1 /*
   2  * Copyright (c) 1998, 2018, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "ci/ciMethodData.hpp"
  27 #include "compiler/compileLog.hpp"
  28 #include "gc/shared/barrierSet.hpp"
  29 #include "gc/shared/c2/barrierSetC2.hpp"
  30 #include "libadt/vectset.hpp"
  31 #include "memory/allocation.inline.hpp"
  32 #include "memory/resourceArea.hpp"
  33 #include "opto/addnode.hpp"
  34 #include "opto/callnode.hpp"
  35 #include "opto/connode.hpp"
  36 #include "opto/convertnode.hpp"
  37 #include "opto/divnode.hpp"
  38 #include "opto/idealGraphPrinter.hpp"
  39 #include "opto/loopnode.hpp"
  40 #include "opto/mulnode.hpp"
  41 #include "opto/rootnode.hpp"
  42 #include "opto/superword.hpp"




  43 
  44 //=============================================================================
  45 //------------------------------is_loop_iv-------------------------------------
  46 // Determine if a node is Counted loop induction variable.
  47 // The method is declared in node.hpp.
  48 const Node* Node::is_loop_iv() const {
  49   if (this->is_Phi() && !this->as_Phi()->is_copy() &&
  50       this->as_Phi()->region()->is_CountedLoop() &&
  51       this->as_Phi()->region()->as_CountedLoop()->phi() == this) {
  52     return this;
  53   } else {
  54     return NULL;
  55   }
  56 }
  57 
  58 //=============================================================================
  59 //------------------------------dump_spec--------------------------------------
  60 // Dump special per-node info
  61 #ifndef PRODUCT
  62 void LoopNode::dump_spec(outputStream *st) const {
  63   if (is_inner_loop()) st->print( "inner " );
  64   if (is_partial_peel_loop()) st->print( "partial_peel " );
  65   if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
  66 }
  67 #endif
  68 
  69 //------------------------------is_valid_counted_loop-------------------------
  70 bool LoopNode::is_valid_counted_loop() const {
  71   if (is_CountedLoop()) {
  72     CountedLoopNode*    l  = as_CountedLoop();
  73     CountedLoopEndNode* le = l->loopexit_or_null();
  74     if (le != NULL &&
  75         le->proj_out_or_null(1 /* true */) == l->in(LoopNode::LoopBackControl)) {
  76       Node* phi  = l->phi();
  77       Node* exit = le->proj_out_or_null(0 /* false */);
  78       if (exit != NULL && exit->Opcode() == Op_IfFalse &&
  79           phi != NULL && phi->is_Phi() &&
  80           phi->in(LoopNode::LoopBackControl) == l->incr() &&
  81           le->loopnode() == l && le->stride_is_con()) {
  82         return true;
  83       }
  84     }
  85   }
  86   return false;
  87 }
  88 
  89 //------------------------------get_early_ctrl---------------------------------
  90 // Compute earliest legal control
  91 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
  92   assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
  93   uint i;
  94   Node *early;
  95   if (n->in(0) && !n->is_expensive()) {
  96     early = n->in(0);
  97     if (!early->is_CFG()) // Might be a non-CFG multi-def
  98       early = get_ctrl(early);        // So treat input as a straight data input
  99     i = 1;
 100   } else {
 101     early = get_ctrl(n->in(1));
 102     i = 2;
 103   }
 104   uint e_d = dom_depth(early);
 105   assert( early, "" );
 106   for (; i < n->req(); i++) {
 107     Node *cin = get_ctrl(n->in(i));
 108     assert( cin, "" );
 109     // Keep deepest dominator depth
 110     uint c_d = dom_depth(cin);
 111     if (c_d > e_d) {           // Deeper guy?
 112       early = cin;              // Keep deepest found so far
 113       e_d = c_d;
 114     } else if (c_d == e_d &&    // Same depth?
 115                early != cin) { // If not equal, must use slower algorithm
 116       // If same depth but not equal, one _must_ dominate the other
 117       // and we want the deeper (i.e., dominated) guy.
 118       Node *n1 = early;
 119       Node *n2 = cin;
 120       while (1) {
 121         n1 = idom(n1);          // Walk up until break cycle
 122         n2 = idom(n2);
 123         if (n1 == cin ||        // Walked early up to cin
 124             dom_depth(n2) < c_d)
 125           break;                // early is deeper; keep him
 126         if (n2 == early ||      // Walked cin up to early
 127             dom_depth(n1) < c_d) {
 128           early = cin;          // cin is deeper; keep him
 129           break;
 130         }
 131       }
 132       e_d = dom_depth(early);   // Reset depth register cache
 133     }
 134   }
 135 
 136   // Return earliest legal location
 137   assert(early == find_non_split_ctrl(early), "unexpected early control");
 138 
 139   if (n->is_expensive() && !_verify_only && !_verify_me) {
 140     assert(n->in(0), "should have control input");
 141     early = get_early_ctrl_for_expensive(n, early);
 142   }
 143 
 144   return early;
 145 }
 146 
 147 //------------------------------get_early_ctrl_for_expensive---------------------------------
 148 // Move node up the dominator tree as high as legal while still beneficial
 149 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {
 150   assert(n->in(0) && n->is_expensive(), "expensive node with control input here");
 151   assert(OptimizeExpensiveOps, "optimization off?");
 152 
 153   Node* ctl = n->in(0);
 154   assert(ctl->is_CFG(), "expensive input 0 must be cfg");
 155   uint min_dom_depth = dom_depth(earliest);
 156 #ifdef ASSERT
 157   if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) {
 158     dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl);
 159     assert(false, "Bad graph detected in get_early_ctrl_for_expensive");
 160   }
 161 #endif
 162   if (dom_depth(ctl) < min_dom_depth) {
 163     return earliest;
 164   }
 165 
 166   while (1) {
 167     Node *next = ctl;
 168     // Moving the node out of a loop on the projection of a If
 169     // confuses loop predication. So once we hit a Loop in a If branch
 170     // that doesn't branch to an UNC, we stop. The code that process
 171     // expensive nodes will notice the loop and skip over it to try to
 172     // move the node further up.
 173     if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) {
 174       if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
 175         break;
 176       }
 177       next = idom(ctl->in(1)->in(0));
 178     } else if (ctl->is_Proj()) {
 179       // We only move it up along a projection if the projection is
 180       // the single control projection for its parent: same code path,
 181       // if it's a If with UNC or fallthrough of a call.
 182       Node* parent_ctl = ctl->in(0);
 183       if (parent_ctl == NULL) {
 184         break;
 185       } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) {
 186         next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control();
 187       } else if (parent_ctl->is_If()) {
 188         if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
 189           break;
 190         }
 191         assert(idom(ctl) == parent_ctl, "strange");
 192         next = idom(parent_ctl);
 193       } else if (ctl->is_CatchProj()) {
 194         if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) {
 195           break;
 196         }
 197         assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph");
 198         next = parent_ctl->in(0)->in(0)->in(0);
 199       } else {
 200         // Check if parent control has a single projection (this
 201         // control is the only possible successor of the parent
 202         // control). If so, we can try to move the node above the
 203         // parent control.
 204         int nb_ctl_proj = 0;
 205         for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) {
 206           Node *p = parent_ctl->fast_out(i);
 207           if (p->is_Proj() && p->is_CFG()) {
 208             nb_ctl_proj++;
 209             if (nb_ctl_proj > 1) {
 210               break;
 211             }
 212           }
 213         }
 214 
 215         if (nb_ctl_proj > 1) {
 216           break;
 217         }
 218         assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call() ||
 219                BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(parent_ctl), "unexpected node");
 220         assert(idom(ctl) == parent_ctl, "strange");
 221         next = idom(parent_ctl);
 222       }
 223     } else {
 224       next = idom(ctl);
 225     }
 226     if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) {
 227       break;
 228     }
 229     ctl = next;
 230   }
 231 
 232   if (ctl != n->in(0)) {
 233     _igvn.replace_input_of(n, 0, ctl);
 234     _igvn.hash_insert(n);
 235   }
 236 
 237   return ctl;
 238 }
 239 
 240 
 241 //------------------------------set_early_ctrl---------------------------------
 242 // Set earliest legal control
 243 void PhaseIdealLoop::set_early_ctrl( Node *n ) {
 244   Node *early = get_early_ctrl(n);
 245 
 246   // Record earliest legal location
 247   set_ctrl(n, early);
 248 }
 249 
 250 //------------------------------set_subtree_ctrl-------------------------------
 251 // set missing _ctrl entries on new nodes
 252 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {
 253   // Already set?  Get out.
 254   if( _nodes[n->_idx] ) return;
 255   // Recursively set _nodes array to indicate where the Node goes
 256   uint i;
 257   for( i = 0; i < n->req(); ++i ) {
 258     Node *m = n->in(i);
 259     if( m && m != C->root() )
 260       set_subtree_ctrl( m );
 261   }
 262 
 263   // Fixup self
 264   set_early_ctrl( n );
 265 }
 266 
 267 // Create a skeleton strip mined outer loop: a Loop head before the
 268 // inner strip mined loop, a safepoint and an exit condition guarded
 269 // by an opaque node after the inner strip mined loop with a backedge
 270 // to the loop head. The inner strip mined loop is left as it is. Only
 271 // once loop optimizations are over, do we adjust the inner loop exit
 272 // condition to limit its number of iterations, set the outer loop
 273 // exit condition and add Phis to the outer loop head. Some loop
 274 // optimizations that operate on the inner strip mined loop need to be
 275 // aware of the outer strip mined loop: loop unswitching needs to
 276 // clone the outer loop as well as the inner, unrolling needs to only
 277 // clone the inner loop etc. No optimizations need to change the outer
 278 // strip mined loop as it is only a skeleton.
 279 IdealLoopTree* PhaseIdealLoop::create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
 280                                                              IdealLoopTree* loop, float cl_prob, float le_fcnt,
 281                                                              Node*& entry_control, Node*& iffalse) {
 282   Node* outer_test = _igvn.intcon(0);
 283   set_ctrl(outer_test, C->root());
 284   Node *orig = iffalse;
 285   iffalse = iffalse->clone();
 286   _igvn.register_new_node_with_optimizer(iffalse);
 287   set_idom(iffalse, idom(orig), dom_depth(orig));
 288 
 289   IfNode *outer_le = new OuterStripMinedLoopEndNode(iffalse, outer_test, cl_prob, le_fcnt);
 290   Node *outer_ift = new IfTrueNode (outer_le);
 291   Node* outer_iff = orig;
 292   _igvn.replace_input_of(outer_iff, 0, outer_le);
 293 
 294   LoopNode *outer_l = new OuterStripMinedLoopNode(C, init_control, outer_ift);
 295   entry_control = outer_l;
 296 
 297   IdealLoopTree* outer_ilt = new IdealLoopTree(this, outer_l, outer_ift);
 298   IdealLoopTree* parent = loop->_parent;
 299   IdealLoopTree* sibling = parent->_child;
 300   if (sibling == loop) {
 301     parent->_child = outer_ilt;
 302   } else {
 303     while (sibling->_next != loop) {
 304       sibling = sibling->_next;
 305     }
 306     sibling->_next = outer_ilt;
 307   }
 308   outer_ilt->_next = loop->_next;
 309   outer_ilt->_parent = parent;
 310   outer_ilt->_child = loop;
 311   outer_ilt->_nest = loop->_nest;
 312   loop->_parent = outer_ilt;
 313   loop->_next = NULL;
 314   loop->_nest++;
 315 
 316   set_loop(iffalse, outer_ilt);
 317   register_control(outer_le, outer_ilt, iffalse);
 318   register_control(outer_ift, outer_ilt, outer_le);
 319   set_idom(outer_iff, outer_le, dom_depth(outer_le));
 320   _igvn.register_new_node_with_optimizer(outer_l);
 321   set_loop(outer_l, outer_ilt);
 322   set_idom(outer_l, init_control, dom_depth(init_control)+1);
 323 
 324   return outer_ilt;
 325 }
 326 
 327 void PhaseIdealLoop::insert_loop_limit_check(ProjNode* limit_check_proj, Node* cmp_limit, Node* bol) {
 328   Node* new_predicate_proj = create_new_if_for_predicate(limit_check_proj, NULL,
 329                                                          Deoptimization::Reason_loop_limit_check,
 330                                                          Op_If);
 331   Node* iff = new_predicate_proj->in(0);
 332   assert(iff->Opcode() == Op_If, "bad graph shape");
 333   Node* conv = iff->in(1);
 334   assert(conv->Opcode() == Op_Conv2B, "bad graph shape");
 335   Node* opaq = conv->in(1);
 336   assert(opaq->Opcode() == Op_Opaque1, "bad graph shape");
 337   cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
 338   bol = _igvn.register_new_node_with_optimizer(bol);
 339   set_subtree_ctrl(bol);
 340   _igvn.replace_input_of(iff, 1, bol);
 341 
 342 #ifndef PRODUCT
 343   // report that the loop predication has been actually performed
 344   // for this loop
 345   if (TraceLoopLimitCheck) {
 346     tty->print_cr("Counted Loop Limit Check generated:");
 347     debug_only( bol->dump(2); )
 348   }
 349 #endif
 350 }
 351 
 352 //------------------------------is_counted_loop--------------------------------
 353 bool PhaseIdealLoop::is_counted_loop(Node* x, IdealLoopTree*& loop) {
 354   PhaseGVN *gvn = &_igvn;
 355 
 356   // Counted loop head must be a good RegionNode with only 3 not NULL
 357   // control input edges: Self, Entry, LoopBack.
 358   if (x->in(LoopNode::Self) == NULL || x->req() != 3 || loop->_irreducible) {
 359     return false;
 360   }
 361   Node *init_control = x->in(LoopNode::EntryControl);
 362   Node *back_control = x->in(LoopNode::LoopBackControl);
 363   if (init_control == NULL || back_control == NULL)    // Partially dead
 364     return false;
 365   // Must also check for TOP when looking for a dead loop
 366   if (init_control->is_top() || back_control->is_top())
 367     return false;
 368 
 369   // Allow funny placement of Safepoint
 370   if (back_control->Opcode() == Op_SafePoint) {
 371     if (LoopStripMiningIter != 0) {
 372       // Leaving the safepoint on the backedge and creating a
 373       // CountedLoop will confuse optimizations. We can't move the
 374       // safepoint around because its jvm state wouldn't match a new
 375       // location. Give up on that loop.
 376       return false;
 377     }
 378     back_control = back_control->in(TypeFunc::Control);
 379   }
 380 
 381   // Controlling test for loop
 382   Node *iftrue = back_control;
 383   uint iftrue_op = iftrue->Opcode();
 384   if (iftrue_op != Op_IfTrue &&
 385       iftrue_op != Op_IfFalse)
 386     // I have a weird back-control.  Probably the loop-exit test is in
 387     // the middle of the loop and I am looking at some trailing control-flow
 388     // merge point.  To fix this I would have to partially peel the loop.
 389     return false; // Obscure back-control
 390 
 391   // Get boolean guarding loop-back test
 392   Node *iff = iftrue->in(0);
 393   if (get_loop(iff) != loop || !iff->in(1)->is_Bool())
 394     return false;
 395   BoolNode *test = iff->in(1)->as_Bool();
 396   BoolTest::mask bt = test->_test._test;
 397   float cl_prob = iff->as_If()->_prob;
 398   if (iftrue_op == Op_IfFalse) {
 399     bt = BoolTest(bt).negate();
 400     cl_prob = 1.0 - cl_prob;
 401   }
 402   // Get backedge compare
 403   Node *cmp = test->in(1);
 404   int cmp_op = cmp->Opcode();
 405   if (cmp_op != Op_CmpI)
 406     return false;                // Avoid pointer & float compares
 407 
 408   // Find the trip-counter increment & limit.  Limit must be loop invariant.
 409   Node *incr  = cmp->in(1);
 410   Node *limit = cmp->in(2);
 411 
 412   // ---------
 413   // need 'loop()' test to tell if limit is loop invariant
 414   // ---------
 415 
 416   if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
 417     Node *tmp = incr;            // Then reverse order into the CmpI
 418     incr = limit;
 419     limit = tmp;
 420     bt = BoolTest(bt).commute(); // And commute the exit test
 421   }
 422   if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant
 423     return false;
 424   if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
 425     return false;
 426 
 427   Node* phi_incr = NULL;
 428   // Trip-counter increment must be commutative & associative.
 429   if (incr->Opcode() == Op_CastII) {
 430     incr = incr->in(1);
 431   }
 432   if (incr->is_Phi()) {
 433     if (incr->as_Phi()->region() != x || incr->req() != 3)
 434       return false; // Not simple trip counter expression
 435     phi_incr = incr;
 436     incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
 437     if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
 438       return false;
 439   }
 440 
 441   Node* trunc1 = NULL;
 442   Node* trunc2 = NULL;
 443   const TypeInt* iv_trunc_t = NULL;
 444   Node* orig_incr = incr;
 445   if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {
 446     return false; // Funny increment opcode
 447   }
 448   assert(incr->Opcode() == Op_AddI, "wrong increment code");
 449 
 450   const TypeInt* limit_t = gvn->type(limit)->is_int();
 451   if (trunc1 != NULL) {
 452     // When there is a truncation, we must be sure that after the truncation
 453     // the trip counter will end up higher than the limit, otherwise we are looking
 454     // at an endless loop. Can happen with range checks.
 455 
 456     // Example:
 457     // int i = 0;
 458     // while (true)
 459     //    sum + = array[i];
 460     //    i++;
 461     //    i = i && 0x7fff;
 462     //  }
 463     //
 464     // If the array is shorter than 0x8000 this exits through a AIOOB
 465     //  - Counted loop transformation is ok
 466     // If the array is longer then this is an endless loop
 467     //  - No transformation can be done.
 468 
 469     const TypeInt* incr_t = gvn->type(orig_incr)->is_int();
 470     if (limit_t->_hi > incr_t->_hi) {
 471       // if the limit can have a higher value than the increment (before the phi)
 472       return false;
 473     }
 474   }
 475 
 476   // Get merge point
 477   Node *xphi = incr->in(1);
 478   Node *stride = incr->in(2);
 479   if (!stride->is_Con()) {     // Oops, swap these
 480     if (!xphi->is_Con())       // Is the other guy a constant?
 481       return false;             // Nope, unknown stride, bail out
 482     Node *tmp = xphi;           // 'incr' is commutative, so ok to swap
 483     xphi = stride;
 484     stride = tmp;
 485   }
 486   if (xphi->Opcode() == Op_CastII) {
 487     xphi = xphi->in(1);
 488   }
 489   // Stride must be constant
 490   int stride_con = stride->get_int();
 491   if (stride_con == 0)
 492     return false; // missed some peephole opt
 493 
 494   if (!xphi->is_Phi())
 495     return false; // Too much math on the trip counter
 496   if (phi_incr != NULL && phi_incr != xphi)
 497     return false;
 498   PhiNode *phi = xphi->as_Phi();
 499 
 500   // Phi must be of loop header; backedge must wrap to increment
 501   if (phi->region() != x)
 502     return false;
 503   if ((trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr) ||
 504       (trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1)) {
 505     return false;
 506   }
 507   Node *init_trip = phi->in(LoopNode::EntryControl);
 508 
 509   // If iv trunc type is smaller than int, check for possible wrap.
 510   if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
 511     assert(trunc1 != NULL, "must have found some truncation");
 512 
 513     // Get a better type for the phi (filtered thru if's)
 514     const TypeInt* phi_ft = filtered_type(phi);
 515 
 516     // Can iv take on a value that will wrap?
 517     //
 518     // Ensure iv's limit is not within "stride" of the wrap value.
 519     //
 520     // Example for "short" type
 521     //    Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
 522     //    If the stride is +10, then the last value of the induction
 523     //    variable before the increment (phi_ft->_hi) must be
 524     //    <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
 525     //    ensure no truncation occurs after the increment.
 526 
 527     if (stride_con > 0) {
 528       if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
 529           iv_trunc_t->_lo > phi_ft->_lo) {
 530         return false;  // truncation may occur
 531       }
 532     } else if (stride_con < 0) {
 533       if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
 534           iv_trunc_t->_hi < phi_ft->_hi) {
 535         return false;  // truncation may occur
 536       }
 537     }
 538     // No possibility of wrap so truncation can be discarded
 539     // Promote iv type to Int
 540   } else {
 541     assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
 542   }
 543 
 544   // If the condition is inverted and we will be rolling
 545   // through MININT to MAXINT, then bail out.
 546   if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
 547       // Odd stride
 548       (bt == BoolTest::ne && stride_con != 1 && stride_con != -1) ||
 549       // Count down loop rolls through MAXINT
 550       ((bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0) ||
 551       // Count up loop rolls through MININT
 552       ((bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0)) {
 553     return false; // Bail out
 554   }
 555 
 556   const TypeInt* init_t = gvn->type(init_trip)->is_int();
 557 
 558   if (stride_con > 0) {
 559     jlong init_p = (jlong)init_t->_lo + stride_con;
 560     if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi)
 561       return false; // cyclic loop or this loop trips only once
 562   } else {
 563     jlong init_p = (jlong)init_t->_hi + stride_con;
 564     if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo)
 565       return false; // cyclic loop or this loop trips only once
 566   }
 567 
 568   if (phi_incr != NULL && bt != BoolTest::ne) {
 569     // check if there is a possiblity of IV overflowing after the first increment
 570     if (stride_con > 0) {
 571       if (init_t->_hi > max_jint - stride_con) {
 572         return false;
 573       }
 574     } else {
 575       if (init_t->_lo < min_jint - stride_con) {
 576         return false;
 577       }
 578     }
 579   }
 580 
 581   // =================================================
 582   // ---- SUCCESS!   Found A Trip-Counted Loop!  -----
 583   //
 584   assert(x->Opcode() == Op_Loop, "regular loops only");
 585   C->print_method(PHASE_BEFORE_CLOOPS, 3);
 586 
 587   Node *hook = new Node(6);
 588 
 589   // ===================================================
 590   // Generate loop limit check to avoid integer overflow
 591   // in cases like next (cyclic loops):
 592   //
 593   // for (i=0; i <= max_jint; i++) {}
 594   // for (i=0; i <  max_jint; i+=2) {}
 595   //
 596   //
 597   // Limit check predicate depends on the loop test:
 598   //
 599   // for(;i != limit; i++)       --> limit <= (max_jint)
 600   // for(;i <  limit; i+=stride) --> limit <= (max_jint - stride + 1)
 601   // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride    )
 602   //
 603 
 604   // Check if limit is excluded to do more precise int overflow check.
