1 /*
   2  * Copyright (c) 1998, 2015, 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 "libadt/vectset.hpp"
  29 #include "memory/allocation.inline.hpp"
  30 #include "opto/addnode.hpp"
  31 #include "opto/callnode.hpp"
  32 #include "opto/connode.hpp"
  33 #include "opto/convertnode.hpp"
  34 #include "opto/divnode.hpp"
  35 #include "opto/idealGraphPrinter.hpp"
  36 #include "opto/loopnode.hpp"
  37 #include "opto/mulnode.hpp"
  38 #include "opto/rootnode.hpp"
  39 #include "opto/superword.hpp"
  40 
  41 //=============================================================================
  42 //------------------------------is_loop_iv-------------------------------------
  43 // Determine if a node is Counted loop induction variable.
  44 // The method is declared in node.hpp.
  45 const Node* Node::is_loop_iv() const {
  46   if (this->is_Phi() && !this->as_Phi()->is_copy() &&
  47       this->as_Phi()->region()->is_CountedLoop() &&
  48       this->as_Phi()->region()->as_CountedLoop()->phi() == this) {
  49     return this;
  50   } else {
  51     return NULL;
  52   }
  53 }
  54 
  55 //=============================================================================
  56 //------------------------------dump_spec--------------------------------------
  57 // Dump special per-node info
  58 #ifndef PRODUCT
  59 void LoopNode::dump_spec(outputStream *st) const {
  60   if (is_inner_loop()) st->print( "inner " );
  61   if (is_partial_peel_loop()) st->print( "partial_peel " );
  62   if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
  63 }
  64 #endif
  65 
  66 //------------------------------is_valid_counted_loop-------------------------
  67 bool LoopNode::is_valid_counted_loop() const {
  68   if (is_CountedLoop()) {
  69     CountedLoopNode*    l  = as_CountedLoop();
  70     CountedLoopEndNode* le = l->loopexit();
  71     if (le != NULL &&
  72         le->proj_out(1 /* true */) == l->in(LoopNode::LoopBackControl)) {
  73       Node* phi  = l->phi();
  74       Node* exit = le->proj_out(0 /* false */);
  75       if (exit != NULL && exit->Opcode() == Op_IfFalse &&
  76           phi != NULL && phi->is_Phi() &&
  77           phi->in(LoopNode::LoopBackControl) == l->incr() &&
  78           le->loopnode() == l && le->stride_is_con()) {
  79         return true;
  80       }
  81     }
  82   }
  83   return false;
  84 }
  85 
  86 //------------------------------get_early_ctrl---------------------------------
  87 // Compute earliest legal control
  88 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {
  89   assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
  90   uint i;
  91   Node *early;
  92   if (n->in(0) && !n->is_expensive()) {
  93     early = n->in(0);
  94     if (!early->is_CFG()) // Might be a non-CFG multi-def
  95       early = get_ctrl(early);        // So treat input as a straight data input
  96     i = 1;
  97   } else {
  98     early = get_ctrl(n->in(1));
  99     i = 2;
 100   }
 101   uint e_d = dom_depth(early);
 102   assert( early, "" );
 103   for (; i < n->req(); i++) {
 104     Node *cin = get_ctrl(n->in(i));
 105     assert( cin, "" );
 106     // Keep deepest dominator depth
 107     uint c_d = dom_depth(cin);
 108     if (c_d > e_d) {           // Deeper guy?
 109       early = cin;              // Keep deepest found so far
 110       e_d = c_d;
 111     } else if (c_d == e_d &&    // Same depth?
 112                early != cin) { // If not equal, must use slower algorithm
 113       // If same depth but not equal, one _must_ dominate the other
 114       // and we want the deeper (i.e., dominated) guy.
 115       Node *n1 = early;
 116       Node *n2 = cin;
 117       while (1) {
 118         n1 = idom(n1);          // Walk up until break cycle
 119         n2 = idom(n2);
 120         if (n1 == cin ||        // Walked early up to cin
 121             dom_depth(n2) < c_d)
 122           break;                // early is deeper; keep him
 123         if (n2 == early ||      // Walked cin up to early
 124             dom_depth(n1) < c_d) {
 125           early = cin;          // cin is deeper; keep him
 126           break;
 127         }
 128       }
 129       e_d = dom_depth(early);   // Reset depth register cache
 130     }
 131   }
 132 
 133   // Return earliest legal location
 134   assert(early == find_non_split_ctrl(early), "unexpected early control");
 135 
 136   if (n->is_expensive() && !_verify_only && !_verify_me) {
 137     assert(n->in(0), "should have control input");
 138     early = get_early_ctrl_for_expensive(n, early);
 139   }
 140 
 141   return early;
 142 }
 143 
 144 //------------------------------get_early_ctrl_for_expensive---------------------------------
 145 // Move node up the dominator tree as high as legal while still beneficial
 146 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {
 147   assert(n->in(0) && n->is_expensive(), "expensive node with control input here");
 148   assert(OptimizeExpensiveOps, "optimization off?");
 149 
 150   Node* ctl = n->in(0);
 151   assert(ctl->is_CFG(), "expensive input 0 must be cfg");
 152   uint min_dom_depth = dom_depth(earliest);
 153 #ifdef ASSERT
 154   if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) {
 155     dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl);
 156     assert(false, "Bad graph detected in get_early_ctrl_for_expensive");
 157   }
 158 #endif
 159   if (dom_depth(ctl) < min_dom_depth) {
 160     return earliest;
 161   }
 162 
 163   while (1) {
 164     Node *next = ctl;
 165     // Moving the node out of a loop on the projection of a If
 166     // confuses loop predication. So once we hit a Loop in a If branch
 167     // that doesn't branch to an UNC, we stop. The code that process
 168     // expensive nodes will notice the loop and skip over it to try to
 169     // move the node further up.
 170     if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) {
 171       if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
 172         break;
 173       }
 174       next = idom(ctl->in(1)->in(0));
 175     } else if (ctl->is_Proj()) {
 176       // We only move it up along a projection if the projection is
 177       // the single control projection for its parent: same code path,
 178       // if it's a If with UNC or fallthrough of a call.
 179       Node* parent_ctl = ctl->in(0);
 180       if (parent_ctl == NULL) {
 181         break;
 182       } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) {
 183         next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control();
 184       } else if (parent_ctl->is_If()) {
 185         if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
 186           break;
 187         }
 188         assert(idom(ctl) == parent_ctl, "strange");
 189         next = idom(parent_ctl);
 190       } else if (ctl->is_CatchProj()) {
 191         if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) {
 192           break;
 193         }
 194         assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph");
 195         next = parent_ctl->in(0)->in(0)->in(0);
 196       } else {
 197         // Check if parent control has a single projection (this
 198         // control is the only possible successor of the parent
 199         // control). If so, we can try to move the node above the
 200         // parent control.
 201         int nb_ctl_proj = 0;
 202         for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) {
 203           Node *p = parent_ctl->fast_out(i);
 204           if (p->is_Proj() && p->is_CFG()) {
 205             nb_ctl_proj++;
 206             if (nb_ctl_proj > 1) {
 207               break;
 208             }
 209           }
 210         }
 211 
 212         if (nb_ctl_proj > 1) {
 213           break;
 214         }
 215         assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call(), "unexpected node");
 216         assert(idom(ctl) == parent_ctl, "strange");
 217         next = idom(parent_ctl);
 218       }
 219     } else {
 220       next = idom(ctl);
 221     }
 222     if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) {
 223       break;
 224     }
 225     ctl = next;
 226   }
 227 
 228   if (ctl != n->in(0)) {
 229     _igvn.replace_input_of(n, 0, ctl);
 230     _igvn.hash_insert(n);
 231   }
 232 
 233   return ctl;
 234 }
 235 
 236 
 237 //------------------------------set_early_ctrl---------------------------------
 238 // Set earliest legal control
 239 void PhaseIdealLoop::set_early_ctrl( Node *n ) {
 240   Node *early = get_early_ctrl(n);
 241 
 242   // Record earliest legal location
 243   set_ctrl(n, early);
 244 }
 245 
 246 //------------------------------set_subtree_ctrl-------------------------------
 247 // set missing _ctrl entries on new nodes
 248 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) {
 249   // Already set?  Get out.
 250   if( _nodes[n->_idx] ) return;
 251   // Recursively set _nodes array to indicate where the Node goes
 252   uint i;
 253   for( i = 0; i < n->req(); ++i ) {
 254     Node *m = n->in(i);
 255     if( m && m != C->root() )
 256       set_subtree_ctrl( m );
 257   }
 258 
 259   // Fixup self
 260   set_early_ctrl( n );
 261 }
 262 
 263 //------------------------------is_counted_loop--------------------------------
 264 bool PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) {
 265   PhaseGVN *gvn = &_igvn;
 266 
 267   // Counted loop head must be a good RegionNode with only 3 not NULL
 268   // control input edges: Self, Entry, LoopBack.
 269   if (x->in(LoopNode::Self) == NULL || x->req() != 3 || loop->_irreducible) {
 270     return false;
 271   }
 272   Node *init_control = x->in(LoopNode::EntryControl);
 273   Node *back_control = x->in(LoopNode::LoopBackControl);
 274   if (init_control == NULL || back_control == NULL)    // Partially dead
 275     return false;
 276   // Must also check for TOP when looking for a dead loop
 277   if (init_control->is_top() || back_control->is_top())
 278     return false;
 279 
 280   // Allow funny placement of Safepoint
 281   if (back_control->Opcode() == Op_SafePoint)
 282     back_control = back_control->in(TypeFunc::Control);
 283 
 284   // Controlling test for loop
 285   Node *iftrue = back_control;
 286   uint iftrue_op = iftrue->Opcode();
 287   if (iftrue_op != Op_IfTrue &&
 288       iftrue_op != Op_IfFalse)
 289     // I have a weird back-control.  Probably the loop-exit test is in
 290     // the middle of the loop and I am looking at some trailing control-flow
 291     // merge point.  To fix this I would have to partially peel the loop.
 292     return false; // Obscure back-control
 293 
 294   // Get boolean guarding loop-back test
 295   Node *iff = iftrue->in(0);
 296   if (get_loop(iff) != loop || !iff->in(1)->is_Bool())
 297     return false;
 298   BoolNode *test = iff->in(1)->as_Bool();
 299   BoolTest::mask bt = test->_test._test;
 300   float cl_prob = iff->as_If()->_prob;
 301   if (iftrue_op == Op_IfFalse) {
 302     bt = BoolTest(bt).negate();
 303     cl_prob = 1.0 - cl_prob;
 304   }
 305   // Get backedge compare
 306   Node *cmp = test->in(1);
 307   int cmp_op = cmp->Opcode();
 308   if (cmp_op != Op_CmpI)
 309     return false;                // Avoid pointer & float compares
 310 
 311   // Find the trip-counter increment & limit.  Limit must be loop invariant.
 312   Node *incr  = cmp->in(1);
 313   Node *limit = cmp->in(2);
 314 
 315   // ---------
 316   // need 'loop()' test to tell if limit is loop invariant
 317   // ---------
 318 
 319   if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?
 320     Node *tmp = incr;            // Then reverse order into the CmpI
 321     incr = limit;
 322     limit = tmp;
 323     bt = BoolTest(bt).commute(); // And commute the exit test
 324   }
 325   if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant
 326     return false;
 327   if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
 328     return false;
 329 
 330   Node* phi_incr = NULL;
 331   // Trip-counter increment must be commutative & associative.
 332   if (incr->Opcode() == Op_CastII) {
 333     incr = incr->in(1);
 334   }
 335   if (incr->is_Phi()) {
 336     if (incr->as_Phi()->region() != x || incr->req() != 3)
 337       return false; // Not simple trip counter expression
 338     phi_incr = incr;
 339     incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
 340     if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant
 341       return false;
 342   }
 343 
 344   Node* trunc1 = NULL;
 345   Node* trunc2 = NULL;
 346   const TypeInt* iv_trunc_t = NULL;
 347   if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) {
 348     return false; // Funny increment opcode
 349   }
 350   assert(incr->Opcode() == Op_AddI, "wrong increment code");
 351 
 352   // Get merge point
 353   Node *xphi = incr->in(1);
 354   Node *stride = incr->in(2);
 355   if (!stride->is_Con()) {     // Oops, swap these
 356     if (!xphi->is_Con())       // Is the other guy a constant?
 357       return false;             // Nope, unknown stride, bail out
 358     Node *tmp = xphi;           // 'incr' is commutative, so ok to swap
 359     xphi = stride;
 360     stride = tmp;
 361   }
 362   if (xphi->Opcode() == Op_CastII) {
 363     xphi = xphi->in(1);
 364   }
 365   // Stride must be constant
 366   int stride_con = stride->get_int();
 367   if (stride_con == 0)
 368     return false; // missed some peephole opt
 369 
 370   if (!xphi->is_Phi())
 371     return false; // Too much math on the trip counter
 372   if (phi_incr != NULL && phi_incr != xphi)
 373     return false;
 374   PhiNode *phi = xphi->as_Phi();
 375 
 376   // Phi must be of loop header; backedge must wrap to increment
 377   if (phi->region() != x)
 378     return false;
 379   if (trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr ||
 380       trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1) {
 381     return false;
 382   }
 383   Node *init_trip = phi->in(LoopNode::EntryControl);
 384 
 385   // If iv trunc type is smaller than int, check for possible wrap.
 386   if (!TypeInt::INT->higher_equal(iv_trunc_t)) {
 387     assert(trunc1 != NULL, "must have found some truncation");
 388 
 389     // Get a better type for the phi (filtered thru if's)
 390     const TypeInt* phi_ft = filtered_type(phi);
 391 
 392     // Can iv take on a value that will wrap?
 393     //
 394     // Ensure iv's limit is not within "stride" of the wrap value.
 395     //
 396     // Example for "short" type
 397     //    Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
 398     //    If the stride is +10, then the last value of the induction
 399     //    variable before the increment (phi_ft->_hi) must be
 400     //    <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
 401     //    ensure no truncation occurs after the increment.
 402 
 403     if (stride_con > 0) {
 404       if (iv_trunc_t->_hi - phi_ft->_hi < stride_con ||
 405           iv_trunc_t->_lo > phi_ft->_lo) {
 406         return false;  // truncation may occur
 407       }
 408     } else if (stride_con < 0) {
 409       if (iv_trunc_t->_lo - phi_ft->_lo > stride_con ||
 410           iv_trunc_t->_hi < phi_ft->_hi) {
 411         return false;  // truncation may occur
 412       }
 413     }
 414     // No possibility of wrap so truncation can be discarded
 415     // Promote iv type to Int
 416   } else {
 417     assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int");
 418   }
 419 
 420   // If the condition is inverted and we will be rolling
 421   // through MININT to MAXINT, then bail out.
 422   if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
 423       // Odd stride
 424       bt == BoolTest::ne && stride_con != 1 && stride_con != -1 ||
 425       // Count down loop rolls through MAXINT
 426       (bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0 ||
 427       // Count up loop rolls through MININT
 428       (bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0) {
 429     return false; // Bail out
 430   }
 431 
 432   const TypeInt* init_t = gvn->type(init_trip)->is_int();
 433   const TypeInt* limit_t = gvn->type(limit)->is_int();
 434 
 435   if (stride_con > 0) {
 436     jlong init_p = (jlong)init_t->_lo + stride_con;
 437     if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi)
 438       return false; // cyclic loop or this loop trips only once
 439   } else {
 440     jlong init_p = (jlong)init_t->_hi + stride_con;
 441     if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo)
 442       return false; // cyclic loop or this loop trips only once
 443   }
 444 
 445   if (phi_incr != NULL) {
 446     // check if there is a possiblity of IV overflowing after the first increment
 447     if (stride_con > 0) {
 448       if (init_t->_hi > max_jint - stride_con) {
 449         return false;
 450       }
 451     } else {
 452       if (init_t->_lo < min_jint - stride_con) {
 453         return false;
 454       }
 455     }
 456   }
 457 
 458   // =================================================
 459   // ---- SUCCESS!   Found A Trip-Counted Loop!  -----
 460   //
 461   assert(x->Opcode() == Op_Loop, "regular loops only");
 462   C->print_method(PHASE_BEFORE_CLOOPS, 3);
 463 
 464   Node *hook = new Node(6);
 465 
 466   if (LoopLimitCheck) {
 467 
 468   // ===================================================
 469   // Generate loop limit check to avoid integer overflow
 470   // in cases like next (cyclic loops):
 471   //
 472   // for (i=0; i <= max_jint; i++) {}
 473   // for (i=0; i <  max_jint; i+=2) {}
 474   //
 475   //
 476   // Limit check predicate depends on the loop test:
 477   //
 478   // for(;i != limit; i++)       --> limit <= (max_jint)
 479   // for(;i <  limit; i+=stride) --> limit <= (max_jint - stride + 1)
 480   // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride    )
 481   //
 482 
 483   // Check if limit is excluded to do more precise int overflow check.
 484   bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge);
 485   int stride_m  = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1));
 486 
 487   // If compare points directly to the phi we need to adjust
 488   // the compare so that it points to the incr. Limit have
 489   // to be adjusted to keep trip count the same and the
 490   // adjusted limit should be checked for int overflow.
 491   if (phi_incr != NULL) {
 492     stride_m  += stride_con;
 493   }
 494 
 495   if (limit->is_Con()) {
 496     int limit_con = limit->get_int();
 497     if ((stride_con > 0 && limit_con > (max_jint - stride_m)) ||
 498         (stride_con < 0 && limit_con < (min_jint - stride_m))) {
 499       // Bailout: it could be integer overflow.
 500       return false;
 501     }
 502   } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) ||
 503              (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) {
 504       // Limit's type may satisfy the condition, for example,
 505       // when it is an array length.
 506   } else {
 507     // Generate loop's limit check.
 508     // Loop limit check predicate should be near the loop.
 509     ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check);
 510     if (!limit_check_proj) {
 511       // The limit check predicate is not generated if this method trapped here before.
 512 #ifdef ASSERT
 513       if (TraceLoopLimitCheck) {
 514         tty->print("missing loop limit check:");
 515         loop->dump_head();
 516         x->dump(1);
 517       }
 518 #endif
 519       return false;
 520     }
 521 
 522     IfNode* check_iff = limit_check_proj->in(0)->as_If();
 523     Node* cmp_limit;
 524     Node* bol;
 525 
 526     if (stride_con > 0) {
 527       cmp_limit = new CmpINode(limit, _igvn.intcon(max_jint - stride_m));
 528       bol = new BoolNode(cmp_limit, BoolTest::le);
 529     } else {
 530       cmp_limit = new CmpINode(limit, _igvn.intcon(min_jint - stride_m));
 531       bol = new BoolNode(cmp_limit, BoolTest::ge);
 532     }
 533     cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
 534     bol = _igvn.register_new_node_with_optimizer(bol);
 535     set_subtree_ctrl(bol);
 536 
 537     // Replace condition in original predicate but preserve Opaque node
 538     // so that previous predicates could be found.
