1 /*
   2  * Copyright (c) 1998, 2019, 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 
  44 //=============================================================================
  45 //--------------------------is_cloop_ind_var-----------------------------------
  46 // Determine if a node is a counted loop induction variable.
  47 // NOTE: The method is declared in "node.hpp".
  48 bool Node::is_cloop_ind_var() const {
  49   return (is_Phi() && !as_Phi()->is_copy() &&
  50           as_Phi()->region()->is_CountedLoop() &&
  51           as_Phi()->region()->as_CountedLoop()->phi() == this);
  52 }
  53 
  54 //=============================================================================
  55 //------------------------------dump_spec--------------------------------------
  56 // Dump special per-node info
  57 #ifndef PRODUCT
  58 void LoopNode::dump_spec(outputStream *st) const {
  59   if (is_inner_loop()) st->print( "inner " );
  60   if (is_partial_peel_loop()) st->print( "partial_peel " );
  61   if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
  62 }
  63 #endif
  64 
  65 //------------------------------is_valid_counted_loop-------------------------
  66 bool LoopNode::is_valid_counted_loop() const {
  67   if (is_CountedLoop()) {
  68     CountedLoopNode*    l  = as_CountedLoop();
  69     CountedLoopEndNode* le = l->loopexit_or_null();
  70     if (le != NULL &&
  71         le->proj_out_or_null(1 /* true */) == l->in(LoopNode::LoopBackControl)) {
  72       Node* phi  = l->phi();
  73       Node* exit = le->proj_out_or_null(0 /* false */);
  74       if (exit != NULL && exit->Opcode() == Op_IfFalse &&
  75           phi != NULL && phi->is_Phi() &&
  76           phi->in(LoopNode::LoopBackControl) == l->incr() &&
  77           le->loopnode() == l && le->stride_is_con()) {
  78         return true;
  79       }
  80     }
  81   }
  82   return false;
  83 }
  84 
  85 //------------------------------get_early_ctrl---------------------------------
  86 // Compute earliest legal control
  87 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
  88   assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
  89   uint i;
  90   Node *early;
  91   if (n->in(0) && !n->is_expensive()) {
  92     early = n->in(0);
  93     if (!early->is_CFG()) // Might be a non-CFG multi-def
  94       early = get_ctrl(early);        // So treat input as a straight data input
  95     i = 1;
  96   } else {
  97     early = get_ctrl(n->in(1));
  98     i = 2;
  99   }
 100   uint e_d = dom_depth(early);
 101   assert( early, "" );
 102   for (; i < n->req(); i++) {
 103     Node *cin = get_ctrl(n->in(i));
 104     assert( cin, "" );
 105     // Keep deepest dominator depth
 106     uint c_d = dom_depth(cin);
 107     if (c_d > e_d) {           // Deeper guy?
 108       early = cin;              // Keep deepest found so far
 109       e_d = c_d;
 110     } else if (c_d == e_d &&    // Same depth?
 111                early != cin) { // If not equal, must use slower algorithm
 112       // If same depth but not equal, one _must_ dominate the other
 113       // and we want the deeper (i.e., dominated) guy.
 114       Node *n1 = early;
 115       Node *n2 = cin;
 116       while (1) {
 117         n1 = idom(n1);          // Walk up until break cycle
 118         n2 = idom(n2);
 119         if (n1 == cin ||        // Walked early up to cin
 120             dom_depth(n2) < c_d)
 121           break;                // early is deeper; keep him
 122         if (n2 == early ||      // Walked cin up to early
 123             dom_depth(n1) < c_d) {
 124           early = cin;          // cin is deeper; keep him
 125           break;
 126         }
 127       }
 128       e_d = dom_depth(early);   // Reset depth register cache
 129     }
 130   }
 131 
 132   // Return earliest legal location
 133   assert(early == find_non_split_ctrl(early), "unexpected early control");
 134 
 135   if (n->is_expensive() && !_verify_only && !_verify_me) {
 136     assert(n->in(0), "should have control input");
 137     early = get_early_ctrl_for_expensive(n, early);
 138   }
 139 
 140   return early;
 141 }
 142 
 143 //------------------------------get_early_ctrl_for_expensive---------------------------------
 144 // Move node up the dominator tree as high as legal while still beneficial
 145 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {
 146   assert(n->in(0) && n->is_expensive(), "expensive node with control input here");
 147   assert(OptimizeExpensiveOps, "optimization off?");
 148 
 149   Node* ctl = n->in(0);
 150   assert(ctl->is_CFG(), "expensive input 0 must be cfg");
 151   uint min_dom_depth = dom_depth(earliest);
 152 #ifdef ASSERT
 153   if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) {
 154     dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl);
 155     assert(false, "Bad graph detected in get_early_ctrl_for_expensive");
 156   }
 157 #endif
 158   if (dom_depth(ctl) < min_dom_depth) {
 159     return earliest;
 160   }
 161 
 162   while (1) {
 163     Node *next = ctl;
 164     // Moving the node out of a loop on the projection of a If
 165     // confuses loop predication. So once we hit a Loop in a If branch
 166     // that doesn't branch to an UNC, we stop. The code that process
 167     // expensive nodes will notice the loop and skip over it to try to
 168     // move the node further up.
 169     if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) {
 170       if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
 171         break;
 172       }
 173       next = idom(ctl->in(1)->in(0));
 174     } else if (ctl->is_Proj()) {
 175       // We only move it up along a projection if the projection is
 176       // the single control projection for its parent: same code path,
 177       // if it's a If with UNC or fallthrough of a call.
 178       Node* parent_ctl = ctl->in(0);
 179       if (parent_ctl == NULL) {
 180         break;
 181       } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) {
 182         next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control();
 183       } else if (parent_ctl->is_If()) {
 184         if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
 185           break;
 186         }
 187         assert(idom(ctl) == parent_ctl, "strange");
 188         next = idom(parent_ctl);
 189       } else if (ctl->is_CatchProj()) {
 190         if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) {
 191           break;
 192         }
 193         assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph");
 194         next = parent_ctl->in(0)->in(0)->in(0);
 195       } else {
 196         // Check if parent control has a single projection (this
 197         // control is the only possible successor of the parent
 198         // control). If so, we can try to move the node above the
 199         // parent control.
 200         int nb_ctl_proj = 0;
 201         for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) {
 202           Node *p = parent_ctl->fast_out(i);
 203           if (p->is_Proj() && p->is_CFG()) {
 204             nb_ctl_proj++;
 205             if (nb_ctl_proj > 1) {
 206               break;
 207             }
 208           }
 209         }
 210 
 211         if (nb_ctl_proj > 1) {
 212           break;
 213         }
 214         assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call() ||
 215                BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(parent_ctl), "unexpected node");
 216         assert(idom(ctl) == parent_ctl, "strange");
 217         next = idom(parent_ctl);
 218       }
 219     } else {
 220       next = idom(ctl);
 221     }
 222     if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) {
 223       break;
 224     }
 225     ctl = next;
 226   }
 227 
 228   if (ctl != n->in(0)) {
 229     _igvn.replace_input_of(n, 0, ctl);
 230     _igvn.hash_insert(n);
 231   }
 232 
 233   return ctl;
 234 }
 235 
 236 
 237 //------------------------------set_early_ctrl---------------------------------
 238 // Set earliest legal control
 239 void PhaseIdealLoop::set_early_ctrl( Node *n ) {
 240   Node *early = get_early_ctrl(n);
 241 
 242   // Record earliest legal location
 243   set_ctrl(n, early);
 244 }
 245 
 246 //------------------------------set_subtree_ctrl-------------------------------
 247 // set missing _ctrl entries on new nodes
 248 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {
 249   // Already set?  Get out.
 250   if( _nodes[n->_idx] ) return;
 251   // Recursively set _nodes array to indicate where the Node goes
 252   uint i;
 253   for( i = 0; i < n->req(); ++i ) {
 254     Node *m = n->in(i);
 255     if( m && m != C->root() )
 256       set_subtree_ctrl( m );
 257   }
 258 
 259   // Fixup self
 260   set_early_ctrl( n );
 261 }
 262 
 263 // Create a skeleton strip mined outer loop: a Loop head before the
 264 // inner strip mined loop, a safepoint and an exit condition guarded
 265 // by an opaque node after the inner strip mined loop with a backedge
 266 // to the loop head. The inner strip mined loop is left as it is. Only
 267 // once loop optimizations are over, do we adjust the inner loop exit
 268 // condition to limit its number of iterations, set the outer loop
 269 // exit condition and add Phis to the outer loop head. Some loop
 270 // optimizations that operate on the inner strip mined loop need to be
 271 // aware of the outer strip mined loop: loop unswitching needs to
 272 // clone the outer loop as well as the inner, unrolling needs to only
 273 // clone the inner loop etc. No optimizations need to change the outer
 274 // strip mined loop as it is only a skeleton.
 275 IdealLoopTree* PhaseIdealLoop::create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
 276                                                              IdealLoopTree* loop, float cl_prob, float le_fcnt,
 277                                                              Node*& entry_control, Node*& iffalse) {
 278   Node* outer_test = _igvn.intcon(0);
 279   set_ctrl(outer_test, C->root());
 280   Node *orig = iffalse;
 281   iffalse = iffalse->clone();
 282   _igvn.register_new_node_with_optimizer(iffalse);
 283   set_idom(iffalse, idom(orig), dom_depth(orig));
 284 
 285   IfNode *outer_le = new OuterStripMinedLoopEndNode(iffalse, outer_test, cl_prob, le_fcnt);
 286   Node *outer_ift = new IfTrueNode (outer_le);
 287   Node* outer_iff = orig;
 288   _igvn.replace_input_of(outer_iff, 0, outer_le);
 289 
 290   LoopNode *outer_l = new OuterStripMinedLoopNode(C, init_control, outer_ift);
 291   entry_control = outer_l;
 292 
 293   IdealLoopTree* outer_ilt = new IdealLoopTree(this, outer_l, outer_ift);
 294   IdealLoopTree* parent = loop->_parent;
 295   IdealLoopTree* sibling = parent->_child;
 296   if (sibling == loop) {
 297     parent->_child = outer_ilt;
 298   } else {
 299     while (sibling->_next != loop) {
 300       sibling = sibling->_next;
 301     }
 302     sibling->_next = outer_ilt;
 303   }
 304   outer_ilt->_next = loop->_next;
 305   outer_ilt->_parent = parent;
 306   outer_ilt->_child = loop;
 307   outer_ilt->_nest = loop->_nest;
 308   loop->_parent = outer_ilt;
 309   loop->_next = NULL;
 310   loop->_nest++;
 311 
 312   set_loop(iffalse, outer_ilt);
 313   register_control(outer_le, outer_ilt, iffalse);
 314   register_control(outer_ift, outer_ilt, outer_le);
 315   set_idom(outer_iff, outer_le, dom_depth(outer_le));
 316   _igvn.register_new_node_with_optimizer(outer_l);
 317   set_loop(outer_l, outer_ilt);
 318   set_idom(outer_l, init_control, dom_depth(init_control)+1);
 319 
 320   return outer_ilt;
 321 }
 322 
 323 void PhaseIdealLoop::insert_loop_limit_check(ProjNode* limit_check_proj, Node* cmp_limit, Node* bol) {
 324   Node* new_predicate_proj = create_new_if_for_predicate(limit_check_proj, NULL,
 325                                                          Deoptimization::Reason_loop_limit_check,
 326                                                          Op_If);
 327   Node* iff = new_predicate_proj->in(0);
 328   assert(iff->Opcode() == Op_If, "bad graph shape");
 329   Node* conv = iff->in(1);
 330   assert(conv->Opcode() == Op_Conv2B, "bad graph shape");
 331   Node* opaq = conv->in(1);
 332   assert(opaq->Opcode() == Op_Opaque1, "bad graph shape");
 333   cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
 334   bol = _igvn.register_new_node_with_optimizer(bol);
 335   set_subtree_ctrl(bol);
 336   _igvn.replace_input_of(iff, 1, bol);
 337 
 338 #ifndef PRODUCT
 339   // report that the loop predication has been actually performed
 340   // for this loop
 341   if (TraceLoopLimitCheck) {
 342     tty->print_cr("Counted Loop Limit Check generated:");
 343     debug_only( bol->dump(2); )
 344   }
 345 #endif
 346 }
 347 
 348 //------------------------------is_counted_loop--------------------------------
 349 bool PhaseIdealLoop::is_counted_loop(Node* x, IdealLoopTree*& loop) {
 350   PhaseGVN *gvn = &_igvn;
 351 
 352   // Counted loop head must be a good RegionNode with only 3 not NULL
 353   // control input edges: Self, Entry, LoopBack.
 354   if (x->in(LoopNode::Self) == NULL || x->req() != 3 || loop->_irreducible) {
 355     return false;
 356   }
 357   Node *init_control = x->in(LoopNode::EntryControl);
 358   Node *back_control = x->in(LoopNode::LoopBackControl);
 359   if (init_control == NULL || back_control == NULL)    // Partially dead
 360     return false;
 361   // Must also check for TOP when looking for a dead loop
 362   if (init_control->is_top() || back_control->is_top())
 363     return false;
 364 
 365   // Allow funny placement of Safepoint
 366   if (back_control->Opcode() == Op_SafePoint) {
 367     if (LoopStripMiningIter != 0) {
 368       // Leaving the safepoint on the backedge and creating a
 369       // CountedLoop will confuse optimizations. We can't move the
 370       // safepoint around because its jvm state wouldn't match a new
 371       // location. Give up on that loop.
 372       return false;
 373     }
 374     back_control = back_control->in(TypeFunc::Control);
 375   }
 376 
 377   // Controlling test for loop
 378   Node *iftrue = back_control;
 379   uint iftrue_op = iftrue->Opcode();
 380   if (iftrue_op != Op_IfTrue &&
 381       iftrue_op != Op_IfFalse)
 382     // I have a weird back-control.  Probably the loop-exit test is in
 383     // the middle of the loop and I am looking at some trailing control-flow
 384     // merge point.  To fix this I would have to partially peel the loop.
 385     return false; // Obscure back-control
 386 
 387   // Get boolean guarding loop-back test
 388   Node *iff = iftrue->in(0);
 389   if (get_loop(iff) != loop || !iff->in(1)->is_Bool())
 390     return false;
 391   BoolNode *test = iff->in(1)->as_Bool();
 392   BoolTest::mask bt = test->_test._test;
 393   float cl_prob = iff->as_If()->_prob;
 394   if (iftrue_op == Op_IfFalse) {
 395     bt = BoolTest(bt).negate();
 396     cl_prob = 1.0 - cl_prob;
 397   }
 398   // Get backedge compare
 399   Node *cmp = test->in(1);
 400   int cmp_op = cmp->Opcode();
 401   if (cmp_op != Op_CmpI)
 402     return false;                // Avoid pointer & float compares
 403 
 404   // Find the trip-counter increment & limit.  Limit must be loop invariant.
 405   Node *incr  = cmp->in(1);
 406   Node *limit = cmp->in(2);
 407 
 408   // ---------
 409   // need 'loop()' test to tell if limit is loop invariant
 410   // ---------
 411 
 412   if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
 413     Node *tmp = incr;            // Then reverse order into the CmpI
 414     incr = limit;
 415     limit = tmp;
 416     bt = BoolTest(bt).commute(); // And commute the exit test
 417   }
 418   if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant
 419     return false;
 420   if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
 421     return false;
 422 
 423   Node* phi_incr = NULL;
 424   // Trip-counter increment must be commutative & associative.
 425   if (incr->Opcode() == Op_CastII) {
 426     incr = incr->in(1);
 427   }
 428   if (incr->is_Phi()) {
 429     if (incr->as_Phi()->region() != x || incr->req() != 3)
 430       return false; // Not simple trip counter expression
 431     phi_incr = incr;
 432     incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
 433     if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
 434       return false;
 435   }
 436 
 437   Node* trunc1 = NULL;
 438   Node* trunc2 = NULL;
 439   const TypeInt* iv_trunc_t = NULL;
 440   Node* orig_incr = incr;
 441   if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {
 442     return false; // Funny increment opcode
 443   }
 444   assert(incr->Opcode() == Op_AddI, "wrong increment code");
 445 
 446   const TypeInt* limit_t = gvn->type(limit)->is_int();
 447   if (trunc1 != NULL) {
 448     // When there is a truncation, we must be sure that after the truncation
 449     // the trip counter will end up higher than the limit, otherwise we are looking
 450     // at an endless loop. Can happen with range checks.
 451 
 452     // Example:
 453     // int i = 0;
 454     // while (true)
 455     //    sum + = array[i];
 456     //    i++;
 457     //    i = i && 0x7fff;
 458     //  }
 459     //
 460     // If the array is shorter than 0x8000 this exits through a AIOOB
 461     //  - Counted loop transformation is ok
 462     // If the array is longer then this is an endless loop
 463     //  - No transformation can be done.
 464 
 465     const TypeInt* incr_t = gvn->type(orig_incr)->is_int();
 466     if (limit_t->_hi > incr_t->_hi) {
 467       // if the limit can have a higher value than the increment (before the phi)
 468       return false;
 469     }
 470   }
 471 
 472   // Get merge point
 473   Node *xphi = incr->in(1);
 474   Node *stride = incr->in(2);
 475   if (!stride->is_Con()) {     // Oops, swap these
 476     if (!xphi->is_Con())       // Is the other guy a constant?
 477       return false;             // Nope, unknown stride, bail out
 478     Node *tmp = xphi;           // 'incr' is commutative, so ok to swap
 479     xphi = stride;
 480     stride = tmp;
 481   }
 482   if (xphi->Opcode() == Op_CastII) {
 483     xphi = xphi->in(1);
 484   }
 485   // Stride must be constant
 486   int stride_con = stride->get_int();
 487   if (stride_con == 0)
 488     return false; // missed some peephole opt
 489 
 490   if (!xphi->is_Phi())
 491     return false; // Too much math on the trip counter
 492   if (phi_incr != NULL && phi_incr != xphi)
 493     return false;
 494   PhiNode *phi = xphi->as_Phi();
 495 
 496   // Phi must be of loop header; backedge must wrap to increment
 497   if (phi->region() != x)
 498     return false;
 499   if ((trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr) ||
 500       (trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1)) {
 501     return false;
 502   }
 503   Node *init_trip = phi->in(LoopNode::EntryControl);
 504 
 505   // If iv trunc type is smaller than int, check for possible wrap.
 506   if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
 507     assert(trunc1 != NULL, "must have found some truncation");
 508 
 509     // Get a better type for the phi (filtered thru if's)
 510     const TypeInt* phi_ft = filtered_type(phi);
 511 
 512     // Can iv take on a value that will wrap?
 513     //
 514     // Ensure iv's limit is not within "stride" of the wrap value.
 515     //
 516     // Example for "short" type
 517     //    Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
 518     //    If the stride is +10, then the last value of the induction
 519     //    variable before the increment (phi_ft->_hi) must be
 520     //    <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
 521     //    ensure no truncation occurs after the increment.
 522 
 523     if (stride_con > 0) {
 524       if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
 525           iv_trunc_t->_lo > phi_ft->_lo) {
 526         return false;  // truncation may occur
 527       }
 528     } else if (stride_con < 0) {
 529       if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
 530           iv_trunc_t->_hi < phi_ft->_hi) {
 531         return false;  // truncation may occur
 532       }
 533     }
 534     // No possibility of wrap so truncation can be discarded
 535     // Promote iv type to Int
 536   } else {
 537     assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
 538   }
 539 
 540   // If the condition is inverted and we will be rolling
 541   // through MININT to MAXINT, then bail out.
 542   if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
 543       // Odd stride
 544       (bt == BoolTest::ne && stride_con != 1 && stride_con != -1) ||
 545       // Count down loop rolls through MAXINT
 546       ((bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0) ||
 547       // Count up loop rolls through MININT
 548       ((bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0)) {
 549     return false; // Bail out
 550   }
 551 
 552   const TypeInt* init_t = gvn->type(init_trip)->is_int();
 553 
 554   if (stride_con > 0) {
 555     jlong init_p = (jlong)init_t->_lo + stride_con;
 556     if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi)
 557       return false; // cyclic loop or this loop trips only once
 558   } else {
 559     jlong init_p = (jlong)init_t->_hi + stride_con;
 560     if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo)
 561       return false; // cyclic loop or this loop trips only once
 562   }
 563 
 564   if (phi_incr != NULL && bt != BoolTest::ne) {
 565     // check if there is a possiblity of IV overflowing after the first increment
 566     if (stride_con > 0) {
 567       if (init_t->_hi > max_jint - stride_con) {
 568         return false;
 569       }
 570     } else {
 571       if (init_t->_lo < min_jint - stride_con) {
 572         return false;
 573       }
 574     }
 575   }
 576 
 577   // =================================================
 578   // ---- SUCCESS!   Found A Trip-Counted Loop!  -----
 579   //
 580   assert(x->Opcode() == Op_Loop, "regular loops only");
 581   C->print_method(PHASE_BEFORE_CLOOPS, 3);
 582 
 583   Node *hook = new Node(6);
 584 
 585   // ===================================================
 586   // Generate loop limit check to avoid integer overflow
 587   // in cases like next (cyclic loops):
 588   //
 589   // for (i=0; i <= max_jint; i++) {}
 590   // for (i=0; i <  max_jint; i+=2) {}
 591   //
 592   //
 593   // Limit check predicate depends on the loop test:
 594   //
 595   // for(;i != limit; i++)       --> limit <= (max_jint)
 596   // for(;i <  limit; i+=stride) --> limit <= (max_jint - stride + 1)
 597   // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride    )
 598   //
 599 
 600   // Check if limit is excluded to do more precise int overflow check.
 601   bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge);
 602   int stride_m  = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1));
 603 
 604   // If compare points directly to the phi we need to adjust
 605   // the compare so that it points to the incr. Limit have
 606   // to be adjusted to keep trip count the same and the
 607   // adjusted limit should be checked for int overflow.
 608   if (phi_incr != NULL) {
 609     stride_m  += stride_con;
 610   }
 611 
 612   if (limit->is_Con()) {
 613     int limit_con = limit->get_int();
 614     if ((stride_con > 0 && limit_con > (max_jint - stride_m)) ||
 615         (stride_con < 0 && limit_con < (min_jint - stride_m))) {
 616       // Bailout: it could be integer overflow.
