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 #ifdef ASSERT
 956 void LoopNode::verify_strip_mined(int expect_skeleton) const {
 957   const OuterStripMinedLoopNode* outer = NULL;
 958   const CountedLoopNode* inner = NULL;
 959   if (is_strip_mined()) {
 960     if (!is_valid_counted_loop()) {
 961       return; // Skip malformed counted loop
 962     }
 963     assert(is_CountedLoop(), "no Loop should be marked strip mined");
 964     inner = as_CountedLoop();
 965     outer = inner->in(LoopNode::EntryControl)->as_OuterStripMinedLoop();
 966   } else if (is_OuterStripMinedLoop()) {
 967     outer = this->as_OuterStripMinedLoop();
 968     inner = outer->unique_ctrl_out()->as_CountedLoop();
 969     assert(inner->is_valid_counted_loop() && inner->is_strip_mined(), "OuterStripMinedLoop should have been removed");
 970     assert(!is_strip_mined(), "outer loop shouldn't be marked strip mined");
 971   }
 972   if (inner != NULL || outer != NULL) {
 973     assert(inner != NULL && outer != NULL, "missing loop in strip mined nest");
 974     Node* outer_tail = outer->in(LoopNode::LoopBackControl);
 975     Node* outer_le = outer_tail->in(0);
 976     assert(outer_le->Opcode() == Op_OuterStripMinedLoopEnd, "tail of outer loop should be an If");
 977     Node* sfpt = outer_le->in(0);
 978     assert(sfpt->Opcode() == Op_SafePoint, "where's the safepoint?");
 979     Node* inner_out = sfpt->in(0);
 980     if (inner_out->outcnt() != 1) {
 981       ResourceMark rm;
 982       Unique_Node_List wq;
 983 
 984       for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
 985         Node* u = inner_out->fast_out(i);
 986         if (u == sfpt) {
 987           continue;
 988         }
 989         wq.clear();
 990         wq.push(u);
 991         bool found_sfpt = false;
 992         for (uint next = 0; next < wq.size() && !found_sfpt; next++) {
 993           Node *n = wq.at(next);
 994           for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && !found_sfpt; i++) {
 995             Node* u = n->fast_out(i);
 996             if (u == sfpt) {
 997               found_sfpt = true;
 998             }
 999             if (!u->is_CFG()) {
1000               wq.push(u);
1001             }
1002           }
1003         }
1004         assert(found_sfpt, "no node in loop that's not input to safepoint");
1005       }
1006     }
1007     CountedLoopEndNode* cle = inner_out->in(0)->as_CountedLoopEnd();
1008     assert(cle == inner->loopexit_or_null(), "mismatch");
1009     bool has_skeleton = outer_le->in(1)->bottom_type()->singleton() && outer_le->in(1)->bottom_type()->is_int()->get_con() == 0;
1010     if (has_skeleton) {
1011       assert(expect_skeleton == 1 || expect_skeleton == -1, "unexpected skeleton node");
1012       assert(outer->outcnt() == 2, "only phis");
1013     } else {
1014       assert(expect_skeleton == 0 || expect_skeleton == -1, "no skeleton node?");
1015       uint phis = 0;
1016       for (DUIterator_Fast imax, i = inner->fast_outs(imax); i < imax; i++) {
1017         Node* u = inner->fast_out(i);
1018         if (u->is_Phi()) {
1019           phis++;
1020         }
1021       }
1022       for (DUIterator_Fast imax, i = outer->fast_outs(imax); i < imax; i++) {
1023         Node* u = outer->fast_out(i);
1024         assert(u == outer || u == inner || u->is_Phi(), "nothing between inner and outer loop");
1025       }
1026       uint stores = 0;
1027       for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
1028         Node* u = inner_out->fast_out(i);
1029         if (u->is_Store()) {
1030           stores++;
1031         }
1032       }
1033       assert(outer->outcnt() >= phis + 2 && outer->outcnt() <= phis + 2 + stores + 1, "only phis");
1034     }
1035     assert(sfpt->outcnt() == 1, "no data node");
1036     assert(outer_tail->outcnt() == 1 || !has_skeleton, "no data node");
1037   }
1038 }
1039 #endif
1040 
1041 //=============================================================================
1042 //------------------------------Ideal------------------------------------------
1043 // Return a node which is more "ideal" than the current node.
1044 // Attempt to convert into a counted-loop.
1045 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1046   return RegionNode::Ideal(phase, can_reshape);
1047 }
1048 
1049 //------------------------------dump_spec--------------------------------------
1050 // Dump special per-node info
1051 #ifndef PRODUCT
1052 void CountedLoopNode::dump_spec(outputStream *st) const {
1053   LoopNode::dump_spec(st);
1054   if (stride_is_con()) {
1055     st->print("stride: %d ",stride_con());
1056   }
1057   if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
1058   if (is_main_loop()) st->print("main of N%d", _idx);
1059   if (is_post_loop()) st->print("post of N%d", _main_idx);
1060   if (is_strip_mined()) st->print(" strip mined");
1061 }
1062 #endif
1063 
1064 //=============================================================================
1065 int CountedLoopEndNode::stride_con() const {
1066   return stride()->bottom_type()->is_int()->get_con();
1067 }
1068 
1069 //=============================================================================
1070 //------------------------------Value-----------------------------------------
1071 const Type* LoopLimitNode::Value(PhaseGVN* phase) const {
1072   const Type* init_t   = phase->type(in(Init));
1073   const Type* limit_t  = phase->type(in(Limit));
1074   const Type* stride_t = phase->type(in(Stride));
1075   // Either input is TOP ==> the result is TOP
1076   if (init_t   == Type::TOP) return Type::TOP;
1077   if (limit_t  == Type::TOP) return Type::TOP;
1078   if (stride_t == Type::TOP) return Type::TOP;
1079 
1080   int stride_con = stride_t->is_int()->get_con();
1081   if (stride_con == 1)
1082     return NULL;  // Identity
1083 
1084   if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
1085     // Use jlongs to avoid integer overflow.
1086     jlong init_con   =  init_t->is_int()->get_con();
1087     jlong limit_con  = limit_t->is_int()->get_con();
1088     int  stride_m   = stride_con - (stride_con > 0 ? 1 : -1);
1089     jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
1090     jlong final_con  = init_con + stride_con*trip_count;
1091     int final_int = (int)final_con;
1092     // The final value should be in integer range since the loop
1093     // is counted and the limit was checked for overflow.
1094     assert(final_con == (jlong)final_int, "final value should be integer");
1095     return TypeInt::make(final_int);
1096   }
1097 
1098   return bottom_type(); // TypeInt::INT
1099 }
1100 
1101 //------------------------------Ideal------------------------------------------
1102 // Return a node which is more "ideal" than the current node.
1103 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1104   if (phase->type(in(Init))   == Type::TOP ||
1105       phase->type(in(Limit))  == Type::TOP ||
1106       phase->type(in(Stride)) == Type::TOP)
1107     return NULL;  // Dead
1108 
1109   int stride_con = phase->type(in(Stride))->is_int()->get_con();
1110   if (stride_con == 1)
1111     return NULL;  // Identity
1112 
1113   if (in(Init)->is_Con() && in(Limit)->is_Con())
1114     return NULL;  // Value
1115 
1116   // Delay following optimizations until all loop optimizations
1117   // done to keep Ideal graph simple.
1118   if (!can_reshape || phase->C->major_progress())
1119     return NULL;
1120 
1121   const TypeInt* init_t  = phase->type(in(Init) )->is_int();
1122   const TypeInt* limit_t = phase->type(in(Limit))->is_int();
1123   int stride_p;
1124   jlong lim, ini;
1125   julong max;
1126   if (stride_con > 0) {
1127     stride_p = stride_con;
1128     lim = limit_t->_hi;
1129     ini = init_t->_lo;
1130     max = (julong)max_jint;
1131   } else {
1132     stride_p = -stride_con;
1133     lim = init_t->_hi;
1134     ini = limit_t->_lo;
1135     max = (julong)min_jint;
1136   }
1137   julong range = lim - ini + stride_p;
1138   if (range <= max) {
1139     // Convert to integer expression if it is not overflow.
1140     Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
1141     Node *range = phase->transform(new SubINode(in(Limit), in(Init)));
1142     Node *bias  = phase->transform(new AddINode(range, stride_m));
1143     Node *trip  = phase->transform(new DivINode(0, bias, in(Stride)));
1144     Node *span  = phase->transform(new MulINode(trip, in(Stride)));
1145     return new AddINode(span, in(Init)); // exact limit
1146   }
1147 
1148   if (is_power_of_2(stride_p) ||                // divisor is 2^n
1149       !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
1150     // Convert to long expression to avoid integer overflow
1151     // and let igvn optimizer convert this division.
1152     //
1153     Node*   init   = phase->transform( new ConvI2LNode(in(Init)));
1154     Node*  limit   = phase->transform( new ConvI2LNode(in(Limit)));
1155     Node* stride   = phase->longcon(stride_con);
1156     Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
1157 
1158     Node *range = phase->transform(new SubLNode(limit, init));
1159     Node *bias  = phase->transform(new AddLNode(range, stride_m));
1160     Node *span;
1161     if (stride_con > 0 && is_power_of_2(stride_p)) {
1162       // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
1163       // and avoid generating rounding for division. Zero trip guard should
1164       // guarantee that init < limit but sometimes the guard is missing and
1165       // we can get situation when init > limit. Note, for the empty loop
1166       // optimization zero trip guard is generated explicitly which leaves
1167       // only RCE predicate where exact limit is used and the predicate
1168       // will simply fail forcing recompilation.
1169       Node* neg_stride   = phase->longcon(-stride_con);
1170       span = phase->transform(new AndLNode(bias, neg_stride));
1171     } else {
1172       Node *trip  = phase->transform(new DivLNode(0, bias, stride));
1173       span = phase->transform(new MulLNode(trip, stride));
1174     }
1175     // Convert back to int
1176     Node *span_int = phase->transform(new ConvL2INode(span));
1177     return new AddINode(span_int, in(Init)); // exact limit
1178   }
1179 
1180   return NULL;    // No progress
1181 }
1182 
1183 //------------------------------Identity---------------------------------------
1184 // If stride == 1 return limit node.
1185 Node* LoopLimitNode::Identity(PhaseGVN* phase) {
1186   int stride_con = phase->type(in(Stride))->is_int()->get_con();
1187   if (stride_con == 1 || stride_con == -1)
1188     return in(Limit);
1189   return this;
1190 }
1191 
1192 //=============================================================================
1193 //----------------------match_incr_with_optional_truncation--------------------
1194 // Match increment with optional truncation:
1195 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
1196 // Return NULL for failure. Success returns the increment node.
1197 Node* CountedLoopNode::match_incr_with_optional_truncation(
1198                       Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
1199   // Quick cutouts:
1200   if (expr == NULL || expr->req() != 3)  return NULL;
1201 
1202   Node *t1 = NULL;
1203   Node *t2 = NULL;
1204   const TypeInt* trunc_t = TypeInt::INT;
1205   Node* n1 = expr;
1206   int   n1op = n1->Opcode();
1207 
1208   // Try to strip (n1 & M) or (n1 << N >> N) from n1.
1209   if (n1op == Op_AndI &&
1210       n1->in(2)->is_Con() &&
1211       n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
1212     // %%% This check should match any mask of 2**K-1.
1213     t1 = n1;
1214     n1 = t1->in(1);
1215     n1op = n1->Opcode();
1216     trunc_t = TypeInt::CHAR;
1217   } else if (n1op == Op_RShiftI &&
1218              n1->in(1) != NULL &&
1219              n1->in(1)->Opcode() == Op_LShiftI &&
1220              n1->in(2) == n1->in(1)->in(2) &&
1221              n1->in(2)->is_Con()) {
1222     jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
1223     // %%% This check should match any shift in [1..31].
1224     if (shift == 16 || shift == 8) {
1225       t1 = n1;
1226       t2 = t1->in(1);
1227       n1 = t2->in(1);
1228       n1op = n1->Opcode();
1229       if (shift == 16) {
1230         trunc_t = TypeInt::SHORT;
1231       } else if (shift == 8) {
1232         trunc_t = TypeInt::BYTE;
1233       }
1234     }
1235   }
1236 
1237   // If (maybe after stripping) it is an AddI, we won:
1238   if (n1op == Op_AddI) {
1239     *trunc1 = t1;
1240     *trunc2 = t2;
1241     *trunc_type = trunc_t;
1242     return n1;
1243   }
1244 
1245   // failed
1246   return NULL;
1247 }
1248 
1249 LoopNode* CountedLoopNode::skip_strip_mined(int expect_skeleton) {
1250   if (is_strip_mined() && is_valid_counted_loop()) {
1251     verify_strip_mined(expect_skeleton);
1252     return in(EntryControl)->as_Loop();
1253   }
1254   return this;
1255 }
1256 
1257 OuterStripMinedLoopNode* CountedLoopNode::outer_loop() const {
1258   assert(is_strip_mined(), "not a strip mined loop");
1259   Node* c = in(EntryControl);
1260   if (c == NULL || c->is_top() || !c->is_OuterStripMinedLoop()) {
1261     return NULL;
1262   }
1263   return c->as_OuterStripMinedLoop();
1264 }
1265 
1266 IfTrueNode* OuterStripMinedLoopNode::outer_loop_tail() const {
1267   Node* c = in(LoopBackControl);
1268   if (c == NULL || c->is_top()) {
1269     return NULL;
1270   }
1271   return c->as_IfTrue();
1272 }
1273 
1274 IfTrueNode* CountedLoopNode::outer_loop_tail() const {
1275   LoopNode* l = outer_loop();
1276   if (l == NULL) {
1277     return NULL;
1278   }
1279   return l->outer_loop_tail();
1280 }
1281 
1282 OuterStripMinedLoopEndNode* OuterStripMinedLoopNode::outer_loop_end() const {
1283   IfTrueNode* proj = outer_loop_tail();
1284   if (proj == NULL) {
1285     return NULL;
1286   }
1287   Node* c = proj->in(0);
1288   if (c == NULL || c->is_top() || c->outcnt() != 2) {
1289     return NULL;
1290   }
1291   return c->as_OuterStripMinedLoopEnd();
1292 }
1293 
1294 OuterStripMinedLoopEndNode* CountedLoopNode::outer_loop_end() const {
1295   LoopNode* l = outer_loop();
1296   if (l == NULL) {
1297     return NULL;
1298   }
1299   return l->outer_loop_end();
1300 }
1301 
1302 IfFalseNode* OuterStripMinedLoopNode::outer_loop_exit() const {
1303   IfNode* le = outer_loop_end();
1304   if (le == NULL) {
1305     return NULL;
1306   }
1307   Node* c = le->proj_out_or_null(false);
1308   if (c == NULL) {
1309     return NULL;
1310   }
1311   return c->as_IfFalse();
1312 }
1313 
1314 IfFalseNode* CountedLoopNode::outer_loop_exit() const {
1315   LoopNode* l = outer_loop();
1316   if (l == NULL) {
1317     return NULL;
1318   }
1319   return l->outer_loop_exit();
1320 }
1321 
1322 SafePointNode* OuterStripMinedLoopNode::outer_safepoint() const {
1323   IfNode* le = outer_loop_end();
1324   if (le == NULL) {
1325     return NULL;
1326   }
1327   Node* c = le->in(0);
1328   if (c == NULL || c->is_top()) {
1329     return NULL;
1330   }
1331   assert(c->Opcode() == Op_SafePoint, "broken outer loop");
1332   return c->as_SafePoint();
1333 }
1334 
1335 SafePointNode* CountedLoopNode::outer_safepoint() const {
1336   LoopNode* l = outer_loop();
1337   if (l == NULL) {
1338     return NULL;
1339   }
1340   return l->outer_safepoint();
1341 }
1342 
1343 Node* CountedLoopNode::skip_predicates_from_entry(Node* ctrl) {
1344     while (ctrl != NULL && ctrl->is_Proj() && ctrl->in(0)->is_If() &&
1345            ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->outcnt() == 1 &&
1346            ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->unique_out()->Opcode() == Op_Halt) {
1347       ctrl = ctrl->in(0)->in(0);
1348     }
1349 
1350     return ctrl;
1351   }
1352 
1353 Node* CountedLoopNode::skip_predicates() {
1354   if (is_main_loop()) {
1355     Node* ctrl = skip_strip_mined()->in(LoopNode::EntryControl);
1356 
1357     return skip_predicates_from_entry(ctrl);
1358   }
1359   return in(LoopNode::EntryControl);
1360 }
1361 
1362 void OuterStripMinedLoopNode::adjust_strip_mined_loop(PhaseIterGVN* igvn) {
1363   // Look for the outer & inner strip mined loop, reduce number of
1364   // iterations of the inner loop, set exit condition of outer loop,
1365   // construct required phi nodes for outer loop.
