1 /*
   2  * Copyright (c) 1999, 2017, 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 "memory/allocation.inline.hpp"
  27 #include "memory/resourceArea.hpp"
  28 #include "opto/addnode.hpp"
  29 #include "opto/castnode.hpp"
  30 #include "opto/connode.hpp"
  31 #include "opto/castnode.hpp"
  32 #include "opto/divnode.hpp"
  33 #include "opto/loopnode.hpp"
  34 #include "opto/matcher.hpp"
  35 #include "opto/mulnode.hpp"
  36 #include "opto/movenode.hpp"
  37 #include "opto/opaquenode.hpp"
  38 #include "opto/rootnode.hpp"
  39 #include "opto/subnode.hpp"
  40 
  41 //=============================================================================
  42 //------------------------------split_thru_phi---------------------------------
  43 // Split Node 'n' through merge point if there is enough win.
  44 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) {
  45   if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
  46     // ConvI2L may have type information on it which is unsafe to push up
  47     // so disable this for now
  48     return NULL;
  49   }
  50 
  51   // Splitting range check CastIIs through a loop induction Phi can
  52   // cause new Phis to be created that are left unrelated to the loop
  53   // induction Phi and prevent optimizations (vectorization)
  54   if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() &&
  55       region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) {
  56     return NULL;
  57   }
  58 
  59   int wins = 0;
  60   assert(!n->is_CFG(), "");
  61   assert(region->is_Region(), "");
  62 
  63   const Type* type = n->bottom_type();
  64   const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr();
  65   Node *phi;
  66   if (t_oop != NULL && t_oop->is_known_instance_field()) {
  67     int iid    = t_oop->instance_id();
  68     int index  = C->get_alias_index(t_oop);
  69     int offset = t_oop->offset();
  70     phi = new PhiNode(region, type, NULL, iid, index, offset);
  71   } else {
  72     phi = PhiNode::make_blank(region, n);
  73   }
  74   uint old_unique = C->unique();
  75   for (uint i = 1; i < region->req(); i++) {
  76     Node *x;
  77     Node* the_clone = NULL;
  78     if (region->in(i) == C->top()) {
  79       x = C->top();             // Dead path?  Use a dead data op
  80     } else {
  81       x = n->clone();           // Else clone up the data op
  82       the_clone = x;            // Remember for possible deletion.
  83       // Alter data node to use pre-phi inputs
  84       if (n->in(0) == region)
  85         x->set_req( 0, region->in(i) );
  86       for (uint j = 1; j < n->req(); j++) {
  87         Node *in = n->in(j);
  88         if (in->is_Phi() && in->in(0) == region)
  89           x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
  90       }
  91     }
  92     // Check for a 'win' on some paths
  93     const Type *t = x->Value(&_igvn);
  94 
  95     bool singleton = t->singleton();
  96 
  97     // A TOP singleton indicates that there are no possible values incoming
  98     // along a particular edge. In most cases, this is OK, and the Phi will
  99     // be eliminated later in an Ideal call. However, we can't allow this to
 100     // happen if the singleton occurs on loop entry, as the elimination of
 101     // the PhiNode may cause the resulting node to migrate back to a previous
 102     // loop iteration.
 103     if (singleton && t == Type::TOP) {
 104       // Is_Loop() == false does not confirm the absence of a loop (e.g., an
 105       // irreducible loop may not be indicated by an affirmative is_Loop());
 106       // therefore, the only top we can split thru a phi is on a backedge of
 107       // a loop.
 108       singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
 109     }
 110 
 111     if (singleton) {
 112       wins++;
 113       x = ((PhaseGVN&)_igvn).makecon(t);
 114     } else {
 115       // We now call Identity to try to simplify the cloned node.
 116       // Note that some Identity methods call phase->type(this).
 117       // Make sure that the type array is big enough for
 118       // our new node, even though we may throw the node away.
 119       // (Note: This tweaking with igvn only works because x is a new node.)
 120       _igvn.set_type(x, t);
 121       // If x is a TypeNode, capture any more-precise type permanently into Node
 122       // otherwise it will be not updated during igvn->transform since
 123       // igvn->type(x) is set to x->Value() already.
 124       x->raise_bottom_type(t);
 125       Node *y = x->Identity(&_igvn);
 126       if (y != x) {
 127         wins++;
 128         x = y;
 129       } else {
 130         y = _igvn.hash_find(x);
 131         if (y) {
 132           wins++;
 133           x = y;
 134         } else {
 135           // Else x is a new node we are keeping
 136           // We do not need register_new_node_with_optimizer
 137           // because set_type has already been called.
 138           _igvn._worklist.push(x);
 139         }
 140       }
 141     }
 142     if (x != the_clone && the_clone != NULL)
 143       _igvn.remove_dead_node(the_clone);
 144     phi->set_req( i, x );
 145   }
 146   // Too few wins?
 147   if (wins <= policy) {
 148     _igvn.remove_dead_node(phi);
 149     return NULL;
 150   }
 151 
 152   // Record Phi
 153   register_new_node( phi, region );
 154 
 155   for (uint i2 = 1; i2 < phi->req(); i2++) {
 156     Node *x = phi->in(i2);
 157     // If we commoned up the cloned 'x' with another existing Node,
 158     // the existing Node picks up a new use.  We need to make the
 159     // existing Node occur higher up so it dominates its uses.
 160     Node *old_ctrl;
 161     IdealLoopTree *old_loop;
 162 
 163     if (x->is_Con()) {
 164       // Constant's control is always root.
 165       set_ctrl(x, C->root());
 166       continue;
 167     }
 168     // The occasional new node
 169     if (x->_idx >= old_unique) {     // Found a new, unplaced node?
 170       old_ctrl = NULL;
 171       old_loop = NULL;               // Not in any prior loop
 172     } else {
 173       old_ctrl = get_ctrl(x);
 174       old_loop = get_loop(old_ctrl); // Get prior loop
 175     }
 176     // New late point must dominate new use
 177     Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
 178     if (new_ctrl == old_ctrl) // Nothing is changed
 179       continue;
 180 
 181     IdealLoopTree *new_loop = get_loop(new_ctrl);
 182 
 183     // Don't move x into a loop if its uses are
 184     // outside of loop. Otherwise x will be cloned
 185     // for each use outside of this loop.
 186     IdealLoopTree *use_loop = get_loop(region);
 187     if (!new_loop->is_member(use_loop) &&
 188         (old_loop == NULL || !new_loop->is_member(old_loop))) {
 189       // Take early control, later control will be recalculated
 190       // during next iteration of loop optimizations.
 191       new_ctrl = get_early_ctrl(x);
 192       new_loop = get_loop(new_ctrl);
 193     }
 194     // Set new location
 195     set_ctrl(x, new_ctrl);
 196     // If changing loop bodies, see if we need to collect into new body
 197     if (old_loop != new_loop) {
 198       if (old_loop && !old_loop->_child)
 199         old_loop->_body.yank(x);
 200       if (!new_loop->_child)
 201         new_loop->_body.push(x);  // Collect body info
 202     }
 203   }
 204 
 205   return phi;
 206 }
 207 
 208 //------------------------------dominated_by------------------------------------
 209 // Replace the dominated test with an obvious true or false.  Place it on the
 210 // IGVN worklist for later cleanup.  Move control-dependent data Nodes on the
 211 // live path up to the dominating control.
 212 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) {
 213   if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
 214 
 215   // prevdom is the dominating projection of the dominating test.
 216   assert( iff->is_If(), "" );
 217   assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added");
 218   int pop = prevdom->Opcode();
 219   assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
 220   if (flip) {
 221     if (pop == Op_IfTrue)
 222       pop = Op_IfFalse;
 223     else
 224       pop = Op_IfTrue;
 225   }
 226   // 'con' is set to true or false to kill the dominated test.
 227   Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
 228   set_ctrl(con, C->root()); // Constant gets a new use
 229   // Hack the dominated test
 230   _igvn.replace_input_of(iff, 1, con);
 231 
 232   // If I dont have a reachable TRUE and FALSE path following the IfNode then
 233   // I can assume this path reaches an infinite loop.  In this case it's not
 234   // important to optimize the data Nodes - either the whole compilation will
 235   // be tossed or this path (and all data Nodes) will go dead.
 236   if (iff->outcnt() != 2) return;
 237 
 238   // Make control-dependent data Nodes on the live path (path that will remain
 239   // once the dominated IF is removed) become control-dependent on the
 240   // dominating projection.
 241   Node* dp = iff->as_If()->proj_out(pop == Op_IfTrue);
 242 
 243   // Loop predicates may have depending checks which should not
 244   // be skipped. For example, range check predicate has two checks
 245   // for lower and upper bounds.
 246   if (dp == NULL)
 247     return;
 248 
 249   ProjNode* dp_proj  = dp->as_Proj();
 250   ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj();
 251   if (exclude_loop_predicate &&
 252       (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL ||
 253        unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) {
 254     // If this is a range check (IfNode::is_range_check), do not
 255     // reorder because Compile::allow_range_check_smearing might have
 256     // changed the check.
 257     return; // Let IGVN transformation change control dependence.
 258   }
 259 
 260   IdealLoopTree *old_loop = get_loop(dp);
 261 
 262   for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
 263     Node* cd = dp->fast_out(i); // Control-dependent node
 264     if (cd->depends_only_on_test()) {
 265       assert(cd->in(0) == dp, "");
 266       _igvn.replace_input_of(cd, 0, prevdom);
 267       set_early_ctrl(cd);
 268       IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
 269       if (old_loop != new_loop) {
 270         if (!old_loop->_child) old_loop->_body.yank(cd);
 271         if (!new_loop->_child) new_loop->_body.push(cd);
 272       }
 273       --i;
 274       --imax;
 275     }
 276   }
 277 }
 278 
 279 //------------------------------has_local_phi_input----------------------------
 280 // Return TRUE if 'n' has Phi inputs from its local block and no other
 281 // block-local inputs (all non-local-phi inputs come from earlier blocks)
 282 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
 283   Node *n_ctrl = get_ctrl(n);
 284   // See if some inputs come from a Phi in this block, or from before
 285   // this block.
 286   uint i;
 287   for( i = 1; i < n->req(); i++ ) {
 288     Node *phi = n->in(i);
 289     if( phi->is_Phi() && phi->in(0) == n_ctrl )
 290       break;
 291   }
 292   if( i >= n->req() )
 293     return NULL;                // No Phi inputs; nowhere to clone thru
 294 
 295   // Check for inputs created between 'n' and the Phi input.  These
 296   // must split as well; they have already been given the chance
 297   // (courtesy of a post-order visit) and since they did not we must
 298   // recover the 'cost' of splitting them by being very profitable
 299   // when splitting 'n'.  Since this is unlikely we simply give up.
 300   for( i = 1; i < n->req(); i++ ) {
 301     Node *m = n->in(i);
 302     if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
 303       // We allow the special case of AddP's with no local inputs.
 304       // This allows us to split-up address expressions.
 305       if (m->is_AddP() &&
 306           get_ctrl(m->in(2)) != n_ctrl &&
 307           get_ctrl(m->in(3)) != n_ctrl) {
 308         // Move the AddP up to dominating point
 309         set_ctrl_and_loop(m, find_non_split_ctrl(idom(n_ctrl)));
 310         continue;
 311       }
 312       return NULL;
 313     }
 314     assert(m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
 315   }
 316 
 317   return n_ctrl;
 318 }
 319 
 320 //------------------------------remix_address_expressions----------------------
 321 // Rework addressing expressions to get the most loop-invariant stuff
 322 // moved out.  We'd like to do all associative operators, but it's especially
 323 // important (common) to do address expressions.
 324 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
 325   if (!has_ctrl(n))  return NULL;
 326   Node *n_ctrl = get_ctrl(n);
 327   IdealLoopTree *n_loop = get_loop(n_ctrl);
 328 
 329   // See if 'n' mixes loop-varying and loop-invariant inputs and
 330   // itself is loop-varying.
 331 
 332   // Only interested in binary ops (and AddP)
 333   if( n->req() < 3 || n->req() > 4 ) return NULL;
 334 
 335   Node *n1_ctrl = get_ctrl(n->in(                    1));
 336   Node *n2_ctrl = get_ctrl(n->in(                    2));
 337   Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
 338   IdealLoopTree *n1_loop = get_loop( n1_ctrl );
 339   IdealLoopTree *n2_loop = get_loop( n2_ctrl );
 340   IdealLoopTree *n3_loop = get_loop( n3_ctrl );
 341 
 342   // Does one of my inputs spin in a tighter loop than self?
