the inner loop and are in the outer loop
1492     for (DUIterator_Fast imax, i = cle_out->fast_outs(imax); i < imax; i++) {
1493       Node* u = cle_out->fast_out(i);
1494       if (u->is_Store()) {
1495         Node* first = u;
1496         for(;;) {
1497           Node* next = first->in(MemNode::Memory);
1498           if (!next->is_Store() || next->in(0) != cle_out) {
1499             break;
1500           }
1501           first = next;
1502         }
1503         Node* last = u;
1504         for(;;) {
1505           Node* next = NULL;
1506           for (DUIterator_Fast jmax, j = last->fast_outs(jmax); j < jmax; j++) {
1507             Node* uu = last->fast_out(j);
1508             if (uu->is_Store() && uu->in(0) == cle_out) {
1509               assert(next == NULL, "only one in the outer loop");
1510               next = uu;
1511             }
1512           }
1513           if (next == NULL) {
1514             break;
1515           }
1516           last = next;
1517         }
1518         Node* phi = NULL;
1519         for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) {
1520           Node* uu = fast_out(j);
1521           if (uu->is_Phi()) {
1522             Node* be = uu->in(LoopNode::LoopBackControl);
1523             if (be->is_Store() && old_new[be->_idx] != NULL) {
1524               assert(false, "store on the backedge + sunk stores: unsupported");
1525               // drop outer loop
1526               IfNode* outer_le = outer_loop_end();
1527               Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1528               igvn->replace_node(outer_le, iff);
1529               inner_cl->clear_strip_mined();
1530               return;
1531             }
1532             if (be == last || be == first->in(MemNode::Memory)) {
1533               assert(phi == NULL, "only one phi");
1534               phi = uu;
1535             }
1536           }
1537         }
1538 #ifdef ASSERT
1539         for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) {
1540           Node* uu = fast_out(j);
1541           if (uu->is_Phi() && uu->bottom_type() == Type::MEMORY) {
1542             if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) {
1543               assert(phi == uu, "what's that phi?");
1544             } else if (uu->adr_type() == TypePtr::BOTTOM) {
1545               Node* n = uu->in(LoopNode::LoopBackControl);
1546               uint limit = igvn->C->live_nodes();
1547               uint i = 0;
1548               while (n != uu) {
1549                 i++;
1550                 assert(i < limit, "infinite loop");
1551                 if (n->is_Proj()) {
1552                   n = n->in(0);
1553                 } else if (n->is_SafePoint() || n->is_MemBar()) {
1554                   n = n->in(TypeFunc::Memory);
1555                 } else if (n->is_Phi()) {
1556                   n = n->in(1);
1557                 } else if (n->is_MergeMem()) {
1558                   n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type()));
1559                 } else if (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) {
1560                   n = n->in(MemNode::Memory);
1561                 } else {
1562                   n->dump();
1563                   ShouldNotReachHere();
1564                 }
1565               }
1566             }
1567           }
1568         }
1569 #endif
1570         if (phi == NULL) {
1571           // If the an entire chains was sunk, the
1572           // inner loop has no phi for that memory
1573           // slice, create one for the outer loop
1574           phi = PhiNode::make(this, first->in(MemNode::Memory), Type::MEMORY,
1575                               igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type())));
1576           phi->set_req(LoopNode::LoopBackControl, last);
1577           phi = igvn->transform(phi);
1578           igvn->replace_input_of(first, MemNode::Memory, phi);
1579         } else {
1580           // Or fix the outer loop fix to include
1581           // that chain of stores.
1582           Node* be = phi->in(LoopNode::LoopBackControl);
1583           assert(!(be->is_Store() && old_new[be->_idx] != NULL), "store on the backedge + sunk stores: unsupported");
1584           if (be == first->in(MemNode::Memory)) {
1585             if (be == phi->in(LoopNode::LoopBackControl)) {
1586               igvn->replace_input_of(phi, LoopNode::LoopBackControl, last);
1587             } else {
1588               igvn->replace_input_of(be, MemNode::Memory, last);
1589             }
1590           } else {
1591 #ifdef ASSERT
1592             if (be == phi->in(LoopNode::LoopBackControl)) {
1593               assert(phi->in(LoopNode::LoopBackControl) == last, "");
1594             } else {
1595               assert(be->in(MemNode::Memory) == last, "");
1596             }
1597 #endif
1598           }
1599         }
1600       }
1601     }
1602   }
1603 
1604   if (iv_phi != NULL) {
1605     // Now adjust the inner loop's exit condition
1606     Node* limit = inner_cl->limit();
1607     Node* sub = NULL;
1608     if (stride > 0) {
1609       sub = igvn->transform(new SubINode(limit, iv_phi));
1610     } else {
1611       sub = igvn->transform(new SubINode(iv_phi, limit));
1612     }
1613     Node* min = igvn->transform(new MinINode(sub, igvn->intcon(scaled_iters)));
1614     Node* new_limit = NULL;
1615     if (stride > 0) {
1616       new_limit = igvn->transform(new AddINode(min, iv_phi));
1617     } else {
1618       new_limit = igvn->transform(new SubINode(iv_phi, min));
1619     }
1620     Node* inner_cmp = inner_cle->cmp_node();
1621     Node* inner_bol = inner_cle->in(CountedLoopEndNode::TestValue);
1622     Node* outer_bol = inner_bol;
1623     // cmp node for inner loop may be shared
1624     inner_cmp = inner_cmp->clone();
1625     inner_cmp->set_req(2, new_limit);
1626     inner_bol = inner_bol->clone();
1627     inner_bol->set_req(1, igvn->transform(inner_cmp));
1628     igvn->replace_input_of(inner_cle, CountedLoopEndNode::TestValue, igvn->transform(inner_bol));
1629     // Set the outer loop's exit condition too
1630     igvn->replace_input_of(outer_loop_end(), 1, outer_bol);
1631   } else {
1632     assert(false, "should be able to adjust outer loop");
1633     IfNode* outer_le = outer_loop_end();
1634     Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1635     igvn->replace_node(outer_le, iff);
1636     inner_cl->clear_strip_mined();
1637   }
1638 }
1639 
1640 const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const {
1641   if (!in(0)) return Type::TOP;
1642   if (phase->type(in(0)) == Type::TOP)
1643     return Type::TOP;
1644 
1645   return TypeTuple::IFBOTH;
1646 }
1647 
1648 Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1649   if (remove_dead_region(phase, can_reshape))  return this;
1650 
1651   return NULL;
1652 }
1653 
1654 //------------------------------filtered_type--------------------------------
1655 // Return a type based on condition control flow
1656 // A successful return will be a type that is restricted due
1657 // to a series of dominating if-tests, such as:
1658 //    if (i < 10) {
1659 //       if (i > 0) {
1660 //          here: "i" type is [1..10)
1661 //       }
1662 //    }
1663 // or a control flow merge
1664 //    if (i < 10) {
1665 //       do {
1666 //          phi( , ) -- at top of loop type is [min_int..10)
1667 //         i = ?
1668 //       } while ( i < 10)
1669 //
1670 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
1671   assert(n && n->bottom_type()->is_int(), "must be int");
1672   const TypeInt* filtered_t = NULL;
1673   if (!n->is_Phi()) {
1674     assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
1675     filtered_t = filtered_type_from_dominators(n, n_ctrl);
1676 
1677   } else {
1678     Node* phi    = n->as_Phi();
1679     Node* region = phi->in(0);
1680     assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
1681     if (region && region != C->top()) {
1682       for (uint i = 1; i < phi->req(); i++) {
1683         Node* val   = phi->in(i);
1684         Node* use_c = region->in(i);
1685         const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
1686         if (val_t != NULL) {
1687           if (filtered_t == NULL) {
1688             filtered_t = val_t;
1689           } else {
1690             filtered_t = filtered_t->meet(val_t)->is_int();
1691           }
1692         }
1693       }
1694     }
1695   }
1696   const TypeInt* n_t = _igvn.type(n)->is_int();
1697   if (filtered_t != NULL) {
1698     n_t = n_t->join(filtered_t)->is_int();
1699   }
1700   return n_t;
1701 }
1702 
1703 
1704 //------------------------------filtered_type_from_dominators--------------------------------
1705 // Return a possibly more restrictive type for val based on condition control flow of dominators
1706 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
1707   if (val->is_Con()) {
1708      return val->bottom_type()->is_int();
1709   }
1710   uint if_limit = 10; // Max number of dominating if's visited
1711   const TypeInt* rtn_t = NULL;
1712 
1713   if (use_ctrl && use_ctrl != C->top()) {
1714     Node* val_ctrl = get_ctrl(val);
1715     uint val_dom_depth = dom_depth(val_ctrl);
1716     Node* pred = use_ctrl;
1717     uint if_cnt = 0;
1718     while (if_cnt < if_limit) {
1719       if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
1720         if_cnt++;
1721         const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
1722         if (if_t != NULL) {
1723           if (rtn_t == NULL) {
1724             rtn_t = if_t;
1725           } else {
1726             rtn_t = rtn_t->join(if_t)->is_int();
1727           }
1728         }
1729       }
1730       pred = idom(pred);
1731       if (pred == NULL || pred == C->top()) {
1732         break;
1733       }
1734       // Stop if going beyond definition block of val
1735       if (dom_depth(pred) < val_dom_depth) {
1736         break;
1737       }
1738     }
1739   }
1740   return rtn_t;
1741 }
1742 
1743 
1744 //------------------------------dump_spec--------------------------------------
1745 // Dump special per-node info
1746 #ifndef PRODUCT
1747 void CountedLoopEndNode::dump_spec(outputStream *st) const {
1748   if( in(TestValue) != NULL && in(TestValue)->is_Bool() ) {
1749     BoolTest bt( test_trip()); // Added this for g++.
1750 
1751     st->print("[");
1752     bt.dump_on(st);
1753     st->print("]");
1754   }
1755   st->print(" ");
1756   IfNode::dump_spec(st);
1757 }
1758 #endif
1759 
1760 //=============================================================================
1761 //------------------------------is_member--------------------------------------
1762 // Is 'l' a member of 'this'?
1763 bool IdealLoopTree::is_member(const IdealLoopTree *l) const {
1764   while( l->_nest > _nest ) l = l->_parent;
1765   return l == this;
1766 }
1767 
1768 //------------------------------set_nest---------------------------------------
1769 // Set loop tree nesting depth.  Accumulate _has_call bits.
1770 int IdealLoopTree::set_nest( uint depth ) {
1771   _nest = depth;
1772   int bits = _has_call;
1773   if( _child ) bits |= _child->set_nest(depth+1);
1774   if( bits ) _has_call = 1;
1775   if( _next  ) bits |= _next ->set_nest(depth  );
1776   return bits;
1777 }
1778 
1779 //------------------------------split_fall_in----------------------------------
1780 // Split out multiple fall-in edges from the loop header.  Move them to a
1781 // private RegionNode before the loop.  This becomes the loop landing pad.
1782 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
1783   PhaseIterGVN &igvn = phase->_igvn;
1784   uint i;
1785 
1786   // Make a new RegionNode to be the landing pad.
1787   Node *landing_pad = new RegionNode( fall_in_cnt+1 );
1788   phase->set_loop(landing_pad,_parent);
1789   // Gather all the fall-in control paths into the landing pad
1790   uint icnt = fall_in_cnt;
1791   uint oreq = _head->req();
1792   for( i = oreq-1; i>0; i-- )
1793     if( !phase->is_member( this, _head->in(i) ) )
1794       landing_pad->set_req(icnt--,_head->in(i));
1795 
1796   // Peel off PhiNode edges as well
1797   for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1798     Node *oj = _head->fast_out(j);
1799     if( oj->is_Phi() ) {
1800       PhiNode* old_phi = oj->as_Phi();
1801       assert( old_phi->region() == _head, "" );
1802       igvn.hash_delete(old_phi);   // Yank from hash before hacking edges
1803       Node *p = PhiNode::make_blank(landing_pad, old_phi);
1804       uint icnt = fall_in_cnt;
1805       for( i = oreq-1; i>0; i-- ) {
1806         if( !phase->is_member( this, _head->in(i) ) ) {
1807           p->init_req(icnt--, old_phi->in(i));
1808           // Go ahead and clean out old edges from old phi
1809           old_phi->del_req(i);
1810         }
1811       }
1812       // Search for CSE's here, because ZKM.jar does a lot of
1813       // loop hackery and we need to be a little incremental
1814       // with the CSE to avoid O(N^2) node blow-up.
1815       Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
1816       if( p2 ) {                // Found CSE
1817         p->destruct();          // Recover useless new node
1818         p = p2;                 // Use old node
1819       } else {
1820         igvn.register_new_node_with_optimizer(p, old_phi);
1821       }
1822       // Make old Phi refer to new Phi.
1823       old_phi->add_req(p);
1824       // Check for the special case of making the old phi useless and
1825       // disappear it.  In JavaGrande I have a case where this useless
1826       // Phi is the loop limit and prevents recognizing a CountedLoop
1827       // which in turn prevents removing an empty loop.
1828       Node *id_old_phi = old_phi->Identity( &igvn );
1829       if( id_old_phi != old_phi ) { // Found a simple identity?
1830         // Note that I cannot call 'replace_node' here, because
1831         // that will yank the edge from old_phi to the Region and
1832         // I'm mid-iteration over the Region's uses.
1833         for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
1834           Node* use = old_phi->last_out(i);
1835           igvn.rehash_node_delayed(use);
1836           uint uses_found = 0;
1837           for (uint j = 0; j < use->len(); j++) {
1838             if (use->in(j) == old_phi) {
1839               if (j < use->req()) use->set_req (j, id_old_phi);
1840               else                use->set_prec(j, id_old_phi);
1841               uses_found++;
1842             }
1843           }
1844           i -= uses_found;    // we deleted 1 or more copies of this edge
1845         }
1846       }
1847       igvn._worklist.push(old_phi);
1848     }
1849   }
1850   // Finally clean out the fall-in edges from the RegionNode
1851   for( i = oreq-1; i>0; i-- ) {
1852     if( !phase->is_member( this, _head->in(i) ) ) {
1853       _head->del_req(i);
1854     }
1855   }
1856   igvn.rehash_node_delayed(_head);
1857   // Transform landing pad
1858   igvn.register_new_node_with_optimizer(landing_pad, _head);
1859   // Insert landing pad into the header
1860   _head->add_req(landing_pad);
1861 }
1862 
1863 //------------------------------split_outer_loop-------------------------------
1864 // Split out the outermost loop from this shared header.
1865 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
1866   PhaseIterGVN &igvn = phase->_igvn;
1867 
1868   // Find index of outermost loop; it should also be my tail.
1869   uint outer_idx = 1;
1870   while( _head->in(outer_idx) != _tail ) outer_idx++;
1871 
1872   // Make a LoopNode for the outermost loop.
1873   Node *ctl = _head->in(LoopNode::EntryControl);
1874   Node *outer = new LoopNode( ctl, _head->in(outer_idx) );
1875   outer = igvn.register_new_node_with_optimizer(outer, _head);
1876   phase->set_created_loop_node();
1877 
1878   // Outermost loop falls into '_head' loop
1879   _head->set_req(LoopNode::EntryControl, outer);
1880   _head->del_req(outer_idx);
1881   // Split all the Phis up between '_head' loop and 'outer' loop.