1366   CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
1367   assert(inner_cl->is_strip_mined(), "inner loop should be strip mined");
1368   Node* inner_iv_phi = inner_cl->phi();
1369   if (inner_iv_phi == NULL) {
1370     IfNode* outer_le = outer_loop_end();
1371     Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1372     igvn->replace_node(outer_le, iff);
1373     inner_cl->clear_strip_mined();
1374     return;
1375   }
1376   CountedLoopEndNode* inner_cle = inner_cl->loopexit();
1377 
1378   int stride = inner_cl->stride_con();
1379   jlong scaled_iters_long = ((jlong)LoopStripMiningIter) * ABS(stride);
1380   int scaled_iters = (int)scaled_iters_long;
1381   int short_scaled_iters = LoopStripMiningIterShortLoop* ABS(stride);
1382   const TypeInt* inner_iv_t = igvn->type(inner_iv_phi)->is_int();
1383   jlong iter_estimate = (jlong)inner_iv_t->_hi - (jlong)inner_iv_t->_lo;
1384   assert(iter_estimate > 0, "broken");
1385   if ((jlong)scaled_iters != scaled_iters_long || iter_estimate <= short_scaled_iters) {
1386     // Remove outer loop and safepoint (too few iterations)
1387     Node* outer_sfpt = outer_safepoint();
1388     Node* outer_out = outer_loop_exit();
1389     igvn->replace_node(outer_out, outer_sfpt->in(0));
1390     igvn->replace_input_of(outer_sfpt, 0, igvn->C->top());
1391     inner_cl->clear_strip_mined();
1392     return;
1393   }
1394   if (iter_estimate <= scaled_iters_long) {
1395     // We would only go through one iteration of
1396     // the outer loop: drop the outer loop but
1397     // keep the safepoint so we don't run for
1398     // too long without a safepoint
1399     IfNode* outer_le = outer_loop_end();
1400     Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1401     igvn->replace_node(outer_le, iff);
1402     inner_cl->clear_strip_mined();
1403     return;
1404   }
1405 
1406   Node* cle_tail = inner_cle->proj_out(true);
1407   ResourceMark rm;
1408   Node_List old_new;
1409   if (cle_tail->outcnt() > 1) {
1410     // Look for nodes on backedge of inner loop and clone them
1411     Unique_Node_List backedge_nodes;
1412     for (DUIterator_Fast imax, i = cle_tail->fast_outs(imax); i < imax; i++) {
1413       Node* u = cle_tail->fast_out(i);
1414       if (u != inner_cl) {
1415         assert(!u->is_CFG(), "control flow on the backedge?");
1416         backedge_nodes.push(u);
1417       }
1418     }
1419     uint last = igvn->C->unique();
1420     for (uint next = 0; next < backedge_nodes.size(); next++) {
1421       Node* n = backedge_nodes.at(next);
1422       old_new.map(n->_idx, n->clone());
1423       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1424         Node* u = n->fast_out(i);
1425         assert(!u->is_CFG(), "broken");
1426         if (u->_idx >= last) {
1427           continue;
1428         }
1429         if (!u->is_Phi()) {
1430           backedge_nodes.push(u);
1431         } else {
1432           assert(u->in(0) == inner_cl, "strange phi on the backedge");
1433         }
1434       }
1435     }
1436     // Put the clones on the outer loop backedge
1437     Node* le_tail = outer_loop_tail();
1438     for (uint next = 0; next < backedge_nodes.size(); next++) {
1439       Node *n = old_new[backedge_nodes.at(next)->_idx];
1440       for (uint i = 1; i < n->req(); i++) {
1441         if (n->in(i) != NULL && old_new[n->in(i)->_idx] != NULL) {
1442           n->set_req(i, old_new[n->in(i)->_idx]);
1443         }
1444       }
1445       if (n->in(0) != NULL && n->in(0) == cle_tail) {
1446         n->set_req(0, le_tail);
1447       }
1448       igvn->register_new_node_with_optimizer(n);
1449     }
1450   }
1451 
1452   Node* iv_phi = NULL;
1453   // Make a clone of each phi in the inner loop
1454   // for the outer loop
1455   for (uint i = 0; i < inner_cl->outcnt(); i++) {
1456     Node* u = inner_cl->raw_out(i);
1457     if (u->is_Phi()) {
1458       assert(u->in(0) == inner_cl, "inconsistent");
1459       Node* phi = u->clone();
1460       phi->set_req(0, this);
1461       Node* be = old_new[phi->in(LoopNode::LoopBackControl)->_idx];
1462       if (be != NULL) {
1463         phi->set_req(LoopNode::LoopBackControl, be);
1464       }
1465       phi = igvn->transform(phi);
1466       igvn->replace_input_of(u, LoopNode::EntryControl, phi);
1467       if (u == inner_iv_phi) {
1468         iv_phi = phi;
1469       }
1470     }
1471   }
1472   Node* cle_out = inner_cle->proj_out(false);
1473   if (cle_out->outcnt() > 1) {
1474     // Look for chains of stores that were sunk
1475     // out of the inner loop and are in the outer loop
1476     for (DUIterator_Fast imax, i = cle_out->fast_outs(imax); i < imax; i++) {
1477       Node* u = cle_out->fast_out(i);
1478       if (u->is_Store()) {
1479         Node* first = u;
1480         for(;;) {
1481           Node* next = first->in(MemNode::Memory);
1482           if (!next->is_Store() || next->in(0) != cle_out) {
1483             break;
1484           }
1485           first = next;
1486         }
1487         Node* last = u;
1488         for(;;) {
1489           Node* next = NULL;
1490           for (DUIterator_Fast jmax, j = last->fast_outs(jmax); j < jmax; j++) {
1491             Node* uu = last->fast_out(j);
1492             if (uu->is_Store() && uu->in(0) == cle_out) {
1493               assert(next == NULL, "only one in the outer loop");
1494               next = uu;
1495             }
1496           }
1497           if (next == NULL) {
1498             break;
1499           }
1500           last = next;
1501         }
1502         Node* phi = NULL;
1503         for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) {
1504           Node* uu = fast_out(j);
1505           if (uu->is_Phi()) {
1506             Node* be = uu->in(LoopNode::LoopBackControl);
1507             if (be->is_Store() && old_new[be->_idx] != NULL) {
1508               assert(false, "store on the backedge + sunk stores: unsupported");
1509               // drop outer loop
1510               IfNode* outer_le = outer_loop_end();
1511               Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1512               igvn->replace_node(outer_le, iff);
1513               inner_cl->clear_strip_mined();
1514               return;
1515             }
1516             if (be == last || be == first->in(MemNode::Memory)) {
1517               assert(phi == NULL, "only one phi");
1518               phi = uu;
1519             }
1520           }
1521         }
1522 #ifdef ASSERT
1523         for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) {
1524           Node* uu = fast_out(j);
1525           if (uu->is_Phi() && uu->bottom_type() == Type::MEMORY) {
1526             if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) {
1527               assert(phi == uu, "what's that phi?");
1528             } else if (uu->adr_type() == TypePtr::BOTTOM) {
1529               Node* n = uu->in(LoopNode::LoopBackControl);
1530               uint limit = igvn->C->live_nodes();
1531               uint i = 0;
1532               while (n != uu) {
1533                 i++;
1534                 assert(i < limit, "infinite loop");
1535                 if (n->is_Proj()) {
1536                   n = n->in(0);
1537                 } else if (n->is_SafePoint() || n->is_MemBar()) {
1538                   n = n->in(TypeFunc::Memory);
1539                 } else if (n->is_Phi()) {
1540                   n = n->in(1);
1541                 } else if (n->is_MergeMem()) {
1542                   n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type()));
1543                 } else if (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) {
1544                   n = n->in(MemNode::Memory);
1545                 } else {
1546                   n->dump();
1547                   ShouldNotReachHere();
1548                 }
1549               }
1550             }
1551           }
1552         }
1553 #endif
1554         if (phi == NULL) {
1555           // If the an entire chains was sunk, the
1556           // inner loop has no phi for that memory
1557           // slice, create one for the outer loop
1558           phi = PhiNode::make(this, first->in(MemNode::Memory), Type::MEMORY,
1559                               igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type())));
1560           phi->set_req(LoopNode::LoopBackControl, last);
1561           phi = igvn->transform(phi);
1562           igvn->replace_input_of(first, MemNode::Memory, phi);
1563         } else {
1564           // Or fix the outer loop fix to include
1565           // that chain of stores.
1566           Node* be = phi->in(LoopNode::LoopBackControl);
1567           assert(!(be->is_Store() && old_new[be->_idx] != NULL), "store on the backedge + sunk stores: unsupported");
1568           if (be == first->in(MemNode::Memory)) {
1569             if (be == phi->in(LoopNode::LoopBackControl)) {
1570               igvn->replace_input_of(phi, LoopNode::LoopBackControl, last);
1571             } else {
1572               igvn->replace_input_of(be, MemNode::Memory, last);
1573             }
1574           } else {
1575 #ifdef ASSERT
1576             if (be == phi->in(LoopNode::LoopBackControl)) {
1577               assert(phi->in(LoopNode::LoopBackControl) == last, "");
1578             } else {
1579               assert(be->in(MemNode::Memory) == last, "");
1580             }
1581 #endif
1582           }
1583         }
1584       }
1585     }
1586   }
1587 
1588   if (iv_phi != NULL) {
1589     // Now adjust the inner loop's exit condition
1590     Node* limit = inner_cl->limit();
1591     Node* sub = NULL;
1592     if (stride > 0) {
1593       sub = igvn->transform(new SubINode(limit, iv_phi));
1594     } else {
1595       sub = igvn->transform(new SubINode(iv_phi, limit));
1596     }
1597     Node* min = igvn->transform(new MinINode(sub, igvn->intcon(scaled_iters)));
1598     Node* new_limit = NULL;
1599     if (stride > 0) {
1600       new_limit = igvn->transform(new AddINode(min, iv_phi));
1601     } else {
1602       new_limit = igvn->transform(new SubINode(iv_phi, min));
1603     }
1604     Node* inner_cmp = inner_cle->cmp_node();
1605     Node* inner_bol = inner_cle->in(CountedLoopEndNode::TestValue);
1606     Node* outer_bol = inner_bol;
1607     // cmp node for inner loop may be shared
1608     inner_cmp = inner_cmp->clone();
1609     inner_cmp->set_req(2, new_limit);
1610     inner_bol = inner_bol->clone();
1611     inner_bol->set_req(1, igvn->transform(inner_cmp));
1612     igvn->replace_input_of(inner_cle, CountedLoopEndNode::TestValue, igvn->transform(inner_bol));
1613     // Set the outer loop's exit condition too
1614     igvn->replace_input_of(outer_loop_end(), 1, outer_bol);
1615   } else {
1616     assert(false, "should be able to adjust outer loop");
1617     IfNode* outer_le = outer_loop_end();
1618     Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1619     igvn->replace_node(outer_le, iff);
1620     inner_cl->clear_strip_mined();
1621   }
1622 }
1623 
1624 const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const {
1625   if (!in(0)) return Type::TOP;
1626   if (phase->type(in(0)) == Type::TOP)
1627     return Type::TOP;
1628 
1629   return TypeTuple::IFBOTH;
1630 }
1631 
1632 Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1633   if (remove_dead_region(phase, can_reshape))  return this;
1634 
1635   return NULL;
1636 }
1637 
1638 //------------------------------filtered_type--------------------------------
1639 // Return a type based on condition control flow
1640 // A successful return will be a type that is restricted due
1641 // to a series of dominating if-tests, such as:
1642 //    if (i < 10) {
1643 //       if (i > 0) {
1644 //          here: "i" type is [1..10)
1645 //       }
1646 //    }
1647 // or a control flow merge
1648 //    if (i < 10) {
1649 //       do {
1650 //          phi( , ) -- at top of loop type is [min_int..10)
1651 //         i = ?
1652 //       } while ( i < 10)
1653 //
1654 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
1655   assert(n && n->bottom_type()->is_int(), "must be int");
1656   const TypeInt* filtered_t = NULL;
1657   if (!n->is_Phi()) {
1658     assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
1659     filtered_t = filtered_type_from_dominators(n, n_ctrl);
1660 
1661   } else {
1662     Node* phi    = n->as_Phi();
1663     Node* region = phi->in(0);
1664     assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
1665     if (region && region != C->top()) {
1666       for (uint i = 1; i < phi->req(); i++) {
1667         Node* val   = phi->in(i);
1668         Node* use_c = region->in(i);
1669         const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
1670         if (val_t != NULL) {
1671           if (filtered_t == NULL) {
1672             filtered_t = val_t;
1673           } else {
1674             filtered_t = filtered_t->meet(val_t)->is_int();
1675           }
1676         }
1677       }
1678     }
1679   }
1680   const TypeInt* n_t = _igvn.type(n)->is_int();
1681   if (filtered_t != NULL) {
1682     n_t = n_t->join(filtered_t)->is_int();
1683   }
1684   return n_t;
1685 }
1686 
1687 
1688 //------------------------------filtered_type_from_dominators--------------------------------
1689 // Return a possibly more restrictive type for val based on condition control flow of dominators
1690 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
1691   if (val->is_Con()) {
1692      return val->bottom_type()->is_int();
1693   }
1694   uint if_limit = 10; // Max number of dominating if's visited
1695   const TypeInt* rtn_t = NULL;
1696 
1697   if (use_ctrl && use_ctrl != C->top()) {
1698     Node* val_ctrl = get_ctrl(val);
1699     uint val_dom_depth = dom_depth(val_ctrl);
1700     Node* pred = use_ctrl;
1701     uint if_cnt = 0;
1702     while (if_cnt < if_limit) {
1703       if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
1704         if_cnt++;
1705         const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
1706         if (if_t != NULL) {
1707           if (rtn_t == NULL) {
1708             rtn_t = if_t;
1709           } else {
1710             rtn_t = rtn_t->join(if_t)->is_int();
1711           }
1712         }
1713       }
1714       pred = idom(pred);
1715       if (pred == NULL || pred == C->top()) {
1716         break;
1717       }
1718       // Stop if going beyond definition block of val
1719       if (dom_depth(pred) < val_dom_depth) {
1720         break;
1721       }
1722     }
1723   }
1724   return rtn_t;
1725 }
1726 
1727 
1728 //------------------------------dump_spec--------------------------------------
1729 // Dump special per-node info
1730 #ifndef PRODUCT
1731 void CountedLoopEndNode::dump_spec(outputStream *st) const {
1732   if( in(TestValue) != NULL && in(TestValue)->is_Bool() ) {
1733     BoolTest bt( test_trip()); // Added this for g++.
1734 
1735     st->print("[");
1736     bt.dump_on(st);
1737     st->print("]");
1738   }
1739   st->print(" ");
1740   IfNode::dump_spec(st);
1741 }
1742 #endif
1743 
1744 //=============================================================================
1745 //------------------------------is_member--------------------------------------
1746 // Is 'l' a member of 'this'?
1747 bool IdealLoopTree::is_member(const IdealLoopTree *l) const {
1748   while( l->_nest > _nest ) l = l->_parent;
1749   return l == this;
1750 }
1751 
1752 //------------------------------set_nest---------------------------------------
1753 // Set loop tree nesting depth.  Accumulate _has_call bits.
1754 int IdealLoopTree::set_nest( uint depth ) {
1755   _nest = depth;
1756   int bits = _has_call;
1757   if( _child ) bits |= _child->set_nest(depth+1);
1758   if( bits ) _has_call = 1;
1759   if( _next  ) bits |= _next ->set_nest(depth  );
1760   return bits;
1761 }
1762 
1763 //------------------------------split_fall_in----------------------------------
1764 // Split out multiple fall-in edges from the loop header.  Move them to a
1765 // private RegionNode before the loop.  This becomes the loop landing pad.
1766 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
1767   PhaseIterGVN &igvn = phase->_igvn;
1768   uint i;
1769 
1770   // Make a new RegionNode to be the landing pad.
1771   Node *landing_pad = new RegionNode( fall_in_cnt+1 );
1772   phase->set_loop(landing_pad,_parent);
1773   // Gather all the fall-in control paths into the landing pad
1774   uint icnt = fall_in_cnt;
1775   uint oreq = _head->req();
1776   for( i = oreq-1; i>0; i-- )
1777     if( !phase->is_member( this, _head->in(i) ) )
1778       landing_pad->set_req(icnt--,_head->in(i));
1779 
1780   // Peel off PhiNode edges as well
1781   for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1782     Node *oj = _head->fast_out(j);
1783     if( oj->is_Phi() ) {
1784       PhiNode* old_phi = oj->as_Phi();
1785       assert( old_phi->region() == _head, "" );
1786       igvn.hash_delete(old_phi);   // Yank from hash before hacking edges
1787       Node *p = PhiNode::make_blank(landing_pad, old_phi);
1788       uint icnt = fall_in_cnt;
1789       for( i = oreq-1; i>0; i-- ) {
1790         if( !phase->is_member( this, _head->in(i) ) ) {
1791           p->init_req(icnt--, old_phi->in(i));
1792           // Go ahead and clean out old edges from old phi
1793           old_phi->del_req(i);
1794         }
1795       }
1796       // Search for CSE's here, because ZKM.jar does a lot of
1797       // loop hackery and we need to be a little incremental
1798       // with the CSE to avoid O(N^2) node blow-up.
1799       Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
1800       if( p2 ) {                // Found CSE
1801         p->destruct();          // Recover useless new node
1802         p = p2;                 // Use old node
1803       } else {
1804         igvn.register_new_node_with_optimizer(p, old_phi);
1805       }
1806       // Make old Phi refer to new Phi.
1807       old_phi->add_req(p);
1808       // Check for the special case of making the old phi useless and
1809       // disappear it.  In JavaGrande I have a case where this useless
1810       // Phi is the loop limit and prevents recognizing a CountedLoop
1811       // which in turn prevents removing an empty loop.
1812       Node *id_old_phi = igvn.apply_identity(old_phi);
1813       if( id_old_phi != old_phi ) { // Found a simple identity?
1814         // Note that I cannot call 'replace_node' here, because
1815         // that will yank the edge from old_phi to the Region and
1816         // I'm mid-iteration over the Region's uses.
1817         for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
1818           Node* use = old_phi->last_out(i);
1819           igvn.rehash_node_delayed(use);
1820           uint uses_found = 0;
1821           for (uint j = 0; j < use->len(); j++) {
1822             if (use->in(j) == old_phi) {
1823               if (j < use->req()) use->set_req (j, id_old_phi);
1824               else                use->set_prec(j, id_old_phi);
1825               uses_found++;
1826             }
1827           }
1828           i -= uses_found;    // we deleted 1 or more copies of this edge
1829         }
1830       }
1831       igvn._worklist.push(old_phi);
1832     }
1833   }
1834   // Finally clean out the fall-in edges from the RegionNode
1835   for( i = oreq-1; i>0; i-- ) {
1836     if( !phase->is_member( this, _head->in(i) ) ) {
1837       _head->del_req(i);
1838     }
1839   }
1840   igvn.rehash_node_delayed(_head);
1841   // Transform landing pad
1842   igvn.register_new_node_with_optimizer(landing_pad, _head);
1843   // Insert landing pad into the header
1844   _head->add_req(landing_pad);
1845 }
1846 
1847 //------------------------------split_outer_loop-------------------------------
1848 // Split out the outermost loop from this shared header.