 539     assert(check_iff->in(1)->Opcode() == Op_Conv2B &&
 540            check_iff->in(1)->in(1)->Opcode() == Op_Opaque1, "");
 541     Node* opq = check_iff->in(1)->in(1);
 542     _igvn.replace_input_of(opq, 1, bol);
 543     // Update ctrl.
 544     set_ctrl(opq, check_iff->in(0));
 545     set_ctrl(check_iff->in(1), check_iff->in(0));
 546 
 547 #ifndef PRODUCT
 548     // report that the loop predication has been actually performed
 549     // for this loop
 550     if (TraceLoopLimitCheck) {
 551       tty->print_cr("Counted Loop Limit Check generated:");
 552       debug_only( bol->dump(2); )
 553     }
 554 #endif
 555   }
 556 
 557   if (phi_incr != NULL) {
 558     // If compare points directly to the phi we need to adjust
 559     // the compare so that it points to the incr. Limit have
 560     // to be adjusted to keep trip count the same and we
 561     // should avoid int overflow.
 562     //
 563     //   i = init; do {} while(i++ < limit);
 564     // is converted to
 565     //   i = init; do {} while(++i < limit+1);
 566     //
 567     limit = gvn->transform(new AddINode(limit, stride));
 568   }
 569 
 570   // Now we need to canonicalize loop condition.
 571   if (bt == BoolTest::ne) {
 572     assert(stride_con == 1 || stride_con == -1, "simple increment only");
 573     // 'ne' can be replaced with 'lt' only when init < limit.
 574     if (stride_con > 0 && init_t->_hi < limit_t->_lo)
 575       bt = BoolTest::lt;
 576     // 'ne' can be replaced with 'gt' only when init > limit.
 577     if (stride_con < 0 && init_t->_lo > limit_t->_hi)
 578       bt = BoolTest::gt;
 579   }
 580 
 581   if (incl_limit) {
 582     // The limit check guaranties that 'limit <= (max_jint - stride)' so
 583     // we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
 584     //
 585     Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1);
 586     limit = gvn->transform(new AddINode(limit, one));
 587     if (bt == BoolTest::le)
 588       bt = BoolTest::lt;
 589     else if (bt == BoolTest::ge)
 590       bt = BoolTest::gt;
 591     else
 592       ShouldNotReachHere();
 593   }
 594   set_subtree_ctrl( limit );
 595 
 596   } else { // LoopLimitCheck
 597 
 598   // If compare points to incr, we are ok.  Otherwise the compare
 599   // can directly point to the phi; in this case adjust the compare so that
 600   // it points to the incr by adjusting the limit.
 601   if (cmp->in(1) == phi || cmp->in(2) == phi)
 602     limit = gvn->transform(new AddINode(limit,stride));
 603 
 604   // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride.
 605   // Final value for iterator should be: trip_count * stride + init_trip.
 606   Node *one_p = gvn->intcon( 1);
 607   Node *one_m = gvn->intcon(-1);
 608 
 609   Node *trip_count = NULL;
 610   switch( bt ) {
 611   case BoolTest::eq:
 612     ShouldNotReachHere();
 613   case BoolTest::ne:            // Ahh, the case we desire
 614     if (stride_con == 1)
 615       trip_count = gvn->transform(new SubINode(limit,init_trip));
 616     else if (stride_con == -1)
 617       trip_count = gvn->transform(new SubINode(init_trip,limit));
 618     else
 619       ShouldNotReachHere();
 620     set_subtree_ctrl(trip_count);
 621     //_loop.map(trip_count->_idx,loop(limit));
 622     break;
 623   case BoolTest::le:            // Maybe convert to '<' case
 624     limit = gvn->transform(new AddINode(limit,one_p));
 625     set_subtree_ctrl( limit );
 626     hook->init_req(4, limit);
 627 
 628     bt = BoolTest::lt;
 629     // Make the new limit be in the same loop nest as the old limit
 630     //_loop.map(limit->_idx,limit_loop);
 631     // Fall into next case
 632   case BoolTest::lt: {          // Maybe convert to '!=' case
 633     if (stride_con < 0) // Count down loop rolls through MAXINT
 634       ShouldNotReachHere();
 635     Node *range = gvn->transform(new SubINode(limit,init_trip));
 636     set_subtree_ctrl( range );
 637     hook->init_req(0, range);
 638 
 639     Node *bias  = gvn->transform(new AddINode(range,stride));
 640     set_subtree_ctrl( bias );
 641     hook->init_req(1, bias);
 642 
 643     Node *bias1 = gvn->transform(new AddINode(bias,one_m));
 644     set_subtree_ctrl( bias1 );
 645     hook->init_req(2, bias1);
 646 
 647     trip_count  = gvn->transform(new DivINode(0,bias1,stride));
 648     set_subtree_ctrl( trip_count );
 649     hook->init_req(3, trip_count);
 650     break;
 651   }
 652 
 653   case BoolTest::ge:            // Maybe convert to '>' case
 654     limit = gvn->transform(new AddINode(limit,one_m));
 655     set_subtree_ctrl( limit );
 656     hook->init_req(4 ,limit);
 657 
 658     bt = BoolTest::gt;
 659     // Make the new limit be in the same loop nest as the old limit
 660     //_loop.map(limit->_idx,limit_loop);
 661     // Fall into next case
 662   case BoolTest::gt: {          // Maybe convert to '!=' case
 663     if (stride_con > 0) // count up loop rolls through MININT
 664       ShouldNotReachHere();
 665     Node *range = gvn->transform(new SubINode(limit,init_trip));
 666     set_subtree_ctrl( range );
 667     hook->init_req(0, range);
 668 
 669     Node *bias  = gvn->transform(new AddINode(range,stride));
 670     set_subtree_ctrl( bias );
 671     hook->init_req(1, bias);
 672 
 673     Node *bias1 = gvn->transform(new AddINode(bias,one_p));
 674     set_subtree_ctrl( bias1 );
 675     hook->init_req(2, bias1);
 676 
 677     trip_count  = gvn->transform(new DivINode(0,bias1,stride));
 678     set_subtree_ctrl( trip_count );
 679     hook->init_req(3, trip_count);
 680     break;
 681   }
 682   } // switch( bt )
 683 
 684   Node *span = gvn->transform(new MulINode(trip_count,stride));
 685   set_subtree_ctrl( span );
 686   hook->init_req(5, span);
 687 
 688   limit = gvn->transform(new AddINode(span,init_trip));
 689   set_subtree_ctrl( limit );
 690 
 691   } // LoopLimitCheck
 692 
 693   if (!UseCountedLoopSafepoints) {
 694     // Check for SafePoint on backedge and remove
 695     Node *sfpt = x->in(LoopNode::LoopBackControl);
 696     if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) {
 697       lazy_replace( sfpt, iftrue );
 698       if (loop->_safepts != NULL) {
 699         loop->_safepts->yank(sfpt);
 700       }
 701       loop->_tail = iftrue;
 702     }
 703   }
 704 
 705   // Build a canonical trip test.
 706   // Clone code, as old values may be in use.
 707   incr = incr->clone();
 708   incr->set_req(1,phi);
 709   incr->set_req(2,stride);
 710   incr = _igvn.register_new_node_with_optimizer(incr);
 711   set_early_ctrl( incr );
 712   _igvn.rehash_node_delayed(phi);
 713   phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
 714 
 715   // If phi type is more restrictive than Int, raise to
 716   // Int to prevent (almost) infinite recursion in igvn
 717   // which can only handle integer types for constants or minint..maxint.
 718   if (!TypeInt::INT->higher_equal(phi->bottom_type())) {
 719     Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT);
 720     nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
 721     nphi = _igvn.register_new_node_with_optimizer(nphi);
 722     set_ctrl(nphi, get_ctrl(phi));
 723     _igvn.replace_node(phi, nphi);
 724     phi = nphi->as_Phi();
 725   }
 726   cmp = cmp->clone();
 727   cmp->set_req(1,incr);
 728   cmp->set_req(2,limit);
 729   cmp = _igvn.register_new_node_with_optimizer(cmp);
 730   set_ctrl(cmp, iff->in(0));
 731 
 732   test = test->clone()->as_Bool();
 733   (*(BoolTest*)&test->_test)._test = bt;
 734   test->set_req(1,cmp);
 735   _igvn.register_new_node_with_optimizer(test);
 736   set_ctrl(test, iff->in(0));
 737 
 738   // Replace the old IfNode with a new LoopEndNode
 739   Node *lex = _igvn.register_new_node_with_optimizer(new CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt ));
 740   IfNode *le = lex->as_If();
 741   uint dd = dom_depth(iff);
 742   set_idom(le, le->in(0), dd); // Update dominance for loop exit
 743   set_loop(le, loop);
 744 
 745   // Get the loop-exit control
 746   Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
 747 
 748   // Need to swap loop-exit and loop-back control?
 749   if (iftrue_op == Op_IfFalse) {
 750     Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le));
 751     Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le));
 752 
 753     loop->_tail = back_control = ift2;
 754     set_loop(ift2, loop);
 755     set_loop(iff2, get_loop(iffalse));
 756 
 757     // Lazy update of 'get_ctrl' mechanism.
 758     lazy_replace(iffalse, iff2);
 759     lazy_replace(iftrue,  ift2);
 760 
 761     // Swap names
 762     iffalse = iff2;
 763     iftrue  = ift2;
 764   } else {
 765     _igvn.rehash_node_delayed(iffalse);
 766     _igvn.rehash_node_delayed(iftrue);
 767     iffalse->set_req_X( 0, le, &_igvn );
 768     iftrue ->set_req_X( 0, le, &_igvn );
 769   }
 770 
 771   set_idom(iftrue,  le, dd+1);
 772   set_idom(iffalse, le, dd+1);
 773   assert(iff->outcnt() == 0, "should be dead now");
 774   lazy_replace( iff, le ); // fix 'get_ctrl'
 775 
 776   // Now setup a new CountedLoopNode to replace the existing LoopNode
 777   CountedLoopNode *l = new CountedLoopNode(init_control, back_control);
 778   l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
 779   // The following assert is approximately true, and defines the intention
 780   // of can_be_counted_loop.  It fails, however, because phase->type
 781   // is not yet initialized for this loop and its parts.
 782   //assert(l->can_be_counted_loop(this), "sanity");
 783   _igvn.register_new_node_with_optimizer(l);
 784   set_loop(l, loop);
 785   loop->_head = l;
 786   // Fix all data nodes placed at the old loop head.
 787   // Uses the lazy-update mechanism of 'get_ctrl'.
 788   lazy_replace( x, l );
 789   set_idom(l, init_control, dom_depth(x));
 790 
 791   if (!UseCountedLoopSafepoints) {
 792     // Check for immediately preceding SafePoint and remove
 793     Node *sfpt2 = le->in(0);
 794     if (sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2)) {
 795       lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control));
 796       if (loop->_safepts != NULL) {
 797         loop->_safepts->yank(sfpt2);
 798       }
 799     }
 800   }
 801 
 802   // Free up intermediate goo
 803   _igvn.remove_dead_node(hook);
 804 
 805 #ifdef ASSERT
 806   assert(l->is_valid_counted_loop(), "counted loop shape is messed up");
 807   assert(l == loop->_head && l->phi() == phi && l->loopexit() == lex, "" );
 808 #endif
 809 #ifndef PRODUCT
 810   if (TraceLoopOpts) {
 811     tty->print("Counted      ");
 812     loop->dump_head();
 813   }
 814 #endif
 815 
 816   C->print_method(PHASE_AFTER_CLOOPS, 3);
 817 
 818   // Capture bounds of the loop in the induction variable Phi before
 819   // subsequent transformation (iteration splitting) obscures the
 820   // bounds
 821   l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn));
 822 
 823   return true;
 824 }
 825 
 826 //----------------------exact_limit-------------------------------------------
 827 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {
 828   assert(loop->_head->is_CountedLoop(), "");
 829   CountedLoopNode *cl = loop->_head->as_CountedLoop();
 830   assert(cl->is_valid_counted_loop(), "");
 831 
 832   if (!LoopLimitCheck || ABS(cl->stride_con()) == 1 ||
 833       cl->limit()->Opcode() == Op_LoopLimit) {
 834     // Old code has exact limit (it could be incorrect in case of int overflow).
 835     // Loop limit is exact with stride == 1. And loop may already have exact limit.
 836     return cl->limit();
 837   }
 838   Node *limit = NULL;
 839 #ifdef ASSERT
 840   BoolTest::mask bt = cl->loopexit()->test_trip();
 841   assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
 842 #endif
 843   if (cl->has_exact_trip_count()) {
 844     // Simple case: loop has constant boundaries.
 845     // Use jlongs to avoid integer overflow.
 846     int stride_con = cl->stride_con();
 847     jlong  init_con = cl->init_trip()->get_int();
 848     jlong limit_con = cl->limit()->get_int();
 849     julong trip_cnt = cl->trip_count();
 850     jlong final_con = init_con + trip_cnt*stride_con;
 851     int final_int = (int)final_con;
 852     // The final value should be in integer range since the loop
 853     // is counted and the limit was checked for overflow.
 854     assert(final_con == (jlong)final_int, "final value should be integer");
 855     limit = _igvn.intcon(final_int);
 856   } else {
 857     // Create new LoopLimit node to get exact limit (final iv value).
 858     limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
 859     register_new_node(limit, cl->in(LoopNode::EntryControl));
 860   }
 861   assert(limit != NULL, "sanity");
 862   return limit;
 863 }
 864 
 865 //------------------------------Ideal------------------------------------------
 866 // Return a node which is more "ideal" than the current node.
 867 // Attempt to convert into a counted-loop.
 868 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
 869   return RegionNode::Ideal(phase, can_reshape);
 870 }
 871 
 872 
 873 //=============================================================================
 874 //------------------------------Ideal------------------------------------------
 875 // Return a node which is more "ideal" than the current node.
 876 // Attempt to convert into a counted-loop.
 877 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {
 878   return RegionNode::Ideal(phase, can_reshape);
 879 }
 880 
 881 //------------------------------dump_spec--------------------------------------
 882 // Dump special per-node info
 883 #ifndef PRODUCT
 884 void CountedLoopNode::dump_spec(outputStream *st) const {
 885   LoopNode::dump_spec(st);
 886   if (stride_is_con()) {
 887     st->print("stride: %d ",stride_con());
 888   }
 889   if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);
 890   if (is_main_loop()) st->print("main of N%d", _idx);
 891   if (is_post_loop()) st->print("post of N%d", _main_idx);
 892 }
 893 #endif
 894 
 895 //=============================================================================
 896 int CountedLoopEndNode::stride_con() const {
 897   return stride()->bottom_type()->is_int()->get_con();
 898 }
 899 
 900 //=============================================================================
 901 //------------------------------Value-----------------------------------------
 902 const Type* LoopLimitNode::Value(PhaseGVN* phase) const {
 903   const Type* init_t   = phase->type(in(Init));
 904   const Type* limit_t  = phase->type(in(Limit));
 905   const Type* stride_t = phase->type(in(Stride));
 906   // Either input is TOP ==> the result is TOP
 907   if (init_t   == Type::TOP) return Type::TOP;
 908   if (limit_t  == Type::TOP) return Type::TOP;
 909   if (stride_t == Type::TOP) return Type::TOP;
 910 
 911   int stride_con = stride_t->is_int()->get_con();
 912   if (stride_con == 1)
 913     return NULL;  // Identity
 914 
 915   if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {
 916     // Use jlongs to avoid integer overflow.
 917     jlong init_con   =  init_t->is_int()->get_con();
 918     jlong limit_con  = limit_t->is_int()->get_con();
 919     int  stride_m   = stride_con - (stride_con > 0 ? 1 : -1);
 920     jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
 921     jlong final_con  = init_con + stride_con*trip_count;
 922     int final_int = (int)final_con;
 923     // The final value should be in integer range since the loop
 924     // is counted and the limit was checked for overflow.
 925     assert(final_con == (jlong)final_int, "final value should be integer");
 926     return TypeInt::make(final_int);
 927   }
 928 
 929   return bottom_type(); // TypeInt::INT
 930 }
 931 
 932 //------------------------------Ideal------------------------------------------
 933 // Return a node which is more "ideal" than the current node.
 934 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {
 935   if (phase->type(in(Init))   == Type::TOP ||
 936       phase->type(in(Limit))  == Type::TOP ||
 937       phase->type(in(Stride)) == Type::TOP)
 938     return NULL;  // Dead
 939 
 940   int stride_con = phase->type(in(Stride))->is_int()->get_con();
 941   if (stride_con == 1)
 942     return NULL;  // Identity
 943 
 944   if (in(Init)->is_Con() && in(Limit)->is_Con())
 945     return NULL;  // Value
 946 
 947   // Delay following optimizations until all loop optimizations
 948   // done to keep Ideal graph simple.
 949   if (!can_reshape || phase->C->major_progress())
 950     return NULL;
 951 
 952   const TypeInt* init_t  = phase->type(in(Init) )->is_int();
 953   const TypeInt* limit_t = phase->type(in(Limit))->is_int();
 954   int stride_p;
 955   jlong lim, ini;
 956   julong max;
 957   if (stride_con > 0) {
 958     stride_p = stride_con;
 959     lim = limit_t->_hi;
 960     ini = init_t->_lo;
 961     max = (julong)max_jint;
 962   } else {
 963     stride_p = -stride_con;
 964     lim = init_t->_hi;
 965     ini = limit_t->_lo;
 966     max = (julong)min_jint;
 967   }
 968   julong range = lim - ini + stride_p;
 969   if (range <= max) {
 970     // Convert to integer expression if it is not overflow.
 971     Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
 972     Node *range = phase->transform(new SubINode(in(Limit), in(Init)));
 973     Node *bias  = phase->transform(new AddINode(range, stride_m));
 974     Node *trip  = phase->transform(new DivINode(0, bias, in(Stride)));
 975     Node *span  = phase->transform(new MulINode(trip, in(Stride)));
 976     return new AddINode(span, in(Init)); // exact limit
 977   }
 978 
 979   if (is_power_of_2(stride_p) ||                // divisor is 2^n
 980       !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?
 981     // Convert to long expression to avoid integer overflow
 982     // and let igvn optimizer convert this division.