1882   for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1883     Node *out = _head->fast_out(j);
1884     if( out->is_Phi() ) {
1885       PhiNode *old_phi = out->as_Phi();
1886       assert( old_phi->region() == _head, "" );
1887       Node *phi = PhiNode::make_blank(outer, old_phi);
1888       phi->init_req(LoopNode::EntryControl,    old_phi->in(LoopNode::EntryControl));
1889       phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
1890       phi = igvn.register_new_node_with_optimizer(phi, old_phi);
1891       // Make old Phi point to new Phi on the fall-in path
1892       igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
1893       old_phi->del_req(outer_idx);
1894     }
1895   }
1896 
1897   // Use the new loop head instead of the old shared one
1898   _head = outer;
1899   phase->set_loop(_head, this);
1900 }
1901 
1902 //------------------------------fix_parent-------------------------------------
1903 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
1904   loop->_parent = parent;
1905   if( loop->_child ) fix_parent( loop->_child, loop   );
1906   if( loop->_next  ) fix_parent( loop->_next , parent );
1907 }
1908 
1909 //------------------------------estimate_path_freq-----------------------------
1910 static float estimate_path_freq( Node *n ) {
1911   // Try to extract some path frequency info
1912   IfNode *iff;
1913   for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
1914     uint nop = n->Opcode();
1915     if( nop == Op_SafePoint ) {   // Skip any safepoint
1916       n = n->in(0);
1917       continue;
1918     }
1919     if( nop == Op_CatchProj ) {   // Get count from a prior call
1920       // Assume call does not always throw exceptions: means the call-site
1921       // count is also the frequency of the fall-through path.
1922       assert( n->is_CatchProj(), "" );
1923       if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
1924         return 0.0f;            // Assume call exception path is rare
1925       Node *call = n->in(0)->in(0)->in(0);
1926       assert( call->is_Call(), "expect a call here" );
1927       const JVMState *jvms = ((CallNode*)call)->jvms();
1928       ciMethodData* methodData = jvms->method()->method_data();
1929       if (!methodData->is_mature())  return 0.0f; // No call-site data
1930       ciProfileData* data = methodData->bci_to_data(jvms->bci());
1931       if ((data == NULL) || !data->is_CounterData()) {
1932         // no call profile available, try call's control input
1933         n = n->in(0);
1934         continue;
1935       }
1936       return data->as_CounterData()->count()/FreqCountInvocations;
1937     }
1938     // See if there's a gating IF test
1939     Node *n_c = n->in(0);
1940     if( !n_c->is_If() ) break;       // No estimate available
1941     iff = n_c->as_If();
1942     if( iff->_fcnt != COUNT_UNKNOWN )   // Have a valid count?
1943       // Compute how much count comes on this path
1944       return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
1945     // Have no count info.  Skip dull uncommon-trap like branches.
1946     if( (nop == Op_IfTrue  && iff->_prob < PROB_LIKELY_MAG(5)) ||
1947         (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
1948       break;
1949     // Skip through never-taken branch; look for a real loop exit.
1950     n = iff->in(0);
1951   }
1952   return 0.0f;                  // No estimate available
1953 }
1954 
1955 //------------------------------merge_many_backedges---------------------------
1956 // Merge all the backedges from the shared header into a private Region.
1957 // Feed that region as the one backedge to this loop.
1958 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
1959   uint i;
1960 
1961   // Scan for the top 2 hottest backedges
1962   float hotcnt = 0.0f;
1963   float warmcnt = 0.0f;
1964   uint hot_idx = 0;
1965   // Loop starts at 2 because slot 1 is the fall-in path
1966   for( i = 2; i < _head->req(); i++ ) {
1967     float cnt = estimate_path_freq(_head->in(i));
1968     if( cnt > hotcnt ) {       // Grab hottest path
1969       warmcnt = hotcnt;
1970       hotcnt = cnt;
1971       hot_idx = i;
1972     } else if( cnt > warmcnt ) { // And 2nd hottest path
1973       warmcnt = cnt;
1974     }
1975   }
1976 
1977   // See if the hottest backedge is worthy of being an inner loop
1978   // by being much hotter than the next hottest backedge.
1979   if( hotcnt <= 0.0001 ||
1980       hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
1981 
1982   // Peel out the backedges into a private merge point; peel
1983   // them all except optionally hot_idx.
1984   PhaseIterGVN &igvn = phase->_igvn;
1985 
1986   Node *hot_tail = NULL;
1987   // Make a Region for the merge point
1988   Node *r = new RegionNode(1);
1989   for( i = 2; i < _head->req(); i++ ) {
1990     if( i != hot_idx )
1991       r->add_req( _head->in(i) );
1992     else hot_tail = _head->in(i);
1993   }
1994   igvn.register_new_node_with_optimizer(r, _head);
1995   // Plug region into end of loop _head, followed by hot_tail
1996   while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
1997   igvn.replace_input_of(_head, 2, r);
1998   if( hot_idx ) _head->add_req(hot_tail);
1999 
2000   // Split all the Phis up between '_head' loop and the Region 'r'
2001   for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
2002     Node *out = _head->fast_out(j);
2003     if( out->is_Phi() ) {
2004       PhiNode* n = out->as_Phi();
2005       igvn.hash_delete(n);      // Delete from hash before hacking edges
2006       Node *hot_phi = NULL;
2007       Node *phi = new PhiNode(r, n->type(), n->adr_type());
2008       // Check all inputs for the ones to peel out
2009       uint j = 1;
2010       for( uint i = 2; i < n->req(); i++ ) {
2011         if( i != hot_idx )
2012           phi->set_req( j++, n->in(i) );
2013         else hot_phi = n->in(i);
2014       }
2015       // Register the phi but do not transform until whole place transforms
2016       igvn.register_new_node_with_optimizer(phi, n);
2017       // Add the merge phi to the old Phi
2018       while( n->req() > 3 ) n->del_req( n->req()-1 );
2019       igvn.replace_input_of(n, 2, phi);
2020       if( hot_idx ) n->add_req(hot_phi);
2021     }
2022   }
2023 
2024 
2025   // Insert a new IdealLoopTree inserted below me.  Turn it into a clone
2026   // of self loop tree.  Turn self into a loop headed by _head and with
2027   // tail being the new merge point.
2028   IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
2029   phase->set_loop(_tail,ilt);   // Adjust tail
2030   _tail = r;                    // Self's tail is new merge point
2031   phase->set_loop(r,this);
2032   ilt->_child = _child;         // New guy has my children
2033   _child = ilt;                 // Self has new guy as only child
2034   ilt->_parent = this;          // new guy has self for parent
2035   ilt->_nest = _nest;           // Same nesting depth (for now)
2036 
2037   // Starting with 'ilt', look for child loop trees using the same shared
2038   // header.  Flatten these out; they will no longer be loops in the end.
2039   IdealLoopTree **pilt = &_child;
2040   while( ilt ) {
2041     if( ilt->_head == _head ) {
2042       uint i;
2043       for( i = 2; i < _head->req(); i++ )
2044         if( _head->in(i) == ilt->_tail )
2045           break;                // Still a loop
2046       if( i == _head->req() ) { // No longer a loop
2047         // Flatten ilt.  Hang ilt's "_next" list from the end of
2048         // ilt's '_child' list.  Move the ilt's _child up to replace ilt.
2049         IdealLoopTree **cp = &ilt->_child;
2050         while( *cp ) cp = &(*cp)->_next;   // Find end of child list
2051         *cp = ilt->_next;       // Hang next list at end of child list
2052         *pilt = ilt->_child;    // Move child up to replace ilt
2053         ilt->_head = NULL;      // Flag as a loop UNIONED into parent
2054         ilt = ilt->_child;      // Repeat using new ilt
2055         continue;               // do not advance over ilt->_child
2056       }
2057       assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
2058       phase->set_loop(_head,ilt);
2059     }
2060     pilt = &ilt->_child;        // Advance to next
2061     ilt = *pilt;
2062   }
2063 
2064   if( _child ) fix_parent( _child, this );
2065 }
2066 
2067 //------------------------------beautify_loops---------------------------------
2068 // Split shared headers and insert loop landing pads.
2069 // Insert a LoopNode to replace the RegionNode.
2070 // Return TRUE if loop tree is structurally changed.
2071 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
2072   bool result = false;
2073   // Cache parts in locals for easy
2074   PhaseIterGVN &igvn = phase->_igvn;
2075 
2076   igvn.hash_delete(_head);      // Yank from hash before hacking edges
2077 
2078   // Check for multiple fall-in paths.  Peel off a landing pad if need be.
2079   int fall_in_cnt = 0;
2080   for( uint i = 1; i < _head->req(); i++ )
2081     if( !phase->is_member( this, _head->in(i) ) )
2082       fall_in_cnt++;
2083   assert( fall_in_cnt, "at least 1 fall-in path" );
2084   if( fall_in_cnt > 1 )         // Need a loop landing pad to merge fall-ins
2085     split_fall_in( phase, fall_in_cnt );
2086 
2087   // Swap inputs to the _head and all Phis to move the fall-in edge to
2088   // the left.
2089   fall_in_cnt = 1;
2090   while( phase->is_member( this, _head->in(fall_in_cnt) ) )
2091     fall_in_cnt++;
2092   if( fall_in_cnt > 1 ) {
2093     // Since I am just swapping inputs I do not need to update def-use info
2094     Node *tmp = _head->in(1);
2095     igvn.rehash_node_delayed(_head);
2096     _head->set_req( 1, _head->in(fall_in_cnt) );
2097     _head->set_req( fall_in_cnt, tmp );
2098     // Swap also all Phis
2099     for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
2100       Node* phi = _head->fast_out(i);
2101       if( phi->is_Phi() ) {
2102         igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges
2103         tmp = phi->in(1);
2104         phi->set_req( 1, phi->in(fall_in_cnt) );
2105         phi->set_req( fall_in_cnt, tmp );
2106       }
2107     }
2108   }
2109   assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
2110   assert(  phase->is_member( this, _head->in(2) ), "right edge is loop" );
2111 
2112   // If I am a shared header (multiple backedges), peel off the many
2113   // backedges into a private merge point and use the merge point as
2114   // the one true backedge.
2115   if( _head->req() > 3 ) {
2116     // Merge the many backedges into a single backedge but leave
2117     // the hottest backedge as separate edge for the following peel.
2118     merge_many_backedges( phase );
2119     result = true;
2120   }
2121 
2122   // If I have one hot backedge, peel off myself loop.
2123   // I better be the outermost loop.
2124   if (_head->req() > 3 && !_irreducible) {
2125     split_outer_loop( phase );
2126     result = true;
2127 
2128   } else if (!_head->is_Loop() && !_irreducible) {
2129     // Make a new LoopNode to replace the old loop head
2130     Node *l = new LoopNode( _head->in(1), _head->in(2) );
2131     l = igvn.register_new_node_with_optimizer(l, _head);
2132     phase->set_created_loop_node();
2133     // Go ahead and replace _head
2134     phase->_igvn.replace_node( _head, l );
2135     _head = l;
2136     phase->set_loop(_head, this);
2137   }
2138 
2139   // Now recursively beautify nested loops
2140   if( _child ) result |= _child->beautify_loops( phase );
2141   if( _next  ) result |= _next ->beautify_loops( phase );
2142   return result;
2143 }
2144 
2145 //------------------------------allpaths_check_safepts----------------------------
2146 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
2147 // encountered.  Helper for check_safepts.
2148 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
2149   assert(stack.size() == 0, "empty stack");
2150   stack.push(_tail);
2151   visited.Clear();
2152   visited.set(_tail->_idx);
2153   while (stack.size() > 0) {
2154     Node* n = stack.pop();
2155     if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
2156       // Terminate this path
2157     } else if (n->Opcode() == Op_SafePoint) {
2158       if (_phase->get_loop(n) != this) {
2159         if (_required_safept == NULL) _required_safept = new Node_List();
2160         _required_safept->push(n);  // save the one closest to the tail
2161       }
2162       // Terminate this path
2163     } else {
2164       uint start = n->is_Region() ? 1 : 0;
2165       uint end   = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
2166       for (uint i = start; i < end; i++) {
2167         Node* in = n->in(i);
2168         assert(in->is_CFG(), "must be");
2169         if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
2170           stack.push(in);
2171         }
2172       }
2173     }
2174   }
2175 }
2176 
2177 //------------------------------check_safepts----------------------------
2178 // Given dominators, try to find loops with calls that must always be
2179 // executed (call dominates loop tail).  These loops do not need non-call
2180 // safepoints (ncsfpt).
2181 //
2182 // A complication is that a safepoint in a inner loop may be needed
2183 // by an outer loop. In the following, the inner loop sees it has a
2184 // call (block 3) on every path from the head (block 2) to the
2185 // backedge (arc 3->2).  So it deletes the ncsfpt (non-call safepoint)
2186 // in block 2, _but_ this leaves the outer loop without a safepoint.
2187 //
2188 //          entry  0
2189 //                 |
2190 //                 v
2191 // outer 1,2    +->1
2192 //              |  |
2193 //              |  v
2194 //              |  2<---+  ncsfpt in 2
2195 //              |_/|\   |
2196 //                 | v  |
2197 // inner 2,3      /  3  |  call in 3
2198 //               /   |  |
2199 //              v    +--+
2200 //        exit  4
2201 //
2202 //
2203 // This method creates a list (_required_safept) of ncsfpt nodes that must
2204 // be protected is created for each loop. When a ncsfpt maybe deleted, it
2205 // is first looked for in the lists for the outer loops of the current loop.
2206 //
2207 // The insights into the problem:
2208 //  A) counted loops are okay
2209 //  B) innermost loops are okay (only an inner loop can delete
2210 //     a ncsfpt needed by an outer loop)
2211 //  C) a loop is immune from an inner loop deleting a safepoint
2212 //     if the loop has a call on the idom-path
2213 //  D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
2214 //     idom-path that is not in a nested loop
2215 //  E) otherwise, an ncsfpt on the idom-path that is nested in an inner
2216 //     loop needs to be prevented from deletion by an inner loop
2217 //
2218 // There are two analyses:
2219 //  1) The first, and cheaper one, scans the loop body from
2220 //     tail to head following the idom (immediate dominator)
2221 //     chain, looking for the cases (C,D,E) above.
2222 //     Since inner loops are scanned before outer loops, there is summary
2223 //     information about inner loops.  Inner loops can be skipped over
2224 //     when the tail of an inner loop is encountered.
2225 //
2226 //  2) The second, invoked if the first fails to find a call or ncsfpt on
2227 //     the idom path (which is rare), scans all predecessor control paths
2228 //     from the tail to the head, terminating a path when a call or sfpt
2229 //     is encountered, to find the ncsfpt's that are closest to the tail.
2230 //
2231 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
2232   // Bottom up traversal
2233   IdealLoopTree* ch = _child;
2234   if (_child) _child->check_safepts(visited, stack);
2235   if (_next)  _next ->check_safepts(visited, stack);
2236 
2237   if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
2238     bool  has_call         = false; // call on dom-path
2239     bool  has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
2240     Node* nonlocal_ncsfpt  = NULL;  // ncsfpt on dom-path at a deeper depth
2241     // Scan the dom-path nodes from tail to head
2242     for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
2243       if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
2244         has_call = true;
2245         _has_sfpt = 1;          // Then no need for a safept!