1849 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
1850   PhaseIterGVN &igvn = phase->_igvn;
1851 
1852   // Find index of outermost loop; it should also be my tail.
1853   uint outer_idx = 1;
1854   while( _head->in(outer_idx) != _tail ) outer_idx++;
1855 
1856   // Make a LoopNode for the outermost loop.
1857   Node *ctl = _head->in(LoopNode::EntryControl);
1858   Node *outer = new LoopNode( ctl, _head->in(outer_idx) );
1859   outer = igvn.register_new_node_with_optimizer(outer, _head);
1860   phase->set_created_loop_node();
1861 
1862   // Outermost loop falls into '_head' loop
1863   _head->set_req(LoopNode::EntryControl, outer);
1864   _head->del_req(outer_idx);
1865   // Split all the Phis up between '_head' loop and 'outer' loop.
1866   for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1867     Node *out = _head->fast_out(j);
1868     if( out->is_Phi() ) {
1869       PhiNode *old_phi = out->as_Phi();
1870       assert( old_phi->region() == _head, "" );
1871       Node *phi = PhiNode::make_blank(outer, old_phi);
1872       phi->init_req(LoopNode::EntryControl,    old_phi->in(LoopNode::EntryControl));
1873       phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
1874       phi = igvn.register_new_node_with_optimizer(phi, old_phi);
1875       // Make old Phi point to new Phi on the fall-in path
1876       igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
1877       old_phi->del_req(outer_idx);
1878     }
1879   }
1880 
1881   // Use the new loop head instead of the old shared one
1882   _head = outer;
1883   phase->set_loop(_head, this);
1884 }
1885 
1886 //------------------------------fix_parent-------------------------------------
1887 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
1888   loop->_parent = parent;
1889   if( loop->_child ) fix_parent( loop->_child, loop   );
1890   if( loop->_next  ) fix_parent( loop->_next , parent );
1891 }
1892 
1893 //------------------------------estimate_path_freq-----------------------------
1894 static float estimate_path_freq( Node *n ) {
1895   // Try to extract some path frequency info
1896   IfNode *iff;
1897   for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
1898     uint nop = n->Opcode();
1899     if( nop == Op_SafePoint ) {   // Skip any safepoint
1900       n = n->in(0);
1901       continue;
1902     }
1903     if( nop == Op_CatchProj ) {   // Get count from a prior call
1904       // Assume call does not always throw exceptions: means the call-site
1905       // count is also the frequency of the fall-through path.
1906       assert( n->is_CatchProj(), "" );
1907       if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
1908         return 0.0f;            // Assume call exception path is rare
1909       Node *call = n->in(0)->in(0)->in(0);
1910       assert( call->is_Call(), "expect a call here" );
1911       const JVMState *jvms = ((CallNode*)call)->jvms();
1912       ciMethodData* methodData = jvms->method()->method_data();
1913       if (!methodData->is_mature())  return 0.0f; // No call-site data
1914       ciProfileData* data = methodData->bci_to_data(jvms->bci());
1915       if ((data == NULL) || !data->is_CounterData()) {
1916         // no call profile available, try call's control input
1917         n = n->in(0);
1918         continue;
1919       }
1920       return data->as_CounterData()->count()/FreqCountInvocations;
1921     }
1922     // See if there's a gating IF test
1923     Node *n_c = n->in(0);
1924     if( !n_c->is_If() ) break;       // No estimate available
1925     iff = n_c->as_If();
1926     if( iff->_fcnt != COUNT_UNKNOWN )   // Have a valid count?
1927       // Compute how much count comes on this path
1928       return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
1929     // Have no count info.  Skip dull uncommon-trap like branches.
1930     if( (nop == Op_IfTrue  && iff->_prob < PROB_LIKELY_MAG(5)) ||
1931         (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
1932       break;
1933     // Skip through never-taken branch; look for a real loop exit.
1934     n = iff->in(0);
1935   }
1936   return 0.0f;                  // No estimate available
1937 }
1938 
1939 //------------------------------merge_many_backedges---------------------------
1940 // Merge all the backedges from the shared header into a private Region.
1941 // Feed that region as the one backedge to this loop.
1942 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
1943   uint i;
1944 
1945   // Scan for the top 2 hottest backedges
1946   float hotcnt = 0.0f;
1947   float warmcnt = 0.0f;
1948   uint hot_idx = 0;
1949   // Loop starts at 2 because slot 1 is the fall-in path
1950   for( i = 2; i < _head->req(); i++ ) {
1951     float cnt = estimate_path_freq(_head->in(i));
1952     if( cnt > hotcnt ) {       // Grab hottest path
1953       warmcnt = hotcnt;
1954       hotcnt = cnt;
1955       hot_idx = i;
1956     } else if( cnt > warmcnt ) { // And 2nd hottest path
1957       warmcnt = cnt;
1958     }
1959   }
1960 
1961   // See if the hottest backedge is worthy of being an inner loop
1962   // by being much hotter than the next hottest backedge.
1963   if( hotcnt <= 0.0001 ||
1964       hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
1965 
1966   // Peel out the backedges into a private merge point; peel
1967   // them all except optionally hot_idx.
1968   PhaseIterGVN &igvn = phase->_igvn;
1969 
1970   Node *hot_tail = NULL;
1971   // Make a Region for the merge point
1972   Node *r = new RegionNode(1);
1973   for( i = 2; i < _head->req(); i++ ) {
1974     if( i != hot_idx )
1975       r->add_req( _head->in(i) );
1976     else hot_tail = _head->in(i);
1977   }
1978   igvn.register_new_node_with_optimizer(r, _head);
1979   // Plug region into end of loop _head, followed by hot_tail
1980   while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
1981   igvn.replace_input_of(_head, 2, r);
1982   if( hot_idx ) _head->add_req(hot_tail);
1983 
1984   // Split all the Phis up between '_head' loop and the Region 'r'
1985   for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1986     Node *out = _head->fast_out(j);
1987     if( out->is_Phi() ) {
1988       PhiNode* n = out->as_Phi();
1989       igvn.hash_delete(n);      // Delete from hash before hacking edges
1990       Node *hot_phi = NULL;
1991       Node *phi = new PhiNode(r, n->type(), n->adr_type());
1992       // Check all inputs for the ones to peel out
1993       uint j = 1;
1994       for( uint i = 2; i < n->req(); i++ ) {
1995         if( i != hot_idx )
1996           phi->set_req( j++, n->in(i) );
1997         else hot_phi = n->in(i);
1998       }
1999       // Register the phi but do not transform until whole place transforms
2000       igvn.register_new_node_with_optimizer(phi, n);
2001       // Add the merge phi to the old Phi
2002       while( n->req() > 3 ) n->del_req( n->req()-1 );
2003       igvn.replace_input_of(n, 2, phi);
2004       if( hot_idx ) n->add_req(hot_phi);
2005     }
2006   }
2007 
2008 
2009   // Insert a new IdealLoopTree inserted below me.  Turn it into a clone
2010   // of self loop tree.  Turn self into a loop headed by _head and with
2011   // tail being the new merge point.
2012   IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
2013   phase->set_loop(_tail,ilt);   // Adjust tail
2014   _tail = r;                    // Self's tail is new merge point
2015   phase->set_loop(r,this);
2016   ilt->_child = _child;         // New guy has my children
2017   _child = ilt;                 // Self has new guy as only child
2018   ilt->_parent = this;          // new guy has self for parent
2019   ilt->_nest = _nest;           // Same nesting depth (for now)
2020 
2021   // Starting with 'ilt', look for child loop trees using the same shared
2022   // header.  Flatten these out; they will no longer be loops in the end.
2023   IdealLoopTree **pilt = &_child;
2024   while( ilt ) {
2025     if( ilt->_head == _head ) {
2026       uint i;
2027       for( i = 2; i < _head->req(); i++ )
2028         if( _head->in(i) == ilt->_tail )
2029           break;                // Still a loop
2030       if( i == _head->req() ) { // No longer a loop
2031         // Flatten ilt.  Hang ilt's "_next" list from the end of
2032         // ilt's '_child' list.  Move the ilt's _child up to replace ilt.
2033         IdealLoopTree **cp = &ilt->_child;
2034         while( *cp ) cp = &(*cp)->_next;   // Find end of child list
2035         *cp = ilt->_next;       // Hang next list at end of child list
2036         *pilt = ilt->_child;    // Move child up to replace ilt
2037         ilt->_head = NULL;      // Flag as a loop UNIONED into parent
2038         ilt = ilt->_child;      // Repeat using new ilt
2039         continue;               // do not advance over ilt->_child
2040       }
2041       assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
2042       phase->set_loop(_head,ilt);
2043     }
2044     pilt = &ilt->_child;        // Advance to next
2045     ilt = *pilt;
2046   }
2047 
2048   if( _child ) fix_parent( _child, this );
2049 }
2050 
2051 //------------------------------beautify_loops---------------------------------
2052 // Split shared headers and insert loop landing pads.
2053 // Insert a LoopNode to replace the RegionNode.
2054 // Return TRUE if loop tree is structurally changed.
2055 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
2056   bool result = false;
2057   // Cache parts in locals for easy
2058   PhaseIterGVN &igvn = phase->_igvn;
2059 
2060   igvn.hash_delete(_head);      // Yank from hash before hacking edges
2061 
2062   // Check for multiple fall-in paths.  Peel off a landing pad if need be.
2063   int fall_in_cnt = 0;
2064   for( uint i = 1; i < _head->req(); i++ )
2065     if( !phase->is_member( this, _head->in(i) ) )
2066       fall_in_cnt++;
2067   assert( fall_in_cnt, "at least 1 fall-in path" );
2068   if( fall_in_cnt > 1 )         // Need a loop landing pad to merge fall-ins
2069     split_fall_in( phase, fall_in_cnt );
2070 
2071   // Swap inputs to the _head and all Phis to move the fall-in edge to
2072   // the left.
2073   fall_in_cnt = 1;
2074   while( phase->is_member( this, _head->in(fall_in_cnt) ) )
2075     fall_in_cnt++;
2076   if( fall_in_cnt > 1 ) {
2077     // Since I am just swapping inputs I do not need to update def-use info
2078     Node *tmp = _head->in(1);
2079     igvn.rehash_node_delayed(_head);
2080     _head->set_req( 1, _head->in(fall_in_cnt) );
2081     _head->set_req( fall_in_cnt, tmp );
2082     // Swap also all Phis
2083     for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
2084       Node* phi = _head->fast_out(i);
2085       if( phi->is_Phi() ) {
2086         igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges
2087         tmp = phi->in(1);
2088         phi->set_req( 1, phi->in(fall_in_cnt) );
2089         phi->set_req( fall_in_cnt, tmp );
2090       }
2091     }
2092   }
2093   assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
2094   assert(  phase->is_member( this, _head->in(2) ), "right edge is loop" );
2095 
2096   // If I am a shared header (multiple backedges), peel off the many
2097   // backedges into a private merge point and use the merge point as
2098   // the one true backedge.
2099   if (_head->req() > 3) {
2100     // Merge the many backedges into a single backedge but leave
2101     // the hottest backedge as separate edge for the following peel.
2102     if (!_irreducible) {
2103       merge_many_backedges( phase );
2104     }
2105 
2106     // When recursively beautify my children, split_fall_in can change
2107     // loop tree structure when I am an irreducible loop. Then the head
2108     // of my children has a req() not bigger than 3. Here we need to set
2109     // result to true to catch that case in order to tell the caller to
2110     // rebuild loop tree. See issue JDK-8244407 for details.
2111     result = true;
2112   }
2113 
2114   // If I have one hot backedge, peel off myself loop.
2115   // I better be the outermost loop.
2116   if (_head->req() > 3 && !_irreducible) {
2117     split_outer_loop( phase );
2118     result = true;
2119 
2120   } else if (!_head->is_Loop() && !_irreducible) {
2121     // Make a new LoopNode to replace the old loop head
2122     Node *l = new LoopNode( _head->in(1), _head->in(2) );
2123     l = igvn.register_new_node_with_optimizer(l, _head);
2124     phase->set_created_loop_node();
2125     // Go ahead and replace _head
2126     phase->_igvn.replace_node( _head, l );
2127     _head = l;
2128     phase->set_loop(_head, this);
2129   }
2130 
2131   // Now recursively beautify nested loops
2132   if( _child ) result |= _child->beautify_loops( phase );
2133   if( _next  ) result |= _next ->beautify_loops( phase );
2134   return result;
2135 }
2136 
2137 //------------------------------allpaths_check_safepts----------------------------
2138 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
2139 // encountered.  Helper for check_safepts.
2140 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
2141   assert(stack.size() == 0, "empty stack");
2142   stack.push(_tail);
2143   visited.Clear();
2144   visited.set(_tail->_idx);
2145   while (stack.size() > 0) {
2146     Node* n = stack.pop();
2147     if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
2148       // Terminate this path
2149     } else if (n->Opcode() == Op_SafePoint) {
2150       if (_phase->get_loop(n) != this) {
2151         if (_required_safept == NULL) _required_safept = new Node_List();
2152         _required_safept->push(n);  // save the one closest to the tail
2153       }
2154       // Terminate this path
2155     } else {
2156       uint start = n->is_Region() ? 1 : 0;
2157       uint end   = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
2158       for (uint i = start; i < end; i++) {
2159         Node* in = n->in(i);
2160         assert(in->is_CFG(), "must be");
2161         if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
2162           stack.push(in);
2163         }
2164       }
2165     }
2166   }
2167 }
2168 
2169 //------------------------------check_safepts----------------------------
2170 // Given dominators, try to find loops with calls that must always be
2171 // executed (call dominates loop tail).  These loops do not need non-call
2172 // safepoints (ncsfpt).
2173 //
2174 // A complication is that a safepoint in a inner loop may be needed
2175 // by an outer loop. In the following, the inner loop sees it has a
2176 // call (block 3) on every path from the head (block 2) to the
2177 // backedge (arc 3->2).  So it deletes the ncsfpt (non-call safepoint)
2178 // in block 2, _but_ this leaves the outer loop without a safepoint.
2179 //
2180 //          entry  0
2181 //                 |
2182 //                 v
2183 // outer 1,2    +->1
2184 //              |  |
2185 //              |  v
2186 //              |  2<---+  ncsfpt in 2
2187 //              |_/|\   |
2188 //                 | v  |
2189 // inner 2,3      /  3  |  call in 3
2190 //               /   |  |
2191 //              v    +--+
2192 //        exit  4
2193 //
2194 //
2195 // This method creates a list (_required_safept) of ncsfpt nodes that must
2196 // be protected is created for each loop. When a ncsfpt maybe deleted, it
2197 // is first looked for in the lists for the outer loops of the current loop.
2198 //
2199 // The insights into the problem:
2200 //  A) counted loops are okay
2201 //  B) innermost loops are okay (only an inner loop can delete
2202 //     a ncsfpt needed by an outer loop)
2203 //  C) a loop is immune from an inner loop deleting a safepoint
2204 //     if the loop has a call on the idom-path
2205 //  D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
2206 //     idom-path that is not in a nested loop
2207 //  E) otherwise, an ncsfpt on the idom-path that is nested in an inner
2208 //     loop needs to be prevented from deletion by an inner loop
2209 //
2210 // There are two analyses:
2211 //  1) The first, and cheaper one, scans the loop body from
2212 //     tail to head following the idom (immediate dominator)
2213 //     chain, looking for the cases (C,D,E) above.
2214 //     Since inner loops are scanned before outer loops, there is summary
2215 //     information about inner loops.  Inner loops can be skipped over
2216 //     when the tail of an inner loop is encountered.
2217 //
2218 //  2) The second, invoked if the first fails to find a call or ncsfpt on
2219 //     the idom path (which is rare), scans all predecessor control paths
2220 //     from the tail to the head, terminating a path when a call or sfpt
2221 //     is encountered, to find the ncsfpt's that are closest to the tail.
2222 //
2223 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
2224   // Bottom up traversal
2225   IdealLoopTree* ch = _child;
2226   if (_child) _child->check_safepts(visited, stack);
2227   if (_next)  _next ->check_safepts(visited, stack);
2228 
2229   if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
2230     bool  has_call         = false; // call on dom-path
2231     bool  has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
2232     Node* nonlocal_ncsfpt  = NULL;  // ncsfpt on dom-path at a deeper depth
2233     // Scan the dom-path nodes from tail to head
2234     for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
2235       if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
2236         has_call = true;
2237         _has_sfpt = 1;          // Then no need for a safept!
2238         break;
2239       } else if (n->Opcode() == Op_SafePoint) {
2240         if (_phase->get_loop(n) == this) {
2241           has_local_ncsfpt = true;
2242           break;
2243         }
2244         if (nonlocal_ncsfpt == NULL) {
2245           nonlocal_ncsfpt = n; // save the one closest to the tail
2246         }
2247       } else {
2248         IdealLoopTree* nlpt = _phase->get_loop(n);
2249         if (this != nlpt) {
2250           // If at an inner loop tail, see if the inner loop has already
2251           // recorded seeing a call on the dom-path (and stop.)  If not,
2252           // jump to the head of the inner loop.
2253           assert(is_member(nlpt), "nested loop");
2254           Node* tail = nlpt->_tail;
2255           if (tail->in(0)->is_If()) tail = tail->in(0);
2256           if (n == tail) {
2257             // If inner loop has call on dom-path, so does outer loop
2258             if (nlpt->_has_sfpt) {
2259               has_call = true;
2260               _has_sfpt = 1;
2261               break;
2262             }
2263             // Skip to head of inner loop
2264             assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
2265             n = nlpt->_head;
2266           }
2267         }
2268       }
2269     }
2270     // Record safept's that this loop needs preserved when an
2271     // inner loop attempts to delete it's safepoints.