 983     //
 984     Node*   init   = phase->transform( new ConvI2LNode(in(Init)));
 985     Node*  limit   = phase->transform( new ConvI2LNode(in(Limit)));
 986     Node* stride   = phase->longcon(stride_con);
 987     Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
 988 
 989     Node *range = phase->transform(new SubLNode(limit, init));
 990     Node *bias  = phase->transform(new AddLNode(range, stride_m));
 991     Node *span;
 992     if (stride_con > 0 && is_power_of_2(stride_p)) {
 993       // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
 994       // and avoid generating rounding for division. Zero trip guard should
 995       // guarantee that init < limit but sometimes the guard is missing and
 996       // we can get situation when init > limit. Note, for the empty loop
 997       // optimization zero trip guard is generated explicitly which leaves
 998       // only RCE predicate where exact limit is used and the predicate
 999       // will simply fail forcing recompilation.
1000       Node* neg_stride   = phase->longcon(-stride_con);
1001       span = phase->transform(new AndLNode(bias, neg_stride));
1002     } else {
1003       Node *trip  = phase->transform(new DivLNode(0, bias, stride));
1004       span = phase->transform(new MulLNode(trip, stride));
1005     }
1006     // Convert back to int
1007     Node *span_int = phase->transform(new ConvL2INode(span));
1008     return new AddINode(span_int, in(Init)); // exact limit
1009   }
1010 
1011   return NULL;    // No progress
1012 }
1013 
1014 //------------------------------Identity---------------------------------------
1015 // If stride == 1 return limit node.
1016 Node* LoopLimitNode::Identity(PhaseGVN* phase) {
1017   int stride_con = phase->type(in(Stride))->is_int()->get_con();
1018   if (stride_con == 1 || stride_con == -1)
1019     return in(Limit);
1020   return this;
1021 }
1022 
1023 //=============================================================================
1024 //----------------------match_incr_with_optional_truncation--------------------
1025 // Match increment with optional truncation:
1026 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
1027 // Return NULL for failure. Success returns the increment node.
1028 Node* CountedLoopNode::match_incr_with_optional_truncation(
1029                       Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) {
1030   // Quick cutouts:
1031   if (expr == NULL || expr->req() != 3)  return NULL;
1032 
1033   Node *t1 = NULL;
1034   Node *t2 = NULL;
1035   const TypeInt* trunc_t = TypeInt::INT;
1036   Node* n1 = expr;
1037   int   n1op = n1->Opcode();
1038 
1039   // Try to strip (n1 & M) or (n1 << N >> N) from n1.
1040   if (n1op == Op_AndI &&
1041       n1->in(2)->is_Con() &&
1042       n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {
1043     // %%% This check should match any mask of 2**K-1.
1044     t1 = n1;
1045     n1 = t1->in(1);
1046     n1op = n1->Opcode();
1047     trunc_t = TypeInt::CHAR;
1048   } else if (n1op == Op_RShiftI &&
1049              n1->in(1) != NULL &&
1050              n1->in(1)->Opcode() == Op_LShiftI &&
1051              n1->in(2) == n1->in(1)->in(2) &&
1052              n1->in(2)->is_Con()) {
1053     jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
1054     // %%% This check should match any shift in [1..31].
1055     if (shift == 16 || shift == 8) {
1056       t1 = n1;
1057       t2 = t1->in(1);
1058       n1 = t2->in(1);
1059       n1op = n1->Opcode();
1060       if (shift == 16) {
1061         trunc_t = TypeInt::SHORT;
1062       } else if (shift == 8) {
1063         trunc_t = TypeInt::BYTE;
1064       }
1065     }
1066   }
1067 
1068   // If (maybe after stripping) it is an AddI, we won:
1069   if (n1op == Op_AddI) {
1070     *trunc1 = t1;
1071     *trunc2 = t2;
1072     *trunc_type = trunc_t;
1073     return n1;
1074   }
1075 
1076   // failed
1077   return NULL;
1078 }
1079 
1080 
1081 //------------------------------filtered_type--------------------------------
1082 // Return a type based on condition control flow
1083 // A successful return will be a type that is restricted due
1084 // to a series of dominating if-tests, such as:
1085 //    if (i < 10) {
1086 //       if (i > 0) {
1087 //          here: "i" type is [1..10)
1088 //       }
1089 //    }
1090 // or a control flow merge
1091 //    if (i < 10) {
1092 //       do {
1093 //          phi( , ) -- at top of loop type is [min_int..10)
1094 //         i = ?
1095 //       } while ( i < 10)
1096 //
1097 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {
1098   assert(n && n->bottom_type()->is_int(), "must be int");
1099   const TypeInt* filtered_t = NULL;
1100   if (!n->is_Phi()) {
1101     assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control");
1102     filtered_t = filtered_type_from_dominators(n, n_ctrl);
1103 
1104   } else {
1105     Node* phi    = n->as_Phi();
1106     Node* region = phi->in(0);
1107     assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region");
1108     if (region && region != C->top()) {
1109       for (uint i = 1; i < phi->req(); i++) {
1110         Node* val   = phi->in(i);
1111         Node* use_c = region->in(i);
1112         const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
1113         if (val_t != NULL) {
1114           if (filtered_t == NULL) {
1115             filtered_t = val_t;
1116           } else {
1117             filtered_t = filtered_t->meet(val_t)->is_int();
1118           }
1119         }
1120       }
1121     }
1122   }
1123   const TypeInt* n_t = _igvn.type(n)->is_int();
1124   if (filtered_t != NULL) {
1125     n_t = n_t->join(filtered_t)->is_int();
1126   }
1127   return n_t;
1128 }
1129 
1130 
1131 //------------------------------filtered_type_from_dominators--------------------------------
1132 // Return a possibly more restrictive type for val based on condition control flow of dominators
1133 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {
1134   if (val->is_Con()) {
1135      return val->bottom_type()->is_int();
1136   }
1137   uint if_limit = 10; // Max number of dominating if's visited
1138   const TypeInt* rtn_t = NULL;
1139 
1140   if (use_ctrl && use_ctrl != C->top()) {
1141     Node* val_ctrl = get_ctrl(val);
1142     uint val_dom_depth = dom_depth(val_ctrl);
1143     Node* pred = use_ctrl;
1144     uint if_cnt = 0;
1145     while (if_cnt < if_limit) {
1146       if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {
1147         if_cnt++;
1148         const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
1149         if (if_t != NULL) {
1150           if (rtn_t == NULL) {
1151             rtn_t = if_t;
1152           } else {
1153             rtn_t = rtn_t->join(if_t)->is_int();
1154           }
1155         }
1156       }
1157       pred = idom(pred);
1158       if (pred == NULL || pred == C->top()) {
1159         break;
1160       }
1161       // Stop if going beyond definition block of val
1162       if (dom_depth(pred) < val_dom_depth) {
1163         break;
1164       }
1165     }
1166   }
1167   return rtn_t;
1168 }
1169 
1170 
1171 //------------------------------dump_spec--------------------------------------
1172 // Dump special per-node info
1173 #ifndef PRODUCT
1174 void CountedLoopEndNode::dump_spec(outputStream *st) const {
1175   if( in(TestValue) != NULL && in(TestValue)->is_Bool() ) {
1176     BoolTest bt( test_trip()); // Added this for g++.
1177 
1178     st->print("[");
1179     bt.dump_on(st);
1180     st->print("]");
1181   }
1182   st->print(" ");
1183   IfNode::dump_spec(st);
1184 }
1185 #endif
1186 
1187 //=============================================================================
1188 //------------------------------is_member--------------------------------------
1189 // Is 'l' a member of 'this'?
1190 bool IdealLoopTree::is_member(const IdealLoopTree *l) const {
1191   while( l->_nest > _nest ) l = l->_parent;
1192   return l == this;
1193 }
1194 
1195 //------------------------------set_nest---------------------------------------
1196 // Set loop tree nesting depth.  Accumulate _has_call bits.
1197 int IdealLoopTree::set_nest( uint depth ) {
1198   _nest = depth;
1199   int bits = _has_call;
1200   if( _child ) bits |= _child->set_nest(depth+1);
1201   if( bits ) _has_call = 1;
1202   if( _next  ) bits |= _next ->set_nest(depth  );
1203   return bits;
1204 }
1205 
1206 //------------------------------split_fall_in----------------------------------
1207 // Split out multiple fall-in edges from the loop header.  Move them to a
1208 // private RegionNode before the loop.  This becomes the loop landing pad.
1209 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {
1210   PhaseIterGVN &igvn = phase->_igvn;
1211   uint i;
1212 
1213   // Make a new RegionNode to be the landing pad.
1214   Node *landing_pad = new RegionNode( fall_in_cnt+1 );
1215   phase->set_loop(landing_pad,_parent);
1216   // Gather all the fall-in control paths into the landing pad
1217   uint icnt = fall_in_cnt;
1218   uint oreq = _head->req();
1219   for( i = oreq-1; i>0; i-- )
1220     if( !phase->is_member( this, _head->in(i) ) )
1221       landing_pad->set_req(icnt--,_head->in(i));
1222 
1223   // Peel off PhiNode edges as well
1224   for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1225     Node *oj = _head->fast_out(j);
1226     if( oj->is_Phi() ) {
1227       PhiNode* old_phi = oj->as_Phi();
1228       assert( old_phi->region() == _head, "" );
1229       igvn.hash_delete(old_phi);   // Yank from hash before hacking edges
1230       Node *p = PhiNode::make_blank(landing_pad, old_phi);
1231       uint icnt = fall_in_cnt;
1232       for( i = oreq-1; i>0; i-- ) {
1233         if( !phase->is_member( this, _head->in(i) ) ) {
1234           p->init_req(icnt--, old_phi->in(i));
1235           // Go ahead and clean out old edges from old phi
1236           old_phi->del_req(i);
1237         }
1238       }
1239       // Search for CSE's here, because ZKM.jar does a lot of
1240       // loop hackery and we need to be a little incremental
1241       // with the CSE to avoid O(N^2) node blow-up.
1242       Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
1243       if( p2 ) {                // Found CSE
1244         p->destruct();          // Recover useless new node
1245         p = p2;                 // Use old node
1246       } else {
1247         igvn.register_new_node_with_optimizer(p, old_phi);
1248       }
1249       // Make old Phi refer to new Phi.
1250       old_phi->add_req(p);
1251       // Check for the special case of making the old phi useless and
1252       // disappear it.  In JavaGrande I have a case where this useless
1253       // Phi is the loop limit and prevents recognizing a CountedLoop
1254       // which in turn prevents removing an empty loop.
1255       Node *id_old_phi = old_phi->Identity( &igvn );
1256       if( id_old_phi != old_phi ) { // Found a simple identity?
1257         // Note that I cannot call 'replace_node' here, because
1258         // that will yank the edge from old_phi to the Region and
1259         // I'm mid-iteration over the Region's uses.
1260         for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {
1261           Node* use = old_phi->last_out(i);
1262           igvn.rehash_node_delayed(use);
1263           uint uses_found = 0;
1264           for (uint j = 0; j < use->len(); j++) {
1265             if (use->in(j) == old_phi) {
1266               if (j < use->req()) use->set_req (j, id_old_phi);
1267               else                use->set_prec(j, id_old_phi);
1268               uses_found++;
1269             }
1270           }
1271           i -= uses_found;    // we deleted 1 or more copies of this edge
1272         }
1273       }
1274       igvn._worklist.push(old_phi);
1275     }
1276   }
1277   // Finally clean out the fall-in edges from the RegionNode
1278   for( i = oreq-1; i>0; i-- ) {
1279     if( !phase->is_member( this, _head->in(i) ) ) {
1280       _head->del_req(i);
1281     }
1282   }
1283   igvn.rehash_node_delayed(_head);
1284   // Transform landing pad
1285   igvn.register_new_node_with_optimizer(landing_pad, _head);
1286   // Insert landing pad into the header
1287   _head->add_req(landing_pad);
1288 }
1289 
1290 //------------------------------split_outer_loop-------------------------------
1291 // Split out the outermost loop from this shared header.
1292 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {
1293   PhaseIterGVN &igvn = phase->_igvn;
1294 
1295   // Find index of outermost loop; it should also be my tail.
1296   uint outer_idx = 1;
1297   while( _head->in(outer_idx) != _tail ) outer_idx++;
1298 
1299   // Make a LoopNode for the outermost loop.
1300   Node *ctl = _head->in(LoopNode::EntryControl);
1301   Node *outer = new LoopNode( ctl, _head->in(outer_idx) );
1302   outer = igvn.register_new_node_with_optimizer(outer, _head);
1303   phase->set_created_loop_node();
1304 
1305   // Outermost loop falls into '_head' loop
1306   _head->set_req(LoopNode::EntryControl, outer);
1307   _head->del_req(outer_idx);
1308   // Split all the Phis up between '_head' loop and 'outer' loop.
1309   for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1310     Node *out = _head->fast_out(j);
1311     if( out->is_Phi() ) {
1312       PhiNode *old_phi = out->as_Phi();
1313       assert( old_phi->region() == _head, "" );
1314       Node *phi = PhiNode::make_blank(outer, old_phi);
1315       phi->init_req(LoopNode::EntryControl,    old_phi->in(LoopNode::EntryControl));
1316       phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
1317       phi = igvn.register_new_node_with_optimizer(phi, old_phi);
1318       // Make old Phi point to new Phi on the fall-in path
1319       igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
1320       old_phi->del_req(outer_idx);
1321     }
1322   }
1323 
1324   // Use the new loop head instead of the old shared one
1325   _head = outer;
1326   phase->set_loop(_head, this);
1327 }
1328 
1329 //------------------------------fix_parent-------------------------------------
1330 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {
1331   loop->_parent = parent;
1332   if( loop->_child ) fix_parent( loop->_child, loop   );
1333   if( loop->_next  ) fix_parent( loop->_next , parent );
1334 }
1335 
1336 //------------------------------estimate_path_freq-----------------------------
1337 static float estimate_path_freq( Node *n ) {
1338   // Try to extract some path frequency info
1339   IfNode *iff;
1340   for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests
1341     uint nop = n->Opcode();
1342     if( nop == Op_SafePoint ) {   // Skip any safepoint
1343       n = n->in(0);
1344       continue;
1345     }
1346     if( nop == Op_CatchProj ) {   // Get count from a prior call
1347       // Assume call does not always throw exceptions: means the call-site
1348       // count is also the frequency of the fall-through path.
1349       assert( n->is_CatchProj(), "" );
1350       if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
1351         return 0.0f;            // Assume call exception path is rare
1352       Node *call = n->in(0)->in(0)->in(0);
1353       assert( call->is_Call(), "expect a call here" );
1354       const JVMState *jvms = ((CallNode*)call)->jvms();
1355       ciMethodData* methodData = jvms->method()->method_data();
1356       if (!methodData->is_mature())  return 0.0f; // No call-site data
1357       ciProfileData* data = methodData->bci_to_data(jvms->bci());
1358       if ((data == NULL) || !data->is_CounterData()) {
1359         // no call profile available, try call's control input
1360         n = n->in(0);
1361         continue;
1362       }
1363       return data->as_CounterData()->count()/FreqCountInvocations;
1364     }
1365     // See if there's a gating IF test
1366     Node *n_c = n->in(0);
1367     if( !n_c->is_If() ) break;       // No estimate available
1368     iff = n_c->as_If();
1369     if( iff->_fcnt != COUNT_UNKNOWN )   // Have a valid count?
1370       // Compute how much count comes on this path
1371       return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
1372     // Have no count info.  Skip dull uncommon-trap like branches.
1373     if( (nop == Op_IfTrue  && iff->_prob < PROB_LIKELY_MAG(5)) ||
1374         (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
1375       break;
1376     // Skip through never-taken branch; look for a real loop exit.
1377     n = iff->in(0);
1378   }
1379   return 0.0f;                  // No estimate available
1380 }
1381 
1382 //------------------------------merge_many_backedges---------------------------
1383 // Merge all the backedges from the shared header into a private Region.
1384 // Feed that region as the one backedge to this loop.
1385 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {
1386   uint i;
1387 
1388   // Scan for the top 2 hottest backedges
1389   float hotcnt = 0.0f;
1390   float warmcnt = 0.0f;
1391   uint hot_idx = 0;
1392   // Loop starts at 2 because slot 1 is the fall-in path
1393   for( i = 2; i < _head->req(); i++ ) {
1394     float cnt = estimate_path_freq(_head->in(i));
1395     if( cnt > hotcnt ) {       // Grab hottest path
1396       warmcnt = hotcnt;
1397       hotcnt = cnt;
1398       hot_idx = i;
1399     } else if( cnt > warmcnt ) { // And 2nd hottest path
1400       warmcnt = cnt;
1401     }
1402   }
1403 
1404   // See if the hottest backedge is worthy of being an inner loop
1405   // by being much hotter than the next hottest backedge.
1406   if( hotcnt <= 0.0001 ||
1407       hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
1408 
1409   // Peel out the backedges into a private merge point; peel
1410   // them all except optionally hot_idx.
1411   PhaseIterGVN &igvn = phase->_igvn;
1412 
1413   Node *hot_tail = NULL;
1414   // Make a Region for the merge point
1415   Node *r = new RegionNode(1);
1416   for( i = 2; i < _head->req(); i++ ) {
1417     if( i != hot_idx )
1418       r->add_req( _head->in(i) );
1419     else hot_tail = _head->in(i);
1420   }
1421   igvn.register_new_node_with_optimizer(r, _head);
1422   // Plug region into end of loop _head, followed by hot_tail
1423   while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
1424   igvn.replace_input_of(_head, 2, r);
1425   if( hot_idx ) _head->add_req(hot_tail);
1426 
1427   // Split all the Phis up between '_head' loop and the Region 'r'
1428   for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {
1429     Node *out = _head->fast_out(j);
1430     if( out->is_Phi() ) {
1431       PhiNode* n = out->as_Phi();
1432       igvn.hash_delete(n);      // Delete from hash before hacking edges
1433       Node *hot_phi = NULL;
1434       Node *phi = new PhiNode(r, n->type(), n->adr_type());
1435       // Check all inputs for the ones to peel out
1436       uint j = 1;
1437       for( uint i = 2; i < n->req(); i++ ) {
1438         if( i != hot_idx )
1439           phi->set_req( j++, n->in(i) );
1440         else hot_phi = n->in(i);
1441       }
1442       // Register the phi but do not transform until whole place transforms
1443       igvn.register_new_node_with_optimizer(phi, n);
1444       // Add the merge phi to the old Phi
1445       while( n->req() > 3 ) n->del_req( n->req()-1 );
1446       igvn.replace_input_of(n, 2, phi);
1447       if( hot_idx ) n->add_req(hot_phi);
1448     }
1449   }
1450 
1451 
1452   // Insert a new IdealLoopTree inserted below me.  Turn it into a clone
1453   // of self loop tree.  Turn self into a loop headed by _head and with
1454   // tail being the new merge point.