 343   if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
 344       (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
 345       (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
 346     return NULL;                // Leave well enough alone
 347 
 348   // Is at least one of my inputs loop-invariant?
 349   if( n1_loop == n_loop &&
 350       n2_loop == n_loop &&
 351       n3_loop == n_loop )
 352     return NULL;                // No loop-invariant inputs
 353 
 354 
 355   int n_op = n->Opcode();
 356 
 357   // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
 358   if( n_op == Op_LShiftI ) {
 359     // Scale is loop invariant
 360     Node *scale = n->in(2);
 361     Node *scale_ctrl = get_ctrl(scale);
 362     IdealLoopTree *scale_loop = get_loop(scale_ctrl );
 363     if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
 364       return NULL;
 365     const TypeInt *scale_t = scale->bottom_type()->isa_int();
 366     if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
 367       return NULL;              // Dont bother with byte/short masking
 368     // Add must vary with loop (else shift would be loop-invariant)
 369     Node *add = n->in(1);
 370     Node *add_ctrl = get_ctrl(add);
 371     IdealLoopTree *add_loop = get_loop(add_ctrl);
 372     //assert( n_loop == add_loop, "" );
 373     if( n_loop != add_loop ) return NULL;  // happens w/ evil ZKM loops
 374 
 375     // Convert I-V into I+ (0-V); same for V-I
 376     if( add->Opcode() == Op_SubI &&
 377         _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
 378       Node *zero = _igvn.intcon(0);
 379       set_ctrl(zero, C->root());
 380       Node *neg = new SubINode( _igvn.intcon(0), add->in(2) );
 381       register_new_node( neg, get_ctrl(add->in(2) ) );
 382       add = new AddINode( add->in(1), neg );
 383       register_new_node( add, add_ctrl );
 384     }
 385     if( add->Opcode() != Op_AddI ) return NULL;
 386     // See if one add input is loop invariant
 387     Node *add_var = add->in(1);
 388     Node *add_var_ctrl = get_ctrl(add_var);
 389     IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
 390     Node *add_invar = add->in(2);
 391     Node *add_invar_ctrl = get_ctrl(add_invar);
 392     IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
 393     if( add_var_loop == n_loop ) {
 394     } else if( add_invar_loop == n_loop ) {
 395       // Swap to find the invariant part
 396       add_invar = add_var;
 397       add_invar_ctrl = add_var_ctrl;
 398       add_invar_loop = add_var_loop;
 399       add_var = add->in(2);
 400       Node *add_var_ctrl = get_ctrl(add_var);
 401       IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
 402     } else                      // Else neither input is loop invariant
 403       return NULL;
 404     if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
 405       return NULL;              // No invariant part of the add?
 406 
 407     // Yes!  Reshape address expression!
 408     Node *inv_scale = new LShiftINode( add_invar, scale );
 409     Node *inv_scale_ctrl =
 410       dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
 411       add_invar_ctrl : scale_ctrl;
 412     register_new_node( inv_scale, inv_scale_ctrl );
 413     Node *var_scale = new LShiftINode( add_var, scale );
 414     register_new_node( var_scale, n_ctrl );
 415     Node *var_add = new AddINode( var_scale, inv_scale );
 416     register_new_node( var_add, n_ctrl );
 417     _igvn.replace_node( n, var_add );
 418     return var_add;
 419   }
 420 
 421   // Replace (I+V) with (V+I)
 422   if( n_op == Op_AddI ||
 423       n_op == Op_AddL ||
 424       n_op == Op_AddF ||
 425       n_op == Op_AddD ||
 426       n_op == Op_MulI ||
 427       n_op == Op_MulL ||
 428       n_op == Op_MulF ||
 429       n_op == Op_MulD ) {
 430     if( n2_loop == n_loop ) {
 431       assert( n1_loop != n_loop, "" );
 432       n->swap_edges(1, 2);
 433     }
 434   }
 435 
 436   // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
 437   // but not if I2 is a constant.
 438   if( n_op == Op_AddP ) {
 439     if( n2_loop == n_loop && n3_loop != n_loop ) {
 440       if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
 441         Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
 442         Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
 443         IdealLoopTree *n22loop = get_loop( n22_ctrl );
 444         IdealLoopTree *n23_loop = get_loop( n23_ctrl );
 445         if( n22loop != n_loop && n22loop->is_member(n_loop) &&
 446             n23_loop == n_loop ) {
 447           Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
 448           // Stuff new AddP in the loop preheader
 449           register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
 450           Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) );
 451           register_new_node( add2, n_ctrl );
 452           _igvn.replace_node( n, add2 );
 453           return add2;
 454         }
 455       }
 456     }
 457 
 458     // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
 459     if (n2_loop != n_loop && n3_loop == n_loop) {
 460       if (n->in(3)->Opcode() == Op_AddX) {
 461         Node *V = n->in(3)->in(1);
 462         Node *I = n->in(3)->in(2);
 463         if (is_member(n_loop,get_ctrl(V))) {
 464         } else {
 465           Node *tmp = V; V = I; I = tmp;
 466         }
 467         if (!is_member(n_loop,get_ctrl(I))) {
 468           Node *add1 = new AddPNode(n->in(1), n->in(2), I);
 469           // Stuff new AddP in the loop preheader
 470           register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl));
 471           Node *add2 = new AddPNode(n->in(1), add1, V);
 472           register_new_node(add2, n_ctrl);
 473           _igvn.replace_node(n, add2);
 474           return add2;
 475         }
 476       }
 477     }
 478   }
 479 
 480   return NULL;
 481 }
 482 
 483 //------------------------------conditional_move-------------------------------
 484 // Attempt to replace a Phi with a conditional move.  We have some pretty
 485 // strict profitability requirements.  All Phis at the merge point must
 486 // be converted, so we can remove the control flow.  We need to limit the
 487 // number of c-moves to a small handful.  All code that was in the side-arms
 488 // of the CFG diamond is now speculatively executed.  This code has to be
 489 // "cheap enough".  We are pretty much limited to CFG diamonds that merge
 490 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
 491 Node *PhaseIdealLoop::conditional_move( Node *region ) {
 492 
 493   assert(region->is_Region(), "sanity check");
 494   if (region->req() != 3) return NULL;
 495 
 496   // Check for CFG diamond
 497   Node *lp = region->in(1);
 498   Node *rp = region->in(2);
 499   if (!lp || !rp) return NULL;
 500   Node *lp_c = lp->in(0);
 501   if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL;
 502   IfNode *iff = lp_c->as_If();
 503 
 504   // Check for ops pinned in an arm of the diamond.
 505   // Can't remove the control flow in this case
 506   if (lp->outcnt() > 1) return NULL;
 507   if (rp->outcnt() > 1) return NULL;
 508 
 509   IdealLoopTree* r_loop = get_loop(region);
 510   assert(r_loop == get_loop(iff), "sanity");
 511   // Always convert to CMOVE if all results are used only outside this loop.
 512   bool used_inside_loop = (r_loop == _ltree_root);
 513 
 514   // Check profitability
 515   int cost = 0;
 516   int phis = 0;
 517   for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
 518     Node *out = region->fast_out(i);
 519     if (!out->is_Phi()) continue; // Ignore other control edges, etc
 520     phis++;
 521     PhiNode* phi = out->as_Phi();
 522     BasicType bt = phi->type()->basic_type();
 523     switch (bt) {
 524     case T_DOUBLE:
 525       if (C->use_cmove()) {
 526         continue; //TODO: maybe we want to add some cost
 527       }
 528     case T_FLOAT: {
 529       cost += Matcher::float_cmove_cost(); // Could be very expensive
 530       break;
 531     }
 532     case T_LONG: {
 533       cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
 534     }
 535     case T_INT:                 // These all CMOV fine
 536     case T_ADDRESS: {           // (RawPtr)
 537       cost++;
 538       break;
 539     }
 540     case T_NARROWOOP: // Fall through
 541     case T_OBJECT: {            // Base oops are OK, but not derived oops
 542       const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
 543       // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
 544       // CMOVE'd derived pointer?  It's a CMOVE'd derived base.  Thus
 545       // CMOVE'ing a derived pointer requires we also CMOVE the base.  If we
 546       // have a Phi for the base here that we convert to a CMOVE all is well
 547       // and good.  But if the base is dead, we'll not make a CMOVE.  Later
 548       // the allocator will have to produce a base by creating a CMOVE of the
 549       // relevant bases.  This puts the allocator in the business of
 550       // manufacturing expensive instructions, generally a bad plan.
 551       // Just Say No to Conditionally-Moved Derived Pointers.
 552       if (tp && tp->offset() != 0)
 553         return NULL;
 554       cost++;
 555       break;
 556     }
 557     default:
 558       return NULL;              // In particular, can't do memory or I/O
 559     }
 560     // Add in cost any speculative ops
 561     for (uint j = 1; j < region->req(); j++) {
 562       Node *proj = region->in(j);
 563       Node *inp = phi->in(j);
 564       if (get_ctrl(inp) == proj) { // Found local op
 565         cost++;
 566         // Check for a chain of dependent ops; these will all become
 567         // speculative in a CMOV.
 568         for (uint k = 1; k < inp->req(); k++)
 569           if (get_ctrl(inp->in(k)) == proj)
 570             cost += ConditionalMoveLimit; // Too much speculative goo
 571       }
 572     }
 573     // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
 574     // This will likely Split-If, a higher-payoff operation.
 575     for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
 576       Node* use = phi->fast_out(k);
 577       if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
 578         cost += ConditionalMoveLimit;
 579       // Is there a use inside the loop?
 580       // Note: check only basic types since CMoveP is pinned.
 581       if (!used_inside_loop && is_java_primitive(bt)) {
 582         IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
 583         if (r_loop == u_loop || r_loop->is_member(u_loop)) {
 584           used_inside_loop = true;
 585         }
 586       }
 587     }
 588   }//for
 589   Node* bol = iff->in(1);
 590   assert(bol->Opcode() == Op_Bool, "");
 591   int cmp_op = bol->in(1)->Opcode();
 592   // It is expensive to generate flags from a float compare.
 593   // Avoid duplicated float compare.
 594   if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
 595 
 596   float infrequent_prob = PROB_UNLIKELY_MAG(3);
 597   // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
 598   if (used_inside_loop) {
 599     if (cost >= ConditionalMoveLimit) return NULL; // Too much goo
 600 
 601     // BlockLayoutByFrequency optimization moves infrequent branch
 602     // from hot path. No point in CMOV'ing in such case (110 is used
 603     // instead of 100 to take into account not exactness of float value).
 604     if (BlockLayoutByFrequency) {
 605       infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
 606     }
 607   }
 608   // Check for highly predictable branch.  No point in CMOV'ing if
 609   // we are going to predict accurately all the time.
 610   if (C->use_cmove() && cmp_op == Op_CmpD) ;//keep going
 611   else if (iff->_prob < infrequent_prob ||
 612       iff->_prob > (1.0f - infrequent_prob))
 613     return NULL;
 614 
 615   // --------------
 616   // Now replace all Phis with CMOV's
 617   Node *cmov_ctrl = iff->in(0);
 618   uint flip = (lp->Opcode() == Op_IfTrue);
 619   Node_List wq;
 620   while (1) {
 621     PhiNode* phi = NULL;
 622     for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
 623       Node *out = region->fast_out(i);
 624       if (out->is_Phi()) {
 625         phi = out->as_Phi();
 626         break;
 627       }
 628     }
 629     if (phi == NULL)  break;
 630     if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); }
 631     // Move speculative ops
 632     wq.push(phi);
 633     while (wq.size() > 0) {
 634       Node *n = wq.pop();
 635       for (uint j = 1; j < n->req(); j++) {
 636         Node* m = n->in(j);
 637         if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) {
 638 #ifndef PRODUCT
 639           if (PrintOpto && VerifyLoopOptimizations) {
 640             tty->print("  speculate: ");
 641             m->dump();
 642           }
 643 #endif
 644           set_ctrl(m, cmov_ctrl);
 645           wq.push(m);
 646         }
 647       }
 648     }
 649     Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
 650     register_new_node( cmov, cmov_ctrl );
 651     _igvn.replace_node( phi, cmov );
 652 #ifndef PRODUCT
 653     if (TraceLoopOpts) {
 654       tty->print("CMOV  ");
 655       r_loop->dump_head();
 656       if (Verbose) {
 657         bol->in(1)->dump(1);
 658         cmov->dump(1);
 659       }
 660     }
 661     if (VerifyLoopOptimizations) verify();
 662 #endif
 663   }
 664 
 665   // The useless CFG diamond will fold up later; see the optimization in
 666   // RegionNode::Ideal.