 605   bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge);
 606   int stride_m  = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1));
 607 
 608   // If compare points directly to the phi we need to adjust
 609   // the compare so that it points to the incr. Limit have
 610   // to be adjusted to keep trip count the same and the
 611   // adjusted limit should be checked for int overflow.
 612   if (phi_incr != NULL) {
 613     stride_m  += stride_con;
 614   }
 615 
 616   if (limit->is_Con()) {
 617     int limit_con = limit->get_int();
 618     if ((stride_con > 0 && limit_con > (max_jint - stride_m)) ||
 619         (stride_con < 0 && limit_con < (min_jint - stride_m))) {
 620       // Bailout: it could be integer overflow.
 621       return false;
 622     }
 623   } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) ||
 624              (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) {
 625       // Limit's type may satisfy the condition, for example,
 626       // when it is an array length.
 627   } else {
 628     // Generate loop's limit check.
 629     // Loop limit check predicate should be near the loop.
 630     ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
 631     if (!limit_check_proj) {
 632       // The limit check predicate is not generated if this method trapped here before.
 633 #ifdef ASSERT
 634       if (TraceLoopLimitCheck) {
 635         tty->print("missing loop limit check:");
 636         loop->dump_head();
 637         x->dump(1);
 638       }
 639 #endif
 640       return false;
 641     }
 642 
 643     IfNode* check_iff = limit_check_proj->in(0)->as_If();
 644 
 645     if (!is_dominator(get_ctrl(limit), check_iff->in(0))) {
 646       return false;
 647     }
 648 
 649     Node* cmp_limit;
 650     Node* bol;
 651 
 652     if (stride_con > 0) {
 653       cmp_limit = new CmpINode(limit, _igvn.intcon(max_jint - stride_m));
 654       bol = new BoolNode(cmp_limit, BoolTest::le);
 655     } else {
 656       cmp_limit = new CmpINode(limit, _igvn.intcon(min_jint - stride_m));
 657       bol = new BoolNode(cmp_limit, BoolTest::ge);
 658     }
 659 
 660     insert_loop_limit_check(limit_check_proj, cmp_limit, bol);
 661   }
 662 
 663   // Now we need to canonicalize loop condition.
 664   if (bt == BoolTest::ne) {
 665     assert(stride_con == 1 || stride_con == -1, "simple increment only");
 666     if (stride_con > 0 && init_t->_hi < limit_t->_lo) {
 667       // 'ne' can be replaced with 'lt' only when init < limit.
 668       bt = BoolTest::lt;
 669     } else if (stride_con < 0 && init_t->_lo > limit_t->_hi) {
 670       // 'ne' can be replaced with 'gt' only when init > limit.
 671       bt = BoolTest::gt;
 672     } else {
 673       ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
 674       if (!limit_check_proj) {
 675         // The limit check predicate is not generated if this method trapped here before.
 676 #ifdef ASSERT
 677         if (TraceLoopLimitCheck) {
 678           tty->print("missing loop limit check:");
 679           loop->dump_head();
 680           x->dump(1);
 681         }
 682 #endif
 683         return false;
 684       }
 685       IfNode* check_iff = limit_check_proj->in(0)->as_If();
 686 
 687       if (!is_dominator(get_ctrl(limit), check_iff->in(0)) ||
 688           !is_dominator(get_ctrl(init_trip), check_iff->in(0))) {
 689         return false;
 690       }
 691 
 692       Node* cmp_limit;
 693       Node* bol;
 694 
 695       if (stride_con > 0) {
 696         cmp_limit = new CmpINode(init_trip, limit);
 697         bol = new BoolNode(cmp_limit, BoolTest::lt);
 698       } else {
 699         cmp_limit = new CmpINode(init_trip, limit);
 700         bol = new BoolNode(cmp_limit, BoolTest::gt);
 701       }
 702 
 703       insert_loop_limit_check(limit_check_proj, cmp_limit, bol);
 704 
 705       if (stride_con > 0) {
 706         // 'ne' can be replaced with 'lt' only when init < limit.
 707         bt = BoolTest::lt;
 708       } else if (stride_con < 0) {
 709         // 'ne' can be replaced with 'gt' only when init > limit.
 710         bt = BoolTest::gt;
 711       }
 712     }
 713   }
 714 
 715   if (phi_incr != NULL) {
 716     // If compare points directly to the phi we need to adjust
 717     // the compare so that it points to the incr. Limit have
 718     // to be adjusted to keep trip count the same and we
 719     // should avoid int overflow.
 720     //
 721     //   i = init; do {} while(i++ < limit);
 722     // is converted to
 723     //   i = init; do {} while(++i < limit+1);
 724     //
 725     limit = gvn->transform(new AddINode(limit, stride));
 726   }
 727 
 728   if (incl_limit) {
 729     // The limit check guaranties that 'limit <= (max_jint - stride)' so
 730     // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
 731     //
 732     Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1);
 733     limit = gvn->transform(new AddINode(limit, one));
 734     if (bt == BoolTest::le)
 735       bt = BoolTest::lt;
 736     else if (bt == BoolTest::ge)
 737       bt = BoolTest::gt;
 738     else
 739       ShouldNotReachHere();
 740   }
 741   set_subtree_ctrl( limit );
 742 
 743   if (LoopStripMiningIter == 0) {
 744     // Check for SafePoint on backedge and remove
 745     Node *sfpt = x->in(LoopNode::LoopBackControl);
 746     if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
 747       lazy_replace( sfpt, iftrue );
 748       if (loop->_safepts != NULL) {
 749         loop->_safepts->yank(sfpt);
 750       }
 751       loop->_tail = iftrue;
 752     }
 753   }
 754 
 755   // Build a canonical trip test.
 756   // Clone code, as old values may be in use.
 757   incr = incr->clone();
 758   incr->set_req(1,phi);
 759   incr->set_req(2,stride);
 760   incr = _igvn.register_new_node_with_optimizer(incr);
 761   set_early_ctrl( incr );
 762   _igvn.rehash_node_delayed(phi);
 763   phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
 764 
 765   // If phi type is more restrictive than Int, raise to
 766   // Int to prevent (almost) infinite recursion in igvn
 767   // which can only handle integer types for constants or minint..maxint.
 768   if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
 769     Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
 770     nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
 771     nphi = _igvn.register_new_node_with_optimizer(nphi);
 772     set_ctrl(nphi, get_ctrl(phi));
 773     _igvn.replace_node(phi, nphi);
 774     phi = nphi->as_Phi();
 775   }
 776   cmp = cmp->clone();
 777   cmp->set_req(1,incr);
 778   cmp->set_req(2,limit);
 779   cmp = _igvn.register_new_node_with_optimizer(cmp);
 780   set_ctrl(cmp, iff->in(0));
 781 
 782   test = test->clone()->as_Bool();
 783   (*(BoolTest*)&test->_test)._test = bt;
 784   test->set_req(1,cmp);
 785   _igvn.register_new_node_with_optimizer(test);
 786   set_ctrl(test, iff->in(0));
 787 
 788   // Replace the old IfNode with a new LoopEndNode
 789   Node *lex = _igvn.register_new_node_with_optimizer(new CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
 790   IfNode *le = lex->as_If();
 791   uint dd = dom_depth(iff);
 792   set_idom(le, le->in(0), dd); // Update dominance for loop exit
 793   set_loop(le, loop);
 794 
 795   // Get the loop-exit control
 796   Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
 797 
 798   // Need to swap loop-exit and loop-back control?
 799   if (iftrue_op == Op_IfFalse) {
 800     Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le));
 801     Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le));
 802 
 803     loop->_tail = back_control = ift2;
 804     set_loop(ift2, loop);
 805     set_loop(iff2, get_loop(iffalse));
 806 
 807     // Lazy update of 'get_ctrl' mechanism.
 808     lazy_replace(iffalse, iff2);
 809     lazy_replace(iftrue,  ift2);
 810 
 811     // Swap names
 812     iffalse = iff2;
 813     iftrue  = ift2;
 814   } else {
 815     _igvn.rehash_node_delayed(iffalse);
 816     _igvn.rehash_node_delayed(iftrue);
 817     iffalse->set_req_X( 0, le, &_igvn );
 818     iftrue ->set_req_X( 0, le, &_igvn );
 819   }
 820 
 821   set_idom(iftrue,  le, dd+1);
 822   set_idom(iffalse, le, dd+1);
 823   assert(iff->outcnt() == 0, "should be dead now");
 824   lazy_replace( iff, le ); // fix 'get_ctrl'
 825 
 826   Node *sfpt2 = le->in(0);
 827 
 828   Node* entry_control = init_control;
 829   bool strip_mine_loop = LoopStripMiningIter > 1 && loop->_child == NULL &&
 830     sfpt2->Opcode() == Op_SafePoint && !loop->_has_call;
 831   IdealLoopTree* outer_ilt = NULL;
 832   if (strip_mine_loop) {
 833     outer_ilt = create_outer_strip_mined_loop(test, cmp, init_control, loop,
 834                                               cl_prob, le->_fcnt, entry_control,
 835                                               iffalse);
 836   }
 837 
 838   // Now setup a new CountedLoopNode to replace the existing LoopNode
 839   CountedLoopNode *l = new CountedLoopNode(entry_control, back_control);
 840   l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
 841   // The following assert is approximately true, and defines the intention
 842   // of can_be_counted_loop.  It fails, however, because phase->type
 843   // is not yet initialized for this loop and its parts.
 844   //assert(l->can_be_counted_loop(this), "sanity");
 845   _igvn.register_new_node_with_optimizer(l);
 846   set_loop(l, loop);
 847   loop->_head = l;
 848   // Fix all data nodes placed at the old loop head.
 849   // Uses the lazy-update mechanism of 'get_ctrl'.
 850   lazy_replace( x, l );
 851   set_idom(l, entry_control, dom_depth(entry_control) + 1);
 852 
 853   if (LoopStripMiningIter == 0 || strip_mine_loop) {
 854     // Check for immediately preceding SafePoint and remove
 855     if (sfpt2->Opcode() == Op_SafePoint && (LoopStripMiningIter != 0 || is_deleteable_safept(sfpt2))) {
 856       if (strip_mine_loop) {
 857         Node* outer_le = outer_ilt->_tail->in(0);
 858         Node* sfpt = sfpt2->clone();
 859         sfpt->set_req(0, iffalse);
 860         outer_le->set_req(0, sfpt);
 861         register_control(sfpt, outer_ilt, iffalse);
 862         set_idom(outer_le, sfpt, dom_depth(sfpt));
 863       }
 864       lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
 865       if (loop->_safepts != NULL) {
 866         loop->_safepts->yank(sfpt2);
 867       }
 868     }
 869   }
 870 
 871   // Free up intermediate goo
 872   _igvn.remove_dead_node(hook);
 873 
 874 #ifdef ASSERT
 875   assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
 876   assert(l == loop->_head && l->phi() == phi && l->loopexit_or_null() == lex, "" );
 877 #endif
 878 #ifndef PRODUCT
 879   if (TraceLoopOpts) {
 880     tty->print("Counted      ");
 881     loop->dump_head();
 882   }
 883 #endif
 884 
 885   C->print_method(PHASE_AFTER_CLOOPS, 3);
 886 
 887   // Capture bounds of the loop in the induction variable Phi before
 888   // subsequent transformation (iteration splitting) obscures the
 889   // bounds
 890   l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn));
 891 
 892   if (strip_mine_loop) {
 893     l->mark_strip_mined();
 894     l->verify_strip_mined(1);
 895     outer_ilt->_head->as_Loop()->verify_strip_mined(1);
 896     loop = outer_ilt;
 897   }
 898 
 899   return true;
 900 }
 901 
 902 //----------------------exact_limit-------------------------------------------
 903 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
 904   assert(loop->_head->is_CountedLoop(), "");
 905   CountedLoopNode *cl = loop->_head->as_CountedLoop();
 906   assert(cl->is_valid_counted_loop(), "");
 907 
 908   if (ABS(cl->stride_con()) == 1 ||
 909       cl->limit()->Opcode() == Op_LoopLimit) {
 910     // Old code has exact limit (it could be incorrect in case of int overflow).
 911     // Loop limit is exact with stride == 1. And loop may already have exact limit.
 912     return cl->limit();
 913   }
 914   Node *limit = NULL;
 915 #ifdef ASSERT
 916   BoolTest::mask bt = cl->loopexit()->test_trip();
 917   assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
 918 #endif
 919   if (cl->has_exact_trip_count()) {
 920     // Simple case: loop has constant boundaries.
 921     // Use jlongs to avoid integer overflow.
 922     int stride_con = cl->stride_con();
 923     jlong  init_con = cl->init_trip()->get_int();
 924     jlong limit_con = cl->limit()->get_int();
 925     julong trip_cnt = cl->trip_count();
 926     jlong final_con = init_con + trip_cnt*stride_con;
 927     int final_int = (int)final_con;
 928     // The final value should be in integer range since the loop
 929     // is counted and the limit was checked for overflow.
 930     assert(final_con == (jlong)final_int, "final value should be integer");
 931     limit = _igvn.intcon(final_int);
 932   } else {
 933     // Create new LoopLimit node to get exact limit (final iv value).
 934     limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
 935     register_new_node(limit, cl->in(LoopNode::EntryControl));
 936   }
 937   assert(limit != NULL, "sanity");
 938   return limit;
 939 }
 940 
 941 //------------------------------Ideal------------------------------------------
 942 // Return a node which is more "ideal" than the current node.
 943 // Attempt to convert into a counted-loop.
 944 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
 945   if (!can_be_counted_loop(phase) && !is_OuterStripMinedLoop()) {
 946     phase->C->set_major_progress();
 947   }
 948   return RegionNode::Ideal(phase, can_reshape);
 949 }
 950 
 951 #ifdef ASSERT
 952 void LoopNode::verify_strip_mined(int expect_skeleton) const {
 953   const OuterStripMinedLoopNode* outer = NULL;
 954   const CountedLoopNode* inner = NULL;
 955   if (is_strip_mined()) {
 956     if (!is_valid_counted_loop()) {
 957       return; // Skip malformed counted loop
 958     }
 959     assert(is_CountedLoop(), "no Loop should be marked strip mined");
 960     inner = as_CountedLoop();
 961     outer = inner->in(LoopNode::EntryControl)->as_OuterStripMinedLoop();
 962   } else if (is_OuterStripMinedLoop()) {
 963     outer = this->as_OuterStripMinedLoop();
 964     inner = outer->unique_ctrl_out()->as_CountedLoop();
 965     assert(inner->is_valid_counted_loop() && inner->is_strip_mined(), "OuterStripMinedLoop should have been removed");
 966     assert(!is_strip_mined(), "outer loop shouldn't be marked strip mined");
 967   }
 968   if (inner != NULL || outer != NULL) {
 969     assert(inner != NULL && outer != NULL, "missing loop in strip mined nest");
 970     Node* outer_tail = outer->in(LoopNode::LoopBackControl);
 971     Node* outer_le = outer_tail->in(0);
 972     assert(outer_le->Opcode() == Op_OuterStripMinedLoopEnd, "tail of outer loop should be an If");
 973     Node* sfpt = outer_le->in(0);
 974     assert(sfpt->Opcode() == Op_SafePoint, "where's the safepoint?");
 975     Node* inner_out = sfpt->in(0);
 976     if (inner_out->outcnt() != 1) {
 977       ResourceMark rm;
 978       Unique_Node_List wq;
 979 
 980       for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
 981         Node* u = inner_out->fast_out(i);
 982         if (u == sfpt) {
 983           continue;
 984         }
 985         wq.clear();
 986         wq.push(u);
 987         bool found_sfpt = false;
 988         for (uint next = 0; next < wq.size() && !found_sfpt; next++) {
 989           Node *n = wq.at(next);
 990           for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && !found_sfpt; i++) {
 991             Node* u = n->fast_out(i);
 992             if (u == sfpt) {
 993               found_sfpt = true;
 994             }
 995             if (!u->is_CFG()) {
 996               wq.push(u);
 997             }
 998           }
 999         }
1000         assert(found_sfpt, "no node in loop that's not input to safepoint");
1001       }
1002     }
1003     CountedLoopEndNode* cle = inner_out->in(0)->as_CountedLoopEnd();
1004     assert(cle == inner->loopexit_or_null(), "mismatch");
1005     bool has_skeleton = outer_le->in(1)->bottom_type()->singleton() && outer_le->in(1)->bottom_type()->is_int()->get_con() == 0;
1006     if (has_skeleton) {
1007       assert(expect_skeleton == 1 || expect_skeleton == -1, "unexpected skeleton node");
1008       assert(outer->outcnt() == 2, "only phis");
1009     } else {
1010       assert(expect_skeleton == 0 || expect_skeleton == -1, "no skeleton node?");
1011       uint phis = 0;
1012       for (DUIterator_Fast imax, i = inner->fast_outs(imax); i < imax; i++) {
1013         Node* u = inner->fast_out(i);
1014         if (u->is_Phi()) {
1015           phis++;
1016         }
1017       }
1018       for (DUIterator_Fast imax, i = outer->fast_outs(imax); i < imax; i++) {
1019         Node* u = outer->fast_out(i);
1020         assert(u == outer || u == inner || u->is_Phi(), "nothing between inner and outer loop");
1021       }
1022       uint stores = 0;
1023       for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {
1024         Node* u = inner_out->fast_out(i);
1025         if (u->is_Store()) {
1026           stores++;
1027         }
1028       }
1029       assert(outer->outcnt() >= phis + 2 && outer->outcnt() <= phis + 2 + stores + 1, "only phis");
1030     }
1031     assert(sfpt->outcnt() == 1, "no data node");
1032     assert(outer_tail->outcnt() == 1 || !has_skeleton, "no data node");
1033   }
1034 }
1035 #endif
1036 
1037 //=============================================================================
1038 //------------------------------Ideal------------------------------------------
1039 // Return a node which is more "ideal" than the current node.
1040 // Attempt to convert into a counted-loop.
1041 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1042   return RegionNode::Ideal(phase, can_reshape);
1043 }
1044 
1045 //------------------------------dump_spec--------------------------------------
1046 // Dump special per-node info
1047 #ifndef PRODUCT
1048 void CountedLoopNode::dump_spec(outputStream *st) const {
1049   LoopNode::dump_spec(st);
1050   if (stride_is_con()) {
1051     st->print("stride: %d ",stride_con());
1052   }
1053   if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
1054   if (is_main_loop()) st->print("main of N%d", _idx);
1055   if (is_post_loop()) st->print("post of N%d", _main_idx);
1056   if (is_strip_mined()) st->print(" strip mined");
1057 }
1058 #endif
1059 
1060 //=============================================================================
1061 int CountedLoopEndNode::stride_con() const {
1062   return stride()->bottom_type()->is_int()->get_con();
1063 }
1064 
1065 //=============================================================================
1066 //------------------------------Value-----------------------------------------
1067 const Type* LoopLimitNode::Value(PhaseGVN* phase) const {
1068   const Type* init_t   = phase->type(in(Init));
1069   const Type* limit_t  = phase->type(in(Limit));
1070   const Type* stride_t = phase->type(in(Stride));
1071   // Either input is TOP ==> the result is TOP
1072   if (init_t   == Type::TOP) return Type::TOP;
1073   if (limit_t  == Type::TOP) return Type::TOP;
1074   if (stride_t == Type::TOP) return Type::TOP;
1075 
1076   int stride_con = stride_t->is_int()->get_con();
1077   if (stride_con == 1)
1078     return NULL;  // Identity
1079 
1080   if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
1081     // Use jlongs to avoid integer overflow.
1082     jlong init_con   =  init_t->is_int()->get_con();
1083     jlong limit_con  = limit_t->is_int()->get_con();
1084     int  stride_m   = stride_con - (stride_con > 0 ? 1 : -1);
1085     jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
1086     jlong final_con  = init_con + stride_con*trip_count;
1087     int final_int = (int)final_con;
1088     // The final value should be in integer range since the loop
1089     // is counted and the limit was checked for overflow.
1090     assert(final_con == (jlong)final_int, "final value should be integer");
1091     return TypeInt::make(final_int);
1092   }
1093 
1094   return bottom_type(); // TypeInt::INT
1095 }
1096 
1097 //------------------------------Ideal------------------------------------------
1098 // Return a node which is more "ideal" than the current node.
1099 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1100   if (phase->type(in(Init))   == Type::TOP ||
1101       phase->type(in(Limit))  == Type::TOP ||
1102       phase->type(in(Stride)) == Type::TOP)
1103     return NULL;  // Dead
1104 
1105   int stride_con = phase->type(in(Stride))->is_int()->get_con();
1106   if (stride_con == 1)
1107     return NULL;  // Identity
1108 
1109   if (in(Init)->is_Con() && in(Limit)->is_Con())
1110     return NULL;  // Value
1111 
1112   // Delay following optimizations until all loop optimizations
1113   // done to keep Ideal graph simple.
1114   if (!can_reshape || phase->C->major_progress())
1115     return NULL;
1116 
1117   const TypeInt* init_t  = phase->type(in(Init) )->is_int();
1118   const TypeInt* limit_t = phase->type(in(Limit))->is_int();
1119   int stride_p;
1120   jlong lim, ini;
1121   julong max;
1122   if (stride_con > 0) {
1123     stride_p = stride_con;
1124     lim = limit_t->_hi;
1125     ini = init_t->_lo;
1126     max = (julong)max_jint;
1127   } else {
1128     stride_p = -stride_con;
1129     lim = init_t->_hi;
1130     ini = limit_t->_lo;
1131     max = (julong)min_jint;
1132   }
1133   julong range = lim - ini + stride_p;
1134   if (range <= max) {
1135     // Convert to integer expression if it is not overflow.