 617       return false;
 618     }
 619   } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) ||
 620              (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) {
 621       // Limit's type may satisfy the condition, for example,
 622       // when it is an array length.
 623   } else {
 624     // Generate loop's limit check.
 625     // Loop limit check predicate should be near the loop.
 626     ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
 627     if (!limit_check_proj) {
 628       // The limit check predicate is not generated if this method trapped here before.
 629 #ifdef ASSERT
 630       if (TraceLoopLimitCheck) {
 631         tty->print("missing loop limit check:");
 632         loop->dump_head();
 633         x->dump(1);
 634       }
 635 #endif
 636       return false;
 637     }
 638 
 639     IfNode* check_iff = limit_check_proj->in(0)->as_If();
 640 
 641     if (!is_dominator(get_ctrl(limit), check_iff->in(0))) {
 642       return false;
 643     }
 644 
 645     Node* cmp_limit;
 646     Node* bol;
 647 
 648     if (stride_con > 0) {
 649       cmp_limit = new CmpINode(limit, _igvn.intcon(max_jint - stride_m));
 650       bol = new BoolNode(cmp_limit, BoolTest::le);
 651     } else {
 652       cmp_limit = new CmpINode(limit, _igvn.intcon(min_jint - stride_m));
 653       bol = new BoolNode(cmp_limit, BoolTest::ge);
 654     }
 655 
 656     insert_loop_limit_check(limit_check_proj, cmp_limit, bol);
 657   }
 658 
 659   // Now we need to canonicalize loop condition.
 660   if (bt == BoolTest::ne) {
 661     assert(stride_con == 1 || stride_con == -1, "simple increment only");
 662     if (stride_con > 0 && init_t->_hi < limit_t->_lo) {
 663       // 'ne' can be replaced with 'lt' only when init < limit.
 664       bt = BoolTest::lt;
 665     } else if (stride_con < 0 && init_t->_lo > limit_t->_hi) {
 666       // 'ne' can be replaced with 'gt' only when init > limit.
 667       bt = BoolTest::gt;
 668     } else {
 669       ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
 670       if (!limit_check_proj) {
 671         // The limit check predicate is not generated if this method trapped here before.
 672 #ifdef ASSERT
 673         if (TraceLoopLimitCheck) {
 674           tty->print("missing loop limit check:");
 675           loop->dump_head();
 676           x->dump(1);
 677         }
 678 #endif
 679         return false;
 680       }
 681       IfNode* check_iff = limit_check_proj->in(0)->as_If();
 682 
 683       if (!is_dominator(get_ctrl(limit), check_iff->in(0)) ||
 684           !is_dominator(get_ctrl(init_trip), check_iff->in(0))) {
 685         return false;
 686       }
 687 
 688       Node* cmp_limit;
 689       Node* bol;
 690 
 691       if (stride_con > 0) {
 692         cmp_limit = new CmpINode(init_trip, limit);
 693         bol = new BoolNode(cmp_limit, BoolTest::lt);
 694       } else {
 695         cmp_limit = new CmpINode(init_trip, limit);
 696         bol = new BoolNode(cmp_limit, BoolTest::gt);
 697       }
 698 
 699       insert_loop_limit_check(limit_check_proj, cmp_limit, bol);
 700 
 701       if (stride_con > 0) {
 702         // 'ne' can be replaced with 'lt' only when init < limit.
 703         bt = BoolTest::lt;
 704       } else if (stride_con < 0) {
 705         // 'ne' can be replaced with 'gt' only when init > limit.
 706         bt = BoolTest::gt;
 707       }
 708     }
 709   }
 710 
 711   if (phi_incr != NULL) {
 712     // If compare points directly to the phi we need to adjust
 713     // the compare so that it points to the incr. Limit have
 714     // to be adjusted to keep trip count the same and we
 715     // should avoid int overflow.
 716     //
 717     //   i = init; do {} while(i++ < limit);
 718     // is converted to
 719     //   i = init; do {} while(++i < limit+1);
 720     //
 721     limit = gvn->transform(new AddINode(limit, stride));
 722   }
 723 
 724   if (incl_limit) {
 725     // The limit check guaranties that 'limit <= (max_jint - stride)' so
 726     // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
 727     //
 728     Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1);
 729     limit = gvn->transform(new AddINode(limit, one));
 730     if (bt == BoolTest::le)
 731       bt = BoolTest::lt;
 732     else if (bt == BoolTest::ge)
 733       bt = BoolTest::gt;
 734     else
 735       ShouldNotReachHere();
 736   }
 737   set_subtree_ctrl( limit );
 738 
 739   if (LoopStripMiningIter == 0) {
 740     // Check for SafePoint on backedge and remove
 741     Node *sfpt = x->in(LoopNode::LoopBackControl);
 742     if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
 743       lazy_replace( sfpt, iftrue );
 744       if (loop->_safepts != NULL) {
 745         loop->_safepts->yank(sfpt);
 746       }
 747       loop->_tail = iftrue;
 748     }
 749   }
 750 
 751   // Build a canonical trip test.
 752   // Clone code, as old values may be in use.
 753   incr = incr->clone();
 754   incr->set_req(1,phi);
 755   incr->set_req(2,stride);
 756   incr = _igvn.register_new_node_with_optimizer(incr);
 757   set_early_ctrl( incr );
 758   _igvn.rehash_node_delayed(phi);
 759   phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
 760 
 761   // If phi type is more restrictive than Int, raise to
 762   // Int to prevent (almost) infinite recursion in igvn
 763   // which can only handle integer types for constants or minint..maxint.
 764   if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
 765     Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
 766     nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
 767     nphi = _igvn.register_new_node_with_optimizer(nphi);
 768     set_ctrl(nphi, get_ctrl(phi));
 769     _igvn.replace_node(phi, nphi);
 770     phi = nphi->as_Phi();
 771   }
 772   cmp = cmp->clone();
 773   cmp->set_req(1,incr);
 774   cmp->set_req(2,limit);
 775   cmp = _igvn.register_new_node_with_optimizer(cmp);
 776   set_ctrl(cmp, iff->in(0));
 777 
 778   test = test->clone()->as_Bool();
 779   (*(BoolTest*)&test->_test)._test = bt;
 780   test->set_req(1,cmp);
 781   _igvn.register_new_node_with_optimizer(test);
 782   set_ctrl(test, iff->in(0));
 783 
 784   // Replace the old IfNode with a new LoopEndNode
 785   Node *lex = _igvn.register_new_node_with_optimizer(new CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
 786   IfNode *le = lex->as_If();
 787   uint dd = dom_depth(iff);
 788   set_idom(le, le->in(0), dd); // Update dominance for loop exit
 789   set_loop(le, loop);
 790 
 791   // Get the loop-exit control
 792   Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
 793 
 794   // Need to swap loop-exit and loop-back control?
 795   if (iftrue_op == Op_IfFalse) {
 796     Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le));
 797     Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le));
 798 
 799     loop->_tail = back_control = ift2;
 800     set_loop(ift2, loop);
 801     set_loop(iff2, get_loop(iffalse));
 802 
 803     // Lazy update of 'get_ctrl' mechanism.
 804     lazy_replace(iffalse, iff2);
 805     lazy_replace(iftrue,  ift2);
 806 
 807     // Swap names
 808     iffalse = iff2;
 809     iftrue  = ift2;
 810   } else {
 811     _igvn.rehash_node_delayed(iffalse);
 812     _igvn.rehash_node_delayed(iftrue);
 813     iffalse->set_req_X( 0, le, &_igvn );
 814     iftrue ->set_req_X( 0, le, &_igvn );
 815   }
 816 
 817   set_idom(iftrue,  le, dd+1);
 818   set_idom(iffalse, le, dd+1);
 819   assert(iff->outcnt() == 0, "should be dead now");
 820   lazy_replace( iff, le ); // fix 'get_ctrl'
 821 
 822   Node *sfpt2 = le->in(0);
 823 
 824   Node* entry_control = init_control;
 825   bool strip_mine_loop = LoopStripMiningIter > 1 && loop->_child == NULL &&
 826     sfpt2->Opcode() == Op_SafePoint && !loop->_has_call;
 827   IdealLoopTree* outer_ilt = NULL;
 828   if (strip_mine_loop) {
 829     outer_ilt = create_outer_strip_mined_loop(test, cmp, init_control, loop,
 830                                               cl_prob, le->_fcnt, entry_control,
 831                                               iffalse);
 832   }
 833 
 834   // Now setup a new CountedLoopNode to replace the existing LoopNode
 835   CountedLoopNode *l = new CountedLoopNode(entry_control, back_control);
 836   l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
 837   // The following assert is approximately true, and defines the intention
 838   // of can_be_counted_loop.  It fails, however, because phase->type
 839   // is not yet initialized for this loop and its parts.
 840   //assert(l->can_be_counted_loop(this), "sanity");
 841   _igvn.register_new_node_with_optimizer(l);
 842   set_loop(l, loop);
 843   loop->_head = l;
 844   // Fix all data nodes placed at the old loop head.
 845   // Uses the lazy-update mechanism of 'get_ctrl'.
 846   lazy_replace( x, l );
 847   set_idom(l, entry_control, dom_depth(entry_control) + 1);
 848 
 849   if (LoopStripMiningIter == 0 || strip_mine_loop) {
 850     // Check for immediately preceding SafePoint and remove
 851     if (sfpt2->Opcode() == Op_SafePoint && (LoopStripMiningIter != 0 || is_deleteable_safept(sfpt2))) {
 852       if (strip_mine_loop) {
 853         Node* outer_le = outer_ilt->_tail->in(0);
 854         Node* sfpt = sfpt2->clone();
 855         sfpt->set_req(0, iffalse);
 856         outer_le->set_req(0, sfpt);
 857         register_control(sfpt, outer_ilt, iffalse);
 858         set_idom(outer_le, sfpt, dom_depth(sfpt));
 859       }
 860       lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
 861       if (loop->_safepts != NULL) {
 862         loop->_safepts->yank(sfpt2);
 863       }
 864     }
 865   }
 866 
 867   // Free up intermediate goo
 868   _igvn.remove_dead_node(hook);
 869 
 870 #ifdef ASSERT
 871   assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
 872   assert(l == loop->_head && l->phi() == phi && l->loopexit_or_null() == lex, "" );
 873 #endif
 874 #ifndef PRODUCT
 875   if (TraceLoopOpts) {
 876     tty->print("Counted      ");
 877     loop->dump_head();
 878   }
 879 #endif
 880 
 881   C->print_method(PHASE_AFTER_CLOOPS, 3);
 882 
 883   // Capture bounds of the loop in the induction variable Phi before
 884   // subsequent transformation (iteration splitting) obscures the
 885   // bounds
 886   l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn));
 887 
 888   if (strip_mine_loop) {
 889     l->mark_strip_mined();
 890     l->verify_strip_mined(1);
 891     outer_ilt->_head->as_Loop()->verify_strip_mined(1);
 892     loop = outer_ilt;
 893   }
 894 
 895   return true;
 896 }
 897 
 898 //----------------------exact_limit-------------------------------------------
 899 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
 900   assert(loop->_head->is_CountedLoop(), "");
 901   CountedLoopNode *cl = loop->_head->as_CountedLoop();
 902   assert(cl->is_valid_counted_loop(), "");
 903 
 904   if (ABS(cl->stride_con()) == 1 ||
 905       cl->limit()->Opcode() == Op_LoopLimit) {
 906     // Old code has exact limit (it could be incorrect in case of int overflow).
 907     // Loop limit is exact with stride == 1. And loop may already have exact limit.
 908     return cl->limit();
 909   }
 910   Node *limit = NULL;
 911 #ifdef ASSERT
 912   BoolTest::mask bt = cl->loopexit()->test_trip();
 913   assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
 914 #endif
 915   if (cl->has_exact_trip_count()) {
 916     // Simple case: loop has constant boundaries.
 917     // Use jlongs to avoid integer overflow.
 918     int stride_con = cl->stride_con();
 919     jlong  init_con = cl->init_trip()->get_int();
 920     jlong limit_con = cl->limit()->get_int();
 921     julong trip_cnt = cl->trip_count();
 922     jlong final_con = init_con + trip_cnt*stride_con;
 923     int final_int = (int)final_con;
 924     // The final value should be in integer range since the loop
 925     // is counted and the limit was checked for overflow.
 926     assert(final_con == (jlong)final_int, "final value should be integer");
 927     limit = _igvn.intcon(final_int);
 928   } else {
 929     // Create new LoopLimit node to get exact limit (final iv value).
 930     limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
 931     register_new_node(limit, cl->in(LoopNode::EntryControl));
 932   }
 933   assert(limit != NULL, "sanity");
 934   return limit;
 935 }
 936 
 937 //------------------------------Ideal------------------------------------------
 938 // Return a node which is more "ideal" than the current node.
 939 // Attempt to convert into a counted-loop.
 940 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
 941   if (!can_be_counted_loop(phase) && !is_OuterStripMinedLoop()) {
 942     phase->C->set_major_progress();
 943   }
 944   return RegionNode::Ideal(phase, can_reshape);
 945 }
 946 
 947 void LoopNode::verify_strip_mined(int expect_skeleton) const {
 948 #ifdef ASSERT
 949   const OuterStripMinedLoopNode* outer = NULL;
 950   const CountedLoopNode* inner = NULL;
 951   if (is_strip_mined()) {
 952     assert(is_CountedLoop(), "no Loop should be marked strip mined");
 953     inner = as_CountedLoop();
 954     outer = inner->in(LoopNode::EntryControl)->as_OuterStripMinedLoop();
 955   } else if (is_OuterStripMinedLoop()) {
 956     outer = this->as_OuterStripMinedLoop();
 957     inner = outer->unique_ctrl_out()->as_CountedLoop();
 958     assert(!is_strip_mined(), "outer loop shouldn't be marked strip mined");
 959   }
 960   if (inner != NULL || outer != NULL) {
 961     assert(inner != NULL && outer != NULL, "missing loop in strip mined nest");
 962     Node* outer_tail = outer->in(LoopNode::LoopBackControl);
 963     Node* outer_le = outer_tail->in(0);
 964     assert(outer_le->Opcode() == Op_OuterStripMinedLoopEnd, "tail of outer loop should be an If");
 965     Node* sfpt = outer_le->in(0);
 966     assert(sfpt->Opcode() == Op_SafePoint, "where's the safepoint?");
 967     Node* inner_out = sfpt->in(0);
 968     if (inner_out->outcnt() != 1) {
 969       ResourceMark rm;
 970       Unique_Node_List wq;
 971 
 972       for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
 973         Node* u = inner_out->fast_out(i);
 974         if (u == sfpt) {
 975           continue;
 976         }
 977         wq.clear();
 978         wq.push(u);
 979         bool found_sfpt = false;
 980         for (uint next = 0; next < wq.size() && !found_sfpt; next++) {
 981           Node *n = wq.at(next);
 982           for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && !found_sfpt; i++) {
 983             Node* u = n->fast_out(i);
 984             if (u == sfpt) {
 985               found_sfpt = true;
 986             }
 987             if (!u->is_CFG()) {
 988               wq.push(u);
 989             }
 990           }
 991         }
 992         assert(found_sfpt, "no node in loop that's not input to safepoint");
 993       }
 994     }
 995     CountedLoopEndNode* cle = inner_out->in(0)->as_CountedLoopEnd();
 996     assert(cle == inner->loopexit_or_null(), "mismatch");
 997     bool has_skeleton = outer_le->in(1)->bottom_type()->singleton() && outer_le->in(1)->bottom_type()->is_int()->get_con() == 0;
 998     if (has_skeleton) {
 999       assert(expect_skeleton == 1 || expect_skeleton == -1, "unexpected skeleton node");
1000       assert(outer->outcnt() == 2, "only phis");
1001     } else {
1002       assert(expect_skeleton == 0 || expect_skeleton == -1, "no skeleton node?");
1003       uint phis = 0;
1004       for (DUIterator_Fast imax, i = inner->fast_outs(imax); i < imax; i++) {
1005         Node* u = inner->fast_out(i);
1006         if (u->is_Phi()) {
1007           phis++;
1008         }
1009       }
1010       for (DUIterator_Fast imax, i = outer->fast_outs(imax); i < imax; i++) {
1011         Node* u = outer->fast_out(i);
1012         assert(u == outer || u == inner || u->is_Phi(), "nothing between inner and outer loop");
1013       }
1014       uint stores = 0;
1015       for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
1016         Node* u = inner_out->fast_out(i);
1017         if (u->is_Store()) {
1018           stores++;
1019         }
1020       }
1021       assert(outer->outcnt() >= phis + 2 && outer->outcnt() <= phis + 2 + stores + 1, "only phis");
1022     }
1023     assert(sfpt->outcnt() == 1, "no data node");
1024     assert(outer_tail->outcnt() == 1 || !has_skeleton, "no data node");
1025   }
1026 #endif
1027 }
1028 
1029 //=============================================================================
1030 //------------------------------Ideal------------------------------------------
1031 // Return a node which is more "ideal" than the current node.
1032 // Attempt to convert into a counted-loop.
1033 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1034   return RegionNode::Ideal(phase, can_reshape);
1035 }
1036 
1037 //------------------------------dump_spec--------------------------------------
1038 // Dump special per-node info
1039 #ifndef PRODUCT
1040 void CountedLoopNode::dump_spec(outputStream *st) const {
1041   LoopNode::dump_spec(st);
1042   if (stride_is_con()) {
1043     st->print("stride: %d ",stride_con());
1044   }
1045   if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
1046   if (is_main_loop()) st->print("main of N%d", _idx);
1047   if (is_post_loop()) st->print("post of N%d", _main_idx);
1048   if (is_strip_mined()) st->print(" strip mined");
1049 }
1050 #endif
1051 
1052 //=============================================================================
1053 int CountedLoopEndNode::stride_con() const {
1054   return stride()->bottom_type()->is_int()->get_con();
1055 }
1056 
1057 //=============================================================================
1058 //------------------------------Value-----------------------------------------
1059 const Type* LoopLimitNode::Value(PhaseGVN* phase) const {
1060   const Type* init_t   = phase->type(in(Init));
1061   const Type* limit_t  = phase->type(in(Limit));
1062   const Type* stride_t = phase->type(in(Stride));
1063   // Either input is TOP ==> the result is TOP
1064   if (init_t   == Type::TOP) return Type::TOP;
1065   if (limit_t  == Type::TOP) return Type::TOP;
1066   if (stride_t == Type::TOP) return Type::TOP;
1067 
1068   int stride_con = stride_t->is_int()->get_con();
1069   if (stride_con == 1)
1070     return NULL;  // Identity
1071 
1072   if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
1073     // Use jlongs to avoid integer overflow.
1074     jlong init_con   =  init_t->is_int()->get_con();
1075     jlong limit_con  = limit_t->is_int()->get_con();
1076     int  stride_m   = stride_con - (stride_con > 0 ? 1 : -1);
1077     jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
1078     jlong final_con  = init_con + stride_con*trip_count;
1079     int final_int = (int)final_con;
1080     // The final value should be in integer range since the loop
1081     // is counted and the limit was checked for overflow.
1082     assert(final_con == (jlong)final_int, "final value should be integer");
1083     return TypeInt::make(final_int);
1084   }
1085 
1086   return bottom_type(); // TypeInt::INT
1087 }
1088 
1089 //------------------------------Ideal------------------------------------------
1090 // Return a node which is more "ideal" than the current node.
1091 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1092   if (phase->type(in(Init))   == Type::TOP ||
1093       phase->type(in(Limit))  == Type::TOP ||
1094       phase->type(in(Stride)) == Type::TOP)
1095     return NULL;  // Dead
1096 
1097   int stride_con = phase->type(in(Stride))->is_int()->get_con();
1098   if (stride_con == 1)
1099     return NULL;  // Identity
1100 
1101   if (in(Init)->is_Con() && in(Limit)->is_Con())
1102     return NULL;  // Value
1103 
1104   // Delay following optimizations until all loop optimizations
1105   // done to keep Ideal graph simple.
1106   if (!can_reshape || phase->C->major_progress())
1107     return NULL;
1108 
1109   const TypeInt* init_t  = phase->type(in(Init) )->is_int();
1110   const TypeInt* limit_t = phase->type(in(Limit))->is_int();
1111   int stride_p;
1112   jlong lim, ini;
1113   julong max;
1114   if (stride_con > 0) {
1115     stride_p = stride_con;
1116     lim = limit_t->_hi;
1117     ini = init_t->_lo;
1118     max = (julong)max_jint;
1119   } else {
1120     stride_p = -stride_con;
1121     lim = init_t->_hi;
1122     ini = limit_t->_lo;
1123     max = (julong)min_jint;
1124   }
1125   julong range = lim - ini + stride_p;
1126   if (range <= max) {
1127     // Convert to integer expression if it is not overflow.
1128     Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
1129     Node *range = phase->transform(new SubINode(in(Limit), in(Init)));
1130     Node *bias  = phase->transform(new AddINode(range, stride_m));
1131     Node *trip  = phase->transform(new DivINode(0, bias, in(Stride)));
1132     Node *span  = phase->transform(new MulINode(trip, in(Stride)));
1133     return new AddINode(span, in(Init)); // exact limit
1134   }
1135 
1136   if (is_power_of_2(stride_p) ||                // divisor is 2^n
1137       !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
1138     // Convert to long expression to avoid integer overflow
1139     // and let igvn optimizer convert this division.
1140     //
1141     Node*   init   = phase->transform( new ConvI2LNode(in(Init)));
1142     Node*  limit   = phase->transform( new ConvI2LNode(in(Limit)));
1143     Node* stride   = phase->longcon(stride_con);
1144     Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
1145 
1146     Node *range = phase->transform(new SubLNode(limit, init));
1147     Node *bias  = phase->transform(new AddLNode(range, stride_m));
1148     Node *span;
1149     if (stride_con > 0 && is_power_of_2(stride_p)) {
1150       // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
1151       // and avoid generating rounding for division. Zero trip guard should
1152       // guarantee that init < limit but sometimes the guard is missing and
1153       // we can get situation when init > limit. Note, for the empty loop
1154       // optimization zero trip guard is generated explicitly which leaves
1155       // only RCE predicate where exact limit is used and the predicate
1156       // will simply fail forcing recompilation.