1455   IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
1456   phase->set_loop(_tail,ilt);   // Adjust tail
1457   _tail = r;                    // Self's tail is new merge point
1458   phase->set_loop(r,this);
1459   ilt->_child = _child;         // New guy has my children
1460   _child = ilt;                 // Self has new guy as only child
1461   ilt->_parent = this;          // new guy has self for parent
1462   ilt->_nest = _nest;           // Same nesting depth (for now)
1463 
1464   // Starting with 'ilt', look for child loop trees using the same shared
1465   // header.  Flatten these out; they will no longer be loops in the end.
1466   IdealLoopTree **pilt = &_child;
1467   while( ilt ) {
1468     if( ilt->_head == _head ) {
1469       uint i;
1470       for( i = 2; i < _head->req(); i++ )
1471         if( _head->in(i) == ilt->_tail )
1472           break;                // Still a loop
1473       if( i == _head->req() ) { // No longer a loop
1474         // Flatten ilt.  Hang ilt's "_next" list from the end of
1475         // ilt's '_child' list.  Move the ilt's _child up to replace ilt.
1476         IdealLoopTree **cp = &ilt->_child;
1477         while( *cp ) cp = &(*cp)->_next;   // Find end of child list
1478         *cp = ilt->_next;       // Hang next list at end of child list
1479         *pilt = ilt->_child;    // Move child up to replace ilt
1480         ilt->_head = NULL;      // Flag as a loop UNIONED into parent
1481         ilt = ilt->_child;      // Repeat using new ilt
1482         continue;               // do not advance over ilt->_child
1483       }
1484       assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
1485       phase->set_loop(_head,ilt);
1486     }
1487     pilt = &ilt->_child;        // Advance to next
1488     ilt = *pilt;
1489   }
1490 
1491   if( _child ) fix_parent( _child, this );
1492 }
1493 
1494 //------------------------------beautify_loops---------------------------------
1495 // Split shared headers and insert loop landing pads.
1496 // Insert a LoopNode to replace the RegionNode.
1497 // Return TRUE if loop tree is structurally changed.
1498 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {
1499   bool result = false;
1500   // Cache parts in locals for easy
1501   PhaseIterGVN &igvn = phase->_igvn;
1502 
1503   igvn.hash_delete(_head);      // Yank from hash before hacking edges
1504 
1505   // Check for multiple fall-in paths.  Peel off a landing pad if need be.
1506   int fall_in_cnt = 0;
1507   for( uint i = 1; i < _head->req(); i++ )
1508     if( !phase->is_member( this, _head->in(i) ) )
1509       fall_in_cnt++;
1510   assert( fall_in_cnt, "at least 1 fall-in path" );
1511   if( fall_in_cnt > 1 )         // Need a loop landing pad to merge fall-ins
1512     split_fall_in( phase, fall_in_cnt );
1513 
1514   // Swap inputs to the _head and all Phis to move the fall-in edge to
1515   // the left.
1516   fall_in_cnt = 1;
1517   while( phase->is_member( this, _head->in(fall_in_cnt) ) )
1518     fall_in_cnt++;
1519   if( fall_in_cnt > 1 ) {
1520     // Since I am just swapping inputs I do not need to update def-use info
1521     Node *tmp = _head->in(1);
1522     igvn.rehash_node_delayed(_head);
1523     _head->set_req( 1, _head->in(fall_in_cnt) );
1524     _head->set_req( fall_in_cnt, tmp );
1525     // Swap also all Phis
1526     for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {
1527       Node* phi = _head->fast_out(i);
1528       if( phi->is_Phi() ) {
1529         igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges
1530         tmp = phi->in(1);
1531         phi->set_req( 1, phi->in(fall_in_cnt) );
1532         phi->set_req( fall_in_cnt, tmp );
1533       }
1534     }
1535   }
1536   assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
1537   assert(  phase->is_member( this, _head->in(2) ), "right edge is loop" );
1538 
1539   // If I am a shared header (multiple backedges), peel off the many
1540   // backedges into a private merge point and use the merge point as
1541   // the one true backedge.
1542   if( _head->req() > 3 ) {
1543     // Merge the many backedges into a single backedge but leave
1544     // the hottest backedge as separate edge for the following peel.
1545     merge_many_backedges( phase );
1546     result = true;
1547   }
1548 
1549   // If I have one hot backedge, peel off myself loop.
1550   // I better be the outermost loop.
1551   if (_head->req() > 3 && !_irreducible) {
1552     split_outer_loop( phase );
1553     result = true;
1554 
1555   } else if (!_head->is_Loop() && !_irreducible) {
1556     // Make a new LoopNode to replace the old loop head
1557     Node *l = new LoopNode( _head->in(1), _head->in(2) );
1558     l = igvn.register_new_node_with_optimizer(l, _head);
1559     phase->set_created_loop_node();
1560     // Go ahead and replace _head
1561     phase->_igvn.replace_node( _head, l );
1562     _head = l;
1563     phase->set_loop(_head, this);
1564   }
1565 
1566   // Now recursively beautify nested loops
1567   if( _child ) result |= _child->beautify_loops( phase );
1568   if( _next  ) result |= _next ->beautify_loops( phase );
1569   return result;
1570 }
1571 
1572 //------------------------------allpaths_check_safepts----------------------------
1573 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
1574 // encountered.  Helper for check_safepts.
1575 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {
1576   assert(stack.size() == 0, "empty stack");
1577   stack.push(_tail);
1578   visited.Clear();
1579   visited.set(_tail->_idx);
1580   while (stack.size() > 0) {
1581     Node* n = stack.pop();
1582     if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1583       // Terminate this path
1584     } else if (n->Opcode() == Op_SafePoint) {
1585       if (_phase->get_loop(n) != this) {
1586         if (_required_safept == NULL) _required_safept = new Node_List();
1587         _required_safept->push(n);  // save the one closest to the tail
1588       }
1589       // Terminate this path
1590     } else {
1591       uint start = n->is_Region() ? 1 : 0;
1592       uint end   = n->is_Region() && !n->is_Loop() ? n->req() : start + 1;
1593       for (uint i = start; i < end; i++) {
1594         Node* in = n->in(i);
1595         assert(in->is_CFG(), "must be");
1596         if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {
1597           stack.push(in);
1598         }
1599       }
1600     }
1601   }
1602 }
1603 
1604 //------------------------------check_safepts----------------------------
1605 // Given dominators, try to find loops with calls that must always be
1606 // executed (call dominates loop tail).  These loops do not need non-call
1607 // safepoints (ncsfpt).
1608 //
1609 // A complication is that a safepoint in a inner loop may be needed
1610 // by an outer loop. In the following, the inner loop sees it has a
1611 // call (block 3) on every path from the head (block 2) to the
1612 // backedge (arc 3->2).  So it deletes the ncsfpt (non-call safepoint)
1613 // in block 2, _but_ this leaves the outer loop without a safepoint.
1614 //
1615 //          entry  0
1616 //                 |
1617 //                 v
1618 // outer 1,2    +->1
1619 //              |  |
1620 //              |  v
1621 //              |  2<---+  ncsfpt in 2
1622 //              |_/|\   |
1623 //                 | v  |
1624 // inner 2,3      /  3  |  call in 3
1625 //               /   |  |
1626 //              v    +--+
1627 //        exit  4
1628 //
1629 //
1630 // This method creates a list (_required_safept) of ncsfpt nodes that must
1631 // be protected is created for each loop. When a ncsfpt maybe deleted, it
1632 // is first looked for in the lists for the outer loops of the current loop.
1633 //
1634 // The insights into the problem:
1635 //  A) counted loops are okay
1636 //  B) innermost loops are okay (only an inner loop can delete
1637 //     a ncsfpt needed by an outer loop)
1638 //  C) a loop is immune from an inner loop deleting a safepoint
1639 //     if the loop has a call on the idom-path
1640 //  D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
1641 //     idom-path that is not in a nested loop
1642 //  E) otherwise, an ncsfpt on the idom-path that is nested in an inner
1643 //     loop needs to be prevented from deletion by an inner loop
1644 //
1645 // There are two analyses:
1646 //  1) The first, and cheaper one, scans the loop body from
1647 //     tail to head following the idom (immediate dominator)
1648 //     chain, looking for the cases (C,D,E) above.
1649 //     Since inner loops are scanned before outer loops, there is summary
1650 //     information about inner loops.  Inner loops can be skipped over
1651 //     when the tail of an inner loop is encountered.
1652 //
1653 //  2) The second, invoked if the first fails to find a call or ncsfpt on
1654 //     the idom path (which is rare), scans all predecessor control paths
1655 //     from the tail to the head, terminating a path when a call or sfpt
1656 //     is encountered, to find the ncsfpt's that are closest to the tail.
1657 //
1658 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {
1659   // Bottom up traversal
1660   IdealLoopTree* ch = _child;
1661   if (_child) _child->check_safepts(visited, stack);
1662   if (_next)  _next ->check_safepts(visited, stack);
1663 
1664   if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) {
1665     bool  has_call         = false; // call on dom-path
1666     bool  has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
1667     Node* nonlocal_ncsfpt  = NULL;  // ncsfpt on dom-path at a deeper depth
1668     // Scan the dom-path nodes from tail to head
1669     for (Node* n = tail(); n != _head; n = _phase->idom(n)) {
1670       if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {
1671         has_call = true;
1672         _has_sfpt = 1;          // Then no need for a safept!
1673         break;
1674       } else if (n->Opcode() == Op_SafePoint) {
1675         if (_phase->get_loop(n) == this) {
1676           has_local_ncsfpt = true;
1677           break;
1678         }
1679         if (nonlocal_ncsfpt == NULL) {
1680           nonlocal_ncsfpt = n; // save the one closest to the tail
1681         }
1682       } else {
1683         IdealLoopTree* nlpt = _phase->get_loop(n);
1684         if (this != nlpt) {
1685           // If at an inner loop tail, see if the inner loop has already
1686           // recorded seeing a call on the dom-path (and stop.)  If not,
1687           // jump to the head of the inner loop.
1688           assert(is_member(nlpt), "nested loop");
1689           Node* tail = nlpt->_tail;
1690           if (tail->in(0)->is_If()) tail = tail->in(0);
1691           if (n == tail) {
1692             // If inner loop has call on dom-path, so does outer loop
1693             if (nlpt->_has_sfpt) {
1694               has_call = true;
1695               _has_sfpt = 1;
1696               break;
1697             }
1698             // Skip to head of inner loop
1699             assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
1700             n = nlpt->_head;
1701           }
1702         }
1703       }
1704     }
1705     // Record safept's that this loop needs preserved when an
1706     // inner loop attempts to delete it's safepoints.
1707     if (_child != NULL && !has_call && !has_local_ncsfpt) {
1708       if (nonlocal_ncsfpt != NULL) {
1709         if (_required_safept == NULL) _required_safept = new Node_List();
1710         _required_safept->push(nonlocal_ncsfpt);
1711       } else {
1712         // Failed to find a suitable safept on the dom-path.  Now use
1713         // an all paths walk from tail to head, looking for safepoints to preserve.
1714         allpaths_check_safepts(visited, stack);
1715       }
1716     }
1717   }
1718 }
1719 
1720 //---------------------------is_deleteable_safept----------------------------
1721 // Is safept not required by an outer loop?
1722 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {
1723   assert(sfpt->Opcode() == Op_SafePoint, "");
1724   IdealLoopTree* lp = get_loop(sfpt)->_parent;
1725   while (lp != NULL) {
1726     Node_List* sfpts = lp->_required_safept;
1727     if (sfpts != NULL) {
1728       for (uint i = 0; i < sfpts->size(); i++) {
1729         if (sfpt == sfpts->at(i))
1730           return false;
1731       }
1732     }
1733     lp = lp->_parent;
1734   }
1735   return true;
1736 }
1737 
1738 //---------------------------replace_parallel_iv-------------------------------
1739 // Replace parallel induction variable (parallel to trip counter)
1740 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {
1741   assert(loop->_head->is_CountedLoop(), "");
1742   CountedLoopNode *cl = loop->_head->as_CountedLoop();
1743   if (!cl->is_valid_counted_loop())
1744     return;         // skip malformed counted loop
1745   Node *incr = cl->incr();
1746   if (incr == NULL)
1747     return;         // Dead loop?
1748   Node *init = cl->init_trip();
1749   Node *phi  = cl->phi();
1750   int stride_con = cl->stride_con();
1751 
1752   // Visit all children, looking for Phis
1753   for (DUIterator i = cl->outs(); cl->has_out(i); i++) {
1754     Node *out = cl->out(i);
1755     // Look for other phis (secondary IVs). Skip dead ones
1756     if (!out->is_Phi() || out == phi || !has_node(out))
1757       continue;
1758     PhiNode* phi2 = out->as_Phi();
1759     Node *incr2 = phi2->in( LoopNode::LoopBackControl );
1760     // Look for induction variables of the form:  X += constant
1761     if (phi2->region() != loop->_head ||
1762         incr2->req() != 3 ||
1763         incr2->in(1) != phi2 ||
1764         incr2 == incr ||
1765         incr2->Opcode() != Op_AddI ||
1766         !incr2->in(2)->is_Con())
1767       continue;
1768 
1769     // Check for parallel induction variable (parallel to trip counter)
1770     // via an affine function.  In particular, count-down loops with
1771     // count-up array indices are common. We only RCE references off
1772     // the trip-counter, so we need to convert all these to trip-counter
1773     // expressions.
1774     Node *init2 = phi2->in( LoopNode::EntryControl );
1775     int stride_con2 = incr2->in(2)->get_int();
1776 
1777     // The general case here gets a little tricky.  We want to find the
1778     // GCD of all possible parallel IV's and make a new IV using this
1779     // GCD for the loop.  Then all possible IVs are simple multiples of
1780     // the GCD.  In practice, this will cover very few extra loops.
1781     // Instead we require 'stride_con2' to be a multiple of 'stride_con',
1782     // where +/-1 is the common case, but other integer multiples are
1783     // also easy to handle.
1784     int ratio_con = stride_con2/stride_con;
1785 
1786     if ((ratio_con * stride_con) == stride_con2) { // Check for exact
1787 #ifndef PRODUCT
1788       if (TraceLoopOpts) {
1789         tty->print("Parallel IV: %d ", phi2->_idx);
1790         loop->dump_head();
1791       }
1792 #endif
1793       // Convert to using the trip counter.  The parallel induction
1794       // variable differs from the trip counter by a loop-invariant
1795       // amount, the difference between their respective initial values.
1796       // It is scaled by the 'ratio_con'.
1797       Node* ratio = _igvn.intcon(ratio_con);
1798       set_ctrl(ratio, C->root());
1799       Node* ratio_init = new MulINode(init, ratio);
1800       _igvn.register_new_node_with_optimizer(ratio_init, init);
1801       set_early_ctrl(ratio_init);
1802       Node* diff = new SubINode(init2, ratio_init);
1803       _igvn.register_new_node_with_optimizer(diff, init2);
1804       set_early_ctrl(diff);
1805       Node* ratio_idx = new MulINode(phi, ratio);
1806       _igvn.register_new_node_with_optimizer(ratio_idx, phi);
1807       set_ctrl(ratio_idx, cl);
1808       Node* add = new AddINode(ratio_idx, diff);
1809       _igvn.register_new_node_with_optimizer(add);
1810       set_ctrl(add, cl);
1811       _igvn.replace_node( phi2, add );
1812       // Sometimes an induction variable is unused
1813       if (add->outcnt() == 0) {
1814         _igvn.remove_dead_node(add);
1815       }
1816       --i; // deleted this phi; rescan starting with next position
1817       continue;
1818     }
1819   }
1820 }
1821 
1822 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {
1823   Node* keep = NULL;
1824   if (keep_one) {
1825     // Look for a safepoint on the idom-path.
1826     for (Node* i = tail(); i != _head; i = phase->idom(i)) {
1827       if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {
1828         keep = i;
1829         break; // Found one
1830       }
1831     }
1832   }
1833 
1834   // Don't remove any safepoints if it is requested to keep a single safepoint and
1835   // no safepoint was found on idom-path. It is not safe to remove any safepoint
1836   // in this case since there's no safepoint dominating all paths in the loop body.
1837   bool prune = !keep_one || keep != NULL;
1838 
1839   // Delete other safepoints in this loop.
1840   Node_List* sfpts = _safepts;
1841   if (prune && sfpts != NULL) {
1842     assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint");
1843     for (uint i = 0; i < sfpts->size(); i++) {
1844       Node* n = sfpts->at(i);
1845       assert(phase->get_loop(n) == this, "");
1846       if (n != keep && phase->is_deleteable_safept(n)) {
1847         phase->lazy_replace(n, n->in(TypeFunc::Control));
1848       }
1849     }
1850   }
1851 }
1852 
1853 //------------------------------counted_loop-----------------------------------
1854 // Convert to counted loops where possible
1855 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {
1856 
1857   // For grins, set the inner-loop flag here
1858   if (!_child) {
1859     if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
1860   }
1861 
1862   if (_head->is_CountedLoop() ||
1863       phase->is_counted_loop(_head, this)) {
1864 
1865     if (!UseCountedLoopSafepoints) {
1866       // Indicate we do not need a safepoint here
1867       _has_sfpt = 1;
1868     }
1869 
1870     // Remove safepoints
1871     bool keep_one_sfpt = !(_has_call || _has_sfpt);
1872     remove_safepoints(phase, keep_one_sfpt);
1873 
1874     // Look for induction variables
1875     phase->replace_parallel_iv(this);
1876 
1877   } else if (_parent != NULL && !_irreducible) {
1878     // Not a counted loop. Keep one safepoint.