 667   _igvn._worklist.push(region);
 668 
 669   return iff->in(1);
 670 }
 671 
 672 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
 673   for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
 674     Node* u = m->fast_out(i);
 675     if (u->is_CFG()) {
 676       if (u->Opcode() == Op_NeverBranch) {
 677         u = ((NeverBranchNode*)u)->proj_out(0);
 678         enqueue_cfg_uses(u, wq);
 679       } else {
 680         wq.push(u);
 681       }
 682     }
 683   }
 684 }
 685 
 686 // Try moving a store out of a loop, right before the loop
 687 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
 688   // Store has to be first in the loop body
 689   IdealLoopTree *n_loop = get_loop(n_ctrl);
 690   if (n->is_Store() && n_loop != _ltree_root && n_loop->is_loop() && n->in(0) != NULL) {
 691     Node* address = n->in(MemNode::Address);
 692     Node* value = n->in(MemNode::ValueIn);
 693     Node* mem = n->in(MemNode::Memory);
 694     IdealLoopTree* address_loop = get_loop(get_ctrl(address));
 695     IdealLoopTree* value_loop = get_loop(get_ctrl(value));
 696 
 697     // - address and value must be loop invariant
 698     // - memory must be a memory Phi for the loop
 699     // - Store must be the only store on this memory slice in the
 700     // loop: if there's another store following this one then value
 701     // written at iteration i by the second store could be overwritten
 702     // at iteration i+n by the first store: it's not safe to move the
 703     // first store out of the loop
 704     // - nothing must observe the memory Phi: it guarantees no read
 705     // before the store, we are also guaranteed the store post
 706     // dominates the loop head (ignoring a possible early
 707     // exit). Otherwise there would be extra Phi involved between the
 708     // loop's Phi and the store.
 709     // - there must be no early exit from the loop before the Store
 710     // (such an exit most of the time would be an extra use of the
 711     // memory Phi but sometimes is a bottom memory Phi that takes the
 712     // store as input).
 713 
 714     if (!n_loop->is_member(address_loop) &&
 715         !n_loop->is_member(value_loop) &&
 716         mem->is_Phi() && mem->in(0) == n_loop->_head &&
 717         mem->outcnt() == 1 &&
 718         mem->in(LoopNode::LoopBackControl) == n) {
 719 
 720       assert(n_loop->_tail != NULL, "need a tail");
 721       assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
 722 
 723       // Verify that there's no early exit of the loop before the store.
 724       bool ctrl_ok = false;
 725       {
 726         // Follow control from loop head until n, we exit the loop or
 727         // we reach the tail
 728         ResourceMark rm;
 729         Unique_Node_List wq;
 730         wq.push(n_loop->_head);
 731 
 732         for (uint next = 0; next < wq.size(); ++next) {
 733           Node *m = wq.at(next);
 734           if (m == n->in(0)) {
 735             ctrl_ok = true;
 736             continue;
 737           }
 738           assert(!has_ctrl(m), "should be CFG");
 739           if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
 740             ctrl_ok = false;
 741             break;
 742           }
 743           enqueue_cfg_uses(m, wq);
 744           if (wq.size() > 10) {
 745             ctrl_ok = false;
 746             break;
 747           }
 748         }
 749       }
 750       if (ctrl_ok) {
 751         // move the Store
 752         _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
 753         _igvn.replace_input_of(n, 0, n_loop->_head->in(LoopNode::EntryControl));
 754         _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
 755         // Disconnect the phi now. An empty phi can confuse other
 756         // optimizations in this pass of loop opts.
 757         _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
 758         n_loop->_body.yank(mem);
 759 
 760         IdealLoopTree* new_loop = get_loop(n->in(0));
 761         set_ctrl_and_loop(n, n->in(0));
 762 
 763         return n;
 764       }
 765     }
 766   }
 767   return NULL;
 768 }
 769 
 770 // Try moving a store out of a loop, right after the loop
 771 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
 772   if (n->is_Store() && n->in(0) != NULL) {
 773     Node *n_ctrl = get_ctrl(n);
 774     IdealLoopTree *n_loop = get_loop(n_ctrl);
 775     // Store must be in a loop
 776     if (n_loop != _ltree_root && !n_loop->_irreducible) {
 777       Node* address = n->in(MemNode::Address);
 778       Node* value = n->in(MemNode::ValueIn);
 779       IdealLoopTree* address_loop = get_loop(get_ctrl(address));
 780       // address must be loop invariant
 781       if (!n_loop->is_member(address_loop)) {
 782         // Store must be last on this memory slice in the loop and
 783         // nothing in the loop must observe it
 784         Node* phi = NULL;
 785         for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 786           Node* u = n->fast_out(i);
 787           if (has_ctrl(u)) { // control use?
 788             IdealLoopTree *u_loop = get_loop(get_ctrl(u));
 789             if (!n_loop->is_member(u_loop)) {
 790               continue;
 791             }
 792             if (u->is_Phi() && u->in(0) == n_loop->_head) {
 793               assert(_igvn.type(u) == Type::MEMORY, "bad phi");
 794               // multiple phis on the same slice are possible
 795               if (phi != NULL) {
 796                 return;
 797               }
 798               phi = u;
 799               continue;
 800             }
 801           }
 802           return;
 803         }
 804         if (phi != NULL) {
 805           // Nothing in the loop before the store (next iteration)
 806           // must observe the stored value
 807           bool mem_ok = true;
 808           {
 809             ResourceMark rm;
 810             Unique_Node_List wq;
 811             wq.push(phi);
 812             for (uint next = 0; next < wq.size() && mem_ok; ++next) {
 813               Node *m = wq.at(next);
 814               for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
 815                 Node* u = m->fast_out(i);
 816                 if (u->is_Store() || u->is_Phi()) {
 817                   if (u != n) {
 818                     wq.push(u);
 819                     mem_ok = (wq.size() <= 10);
 820                   }
 821                 } else {
 822                   mem_ok = false;
 823                   break;
 824                 }
 825               }
 826             }
 827           }
 828           if (mem_ok) {
 829             // Move the Store out of the loop creating clones along
 830             // all paths out of the loop that observe the stored value
 831             _igvn.rehash_node_delayed(phi);
 832             int count = phi->replace_edge(n, n->in(MemNode::Memory));
 833             assert(count > 0, "inconsistent phi");
 834             for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 835               Node* u = n->fast_out(i);
 836               Node* c = get_ctrl(u);
 837 
 838               if (u->is_Phi()) {
 839                 c = u->in(0)->in(u->find_edge(n));
 840               }
 841               IdealLoopTree *u_loop = get_loop(c);
 842               assert (!n_loop->is_member(u_loop), "only the phi should have been a use in the loop");
 843               while(true) {
 844                 Node* next_c = find_non_split_ctrl(idom(c));
 845                 if (n_loop->is_member(get_loop(next_c))) {
 846                   break;
 847                 }
 848                 c = next_c;
 849               }
 850 
 851               Node* st = n->clone();
 852               st->set_req(0, c);
 853               _igvn.register_new_node_with_optimizer(st);
 854 
 855               set_ctrl(st, c);
 856               IdealLoopTree* new_loop = get_loop(c);
 857               assert(new_loop != n_loop, "should be moved out of loop");
 858               if (new_loop->_child == NULL) new_loop->_body.push(st);
 859 
 860               _igvn.replace_input_of(u, u->find_edge(n), st);
 861               --imax;
 862               --i;
 863             }
 864 
 865 
 866             assert(n->outcnt() == 0, "all uses should be gone");
 867             _igvn.replace_input_of(n, MemNode::Memory, C->top());
 868             // Disconnect the phi now. An empty phi can confuse other
 869             // optimizations in this pass of loop opts..
 870             if (phi->in(LoopNode::LoopBackControl) == phi) {
 871               _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
 872               n_loop->_body.yank(phi);
 873             }
 874           }
 875         }
 876       }
 877     }
 878   }
 879 }
 880 
 881 //------------------------------split_if_with_blocks_pre-----------------------
 882 // Do the real work in a non-recursive function.  Data nodes want to be
 883 // cloned in the pre-order so they can feed each other nicely.
 884 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
 885   // Cloning these guys is unlikely to win
 886   int n_op = n->Opcode();
 887   if( n_op == Op_MergeMem ) return n;
 888   if( n->is_Proj() ) return n;
 889   // Do not clone-up CmpFXXX variations, as these are always
 890   // followed by a CmpI
 891   if( n->is_Cmp() ) return n;
 892   // Attempt to use a conditional move instead of a phi/branch
 893   if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
 894     Node *cmov = conditional_move( n );
 895     if( cmov ) return cmov;
 896   }
 897   if( n->is_CFG() || n->is_LoadStore() )
 898     return n;
 899   if( n_op == Op_Opaque1 ||     // Opaque nodes cannot be mod'd
 900       n_op == Op_Opaque2 ) {
 901     if( !C->major_progress() )   // If chance of no more loop opts...
 902       _igvn._worklist.push(n);  // maybe we'll remove them
 903     return n;
 904   }
 905 
 906   if( n->is_Con() ) return n;   // No cloning for Con nodes
 907 
 908   Node *n_ctrl = get_ctrl(n);
 909   if( !n_ctrl ) return n;       // Dead node
 910 
 911   Node* res = try_move_store_before_loop(n, n_ctrl);
 912   if (res != NULL) {
 913     return n;
 914   }
 915 
 916   // Attempt to remix address expressions for loop invariants
 917   Node *m = remix_address_expressions( n );
 918   if( m ) return m;
 919 
 920   if (n->is_ConstraintCast()) {
 921     Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
 922     // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
 923     // Node control inputs don't necessarily agree with loop control info (due to
 924     // transformations happened in between), thus additional dominance check is needed
 925     // to keep loop info valid.
 926     if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
 927       _igvn.replace_node(n, dom_cast);
 928       return dom_cast;
 929     }
 930   }
 931 
 932   // Determine if the Node has inputs from some local Phi.
 933   // Returns the block to clone thru.
 934   Node *n_blk = has_local_phi_input( n );
 935   if( !n_blk ) return n;
 936 
 937   // Do not clone the trip counter through on a CountedLoop
 938   // (messes up the canonical shape).
 939   if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
 940 
 941   // Check for having no control input; not pinned.  Allow
 942   // dominating control.
 943   if (n->in(0)) {
 944     Node *dom = idom(n_blk);
 945     if (dom_lca(n->in(0), dom) != n->in(0)) {
 946       return n;
 947     }
 948   }
 949   // Policy: when is it profitable.  You must get more wins than
 950   // policy before it is considered profitable.  Policy is usually 0,
 951   // so 1 win is considered profitable.  Big merges will require big
 952   // cloning, so get a larger policy.
 953   int policy = n_blk->req() >> 2;
 954 
 955   // If the loop is a candidate for range check elimination,
 956   // delay splitting through it's phi until a later loop optimization
 957   if (n_blk->is_CountedLoop()) {
 958     IdealLoopTree *lp = get_loop(n_blk);
 959     if (lp && lp->_rce_candidate) {
 960       return n;
 961     }
 962   }
 963 
 964   // Use same limit as split_if_with_blocks_post
 965   if( C->live_nodes() > 35000 ) return n; // Method too big
 966 
 967   // Split 'n' through the merge point if it is profitable
 968   Node *phi = split_thru_phi( n, n_blk, policy );
 969   if (!phi) return n;
 970 
 971   // Found a Phi to split thru!
 972   // Replace 'n' with the new phi
 973   _igvn.replace_node( n, phi );
 974   // Moved a load around the loop, 'en-registering' something.
 975   if (n_blk->is_Loop() && n->is_Load() &&
 976       !phi->in(LoopNode::LoopBackControl)->is_Load())
 977     C->set_major_progress();
 978 
 979   return phi;
 980 }
 981 
 982 static bool merge_point_too_heavy(Compile* C, Node* region) {
 983   // Bail out if the region and its phis have too many users.
 984   int weight = 0;
 985   for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
 986     weight += region->fast_out(i)->outcnt();
 987   }
 988   int nodes_left = C->max_node_limit() - C->live_nodes();
 989   if (weight * 8 > nodes_left) {
 990     if (PrintOpto) {
 991       tty->print_cr("*** Split-if bails out:  %d nodes, region weight %d", C->unique(), weight);
 992     }
 993     return true;
 994   } else {
 995     return false;
 996   }
 997 }
 998 
 999 static bool merge_point_safe(Node* region) {
1000   // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1001   // having a PhiNode input. This sidesteps the dangerous case where the split
1002   // ConvI2LNode may become TOP if the input Value() does not
1003   // overlap the ConvI2L range, leaving a node which may not dominate its
1004   // uses.