1157       Node* neg_stride   = phase->longcon(-stride_con);
1158       span = phase->transform(new AndLNode(bias, neg_stride));
1159     } else {
1160       Node *trip  = phase->transform(new DivLNode(0, bias, stride));
1161       span = phase->transform(new MulLNode(trip, stride));
1162     }
1163     // Convert back to int
1164     Node *span_int = phase->transform(new ConvL2INode(span));
1165     return new AddINode(span_int, in(Init)); // exact limit
1166   }
1167 
1168   return NULL;    // No progress
1169 }
1170 
1171 //------------------------------Identity---------------------------------------
1172 // If stride == 1 return limit node.
1173 Node* LoopLimitNode::Identity(PhaseGVN* phase) {
1174   int stride_con = phase->type(in(Stride))->is_int()->get_con();
1175   if (stride_con == 1 || stride_con == -1)
1176     return in(Limit);
1177   return this;
1178 }
1179 
1180 //=============================================================================
1181 //----------------------match_incr_with_optional_truncation--------------------
1182 // Match increment with optional truncation:
1183 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
1184 // Return NULL for failure. Success returns the increment node.
1185 Node* CountedLoopNode::match_incr_with_optional_truncation(
1186                       Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
1187   // Quick cutouts:
1188   if (expr == NULL || expr->req() != 3)  return NULL;
1189 
1190   Node *t1 = NULL;
1191   Node *t2 = NULL;
1192   const TypeInt* trunc_t = TypeInt::INT;
1193   Node* n1 = expr;
1194   int   n1op = n1->Opcode();
1195 
1196   // Try to strip (n1 & M) or (n1 << N >> N) from n1.
1197   if (n1op == Op_AndI &&
1198       n1->in(2)->is_Con() &&
1199       n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
1200     // %%% This check should match any mask of 2**K-1.
1201     t1 = n1;
1202     n1 = t1->in(1);
1203     n1op = n1->Opcode();
1204     trunc_t = TypeInt::CHAR;
1205   } else if (n1op == Op_RShiftI &&
1206              n1->in(1) != NULL &&
1207              n1->in(1)->Opcode() == Op_LShiftI &&
1208              n1->in(2) == n1->in(1)->in(2) &&
1209              n1->in(2)->is_Con()) {
1210     jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
1211     // %%% This check should match any shift in [1..31].
1212     if (shift == 16 || shift == 8) {
1213       t1 = n1;
1214       t2 = t1->in(1);
1215       n1 = t2->in(1);
1216       n1op = n1->Opcode();
1217       if (shift == 16) {
1218         trunc_t = TypeInt::SHORT;
1219       } else if (shift == 8) {
1220         trunc_t = TypeInt::BYTE;
1221       }
1222     }
1223   }
1224 
1225   // If (maybe after stripping) it is an AddI, we won:
1226   if (n1op == Op_AddI) {
1227     *trunc1 = t1;
1228     *trunc2 = t2;
1229     *trunc_type = trunc_t;
1230     return n1;
1231   }
1232 
1233   // failed
1234   return NULL;
1235 }
1236 
1237 LoopNode* CountedLoopNode::skip_strip_mined(int expect_skeleton) {
1238   if (is_strip_mined()) {
1239     verify_strip_mined(expect_skeleton);
1240     return in(EntryControl)->as_Loop();
1241   }
1242   return this;
1243 }
1244 
1245 OuterStripMinedLoopNode* CountedLoopNode::outer_loop() const {
1246   assert(is_strip_mined(), "not a strip mined loop");
1247   Node* c = in(EntryControl);
1248   if (c == NULL || c->is_top() || !c->is_OuterStripMinedLoop()) {
1249     return NULL;
1250   }
1251   return c->as_OuterStripMinedLoop();
1252 }
1253 
1254 IfTrueNode* OuterStripMinedLoopNode::outer_loop_tail() const {
1255   Node* c = in(LoopBackControl);
1256   if (c == NULL || c->is_top()) {
1257     return NULL;
1258   }
1259   return c->as_IfTrue();
1260 }
1261 
1262 IfTrueNode* CountedLoopNode::outer_loop_tail() const {
1263   LoopNode* l = outer_loop();
1264   if (l == NULL) {
1265     return NULL;
1266   }
1267   return l->outer_loop_tail();
1268 }
1269 
1270 OuterStripMinedLoopEndNode* OuterStripMinedLoopNode::outer_loop_end() const {
1271   IfTrueNode* proj = outer_loop_tail();
1272   if (proj == NULL) {
1273     return NULL;
1274   }
1275   Node* c = proj->in(0);
1276   if (c == NULL || c->is_top() || c->outcnt() != 2) {
1277     return NULL;
1278   }
1279   return c->as_OuterStripMinedLoopEnd();
1280 }
1281 
1282 OuterStripMinedLoopEndNode* CountedLoopNode::outer_loop_end() const {
1283   LoopNode* l = outer_loop();
1284   if (l == NULL) {
1285     return NULL;
1286   }
1287   return l->outer_loop_end();
1288 }
1289 
1290 IfFalseNode* OuterStripMinedLoopNode::outer_loop_exit() const {
1291   IfNode* le = outer_loop_end();
1292   if (le == NULL) {
1293     return NULL;
1294   }
1295   Node* c = le->proj_out_or_null(false);
1296   if (c == NULL) {
1297     return NULL;
1298   }
1299   return c->as_IfFalse();
1300 }
1301 
1302 IfFalseNode* CountedLoopNode::outer_loop_exit() const {
1303   LoopNode* l = outer_loop();
1304   if (l == NULL) {
1305     return NULL;
1306   }
1307   return l->outer_loop_exit();
1308 }
1309 
1310 SafePointNode* OuterStripMinedLoopNode::outer_safepoint() const {
1311   IfNode* le = outer_loop_end();
1312   if (le == NULL) {
1313     return NULL;
1314   }
1315   Node* c = le->in(0);
1316   if (c == NULL || c->is_top()) {
1317     return NULL;
1318   }
1319   assert(c->Opcode() == Op_SafePoint, "broken outer loop");
1320   return c->as_SafePoint();
1321 }
1322 
1323 SafePointNode* CountedLoopNode::outer_safepoint() const {
1324   LoopNode* l = outer_loop();
1325   if (l == NULL) {
1326     return NULL;
1327   }
1328   return l->outer_safepoint();
1329 }
1330 
1331 Node* CountedLoopNode::skip_predicates_from_entry(Node* ctrl) {
1332     while (ctrl != NULL && ctrl->is_Proj() && ctrl->in(0)->is_If() &&
1333            ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->outcnt() == 1 &&
1334            ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->unique_out()->Opcode() == Op_Halt) {
1335       ctrl = ctrl->in(0)->in(0);
1336     }
1337 
1338     return ctrl;
1339   }
1340 
1341 Node* CountedLoopNode::skip_predicates() {
1342   if (is_main_loop()) {
1343     Node* ctrl = skip_strip_mined()->in(LoopNode::EntryControl);
1344 
1345     return skip_predicates_from_entry(ctrl);
1346   }
1347   return in(LoopNode::EntryControl);
1348 }
1349 
1350 void OuterStripMinedLoopNode::adjust_strip_mined_loop(PhaseIterGVN* igvn) {
1351   // Look for the outer & inner strip mined loop, reduce number of
1352   // iterations of the inner loop, set exit condition of outer loop,
1353   // construct required phi nodes for outer loop.
1354   CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
1355   assert(inner_cl->is_strip_mined(), "inner loop should be strip mined");
1356   Node* inner_iv_phi = inner_cl->phi();
1357   if (inner_iv_phi == NULL) {
1358     IfNode* outer_le = outer_loop_end();
1359     Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1360     igvn->replace_node(outer_le, iff);
1361     inner_cl->clear_strip_mined();
1362     return;
1363   }
1364   CountedLoopEndNode* inner_cle = inner_cl->loopexit();
1365 
1366   int stride = inner_cl->stride_con();
1367   jlong scaled_iters_long = ((jlong)LoopStripMiningIter) * ABS(stride);
1368   int scaled_iters = (int)scaled_iters_long;
1369   int short_scaled_iters = LoopStripMiningIterShortLoop* ABS(stride);
1370   const TypeInt* inner_iv_t = igvn->type(inner_iv_phi)->is_int();
1371   jlong iter_estimate = (jlong)inner_iv_t->_hi - (jlong)inner_iv_t->_lo;
1372   assert(iter_estimate > 0, "broken");
1373   if ((jlong)scaled_iters != scaled_iters_long || iter_estimate <= short_scaled_iters) {
1374     // Remove outer loop and safepoint (too few iterations)
1375     Node* outer_sfpt = outer_safepoint();
1376     Node* outer_out = outer_loop_exit();
1377     igvn->replace_node(outer_out, outer_sfpt->in(0));
1378     igvn->replace_input_of(outer_sfpt, 0, igvn->C->top());
1379     inner_cl->clear_strip_mined();
1380     return;
1381   }
1382   if (iter_estimate <= scaled_iters_long) {
1383     // We would only go through one iteration of
1384     // the outer loop: drop the outer loop but
1385     // keep the safepoint so we don't run for
1386     // too long without a safepoint
1387     IfNode* outer_le = outer_loop_end();
1388     Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1389     igvn->replace_node(outer_le, iff);
1390     inner_cl->clear_strip_mined();
1391     return;
1392   }
1393 
1394   Node* cle_tail = inner_cle->proj_out(true);
1395   ResourceMark rm;
1396   Node_List old_new;
1397   if (cle_tail->outcnt() > 1) {
1398     // Look for nodes on backedge of inner loop and clone them
1399     Unique_Node_List backedge_nodes;
1400     for (DUIterator_Fast imax, i = cle_tail->fast_outs(imax); i < imax; i++) {
1401       Node* u = cle_tail->fast_out(i);
1402       if (u != inner_cl) {
1403         assert(!u->is_CFG(), "control flow on the backedge?");
1404         backedge_nodes.push(u);
1405       }
1406     }
1407     uint last = igvn->C->unique();
1408     for (uint next = 0; next < backedge_nodes.size(); next++) {
1409       Node* n = backedge_nodes.at(next);
1410       old_new.map(n->_idx, n->clone());
1411       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1412         Node* u = n->fast_out(i);
1413         assert(!u->is_CFG(), "broken");
1414         if (u->_idx >= last) {
1415           continue;
1416         }
1417         if (!u->is_Phi()) {
1418           backedge_nodes.push(u);
1419         } else {
1420           assert(u->in(0) == inner_cl, "strange phi on the backedge");
1421         }
1422       }
1423     }
1424     // Put the clones on the outer loop backedge
1425     Node* le_tail = outer_loop_tail();
1426     for (uint next = 0; next < backedge_nodes.size(); next++) {
1427       Node *n = old_new[backedge_nodes.at(next)->_idx];
1428       for (uint i = 1; i < n->req(); i++) {
1429         if (n->in(i) != NULL && old_new[n->in(i)->_idx] != NULL) {
1430           n->set_req(i, old_new[n->in(i)->_idx]);
1431         }
1432       }
1433       if (n->in(0) != NULL && n->in(0) == cle_tail) {
1434         n->set_req(0, le_tail);
1435       }
1436       igvn->register_new_node_with_optimizer(n);
1437     }
1438   }
1439 
1440   Node* iv_phi = NULL;
1441   // Make a clone of each phi in the inner loop
1442   // for the outer loop
1443   for (uint i = 0; i < inner_cl->outcnt(); i++) {
1444     Node* u = inner_cl->raw_out(i);
1445     if (u->is_Phi()) {
1446       assert(u->in(0) == inner_cl, "inconsistent");
1447       Node* phi = u->clone();
1448       phi->set_req(0, this);
1449       Node* be = old_new[phi->in(LoopNode::LoopBackControl)->_idx];
1450       if (be != NULL) {
1451         phi->set_req(LoopNode::LoopBackControl, be);
1452       }
1453       phi = igvn->transform(phi);
1454       igvn->replace_input_of(u, LoopNode::EntryControl, phi);
1455       if (u == inner_iv_phi) {
1456         iv_phi = phi;
1457       }
1458     }
1459   }
1460   Node* cle_out = inner_cle->proj_out(false);
1461   if (cle_out->outcnt() > 1) {
1462     // Look for chains of stores that were sunk
1463     // out of the inner loop and are in the outer loop
1464     for (DUIterator_Fast imax, i = cle_out->fast_outs(imax); i < imax; i++) {
1465       Node* u = cle_out->fast_out(i);
1466       if (u->is_Store()) {
1467         Node* first = u;
1468         for(;;) {
1469           Node* next = first->in(MemNode::Memory);
1470           if (!next->is_Store() || next->in(0) != cle_out) {
1471             break;
1472           }
1473           first = next;
1474         }
1475         Node* last = u;
1476         for(;;) {
1477           Node* next = NULL;
1478           for (DUIterator_Fast jmax, j = last->fast_outs(jmax); j < jmax; j++) {
1479             Node* uu = last->fast_out(j);
1480             if (uu->is_Store() && uu->in(0) == cle_out) {
1481               assert(next == NULL, "only one in the outer loop");
1482               next = uu;
1483             }
1484           }
1485           if (next == NULL) {
1486             break;
1487           }
1488           last = next;
1489         }
1490         Node* phi = NULL;
1491         for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) {
1492           Node* uu = fast_out(j);
1493           if (uu->is_Phi()) {
1494             Node* be = uu->in(LoopNode::LoopBackControl);
1495             if (be->is_Store() && old_new[be->_idx] != NULL) {
1496               assert(false, "store on the backedge + sunk stores: unsupported");
1497               // drop outer loop
1498               IfNode* outer_le = outer_loop_end();
1499               Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1500               igvn->replace_node(outer_le, iff);
1501               inner_cl->clear_strip_mined();
1502               return;
1503             }
1504             if (be == last || be == first->in(MemNode::Memory)) {
1505               assert(phi == NULL, "only one phi");
1506               phi = uu;
1507             }
1508           }
1509         }
1510 #ifdef ASSERT
1511         for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) {
1512           Node* uu = fast_out(j);
1513           if (uu->is_Phi() && uu->bottom_type() == Type::MEMORY) {
1514             if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) {
1515               assert(phi == uu, "what's that phi?");
1516             } else if (uu->adr_type() == TypePtr::BOTTOM) {
1517               Node* n = uu->in(LoopNode::LoopBackControl);
1518               uint limit = igvn->C->live_nodes();
1519               uint i = 0;
1520               while (n != uu) {
1521                 i++;
1522                 assert(i < limit, "infinite loop");
1523                 if (n->is_Proj()) {
1524                   n = n->in(0);
1525                 } else if (n->is_SafePoint() || n->is_MemBar()) {
1526                   n = n->in(TypeFunc::Memory);
1527                 } else if (n->is_Phi()) {
1528                   n = n->in(1);
1529                 } else if (n->is_MergeMem()) {
1530                   n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type()));
1531                 } else if (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) {
1532                   n = n->in(MemNode::Memory);
1533                 } else {
1534                   n->dump();
1535                   ShouldNotReachHere();
1536                 }
1537               }
1538             }
1539           }
1540         }
1541 #endif
1542         if (phi == NULL) {
1543           // If the an entire chains was sunk, the
1544           // inner loop has no phi for that memory
1545           // slice, create one for the outer loop
1546           phi = PhiNode::make(this, first->in(MemNode::Memory), Type::MEMORY,
1547                               igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type())));
1548           phi->set_req(LoopNode::LoopBackControl, last);
1549           phi = igvn->transform(phi);
1550           igvn->replace_input_of(first, MemNode::Memory, phi);
1551         } else {
1552           // Or fix the outer loop fix to include
1553           // that chain of stores.
1554           Node* be = phi->in(LoopNode::LoopBackControl);
1555           assert(!(be->is_Store() && old_new[be->_idx] != NULL), "store on the backedge + sunk stores: unsupported");
1556           if (be == first->in(MemNode::Memory)) {
1557             if (be == phi->in(LoopNode::LoopBackControl)) {
1558               igvn->replace_input_of(phi, LoopNode::LoopBackControl, last);
1559             } else {
1560               igvn->replace_input_of(be, MemNode::Memory, last);
1561             }
1562           } else {
1563 #ifdef ASSERT
1564             if (be == phi->in(LoopNode::LoopBackControl)) {
1565               assert(phi->in(LoopNode::LoopBackControl) == last, "");
1566             } else {
1567               assert(be->in(MemNode::Memory) == last, "");
1568             }
1569 #endif
1570           }
1571         }
1572       }
1573     }
1574   }
1575 
1576   if (iv_phi != NULL) {
1577     // Now adjust the inner loop's exit condition
1578     Node* limit = inner_cl->limit();
1579     Node* sub = NULL;
1580     if (stride > 0) {
1581       sub = igvn->transform(new SubINode(limit, iv_phi));
1582     } else {
1583       sub = igvn->transform(new SubINode(iv_phi, limit));
1584     }
1585     Node* min = igvn->transform(new MinINode(sub, igvn->intcon(scaled_iters)));
1586     Node* new_limit = NULL;
1587     if (stride > 0) {
1588       new_limit = igvn->transform(new AddINode(min, iv_phi));
1589     } else {
1590       new_limit = igvn->transform(new SubINode(iv_phi, min));
1591     }
1592     Node* inner_cmp = inner_cle->cmp_node();
1593     Node* inner_bol = inner_cle->in(CountedLoopEndNode::TestValue);
1594     Node* outer_bol = inner_bol;
1595     // cmp node for inner loop may be shared
1596     inner_cmp = inner_cmp->clone();
1597     inner_cmp->set_req(2, new_limit);
1598     inner_bol = inner_bol->clone();
1599     inner_bol->set_req(1, igvn->transform(inner_cmp));
1600     igvn->replace_input_of(inner_cle, CountedLoopEndNode::TestValue, igvn->transform(inner_bol));
1601     // Set the outer loop's exit condition too
1602     igvn->replace_input_of(outer_loop_end(), 1, outer_bol);
1603   } else {
1604     assert(false, "should be able to adjust outer loop");
1605     IfNode* outer_le = outer_loop_end();
1606     Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1607     igvn->replace_node(outer_le, iff);
1608     inner_cl->clear_strip_mined();
1609   }
1610 }
1611 
1612 const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const {
1613   if (!in(0)) return Type::TOP;
1614   if (phase->type(in(0)) == Type::TOP)
1615     return Type::TOP;
1616 
1617   return TypeTuple::IFBOTH;
1618 }
1619 
1620 Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1621   if (remove_dead_region(phase, can_reshape))  return this;
1622 
1623   return NULL;
1624 }
1625 
1626 //------------------------------filtered_type--------------------------------
1627 // Return a type based on condition control flow
1628 // A successful return will be a type that is restricted due
1629 // to a series of dominating if-tests, such as:
1630 //    if (i < 10) {
1631 //       if (i > 0) {
1632 //          here: "i" type is [1..10)
1633 //       }
1634 //    }
1635 // or a control flow merge
1636 //    if (i < 10) {
1637 //       do {
1638 //          phi( , ) -- at top of loop type is [min_int..10)
1639 //         i = ?
1640 //       } while ( i < 10)
1641 //
1642 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
1643   assert(n && n->bottom_type()->is_int(), "must be int");
1644   const TypeInt* filtered_t = NULL;
1645   if (!n->is_Phi()) {
1646     assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
1647     filtered_t = filtered_type_from_dominators(n, n_ctrl);
1648 
1649   } else {
1650     Node* phi    = n->as_Phi();
1651     Node* region = phi->in(0);
1652     assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
1653     if (region && region != C->top()) {
1654       for (uint i = 1; i < phi->req(); i++) {
1655         Node* val   = phi->in(i);
1656         Node* use_c = region->in(i);
1657         const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
1658         if (val_t != NULL) {
1659           if (filtered_t == NULL) {
1660             filtered_t = val_t;
1661           } else {
1662             filtered_t = filtered_t->meet(val_t)->is_int();
1663           }
1664         }
1665       }
1666     }
1667   }
1668   const TypeInt* n_t = _igvn.type(n)->is_int();
1669   if (filtered_t != NULL) {
1670     n_t = n_t->join(filtered_t)->is_int();
1671   }
1672   return n_t;
1673 }
1674 
1675 
1676 //------------------------------filtered_type_from_dominators--------------------------------
1677 // Return a possibly more restrictive type for val based on condition control flow of dominators
1678 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
1679   if (val->is_Con()) {
1680      return val->bottom_type()->is_int();
1681   }
1682   uint if_limit = 10; // Max number of dominating if's visited
1683   const TypeInt* rtn_t = NULL;
1684 
1685   if (use_ctrl && use_ctrl != C->top()) {
1686     Node* val_ctrl = get_ctrl(val);
1687     uint val_dom_depth = dom_depth(val_ctrl);
1688     Node* pred = use_ctrl;
1689     uint if_cnt = 0;
1690     while (if_cnt < if_limit) {
1691       if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
1692         if_cnt++;
1693         const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
1694         if (if_t != NULL) {
1695           if (rtn_t == NULL) {
1696             rtn_t = if_t;
1697           } else {
1698             rtn_t = rtn_t->join(if_t)->is_int();
1699           }
1700         }
1701       }
1702       pred = idom(pred);
1703       if (pred == NULL || pred == C->top()) {
1704         break;
1705       }
1706       // Stop if going beyond definition block of val
1707       if (dom_depth(pred) < val_dom_depth) {
1708         break;
1709       }
1710     }
1711   }
1712   return rtn_t;
1713 }
1714 
1715 
1716 //------------------------------dump_spec--------------------------------------
1717 // Dump special per-node info
1718 #ifndef PRODUCT
1719 void CountedLoopEndNode::dump_spec(outputStream *st) const {
1720   if( in(TestValue) != NULL && in(TestValue)->is_Bool() ) {
1721     BoolTest bt( test_trip()); // Added this for g++.
1722 
1723     st->print("[");
1724     bt.dump_on(st);
1725     st->print("]");
1726   }
1727   st->print(" ");
1728   IfNode::dump_spec(st);
1729 }
1730 #endif
1731 
1732 //=============================================================================
1733 //------------------------------is_member--------------------------------------
1734 // Is 'l' a member of 'this'?
1735 bool IdealLoopTree::is_member(const IdealLoopTree *l) const {
1736   while( l->_nest > _nest ) l = l->_parent;
1737   return l == this;
1738 }
1739 
1740 //------------------------------set_nest---------------------------------------
1741 // Set loop tree nesting depth.  Accumulate _has_call bits.