1136     Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
1137     Node *range = phase->transform(new SubINode(in(Limit), in(Init)));
1138     Node *bias  = phase->transform(new AddINode(range, stride_m));
1139     Node *trip  = phase->transform(new DivINode(0, bias, in(Stride)));
1140     Node *span  = phase->transform(new MulINode(trip, in(Stride)));
1141     return new AddINode(span, in(Init)); // exact limit
1142   }
1143 
1144   if (is_power_of_2(stride_p) ||                // divisor is 2^n
1145       !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
1146     // Convert to long expression to avoid integer overflow
1147     // and let igvn optimizer convert this division.
1148     //
1149     Node*   init   = phase->transform( new ConvI2LNode(in(Init)));
1150     Node*  limit   = phase->transform( new ConvI2LNode(in(Limit)));
1151     Node* stride   = phase->longcon(stride_con);
1152     Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
1153 
1154     Node *range = phase->transform(new SubLNode(limit, init));
1155     Node *bias  = phase->transform(new AddLNode(range, stride_m));
1156     Node *span;
1157     if (stride_con > 0 && is_power_of_2(stride_p)) {
1158       // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
1159       // and avoid generating rounding for division. Zero trip guard should
1160       // guarantee that init < limit but sometimes the guard is missing and
1161       // we can get situation when init > limit. Note, for the empty loop
1162       // optimization zero trip guard is generated explicitly which leaves
1163       // only RCE predicate where exact limit is used and the predicate
1164       // will simply fail forcing recompilation.
1165       Node* neg_stride   = phase->longcon(-stride_con);
1166       span = phase->transform(new AndLNode(bias, neg_stride));
1167     } else {
1168       Node *trip  = phase->transform(new DivLNode(0, bias, stride));
1169       span = phase->transform(new MulLNode(trip, stride));
1170     }
1171     // Convert back to int
1172     Node *span_int = phase->transform(new ConvL2INode(span));
1173     return new AddINode(span_int, in(Init)); // exact limit
1174   }
1175 
1176   return NULL;    // No progress
1177 }
1178 
1179 //------------------------------Identity---------------------------------------
1180 // If stride == 1 return limit node.
1181 Node* LoopLimitNode::Identity(PhaseGVN* phase) {
1182   int stride_con = phase->type(in(Stride))->is_int()->get_con();
1183   if (stride_con == 1 || stride_con == -1)
1184     return in(Limit);
1185   return this;
1186 }
1187 
1188 //=============================================================================
1189 //----------------------match_incr_with_optional_truncation--------------------
1190 // Match increment with optional truncation:
1191 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
1192 // Return NULL for failure. Success returns the increment node.
1193 Node* CountedLoopNode::match_incr_with_optional_truncation(
1194                       Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
1195   // Quick cutouts:
1196   if (expr == NULL || expr->req() != 3)  return NULL;
1197 
1198   Node *t1 = NULL;
1199   Node *t2 = NULL;
1200   const TypeInt* trunc_t = TypeInt::INT;
1201   Node* n1 = expr;
1202   int   n1op = n1->Opcode();
1203 
1204   // Try to strip (n1 & M) or (n1 << N >> N) from n1.
1205   if (n1op == Op_AndI &&
1206       n1->in(2)->is_Con() &&
1207       n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
1208     // %%% This check should match any mask of 2**K-1.
1209     t1 = n1;
1210     n1 = t1->in(1);
1211     n1op = n1->Opcode();
1212     trunc_t = TypeInt::CHAR;
1213   } else if (n1op == Op_RShiftI &&
1214              n1->in(1) != NULL &&
1215              n1->in(1)->Opcode() == Op_LShiftI &&
1216              n1->in(2) == n1->in(1)->in(2) &&
1217              n1->in(2)->is_Con()) {
1218     jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
1219     // %%% This check should match any shift in [1..31].
1220     if (shift == 16 || shift == 8) {
1221       t1 = n1;
1222       t2 = t1->in(1);
1223       n1 = t2->in(1);
1224       n1op = n1->Opcode();
1225       if (shift == 16) {
1226         trunc_t = TypeInt::SHORT;
1227       } else if (shift == 8) {
1228         trunc_t = TypeInt::BYTE;
1229       }
1230     }
1231   }
1232 
1233   // If (maybe after stripping) it is an AddI, we won:
1234   if (n1op == Op_AddI) {
1235     *trunc1 = t1;
1236     *trunc2 = t2;
1237     *trunc_type = trunc_t;
1238     return n1;
1239   }
1240 
1241   // failed
1242   return NULL;
1243 }
1244 
1245 LoopNode* CountedLoopNode::skip_strip_mined(int expect_skeleton) {
1246   if (is_strip_mined() && is_valid_counted_loop()) {
1247     verify_strip_mined(expect_skeleton);
1248     return in(EntryControl)->as_Loop();
1249   }
1250   return this;
1251 }
1252 
1253 OuterStripMinedLoopNode* CountedLoopNode::outer_loop() const {
1254   assert(is_strip_mined(), "not a strip mined loop");
1255   Node* c = in(EntryControl);
1256   if (c == NULL || c->is_top() || !c->is_OuterStripMinedLoop()) {
1257     return NULL;
1258   }
1259   return c->as_OuterStripMinedLoop();
1260 }
1261 
1262 IfTrueNode* OuterStripMinedLoopNode::outer_loop_tail() const {
1263   Node* c = in(LoopBackControl);
1264   if (c == NULL || c->is_top()) {
1265     return NULL;
1266   }
1267   return c->as_IfTrue();
1268 }
1269 
1270 IfTrueNode* CountedLoopNode::outer_loop_tail() const {
1271   LoopNode* l = outer_loop();
1272   if (l == NULL) {
1273     return NULL;
1274   }
1275   return l->outer_loop_tail();
1276 }
1277 
1278 OuterStripMinedLoopEndNode* OuterStripMinedLoopNode::outer_loop_end() const {
1279   IfTrueNode* proj = outer_loop_tail();
1280   if (proj == NULL) {
1281     return NULL;
1282   }
1283   Node* c = proj->in(0);
1284   if (c == NULL || c->is_top() || c->outcnt() != 2) {
1285     return NULL;
1286   }
1287   return c->as_OuterStripMinedLoopEnd();
1288 }
1289 
1290 OuterStripMinedLoopEndNode* CountedLoopNode::outer_loop_end() const {
1291   LoopNode* l = outer_loop();
1292   if (l == NULL) {
1293     return NULL;
1294   }
1295   return l->outer_loop_end();
1296 }
1297 
1298 IfFalseNode* OuterStripMinedLoopNode::outer_loop_exit() const {
1299   IfNode* le = outer_loop_end();
1300   if (le == NULL) {
1301     return NULL;
1302   }
1303   Node* c = le->proj_out_or_null(false);
1304   if (c == NULL) {
1305     return NULL;
1306   }
1307   return c->as_IfFalse();
1308 }
1309 
1310 IfFalseNode* CountedLoopNode::outer_loop_exit() const {
1311   LoopNode* l = outer_loop();
1312   if (l == NULL) {
1313     return NULL;
1314   }
1315   return l->outer_loop_exit();
1316 }
1317 
1318 SafePointNode* OuterStripMinedLoopNode::outer_safepoint() const {
1319   IfNode* le = outer_loop_end();
1320   if (le == NULL) {
1321     return NULL;
1322   }
1323   Node* c = le->in(0);
1324   if (c == NULL || c->is_top()) {
1325     return NULL;
1326   }
1327   assert(c->Opcode() == Op_SafePoint, "broken outer loop");
1328   return c->as_SafePoint();
1329 }
1330 
1331 SafePointNode* CountedLoopNode::outer_safepoint() const {
1332   LoopNode* l = outer_loop();
1333   if (l == NULL) {
1334     return NULL;
1335   }
1336   return l->outer_safepoint();
1337 }
1338 
1339 Node* CountedLoopNode::skip_predicates_from_entry(Node* ctrl) {
1340     while (ctrl != NULL && ctrl->is_Proj() && ctrl->in(0)->is_If() &&
1341            ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->outcnt() == 1 &&
1342            ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->unique_out()->Opcode() == Op_Halt) {
1343       ctrl = ctrl->in(0)->in(0);
1344     }
1345 
1346     return ctrl;
1347   }
1348 
1349 Node* CountedLoopNode::skip_predicates() {
1350   if (is_main_loop()) {
1351     Node* ctrl = skip_strip_mined()->in(LoopNode::EntryControl);
1352 
1353     return skip_predicates_from_entry(ctrl);
1354   }
1355   return in(LoopNode::EntryControl);
1356 }
1357 
1358 void OuterStripMinedLoopNode::adjust_strip_mined_loop(PhaseIterGVN* igvn) {
1359   // Look for the outer & inner strip mined loop, reduce number of
1360   // iterations of the inner loop, set exit condition of outer loop,
1361   // construct required phi nodes for outer loop.
1362   CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
1363   assert(inner_cl->is_strip_mined(), "inner loop should be strip mined");
1364   Node* inner_iv_phi = inner_cl->phi();
1365   if (inner_iv_phi == NULL) {
1366     IfNode* outer_le = outer_loop_end();
1367     Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1368     igvn->replace_node(outer_le, iff);
1369     inner_cl->clear_strip_mined();
1370     return;
1371   }
1372   CountedLoopEndNode* inner_cle = inner_cl->loopexit();
1373 
1374   int stride = inner_cl->stride_con();
1375   jlong scaled_iters_long = ((jlong)LoopStripMiningIter) * ABS(stride);
1376   int scaled_iters = (int)scaled_iters_long;
1377   int short_scaled_iters = LoopStripMiningIterShortLoop* ABS(stride);
1378   const TypeInt* inner_iv_t = igvn->type(inner_iv_phi)->is_int();
1379   jlong iter_estimate = (jlong)inner_iv_t->_hi - (jlong)inner_iv_t->_lo;
1380   assert(iter_estimate > 0, "broken");
1381   if ((jlong)scaled_iters != scaled_iters_long || iter_estimate <= short_scaled_iters) {
1382     // Remove outer loop and safepoint (too few iterations)
1383     Node* outer_sfpt = outer_safepoint();
1384     Node* outer_out = outer_loop_exit();
1385     igvn->replace_node(outer_out, outer_sfpt->in(0));
1386     igvn->replace_input_of(outer_sfpt, 0, igvn->C->top());
1387     inner_cl->clear_strip_mined();
1388     return;
1389   }
1390   if (iter_estimate <= scaled_iters_long) {
1391     // We would only go through one iteration of
1392     // the outer loop: drop the outer loop but
1393     // keep the safepoint so we don't run for
1394     // too long without a safepoint
1395     IfNode* outer_le = outer_loop_end();
1396     Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1397     igvn->replace_node(outer_le, iff);
1398     inner_cl->clear_strip_mined();
1399     return;
1400   }
1401 
1402   Node* cle_tail = inner_cle->proj_out(true);
1403   ResourceMark rm;
1404   Node_List old_new;
1405   if (cle_tail->outcnt() > 1) {
1406     // Look for nodes on backedge of inner loop and clone them
1407     Unique_Node_List backedge_nodes;
1408     for (DUIterator_Fast imax, i = cle_tail->fast_outs(imax); i < imax; i++) {
1409       Node* u = cle_tail->fast_out(i);
1410       if (u != inner_cl) {
1411         assert(!u->is_CFG(), "control flow on the backedge?");
1412         backedge_nodes.push(u);
1413       }
1414     }
1415     uint last = igvn->C->unique();
1416     for (uint next = 0; next < backedge_nodes.size(); next++) {
1417       Node* n = backedge_nodes.at(next);
1418       old_new.map(n->_idx, n->clone());
1419       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1420         Node* u = n->fast_out(i);
1421         assert(!u->is_CFG(), "broken");
1422         if (u->_idx >= last) {
1423           continue;
1424         }
1425         if (!u->is_Phi()) {
1426           backedge_nodes.push(u);
1427         } else {
1428           assert(u->in(0) == inner_cl, "strange phi on the backedge");
1429         }
1430       }
1431     }
1432     // Put the clones on the outer loop backedge
1433     Node* le_tail = outer_loop_tail();
1434     for (uint next = 0; next < backedge_nodes.size(); next++) {
1435       Node *n = old_new[backedge_nodes.at(next)->_idx];
1436       for (uint i = 1; i < n->req(); i++) {
1437         if (n->in(i) != NULL && old_new[n->in(i)->_idx] != NULL) {
1438           n->set_req(i, old_new[n->in(i)->_idx]);
1439         }
1440       }
1441       if (n->in(0) != NULL && n->in(0) == cle_tail) {
1442         n->set_req(0, le_tail);
1443       }
1444       igvn->register_new_node_with_optimizer(n);
1445     }
1446   }
1447 
1448   Node* iv_phi = NULL;
1449   // Make a clone of each phi in the inner loop
1450   // for the outer loop
1451   for (uint i = 0; i < inner_cl->outcnt(); i++) {
1452     Node* u = inner_cl->raw_out(i);
1453     if (u->is_Phi()) {
1454       assert(u->in(0) == inner_cl, "inconsistent");
1455       Node* phi = u->clone();
1456       phi->set_req(0, this);
1457       Node* be = old_new[phi->in(LoopNode::LoopBackControl)->_idx];
1458       if (be != NULL) {
1459         phi->set_req(LoopNode::LoopBackControl, be);
1460       }
1461       phi = igvn->transform(phi);
1462       igvn->replace_input_of(u, LoopNode::EntryControl, phi);
1463       if (u == inner_iv_phi) {
1464         iv_phi = phi;
1465       }
1466     }
1467   }
1468   Node* cle_out = inner_cle->proj_out(false);
1469   if (cle_out->outcnt() > 1) {
1470     // Look for chains of stores that were sunk
1471     // out of the inner loop and are in the outer loop
1472     for (DUIterator_Fast imax, i = cle_out->fast_outs(imax); i < imax; i++) {
1473       Node* u = cle_out->fast_out(i);
1474       if (u->is_Store()) {
1475         Node* first = u;
1476         for(;;) {
1477           Node* next = first->in(MemNode::Memory);
1478           if (!next->is_Store() || next->in(0) != cle_out) {
1479             break;
1480           }
1481           first = next;
1482         }
1483         Node* last = u;
1484         for(;;) {
1485           Node* next = NULL;
1486           for (DUIterator_Fast jmax, j = last->fast_outs(jmax); j < jmax; j++) {
1487             Node* uu = last->fast_out(j);
1488             if (uu->is_Store() && uu->in(0) == cle_out) {
1489               assert(next == NULL, "only one in the outer loop");
1490               next = uu;
1491             }
1492           }
1493           if (next == NULL) {
1494             break;
1495           }
1496           last = next;
1497         }
1498         Node* phi = NULL;
1499         for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) {
1500           Node* uu = fast_out(j);
1501           if (uu->is_Phi()) {
1502             Node* be = uu->in(LoopNode::LoopBackControl);
1503             if (be->is_Store() && old_new[be->_idx] != NULL) {
1504               assert(false, "store on the backedge + sunk stores: unsupported");
1505               // drop outer loop
1506               IfNode* outer_le = outer_loop_end();
1507               Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1508               igvn->replace_node(outer_le, iff);
1509               inner_cl->clear_strip_mined();
1510               return;
1511             }
1512             if (be == last || be == first->in(MemNode::Memory)) {
1513               assert(phi == NULL, "only one phi");
1514               phi = uu;
1515             }
1516           }
1517         }
1518 #ifdef ASSERT
1519         for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) {
1520           Node* uu = fast_out(j);
1521           if (uu->is_Phi() && uu->bottom_type() == Type::MEMORY) {
1522             if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) {
1523               assert(phi == uu, "what's that phi?");
1524             } else if (uu->adr_type() == TypePtr::BOTTOM) {
1525               Node* n = uu->in(LoopNode::LoopBackControl);
1526               uint limit = igvn->C->live_nodes();
1527               uint i = 0;
1528               while (n != uu) {
1529                 i++;
1530                 assert(i < limit, "infinite loop");
1531                 if (n->is_Proj()) {
1532                   n = n->in(0);
1533                 } else if (n->is_SafePoint() || n->is_MemBar()) {
1534                   n = n->in(TypeFunc::Memory);
1535                 } else if (n->is_Phi()) {
1536                   n = n->in(1);
1537                 } else if (n->is_MergeMem()) {
1538                   n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type()));
1539                 } else if (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) {
1540                   n = n->in(MemNode::Memory);
1541                 } else {
1542                   n->dump();
1543                   ShouldNotReachHere();
1544                 }
1545               }
1546             }
1547           }
1548         }
1549 #endif
1550         if (phi == NULL) {
1551           // If the an entire chains was sunk, the
1552           // inner loop has no phi for that memory
1553           // slice, create one for the outer loop
1554           phi = PhiNode::make(this, first->in(MemNode::Memory), Type::MEMORY,
1555                               igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type())));
1556           phi->set_req(LoopNode::LoopBackControl, last);
1557           phi = igvn->transform(phi);
1558           igvn->replace_input_of(first, MemNode::Memory, phi);
1559         } else {
1560           // Or fix the outer loop fix to include
1561           // that chain of stores.
1562           Node* be = phi->in(LoopNode::LoopBackControl);
1563           assert(!(be->is_Store() && old_new[be->_idx] != NULL), "store on the backedge + sunk stores: unsupported");
1564           if (be == first->in(MemNode::Memory)) {
1565             if (be == phi->in(LoopNode::LoopBackControl)) {
1566               igvn->replace_input_of(phi, LoopNode::LoopBackControl, last);
1567             } else {
1568               igvn->replace_input_of(be, MemNode::Memory, last);
1569             }
1570           } else {
1571 #ifdef ASSERT
1572             if (be == phi->in(LoopNode::LoopBackControl)) {
1573               assert(phi->in(LoopNode::LoopBackControl) == last, "");
1574             } else {
1575               assert(be->in(MemNode::Memory) == last, "");
1576             }
1577 #endif
1578           }
1579         }
1580       }
1581     }
1582   }
1583 
1584   if (iv_phi != NULL) {
1585     // Now adjust the inner loop's exit condition
1586     Node* limit = inner_cl->limit();
1587     Node* sub = NULL;
1588     if (stride > 0) {
1589       sub = igvn->transform(new SubINode(limit, iv_phi));
1590     } else {
1591       sub = igvn->transform(new SubINode(iv_phi, limit));
1592     }
1593     Node* min = igvn->transform(new MinINode(sub, igvn->intcon(scaled_iters)));
1594     Node* new_limit = NULL;
1595     if (stride > 0) {
1596       new_limit = igvn->transform(new AddINode(min, iv_phi));
1597     } else {
1598       new_limit = igvn->transform(new SubINode(iv_phi, min));
1599     }
1600     Node* inner_cmp = inner_cle->cmp_node();
1601     Node* inner_bol = inner_cle->in(CountedLoopEndNode::TestValue);
1602     Node* outer_bol = inner_bol;
1603     // cmp node for inner loop may be shared
1604     inner_cmp = inner_cmp->clone();
1605     inner_cmp->set_req(2, new_limit);
1606     inner_bol = inner_bol->clone();
1607     inner_bol->set_req(1, igvn->transform(inner_cmp));
1608     igvn->replace_input_of(inner_cle, CountedLoopEndNode::TestValue, igvn->transform(inner_bol));
1609     // Set the outer loop's exit condition too
1610     igvn->replace_input_of(outer_loop_end(), 1, outer_bol);
1611   } else {
1612     assert(false, "should be able to adjust outer loop");
1613     IfNode* outer_le = outer_loop_end();
1614     Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
1615     igvn->replace_node(outer_le, iff);
1616     inner_cl->clear_strip_mined();
1617   }
1618 }
1619 
1620 const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const {
1621   if (!in(0)) return Type::TOP;
1622   if (phase->type(in(0)) == Type::TOP)
1623     return Type::TOP;
1624 
1625   return TypeTuple::IFBOTH;
1626 }
1627 
1628 Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1629   if (remove_dead_region(phase, can_reshape))  return this;
1630 
1631   return NULL;
1632 }
1633 
1634 //------------------------------filtered_type--------------------------------
1635 // Return a type based on condition control flow
1636 // A successful return will be a type that is restricted due
1637 // to a series of dominating if-tests, such as:
1638 //    if (i < 10) {
1639 //       if (i > 0) {
1640 //          here: "i" type is [1..10)
1641 //       }
1642 //    }
1643 // or a control flow merge
1644 //    if (i < 10) {
1645 //       do {
1646 //          phi( , ) -- at top of loop type is [min_int..10)
1647 //         i = ?
1648 //       } while ( i < 10)
1649 //
1650 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
1651   assert(n && n->bottom_type()->is_int(), "must be int");
1652   const TypeInt* filtered_t = NULL;
1653   if (!n->is_Phi()) {
1654     assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
1655     filtered_t = filtered_type_from_dominators(n, n_ctrl);
1656 
1657   } else {
1658     Node* phi    = n->as_Phi();
1659     Node* region = phi->in(0);
1660     assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
1661     if (region && region != C->top()) {
1662       for (uint i = 1; i < phi->req(); i++) {
1663         Node* val   = phi->in(i);
1664         Node* use_c = region->in(i);
1665         const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
1666         if (val_t != NULL) {
1667           if (filtered_t == NULL) {
1668             filtered_t = val_t;
1669           } else {
1670             filtered_t = filtered_t->meet(val_t)->is_int();
1671           }
1672         }
1673       }
1674     }
1675   }
1676   const TypeInt* n_t = _igvn.type(n)->is_int();
1677   if (filtered_t != NULL) {
1678     n_t = n_t->join(filtered_t)->is_int();
1679   }
1680   return n_t;
1681 }
1682 
1683 
1684 //------------------------------filtered_type_from_dominators--------------------------------
1685 // Return a possibly more restrictive type for val based on condition control flow of dominators
1686 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
1687   if (val->is_Con()) {
1688      return val->bottom_type()->is_int();
1689   }
1690   uint if_limit = 10; // Max number of dominating if's visited
1691   const TypeInt* rtn_t = NULL;
1692 
1693   if (use_ctrl && use_ctrl != C->top()) {
1694     Node* val_ctrl = get_ctrl(val);
1695     uint val_dom_depth = dom_depth(val_ctrl);
1696     Node* pred = use_ctrl;
1697     uint if_cnt = 0;
1698     while (if_cnt < if_limit) {
1699       if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
1700         if_cnt++;
1701         const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
1702         if (if_t != NULL) {
1703           if (rtn_t == NULL) {
1704             rtn_t = if_t;
1705           } else {
1706             rtn_t = rtn_t->join(if_t)->is_int();
1707           }
1708         }
1709       }
1710       pred = idom(pred);
1711       if (pred == NULL || pred == C->top()) {
1712         break;
1713       }
1714       // Stop if going beyond definition block of val
1715       if (dom_depth(pred) < val_dom_depth) {
1716         break;
1717       }
1718     }
1719   }
1720   return rtn_t;
1721 }
1722 
1723 
1724 //------------------------------dump_spec--------------------------------------
1725 // Dump special per-node info
1726 #ifndef PRODUCT
1727 void CountedLoopEndNode::dump_spec(outputStream *st) const {
1728   if( in(TestValue) != NULL && in(TestValue)->is_Bool() ) {
1729     BoolTest bt( test_trip()); // Added this for g++.
1730 
1731     st->print("[");
1732     bt.dump_on(st);
1733     st->print("]");
1734   }
1735   st->print(" ");
1736   IfNode::dump_spec(st);
1737 }
1738 #endif
1739 
1740 //=============================================================================
1741 //------------------------------is_member--------------------------------------
1742 // Is 'l' a member of 'this'?