1005   // A better fix for this problem can be found in the BugTraq entry, but
1006   // expediency for Mantis demands this hack.
1007   // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop
1008   // split_if_with_blocks from splitting a block because we could not move around
1009   // the FastLockNode.
1010   for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1011     Node* n = region->fast_out(i);
1012     if (n->is_Phi()) {
1013       for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1014         Node* m = n->fast_out(j);
1015         if (m->is_FastLock())
1016           return false;
1017 #ifdef _LP64
1018         if (m->Opcode() == Op_ConvI2L)
1019           return false;
1020         if (m->is_CastII() && m->isa_CastII()->has_range_check()) {
1021           return false;
1022         }
1023 #endif
1024       }
1025     }
1026   }
1027   return true;
1028 }
1029 
1030 
1031 //------------------------------place_near_use---------------------------------
1032 // Place some computation next to use but not inside inner loops.
1033 // For inner loop uses move it to the preheader area.
1034 Node *PhaseIdealLoop::place_near_use( Node *useblock ) const {
1035   IdealLoopTree *u_loop = get_loop( useblock );
1036   return (u_loop->_irreducible || u_loop->_child)
1037     ? useblock
1038     : u_loop->_head->in(LoopNode::EntryControl);
1039 }
1040 
1041 
1042 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1043   if (!n->is_If()) {
1044     return false;
1045   }
1046   if (!n->in(0)->is_Region()) {
1047     return false;
1048   }
1049   Node* region = n->in(0);
1050   Node* dom = idom(region);
1051   if (!dom->is_If() || dom->in(1) != n->in(1)) {
1052     return false;
1053   }
1054   IfNode* dom_if = dom->as_If();
1055   Node* proj_true = dom_if->proj_out(1);
1056   Node* proj_false = dom_if->proj_out(0);
1057 
1058   for (uint i = 1; i < region->req(); i++) {
1059     if (is_dominator(proj_true, region->in(i))) {
1060       continue;
1061     }
1062     if (is_dominator(proj_false, region->in(i))) {
1063       continue;
1064     }
1065     return false;
1066   }
1067 
1068   return true;
1069 }
1070 
1071 bool PhaseIdealLoop::can_split_if(Node *n_ctrl) {
1072   if (C->live_nodes() > 35000) {
1073     return false; // Method too big
1074   }
1075 
1076   // Do not do 'split-if' if irreducible loops are present.
1077   if (_has_irreducible_loops) {
1078     return false;
1079   }
1080 
1081   if (merge_point_too_heavy(C, n_ctrl)) {
1082     return false;
1083   }
1084 
1085   // Do not do 'split-if' if some paths are dead.  First do dead code
1086   // elimination and then see if its still profitable.
1087   for (uint i = 1; i < n_ctrl->req(); i++) {
1088     if (n_ctrl->in(i) == C->top()) {
1089       return false;
1090     }
1091   }
1092 
1093   // If trying to do a 'Split-If' at the loop head, it is only
1094   // profitable if the cmp folds up on BOTH paths.  Otherwise we
1095   // risk peeling a loop forever.
1096 
1097   // CNC - Disabled for now.  Requires careful handling of loop
1098   // body selection for the cloned code.  Also, make sure we check
1099   // for any input path not being in the same loop as n_ctrl.  For
1100   // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1101   // because the alternative loop entry points won't be converted
1102   // into LoopNodes.
1103   IdealLoopTree *n_loop = get_loop(n_ctrl);
1104   for (uint j = 1; j < n_ctrl->req(); j++) {
1105     if (get_loop(n_ctrl->in(j)) != n_loop) {
1106       return false;
1107     }
1108   }
1109 
1110   // Check for safety of the merge point.
1111   if (!merge_point_safe(n_ctrl)) {
1112     return false;
1113   }
1114 
1115   return true;
1116 }
1117 
1118 //------------------------------split_if_with_blocks_post----------------------
1119 // Do the real work in a non-recursive function.  CFG hackery wants to be
1120 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1121 // info.
1122 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1123 
1124   // Cloning Cmp through Phi's involves the split-if transform.
1125   // FastLock is not used by an If
1126   if (n->is_Cmp() && !n->is_FastLock()) {
1127     Node *n_ctrl = get_ctrl(n);
1128     // Determine if the Node has inputs from some local Phi.
1129     // Returns the block to clone thru.
1130     Node *n_blk = has_local_phi_input(n);
1131     if (n_blk != n_ctrl) {
1132       return;
1133     }
1134 
1135     if (!can_split_if(n_ctrl)) {
1136       return;
1137     }
1138 
1139     if (n->outcnt() != 1) {
1140       return; // Multiple bool's from 1 compare?
1141     }
1142     Node *bol = n->unique_out();
1143     assert(bol->is_Bool(), "expect a bool here");
1144     if (bol->outcnt() != 1) {
1145       return;// Multiple branches from 1 compare?
1146     }
1147     Node *iff = bol->unique_out();
1148 
1149     // Check some safety conditions
1150     if (iff->is_If()) {        // Classic split-if?
1151       if (iff->in(0) != n_ctrl) {
1152         return; // Compare must be in same blk as if
1153       }
1154     } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1155       // Can't split CMove with different control edge.
1156       if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) {
1157         return;
1158       }
1159       if (get_ctrl(iff->in(2)) == n_ctrl ||
1160           get_ctrl(iff->in(3)) == n_ctrl) {
1161         return;                 // Inputs not yet split-up
1162       }
1163       if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1164         return;                 // Loop-invar test gates loop-varying CMOVE
1165       }
1166     } else {
1167       return;  // some other kind of node, such as an Allocate
1168     }
1169 
1170     // When is split-if profitable?  Every 'win' on means some control flow
1171     // goes dead, so it's almost always a win.
1172     int policy = 0;
1173     // Split compare 'n' through the merge point if it is profitable
1174     Node *phi = split_thru_phi( n, n_ctrl, policy);
1175     if (!phi) {
1176       return;
1177     }
1178 
1179     // Found a Phi to split thru!
1180     // Replace 'n' with the new phi
1181     _igvn.replace_node(n, phi);
1182 
1183     // Now split the bool up thru the phi
1184     Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1185     guarantee(bolphi != NULL, "null boolean phi node");
1186 
1187     _igvn.replace_node(bol, bolphi);
1188     assert(iff->in(1) == bolphi, "");
1189 
1190     if (bolphi->Value(&_igvn)->singleton()) {
1191       return;
1192     }
1193 
1194     // Conditional-move?  Must split up now
1195     if (!iff->is_If()) {
1196       Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1197       _igvn.replace_node(iff, cmovphi);
1198       return;
1199     }
1200 
1201     // Now split the IF
1202     do_split_if(iff);
1203     return;
1204   }
1205 
1206   // Two identical ifs back to back can be merged
1207   if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1208     Node *n_ctrl = n->in(0);
1209     PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1210     IfNode* dom_if = idom(n_ctrl)->as_If();
1211     Node* proj_true = dom_if->proj_out(1);
1212     Node* proj_false = dom_if->proj_out(0);
1213     Node* con_true = _igvn.makecon(TypeInt::ONE);
1214     Node* con_false = _igvn.makecon(TypeInt::ZERO);
1215 
1216     for (uint i = 1; i < n_ctrl->req(); i++) {
1217       if (is_dominator(proj_true, n_ctrl->in(i))) {
1218         bolphi->init_req(i, con_true);
1219       } else {
1220         assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1221         bolphi->init_req(i, con_false);
1222       }
1223     }
1224     register_new_node(bolphi, n_ctrl);
1225     _igvn.replace_input_of(n, 1, bolphi);
1226 
1227     // Now split the IF
1228     do_split_if(n);
1229     return;
1230   }
1231 
1232   // Check for an IF ready to split; one that has its
1233   // condition codes input coming from a Phi at the block start.
1234   int n_op = n->Opcode();
1235 
1236   // Check for an IF being dominated by another IF same test
1237   if (n_op == Op_If ||
1238       n_op == Op_RangeCheck) {
1239     Node *bol = n->in(1);
1240     uint max = bol->outcnt();
1241     // Check for same test used more than once?
1242     if (max > 1 && bol->is_Bool()) {
1243       // Search up IDOMs to see if this IF is dominated.
1244       Node *cutoff = get_ctrl(bol);
1245 
1246       // Now search up IDOMs till cutoff, looking for a dominating test
1247       Node *prevdom = n;
1248       Node *dom = idom(prevdom);
1249       while (dom != cutoff) {
1250         if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) {
1251           // Replace the dominated test with an obvious true or false.
1252           // Place it on the IGVN worklist for later cleanup.
1253           C->set_major_progress();
1254           dominated_by(prevdom, n, false, true);
1255 #ifndef PRODUCT
1256           if( VerifyLoopOptimizations ) verify();
1257 #endif
1258           return;
1259         }
1260         prevdom = dom;
1261         dom = idom(prevdom);
1262       }
1263     }
1264   }
1265 
1266   // See if a shared loop-varying computation has no loop-varying uses.
1267   // Happens if something is only used for JVM state in uncommon trap exits,
1268   // like various versions of induction variable+offset.  Clone the
1269   // computation per usage to allow it to sink out of the loop.
1270   if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
1271     Node *n_ctrl = get_ctrl(n);
1272     IdealLoopTree *n_loop = get_loop(n_ctrl);
1273     if( n_loop != _ltree_root ) {
1274       DUIterator_Fast imax, i = n->fast_outs(imax);
1275       for (; i < imax; i++) {
1276         Node* u = n->fast_out(i);
1277         if( !has_ctrl(u) )     break; // Found control user
1278         IdealLoopTree *u_loop = get_loop(get_ctrl(u));
1279         if( u_loop == n_loop ) break; // Found loop-varying use
1280         if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
1281         if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
1282       }
1283       bool did_break = (i < imax);  // Did we break out of the previous loop?
1284       if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
1285         Node *late_load_ctrl = NULL;
1286         if (n->is_Load()) {
1287           // If n is a load, get and save the result from get_late_ctrl(),
1288           // to be later used in calculating the control for n's clones.
1289           clear_dom_lca_tags();
1290           late_load_ctrl = get_late_ctrl(n, n_ctrl);
1291         }
1292         // If n is a load, and the late control is the same as the current
1293         // control, then the cloning of n is a pointless exercise, because
1294         // GVN will ensure that we end up where we started.
1295         if (!n->is_Load() || late_load_ctrl != n_ctrl) {
1296           for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
1297             Node *u = n->last_out(j); // Clone private computation per use
1298             _igvn.rehash_node_delayed(u);
1299             Node *x = n->clone(); // Clone computation
1300             Node *x_ctrl = NULL;
1301             if( u->is_Phi() ) {
1302               // Replace all uses of normal nodes.  Replace Phi uses
1303               // individually, so the separate Nodes can sink down
1304               // different paths.
1305               uint k = 1;
1306               while( u->in(k) != n ) k++;
1307               u->set_req( k, x );
1308               // x goes next to Phi input path
1309               x_ctrl = u->in(0)->in(k);
1310               --j;
1311             } else {              // Normal use
1312               // Replace all uses
1313               for( uint k = 0; k < u->req(); k++ ) {
1314                 if( u->in(k) == n ) {
1315                   u->set_req( k, x );
1316                   --j;
1317                 }
1318               }
1319               x_ctrl = get_ctrl(u);
1320             }
1321 
1322             // Find control for 'x' next to use but not inside inner loops.
1323             // For inner loop uses get the preheader area.
1324             x_ctrl = place_near_use(x_ctrl);
1325 
1326             if (n->is_Load()) {
1327               // For loads, add a control edge to a CFG node outside of the loop
1328               // to force them to not combine and return back inside the loop
1329               // during GVN optimization (4641526).
1330               //
1331               // Because we are setting the actual control input, factor in
1332               // the result from get_late_ctrl() so we respect any
1333               // anti-dependences. (6233005).
1334               x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
1335 
1336               // Don't allow the control input to be a CFG splitting node.
1337               // Such nodes should only have ProjNodes as outs, e.g. IfNode
1338               // should only have IfTrueNode and IfFalseNode (4985384).