1742 int IdealLoopTree::set_nest( uint depth ) {
1743   _nest = depth;
1744   int bits = _has_call;
1745   if( _child ) bits |= _child->set_nest(depth+1);
1746   if( bits ) _has_call = 1;
1747   if( _next  ) bits |= _next ->set_nest(depth  );
1748   return bits;
1749 }
1750 
1751 //------------------------------split_fall_in----------------------------------
1752 // Split out multiple fall-in edges from the loop header.  Move them to a
1753 // private RegionNode before the loop.  This becomes the loop landing pad.
1754 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
1755   PhaseIterGVN &igvn = phase->_igvn;
1756   uint i;
1757 
1758   // Make a new RegionNode to be the landing pad.
1759   Node *landing_pad = new RegionNode( fall_in_cnt+1 );
1760   phase->set_loop(landing_pad,_parent);
1761   // Gather all the fall-in control paths into the landing pad
1762   uint icnt = fall_in_cnt;
1763   uint oreq = _head->req();
1764   for( i = oreq-1; i>0; i-- )
1765     if( !phase->is_member( this, _head->in(i) ) )
1766       landing_pad->set_req(icnt--,_head->in(i));
1767 
1768   // Peel off PhiNode edges as well
1769   for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1770     Node *oj = _head->fast_out(j);
1771     if( oj->is_Phi() ) {
1772       PhiNode* old_phi = oj->as_Phi();
1773       assert( old_phi->region() == _head, "" );
1774       igvn.hash_delete(old_phi);   // Yank from hash before hacking edges
1775       Node *p = PhiNode::make_blank(landing_pad, old_phi);
1776       uint icnt = fall_in_cnt;
1777       for( i = oreq-1; i>0; i-- ) {
1778         if( !phase->is_member( this, _head->in(i) ) ) {
1779           p->init_req(icnt--, old_phi->in(i));
1780           // Go ahead and clean out old edges from old phi
1781           old_phi->del_req(i);
1782         }
1783       }
1784       // Search for CSE's here, because ZKM.jar does a lot of
1785       // loop hackery and we need to be a little incremental
1786       // with the CSE to avoid O(N^2) node blow-up.
1787       Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
1788       if( p2 ) {                // Found CSE
1789         p->destruct();          // Recover useless new node
1790         p = p2;                 // Use old node
1791       } else {
1792         igvn.register_new_node_with_optimizer(p, old_phi);
1793       }
1794       // Make old Phi refer to new Phi.
1795       old_phi->add_req(p);
1796       // Check for the special case of making the old phi useless and
1797       // disappear it.  In JavaGrande I have a case where this useless
1798       // Phi is the loop limit and prevents recognizing a CountedLoop
1799       // which in turn prevents removing an empty loop.
1800       Node *id_old_phi = igvn.apply_identity(old_phi);
1801       if( id_old_phi != old_phi ) { // Found a simple identity?
1802         // Note that I cannot call 'replace_node' here, because
1803         // that will yank the edge from old_phi to the Region and
1804         // I'm mid-iteration over the Region's uses.
1805         for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
1806           Node* use = old_phi->last_out(i);
1807           igvn.rehash_node_delayed(use);
1808           uint uses_found = 0;
1809           for (uint j = 0; j < use->len(); j++) {
1810             if (use->in(j) == old_phi) {
1811               if (j < use->req()) use->set_req (j, id_old_phi);
1812               else                use->set_prec(j, id_old_phi);
1813               uses_found++;
1814             }
1815           }
1816           i -= uses_found;    // we deleted 1 or more copies of this edge
1817         }
1818       }
1819       igvn._worklist.push(old_phi);
1820     }
1821   }
1822   // Finally clean out the fall-in edges from the RegionNode
1823   for( i = oreq-1; i>0; i-- ) {
1824     if( !phase->is_member( this, _head->in(i) ) ) {
1825       _head->del_req(i);
1826     }
1827   }
1828   igvn.rehash_node_delayed(_head);
1829   // Transform landing pad
1830   igvn.register_new_node_with_optimizer(landing_pad, _head);
1831   // Insert landing pad into the header
1832   _head->add_req(landing_pad);
1833 }
1834 
1835 //------------------------------split_outer_loop-------------------------------
1836 // Split out the outermost loop from this shared header.
1837 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
1838   PhaseIterGVN &igvn = phase->_igvn;
1839 
1840   // Find index of outermost loop; it should also be my tail.
1841   uint outer_idx = 1;
1842   while( _head->in(outer_idx) != _tail ) outer_idx++;
1843 
1844   // Make a LoopNode for the outermost loop.
1845   Node *ctl = _head->in(LoopNode::EntryControl);
1846   Node *outer = new LoopNode( ctl, _head->in(outer_idx) );
1847   outer = igvn.register_new_node_with_optimizer(outer, _head);
1848   phase->set_created_loop_node();
1849 
1850   // Outermost loop falls into '_head' loop
1851   _head->set_req(LoopNode::EntryControl, outer);
1852   _head->del_req(outer_idx);
1853   // Split all the Phis up between '_head' loop and 'outer' loop.
1854   for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1855     Node *out = _head->fast_out(j);
1856     if( out->is_Phi() ) {
1857       PhiNode *old_phi = out->as_Phi();
1858       assert( old_phi->region() == _head, "" );
1859       Node *phi = PhiNode::make_blank(outer, old_phi);
1860       phi->init_req(LoopNode::EntryControl,    old_phi->in(LoopNode::EntryControl));
1861       phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
1862       phi = igvn.register_new_node_with_optimizer(phi, old_phi);
1863       // Make old Phi point to new Phi on the fall-in path
1864       igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
1865       old_phi->del_req(outer_idx);
1866     }
1867   }
1868 
1869   // Use the new loop head instead of the old shared one
1870   _head = outer;
1871   phase->set_loop(_head, this);
1872 }
1873 
1874 //------------------------------fix_parent-------------------------------------
1875 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
1876   loop->_parent = parent;
1877   if( loop->_child ) fix_parent( loop->_child, loop   );
1878   if( loop->_next  ) fix_parent( loop->_next , parent );
1879 }
1880 
1881 //------------------------------estimate_path_freq-----------------------------
1882 static float estimate_path_freq( Node *n ) {
1883   // Try to extract some path frequency info
1884   IfNode *iff;
1885   for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
1886     uint nop = n->Opcode();
1887     if( nop == Op_SafePoint ) {   // Skip any safepoint
1888       n = n->in(0);
1889       continue;
1890     }
1891     if( nop == Op_CatchProj ) {   // Get count from a prior call
1892       // Assume call does not always throw exceptions: means the call-site
1893       // count is also the frequency of the fall-through path.
1894       assert( n->is_CatchProj(), "" );
1895       if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
1896         return 0.0f;            // Assume call exception path is rare
1897       Node *call = n->in(0)->in(0)->in(0);
1898       assert( call->is_Call(), "expect a call here" );
1899       const JVMState *jvms = ((CallNode*)call)->jvms();
1900       ciMethodData* methodData = jvms->method()->method_data();
1901       if (!methodData->is_mature())  return 0.0f; // No call-site data
1902       ciProfileData* data = methodData->bci_to_data(jvms->bci());
1903       if ((data == NULL) || !data->is_CounterData()) {
1904         // no call profile available, try call's control input
1905         n = n->in(0);
1906         continue;
1907       }
1908       return data->as_CounterData()->count()/FreqCountInvocations;
1909     }
1910     // See if there's a gating IF test
1911     Node *n_c = n->in(0);
1912     if( !n_c->is_If() ) break;       // No estimate available
1913     iff = n_c->as_If();
1914     if( iff->_fcnt != COUNT_UNKNOWN )   // Have a valid count?
1915       // Compute how much count comes on this path
1916       return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
1917     // Have no count info.  Skip dull uncommon-trap like branches.
1918     if( (nop == Op_IfTrue  && iff->_prob < PROB_LIKELY_MAG(5)) ||
1919         (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
1920       break;
1921     // Skip through never-taken branch; look for a real loop exit.
1922     n = iff->in(0);
1923   }
1924   return 0.0f;                  // No estimate available
1925 }
1926 
1927 //------------------------------merge_many_backedges---------------------------
1928 // Merge all the backedges from the shared header into a private Region.
1929 // Feed that region as the one backedge to this loop.
1930 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
1931   uint i;
1932 
1933   // Scan for the top 2 hottest backedges
1934   float hotcnt = 0.0f;
1935   float warmcnt = 0.0f;
1936   uint hot_idx = 0;
1937   // Loop starts at 2 because slot 1 is the fall-in path
1938   for( i = 2; i < _head->req(); i++ ) {
1939     float cnt = estimate_path_freq(_head->in(i));
1940     if( cnt > hotcnt ) {       // Grab hottest path
1941       warmcnt = hotcnt;
1942       hotcnt = cnt;
1943       hot_idx = i;
1944     } else if( cnt > warmcnt ) { // And 2nd hottest path
1945       warmcnt = cnt;
1946     }
1947   }
1948 
1949   // See if the hottest backedge is worthy of being an inner loop
1950   // by being much hotter than the next hottest backedge.
1951   if( hotcnt <= 0.0001 ||
1952       hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
1953 
1954   // Peel out the backedges into a private merge point; peel
1955   // them all except optionally hot_idx.
1956   PhaseIterGVN &igvn = phase->_igvn;
1957 
1958   Node *hot_tail = NULL;
1959   // Make a Region for the merge point
1960   Node *r = new RegionNode(1);
1961   for( i = 2; i < _head->req(); i++ ) {
1962     if( i != hot_idx )
1963       r->add_req( _head->in(i) );
1964     else hot_tail = _head->in(i);
1965   }
1966   igvn.register_new_node_with_optimizer(r, _head);
1967   // Plug region into end of loop _head, followed by hot_tail
1968   while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
1969   igvn.replace_input_of(_head, 2, r);
1970   if( hot_idx ) _head->add_req(hot_tail);
1971 
1972   // Split all the Phis up between '_head' loop and the Region 'r'
1973   for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1974     Node *out = _head->fast_out(j);
1975     if( out->is_Phi() ) {
1976       PhiNode* n = out->as_Phi();
1977       igvn.hash_delete(n);      // Delete from hash before hacking edges
1978       Node *hot_phi = NULL;
1979       Node *phi = new PhiNode(r, n->type(), n->adr_type());
1980       // Check all inputs for the ones to peel out
1981       uint j = 1;
1982       for( uint i = 2; i < n->req(); i++ ) {
1983         if( i != hot_idx )
1984           phi->set_req( j++, n->in(i) );
1985         else hot_phi = n->in(i);
1986       }
1987       // Register the phi but do not transform until whole place transforms
1988       igvn.register_new_node_with_optimizer(phi, n);
1989       // Add the merge phi to the old Phi
1990       while( n->req() > 3 ) n->del_req( n->req()-1 );
1991       igvn.replace_input_of(n, 2, phi);
1992       if( hot_idx ) n->add_req(hot_phi);
1993     }
1994   }
1995 
1996 
1997   // Insert a new IdealLoopTree inserted below me.  Turn it into a clone
1998   // of self loop tree.  Turn self into a loop headed by _head and with
1999   // tail being the new merge point.
2000   IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
2001   phase->set_loop(_tail,ilt);   // Adjust tail
2002   _tail = r;                    // Self's tail is new merge point
2003   phase->set_loop(r,this);
2004   ilt->_child = _child;         // New guy has my children
2005   _child = ilt;                 // Self has new guy as only child
2006   ilt->_parent = this;          // new guy has self for parent
2007   ilt->_nest = _nest;           // Same nesting depth (for now)
2008 
2009   // Starting with 'ilt', look for child loop trees using the same shared
2010   // header.  Flatten these out; they will no longer be loops in the end.
2011   IdealLoopTree **pilt = &_child;
2012   while( ilt ) {
2013     if( ilt->_head == _head ) {
2014       uint i;
2015       for( i = 2; i < _head->req(); i++ )
2016         if( _head->in(i) == ilt->_tail )
2017           break;                // Still a loop
2018       if( i == _head->req() ) { // No longer a loop
2019         // Flatten ilt.  Hang ilt's "_next" list from the end of
2020         // ilt's '_child' list.  Move the ilt's _child up to replace ilt.
2021         IdealLoopTree **cp = &ilt->_child;
2022         while( *cp ) cp = &(*cp)->_next;   // Find end of child list
2023         *cp = ilt->_next;       // Hang next list at end of child list
2024         *pilt = ilt->_child;    // Move child up to replace ilt
2025         ilt->_head = NULL;      // Flag as a loop UNIONED into parent
2026         ilt = ilt->_child;      // Repeat using new ilt
2027         continue;               // do not advance over ilt->_child
2028       }
2029       assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
2030       phase->set_loop(_head,ilt);
2031     }
2032     pilt = &ilt->_child;        // Advance to next
2033     ilt = *pilt;
2034   }
2035 
2036   if( _child ) fix_parent( _child, this );
2037 }
2038 
2039 //------------------------------beautify_loops---------------------------------
2040 // Split shared headers and insert loop landing pads.
2041 // Insert a LoopNode to replace the RegionNode.
2042 // Return TRUE if loop tree is structurally changed.
2043 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
2044   bool result = false;
2045   // Cache parts in locals for easy
2046   PhaseIterGVN &igvn = phase->_igvn;
2047 
2048   igvn.hash_delete(_head);      // Yank from hash before hacking edges
2049 
2050   // Check for multiple fall-in paths.  Peel off a landing pad if need be.
2051   int fall_in_cnt = 0;
2052   for( uint i = 1; i < _head->req(); i++ )
2053     if( !phase->is_member( this, _head->in(i) ) )
2054       fall_in_cnt++;
2055   assert( fall_in_cnt, "at least 1 fall-in path" );
2056   if( fall_in_cnt > 1 )         // Need a loop landing pad to merge fall-ins
2057     split_fall_in( phase, fall_in_cnt );
2058 
2059   // Swap inputs to the _head and all Phis to move the fall-in edge to
2060   // the left.
2061   fall_in_cnt = 1;
2062   while( phase->is_member( this, _head->in(fall_in_cnt) ) )
2063     fall_in_cnt++;
2064   if( fall_in_cnt > 1 ) {
2065     // Since I am just swapping inputs I do not need to update def-use info
2066     Node *tmp = _head->in(1);
2067     igvn.rehash_node_delayed(_head);
2068     _head->set_req( 1, _head->in(fall_in_cnt) );
2069     _head->set_req( fall_in_cnt, tmp );
2070     // Swap also all Phis
2071     for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
2072       Node* phi = _head->fast_out(i);
2073       if( phi->is_Phi() ) {
2074         igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges
2075         tmp = phi->in(1);
2076         phi->set_req( 1, phi->in(fall_in_cnt) );
2077         phi->set_req( fall_in_cnt, tmp );
2078       }
2079     }
2080   }
2081   assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
2082   assert(  phase->is_member( this, _head->in(2) ), "right edge is loop" );
2083 
2084   // If I am a shared header (multiple backedges), peel off the many
2085   // backedges into a private merge point and use the merge point as
2086   // the one true backedge.
2087   if( _head->req() > 3 ) {
2088     // Merge the many backedges into a single backedge but leave
2089     // the hottest backedge as separate edge for the following peel.
2090     merge_many_backedges( phase );
2091     result = true;
2092   }
2093 
2094   // If I have one hot backedge, peel off myself loop.
2095   // I better be the outermost loop.
2096   if (_head->req() > 3 && !_irreducible) {
2097     split_outer_loop( phase );
2098     result = true;
2099 
2100   } else if (!_head->is_Loop() && !_irreducible) {
2101     // Make a new LoopNode to replace the old loop head
2102     Node *l = new LoopNode( _head->in(1), _head->in(2) );
2103     l = igvn.register_new_node_with_optimizer(l, _head);
2104     phase->set_created_loop_node();
2105     // Go ahead and replace _head
2106     phase->_igvn.replace_node( _head, l );
2107     _head = l;
2108     phase->set_loop(_head, this);
2109   }
2110 
2111   // Now recursively beautify nested loops
2112   if( _child ) result |= _child->beautify_loops( phase );
2113   if( _next  ) result |= _next ->beautify_loops( phase );
2114   return result;
2115 }
2116 
2117 //------------------------------allpaths_check_safepts----------------------------
2118 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
2119 // encountered.  Helper for check_safepts.
2120 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
2121   assert(stack.size() == 0, "empty stack");
2122   stack.push(_tail);
2123   visited.Clear();
2124   visited.set(_tail->_idx);
2125   while (stack.size() > 0) {
2126     Node* n = stack.pop();
2127     if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
2128       // Terminate this path
2129     } else if (n->Opcode() == Op_SafePoint) {
2130       if (_phase->get_loop(n) != this) {
2131         if (_required_safept == NULL) _required_safept = new Node_List();
2132         _required_safept->push(n);  // save the one closest to the tail
2133       }
2134       // Terminate this path
2135     } else {
2136       uint start = n->is_Region() ? 1 : 0;
2137       uint end   = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
2138       for (uint i = start; i < end; i++) {
2139         Node* in = n->in(i);
2140         assert(in->is_CFG(), "must be");
2141         if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
2142           stack.push(in);
2143         }
2144       }
2145     }
2146   }
2147 }
2148 
2149 //------------------------------check_safepts----------------------------
2150 // Given dominators, try to find loops with calls that must always be
2151 // executed (call dominates loop tail).  These loops do not need non-call
2152 // safepoints (ncsfpt).
2153 //
2154 // A complication is that a safepoint in a inner loop may be needed
2155 // by an outer loop. In the following, the inner loop sees it has a
2156 // call (block 3) on every path from the head (block 2) to the
2157 // backedge (arc 3->2).  So it deletes the ncsfpt (non-call safepoint)
2158 // in block 2, _but_ this leaves the outer loop without a safepoint.
2159 //
2160 //          entry  0
2161 //                 |
2162 //                 v
2163 // outer 1,2    +->1
2164 //              |  |
2165 //              |  v
2166 //              |  2<---+  ncsfpt in 2
2167 //              |_/|\   |
2168 //                 | v  |
2169 // inner 2,3      /  3  |  call in 3
2170 //               /   |  |
2171 //              v    +--+
2172 //        exit  4
2173 //
2174 //
2175 // This method creates a list (_required_safept) of ncsfpt nodes that must
2176 // be protected is created for each loop. When a ncsfpt maybe deleted, it
2177 // is first looked for in the lists for the outer loops of the current loop.
2178 //
2179 // The insights into the problem:
2180 //  A) counted loops are okay
2181 //  B) innermost loops are okay (only an inner loop can delete
2182 //     a ncsfpt needed by an outer loop)
2183 //  C) a loop is immune from an inner loop deleting a safepoint
2184 //     if the loop has a call on the idom-path
2185 //  D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
2186 //     idom-path that is not in a nested loop
2187 //  E) otherwise, an ncsfpt on the idom-path that is nested in an inner
2188 //     loop needs to be prevented from deletion by an inner loop
2189 //
2190 // There are two analyses:
2191 //  1) The first, and cheaper one, scans the loop body from
2192 //     tail to head following the idom (immediate dominator)
2193 //     chain, looking for the cases (C,D,E) above.
2194 //     Since inner loops are scanned before outer loops, there is summary
2195 //     information about inner loops.  Inner loops can be skipped over
2196 //     when the tail of an inner loop is encountered.
2197 //
2198 //  2) The second, invoked if the first fails to find a call or ncsfpt on
2199 //     the idom path (which is rare), scans all predecessor control paths
2200 //     from the tail to the head, terminating a path when a call or sfpt
2201 //     is encountered, to find the ncsfpt's that are closest to the tail.
2202 //
2203 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
2204   // Bottom up traversal
2205   IdealLoopTree* ch = _child;
2206   if (_child) _child->check_safepts(visited, stack);
2207   if (_next)  _next ->check_safepts(visited, stack);
2208 
2209   if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
2210     bool  has_call         = false; // call on dom-path
2211     bool  has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
2212     Node* nonlocal_ncsfpt  = NULL;  // ncsfpt on dom-path at a deeper depth
2213     // Scan the dom-path nodes from tail to head
2214     for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
2215       if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
2216         has_call = true;
2217         _has_sfpt = 1;          // Then no need for a safept!
2218         break;
2219       } else if (n->Opcode() == Op_SafePoint) {
2220         if (_phase->get_loop(n) == this) {
2221           has_local_ncsfpt = true;
2222           break;
2223         }
2224         if (nonlocal_ncsfpt == NULL) {
2225           nonlocal_ncsfpt = n; // save the one closest to the tail
2226         }
2227       } else {
2228         IdealLoopTree* nlpt = _phase->get_loop(n);
2229         if (this != nlpt) {
2230           // If at an inner loop tail, see if the inner loop has already
2231           // recorded seeing a call on the dom-path (and stop.)  If not,
2232           // jump to the head of the inner loop.
2233           assert(is_member(nlpt), "nested loop");
2234           Node* tail = nlpt->_tail;
2235           if (tail->in(0)->is_If()) tail = tail->in(0);
2236           if (n == tail) {
2237             // If inner loop has call on dom-path, so does outer loop
2238             if (nlpt->_has_sfpt) {
2239               has_call = true;
2240               _has_sfpt = 1;
2241               break;
2242             }
2243             // Skip to head of inner loop
2244             assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
2245             n = nlpt->_head;
2246           }
2247         }
2248       }
2249     }
2250     // Record safept's that this loop needs preserved when an
2251     // inner loop attempts to delete it's safepoints.
2252     if (_child != NULL && !has_call && !has_local_ncsfpt) {
2253       if (nonlocal_ncsfpt != NULL) {
2254         if (_required_safept == NULL) _required_safept = new Node_List();
2255         _required_safept->push(nonlocal_ncsfpt);
2256       } else {
2257         // Failed to find a suitable safept on the dom-path.  Now use
2258         // an all paths walk from tail to head, looking for safepoints to preserve.
2259         allpaths_check_safepts(visited, stack);
2260       }
2261     }
2262   }
2263 }
2264 
2265 //---------------------------is_deleteable_safept----------------------------
2266 // Is safept not required by an outer loop?
2267 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
2268   assert(sfpt->Opcode() == Op_SafePoint, "");
2269   IdealLoopTree* lp = get_loop(sfpt)->_parent;
2270   while (lp != NULL) {
2271     Node_List* sfpts = lp->_required_safept;
2272     if (sfpts != NULL) {
2273       for (uint i = 0; i < sfpts->size(); i++) {
2274         if (sfpt == sfpts->at(i))
2275           return false;
2276       }
2277     }
2278     lp = lp->_parent;
2279   }
2280   return true;
2281 }
2282 
2283 //---------------------------replace_parallel_iv-------------------------------
2284 // Replace parallel induction variable (parallel to trip counter)
2285 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
2286   assert(loop->_head->is_CountedLoop(), "");
2287   CountedLoopNode *cl = loop->_head->as_CountedLoop();
2288   if (!cl->is_valid_counted_loop())
2289     return;         // skip malformed counted loop
2290   Node *incr = cl->incr();
2291   if (incr == NULL)
2292     return;         // Dead loop?