1743 bool IdealLoopTree::is_member(const IdealLoopTree *l) const {
1744   while( l->_nest > _nest ) l = l->_parent;
1745   return l == this;
1746 }
1747 
1748 //------------------------------set_nest---------------------------------------
1749 // Set loop tree nesting depth.  Accumulate _has_call bits.
1750 int IdealLoopTree::set_nest( uint depth ) {
1751   _nest = depth;
1752   int bits = _has_call;
1753   if( _child ) bits |= _child->set_nest(depth+1);
1754   if( bits ) _has_call = 1;
1755   if( _next  ) bits |= _next ->set_nest(depth  );
1756   return bits;
1757 }
1758 
1759 //------------------------------split_fall_in----------------------------------
1760 // Split out multiple fall-in edges from the loop header.  Move them to a
1761 // private RegionNode before the loop.  This becomes the loop landing pad.
1762 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
1763   PhaseIterGVN &igvn = phase->_igvn;
1764   uint i;
1765 
1766   // Make a new RegionNode to be the landing pad.
1767   Node *landing_pad = new RegionNode( fall_in_cnt+1 );
1768   phase->set_loop(landing_pad,_parent);
1769   // Gather all the fall-in control paths into the landing pad
1770   uint icnt = fall_in_cnt;
1771   uint oreq = _head->req();
1772   for( i = oreq-1; i>0; i-- )
1773     if( !phase->is_member( this, _head->in(i) ) )
1774       landing_pad->set_req(icnt--,_head->in(i));
1775 
1776   // Peel off PhiNode edges as well
1777   for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1778     Node *oj = _head->fast_out(j);
1779     if( oj->is_Phi() ) {
1780       PhiNode* old_phi = oj->as_Phi();
1781       assert( old_phi->region() == _head, "" );
1782       igvn.hash_delete(old_phi);   // Yank from hash before hacking edges
1783       Node *p = PhiNode::make_blank(landing_pad, old_phi);
1784       uint icnt = fall_in_cnt;
1785       for( i = oreq-1; i>0; i-- ) {
1786         if( !phase->is_member( this, _head->in(i) ) ) {
1787           p->init_req(icnt--, old_phi->in(i));
1788           // Go ahead and clean out old edges from old phi
1789           old_phi->del_req(i);
1790         }
1791       }
1792       // Search for CSE's here, because ZKM.jar does a lot of
1793       // loop hackery and we need to be a little incremental
1794       // with the CSE to avoid O(N^2) node blow-up.
1795       Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
1796       if( p2 ) {                // Found CSE
1797         p->destruct();          // Recover useless new node
1798         p = p2;                 // Use old node
1799       } else {
1800         igvn.register_new_node_with_optimizer(p, old_phi);
1801       }
1802       // Make old Phi refer to new Phi.
1803       old_phi->add_req(p);
1804       // Check for the special case of making the old phi useless and
1805       // disappear it.  In JavaGrande I have a case where this useless
1806       // Phi is the loop limit and prevents recognizing a CountedLoop
1807       // which in turn prevents removing an empty loop.
1808       Node *id_old_phi = igvn.apply_identity(old_phi);
1809       if( id_old_phi != old_phi ) { // Found a simple identity?
1810         // Note that I cannot call 'replace_node' here, because
1811         // that will yank the edge from old_phi to the Region and
1812         // I'm mid-iteration over the Region's uses.
1813         for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
1814           Node* use = old_phi->last_out(i);
1815           igvn.rehash_node_delayed(use);
1816           uint uses_found = 0;
1817           for (uint j = 0; j < use->len(); j++) {
1818             if (use->in(j) == old_phi) {
1819               if (j < use->req()) use->set_req (j, id_old_phi);
1820               else                use->set_prec(j, id_old_phi);
1821               uses_found++;
1822             }
1823           }
1824           i -= uses_found;    // we deleted 1 or more copies of this edge
1825         }
1826       }
1827       igvn._worklist.push(old_phi);
1828     }
1829   }
1830   // Finally clean out the fall-in edges from the RegionNode
1831   for( i = oreq-1; i>0; i-- ) {
1832     if( !phase->is_member( this, _head->in(i) ) ) {
1833       _head->del_req(i);
1834     }
1835   }
1836   igvn.rehash_node_delayed(_head);
1837   // Transform landing pad
1838   igvn.register_new_node_with_optimizer(landing_pad, _head);
1839   // Insert landing pad into the header
1840   _head->add_req(landing_pad);
1841 }
1842 
1843 //------------------------------split_outer_loop-------------------------------
1844 // Split out the outermost loop from this shared header.
1845 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
1846   PhaseIterGVN &igvn = phase->_igvn;
1847 
1848   // Find index of outermost loop; it should also be my tail.
1849   uint outer_idx = 1;
1850   while( _head->in(outer_idx) != _tail ) outer_idx++;
1851 
1852   // Make a LoopNode for the outermost loop.
1853   Node *ctl = _head->in(LoopNode::EntryControl);
1854   Node *outer = new LoopNode( ctl, _head->in(outer_idx) );
1855   outer = igvn.register_new_node_with_optimizer(outer, _head);
1856   phase->set_created_loop_node();
1857 
1858   // Outermost loop falls into '_head' loop
1859   _head->set_req(LoopNode::EntryControl, outer);
1860   _head->del_req(outer_idx);
1861   // Split all the Phis up between '_head' loop and 'outer' loop.
1862   for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1863     Node *out = _head->fast_out(j);
1864     if( out->is_Phi() ) {
1865       PhiNode *old_phi = out->as_Phi();
1866       assert( old_phi->region() == _head, "" );
1867       Node *phi = PhiNode::make_blank(outer, old_phi);
1868       phi->init_req(LoopNode::EntryControl,    old_phi->in(LoopNode::EntryControl));
1869       phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
1870       phi = igvn.register_new_node_with_optimizer(phi, old_phi);
1871       // Make old Phi point to new Phi on the fall-in path
1872       igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
1873       old_phi->del_req(outer_idx);
1874     }
1875   }
1876 
1877   // Use the new loop head instead of the old shared one
1878   _head = outer;
1879   phase->set_loop(_head, this);
1880 }
1881 
1882 //------------------------------fix_parent-------------------------------------
1883 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
1884   loop->_parent = parent;
1885   if( loop->_child ) fix_parent( loop->_child, loop   );
1886   if( loop->_next  ) fix_parent( loop->_next , parent );
1887 }
1888 
1889 //------------------------------estimate_path_freq-----------------------------
1890 static float estimate_path_freq( Node *n ) {
1891   // Try to extract some path frequency info
1892   IfNode *iff;
1893   for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
1894     uint nop = n->Opcode();
1895     if( nop == Op_SafePoint ) {   // Skip any safepoint
1896       n = n->in(0);
1897       continue;
1898     }
1899     if( nop == Op_CatchProj ) {   // Get count from a prior call
1900       // Assume call does not always throw exceptions: means the call-site
1901       // count is also the frequency of the fall-through path.
1902       assert( n->is_CatchProj(), "" );
1903       if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
1904         return 0.0f;            // Assume call exception path is rare
1905       Node *call = n->in(0)->in(0)->in(0);
1906       assert( call->is_Call(), "expect a call here" );
1907       const JVMState *jvms = ((CallNode*)call)->jvms();
1908       ciMethodData* methodData = jvms->method()->method_data();
1909       if (!methodData->is_mature())  return 0.0f; // No call-site data
1910       ciProfileData* data = methodData->bci_to_data(jvms->bci());
1911       if ((data == NULL) || !data->is_CounterData()) {
1912         // no call profile available, try call's control input
1913         n = n->in(0);
1914         continue;
1915       }
1916       return data->as_CounterData()->count()/FreqCountInvocations;
1917     }
1918     // See if there's a gating IF test
1919     Node *n_c = n->in(0);
1920     if( !n_c->is_If() ) break;       // No estimate available
1921     iff = n_c->as_If();
1922     if( iff->_fcnt != COUNT_UNKNOWN )   // Have a valid count?
1923       // Compute how much count comes on this path
1924       return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
1925     // Have no count info.  Skip dull uncommon-trap like branches.
1926     if( (nop == Op_IfTrue  && iff->_prob < PROB_LIKELY_MAG(5)) ||
1927         (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
1928       break;
1929     // Skip through never-taken branch; look for a real loop exit.
1930     n = iff->in(0);
1931   }
1932   return 0.0f;                  // No estimate available
1933 }
1934 
1935 //------------------------------merge_many_backedges---------------------------
1936 // Merge all the backedges from the shared header into a private Region.
1937 // Feed that region as the one backedge to this loop.
1938 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
1939   uint i;
1940 
1941   // Scan for the top 2 hottest backedges
1942   float hotcnt = 0.0f;
1943   float warmcnt = 0.0f;
1944   uint hot_idx = 0;
1945   // Loop starts at 2 because slot 1 is the fall-in path
1946   for( i = 2; i < _head->req(); i++ ) {
1947     float cnt = estimate_path_freq(_head->in(i));
1948     if( cnt > hotcnt ) {       // Grab hottest path
1949       warmcnt = hotcnt;
1950       hotcnt = cnt;
1951       hot_idx = i;
1952     } else if( cnt > warmcnt ) { // And 2nd hottest path
1953       warmcnt = cnt;
1954     }
1955   }
1956 
1957   // See if the hottest backedge is worthy of being an inner loop
1958   // by being much hotter than the next hottest backedge.
1959   if( hotcnt <= 0.0001 ||
1960       hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
1961 
1962   // Peel out the backedges into a private merge point; peel
1963   // them all except optionally hot_idx.
1964   PhaseIterGVN &igvn = phase->_igvn;
1965 
1966   Node *hot_tail = NULL;
1967   // Make a Region for the merge point
1968   Node *r = new RegionNode(1);
1969   for( i = 2; i < _head->req(); i++ ) {
1970     if( i != hot_idx )
1971       r->add_req( _head->in(i) );
1972     else hot_tail = _head->in(i);
1973   }
1974   igvn.register_new_node_with_optimizer(r, _head);
1975   // Plug region into end of loop _head, followed by hot_tail
1976   while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
1977   igvn.replace_input_of(_head, 2, r);
1978   if( hot_idx ) _head->add_req(hot_tail);
1979 
1980   // Split all the Phis up between '_head' loop and the Region 'r'
1981   for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1982     Node *out = _head->fast_out(j);
1983     if( out->is_Phi() ) {
1984       PhiNode* n = out->as_Phi();
1985       igvn.hash_delete(n);      // Delete from hash before hacking edges
1986       Node *hot_phi = NULL;
1987       Node *phi = new PhiNode(r, n->type(), n->adr_type());
1988       // Check all inputs for the ones to peel out
1989       uint j = 1;
1990       for( uint i = 2; i < n->req(); i++ ) {
1991         if( i != hot_idx )
1992           phi->set_req( j++, n->in(i) );
1993         else hot_phi = n->in(i);
1994       }
1995       // Register the phi but do not transform until whole place transforms
1996       igvn.register_new_node_with_optimizer(phi, n);
1997       // Add the merge phi to the old Phi
1998       while( n->req() > 3 ) n->del_req( n->req()-1 );
1999       igvn.replace_input_of(n, 2, phi);
2000       if( hot_idx ) n->add_req(hot_phi);
2001     }
2002   }
2003 
2004 
2005   // Insert a new IdealLoopTree inserted below me.  Turn it into a clone
2006   // of self loop tree.  Turn self into a loop headed by _head and with
2007   // tail being the new merge point.
2008   IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
2009   phase->set_loop(_tail,ilt);   // Adjust tail
2010   _tail = r;                    // Self's tail is new merge point
2011   phase->set_loop(r,this);
2012   ilt->_child = _child;         // New guy has my children
2013   _child = ilt;                 // Self has new guy as only child
2014   ilt->_parent = this;          // new guy has self for parent
2015   ilt->_nest = _nest;           // Same nesting depth (for now)
2016 
2017   // Starting with 'ilt', look for child loop trees using the same shared
2018   // header.  Flatten these out; they will no longer be loops in the end.
2019   IdealLoopTree **pilt = &_child;
2020   while( ilt ) {
2021     if( ilt->_head == _head ) {
2022       uint i;
2023       for( i = 2; i < _head->req(); i++ )
2024         if( _head->in(i) == ilt->_tail )
2025           break;                // Still a loop
2026       if( i == _head->req() ) { // No longer a loop
2027         // Flatten ilt.  Hang ilt's "_next" list from the end of
2028         // ilt's '_child' list.  Move the ilt's _child up to replace ilt.
2029         IdealLoopTree **cp = &ilt->_child;
2030         while( *cp ) cp = &(*cp)->_next;   // Find end of child list
2031         *cp = ilt->_next;       // Hang next list at end of child list
2032         *pilt = ilt->_child;    // Move child up to replace ilt
2033         ilt->_head = NULL;      // Flag as a loop UNIONED into parent
2034         ilt = ilt->_child;      // Repeat using new ilt
2035         continue;               // do not advance over ilt->_child
2036       }
2037       assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
2038       phase->set_loop(_head,ilt);
2039     }
2040     pilt = &ilt->_child;        // Advance to next
2041     ilt = *pilt;
2042   }
2043 
2044   if( _child ) fix_parent( _child, this );
2045 }
2046 
2047 //------------------------------beautify_loops---------------------------------
2048 // Split shared headers and insert loop landing pads.
2049 // Insert a LoopNode to replace the RegionNode.
2050 // Return TRUE if loop tree is structurally changed.
2051 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
2052   bool result = false;
2053   // Cache parts in locals for easy
2054   PhaseIterGVN &igvn = phase->_igvn;
2055 
2056   igvn.hash_delete(_head);      // Yank from hash before hacking edges
2057 
2058   // Check for multiple fall-in paths.  Peel off a landing pad if need be.
2059   int fall_in_cnt = 0;
2060   for( uint i = 1; i < _head->req(); i++ )
2061     if( !phase->is_member( this, _head->in(i) ) )
2062       fall_in_cnt++;
2063   assert( fall_in_cnt, "at least 1 fall-in path" );
2064   if( fall_in_cnt > 1 )         // Need a loop landing pad to merge fall-ins
2065     split_fall_in( phase, fall_in_cnt );
2066 
2067   // Swap inputs to the _head and all Phis to move the fall-in edge to
2068   // the left.
2069   fall_in_cnt = 1;
2070   while( phase->is_member( this, _head->in(fall_in_cnt) ) )
2071     fall_in_cnt++;
2072   if( fall_in_cnt > 1 ) {
2073     // Since I am just swapping inputs I do not need to update def-use info
2074     Node *tmp = _head->in(1);
2075     igvn.rehash_node_delayed(_head);
2076     _head->set_req( 1, _head->in(fall_in_cnt) );
2077     _head->set_req( fall_in_cnt, tmp );
2078     // Swap also all Phis
2079     for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
2080       Node* phi = _head->fast_out(i);
2081       if( phi->is_Phi() ) {
2082         igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges
2083         tmp = phi->in(1);
2084         phi->set_req( 1, phi->in(fall_in_cnt) );
2085         phi->set_req( fall_in_cnt, tmp );
2086       }
2087     }
2088   }
2089   assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
2090   assert(  phase->is_member( this, _head->in(2) ), "right edge is loop" );
2091 
2092   // If I am a shared header (multiple backedges), peel off the many
2093   // backedges into a private merge point and use the merge point as
2094   // the one true backedge.
2095   if (_head->req() > 3) {
2096     // Merge the many backedges into a single backedge but leave
2097     // the hottest backedge as separate edge for the following peel.
2098     if (!_irreducible) {
2099       merge_many_backedges( phase );
2100     }
2101 
2102     // When recursively beautify my children, split_fall_in can change
2103     // loop tree structure when I am an irreducible loop. Then the head
2104     // of my children has a req() not bigger than 3. Here we need to set
2105     // result to true to catch that case in order to tell the caller to
2106     // rebuild loop tree. See issue JDK-8244407 for details.
2107     result = true;
2108   }
2109 
2110   // If I have one hot backedge, peel off myself loop.
2111   // I better be the outermost loop.
2112   if (_head->req() > 3 && !_irreducible) {
2113     split_outer_loop( phase );
2114     result = true;
2115 
2116   } else if (!_head->is_Loop() && !_irreducible) {
2117     // Make a new LoopNode to replace the old loop head
2118     Node *l = new LoopNode( _head->in(1), _head->in(2) );
2119     l = igvn.register_new_node_with_optimizer(l, _head);
2120     phase->set_created_loop_node();
2121     // Go ahead and replace _head
2122     phase->_igvn.replace_node( _head, l );
2123     _head = l;
2124     phase->set_loop(_head, this);
2125   }
2126 
2127   // Now recursively beautify nested loops
2128   if( _child ) result |= _child->beautify_loops( phase );
2129   if( _next  ) result |= _next ->beautify_loops( phase );
2130   return result;
2131 }
2132 
2133 //------------------------------allpaths_check_safepts----------------------------
2134 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
2135 // encountered.  Helper for check_safepts.
2136 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
2137   assert(stack.size() == 0, "empty stack");
2138   stack.push(_tail);
2139   visited.Clear();
2140   visited.set(_tail->_idx);
2141   while (stack.size() > 0) {
2142     Node* n = stack.pop();
2143     if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
2144       // Terminate this path
2145     } else if (n->Opcode() == Op_SafePoint) {
2146       if (_phase->get_loop(n) != this) {
2147         if (_required_safept == NULL) _required_safept = new Node_List();
2148         _required_safept->push(n);  // save the one closest to the tail
2149       }
2150       // Terminate this path
2151     } else {
2152       uint start = n->is_Region() ? 1 : 0;
2153       uint end   = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
2154       for (uint i = start; i < end; i++) {
2155         Node* in = n->in(i);
2156         assert(in->is_CFG(), "must be");
2157         if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
2158           stack.push(in);
2159         }
2160       }
2161     }
2162   }
2163 }
2164 
2165 //------------------------------check_safepts----------------------------
2166 // Given dominators, try to find loops with calls that must always be
2167 // executed (call dominates loop tail).  These loops do not need non-call
2168 // safepoints (ncsfpt).
2169 //
2170 // A complication is that a safepoint in a inner loop may be needed
2171 // by an outer loop. In the following, the inner loop sees it has a
2172 // call (block 3) on every path from the head (block 2) to the
2173 // backedge (arc 3->2).  So it deletes the ncsfpt (non-call safepoint)
2174 // in block 2, _but_ this leaves the outer loop without a safepoint.
2175 //
2176 //          entry  0
2177 //                 |
2178 //                 v
2179 // outer 1,2    +->1
2180 //              |  |
2181 //              |  v
2182 //              |  2<---+  ncsfpt in 2
2183 //              |_/|\   |
2184 //                 | v  |
2185 // inner 2,3      /  3  |  call in 3
2186 //               /   |  |
2187 //              v    +--+
2188 //        exit  4
2189 //
2190 //
2191 // This method creates a list (_required_safept) of ncsfpt nodes that must
2192 // be protected is created for each loop. When a ncsfpt maybe deleted, it
2193 // is first looked for in the lists for the outer loops of the current loop.
2194 //
2195 // The insights into the problem:
2196 //  A) counted loops are okay
2197 //  B) innermost loops are okay (only an inner loop can delete
2198 //     a ncsfpt needed by an outer loop)
2199 //  C) a loop is immune from an inner loop deleting a safepoint
2200 //     if the loop has a call on the idom-path
2201 //  D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
2202 //     idom-path that is not in a nested loop
2203 //  E) otherwise, an ncsfpt on the idom-path that is nested in an inner
2204 //     loop needs to be prevented from deletion by an inner loop
2205 //
2206 // There are two analyses:
2207 //  1) The first, and cheaper one, scans the loop body from
2208 //     tail to head following the idom (immediate dominator)
2209 //     chain, looking for the cases (C,D,E) above.
2210 //     Since inner loops are scanned before outer loops, there is summary
2211 //     information about inner loops.  Inner loops can be skipped over
2212 //     when the tail of an inner loop is encountered.
2213 //
2214 //  2) The second, invoked if the first fails to find a call or ncsfpt on
2215 //     the idom path (which is rare), scans all predecessor control paths
2216 //     from the tail to the head, terminating a path when a call or sfpt
2217 //     is encountered, to find the ncsfpt's that are closest to the tail.
2218 //
2219 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
2220   // Bottom up traversal
2221   IdealLoopTree* ch = _child;
2222   if (_child) _child->check_safepts(visited, stack);
2223   if (_next)  _next ->check_safepts(visited, stack);
2224 
2225   if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
2226     bool  has_call         = false; // call on dom-path
2227     bool  has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
2228     Node* nonlocal_ncsfpt  = NULL;  // ncsfpt on dom-path at a deeper depth
2229     // Scan the dom-path nodes from tail to head
2230     for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
2231       if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
2232         has_call = true;
2233         _has_sfpt = 1;          // Then no need for a safept!
2234         break;
2235       } else if (n->Opcode() == Op_SafePoint) {
2236         if (_phase->get_loop(n) == this) {
2237           has_local_ncsfpt = true;
2238           break;
2239         }
2240         if (nonlocal_ncsfpt == NULL) {
2241           nonlocal_ncsfpt = n; // save the one closest to the tail
2242         }
2243       } else {
2244         IdealLoopTree* nlpt = _phase->get_loop(n);
2245         if (this != nlpt) {
2246           // If at an inner loop tail, see if the inner loop has already
2247           // recorded seeing a call on the dom-path (and stop.)  If not,
2248           // jump to the head of the inner loop.