1339               x_ctrl = find_non_split_ctrl(x_ctrl);
1340               assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1341 
1342               x->set_req(0, x_ctrl);
1343             }
1344             register_new_node(x, x_ctrl);
1345 
1346             // Some institutional knowledge is needed here: 'x' is
1347             // yanked because if the optimizer runs GVN on it all the
1348             // cloned x's will common up and undo this optimization and
1349             // be forced back in the loop.  This is annoying because it
1350             // makes +VerifyOpto report false-positives on progress.  I
1351             // tried setting control edges on the x's to force them to
1352             // not combine, but the matching gets worried when it tries
1353             // to fold a StoreP and an AddP together (as part of an
1354             // address expression) and the AddP and StoreP have
1355             // different controls.
1356             if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x);
1357           }
1358           _igvn.remove_dead_node(n);
1359         }
1360       }
1361     }
1362   }
1363 
1364   try_move_store_after_loop(n);
1365 
1366   // Check for Opaque2's who's loop has disappeared - who's input is in the
1367   // same loop nest as their output.  Remove 'em, they are no longer useful.
1368   if( n_op == Op_Opaque2 &&
1369       n->in(1) != NULL &&
1370       get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1371     _igvn.replace_node( n, n->in(1) );
1372   }
1373 }
1374 
1375 //------------------------------split_if_with_blocks---------------------------
1376 // Check for aggressive application of 'split-if' optimization,
1377 // using basic block level info.
1378 void PhaseIdealLoop::split_if_with_blocks( VectorSet &visited, Node_Stack &nstack ) {
1379   Node *n = C->root();
1380   visited.set(n->_idx); // first, mark node as visited
1381   // Do pre-visit work for root
1382   n = split_if_with_blocks_pre( n );
1383   uint cnt = n->outcnt();
1384   uint i   = 0;
1385   while (true) {
1386     // Visit all children
1387     if (i < cnt) {
1388       Node* use = n->raw_out(i);
1389       ++i;
1390       if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1391         // Now do pre-visit work for this use
1392         use = split_if_with_blocks_pre( use );
1393         nstack.push(n, i); // Save parent and next use's index.
1394         n   = use;         // Process all children of current use.
1395         cnt = use->outcnt();
1396         i   = 0;
1397       }
1398     }
1399     else {
1400       // All of n's children have been processed, complete post-processing.
1401       if (cnt != 0 && !n->is_Con()) {
1402         assert(has_node(n), "no dead nodes");
1403         split_if_with_blocks_post( n );
1404       }
1405       if (nstack.is_empty()) {
1406         // Finished all nodes on stack.
1407         break;
1408       }
1409       // Get saved parent node and next use's index. Visit the rest of uses.
1410       n   = nstack.node();
1411       cnt = n->outcnt();
1412       i   = nstack.index();
1413       nstack.pop();
1414     }
1415   }
1416 }
1417 
1418 
1419 //=============================================================================
1420 //
1421 //                   C L O N E   A   L O O P   B O D Y
1422 //
1423 
1424 //------------------------------clone_iff--------------------------------------
1425 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1426 // "Nearly" because all Nodes have been cloned from the original in the loop,
1427 // but the fall-in edges to the Cmp are different.  Clone bool/Cmp pairs
1428 // through the Phi recursively, and return a Bool.
1429 BoolNode *PhaseIdealLoop::clone_iff( PhiNode *phi, IdealLoopTree *loop ) {
1430 
1431   // Convert this Phi into a Phi merging Bools
1432   uint i;
1433   for( i = 1; i < phi->req(); i++ ) {
1434     Node *b = phi->in(i);
1435     if( b->is_Phi() ) {
1436       _igvn.replace_input_of(phi, i, clone_iff( b->as_Phi(), loop ));
1437     } else {
1438       assert( b->is_Bool(), "" );
1439     }
1440   }
1441 
1442   Node *sample_bool = phi->in(1);
1443   Node *sample_cmp  = sample_bool->in(1);
1444 
1445   // Make Phis to merge the Cmp's inputs.
1446   PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
1447   PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
1448   for( i = 1; i < phi->req(); i++ ) {
1449     Node *n1 = phi->in(i)->in(1)->in(1);
1450     Node *n2 = phi->in(i)->in(1)->in(2);
1451     phi1->set_req( i, n1 );
1452     phi2->set_req( i, n2 );
1453     phi1->set_type( phi1->type()->meet_speculative(n1->bottom_type()));
1454     phi2->set_type( phi2->type()->meet_speculative(n2->bottom_type()));
1455   }
1456   // See if these Phis have been made before.
1457   // Register with optimizer
1458   Node *hit1 = _igvn.hash_find_insert(phi1);
1459   if( hit1 ) {                  // Hit, toss just made Phi
1460     _igvn.remove_dead_node(phi1); // Remove new phi
1461     assert( hit1->is_Phi(), "" );
1462     phi1 = (PhiNode*)hit1;      // Use existing phi
1463   } else {                      // Miss
1464     _igvn.register_new_node_with_optimizer(phi1);
1465   }
1466   Node *hit2 = _igvn.hash_find_insert(phi2);
1467   if( hit2 ) {                  // Hit, toss just made Phi
1468     _igvn.remove_dead_node(phi2); // Remove new phi
1469     assert( hit2->is_Phi(), "" );
1470     phi2 = (PhiNode*)hit2;      // Use existing phi
1471   } else {                      // Miss
1472     _igvn.register_new_node_with_optimizer(phi2);
1473   }
1474   // Register Phis with loop/block info
1475   set_ctrl(phi1, phi->in(0));
1476   set_ctrl(phi2, phi->in(0));
1477   // Make a new Cmp
1478   Node *cmp = sample_cmp->clone();
1479   cmp->set_req( 1, phi1 );
1480   cmp->set_req( 2, phi2 );
1481   _igvn.register_new_node_with_optimizer(cmp);
1482   set_ctrl(cmp, phi->in(0));
1483 
1484   // Make a new Bool
1485   Node *b = sample_bool->clone();
1486   b->set_req(1,cmp);
1487   _igvn.register_new_node_with_optimizer(b);
1488   set_ctrl(b, phi->in(0));
1489 
1490   assert( b->is_Bool(), "" );
1491   return (BoolNode*)b;
1492 }
1493 
1494 //------------------------------clone_bool-------------------------------------
1495 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1496 // "Nearly" because all Nodes have been cloned from the original in the loop,
1497 // but the fall-in edges to the Cmp are different.  Clone bool/Cmp pairs
1498 // through the Phi recursively, and return a Bool.
1499 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1500   uint i;
1501   // Convert this Phi into a Phi merging Bools
1502   for( i = 1; i < phi->req(); i++ ) {
1503     Node *b = phi->in(i);
1504     if( b->is_Phi() ) {
1505       _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop ));
1506     } else {
1507       assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1508     }
1509   }
1510 
1511   Node *sample_cmp = phi->in(1);
1512 
1513   // Make Phis to merge the Cmp's inputs.
1514   PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
1515   PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
1516   for( uint j = 1; j < phi->req(); j++ ) {
1517     Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1518     Node *n1, *n2;
1519     if( cmp_top->is_Cmp() ) {
1520       n1 = cmp_top->in(1);
1521       n2 = cmp_top->in(2);
1522     } else {
1523       n1 = n2 = cmp_top;
1524     }
1525     phi1->set_req( j, n1 );
1526     phi2->set_req( j, n2 );
1527     phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1528     phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1529   }
1530 
1531   // See if these Phis have been made before.
1532   // Register with optimizer
1533   Node *hit1 = _igvn.hash_find_insert(phi1);
1534   if( hit1 ) {                  // Hit, toss just made Phi
1535     _igvn.remove_dead_node(phi1); // Remove new phi
1536     assert( hit1->is_Phi(), "" );
1537     phi1 = (PhiNode*)hit1;      // Use existing phi
1538   } else {                      // Miss
1539     _igvn.register_new_node_with_optimizer(phi1);
1540   }
1541   Node *hit2 = _igvn.hash_find_insert(phi2);
1542   if( hit2 ) {                  // Hit, toss just made Phi
1543     _igvn.remove_dead_node(phi2); // Remove new phi
1544     assert( hit2->is_Phi(), "" );
1545     phi2 = (PhiNode*)hit2;      // Use existing phi
1546   } else {                      // Miss
1547     _igvn.register_new_node_with_optimizer(phi2);
1548   }
1549   // Register Phis with loop/block info
1550   set_ctrl(phi1, phi->in(0));
1551   set_ctrl(phi2, phi->in(0));
1552   // Make a new Cmp
1553   Node *cmp = sample_cmp->clone();
1554   cmp->set_req( 1, phi1 );
1555   cmp->set_req( 2, phi2 );
1556   _igvn.register_new_node_with_optimizer(cmp);
1557   set_ctrl(cmp, phi->in(0));
1558 
1559   assert( cmp->is_Cmp(), "" );
1560   return (CmpNode*)cmp;
1561 }
1562 
1563 //------------------------------sink_use---------------------------------------
1564 // If 'use' was in the loop-exit block, it now needs to be sunk
1565 // below the post-loop merge point.
1566 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1567   if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1568     set_ctrl(use, post_loop);
1569     for (DUIterator j = use->outs(); use->has_out(j); j++)
1570       sink_use(use->out(j), post_loop);
1571   }
1572 }
1573 
1574 //------------------------------clone_loop-------------------------------------
1575 //
1576 //                   C L O N E   A   L O O P   B O D Y
1577 //
1578 // This is the basic building block of the loop optimizations.  It clones an
1579 // entire loop body.  It makes an old_new loop body mapping; with this mapping
1580 // you can find the new-loop equivalent to an old-loop node.  All new-loop
1581 // nodes are exactly equal to their old-loop counterparts, all edges are the
1582 // same.  All exits from the old-loop now have a RegionNode that merges the
1583 // equivalent new-loop path.  This is true even for the normal "loop-exit"
1584 // condition.  All uses of loop-invariant old-loop values now come from (one
1585 // or more) Phis that merge their new-loop equivalents.
1586 //
1587 // This operation leaves the graph in an illegal state: there are two valid
1588 // control edges coming from the loop pre-header to both loop bodies.  I'll
1589 // definitely have to hack the graph after running this transform.
1590 //
1591 // From this building block I will further edit edges to perform loop peeling
1592 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
1593 //
1594 // Parameter side_by_size_idom:
1595 //   When side_by_size_idom is NULL, the dominator tree is constructed for
1596 //      the clone loop to dominate the original.  Used in construction of
1597 //      pre-main-post loop sequence.
1598 //   When nonnull, the clone and original are side-by-side, both are
1599 //      dominated by the side_by_side_idom node.  Used in construction of
1600 //      unswitched loops.
1601 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
1602                                  Node* side_by_side_idom) {
1603 
1604   if (C->do_vector_loop() && PrintOpto) {
1605     const char* mname = C->method()->name()->as_quoted_ascii();
1606     if (mname != NULL) {
1607       tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
1608     }
1609   }
1610 
1611   CloneMap& cm = C->clone_map();
1612   Dict* dict = cm.dict();
1613   if (C->do_vector_loop()) {
1614     cm.set_clone_idx(cm.max_gen()+1);
1615 #ifndef PRODUCT
1616     if (PrintOpto) {
1617       tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
1618       loop->dump_head();
1619     }
1620 #endif
1621   }
1622 
1623   // Step 1: Clone the loop body.  Make the old->new mapping.
1624   uint i;
1625   for( i = 0; i < loop->_body.size(); i++ ) {
1626     Node *old = loop->_body.at(i);
1627     Node *nnn = old->clone();
1628     old_new.map( old->_idx, nnn );
1629     if (C->do_vector_loop()) {
1630       cm.verify_insert_and_clone(old, nnn, cm.clone_idx());
1631     }
1632     _igvn.register_new_node_with_optimizer(nnn);
1633   }
1634 
1635 
1636   // Step 2: Fix the edges in the new body.  If the old input is outside the
1637   // loop use it.  If the old input is INside the loop, use the corresponding
1638   // new node instead.
1639   for( i = 0; i < loop->_body.size(); i++ ) {
1640     Node *old = loop->_body.at(i);
1641     Node *nnn = old_new[old->_idx];
1642     // Fix CFG/Loop controlling the new node
1643     if (has_ctrl(old)) {
1644       set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
1645     } else {
1646       set_loop(nnn, loop->_parent);
1647       if (old->outcnt() > 0) {
1648         set_idom( nnn, old_new[idom(old)->_idx], dd );
1649       }
1650     }
1651     // Correct edges to the new node
1652     for( uint j = 0; j < nnn->req(); j++ ) {
1653         Node *n = nnn->in(j);
1654         if( n ) {
1655           IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
1656           if( loop->is_member( old_in_loop ) )
1657             nnn->set_req(j, old_new[n->_idx]);
1658         }
1659     }
1660     _igvn.hash_find_insert(nnn);
1661   }
1662   Node *newhead = old_new[loop->_head->_idx];
1663   set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1664 
1665 
1666   // Step 3: Now fix control uses.  Loop varying control uses have already
1667   // been fixed up (as part of all input edges in Step 2).  Loop invariant
1668   // control uses must be either an IfFalse or an IfTrue.  Make a merge
1669   // point to merge the old and new IfFalse/IfTrue nodes; make the use
1670   // refer to this.