2293   Node *init = cl->init_trip();
2294   Node *phi  = cl->phi();
2295   int stride_con = cl->stride_con();
2296 
2297   // Visit all children, looking for Phis
2298   for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
2299     Node *out = cl->out(i);
2300     // Look for other phis (secondary IVs). Skip dead ones
2301     if (!out->is_Phi() || out == phi || !has_node(out))
2302       continue;
2303     PhiNode* phi2 = out->as_Phi();
2304     Node *incr2 = phi2->in( LoopNode::LoopBackControl );
2305     // Look for induction variables of the form:  X += constant
2306     if (phi2->region() != loop->_head ||
2307         incr2->req() != 3 ||
2308         incr2->in(1) != phi2 ||
2309         incr2 == incr ||
2310         incr2->Opcode() != Op_AddI ||
2311         !incr2->in(2)->is_Con())
2312       continue;
2313 
2314     // Check for parallel induction variable (parallel to trip counter)
2315     // via an affine function.  In particular, count-down loops with
2316     // count-up array indices are common. We only RCE references off
2317     // the trip-counter, so we need to convert all these to trip-counter
2318     // expressions.
2319     Node *init2 = phi2->in( LoopNode::EntryControl );
2320     int stride_con2 = incr2->in(2)->get_int();
2321 
2322     // The ratio of the two strides cannot be represented as an int
2323     // if stride_con2 is min_int and stride_con is -1.
2324     if (stride_con2 == min_jint && stride_con == -1) {
2325       continue;
2326     }
2327 
2328     // The general case here gets a little tricky.  We want to find the
2329     // GCD of all possible parallel IV's and make a new IV using this
2330     // GCD for the loop.  Then all possible IVs are simple multiples of
2331     // the GCD.  In practice, this will cover very few extra loops.
2332     // Instead we require 'stride_con2' to be a multiple of 'stride_con',
2333     // where +/-1 is the common case, but other integer multiples are
2334     // also easy to handle.
2335     int ratio_con = stride_con2/stride_con;
2336 
2337     if ((ratio_con * stride_con) == stride_con2) { // Check for exact
2338 #ifndef PRODUCT
2339       if (TraceLoopOpts) {
2340         tty->print("Parallel IV: %d ", phi2->_idx);
2341         loop->dump_head();
2342       }
2343 #endif
2344       // Convert to using the trip counter.  The parallel induction
2345       // variable differs from the trip counter by a loop-invariant
2346       // amount, the difference between their respective initial values.
2347       // It is scaled by the 'ratio_con'.
2348       Node* ratio = _igvn.intcon(ratio_con);
2349       set_ctrl(ratio, C->root());
2350       Node* ratio_init = new MulINode(init, ratio);
2351       _igvn.register_new_node_with_optimizer(ratio_init, init);
2352       set_early_ctrl(ratio_init);
2353       Node* diff = new SubINode(init2, ratio_init);
2354       _igvn.register_new_node_with_optimizer(diff, init2);
2355       set_early_ctrl(diff);
2356       Node* ratio_idx = new MulINode(phi, ratio);
2357       _igvn.register_new_node_with_optimizer(ratio_idx, phi);
2358       set_ctrl(ratio_idx, cl);
2359       Node* add = new AddINode(ratio_idx, diff);
2360       _igvn.register_new_node_with_optimizer(add);
2361       set_ctrl(add, cl);
2362       _igvn.replace_node( phi2, add );
2363       // Sometimes an induction variable is unused
2364       if (add->outcnt() == 0) {
2365         _igvn.remove_dead_node(add);
2366       }
2367       --i; // deleted this phi; rescan starting with next position
2368       continue;
2369     }
2370   }
2371 }
2372 
2373 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
2374   Node* keep = NULL;
2375   if (keep_one) {
2376     // Look for a safepoint on the idom-path.
2377     for (Node* i = tail(); i != _head; i = phase->idom(i)) {
2378       if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
2379         keep = i;
2380         break; // Found one
2381       }
2382     }
2383   }
2384 
2385   // Don't remove any safepoints if it is requested to keep a single safepoint and
2386   // no safepoint was found on idom-path. It is not safe to remove any safepoint
2387   // in this case since there's no safepoint dominating all paths in the loop body.
2388   bool prune = !keep_one || keep != NULL;
2389 
2390   // Delete other safepoints in this loop.
2391   Node_List* sfpts = _safepts;
2392   if (prune && sfpts != NULL) {
2393     assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint");
2394     for (uint i = 0; i < sfpts->size(); i++) {
2395       Node* n = sfpts->at(i);
2396       assert(phase->get_loop(n) == this, "");
2397       if (n != keep && phase->is_deleteable_safept(n)) {
2398         phase->lazy_replace(n, n->in(TypeFunc::Control));
2399       }
2400     }
2401   }
2402 }
2403 
2404 //------------------------------counted_loop-----------------------------------
2405 // Convert to counted loops where possible
2406 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
2407 
2408   // For grins, set the inner-loop flag here
2409   if (!_child) {
2410     if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
2411   }
2412 
2413   IdealLoopTree* loop = this;
2414   if (_head->is_CountedLoop() ||
2415       phase->is_counted_loop(_head, loop)) {
2416 
2417     if (LoopStripMiningIter == 0 || (LoopStripMiningIter > 1 && _child == NULL)) {
2418       // Indicate we do not need a safepoint here
2419       _has_sfpt = 1;
2420     }
2421 
2422     // Remove safepoints
2423     bool keep_one_sfpt = !(_has_call || _has_sfpt);
2424     remove_safepoints(phase, keep_one_sfpt);
2425 
2426     // Look for induction variables
2427     phase->replace_parallel_iv(this);
2428 
2429   } else if (_parent != NULL && !_irreducible) {
2430     // Not a counted loop. Keep one safepoint.
2431     bool keep_one_sfpt = true;
2432     remove_safepoints(phase, keep_one_sfpt);
2433   }
2434 
2435   // Recursively
2436   assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?");
2437   assert(loop->_child != this || (loop->_child->_child == NULL && loop->_child->_next == NULL), "would miss some loops");
2438   if (loop->_child && loop->_child != this) loop->_child->counted_loop(phase);
2439   if (loop->_next)  loop->_next ->counted_loop(phase);
2440 }
2441 
2442 
2443 // The Estimated Loop Clone Size:
2444 //   CloneFactor * (~112% * BodySize + BC) + CC + FanOutTerm,
2445 // where  BC and  CC are  totally ad-hoc/magic  "body" and "clone" constants,
2446 // respectively, used to ensure that the node usage estimates made are on the
2447 // safe side, for the most part. The FanOutTerm is an attempt to estimate the
2448 // possible additional/excessive nodes generated due to data and control flow
2449 // merging, for edges reaching outside the loop.
2450 uint IdealLoopTree::est_loop_clone_sz(uint factor) const {
2451 
2452   precond(0 < factor && factor < 16);
2453 
2454   uint const bc = 13;
2455   uint const cc = 17;
2456   uint const sz = _body.size() + (_body.size() + 7) / 8;
2457   uint estimate = factor * (sz + bc) + cc;
2458 
2459   assert((estimate - cc) / factor == sz + bc, "overflow");
2460 
2461   uint ctrl_edge_out_cnt = 0;
2462   uint data_edge_out_cnt = 0;
2463 
2464   for (uint i = 0; i < _body.size(); i++) {
2465     Node* node = _body.at(i);
2466     uint outcnt = node->outcnt();
2467 
2468     for (uint k = 0; k < outcnt; k++) {
2469       Node* out = node->raw_out(k);
2470 
2471       if (out->is_CFG()) {
2472         if (!is_member(_phase->get_loop(out))) {
2473           ctrl_edge_out_cnt++;
2474         }
2475       } else {
2476         Node* ctrl = _phase->get_ctrl(out);
2477         assert(ctrl->is_CFG(), "must be");
2478         if (!is_member(_phase->get_loop(ctrl))) {
2479           data_edge_out_cnt++;
2480         }
2481       }
2482     }
2483   }
2484   // Add data (x1.5) and control (x1.0) count to estimate iff both are > 0.
2485   if (ctrl_edge_out_cnt > 0 && data_edge_out_cnt > 0) {
2486     estimate += ctrl_edge_out_cnt + data_edge_out_cnt + data_edge_out_cnt / 2;
2487   }
2488 
2489   return estimate;
2490 }
2491 
2492 #ifndef PRODUCT
2493 //------------------------------dump_head--------------------------------------
2494 // Dump 1 liner for loop header info
2495 void IdealLoopTree::dump_head() const {
2496   for (uint i = 0; i < _nest; i++) {
2497     tty->print("  ");
2498   }
2499   tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
2500   if (_irreducible) tty->print(" IRREDUCIBLE");
2501   Node* entry = _head->is_Loop() ? _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl) : _head->in(LoopNode::EntryControl);
2502   Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
2503   if (predicate != NULL ) {
2504     tty->print(" limit_check");
2505     entry = PhaseIdealLoop::skip_loop_predicates(entry);
2506   }
2507   if (UseLoopPredicate) {
2508     entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
2509     if (entry != NULL) {
2510       tty->print(" predicated");
2511       entry = PhaseIdealLoop::skip_loop_predicates(entry);
2512     }
2513   }
2514   if (UseProfiledLoopPredicate) {
2515     entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate);
2516     if (entry != NULL) {
2517       tty->print(" profile_predicated");
2518     }
2519   }
2520   if (_head->is_CountedLoop()) {
2521     CountedLoopNode *cl = _head->as_CountedLoop();
2522     tty->print(" counted");
2523 
2524     Node* init_n = cl->init_trip();
2525     if (init_n  != NULL &&  init_n->is_Con())
2526       tty->print(" [%d,", cl->init_trip()->get_int());
2527     else
2528       tty->print(" [int,");
2529     Node* limit_n = cl->limit();
2530     if (limit_n  != NULL &&  limit_n->is_Con())
2531       tty->print("%d),", cl->limit()->get_int());
2532     else
2533       tty->print("int),");
2534     int stride_con  = cl->stride_con();
2535     if (stride_con > 0) tty->print("+");
2536     tty->print("%d", stride_con);
2537 
2538     tty->print(" (%0.f iters) ", cl->profile_trip_cnt());
2539 
2540     if (cl->is_pre_loop ()) tty->print(" pre" );
2541     if (cl->is_main_loop()) tty->print(" main");
2542     if (cl->is_post_loop()) tty->print(" post");
2543     if (cl->is_vectorized_loop()) tty->print(" vector");
2544     if (cl->range_checks_present()) tty->print(" rc ");
2545     if (cl->is_multiversioned()) tty->print(" multi ");
2546   }
2547   if (_has_call) tty->print(" has_call");
2548   if (_has_sfpt) tty->print(" has_sfpt");
2549   if (_rce_candidate) tty->print(" rce");
2550   if (_safepts != NULL && _safepts->size() > 0) {
2551     tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
2552   }
2553   if (_required_safept != NULL && _required_safept->size() > 0) {
2554     tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
2555   }
2556   if (Verbose) {
2557     tty->print(" body={"); _body.dump_simple(); tty->print(" }");
2558   }
2559   if (_head->is_Loop() && _head->as_Loop()->is_strip_mined()) {
2560     tty->print(" strip_mined");
2561   }
2562   tty->cr();
2563 }
2564 
2565 //------------------------------dump-------------------------------------------
2566 // Dump loops by loop tree
2567 void IdealLoopTree::dump() const {
2568   dump_head();
2569   if (_child) _child->dump();
2570   if (_next)  _next ->dump();
2571 }
2572 
2573 #endif
2574 
2575 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
2576   if (loop == root) {
2577     if (loop->_child != NULL) {
2578       log->begin_head("loop_tree");
2579       log->end_head();
2580       if( loop->_child ) log_loop_tree(root, loop->_child, log);
2581       log->tail("loop_tree");
2582       assert(loop->_next == NULL, "what?");
2583     }
2584   } else {
2585     Node* head = loop->_head;
2586     log->begin_head("loop");
2587     log->print(" idx='%d' ", head->_idx);
2588     if (loop->_irreducible) log->print("irreducible='1' ");
2589     if (head->is_Loop()) {
2590       if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
2591       if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
2592     }
2593     if (head->is_CountedLoop()) {
2594       CountedLoopNode* cl = head->as_CountedLoop();
2595       if (cl->is_pre_loop())  log->print("pre_loop='%d' ",  cl->main_idx());
2596       if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
2597       if (cl->is_post_loop()) log->print("post_loop='%d' ",  cl->main_idx());
2598     }
2599     log->end_head();
2600     if( loop->_child ) log_loop_tree(root, loop->_child, log);
2601     log->tail("loop");
2602     if( loop->_next  ) log_loop_tree(root, loop->_next, log);
2603   }
2604 }
2605 
2606 //---------------------collect_potentially_useful_predicates-----------------------
2607 // Helper function to collect potentially useful predicates to prevent them from
2608 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
2609 void PhaseIdealLoop::collect_potentially_useful_predicates(
2610                          IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
2611   if (loop->_child) { // child
2612     collect_potentially_useful_predicates(loop->_child, useful_predicates);
2613   }
2614 
2615   // self (only loops that we can apply loop predication may use their predicates)
2616   if (loop->_head->is_Loop() &&
2617       !loop->_irreducible    &&
2618       !loop->tail()->is_top()) {
2619     LoopNode* lpn = loop->_head->as_Loop();
2620     Node* entry = lpn->in(LoopNode::EntryControl);
2621     Node* predicate_proj = find_predicate(entry); // loop_limit_check first
2622     if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
2623       assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
2624       useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2625       entry = skip_loop_predicates(entry);
2626     }
2627     predicate_proj = find_predicate(entry); // Predicate
2628     if (predicate_proj != NULL ) {
2629       useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2630       entry = skip_loop_predicates(entry);
2631     }
2632     if (UseProfiledLoopPredicate) {
2633       predicate_proj = find_predicate(entry); // Predicate
2634       if (predicate_proj != NULL ) {
2635         useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2636       }
2637     }
2638   }
2639 
2640   if (loop->_next) { // sibling
2641     collect_potentially_useful_predicates(loop->_next, useful_predicates);
2642   }
2643 }
2644 
2645 //------------------------eliminate_useless_predicates-----------------------------
2646 // Eliminate all inserted predicates if they could not be used by loop predication.
2647 // Note: it will also eliminates loop limits check predicate since it also uses
2648 // Opaque1 node (see Parse::add_predicate()).
2649 void PhaseIdealLoop::eliminate_useless_predicates() {
2650   if (C->predicate_count() == 0)
2651     return; // no predicate left
2652 
2653   Unique_Node_List useful_predicates; // to store useful predicates
2654   if (C->has_loops()) {
2655     collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
2656   }
2657 
2658   for (int i = C->predicate_count(); i > 0; i--) {
2659      Node * n = C->predicate_opaque1_node(i-1);
2660      assert(n->Opcode() == Op_Opaque1, "must be");
2661      if (!useful_predicates.member(n)) { // not in the useful list
2662        _igvn.replace_node(n, n->in(1));
2663      }
2664   }
2665 }
2666 
2667 //------------------------process_expensive_nodes-----------------------------
2668 // Expensive nodes have their control input set to prevent the GVN
2669 // from commoning them and as a result forcing the resulting node to
2670 // be in a more frequent path. Use CFG information here, to change the
2671 // control inputs so that some expensive nodes can be commoned while
2672 // not executed more frequently.
2673 bool PhaseIdealLoop::process_expensive_nodes() {
2674   assert(OptimizeExpensiveOps, "optimization off?");
2675 
2676   // Sort nodes to bring similar nodes together
2677   C->sort_expensive_nodes();
2678 
2679   bool progress = false;
2680 
2681   for (int i = 0; i < C->expensive_count(); ) {
2682     Node* n = C->expensive_node(i);
2683     int start = i;
2684     // Find nodes similar to n
2685     i++;
2686     for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
2687     int end = i;
2688     // And compare them two by two
2689     for (int j = start; j < end; j++) {
2690       Node* n1 = C->expensive_node(j);
2691       if (is_node_unreachable(n1)) {
2692         continue;
2693       }
2694       for (int k = j+1; k < end; k++) {
2695         Node* n2 = C->expensive_node(k);
2696         if (is_node_unreachable(n2)) {
2697           continue;
2698         }
2699 
2700         assert(n1 != n2, "should be pair of nodes");
2701 
2702         Node* c1 = n1->in(0);
2703         Node* c2 = n2->in(0);
2704 
2705         Node* parent_c1 = c1;
2706         Node* parent_c2 = c2;
2707 
2708         // The call to get_early_ctrl_for_expensive() moves the
2709         // expensive nodes up but stops at loops that are in a if
2710         // branch. See whether we can exit the loop and move above the
2711         // If.
2712         if (c1->is_Loop()) {
2713           parent_c1 = c1->in(1);
2714         }
2715         if (c2->is_Loop()) {
2716           parent_c2 = c2->in(1);
2717         }
2718 
2719         if (parent_c1 == parent_c2) {
2720           _igvn._worklist.push(n1);
2721           _igvn._worklist.push(n2);
2722           continue;
2723         }
2724 
2725         // Look for identical expensive node up the dominator chain.
2726         if (is_dominator(c1, c2)) {
2727           c2 = c1;
2728         } else if (is_dominator(c2, c1)) {
2729           c1 = c2;
2730         } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
2731                    parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
2732           // Both branches have the same expensive node so move it up
2733           // before the if.
2734           c1 = c2 = idom(parent_c1->in(0));
2735         }
2736         // Do the actual moves
2737         if (n1->in(0) != c1) {
2738           _igvn.hash_delete(n1);
2739           n1->set_req(0, c1);
2740           _igvn.hash_insert(n1);
2741           _igvn._worklist.push(n1);
2742           progress = true;
2743         }
2744         if (n2->in(0) != c2) {
2745           _igvn.hash_delete(n2);
2746           n2->set_req(0, c2);
2747           _igvn.hash_insert(n2);
2748           _igvn._worklist.push(n2);
2749           progress = true;
2750         }
2751       }
2752     }
2753   }
2754 
2755   return progress;
2756 }
2757 
2758 
2759 //=============================================================================
2760 //----------------------------build_and_optimize-------------------------------
2761 // Create a PhaseLoop.  Build the ideal Loop tree.  Map each Ideal Node to
2762 // its corresponding LoopNode.  If 'optimize' is true, do some loop cleanups.
2763 void PhaseIdealLoop::build_and_optimize(LoopOptsMode mode) {
2764   bool do_split_ifs = (mode == LoopOptsDefault || mode == LoopOptsLastRound);
2765   bool skip_loop_opts = (mode == LoopOptsNone);
2766 
2767   int old_progress = C->major_progress();
2768   uint orig_worklist_size = _igvn._worklist.size();
2769 
2770   // Reset major-progress flag for the driver's heuristics
2771   C->clear_major_progress();
2772 
2773 #ifndef PRODUCT
2774   // Capture for later assert
2775   uint unique = C->unique();
2776   _loop_invokes++;
2777   _loop_work += unique;
2778 #endif
2779 
2780   // True if the method has at least 1 irreducible loop
2781   _has_irreducible_loops = false;
2782 
2783   _created_loop_node = false;
2784 
2785   Arena *a = Thread::current()->resource_area();
2786   VectorSet visited(a);
2787   // Pre-grow the mapping from Nodes to IdealLoopTrees.
2788   _nodes.map(C->unique(), NULL);
2789   memset(_nodes.adr(), 0, wordSize * C->unique());
2790 
2791   // Pre-build the top-level outermost loop tree entry
2792   _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
2793   // Do not need a safepoint at the top level
2794   _ltree_root->_has_sfpt = 1;
2795 
2796   // Initialize Dominators.
2797   // Checked in clone_loop_predicate() during beautify_loops().
2798   _idom_size = 0;
2799   _idom      = NULL;
2800   _dom_depth = NULL;
2801   _dom_stk   = NULL;
2802 
2803   // Empty pre-order array
2804   allocate_preorders();
2805 
2806   // Build a loop tree on the fly.  Build a mapping from CFG nodes to
2807   // IdealLoopTree entries.  Data nodes are NOT walked.
2808   build_loop_tree();
2809   // Check for bailout, and return
2810   if (C->failing()) {
2811     return;
2812   }
2813 
2814   // No loops after all
2815   if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
2816 
2817   // There should always be an outer loop containing the Root and Return nodes.
2818   // If not, we have a degenerate empty program.  Bail out in this case.
2819   if (!has_node(C->root())) {
2820     if (!_verify_only) {
2821       C->clear_major_progress();
2822       C->record_method_not_compilable("empty program detected during loop optimization");
2823     }
2824     return;
2825   }
2826 
2827   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2828   // Nothing to do, so get out
2829   bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only &&
2830     !bs->is_gc_specific_loop_opts_pass(mode);
2831   bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
2832   bool strip_mined_loops_expanded = bs->strip_mined_loops_expanded(mode);
2833   if (stop_early && !do_expensive_nodes) {
2834     _igvn.optimize();           // Cleanup NeverBranches
2835     return;
2836   }
2837 
2838   // Set loop nesting depth
2839   _ltree_root->set_nest( 0 );
2840 
2841   // Split shared headers and insert loop landing pads.
2842   // Do not bother doing this on the Root loop of course.
2843   if( !_verify_me && !_verify_only && _ltree_root->_child ) {
2844     C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
2845     if( _ltree_root->_child->beautify_loops( this ) ) {
2846       // Re-build loop tree!
2847       _ltree_root->_child = NULL;
2848       _nodes.clear();
2849       reallocate_preorders();
2850       build_loop_tree();
2851       // Check for bailout, and return
2852       if (C->failing()) {
2853         return;
2854       }
2855       // Reset loop nesting depth
2856       _ltree_root->set_nest( 0 );
2857 
2858       C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
2859     }
2860   }
2861 
2862   // Build Dominators for elision of NULL checks & loop finding.
2863   // Since nodes do not have a slot for immediate dominator, make
2864   // a persistent side array for that info indexed on node->_idx.