2249           assert(is_member(nlpt), "nested loop");
2250           Node* tail = nlpt->_tail;
2251           if (tail->in(0)->is_If()) tail = tail->in(0);
2252           if (n == tail) {
2253             // If inner loop has call on dom-path, so does outer loop
2254             if (nlpt->_has_sfpt) {
2255               has_call = true;
2256               _has_sfpt = 1;
2257               break;
2258             }
2259             // Skip to head of inner loop
2260             assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
2261             n = nlpt->_head;
2262           }
2263         }
2264       }
2265     }
2266     // Record safept's that this loop needs preserved when an
2267     // inner loop attempts to delete it's safepoints.
2268     if (_child != NULL && !has_call && !has_local_ncsfpt) {
2269       if (nonlocal_ncsfpt != NULL) {
2270         if (_required_safept == NULL) _required_safept = new Node_List();
2271         _required_safept->push(nonlocal_ncsfpt);
2272       } else {
2273         // Failed to find a suitable safept on the dom-path.  Now use
2274         // an all paths walk from tail to head, looking for safepoints to preserve.
2275         allpaths_check_safepts(visited, stack);
2276       }
2277     }
2278   }
2279 }
2280 
2281 //---------------------------is_deleteable_safept----------------------------
2282 // Is safept not required by an outer loop?
2283 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
2284   assert(sfpt->Opcode() == Op_SafePoint, "");
2285   IdealLoopTree* lp = get_loop(sfpt)->_parent;
2286   while (lp != NULL) {
2287     Node_List* sfpts = lp->_required_safept;
2288     if (sfpts != NULL) {
2289       for (uint i = 0; i < sfpts->size(); i++) {
2290         if (sfpt == sfpts->at(i))
2291           return false;
2292       }
2293     }
2294     lp = lp->_parent;
2295   }
2296   return true;
2297 }
2298 
2299 //---------------------------replace_parallel_iv-------------------------------
2300 // Replace parallel induction variable (parallel to trip counter)
2301 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
2302   assert(loop->_head->is_CountedLoop(), "");
2303   CountedLoopNode *cl = loop->_head->as_CountedLoop();
2304   if (!cl->is_valid_counted_loop())
2305     return;         // skip malformed counted loop
2306   Node *incr = cl->incr();
2307   if (incr == NULL)
2308     return;         // Dead loop?
2309   Node *init = cl->init_trip();
2310   Node *phi  = cl->phi();
2311   int stride_con = cl->stride_con();
2312 
2313   // Visit all children, looking for Phis
2314   for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
2315     Node *out = cl->out(i);
2316     // Look for other phis (secondary IVs). Skip dead ones
2317     if (!out->is_Phi() || out == phi || !has_node(out))
2318       continue;
2319     PhiNode* phi2 = out->as_Phi();
2320     Node *incr2 = phi2->in( LoopNode::LoopBackControl );
2321     // Look for induction variables of the form:  X += constant
2322     if (phi2->region() != loop->_head ||
2323         incr2->req() != 3 ||
2324         incr2->in(1) != phi2 ||
2325         incr2 == incr ||
2326         incr2->Opcode() != Op_AddI ||
2327         !incr2->in(2)->is_Con())
2328       continue;
2329 
2330     // Check for parallel induction variable (parallel to trip counter)
2331     // via an affine function.  In particular, count-down loops with
2332     // count-up array indices are common. We only RCE references off
2333     // the trip-counter, so we need to convert all these to trip-counter
2334     // expressions.
2335     Node *init2 = phi2->in( LoopNode::EntryControl );
2336     int stride_con2 = incr2->in(2)->get_int();
2337 
2338     // The ratio of the two strides cannot be represented as an int
2339     // if stride_con2 is min_int and stride_con is -1.
2340     if (stride_con2 == min_jint && stride_con == -1) {
2341       continue;
2342     }
2343 
2344     // The general case here gets a little tricky.  We want to find the
2345     // GCD of all possible parallel IV's and make a new IV using this
2346     // GCD for the loop.  Then all possible IVs are simple multiples of
2347     // the GCD.  In practice, this will cover very few extra loops.
2348     // Instead we require 'stride_con2' to be a multiple of 'stride_con',
2349     // where +/-1 is the common case, but other integer multiples are
2350     // also easy to handle.
2351     int ratio_con = stride_con2/stride_con;
2352 
2353     if ((ratio_con * stride_con) == stride_con2) { // Check for exact
2354 #ifndef PRODUCT
2355       if (TraceLoopOpts) {
2356         tty->print("Parallel IV: %d ", phi2->_idx);
2357         loop->dump_head();
2358       }
2359 #endif
2360       // Convert to using the trip counter.  The parallel induction
2361       // variable differs from the trip counter by a loop-invariant
2362       // amount, the difference between their respective initial values.
2363       // It is scaled by the 'ratio_con'.
2364       Node* ratio = _igvn.intcon(ratio_con);
2365       set_ctrl(ratio, C->root());
2366       Node* ratio_init = new MulINode(init, ratio);
2367       _igvn.register_new_node_with_optimizer(ratio_init, init);
2368       set_early_ctrl(ratio_init);
2369       Node* diff = new SubINode(init2, ratio_init);
2370       _igvn.register_new_node_with_optimizer(diff, init2);
2371       set_early_ctrl(diff);
2372       Node* ratio_idx = new MulINode(phi, ratio);
2373       _igvn.register_new_node_with_optimizer(ratio_idx, phi);
2374       set_ctrl(ratio_idx, cl);
2375       Node* add = new AddINode(ratio_idx, diff);
2376       _igvn.register_new_node_with_optimizer(add);
2377       set_ctrl(add, cl);
2378       _igvn.replace_node( phi2, add );
2379       // Sometimes an induction variable is unused
2380       if (add->outcnt() == 0) {
2381         _igvn.remove_dead_node(add);
2382       }
2383       --i; // deleted this phi; rescan starting with next position
2384       continue;
2385     }
2386   }
2387 }
2388 
2389 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
2390   Node* keep = NULL;
2391   if (keep_one) {
2392     // Look for a safepoint on the idom-path.
2393     for (Node* i = tail(); i != _head; i = phase->idom(i)) {
2394       if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
2395         keep = i;
2396         break; // Found one
2397       }
2398     }
2399   }
2400 
2401   // Don't remove any safepoints if it is requested to keep a single safepoint and
2402   // no safepoint was found on idom-path. It is not safe to remove any safepoint
2403   // in this case since there's no safepoint dominating all paths in the loop body.
2404   bool prune = !keep_one || keep != NULL;
2405 
2406   // Delete other safepoints in this loop.
2407   Node_List* sfpts = _safepts;
2408   if (prune && sfpts != NULL) {
2409     assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint");
2410     for (uint i = 0; i < sfpts->size(); i++) {
2411       Node* n = sfpts->at(i);
2412       assert(phase->get_loop(n) == this, "");
2413       if (n != keep && phase->is_deleteable_safept(n)) {
2414         phase->lazy_replace(n, n->in(TypeFunc::Control));
2415       }
2416     }
2417   }
2418 }
2419 
2420 //------------------------------counted_loop-----------------------------------
2421 // Convert to counted loops where possible
2422 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
2423 
2424   // For grins, set the inner-loop flag here
2425   if (!_child) {
2426     if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
2427   }
2428 
2429   IdealLoopTree* loop = this;
2430   if (_head->is_CountedLoop() ||
2431       phase->is_counted_loop(_head, loop)) {
2432 
2433     if (LoopStripMiningIter == 0 || (LoopStripMiningIter > 1 && _child == NULL)) {
2434       // Indicate we do not need a safepoint here
2435       _has_sfpt = 1;
2436     }
2437 
2438     // Remove safepoints
2439     bool keep_one_sfpt = !(_has_call || _has_sfpt);
2440     remove_safepoints(phase, keep_one_sfpt);
2441 
2442     // Look for induction variables
2443     phase->replace_parallel_iv(this);
2444 
2445   } else if (_parent != NULL && !_irreducible) {
2446     // Not a counted loop. Keep one safepoint.
2447     bool keep_one_sfpt = true;
2448     remove_safepoints(phase, keep_one_sfpt);
2449   }
2450 
2451   // Recursively
2452   assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?");
2453   assert(loop->_child != this || (loop->_child->_child == NULL && loop->_child->_next == NULL), "would miss some loops");
2454   if (loop->_child && loop->_child != this) loop->_child->counted_loop(phase);
2455   if (loop->_next)  loop->_next ->counted_loop(phase);
2456 }
2457 
2458 #ifndef PRODUCT
2459 //------------------------------dump_head--------------------------------------
2460 // Dump 1 liner for loop header info
2461 void IdealLoopTree::dump_head( ) const {
2462   for (uint i=0; i<_nest; i++)
2463     tty->print("  ");
2464   tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
2465   if (_irreducible) tty->print(" IRREDUCIBLE");
2466   Node* entry = _head->is_Loop() ? _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl) : _head->in(LoopNode::EntryControl);
2467   Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
2468   if (predicate != NULL ) {
2469     tty->print(" limit_check");
2470     entry = PhaseIdealLoop::skip_loop_predicates(entry);
2471   }
2472   if (UseProfiledLoopPredicate) {
2473     predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate);
2474     if (predicate != NULL) {
2475       tty->print(" profile_predicated");
2476       entry = PhaseIdealLoop::skip_loop_predicates(entry);
2477     }
2478   }
2479   if (UseLoopPredicate) {
2480     predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
2481     if (predicate != NULL) {
2482       tty->print(" predicated");
2483     }
2484   }
2485   if (_head->is_CountedLoop()) {
2486     CountedLoopNode *cl = _head->as_CountedLoop();
2487     tty->print(" counted");
2488 
2489     Node* init_n = cl->init_trip();
2490     if (init_n  != NULL &&  init_n->is_Con())
2491       tty->print(" [%d,", cl->init_trip()->get_int());
2492     else
2493       tty->print(" [int,");
2494     Node* limit_n = cl->limit();
2495     if (limit_n  != NULL &&  limit_n->is_Con())
2496       tty->print("%d),", cl->limit()->get_int());
2497     else
2498       tty->print("int),");
2499     int stride_con  = cl->stride_con();
2500     if (stride_con > 0) tty->print("+");
2501     tty->print("%d", stride_con);
2502 
2503     tty->print(" (%0.f iters) ", cl->profile_trip_cnt());
2504 
2505     if (cl->is_pre_loop ()) tty->print(" pre" );
2506     if (cl->is_main_loop()) tty->print(" main");
2507     if (cl->is_post_loop()) tty->print(" post");
2508     if (cl->is_vectorized_loop()) tty->print(" vector");
2509     if (cl->range_checks_present()) tty->print(" rc ");
2510     if (cl->is_multiversioned()) tty->print(" multi ");
2511   }
2512   if (_has_call) tty->print(" has_call");
2513   if (_has_sfpt) tty->print(" has_sfpt");
2514   if (_rce_candidate) tty->print(" rce");
2515   if (_safepts != NULL && _safepts->size() > 0) {
2516     tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
2517   }
2518   if (_required_safept != NULL && _required_safept->size() > 0) {
2519     tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
2520   }
2521   if (Verbose) {
2522     tty->print(" body={"); _body.dump_simple(); tty->print(" }");
2523   }
2524   if (_head->is_Loop() && _head->as_Loop()->is_strip_mined()) {
2525     tty->print(" strip_mined");
2526   }
2527   tty->cr();
2528 }
2529 
2530 //------------------------------dump-------------------------------------------
2531 // Dump loops by loop tree
2532 void IdealLoopTree::dump( ) const {
2533   dump_head();
2534   if (_child) _child->dump();
2535   if (_next)  _next ->dump();
2536 }
2537 
2538 #endif
2539 
2540 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
2541   if (loop == root) {
2542     if (loop->_child != NULL) {
2543       log->begin_head("loop_tree");
2544       log->end_head();
2545       if( loop->_child ) log_loop_tree(root, loop->_child, log);
2546       log->tail("loop_tree");
2547       assert(loop->_next == NULL, "what?");
2548     }
2549   } else {
2550     Node* head = loop->_head;
2551     log->begin_head("loop");
2552     log->print(" idx='%d' ", head->_idx);
2553     if (loop->_irreducible) log->print("irreducible='1' ");
2554     if (head->is_Loop()) {
2555       if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
2556       if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
2557     }
2558     if (head->is_CountedLoop()) {
2559       CountedLoopNode* cl = head->as_CountedLoop();
2560       if (cl->is_pre_loop())  log->print("pre_loop='%d' ",  cl->main_idx());
2561       if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
2562       if (cl->is_post_loop()) log->print("post_loop='%d' ",  cl->main_idx());
2563     }
2564     log->end_head();
2565     if( loop->_child ) log_loop_tree(root, loop->_child, log);
2566     log->tail("loop");
2567     if( loop->_next  ) log_loop_tree(root, loop->_next, log);
2568   }
2569 }
2570 
2571 //---------------------collect_potentially_useful_predicates-----------------------
2572 // Helper function to collect potentially useful predicates to prevent them from
2573 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
2574 void PhaseIdealLoop::collect_potentially_useful_predicates(
2575                          IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
2576   if (loop->_child) { // child
2577     collect_potentially_useful_predicates(loop->_child, useful_predicates);
2578   }
2579 
2580   // self (only loops that we can apply loop predication may use their predicates)
2581   if (loop->_head->is_Loop() &&
2582       !loop->_irreducible    &&
2583       !loop->tail()->is_top()) {
2584     LoopNode* lpn = loop->_head->as_Loop();
2585     Node* entry = lpn->in(LoopNode::EntryControl);
2586     Node* predicate_proj = find_predicate(entry); // loop_limit_check first
2587     if (predicate_proj != NULL) { // right pattern that can be used by loop predication
2588       assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
2589       useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2590       entry = skip_loop_predicates(entry);
2591     }
2592     if (UseProfiledLoopPredicate) {
2593       predicate_proj = find_predicate(entry); // Predicate
2594       if (predicate_proj != NULL) {
2595         useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2596         entry = skip_loop_predicates(entry);
2597       }
2598     }
2599     predicate_proj = find_predicate(entry); // Predicate
2600     if (predicate_proj != NULL) {
2601       useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2602     }
2603   }
2604 
2605   if (loop->_next) { // sibling
2606     collect_potentially_useful_predicates(loop->_next, useful_predicates);
2607   }
2608 }
2609 
2610 //------------------------eliminate_useless_predicates-----------------------------
2611 // Eliminate all inserted predicates if they could not be used by loop predication.
2612 // Note: it will also eliminates loop limits check predicate since it also uses
2613 // Opaque1 node (see Parse::add_predicate()).
2614 void PhaseIdealLoop::eliminate_useless_predicates() {
2615   if (C->predicate_count() == 0)
2616     return; // no predicate left
2617 
2618   Unique_Node_List useful_predicates; // to store useful predicates
2619   if (C->has_loops()) {
2620     collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
2621   }
2622 
2623   for (int i = C->predicate_count(); i > 0; i--) {
2624      Node * n = C->predicate_opaque1_node(i-1);
2625      assert(n->Opcode() == Op_Opaque1, "must be");
2626      if (!useful_predicates.member(n)) { // not in the useful list
2627        _igvn.replace_node(n, n->in(1));
2628      }
2629   }
2630 }
2631 
2632 //------------------------process_expensive_nodes-----------------------------
2633 // Expensive nodes have their control input set to prevent the GVN
2634 // from commoning them and as a result forcing the resulting node to
2635 // be in a more frequent path. Use CFG information here, to change the
2636 // control inputs so that some expensive nodes can be commoned while
2637 // not executed more frequently.
2638 bool PhaseIdealLoop::process_expensive_nodes() {
2639   assert(OptimizeExpensiveOps, "optimization off?");
2640 
2641   // Sort nodes to bring similar nodes together
2642   C->sort_expensive_nodes();
2643 
2644   bool progress = false;
2645 
2646   for (int i = 0; i < C->expensive_count(); ) {
2647     Node* n = C->expensive_node(i);
2648     int start = i;
2649     // Find nodes similar to n
2650     i++;
2651     for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
2652     int end = i;
2653     // And compare them two by two
2654     for (int j = start; j < end; j++) {
2655       Node* n1 = C->expensive_node(j);
2656       if (is_node_unreachable(n1)) {
2657         continue;
2658       }
2659       for (int k = j+1; k < end; k++) {
2660         Node* n2 = C->expensive_node(k);
2661         if (is_node_unreachable(n2)) {
2662           continue;
2663         }
2664 
2665         assert(n1 != n2, "should be pair of nodes");
2666 
2667         Node* c1 = n1->in(0);
2668         Node* c2 = n2->in(0);
2669 
2670         Node* parent_c1 = c1;
2671         Node* parent_c2 = c2;
2672 
2673         // The call to get_early_ctrl_for_expensive() moves the
2674         // expensive nodes up but stops at loops that are in a if
2675         // branch. See whether we can exit the loop and move above the
2676         // If.
2677         if (c1->is_Loop()) {
2678           parent_c1 = c1->in(1);
2679         }
2680         if (c2->is_Loop()) {
2681           parent_c2 = c2->in(1);
2682         }
2683 
2684         if (parent_c1 == parent_c2) {
2685           _igvn._worklist.push(n1);
2686           _igvn._worklist.push(n2);
2687           continue;
2688         }
2689 
2690         // Look for identical expensive node up the dominator chain.
2691         if (is_dominator(c1, c2)) {
2692           c2 = c1;
2693         } else if (is_dominator(c2, c1)) {
2694           c1 = c2;
2695         } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
2696                    parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
2697           // Both branches have the same expensive node so move it up
2698           // before the if.
2699           c1 = c2 = idom(parent_c1->in(0));
2700         }
2701         // Do the actual moves
2702         if (n1->in(0) != c1) {
2703           _igvn.hash_delete(n1);
2704           n1->set_req(0, c1);
2705           _igvn.hash_insert(n1);
2706           _igvn._worklist.push(n1);
2707           progress = true;
2708         }
2709         if (n2->in(0) != c2) {
2710           _igvn.hash_delete(n2);
2711           n2->set_req(0, c2);
2712           _igvn.hash_insert(n2);
2713           _igvn._worklist.push(n2);
2714           progress = true;
2715         }
2716       }
2717     }
2718   }
2719 
2720   return progress;
2721 }
2722 
2723 
2724 //=============================================================================
2725 //----------------------------build_and_optimize-------------------------------
2726 // Create a PhaseLoop.  Build the ideal Loop tree.  Map each Ideal Node to
2727 // its corresponding LoopNode.  If 'optimize' is true, do some loop cleanups.
2728 void PhaseIdealLoop::build_and_optimize(LoopOptsMode mode) {
2729   bool do_split_ifs = (mode == LoopOptsDefault || mode == LoopOptsLastRound);
2730   bool skip_loop_opts = (mode == LoopOptsNone);




2731 
2732   ResourceMark rm;
2733 
2734   int old_progress = C->major_progress();
2735   uint orig_worklist_size = _igvn._worklist.size();
2736 
2737   // Reset major-progress flag for the driver's heuristics
2738   C->clear_major_progress();
2739 
2740 #ifndef PRODUCT
2741   // Capture for later assert
2742   uint unique = C->unique();
2743   _loop_invokes++;
2744   _loop_work += unique;
2745 #endif
2746 
2747   // True if the method has at least 1 irreducible loop
2748   _has_irreducible_loops = false;
2749 
2750   _created_loop_node = false;
2751 
2752   Arena *a = Thread::current()->resource_area();
2753   VectorSet visited(a);
2754   // Pre-grow the mapping from Nodes to IdealLoopTrees.
2755   _nodes.map(C->unique(), NULL);
2756   memset(_nodes.adr(), 0, wordSize * C->unique());
2757 
2758   // Pre-build the top-level outermost loop tree entry
2759   _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
2760   // Do not need a safepoint at the top level
2761   _ltree_root->_has_sfpt = 1;
2762 
2763   // Initialize Dominators.
2764   // Checked in clone_loop_predicate() during beautify_loops().
2765   _idom_size = 0;
2766   _idom      = NULL;
2767   _dom_depth = NULL;
2768   _dom_stk   = NULL;
2769 
2770   // Empty pre-order array
2771   allocate_preorders();
2772 
2773   // Build a loop tree on the fly.  Build a mapping from CFG nodes to
2774   // IdealLoopTree entries.  Data nodes are NOT walked.
2775   build_loop_tree();
2776   // Check for bailout, and return
2777   if (C->failing()) {
2778     return;
2779   }
2780 
2781   // No loops after all
2782   if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
2783 
2784   // There should always be an outer loop containing the Root and Return nodes.
2785   // If not, we have a degenerate empty program.  Bail out in this case.
2786   if (!has_node(C->root())) {
2787     if (!_verify_only) {
2788       C->clear_major_progress();
2789       C->record_method_not_compilable("empty program detected during loop optimization");
2790     }
2791     return;
2792   }
2793 
2794   // Nothing to do, so get out
2795   bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only;
2796   bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
2797   if (stop_early && !do_expensive_nodes) {
2798     _igvn.optimize();           // Cleanup NeverBranches
2799     return;
2800   }
2801 
2802   // Set loop nesting depth
2803   _ltree_root->set_nest( 0 );
2804 
2805   // Split shared headers and insert loop landing pads.
2806   // Do not bother doing this on the Root loop of course.
2807   if( !_verify_me && !_verify_only && _ltree_root->_child ) {
2808     C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
2809     if( _ltree_root->_child->beautify_loops( this ) ) {
2810       // Re-build loop tree!
2811       _ltree_root->_child = NULL;
2812       _nodes.clear();
2813       reallocate_preorders();
2814       build_loop_tree();
2815       // Check for bailout, and return
2816       if (C->failing()) {
2817         return;
2818       }
2819       // Reset loop nesting depth
2820       _ltree_root->set_nest( 0 );
2821 
2822       C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
2823     }
2824   }
2825 
2826   // Build Dominators for elision of NULL checks & loop finding.