2272     if (_child != NULL && !has_call && !has_local_ncsfpt) {
2273       if (nonlocal_ncsfpt != NULL) {
2274         if (_required_safept == NULL) _required_safept = new Node_List();
2275         _required_safept->push(nonlocal_ncsfpt);
2276       } else {
2277         // Failed to find a suitable safept on the dom-path.  Now use
2278         // an all paths walk from tail to head, looking for safepoints to preserve.
2279         allpaths_check_safepts(visited, stack);
2280       }
2281     }
2282   }
2283 }
2284 
2285 //---------------------------is_deleteable_safept----------------------------
2286 // Is safept not required by an outer loop?
2287 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
2288   assert(sfpt->Opcode() == Op_SafePoint, "");
2289   IdealLoopTree* lp = get_loop(sfpt)->_parent;
2290   while (lp != NULL) {
2291     Node_List* sfpts = lp->_required_safept;
2292     if (sfpts != NULL) {
2293       for (uint i = 0; i < sfpts->size(); i++) {
2294         if (sfpt == sfpts->at(i))
2295           return false;
2296       }
2297     }
2298     lp = lp->_parent;
2299   }
2300   return true;
2301 }
2302 
2303 //---------------------------replace_parallel_iv-------------------------------
2304 // Replace parallel induction variable (parallel to trip counter)
2305 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
2306   assert(loop->_head->is_CountedLoop(), "");
2307   CountedLoopNode *cl = loop->_head->as_CountedLoop();
2308   if (!cl->is_valid_counted_loop())
2309     return;         // skip malformed counted loop
2310   Node *incr = cl->incr();
2311   if (incr == NULL)
2312     return;         // Dead loop?
2313   Node *init = cl->init_trip();
2314   Node *phi  = cl->phi();
2315   int stride_con = cl->stride_con();
2316 
2317   // Visit all children, looking for Phis
2318   for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
2319     Node *out = cl->out(i);
2320     // Look for other phis (secondary IVs). Skip dead ones
2321     if (!out->is_Phi() || out == phi || !has_node(out))
2322       continue;
2323     PhiNode* phi2 = out->as_Phi();
2324     Node *incr2 = phi2->in( LoopNode::LoopBackControl );
2325     // Look for induction variables of the form:  X += constant
2326     if (phi2->region() != loop->_head ||
2327         incr2->req() != 3 ||
2328         incr2->in(1) != phi2 ||
2329         incr2 == incr ||
2330         incr2->Opcode() != Op_AddI ||
2331         !incr2->in(2)->is_Con())
2332       continue;
2333 
2334     // Check for parallel induction variable (parallel to trip counter)
2335     // via an affine function.  In particular, count-down loops with
2336     // count-up array indices are common. We only RCE references off
2337     // the trip-counter, so we need to convert all these to trip-counter
2338     // expressions.
2339     Node *init2 = phi2->in( LoopNode::EntryControl );
2340     int stride_con2 = incr2->in(2)->get_int();
2341 
2342     // The ratio of the two strides cannot be represented as an int
2343     // if stride_con2 is min_int and stride_con is -1.
2344     if (stride_con2 == min_jint && stride_con == -1) {
2345       continue;
2346     }
2347 
2348     // The general case here gets a little tricky.  We want to find the
2349     // GCD of all possible parallel IV's and make a new IV using this
2350     // GCD for the loop.  Then all possible IVs are simple multiples of
2351     // the GCD.  In practice, this will cover very few extra loops.
2352     // Instead we require 'stride_con2' to be a multiple of 'stride_con',
2353     // where +/-1 is the common case, but other integer multiples are
2354     // also easy to handle.
2355     int ratio_con = stride_con2/stride_con;
2356 
2357     if ((ratio_con * stride_con) == stride_con2) { // Check for exact
2358 #ifndef PRODUCT
2359       if (TraceLoopOpts) {
2360         tty->print("Parallel IV: %d ", phi2->_idx);
2361         loop->dump_head();
2362       }
2363 #endif
2364       // Convert to using the trip counter.  The parallel induction
2365       // variable differs from the trip counter by a loop-invariant
2366       // amount, the difference between their respective initial values.
2367       // It is scaled by the 'ratio_con'.
2368       Node* ratio = _igvn.intcon(ratio_con);
2369       set_ctrl(ratio, C->root());
2370       Node* ratio_init = new MulINode(init, ratio);
2371       _igvn.register_new_node_with_optimizer(ratio_init, init);
2372       set_early_ctrl(ratio_init);
2373       Node* diff = new SubINode(init2, ratio_init);
2374       _igvn.register_new_node_with_optimizer(diff, init2);
2375       set_early_ctrl(diff);
2376       Node* ratio_idx = new MulINode(phi, ratio);
2377       _igvn.register_new_node_with_optimizer(ratio_idx, phi);
2378       set_ctrl(ratio_idx, cl);
2379       Node* add = new AddINode(ratio_idx, diff);
2380       _igvn.register_new_node_with_optimizer(add);
2381       set_ctrl(add, cl);
2382       _igvn.replace_node( phi2, add );
2383       // Sometimes an induction variable is unused
2384       if (add->outcnt() == 0) {
2385         _igvn.remove_dead_node(add);
2386       }
2387       --i; // deleted this phi; rescan starting with next position
2388       continue;
2389     }
2390   }
2391 }
2392 
2393 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
2394   Node* keep = NULL;
2395   if (keep_one) {
2396     // Look for a safepoint on the idom-path.
2397     for (Node* i = tail(); i != _head; i = phase->idom(i)) {
2398       if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
2399         keep = i;
2400         break; // Found one
2401       }
2402     }
2403   }
2404 
2405   // Don't remove any safepoints if it is requested to keep a single safepoint and
2406   // no safepoint was found on idom-path. It is not safe to remove any safepoint
2407   // in this case since there's no safepoint dominating all paths in the loop body.
2408   bool prune = !keep_one || keep != NULL;
2409 
2410   // Delete other safepoints in this loop.
2411   Node_List* sfpts = _safepts;
2412   if (prune && sfpts != NULL) {
2413     assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint");
2414     for (uint i = 0; i < sfpts->size(); i++) {
2415       Node* n = sfpts->at(i);
2416       assert(phase->get_loop(n) == this, "");
2417       if (n != keep && phase->is_deleteable_safept(n)) {
2418         phase->lazy_replace(n, n->in(TypeFunc::Control));
2419       }
2420     }
2421   }
2422 }
2423 
2424 //------------------------------counted_loop-----------------------------------
2425 // Convert to counted loops where possible
2426 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
2427 
2428   // For grins, set the inner-loop flag here
2429   if (!_child) {
2430     if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
2431   }
2432 
2433   IdealLoopTree* loop = this;
2434   if (_head->is_CountedLoop() ||
2435       phase->is_counted_loop(_head, loop)) {
2436 
2437     if (LoopStripMiningIter == 0 || (LoopStripMiningIter > 1 && _child == NULL)) {
2438       // Indicate we do not need a safepoint here
2439       _has_sfpt = 1;
2440     }
2441 
2442     // Remove safepoints
2443     bool keep_one_sfpt = !(_has_call || _has_sfpt);
2444     remove_safepoints(phase, keep_one_sfpt);
2445 
2446     // Look for induction variables
2447     phase->replace_parallel_iv(this);
2448 
2449   } else if (_parent != NULL && !_irreducible) {
2450     // Not a counted loop. Keep one safepoint.
2451     bool keep_one_sfpt = true;
2452     remove_safepoints(phase, keep_one_sfpt);
2453   }
2454 
2455   // Recursively
2456   assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?");
2457   assert(loop->_child != this || (loop->_child->_child == NULL && loop->_child->_next == NULL), "would miss some loops");
2458   if (loop->_child && loop->_child != this) loop->_child->counted_loop(phase);
2459   if (loop->_next)  loop->_next ->counted_loop(phase);
2460 }
2461 
2462 #ifndef PRODUCT
2463 //------------------------------dump_head--------------------------------------
2464 // Dump 1 liner for loop header info
2465 void IdealLoopTree::dump_head( ) const {
2466   for (uint i=0; i<_nest; i++)
2467     tty->print("  ");
2468   tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
2469   if (_irreducible) tty->print(" IRREDUCIBLE");
2470   Node* entry = _head->is_Loop() ? _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl) : _head->in(LoopNode::EntryControl);
2471   Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
2472   if (predicate != NULL ) {
2473     tty->print(" limit_check");
2474     entry = PhaseIdealLoop::skip_loop_predicates(entry);
2475   }
2476   if (UseProfiledLoopPredicate) {
2477     predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate);
2478     if (predicate != NULL) {
2479       tty->print(" profile_predicated");
2480       entry = PhaseIdealLoop::skip_loop_predicates(entry);
2481     }
2482   }
2483   if (UseLoopPredicate) {
2484     predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
2485     if (predicate != NULL) {
2486       tty->print(" predicated");
2487     }
2488   }
2489   if (_head->is_CountedLoop()) {
2490     CountedLoopNode *cl = _head->as_CountedLoop();
2491     tty->print(" counted");
2492 
2493     Node* init_n = cl->init_trip();
2494     if (init_n  != NULL &&  init_n->is_Con())
2495       tty->print(" [%d,", cl->init_trip()->get_int());
2496     else
2497       tty->print(" [int,");
2498     Node* limit_n = cl->limit();
2499     if (limit_n  != NULL &&  limit_n->is_Con())
2500       tty->print("%d),", cl->limit()->get_int());
2501     else
2502       tty->print("int),");
2503     int stride_con  = cl->stride_con();
2504     if (stride_con > 0) tty->print("+");
2505     tty->print("%d", stride_con);
2506 
2507     tty->print(" (%0.f iters) ", cl->profile_trip_cnt());
2508 
2509     if (cl->is_pre_loop ()) tty->print(" pre" );
2510     if (cl->is_main_loop()) tty->print(" main");
2511     if (cl->is_post_loop()) tty->print(" post");
2512     if (cl->is_vectorized_loop()) tty->print(" vector");
2513     if (cl->range_checks_present()) tty->print(" rc ");
2514     if (cl->is_multiversioned()) tty->print(" multi ");
2515   }
2516   if (_has_call) tty->print(" has_call");
2517   if (_has_sfpt) tty->print(" has_sfpt");
2518   if (_rce_candidate) tty->print(" rce");
2519   if (_safepts != NULL && _safepts->size() > 0) {
2520     tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
2521   }
2522   if (_required_safept != NULL && _required_safept->size() > 0) {
2523     tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
2524   }
2525   if (Verbose) {
2526     tty->print(" body={"); _body.dump_simple(); tty->print(" }");
2527   }
2528   if (_head->is_Loop() && _head->as_Loop()->is_strip_mined()) {
2529     tty->print(" strip_mined");
2530   }
2531   tty->cr();
2532 }
2533 
2534 //------------------------------dump-------------------------------------------
2535 // Dump loops by loop tree
2536 void IdealLoopTree::dump( ) const {
2537   dump_head();
2538   if (_child) _child->dump();
2539   if (_next)  _next ->dump();
2540 }
2541 
2542 #endif
2543 
2544 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
2545   if (loop == root) {
2546     if (loop->_child != NULL) {
2547       log->begin_head("loop_tree");
2548       log->end_head();
2549       if( loop->_child ) log_loop_tree(root, loop->_child, log);
2550       log->tail("loop_tree");
2551       assert(loop->_next == NULL, "what?");
2552     }
2553   } else {
2554     Node* head = loop->_head;
2555     log->begin_head("loop");
2556     log->print(" idx='%d' ", head->_idx);
2557     if (loop->_irreducible) log->print("irreducible='1' ");
2558     if (head->is_Loop()) {
2559       if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
2560       if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
2561     }
2562     if (head->is_CountedLoop()) {
2563       CountedLoopNode* cl = head->as_CountedLoop();
2564       if (cl->is_pre_loop())  log->print("pre_loop='%d' ",  cl->main_idx());
2565       if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
2566       if (cl->is_post_loop()) log->print("post_loop='%d' ",  cl->main_idx());
2567     }
2568     log->end_head();
2569     if( loop->_child ) log_loop_tree(root, loop->_child, log);
2570     log->tail("loop");
2571     if( loop->_next  ) log_loop_tree(root, loop->_next, log);
2572   }
2573 }
2574 
2575 //---------------------collect_potentially_useful_predicates-----------------------
2576 // Helper function to collect potentially useful predicates to prevent them from
2577 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
2578 void PhaseIdealLoop::collect_potentially_useful_predicates(
2579                          IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
2580   if (loop->_child) { // child
2581     collect_potentially_useful_predicates(loop->_child, useful_predicates);
2582   }
2583 
2584   // self (only loops that we can apply loop predication may use their predicates)
2585   if (loop->_head->is_Loop() &&
2586       !loop->_irreducible    &&
2587       !loop->tail()->is_top()) {
2588     LoopNode* lpn = loop->_head->as_Loop();
2589     Node* entry = lpn->in(LoopNode::EntryControl);
2590     Node* predicate_proj = find_predicate(entry); // loop_limit_check first
2591     if (predicate_proj != NULL) { // right pattern that can be used by loop predication
2592       assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
2593       useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2594       entry = skip_loop_predicates(entry);
2595     }
2596     if (UseProfiledLoopPredicate) {
2597       predicate_proj = find_predicate(entry); // Predicate
2598       if (predicate_proj != NULL) {
2599         useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2600         entry = skip_loop_predicates(entry);
2601       }
2602     }
2603     predicate_proj = find_predicate(entry); // Predicate
2604     if (predicate_proj != NULL) {
2605       useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2606     }
2607   }
2608 
2609   if (loop->_next) { // sibling
2610     collect_potentially_useful_predicates(loop->_next, useful_predicates);
2611   }
2612 }
2613 
2614 //------------------------eliminate_useless_predicates-----------------------------
2615 // Eliminate all inserted predicates if they could not be used by loop predication.
2616 // Note: it will also eliminates loop limits check predicate since it also uses
2617 // Opaque1 node (see Parse::add_predicate()).
2618 void PhaseIdealLoop::eliminate_useless_predicates() {
2619   if (C->predicate_count() == 0)
2620     return; // no predicate left
2621 
2622   Unique_Node_List useful_predicates; // to store useful predicates
2623   if (C->has_loops()) {
2624     collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
2625   }
2626 
2627   for (int i = C->predicate_count(); i > 0; i--) {
2628      Node * n = C->predicate_opaque1_node(i-1);
2629      assert(n->Opcode() == Op_Opaque1, "must be");
2630      if (!useful_predicates.member(n)) { // not in the useful list
2631        _igvn.replace_node(n, n->in(1));
2632      }
2633   }
2634 }
2635 
2636 //------------------------process_expensive_nodes-----------------------------
2637 // Expensive nodes have their control input set to prevent the GVN
2638 // from commoning them and as a result forcing the resulting node to
2639 // be in a more frequent path. Use CFG information here, to change the
2640 // control inputs so that some expensive nodes can be commoned while
2641 // not executed more frequently.
2642 bool PhaseIdealLoop::process_expensive_nodes() {
2643   assert(OptimizeExpensiveOps, "optimization off?");
2644 
2645   // Sort nodes to bring similar nodes together
2646   C->sort_expensive_nodes();
2647 
2648   bool progress = false;
2649 
2650   for (int i = 0; i < C->expensive_count(); ) {
2651     Node* n = C->expensive_node(i);
2652     int start = i;
2653     // Find nodes similar to n
2654     i++;
2655     for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
2656     int end = i;
2657     // And compare them two by two
2658     for (int j = start; j < end; j++) {
2659       Node* n1 = C->expensive_node(j);
2660       if (is_node_unreachable(n1)) {
2661         continue;
2662       }
2663       for (int k = j+1; k < end; k++) {
2664         Node* n2 = C->expensive_node(k);
2665         if (is_node_unreachable(n2)) {
2666           continue;
2667         }
2668 
2669         assert(n1 != n2, "should be pair of nodes");
2670 
2671         Node* c1 = n1->in(0);
2672         Node* c2 = n2->in(0);
2673 
2674         Node* parent_c1 = c1;
2675         Node* parent_c2 = c2;
2676 
2677         // The call to get_early_ctrl_for_expensive() moves the
2678         // expensive nodes up but stops at loops that are in a if
2679         // branch. See whether we can exit the loop and move above the
2680         // If.
2681         if (c1->is_Loop()) {
2682           parent_c1 = c1->in(1);
2683         }
2684         if (c2->is_Loop()) {
2685           parent_c2 = c2->in(1);
2686         }
2687 
2688         if (parent_c1 == parent_c2) {
2689           _igvn._worklist.push(n1);
2690           _igvn._worklist.push(n2);
2691           continue;
2692         }
2693 
2694         // Look for identical expensive node up the dominator chain.
2695         if (is_dominator(c1, c2)) {
2696           c2 = c1;
2697         } else if (is_dominator(c2, c1)) {
2698           c1 = c2;
2699         } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
2700                    parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
2701           // Both branches have the same expensive node so move it up
2702           // before the if.
2703           c1 = c2 = idom(parent_c1->in(0));
2704         }
2705         // Do the actual moves
2706         if (n1->in(0) != c1) {
2707           _igvn.hash_delete(n1);
2708           n1->set_req(0, c1);
2709           _igvn.hash_insert(n1);
2710           _igvn._worklist.push(n1);
2711           progress = true;
2712         }
2713         if (n2->in(0) != c2) {
2714           _igvn.hash_delete(n2);
2715           n2->set_req(0, c2);
2716           _igvn.hash_insert(n2);
2717           _igvn._worklist.push(n2);
2718           progress = true;
2719         }
2720       }
2721     }
2722   }
2723 
2724   return progress;
2725 }
2726 
2727 
2728 //=============================================================================
2729 //----------------------------build_and_optimize-------------------------------
2730 // Create a PhaseLoop.  Build the ideal Loop tree.  Map each Ideal Node to
2731 // its corresponding LoopNode.  If 'optimize' is true, do some loop cleanups.