1879     bool keep_one_sfpt = true;
1880     remove_safepoints(phase, keep_one_sfpt);
1881   }
1882 
1883   // Recursively
1884   if (_child) _child->counted_loop( phase );
1885   if (_next)  _next ->counted_loop( phase );
1886 }
1887 
1888 #ifndef PRODUCT
1889 //------------------------------dump_head--------------------------------------
1890 // Dump 1 liner for loop header info
1891 void IdealLoopTree::dump_head( ) const {
1892   for (uint i=0; i<_nest; i++)
1893     tty->print("  ");
1894   tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx);
1895   if (_irreducible) tty->print(" IRREDUCIBLE");
1896   Node* entry = _head->in(LoopNode::EntryControl);
1897   if (LoopLimitCheck) {
1898     Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
1899     if (predicate != NULL ) {
1900       tty->print(" limit_check");
1901       entry = entry->in(0)->in(0);
1902     }
1903   }
1904   if (UseLoopPredicate) {
1905     entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
1906     if (entry != NULL) {
1907       tty->print(" predicated");
1908     }
1909   }
1910   if (_head->is_CountedLoop()) {
1911     CountedLoopNode *cl = _head->as_CountedLoop();
1912     tty->print(" counted");
1913 
1914     Node* init_n = cl->init_trip();
1915     if (init_n  != NULL &&  init_n->is_Con())
1916       tty->print(" [%d,", cl->init_trip()->get_int());
1917     else
1918       tty->print(" [int,");
1919     Node* limit_n = cl->limit();
1920     if (limit_n  != NULL &&  limit_n->is_Con())
1921       tty->print("%d),", cl->limit()->get_int());
1922     else
1923       tty->print("int),");
1924     int stride_con  = cl->stride_con();
1925     if (stride_con > 0) tty->print("+");
1926     tty->print("%d", stride_con);
1927 
1928     tty->print(" (%0.f iters) ", cl->profile_trip_cnt());
1929 
1930     if (cl->is_pre_loop ()) tty->print(" pre" );
1931     if (cl->is_main_loop()) tty->print(" main");
1932     if (cl->is_post_loop()) tty->print(" post");
1933   }
1934   if (_has_call) tty->print(" has_call");
1935   if (_has_sfpt) tty->print(" has_sfpt");
1936   if (_rce_candidate) tty->print(" rce");
1937   if (_safepts != NULL && _safepts->size() > 0) {
1938     tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");
1939   }
1940   if (_required_safept != NULL && _required_safept->size() > 0) {
1941     tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");
1942   }
1943   tty->cr();
1944 }
1945 
1946 //------------------------------dump-------------------------------------------
1947 // Dump loops by loop tree
1948 void IdealLoopTree::dump( ) const {
1949   dump_head();
1950   if (_child) _child->dump();
1951   if (_next)  _next ->dump();
1952 }
1953 
1954 #endif
1955 
1956 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {
1957   if (loop == root) {
1958     if (loop->_child != NULL) {
1959       log->begin_head("loop_tree");
1960       log->end_head();
1961       if( loop->_child ) log_loop_tree(root, loop->_child, log);
1962       log->tail("loop_tree");
1963       assert(loop->_next == NULL, "what?");
1964     }
1965   } else {
1966     Node* head = loop->_head;
1967     log->begin_head("loop");
1968     log->print(" idx='%d' ", head->_idx);
1969     if (loop->_irreducible) log->print("irreducible='1' ");
1970     if (head->is_Loop()) {
1971       if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");
1972       if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");
1973     }
1974     if (head->is_CountedLoop()) {
1975       CountedLoopNode* cl = head->as_CountedLoop();
1976       if (cl->is_pre_loop())  log->print("pre_loop='%d' ",  cl->main_idx());
1977       if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);
1978       if (cl->is_post_loop()) log->print("post_loop='%d' ",  cl->main_idx());
1979     }
1980     log->end_head();
1981     if( loop->_child ) log_loop_tree(root, loop->_child, log);
1982     log->tail("loop");
1983     if( loop->_next  ) log_loop_tree(root, loop->_next, log);
1984   }
1985 }
1986 
1987 //---------------------collect_potentially_useful_predicates-----------------------
1988 // Helper function to collect potentially useful predicates to prevent them from
1989 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates
1990 void PhaseIdealLoop::collect_potentially_useful_predicates(
1991                          IdealLoopTree * loop, Unique_Node_List &useful_predicates) {
1992   if (loop->_child) { // child
1993     collect_potentially_useful_predicates(loop->_child, useful_predicates);
1994   }
1995 
1996   // self (only loops that we can apply loop predication may use their predicates)
1997   if (loop->_head->is_Loop() &&
1998       !loop->_irreducible    &&
1999       !loop->tail()->is_top()) {
2000     LoopNode* lpn = loop->_head->as_Loop();
2001     Node* entry = lpn->in(LoopNode::EntryControl);
2002     Node* predicate_proj = find_predicate(entry); // loop_limit_check first
2003     if (predicate_proj != NULL ) { // right pattern that can be used by loop predication
2004       assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be");
2005       useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2006       entry = entry->in(0)->in(0);
2007     }
2008     predicate_proj = find_predicate(entry); // Predicate
2009     if (predicate_proj != NULL ) {
2010       useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one
2011     }
2012   }
2013 
2014   if (loop->_next) { // sibling
2015     collect_potentially_useful_predicates(loop->_next, useful_predicates);
2016   }
2017 }
2018 
2019 //------------------------eliminate_useless_predicates-----------------------------
2020 // Eliminate all inserted predicates if they could not be used by loop predication.
2021 // Note: it will also eliminates loop limits check predicate since it also uses
2022 // Opaque1 node (see Parse::add_predicate()).
2023 void PhaseIdealLoop::eliminate_useless_predicates() {
2024   if (C->predicate_count() == 0)
2025     return; // no predicate left
2026 
2027   Unique_Node_List useful_predicates; // to store useful predicates
2028   if (C->has_loops()) {
2029     collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates);
2030   }
2031 
2032   for (int i = C->predicate_count(); i > 0; i--) {
2033      Node * n = C->predicate_opaque1_node(i-1);
2034      assert(n->Opcode() == Op_Opaque1, "must be");
2035      if (!useful_predicates.member(n)) { // not in the useful list
2036        _igvn.replace_node(n, n->in(1));
2037      }
2038   }
2039 }
2040 
2041 //------------------------process_expensive_nodes-----------------------------
2042 // Expensive nodes have their control input set to prevent the GVN
2043 // from commoning them and as a result forcing the resulting node to
2044 // be in a more frequent path. Use CFG information here, to change the
2045 // control inputs so that some expensive nodes can be commoned while
2046 // not executed more frequently.
2047 bool PhaseIdealLoop::process_expensive_nodes() {
2048   assert(OptimizeExpensiveOps, "optimization off?");
2049 
2050   // Sort nodes to bring similar nodes together
2051   C->sort_expensive_nodes();
2052 
2053   bool progress = false;
2054 
2055   for (int i = 0; i < C->expensive_count(); ) {
2056     Node* n = C->expensive_node(i);
2057     int start = i;
2058     // Find nodes similar to n
2059     i++;
2060     for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
2061     int end = i;
2062     // And compare them two by two
2063     for (int j = start; j < end; j++) {
2064       Node* n1 = C->expensive_node(j);
2065       if (is_node_unreachable(n1)) {
2066         continue;
2067       }
2068       for (int k = j+1; k < end; k++) {
2069         Node* n2 = C->expensive_node(k);
2070         if (is_node_unreachable(n2)) {
2071           continue;
2072         }
2073 
2074         assert(n1 != n2, "should be pair of nodes");
2075 
2076         Node* c1 = n1->in(0);
2077         Node* c2 = n2->in(0);
2078 
2079         Node* parent_c1 = c1;
2080         Node* parent_c2 = c2;
2081 
2082         // The call to get_early_ctrl_for_expensive() moves the
2083         // expensive nodes up but stops at loops that are in a if
2084         // branch. See whether we can exit the loop and move above the
2085         // If.
2086         if (c1->is_Loop()) {
2087           parent_c1 = c1->in(1);
2088         }
2089         if (c2->is_Loop()) {
2090           parent_c2 = c2->in(1);
2091         }
2092 
2093         if (parent_c1 == parent_c2) {
2094           _igvn._worklist.push(n1);
2095           _igvn._worklist.push(n2);
2096           continue;
2097         }
2098 
2099         // Look for identical expensive node up the dominator chain.
2100         if (is_dominator(c1, c2)) {
2101           c2 = c1;
2102         } else if (is_dominator(c2, c1)) {
2103           c1 = c2;
2104         } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
2105                    parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {
2106           // Both branches have the same expensive node so move it up
2107           // before the if.
2108           c1 = c2 = idom(parent_c1->in(0));
2109         }
2110         // Do the actual moves
2111         if (n1->in(0) != c1) {
2112           _igvn.hash_delete(n1);
2113           n1->set_req(0, c1);
2114           _igvn.hash_insert(n1);
2115           _igvn._worklist.push(n1);
2116           progress = true;
2117         }
2118         if (n2->in(0) != c2) {
2119           _igvn.hash_delete(n2);
2120           n2->set_req(0, c2);
2121           _igvn.hash_insert(n2);
2122           _igvn._worklist.push(n2);
2123           progress = true;
2124         }
2125       }
2126     }
2127   }
2128 
2129   return progress;
2130 }
2131 
2132 
2133 //=============================================================================
2134 //----------------------------build_and_optimize-------------------------------
2135 // Create a PhaseLoop.  Build the ideal Loop tree.  Map each Ideal Node to
2136 // its corresponding LoopNode.
2137 void PhaseIdealLoop::build_and_optimize(bool do_split_ifs, bool skip_loop_opts) {
2138   ResourceMark rm;
2139 
2140   bool old_progress = C->major_progress();
2141   uint orig_worklist_size = _igvn._worklist.size();
2142 
2143   // Reset major-progress flag for the driver's heuristics
2144   C->clear_major_progress();
2145 
2146 #ifndef PRODUCT
2147   // Capture for later assert
2148   uint unique = C->unique();
2149   _loop_invokes++;
2150   _loop_work += unique;
2151 #endif
2152 
2153   // True if the method has at least 1 irreducible loop
2154   _has_irreducible_loops = false;
2155 
2156   _created_loop_node = false;
2157 
2158   Arena *a = Thread::current()->resource_area();
2159   VectorSet visited(a);
2160   // Pre-grow the mapping from Nodes to IdealLoopTrees.
2161   _nodes.map(C->unique(), NULL);
2162   memset(_nodes.adr(), 0, wordSize * C->unique());
2163 
2164   // Pre-build the top-level outermost loop tree entry
2165   _ltree_root = new IdealLoopTree( this, C->root(), C->root() );
2166   // Do not need a safepoint at the top level
2167   _ltree_root->_has_sfpt = 1;
2168 
2169   // Initialize Dominators.
2170   // Checked in clone_loop_predicate() during beautify_loops().
2171   _idom_size = 0;
2172   _idom      = NULL;
2173   _dom_depth = NULL;
2174   _dom_stk   = NULL;
2175 
2176   // Empty pre-order array
2177   allocate_preorders();
2178 
2179   // Build a loop tree on the fly.  Build a mapping from CFG nodes to
2180   // IdealLoopTree entries.  Data nodes are NOT walked.
2181   build_loop_tree();
2182   // Check for bailout, and return
2183   if (C->failing()) {
2184     return;
2185   }
2186 
2187   // No loops after all
2188   if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
2189 
2190   // There should always be an outer loop containing the Root and Return nodes.
2191   // If not, we have a degenerate empty program.  Bail out in this case.
2192   if (!has_node(C->root())) {
2193     if (!_verify_only) {
2194       C->clear_major_progress();
2195       C->record_method_not_compilable("empty program detected during loop optimization");
2196     }
2197     return;
2198   }
2199 
2200   // Nothing to do, so get out
2201   bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only;
2202   bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
2203   if (stop_early && !do_expensive_nodes) {
2204     _igvn.optimize();           // Cleanup NeverBranches
2205     return;
2206   }
2207 
2208   // Set loop nesting depth
2209   _ltree_root->set_nest( 0 );
2210 
2211   // Split shared headers and insert loop landing pads.
2212   // Do not bother doing this on the Root loop of course.
2213   if( !_verify_me && !_verify_only && _ltree_root->_child ) {
2214     C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
2215     if( _ltree_root->_child->beautify_loops( this ) ) {
2216       // Re-build loop tree!
2217       _ltree_root->_child = NULL;
2218       _nodes.clear();
2219       reallocate_preorders();
2220       build_loop_tree();
2221       // Check for bailout, and return
2222       if (C->failing()) {
2223         return;
2224       }
2225       // Reset loop nesting depth
2226       _ltree_root->set_nest( 0 );
2227 
2228       C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
2229     }
2230   }
2231 
2232   // Build Dominators for elision of NULL checks & loop finding.
2233   // Since nodes do not have a slot for immediate dominator, make
2234   // a persistent side array for that info indexed on node->_idx.
2235   _idom_size = C->unique();
2236   _idom      = NEW_RESOURCE_ARRAY( Node*, _idom_size );
2237   _dom_depth = NEW_RESOURCE_ARRAY( uint,  _idom_size );
2238   _dom_stk   = NULL; // Allocated on demand in recompute_dom_depth
2239   memset( _dom_depth, 0, _idom_size * sizeof(uint) );
2240 
2241   Dominators();
2242 
2243   if (!_verify_only) {
2244     // As a side effect, Dominators removed any unreachable CFG paths
2245     // into RegionNodes.  It doesn't do this test against Root, so
2246     // we do it here.
2247     for( uint i = 1; i < C->root()->req(); i++ ) {
2248       if( !_nodes[C->root()->in(i)->_idx] ) {    // Dead path into Root?
2249         _igvn.delete_input_of(C->root(), i);
2250         i--;                      // Rerun same iteration on compressed edges
2251       }
2252     }
2253 
2254     // Given dominators, try to find inner loops with calls that must
2255     // always be executed (call dominates loop tail).  These loops do
2256     // not need a separate safepoint.
2257     Node_List cisstack(a);
2258     _ltree_root->check_safepts(visited, cisstack);
2259   }
2260 
2261   // Walk the DATA nodes and place into loops.  Find earliest control
2262   // node.  For CFG nodes, the _nodes array starts out and remains
2263   // holding the associated IdealLoopTree pointer.  For DATA nodes, the
2264   // _nodes array holds the earliest legal controlling CFG node.
2265 
2266   // Allocate stack with enough space to avoid frequent realloc
2267   int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
2268   Node_Stack nstack( a, stack_size );
2269 
2270   visited.Clear();
2271   Node_List worklist(a);
2272   // Don't need C->root() on worklist since
2273   // it will be processed among C->top() inputs
2274   worklist.push( C->top() );
2275   visited.set( C->top()->_idx ); // Set C->top() as visited now
2276   build_loop_early( visited, worklist, nstack );
2277 
2278   // Given early legal placement, try finding counted loops.  This placement
2279   // is good enough to discover most loop invariants.
2280   if( !_verify_me && !_verify_only )
2281     _ltree_root->counted_loop( this );
2282 
2283   // Find latest loop placement.  Find ideal loop placement.
2284   visited.Clear();
2285   init_dom_lca_tags();
2286   // Need C->root() on worklist when processing outs
2287   worklist.push( C->root() );
2288   NOT_PRODUCT( C->verify_graph_edges(); )
2289   worklist.push( C->top() );
2290   build_loop_late( visited, worklist, nstack );
2291 
2292   if (_verify_only) {
2293     if (old_progress) {
2294       C->set_major_progress();
2295     }
2296     assert(C->unique() == unique, "verification mode made Nodes? ? ?");
2297     assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
2298     return;
2299   }
2300 
2301   // clear out the dead code after build_loop_late
2302   while (_deadlist.size()) {
2303     _igvn.remove_globally_dead_node(_deadlist.pop());
2304   }
2305 
2306   if (stop_early) {
2307     assert(do_expensive_nodes, "why are we here?");
2308     if (process_expensive_nodes()) {
2309       // If we made some progress when processing expensive nodes then
2310       // the IGVN may modify the graph in a way that will allow us to
2311       // make some more progress: we need to try processing expensive
2312       // nodes again.
2313       C->set_major_progress();
2314     }
2315     _igvn.optimize();
2316     return;
2317   }
2318 
2319   // Some parser-inserted loop predicates could never be used by loop
2320   // predication or they were moved away from loop during some optimizations.
2321   // For example, peeling. Eliminate them before next loop optimizations.
2322   if (UseLoopPredicate || LoopLimitCheck) {
2323     eliminate_useless_predicates();
2324   }
2325 
2326 #ifndef PRODUCT
2327   C->verify_graph_edges();
2328   if (_verify_me) {             // Nested verify pass?
2329     // Check to see if the verify mode is broken
2330     assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?");
2331     return;
2332   }
2333   if(VerifyLoopOptimizations) verify();
2334   if(TraceLoopOpts && C->has_loops()) {
2335     _ltree_root->dump();
2336   }
2337 #endif
2338 
2339   if (skip_loop_opts) {
2340     // restore major progress flag
2341     if (old_progress) {
2342       C->set_major_progress();
2343     }
2344 
2345     // Cleanup any modified bits
2346     _igvn.optimize();
2347 
2348     if (C->log() != NULL) {
2349       log_loop_tree(_ltree_root, _ltree_root, C->log());
2350     }
2351     return;
2352   }
2353 
2354   if (ReassociateInvariants) {
2355     // Reassociate invariants and prep for split_thru_phi
2356     for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2357       IdealLoopTree* lpt = iter.current();
2358       bool is_counted = lpt->is_counted();
2359       if (!is_counted || !lpt->is_inner()) continue;
2360 
2361       // check for vectorized loops, any reassociation of invariants was already done
2362       if (is_counted && lpt->_head->as_CountedLoop()->do_unroll_only()) continue;
2363 
2364       lpt->reassociate_invariants(this);
2365 
2366       // Because RCE opportunities can be masked by split_thru_phi,
2367       // look for RCE candidates and inhibit split_thru_phi
2368       // on just their loop-phi's for this pass of loop opts
2369       if (SplitIfBlocks && do_split_ifs) {
2370         if (lpt->policy_range_check(this)) {
2371           lpt->_rce_candidate = 1; // = true
2372         }
2373       }
2374     }
2375   }
2376 
2377   // Check for aggressive application of split-if and other transforms
2378   // that require basic-block info (like cloning through Phi's)
2379   if( SplitIfBlocks && do_split_ifs ) {
2380     visited.Clear();
2381     split_if_with_blocks( visited, nstack );
2382     NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); );
2383   }
2384 
2385   if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {
2386     C->set_major_progress();
2387   }
2388 
2389   // Perform loop predication before iteration splitting
2390   if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) {
2391     _ltree_root->_child->loop_predication(this);
2392   }
2393 
2394   if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {
2395     if (do_intrinsify_fill()) {
2396       C->set_major_progress();
2397     }
2398   }
2399 
2400   // Perform iteration-splitting on inner loops.  Split iterations to avoid
2401   // range checks or one-shot null checks.