1671   ResourceArea *area = Thread::current()->resource_area();
1672   Node_List worklist(area);
1673   uint new_counter = C->unique();
1674   for( i = 0; i < loop->_body.size(); i++ ) {
1675     Node* old = loop->_body.at(i);
1676     if( !old->is_CFG() ) continue;
1677     Node* nnn = old_new[old->_idx];
1678 
1679     // Copy uses to a worklist, so I can munge the def-use info
1680     // with impunity.
1681     for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1682       worklist.push(old->fast_out(j));
1683 
1684     while( worklist.size() ) {  // Visit all uses
1685       Node *use = worklist.pop();
1686       if (!has_node(use))  continue; // Ignore dead nodes
1687       IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1688       if( !loop->is_member( use_loop ) && use->is_CFG() ) {
1689         // Both OLD and USE are CFG nodes here.
1690         assert( use->is_Proj(), "" );
1691 
1692         // Clone the loop exit control projection
1693         Node *newuse = use->clone();
1694         if (C->do_vector_loop()) {
1695           cm.verify_insert_and_clone(use, newuse, cm.clone_idx());
1696         }
1697         newuse->set_req(0,nnn);
1698         _igvn.register_new_node_with_optimizer(newuse);
1699         set_loop(newuse, use_loop);
1700         set_idom(newuse, nnn, dom_depth(nnn) + 1 );
1701 
1702         // We need a Region to merge the exit from the peeled body and the
1703         // exit from the old loop body.
1704         RegionNode *r = new RegionNode(3);
1705         // Map the old use to the new merge point
1706         old_new.map( use->_idx, r );
1707         uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
1708         assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
1709 
1710         // The original user of 'use' uses 'r' instead.
1711         for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
1712           Node* useuse = use->last_out(l);
1713           _igvn.rehash_node_delayed(useuse);
1714           uint uses_found = 0;
1715           if( useuse->in(0) == use ) {
1716             useuse->set_req(0, r);
1717             uses_found++;
1718             if( useuse->is_CFG() ) {
1719               assert( dom_depth(useuse) > dd_r, "" );
1720               set_idom(useuse, r, dom_depth(useuse));
1721             }
1722           }
1723           for( uint k = 1; k < useuse->req(); k++ ) {
1724             if( useuse->in(k) == use ) {
1725               useuse->set_req(k, r);
1726               uses_found++;
1727             }
1728           }
1729           l -= uses_found;    // we deleted 1 or more copies of this edge
1730         }
1731 
1732         // Now finish up 'r'
1733         r->set_req( 1, newuse );
1734         r->set_req( 2,    use );
1735         _igvn.register_new_node_with_optimizer(r);
1736         set_loop(r, use_loop);
1737         set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
1738       } // End of if a loop-exit test
1739     }
1740   }
1741 
1742   // Step 4: If loop-invariant use is not control, it must be dominated by a
1743   // loop exit IfFalse/IfTrue.  Find "proper" loop exit.  Make a Region
1744   // there if needed.  Make a Phi there merging old and new used values.
1745   Node_List *split_if_set = NULL;
1746   Node_List *split_bool_set = NULL;
1747   Node_List *split_cex_set = NULL;
1748   for( i = 0; i < loop->_body.size(); i++ ) {
1749     Node* old = loop->_body.at(i);
1750     Node* nnn = old_new[old->_idx];
1751     // Copy uses to a worklist, so I can munge the def-use info
1752     // with impunity.
1753     for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1754       worklist.push(old->fast_out(j));
1755 
1756     while( worklist.size() ) {
1757       Node *use = worklist.pop();
1758       if (!has_node(use))  continue; // Ignore dead nodes
1759       if (use->in(0) == C->top())  continue;
1760       IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1761       // Check for data-use outside of loop - at least one of OLD or USE
1762       // must not be a CFG node.
1763       if( !loop->is_member( use_loop ) && (!old->is_CFG() || !use->is_CFG())) {
1764 
1765         // If the Data use is an IF, that means we have an IF outside of the
1766         // loop that is switching on a condition that is set inside of the
1767         // loop.  Happens if people set a loop-exit flag; then test the flag
1768         // in the loop to break the loop, then test is again outside of the
1769         // loop to determine which way the loop exited.
1770         // Loop predicate If node connects to Bool node through Opaque1 node.
1771         if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use)) {
1772           // Since this code is highly unlikely, we lazily build the worklist
1773           // of such Nodes to go split.
1774           if( !split_if_set )
1775             split_if_set = new Node_List(area);
1776           split_if_set->push(use);
1777         }
1778         if( use->is_Bool() ) {
1779           if( !split_bool_set )
1780             split_bool_set = new Node_List(area);
1781           split_bool_set->push(use);
1782         }
1783         if( use->Opcode() == Op_CreateEx ) {
1784           if( !split_cex_set )
1785             split_cex_set = new Node_List(area);
1786           split_cex_set->push(use);
1787         }
1788 
1789 
1790         // Get "block" use is in
1791         uint idx = 0;
1792         while( use->in(idx) != old ) idx++;
1793         Node *prev = use->is_CFG() ? use : get_ctrl(use);
1794         assert( !loop->is_member( get_loop( prev ) ), "" );
1795         Node *cfg = prev->_idx >= new_counter
1796           ? prev->in(2)
1797           : idom(prev);
1798         if( use->is_Phi() )     // Phi use is in prior block
1799           cfg = prev->in(idx);  // NOT in block of Phi itself
1800         if (cfg->is_top()) {    // Use is dead?
1801           _igvn.replace_input_of(use, idx, C->top());
1802           continue;
1803         }
1804 
1805         while( !loop->is_member( get_loop( cfg ) ) ) {
1806           prev = cfg;
1807           cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1808         }
1809         // If the use occurs after merging several exits from the loop, then
1810         // old value must have dominated all those exits.  Since the same old
1811         // value was used on all those exits we did not need a Phi at this
1812         // merge point.  NOW we do need a Phi here.  Each loop exit value
1813         // is now merged with the peeled body exit; each exit gets its own
1814         // private Phi and those Phis need to be merged here.
1815         Node *phi;
1816         if( prev->is_Region() ) {
1817           if( idx == 0 ) {      // Updating control edge?
1818             phi = prev;         // Just use existing control
1819           } else {              // Else need a new Phi
1820             phi = PhiNode::make( prev, old );
1821             // Now recursively fix up the new uses of old!
1822             for( uint i = 1; i < prev->req(); i++ ) {
1823               worklist.push(phi); // Onto worklist once for each 'old' input
1824             }
1825           }
1826         } else {
1827           // Get new RegionNode merging old and new loop exits
1828           prev = old_new[prev->_idx];
1829           assert( prev, "just made this in step 7" );
1830           if( idx == 0 ) {      // Updating control edge?
1831             phi = prev;         // Just use existing control
1832           } else {              // Else need a new Phi
1833             // Make a new Phi merging data values properly
1834             phi = PhiNode::make( prev, old );
1835             phi->set_req( 1, nnn );
1836           }
1837         }
1838         // If inserting a new Phi, check for prior hits
1839         if( idx != 0 ) {
1840           Node *hit = _igvn.hash_find_insert(phi);
1841           if( hit == NULL ) {
1842            _igvn.register_new_node_with_optimizer(phi); // Register new phi
1843           } else {                                      // or
1844             // Remove the new phi from the graph and use the hit
1845             _igvn.remove_dead_node(phi);
1846             phi = hit;                                  // Use existing phi
1847           }
1848           set_ctrl(phi, prev);
1849         }
1850         // Make 'use' use the Phi instead of the old loop body exit value
1851         _igvn.replace_input_of(use, idx, phi);
1852         if( use->_idx >= new_counter ) { // If updating new phis
1853           // Not needed for correctness, but prevents a weak assert
1854           // in AddPNode from tripping (when we end up with different
1855           // base & derived Phis that will become the same after
1856           // IGVN does CSE).
1857           Node *hit = _igvn.hash_find_insert(use);
1858           if( hit )             // Go ahead and re-hash for hits.
1859             _igvn.replace_node( use, hit );
1860         }
1861 
1862         // If 'use' was in the loop-exit block, it now needs to be sunk
1863         // below the post-loop merge point.
1864         sink_use( use, prev );
1865       }
1866     }
1867   }
1868 
1869   // Check for IFs that need splitting/cloning.  Happens if an IF outside of
1870   // the loop uses a condition set in the loop.  The original IF probably
1871   // takes control from one or more OLD Regions (which in turn get from NEW
1872   // Regions).  In any case, there will be a set of Phis for each merge point
1873   // from the IF up to where the original BOOL def exists the loop.
1874   if( split_if_set ) {
1875     while( split_if_set->size() ) {
1876       Node *iff = split_if_set->pop();
1877       if( iff->in(1)->is_Phi() ) {
1878         BoolNode *b = clone_iff( iff->in(1)->as_Phi(), loop );
1879         _igvn.replace_input_of(iff, 1, b);
1880       }
1881     }
1882   }
1883   if( split_bool_set ) {
1884     while( split_bool_set->size() ) {
1885       Node *b = split_bool_set->pop();
1886       Node *phi = b->in(1);
1887       assert( phi->is_Phi(), "" );
1888       CmpNode *cmp = clone_bool( (PhiNode*)phi, loop );
1889       _igvn.replace_input_of(b, 1, cmp);
1890     }
1891   }
1892   if( split_cex_set ) {
1893     while( split_cex_set->size() ) {
1894       Node *b = split_cex_set->pop();
1895       assert( b->in(0)->is_Region(), "" );
1896       assert( b->in(1)->is_Phi(), "" );
1897       assert( b->in(0)->in(0) == b->in(1)->in(0), "" );
1898       split_up( b, b->in(0), NULL );
1899     }
1900   }
1901 
1902 }
1903 
1904 
1905 //---------------------- stride_of_possible_iv -------------------------------------
1906 // Looks for an iff/bool/comp with one operand of the compare
1907 // being a cycle involving an add and a phi,
1908 // with an optional truncation (left-shift followed by a right-shift)
1909 // of the add. Returns zero if not an iv.
1910 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
1911   Node* trunc1 = NULL;
1912   Node* trunc2 = NULL;
1913   const TypeInt* ttype = NULL;
1914   if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
1915     return 0;
1916   }
1917   BoolNode* bl = iff->in(1)->as_Bool();
1918   Node* cmp = bl->in(1);
1919   if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
1920     return 0;
1921   }
1922   // Must have an invariant operand
1923   if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
1924     return 0;
1925   }
1926   Node* add2 = NULL;
1927   Node* cmp1 = cmp->in(1);
1928   if (cmp1->is_Phi()) {
1929     // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
1930     Node* phi = cmp1;
1931     for (uint i = 1; i < phi->req(); i++) {
1932       Node* in = phi->in(i);
1933       Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
1934                                 &trunc1, &trunc2, &ttype);
1935       if (add && add->in(1) == phi) {
1936         add2 = add->in(2);
1937         break;
1938       }
1939     }
1940   } else {
1941     // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
1942     Node* addtrunc = cmp1;
1943     Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
1944                                 &trunc1, &trunc2, &ttype);
1945     if (add && add->in(1)->is_Phi()) {
1946       Node* phi = add->in(1);
1947       for (uint i = 1; i < phi->req(); i++) {
1948         if (phi->in(i) == addtrunc) {
1949           add2 = add->in(2);
1950           break;
1951         }
1952       }
1953     }
1954   }
1955   if (add2 != NULL) {
1956     const TypeInt* add2t = _igvn.type(add2)->is_int();
1957     if (add2t->is_con()) {
1958       return add2t->get_con();
1959     }
1960   }
1961   return 0;
1962 }
1963 
1964 
1965 //---------------------- stay_in_loop -------------------------------------
1966 // Return the (unique) control output node that's in the loop (if it exists.)