2732 void PhaseIdealLoop::build_and_optimize(LoopOptsMode mode) {
2733   bool do_split_ifs = (mode == LoopOptsDefault || mode == LoopOptsLastRound);
2734   bool skip_loop_opts = (mode == LoopOptsNone);
2735 #if INCLUDE_SHENANDOAHGC
2736   bool shenandoah_opts = (mode == LoopOptsShenandoahExpand ||
2737                           mode == LoopOptsShenandoahPostExpand);
2738 #endif
2739 
2740   ResourceMark rm;
2741 
2742   int old_progress = C->major_progress();
2743   uint orig_worklist_size = _igvn._worklist.size();
2744 
2745   // Reset major-progress flag for the driver's heuristics
2746   C->clear_major_progress();
2747 
2748 #ifndef PRODUCT
2749   // Capture for later assert
2750   uint unique = C->unique();
2751   _loop_invokes++;
2752   _loop_work += unique;
2753 #endif
2754 
2755   // True if the method has at least 1 irreducible loop
2756   _has_irreducible_loops = false;
2757 
2758   _created_loop_node = false;
2759 
2760   Arena *a = Thread::current()->resource_area();
2761   VectorSet visited(a);
2762   // Pre-grow the mapping from Nodes to IdealLoopTrees.
2763   _nodes.map(C->unique(), NULL);
2764   memset(_nodes.adr(), 0, wordSize * C->unique());
2765 
2766   // Pre-build the top-level outermost loop tree entry
2767   _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
2768   // Do not need a safepoint at the top level
2769   _ltree_root->_has_sfpt = 1;
2770 
2771   // Initialize Dominators.
2772   // Checked in clone_loop_predicate() during beautify_loops().
2773   _idom_size = 0;
2774   _idom      = NULL;
2775   _dom_depth = NULL;
2776   _dom_stk   = NULL;
2777 
2778   // Empty pre-order array
2779   allocate_preorders();
2780 
2781   // Build a loop tree on the fly.  Build a mapping from CFG nodes to
2782   // IdealLoopTree entries.  Data nodes are NOT walked.
2783   build_loop_tree();
2784   // Check for bailout, and return
2785   if (C->failing()) {
2786     return;
2787   }
2788 
2789   // No loops after all
2790   if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
2791 
2792   // There should always be an outer loop containing the Root and Return nodes.
2793   // If not, we have a degenerate empty program.  Bail out in this case.
2794   if (!has_node(C->root())) {
2795     if (!_verify_only) {
2796       C->clear_major_progress();
2797       C->record_method_not_compilable("empty program detected during loop optimization");
2798     }
2799     return;
2800   }
2801 
2802   // Nothing to do, so get out
2803   bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only SHENANDOAHGC_ONLY(&& !shenandoah_opts);
2804   bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
2805   if (stop_early && !do_expensive_nodes) {
2806     _igvn.optimize();           // Cleanup NeverBranches
2807     return;
2808   }
2809 
2810   // Set loop nesting depth
2811   _ltree_root->set_nest( 0 );
2812 
2813   // Split shared headers and insert loop landing pads.
2814   // Do not bother doing this on the Root loop of course.
2815   if( !_verify_me && !_verify_only && _ltree_root->_child ) {
2816     C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
2817     if( _ltree_root->_child->beautify_loops( this ) ) {
2818       // Re-build loop tree!
2819       _ltree_root->_child = NULL;
2820       _nodes.clear();
2821       reallocate_preorders();
2822       build_loop_tree();
2823       // Check for bailout, and return
2824       if (C->failing()) {
2825         return;
2826       }
2827       // Reset loop nesting depth
2828       _ltree_root->set_nest( 0 );
2829 
2830       C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
2831     }
2832   }
2833 
2834   // Build Dominators for elision of NULL checks & loop finding.
2835   // Since nodes do not have a slot for immediate dominator, make
2836   // a persistent side array for that info indexed on node->_idx.
2837   _idom_size = C->unique();
2838   _idom      = NEW_RESOURCE_ARRAY( Node*, _idom_size );
2839   _dom_depth = NEW_RESOURCE_ARRAY( uint,  _idom_size );
2840   _dom_stk   = NULL; // Allocated on demand in recompute_dom_depth
2841   memset( _dom_depth, 0, _idom_size * sizeof(uint) );
2842 
2843   Dominators();
2844 
2845   if (!_verify_only) {
2846     // As a side effect, Dominators removed any unreachable CFG paths
2847     // into RegionNodes.  It doesn't do this test against Root, so
2848     // we do it here.
2849     for( uint i = 1; i < C->root()->req(); i++ ) {
2850       if( !_nodes[C->root()->in(i)->_idx] ) {    // Dead path into Root?
2851         _igvn.delete_input_of(C->root(), i);
2852         i--;                      // Rerun same iteration on compressed edges
2853       }
2854     }
2855 
2856     // Given dominators, try to find inner loops with calls that must
2857     // always be executed (call dominates loop tail).  These loops do
2858     // not need a separate safepoint.
2859     Node_List cisstack(a);
2860     _ltree_root->check_safepts(visited, cisstack);
2861   }
2862 
2863   // Walk the DATA nodes and place into loops.  Find earliest control
2864   // node.  For CFG nodes, the _nodes array starts out and remains
2865   // holding the associated IdealLoopTree pointer.  For DATA nodes, the
2866   // _nodes array holds the earliest legal controlling CFG node.
2867 
2868   // Allocate stack with enough space to avoid frequent realloc
2869   int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
2870   Node_Stack nstack( a, stack_size );
2871 
2872   visited.Clear();
2873   Node_List worklist(a);
2874   // Don't need C->root() on worklist since
2875   // it will be processed among C->top() inputs
2876   worklist.push( C->top() );
2877   visited.set( C->top()->_idx ); // Set C->top() as visited now
2878   build_loop_early( visited, worklist, nstack );
2879 
2880   // Given early legal placement, try finding counted loops.  This placement
2881   // is good enough to discover most loop invariants.
2882   if( !_verify_me && !_verify_only SHENANDOAHGC_ONLY(&& !shenandoah_opts))
2883     _ltree_root->counted_loop( this );
2884 
2885   // Find latest loop placement.  Find ideal loop placement.
2886   visited.Clear();
2887   init_dom_lca_tags();
2888   // Need C->root() on worklist when processing outs
2889   worklist.push( C->root() );
2890   NOT_PRODUCT( C->verify_graph_edges(); )
2891   worklist.push( C->top() );
2892   build_loop_late( visited, worklist, nstack );
2893 
2894   if (_verify_only) {
2895     // restore major progress flag
2896     for (int i = 0; i < old_progress; i++)
2897       C->set_major_progress();
2898     assert(C->unique() == unique, "verification mode made Nodes? ? ?");
2899     assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
2900     return;
2901   }
2902 
2903   // clear out the dead code after build_loop_late
2904   while (_deadlist.size()) {
2905     _igvn.remove_globally_dead_node(_deadlist.pop());
2906   }
2907 
2908   if (stop_early) {
2909     assert(do_expensive_nodes, "why are we here?");
2910     if (process_expensive_nodes()) {
2911       // If we made some progress when processing expensive nodes then
2912       // the IGVN may modify the graph in a way that will allow us to
2913       // make some more progress: we need to try processing expensive
2914       // nodes again.
2915       C->set_major_progress();
2916     }
2917     _igvn.optimize();
2918     return;
2919   }
2920 
2921   // Some parser-inserted loop predicates could never be used by loop
2922   // predication or they were moved away from loop during some optimizations.
2923   // For example, peeling. Eliminate them before next loop optimizations.
2924   eliminate_useless_predicates();
2925 
2926 #ifndef PRODUCT
2927   C->verify_graph_edges();
2928   if (_verify_me) {             // Nested verify pass?
2929     // Check to see if the verify mode is broken
2930     assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
2931     return;
2932   }
2933   if(VerifyLoopOptimizations) verify();
2934   if(TraceLoopOpts && C->has_loops()) {
2935     _ltree_root->dump();
2936   }
2937 #endif
2938 
2939   if (skip_loop_opts) {
2940     // restore major progress flag
2941     for (int i = 0; i < old_progress; i++) {
2942       C->set_major_progress();
2943     }
2944 
2945     // Cleanup any modified bits
2946     _igvn.optimize();
2947 
2948     if (C->log() != NULL) {
2949       log_loop_tree(_ltree_root, _ltree_root, C->log());
2950     }
2951     return;
2952   }
2953 
2954 #if INCLUDE_SHENANDOAHGC
2955   if (UseShenandoahGC && ((ShenandoahBarrierSetC2*) BarrierSet::barrier_set()->barrier_set_c2())->optimize_loops(this, mode, visited, nstack, worklist)) {
2956     _igvn.optimize();
2957     if (C->log() != NULL) {
2958       log_loop_tree(_ltree_root, _ltree_root, C->log());
2959     }
2960     return;
2961   }
2962 #endif
2963 
2964   if (ReassociateInvariants) {
2965     // Reassociate invariants and prep for split_thru_phi
2966     for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2967       IdealLoopTree* lpt = iter.current();
2968       bool is_counted = lpt->is_counted();
2969       if (!is_counted || !lpt->is_inner()) continue;
2970 
2971       // check for vectorized loops, any reassociation of invariants was already done
2972       if (is_counted && lpt->_head->as_CountedLoop()->do_unroll_only()) continue;
2973 
2974       lpt->reassociate_invariants(this);
2975 
2976       // Because RCE opportunities can be masked by split_thru_phi,
2977       // look for RCE candidates and inhibit split_thru_phi
2978       // on just their loop-phi's for this pass of loop opts
2979       if (SplitIfBlocks && do_split_ifs) {
2980         if (lpt->policy_range_check(this)) {
2981           lpt->_rce_candidate = 1; // = true
2982         }
2983       }
2984     }
2985   }
2986 
2987   // Check for aggressive application of split-if and other transforms
2988   // that require basic-block info (like cloning through Phi's)
2989   if( SplitIfBlocks && do_split_ifs ) {
2990     visited.Clear();
2991     split_if_with_blocks( visited, nstack, mode == LoopOptsLastRound );
2992     NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
2993     if (mode == LoopOptsLastRound) {
2994       C->set_major_progress();
2995     }
2996   }
2997 
2998   if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
2999     C->set_major_progress();
3000   }
3001 
3002   // Perform loop predication before iteration splitting
3003   if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
3004     _ltree_root->_child->loop_predication(this);
3005   }
3006 
3007   if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
3008     if (do_intrinsify_fill()) {
3009       C->set_major_progress();
3010     }
3011   }
3012 
3013   // Perform iteration-splitting on inner loops.  Split iterations to avoid
3014   // range checks or one-shot null checks.
3015 
3016   // If split-if's didn't hack the graph too bad (no CFG changes)
3017   // then do loop opts.
3018   if (C->has_loops() && !C->major_progress()) {
3019     memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
3020     _ltree_root->_child->iteration_split( this, worklist );
3021     // No verify after peeling!  GCM has hoisted code out of the loop.
3022     // After peeling, the hoisted code could sink inside the peeled area.
3023     // The peeling code does not try to recompute the best location for
3024     // all the code before the peeled area, so the verify pass will always
3025     // complain about it.
3026   }
3027   // Do verify graph edges in any case
3028   NOT_PRODUCT( C->verify_graph_edges(); );
3029 
3030   if (!do_split_ifs) {
3031     // We saw major progress in Split-If to get here.  We forced a
3032     // pass with unrolling and not split-if, however more split-if's
3033     // might make progress.  If the unrolling didn't make progress
3034     // then the major-progress flag got cleared and we won't try
3035     // another round of Split-If.  In particular the ever-common
3036     // instance-of/check-cast pattern requires at least 2 rounds of
3037     // Split-If to clear out.
3038     C->set_major_progress();
3039   }
3040 
3041   // Repeat loop optimizations if new loops were seen
3042   if (created_loop_node()) {
3043     C->set_major_progress();
3044   }
3045 
3046   // Keep loop predicates and perform optimizations with them
3047   // until no more loop optimizations could be done.
3048   // After that switch predicates off and do more loop optimizations.
3049   if (!C->major_progress() && (C->predicate_count() > 0)) {
3050      C->cleanup_loop_predicates(_igvn);
3051      if (TraceLoopOpts) {
3052        tty->print_cr("PredicatesOff");
3053      }
3054      C->set_major_progress();
3055   }
3056 
3057   // Convert scalar to superword operations at the end of all loop opts.
3058   if (UseSuperWord && C->has_loops() && !C->major_progress()) {
3059     // SuperWord transform
3060     SuperWord sw(this);
3061     for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
3062       IdealLoopTree* lpt = iter.current();
3063       if (lpt->is_counted()) {
3064         CountedLoopNode *cl = lpt->_head->as_CountedLoop();
3065 
3066         if (PostLoopMultiversioning && cl->is_rce_post_loop() && !cl->is_vectorized_loop()) {
3067           // Check that the rce'd post loop is encountered first, multiversion after all
3068           // major main loop optimization are concluded
3069           if (!C->major_progress()) {
3070             IdealLoopTree *lpt_next = lpt->_next;
3071             if (lpt_next && lpt_next->is_counted()) {
3072               CountedLoopNode *cl = lpt_next->_head->as_CountedLoop();
3073               has_range_checks(lpt_next);
3074               if (cl->is_post_loop() && cl->range_checks_present()) {
3075                 if (!cl->is_multiversioned()) {
3076                   if (multi_version_post_loops(lpt, lpt_next) == false) {
3077                     // Cause the rce loop to be optimized away if we fail
3078                     cl->mark_is_multiversioned();
3079                     cl->set_slp_max_unroll(0);
3080                     poison_rce_post_loop(lpt);
3081                   }
3082                 }
3083               }
3084             }
3085             sw.transform_loop(lpt, true);
3086           }
3087         } else if (cl->is_main_loop()) {
3088           sw.transform_loop(lpt, true);
3089         }
3090       }
3091     }
3092   }
3093 
3094   // Cleanup any modified bits
3095   _igvn.optimize();
3096 
3097   // disable assert until issue with split_flow_path is resolved (6742111)
3098   // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
3099   //        "shouldn't introduce irreducible loops");
3100 
3101   if (C->log() != NULL) {
3102     log_loop_tree(_ltree_root, _ltree_root, C->log());
3103   }
3104 }
3105 
3106 #ifndef PRODUCT
3107 //------------------------------print_statistics-------------------------------
3108 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
3109 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
3110 void PhaseIdealLoop::print_statistics() {
3111   tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
3112 }
3113 
3114 //------------------------------verify-----------------------------------------
3115 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
3116 static int fail;                // debug only, so its multi-thread dont care
3117 void PhaseIdealLoop::verify() const {
3118   int old_progress = C->major_progress();
3119   ResourceMark rm;
3120   PhaseIdealLoop loop_verify( _igvn, this );
3121   VectorSet visited(Thread::current()->resource_area());
3122 
3123   fail = 0;
3124   verify_compare( C->root(), &loop_verify, visited );
3125   assert( fail == 0, "verify loops failed" );
3126   // Verify loop structure is the same
3127   _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
3128   // Reset major-progress.  It was cleared by creating a verify version of
3129   // PhaseIdealLoop.
3130   for( int i=0; i<old_progress; i++ )
3131     C->set_major_progress();
3132 }
3133 
3134 //------------------------------verify_compare---------------------------------
3135 // Make sure me and the given PhaseIdealLoop agree on key data structures
3136 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
3137   if( !n ) return;
3138   if( visited.test_set( n->_idx ) ) return;
3139   if( !_nodes[n->_idx] ) {      // Unreachable
3140     assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
3141     return;
3142   }
3143 
3144   uint i;
3145   for( i = 0; i < n->req(); i++ )
3146     verify_compare( n->in(i), loop_verify, visited );
3147 
3148   // Check the '_nodes' block/loop structure
3149   i = n->_idx;
3150   if( has_ctrl(n) ) {           // We have control; verify has loop or ctrl
3151     if( _nodes[i] != loop_verify->_nodes[i] &&
3152         get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
3153       tty->print("Mismatched control setting for: ");
3154       n->dump();
3155       if( fail++ > 10 ) return;
3156       Node *c = get_ctrl_no_update(n);
3157       tty->print("We have it as: ");
3158       if( c->in(0) ) c->dump();
3159         else tty->print_cr("N%d",c->_idx);
3160       tty->print("Verify thinks: ");
3161       if( loop_verify->has_ctrl(n) )
3162         loop_verify->get_ctrl_no_update(n)->dump();
3163       else
3164         loop_verify->get_loop_idx(n)->dump();
3165       tty->cr();
3166     }
3167   } else {                    // We have a loop
3168     IdealLoopTree *us = get_loop_idx(n);
3169     if( loop_verify->has_ctrl(n) ) {
3170       tty->print("Mismatched loop setting for: ");
3171       n->dump();
3172       if( fail++ > 10 ) return;
3173       tty->print("We have it as: ");
3174       us->dump();
3175       tty->print("Verify thinks: ");
3176       loop_verify->get_ctrl_no_update(n)->dump();
3177       tty->cr();
3178     } else if (!C->major_progress()) {
3179       // Loop selection can be messed up if we did a major progress
3180       // operation, like split-if.  Do not verify in that case.