2827   // Since nodes do not have a slot for immediate dominator, make
2828   // a persistent side array for that info indexed on node->_idx.
2829   _idom_size = C->unique();
2830   _idom      = NEW_RESOURCE_ARRAY( Node*, _idom_size );
2831   _dom_depth = NEW_RESOURCE_ARRAY( uint,  _idom_size );
2832   _dom_stk   = NULL; // Allocated on demand in recompute_dom_depth
2833   memset( _dom_depth, 0, _idom_size * sizeof(uint) );
2834 
2835   Dominators();
2836 
2837   if (!_verify_only) {
2838     // As a side effect, Dominators removed any unreachable CFG paths
2839     // into RegionNodes.  It doesn't do this test against Root, so
2840     // we do it here.
2841     for( uint i = 1; i < C->root()->req(); i++ ) {
2842       if( !_nodes[C->root()->in(i)->_idx] ) {    // Dead path into Root?
2843         _igvn.delete_input_of(C->root(), i);
2844         i--;                      // Rerun same iteration on compressed edges
2845       }
2846     }
2847 
2848     // Given dominators, try to find inner loops with calls that must
2849     // always be executed (call dominates loop tail).  These loops do
2850     // not need a separate safepoint.
2851     Node_List cisstack(a);
2852     _ltree_root->check_safepts(visited, cisstack);
2853   }
2854 
2855   // Walk the DATA nodes and place into loops.  Find earliest control
2856   // node.  For CFG nodes, the _nodes array starts out and remains
2857   // holding the associated IdealLoopTree pointer.  For DATA nodes, the
2858   // _nodes array holds the earliest legal controlling CFG node.
2859 
2860   // Allocate stack with enough space to avoid frequent realloc
2861   int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
2862   Node_Stack nstack( a, stack_size );
2863 
2864   visited.Clear();
2865   Node_List worklist(a);
2866   // Don't need C->root() on worklist since
2867   // it will be processed among C->top() inputs
2868   worklist.push( C->top() );
2869   visited.set( C->top()->_idx ); // Set C->top() as visited now
2870   build_loop_early( visited, worklist, nstack );
2871 
2872   // Given early legal placement, try finding counted loops.  This placement
2873   // is good enough to discover most loop invariants.
2874   if( !_verify_me && !_verify_only )
2875     _ltree_root->counted_loop( this );
2876 
2877   // Find latest loop placement.  Find ideal loop placement.
2878   visited.Clear();
2879   init_dom_lca_tags();
2880   // Need C->root() on worklist when processing outs
2881   worklist.push( C->root() );
2882   NOT_PRODUCT( C->verify_graph_edges(); )
2883   worklist.push( C->top() );
2884   build_loop_late( visited, worklist, nstack );
2885 
2886   if (_verify_only) {
2887     // restore major progress flag
2888     for (int i = 0; i < old_progress; i++)
2889       C->set_major_progress();
2890     assert(C->unique() == unique, "verification mode made Nodes? ? ?");
2891     assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
2892     return;
2893   }
2894 
2895   // clear out the dead code after build_loop_late
2896   while (_deadlist.size()) {
2897     _igvn.remove_globally_dead_node(_deadlist.pop());
2898   }
2899 
2900   if (stop_early) {
2901     assert(do_expensive_nodes, "why are we here?");
2902     if (process_expensive_nodes()) {
2903       // If we made some progress when processing expensive nodes then
2904       // the IGVN may modify the graph in a way that will allow us to
2905       // make some more progress: we need to try processing expensive
2906       // nodes again.
2907       C->set_major_progress();
2908     }
2909     _igvn.optimize();
2910     return;
2911   }
2912 
2913   // Some parser-inserted loop predicates could never be used by loop
2914   // predication or they were moved away from loop during some optimizations.
2915   // For example, peeling. Eliminate them before next loop optimizations.
2916   eliminate_useless_predicates();
2917 
2918 #ifndef PRODUCT
2919   C->verify_graph_edges();
2920   if (_verify_me) {             // Nested verify pass?
2921     // Check to see if the verify mode is broken
2922     assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
2923     return;
2924   }
2925   if(VerifyLoopOptimizations) verify();
2926   if(TraceLoopOpts && C->has_loops()) {
2927     _ltree_root->dump();
2928   }
2929 #endif
2930 
2931   if (skip_loop_opts) {
2932     // restore major progress flag
2933     for (int i = 0; i < old_progress; i++) {
2934       C->set_major_progress();
2935     }
2936 
2937     // Cleanup any modified bits
2938     _igvn.optimize();
2939 
2940     if (C->log() != NULL) {
2941       log_loop_tree(_ltree_root, _ltree_root, C->log());
2942     }
2943     return;
2944   }
2945 










2946   if (ReassociateInvariants) {
2947     // Reassociate invariants and prep for split_thru_phi
2948     for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2949       IdealLoopTree* lpt = iter.current();
2950       bool is_counted = lpt->is_counted();
2951       if (!is_counted || !lpt->is_inner()) continue;
2952 
2953       // check for vectorized loops, any reassociation of invariants was already done
2954       if (is_counted && lpt->_head->as_CountedLoop()->do_unroll_only()) continue;
2955 
2956       lpt->reassociate_invariants(this);
2957 
2958       // Because RCE opportunities can be masked by split_thru_phi,
2959       // look for RCE candidates and inhibit split_thru_phi
2960       // on just their loop-phi's for this pass of loop opts
2961       if (SplitIfBlocks && do_split_ifs) {
2962         if (lpt->policy_range_check(this)) {
2963           lpt->_rce_candidate = 1; // = true
2964         }
2965       }
2966     }
2967   }
2968 
2969   // Check for aggressive application of split-if and other transforms
2970   // that require basic-block info (like cloning through Phi's)
2971   if( SplitIfBlocks && do_split_ifs ) {
2972     visited.Clear();
2973     split_if_with_blocks( visited, nstack, mode == LoopOptsLastRound );
2974     NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
2975     if (mode == LoopOptsLastRound) {
2976       C->set_major_progress();
2977     }
2978   }
2979 
2980   if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
2981     C->set_major_progress();
2982   }
2983 
2984   // Perform loop predication before iteration splitting
2985   if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
2986     _ltree_root->_child->loop_predication(this);
2987   }
2988 
2989   if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
2990     if (do_intrinsify_fill()) {
2991       C->set_major_progress();
2992     }
2993   }
2994 
2995   // Perform iteration-splitting on inner loops.  Split iterations to avoid
2996   // range checks or one-shot null checks.
2997 
2998   // If split-if's didn't hack the graph too bad (no CFG changes)
2999   // then do loop opts.
3000   if (C->has_loops() && !C->major_progress()) {
3001     memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
3002     _ltree_root->_child->iteration_split( this, worklist );
3003     // No verify after peeling!  GCM has hoisted code out of the loop.
3004     // After peeling, the hoisted code could sink inside the peeled area.
3005     // The peeling code does not try to recompute the best location for
3006     // all the code before the peeled area, so the verify pass will always
3007     // complain about it.
3008   }
3009   // Do verify graph edges in any case
3010   NOT_PRODUCT( C->verify_graph_edges(); );
3011 
3012   if (!do_split_ifs) {
3013     // We saw major progress in Split-If to get here.  We forced a
3014     // pass with unrolling and not split-if, however more split-if's
3015     // might make progress.  If the unrolling didn't make progress
3016     // then the major-progress flag got cleared and we won't try
3017     // another round of Split-If.  In particular the ever-common
3018     // instance-of/check-cast pattern requires at least 2 rounds of
3019     // Split-If to clear out.
3020     C->set_major_progress();
3021   }
3022 
3023   // Repeat loop optimizations if new loops were seen
3024   if (created_loop_node()) {
3025     C->set_major_progress();
3026   }
3027 
3028   // Keep loop predicates and perform optimizations with them
3029   // until no more loop optimizations could be done.
3030   // After that switch predicates off and do more loop optimizations.
3031   if (!C->major_progress() && (C->predicate_count() > 0)) {
3032      C->cleanup_loop_predicates(_igvn);
3033      if (TraceLoopOpts) {
3034        tty->print_cr("PredicatesOff");
3035      }
3036      C->set_major_progress();
3037   }
3038 
3039   // Convert scalar to superword operations at the end of all loop opts.
3040   if (UseSuperWord && C->has_loops() && !C->major_progress()) {
3041     // SuperWord transform
3042     SuperWord sw(this);
3043     for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
3044       IdealLoopTree* lpt = iter.current();
3045       if (lpt->is_counted()) {
3046         CountedLoopNode *cl = lpt->_head->as_CountedLoop();
3047 
3048         if (PostLoopMultiversioning && cl->is_rce_post_loop() && !cl->is_vectorized_loop()) {
3049           // Check that the rce'd post loop is encountered first, multiversion after all
3050           // major main loop optimization are concluded
3051           if (!C->major_progress()) {
3052             IdealLoopTree *lpt_next = lpt->_next;
3053             if (lpt_next && lpt_next->is_counted()) {
3054               CountedLoopNode *cl = lpt_next->_head->as_CountedLoop();
3055               has_range_checks(lpt_next);
3056               if (cl->is_post_loop() && cl->range_checks_present()) {
3057                 if (!cl->is_multiversioned()) {
3058                   if (multi_version_post_loops(lpt, lpt_next) == false) {
3059                     // Cause the rce loop to be optimized away if we fail
3060                     cl->mark_is_multiversioned();
3061                     cl->set_slp_max_unroll(0);
3062                     poison_rce_post_loop(lpt);
3063                   }
3064                 }
3065               }
3066             }
3067             sw.transform_loop(lpt, true);
3068           }
3069         } else if (cl->is_main_loop()) {
3070           sw.transform_loop(lpt, true);
3071         }
3072       }
3073     }
3074   }
3075 
3076   // Cleanup any modified bits
3077   _igvn.optimize();
3078 
3079   // disable assert until issue with split_flow_path is resolved (6742111)
3080   // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
3081   //        "shouldn't introduce irreducible loops");
3082 
3083   if (C->log() != NULL) {
3084     log_loop_tree(_ltree_root, _ltree_root, C->log());
3085   }
3086 }
3087 
3088 #ifndef PRODUCT
3089 //------------------------------print_statistics-------------------------------
3090 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
3091 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
3092 void PhaseIdealLoop::print_statistics() {
3093   tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
3094 }
3095 
3096 //------------------------------verify-----------------------------------------
3097 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
3098 static int fail;                // debug only, so its multi-thread dont care
3099 void PhaseIdealLoop::verify() const {
3100   int old_progress = C->major_progress();
3101   ResourceMark rm;
3102   PhaseIdealLoop loop_verify( _igvn, this );
3103   VectorSet visited(Thread::current()->resource_area());
3104 
3105   fail = 0;
3106   verify_compare( C->root(), &loop_verify, visited );
3107   assert( fail == 0, "verify loops failed" );
3108   // Verify loop structure is the same
3109   _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
3110   // Reset major-progress.  It was cleared by creating a verify version of
3111   // PhaseIdealLoop.
3112   for( int i=0; i<old_progress; i++ )
3113     C->set_major_progress();
3114 }
3115 
3116 //------------------------------verify_compare---------------------------------
3117 // Make sure me and the given PhaseIdealLoop agree on key data structures
3118 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
3119   if( !n ) return;
3120   if( visited.test_set( n->_idx ) ) return;
3121   if( !_nodes[n->_idx] ) {      // Unreachable
3122     assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
3123     return;
3124   }
3125 
3126   uint i;
3127   for( i = 0; i < n->req(); i++ )
3128     verify_compare( n->in(i), loop_verify, visited );
3129 
3130   // Check the '_nodes' block/loop structure
3131   i = n->_idx;
3132   if( has_ctrl(n) ) {           // We have control; verify has loop or ctrl
3133     if( _nodes[i] != loop_verify->_nodes[i] &&
3134         get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
3135       tty->print("Mismatched control setting for: ");
3136       n->dump();
3137       if( fail++ > 10 ) return;
3138       Node *c = get_ctrl_no_update(n);
3139       tty->print("We have it as: ");
3140       if( c->in(0) ) c->dump();
3141         else tty->print_cr("N%d",c->_idx);
3142       tty->print("Verify thinks: ");
3143       if( loop_verify->has_ctrl(n) )
3144         loop_verify->get_ctrl_no_update(n)->dump();
3145       else
3146         loop_verify->get_loop_idx(n)->dump();
3147       tty->cr();
3148     }
3149   } else {                    // We have a loop
3150     IdealLoopTree *us = get_loop_idx(n);
3151     if( loop_verify->has_ctrl(n) ) {
3152       tty->print("Mismatched loop setting for: ");
3153       n->dump();
3154       if( fail++ > 10 ) return;
3155       tty->print("We have it as: ");
3156       us->dump();
3157       tty->print("Verify thinks: ");
3158       loop_verify->get_ctrl_no_update(n)->dump();
3159       tty->cr();
3160     } else if (!C->major_progress()) {
3161       // Loop selection can be messed up if we did a major progress
3162       // operation, like split-if.  Do not verify in that case.
3163       IdealLoopTree *them = loop_verify->get_loop_idx(n);
3164       if( us->_head != them->_head ||  us->_tail != them->_tail ) {
3165         tty->print("Unequals loops for: ");
3166         n->dump();
3167         if( fail++ > 10 ) return;
3168         tty->print("We have it as: ");
3169         us->dump();
3170         tty->print("Verify thinks: ");
3171         them->dump();
3172         tty->cr();
3173       }
3174     }
3175   }
3176 
3177   // Check for immediate dominators being equal
3178   if( i >= _idom_size ) {
3179     if( !n->is_CFG() ) return;
3180     tty->print("CFG Node with no idom: ");
3181     n->dump();
3182     return;
3183   }
3184   if( !n->is_CFG() ) return;
3185   if( n == C->root() ) return; // No IDOM here
3186 
3187   assert(n->_idx == i, "sanity");
3188   Node *id = idom_no_update(n);
3189   if( id != loop_verify->idom_no_update(n) ) {
3190     tty->print("Unequals idoms for: ");
3191     n->dump();
3192     if( fail++ > 10 ) return;
3193     tty->print("We have it as: ");
3194     id->dump();
3195     tty->print("Verify thinks: ");
3196     loop_verify->idom_no_update(n)->dump();
3197     tty->cr();
3198   }
3199 
3200 }
3201 
3202 //------------------------------verify_tree------------------------------------
3203 // Verify that tree structures match.  Because the CFG can change, siblings
3204 // within the loop tree can be reordered.  We attempt to deal with that by
3205 // reordering the verify's loop tree if possible.
3206 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
3207   assert( _parent == parent, "Badly formed loop tree" );
3208 
3209   // Siblings not in same order?  Attempt to re-order.
3210   if( _head != loop->_head ) {
3211     // Find _next pointer to update
3212     IdealLoopTree **pp = &loop->_parent->_child;
3213     while( *pp != loop )
3214       pp = &((*pp)->_next);
3215     // Find proper sibling to be next
3216     IdealLoopTree **nn = &loop->_next;
3217     while( (*nn) && (*nn)->_head != _head )
3218       nn = &((*nn)->_next);
3219 
3220     // Check for no match.
3221     if( !(*nn) ) {
3222       // Annoyingly, irreducible loops can pick different headers
3223       // after a major_progress operation, so the rest of the loop
3224       // tree cannot be matched.
3225       if (_irreducible && Compile::current()->major_progress())  return;
3226       assert( 0, "failed to match loop tree" );
3227     }
3228 
3229     // Move (*nn) to (*pp)
3230     IdealLoopTree *hit = *nn;
3231     *nn = hit->_next;
3232     hit->_next = loop;
3233     *pp = loop;
3234     loop = hit;
3235     // Now try again to verify
3236   }
3237 
3238   assert( _head  == loop->_head , "mismatched loop head" );
3239   Node *tail = _tail;           // Inline a non-updating version of
3240   while( !tail->in(0) )         // the 'tail()' call.
3241     tail = tail->in(1);
3242   assert( tail == loop->_tail, "mismatched loop tail" );
3243 
3244   // Counted loops that are guarded should be able to find their guards
3245   if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
3246     CountedLoopNode *cl = _head->as_CountedLoop();
3247     Node *init = cl->init_trip();
3248     Node *ctrl = cl->in(LoopNode::EntryControl);
3249     assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
3250     Node *iff  = ctrl->in(0);
3251     assert( iff->Opcode() == Op_If, "" );
3252     Node *bol  = iff->in(1);
3253     assert( bol->Opcode() == Op_Bool, "" );
3254     Node *cmp  = bol->in(1);
3255     assert( cmp->Opcode() == Op_CmpI, "" );
3256     Node *add  = cmp->in(1);
3257     Node *opaq;
3258     if( add->Opcode() == Op_Opaque1 ) {
3259       opaq = add;
3260     } else {
3261       assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
3262       assert( add == init, "" );
3263       opaq = cmp->in(2);
3264     }
3265     assert( opaq->Opcode() == Op_Opaque1, "" );
3266 
3267   }
3268 
3269   if (_child != NULL)  _child->verify_tree(loop->_child, this);
3270   if (_next  != NULL)  _next ->verify_tree(loop->_next,  parent);
3271   // Innermost loops need to verify loop bodies,
3272   // but only if no 'major_progress'
3273   int fail = 0;
3274   if (!Compile::current()->major_progress() && _child == NULL) {
3275     for( uint i = 0; i < _body.size(); i++ ) {
3276       Node *n = _body.at(i);
3277       if (n->outcnt() == 0)  continue; // Ignore dead
3278       uint j;
3279       for( j = 0; j < loop->_body.size(); j++ )
3280         if( loop->_body.at(j) == n )
3281           break;
3282       if( j == loop->_body.size() ) { // Not found in loop body
3283         // Last ditch effort to avoid assertion: Its possible that we
3284         // have some users (so outcnt not zero) but are still dead.
3285         // Try to find from root.
3286         if (Compile::current()->root()->find(n->_idx)) {
3287           fail++;
3288           tty->print("We have that verify does not: ");
3289           n->dump();
3290         }
3291       }
3292     }
3293     for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
3294       Node *n = loop->_body.at(i2);
3295       if (n->outcnt() == 0)  continue; // Ignore dead
3296       uint j;
3297       for( j = 0; j < _body.size(); j++ )
3298         if( _body.at(j) == n )
3299           break;
3300       if( j == _body.size() ) { // Not found in loop body
3301         // Last ditch effort to avoid assertion: Its possible that we
3302         // have some users (so outcnt not zero) but are still dead.
3303         // Try to find from root.
3304         if (Compile::current()->root()->find(n->_idx)) {
3305           fail++;
3306           tty->print("Verify has that we do not: ");
3307           n->dump();
3308         }
3309       }
3310     }
3311     assert( !fail, "loop body mismatch" );
3312   }
3313 }
3314 
3315 #endif
3316 
3317 //------------------------------set_idom---------------------------------------
3318 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
3319   uint idx = d->_idx;
3320   if (idx >= _idom_size) {
3321     uint newsize = _idom_size<<1;
3322     while( idx >= newsize ) {
3323       newsize <<= 1;
3324     }
3325     _idom      = REALLOC_RESOURCE_ARRAY( Node*,     _idom,_idom_size,newsize);
3326     _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
3327     memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
3328     _idom_size = newsize;
3329   }
3330   _idom[idx] = n;
3331   _dom_depth[idx] = dom_depth;
3332 }
3333 
3334 //------------------------------recompute_dom_depth---------------------------------------
3335 // The dominator tree is constructed with only parent pointers.
3336 // This recomputes the depth in the tree by first tagging all
3337 // nodes as "no depth yet" marker.  The next pass then runs up
3338 // the dom tree from each node marked "no depth yet", and computes
3339 // the depth on the way back down.
3340 void PhaseIdealLoop::recompute_dom_depth() {
3341   uint no_depth_marker = C->unique();
3342   uint i;
3343   // Initialize depth to "no depth yet" and realize all lazy updates
3344   for (i = 0; i < _idom_size; i++) {
3345     // Only indices with a _dom_depth has a Node* or NULL (otherwise uninitalized).
3346     if (_dom_depth[i] > 0 && _idom[i] != NULL) {
3347       _dom_depth[i] = no_depth_marker;
3348 
3349       // heal _idom if it has a fwd mapping in _nodes
3350       if (_idom[i]->in(0) == NULL) {
3351         idom(i);
3352       }
3353     }
3354   }
3355   if (_dom_stk == NULL) {
3356     uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
3357     if (init_size < 10) init_size = 10;
3358     _dom_stk = new GrowableArray<uint>(init_size);
3359   }
3360   // Compute new depth for each node.
3361   for (i = 0; i < _idom_size; i++) {
3362     uint j = i;
3363     // Run up the dom tree to find a node with a depth
3364     while (_dom_depth[j] == no_depth_marker) {
3365       _dom_stk->push(j);
3366       j = _idom[j]->_idx;
3367     }
3368     // Compute the depth on the way back down this tree branch
3369     uint dd = _dom_depth[j] + 1;
3370     while (_dom_stk->length() > 0) {
3371       uint j = _dom_stk->pop();
3372       _dom_depth[j] = dd;
3373       dd++;
3374     }
3375   }
3376 }
3377 
3378 //------------------------------sort-------------------------------------------
3379 // Insert 'loop' into the existing loop tree.  'innermost' is a leaf of the
3380 // loop tree, not the root.
3381 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
3382   if( !innermost ) return loop; // New innermost loop
3383 
3384   int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
3385   assert( loop_preorder, "not yet post-walked loop" );
3386   IdealLoopTree **pp = &innermost;      // Pointer to previous next-pointer
3387   IdealLoopTree *l = *pp;               // Do I go before or after 'l'?
3388 
3389   // Insert at start of list
3390   while( l ) {                  // Insertion sort based on pre-order
3391     if( l == loop ) return innermost; // Already on list!
3392     int l_preorder = get_preorder(l->_head); // Cache pre-order number
3393     assert( l_preorder, "not yet post-walked l" );
3394     // Check header pre-order number to figure proper nesting
3395     if( loop_preorder > l_preorder )
3396       break;                    // End of insertion
3397     // If headers tie (e.g., shared headers) check tail pre-order numbers.
3398     // Since I split shared headers, you'd think this could not happen.
