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