3181       IdealLoopTree *them = loop_verify->get_loop_idx(n);
3182       if( us->_head != them->_head ||  us->_tail != them->_tail ) {
3183         tty->print("Unequals loops for: ");
3184         n->dump();
3185         if( fail++ > 10 ) return;
3186         tty->print("We have it as: ");
3187         us->dump();
3188         tty->print("Verify thinks: ");
3189         them->dump();
3190         tty->cr();
3191       }
3192     }
3193   }
3194 
3195   // Check for immediate dominators being equal
3196   if( i >= _idom_size ) {
3197     if( !n->is_CFG() ) return;
3198     tty->print("CFG Node with no idom: ");
3199     n->dump();
3200     return;
3201   }
3202   if( !n->is_CFG() ) return;
3203   if( n == C->root() ) return; // No IDOM here
3204 
3205   assert(n->_idx == i, "sanity");
3206   Node *id = idom_no_update(n);
3207   if( id != loop_verify->idom_no_update(n) ) {
3208     tty->print("Unequals idoms for: ");
3209     n->dump();
3210     if( fail++ > 10 ) return;
3211     tty->print("We have it as: ");
3212     id->dump();
3213     tty->print("Verify thinks: ");
3214     loop_verify->idom_no_update(n)->dump();
3215     tty->cr();
3216   }
3217 
3218 }
3219 
3220 //------------------------------verify_tree------------------------------------
3221 // Verify that tree structures match.  Because the CFG can change, siblings
3222 // within the loop tree can be reordered.  We attempt to deal with that by
3223 // reordering the verify's loop tree if possible.
3224 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
3225   assert( _parent == parent, "Badly formed loop tree" );
3226 
3227   // Siblings not in same order?  Attempt to re-order.
3228   if( _head != loop->_head ) {
3229     // Find _next pointer to update
3230     IdealLoopTree **pp = &loop->_parent->_child;
3231     while( *pp != loop )
3232       pp = &((*pp)->_next);
3233     // Find proper sibling to be next
3234     IdealLoopTree **nn = &loop->_next;
3235     while( (*nn) && (*nn)->_head != _head )
3236       nn = &((*nn)->_next);
3237 
3238     // Check for no match.
3239     if( !(*nn) ) {
3240       // Annoyingly, irreducible loops can pick different headers
3241       // after a major_progress operation, so the rest of the loop
3242       // tree cannot be matched.
3243       if (_irreducible && Compile::current()->major_progress())  return;
3244       assert( 0, "failed to match loop tree" );
3245     }
3246 
3247     // Move (*nn) to (*pp)
3248     IdealLoopTree *hit = *nn;
3249     *nn = hit->_next;
3250     hit->_next = loop;
3251     *pp = loop;
3252     loop = hit;
3253     // Now try again to verify
3254   }
3255 
3256   assert( _head  == loop->_head , "mismatched loop head" );
3257   Node *tail = _tail;           // Inline a non-updating version of
3258   while( !tail->in(0) )         // the 'tail()' call.
3259     tail = tail->in(1);
3260   assert( tail == loop->_tail, "mismatched loop tail" );
3261 
3262   // Counted loops that are guarded should be able to find their guards
3263   if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
3264     CountedLoopNode *cl = _head->as_CountedLoop();
3265     Node *init = cl->init_trip();
3266     Node *ctrl = cl->in(LoopNode::EntryControl);
3267     assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
3268     Node *iff  = ctrl->in(0);
3269     assert( iff->Opcode() == Op_If, "" );
3270     Node *bol  = iff->in(1);
3271     assert( bol->Opcode() == Op_Bool, "" );
3272     Node *cmp  = bol->in(1);
3273     assert( cmp->Opcode() == Op_CmpI, "" );
3274     Node *add  = cmp->in(1);
3275     Node *opaq;
3276     if( add->Opcode() == Op_Opaque1 ) {
3277       opaq = add;
3278     } else {
3279       assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
3280       assert( add == init, "" );
3281       opaq = cmp->in(2);
3282     }
3283     assert( opaq->Opcode() == Op_Opaque1, "" );
3284 
3285   }
3286 
3287   if (_child != NULL)  _child->verify_tree(loop->_child, this);
3288   if (_next  != NULL)  _next ->verify_tree(loop->_next,  parent);
3289   // Innermost loops need to verify loop bodies,
3290   // but only if no 'major_progress'
3291   int fail = 0;
3292   if (!Compile::current()->major_progress() && _child == NULL) {
3293     for( uint i = 0; i < _body.size(); i++ ) {
3294       Node *n = _body.at(i);
3295       if (n->outcnt() == 0)  continue; // Ignore dead
3296       uint j;
3297       for( j = 0; j < loop->_body.size(); j++ )
3298         if( loop->_body.at(j) == n )
3299           break;
3300       if( j == loop->_body.size() ) { // Not found in loop body
3301         // Last ditch effort to avoid assertion: Its possible that we
3302         // have some users (so outcnt not zero) but are still dead.
3303         // Try to find from root.
3304         if (Compile::current()->root()->find(n->_idx)) {
3305           fail++;
3306           tty->print("We have that verify does not: ");
3307           n->dump();
3308         }
3309       }
3310     }
3311     for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
3312       Node *n = loop->_body.at(i2);
3313       if (n->outcnt() == 0)  continue; // Ignore dead
3314       uint j;
3315       for( j = 0; j < _body.size(); j++ )
3316         if( _body.at(j) == n )
3317           break;
3318       if( j == _body.size() ) { // Not found in loop body
3319         // Last ditch effort to avoid assertion: Its possible that we
3320         // have some users (so outcnt not zero) but are still dead.
3321         // Try to find from root.
3322         if (Compile::current()->root()->find(n->_idx)) {
3323           fail++;
3324           tty->print("Verify has that we do not: ");
3325           n->dump();
3326         }
3327       }
3328     }
3329     assert( !fail, "loop body mismatch" );
3330   }
3331 }
3332 
3333 #endif
3334 
3335 //------------------------------set_idom---------------------------------------
3336 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
3337   uint idx = d->_idx;
3338   if (idx >= _idom_size) {
3339     uint newsize = _idom_size<<1;
3340     while( idx >= newsize ) {
3341       newsize <<= 1;
3342     }
3343     _idom      = REALLOC_RESOURCE_ARRAY( Node*,     _idom,_idom_size,newsize);
3344     _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
3345     memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
3346     _idom_size = newsize;
3347   }
3348   _idom[idx] = n;
3349   _dom_depth[idx] = dom_depth;
3350 }
3351 
3352 //------------------------------recompute_dom_depth---------------------------------------
3353 // The dominator tree is constructed with only parent pointers.
3354 // This recomputes the depth in the tree by first tagging all
3355 // nodes as "no depth yet" marker.  The next pass then runs up
3356 // the dom tree from each node marked "no depth yet", and computes
3357 // the depth on the way back down.
3358 void PhaseIdealLoop::recompute_dom_depth() {
3359   uint no_depth_marker = C->unique();
3360   uint i;
3361   // Initialize depth to "no depth yet" and realize all lazy updates
3362   for (i = 0; i < _idom_size; i++) {
3363     // Only indices with a _dom_depth has a Node* or NULL (otherwise uninitalized).
3364     if (_dom_depth[i] > 0 && _idom[i] != NULL) {
3365       _dom_depth[i] = no_depth_marker;
3366 
3367       // heal _idom if it has a fwd mapping in _nodes
3368       if (_idom[i]->in(0) == NULL) {
3369         idom(i);
3370       }
3371     }
3372   }
3373   if (_dom_stk == NULL) {
3374     uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
3375     if (init_size < 10) init_size = 10;
3376     _dom_stk = new GrowableArray<uint>(init_size);
3377   }
3378   // Compute new depth for each node.
3379   for (i = 0; i < _idom_size; i++) {
3380     uint j = i;
3381     // Run up the dom tree to find a node with a depth
3382     while (_dom_depth[j] == no_depth_marker) {
3383       _dom_stk->push(j);
3384       j = _idom[j]->_idx;
3385     }
3386     // Compute the depth on the way back down this tree branch
3387     uint dd = _dom_depth[j] + 1;
3388     while (_dom_stk->length() > 0) {
3389       uint j = _dom_stk->pop();
3390       _dom_depth[j] = dd;
3391       dd++;
3392     }
3393   }
3394 }
3395 
3396 //------------------------------sort-------------------------------------------
3397 // Insert 'loop' into the existing loop tree.  'innermost' is a leaf of the
3398 // loop tree, not the root.
3399 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
3400   if( !innermost ) return loop; // New innermost loop
3401 
3402   int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
3403   assert( loop_preorder, "not yet post-walked loop" );
3404   IdealLoopTree **pp = &innermost;      // Pointer to previous next-pointer
3405   IdealLoopTree *l = *pp;               // Do I go before or after 'l'?
3406 
3407   // Insert at start of list
3408   while( l ) {                  // Insertion sort based on pre-order
3409     if( l == loop ) return innermost; // Already on list!
3410     int l_preorder = get_preorder(l->_head); // Cache pre-order number
3411     assert( l_preorder, "not yet post-walked l" );
3412     // Check header pre-order number to figure proper nesting
3413     if( loop_preorder > l_preorder )
3414       break;                    // End of insertion
3415     // If headers tie (e.g., shared headers) check tail pre-order numbers.
3416     // Since I split shared headers, you'd think this could not happen.
3417     // BUT: I must first do the preorder numbering before I can discover I
3418     // have shared headers, so the split headers all get the same preorder
3419     // number as the RegionNode they split from.
3420     if( loop_preorder == l_preorder &&
3421         get_preorder(loop->_tail) < get_preorder(l->_tail) )
3422       break;                    // Also check for shared headers (same pre#)
3423     pp = &l->_parent;           // Chain up list
3424     l = *pp;
3425   }
3426   // Link into list
3427   // Point predecessor to me
3428   *pp = loop;
3429   // Point me to successor
3430   IdealLoopTree *p = loop->_parent;
3431   loop->_parent = l;            // Point me to successor
3432   if( p ) sort( p, innermost ); // Insert my parents into list as well
3433   return innermost;
3434 }
3435 
3436 //------------------------------build_loop_tree--------------------------------
3437 // I use a modified Vick/Tarjan algorithm.  I need pre- and a post- visit
3438 // bits.  The _nodes[] array is mapped by Node index and holds a NULL for
3439 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
3440 // tightest enclosing IdealLoopTree for post-walked.
3441 //
3442 // During my forward walk I do a short 1-layer lookahead to see if I can find
3443 // a loop backedge with that doesn't have any work on the backedge.  This
3444 // helps me construct nested loops with shared headers better.
3445 //
3446 // Once I've done the forward recursion, I do the post-work.  For each child
3447 // I check to see if there is a backedge.  Backedges define a loop!  I
3448 // insert an IdealLoopTree at the target of the backedge.
3449 //
3450 // During the post-work I also check to see if I have several children
3451 // belonging to different loops.  If so, then this Node is a decision point
3452 // where control flow can choose to change loop nests.  It is at this
3453 // decision point where I can figure out how loops are nested.  At this
3454 // time I can properly order the different loop nests from my children.
3455 // Note that there may not be any backedges at the decision point!
3456 //
3457 // Since the decision point can be far removed from the backedges, I can't
3458 // order my loops at the time I discover them.  Thus at the decision point
3459 // I need to inspect loop header pre-order numbers to properly nest my
3460 // loops.  This means I need to sort my childrens' loops by pre-order.
3461 // The sort is of size number-of-control-children, which generally limits
3462 // it to size 2 (i.e., I just choose between my 2 target loops).
3463 void PhaseIdealLoop::build_loop_tree() {
3464   // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3465   GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
3466   Node *n = C->root();
3467   bltstack.push(n);
3468   int pre_order = 1;
3469   int stack_size;
3470 
3471   while ( ( stack_size = bltstack.length() ) != 0 ) {
3472     n = bltstack.top(); // Leave node on stack
3473     if ( !is_visited(n) ) {
3474       // ---- Pre-pass Work ----
3475       // Pre-walked but not post-walked nodes need a pre_order number.
3476 
3477       set_preorder_visited( n, pre_order ); // set as visited
3478 
3479       // ---- Scan over children ----
3480       // Scan first over control projections that lead to loop headers.
3481       // This helps us find inner-to-outer loops with shared headers better.
3482 
3483       // Scan children's children for loop headers.
3484       for ( int i = n->outcnt() - 1; i >= 0; --i ) {
3485         Node* m = n->raw_out(i);       // Child
3486         if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
3487           // Scan over children's children to find loop
3488           for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3489             Node* l = m->fast_out(j);
3490             if( is_visited(l) &&       // Been visited?
3491                 !is_postvisited(l) &&  // But not post-visited
3492                 get_preorder(l) < pre_order ) { // And smaller pre-order
3493               // Found!  Scan the DFS down this path before doing other paths
3494               bltstack.push(m);
3495               break;
3496             }
3497           }
3498         }
3499       }
3500       pre_order++;
3501     }
3502     else if ( !is_postvisited(n) ) {
3503       // Note: build_loop_tree_impl() adds out edges on rare occasions,
3504       // such as com.sun.rsasign.am::a.
3505       // For non-recursive version, first, process current children.
3506       // On next iteration, check if additional children were added.
3507       for ( int k = n->outcnt() - 1; k >= 0; --k ) {
3508         Node* u = n->raw_out(k);
3509         if ( u->is_CFG() && !is_visited(u) ) {
3510           bltstack.push(u);
3511         }
3512       }
3513       if ( bltstack.length() == stack_size ) {
3514         // There were no additional children, post visit node now
3515         (void)bltstack.pop(); // Remove node from stack
3516         pre_order = build_loop_tree_impl( n, pre_order );
3517         // Check for bailout
3518         if (C->failing()) {
3519           return;
3520         }
3521         // Check to grow _preorders[] array for the case when
3522         // build_loop_tree_impl() adds new nodes.
3523         check_grow_preorders();
3524       }
3525     }
3526     else {
3527       (void)bltstack.pop(); // Remove post-visited node from stack
3528     }
3529   }
3530 }
3531 
3532 //------------------------------build_loop_tree_impl---------------------------
3533 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
3534   // ---- Post-pass Work ----
3535   // Pre-walked but not post-walked nodes need a pre_order number.
3536 
3537   // Tightest enclosing loop for this Node
3538   IdealLoopTree *innermost = NULL;
3539 
3540   // For all children, see if any edge is a backedge.  If so, make a loop
3541   // for it.  Then find the tightest enclosing loop for the self Node.
3542   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3543     Node* m = n->fast_out(i);   // Child
3544     if( n == m ) continue;      // Ignore control self-cycles
3545     if( !m->is_CFG() ) continue;// Ignore non-CFG edges
3546 
3547     IdealLoopTree *l;           // Child's loop
3548     if( !is_postvisited(m) ) {  // Child visited but not post-visited?
3549       // Found a backedge
3550       assert( get_preorder(m) < pre_order, "should be backedge" );
3551       // Check for the RootNode, which is already a LoopNode and is allowed
3552       // to have multiple "backedges".
3553       if( m == C->root()) {     // Found the root?
3554         l = _ltree_root;        // Root is the outermost LoopNode
3555       } else {                  // Else found a nested loop
3556         // Insert a LoopNode to mark this loop.
3557         l = new IdealLoopTree(this, m, n);
3558       } // End of Else found a nested loop
3559       if( !has_loop(m) )        // If 'm' does not already have a loop set
3560         set_loop(m, l);         // Set loop header to loop now
3561 
3562     } else {                    // Else not a nested loop
3563       if( !_nodes[m->_idx] ) continue; // Dead code has no loop
3564       l = get_loop(m);          // Get previously determined loop
3565       // If successor is header of a loop (nest), move up-loop till it
3566       // is a member of some outer enclosing loop.  Since there are no
3567       // shared headers (I've split them already) I only need to go up
3568       // at most 1 level.
3569       while( l && l->_head == m ) // Successor heads loop?
3570         l = l->_parent;         // Move up 1 for me
3571       // If this loop is not properly parented, then this loop
3572       // has no exit path out, i.e. its an infinite loop.
3573       if( !l ) {
3574         // Make loop "reachable" from root so the CFG is reachable.  Basically
3575         // insert a bogus loop exit that is never taken.  'm', the loop head,
3576         // points to 'n', one (of possibly many) fall-in paths.  There may be
3577         // many backedges as well.
3578 
3579         // Here I set the loop to be the root loop.  I could have, after
3580         // inserting a bogus loop exit, restarted the recursion and found my
3581         // new loop exit.  This would make the infinite loop a first-class
3582         // loop and it would then get properly optimized.  What's the use of
3583         // optimizing an infinite loop?
3584         l = _ltree_root;        // Oops, found infinite loop
3585 
3586         if (!_verify_only) {
3587           // Insert the NeverBranch between 'm' and it's control user.
3588           NeverBranchNode *iff = new NeverBranchNode( m );
3589           _igvn.register_new_node_with_optimizer(iff);
3590           set_loop(iff, l);
3591           Node *if_t = new CProjNode( iff, 0 );
3592           _igvn.register_new_node_with_optimizer(if_t);
3593           set_loop(if_t, l);
3594 
3595           Node* cfg = NULL;       // Find the One True Control User of m
3596           for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3597             Node* x = m->fast_out(j);
3598             if (x->is_CFG() && x != m && x != iff)
3599               { cfg = x; break; }
3600           }
3601           assert(cfg != NULL, "must find the control user of m");
3602           uint k = 0;             // Probably cfg->in(0)
3603           while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
3604           cfg->set_req( k, if_t ); // Now point to NeverBranch
3605           _igvn._worklist.push(cfg);
3606 
3607           // Now create the never-taken loop exit
3608           Node *if_f = new CProjNode( iff, 1 );
3609           _igvn.register_new_node_with_optimizer(if_f);
3610           set_loop(if_f, l);
3611           // Find frame ptr for Halt.  Relies on the optimizer
3612           // V-N'ing.  Easier and quicker than searching through
3613           // the program structure.
3614           Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr );
3615           _igvn.register_new_node_with_optimizer(frame);
3616           // Halt & Catch Fire
3617           Node* halt = new HaltNode(if_f, frame, "never-taken loop exit reached");
3618           _igvn.register_new_node_with_optimizer(halt);
3619           set_loop(halt, l);
3620           C->root()->add_req(halt);
3621         }
3622         set_loop(C->root(), _ltree_root);
3623       }
3624     }
3625     // Weeny check for irreducible.  This child was already visited (this
3626     // IS the post-work phase).  Is this child's loop header post-visited
3627     // as well?  If so, then I found another entry into the loop.