2246         break;
2247       } else if (n->Opcode() == Op_SafePoint) {
2248         if (_phase->get_loop(n) == this) {
2249           has_local_ncsfpt = true;
2250           break;
2251         }
2252         if (nonlocal_ncsfpt == NULL) {
2253           nonlocal_ncsfpt = n; // save the one closest to the tail
2254         }
2255       } else {
2256         IdealLoopTree* nlpt = _phase->get_loop(n);
2257         if (this != nlpt) {
2258           // If at an inner loop tail, see if the inner loop has already
2259           // recorded seeing a call on the dom-path (and stop.)  If not,
2260           // jump to the head of the inner loop.
2261           assert(is_member(nlpt), "nested loop");
2262           Node* tail = nlpt->_tail;
2263           if (tail->in(0)->is_If()) tail = tail->in(0);
2264           if (n == tail) {
2265             // If inner loop has call on dom-path, so does outer loop
2266             if (nlpt->_has_sfpt) {
2267               has_call = true;
2268               _has_sfpt = 1;
2269               break;
2270             }
2271             // Skip to head of inner loop
2272             assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
2273             n = nlpt->_head;
2274           }
2275         }
2276       }
2277     }
2278     // Record safept's that this loop needs preserved when an
2279     // inner loop attempts to delete it's safepoints.
2280     if (_child != NULL && !has_call && !has_local_ncsfpt) {
2281       if (nonlocal_ncsfpt != NULL) {
2282         if (_required_safept == NULL) _required_safept = new Node_List();
2283         _required_safept->push(nonlocal_ncsfpt);
2284       } else {
2285         // Failed to find a suitable safept on the dom-path.  Now use
2286         // an all paths walk from tail to head, looking for safepoints to preserve.
2287         allpaths_check_safepts(visited, stack);
2288       }
2289     }
2290   }
2291 }
2292 
2293 //---------------------------is_deleteable_safept----------------------------
2294 // Is safept not required by an outer loop?
2295 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
2296   assert(sfpt->Opcode() == Op_SafePoint, "");
2297   IdealLoopTree* lp = get_loop(sfpt)->_parent;
2298   while (lp != NULL) {
2299     Node_List* sfpts = lp->_required_safept;
2300     if (sfpts != NULL) {
2301       for (uint i = 0; i < sfpts->size(); i++) {
2302         if (sfpt == sfpts->at(i))
2303           return false;
2304       }
2305     }
2306     lp = lp->_parent;
2307   }
2308   return true;
2309 }
2310 
2311 //---------------------------replace_parallel_iv-------------------------------
2312 // Replace parallel induction variable (parallel to trip counter)
2313 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
2314   assert(loop->_head->is_CountedLoop(), "");
2315   CountedLoopNode *cl = loop->_head->as_CountedLoop();
2316   if (!cl->is_valid_counted_loop())
2317     return;         // skip malformed counted loop
2318   Node *incr = cl->incr();
2319   if (incr == NULL)
2320     return;         // Dead loop?
2321   Node *init = cl->init_trip();
2322   Node *phi  = cl->phi();
2323   int stride_con = cl->stride_con();
2324 
2325   // Visit all children, looking for Phis
2326   for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
2327     Node *out = cl->out(i);
2328     // Look for other phis (secondary IVs). Skip dead ones
2329     if (!out->is_Phi() || out == phi || !has_node(out))
2330       continue;
2331     PhiNode* phi2 = out->as_Phi();
2332     Node *incr2 = phi2->in( LoopNode::LoopBackControl );
2333     // Look for induction variables of the form:  X += constant
2334     if (phi2->region() != loop->_head ||
2335         incr2->req() != 3 ||
2336         incr2->in(1) != phi2 ||
2337         incr2 == incr ||
2338         incr2->Opcode() != Op_AddI ||
2339         !incr2->in(2)->is_Con())
2340       continue;
2341 
2342     // Check for parallel induction variable (parallel to trip counter)
2343     // via an affine function.  In particular, count-down loops with
2344     // count-up array indices are common. We only RCE references off
2345     // the trip-counter, so we need to convert all these to trip-counter
2346     // expressions.
2347     Node *init2 = phi2->in( LoopNode::EntryControl );
2348     int stride_con2 = incr2->in(2)->get_int();
2349 
2350     // The ratio of the two strides cannot be represented as an int
2351     // if stride_con2 is min_int and stride_con is -1.
2352     if (stride_con2 == min_jint && stride_con == -1) {
2353       continue;
2354     }
2355 
2356     // The general case here gets a little tricky.  We want to find the
2357     // GCD of all possible parallel IV's and make a new IV using this
2358     // GCD for the loop.  Then all possible IVs are simple multiples of
2359     // the GCD.  In practice, this will cover very few extra loops.
2360     // Instead we require 'stride_con2' to be a multiple of 'stride_con',
2361     // where +/-1 is the common case, but other integer multiples are
2362     // also easy to handle.
2363     int ratio_con = stride_con2/stride_con;
2364 
2365     if ((ratio_con * stride_con) == stride_con2) { // Check for exact
2366 #ifndef PRODUCT
2367       if (TraceLoopOpts) {
2368         tty->print("Parallel IV: %d ", phi2->_idx);
2369         loop->dump_head();
2370       }
2371 #endif
2372       // Convert to using the trip counter.  The parallel induction
2373       // variable differs from the trip counter by a loop-invariant
2374       // amount, the difference between their respective initial values.
2375       // It is scaled by the 'ratio_con'.
2376       Node* ratio = _igvn.intcon(ratio_con);
2377       set_ctrl(ratio, C->root());
2378       Node* ratio_init = new MulINode(init, ratio);
2379       _igvn.register_new_node_with_optimizer(ratio_init, init);
2380       set_early_ctrl(ratio_init);
2381       Node* diff = new SubINode(init2, ratio_init);
2382       _igvn.register_new_node_with_optimizer(diff, init2);
2383       set_early_ctrl(diff);
2384       Node* ratio_idx = new MulINode(phi, ratio);
2385       _igvn.register_new_node_with_optimizer(ratio_idx, phi);
2386       set_ctrl(ratio_idx, cl);
2387       Node* add = new AddINode(ratio_idx, diff);
2388       _igvn.register_new_node_with_optimizer(add);
2389       set_ctrl(add, cl);
2390       _igvn.replace_node( phi2, add );
2391       // Sometimes an induction variable is unused
2392       if (add->outcnt() == 0) {
2393         _igvn.remove_dead_node(add);
2394       }
2395       --i; // deleted this phi; rescan starting with next position
2396       continue;
2397     }
2398   }
2399 }
2400 
2401 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
2402   Node* keep = NULL;
2403   if (keep_one) {
2404     // Look for a safepoint on the idom-path.
2405     for (Node* i = tail(); i != _head; i = phase->idom(i)) {
2406       if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
2407         keep = i;
2408         break; // Found one
2409       }
2410     }
2411   }
2412 
2413   // Don't remove any safepoints if it is requested to keep a single safepoint and
2414   // no safepoint was found on idom-path. It is not safe to remove any safepoint
2415   // in this case since there's no safepoint dominating all paths in the loop body.
2416   bool prune = !keep_one || keep != NULL;
2417 
2418   // Delete other safepoints in this loop.
2419   Node_List* sfpts = _safepts;
2420   if (prune && sfpts != NULL) {
2421     assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint");
2422     for (uint i = 0; i < sfpts->size(); i++) {
2423       Node* n = sfpts->at(i);
2424       assert(phase->get_loop(n) == this, "");
2425       if (n != keep && phase->is_deleteable_safept(n)) {
2426         phase->lazy_replace(n, n->in(TypeFunc::Control));
2427       }
2428     }
2429   }
2430 }
2431 
2432 //------------------------------counted_loop-----------------------------------
2433 // Convert to counted loops where possible
2434 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
2435 
2436   // For grins, set the inner-loop flag here
2437   if (!_child) {
2438     if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
2439   }
2440 
2441   IdealLoopTree* loop = this;
2442   if (_head->is_CountedLoop() ||
2443       phase->is_counted_loop(_head, loop)) {
2444 
2445     if (LoopStripMiningIter == 0 || (LoopStripMiningIter > 1 && _child == NULL)) {
2446       // Indicate we do not need a safepoint here
2447       _has_sfpt = 1;
2448     }
2449 
2450     // Remove safepoints
2451     bool keep_one_sfpt = !(_has_call || _has_sfpt);
2452     remove_safepoints(phase, keep_one_sfpt);
2453 
2454     // Look for induction variables
2455     phase->replace_parallel_iv(this);
2456 
2457   } else if (_parent != NULL && !_irreducible) {
2458     // Not a counted loop. Keep one safepoint.
2459     bool keep_one_sfpt = true;
2460     remove_safepoints(phase, keep_one_sfpt);
2461   }
2462 
2463   // Recursively
2464   assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?");
2465   assert(loop->_child != this || (loop->_child->_child == NULL && loop->_child->_next == NULL), "would miss some loops");
2466   if (loop->_child && loop->_child != this) loop->_child->counted_loop(phase);
2467   if (loop->_next)  loop->_next ->counted_loop(phase);
2468 }
2469 
2470 #ifndef PRODUCT
2471 //------------------------------dump_head--------------------------------------
2472 // Dump 1 liner for loop header info
2473 void IdealLoopTree::dump_head( ) const {
2474   for (uint i=0; i<_nest; i++)
2475     tty->print("  ");
2476   tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
2477   if (_irreducible) tty->print(" IRREDUCIBLE");
2478   Node* entry = _head->is_Loop() ? _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl) : _head->in(LoopNode::EntryControl);
2479   Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
2480   if (predicate != NULL ) {
2481     tty->print(" limit_check");
2482     entry = PhaseIdealLoop::skip_loop_predicates(entry);
2483   }
2484   if (UseLoopPredicate) {
2485     entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
2486     if (entry != NULL) {
2487       tty->print(" predicated");
2488       entry = PhaseIdealLoop::skip_loop_predicates(entry);
2489     }
2490   }
2491   if (UseProfiledLoopPredicate) {
2492     entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate);
2493     if (entry != NULL) {
2494       tty->print(" profile_predicated");
2495     }
2496   }
2497   if (_head->is_CountedLoop()) {
2498     CountedLoopNode *cl = _head->as_CountedLoop();
2499     tty->print(" counted");
2500 
2501     Node* init_n = cl->init_trip();
2502     if (init_n  != NULL &&  init_n->is_Con())
2503       tty->print(" [%d,", cl->init_trip()->get_int());
2504     else
2505       tty->print(" [int,");
2506     Node* limit_n = cl->limit();
2507     if (limit_n  != NULL &&  limit_n->is_Con())
2508       tty->print("%d),", cl->limit()->get_int());
2509     else
2510       tty->print("int),");
2511     int stride_con  = cl->stride_con();
2512     if (stride_con > 0) tty->print("+");
2513     tty->print("%d", stride_con);
2514 
2515     tty->print(" (%0.f iters) ", cl->profile_trip_cnt());
2516 
2517     if (cl->is_pre_loop ()) tty->print(" pre" );
2518     if (cl->is_main_loop()) tty->print(" main");
2519     if (cl->is_post_loop()) tty->print(" post");
2520     if (cl->is_vectorized_loop()) tty->print(" vector");
2521     if (cl->range_checks_present()) tty->print(" rc ");
2522     if (cl->is_multiversioned()) tty->print(" multi ");
2523   }
2524   if (_has_call) tty->print(" has_call");
2525   if (_has_sfpt) tty->print(" has_sfpt");
2526   if (_rce_candidate) tty->print(" rce");
2527   if (_safepts != NULL && _safepts->size() > 0) {
2528     tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
2529   }
2530   if (_required_safept != NULL && _required_safept->size() > 0) {
2531     tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
2532   }
2533   if (Verbose) {
2534     tty->print(" body={"); _body.dump_simple(); tty->print(" }");
2535   }
2536   if (_head->is_Loop() && _head->as_Loop()->is_strip_mined()) {
2537     tty->print(" strip_mined");
2538   }
2539   tty->cr();
2540 }
2541 
2542 //------------------------------dump-------------------------------------------
2543 // Dump loops by loop tree
2544 void IdealLoopTree::dump( ) const {
2545   dump_head();
2546   if (_child) _child->dump();
2547   if (_next)  _next ->dump();
2548 }
2549 
2550 #endif
2551 
2552 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
2553   if (loop == root) {
2554     if (loop->_child != NULL) {
2555       log->begin_head("loop_tree");
2556       log->end_head();
2557       if( loop->_child ) log_loop_tree(root, loop->_child, log);
2558       log->tail("loop_tree");
2559       assert(loop->_next == NULL, "what?");
2560     }
2561   } else {
2562     Node* head = loop->_head;
2563     log->begin_head("loop");
2564     log->print(" idx='%d' ", head->_idx);
2565     if (loop->_irreducible) log->print("irreducible='1' ");
2566     if (head->is_Loop()) {
2567       if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
2568       if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
2569     }
2570     if (head->is_CountedLoop()) {
2571       CountedLoopNode* cl = head->as_CountedLoop();
2572       if (cl->is_pre_loop())  log->print("pre_loop='%d' ",  cl->main_idx());
2573       if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
2574       if (cl->is_post_loop()) log->print("post_loop='%d' ",  cl->main_idx());
2575     }
2576     log->end_head();
2577     if( loop->_child ) log_loop_tree(root, loop->_child, log);
2578     log->tail("loop");
2579     if( loop->_next  ) log_loop_tree(root, loop->_next, log);
2580   }
2581 }
2582 
2583 //---------------------collect_potentially_useful_predicates-----------------------
2584 // Helper function to collect potentially useful predicates to prevent them from
2585 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
2586 void PhaseIdealLoop::collect_potentially_useful_predicates(
2587                          IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
2588   if (loop->_child) { // child
2589     collect_potentially_useful_predicates(loop->_child, useful_predicates);
2590   }
2591 
2592   // self (only loops that we can apply loop predication may use their predicates)
2593   if (loop->_head->is_Loop() &&
2594       !loop->_irreducible    &&
2595       !loop->tail()->is_top()) {
2596     LoopNode* lpn = loop->_head->as_Loop();
2597     Node* entry = lpn->in(LoopNode::EntryControl);
2598     Node* predicate_proj = find_predicate(entry); // loop_limit_check first
2599     if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
2600       assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
2601       useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2602       entry = skip_loop_predicates(entry);
2603     }
2604     predicate_proj = find_predicate(entry); // Predicate
2605     if (predicate_proj != NULL ) {
2606       useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2607       entry = skip_loop_predicates(entry);
2608     }
2609     if (UseProfiledLoopPredicate) {
2610       predicate_proj = find_predicate(entry); // Predicate
2611       if (predicate_proj != NULL ) {
2612         useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2613       }
2614     }
2615   }
2616 
2617   if (loop->_next) { // sibling
2618     collect_potentially_useful_predicates(loop->_next, useful_predicates);
2619   }
2620 }
2621 
2622 //------------------------eliminate_useless_predicates-----------------------------
2623 // Eliminate all inserted predicates if they could not be used by loop predication.
2624 // Note: it will also eliminates loop limits check predicate since it also uses
2625 // Opaque1 node (see Parse::add_predicate()).