2865   _idom_size = C->unique();
2866   _idom      = NEW_RESOURCE_ARRAY( Node*, _idom_size );
2867   _dom_depth = NEW_RESOURCE_ARRAY( uint,  _idom_size );
2868   _dom_stk   = NULL; // Allocated on demand in recompute_dom_depth
2869   memset( _dom_depth, 0, _idom_size * sizeof(uint) );
2870 
2871   Dominators();
2872 
2873   if (!_verify_only) {
2874     // As a side effect, Dominators removed any unreachable CFG paths
2875     // into RegionNodes.  It doesn't do this test against Root, so
2876     // we do it here.
2877     for( uint i = 1; i < C->root()->req(); i++ ) {
2878       if( !_nodes[C->root()->in(i)->_idx] ) {    // Dead path into Root?
2879         _igvn.delete_input_of(C->root(), i);
2880         i--;                      // Rerun same iteration on compressed edges
2881       }
2882     }
2883 
2884     // Given dominators, try to find inner loops with calls that must
2885     // always be executed (call dominates loop tail).  These loops do
2886     // not need a separate safepoint.
2887     Node_List cisstack(a);
2888     _ltree_root->check_safepts(visited, cisstack);
2889   }
2890 
2891   // Walk the DATA nodes and place into loops.  Find earliest control
2892   // node.  For CFG nodes, the _nodes array starts out and remains
2893   // holding the associated IdealLoopTree pointer.  For DATA nodes, the
2894   // _nodes array holds the earliest legal controlling CFG node.
2895 
2896   // Allocate stack with enough space to avoid frequent realloc
2897   int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
2898   Node_Stack nstack( a, stack_size );
2899 
2900   visited.Clear();
2901   Node_List worklist(a);
2902   // Don't need C->root() on worklist since
2903   // it will be processed among C->top() inputs
2904   worklist.push( C->top() );
2905   visited.set( C->top()->_idx ); // Set C->top() as visited now
2906   build_loop_early( visited, worklist, nstack );
2907 
2908   // Given early legal placement, try finding counted loops.  This placement
2909   // is good enough to discover most loop invariants.
2910   if (!_verify_me && !_verify_only && !strip_mined_loops_expanded) {
2911     _ltree_root->counted_loop( this );
2912   }
2913 
2914   // Find latest loop placement.  Find ideal loop placement.
2915   visited.Clear();
2916   init_dom_lca_tags();
2917   // Need C->root() on worklist when processing outs
2918   worklist.push( C->root() );
2919   NOT_PRODUCT( C->verify_graph_edges(); )
2920   worklist.push( C->top() );
2921   build_loop_late( visited, worklist, nstack );
2922 
2923   if (_verify_only) {
2924     // restore major progress flag
2925     for (int i = 0; i < old_progress; i++)
2926       C->set_major_progress();
2927     assert(C->unique() == unique, "verification mode made Nodes? ? ?");
2928     assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
2929     return;
2930   }
2931 
2932   // clear out the dead code after build_loop_late
2933   while (_deadlist.size()) {
2934     _igvn.remove_globally_dead_node(_deadlist.pop());
2935   }
2936 
2937   if (stop_early) {
2938     assert(do_expensive_nodes, "why are we here?");
2939     if (process_expensive_nodes()) {
2940       // If we made some progress when processing expensive nodes then
2941       // the IGVN may modify the graph in a way that will allow us to
2942       // make some more progress: we need to try processing expensive
2943       // nodes again.
2944       C->set_major_progress();
2945     }
2946     _igvn.optimize();
2947     return;
2948   }
2949 
2950   // Some parser-inserted loop predicates could never be used by loop
2951   // predication or they were moved away from loop during some optimizations.
2952   // For example, peeling. Eliminate them before next loop optimizations.
2953   eliminate_useless_predicates();
2954 
2955 #ifndef PRODUCT
2956   C->verify_graph_edges();
2957   if (_verify_me) {             // Nested verify pass?
2958     // Check to see if the verify mode is broken
2959     assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
2960     return;
2961   }
2962   if (VerifyLoopOptimizations) verify();
2963   if (TraceLoopOpts && C->has_loops()) {
2964     _ltree_root->dump();
2965   }
2966 #endif
2967 
2968   if (skip_loop_opts) {
2969     // restore major progress flag
2970     for (int i = 0; i < old_progress; i++) {
2971       C->set_major_progress();
2972     }
2973 
2974     // Cleanup any modified bits
2975     _igvn.optimize();
2976 
2977     if (C->log() != NULL) {
2978       log_loop_tree(_ltree_root, _ltree_root, C->log());
2979     }
2980     return;
2981   }
2982 
2983   if (bs->optimize_loops(this, mode, visited, nstack, worklist)) {
2984     _igvn.optimize();
2985     if (C->log() != NULL) {
2986       log_loop_tree(_ltree_root, _ltree_root, C->log());
2987     }
2988     return;
2989   }
2990 
2991   if (ReassociateInvariants) {

2992     // Reassociate invariants and prep for split_thru_phi
2993     for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2994       IdealLoopTree* lpt = iter.current();
2995       bool is_counted = lpt->is_counted();
2996       if (!is_counted || !lpt->is_innermost()) continue;
2997 
2998       // check for vectorized loops, any reassociation of invariants was already done
2999       if (is_counted && lpt->_head->as_CountedLoop()->is_unroll_only()) {
3000         continue;
3001       } else {
3002         AutoNodeBudget node_budget(this);
3003         lpt->reassociate_invariants(this);
3004       }
3005       // Because RCE opportunities can be masked by split_thru_phi,
3006       // look for RCE candidates and inhibit split_thru_phi
3007       // on just their loop-phi's for this pass of loop opts
3008       if (SplitIfBlocks && do_split_ifs) {
3009         AutoNodeBudget node_budget(this, AutoNodeBudget::NO_BUDGET_CHECK);
3010         if (lpt->policy_range_check(this)) {
3011           lpt->_rce_candidate = 1; // = true
3012         }
3013       }
3014     }
3015   }
3016 
3017   // Check for aggressive application of split-if and other transforms
3018   // that require basic-block info (like cloning through Phi's)
3019   if( SplitIfBlocks && do_split_ifs ) {
3020     visited.Clear();
3021     split_if_with_blocks( visited, nstack, mode == LoopOptsLastRound );
3022     NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
3023     if (mode == LoopOptsLastRound) {
3024       C->set_major_progress();
3025     }
3026   }
3027 
3028   if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
3029     C->set_major_progress();
3030   }
3031 
3032   // Perform loop predication before iteration splitting
3033   if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
3034     _ltree_root->_child->loop_predication(this);
3035   }
3036 
3037   if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
3038     if (do_intrinsify_fill()) {
3039       C->set_major_progress();
3040     }
3041   }
3042 
3043   // Perform iteration-splitting on inner loops.  Split iterations to avoid
3044   // range checks or one-shot null checks.
3045 
3046   // If split-if's didn't hack the graph too bad (no CFG changes)
3047   // then do loop opts.
3048   if (C->has_loops() && !C->major_progress()) {
3049     memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
3050     _ltree_root->_child->iteration_split( this, worklist );
3051     // No verify after peeling!  GCM has hoisted code out of the loop.
3052     // After peeling, the hoisted code could sink inside the peeled area.
3053     // The peeling code does not try to recompute the best location for
3054     // all the code before the peeled area, so the verify pass will always
3055     // complain about it.
3056   }
3057   // Do verify graph edges in any case
3058   NOT_PRODUCT( C->verify_graph_edges(); );
3059 
3060   if (!do_split_ifs) {
3061     // We saw major progress in Split-If to get here.  We forced a
3062     // pass with unrolling and not split-if, however more split-if's
3063     // might make progress.  If the unrolling didn't make progress
3064     // then the major-progress flag got cleared and we won't try
3065     // another round of Split-If.  In particular the ever-common
3066     // instance-of/check-cast pattern requires at least 2 rounds of
3067     // Split-If to clear out.
3068     C->set_major_progress();
3069   }
3070 
3071   // Repeat loop optimizations if new loops were seen
3072   if (created_loop_node()) {
3073     C->set_major_progress();
3074   }
3075 
3076   // Keep loop predicates and perform optimizations with them
3077   // until no more loop optimizations could be done.
3078   // After that switch predicates off and do more loop optimizations.
3079   if (!C->major_progress() && (C->predicate_count() > 0)) {
3080      C->cleanup_loop_predicates(_igvn);
3081      if (TraceLoopOpts) {
3082        tty->print_cr("PredicatesOff");
3083      }
3084      C->set_major_progress();
3085   }
3086 
3087   // Convert scalar to superword operations at the end of all loop opts.
3088   if (UseSuperWord && C->has_loops() && !C->major_progress()) {
3089     // SuperWord transform
3090     SuperWord sw(this);
3091     for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
3092       IdealLoopTree* lpt = iter.current();
3093       if (lpt->is_counted()) {
3094         CountedLoopNode *cl = lpt->_head->as_CountedLoop();
3095 
3096         if (PostLoopMultiversioning && cl->is_rce_post_loop() && !cl->is_vectorized_loop()) {
3097           // Check that the rce'd post loop is encountered first, multiversion after all
3098           // major main loop optimization are concluded
3099           if (!C->major_progress()) {
3100             IdealLoopTree *lpt_next = lpt->_next;
3101             if (lpt_next && lpt_next->is_counted()) {
3102               CountedLoopNode *cl = lpt_next->_head->as_CountedLoop();
3103               has_range_checks(lpt_next);
3104               if (cl->is_post_loop() && cl->range_checks_present()) {
3105                 if (!cl->is_multiversioned()) {
3106                   if (multi_version_post_loops(lpt, lpt_next) == false) {
3107                     // Cause the rce loop to be optimized away if we fail
3108                     cl->mark_is_multiversioned();
3109                     cl->set_slp_max_unroll(0);
3110                     poison_rce_post_loop(lpt);
3111                   }
3112                 }
3113               }
3114             }
3115             sw.transform_loop(lpt, true);
3116           }
3117         } else if (cl->is_main_loop()) {
3118           sw.transform_loop(lpt, true);
3119         }
3120       }
3121     }
3122   }
3123 
3124   // Cleanup any modified bits
3125   _igvn.optimize();
3126 
3127   // disable assert until issue with split_flow_path is resolved (6742111)
3128   // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
3129   //        "shouldn't introduce irreducible loops");
3130 
3131   if (C->log() != NULL) {
3132     log_loop_tree(_ltree_root, _ltree_root, C->log());
3133   }
3134 }
3135 
3136 #ifndef PRODUCT
3137 //------------------------------print_statistics-------------------------------
3138 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
3139 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
3140 void PhaseIdealLoop::print_statistics() {
3141   tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
3142 }
3143 
3144 //------------------------------verify-----------------------------------------
3145 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
3146 static int fail;                // debug only, so its multi-thread dont care
3147 void PhaseIdealLoop::verify() const {
3148   int old_progress = C->major_progress();
3149   ResourceMark rm;
3150   PhaseIdealLoop loop_verify( _igvn, this );
3151   VectorSet visited(Thread::current()->resource_area());
3152 
3153   fail = 0;
3154   verify_compare( C->root(), &loop_verify, visited );
3155   assert( fail == 0, "verify loops failed" );
3156   // Verify loop structure is the same
3157   _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
3158   // Reset major-progress.  It was cleared by creating a verify version of
3159   // PhaseIdealLoop.
3160   for( int i=0; i<old_progress; i++ )
3161     C->set_major_progress();
3162 }
3163 
3164 //------------------------------verify_compare---------------------------------
3165 // Make sure me and the given PhaseIdealLoop agree on key data structures
3166 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
3167   if( !n ) return;
3168   if( visited.test_set( n->_idx ) ) return;
3169   if( !_nodes[n->_idx] ) {      // Unreachable
3170     assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
3171     return;
3172   }
3173 
3174   uint i;
3175   for( i = 0; i < n->req(); i++ )
3176     verify_compare( n->in(i), loop_verify, visited );
3177 
3178   // Check the '_nodes' block/loop structure
3179   i = n->_idx;
3180   if( has_ctrl(n) ) {           // We have control; verify has loop or ctrl
3181     if( _nodes[i] != loop_verify->_nodes[i] &&
3182         get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
3183       tty->print("Mismatched control setting for: ");
3184       n->dump();
3185       if( fail++ > 10 ) return;
3186       Node *c = get_ctrl_no_update(n);
3187       tty->print("We have it as: ");
3188       if( c->in(0) ) c->dump();
3189         else tty->print_cr("N%d",c->_idx);
3190       tty->print("Verify thinks: ");
3191       if( loop_verify->has_ctrl(n) )
3192         loop_verify->get_ctrl_no_update(n)->dump();
3193       else
3194         loop_verify->get_loop_idx(n)->dump();
3195       tty->cr();
3196     }
3197   } else {                    // We have a loop
3198     IdealLoopTree *us = get_loop_idx(n);
3199     if( loop_verify->has_ctrl(n) ) {
3200       tty->print("Mismatched loop setting for: ");
3201       n->dump();
3202       if( fail++ > 10 ) return;
3203       tty->print("We have it as: ");
3204       us->dump();
3205       tty->print("Verify thinks: ");
3206       loop_verify->get_ctrl_no_update(n)->dump();
3207       tty->cr();
3208     } else if (!C->major_progress()) {
3209       // Loop selection can be messed up if we did a major progress
3210       // operation, like split-if.  Do not verify in that case.
3211       IdealLoopTree *them = loop_verify->get_loop_idx(n);
3212       if( us->_head != them->_head ||  us->_tail != them->_tail ) {
3213         tty->print("Unequals loops for: ");
3214         n->dump();
3215         if( fail++ > 10 ) return;
3216         tty->print("We have it as: ");
3217         us->dump();
3218         tty->print("Verify thinks: ");
3219         them->dump();
3220         tty->cr();
3221       }
3222     }
3223   }
3224 
3225   // Check for immediate dominators being equal
3226   if( i >= _idom_size ) {
3227     if( !n->is_CFG() ) return;
3228     tty->print("CFG Node with no idom: ");
3229     n->dump();
3230     return;
3231   }
3232   if( !n->is_CFG() ) return;
3233   if( n == C->root() ) return; // No IDOM here
3234 
3235   assert(n->_idx == i, "sanity");
3236   Node *id = idom_no_update(n);
3237   if( id != loop_verify->idom_no_update(n) ) {
3238     tty->print("Unequals idoms for: ");
3239     n->dump();
3240     if( fail++ > 10 ) return;
3241     tty->print("We have it as: ");
3242     id->dump();
3243     tty->print("Verify thinks: ");
3244     loop_verify->idom_no_update(n)->dump();
3245     tty->cr();
3246   }
3247 
3248 }
3249 
3250 //------------------------------verify_tree------------------------------------
3251 // Verify that tree structures match.  Because the CFG can change, siblings
3252 // within the loop tree can be reordered.  We attempt to deal with that by
3253 // reordering the verify's loop tree if possible.
3254 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
3255   assert( _parent == parent, "Badly formed loop tree" );
3256 
3257   // Siblings not in same order?  Attempt to re-order.
3258   if( _head != loop->_head ) {
3259     // Find _next pointer to update
3260     IdealLoopTree **pp = &loop->_parent->_child;
3261     while( *pp != loop )
3262       pp = &((*pp)->_next);
3263     // Find proper sibling to be next
3264     IdealLoopTree **nn = &loop->_next;
3265     while( (*nn) && (*nn)->_head != _head )
3266       nn = &((*nn)->_next);
3267 
3268     // Check for no match.
3269     if( !(*nn) ) {
3270       // Annoyingly, irreducible loops can pick different headers
3271       // after a major_progress operation, so the rest of the loop
3272       // tree cannot be matched.
3273       if (_irreducible && Compile::current()->major_progress())  return;
3274       assert( 0, "failed to match loop tree" );
3275     }
3276 
3277     // Move (*nn) to (*pp)
3278     IdealLoopTree *hit = *nn;
3279     *nn = hit->_next;
3280     hit->_next = loop;
3281     *pp = loop;
3282     loop = hit;
3283     // Now try again to verify
3284   }
3285 
3286   assert( _head  == loop->_head , "mismatched loop head" );
3287   Node *tail = _tail;           // Inline a non-updating version of
3288   while( !tail->in(0) )         // the 'tail()' call.
3289     tail = tail->in(1);
3290   assert( tail == loop->_tail, "mismatched loop tail" );
3291 
3292   // Counted loops that are guarded should be able to find their guards
3293   if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
3294     CountedLoopNode *cl = _head->as_CountedLoop();
3295     Node *init = cl->init_trip();
3296     Node *ctrl = cl->in(LoopNode::EntryControl);
3297     assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
3298     Node *iff  = ctrl->in(0);
3299     assert( iff->Opcode() == Op_If, "" );
3300     Node *bol  = iff->in(1);
3301     assert( bol->Opcode() == Op_Bool, "" );
3302     Node *cmp  = bol->in(1);
3303     assert( cmp->Opcode() == Op_CmpI, "" );
3304     Node *add  = cmp->in(1);
3305     Node *opaq;
3306     if( add->Opcode() == Op_Opaque1 ) {
3307       opaq = add;
3308     } else {
3309       assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
3310       assert( add == init, "" );
3311       opaq = cmp->in(2);
3312     }
3313     assert( opaq->Opcode() == Op_Opaque1, "" );
3314 
3315   }
3316 
3317   if (_child != NULL)  _child->verify_tree(loop->_child, this);
3318   if (_next  != NULL)  _next ->verify_tree(loop->_next,  parent);
3319   // Innermost loops need to verify loop bodies,
3320   // but only if no 'major_progress'
3321   int fail = 0;
3322   if (!Compile::current()->major_progress() && _child == NULL) {
3323     for( uint i = 0; i < _body.size(); i++ ) {
3324       Node *n = _body.at(i);
3325       if (n->outcnt() == 0)  continue; // Ignore dead
3326       uint j;
3327       for( j = 0; j < loop->_body.size(); j++ )
3328         if( loop->_body.at(j) == n )
3329           break;
3330       if( j == loop->_body.size() ) { // Not found in loop body
3331         // Last ditch effort to avoid assertion: Its possible that we
3332         // have some users (so outcnt not zero) but are still dead.
3333         // Try to find from root.
3334         if (Compile::current()->root()->find(n->_idx)) {
3335           fail++;
3336           tty->print("We have that verify does not: ");
3337           n->dump();
3338         }
3339       }
3340     }
3341     for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
3342       Node *n = loop->_body.at(i2);
3343       if (n->outcnt() == 0)  continue; // Ignore dead
3344       uint j;
3345       for( j = 0; j < _body.size(); j++ )
3346         if( _body.at(j) == n )
3347           break;
3348       if( j == _body.size() ) { // Not found in loop body
3349         // Last ditch effort to avoid assertion: Its possible that we
3350         // have some users (so outcnt not zero) but are still dead.
3351         // Try to find from root.
3352         if (Compile::current()->root()->find(n->_idx)) {
3353           fail++;
3354           tty->print("Verify has that we do not: ");
3355           n->dump();
3356         }
3357       }
3358     }
3359     assert( !fail, "loop body mismatch" );
3360   }
3361 }
3362 
3363 #endif
3364 
3365 //------------------------------set_idom---------------------------------------
3366 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
3367   uint idx = d->_idx;
3368   if (idx >= _idom_size) {
3369     uint newsize = _idom_size<<1;
3370     while( idx >= newsize ) {
3371       newsize <<= 1;
3372     }
3373     _idom      = REALLOC_RESOURCE_ARRAY( Node*,     _idom,_idom_size,newsize);
3374     _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
3375     memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
3376     _idom_size = newsize;
3377   }
3378   _idom[idx] = n;
3379   _dom_depth[idx] = dom_depth;
3380 }
3381 
3382 //------------------------------recompute_dom_depth---------------------------------------
3383 // The dominator tree is constructed with only parent pointers.
3384 // This recomputes the depth in the tree by first tagging all
3385 // nodes as "no depth yet" marker.  The next pass then runs up
3386 // the dom tree from each node marked "no depth yet", and computes
3387 // the depth on the way back down.
3388 void PhaseIdealLoop::recompute_dom_depth() {
3389   uint no_depth_marker = C->unique();
3390   uint i;
3391   // Initialize depth to "no depth yet" and realize all lazy updates
3392   for (i = 0; i < _idom_size; i++) {
3393     // Only indices with a _dom_depth has a Node* or NULL (otherwise uninitalized).
3394     if (_dom_depth[i] > 0 && _idom[i] != NULL) {
3395       _dom_depth[i] = no_depth_marker;
3396 
3397       // heal _idom if it has a fwd mapping in _nodes
3398       if (_idom[i]->in(0) == NULL) {
3399         idom(i);
3400       }
3401     }
3402   }
3403   if (_dom_stk == NULL) {
3404     uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
3405     if (init_size < 10) init_size = 10;
3406     _dom_stk = new GrowableArray<uint>(init_size);
3407   }
3408   // Compute new depth for each node.
3409   for (i = 0; i < _idom_size; i++) {
3410     uint j = i;
3411     // Run up the dom tree to find a node with a depth
3412     while (_dom_depth[j] == no_depth_marker) {
3413       _dom_stk->push(j);
3414       j = _idom[j]->_idx;
3415     }
3416     // Compute the depth on the way back down this tree branch
3417     uint dd = _dom_depth[j] + 1;
3418     while (_dom_stk->length() > 0) {
3419       uint j = _dom_stk->pop();
3420       _dom_depth[j] = dd;
3421       dd++;
3422     }
3423   }
3424 }
3425 
3426 //------------------------------sort-------------------------------------------
3427 // Insert 'loop' into the existing loop tree.  'innermost' is a leaf of the
3428 // loop tree, not the root.
3429 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
3430   if( !innermost ) return loop; // New innermost loop
3431 
3432   int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
3433   assert( loop_preorder, "not yet post-walked loop" );
3434   IdealLoopTree **pp = &innermost;      // Pointer to previous next-pointer
3435   IdealLoopTree *l = *pp;               // Do I go before or after 'l'?
3436 
3437   // Insert at start of list
3438   while( l ) {                  // Insertion sort based on pre-order
3439     if( l == loop ) return innermost; // Already on list!
3440     int l_preorder = get_preorder(l->_head); // Cache pre-order number
3441     assert( l_preorder, "not yet post-walked l" );
3442     // Check header pre-order number to figure proper nesting
3443     if( loop_preorder > l_preorder )
3444       break;                    // End of insertion
3445     // If headers tie (e.g., shared headers) check tail pre-order numbers.