3399     // BUT: I must first do the preorder numbering before I can discover I
3400     // have shared headers, so the split headers all get the same preorder
3401     // number as the RegionNode they split from.
3402     if( loop_preorder == l_preorder &&
3403         get_preorder(loop->_tail) < get_preorder(l->_tail) )
3404       break;                    // Also check for shared headers (same pre#)
3405     pp = &l->_parent;           // Chain up list
3406     l = *pp;
3407   }
3408   // Link into list
3409   // Point predecessor to me
3410   *pp = loop;
3411   // Point me to successor
3412   IdealLoopTree *p = loop->_parent;
3413   loop->_parent = l;            // Point me to successor
3414   if( p ) sort( p, innermost ); // Insert my parents into list as well
3415   return innermost;
3416 }
3417 
3418 //------------------------------build_loop_tree--------------------------------
3419 // I use a modified Vick/Tarjan algorithm.  I need pre- and a post- visit
3420 // bits.  The _nodes[] array is mapped by Node index and holds a NULL for
3421 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
3422 // tightest enclosing IdealLoopTree for post-walked.
3423 //
3424 // During my forward walk I do a short 1-layer lookahead to see if I can find
3425 // a loop backedge with that doesn't have any work on the backedge.  This
3426 // helps me construct nested loops with shared headers better.
3427 //
3428 // Once I've done the forward recursion, I do the post-work.  For each child
3429 // I check to see if there is a backedge.  Backedges define a loop!  I
3430 // insert an IdealLoopTree at the target of the backedge.
3431 //
3432 // During the post-work I also check to see if I have several children
3433 // belonging to different loops.  If so, then this Node is a decision point
3434 // where control flow can choose to change loop nests.  It is at this
3435 // decision point where I can figure out how loops are nested.  At this
3436 // time I can properly order the different loop nests from my children.
3437 // Note that there may not be any backedges at the decision point!
3438 //
3439 // Since the decision point can be far removed from the backedges, I can't
3440 // order my loops at the time I discover them.  Thus at the decision point
3441 // I need to inspect loop header pre-order numbers to properly nest my
3442 // loops.  This means I need to sort my childrens' loops by pre-order.
3443 // The sort is of size number-of-control-children, which generally limits
3444 // it to size 2 (i.e., I just choose between my 2 target loops).
3445 void PhaseIdealLoop::build_loop_tree() {
3446   // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3447   GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
3448   Node *n = C->root();
3449   bltstack.push(n);
3450   int pre_order = 1;
3451   int stack_size;
3452 
3453   while ( ( stack_size = bltstack.length() ) != 0 ) {
3454     n = bltstack.top(); // Leave node on stack
3455     if ( !is_visited(n) ) {
3456       // ---- Pre-pass Work ----
3457       // Pre-walked but not post-walked nodes need a pre_order number.
3458 
3459       set_preorder_visited( n, pre_order ); // set as visited
3460 
3461       // ---- Scan over children ----
3462       // Scan first over control projections that lead to loop headers.
3463       // This helps us find inner-to-outer loops with shared headers better.
3464 
3465       // Scan children's children for loop headers.
3466       for ( int i = n->outcnt() - 1; i >= 0; --i ) {
3467         Node* m = n->raw_out(i);       // Child
3468         if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
3469           // Scan over children's children to find loop
3470           for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3471             Node* l = m->fast_out(j);
3472             if( is_visited(l) &&       // Been visited?
3473                 !is_postvisited(l) &&  // But not post-visited
3474                 get_preorder(l) < pre_order ) { // And smaller pre-order
3475               // Found!  Scan the DFS down this path before doing other paths
3476               bltstack.push(m);
3477               break;
3478             }
3479           }
3480         }
3481       }
3482       pre_order++;
3483     }
3484     else if ( !is_postvisited(n) ) {
3485       // Note: build_loop_tree_impl() adds out edges on rare occasions,
3486       // such as com.sun.rsasign.am::a.
3487       // For non-recursive version, first, process current children.
3488       // On next iteration, check if additional children were added.
3489       for ( int k = n->outcnt() - 1; k >= 0; --k ) {
3490         Node* u = n->raw_out(k);
3491         if ( u->is_CFG() && !is_visited(u) ) {
3492           bltstack.push(u);
3493         }
3494       }
3495       if ( bltstack.length() == stack_size ) {
3496         // There were no additional children, post visit node now
3497         (void)bltstack.pop(); // Remove node from stack
3498         pre_order = build_loop_tree_impl( n, pre_order );
3499         // Check for bailout
3500         if (C->failing()) {
3501           return;
3502         }
3503         // Check to grow _preorders[] array for the case when
3504         // build_loop_tree_impl() adds new nodes.
3505         check_grow_preorders();
3506       }
3507     }
3508     else {
3509       (void)bltstack.pop(); // Remove post-visited node from stack
3510     }
3511   }
3512 }
3513 
3514 //------------------------------build_loop_tree_impl---------------------------
3515 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
3516   // ---- Post-pass Work ----
3517   // Pre-walked but not post-walked nodes need a pre_order number.
3518 
3519   // Tightest enclosing loop for this Node
3520   IdealLoopTree *innermost = NULL;
3521 
3522   // For all children, see if any edge is a backedge.  If so, make a loop
3523   // for it.  Then find the tightest enclosing loop for the self Node.
3524   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3525     Node* m = n->fast_out(i);   // Child
3526     if( n == m ) continue;      // Ignore control self-cycles
3527     if( !m->is_CFG() ) continue;// Ignore non-CFG edges
3528 
3529     IdealLoopTree *l;           // Child's loop
3530     if( !is_postvisited(m) ) {  // Child visited but not post-visited?
3531       // Found a backedge
3532       assert( get_preorder(m) < pre_order, "should be backedge" );
3533       // Check for the RootNode, which is already a LoopNode and is allowed
3534       // to have multiple "backedges".
3535       if( m == C->root()) {     // Found the root?
3536         l = _ltree_root;        // Root is the outermost LoopNode
3537       } else {                  // Else found a nested loop
3538         // Insert a LoopNode to mark this loop.
3539         l = new IdealLoopTree(this, m, n);
3540       } // End of Else found a nested loop
3541       if( !has_loop(m) )        // If 'm' does not already have a loop set
3542         set_loop(m, l);         // Set loop header to loop now
3543 
3544     } else {                    // Else not a nested loop
3545       if( !_nodes[m->_idx] ) continue; // Dead code has no loop
3546       l = get_loop(m);          // Get previously determined loop
3547       // If successor is header of a loop (nest), move up-loop till it
3548       // is a member of some outer enclosing loop.  Since there are no
3549       // shared headers (I've split them already) I only need to go up
3550       // at most 1 level.
3551       while( l && l->_head == m ) // Successor heads loop?
3552         l = l->_parent;         // Move up 1 for me
3553       // If this loop is not properly parented, then this loop
3554       // has no exit path out, i.e. its an infinite loop.
3555       if( !l ) {
3556         // Make loop "reachable" from root so the CFG is reachable.  Basically
3557         // insert a bogus loop exit that is never taken.  'm', the loop head,
3558         // points to 'n', one (of possibly many) fall-in paths.  There may be
3559         // many backedges as well.
3560 
3561         // Here I set the loop to be the root loop.  I could have, after
3562         // inserting a bogus loop exit, restarted the recursion and found my
3563         // new loop exit.  This would make the infinite loop a first-class
3564         // loop and it would then get properly optimized.  What's the use of
3565         // optimizing an infinite loop?
3566         l = _ltree_root;        // Oops, found infinite loop
3567 
3568         if (!_verify_only) {
3569           // Insert the NeverBranch between 'm' and it's control user.
3570           NeverBranchNode *iff = new NeverBranchNode( m );
3571           _igvn.register_new_node_with_optimizer(iff);
3572           set_loop(iff, l);
3573           Node *if_t = new CProjNode( iff, 0 );
3574           _igvn.register_new_node_with_optimizer(if_t);
3575           set_loop(if_t, l);
3576 
3577           Node* cfg = NULL;       // Find the One True Control User of m
3578           for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3579             Node* x = m->fast_out(j);
3580             if (x->is_CFG() && x != m && x != iff)
3581               { cfg = x; break; }
3582           }
3583           assert(cfg != NULL, "must find the control user of m");
3584           uint k = 0;             // Probably cfg->in(0)
3585           while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
3586           cfg->set_req( k, if_t ); // Now point to NeverBranch
3587           _igvn._worklist.push(cfg);
3588 
3589           // Now create the never-taken loop exit
3590           Node *if_f = new CProjNode( iff, 1 );
3591           _igvn.register_new_node_with_optimizer(if_f);
3592           set_loop(if_f, l);
3593           // Find frame ptr for Halt.  Relies on the optimizer
3594           // V-N'ing.  Easier and quicker than searching through
3595           // the program structure.
3596           Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr );
3597           _igvn.register_new_node_with_optimizer(frame);
3598           // Halt & Catch Fire
3599           Node* halt = new HaltNode(if_f, frame, "never-taken loop exit reached");
3600           _igvn.register_new_node_with_optimizer(halt);
3601           set_loop(halt, l);
3602           C->root()->add_req(halt);
3603         }
3604         set_loop(C->root(), _ltree_root);
3605       }
3606     }
3607     // Weeny check for irreducible.  This child was already visited (this
3608     // IS the post-work phase).  Is this child's loop header post-visited
3609     // as well?  If so, then I found another entry into the loop.
3610     if (!_verify_only) {
3611       while( is_postvisited(l->_head) ) {
3612         // found irreducible
3613         l->_irreducible = 1; // = true
3614         l = l->_parent;
3615         _has_irreducible_loops = true;
3616         // Check for bad CFG here to prevent crash, and bailout of compile
3617         if (l == NULL) {
3618           C->record_method_not_compilable("unhandled CFG detected during loop optimization");
3619           return pre_order;
3620         }
3621       }
3622       C->set_has_irreducible_loop(_has_irreducible_loops);
3623     }
3624 
3625     // This Node might be a decision point for loops.  It is only if
3626     // it's children belong to several different loops.  The sort call
3627     // does a trivial amount of work if there is only 1 child or all
3628     // children belong to the same loop.  If however, the children
3629     // belong to different loops, the sort call will properly set the
3630     // _parent pointers to show how the loops nest.
3631     //
3632     // In any case, it returns the tightest enclosing loop.
3633     innermost = sort( l, innermost );
3634   }
3635 
3636   // Def-use info will have some dead stuff; dead stuff will have no
3637   // loop decided on.
3638 
3639   // Am I a loop header?  If so fix up my parent's child and next ptrs.
3640   if( innermost && innermost->_head == n ) {
3641     assert( get_loop(n) == innermost, "" );
3642     IdealLoopTree *p = innermost->_parent;
3643     IdealLoopTree *l = innermost;
3644     while( p && l->_head == n ) {
3645       l->_next = p->_child;     // Put self on parents 'next child'
3646       p->_child = l;            // Make self as first child of parent
3647       l = p;                    // Now walk up the parent chain
3648       p = l->_parent;
3649     }
3650   } else {
3651     // Note that it is possible for a LoopNode to reach here, if the
3652     // backedge has been made unreachable (hence the LoopNode no longer
3653     // denotes a Loop, and will eventually be removed).
3654 
3655     // Record tightest enclosing loop for self.  Mark as post-visited.
3656     set_loop(n, innermost);
3657     // Also record has_call flag early on
3658     if( innermost ) {
3659       if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
3660         // Do not count uncommon calls
3661         if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
3662           Node *iff = n->in(0)->in(0);
3663           // No any calls for vectorized loops.
3664           if( UseSuperWord || !iff->is_If() ||
3665               (n->in(0)->Opcode() == Op_IfFalse &&
3666                (1.0 - iff->as_If()->_prob) >= 0.01) ||
3667               (iff->as_If()->_prob >= 0.01) )
3668             innermost->_has_call = 1;
3669         }
3670       } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
3671         // Disable loop optimizations if the loop has a scalar replaceable
3672         // allocation. This disabling may cause a potential performance lost
3673         // if the allocation is not eliminated for some reason.
3674         innermost->_allow_optimizations = false;
3675         innermost->_has_call = 1; // = true
3676       } else if (n->Opcode() == Op_SafePoint) {
3677         // Record all safepoints in this loop.
3678         if (innermost->_safepts == NULL) innermost->_safepts = new Node_List();
3679         innermost->_safepts->push(n);
3680       }
3681     }
3682   }
3683 
3684   // Flag as post-visited now
3685   set_postvisited(n);
3686   return pre_order;
3687 }
3688 
3689 
3690 //------------------------------build_loop_early-------------------------------
3691 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3692 // First pass computes the earliest controlling node possible.  This is the
3693 // controlling input with the deepest dominating depth.
3694 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3695   while (worklist.size() != 0) {
3696     // Use local variables nstack_top_n & nstack_top_i to cache values
3697     // on nstack's top.
3698     Node *nstack_top_n = worklist.pop();
3699     uint  nstack_top_i = 0;
3700 //while_nstack_nonempty:
3701     while (true) {
3702       // Get parent node and next input's index from stack's top.
3703       Node  *n = nstack_top_n;
3704       uint   i = nstack_top_i;
3705       uint cnt = n->req(); // Count of inputs
3706       if (i == 0) {        // Pre-process the node.
3707         if( has_node(n) &&            // Have either loop or control already?
3708             !has_ctrl(n) ) {          // Have loop picked out already?
3709           // During "merge_many_backedges" we fold up several nested loops
3710           // into a single loop.  This makes the members of the original
3711           // loop bodies pointing to dead loops; they need to move up
3712           // to the new UNION'd larger loop.  I set the _head field of these
3713           // dead loops to NULL and the _parent field points to the owning
3714           // loop.  Shades of UNION-FIND algorithm.
3715           IdealLoopTree *ilt;
3716           while( !(ilt = get_loop(n))->_head ) {
3717             // Normally I would use a set_loop here.  But in this one special
3718             // case, it is legal (and expected) to change what loop a Node
3719             // belongs to.
3720             _nodes.map(n->_idx, (Node*)(ilt->_parent) );
3721           }
3722           // Remove safepoints ONLY if I've already seen I don't need one.
3723           // (the old code here would yank a 2nd safepoint after seeing a
3724           // first one, even though the 1st did not dominate in the loop body
3725           // and thus could be avoided indefinitely)
3726           if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
3727               is_deleteable_safept(n)) {
3728             Node *in = n->in(TypeFunc::Control);
3729             lazy_replace(n,in);       // Pull safepoint now
3730             if (ilt->_safepts != NULL) {
3731               ilt->_safepts->yank(n);
3732             }
3733             // Carry on with the recursion "as if" we are walking
3734             // only the control input
3735             if( !visited.test_set( in->_idx ) ) {
3736               worklist.push(in);      // Visit this guy later, using worklist
3737             }
3738             // Get next node from nstack:
3739             // - skip n's inputs processing by setting i > cnt;
3740             // - we also will not call set_early_ctrl(n) since
3741             //   has_node(n) == true (see the condition above).
3742             i = cnt + 1;
3743           }
3744         }
3745       } // if (i == 0)
3746 
3747       // Visit all inputs
3748       bool done = true;       // Assume all n's inputs will be processed
3749       while (i < cnt) {
3750         Node *in = n->in(i);
3751         ++i;
3752         if (in == NULL) continue;
3753         if (in->pinned() && !in->is_CFG())
3754           set_ctrl(in, in->in(0));
3755         int is_visited = visited.test_set( in->_idx );
3756         if (!has_node(in)) {  // No controlling input yet?
3757           assert( !in->is_CFG(), "CFG Node with no controlling input?" );
3758           assert( !is_visited, "visit only once" );
3759           nstack.push(n, i);  // Save parent node and next input's index.
3760           nstack_top_n = in;  // Process current input now.
3761           nstack_top_i = 0;
3762           done = false;       // Not all n's inputs processed.
3763           break; // continue while_nstack_nonempty;
3764         } else if (!is_visited) {
3765           // This guy has a location picked out for him, but has not yet
3766           // been visited.  Happens to all CFG nodes, for instance.
3767           // Visit him using the worklist instead of recursion, to break
3768           // cycles.  Since he has a location already we do not need to
3769           // find his location before proceeding with the current Node.
3770           worklist.push(in);  // Visit this guy later, using worklist
3771         }
3772       }
3773       if (done) {
3774         // All of n's inputs have been processed, complete post-processing.
3775 
3776         // Compute earliest point this Node can go.
3777         // CFG, Phi, pinned nodes already know their controlling input.
3778         if (!has_node(n)) {
3779           // Record earliest legal location
3780           set_early_ctrl( n );
3781         }
3782         if (nstack.is_empty()) {
3783           // Finished all nodes on stack.
3784           // Process next node on the worklist.
3785           break;
3786         }
3787         // Get saved parent node and next input's index.
3788         nstack_top_n = nstack.node();
3789         nstack_top_i = nstack.index();
3790         nstack.pop();
3791       }
3792     } // while (true)
3793   }
3794 }
3795 
3796 //------------------------------dom_lca_internal--------------------------------
3797 // Pair-wise LCA
3798 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
3799   if( !n1 ) return n2;          // Handle NULL original LCA
3800   assert( n1->is_CFG(), "" );
3801   assert( n2->is_CFG(), "" );
3802   // find LCA of all uses
3803   uint d1 = dom_depth(n1);
3804   uint d2 = dom_depth(n2);
3805   while (n1 != n2) {
3806     if (d1 > d2) {
3807       n1 =      idom(n1);
3808       d1 = dom_depth(n1);
3809     } else if (d1 < d2) {
3810       n2 =      idom(n2);
3811       d2 = dom_depth(n2);
3812     } else {
3813       // Here d1 == d2.  Due to edits of the dominator-tree, sections
3814       // of the tree might have the same depth.  These sections have
3815       // to be searched more carefully.
3816 
3817       // Scan up all the n1's with equal depth, looking for n2.
3818       Node *t1 = idom(n1);
3819       while (dom_depth(t1) == d1) {
3820         if (t1 == n2)  return n2;
3821         t1 = idom(t1);
3822       }
3823       // Scan up all the n2's with equal depth, looking for n1.
3824       Node *t2 = idom(n2);
3825       while (dom_depth(t2) == d2) {
3826         if (t2 == n1)  return n1;
3827         t2 = idom(t2);
3828       }
3829       // Move up to a new dominator-depth value as well as up the dom-tree.
3830       n1 = t1;
3831       n2 = t2;
3832       d1 = dom_depth(n1);
3833       d2 = dom_depth(n2);
3834     }
3835   }
3836   return n1;
3837 }
3838 
3839 //------------------------------compute_idom-----------------------------------
3840 // Locally compute IDOM using dom_lca call.  Correct only if the incoming
3841 // IDOMs are correct.
3842 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
3843   assert( region->is_Region(), "" );
3844   Node *LCA = NULL;
3845   for( uint i = 1; i < region->req(); i++ ) {
3846     if( region->in(i) != C->top() )
3847       LCA = dom_lca( LCA, region->in(i) );
3848   }
3849   return LCA;
3850 }
3851 
3852 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
3853   bool had_error = false;
3854 #ifdef ASSERT
3855   if (early != C->root()) {
3856     // Make sure that there's a dominance path from LCA to early
3857     Node* d = LCA;
3858     while (d != early) {
3859       if (d == C->root()) {
3860         dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
3861         tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
3862         had_error = true;
3863         break;
3864       }
3865       d = idom(d);
3866     }
3867   }
3868 #endif
3869   return had_error;
3870 }
3871 
3872 
3873 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
3874   // Compute LCA over list of uses
3875   bool had_error = false;
3876   Node *LCA = NULL;
3877   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
3878     Node* c = n->fast_out(i);
3879     if (_nodes[c->_idx] == NULL)
3880       continue;                 // Skip the occasional dead node
3881     if( c->is_Phi() ) {         // For Phis, we must land above on the path
3882       for( uint j=1; j<c->req(); j++ ) {// For all inputs
3883         if( c->in(j) == n ) {   // Found matching input?
3884           Node *use = c->in(0)->in(j);
3885           if (_verify_only && use->is_top()) continue;
3886           LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3887           if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3888         }
3889       }
3890     } else {
3891       // For CFG data-users, use is in the block just prior
3892       Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
3893       LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3894       if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3895     }
3896   }
3897   assert(!had_error, "bad dominance");
3898   return LCA;
3899 }
3900 
3901 // Check the shape of the graph at the loop entry. In some cases,
3902 // the shape of the graph does not match the shape outlined below.
3903 // That is caused by the Opaque1 node "protecting" the shape of
3904 // the graph being removed by, for example, the IGVN performed
3905 // in PhaseIdealLoop::build_and_optimize().
3906 //
3907 // After the Opaque1 node has been removed, optimizations (e.g., split-if,
3908 // loop unswitching, and IGVN, or a combination of them) can freely change
3909 // the graph's shape. As a result, the graph shape outlined below cannot
3910 // be guaranteed anymore.
3911 bool PhaseIdealLoop::is_canonical_loop_entry(CountedLoopNode* cl) {
3912   if (!cl->is_main_loop() && !cl->is_post_loop()) {
3913     return false;
3914   }
3915   Node* ctrl = cl->skip_predicates();
3916 
3917   if (ctrl == NULL || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) {
3918     return false;
3919   }
3920   Node* iffm = ctrl->in(0);
3921   if (iffm == NULL || !iffm->is_If()) {
3922     return false;
3923   }
3924   Node* bolzm = iffm->in(1);
3925   if (bolzm == NULL || !bolzm->is_Bool()) {
3926     return false;
3927   }
3928   Node* cmpzm = bolzm->in(1);
3929   if (cmpzm == NULL || !cmpzm->is_Cmp()) {
3930     return false;
3931   }
3932   // compares can get conditionally flipped
3933   bool found_opaque = false;
3934   for (uint i = 1; i < cmpzm->req(); i++) {
3935     Node* opnd = cmpzm->in(i);
3936     if (opnd && opnd->Opcode() == Op_Opaque1) {
3937       found_opaque = true;
3938       break;
3939     }
3940   }
3941   if (!found_opaque) {
3942     return false;
3943   }
3944   return true;
3945 }
3946 
3947 //------------------------------get_late_ctrl----------------------------------
3948 // Compute latest legal control.