2626 void PhaseIdealLoop::eliminate_useless_predicates() {
2627   if (C->predicate_count() == 0)
2628     return; // no predicate left
2629 
2630   Unique_Node_List useful_predicates; // to store useful predicates
2631   if (C->has_loops()) {
2632     collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
2633   }
2634 
2635   for (int i = C->predicate_count(); i > 0; i--) {
2636      Node * n = C->predicate_opaque1_node(i-1);
2637      assert(n->Opcode() == Op_Opaque1, "must be");
2638      if (!useful_predicates.member(n)) { // not in the useful list
2639        _igvn.replace_node(n, n->in(1));
2640      }
2641   }
2642 }
2643 
2644 //------------------------process_expensive_nodes-----------------------------
2645 // Expensive nodes have their control input set to prevent the GVN
2646 // from commoning them and as a result forcing the resulting node to
2647 // be in a more frequent path. Use CFG information here, to change the
2648 // control inputs so that some expensive nodes can be commoned while
2649 // not executed more frequently.
2650 bool PhaseIdealLoop::process_expensive_nodes() {
2651   assert(OptimizeExpensiveOps, "optimization off?");
2652 
2653   // Sort nodes to bring similar nodes together
2654   C->sort_expensive_nodes();
2655 
2656   bool progress = false;
2657 
2658   for (int i = 0; i < C->expensive_count(); ) {
2659     Node* n = C->expensive_node(i);
2660     int start = i;
2661     // Find nodes similar to n
2662     i++;
2663     for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
2664     int end = i;
2665     // And compare them two by two
2666     for (int j = start; j < end; j++) {
2667       Node* n1 = C->expensive_node(j);
2668       if (is_node_unreachable(n1)) {
2669         continue;
2670       }
2671       for (int k = j+1; k < end; k++) {
2672         Node* n2 = C->expensive_node(k);
2673         if (is_node_unreachable(n2)) {
2674           continue;
2675         }
2676 
2677         assert(n1 != n2, "should be pair of nodes");
2678 
2679         Node* c1 = n1->in(0);
2680         Node* c2 = n2->in(0);
2681 
2682         Node* parent_c1 = c1;
2683         Node* parent_c2 = c2;
2684 
2685         // The call to get_early_ctrl_for_expensive() moves the
2686         // expensive nodes up but stops at loops that are in a if
2687         // branch. See whether we can exit the loop and move above the
2688         // If.
2689         if (c1->is_Loop()) {
2690           parent_c1 = c1->in(1);
2691         }
2692         if (c2->is_Loop()) {
2693           parent_c2 = c2->in(1);
2694         }
2695 
2696         if (parent_c1 == parent_c2) {
2697           _igvn._worklist.push(n1);
2698           _igvn._worklist.push(n2);
2699           continue;
2700         }
2701 
2702         // Look for identical expensive node up the dominator chain.
2703         if (is_dominator(c1, c2)) {
2704           c2 = c1;
2705         } else if (is_dominator(c2, c1)) {
2706           c1 = c2;
2707         } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
2708                    parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
2709           // Both branches have the same expensive node so move it up
2710           // before the if.
2711           c1 = c2 = idom(parent_c1->in(0));
2712         }
2713         // Do the actual moves
2714         if (n1->in(0) != c1) {
2715           _igvn.hash_delete(n1);
2716           n1->set_req(0, c1);
2717           _igvn.hash_insert(n1);
2718           _igvn._worklist.push(n1);
2719           progress = true;
2720         }
2721         if (n2->in(0) != c2) {
2722           _igvn.hash_delete(n2);
2723           n2->set_req(0, c2);
2724           _igvn.hash_insert(n2);
2725           _igvn._worklist.push(n2);
2726           progress = true;
2727         }
2728       }
2729     }
2730   }
2731 
2732   return progress;
2733 }
2734 
2735 
2736 //=============================================================================
2737 //----------------------------build_and_optimize-------------------------------
2738 // Create a PhaseLoop.  Build the ideal Loop tree.  Map each Ideal Node to
2739 // its corresponding LoopNode.  If 'optimize' is true, do some loop cleanups.
2740 void PhaseIdealLoop::build_and_optimize(LoopOptsMode mode) {
2741   bool do_split_ifs = (mode == LoopOptsDefault || mode == LoopOptsZgcLastRound);
2742   bool skip_loop_opts = (mode == LoopOptsNone) ;
2743   bool shenandoah_opts = (mode == LoopOptsShenandoahExpand ||
2744                           mode == LoopOptsShenandoahPostExpand);
2745 
2746   ResourceMark rm;
2747 
2748   int old_progress = C->major_progress();
2749   uint orig_worklist_size = _igvn._worklist.size();
2750 
2751   // Reset major-progress flag for the driver's heuristics
2752   C->clear_major_progress();
2753 
2754 #ifndef PRODUCT
2755   // Capture for later assert
2756   uint unique = C->unique();
2757   _loop_invokes++;
2758   _loop_work += unique;
2759 #endif
2760 
2761   // True if the method has at least 1 irreducible loop
2762   _has_irreducible_loops = false;
2763 
2764   _created_loop_node = false;
2765 
2766   Arena *a = Thread::current()->resource_area();
2767   VectorSet visited(a);
2768   // Pre-grow the mapping from Nodes to IdealLoopTrees.
2769   _nodes.map(C->unique(), NULL);
2770   memset(_nodes.adr(), 0, wordSize * C->unique());
2771 
2772   // Pre-build the top-level outermost loop tree entry
2773   _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
2774   // Do not need a safepoint at the top level
2775   _ltree_root->_has_sfpt = 1;
2776 
2777   // Initialize Dominators.
2778   // Checked in clone_loop_predicate() during beautify_loops().
2779   _idom_size = 0;
2780   _idom      = NULL;
2781   _dom_depth = NULL;
2782   _dom_stk   = NULL;
2783 
2784   // Empty pre-order array
2785   allocate_preorders();
2786 
2787   // Build a loop tree on the fly.  Build a mapping from CFG nodes to
2788   // IdealLoopTree entries.  Data nodes are NOT walked.
2789   build_loop_tree();
2790   // Check for bailout, and return
2791   if (C->failing()) {
2792     return;
2793   }
2794 
2795   // No loops after all
2796   if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
2797 
2798   // There should always be an outer loop containing the Root and Return nodes.
2799   // If not, we have a degenerate empty program.  Bail out in this case.
2800   if (!has_node(C->root())) {
2801     if (!_verify_only) {
2802       C->clear_major_progress();
2803       C->record_method_not_compilable("empty program detected during loop optimization");
2804     }
2805     return;
2806   }
2807 
2808   // Nothing to do, so get out
2809   bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only && !shenandoah_opts;
2810   bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
2811   if (stop_early && !do_expensive_nodes) {
2812     _igvn.optimize();           // Cleanup NeverBranches
2813     return;
2814   }
2815 
2816   // Set loop nesting depth
2817   _ltree_root->set_nest( 0 );
2818 
2819   // Split shared headers and insert loop landing pads.
2820   // Do not bother doing this on the Root loop of course.
2821   if( !_verify_me && !_verify_only && _ltree_root->_child ) {
2822     C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
2823     if( _ltree_root->_child->beautify_loops( this ) ) {
2824       // Re-build loop tree!
2825       _ltree_root->_child = NULL;
2826       _nodes.clear();
2827       reallocate_preorders();
2828       build_loop_tree();
2829       // Check for bailout, and return
2830       if (C->failing()) {
2831         return;
2832       }
2833       // Reset loop nesting depth
2834       _ltree_root->set_nest( 0 );
2835 
2836       C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
2837     }
2838   }
2839 
2840   // Build Dominators for elision of NULL checks & loop finding.
2841   // Since nodes do not have a slot for immediate dominator, make
2842   // a persistent side array for that info indexed on node->_idx.
2843   _idom_size = C->unique();
2844   _idom      = NEW_RESOURCE_ARRAY( Node*, _idom_size );
2845   _dom_depth = NEW_RESOURCE_ARRAY( uint,  _idom_size );
2846   _dom_stk   = NULL; // Allocated on demand in recompute_dom_depth
2847   memset( _dom_depth, 0, _idom_size * sizeof(uint) );
2848 
2849   Dominators();
2850 
2851   if (!_verify_only) {
2852     // As a side effect, Dominators removed any unreachable CFG paths
2853     // into RegionNodes.  It doesn't do this test against Root, so
2854     // we do it here.
2855     for( uint i = 1; i < C->root()->req(); i++ ) {
2856       if( !_nodes[C->root()->in(i)->_idx] ) {    // Dead path into Root?
2857         _igvn.delete_input_of(C->root(), i);
2858         i--;                      // Rerun same iteration on compressed edges
2859       }
2860     }
2861 
2862     // Given dominators, try to find inner loops with calls that must
2863     // always be executed (call dominates loop tail).  These loops do
2864     // not need a separate safepoint.
2865     Node_List cisstack(a);
2866     _ltree_root->check_safepts(visited, cisstack);
2867   }
2868 
2869   // Walk the DATA nodes and place into loops.  Find earliest control
2870   // node.  For CFG nodes, the _nodes array starts out and remains
2871   // holding the associated IdealLoopTree pointer.  For DATA nodes, the
2872   // _nodes array holds the earliest legal controlling CFG node.
2873 
2874   // Allocate stack with enough space to avoid frequent realloc
2875   int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
2876   Node_Stack nstack( a, stack_size );
2877 
2878   visited.Clear();
2879   Node_List worklist(a);
2880   // Don't need C->root() on worklist since
2881   // it will be processed among C->top() inputs
2882   worklist.push( C->top() );
2883   visited.set( C->top()->_idx ); // Set C->top() as visited now
2884   build_loop_early( visited, worklist, nstack );
2885 
2886   // Given early legal placement, try finding counted loops.  This placement
2887   // is good enough to discover most loop invariants.
2888   if (!_verify_me && !_verify_only && !shenandoah_opts) {
2889     _ltree_root->counted_loop(this);
2890   }
2891 
2892   // Find latest loop placement.  Find ideal loop placement.
2893   visited.Clear();
2894   init_dom_lca_tags();
2895   // Need C->root() on worklist when processing outs
2896   worklist.push( C->root() );
2897   NOT_PRODUCT( C->verify_graph_edges(); )
2898   worklist.push( C->top() );
2899   build_loop_late(visited, worklist, nstack, !shenandoah_opts);
2900 
2901   if (_verify_only) {
2902     // restore major progress flag
2903     for (int i = 0; i < old_progress; i++)
2904       C->set_major_progress();
2905     assert(C->unique() == unique, "verification mode made Nodes? ? ?");
2906     assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
2907     return;
2908   }
2909 
2910   // clear out the dead code after build_loop_late
2911   while (_deadlist.size()) {
2912     _igvn.remove_globally_dead_node(_deadlist.pop());
2913   }
2914 
2915   if (stop_early) {
2916     assert(do_expensive_nodes, "why are we here?");
2917     if (process_expensive_nodes()) {
2918       // If we made some progress when processing expensive nodes then
2919       // the IGVN may modify the graph in a way that will allow us to
2920       // make some more progress: we need to try processing expensive
2921       // nodes again.
2922       C->set_major_progress();
2923     }
2924     _igvn.optimize();
2925     return;
2926   }
2927 
2928   // Some parser-inserted loop predicates could never be used by loop
2929   // predication or they were moved away from loop during some optimizations.
2930   // For example, peeling. Eliminate them before next loop optimizations.
2931   eliminate_useless_predicates();
2932 
2933 #ifndef PRODUCT
2934   C->verify_graph_edges();
2935   if (_verify_me) {             // Nested verify pass?
2936     // Check to see if the verify mode is broken
2937     assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
2938     return;
2939   }
2940   if(VerifyLoopOptimizations) verify();
2941   if(TraceLoopOpts && C->has_loops()) {
2942     _ltree_root->dump();
2943   }
2944 #endif
2945 
2946 #if INCLUDE_SHENANDOAHGC
2947   if (mode == LoopOptsShenandoahExpand) {
2948     assert(UseShenandoahGC, "only for shenandoah");
2949     ShenandoahWriteBarrierNode::pin_and_expand(this);
2950   } else if (mode == LoopOptsShenandoahPostExpand) {
2951     assert(UseShenandoahGC, "only for shenandoah");
2952     visited.Clear();
2953     ShenandoahWriteBarrierNode::optimize_after_expansion(visited, nstack, worklist, this);
2954   }
2955 
2956   if (shenandoah_opts) {
2957     _igvn.optimize();
2958     if (C->log() != NULL) {
2959       log_loop_tree(_ltree_root, _ltree_root, C->log());
2960     }
2961     return;
2962   }
2963 #endif
2964 
2965   if (skip_loop_opts) {
2966     // restore major progress flag
2967     for (int i = 0; i < old_progress; i++) {
2968       C->set_major_progress();
2969     }
2970 
2971     // Cleanup any modified bits
2972     _igvn.optimize();
2973 
2974     if (C->log() != NULL) {
2975       log_loop_tree(_ltree_root, _ltree_root, C->log());
2976     }
2977     return;
2978   }
2979 
2980 #if INCLUDE_SHENANDOAHGC
2981   if (UseShenandoahGC) {
2982     GrowableArray<MemoryGraphFixer*> memory_graph_fixers;
2983     ShenandoahWriteBarrierNode::optimize_before_expansion(this, memory_graph_fixers, false);
2984   }
2985 #endif
2986 
2987   if (ReassociateInvariants) {
2988     // Reassociate invariants and prep for split_thru_phi
2989     for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2990       IdealLoopTree* lpt = iter.current();
2991       bool is_counted = lpt->is_counted();
2992       if (!is_counted || !lpt->is_inner()) continue;
2993 
2994       // check for vectorized loops, any reassociation of invariants was already done
2995       if (is_counted && lpt->_head->as_CountedLoop()->do_unroll_only()) continue;
2996 
2997       lpt->reassociate_invariants(this);
2998 
2999       // Because RCE opportunities can be masked by split_thru_phi,
3000       // look for RCE candidates and inhibit split_thru_phi
3001       // on just their loop-phi's for this pass of loop opts
3002       if (SplitIfBlocks && do_split_ifs) {
3003         if (lpt->policy_range_check(this)) {
3004           lpt->_rce_candidate = 1; // = true
3005         }
3006       }
3007     }
3008   }
3009 
3010   // Check for aggressive application of split-if and other transforms
3011   // that require basic-block info (like cloning through Phi's)
3012   if( SplitIfBlocks && do_split_ifs ) {
3013     visited.Clear();
3014     split_if_with_blocks( visited, nstack, mode == LoopOptsZgcLastRound );
3015     NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
3016     if (mode == LoopOptsZgcLastRound) {
3017       C->set_major_progress();
3018     }
3019   }
3020 
3021   if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
3022     C->set_major_progress();
3023   }
3024 
3025   // Perform loop predication before iteration splitting
3026   if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
3027     _ltree_root->_child->loop_predication(this);
3028   }
3029 
3030   if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
3031     if (do_intrinsify_fill()) {
3032       C->set_major_progress();
3033     }
3034   }
3035 
3036   // Perform iteration-splitting on inner loops.  Split iterations to avoid
3037   // range checks or one-shot null checks.
3038 
3039   // If split-if's didn't hack the graph too bad (no CFG changes)
3040   // then do loop opts.