3628     if (!_verify_only) {
3629       while( is_postvisited(l->_head) ) {
3630         // found irreducible
3631         l->_irreducible = 1; // = true
3632         l = l->_parent;
3633         _has_irreducible_loops = true;
3634         // Check for bad CFG here to prevent crash, and bailout of compile
3635         if (l == NULL) {
3636           C->record_method_not_compilable("unhandled CFG detected during loop optimization");
3637           return pre_order;
3638         }
3639       }
3640       C->set_has_irreducible_loop(_has_irreducible_loops);
3641     }
3642 
3643     // This Node might be a decision point for loops.  It is only if
3644     // it's children belong to several different loops.  The sort call
3645     // does a trivial amount of work if there is only 1 child or all
3646     // children belong to the same loop.  If however, the children
3647     // belong to different loops, the sort call will properly set the
3648     // _parent pointers to show how the loops nest.
3649     //
3650     // In any case, it returns the tightest enclosing loop.
3651     innermost = sort( l, innermost );
3652   }
3653 
3654   // Def-use info will have some dead stuff; dead stuff will have no
3655   // loop decided on.
3656 
3657   // Am I a loop header?  If so fix up my parent's child and next ptrs.
3658   if( innermost && innermost->_head == n ) {
3659     assert( get_loop(n) == innermost, "" );
3660     IdealLoopTree *p = innermost->_parent;
3661     IdealLoopTree *l = innermost;
3662     while( p && l->_head == n ) {
3663       l->_next = p->_child;     // Put self on parents 'next child'
3664       p->_child = l;            // Make self as first child of parent
3665       l = p;                    // Now walk up the parent chain
3666       p = l->_parent;
3667     }
3668   } else {
3669     // Note that it is possible for a LoopNode to reach here, if the
3670     // backedge has been made unreachable (hence the LoopNode no longer
3671     // denotes a Loop, and will eventually be removed).
3672 
3673     // Record tightest enclosing loop for self.  Mark as post-visited.
3674     set_loop(n, innermost);
3675     // Also record has_call flag early on
3676     if( innermost ) {
3677       if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
3678         // Do not count uncommon calls
3679         if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
3680           Node *iff = n->in(0)->in(0);
3681           // No any calls for vectorized loops.
3682           if( UseSuperWord || !iff->is_If() ||
3683               (n->in(0)->Opcode() == Op_IfFalse &&
3684                (1.0 - iff->as_If()->_prob) >= 0.01) ||
3685               (iff->as_If()->_prob >= 0.01) )
3686             innermost->_has_call = 1;
3687         }
3688       } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
3689         // Disable loop optimizations if the loop has a scalar replaceable
3690         // allocation. This disabling may cause a potential performance lost
3691         // if the allocation is not eliminated for some reason.
3692         innermost->_allow_optimizations = false;
3693         innermost->_has_call = 1; // = true
3694       } else if (n->Opcode() == Op_SafePoint) {
3695         // Record all safepoints in this loop.
3696         if (innermost->_safepts == NULL) innermost->_safepts = new Node_List();
3697         innermost->_safepts->push(n);
3698       }
3699     }
3700   }
3701 
3702   // Flag as post-visited now
3703   set_postvisited(n);
3704   return pre_order;
3705 }
3706 
3707 
3708 //------------------------------build_loop_early-------------------------------
3709 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3710 // First pass computes the earliest controlling node possible.  This is the
3711 // controlling input with the deepest dominating depth.
3712 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3713   while (worklist.size() != 0) {
3714     // Use local variables nstack_top_n & nstack_top_i to cache values
3715     // on nstack's top.
3716     Node *nstack_top_n = worklist.pop();
3717     uint  nstack_top_i = 0;
3718 //while_nstack_nonempty:
3719     while (true) {
3720       // Get parent node and next input's index from stack's top.
3721       Node  *n = nstack_top_n;
3722       uint   i = nstack_top_i;
3723       uint cnt = n->req(); // Count of inputs
3724       if (i == 0) {        // Pre-process the node.
3725         if( has_node(n) &&            // Have either loop or control already?
3726             !has_ctrl(n) ) {          // Have loop picked out already?
3727           // During "merge_many_backedges" we fold up several nested loops
3728           // into a single loop.  This makes the members of the original
3729           // loop bodies pointing to dead loops; they need to move up
3730           // to the new UNION'd larger loop.  I set the _head field of these
3731           // dead loops to NULL and the _parent field points to the owning
3732           // loop.  Shades of UNION-FIND algorithm.
3733           IdealLoopTree *ilt;
3734           while( !(ilt = get_loop(n))->_head ) {
3735             // Normally I would use a set_loop here.  But in this one special
3736             // case, it is legal (and expected) to change what loop a Node
3737             // belongs to.
3738             _nodes.map(n->_idx, (Node*)(ilt->_parent) );
3739           }
3740           // Remove safepoints ONLY if I've already seen I don't need one.
3741           // (the old code here would yank a 2nd safepoint after seeing a
3742           // first one, even though the 1st did not dominate in the loop body
3743           // and thus could be avoided indefinitely)
3744           if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
3745               is_deleteable_safept(n)) {
3746             Node *in = n->in(TypeFunc::Control);
3747             lazy_replace(n,in);       // Pull safepoint now
3748             if (ilt->_safepts != NULL) {
3749               ilt->_safepts->yank(n);
3750             }
3751             // Carry on with the recursion "as if" we are walking
3752             // only the control input
3753             if( !visited.test_set( in->_idx ) ) {
3754               worklist.push(in);      // Visit this guy later, using worklist
3755             }
3756             // Get next node from nstack:
3757             // - skip n's inputs processing by setting i > cnt;
3758             // - we also will not call set_early_ctrl(n) since
3759             //   has_node(n) == true (see the condition above).
3760             i = cnt + 1;
3761           }
3762         }
3763       } // if (i == 0)
3764 
3765       // Visit all inputs
3766       bool done = true;       // Assume all n's inputs will be processed
3767       while (i < cnt) {
3768         Node *in = n->in(i);
3769         ++i;
3770         if (in == NULL) continue;
3771         if (in->pinned() && !in->is_CFG())
3772           set_ctrl(in, in->in(0));
3773         int is_visited = visited.test_set( in->_idx );
3774         if (!has_node(in)) {  // No controlling input yet?
3775           assert( !in->is_CFG(), "CFG Node with no controlling input?" );
3776           assert( !is_visited, "visit only once" );
3777           nstack.push(n, i);  // Save parent node and next input's index.
3778           nstack_top_n = in;  // Process current input now.
3779           nstack_top_i = 0;
3780           done = false;       // Not all n's inputs processed.
3781           break; // continue while_nstack_nonempty;
3782         } else if (!is_visited) {
3783           // This guy has a location picked out for him, but has not yet
3784           // been visited.  Happens to all CFG nodes, for instance.
3785           // Visit him using the worklist instead of recursion, to break
3786           // cycles.  Since he has a location already we do not need to
3787           // find his location before proceeding with the current Node.
3788           worklist.push(in);  // Visit this guy later, using worklist
3789         }
3790       }
3791       if (done) {
3792         // All of n's inputs have been processed, complete post-processing.
3793 
3794         // Compute earliest point this Node can go.
3795         // CFG, Phi, pinned nodes already know their controlling input.
3796         if (!has_node(n)) {
3797           // Record earliest legal location
3798           set_early_ctrl( n );
3799         }
3800         if (nstack.is_empty()) {
3801           // Finished all nodes on stack.
3802           // Process next node on the worklist.
3803           break;
3804         }
3805         // Get saved parent node and next input's index.
3806         nstack_top_n = nstack.node();
3807         nstack_top_i = nstack.index();
3808         nstack.pop();
3809       }
3810     } // while (true)
3811   }
3812 }
3813 
3814 //------------------------------dom_lca_internal--------------------------------
3815 // Pair-wise LCA
3816 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
3817   if( !n1 ) return n2;          // Handle NULL original LCA
3818   assert( n1->is_CFG(), "" );
3819   assert( n2->is_CFG(), "" );
3820   // find LCA of all uses
3821   uint d1 = dom_depth(n1);
3822   uint d2 = dom_depth(n2);
3823   while (n1 != n2) {
3824     if (d1 > d2) {
3825       n1 =      idom(n1);
3826       d1 = dom_depth(n1);
3827     } else if (d1 < d2) {
3828       n2 =      idom(n2);
3829       d2 = dom_depth(n2);
3830     } else {
3831       // Here d1 == d2.  Due to edits of the dominator-tree, sections
3832       // of the tree might have the same depth.  These sections have
3833       // to be searched more carefully.
3834 
3835       // Scan up all the n1's with equal depth, looking for n2.
3836       Node *t1 = idom(n1);
3837       while (dom_depth(t1) == d1) {
3838         if (t1 == n2)  return n2;
3839         t1 = idom(t1);
3840       }
3841       // Scan up all the n2's with equal depth, looking for n1.
3842       Node *t2 = idom(n2);
3843       while (dom_depth(t2) == d2) {
3844         if (t2 == n1)  return n1;
3845         t2 = idom(t2);
3846       }
3847       // Move up to a new dominator-depth value as well as up the dom-tree.
3848       n1 = t1;
3849       n2 = t2;
3850       d1 = dom_depth(n1);
3851       d2 = dom_depth(n2);
3852     }
3853   }
3854   return n1;
3855 }
3856 
3857 //------------------------------compute_idom-----------------------------------
3858 // Locally compute IDOM using dom_lca call.  Correct only if the incoming
3859 // IDOMs are correct.
3860 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
3861   assert( region->is_Region(), "" );
3862   Node *LCA = NULL;
3863   for( uint i = 1; i < region->req(); i++ ) {
3864     if( region->in(i) != C->top() )
3865       LCA = dom_lca( LCA, region->in(i) );
3866   }
3867   return LCA;
3868 }
3869 
3870 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
3871   bool had_error = false;
3872 #ifdef ASSERT
3873   if (early != C->root()) {
3874     // Make sure that there's a dominance path from LCA to early
3875     Node* d = LCA;
3876     while (d != early) {
3877       if (d == C->root()) {
3878         dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
3879         tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
3880         had_error = true;
3881         break;
3882       }
3883       d = idom(d);
3884     }
3885   }
3886 #endif
3887   return had_error;
3888 }
3889 
3890 
3891 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
3892   // Compute LCA over list of uses
3893   bool had_error = false;
3894   Node *LCA = NULL;
3895   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
3896     Node* c = n->fast_out(i);
3897     if (_nodes[c->_idx] == NULL)
3898       continue;                 // Skip the occasional dead node
3899     if( c->is_Phi() ) {         // For Phis, we must land above on the path
3900       for( uint j=1; j<c->req(); j++ ) {// For all inputs
3901         if( c->in(j) == n ) {   // Found matching input?
3902           Node *use = c->in(0)->in(j);
3903           if (_verify_only && use->is_top()) continue;
3904           LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3905           if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3906         }
3907       }
3908     } else {
3909       // For CFG data-users, use is in the block just prior
3910       Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
3911       LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3912       if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3913     }
3914   }
3915   assert(!had_error, "bad dominance");
3916   return LCA;
3917 }
3918 
3919 // Check the shape of the graph at the loop entry. In some cases,
3920 // the shape of the graph does not match the shape outlined below.
3921 // That is caused by the Opaque1 node "protecting" the shape of
3922 // the graph being removed by, for example, the IGVN performed
3923 // in PhaseIdealLoop::build_and_optimize().
3924 //
3925 // After the Opaque1 node has been removed, optimizations (e.g., split-if,
3926 // loop unswitching, and IGVN, or a combination of them) can freely change
3927 // the graph's shape. As a result, the graph shape outlined below cannot
3928 // be guaranteed anymore.
3929 bool PhaseIdealLoop::is_canonical_loop_entry(CountedLoopNode* cl) {
3930   if (!cl->is_main_loop() && !cl->is_post_loop()) {
3931     return false;
3932   }
3933   Node* ctrl = cl->skip_predicates();
3934 
3935   if (ctrl == NULL || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) {
3936     return false;
3937   }
3938   Node* iffm = ctrl->in(0);
3939   if (iffm == NULL || !iffm->is_If()) {
3940     return false;
3941   }
3942   Node* bolzm = iffm->in(1);
3943   if (bolzm == NULL || !bolzm->is_Bool()) {
3944     return false;
3945   }
3946   Node* cmpzm = bolzm->in(1);
3947   if (cmpzm == NULL || !cmpzm->is_Cmp()) {
3948     return false;
3949   }
3950   // compares can get conditionally flipped
3951   bool found_opaque = false;
3952   for (uint i = 1; i < cmpzm->req(); i++) {
3953     Node* opnd = cmpzm->in(i);
3954     if (opnd && opnd->Opcode() == Op_Opaque1) {
3955       found_opaque = true;
3956       break;
3957     }
3958   }
3959   if (!found_opaque) {
3960     return false;
3961   }
3962   return true;
3963 }
3964 
3965 //------------------------------get_late_ctrl----------------------------------
3966 // Compute latest legal control.
3967 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
3968   assert(early != NULL, "early control should not be NULL");
3969 
3970   Node* LCA = compute_lca_of_uses(n, early);
3971 #ifdef ASSERT
3972   if (LCA == C->root() && LCA != early) {
3973     // def doesn't dominate uses so print some useful debugging output
3974     compute_lca_of_uses(n, early, true);
3975   }
3976 #endif
3977 
3978   // if this is a load, check for anti-dependent stores
3979   // We use a conservative algorithm to identify potential interfering
3980   // instructions and for rescheduling the load.  The users of the memory
3981   // input of this load are examined.  Any use which is not a load and is
3982   // dominated by early is considered a potentially interfering store.
3983   // This can produce false positives.
3984   if (n->is_Load() && LCA != early) {
3985     int load_alias_idx = C->get_alias_index(n->adr_type());
3986     if (C->alias_type(load_alias_idx)->is_rewritable()) {
3987 
3988       Node_List worklist;
3989 
3990       Node *mem = n->in(MemNode::Memory);
3991       for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
3992         Node* s = mem->fast_out(i);
3993         worklist.push(s);
3994       }
3995       while(worklist.size() != 0 && LCA != early) {
3996         Node* s = worklist.pop();
3997         if (s->is_Load() || s->Opcode() == Op_SafePoint ||
3998             (s->is_CallStaticJava() && s->as_CallStaticJava()->uncommon_trap_request() != 0)) {
3999           continue;
4000         } else if (s->is_MergeMem()) {
4001           for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
4002             Node* s1 = s->fast_out(i);
4003             worklist.push(s1);
4004           }
4005         } else {
4006           Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
4007           assert(sctrl != NULL || s->outcnt() == 0, "must have control");
4008           if (sctrl != NULL && !sctrl->is_top() && C->can_alias(s->adr_type(), load_alias_idx) && is_dominator(early, sctrl)) {
4009             LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
4010           }
4011         }
4012       }
4013     }
4014   }
4015 
4016   assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
4017   return LCA;
4018 }
4019 
4020 // true if CFG node d dominates CFG node n
4021 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
4022   if (d == n)
4023     return true;
4024   assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
4025   uint dd = dom_depth(d);
4026   while (dom_depth(n) >= dd) {
4027     if (n == d)
4028       return true;
4029     n = idom(n);
4030   }
4031   return false;
4032 }
4033 
4034 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
4035 // Pair-wise LCA with tags.
4036 // Tag each index with the node 'tag' currently being processed
4037 // before advancing up the dominator chain using idom().
4038 // Later calls that find a match to 'tag' know that this path has already
4039 // been considered in the current LCA (which is input 'n1' by convention).
4040 // Since get_late_ctrl() is only called once for each node, the tag array
4041 // does not need to be cleared between calls to get_late_ctrl().
4042 // Algorithm trades a larger constant factor for better asymptotic behavior
4043 //
4044 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
4045   uint d1 = dom_depth(n1);
4046   uint d2 = dom_depth(n2);
4047 
4048   do {
4049     if (d1 > d2) {
4050       // current lca is deeper than n2
4051       _dom_lca_tags.map(n1->_idx, tag);
4052       n1 =      idom(n1);
4053       d1 = dom_depth(n1);
4054     } else if (d1 < d2) {
4055       // n2 is deeper than current lca
4056       Node *memo = _dom_lca_tags[n2->_idx];
4057       if( memo == tag ) {
4058         return n1;    // Return the current LCA
4059       }
4060       _dom_lca_tags.map(n2->_idx, tag);
4061       n2 =      idom(n2);
4062       d2 = dom_depth(n2);
4063     } else {
4064       // Here d1 == d2.  Due to edits of the dominator-tree, sections
4065       // of the tree might have the same depth.  These sections have
4066       // to be searched more carefully.
4067 
4068       // Scan up all the n1's with equal depth, looking for n2.
4069       _dom_lca_tags.map(n1->_idx, tag);
4070       Node *t1 = idom(n1);
4071       while (dom_depth(t1) == d1) {
4072         if (t1 == n2)  return n2;
4073         _dom_lca_tags.map(t1->_idx, tag);
4074         t1 = idom(t1);
4075       }
4076       // Scan up all the n2's with equal depth, looking for n1.
4077       _dom_lca_tags.map(n2->_idx, tag);
4078       Node *t2 = idom(n2);
4079       while (dom_depth(t2) == d2) {
4080         if (t2 == n1)  return n1;
4081         _dom_lca_tags.map(t2->_idx, tag);
4082         t2 = idom(t2);
4083       }
4084       // Move up to a new dominator-depth value as well as up the dom-tree.
4085       n1 = t1;
4086       n2 = t2;
4087       d1 = dom_depth(n1);
4088       d2 = dom_depth(n2);
4089     }
4090   } while (n1 != n2);
4091   return n1;
4092 }
4093 
4094 //------------------------------init_dom_lca_tags------------------------------
4095 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
4096 // Intended use does not involve any growth for the array, so it could
4097 // be of fixed size.