2402 
2403   // If split-if's didn't hack the graph too bad (no CFG changes)
2404   // then do loop opts.
2405   if (C->has_loops() && !C->major_progress()) {
2406     memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) );
2407     _ltree_root->_child->iteration_split( this, worklist );
2408     // No verify after peeling!  GCM has hoisted code out of the loop.
2409     // After peeling, the hoisted code could sink inside the peeled area.
2410     // The peeling code does not try to recompute the best location for
2411     // all the code before the peeled area, so the verify pass will always
2412     // complain about it.
2413   }
2414   // Do verify graph edges in any case
2415   NOT_PRODUCT( C->verify_graph_edges(); );
2416 
2417   if (!do_split_ifs) {
2418     // We saw major progress in Split-If to get here.  We forced a
2419     // pass with unrolling and not split-if, however more split-if's
2420     // might make progress.  If the unrolling didn't make progress
2421     // then the major-progress flag got cleared and we won't try
2422     // another round of Split-If.  In particular the ever-common
2423     // instance-of/check-cast pattern requires at least 2 rounds of
2424     // Split-If to clear out.
2425     C->set_major_progress();
2426   }
2427 
2428   // Repeat loop optimizations if new loops were seen
2429   if (created_loop_node()) {
2430     C->set_major_progress();
2431   }
2432 
2433   // Keep loop predicates and perform optimizations with them
2434   // until no more loop optimizations could be done.
2435   // After that switch predicates off and do more loop optimizations.
2436   if (!C->major_progress() && (C->predicate_count() > 0)) {
2437      C->cleanup_loop_predicates(_igvn);
2438      if (TraceLoopOpts) {
2439        tty->print_cr("PredicatesOff");
2440      }
2441      C->set_major_progress();
2442   }
2443 
2444   // Convert scalar to superword operations at the end of all loop opts.
2445   if (UseSuperWord && C->has_loops() && !C->major_progress()) {
2446     // SuperWord transform
2447     SuperWord sw(this);
2448     for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2449       IdealLoopTree* lpt = iter.current();
2450       if (lpt->is_counted()) {
2451         sw.transform_loop(lpt, true);
2452       }
2453     }
2454   }
2455 
2456   // Cleanup any modified bits
2457   _igvn.optimize();
2458 
2459   // disable assert until issue with split_flow_path is resolved (6742111)
2460   // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(),
2461   //        "shouldn't introduce irreducible loops");
2462 
2463   if (C->log() != NULL) {
2464     log_loop_tree(_ltree_root, _ltree_root, C->log());
2465   }
2466 }
2467 
2468 #ifndef PRODUCT
2469 //------------------------------print_statistics-------------------------------
2470 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
2471 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
2472 void PhaseIdealLoop::print_statistics() {
2473   tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work);
2474 }
2475 
2476 //------------------------------verify-----------------------------------------
2477 // Build a verify-only PhaseIdealLoop, and see that it agrees with me.
2478 static int fail;                // debug only, so its multi-thread dont care
2479 void PhaseIdealLoop::verify() const {
2480   int old_progress = C->major_progress();
2481   ResourceMark rm;
2482   PhaseIdealLoop loop_verify( _igvn, this );
2483   VectorSet visited(Thread::current()->resource_area());
2484 
2485   fail = 0;
2486   verify_compare( C->root(), &loop_verify, visited );
2487   assert( fail == 0, "verify loops failed" );
2488   // Verify loop structure is the same
2489   _ltree_root->verify_tree(loop_verify._ltree_root, NULL);
2490   // Reset major-progress.  It was cleared by creating a verify version of
2491   // PhaseIdealLoop.
2492   if (old_progress) {
2493     C->set_major_progress();
2494   }
2495 }
2496 
2497 //------------------------------verify_compare---------------------------------
2498 // Make sure me and the given PhaseIdealLoop agree on key data structures
2499 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const {
2500   if( !n ) return;
2501   if( visited.test_set( n->_idx ) ) return;
2502   if( !_nodes[n->_idx] ) {      // Unreachable
2503     assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" );
2504     return;
2505   }
2506 
2507   uint i;
2508   for( i = 0; i < n->req(); i++ )
2509     verify_compare( n->in(i), loop_verify, visited );
2510 
2511   // Check the '_nodes' block/loop structure
2512   i = n->_idx;
2513   if( has_ctrl(n) ) {           // We have control; verify has loop or ctrl
2514     if( _nodes[i] != loop_verify->_nodes[i] &&
2515         get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {
2516       tty->print("Mismatched control setting for: ");
2517       n->dump();
2518       if( fail++ > 10 ) return;
2519       Node *c = get_ctrl_no_update(n);
2520       tty->print("We have it as: ");
2521       if( c->in(0) ) c->dump();
2522         else tty->print_cr("N%d",c->_idx);
2523       tty->print("Verify thinks: ");
2524       if( loop_verify->has_ctrl(n) )
2525         loop_verify->get_ctrl_no_update(n)->dump();
2526       else
2527         loop_verify->get_loop_idx(n)->dump();
2528       tty->cr();
2529     }
2530   } else {                    // We have a loop
2531     IdealLoopTree *us = get_loop_idx(n);
2532     if( loop_verify->has_ctrl(n) ) {
2533       tty->print("Mismatched loop setting for: ");
2534       n->dump();
2535       if( fail++ > 10 ) return;
2536       tty->print("We have it as: ");
2537       us->dump();
2538       tty->print("Verify thinks: ");
2539       loop_verify->get_ctrl_no_update(n)->dump();
2540       tty->cr();
2541     } else if (!C->major_progress()) {
2542       // Loop selection can be messed up if we did a major progress
2543       // operation, like split-if.  Do not verify in that case.
2544       IdealLoopTree *them = loop_verify->get_loop_idx(n);
2545       if( us->_head != them->_head ||  us->_tail != them->_tail ) {
2546         tty->print("Unequals loops for: ");
2547         n->dump();
2548         if( fail++ > 10 ) return;
2549         tty->print("We have it as: ");
2550         us->dump();
2551         tty->print("Verify thinks: ");
2552         them->dump();
2553         tty->cr();
2554       }
2555     }
2556   }
2557 
2558   // Check for immediate dominators being equal
2559   if( i >= _idom_size ) {
2560     if( !n->is_CFG() ) return;
2561     tty->print("CFG Node with no idom: ");
2562     n->dump();
2563     return;
2564   }
2565   if( !n->is_CFG() ) return;
2566   if( n == C->root() ) return; // No IDOM here
2567 
2568   assert(n->_idx == i, "sanity");
2569   Node *id = idom_no_update(n);
2570   if( id != loop_verify->idom_no_update(n) ) {
2571     tty->print("Unequals idoms for: ");
2572     n->dump();
2573     if( fail++ > 10 ) return;
2574     tty->print("We have it as: ");
2575     id->dump();
2576     tty->print("Verify thinks: ");
2577     loop_verify->idom_no_update(n)->dump();
2578     tty->cr();
2579   }
2580 
2581 }
2582 
2583 //------------------------------verify_tree------------------------------------
2584 // Verify that tree structures match.  Because the CFG can change, siblings
2585 // within the loop tree can be reordered.  We attempt to deal with that by
2586 // reordering the verify's loop tree if possible.
2587 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const {
2588   assert( _parent == parent, "Badly formed loop tree" );
2589 
2590   // Siblings not in same order?  Attempt to re-order.
2591   if( _head != loop->_head ) {
2592     // Find _next pointer to update
2593     IdealLoopTree **pp = &loop->_parent->_child;
2594     while( *pp != loop )
2595       pp = &((*pp)->_next);
2596     // Find proper sibling to be next
2597     IdealLoopTree **nn = &loop->_next;
2598     while( (*nn) && (*nn)->_head != _head )
2599       nn = &((*nn)->_next);
2600 
2601     // Check for no match.
2602     if( !(*nn) ) {
2603       // Annoyingly, irreducible loops can pick different headers
2604       // after a major_progress operation, so the rest of the loop
2605       // tree cannot be matched.
2606       if (_irreducible && Compile::current()->major_progress())  return;
2607       assert( 0, "failed to match loop tree" );
2608     }
2609 
2610     // Move (*nn) to (*pp)
2611     IdealLoopTree *hit = *nn;
2612     *nn = hit->_next;
2613     hit->_next = loop;
2614     *pp = loop;
2615     loop = hit;
2616     // Now try again to verify
2617   }
2618 
2619   assert( _head  == loop->_head , "mismatched loop head" );
2620   Node *tail = _tail;           // Inline a non-updating version of
2621   while( !tail->in(0) )         // the 'tail()' call.
2622     tail = tail->in(1);
2623   assert( tail == loop->_tail, "mismatched loop tail" );
2624 
2625   // Counted loops that are guarded should be able to find their guards
2626   if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) {
2627     CountedLoopNode *cl = _head->as_CountedLoop();
2628     Node *init = cl->init_trip();
2629     Node *ctrl = cl->in(LoopNode::EntryControl);
2630     assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
2631     Node *iff  = ctrl->in(0);
2632     assert( iff->Opcode() == Op_If, "" );
2633     Node *bol  = iff->in(1);
2634     assert( bol->Opcode() == Op_Bool, "" );
2635     Node *cmp  = bol->in(1);
2636     assert( cmp->Opcode() == Op_CmpI, "" );
2637     Node *add  = cmp->in(1);
2638     Node *opaq;
2639     if( add->Opcode() == Op_Opaque1 ) {
2640       opaq = add;
2641     } else {
2642       assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" );
2643       assert( add == init, "" );
2644       opaq = cmp->in(2);
2645     }
2646     assert( opaq->Opcode() == Op_Opaque1, "" );
2647 
2648   }
2649 
2650   if (_child != NULL)  _child->verify_tree(loop->_child, this);
2651   if (_next  != NULL)  _next ->verify_tree(loop->_next,  parent);
2652   // Innermost loops need to verify loop bodies,
2653   // but only if no 'major_progress'
2654   int fail = 0;
2655   if (!Compile::current()->major_progress() && _child == NULL) {
2656     for( uint i = 0; i < _body.size(); i++ ) {
2657       Node *n = _body.at(i);
2658       if (n->outcnt() == 0)  continue; // Ignore dead
2659       uint j;
2660       for( j = 0; j < loop->_body.size(); j++ )
2661         if( loop->_body.at(j) == n )
2662           break;
2663       if( j == loop->_body.size() ) { // Not found in loop body
2664         // Last ditch effort to avoid assertion: Its possible that we
2665         // have some users (so outcnt not zero) but are still dead.
2666         // Try to find from root.
2667         if (Compile::current()->root()->find(n->_idx)) {
2668           fail++;
2669           tty->print("We have that verify does not: ");
2670           n->dump();
2671         }
2672       }
2673     }
2674     for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {
2675       Node *n = loop->_body.at(i2);
2676       if (n->outcnt() == 0)  continue; // Ignore dead
2677       uint j;
2678       for( j = 0; j < _body.size(); j++ )
2679         if( _body.at(j) == n )
2680           break;
2681       if( j == _body.size() ) { // Not found in loop body
2682         // Last ditch effort to avoid assertion: Its possible that we
2683         // have some users (so outcnt not zero) but are still dead.
2684         // Try to find from root.
2685         if (Compile::current()->root()->find(n->_idx)) {
2686           fail++;
2687           tty->print("Verify has that we do not: ");
2688           n->dump();
2689         }
2690       }
2691     }
2692     assert( !fail, "loop body mismatch" );
2693   }
2694 }
2695 
2696 #endif
2697 
2698 //------------------------------set_idom---------------------------------------
2699 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {
2700   uint idx = d->_idx;
2701   if (idx >= _idom_size) {
2702     uint newsize = _idom_size<<1;
2703     while( idx >= newsize ) {
2704       newsize <<= 1;
2705     }
2706     _idom      = REALLOC_RESOURCE_ARRAY( Node*,     _idom,_idom_size,newsize);
2707     _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
2708     memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
2709     _idom_size = newsize;
2710   }
2711   _idom[idx] = n;
2712   _dom_depth[idx] = dom_depth;
2713 }
2714 
2715 //------------------------------recompute_dom_depth---------------------------------------
2716 // The dominator tree is constructed with only parent pointers.
2717 // This recomputes the depth in the tree by first tagging all
2718 // nodes as "no depth yet" marker.  The next pass then runs up
2719 // the dom tree from each node marked "no depth yet", and computes
2720 // the depth on the way back down.
2721 void PhaseIdealLoop::recompute_dom_depth() {
2722   uint no_depth_marker = C->unique();
2723   uint i;
2724   // Initialize depth to "no depth yet"
2725   for (i = 0; i < _idom_size; i++) {
2726     if (_dom_depth[i] > 0 && _idom[i] != NULL) {
2727      _dom_depth[i] = no_depth_marker;
2728     }
2729   }
2730   if (_dom_stk == NULL) {
2731     uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
2732     if (init_size < 10) init_size = 10;
2733     _dom_stk = new GrowableArray<uint>(init_size);
2734   }
2735   // Compute new depth for each node.
2736   for (i = 0; i < _idom_size; i++) {
2737     uint j = i;
2738     // Run up the dom tree to find a node with a depth
2739     while (_dom_depth[j] == no_depth_marker) {
2740       _dom_stk->push(j);
2741       j = _idom[j]->_idx;
2742     }
2743     // Compute the depth on the way back down this tree branch
2744     uint dd = _dom_depth[j] + 1;
2745     while (_dom_stk->length() > 0) {
2746       uint j = _dom_stk->pop();
2747       _dom_depth[j] = dd;
2748       dd++;
2749     }
2750   }
2751 }
2752 
2753 //------------------------------sort-------------------------------------------
2754 // Insert 'loop' into the existing loop tree.  'innermost' is a leaf of the
2755 // loop tree, not the root.
2756 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {
2757   if( !innermost ) return loop; // New innermost loop
2758 
2759   int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
2760   assert( loop_preorder, "not yet post-walked loop" );
2761   IdealLoopTree **pp = &innermost;      // Pointer to previous next-pointer
2762   IdealLoopTree *l = *pp;               // Do I go before or after 'l'?
2763 
2764   // Insert at start of list
2765   while( l ) {                  // Insertion sort based on pre-order
2766     if( l == loop ) return innermost; // Already on list!
2767     int l_preorder = get_preorder(l->_head); // Cache pre-order number
2768     assert( l_preorder, "not yet post-walked l" );
2769     // Check header pre-order number to figure proper nesting
2770     if( loop_preorder > l_preorder )
2771       break;                    // End of insertion
2772     // If headers tie (e.g., shared headers) check tail pre-order numbers.
2773     // Since I split shared headers, you'd think this could not happen.
2774     // BUT: I must first do the preorder numbering before I can discover I
2775     // have shared headers, so the split headers all get the same preorder
2776     // number as the RegionNode they split from.
2777     if( loop_preorder == l_preorder &&
2778         get_preorder(loop->_tail) < get_preorder(l->_tail) )
2779       break;                    // Also check for shared headers (same pre#)
2780     pp = &l->_parent;           // Chain up list
2781     l = *pp;
2782   }
2783   // Link into list
2784   // Point predecessor to me
2785   *pp = loop;
2786   // Point me to successor
2787   IdealLoopTree *p = loop->_parent;
2788   loop->_parent = l;            // Point me to successor
2789   if( p ) sort( p, innermost ); // Insert my parents into list as well
2790   return innermost;
2791 }
2792 
2793 //------------------------------build_loop_tree--------------------------------
2794 // I use a modified Vick/Tarjan algorithm.  I need pre- and a post- visit
2795 // bits.  The _nodes[] array is mapped by Node index and holds a NULL for
2796 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
2797 // tightest enclosing IdealLoopTree for post-walked.
2798 //
2799 // During my forward walk I do a short 1-layer lookahead to see if I can find
2800 // a loop backedge with that doesn't have any work on the backedge.  This
2801 // helps me construct nested loops with shared headers better.
2802 //
2803 // Once I've done the forward recursion, I do the post-work.  For each child
2804 // I check to see if there is a backedge.  Backedges define a loop!  I
2805 // insert an IdealLoopTree at the target of the backedge.
2806 //
2807 // During the post-work I also check to see if I have several children
2808 // belonging to different loops.  If so, then this Node is a decision point
2809 // where control flow can choose to change loop nests.  It is at this
2810 // decision point where I can figure out how loops are nested.  At this
2811 // time I can properly order the different loop nests from my children.
2812 // Note that there may not be any backedges at the decision point!
2813 //
2814 // Since the decision point can be far removed from the backedges, I can't
2815 // order my loops at the time I discover them.  Thus at the decision point
2816 // I need to inspect loop header pre-order numbers to properly nest my
2817 // loops.  This means I need to sort my childrens' loops by pre-order.
2818 // The sort is of size number-of-control-children, which generally limits
2819 // it to size 2 (i.e., I just choose between my 2 target loops).
2820 void PhaseIdealLoop::build_loop_tree() {
2821   // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
2822   GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
2823   Node *n = C->root();
2824   bltstack.push(n);
2825   int pre_order = 1;
2826   int stack_size;
2827 
2828   while ( ( stack_size = bltstack.length() ) != 0 ) {
2829     n = bltstack.top(); // Leave node on stack
2830     if ( !is_visited(n) ) {
2831       // ---- Pre-pass Work ----
2832       // Pre-walked but not post-walked nodes need a pre_order number.
2833 
2834       set_preorder_visited( n, pre_order ); // set as visited
2835 
2836       // ---- Scan over children ----
2837       // Scan first over control projections that lead to loop headers.
2838       // This helps us find inner-to-outer loops with shared headers better.
2839 
2840       // Scan children's children for loop headers.
2841       for ( int i = n->outcnt() - 1; i >= 0; --i ) {
2842         Node* m = n->raw_out(i);       // Child
2843         if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children
2844           // Scan over children's children to find loop
2845           for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2846             Node* l = m->fast_out(j);
2847             if( is_visited(l) &&       // Been visited?
2848                 !is_postvisited(l) &&  // But not post-visited
2849                 get_preorder(l) < pre_order ) { // And smaller pre-order
2850               // Found!  Scan the DFS down this path before doing other paths
2851               bltstack.push(m);
2852               break;
2853             }
2854           }
2855         }
2856       }
2857       pre_order++;
2858     }
2859     else if ( !is_postvisited(n) ) {
2860       // Note: build_loop_tree_impl() adds out edges on rare occasions,
2861       // such as com.sun.rsasign.am::a.
2862       // For non-recursive version, first, process current children.
2863       // On next iteration, check if additional children were added.