3041   if (C->has_loops() && !C->major_progress()) {
3042     memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
3043     _ltree_root->_child->iteration_split( this, worklist );
3044     // No verify after peeling!  GCM has hoisted code out of the loop.
3045     // After peeling, the hoisted code could sink inside the peeled area.
3046     // The peeling code does not try to recompute the best location for
3047     // all the code before the peeled area, so the verify pass will always
3048     // complain about it.
3049   }
3050   // Do verify graph edges in any case
3051   NOT_PRODUCT( C->verify_graph_edges(); );
3052 
3053   if (!do_split_ifs) {
3054     // We saw major progress in Split-If to get here.  We forced a
3055     // pass with unrolling and not split-if, however more split-if's
3056     // might make progress.  If the unrolling didn't make progress
3057     // then the major-progress flag got cleared and we won't try
3058     // another round of Split-If.  In particular the ever-common
3059     // instance-of/check-cast pattern requires at least 2 rounds of
3060     // Split-If to clear out.
3061     C->set_major_progress();
3062   }
3063 
3064   // Repeat loop optimizations if new loops were seen
3065   if (created_loop_node()) {
3066     C->set_major_progress();
3067   }
3068 
3069   // Keep loop predicates and perform optimizations with them
3070   // until no more loop optimizations could be done.
3071   // After that switch predicates off and do more loop optimizations.
3072   if (!C->major_progress() && (C->predicate_count() > 0)) {
3073      C->cleanup_loop_predicates(_igvn);
3074      if (TraceLoopOpts) {
3075        tty->print_cr("PredicatesOff");
3076      }
3077      C->set_major_progress();
3078   }
3079 
3080   // Convert scalar to superword operations at the end of all loop opts.
3081   if (UseSuperWord && C->has_loops() && !C->major_progress()) {
3082     // SuperWord transform
3083     SuperWord sw(this);
3084     for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
3085       IdealLoopTree* lpt = iter.current();
3086       if (lpt->is_counted()) {
3087         CountedLoopNode *cl = lpt->_head->as_CountedLoop();
3088 
3089         if (PostLoopMultiversioning && cl->is_rce_post_loop() && !cl->is_vectorized_loop()) {
3090           // Check that the rce'd post loop is encountered first, multiversion after all
3091           // major main loop optimization are concluded
3092           if (!C->major_progress()) {
3093             IdealLoopTree *lpt_next = lpt->_next;
3094             if (lpt_next && lpt_next->is_counted()) {
3095               CountedLoopNode *cl = lpt_next->_head->as_CountedLoop();
3096               has_range_checks(lpt_next);
3097               if (cl->is_post_loop() && cl->range_checks_present()) {
3098                 if (!cl->is_multiversioned()) {
3099                   if (multi_version_post_loops(lpt, lpt_next) == false) {
3100                     // Cause the rce loop to be optimized away if we fail
3101                     cl->mark_is_multiversioned();
3102                     cl->set_slp_max_unroll(0);
3103                     poison_rce_post_loop(lpt);
3104                   }
3105                 }
3106               }
3107             }
3108             sw.transform_loop(lpt, true);
3109           }
3110         } else if (cl->is_main_loop()) {
3111           sw.transform_loop(lpt, true);
3112         }
3113       }
3114     }
3115   }
3116 
3117   // Cleanup any modified bits
3118   _igvn.optimize();
3119 
3120   // disable assert until issue with split_flow_path is resolved (6742111)
3121   // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
3122   //        "shouldn't introduce irreducible loops");
3123 
3124   if (C->log() != NULL) {
3125     log_loop_tree(_ltree_root, _ltree_root, C->log());
3126   }
3127 }
3128 
3129 #ifndef PRODUCT
3130 //------------------------------print_statistics-------------------------------
3131 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
3132 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
3133 void PhaseIdealLoop::print_statistics() {
3134   tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
3135 }
3136 
3137 //------------------------------verify-----------------------------------------
3138 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
3139 static int fail;                // debug only, so its multi-thread dont care
3140 void PhaseIdealLoop::verify() const {
3141   int old_progress = C->major_progress();
3142   ResourceMark rm;
3143   PhaseIdealLoop loop_verify( _igvn, this );
3144   VectorSet visited(Thread::current()->resource_area());
3145 
3146   fail = 0;
3147   verify_compare( C->root(), &loop_verify, visited );
3148   assert( fail == 0, "verify loops failed" );
3149   // Verify loop structure is the same
3150   _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
3151   // Reset major-progress.  It was cleared by creating a verify version of
3152   // PhaseIdealLoop.
3153   for( int i=0; i<old_progress; i++ )
3154     C->set_major_progress();
3155 }
3156 
3157 //------------------------------verify_compare---------------------------------
3158 // Make sure me and the given PhaseIdealLoop agree on key data structures
3159 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
3160   if( !n ) return;
3161   if( visited.test_set( n->_idx ) ) return;
3162   if( !_nodes[n->_idx] ) {      // Unreachable
3163     assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
3164     return;
3165   }
3166 
3167   uint i;
3168   for( i = 0; i < n->req(); i++ )
3169     verify_compare( n->in(i), loop_verify, visited );
3170 
3171   // Check the '_nodes' block/loop structure
3172   i = n->_idx;
3173   if( has_ctrl(n) ) {           // We have control; verify has loop or ctrl
3174     if( _nodes[i] != loop_verify->_nodes[i] &&
3175         get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
3176       tty->print("Mismatched control setting for: ");
3177       n->dump();
3178       if( fail++ > 10 ) return;
3179       Node *c = get_ctrl_no_update(n);
3180       tty->print("We have it as: ");
3181       if( c->in(0) ) c->dump();
3182         else tty->print_cr("N%d",c->_idx);
3183       tty->print("Verify thinks: ");
3184       if( loop_verify->has_ctrl(n) )
3185         loop_verify->get_ctrl_no_update(n)->dump();
3186       else
3187         loop_verify->get_loop_idx(n)->dump();
3188       tty->cr();
3189     }
3190   } else {                    // We have a loop
3191     IdealLoopTree *us = get_loop_idx(n);
3192     if( loop_verify->has_ctrl(n) ) {
3193       tty->print("Mismatched loop setting for: ");
3194       n->dump();
3195       if( fail++ > 10 ) return;
3196       tty->print("We have it as: ");
3197       us->dump();
3198       tty->print("Verify thinks: ");
3199       loop_verify->get_ctrl_no_update(n)->dump();
3200       tty->cr();
3201     } else if (!C->major_progress()) {
3202       // Loop selection can be messed up if we did a major progress
3203       // operation, like split-if.  Do not verify in that case.
3204       IdealLoopTree *them = loop_verify->get_loop_idx(n);
3205       if( us->_head != them->_head ||  us->_tail != them->_tail ) {
3206         tty->print("Unequals loops for: ");
3207         n->dump();
3208         if( fail++ > 10 ) return;
3209         tty->print("We have it as: ");
3210         us->dump();
3211         tty->print("Verify thinks: ");
3212         them->dump();
3213         tty->cr();
3214       }
3215     }
3216   }
3217 
3218   // Check for immediate dominators being equal
3219   if( i >= _idom_size ) {
3220     if( !n->is_CFG() ) return;
3221     tty->print("CFG Node with no idom: ");
3222     n->dump();
3223     return;
3224   }
3225   if( !n->is_CFG() ) return;
3226   if( n == C->root() ) return; // No IDOM here
3227 
3228   assert(n->_idx == i, "sanity");
3229   Node *id = idom_no_update(n);
3230   if( id != loop_verify->idom_no_update(n) ) {
3231     tty->print("Unequals idoms for: ");
3232     n->dump();
3233     if( fail++ > 10 ) return;
3234     tty->print("We have it as: ");
3235     id->dump();
3236     tty->print("Verify thinks: ");
3237     loop_verify->idom_no_update(n)->dump();
3238     tty->cr();
3239   }
3240 
3241 }
3242 
3243 //------------------------------verify_tree------------------------------------
3244 // Verify that tree structures match.  Because the CFG can change, siblings
3245 // within the loop tree can be reordered.  We attempt to deal with that by
3246 // reordering the verify's loop tree if possible.
3247 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
3248   assert( _parent == parent, "Badly formed loop tree" );
3249 
3250   // Siblings not in same order?  Attempt to re-order.
3251   if( _head != loop->_head ) {
3252     // Find _next pointer to update
3253     IdealLoopTree **pp = &loop->_parent->_child;
3254     while( *pp != loop )
3255       pp = &((*pp)->_next);
3256     // Find proper sibling to be next
3257     IdealLoopTree **nn = &loop->_next;
3258     while( (*nn) && (*nn)->_head != _head )
3259       nn = &((*nn)->_next);
3260 
3261     // Check for no match.
3262     if( !(*nn) ) {
3263       // Annoyingly, irreducible loops can pick different headers
3264       // after a major_progress operation, so the rest of the loop
3265       // tree cannot be matched.
3266       if (_irreducible && Compile::current()->major_progress())  return;
3267       assert( 0, "failed to match loop tree" );
3268     }
3269 
3270     // Move (*nn) to (*pp)
3271     IdealLoopTree *hit = *nn;
3272     *nn = hit->_next;
3273     hit->_next = loop;
3274     *pp = loop;
3275     loop = hit;
3276     // Now try again to verify
3277   }
3278 
3279   assert( _head  == loop->_head , "mismatched loop head" );
3280   Node *tail = _tail;           // Inline a non-updating version of
3281   while( !tail->in(0) )         // the 'tail()' call.
3282     tail = tail->in(1);
3283   assert( tail == loop->_tail, "mismatched loop tail" );
3284 
3285   // Counted loops that are guarded should be able to find their guards
3286   if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
3287     CountedLoopNode *cl = _head->as_CountedLoop();
3288     Node *init = cl->init_trip();
3289     Node *ctrl = cl->in(LoopNode::EntryControl);
3290     assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
3291     Node *iff  = ctrl->in(0);
3292     assert( iff->Opcode() == Op_If, "" );
3293     Node *bol  = iff->in(1);
3294     assert( bol->Opcode() == Op_Bool, "" );
3295     Node *cmp  = bol->in(1);
3296     assert( cmp->Opcode() == Op_CmpI, "" );
3297     Node *add  = cmp->in(1);
3298     Node *opaq;
3299     if( add->Opcode() == Op_Opaque1 ) {
3300       opaq = add;
3301     } else {
3302       assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
3303       assert( add == init, "" );
3304       opaq = cmp->in(2);
3305     }
3306     assert( opaq->Opcode() == Op_Opaque1, "" );
3307 
3308   }
3309 
3310   if (_child != NULL)  _child->verify_tree(loop->_child, this);
3311   if (_next  != NULL)  _next ->verify_tree(loop->_next,  parent);
3312   // Innermost loops need to verify loop bodies,
3313   // but only if no 'major_progress'
3314   int fail = 0;
3315   if (!Compile::current()->major_progress() && _child == NULL) {
3316     for( uint i = 0; i < _body.size(); i++ ) {
3317       Node *n = _body.at(i);
3318       if (n->outcnt() == 0)  continue; // Ignore dead
3319       uint j;
3320       for( j = 0; j < loop->_body.size(); j++ )
3321         if( loop->_body.at(j) == n )
3322           break;
3323       if( j == loop->_body.size() ) { // Not found in loop body
3324         // Last ditch effort to avoid assertion: Its possible that we
3325         // have some users (so outcnt not zero) but are still dead.
3326         // Try to find from root.
3327         if (Compile::current()->root()->find(n->_idx)) {
3328           fail++;
3329           tty->print("We have that verify does not: ");
3330           n->dump();
3331         }
3332       }
3333     }
3334     for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
3335       Node *n = loop->_body.at(i2);
3336       if (n->outcnt() == 0)  continue; // Ignore dead
3337       uint j;
3338       for( j = 0; j < _body.size(); j++ )
3339         if( _body.at(j) == n )
3340           break;
3341       if( j == _body.size() ) { // Not found in loop body
3342         // Last ditch effort to avoid assertion: Its possible that we
3343         // have some users (so outcnt not zero) but are still dead.
3344         // Try to find from root.
3345         if (Compile::current()->root()->find(n->_idx)) {
3346           fail++;
3347           tty->print("Verify has that we do not: ");
3348           n->dump();
3349         }
3350       }
3351     }
3352     assert( !fail, "loop body mismatch" );
3353   }
3354 }
3355 
3356 #endif
3357 
3358 //------------------------------set_idom---------------------------------------
3359 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
3360   uint idx = d->_idx;
3361   if (idx >= _idom_size) {
3362     uint newsize = _idom_size<<1;
3363     while( idx >= newsize ) {
3364       newsize <<= 1;
3365     }
3366     _idom      = REALLOC_RESOURCE_ARRAY( Node*,     _idom,_idom_size,newsize);
3367     _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
3368     memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
3369     _idom_size = newsize;
3370   }
3371   _idom[idx] = n;
3372   _dom_depth[idx] = dom_depth;
3373 }
3374 
3375 //------------------------------recompute_dom_depth---------------------------------------
3376 // The dominator tree is constructed with only parent pointers.
3377 // This recomputes the depth in the tree by first tagging all
3378 // nodes as "no depth yet" marker.  The next pass then runs up
3379 // the dom tree from each node marked "no depth yet", and computes
3380 // the depth on the way back down.
3381 void PhaseIdealLoop::recompute_dom_depth() {
3382   uint no_depth_marker = C->unique();
3383   uint i;
3384   // Initialize depth to "no depth yet" and realize all lazy updates
3385   for (i = 0; i < _idom_size; i++) {
3386     // Only indices with a _dom_depth has a Node* or NULL (otherwise uninitalized).
3387     if (_dom_depth[i] > 0 && _idom[i] != NULL) {
3388       _dom_depth[i] = no_depth_marker;
3389 
3390       // heal _idom if it has a fwd mapping in _nodes
3391       if (_idom[i]->in(0) == NULL) {
3392         idom(i);
3393       }
3394     }
3395   }
3396   if (_dom_stk == NULL) {
3397     uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
3398     if (init_size < 10) init_size = 10;
3399     _dom_stk = new GrowableArray<uint>(init_size);
3400   }
3401   // Compute new depth for each node.
3402   for (i = 0; i < _idom_size; i++) {
3403     uint j = i;
3404     // Run up the dom tree to find a node with a depth
3405     while (_dom_depth[j] == no_depth_marker) {
3406       _dom_stk->push(j);
3407       j = _idom[j]->_idx;
3408     }
3409     // Compute the depth on the way back down this tree branch
3410     uint dd = _dom_depth[j] + 1;
3411     while (_dom_stk->length() > 0) {
3412       uint j = _dom_stk->pop();
3413       _dom_depth[j] = dd;
3414       dd++;
3415     }
3416   }
3417 }
3418 
3419 //------------------------------sort-------------------------------------------
3420 // Insert 'loop' into the existing loop tree.  'innermost' is a leaf of the
3421 // loop tree, not the root.
3422 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
3423   if( !innermost ) return loop; // New innermost loop
3424 
3425   int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
3426   assert( loop_preorder, "not yet post-walked loop" );
3427   IdealLoopTree **pp = &innermost;      // Pointer to previous next-pointer
3428   IdealLoopTree *l = *pp;               // Do I go before or after 'l'?
3429 
3430   // Insert at start of list
3431   while( l ) {                  // Insertion sort based on pre-order
3432     if( l == loop ) return innermost; // Already on list!