1967 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
1968   Node* unique = NULL;
1969   if (!n) return NULL;
1970   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1971     Node* use = n->fast_out(i);
1972     if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
1973       if (unique != NULL) {
1974         return NULL;
1975       }
1976       unique = use;
1977     }
1978   }
1979   return unique;
1980 }
1981 
1982 //------------------------------ register_node -------------------------------------
1983 // Utility to register node "n" with PhaseIdealLoop
1984 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
1985   _igvn.register_new_node_with_optimizer(n);
1986   loop->_body.push(n);
1987   if (n->is_CFG()) {
1988     set_loop(n, loop);
1989     set_idom(n, pred, ddepth);
1990   } else {
1991     set_ctrl(n, pred);
1992   }
1993 }
1994 
1995 //------------------------------ proj_clone -------------------------------------
1996 // Utility to create an if-projection
1997 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
1998   ProjNode* c = p->clone()->as_Proj();
1999   c->set_req(0, iff);
2000   return c;
2001 }
2002 
2003 //------------------------------ short_circuit_if -------------------------------------
2004 // Force the iff control output to be the live_proj
2005 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2006   guarantee(live_proj != NULL, "null projection");
2007   int proj_con = live_proj->_con;
2008   assert(proj_con == 0 || proj_con == 1, "false or true projection");
2009   Node *con = _igvn.intcon(proj_con);
2010   set_ctrl(con, C->root());
2011   if (iff) {
2012     iff->set_req(1, con);
2013   }
2014   return con;
2015 }
2016 
2017 //------------------------------ insert_if_before_proj -------------------------------------
2018 // Insert a new if before an if projection (* - new node)
2019 //
2020 // before
2021 //           if(test)
2022 //           /     \
2023 //          v       v
2024 //    other-proj   proj (arg)
2025 //
2026 // after
2027 //           if(test)
2028 //           /     \
2029 //          /       v
2030 //         |      * proj-clone
2031 //         v          |
2032 //    other-proj      v
2033 //                * new_if(relop(cmp[IU](left,right)))
2034 //                  /  \
2035 //                 v    v
2036 //         * new-proj  proj
2037 //         (returned)
2038 //
2039 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2040   IfNode* iff = proj->in(0)->as_If();
2041   IdealLoopTree *loop = get_loop(proj);
2042   ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2043   int ddepth = dom_depth(proj);
2044 
2045   _igvn.rehash_node_delayed(iff);
2046   _igvn.rehash_node_delayed(proj);
2047 
2048   proj->set_req(0, NULL);  // temporary disconnect
2049   ProjNode* proj2 = proj_clone(proj, iff);
2050   register_node(proj2, loop, iff, ddepth);
2051 
2052   Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2053   register_node(cmp, loop, proj2, ddepth);
2054 
2055   BoolNode* bol = new BoolNode(cmp, relop);
2056   register_node(bol, loop, proj2, ddepth);
2057 
2058   int opcode = iff->Opcode();
2059   assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2060   IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt):
2061     new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt);
2062   register_node(new_if, loop, proj2, ddepth);
2063 
2064   proj->set_req(0, new_if); // reattach
2065   set_idom(proj, new_if, ddepth);
2066 
2067   ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2068   guarantee(new_exit != NULL, "null exit node");
2069   register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2070 
2071   return new_exit;
2072 }
2073 
2074 //------------------------------ insert_region_before_proj -------------------------------------
2075 // Insert a region before an if projection (* - new node)
2076 //
2077 // before
2078 //           if(test)
2079 //          /      |
2080 //         v       |
2081 //       proj      v
2082 //               other-proj
2083 //
2084 // after
2085 //           if(test)
2086 //          /      |
2087 //         v       |
2088 // * proj-clone    v
2089 //         |     other-proj
2090 //         v
2091 // * new-region
2092 //         |
2093 //         v
2094 // *      dum_if
2095 //       /     \
2096 //      v       \
2097 // * dum-proj    v
2098 //              proj
2099 //
2100 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
2101   IfNode* iff = proj->in(0)->as_If();
2102   IdealLoopTree *loop = get_loop(proj);
2103   ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2104   int ddepth = dom_depth(proj);
2105 
2106   _igvn.rehash_node_delayed(iff);
2107   _igvn.rehash_node_delayed(proj);
2108 
2109   proj->set_req(0, NULL);  // temporary disconnect
2110   ProjNode* proj2 = proj_clone(proj, iff);
2111   register_node(proj2, loop, iff, ddepth);
2112 
2113   RegionNode* reg = new RegionNode(2);
2114   reg->set_req(1, proj2);
2115   register_node(reg, loop, iff, ddepth);
2116 
2117   IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
2118   register_node(dum_if, loop, reg, ddepth);
2119 
2120   proj->set_req(0, dum_if); // reattach
2121   set_idom(proj, dum_if, ddepth);
2122 
2123   ProjNode* dum_proj = proj_clone(other_proj, dum_if);
2124   register_node(dum_proj, loop, dum_if, ddepth);
2125 
2126   return reg;
2127 }
2128 
2129 //------------------------------ insert_cmpi_loop_exit -------------------------------------
2130 // Clone a signed compare loop exit from an unsigned compare and
2131 // insert it before the unsigned cmp on the stay-in-loop path.
2132 // All new nodes inserted in the dominator tree between the original
2133 // if and it's projections.  The original if test is replaced with
2134 // a constant to force the stay-in-loop path.
2135 //
2136 // This is done to make sure that the original if and it's projections
2137 // still dominate the same set of control nodes, that the ctrl() relation
2138 // from data nodes to them is preserved, and that their loop nesting is
2139 // preserved.
2140 //
2141 // before
2142 //          if(i <u limit)    unsigned compare loop exit
2143 //         /       |
2144 //        v        v
2145 //   exit-proj   stay-in-loop-proj
2146 //
2147 // after
2148 //          if(stay-in-loop-const)  original if
2149 //         /       |
2150 //        /        v
2151 //       /  if(i <  limit)    new signed test
2152 //      /  /       |
2153 //     /  /        v
2154 //    /  /  if(i <u limit)    new cloned unsigned test
2155 //   /  /   /      |
2156 //   v  v  v       |
2157 //    region       |
2158 //        |        |
2159 //      dum-if     |
2160 //     /  |        |
2161 // ether  |        |
2162 //        v        v
2163 //   exit-proj   stay-in-loop-proj
2164 //
2165 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
2166   const bool Signed   = true;
2167   const bool Unsigned = false;
2168 
2169   BoolNode* bol = if_cmpu->in(1)->as_Bool();
2170   if (bol->_test._test != BoolTest::lt) return NULL;
2171   CmpNode* cmpu = bol->in(1)->as_Cmp();
2172   if (cmpu->Opcode() != Op_CmpU) return NULL;
2173   int stride = stride_of_possible_iv(if_cmpu);
2174   if (stride == 0) return NULL;
2175 
2176   Node* lp_proj = stay_in_loop(if_cmpu, loop);
2177   guarantee(lp_proj != NULL, "null loop node");
2178 
2179   ProjNode* lp_continue = lp_proj->as_Proj();
2180   ProjNode* lp_exit     = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
2181 
2182   Node* limit = NULL;
2183   if (stride > 0) {
2184     limit = cmpu->in(2);
2185   } else {
2186     limit = _igvn.makecon(TypeInt::ZERO);
2187     set_ctrl(limit, C->root());
2188   }
2189   // Create a new region on the exit path
2190   RegionNode* reg = insert_region_before_proj(lp_exit);
2191   guarantee(reg != NULL, "null region node");
2192 
2193   // Clone the if-cmpu-true-false using a signed compare
2194   BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
2195   ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
2196   reg->add_req(cmpi_exit);
2197 
2198   // Clone the if-cmpu-true-false
2199   BoolTest::mask rel_u = bol->_test._test;
2200   ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
2201   reg->add_req(cmpu_exit);
2202 
2203   // Force original if to stay in loop.
2204   short_circuit_if(if_cmpu, lp_continue);
2205 
2206   return cmpi_exit->in(0)->as_If();
2207 }
2208 
2209 //------------------------------ remove_cmpi_loop_exit -------------------------------------
2210 // Remove a previously inserted signed compare loop exit.
2211 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
2212   Node* lp_proj = stay_in_loop(if_cmp, loop);
2213   assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
2214          stay_in_loop(lp_proj, loop)->is_If() &&
2215          stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
2216   Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
2217   set_ctrl(con, C->root());
2218   if_cmp->set_req(1, con);
2219 }
2220 
2221 //------------------------------ scheduled_nodelist -------------------------------------
2222 // Create a post order schedule of nodes that are in the
2223 // "member" set.  The list is returned in "sched".
2224 // The first node in "sched" is the loop head, followed by
2225 // nodes which have no inputs in the "member" set, and then
2226 // followed by the nodes that have an immediate input dependence
2227 // on a node in "sched".
2228 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
2229 
2230   assert(member.test(loop->_head->_idx), "loop head must be in member set");
2231   Arena *a = Thread::current()->resource_area();
2232   VectorSet visited(a);
2233   Node_Stack nstack(a, loop->_body.size());
2234 
2235   Node* n  = loop->_head;  // top of stack is cached in "n"
2236   uint idx = 0;
2237   visited.set(n->_idx);
2238 
2239   // Initially push all with no inputs from within member set
2240   for(uint i = 0; i < loop->_body.size(); i++ ) {
2241     Node *elt = loop->_body.at(i);
2242     if (member.test(elt->_idx)) {
2243       bool found = false;
2244       for (uint j = 0; j < elt->req(); j++) {
2245         Node* def = elt->in(j);
2246         if (def && member.test(def->_idx) && def != elt) {
2247           found = true;
2248           break;
2249         }
2250       }
2251       if (!found && elt != loop->_head) {
2252         nstack.push(n, idx);
2253         n = elt;
2254         assert(!visited.test(n->_idx), "not seen yet");
2255         visited.set(n->_idx);
2256       }
2257     }
2258   }
2259 
2260   // traverse out's that are in the member set
2261   while (true) {
2262     if (idx < n->outcnt()) {
2263       Node* use = n->raw_out(idx);
2264       idx++;
2265       if (!visited.test_set(use->_idx)) {
2266         if (member.test(use->_idx)) {
2267           nstack.push(n, idx);
2268           n = use;
2269           idx = 0;
2270         }
2271       }
2272     } else {
2273       // All outputs processed
2274       sched.push(n);
2275       if (nstack.is_empty()) break;
2276       n   = nstack.node();
2277       idx = nstack.index();
2278       nstack.pop();
2279     }
2280   }
2281 }
2282 
2283 
2284 //------------------------------ has_use_in_set -------------------------------------
2285 // Has a use in the vector set
2286 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
2287   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2288     Node* use = n->fast_out(j);
2289     if (vset.test(use->_idx)) {
2290       return true;
2291     }
2292   }
2293   return false;
2294 }
2295 
2296 
2297 //------------------------------ has_use_internal_to_set -------------------------------------
2298 // Has use internal to the vector set (ie. not in a phi at the loop head)
2299 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
2300   Node* head  = loop->_head;
2301   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2302     Node* use = n->fast_out(j);
2303     if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
2304       return true;
2305     }
2306   }
2307   return false;
2308 }
2309 
2310 
2311 //------------------------------ clone_for_use_outside_loop -------------------------------------
2312 // clone "n" for uses that are outside of loop
2313 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
2314   int cloned = 0;
2315   assert(worklist.size() == 0, "should be empty");
2316   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2317     Node* use = n->fast_out(j);
2318     if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
2319       worklist.push(use);
2320     }
2321   }
2322   while( worklist.size() ) {
2323     Node *use = worklist.pop();
2324     if (!has_node(use) || use->in(0) == C->top()) continue;
2325     uint j;
2326     for (j = 0; j < use->req(); j++) {
2327       if (use->in(j) == n) break;
2328     }
2329     assert(j < use->req(), "must be there");
2330 
2331     // clone "n" and insert it between the inputs of "n" and the use outside the loop
2332     Node* n_clone = n->clone();
2333     _igvn.replace_input_of(use, j, n_clone);
2334     cloned++;
2335     Node* use_c;
2336     if (!use->is_Phi()) {
2337       use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
2338     } else {
2339       // Use in a phi is considered a use in the associated predecessor block
2340       use_c = use->in(0)->in(j);
2341     }
2342     set_ctrl(n_clone, use_c);
2343     assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
2344     get_loop(use_c)->_body.push(n_clone);
2345     _igvn.register_new_node_with_optimizer(n_clone);
2346 #if !defined(PRODUCT)
2347     if (TracePartialPeeling) {
2348       tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
2349     }
2350 #endif
2351   }
2352   return cloned;
2353 }
2354 
2355 
2356 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
2357 // clone "n" for special uses that are in the not_peeled region.