4098 void PhaseIdealLoop::init_dom_lca_tags() {
4099   uint limit = C->unique() + 1;
4100   _dom_lca_tags.map( limit, NULL );
4101 #ifdef ASSERT
4102   for( uint i = 0; i < limit; ++i ) {
4103     assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
4104   }
4105 #endif // ASSERT
4106 }
4107 
4108 //------------------------------clear_dom_lca_tags------------------------------
4109 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
4110 // Intended use does not involve any growth for the array, so it could
4111 // be of fixed size.
4112 void PhaseIdealLoop::clear_dom_lca_tags() {
4113   uint limit = C->unique() + 1;
4114   _dom_lca_tags.map( limit, NULL );
4115   _dom_lca_tags.clear();
4116 #ifdef ASSERT
4117   for( uint i = 0; i < limit; ++i ) {
4118     assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
4119   }
4120 #endif // ASSERT
4121 }
4122 
4123 //------------------------------build_loop_late--------------------------------
4124 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
4125 // Second pass finds latest legal placement, and ideal loop placement.
4126 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
4127   while (worklist.size() != 0) {
4128     Node *n = worklist.pop();
4129     // Only visit once
4130     if (visited.test_set(n->_idx)) continue;
4131     uint cnt = n->outcnt();
4132     uint   i = 0;
4133     while (true) {
4134       assert( _nodes[n->_idx], "no dead nodes" );
4135       // Visit all children
4136       if (i < cnt) {
4137         Node* use = n->raw_out(i);
4138         ++i;
4139         // Check for dead uses.  Aggressively prune such junk.  It might be
4140         // dead in the global sense, but still have local uses so I cannot
4141         // easily call 'remove_dead_node'.
4142         if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
4143           // Due to cycles, we might not hit the same fixed point in the verify
4144           // pass as we do in the regular pass.  Instead, visit such phis as
4145           // simple uses of the loop head.
4146           if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
4147             if( !visited.test(use->_idx) )
4148               worklist.push(use);
4149           } else if( !visited.test_set(use->_idx) ) {
4150             nstack.push(n, i); // Save parent and next use's index.
4151             n   = use;         // Process all children of current use.
4152             cnt = use->outcnt();
4153             i   = 0;
4154           }
4155         } else {
4156           // Do not visit around the backedge of loops via data edges.
4157           // push dead code onto a worklist
4158           _deadlist.push(use);
4159         }
4160       } else {
4161         // All of n's children have been processed, complete post-processing.
4162         build_loop_late_post(n);
4163         if (nstack.is_empty()) {
4164           // Finished all nodes on stack.
4165           // Process next node on the worklist.
4166           break;
4167         }
4168         // Get saved parent node and next use's index. Visit the rest of uses.
4169         n   = nstack.node();
4170         cnt = n->outcnt();
4171         i   = nstack.index();
4172         nstack.pop();
4173       }
4174     }
4175   }
4176 }
4177 
4178 // Verify that no data node is schedules in the outer loop of a strip
4179 // mined loop.
4180 void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) {
4181 #ifdef ASSERT
4182   if (get_loop(least)->_nest == 0) {
4183     return;
4184   }
4185   IdealLoopTree* loop = get_loop(least);
4186   Node* head = loop->_head;
4187   if (head->is_OuterStripMinedLoop()
4188 #if INCLUDE_SHENANDOAHGC
4189       && // Verification can't be applied to fully built strip mined loops
4190       head->as_Loop()->outer_loop_end()->in(1)->find_int_con(-1) == 0
4191 #endif
4192       ) {
4193     Node* sfpt = head->as_Loop()->outer_safepoint();
4194     ResourceMark rm;
4195     Unique_Node_List wq;
4196     wq.push(sfpt);
4197     for (uint i = 0; i < wq.size(); i++) {
4198       Node *m = wq.at(i);
4199       for (uint i = 1; i < m->req(); i++) {
4200         Node* nn = m->in(i);
4201         if (nn == n) {
4202           return;
4203         }
4204         if (nn != NULL && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) {
4205           wq.push(nn);
4206         }
4207       }
4208     }
4209     ShouldNotReachHere();
4210   }
4211 #endif
4212 }
4213 
4214 
4215 //------------------------------build_loop_late_post---------------------------
4216 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
4217 // Second pass finds latest legal placement, and ideal loop placement.
4218 void PhaseIdealLoop::build_loop_late_post( Node *n ) {
4219 
4220   if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) {
4221     _igvn._worklist.push(n);  // Maybe we'll normalize it, if no more loops.
4222   }
4223 
4224 #ifdef ASSERT
4225   if (_verify_only && !n->is_CFG()) {
4226     // Check def-use domination.
4227     compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
4228   }
4229 #endif
4230 
4231   // CFG and pinned nodes already handled
4232   if( n->in(0) ) {
4233     if( n->in(0)->is_top() ) return; // Dead?
4234 
4235     // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
4236     // _must_ be pinned (they have to observe their control edge of course).
4237     // Unlike Stores (which modify an unallocable resource, the memory
4238     // state), Mods/Loads can float around.  So free them up.
4239     bool pinned = true;
4240     switch( n->Opcode() ) {
4241     case Op_DivI:
4242     case Op_DivF:
4243     case Op_DivD:
4244     case Op_ModI:
4245     case Op_ModF:
4246     case Op_ModD:
4247     case Op_LoadB:              // Same with Loads; they can sink
4248     case Op_LoadUB:             // during loop optimizations.
4249     case Op_LoadUS:
4250     case Op_LoadD:
4251     case Op_LoadF:
4252     case Op_LoadI:
4253     case Op_LoadKlass:
4254     case Op_LoadNKlass:
4255     case Op_LoadL:
4256     case Op_LoadS:
4257     case Op_LoadP:
4258     case Op_LoadBarrierSlowReg:
4259     case Op_LoadBarrierWeakSlowReg:
4260     case Op_LoadN:
4261     case Op_LoadRange:
4262     case Op_LoadD_unaligned:
4263     case Op_LoadL_unaligned:
4264     case Op_StrComp:            // Does a bunch of load-like effects
4265     case Op_StrEquals:
4266     case Op_StrIndexOf:
4267     case Op_StrIndexOfChar:
4268     case Op_AryEq:
4269     case Op_HasNegatives:
4270       pinned = false;
4271     }
4272 #if INCLUDE_SHENANDOAHGC
4273     if (UseShenandoahGC && n->is_CMove()) {
4274       pinned = false;
4275     }
4276 #endif
4277     if( pinned ) {
4278       IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
4279       if( !chosen_loop->_child )       // Inner loop?
4280         chosen_loop->_body.push(n); // Collect inner loops
4281       return;
4282     }
4283   } else {                      // No slot zero
4284     if( n->is_CFG() ) {         // CFG with no slot 0 is dead
4285       _nodes.map(n->_idx,0);    // No block setting, it's globally dead
4286       return;
4287     }
4288     assert(!n->is_CFG() || n->outcnt() == 0, "");
4289   }
4290 
4291   // Do I have a "safe range" I can select over?
4292   Node *early = get_ctrl(n);// Early location already computed
4293 
4294   // Compute latest point this Node can go
4295   Node *LCA = get_late_ctrl( n, early );
4296   // LCA is NULL due to uses being dead
4297   if( LCA == NULL ) {
4298 #ifdef ASSERT
4299     for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
4300       assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
4301     }
4302 #endif
4303     _nodes.map(n->_idx, 0);     // This node is useless
4304     _deadlist.push(n);
4305     return;
4306   }
4307   assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
4308 
4309   Node *legal = LCA;            // Walk 'legal' up the IDOM chain
4310   Node *least = legal;          // Best legal position so far
4311   while( early != legal ) {     // While not at earliest legal
4312 #ifdef ASSERT
4313     if (legal->is_Start() && !early->is_Root()) {
4314       // Bad graph. Print idom path and fail.
4315       dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
4316       assert(false, "Bad graph detected in build_loop_late");
4317     }
4318 #endif
4319     // Find least loop nesting depth
4320     legal = idom(legal);        // Bump up the IDOM tree
4321     // Check for lower nesting depth
4322     if( get_loop(legal)->_nest < get_loop(least)->_nest )
4323       least = legal;
4324   }
4325   assert(early == legal || legal != C->root(), "bad dominance of inputs");
4326 
4327   // Try not to place code on a loop entry projection
4328   // which can inhibit range check elimination.
4329   if (least != early) {
4330     Node* ctrl_out = least->unique_ctrl_out();
4331     if (ctrl_out && ctrl_out->is_Loop() &&
4332         least == ctrl_out->in(LoopNode::EntryControl)) {
4333       // Move the node above predicates as far up as possible so a
4334       // following pass of loop predication doesn't hoist a predicate
4335       // that depends on it above that node.
4336       Node* new_ctrl = least;
4337       for (;;) {
4338         if (!new_ctrl->is_Proj()) {
4339           break;
4340         }
4341         CallStaticJavaNode* call = new_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
4342         if (call == NULL) {
4343           break;
4344         }
4345         int req = call->uncommon_trap_request();
4346         Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req);
4347         if (trap_reason != Deoptimization::Reason_loop_limit_check &&
4348             trap_reason != Deoptimization::Reason_predicate &&
4349             trap_reason != Deoptimization::Reason_profile_predicate) {
4350           break;
4351         }
4352         Node* c = new_ctrl->in(0)->in(0);
4353         if (is_dominator(c, early) && c != early) {
4354           break;
4355         }
4356         new_ctrl = c;
4357       }
4358       least = new_ctrl;
4359     }
4360   }
4361 
4362 #ifdef ASSERT
4363   // If verifying, verify that 'verify_me' has a legal location
4364   // and choose it as our location.
4365   if( _verify_me ) {
4366     Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
4367     Node *legal = LCA;
4368     while( early != legal ) {   // While not at earliest legal
4369       if( legal == v_ctrl ) break;  // Check for prior good location
4370       legal = idom(legal)      ;// Bump up the IDOM tree
4371     }
4372     // Check for prior good location
4373     if( legal == v_ctrl ) least = legal; // Keep prior if found
4374   }
4375 #endif
4376 
4377   // Assign discovered "here or above" point
4378   least = find_non_split_ctrl(least);
4379   verify_strip_mined_scheduling(n, least);
4380   set_ctrl(n, least);
4381 
4382   // Collect inner loop bodies
4383   IdealLoopTree *chosen_loop = get_loop(least);
4384   if( !chosen_loop->_child )   // Inner loop?
4385     chosen_loop->_body.push(n);// Collect inner loops
4386 }
4387 
4388 #ifdef ASSERT
4389 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
4390   tty->print_cr("%s", msg);
4391   tty->print("n: "); n->dump();
4392   tty->print("early(n): "); early->dump();
4393   if (n->in(0) != NULL  && !n->in(0)->is_top() &&
4394       n->in(0) != early && !n->in(0)->is_Root()) {
4395     tty->print("n->in(0): "); n->in(0)->dump();
4396   }
4397   for (uint i = 1; i < n->req(); i++) {
4398     Node* in1 = n->in(i);
4399     if (in1 != NULL && in1 != n && !in1->is_top()) {
4400       tty->print("n->in(%d): ", i); in1->dump();
4401       Node* in1_early = get_ctrl(in1);
4402       tty->print("early(n->in(%d)): ", i); in1_early->dump();
4403       if (in1->in(0) != NULL     && !in1->in(0)->is_top() &&
4404           in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
4405         tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
4406       }
4407       for (uint j = 1; j < in1->req(); j++) {
4408         Node* in2 = in1->in(j);
4409         if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) {
4410           tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
4411           Node* in2_early = get_ctrl(in2);
4412           tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
4413           if (in2->in(0) != NULL     && !in2->in(0)->is_top() &&
4414               in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
4415             tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
4416           }
4417         }
4418       }
4419     }
4420   }
4421   tty->cr();
4422   tty->print("LCA(n): "); LCA->dump();
4423   for (uint i = 0; i < n->outcnt(); i++) {
4424     Node* u1 = n->raw_out(i);
4425     if (u1 == n)
4426       continue;
4427     tty->print("n->out(%d): ", i); u1->dump();
4428     if (u1->is_CFG()) {
4429       for (uint j = 0; j < u1->outcnt(); j++) {
4430         Node* u2 = u1->raw_out(j);
4431         if (u2 != u1 && u2 != n && u2->is_CFG()) {
4432           tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
4433         }
4434       }
4435     } else {
4436       Node* u1_later = get_ctrl(u1);
4437       tty->print("later(n->out(%d)): ", i); u1_later->dump();
4438       if (u1->in(0) != NULL     && !u1->in(0)->is_top() &&
4439           u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
4440         tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
4441       }
4442       for (uint j = 0; j < u1->outcnt(); j++) {
4443         Node* u2 = u1->raw_out(j);
4444         if (u2 == n || u2 == u1)
4445           continue;
4446         tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
4447         if (!u2->is_CFG()) {
4448           Node* u2_later = get_ctrl(u2);
4449           tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
4450           if (u2->in(0) != NULL     && !u2->in(0)->is_top() &&
4451               u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
4452             tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
4453           }
4454         }
4455       }
4456     }
4457   }
4458   tty->cr();
4459   int ct = 0;
4460   Node *dbg_legal = LCA;
4461   while(!dbg_legal->is_Start() && ct < 100) {
4462     tty->print("idom[%d] ",ct); dbg_legal->dump();
4463     ct++;
4464     dbg_legal = idom(dbg_legal);
4465   }
4466   tty->cr();
4467 }
4468 #endif
4469 
4470 #ifndef PRODUCT
4471 //------------------------------dump-------------------------------------------
4472 void PhaseIdealLoop::dump( ) const {
4473   ResourceMark rm;
4474   Arena* arena = Thread::current()->resource_area();
4475   Node_Stack stack(arena, C->live_nodes() >> 2);
4476   Node_List rpo_list;
4477   VectorSet visited(arena);
4478   visited.set(C->top()->_idx);
4479   rpo( C->root(), stack, visited, rpo_list );
4480   // Dump root loop indexed by last element in PO order
4481   dump( _ltree_root, rpo_list.size(), rpo_list );
4482 }
4483 
4484 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
4485   loop->dump_head();
4486 
4487   // Now scan for CFG nodes in the same loop
4488   for( uint j=idx; j > 0;  j-- ) {
4489     Node *n = rpo_list[j-1];
4490     if( !_nodes[n->_idx] )      // Skip dead nodes
4491       continue;
4492     if( get_loop(n) != loop ) { // Wrong loop nest
4493       if( get_loop(n)->_head == n &&    // Found nested loop?
4494           get_loop(n)->_parent == loop )
4495         dump(get_loop(n),rpo_list.size(),rpo_list);     // Print it nested-ly
4496       continue;
4497     }
4498 
4499     // Dump controlling node
4500     for( uint x = 0; x < loop->_nest; x++ )
4501       tty->print("  ");
4502     tty->print("C");
4503     if( n == C->root() ) {
4504       n->dump();
4505     } else {
4506       Node* cached_idom   = idom_no_update(n);
4507       Node *computed_idom = n->in(0);
4508       if( n->is_Region() ) {
4509         computed_idom = compute_idom(n);
4510         // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
4511         // any MultiBranch ctrl node), so apply a similar transform to
4512         // the cached idom returned from idom_no_update.
4513         cached_idom = find_non_split_ctrl(cached_idom);
4514       }
4515       tty->print(" ID:%d",computed_idom->_idx);
4516       n->dump();
4517       if( cached_idom != computed_idom ) {
4518         tty->print_cr("*** BROKEN IDOM!  Computed as: %d, cached as: %d",
4519                       computed_idom->_idx, cached_idom->_idx);
4520       }
4521     }
4522     // Dump nodes it controls
4523     for( uint k = 0; k < _nodes.Size(); k++ ) {
4524       // (k < C->unique() && get_ctrl(find(k)) == n)
4525       if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
4526         Node *m = C->root()->find(k);
4527         if( m && m->outcnt() > 0 ) {
4528           if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
4529             tty->print_cr("*** BROKEN CTRL ACCESSOR!  _nodes[k] is %p, ctrl is %p",
4530                           _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
4531           }
4532           for( uint j = 0; j < loop->_nest; j++ )
4533             tty->print("  ");
4534           tty->print(" ");
4535           m->dump();
4536         }
4537       }
4538     }
4539   }
4540 }
4541 #endif
4542 
4543 #if !defined(PRODUCT) || INCLUDE_SHENANDOAHGC
4544 // Collect a R-P-O for the whole CFG.
4545 // Result list is in post-order (scan backwards for RPO)
4546 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
4547   stk.push(start, 0);
4548   visited.set(start->_idx);
4549 
4550   while (stk.is_nonempty()) {
4551     Node* m   = stk.node();
4552     uint  idx = stk.index();
4553     if (idx < m->outcnt()) {
4554       stk.set_index(idx + 1);
4555       Node* n = m->raw_out(idx);
4556       if (n->is_CFG() && !visited.test_set(n->_idx)) {
4557         stk.push(n, 0);
4558       }
4559     } else {
4560       rpo_list.push(m);
4561       stk.pop();
4562     }
4563   }
4564 }
4565 #endif
4566 
4567 
4568 //=============================================================================
4569 //------------------------------LoopTreeIterator-----------------------------------
4570 
4571 // Advance to next loop tree using a preorder, left-to-right traversal.
4572 void LoopTreeIterator::next() {
4573   assert(!done(), "must not be done.");
4574   if (_curnt->_child != NULL) {
4575     _curnt = _curnt->_child;
4576   } else if (_curnt->_next != NULL) {
4577     _curnt = _curnt->_next;
4578   } else {
4579     while (_curnt != _root && _curnt->_next == NULL) {
4580       _curnt = _curnt->_parent;
4581     }
4582     if (_curnt == _root) {
4583       _curnt = NULL;
4584       assert(done(), "must be done.");
4585     } else {
4586       assert(_curnt->_next != NULL, "must be more to do");
4587       _curnt = _curnt->_next;
4588     }
4589   }
4590 }