3433     int l_preorder = get_preorder(l->_head); // Cache pre-order number
3434     assert( l_preorder, "not yet post-walked l" );
3435     // Check header pre-order number to figure proper nesting
3436     if( loop_preorder > l_preorder )
3437       break;                    // End of insertion
3438     // If headers tie (e.g., shared headers) check tail pre-order numbers.
3439     // Since I split shared headers, you'd think this could not happen.
3440     // BUT: I must first do the preorder numbering before I can discover I
3441     // have shared headers, so the split headers all get the same preorder
3442     // number as the RegionNode they split from.
3443     if( loop_preorder == l_preorder &&
3444         get_preorder(loop->_tail) < get_preorder(l->_tail) )
3445       break;                    // Also check for shared headers (same pre#)
3446     pp = &l->_parent;           // Chain up list
3447     l = *pp;
3448   }
3449   // Link into list
3450   // Point predecessor to me
3451   *pp = loop;
3452   // Point me to successor
3453   IdealLoopTree *p = loop->_parent;
3454   loop->_parent = l;            // Point me to successor
3455   if( p ) sort( p, innermost ); // Insert my parents into list as well
3456   return innermost;
3457 }
3458 
3459 //------------------------------build_loop_tree--------------------------------
3460 // I use a modified Vick/Tarjan algorithm.  I need pre- and a post- visit
3461 // bits.  The _nodes[] array is mapped by Node index and holds a NULL for
3462 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
3463 // tightest enclosing IdealLoopTree for post-walked.
3464 //
3465 // During my forward walk I do a short 1-layer lookahead to see if I can find
3466 // a loop backedge with that doesn't have any work on the backedge.  This
3467 // helps me construct nested loops with shared headers better.
3468 //
3469 // Once I've done the forward recursion, I do the post-work.  For each child
3470 // I check to see if there is a backedge.  Backedges define a loop!  I
3471 // insert an IdealLoopTree at the target of the backedge.
3472 //
3473 // During the post-work I also check to see if I have several children
3474 // belonging to different loops.  If so, then this Node is a decision point
3475 // where control flow can choose to change loop nests.  It is at this
3476 // decision point where I can figure out how loops are nested.  At this
3477 // time I can properly order the different loop nests from my children.
3478 // Note that there may not be any backedges at the decision point!
3479 //
3480 // Since the decision point can be far removed from the backedges, I can't
3481 // order my loops at the time I discover them.  Thus at the decision point
3482 // I need to inspect loop header pre-order numbers to properly nest my
3483 // loops.  This means I need to sort my childrens' loops by pre-order.
3484 // The sort is of size number-of-control-children, which generally limits
3485 // it to size 2 (i.e., I just choose between my 2 target loops).
3486 void PhaseIdealLoop::build_loop_tree() {
3487   // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3488   GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
3489   Node *n = C->root();
3490   bltstack.push(n);
3491   int pre_order = 1;
3492   int stack_size;
3493 
3494   while ( ( stack_size = bltstack.length() ) != 0 ) {
3495     n = bltstack.top(); // Leave node on stack
3496     if ( !is_visited(n) ) {
3497       // ---- Pre-pass Work ----
3498       // Pre-walked but not post-walked nodes need a pre_order number.
3499 
3500       set_preorder_visited( n, pre_order ); // set as visited
3501 
3502       // ---- Scan over children ----
3503       // Scan first over control projections that lead to loop headers.
3504       // This helps us find inner-to-outer loops with shared headers better.
3505 
3506       // Scan children's children for loop headers.
3507       for ( int i = n->outcnt() - 1; i >= 0; --i ) {
3508         Node* m = n->raw_out(i);       // Child
3509         if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
3510           // Scan over children's children to find loop
3511           for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3512             Node* l = m->fast_out(j);
3513             if( is_visited(l) &&       // Been visited?
3514                 !is_postvisited(l) &&  // But not post-visited
3515                 get_preorder(l) < pre_order ) { // And smaller pre-order
3516               // Found!  Scan the DFS down this path before doing other paths
3517               bltstack.push(m);
3518               break;
3519             }
3520           }
3521         }
3522       }
3523       pre_order++;
3524     }
3525     else if ( !is_postvisited(n) ) {
3526       // Note: build_loop_tree_impl() adds out edges on rare occasions,
3527       // such as com.sun.rsasign.am::a.
3528       // For non-recursive version, first, process current children.
3529       // On next iteration, check if additional children were added.
3530       for ( int k = n->outcnt() - 1; k >= 0; --k ) {
3531         Node* u = n->raw_out(k);
3532         if ( u->is_CFG() && !is_visited(u) ) {
3533           bltstack.push(u);
3534         }
3535       }
3536       if ( bltstack.length() == stack_size ) {
3537         // There were no additional children, post visit node now
3538         (void)bltstack.pop(); // Remove node from stack
3539         pre_order = build_loop_tree_impl( n, pre_order );
3540         // Check for bailout
3541         if (C->failing()) {
3542           return;
3543         }
3544         // Check to grow _preorders[] array for the case when
3545         // build_loop_tree_impl() adds new nodes.
3546         check_grow_preorders();
3547       }
3548     }
3549     else {
3550       (void)bltstack.pop(); // Remove post-visited node from stack
3551     }
3552   }
3553 }
3554 
3555 //------------------------------build_loop_tree_impl---------------------------
3556 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
3557   // ---- Post-pass Work ----
3558   // Pre-walked but not post-walked nodes need a pre_order number.
3559 
3560   // Tightest enclosing loop for this Node
3561   IdealLoopTree *innermost = NULL;
3562 
3563   // For all children, see if any edge is a backedge.  If so, make a loop
3564   // for it.  Then find the tightest enclosing loop for the self Node.
3565   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3566     Node* m = n->fast_out(i);   // Child
3567     if( n == m ) continue;      // Ignore control self-cycles
3568     if( !m->is_CFG() ) continue;// Ignore non-CFG edges
3569 
3570     IdealLoopTree *l;           // Child's loop
3571     if( !is_postvisited(m) ) {  // Child visited but not post-visited?
3572       // Found a backedge
3573       assert( get_preorder(m) < pre_order, "should be backedge" );
3574       // Check for the RootNode, which is already a LoopNode and is allowed
3575       // to have multiple "backedges".
3576       if( m == C->root()) {     // Found the root?
3577         l = _ltree_root;        // Root is the outermost LoopNode
3578       } else {                  // Else found a nested loop
3579         // Insert a LoopNode to mark this loop.
3580         l = new IdealLoopTree(this, m, n);
3581       } // End of Else found a nested loop
3582       if( !has_loop(m) )        // If 'm' does not already have a loop set
3583         set_loop(m, l);         // Set loop header to loop now
3584 
3585     } else {                    // Else not a nested loop
3586       if( !_nodes[m->_idx] ) continue; // Dead code has no loop
3587       l = get_loop(m);          // Get previously determined loop
3588       // If successor is header of a loop (nest), move up-loop till it
3589       // is a member of some outer enclosing loop.  Since there are no
3590       // shared headers (I've split them already) I only need to go up
3591       // at most 1 level.
3592       while( l && l->_head == m ) // Successor heads loop?
3593         l = l->_parent;         // Move up 1 for me
3594       // If this loop is not properly parented, then this loop
3595       // has no exit path out, i.e. its an infinite loop.
3596       if( !l ) {
3597         // Make loop "reachable" from root so the CFG is reachable.  Basically
3598         // insert a bogus loop exit that is never taken.  'm', the loop head,
3599         // points to 'n', one (of possibly many) fall-in paths.  There may be
3600         // many backedges as well.
3601 
3602         // Here I set the loop to be the root loop.  I could have, after
3603         // inserting a bogus loop exit, restarted the recursion and found my
3604         // new loop exit.  This would make the infinite loop a first-class
3605         // loop and it would then get properly optimized.  What's the use of
3606         // optimizing an infinite loop?
3607         l = _ltree_root;        // Oops, found infinite loop
3608 
3609         if (!_verify_only) {
3610           // Insert the NeverBranch between 'm' and it's control user.
3611           NeverBranchNode *iff = new NeverBranchNode( m );
3612           _igvn.register_new_node_with_optimizer(iff);
3613           set_loop(iff, l);
3614           Node *if_t = new CProjNode( iff, 0 );
3615           _igvn.register_new_node_with_optimizer(if_t);
3616           set_loop(if_t, l);
3617 
3618           Node* cfg = NULL;       // Find the One True Control User of m
3619           for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3620             Node* x = m->fast_out(j);
3621             if (x->is_CFG() && x != m && x != iff)
3622               { cfg = x; break; }
3623           }
3624           assert(cfg != NULL, "must find the control user of m");
3625           uint k = 0;             // Probably cfg->in(0)
3626           while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
3627           cfg->set_req( k, if_t ); // Now point to NeverBranch
3628           _igvn._worklist.push(cfg);
3629 
3630           // Now create the never-taken loop exit
3631           Node *if_f = new CProjNode( iff, 1 );
3632           _igvn.register_new_node_with_optimizer(if_f);
3633           set_loop(if_f, l);
3634           // Find frame ptr for Halt.  Relies on the optimizer
3635           // V-N'ing.  Easier and quicker than searching through
3636           // the program structure.
3637           Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr );
3638           _igvn.register_new_node_with_optimizer(frame);
3639           // Halt & Catch Fire
3640           Node *halt = new HaltNode( if_f, frame );
3641           _igvn.register_new_node_with_optimizer(halt);
3642           set_loop(halt, l);
3643           C->root()->add_req(halt);
3644         }
3645         set_loop(C->root(), _ltree_root);
3646       }
3647     }
3648     // Weeny check for irreducible.  This child was already visited (this
3649     // IS the post-work phase).  Is this child's loop header post-visited
3650     // as well?  If so, then I found another entry into the loop.
3651     if (!_verify_only) {
3652       while( is_postvisited(l->_head) ) {
3653         // found irreducible
3654         l->_irreducible = 1; // = true
3655         l = l->_parent;
3656         _has_irreducible_loops = true;
3657         // Check for bad CFG here to prevent crash, and bailout of compile
3658         if (l == NULL) {
3659           C->record_method_not_compilable("unhandled CFG detected during loop optimization");
3660           return pre_order;
3661         }
3662       }
3663       C->set_has_irreducible_loop(_has_irreducible_loops);
3664     }
3665 
3666     // This Node might be a decision point for loops.  It is only if
3667     // it's children belong to several different loops.  The sort call
3668     // does a trivial amount of work if there is only 1 child or all
3669     // children belong to the same loop.  If however, the children
3670     // belong to different loops, the sort call will properly set the
3671     // _parent pointers to show how the loops nest.
3672     //
3673     // In any case, it returns the tightest enclosing loop.
3674     innermost = sort( l, innermost );
3675   }
3676 
3677   // Def-use info will have some dead stuff; dead stuff will have no
3678   // loop decided on.
3679 
3680   // Am I a loop header?  If so fix up my parent's child and next ptrs.
3681   if( innermost && innermost->_head == n ) {
3682     assert( get_loop(n) == innermost, "" );
3683     IdealLoopTree *p = innermost->_parent;
3684     IdealLoopTree *l = innermost;
3685     while( p && l->_head == n ) {
3686       l->_next = p->_child;     // Put self on parents 'next child'
3687       p->_child = l;            // Make self as first child of parent
3688       l = p;                    // Now walk up the parent chain
3689       p = l->_parent;
3690     }
3691   } else {
3692     // Note that it is possible for a LoopNode to reach here, if the
3693     // backedge has been made unreachable (hence the LoopNode no longer
3694     // denotes a Loop, and will eventually be removed).
3695 
3696     // Record tightest enclosing loop for self.  Mark as post-visited.
3697     set_loop(n, innermost);
3698     // Also record has_call flag early on
3699     if( innermost ) {
3700       if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
3701         // Do not count uncommon calls
3702         if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
3703           Node *iff = n->in(0)->in(0);
3704           // No any calls for vectorized loops.
3705           if( UseSuperWord || !iff->is_If() ||
3706               (n->in(0)->Opcode() == Op_IfFalse &&
3707                (1.0 - iff->as_If()->_prob) >= 0.01) ||
3708               (iff->as_If()->_prob >= 0.01) )
3709             innermost->_has_call = 1;
3710         }
3711       } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
3712         // Disable loop optimizations if the loop has a scalar replaceable
3713         // allocation. This disabling may cause a potential performance lost
3714         // if the allocation is not eliminated for some reason.
3715         innermost->_allow_optimizations = false;
3716         innermost->_has_call = 1; // = true
3717       } else if (n->Opcode() == Op_SafePoint) {
3718         // Record all safepoints in this loop.
3719         if (innermost->_safepts == NULL) innermost->_safepts = new Node_List();
3720         innermost->_safepts->push(n);
3721       }
3722     }
3723   }
3724 
3725   // Flag as post-visited now
3726   set_postvisited(n);
3727   return pre_order;
3728 }
3729 
3730 
3731 //------------------------------build_loop_early-------------------------------
3732 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3733 // First pass computes the earliest controlling node possible.  This is the
3734 // controlling input with the deepest dominating depth.
3735 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3736   while (worklist.size() != 0) {
3737     // Use local variables nstack_top_n & nstack_top_i to cache values
3738     // on nstack's top.
3739     Node *nstack_top_n = worklist.pop();
3740     uint  nstack_top_i = 0;
3741 //while_nstack_nonempty:
3742     while (true) {
3743       // Get parent node and next input's index from stack's top.
3744       Node  *n = nstack_top_n;
3745       uint   i = nstack_top_i;
3746       uint cnt = n->req(); // Count of inputs
3747       if (i == 0) {        // Pre-process the node.
3748         if( has_node(n) &&            // Have either loop or control already?
3749             !has_ctrl(n) ) {          // Have loop picked out already?
3750           // During "merge_many_backedges" we fold up several nested loops
3751           // into a single loop.  This makes the members of the original
3752           // loop bodies pointing to dead loops; they need to move up
3753           // to the new UNION'd larger loop.  I set the _head field of these
3754           // dead loops to NULL and the _parent field points to the owning
3755           // loop.  Shades of UNION-FIND algorithm.
3756           IdealLoopTree *ilt;
3757           while( !(ilt = get_loop(n))->_head ) {
3758             // Normally I would use a set_loop here.  But in this one special
3759             // case, it is legal (and expected) to change what loop a Node
3760             // belongs to.
3761             _nodes.map(n->_idx, (Node*)(ilt->_parent) );
3762           }
3763           // Remove safepoints ONLY if I've already seen I don't need one.
3764           // (the old code here would yank a 2nd safepoint after seeing a
3765           // first one, even though the 1st did not dominate in the loop body
3766           // and thus could be avoided indefinitely)
3767           if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
3768               is_deleteable_safept(n)) {
3769             Node *in = n->in(TypeFunc::Control);
3770             lazy_replace(n,in);       // Pull safepoint now
3771             if (ilt->_safepts != NULL) {
3772               ilt->_safepts->yank(n);
3773             }
3774             // Carry on with the recursion "as if" we are walking
3775             // only the control input
3776             if( !visited.test_set( in->_idx ) ) {
3777               worklist.push(in);      // Visit this guy later, using worklist
3778             }
3779             // Get next node from nstack:
3780             // - skip n's inputs processing by setting i > cnt;
3781             // - we also will not call set_early_ctrl(n) since
3782             //   has_node(n) == true (see the condition above).