3446     // Since I split shared headers, you'd think this could not happen.
3447     // BUT: I must first do the preorder numbering before I can discover I
3448     // have shared headers, so the split headers all get the same preorder
3449     // number as the RegionNode they split from.
3450     if( loop_preorder == l_preorder &&
3451         get_preorder(loop->_tail) < get_preorder(l->_tail) )
3452       break;                    // Also check for shared headers (same pre#)
3453     pp = &l->_parent;           // Chain up list
3454     l = *pp;
3455   }
3456   // Link into list
3457   // Point predecessor to me
3458   *pp = loop;
3459   // Point me to successor
3460   IdealLoopTree *p = loop->_parent;
3461   loop->_parent = l;            // Point me to successor
3462   if( p ) sort( p, innermost ); // Insert my parents into list as well
3463   return innermost;
3464 }
3465 
3466 //------------------------------build_loop_tree--------------------------------
3467 // I use a modified Vick/Tarjan algorithm.  I need pre- and a post- visit
3468 // bits.  The _nodes[] array is mapped by Node index and holds a NULL for
3469 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
3470 // tightest enclosing IdealLoopTree for post-walked.
3471 //
3472 // During my forward walk I do a short 1-layer lookahead to see if I can find
3473 // a loop backedge with that doesn't have any work on the backedge.  This
3474 // helps me construct nested loops with shared headers better.
3475 //
3476 // Once I've done the forward recursion, I do the post-work.  For each child
3477 // I check to see if there is a backedge.  Backedges define a loop!  I
3478 // insert an IdealLoopTree at the target of the backedge.
3479 //
3480 // During the post-work I also check to see if I have several children
3481 // belonging to different loops.  If so, then this Node is a decision point
3482 // where control flow can choose to change loop nests.  It is at this
3483 // decision point where I can figure out how loops are nested.  At this
3484 // time I can properly order the different loop nests from my children.
3485 // Note that there may not be any backedges at the decision point!
3486 //
3487 // Since the decision point can be far removed from the backedges, I can't
3488 // order my loops at the time I discover them.  Thus at the decision point
3489 // I need to inspect loop header pre-order numbers to properly nest my
3490 // loops.  This means I need to sort my childrens' loops by pre-order.
3491 // The sort is of size number-of-control-children, which generally limits
3492 // it to size 2 (i.e., I just choose between my 2 target loops).
3493 void PhaseIdealLoop::build_loop_tree() {
3494   // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3495   GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
3496   Node *n = C->root();
3497   bltstack.push(n);
3498   int pre_order = 1;
3499   int stack_size;
3500 
3501   while ( ( stack_size = bltstack.length() ) != 0 ) {
3502     n = bltstack.top(); // Leave node on stack
3503     if ( !is_visited(n) ) {
3504       // ---- Pre-pass Work ----
3505       // Pre-walked but not post-walked nodes need a pre_order number.
3506 
3507       set_preorder_visited( n, pre_order ); // set as visited
3508 
3509       // ---- Scan over children ----
3510       // Scan first over control projections that lead to loop headers.
3511       // This helps us find inner-to-outer loops with shared headers better.
3512 
3513       // Scan children's children for loop headers.
3514       for ( int i = n->outcnt() - 1; i >= 0; --i ) {
3515         Node* m = n->raw_out(i);       // Child
3516         if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
3517           // Scan over children's children to find loop
3518           for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3519             Node* l = m->fast_out(j);
3520             if( is_visited(l) &&       // Been visited?
3521                 !is_postvisited(l) &&  // But not post-visited
3522                 get_preorder(l) < pre_order ) { // And smaller pre-order
3523               // Found!  Scan the DFS down this path before doing other paths
3524               bltstack.push(m);
3525               break;
3526             }
3527           }
3528         }
3529       }
3530       pre_order++;
3531     }
3532     else if ( !is_postvisited(n) ) {
3533       // Note: build_loop_tree_impl() adds out edges on rare occasions,
3534       // such as com.sun.rsasign.am::a.
3535       // For non-recursive version, first, process current children.
3536       // On next iteration, check if additional children were added.
3537       for ( int k = n->outcnt() - 1; k >= 0; --k ) {
3538         Node* u = n->raw_out(k);
3539         if ( u->is_CFG() && !is_visited(u) ) {
3540           bltstack.push(u);
3541         }
3542       }
3543       if ( bltstack.length() == stack_size ) {
3544         // There were no additional children, post visit node now
3545         (void)bltstack.pop(); // Remove node from stack
3546         pre_order = build_loop_tree_impl( n, pre_order );
3547         // Check for bailout
3548         if (C->failing()) {
3549           return;
3550         }
3551         // Check to grow _preorders[] array for the case when
3552         // build_loop_tree_impl() adds new nodes.
3553         check_grow_preorders();
3554       }
3555     }
3556     else {
3557       (void)bltstack.pop(); // Remove post-visited node from stack
3558     }
3559   }
3560 }
3561 
3562 //------------------------------build_loop_tree_impl---------------------------
3563 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
3564   // ---- Post-pass Work ----
3565   // Pre-walked but not post-walked nodes need a pre_order number.
3566 
3567   // Tightest enclosing loop for this Node
3568   IdealLoopTree *innermost = NULL;
3569 
3570   // For all children, see if any edge is a backedge.  If so, make a loop
3571   // for it.  Then find the tightest enclosing loop for the self Node.
3572   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3573     Node* m = n->fast_out(i);   // Child
3574     if( n == m ) continue;      // Ignore control self-cycles
3575     if( !m->is_CFG() ) continue;// Ignore non-CFG edges
3576 
3577     IdealLoopTree *l;           // Child's loop
3578     if( !is_postvisited(m) ) {  // Child visited but not post-visited?
3579       // Found a backedge
3580       assert( get_preorder(m) < pre_order, "should be backedge" );
3581       // Check for the RootNode, which is already a LoopNode and is allowed
3582       // to have multiple "backedges".
3583       if( m == C->root()) {     // Found the root?
3584         l = _ltree_root;        // Root is the outermost LoopNode
3585       } else {                  // Else found a nested loop
3586         // Insert a LoopNode to mark this loop.
3587         l = new IdealLoopTree(this, m, n);
3588       } // End of Else found a nested loop
3589       if( !has_loop(m) )        // If 'm' does not already have a loop set
3590         set_loop(m, l);         // Set loop header to loop now
3591 
3592     } else {                    // Else not a nested loop
3593       if( !_nodes[m->_idx] ) continue; // Dead code has no loop
3594       l = get_loop(m);          // Get previously determined loop
3595       // If successor is header of a loop (nest), move up-loop till it
3596       // is a member of some outer enclosing loop.  Since there are no
3597       // shared headers (I've split them already) I only need to go up
3598       // at most 1 level.
3599       while( l && l->_head == m ) // Successor heads loop?
3600         l = l->_parent;         // Move up 1 for me
3601       // If this loop is not properly parented, then this loop
3602       // has no exit path out, i.e. its an infinite loop.
3603       if( !l ) {
3604         // Make loop "reachable" from root so the CFG is reachable.  Basically
3605         // insert a bogus loop exit that is never taken.  'm', the loop head,
3606         // points to 'n', one (of possibly many) fall-in paths.  There may be
3607         // many backedges as well.
3608 
3609         // Here I set the loop to be the root loop.  I could have, after
3610         // inserting a bogus loop exit, restarted the recursion and found my
3611         // new loop exit.  This would make the infinite loop a first-class
3612         // loop and it would then get properly optimized.  What's the use of
3613         // optimizing an infinite loop?
3614         l = _ltree_root;        // Oops, found infinite loop
3615 
3616         if (!_verify_only) {
3617           // Insert the NeverBranch between 'm' and it's control user.
3618           NeverBranchNode *iff = new NeverBranchNode( m );
3619           _igvn.register_new_node_with_optimizer(iff);
3620           set_loop(iff, l);
3621           Node *if_t = new CProjNode( iff, 0 );
3622           _igvn.register_new_node_with_optimizer(if_t);
3623           set_loop(if_t, l);
3624 
3625           Node* cfg = NULL;       // Find the One True Control User of m
3626           for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3627             Node* x = m->fast_out(j);
3628             if (x->is_CFG() && x != m && x != iff)
3629               { cfg = x; break; }
3630           }
3631           assert(cfg != NULL, "must find the control user of m");
3632           uint k = 0;             // Probably cfg->in(0)
3633           while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
3634           cfg->set_req( k, if_t ); // Now point to NeverBranch
3635           _igvn._worklist.push(cfg);
3636 
3637           // Now create the never-taken loop exit
3638           Node *if_f = new CProjNode( iff, 1 );
3639           _igvn.register_new_node_with_optimizer(if_f);
3640           set_loop(if_f, l);
3641           // Find frame ptr for Halt.  Relies on the optimizer
3642           // V-N'ing.  Easier and quicker than searching through
3643           // the program structure.
3644           Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr );
3645           _igvn.register_new_node_with_optimizer(frame);
3646           // Halt & Catch Fire
3647           Node *halt = new HaltNode( if_f, frame );
3648           _igvn.register_new_node_with_optimizer(halt);
3649           set_loop(halt, l);
3650           C->root()->add_req(halt);
3651         }
3652         set_loop(C->root(), _ltree_root);
3653       }
3654     }
3655     // Weeny check for irreducible.  This child was already visited (this
3656     // IS the post-work phase).  Is this child's loop header post-visited
3657     // as well?  If so, then I found another entry into the loop.
3658     if (!_verify_only) {
3659       while( is_postvisited(l->_head) ) {
3660         // found irreducible
3661         l->_irreducible = 1; // = true
3662         l = l->_parent;
3663         _has_irreducible_loops = true;
3664         // Check for bad CFG here to prevent crash, and bailout of compile
3665         if (l == NULL) {
3666           C->record_method_not_compilable("unhandled CFG detected during loop optimization");
3667           return pre_order;
3668         }
3669       }
3670       C->set_has_irreducible_loop(_has_irreducible_loops);
3671     }
3672 
3673     // This Node might be a decision point for loops.  It is only if
3674     // it's children belong to several different loops.  The sort call
3675     // does a trivial amount of work if there is only 1 child or all
3676     // children belong to the same loop.  If however, the children
3677     // belong to different loops, the sort call will properly set the
3678     // _parent pointers to show how the loops nest.
3679     //
3680     // In any case, it returns the tightest enclosing loop.
3681     innermost = sort( l, innermost );
3682   }
3683 
3684   // Def-use info will have some dead stuff; dead stuff will have no
3685   // loop decided on.
3686 
3687   // Am I a loop header?  If so fix up my parent's child and next ptrs.
3688   if( innermost && innermost->_head == n ) {
3689     assert( get_loop(n) == innermost, "" );
3690     IdealLoopTree *p = innermost->_parent;
3691     IdealLoopTree *l = innermost;
3692     while( p && l->_head == n ) {
3693       l->_next = p->_child;     // Put self on parents 'next child'
3694       p->_child = l;            // Make self as first child of parent
3695       l = p;                    // Now walk up the parent chain
3696       p = l->_parent;
3697     }
3698   } else {
3699     // Note that it is possible for a LoopNode to reach here, if the
3700     // backedge has been made unreachable (hence the LoopNode no longer
3701     // denotes a Loop, and will eventually be removed).
3702 
3703     // Record tightest enclosing loop for self.  Mark as post-visited.
3704     set_loop(n, innermost);
3705     // Also record has_call flag early on
3706     if( innermost ) {
3707       if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
3708         // Do not count uncommon calls
3709         if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
3710           Node *iff = n->in(0)->in(0);
3711           // No any calls for vectorized loops.
3712           if( UseSuperWord || !iff->is_If() ||
3713               (n->in(0)->Opcode() == Op_IfFalse &&
3714                (1.0 - iff->as_If()->_prob) >= 0.01) ||
3715               (iff->as_If()->_prob >= 0.01) )
3716             innermost->_has_call = 1;
3717         }
3718       } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
3719         // Disable loop optimizations if the loop has a scalar replaceable
3720         // allocation. This disabling may cause a potential performance lost
3721         // if the allocation is not eliminated for some reason.
3722         innermost->_allow_optimizations = false;
3723         innermost->_has_call = 1; // = true
3724       } else if (n->Opcode() == Op_SafePoint) {
3725         // Record all safepoints in this loop.
3726         if (innermost->_safepts == NULL) innermost->_safepts = new Node_List();
3727         innermost->_safepts->push(n);
3728       }
3729     }
3730   }
3731 
3732   // Flag as post-visited now
3733   set_postvisited(n);
3734   return pre_order;
3735 }
3736 
3737 
3738 //------------------------------build_loop_early-------------------------------
3739 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3740 // First pass computes the earliest controlling node possible.  This is the
3741 // controlling input with the deepest dominating depth.
3742 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3743   while (worklist.size() != 0) {
3744     // Use local variables nstack_top_n & nstack_top_i to cache values
3745     // on nstack's top.
3746     Node *nstack_top_n = worklist.pop();
3747     uint  nstack_top_i = 0;
3748 //while_nstack_nonempty:
3749     while (true) {
3750       // Get parent node and next input's index from stack's top.
3751       Node  *n = nstack_top_n;
3752       uint   i = nstack_top_i;
3753       uint cnt = n->req(); // Count of inputs
3754       if (i == 0) {        // Pre-process the node.
3755         if( has_node(n) &&            // Have either loop or control already?
3756             !has_ctrl(n) ) {          // Have loop picked out already?
3757           // During "merge_many_backedges" we fold up several nested loops
3758           // into a single loop.  This makes the members of the original
3759           // loop bodies pointing to dead loops; they need to move up
3760           // to the new UNION'd larger loop.  I set the _head field of these
3761           // dead loops to NULL and the _parent field points to the owning
3762           // loop.  Shades of UNION-FIND algorithm.
3763           IdealLoopTree *ilt;
3764           while( !(ilt = get_loop(n))->_head ) {
3765             // Normally I would use a set_loop here.  But in this one special
3766             // case, it is legal (and expected) to change what loop a Node
3767             // belongs to.
3768             _nodes.map(n->_idx, (Node*)(ilt->_parent) );
3769           }
3770           // Remove safepoints ONLY if I've already seen I don't need one.
3771           // (the old code here would yank a 2nd safepoint after seeing a
3772           // first one, even though the 1st did not dominate in the loop body
3773           // and thus could be avoided indefinitely)
3774           if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
3775               is_deleteable_safept(n)) {
3776             Node *in = n->in(TypeFunc::Control);
3777             lazy_replace(n,in);       // Pull safepoint now
3778             if (ilt->_safepts != NULL) {
3779               ilt->_safepts->yank(n);
3780             }
3781             // Carry on with the recursion "as if" we are walking
3782             // only the control input
3783             if( !visited.test_set( in->_idx ) ) {
3784               worklist.push(in);      // Visit this guy later, using worklist
3785             }
3786             // Get next node from nstack:
3787             // - skip n's inputs processing by setting i > cnt;
3788             // - we also will not call set_early_ctrl(n) since
3789             //   has_node(n) == true (see the condition above).
3790             i = cnt + 1;
3791           }
3792         }
3793       } // if (i == 0)
3794 
3795       // Visit all inputs
3796       bool done = true;       // Assume all n's inputs will be processed
3797       while (i < cnt) {
3798         Node *in = n->in(i);
3799         ++i;
3800         if (in == NULL) continue;
3801         if (in->pinned() && !in->is_CFG())
3802           set_ctrl(in, in->in(0));
3803         int is_visited = visited.test_set( in->_idx );
3804         if (!has_node(in)) {  // No controlling input yet?
3805           assert( !in->is_CFG(), "CFG Node with no controlling input?" );
3806           assert( !is_visited, "visit only once" );
3807           nstack.push(n, i);  // Save parent node and next input's index.
3808           nstack_top_n = in;  // Process current input now.
3809           nstack_top_i = 0;
3810           done = false;       // Not all n's inputs processed.
3811           break; // continue while_nstack_nonempty;
3812         } else if (!is_visited) {
3813           // This guy has a location picked out for him, but has not yet
3814           // been visited.  Happens to all CFG nodes, for instance.
3815           // Visit him using the worklist instead of recursion, to break
3816           // cycles.  Since he has a location already we do not need to
3817           // find his location before proceeding with the current Node.
3818           worklist.push(in);  // Visit this guy later, using worklist
3819         }
3820       }
3821       if (done) {
3822         // All of n's inputs have been processed, complete post-processing.
3823 
3824         // Compute earliest point this Node can go.
3825         // CFG, Phi, pinned nodes already know their controlling input.
3826         if (!has_node(n)) {
3827           // Record earliest legal location
3828           set_early_ctrl( n );
3829         }
3830         if (nstack.is_empty()) {
3831           // Finished all nodes on stack.
3832           // Process next node on the worklist.
3833           break;
3834         }
3835         // Get saved parent node and next input's index.
3836         nstack_top_n = nstack.node();
3837         nstack_top_i = nstack.index();
3838         nstack.pop();
3839       }
3840     } // while (true)
3841   }
3842 }
3843 
3844 //------------------------------dom_lca_internal--------------------------------
3845 // Pair-wise LCA
3846 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
3847   if( !n1 ) return n2;          // Handle NULL original LCA
3848   assert( n1->is_CFG(), "" );
3849   assert( n2->is_CFG(), "" );
3850   // find LCA of all uses
3851   uint d1 = dom_depth(n1);
3852   uint d2 = dom_depth(n2);
3853   while (n1 != n2) {
3854     if (d1 > d2) {
3855       n1 =      idom(n1);
3856       d1 = dom_depth(n1);
3857     } else if (d1 < d2) {
3858       n2 =      idom(n2);
3859       d2 = dom_depth(n2);
3860     } else {
3861       // Here d1 == d2.  Due to edits of the dominator-tree, sections
3862       // of the tree might have the same depth.  These sections have
3863       // to be searched more carefully.
3864 
3865       // Scan up all the n1's with equal depth, looking for n2.
3866       Node *t1 = idom(n1);
3867       while (dom_depth(t1) == d1) {
3868         if (t1 == n2)  return n2;
3869         t1 = idom(t1);
3870       }
3871       // Scan up all the n2's with equal depth, looking for n1.
3872       Node *t2 = idom(n2);
3873       while (dom_depth(t2) == d2) {
3874         if (t2 == n1)  return n1;
3875         t2 = idom(t2);
3876       }
3877       // Move up to a new dominator-depth value as well as up the dom-tree.
3878       n1 = t1;
3879       n2 = t2;
3880       d1 = dom_depth(n1);
3881       d2 = dom_depth(n2);
3882     }
3883   }
3884   return n1;
3885 }
3886 
3887 //------------------------------compute_idom-----------------------------------
3888 // Locally compute IDOM using dom_lca call.  Correct only if the incoming
3889 // IDOMs are correct.
3890 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
3891   assert( region->is_Region(), "" );
3892   Node *LCA = NULL;
3893   for( uint i = 1; i < region->req(); i++ ) {
3894     if( region->in(i) != C->top() )
3895       LCA = dom_lca( LCA, region->in(i) );
3896   }
3897   return LCA;
3898 }
3899 
3900 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
3901   bool had_error = false;
3902 #ifdef ASSERT
3903   if (early != C->root()) {
3904     // Make sure that there's a dominance path from LCA to early
3905     Node* d = LCA;
3906     while (d != early) {
3907       if (d == C->root()) {
3908         dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
3909         tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
3910         had_error = true;
3911         break;
3912       }
3913       d = idom(d);
3914     }
3915   }
3916 #endif
3917   return had_error;
3918 }
3919 
3920 
3921 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
3922   // Compute LCA over list of uses
3923   bool had_error = false;
3924   Node *LCA = NULL;
3925   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
3926     Node* c = n->fast_out(i);
3927     if (_nodes[c->_idx] == NULL)
3928       continue;                 // Skip the occasional dead node
3929     if( c->is_Phi() ) {         // For Phis, we must land above on the path
3930       for( uint j=1; j<c->req(); j++ ) {// For all inputs
3931         if( c->in(j) == n ) {   // Found matching input?
3932           Node *use = c->in(0)->in(j);
3933           if (_verify_only && use->is_top()) continue;
3934           LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3935           if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3936         }
3937       }
3938     } else {
3939       // For CFG data-users, use is in the block just prior
3940       Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
3941       LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3942       if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3943     }
3944   }
3945   assert(!had_error, "bad dominance");
3946   return LCA;
3947 }
3948 
3949 // Check the shape of the graph at the loop entry. In some cases,
3950 // the shape of the graph does not match the shape outlined below.
3951 // That is caused by the Opaque1 node "protecting" the shape of
3952 // the graph being removed by, for example, the IGVN performed
3953 // in PhaseIdealLoop::build_and_optimize().
3954 //
3955 // After the Opaque1 node has been removed, optimizations (e.g., split-if,
3956 // loop unswitching, and IGVN, or a combination of them) can freely change
3957 // the graph's shape. As a result, the graph shape outlined below cannot
3958 // be guaranteed anymore.
3959 bool PhaseIdealLoop::is_canonical_loop_entry(CountedLoopNode* cl) {
3960   if (!cl->is_main_loop() && !cl->is_post_loop()) {
3961     return false;
3962   }
3963   Node* ctrl = cl->skip_predicates();
3964 
3965   if (ctrl == NULL || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) {
3966     return false;
3967   }
3968   Node* iffm = ctrl->in(0);
3969   if (iffm == NULL || !iffm->is_If()) {
3970     return false;
3971   }
3972   Node* bolzm = iffm->in(1);
3973   if (bolzm == NULL || !bolzm->is_Bool()) {
3974     return false;
3975   }
3976   Node* cmpzm = bolzm->in(1);
3977   if (cmpzm == NULL || !cmpzm->is_Cmp()) {
3978     return false;
3979   }
3980   // compares can get conditionally flipped
3981   bool found_opaque = false;
3982   for (uint i = 1; i < cmpzm->req(); i++) {
3983     Node* opnd = cmpzm->in(i);
3984     if (opnd && opnd->Opcode() == Op_Opaque1) {
3985       found_opaque = true;
3986       break;
3987     }
3988   }
3989   if (!found_opaque) {
3990     return false;
3991   }
3992   return true;
3993 }
3994 
3995 //------------------------------get_late_ctrl----------------------------------
3996 // Compute latest legal control.