2864       for ( int k = n->outcnt() - 1; k >= 0; --k ) {
2865         Node* u = n->raw_out(k);
2866         if ( u->is_CFG() && !is_visited(u) ) {
2867           bltstack.push(u);
2868         }
2869       }
2870       if ( bltstack.length() == stack_size ) {
2871         // There were no additional children, post visit node now
2872         (void)bltstack.pop(); // Remove node from stack
2873         pre_order = build_loop_tree_impl( n, pre_order );
2874         // Check for bailout
2875         if (C->failing()) {
2876           return;
2877         }
2878         // Check to grow _preorders[] array for the case when
2879         // build_loop_tree_impl() adds new nodes.
2880         check_grow_preorders();
2881       }
2882     }
2883     else {
2884       (void)bltstack.pop(); // Remove post-visited node from stack
2885     }
2886   }
2887 }
2888 
2889 //------------------------------build_loop_tree_impl---------------------------
2890 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {
2891   // ---- Post-pass Work ----
2892   // Pre-walked but not post-walked nodes need a pre_order number.
2893 
2894   // Tightest enclosing loop for this Node
2895   IdealLoopTree *innermost = NULL;
2896 
2897   // For all children, see if any edge is a backedge.  If so, make a loop
2898   // for it.  Then find the tightest enclosing loop for the self Node.
2899   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2900     Node* m = n->fast_out(i);   // Child
2901     if( n == m ) continue;      // Ignore control self-cycles
2902     if( !m->is_CFG() ) continue;// Ignore non-CFG edges
2903 
2904     IdealLoopTree *l;           // Child's loop
2905     if( !is_postvisited(m) ) {  // Child visited but not post-visited?
2906       // Found a backedge
2907       assert( get_preorder(m) < pre_order, "should be backedge" );
2908       // Check for the RootNode, which is already a LoopNode and is allowed
2909       // to have multiple "backedges".
2910       if( m == C->root()) {     // Found the root?
2911         l = _ltree_root;        // Root is the outermost LoopNode
2912       } else {                  // Else found a nested loop
2913         // Insert a LoopNode to mark this loop.
2914         l = new IdealLoopTree(this, m, n);
2915       } // End of Else found a nested loop
2916       if( !has_loop(m) )        // If 'm' does not already have a loop set
2917         set_loop(m, l);         // Set loop header to loop now
2918 
2919     } else {                    // Else not a nested loop
2920       if( !_nodes[m->_idx] ) continue; // Dead code has no loop
2921       l = get_loop(m);          // Get previously determined loop
2922       // If successor is header of a loop (nest), move up-loop till it
2923       // is a member of some outer enclosing loop.  Since there are no
2924       // shared headers (I've split them already) I only need to go up
2925       // at most 1 level.
2926       while( l && l->_head == m ) // Successor heads loop?
2927         l = l->_parent;         // Move up 1 for me
2928       // If this loop is not properly parented, then this loop
2929       // has no exit path out, i.e. its an infinite loop.
2930       if( !l ) {
2931         // Make loop "reachable" from root so the CFG is reachable.  Basically
2932         // insert a bogus loop exit that is never taken.  'm', the loop head,
2933         // points to 'n', one (of possibly many) fall-in paths.  There may be
2934         // many backedges as well.
2935 
2936         // Here I set the loop to be the root loop.  I could have, after
2937         // inserting a bogus loop exit, restarted the recursion and found my
2938         // new loop exit.  This would make the infinite loop a first-class
2939         // loop and it would then get properly optimized.  What's the use of
2940         // optimizing an infinite loop?
2941         l = _ltree_root;        // Oops, found infinite loop
2942 
2943         if (!_verify_only) {
2944           // Insert the NeverBranch between 'm' and it's control user.
2945           NeverBranchNode *iff = new NeverBranchNode( m );
2946           _igvn.register_new_node_with_optimizer(iff);
2947           set_loop(iff, l);
2948           Node *if_t = new CProjNode( iff, 0 );
2949           _igvn.register_new_node_with_optimizer(if_t);
2950           set_loop(if_t, l);
2951 
2952           Node* cfg = NULL;       // Find the One True Control User of m
2953           for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
2954             Node* x = m->fast_out(j);
2955             if (x->is_CFG() && x != m && x != iff)
2956               { cfg = x; break; }
2957           }
2958           assert(cfg != NULL, "must find the control user of m");
2959           uint k = 0;             // Probably cfg->in(0)
2960           while( cfg->in(k) != m ) k++; // But check incase cfg is a Region
2961           cfg->set_req( k, if_t ); // Now point to NeverBranch
2962           _igvn._worklist.push(cfg);
2963 
2964           // Now create the never-taken loop exit
2965           Node *if_f = new CProjNode( iff, 1 );
2966           _igvn.register_new_node_with_optimizer(if_f);
2967           set_loop(if_f, l);
2968           // Find frame ptr for Halt.  Relies on the optimizer
2969           // V-N'ing.  Easier and quicker than searching through
2970           // the program structure.
2971           Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr );
2972           _igvn.register_new_node_with_optimizer(frame);
2973           // Halt & Catch Fire
2974           Node *halt = new HaltNode( if_f, frame );
2975           _igvn.register_new_node_with_optimizer(halt);
2976           set_loop(halt, l);
2977           C->root()->add_req(halt);
2978         }
2979         set_loop(C->root(), _ltree_root);
2980       }
2981     }
2982     // Weeny check for irreducible.  This child was already visited (this
2983     // IS the post-work phase).  Is this child's loop header post-visited
2984     // as well?  If so, then I found another entry into the loop.
2985     if (!_verify_only) {
2986       while( is_postvisited(l->_head) ) {
2987         // found irreducible
2988         l->_irreducible = 1; // = true
2989         l = l->_parent;
2990         _has_irreducible_loops = true;
2991         // Check for bad CFG here to prevent crash, and bailout of compile
2992         if (l == NULL) {
2993           C->record_method_not_compilable("unhandled CFG detected during loop optimization");
2994           return pre_order;
2995         }
2996       }
2997       C->set_has_irreducible_loop(_has_irreducible_loops);
2998     }
2999 
3000     // This Node might be a decision point for loops.  It is only if
3001     // it's children belong to several different loops.  The sort call
3002     // does a trivial amount of work if there is only 1 child or all
3003     // children belong to the same loop.  If however, the children
3004     // belong to different loops, the sort call will properly set the
3005     // _parent pointers to show how the loops nest.
3006     //
3007     // In any case, it returns the tightest enclosing loop.
3008     innermost = sort( l, innermost );
3009   }
3010 
3011   // Def-use info will have some dead stuff; dead stuff will have no
3012   // loop decided on.
3013 
3014   // Am I a loop header?  If so fix up my parent's child and next ptrs.
3015   if( innermost && innermost->_head == n ) {
3016     assert( get_loop(n) == innermost, "" );
3017     IdealLoopTree *p = innermost->_parent;
3018     IdealLoopTree *l = innermost;
3019     while( p && l->_head == n ) {
3020       l->_next = p->_child;     // Put self on parents 'next child'
3021       p->_child = l;            // Make self as first child of parent
3022       l = p;                    // Now walk up the parent chain
3023       p = l->_parent;
3024     }
3025   } else {
3026     // Note that it is possible for a LoopNode to reach here, if the
3027     // backedge has been made unreachable (hence the LoopNode no longer
3028     // denotes a Loop, and will eventually be removed).
3029 
3030     // Record tightest enclosing loop for self.  Mark as post-visited.
3031     set_loop(n, innermost);
3032     // Also record has_call flag early on
3033     if( innermost ) {
3034       if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {
3035         // Do not count uncommon calls
3036         if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {
3037           Node *iff = n->in(0)->in(0);
3038           // No any calls for vectorized loops.
3039           if( UseSuperWord || !iff->is_If() ||
3040               (n->in(0)->Opcode() == Op_IfFalse &&
3041                (1.0 - iff->as_If()->_prob) >= 0.01) ||
3042               (iff->as_If()->_prob >= 0.01) )
3043             innermost->_has_call = 1;
3044         }
3045       } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {
3046         // Disable loop optimizations if the loop has a scalar replaceable
3047         // allocation. This disabling may cause a potential performance lost
3048         // if the allocation is not eliminated for some reason.
3049         innermost->_allow_optimizations = false;
3050         innermost->_has_call = 1; // = true
3051       } else if (n->Opcode() == Op_SafePoint) {
3052         // Record all safepoints in this loop.
3053         if (innermost->_safepts == NULL) innermost->_safepts = new Node_List();
3054         innermost->_safepts->push(n);
3055       }
3056     }
3057   }
3058 
3059   // Flag as post-visited now
3060   set_postvisited(n);
3061   return pre_order;
3062 }
3063 
3064 
3065 //------------------------------build_loop_early-------------------------------
3066 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3067 // First pass computes the earliest controlling node possible.  This is the
3068 // controlling input with the deepest dominating depth.
3069 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3070   while (worklist.size() != 0) {
3071     // Use local variables nstack_top_n & nstack_top_i to cache values
3072     // on nstack's top.
3073     Node *nstack_top_n = worklist.pop();
3074     uint  nstack_top_i = 0;
3075 //while_nstack_nonempty:
3076     while (true) {
3077       // Get parent node and next input's index from stack's top.
3078       Node  *n = nstack_top_n;
3079       uint   i = nstack_top_i;
3080       uint cnt = n->req(); // Count of inputs
3081       if (i == 0) {        // Pre-process the node.
3082         if( has_node(n) &&            // Have either loop or control already?
3083             !has_ctrl(n) ) {          // Have loop picked out already?
3084           // During "merge_many_backedges" we fold up several nested loops
3085           // into a single loop.  This makes the members of the original
3086           // loop bodies pointing to dead loops; they need to move up
3087           // to the new UNION'd larger loop.  I set the _head field of these
3088           // dead loops to NULL and the _parent field points to the owning
3089           // loop.  Shades of UNION-FIND algorithm.
3090           IdealLoopTree *ilt;
3091           while( !(ilt = get_loop(n))->_head ) {
3092             // Normally I would use a set_loop here.  But in this one special
3093             // case, it is legal (and expected) to change what loop a Node
3094             // belongs to.
3095             _nodes.map(n->_idx, (Node*)(ilt->_parent) );
3096           }
3097           // Remove safepoints ONLY if I've already seen I don't need one.
3098           // (the old code here would yank a 2nd safepoint after seeing a
3099           // first one, even though the 1st did not dominate in the loop body
3100           // and thus could be avoided indefinitely)
3101           if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
3102               is_deleteable_safept(n)) {
3103             Node *in = n->in(TypeFunc::Control);
3104             lazy_replace(n,in);       // Pull safepoint now
3105             if (ilt->_safepts != NULL) {
3106               ilt->_safepts->yank(n);
3107             }
3108             // Carry on with the recursion "as if" we are walking
3109             // only the control input
3110             if( !visited.test_set( in->_idx ) ) {
3111               worklist.push(in);      // Visit this guy later, using worklist
3112             }
3113             // Get next node from nstack:
3114             // - skip n's inputs processing by setting i > cnt;
3115             // - we also will not call set_early_ctrl(n) since
3116             //   has_node(n) == true (see the condition above).
3117             i = cnt + 1;
3118           }
3119         }
3120       } // if (i == 0)
3121 
3122       // Visit all inputs
3123       bool done = true;       // Assume all n's inputs will be processed
3124       while (i < cnt) {
3125         Node *in = n->in(i);
3126         ++i;
3127         if (in == NULL) continue;
3128         if (in->pinned() && !in->is_CFG())
3129           set_ctrl(in, in->in(0));
3130         int is_visited = visited.test_set( in->_idx );
3131         if (!has_node(in)) {  // No controlling input yet?
3132           assert( !in->is_CFG(), "CFG Node with no controlling input?" );
3133           assert( !is_visited, "visit only once" );
3134           nstack.push(n, i);  // Save parent node and next input's index.
3135           nstack_top_n = in;  // Process current input now.
3136           nstack_top_i = 0;
3137           done = false;       // Not all n's inputs processed.
3138           break; // continue while_nstack_nonempty;
3139         } else if (!is_visited) {
3140           // This guy has a location picked out for him, but has not yet
3141           // been visited.  Happens to all CFG nodes, for instance.
3142           // Visit him using the worklist instead of recursion, to break
3143           // cycles.  Since he has a location already we do not need to
3144           // find his location before proceeding with the current Node.
3145           worklist.push(in);  // Visit this guy later, using worklist
3146         }
3147       }
3148       if (done) {
3149         // All of n's inputs have been processed, complete post-processing.
3150 
3151         // Compute earliest point this Node can go.
3152         // CFG, Phi, pinned nodes already know their controlling input.
3153         if (!has_node(n)) {
3154           // Record earliest legal location
3155           set_early_ctrl( n );
3156         }
3157         if (nstack.is_empty()) {
3158           // Finished all nodes on stack.
3159           // Process next node on the worklist.
3160           break;
3161         }
3162         // Get saved parent node and next input's index.
3163         nstack_top_n = nstack.node();
3164         nstack_top_i = nstack.index();
3165         nstack.pop();
3166       }
3167     } // while (true)
3168   }
3169 }
3170 
3171 //------------------------------dom_lca_internal--------------------------------
3172 // Pair-wise LCA
3173 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {
3174   if( !n1 ) return n2;          // Handle NULL original LCA
3175   assert( n1->is_CFG(), "" );
3176   assert( n2->is_CFG(), "" );
3177   // find LCA of all uses
3178   uint d1 = dom_depth(n1);
3179   uint d2 = dom_depth(n2);
3180   while (n1 != n2) {
3181     if (d1 > d2) {
3182       n1 =      idom(n1);
3183       d1 = dom_depth(n1);
3184     } else if (d1 < d2) {
3185       n2 =      idom(n2);
3186       d2 = dom_depth(n2);
3187     } else {
3188       // Here d1 == d2.  Due to edits of the dominator-tree, sections
3189       // of the tree might have the same depth.  These sections have
3190       // to be searched more carefully.
3191 
3192       // Scan up all the n1's with equal depth, looking for n2.
3193       Node *t1 = idom(n1);
3194       while (dom_depth(t1) == d1) {
3195         if (t1 == n2)  return n2;
3196         t1 = idom(t1);
3197       }
3198       // Scan up all the n2's with equal depth, looking for n1.
3199       Node *t2 = idom(n2);
3200       while (dom_depth(t2) == d2) {
3201         if (t2 == n1)  return n1;
3202         t2 = idom(t2);
3203       }
3204       // Move up to a new dominator-depth value as well as up the dom-tree.
3205       n1 = t1;
3206       n2 = t2;
3207       d1 = dom_depth(n1);
3208       d2 = dom_depth(n2);
3209     }
3210   }
3211   return n1;
3212 }
3213 
3214 //------------------------------compute_idom-----------------------------------
3215 // Locally compute IDOM using dom_lca call.  Correct only if the incoming
3216 // IDOMs are correct.
3217 Node *PhaseIdealLoop::compute_idom( Node *region ) const {
3218   assert( region->is_Region(), "" );
3219   Node *LCA = NULL;
3220   for( uint i = 1; i < region->req(); i++ ) {
3221     if( region->in(i) != C->top() )
3222       LCA = dom_lca( LCA, region->in(i) );
3223   }
3224   return LCA;
3225 }
3226 
3227 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {
3228   bool had_error = false;
3229 #ifdef ASSERT
3230   if (early != C->root()) {
3231     // Make sure that there's a dominance path from LCA to early
3232     Node* d = LCA;
3233     while (d != early) {
3234       if (d == C->root()) {
3235         dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);
3236         tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);
3237         had_error = true;
3238         break;
3239       }
3240       d = idom(d);
3241     }
3242   }
3243 #endif
3244   return had_error;
3245 }
3246 
3247 
3248 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {
3249   // Compute LCA over list of uses
3250   bool had_error = false;
3251   Node *LCA = NULL;
3252   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {
3253     Node* c = n->fast_out(i);
3254     if (_nodes[c->_idx] == NULL)
3255       continue;                 // Skip the occasional dead node
3256     if( c->is_Phi() ) {         // For Phis, we must land above on the path
3257       for( uint j=1; j<c->req(); j++ ) {// For all inputs
3258         if( c->in(j) == n ) {   // Found matching input?
3259           Node *use = c->in(0)->in(j);
3260           if (_verify_only && use->is_top()) continue;
3261           LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3262           if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3263         }
3264       }
3265     } else {
3266       // For CFG data-users, use is in the block just prior
3267       Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
3268       LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
3269       if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
3270     }
3271   }
3272   assert(!had_error, "bad dominance");
3273   return LCA;
3274 }
3275 
3276 //------------------------------get_late_ctrl----------------------------------
3277 // Compute latest legal control.
3278 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {
3279   assert(early != NULL, "early control should not be NULL");
3280 
3281   Node* LCA = compute_lca_of_uses(n, early);
3282 #ifdef ASSERT
3283   if (LCA == C->root() && LCA != early) {
3284     // def doesn't dominate uses so print some useful debugging output
3285     compute_lca_of_uses(n, early, true);
3286   }
3287 #endif
3288 
3289   // if this is a load, check for anti-dependent stores
3290   // We use a conservative algorithm to identify potential interfering
3291   // instructions and for rescheduling the load.  The users of the memory
3292   // input of this load are examined.  Any use which is not a load and is
3293   // dominated by early is considered a potentially interfering store.
3294   // This can produce false positives.
3295   if (n->is_Load() && LCA != early) {
3296     Node_List worklist;
3297 
3298     Node *mem = n->in(MemNode::Memory);
3299     for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
3300       Node* s = mem->fast_out(i);
3301       worklist.push(s);
3302     }
3303     while(worklist.size() != 0 && LCA != early) {
3304       Node* s = worklist.pop();
3305       if (s->is_Load()) {
3306         continue;
3307       } else if (s->is_MergeMem()) {
3308         for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {
3309           Node* s1 = s->fast_out(i);
3310           worklist.push(s1);
3311         }
3312       } else {
3313         Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
3314         assert(sctrl != NULL || s->outcnt() == 0, "must have control");
3315         if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) {
3316           LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
3317         }
3318       }
3319     }
3320   }
3321 
3322   assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
3323   return LCA;
3324 }
3325 
3326 // true if CFG node d dominates CFG node n
3327 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) {
3328   if (d == n)
3329     return true;
3330   assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes");
3331   uint dd = dom_depth(d);
3332   while (dom_depth(n) >= dd) {
3333     if (n == d)
3334       return true;
3335     n = idom(n);
3336   }
3337   return false;
3338 }
3339 
3340 //------------------------------dom_lca_for_get_late_ctrl_internal-------------
3341 // Pair-wise LCA with tags.