3783             i = cnt + 1;
3784           }
3785         }
3786       } // if (i == 0)
3787 
3788       // Visit all inputs
3789       bool done = true;       // Assume all n's inputs will be processed
3790       while (i < cnt) {
3791         Node *in = n->in(i);
3792         ++i;
3793         if (in == NULL) continue;
3794         if (in->pinned() && !in->is_CFG())
3795           set_ctrl(in, in->in(0));
3796         int is_visited = visited.test_set( in->_idx );
3797         if (!has_node(in)) {  // No controlling input yet?
3798           assert( !in->is_CFG(), "CFG Node with no controlling input?" );
3799           assert( !is_visited, "visit only once" );
3800           nstack.push(n, i);  // Save parent node and next input's index.
3801           nstack_top_n = in;  // Process current input now.
3802           nstack_top_i = 0;
3803           done = false;       // Not all n's inputs processed.
3804           break; // continue while_nstack_nonempty;
3805         } else if (!is_visited) {
3806           // This guy has a location picked out for him, but has not yet
3807           // been visited.  Happens to all CFG nodes, for instance.
3808           // Visit him using the worklist instead of recursion, to break
3809           // cycles.  Since he has a location already we do not need to
3810           // find his location before proceeding with the current Node.
3811           worklist.push(in);  // Visit this guy later, using worklist
3812         }
3813       }
3814       if (done) {
3815         // All of n's inputs have been processed, complete post-processing.
3816 
3817         // Compute earliest point this Node can go.
3818         // CFG, Phi, pinned nodes already know their controlling input.
3819         if (!has_node(n)) {
3820           // Record earliest legal location
3821           set_early_ctrl( n );
3822         }
3823         if (nstack.is_empty()) {
3824           // Finished all nodes on stack.
3825           // Process next node on the worklist.
3826           break;
3827         }
3828         // Get saved parent node and next input's index.
3829         nstack_top_n = nstack.node();
3830         nstack_top_i = nstack.index();
3831         nstack.pop();
3832       }
3833     } // while (true)
3834   }
3835 }
3836 
3837 //------------------------------dom_lca_internal--------------------------------
3838 // Pair-wise LCA
3839 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
3840   if( !n1 ) return n2;          // Handle NULL original LCA
3841   assert( n1->is_CFG(), "" );
3842   assert( n2->is_CFG(), "" );
3843   // find LCA of all uses
3844   uint d1 = dom_depth(n1);
3845   uint d2 = dom_depth(n2);
3846   while (n1 != n2) {
3847     if (d1 > d2) {
3848       n1 =      idom(n1);
3849       d1 = dom_depth(n1);
3850     } else if (d1 < d2) {
3851       n2 =      idom(n2);
3852       d2 = dom_depth(n2);
3853     } else {
3854       // Here d1 == d2.  Due to edits of the dominator-tree, sections
3855       // of the tree might have the same depth.  These sections have
3856       // to be searched more carefully.
3857 
3858       // Scan up all the n1's with equal depth, looking for n2.
3859       Node *t1 = idom(n1);
3860       while (dom_depth(t1) == d1) {
3861         if (t1 == n2)  return n2;
3862         t1 = idom(t1);
3863       }
3864       // Scan up all the n2's with equal depth, looking for n1.
3865       Node *t2 = idom(n2);
3866       while (dom_depth(t2) == d2) {
3867         if (t2 == n1)  return n1;
3868         t2 = idom(t2);
3869       }
3870       // Move up to a new dominator-depth value as well as up the dom-tree.
3871       n1 = t1;
3872       n2 = t2;
3873       d1 = dom_depth(n1);
3874       d2 = dom_depth(n2);
3875     }
3876   }
3877   return n1;
3878 }
3879 
3880 //------------------------------compute_idom-----------------------------------
3881 // Locally compute IDOM using dom_lca call.  Correct only if the incoming
3882 // IDOMs are correct.
3883 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
3884   assert( region->is_Region(), "" );
3885   Node *LCA = NULL;
3886   for( uint i = 1; i < region->req(); i++ ) {
3887     if( region->in(i) != C->top() )
3888       LCA = dom_lca( LCA, region->in(i) );
3889   }
3890   return LCA;
3891 }
3892 
3893 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
3894   bool had_error = false;
3895 #ifdef ASSERT
3896   if (early != C->root()) {
3897     // Make sure that there's a dominance path from LCA to early
3898     Node* d = LCA;
3899     while (d != early) {
3900       if (d == C->root()) {
3901         dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
3902         tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
3903         had_error = true;
3904         break;
3905       }
3906       d = idom(d);
3907     }
3908   }
3909 #endif
3910   return had_error;
3911 }
3912 
3913 
3914 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
3915   // Compute LCA over list of uses
3916   bool had_error = false;
3917   Node *LCA = NULL;
3918   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
3919     Node* c = n->fast_out(i);
3920     if (_nodes[c->_idx] == NULL)
3921       continue;                 // Skip the occasional dead node
3922     if( c->is_Phi() ) {         // For Phis, we must land above on the path
3923       for( uint j=1; j<c->req(); j++ ) {// For all inputs
3924         if( c->in(j) == n ) {   // Found matching input?
3925           Node *use = c->in(0)->in(j);
3926           if (_verify_only && use->is_top()) continue;
3927           LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3928           if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3929         }
3930       }
3931     } else {
3932       // For CFG data-users, use is in the block just prior
3933       Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
3934       LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3935       if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3936     }
3937   }
3938   assert(!had_error, "bad dominance");
3939   return LCA;
3940 }
3941 
3942 // Check the shape of the graph at the loop entry. In some cases,
3943 // the shape of the graph does not match the shape outlined below.
3944 // That is caused by the Opaque1 node "protecting" the shape of
3945 // the graph being removed by, for example, the IGVN performed
3946 // in PhaseIdealLoop::build_and_optimize().
3947 //
3948 // After the Opaque1 node has been removed, optimizations (e.g., split-if,
3949 // loop unswitching, and IGVN, or a combination of them) can freely change
3950 // the graph's shape. As a result, the graph shape outlined below cannot
3951 // be guaranteed anymore.
3952 bool PhaseIdealLoop::is_canonical_loop_entry(CountedLoopNode* cl) {
3953   if (!cl->is_main_loop() && !cl->is_post_loop()) {
3954     return false;
3955   }
3956   Node* ctrl = cl->skip_predicates();
3957 
3958   if (ctrl == NULL || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) {
3959     return false;
3960   }
3961   Node* iffm = ctrl->in(0);
3962   if (iffm == NULL || !iffm->is_If()) {
3963     return false;
3964   }
3965   Node* bolzm = iffm->in(1);
3966   if (bolzm == NULL || !bolzm->is_Bool()) {
3967     return false;
3968   }
3969   Node* cmpzm = bolzm->in(1);
3970   if (cmpzm == NULL || !cmpzm->is_Cmp()) {
3971     return false;
3972   }
3973   // compares can get conditionally flipped
3974   bool found_opaque = false;
3975   for (uint i = 1; i < cmpzm->req(); i++) {
3976     Node* opnd = cmpzm->in(i);
3977     if (opnd && opnd->Opcode() == Op_Opaque1) {
3978       found_opaque = true;
3979       break;
3980     }
3981   }
3982   if (!found_opaque) {
3983     return false;
3984   }
3985   return true;
3986 }
3987 
3988 //------------------------------get_late_ctrl----------------------------------
3989 // Compute latest legal control.
3990 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
3991   assert(early != NULL, "early control should not be NULL");
3992 
3993   Node* LCA = compute_lca_of_uses(n, early);
3994 #ifdef ASSERT
3995   if (LCA == C->root() && LCA != early) {
3996     // def doesn't dominate uses so print some useful debugging output
3997     compute_lca_of_uses(n, early, true);
3998   }
3999 #endif
4000 
4001   // if this is a load, check for anti-dependent stores
4002   // We use a conservative algorithm to identify potential interfering
4003   // instructions and for rescheduling the load.  The users of the memory
4004   // input of this load are examined.  Any use which is not a load and is
4005   // dominated by early is considered a potentially interfering store.
4006   // This can produce false positives.
4007   if (n->is_Load() && LCA != early) {
4008     Node_List worklist;
4009 
4010     Node *mem = n->in(MemNode::Memory);
4011     for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
4012       Node* s = mem->fast_out(i);
4013       worklist.push(s);
4014     }
4015     while(worklist.size() != 0 && LCA != early) {
4016       Node* s = worklist.pop();
4017       if (s->is_Load() || s->is_ShenandoahBarrier() || s->Opcode() == Op_SafePoint ||
4018           (UseShenandoahGC && s->is_CallStaticJava() && s->as_CallStaticJava()->uncommon_trap_request() != 0)) {
4019         continue;
4020       } else if (s->is_MergeMem()) {
4021         for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
4022           Node* s1 = s->fast_out(i);
4023           worklist.push(s1);
4024         }
4025       } else {
4026         Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
4027         assert(sctrl != NULL || s->outcnt() == 0, "must have control");
4028         if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
4029           LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
4030         }
4031       }
4032     }
4033   }
4034 
4035   assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
4036   return LCA;
4037 }
4038 
4039 // true if CFG node d dominates CFG node n
4040 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
4041   if (d == n)
4042     return true;
4043   assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
4044   uint dd = dom_depth(d);
4045   while (dom_depth(n) >= dd) {
4046     if (n == d)
4047       return true;
4048     n = idom(n);
4049   }
4050   return false;
4051 }
4052 
4053 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
4054 // Pair-wise LCA with tags.
4055 // Tag each index with the node 'tag' currently being processed
4056 // before advancing up the dominator chain using idom().
4057 // Later calls that find a match to 'tag' know that this path has already
4058 // been considered in the current LCA (which is input 'n1' by convention).
4059 // Since get_late_ctrl() is only called once for each node, the tag array
4060 // does not need to be cleared between calls to get_late_ctrl().
4061 // Algorithm trades a larger constant factor for better asymptotic behavior
4062 //
4063 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
4064   uint d1 = dom_depth(n1);
4065   uint d2 = dom_depth(n2);
4066 
4067   do {
4068     if (d1 > d2) {
4069       // current lca is deeper than n2
4070       _dom_lca_tags.map(n1->_idx, tag);
4071       n1 =      idom(n1);
4072       d1 = dom_depth(n1);
4073     } else if (d1 < d2) {
4074       // n2 is deeper than current lca
4075       Node *memo = _dom_lca_tags[n2->_idx];
4076       if( memo == tag ) {
4077         return n1;    // Return the current LCA
4078       }
4079       _dom_lca_tags.map(n2->_idx, tag);
4080       n2 =      idom(n2);
4081       d2 = dom_depth(n2);
4082     } else {
4083       // Here d1 == d2.  Due to edits of the dominator-tree, sections
4084       // of the tree might have the same depth.  These sections have
4085       // to be searched more carefully.
4086 
4087       // Scan up all the n1's with equal depth, looking for n2.
4088       _dom_lca_tags.map(n1->_idx, tag);
4089       Node *t1 = idom(n1);
4090       while (dom_depth(t1) == d1) {
4091         if (t1 == n2)  return n2;
4092         _dom_lca_tags.map(t1->_idx, tag);
4093         t1 = idom(t1);
4094       }
4095       // Scan up all the n2's with equal depth, looking for n1.
4096       _dom_lca_tags.map(n2->_idx, tag);
4097       Node *t2 = idom(n2);
4098       while (dom_depth(t2) == d2) {
4099         if (t2 == n1)  return n1;
4100         _dom_lca_tags.map(t2->_idx, tag);
4101         t2 = idom(t2);
4102       }
4103       // Move up to a new dominator-depth value as well as up the dom-tree.
4104       n1 = t1;
4105       n2 = t2;
4106       d1 = dom_depth(n1);
4107       d2 = dom_depth(n2);
4108     }
4109   } while (n1 != n2);
4110   return n1;
4111 }
4112 
4113 //------------------------------init_dom_lca_tags------------------------------
4114 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
4115 // Intended use does not involve any growth for the array, so it could
4116 // be of fixed size.
4117 void PhaseIdealLoop::init_dom_lca_tags() {
4118   uint limit = C->unique() + 1;
4119   _dom_lca_tags.map( limit, NULL );
4120 #ifdef ASSERT
4121   for( uint i = 0; i < limit; ++i ) {
4122     assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
4123   }
4124 #endif // ASSERT
4125 }
4126 
4127 //------------------------------clear_dom_lca_tags------------------------------
4128 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
4129 // Intended use does not involve any growth for the array, so it could
4130 // be of fixed size.
4131 void PhaseIdealLoop::clear_dom_lca_tags() {
4132   uint limit = C->unique() + 1;
4133   _dom_lca_tags.map( limit, NULL );
4134   _dom_lca_tags.clear();
4135 #ifdef ASSERT
4136   for( uint i = 0; i < limit; ++i ) {
4137     assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
4138   }
4139 #endif // ASSERT
4140 }
4141 
4142 //------------------------------build_loop_late--------------------------------
4143 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
4144 // Second pass finds latest legal placement, and ideal loop placement.
4145 void PhaseIdealLoop::build_loop_late(VectorSet &visited, Node_List &worklist, Node_Stack &nstack, bool verify_strip_mined) {
4146   while (worklist.size() != 0) {
4147     Node *n = worklist.pop();
4148     // Only visit once
4149     if (visited.test_set(n->_idx)) continue;
4150     uint cnt = n->outcnt();
4151     uint   i = 0;
4152     while (true) {
4153       assert( _nodes[n->_idx], "no dead nodes" );
4154       // Visit all children
4155       if (i < cnt) {
4156         Node* use = n->raw_out(i);
4157         ++i;
4158         // Check for dead uses.  Aggressively prune such junk.  It might be
4159         // dead in the global sense, but still have local uses so I cannot
4160         // easily call 'remove_dead_node'.
4161         if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
4162           // Due to cycles, we might not hit the same fixed point in the verify
4163           // pass as we do in the regular pass.  Instead, visit such phis as
4164           // simple uses of the loop head.
4165           if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
4166             if( !visited.test(use->_idx) )
4167               worklist.push(use);
4168           } else if( !visited.test_set(use->_idx) ) {
4169             nstack.push(n, i); // Save parent and next use's index.
4170             n   = use;         // Process all children of current use.
4171             cnt = use->outcnt();
4172             i   = 0;
4173           }
4174         } else {
4175           // Do not visit around the backedge of loops via data edges.
4176           // push dead code onto a worklist
4177           _deadlist.push(use);
4178         }
4179       } else {
4180         // All of n's children have been processed, complete post-processing.
4181         build_loop_late_post(n, verify_strip_mined);
4182         if (nstack.is_empty()) {
4183           // Finished all nodes on stack.
4184           // Process next node on the worklist.
4185           break;
4186         }
4187         // Get saved parent node and next use's index. Visit the rest of uses.
4188         n   = nstack.node();
4189         cnt = n->outcnt();
4190         i   = nstack.index();
4191         nstack.pop();
4192       }
4193     }
4194   }
4195 }
4196 
4197 // Verify that no data node is schedules in the outer loop of a strip
4198 // mined loop.