3997 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
3998   assert(early != NULL, "early control should not be NULL");
3999 
4000   Node* LCA = compute_lca_of_uses(n, early);
4001 #ifdef ASSERT
4002   if (LCA == C->root() && LCA != early) {
4003     // def doesn't dominate uses so print some useful debugging output
4004     compute_lca_of_uses(n, early, true);
4005   }
4006 #endif
4007 
4008   // if this is a load, check for anti-dependent stores
4009   // We use a conservative algorithm to identify potential interfering
4010   // instructions and for rescheduling the load.  The users of the memory
4011   // input of this load are examined.  Any use which is not a load and is
4012   // dominated by early is considered a potentially interfering store.
4013   // This can produce false positives.
4014   if (n->is_Load() && LCA != early) {
4015     Node_List worklist;
4016 
4017     Node *mem = n->in(MemNode::Memory);
4018     for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
4019       Node* s = mem->fast_out(i);
4020       worklist.push(s);
4021     }
4022     while(worklist.size() != 0 && LCA != early) {
4023       Node* s = worklist.pop();
4024       if (s->is_Load() || s->Opcode() == Op_SafePoint ||
4025           (s->is_CallStaticJava() && s->as_CallStaticJava()->uncommon_trap_request() != 0)) {
4026         continue;
4027       } else if (s->is_MergeMem()) {
4028         for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
4029           Node* s1 = s->fast_out(i);
4030           worklist.push(s1);
4031         }
4032       } else {
4033         Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
4034         assert(sctrl != NULL || s->outcnt() == 0, "must have control");
4035         if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
4036           LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
4037         }
4038       }
4039     }
4040   }
4041 
4042   assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
4043   return LCA;
4044 }
4045 
4046 // true if CFG node d dominates CFG node n
4047 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
4048   if (d == n)
4049     return true;
4050   assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
4051   uint dd = dom_depth(d);
4052   while (dom_depth(n) >= dd) {
4053     if (n == d)
4054       return true;
4055     n = idom(n);
4056   }
4057   return false;
4058 }
4059 
4060 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
4061 // Pair-wise LCA with tags.
4062 // Tag each index with the node 'tag' currently being processed
4063 // before advancing up the dominator chain using idom().
4064 // Later calls that find a match to 'tag' know that this path has already
4065 // been considered in the current LCA (which is input 'n1' by convention).
4066 // Since get_late_ctrl() is only called once for each node, the tag array
4067 // does not need to be cleared between calls to get_late_ctrl().
4068 // Algorithm trades a larger constant factor for better asymptotic behavior
4069 //
4070 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
4071   uint d1 = dom_depth(n1);
4072   uint d2 = dom_depth(n2);
4073 
4074   do {
4075     if (d1 > d2) {
4076       // current lca is deeper than n2
4077       _dom_lca_tags.map(n1->_idx, tag);
4078       n1 =      idom(n1);
4079       d1 = dom_depth(n1);
4080     } else if (d1 < d2) {
4081       // n2 is deeper than current lca
4082       Node *memo = _dom_lca_tags[n2->_idx];
4083       if( memo == tag ) {
4084         return n1;    // Return the current LCA
4085       }
4086       _dom_lca_tags.map(n2->_idx, tag);
4087       n2 =      idom(n2);
4088       d2 = dom_depth(n2);
4089     } else {
4090       // Here d1 == d2.  Due to edits of the dominator-tree, sections
4091       // of the tree might have the same depth.  These sections have
4092       // to be searched more carefully.
4093 
4094       // Scan up all the n1's with equal depth, looking for n2.
4095       _dom_lca_tags.map(n1->_idx, tag);
4096       Node *t1 = idom(n1);
4097       while (dom_depth(t1) == d1) {
4098         if (t1 == n2)  return n2;
4099         _dom_lca_tags.map(t1->_idx, tag);
4100         t1 = idom(t1);
4101       }
4102       // Scan up all the n2's with equal depth, looking for n1.
4103       _dom_lca_tags.map(n2->_idx, tag);
4104       Node *t2 = idom(n2);
4105       while (dom_depth(t2) == d2) {
4106         if (t2 == n1)  return n1;
4107         _dom_lca_tags.map(t2->_idx, tag);
4108         t2 = idom(t2);
4109       }
4110       // Move up to a new dominator-depth value as well as up the dom-tree.
4111       n1 = t1;
4112       n2 = t2;
4113       d1 = dom_depth(n1);
4114       d2 = dom_depth(n2);
4115     }
4116   } while (n1 != n2);
4117   return n1;
4118 }
4119 
4120 //------------------------------init_dom_lca_tags------------------------------
4121 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
4122 // Intended use does not involve any growth for the array, so it could
4123 // be of fixed size.
4124 void PhaseIdealLoop::init_dom_lca_tags() {
4125   uint limit = C->unique() + 1;
4126   _dom_lca_tags.map( limit, NULL );
4127 #ifdef ASSERT
4128   for( uint i = 0; i < limit; ++i ) {
4129     assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
4130   }
4131 #endif // ASSERT
4132 }
4133 
4134 //------------------------------clear_dom_lca_tags------------------------------
4135 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
4136 // Intended use does not involve any growth for the array, so it could
4137 // be of fixed size.
4138 void PhaseIdealLoop::clear_dom_lca_tags() {
4139   uint limit = C->unique() + 1;
4140   _dom_lca_tags.map( limit, NULL );
4141   _dom_lca_tags.clear();
4142 #ifdef ASSERT
4143   for( uint i = 0; i < limit; ++i ) {
4144     assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
4145   }
4146 #endif // ASSERT
4147 }
4148 
4149 //------------------------------build_loop_late--------------------------------
4150 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
4151 // Second pass finds latest legal placement, and ideal loop placement.
4152 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
4153   while (worklist.size() != 0) {
4154     Node *n = worklist.pop();
4155     // Only visit once
4156     if (visited.test_set(n->_idx)) continue;
4157     uint cnt = n->outcnt();
4158     uint   i = 0;
4159     while (true) {
4160       assert( _nodes[n->_idx], "no dead nodes" );
4161       // Visit all children
4162       if (i < cnt) {
4163         Node* use = n->raw_out(i);
4164         ++i;
4165         // Check for dead uses.  Aggressively prune such junk.  It might be
4166         // dead in the global sense, but still have local uses so I cannot
4167         // easily call 'remove_dead_node'.
4168         if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
4169           // Due to cycles, we might not hit the same fixed point in the verify
4170           // pass as we do in the regular pass.  Instead, visit such phis as
4171           // simple uses of the loop head.
4172           if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
4173             if( !visited.test(use->_idx) )
4174               worklist.push(use);
4175           } else if( !visited.test_set(use->_idx) ) {
4176             nstack.push(n, i); // Save parent and next use's index.
4177             n   = use;         // Process all children of current use.
4178             cnt = use->outcnt();
4179             i   = 0;
4180           }
4181         } else {
4182           // Do not visit around the backedge of loops via data edges.
4183           // push dead code onto a worklist
4184           _deadlist.push(use);
4185         }
4186       } else {
4187         // All of n's children have been processed, complete post-processing.
4188         build_loop_late_post(n);
4189         if (nstack.is_empty()) {
4190           // Finished all nodes on stack.
4191           // Process next node on the worklist.
4192           break;
4193         }
4194         // Get saved parent node and next use's index. Visit the rest of uses.
4195         n   = nstack.node();
4196         cnt = n->outcnt();
4197         i   = nstack.index();
4198         nstack.pop();
4199       }
4200     }
4201   }
4202 }
4203 
4204 // Verify that no data node is scheduled in the outer loop of a strip
4205 // mined loop.
4206 void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) {
4207 #ifdef ASSERT
4208   if (get_loop(least)->_nest == 0) {
4209     return;
4210   }
4211   IdealLoopTree* loop = get_loop(least);
4212   Node* head = loop->_head;
4213   if (head->is_OuterStripMinedLoop() &&
4214       // Verification can't be applied to fully built strip mined loops
4215       head->as_Loop()->outer_loop_end()->in(1)->find_int_con(-1) == 0) {
4216     Node* sfpt = head->as_Loop()->outer_safepoint();
4217     ResourceMark rm;
4218     Unique_Node_List wq;
4219     wq.push(sfpt);
4220     for (uint i = 0; i < wq.size(); i++) {
4221       Node *m = wq.at(i);
4222       for (uint i = 1; i < m->req(); i++) {
4223         Node* nn = m->in(i);
4224         if (nn == n) {
4225           return;
4226         }
4227         if (nn != NULL && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) {
4228           wq.push(nn);
4229         }
4230       }
4231     }
4232     ShouldNotReachHere();
4233   }
4234 #endif
4235 }
4236 
4237 
4238 //------------------------------build_loop_late_post---------------------------
4239 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
4240 // Second pass finds latest legal placement, and ideal loop placement.
4241 void PhaseIdealLoop::build_loop_late_post(Node *n) {
4242   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4243 
4244   if (bs->build_loop_late_post(this, n)) {
4245     return;
4246   }
4247 
4248   build_loop_late_post_work(n, true);
4249 }
4250 
4251 void PhaseIdealLoop::build_loop_late_post_work(Node *n, bool pinned) {
4252 
4253   if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) {
4254     _igvn._worklist.push(n);  // Maybe we'll normalize it, if no more loops.
4255   }
4256 
4257 #ifdef ASSERT
4258   if (_verify_only && !n->is_CFG()) {
4259     // Check def-use domination.
4260     compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
4261   }
4262 #endif
4263 
4264   // CFG and pinned nodes already handled
4265   if( n->in(0) ) {
4266     if( n->in(0)->is_top() ) return; // Dead?
4267 
4268     // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
4269     // _must_ be pinned (they have to observe their control edge of course).
4270     // Unlike Stores (which modify an unallocable resource, the memory
4271     // state), Mods/Loads can float around.  So free them up.
4272     switch( n->Opcode() ) {
4273     case Op_DivI:
4274     case Op_DivF:
4275     case Op_DivD:
4276     case Op_ModI:
4277     case Op_ModF:
4278     case Op_ModD:
4279     case Op_LoadB:              // Same with Loads; they can sink
4280     case Op_LoadUB:             // during loop optimizations.
4281     case Op_LoadUS:
4282     case Op_LoadD:
4283     case Op_LoadF:
4284     case Op_LoadI:
4285     case Op_LoadKlass:
4286     case Op_LoadNKlass:
4287     case Op_LoadL:
4288     case Op_LoadS:
4289     case Op_LoadP:
4290     case Op_LoadBarrierSlowReg:
4291     case Op_LoadBarrierWeakSlowReg:
4292     case Op_LoadN:
4293     case Op_LoadRange:
4294     case Op_LoadD_unaligned:
4295     case Op_LoadL_unaligned:
4296     case Op_StrComp:            // Does a bunch of load-like effects
4297     case Op_StrEquals:
4298     case Op_StrIndexOf:
4299     case Op_StrIndexOfChar:
4300     case Op_AryEq:
4301     case Op_HasNegatives:
4302       pinned = false;
4303     }
4304     if( pinned ) {
4305       IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
4306       if( !chosen_loop->_child )       // Inner loop?
4307         chosen_loop->_body.push(n); // Collect inner loops
4308       return;
4309     }
4310   } else {                      // No slot zero
4311     if( n->is_CFG() ) {         // CFG with no slot 0 is dead
4312       _nodes.map(n->_idx,0);    // No block setting, it's globally dead
4313       return;
4314     }
4315     assert(!n->is_CFG() || n->outcnt() == 0, "");
4316   }
4317 
4318   // Do I have a "safe range" I can select over?
4319   Node *early = get_ctrl(n);// Early location already computed
4320 
4321   // Compute latest point this Node can go
4322   Node *LCA = get_late_ctrl( n, early );
4323   // LCA is NULL due to uses being dead
4324   if( LCA == NULL ) {
4325 #ifdef ASSERT
4326     for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
4327       assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
4328     }
4329 #endif
4330     _nodes.map(n->_idx, 0);     // This node is useless
4331     _deadlist.push(n);
4332     return;
4333   }
4334   assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
4335 
4336   Node *legal = LCA;            // Walk 'legal' up the IDOM chain
4337   Node *least = legal;          // Best legal position so far
4338   while( early != legal ) {     // While not at earliest legal
4339 #ifdef ASSERT
4340     if (legal->is_Start() && !early->is_Root()) {
4341       // Bad graph. Print idom path and fail.
4342       dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
4343       assert(false, "Bad graph detected in build_loop_late");
4344     }
4345 #endif
4346     // Find least loop nesting depth
4347     legal = idom(legal);        // Bump up the IDOM tree
4348     // Check for lower nesting depth
4349     if( get_loop(legal)->_nest < get_loop(least)->_nest )
4350       least = legal;
4351   }
4352   assert(early == legal || legal != C->root(), "bad dominance of inputs");
4353 
4354   // Try not to place code on a loop entry projection
4355   // which can inhibit range check elimination.
4356   if (least != early) {
4357     Node* ctrl_out = least->unique_ctrl_out();
4358     if (ctrl_out && ctrl_out->is_Loop() &&
4359         least == ctrl_out->in(LoopNode::EntryControl)) {
4360       // Move the node above predicates as far up as possible so a
4361       // following pass of loop predication doesn't hoist a predicate
4362       // that depends on it above that node.
4363       Node* new_ctrl = least;
4364       for (;;) {
4365         if (!new_ctrl->is_Proj()) {
4366           break;
4367         }
4368         CallStaticJavaNode* call = new_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
4369         if (call == NULL) {
4370           break;
4371         }
4372         int req = call->uncommon_trap_request();
4373         Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req);
4374         if (trap_reason != Deoptimization::Reason_loop_limit_check &&
4375             trap_reason != Deoptimization::Reason_predicate &&
4376             trap_reason != Deoptimization::Reason_profile_predicate) {
4377           break;
4378         }
4379         Node* c = new_ctrl->in(0)->in(0);
4380         if (is_dominator(c, early) && c != early) {
4381           break;
4382         }
4383         new_ctrl = c;
4384       }
4385       least = new_ctrl;
4386     }
4387   }
4388 
4389 #ifdef ASSERT
4390   // If verifying, verify that 'verify_me' has a legal location
4391   // and choose it as our location.
4392   if( _verify_me ) {
4393     Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
4394     Node *legal = LCA;
4395     while( early != legal ) {   // While not at earliest legal
4396       if( legal == v_ctrl ) break;  // Check for prior good location
4397       legal = idom(legal)      ;// Bump up the IDOM tree
4398     }
4399     // Check for prior good location
4400     if( legal == v_ctrl ) least = legal; // Keep prior if found
4401   }
4402 #endif
4403 
4404   // Assign discovered "here or above" point
4405   least = find_non_split_ctrl(least);
4406   verify_strip_mined_scheduling(n, least);
4407   set_ctrl(n, least);
4408 
4409   // Collect inner loop bodies
4410   IdealLoopTree *chosen_loop = get_loop(least);
4411   if( !chosen_loop->_child )   // Inner loop?
4412     chosen_loop->_body.push(n);// Collect inner loops
4413 }
4414 
4415 #ifdef ASSERT
4416 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
4417   tty->print_cr("%s", msg);
4418   tty->print("n: "); n->dump();
4419   tty->print("early(n): "); early->dump();
4420   if (n->in(0) != NULL  && !n->in(0)->is_top() &&
4421       n->in(0) != early && !n->in(0)->is_Root()) {
4422     tty->print("n->in(0): "); n->in(0)->dump();
4423   }
4424   for (uint i = 1; i < n->req(); i++) {
4425     Node* in1 = n->in(i);
4426     if (in1 != NULL && in1 != n && !in1->is_top()) {
4427       tty->print("n->in(%d): ", i); in1->dump();
4428       Node* in1_early = get_ctrl(in1);
4429       tty->print("early(n->in(%d)): ", i); in1_early->dump();
4430       if (in1->in(0) != NULL     && !in1->in(0)->is_top() &&
4431           in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
4432         tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
4433       }
4434       for (uint j = 1; j < in1->req(); j++) {
4435         Node* in2 = in1->in(j);
4436         if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) {
4437           tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
4438           Node* in2_early = get_ctrl(in2);
4439           tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
4440           if (in2->in(0) != NULL     && !in2->in(0)->is_top() &&
4441               in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
4442             tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
4443           }
4444         }
4445       }
4446     }
4447   }
4448   tty->cr();
4449   tty->print("LCA(n): "); LCA->dump();
4450   for (uint i = 0; i < n->outcnt(); i++) {
4451     Node* u1 = n->raw_out(i);
4452     if (u1 == n)
4453       continue;
4454     tty->print("n->out(%d): ", i); u1->dump();
4455     if (u1->is_CFG()) {
4456       for (uint j = 0; j < u1->outcnt(); j++) {
4457         Node* u2 = u1->raw_out(j);
4458         if (u2 != u1 && u2 != n && u2->is_CFG()) {
4459           tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
4460         }
4461       }
4462     } else {
4463       Node* u1_later = get_ctrl(u1);
4464       tty->print("later(n->out(%d)): ", i); u1_later->dump();
4465       if (u1->in(0) != NULL     && !u1->in(0)->is_top() &&
4466           u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
4467         tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
4468       }
4469       for (uint j = 0; j < u1->outcnt(); j++) {
4470         Node* u2 = u1->raw_out(j);
4471         if (u2 == n || u2 == u1)
4472           continue;
4473         tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
4474         if (!u2->is_CFG()) {
4475           Node* u2_later = get_ctrl(u2);
4476           tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
4477           if (u2->in(0) != NULL     && !u2->in(0)->is_top() &&
4478               u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
4479             tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
4480           }
4481         }
4482       }
4483     }
4484   }
4485   tty->cr();
4486   int ct = 0;
4487   Node *dbg_legal = LCA;
4488   while(!dbg_legal->is_Start() && ct < 100) {
4489     tty->print("idom[%d] ",ct); dbg_legal->dump();
4490     ct++;
4491     dbg_legal = idom(dbg_legal);
4492   }
4493   tty->cr();
4494 }
4495 #endif
4496 
4497 #ifndef PRODUCT
4498 //------------------------------dump-------------------------------------------
4499 void PhaseIdealLoop::dump( ) const {
4500   ResourceMark rm;
4501   Arena* arena = Thread::current()->resource_area();
4502   Node_Stack stack(arena, C->live_nodes() >> 2);
4503   Node_List rpo_list;
4504   VectorSet visited(arena);
4505   visited.set(C->top()->_idx);
4506   rpo( C->root(), stack, visited, rpo_list );
4507   // Dump root loop indexed by last element in PO order
4508   dump( _ltree_root, rpo_list.size(), rpo_list );
4509 }
4510 
4511 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
4512   loop->dump_head();
4513 
4514   // Now scan for CFG nodes in the same loop
4515   for( uint j=idx; j > 0;  j-- ) {
4516     Node *n = rpo_list[j-1];
4517     if( !_nodes[n->_idx] )      // Skip dead nodes
4518       continue;
4519     if( get_loop(n) != loop ) { // Wrong loop nest
4520       if( get_loop(n)->_head == n &&    // Found nested loop?
4521           get_loop(n)->_parent == loop )
4522         dump(get_loop(n),rpo_list.size(),rpo_list);     // Print it nested-ly
4523       continue;
4524     }
4525 
4526     // Dump controlling node
4527     for( uint x = 0; x < loop->_nest; x++ )
4528       tty->print("  ");
4529     tty->print("C");
4530     if( n == C->root() ) {
4531       n->dump();
4532     } else {
4533       Node* cached_idom   = idom_no_update(n);
4534       Node *computed_idom = n->in(0);
4535       if( n->is_Region() ) {
4536         computed_idom = compute_idom(n);
4537         // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
4538         // any MultiBranch ctrl node), so apply a similar transform to
4539         // the cached idom returned from idom_no_update.
4540         cached_idom = find_non_split_ctrl(cached_idom);
4541       }
4542       tty->print(" ID:%d",computed_idom->_idx);
4543       n->dump();
4544       if( cached_idom != computed_idom ) {
4545         tty->print_cr("*** BROKEN IDOM!  Computed as: %d, cached as: %d",
4546                       computed_idom->_idx, cached_idom->_idx);
4547       }
4548     }
4549     // Dump nodes it controls
4550     for( uint k = 0; k < _nodes.Size(); k++ ) {
4551       // (k < C->unique() && get_ctrl(find(k)) == n)
4552       if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
4553         Node *m = C->root()->find(k);
4554         if( m && m->outcnt() > 0 ) {
4555           if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
4556             tty->print_cr("*** BROKEN CTRL ACCESSOR!  _nodes[k] is %p, ctrl is %p",
4557                           _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
4558           }
4559           for( uint j = 0; j < loop->_nest; j++ )
4560             tty->print("  ");
4561           tty->print(" ");
4562           m->dump();
4563         }
4564       }
4565     }
4566   }
4567 }
4568 #endif
4569 
4570 // Collect a R-P-O for the whole CFG.
4571 // Result list is in post-order (scan backwards for RPO)
4572 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
4573   stk.push(start, 0);
4574   visited.set(start->_idx);
4575 
4576   while (stk.is_nonempty()) {
4577     Node* m   = stk.node();
4578     uint  idx = stk.index();
4579     if (idx < m->outcnt()) {
4580       stk.set_index(idx + 1);
4581       Node* n = m->raw_out(idx);
4582       if (n->is_CFG() && !visited.test_set(n->_idx)) {
4583         stk.push(n, 0);
4584       }
4585     } else {
4586       rpo_list.push(m);
4587       stk.pop();
4588     }
4589   }
4590 }
4591 
4592 
4593 //=============================================================================
4594 //------------------------------LoopTreeIterator-----------------------------------
4595 
4596 // Advance to next loop tree using a preorder, left-to-right traversal.
4597 void LoopTreeIterator::next() {
4598   assert(!done(), "must not be done.");
4599   if (_curnt->_child != NULL) {
4600     _curnt = _curnt->_child;
4601   } else if (_curnt->_next != NULL) {
4602     _curnt = _curnt->_next;
4603   } else {
4604     while (_curnt != _root && _curnt->_next == NULL) {
4605       _curnt = _curnt->_parent;
4606     }
4607     if (_curnt == _root) {
4608       _curnt = NULL;
4609       assert(done(), "must be done.");
4610     } else {
4611       assert(_curnt->_next != NULL, "must be more to do");
4612       _curnt = _curnt->_next;
4613     }
4614   }
4615 }
--- EOF ---