3949 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
3950   assert(early != NULL, "early control should not be NULL");
3951 
3952   Node* LCA = compute_lca_of_uses(n, early);
3953 #ifdef ASSERT
3954   if (LCA == C->root() && LCA != early) {
3955     // def doesn't dominate uses so print some useful debugging output
3956     compute_lca_of_uses(n, early, true);
3957   }
3958 #endif
3959 
3960   // if this is a load, check for anti-dependent stores
3961   // We use a conservative algorithm to identify potential interfering
3962   // instructions and for rescheduling the load.  The users of the memory
3963   // input of this load are examined.  Any use which is not a load and is
3964   // dominated by early is considered a potentially interfering store.
3965   // This can produce false positives.
3966   if (n->is_Load() && LCA != early) {
3967     int load_alias_idx = C->get_alias_index(n->adr_type());
3968     if (C->alias_type(load_alias_idx)->is_rewritable()) {
3969 
3970       Node_List worklist;
3971 
3972       Node *mem = n->in(MemNode::Memory);
3973       for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
3974         Node* s = mem->fast_out(i);
3975         worklist.push(s);
3976       }
3977       while(worklist.size() != 0 && LCA != early) {
3978         Node* s = worklist.pop();
3979         if (s->is_Load() || s->Opcode() == Op_SafePoint ||
3980             (s->is_CallStaticJava() && s->as_CallStaticJava()->uncommon_trap_request() != 0)) {
3981           continue;
3982         } else if (s->is_MergeMem()) {
3983           for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
3984             Node* s1 = s->fast_out(i);
3985             worklist.push(s1);
3986           }
3987         } else {
3988           Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
3989           assert(sctrl != NULL || s->outcnt() == 0, "must have control");
3990           if (sctrl != NULL && !sctrl->is_top() && C->can_alias(s->adr_type(), load_alias_idx) && is_dominator(early, sctrl)) {
3991             LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
3992           }
3993         }
3994       }
3995     }
3996   }
3997 
3998   assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
3999   return LCA;
4000 }
4001 
4002 // true if CFG node d dominates CFG node n
4003 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
4004   if (d == n)
4005     return true;
4006   assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
4007   uint dd = dom_depth(d);
4008   while (dom_depth(n) >= dd) {
4009     if (n == d)
4010       return true;
4011     n = idom(n);
4012   }
4013   return false;
4014 }
4015 
4016 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
4017 // Pair-wise LCA with tags.
4018 // Tag each index with the node 'tag' currently being processed
4019 // before advancing up the dominator chain using idom().
4020 // Later calls that find a match to 'tag' know that this path has already
4021 // been considered in the current LCA (which is input 'n1' by convention).
4022 // Since get_late_ctrl() is only called once for each node, the tag array
4023 // does not need to be cleared between calls to get_late_ctrl().
4024 // Algorithm trades a larger constant factor for better asymptotic behavior
4025 //
4026 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
4027   uint d1 = dom_depth(n1);
4028   uint d2 = dom_depth(n2);
4029 
4030   do {
4031     if (d1 > d2) {
4032       // current lca is deeper than n2
4033       _dom_lca_tags.map(n1->_idx, tag);
4034       n1 =      idom(n1);
4035       d1 = dom_depth(n1);
4036     } else if (d1 < d2) {
4037       // n2 is deeper than current lca
4038       Node *memo = _dom_lca_tags[n2->_idx];
4039       if( memo == tag ) {
4040         return n1;    // Return the current LCA
4041       }
4042       _dom_lca_tags.map(n2->_idx, tag);
4043       n2 =      idom(n2);
4044       d2 = dom_depth(n2);
4045     } else {
4046       // Here d1 == d2.  Due to edits of the dominator-tree, sections
4047       // of the tree might have the same depth.  These sections have
4048       // to be searched more carefully.
4049 
4050       // Scan up all the n1's with equal depth, looking for n2.
4051       _dom_lca_tags.map(n1->_idx, tag);
4052       Node *t1 = idom(n1);
4053       while (dom_depth(t1) == d1) {
4054         if (t1 == n2)  return n2;
4055         _dom_lca_tags.map(t1->_idx, tag);
4056         t1 = idom(t1);
4057       }
4058       // Scan up all the n2's with equal depth, looking for n1.
4059       _dom_lca_tags.map(n2->_idx, tag);
4060       Node *t2 = idom(n2);
4061       while (dom_depth(t2) == d2) {
4062         if (t2 == n1)  return n1;
4063         _dom_lca_tags.map(t2->_idx, tag);
4064         t2 = idom(t2);
4065       }
4066       // Move up to a new dominator-depth value as well as up the dom-tree.
4067       n1 = t1;
4068       n2 = t2;
4069       d1 = dom_depth(n1);
4070       d2 = dom_depth(n2);
4071     }
4072   } while (n1 != n2);
4073   return n1;
4074 }
4075 
4076 //------------------------------init_dom_lca_tags------------------------------
4077 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
4078 // Intended use does not involve any growth for the array, so it could
4079 // be of fixed size.
4080 void PhaseIdealLoop::init_dom_lca_tags() {
4081   uint limit = C->unique() + 1;
4082   _dom_lca_tags.map( limit, NULL );
4083 #ifdef ASSERT
4084   for( uint i = 0; i < limit; ++i ) {
4085     assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
4086   }
4087 #endif // ASSERT
4088 }
4089 
4090 //------------------------------clear_dom_lca_tags------------------------------
4091 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
4092 // Intended use does not involve any growth for the array, so it could
4093 // be of fixed size.
4094 void PhaseIdealLoop::clear_dom_lca_tags() {
4095   uint limit = C->unique() + 1;
4096   _dom_lca_tags.map( limit, NULL );
4097   _dom_lca_tags.clear();
4098 #ifdef ASSERT
4099   for( uint i = 0; i < limit; ++i ) {
4100     assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
4101   }
4102 #endif // ASSERT
4103 }
4104 
4105 //------------------------------build_loop_late--------------------------------
4106 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
4107 // Second pass finds latest legal placement, and ideal loop placement.
4108 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
4109   while (worklist.size() != 0) {
4110     Node *n = worklist.pop();
4111     // Only visit once
4112     if (visited.test_set(n->_idx)) continue;
4113     uint cnt = n->outcnt();
4114     uint   i = 0;
4115     while (true) {
4116       assert( _nodes[n->_idx], "no dead nodes" );
4117       // Visit all children
4118       if (i < cnt) {
4119         Node* use = n->raw_out(i);
4120         ++i;
4121         // Check for dead uses.  Aggressively prune such junk.  It might be
4122         // dead in the global sense, but still have local uses so I cannot
4123         // easily call 'remove_dead_node'.
4124         if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
4125           // Due to cycles, we might not hit the same fixed point in the verify
4126           // pass as we do in the regular pass.  Instead, visit such phis as
4127           // simple uses of the loop head.
4128           if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
4129             if( !visited.test(use->_idx) )
4130               worklist.push(use);
4131           } else if( !visited.test_set(use->_idx) ) {
4132             nstack.push(n, i); // Save parent and next use's index.
4133             n   = use;         // Process all children of current use.
4134             cnt = use->outcnt();
4135             i   = 0;
4136           }
4137         } else {
4138           // Do not visit around the backedge of loops via data edges.
4139           // push dead code onto a worklist
4140           _deadlist.push(use);
4141         }
4142       } else {
4143         // All of n's children have been processed, complete post-processing.
4144         build_loop_late_post(n);
4145         if (nstack.is_empty()) {
4146           // Finished all nodes on stack.
4147           // Process next node on the worklist.
4148           break;
4149         }
4150         // Get saved parent node and next use's index. Visit the rest of uses.
4151         n   = nstack.node();
4152         cnt = n->outcnt();
4153         i   = nstack.index();
4154         nstack.pop();
4155       }
4156     }
4157   }
4158 }
4159 
4160 // Verify that no data node is schedules in the outer loop of a strip
4161 // mined loop.
4162 void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) {
4163 #ifdef ASSERT
4164   if (get_loop(least)->_nest == 0) {
4165     return;
4166   }
4167   IdealLoopTree* loop = get_loop(least);
4168   Node* head = loop->_head;
4169   if (head->is_OuterStripMinedLoop()) {





4170     Node* sfpt = head->as_Loop()->outer_safepoint();
4171     ResourceMark rm;
4172     Unique_Node_List wq;
4173     wq.push(sfpt);
4174     for (uint i = 0; i < wq.size(); i++) {
4175       Node *m = wq.at(i);
4176       for (uint i = 1; i < m->req(); i++) {
4177         Node* nn = m->in(i);
4178         if (nn == n) {
4179           return;
4180         }
4181         if (nn != NULL && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) {
4182           wq.push(nn);
4183         }
4184       }
4185     }
4186     ShouldNotReachHere();
4187   }
4188 #endif
4189 }
4190 
4191 
4192 //------------------------------build_loop_late_post---------------------------
4193 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
4194 // Second pass finds latest legal placement, and ideal loop placement.
4195 void PhaseIdealLoop::build_loop_late_post( Node *n ) {
4196 
4197   if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) {
4198     _igvn._worklist.push(n);  // Maybe we'll normalize it, if no more loops.
4199   }
4200 
4201 #ifdef ASSERT
4202   if (_verify_only && !n->is_CFG()) {
4203     // Check def-use domination.
4204     compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
4205   }
4206 #endif
4207 
4208   // CFG and pinned nodes already handled
4209   if( n->in(0) ) {
4210     if( n->in(0)->is_top() ) return; // Dead?
4211 
4212     // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
4213     // _must_ be pinned (they have to observe their control edge of course).
4214     // Unlike Stores (which modify an unallocable resource, the memory
4215     // state), Mods/Loads can float around.  So free them up.
4216     bool pinned = true;
4217     switch( n->Opcode() ) {
4218     case Op_DivI:
4219     case Op_DivF:
4220     case Op_DivD:
4221     case Op_ModI:
4222     case Op_ModF:
4223     case Op_ModD:
4224     case Op_LoadB:              // Same with Loads; they can sink
4225     case Op_LoadUB:             // during loop optimizations.
4226     case Op_LoadUS:
4227     case Op_LoadD:
4228     case Op_LoadF:
4229     case Op_LoadI:
4230     case Op_LoadKlass:
4231     case Op_LoadNKlass:
4232     case Op_LoadL:
4233     case Op_LoadS:
4234     case Op_LoadP:
4235     case Op_LoadBarrierSlowReg:
4236     case Op_LoadBarrierWeakSlowReg:
4237     case Op_LoadN:
4238     case Op_LoadRange:
4239     case Op_LoadD_unaligned:
4240     case Op_LoadL_unaligned:
4241     case Op_StrComp:            // Does a bunch of load-like effects
4242     case Op_StrEquals:
4243     case Op_StrIndexOf:
4244     case Op_StrIndexOfChar:
4245     case Op_AryEq:
4246     case Op_HasNegatives:
4247       pinned = false;
4248     }





4249     if( pinned ) {
4250       IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
4251       if( !chosen_loop->_child )       // Inner loop?
4252         chosen_loop->_body.push(n); // Collect inner loops
4253       return;
4254     }
4255   } else {                      // No slot zero
4256     if( n->is_CFG() ) {         // CFG with no slot 0 is dead
4257       _nodes.map(n->_idx,0);    // No block setting, it's globally dead
4258       return;
4259     }
4260     assert(!n->is_CFG() || n->outcnt() == 0, "");
4261   }
4262 
4263   // Do I have a "safe range" I can select over?
4264   Node *early = get_ctrl(n);// Early location already computed
4265 
4266   // Compute latest point this Node can go
4267   Node *LCA = get_late_ctrl( n, early );
4268   // LCA is NULL due to uses being dead
4269   if( LCA == NULL ) {
4270 #ifdef ASSERT
4271     for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
4272       assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
4273     }
4274 #endif
4275     _nodes.map(n->_idx, 0);     // This node is useless
4276     _deadlist.push(n);
4277     return;
4278   }
4279   assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
4280 
4281   Node *legal = LCA;            // Walk 'legal' up the IDOM chain
4282   Node *least = legal;          // Best legal position so far
4283   while( early != legal ) {     // While not at earliest legal
4284 #ifdef ASSERT
4285     if (legal->is_Start() && !early->is_Root()) {
4286       // Bad graph. Print idom path and fail.
4287       dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
4288       assert(false, "Bad graph detected in build_loop_late");
4289     }
4290 #endif
4291     // Find least loop nesting depth
4292     legal = idom(legal);        // Bump up the IDOM tree
4293     // Check for lower nesting depth
4294     if( get_loop(legal)->_nest < get_loop(least)->_nest )
4295       least = legal;
4296   }
4297   assert(early == legal || legal != C->root(), "bad dominance of inputs");
4298 
4299   // Try not to place code on a loop entry projection
4300   // which can inhibit range check elimination.
4301   if (least != early) {
4302     Node* ctrl_out = least->unique_ctrl_out();
4303     if (ctrl_out && ctrl_out->is_Loop() &&
4304         least == ctrl_out->in(LoopNode::EntryControl)) {
4305       // Move the node above predicates as far up as possible so a
4306       // following pass of loop predication doesn't hoist a predicate
4307       // that depends on it above that node.
4308       Node* new_ctrl = least;
4309       for (;;) {
4310         if (!new_ctrl->is_Proj()) {
4311           break;
4312         }
4313         CallStaticJavaNode* call = new_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
4314         if (call == NULL) {
4315           break;
4316         }
4317         int req = call->uncommon_trap_request();
4318         Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req);
4319         if (trap_reason != Deoptimization::Reason_loop_limit_check &&
4320             trap_reason != Deoptimization::Reason_predicate &&
4321             trap_reason != Deoptimization::Reason_profile_predicate) {
4322           break;
4323         }
4324         Node* c = new_ctrl->in(0)->in(0);
4325         if (is_dominator(c, early) && c != early) {
4326           break;
4327         }
4328         new_ctrl = c;
4329       }
4330       least = new_ctrl;
4331     }
4332   }
4333 
4334 #ifdef ASSERT
4335   // If verifying, verify that 'verify_me' has a legal location
4336   // and choose it as our location.
4337   if( _verify_me ) {
4338     Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
4339     Node *legal = LCA;
4340     while( early != legal ) {   // While not at earliest legal
4341       if( legal == v_ctrl ) break;  // Check for prior good location
4342       legal = idom(legal)      ;// Bump up the IDOM tree
4343     }
4344     // Check for prior good location
4345     if( legal == v_ctrl ) least = legal; // Keep prior if found
4346   }
4347 #endif
4348 
4349   // Assign discovered "here or above" point
4350   least = find_non_split_ctrl(least);
4351   verify_strip_mined_scheduling(n, least);
4352   set_ctrl(n, least);
4353 
4354   // Collect inner loop bodies
4355   IdealLoopTree *chosen_loop = get_loop(least);
4356   if( !chosen_loop->_child )   // Inner loop?
4357     chosen_loop->_body.push(n);// Collect inner loops
4358 }
4359 
4360 #ifdef ASSERT
4361 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
4362   tty->print_cr("%s", msg);
4363   tty->print("n: "); n->dump();
4364   tty->print("early(n): "); early->dump();
4365   if (n->in(0) != NULL  && !n->in(0)->is_top() &&
4366       n->in(0) != early && !n->in(0)->is_Root()) {
4367     tty->print("n->in(0): "); n->in(0)->dump();
4368   }
4369   for (uint i = 1; i < n->req(); i++) {
4370     Node* in1 = n->in(i);
4371     if (in1 != NULL && in1 != n && !in1->is_top()) {
4372       tty->print("n->in(%d): ", i); in1->dump();
4373       Node* in1_early = get_ctrl(in1);
4374       tty->print("early(n->in(%d)): ", i); in1_early->dump();
4375       if (in1->in(0) != NULL     && !in1->in(0)->is_top() &&
4376           in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
4377         tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
4378       }
4379       for (uint j = 1; j < in1->req(); j++) {
4380         Node* in2 = in1->in(j);
4381         if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) {
4382           tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
4383           Node* in2_early = get_ctrl(in2);
4384           tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
4385           if (in2->in(0) != NULL     && !in2->in(0)->is_top() &&
4386               in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
4387             tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
4388           }
4389         }
4390       }
4391     }
4392   }
4393   tty->cr();
4394   tty->print("LCA(n): "); LCA->dump();
4395   for (uint i = 0; i < n->outcnt(); i++) {
4396     Node* u1 = n->raw_out(i);
4397     if (u1 == n)
4398       continue;
4399     tty->print("n->out(%d): ", i); u1->dump();
4400     if (u1->is_CFG()) {
4401       for (uint j = 0; j < u1->outcnt(); j++) {
4402         Node* u2 = u1->raw_out(j);
4403         if (u2 != u1 && u2 != n && u2->is_CFG()) {
4404           tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
4405         }
4406       }
4407     } else {
4408       Node* u1_later = get_ctrl(u1);
4409       tty->print("later(n->out(%d)): ", i); u1_later->dump();
4410       if (u1->in(0) != NULL     && !u1->in(0)->is_top() &&
4411           u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
4412         tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
4413       }
4414       for (uint j = 0; j < u1->outcnt(); j++) {
4415         Node* u2 = u1->raw_out(j);
4416         if (u2 == n || u2 == u1)
4417           continue;
4418         tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
4419         if (!u2->is_CFG()) {
4420           Node* u2_later = get_ctrl(u2);
4421           tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
4422           if (u2->in(0) != NULL     && !u2->in(0)->is_top() &&
4423               u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
4424             tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
4425           }
4426         }
4427       }
4428     }
4429   }
4430   tty->cr();
4431   int ct = 0;
4432   Node *dbg_legal = LCA;
4433   while(!dbg_legal->is_Start() && ct < 100) {
4434     tty->print("idom[%d] ",ct); dbg_legal->dump();
4435     ct++;
4436     dbg_legal = idom(dbg_legal);
4437   }
4438   tty->cr();
4439 }
4440 #endif
4441 
4442 #ifndef PRODUCT
4443 //------------------------------dump-------------------------------------------
4444 void PhaseIdealLoop::dump( ) const {
4445   ResourceMark rm;
4446   Arena* arena = Thread::current()->resource_area();
4447   Node_Stack stack(arena, C->live_nodes() >> 2);
4448   Node_List rpo_list;
4449   VectorSet visited(arena);
4450   visited.set(C->top()->_idx);
4451   rpo( C->root(), stack, visited, rpo_list );
4452   // Dump root loop indexed by last element in PO order
4453   dump( _ltree_root, rpo_list.size(), rpo_list );
4454 }
4455 
4456 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
4457   loop->dump_head();
4458 
4459   // Now scan for CFG nodes in the same loop
4460   for( uint j=idx; j > 0;  j-- ) {
4461     Node *n = rpo_list[j-1];
4462     if( !_nodes[n->_idx] )      // Skip dead nodes
4463       continue;
4464     if( get_loop(n) != loop ) { // Wrong loop nest
4465       if( get_loop(n)->_head == n &&    // Found nested loop?
4466           get_loop(n)->_parent == loop )
4467         dump(get_loop(n),rpo_list.size(),rpo_list);     // Print it nested-ly
4468       continue;
4469     }
4470 
4471     // Dump controlling node
4472     for( uint x = 0; x < loop->_nest; x++ )
4473       tty->print("  ");
4474     tty->print("C");
4475     if( n == C->root() ) {
4476       n->dump();
4477     } else {
4478       Node* cached_idom   = idom_no_update(n);
4479       Node *computed_idom = n->in(0);
4480       if( n->is_Region() ) {
4481         computed_idom = compute_idom(n);
4482         // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
4483         // any MultiBranch ctrl node), so apply a similar transform to
4484         // the cached idom returned from idom_no_update.
4485         cached_idom = find_non_split_ctrl(cached_idom);
4486       }
4487       tty->print(" ID:%d",computed_idom->_idx);
4488       n->dump();
4489       if( cached_idom != computed_idom ) {
4490         tty->print_cr("*** BROKEN IDOM!  Computed as: %d, cached as: %d",
4491                       computed_idom->_idx, cached_idom->_idx);
4492       }
4493     }
4494     // Dump nodes it controls
4495     for( uint k = 0; k < _nodes.Size(); k++ ) {
4496       // (k < C->unique() && get_ctrl(find(k)) == n)
4497       if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
4498         Node *m = C->root()->find(k);
4499         if( m && m->outcnt() > 0 ) {
4500           if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
4501             tty->print_cr("*** BROKEN CTRL ACCESSOR!  _nodes[k] is %p, ctrl is %p",
4502                           _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
4503           }
4504           for( uint j = 0; j < loop->_nest; j++ )
4505             tty->print("  ");
4506           tty->print(" ");
4507           m->dump();
4508         }
4509       }
4510     }
4511   }
4512 }

4513 

4514 // Collect a R-P-O for the whole CFG.
4515 // Result list is in post-order (scan backwards for RPO)
4516 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
4517   stk.push(start, 0);
4518   visited.set(start->_idx);
4519 
4520   while (stk.is_nonempty()) {
4521     Node* m   = stk.node();
4522     uint  idx = stk.index();
4523     if (idx < m->outcnt()) {
4524       stk.set_index(idx + 1);
4525       Node* n = m->raw_out(idx);
4526       if (n->is_CFG() && !visited.test_set(n->_idx)) {
4527         stk.push(n, 0);
4528       }
4529     } else {
4530       rpo_list.push(m);
4531       stk.pop();
4532     }
4533   }
4534 }
4535 #endif
4536 
4537 
4538 //=============================================================================
4539 //------------------------------LoopTreeIterator-----------------------------------
4540 
4541 // Advance to next loop tree using a preorder, left-to-right traversal.
4542 void LoopTreeIterator::next() {
4543   assert(!done(), "must not be done.");
4544   if (_curnt->_child != NULL) {
4545     _curnt = _curnt->_child;
4546   } else if (_curnt->_next != NULL) {
4547     _curnt = _curnt->_next;
4548   } else {
4549     while (_curnt != _root && _curnt->_next == NULL) {
4550       _curnt = _curnt->_parent;
4551     }
4552     if (_curnt == _root) {
4553       _curnt = NULL;
4554       assert(done(), "must be done.");
4555     } else {
4556       assert(_curnt->_next != NULL, "must be more to do");
4557       _curnt = _curnt->_next;
4558     }
4559   }
4560 }
--- EOF ---