4199 void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) {
4200 #ifdef ASSERT
4201   if (get_loop(least)->_nest == 0) {
4202     return;
4203   }
4204   IdealLoopTree* loop = get_loop(least);
4205   Node* head = loop->_head;
4206   if (head->is_OuterStripMinedLoop()) {
4207     Node* sfpt = head->as_Loop()->outer_safepoint();
4208     ResourceMark rm;
4209     Unique_Node_List wq;
4210     wq.push(sfpt);
4211     for (uint i = 0; i < wq.size(); i++) {
4212       Node *m = wq.at(i);
4213       for (uint i = 1; i < m->req(); i++) {
4214         Node* nn = m->in(i);
4215         if (nn == n) {
4216           return;
4217         }
4218         if (nn != NULL && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) {
4219           wq.push(nn);
4220         }
4221       }
4222     }
4223     ShouldNotReachHere();
4224   }
4225 #endif
4226 }
4227 
4228 
4229 //------------------------------build_loop_late_post---------------------------
4230 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
4231 // Second pass finds latest legal placement, and ideal loop placement.
4232 void PhaseIdealLoop::build_loop_late_post(Node *n, bool verify_strip_mined) {
4233 
4234   if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) {
4235     _igvn._worklist.push(n);  // Maybe we'll normalize it, if no more loops.
4236   }
4237 
4238 #ifdef ASSERT
4239   if (_verify_only && !n->is_CFG()) {
4240     // Check def-use domination.
4241     compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
4242   }
4243 #endif
4244 
4245   // CFG and pinned nodes already handled
4246   if( n->in(0) ) {
4247     if( n->in(0)->is_top() ) return; // Dead?
4248 
4249     // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
4250     // _must_ be pinned (they have to observe their control edge of course).
4251     // Unlike Stores (which modify an unallocable resource, the memory
4252     // state), Mods/Loads can float around.  So free them up.
4253     bool pinned = true;
4254     switch( n->Opcode() ) {
4255     case Op_DivI:
4256     case Op_DivF:
4257     case Op_DivD:
4258     case Op_ModI:
4259     case Op_ModF:
4260     case Op_ModD:
4261     case Op_LoadB:              // Same with Loads; they can sink
4262     case Op_LoadUB:             // during loop optimizations.
4263     case Op_LoadUS:
4264     case Op_LoadD:
4265     case Op_LoadF:
4266     case Op_LoadI:
4267     case Op_LoadKlass:
4268     case Op_LoadNKlass:
4269     case Op_LoadL:
4270     case Op_LoadS:
4271     case Op_LoadP:
4272     case Op_LoadBarrierSlowReg:
4273     case Op_LoadBarrierWeakSlowReg:
4274     case Op_LoadN:
4275     case Op_LoadRange:
4276     case Op_LoadD_unaligned:
4277     case Op_LoadL_unaligned:
4278     case Op_StrComp:            // Does a bunch of load-like effects
4279     case Op_StrEquals:
4280     case Op_StrIndexOf:
4281     case Op_StrIndexOfChar:
4282     case Op_AryEq:
4283 #if INCLUDE_SHENANDOAHGC
4284     case Op_ShenandoahReadBarrier:
4285     case Op_ShenandoahWriteBarrier:
4286     case Op_ShenandoahWBMemProj:
4287 #endif
4288     case Op_HasNegatives:
4289       pinned = false;
4290     }
4291     if( pinned ) {
4292       IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
4293       if( !chosen_loop->_child )       // Inner loop?
4294         chosen_loop->_body.push(n); // Collect inner loops
4295       return;
4296     }
4297   } else {                      // No slot zero
4298     if( n->is_CFG() ) {         // CFG with no slot 0 is dead
4299       _nodes.map(n->_idx,0);    // No block setting, it's globally dead
4300       return;
4301     }
4302     assert(!n->is_CFG() || n->outcnt() == 0, "");
4303   }
4304 
4305   // Do I have a "safe range" I can select over?
4306   Node *early = get_ctrl(n);// Early location already computed
4307 
4308   // Compute latest point this Node can go
4309   Node *LCA = get_late_ctrl( n, early );
4310   // LCA is NULL due to uses being dead
4311   if( LCA == NULL ) {
4312 #ifdef ASSERT
4313     for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
4314       assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
4315     }
4316 #endif
4317     _nodes.map(n->_idx, 0);     // This node is useless
4318     _deadlist.push(n);
4319     return;
4320   }
4321   assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
4322 
4323   Node *legal = LCA;            // Walk 'legal' up the IDOM chain
4324   Node *least = legal;          // Best legal position so far
4325   while( early != legal ) {     // While not at earliest legal
4326 #ifdef ASSERT
4327     if (legal->is_Start() && !early->is_Root()) {
4328       // Bad graph. Print idom path and fail.
4329       dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
4330       assert(false, "Bad graph detected in build_loop_late");
4331     }
4332 #endif
4333     // Find least loop nesting depth
4334     legal = idom(legal);        // Bump up the IDOM tree
4335     // Check for lower nesting depth
4336     if( get_loop(legal)->_nest < get_loop(least)->_nest )
4337       least = legal;
4338   }
4339   assert(early == legal || legal != C->root(), "bad dominance of inputs");
4340 
4341   // Try not to place code on a loop entry projection
4342   // which can inhibit range check elimination.
4343   if (least != early) {
4344     Node* ctrl_out = least->unique_ctrl_out();
4345     if (ctrl_out && ctrl_out->is_Loop() &&
4346         least == ctrl_out->in(LoopNode::EntryControl)) {
4347       // Move the node above predicates as far up as possible so a
4348       // following pass of loop predication doesn't hoist a predicate
4349       // that depends on it above that node.
4350       Node* new_ctrl = least;
4351       for (;;) {
4352         if (!new_ctrl->is_Proj()) {
4353           break;
4354         }
4355         CallStaticJavaNode* call = new_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
4356         if (call == NULL) {
4357           break;
4358         }
4359         int req = call->uncommon_trap_request();
4360         Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req);
4361         if (trap_reason != Deoptimization::Reason_loop_limit_check &&
4362             trap_reason != Deoptimization::Reason_predicate &&
4363             trap_reason != Deoptimization::Reason_profile_predicate) {
4364           break;
4365         }
4366         Node* c = new_ctrl->in(0)->in(0);
4367         if (is_dominator(c, early) && c != early) {
4368           break;
4369         }
4370         new_ctrl = c;
4371       }
4372       least = new_ctrl;
4373     }
4374   }
4375 
4376 #ifdef ASSERT
4377   // If verifying, verify that 'verify_me' has a legal location
4378   // and choose it as our location.
4379   if( _verify_me ) {
4380     Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
4381     Node *legal = LCA;
4382     while( early != legal ) {   // While not at earliest legal
4383       if( legal == v_ctrl ) break;  // Check for prior good location
4384       legal = idom(legal)      ;// Bump up the IDOM tree
4385     }
4386     // Check for prior good location
4387     if( legal == v_ctrl ) least = legal; // Keep prior if found
4388   }
4389 #endif
4390 
4391   // Assign discovered "here or above" point
4392   least = find_non_split_ctrl(least);
4393   if (verify_strip_mined && !_verify_only) {
4394     verify_strip_mined_scheduling(n, least);
4395   }
4396   set_ctrl(n, least);
4397 
4398   // Collect inner loop bodies
4399   IdealLoopTree *chosen_loop = get_loop(least);
4400   if( !chosen_loop->_child )   // Inner loop?
4401     chosen_loop->_body.push(n);// Collect inner loops
4402 
4403   if (n->Opcode() == Op_ShenandoahWriteBarrier) {
4404     // The write barrier and its memory proj must have the same
4405     // control otherwise some loop opts could put nodes (Phis) between
4406     // them
4407     Node* proj = n->find_out_with(Op_ShenandoahWBMemProj);
4408     if (proj != NULL) {
4409       set_ctrl_and_loop(proj, least);
4410     }
4411   }
4412 }
4413 
4414 #ifdef ASSERT
4415 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
4416   tty->print_cr("%s", msg);
4417   tty->print("n: "); n->dump();
4418   tty->print("early(n): "); early->dump();
4419   if (n->in(0) != NULL  && !n->in(0)->is_top() &&
4420       n->in(0) != early && !n->in(0)->is_Root()) {
4421     tty->print("n->in(0): "); n->in(0)->dump();
4422   }
4423   for (uint i = 1; i < n->req(); i++) {
4424     Node* in1 = n->in(i);
4425     if (in1 != NULL && in1 != n && !in1->is_top()) {
4426       tty->print("n->in(%d): ", i); in1->dump();
4427       Node* in1_early = get_ctrl(in1);
4428       tty->print("early(n->in(%d)): ", i); in1_early->dump();
4429       if (in1->in(0) != NULL     && !in1->in(0)->is_top() &&
4430           in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
4431         tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
4432       }
4433       for (uint j = 1; j < in1->req(); j++) {
4434         Node* in2 = in1->in(j);
4435         if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) {
4436           tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
4437           Node* in2_early = get_ctrl(in2);
4438           tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
4439           if (in2->in(0) != NULL     && !in2->in(0)->is_top() &&
4440               in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
4441             tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
4442           }
4443         }
4444       }
4445     }
4446   }
4447   tty->cr();
4448   tty->print("LCA(n): "); LCA->dump();
4449   for (uint i = 0; i < n->outcnt(); i++) {
4450     Node* u1 = n->raw_out(i);
4451     if (u1 == n)
4452       continue;
4453     tty->print("n->out(%d): ", i); u1->dump();
4454     if (u1->is_CFG()) {
4455       for (uint j = 0; j < u1->outcnt(); j++) {
4456         Node* u2 = u1->raw_out(j);
4457         if (u2 != u1 && u2 != n && u2->is_CFG()) {
4458           tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
4459         }
4460       }
4461     } else {
4462       Node* u1_later = get_ctrl(u1);
4463       tty->print("later(n->out(%d)): ", i); u1_later->dump();
4464       if (u1->in(0) != NULL     && !u1->in(0)->is_top() &&
4465           u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
4466         tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
4467       }
4468       for (uint j = 0; j < u1->outcnt(); j++) {
4469         Node* u2 = u1->raw_out(j);
4470         if (u2 == n || u2 == u1)
4471           continue;
4472         tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
4473         if (!u2->is_CFG()) {
4474           Node* u2_later = get_ctrl(u2);
4475           tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
4476           if (u2->in(0) != NULL     && !u2->in(0)->is_top() &&
4477               u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
4478             tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
4479           }
4480         }
4481       }
4482     }
4483   }
4484   tty->cr();
4485   int ct = 0;
4486   Node *dbg_legal = LCA;
4487   while(!dbg_legal->is_Start() && ct < 100) {
4488     tty->print("idom[%d] ",ct); dbg_legal->dump();
4489     ct++;
4490     dbg_legal = idom(dbg_legal);
4491   }
4492   tty->cr();
4493 }
4494 #endif
4495 
4496 #ifndef PRODUCT
4497 //------------------------------dump-------------------------------------------
4498 void PhaseIdealLoop::dump( ) const {
4499   ResourceMark rm;
4500   Arena* arena = Thread::current()->resource_area();
4501   Node_Stack stack(arena, C->live_nodes() >> 2);
4502   Node_List rpo_list;
4503   VectorSet visited(arena);
4504   visited.set(C->top()->_idx);
4505   rpo( C->root(), stack, visited, rpo_list );
4506   // Dump root loop indexed by last element in PO order
4507   dump( _ltree_root, rpo_list.size(), rpo_list );
4508 }
4509 
4510 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
4511   loop->dump_head();
4512 
4513   // Now scan for CFG nodes in the same loop
4514   for( uint j=idx; j > 0;  j-- ) {
4515     Node *n = rpo_list[j-1];
4516     if( !_nodes[n->_idx] )      // Skip dead nodes
4517       continue;
4518     if( get_loop(n) != loop ) { // Wrong loop nest
4519       if( get_loop(n)->_head == n &&    // Found nested loop?
4520           get_loop(n)->_parent == loop )
4521         dump(get_loop(n),rpo_list.size(),rpo_list);     // Print it nested-ly
4522       continue;
4523     }
4524 
4525     // Dump controlling node
4526     for( uint x = 0; x < loop->_nest; x++ )
4527       tty->print("  ");
4528     tty->print("C");
4529     if( n == C->root() ) {
4530       n->dump();
4531     } else {
4532       Node* cached_idom   = idom_no_update(n);
4533       Node *computed_idom = n->in(0);
4534       if( n->is_Region() ) {
4535         computed_idom = compute_idom(n);
4536         // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
4537         // any MultiBranch ctrl node), so apply a similar transform to
4538         // the cached idom returned from idom_no_update.
4539         cached_idom = find_non_split_ctrl(cached_idom);
4540       }
4541       tty->print(" ID:%d",computed_idom->_idx);
4542       n->dump();
4543       if( cached_idom != computed_idom ) {
4544         tty->print_cr("*** BROKEN IDOM!  Computed as: %d, cached as: %d",
4545                       computed_idom->_idx, cached_idom->_idx);
4546       }
4547     }
4548     // Dump nodes it controls
4549     for( uint k = 0; k < _nodes.Size(); k++ ) {
4550       // (k < C->unique() && get_ctrl(find(k)) == n)
4551       if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
4552         Node *m = C->root()->find(k);
4553         if( m && m->outcnt() > 0 ) {
4554           if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
4555             tty->print_cr("*** BROKEN CTRL ACCESSOR!  _nodes[k] is %p, ctrl is %p",
4556                           _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
4557           }
4558           for( uint j = 0; j < loop->_nest; j++ )
4559             tty->print("  ");
4560           tty->print(" ");
4561           m->dump();
4562         }
4563       }
4564     }
4565   }
4566 }
4567 #endif
4568 
4569 // Collect a R-P-O for the whole CFG.
4570 // Result list is in post-order (scan backwards for RPO)
4571 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
4572   stk.push(start, 0);
4573   visited.set(start->_idx);
4574 
4575   while (stk.is_nonempty()) {
4576     Node* m   = stk.node();
4577     uint  idx = stk.index();
4578     if (idx < m->outcnt()) {
4579       stk.set_index(idx + 1);
4580       Node* n = m->raw_out(idx);
4581       if (n->is_CFG() && !visited.test_set(n->_idx)) {
4582         stk.push(n, 0);
4583       }
4584     } else {
4585       rpo_list.push(m);
4586       stk.pop();
4587     }
4588   }
4589 }
4590 
4591 
4592 //=============================================================================
4593 //------------------------------LoopTreeIterator-----------------------------------
4594 
4595 // Advance to next loop tree using a preorder, left-to-right traversal.
4596 void LoopTreeIterator::next() {
4597   assert(!done(), "must not be done.");
4598   if (_curnt->_child != NULL) {
4599     _curnt = _curnt->_child;
4600   } else if (_curnt->_next != NULL) {
4601     _curnt = _curnt->_next;
4602   } else {
4603     while (_curnt != _root && _curnt->_next == NULL) {
4604       _curnt = _curnt->_parent;
4605     }
4606     if (_curnt == _root) {
4607       _curnt = NULL;
4608       assert(done(), "must be done.");
4609     } else {
4610       assert(_curnt->_next != NULL, "must be more to do");
4611       _curnt = _curnt->_next;
4612     }
4613   }
4614 }