3342 // Tag each index with the node 'tag' currently being processed
3343 // before advancing up the dominator chain using idom().
3344 // Later calls that find a match to 'tag' know that this path has already
3345 // been considered in the current LCA (which is input 'n1' by convention).
3346 // Since get_late_ctrl() is only called once for each node, the tag array
3347 // does not need to be cleared between calls to get_late_ctrl().
3348 // Algorithm trades a larger constant factor for better asymptotic behavior
3349 //
3350 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) {
3351   uint d1 = dom_depth(n1);
3352   uint d2 = dom_depth(n2);
3353 
3354   do {
3355     if (d1 > d2) {
3356       // current lca is deeper than n2
3357       _dom_lca_tags.map(n1->_idx, tag);
3358       n1 =      idom(n1);
3359       d1 = dom_depth(n1);
3360     } else if (d1 < d2) {
3361       // n2 is deeper than current lca
3362       Node *memo = _dom_lca_tags[n2->_idx];
3363       if( memo == tag ) {
3364         return n1;    // Return the current LCA
3365       }
3366       _dom_lca_tags.map(n2->_idx, tag);
3367       n2 =      idom(n2);
3368       d2 = dom_depth(n2);
3369     } else {
3370       // Here d1 == d2.  Due to edits of the dominator-tree, sections
3371       // of the tree might have the same depth.  These sections have
3372       // to be searched more carefully.
3373 
3374       // Scan up all the n1's with equal depth, looking for n2.
3375       _dom_lca_tags.map(n1->_idx, tag);
3376       Node *t1 = idom(n1);
3377       while (dom_depth(t1) == d1) {
3378         if (t1 == n2)  return n2;
3379         _dom_lca_tags.map(t1->_idx, tag);
3380         t1 = idom(t1);
3381       }
3382       // Scan up all the n2's with equal depth, looking for n1.
3383       _dom_lca_tags.map(n2->_idx, tag);
3384       Node *t2 = idom(n2);
3385       while (dom_depth(t2) == d2) {
3386         if (t2 == n1)  return n1;
3387         _dom_lca_tags.map(t2->_idx, tag);
3388         t2 = idom(t2);
3389       }
3390       // Move up to a new dominator-depth value as well as up the dom-tree.
3391       n1 = t1;
3392       n2 = t2;
3393       d1 = dom_depth(n1);
3394       d2 = dom_depth(n2);
3395     }
3396   } while (n1 != n2);
3397   return n1;
3398 }
3399 
3400 //------------------------------init_dom_lca_tags------------------------------
3401 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3402 // Intended use does not involve any growth for the array, so it could
3403 // be of fixed size.
3404 void PhaseIdealLoop::init_dom_lca_tags() {
3405   uint limit = C->unique() + 1;
3406   _dom_lca_tags.map( limit, NULL );
3407 #ifdef ASSERT
3408   for( uint i = 0; i < limit; ++i ) {
3409     assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3410   }
3411 #endif // ASSERT
3412 }
3413 
3414 //------------------------------clear_dom_lca_tags------------------------------
3415 // Tag could be a node's integer index, 32bits instead of 64bits in some cases
3416 // Intended use does not involve any growth for the array, so it could
3417 // be of fixed size.
3418 void PhaseIdealLoop::clear_dom_lca_tags() {
3419   uint limit = C->unique() + 1;
3420   _dom_lca_tags.map( limit, NULL );
3421   _dom_lca_tags.clear();
3422 #ifdef ASSERT
3423   for( uint i = 0; i < limit; ++i ) {
3424     assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer");
3425   }
3426 #endif // ASSERT
3427 }
3428 
3429 //------------------------------build_loop_late--------------------------------
3430 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3431 // Second pass finds latest legal placement, and ideal loop placement.
3432 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {
3433   while (worklist.size() != 0) {
3434     Node *n = worklist.pop();
3435     // Only visit once
3436     if (visited.test_set(n->_idx)) continue;
3437     uint cnt = n->outcnt();
3438     uint   i = 0;
3439     while (true) {
3440       assert( _nodes[n->_idx], "no dead nodes" );
3441       // Visit all children
3442       if (i < cnt) {
3443         Node* use = n->raw_out(i);
3444         ++i;
3445         // Check for dead uses.  Aggressively prune such junk.  It might be
3446         // dead in the global sense, but still have local uses so I cannot
3447         // easily call 'remove_dead_node'.
3448         if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead?
3449           // Due to cycles, we might not hit the same fixed point in the verify
3450           // pass as we do in the regular pass.  Instead, visit such phis as
3451           // simple uses of the loop head.
3452           if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {
3453             if( !visited.test(use->_idx) )
3454               worklist.push(use);
3455           } else if( !visited.test_set(use->_idx) ) {
3456             nstack.push(n, i); // Save parent and next use's index.
3457             n   = use;         // Process all children of current use.
3458             cnt = use->outcnt();
3459             i   = 0;
3460           }
3461         } else {
3462           // Do not visit around the backedge of loops via data edges.
3463           // push dead code onto a worklist
3464           _deadlist.push(use);
3465         }
3466       } else {
3467         // All of n's children have been processed, complete post-processing.
3468         build_loop_late_post(n);
3469         if (nstack.is_empty()) {
3470           // Finished all nodes on stack.
3471           // Process next node on the worklist.
3472           break;
3473         }
3474         // Get saved parent node and next use's index. Visit the rest of uses.
3475         n   = nstack.node();
3476         cnt = n->outcnt();
3477         i   = nstack.index();
3478         nstack.pop();
3479       }
3480     }
3481   }
3482 }
3483 
3484 //------------------------------build_loop_late_post---------------------------
3485 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping.
3486 // Second pass finds latest legal placement, and ideal loop placement.
3487 void PhaseIdealLoop::build_loop_late_post( Node *n ) {
3488 
3489   if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) {
3490     _igvn._worklist.push(n);  // Maybe we'll normalize it, if no more loops.
3491   }
3492 
3493 #ifdef ASSERT
3494   if (_verify_only && !n->is_CFG()) {
3495     // Check def-use domination.
3496     compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
3497   }
3498 #endif
3499 
3500   // CFG and pinned nodes already handled
3501   if( n->in(0) ) {
3502     if( n->in(0)->is_top() ) return; // Dead?
3503 
3504     // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
3505     // _must_ be pinned (they have to observe their control edge of course).
3506     // Unlike Stores (which modify an unallocable resource, the memory
3507     // state), Mods/Loads can float around.  So free them up.
3508     bool pinned = true;
3509     switch( n->Opcode() ) {
3510     case Op_DivI:
3511     case Op_DivF:
3512     case Op_DivD:
3513     case Op_ModI:
3514     case Op_ModF:
3515     case Op_ModD:
3516     case Op_LoadB:              // Same with Loads; they can sink
3517     case Op_LoadUB:             // during loop optimizations.
3518     case Op_LoadUS:
3519     case Op_LoadD:
3520     case Op_LoadF:
3521     case Op_LoadI:
3522     case Op_LoadKlass:
3523     case Op_LoadNKlass:
3524     case Op_LoadL:
3525     case Op_LoadS:
3526     case Op_LoadP:
3527     case Op_LoadN:
3528     case Op_LoadRange:
3529     case Op_LoadD_unaligned:
3530     case Op_LoadL_unaligned:
3531     case Op_StrComp:            // Does a bunch of load-like effects
3532     case Op_StrEquals:
3533     case Op_StrIndexOf:
3534     case Op_StrIndexOfChar:
3535     case Op_AryEq:
3536     case Op_HasNegatives:
3537       pinned = false;
3538     }
3539     if( pinned ) {
3540       IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
3541       if( !chosen_loop->_child )       // Inner loop?
3542         chosen_loop->_body.push(n); // Collect inner loops
3543       return;
3544     }
3545   } else {                      // No slot zero
3546     if( n->is_CFG() ) {         // CFG with no slot 0 is dead
3547       _nodes.map(n->_idx,0);    // No block setting, it's globally dead
3548       return;
3549     }
3550     assert(!n->is_CFG() || n->outcnt() == 0, "");
3551   }
3552 
3553   // Do I have a "safe range" I can select over?
3554   Node *early = get_ctrl(n);// Early location already computed
3555 
3556   // Compute latest point this Node can go
3557   Node *LCA = get_late_ctrl( n, early );
3558   // LCA is NULL due to uses being dead
3559   if( LCA == NULL ) {
3560 #ifdef ASSERT
3561     for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {
3562       assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead");
3563     }
3564 #endif
3565     _nodes.map(n->_idx, 0);     // This node is useless
3566     _deadlist.push(n);
3567     return;
3568   }
3569   assert(LCA != NULL && !LCA->is_top(), "no dead nodes");
3570 
3571   Node *legal = LCA;            // Walk 'legal' up the IDOM chain
3572   Node *least = legal;          // Best legal position so far
3573   while( early != legal ) {     // While not at earliest legal
3574 #ifdef ASSERT
3575     if (legal->is_Start() && !early->is_Root()) {
3576       // Bad graph. Print idom path and fail.
3577       dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);
3578       assert(false, "Bad graph detected in build_loop_late");
3579     }
3580 #endif
3581     // Find least loop nesting depth
3582     legal = idom(legal);        // Bump up the IDOM tree
3583     // Check for lower nesting depth
3584     if( get_loop(legal)->_nest < get_loop(least)->_nest )
3585       least = legal;
3586   }
3587   assert(early == legal || legal != C->root(), "bad dominance of inputs");
3588 
3589   // Try not to place code on a loop entry projection
3590   // which can inhibit range check elimination.
3591   if (least != early) {
3592     Node* ctrl_out = least->unique_ctrl_out();
3593     if (ctrl_out && ctrl_out->is_CountedLoop() &&
3594         least == ctrl_out->in(LoopNode::EntryControl)) {
3595       Node* least_dom = idom(least);
3596       if (get_loop(least_dom)->is_member(get_loop(least))) {
3597         least = least_dom;
3598       }
3599     }
3600   }
3601 
3602 #ifdef ASSERT
3603   // If verifying, verify that 'verify_me' has a legal location
3604   // and choose it as our location.
3605   if( _verify_me ) {
3606     Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
3607     Node *legal = LCA;
3608     while( early != legal ) {   // While not at earliest legal
3609       if( legal == v_ctrl ) break;  // Check for prior good location
3610       legal = idom(legal)      ;// Bump up the IDOM tree
3611     }
3612     // Check for prior good location
3613     if( legal == v_ctrl ) least = legal; // Keep prior if found
3614   }
3615 #endif
3616 
3617   // Assign discovered "here or above" point
3618   least = find_non_split_ctrl(least);
3619   set_ctrl(n, least);
3620 
3621   // Collect inner loop bodies
3622   IdealLoopTree *chosen_loop = get_loop(least);
3623   if( !chosen_loop->_child )   // Inner loop?
3624     chosen_loop->_body.push(n);// Collect inner loops
3625 }
3626 
3627 #ifdef ASSERT
3628 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {
3629   tty->print_cr("%s", msg);
3630   tty->print("n: "); n->dump();
3631   tty->print("early(n): "); early->dump();
3632   if (n->in(0) != NULL  && !n->in(0)->is_top() &&
3633       n->in(0) != early && !n->in(0)->is_Root()) {
3634     tty->print("n->in(0): "); n->in(0)->dump();
3635   }
3636   for (uint i = 1; i < n->req(); i++) {
3637     Node* in1 = n->in(i);
3638     if (in1 != NULL && in1 != n && !in1->is_top()) {
3639       tty->print("n->in(%d): ", i); in1->dump();
3640       Node* in1_early = get_ctrl(in1);
3641       tty->print("early(n->in(%d)): ", i); in1_early->dump();
3642       if (in1->in(0) != NULL     && !in1->in(0)->is_top() &&
3643           in1->in(0) != in1_early && !in1->in(0)->is_Root()) {
3644         tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();
3645       }
3646       for (uint j = 1; j < in1->req(); j++) {
3647         Node* in2 = in1->in(j);
3648         if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) {
3649           tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();
3650           Node* in2_early = get_ctrl(in2);
3651           tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();
3652           if (in2->in(0) != NULL     && !in2->in(0)->is_top() &&
3653               in2->in(0) != in2_early && !in2->in(0)->is_Root()) {
3654             tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();
3655           }
3656         }
3657       }
3658     }
3659   }
3660   tty->cr();
3661   tty->print("LCA(n): "); LCA->dump();
3662   for (uint i = 0; i < n->outcnt(); i++) {
3663     Node* u1 = n->raw_out(i);
3664     if (u1 == n)
3665       continue;
3666     tty->print("n->out(%d): ", i); u1->dump();
3667     if (u1->is_CFG()) {
3668       for (uint j = 0; j < u1->outcnt(); j++) {
3669         Node* u2 = u1->raw_out(j);
3670         if (u2 != u1 && u2 != n && u2->is_CFG()) {
3671           tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
3672         }
3673       }
3674     } else {
3675       Node* u1_later = get_ctrl(u1);
3676       tty->print("later(n->out(%d)): ", i); u1_later->dump();
3677       if (u1->in(0) != NULL     && !u1->in(0)->is_top() &&
3678           u1->in(0) != u1_later && !u1->in(0)->is_Root()) {
3679         tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();
3680       }
3681       for (uint j = 0; j < u1->outcnt(); j++) {
3682         Node* u2 = u1->raw_out(j);
3683         if (u2 == n || u2 == u1)
3684           continue;
3685         tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();
3686         if (!u2->is_CFG()) {
3687           Node* u2_later = get_ctrl(u2);
3688           tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();
3689           if (u2->in(0) != NULL     && !u2->in(0)->is_top() &&
3690               u2->in(0) != u2_later && !u2->in(0)->is_Root()) {
3691             tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();
3692           }
3693         }
3694       }
3695     }
3696   }
3697   tty->cr();
3698   int ct = 0;
3699   Node *dbg_legal = LCA;
3700   while(!dbg_legal->is_Start() && ct < 100) {
3701     tty->print("idom[%d] ",ct); dbg_legal->dump();
3702     ct++;
3703     dbg_legal = idom(dbg_legal);
3704   }
3705   tty->cr();
3706 }
3707 #endif
3708 
3709 #ifndef PRODUCT
3710 //------------------------------dump-------------------------------------------
3711 void PhaseIdealLoop::dump( ) const {
3712   ResourceMark rm;
3713   Arena* arena = Thread::current()->resource_area();
3714   Node_Stack stack(arena, C->live_nodes() >> 2);
3715   Node_List rpo_list;
3716   VectorSet visited(arena);
3717   visited.set(C->top()->_idx);
3718   rpo( C->root(), stack, visited, rpo_list );
3719   // Dump root loop indexed by last element in PO order
3720   dump( _ltree_root, rpo_list.size(), rpo_list );
3721 }
3722 
3723 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const {
3724   loop->dump_head();
3725 
3726   // Now scan for CFG nodes in the same loop
3727   for( uint j=idx; j > 0;  j-- ) {
3728     Node *n = rpo_list[j-1];
3729     if( !_nodes[n->_idx] )      // Skip dead nodes
3730       continue;
3731     if( get_loop(n) != loop ) { // Wrong loop nest
3732       if( get_loop(n)->_head == n &&    // Found nested loop?
3733           get_loop(n)->_parent == loop )
3734         dump(get_loop(n),rpo_list.size(),rpo_list);     // Print it nested-ly
3735       continue;
3736     }
3737 
3738     // Dump controlling node
3739     for( uint x = 0; x < loop->_nest; x++ )
3740       tty->print("  ");
3741     tty->print("C");
3742     if( n == C->root() ) {
3743       n->dump();
3744     } else {
3745       Node* cached_idom   = idom_no_update(n);
3746       Node *computed_idom = n->in(0);
3747       if( n->is_Region() ) {
3748         computed_idom = compute_idom(n);
3749         // computed_idom() will return n->in(0) when idom(n) is an IfNode (or
3750         // any MultiBranch ctrl node), so apply a similar transform to
3751         // the cached idom returned from idom_no_update.
3752         cached_idom = find_non_split_ctrl(cached_idom);
3753       }
3754       tty->print(" ID:%d",computed_idom->_idx);
3755       n->dump();
3756       if( cached_idom != computed_idom ) {
3757         tty->print_cr("*** BROKEN IDOM!  Computed as: %d, cached as: %d",
3758                       computed_idom->_idx, cached_idom->_idx);
3759       }
3760     }
3761     // Dump nodes it controls
3762     for( uint k = 0; k < _nodes.Size(); k++ ) {
3763       // (k < C->unique() && get_ctrl(find(k)) == n)
3764       if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) {
3765         Node *m = C->root()->find(k);
3766         if( m && m->outcnt() > 0 ) {
3767           if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {
3768             tty->print_cr("*** BROKEN CTRL ACCESSOR!  _nodes[k] is %p, ctrl is %p",
3769                           _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL);
3770           }
3771           for( uint j = 0; j < loop->_nest; j++ )
3772             tty->print("  ");
3773           tty->print(" ");
3774           m->dump();
3775         }
3776       }
3777     }
3778   }
3779 }
3780 
3781 // Collect a R-P-O for the whole CFG.
3782 // Result list is in post-order (scan backwards for RPO)
3783 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const {
3784   stk.push(start, 0);
3785   visited.set(start->_idx);
3786 
3787   while (stk.is_nonempty()) {
3788     Node* m   = stk.node();
3789     uint  idx = stk.index();
3790     if (idx < m->outcnt()) {
3791       stk.set_index(idx + 1);
3792       Node* n = m->raw_out(idx);
3793       if (n->is_CFG() && !visited.test_set(n->_idx)) {
3794         stk.push(n, 0);
3795       }
3796     } else {
3797       rpo_list.push(m);
3798       stk.pop();
3799     }
3800   }
3801 }
3802 #endif
3803 
3804 
3805 //=============================================================================
3806 //------------------------------LoopTreeIterator-----------------------------------
3807 
3808 // Advance to next loop tree using a preorder, left-to-right traversal.
3809 void LoopTreeIterator::next() {
3810   assert(!done(), "must not be done.");
3811   if (_curnt->_child != NULL) {
3812     _curnt = _curnt->_child;
3813   } else if (_curnt->_next != NULL) {
3814     _curnt = _curnt->_next;
3815   } else {
3816     while (_curnt != _root && _curnt->_next == NULL) {
3817       _curnt = _curnt->_parent;
3818     }
3819     if (_curnt == _root) {
3820       _curnt = NULL;
3821       assert(done(), "must be done.");
3822     } else {
3823       assert(_curnt->_next != NULL, "must be more to do");
3824       _curnt = _curnt->_next;
3825     }
3826   }
3827 }