2358 // If these def-uses occur in separate blocks, the code generator
2359 // marks the method as not compilable.  For example, if a "BoolNode"
2360 // is in a different basic block than the "IfNode" that uses it, then
2361 // the compilation is aborted in the code generator.
2362 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
2363                                                         VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
2364   if (n->is_Phi() || n->is_Load()) {
2365     return;
2366   }
2367   assert(worklist.size() == 0, "should be empty");
2368   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2369     Node* use = n->fast_out(j);
2370     if ( not_peel.test(use->_idx) &&
2371          (use->is_If() || use->is_CMove() || use->is_Bool()) &&
2372          use->in(1) == n)  {
2373       worklist.push(use);
2374     }
2375   }
2376   if (worklist.size() > 0) {
2377     // clone "n" and insert it between inputs of "n" and the use
2378     Node* n_clone = n->clone();
2379     loop->_body.push(n_clone);
2380     _igvn.register_new_node_with_optimizer(n_clone);
2381     set_ctrl(n_clone, get_ctrl(n));
2382     sink_list.push(n_clone);
2383     not_peel <<= n_clone->_idx;  // add n_clone to not_peel set.
2384 #if !defined(PRODUCT)
2385     if (TracePartialPeeling) {
2386       tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
2387     }
2388 #endif
2389     while( worklist.size() ) {
2390       Node *use = worklist.pop();
2391       _igvn.rehash_node_delayed(use);
2392       for (uint j = 1; j < use->req(); j++) {
2393         if (use->in(j) == n) {
2394           use->set_req(j, n_clone);
2395         }
2396       }
2397     }
2398   }
2399 }
2400 
2401 
2402 //------------------------------ insert_phi_for_loop -------------------------------------
2403 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
2404 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
2405   Node *phi = PhiNode::make(lp, back_edge_val);
2406   phi->set_req(LoopNode::EntryControl, lp_entry_val);
2407   // Use existing phi if it already exists
2408   Node *hit = _igvn.hash_find_insert(phi);
2409   if( hit == NULL ) {
2410     _igvn.register_new_node_with_optimizer(phi);
2411     set_ctrl(phi, lp);
2412   } else {
2413     // Remove the new phi from the graph and use the hit
2414     _igvn.remove_dead_node(phi);
2415     phi = hit;
2416   }
2417   _igvn.replace_input_of(use, idx, phi);
2418 }
2419 
2420 #ifdef ASSERT
2421 //------------------------------ is_valid_loop_partition -------------------------------------
2422 // Validate the loop partition sets: peel and not_peel
2423 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
2424                                               VectorSet& not_peel ) {
2425   uint i;
2426   // Check that peel_list entries are in the peel set
2427   for (i = 0; i < peel_list.size(); i++) {
2428     if (!peel.test(peel_list.at(i)->_idx)) {
2429       return false;
2430     }
2431   }
2432   // Check at loop members are in one of peel set or not_peel set
2433   for (i = 0; i < loop->_body.size(); i++ ) {
2434     Node *def  = loop->_body.at(i);
2435     uint di = def->_idx;
2436     // Check that peel set elements are in peel_list
2437     if (peel.test(di)) {
2438       if (not_peel.test(di)) {
2439         return false;
2440       }
2441       // Must be in peel_list also
2442       bool found = false;
2443       for (uint j = 0; j < peel_list.size(); j++) {
2444         if (peel_list.at(j)->_idx == di) {
2445           found = true;
2446           break;
2447         }
2448       }
2449       if (!found) {
2450         return false;
2451       }
2452     } else if (not_peel.test(di)) {
2453       if (peel.test(di)) {
2454         return false;
2455       }
2456     } else {
2457       return false;
2458     }
2459   }
2460   return true;
2461 }
2462 
2463 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2464 // Ensure a use outside of loop is of the right form
2465 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2466   Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2467   return (use->is_Phi() &&
2468           use_c->is_Region() && use_c->req() == 3 &&
2469           (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2470            use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2471            use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2472           loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2473 }
2474 
2475 //------------------------------ is_valid_clone_loop_form -------------------------------------
2476 // Ensure that all uses outside of loop are of the right form
2477 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2478                                                uint orig_exit_idx, uint clone_exit_idx) {
2479   uint len = peel_list.size();
2480   for (uint i = 0; i < len; i++) {
2481     Node *def = peel_list.at(i);
2482 
2483     for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2484       Node *use = def->fast_out(j);
2485       Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2486       if (!loop->is_member(get_loop(use_c))) {
2487         // use is not in the loop, check for correct structure
2488         if (use->in(0) == def) {
2489           // Okay
2490         } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2491           return false;
2492         }
2493       }
2494     }
2495   }
2496   return true;
2497 }
2498 #endif
2499 
2500 //------------------------------ partial_peel -------------------------------------
2501 // Partially peel (aka loop rotation) the top portion of a loop (called
2502 // the peel section below) by cloning it and placing one copy just before
2503 // the new loop head and the other copy at the bottom of the new loop.
2504 //
2505 //    before                       after                where it came from
2506 //
2507 //    stmt1                        stmt1
2508 //  loop:                          stmt2                     clone
2509 //    stmt2                        if condA goto exitA       clone
2510 //    if condA goto exitA        new_loop:                   new
2511 //    stmt3                        stmt3                     clone
2512 //    if !condB goto loop          if condB goto exitB       clone
2513 //  exitB:                         stmt2                     orig
2514 //    stmt4                        if !condA goto new_loop   orig
2515 //  exitA:                         goto exitA
2516 //                               exitB:
2517 //                                 stmt4
2518 //                               exitA:
2519 //
2520 // Step 1: find the cut point: an exit test on probable
2521 //         induction variable.
2522 // Step 2: schedule (with cloning) operations in the peel
2523 //         section that can be executed after the cut into
2524 //         the section that is not peeled.  This may need
2525 //         to clone operations into exit blocks.  For
2526 //         instance, a reference to A[i] in the not-peel
2527 //         section and a reference to B[i] in an exit block
2528 //         may cause a left-shift of i by 2 to be placed
2529 //         in the peel block.  This step will clone the left
2530 //         shift into the exit block and sink the left shift
2531 //         from the peel to the not-peel section.
2532 // Step 3: clone the loop, retarget the control, and insert
2533 //         phis for values that are live across the new loop
2534 //         head.  This is very dependent on the graph structure
2535 //         from clone_loop.  It creates region nodes for
2536 //         exit control and associated phi nodes for values
2537 //         flow out of the loop through that exit.  The region
2538 //         node is dominated by the clone's control projection.
2539 //         So the clone's peel section is placed before the
2540 //         new loop head, and the clone's not-peel section is
2541 //         forms the top part of the new loop.  The original
2542 //         peel section forms the tail of the new loop.
2543 // Step 4: update the dominator tree and recompute the
2544 //         dominator depth.
2545 //
2546 //                   orig
2547 //
2548 //                   stmt1
2549 //                     |
2550 //                     v
2551 //               loop predicate
2552 //                     |
2553 //                     v
2554 //                   loop<----+
2555 //                     |      |
2556 //                   stmt2    |
2557 //                     |      |
2558 //                     v      |
2559 //                    ifA     |
2560 //                   / |      |
2561 //                  v  v      |
2562 //               false true   ^  <-- last_peel
2563 //               /     |      |
2564 //              /   ===|==cut |
2565 //             /     stmt3    |  <-- first_not_peel
2566 //            /        |      |
2567 //            |        v      |
2568 //            v       ifB     |
2569 //          exitA:   / \      |
2570 //                  /   \     |
2571 //                 v     v    |
2572 //               false true   |
2573 //               /       \    |
2574 //              /         ----+
2575 //             |
2576 //             v
2577 //           exitB:
2578 //           stmt4
2579 //
2580 //
2581 //            after clone loop
2582 //
2583 //                   stmt1
2584 //                     |
2585 //                     v
2586 //               loop predicate
2587 //                 /       \
2588 //        clone   /         \   orig
2589 //               /           \
2590 //              /             \
2591 //             v               v
2592 //   +---->loop                loop<----+
2593 //   |      |                    |      |
2594 //   |    stmt2                stmt2    |
2595 //   |      |                    |      |
2596 //   |      v                    v      |
2597 //   |      ifA                 ifA     |
2598 //   |      | \                / |      |
2599 //   |      v  v              v  v      |
2600 //   ^    true  false      false true   ^  <-- last_peel
2601 //   |      |   ^   \       /    |      |
2602 //   | cut==|==  \   \     /  ===|==cut |
2603 //   |    stmt3   \   \   /    stmt3    |  <-- first_not_peel
2604 //   |      |    dom   | |       |      |
2605 //   |      v      \  1v v2      v      |
2606 //   |      ifB     regionA     ifB     |
2607 //   |      / \        |       / \      |
2608 //   |     /   \       v      /   \     |
2609 //   |    v     v    exitA:  v     v    |
2610 //   |    true  false      false true   |
2611 //   |    /     ^   \      /       \    |
2612 //   +----       \   \    /         ----+
2613 //               dom  \  /
2614 //                 \  1v v2
2615 //                  regionB
2616 //                     |
2617 //                     v
2618 //                   exitB:
2619 //                   stmt4
2620 //
2621 //
2622 //           after partial peel
2623 //
2624 //                  stmt1
2625 //                     |
2626 //                     v
2627 //               loop predicate
2628 //                 /
2629 //        clone   /             orig
2630 //               /          TOP
2631 //              /             \
2632 //             v               v
2633 //    TOP->loop                loop----+
2634 //          |                    |      |
2635 //        stmt2                stmt2    |
2636 //          |                    |      |
2637 //          v                    v      |
2638 //          ifA                 ifA     |
2639 //          | \                / |      |
2640 //          v  v              v  v      |
2641 //        true  false      false true   |     <-- last_peel
2642 //          |   ^   \       /    +------|---+
2643 //  +->newloop   \   \     /  === ==cut |   |
2644 //  |     stmt3   \   \   /     TOP     |   |
2645 //  |       |    dom   | |      stmt3   |   | <-- first_not_peel
2646 //  |       v      \  1v v2      v      |   |
2647 //  |       ifB     regionA     ifB     ^   v
2648 //  |       / \        |       / \      |   |
2649 //  |      /   \       v      /   \     |   |
2650 //  |     v     v    exitA:  v     v    |   |
2651 //  |     true  false      false true   |   |
2652 //  |     /     ^   \      /       \    |   |
2653 //  |    |       \   \    /         v   |   |
2654 //  |    |       dom  \  /         TOP  |   |
2655 //  |    |         \  1v v2             |   |
2656 //  ^    v          regionB             |   |
2657 //  |    |             |                |   |
2658 //  |    |             v                ^   v
2659 //  |    |           exitB:             |   |
2660 //  |    |           stmt4              |   |
2661 //  |    +------------>-----------------+   |
2662 //  |                                       |
2663 //  +-----------------<---------------------+
2664 //
2665 //
2666 //              final graph
2667 //
2668 //                  stmt1
2669 //                    |
2670 //                    v
2671 //               loop predicate
2672 //                    |
2673 //                    v
2674 //                  stmt2 clone
2675 //                    |
2676 //                    v
2677 //         ........> ifA clone
2678 //         :        / |
2679 //        dom      /  |
2680 //         :      v   v
2681 //         :  false   true
2682 //         :  |       |
2683 //         :  |       v
2684 //         :  |    newloop<-----+
2685 //         :  |        |        |
2686 //         :  |     stmt3 clone |
2687 //         :  |        |        |
2688 //         :  |        v        |
2689 //         :  |       ifB       |
2690 //         :  |      / \        |
2691 //         :  |     v   v       |
2692 //         :  |  false true     |
2693 //         :  |   |     |       |
2694 //         :  |   v    stmt2    |
2695 //         :  | exitB:  |       |
2696 //         :  | stmt4   v       |
2697 //         :  |       ifA orig  |
2698 //         :  |      /  \       |
2699 //         :  |     /    \      |
2700 //         :  |    v     v      |
2701 //         :  |  false  true    |
2702 //         :  |  /        \     |
2703 //         :  v  v         -----+
2704 //          RegionA
2705 //             |
2706 //             v
2707 //           exitA
2708 //
2709 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
2710 
2711   assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
2712   if (!loop->_head->is_Loop()) {
2713     return false;  }
2714 
2715   LoopNode *head  = loop->_head->as_Loop();
2716 
2717   if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
2718     return false;
2719   }
2720 
2721   // Check for complex exit control
2722   for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
2723     Node *n = loop->_body.at(ii);
2724     int opc = n->Opcode();
2725     if (n->is_Call()        ||
2726         opc == Op_Catch     ||
2727         opc == Op_CatchProj ||
2728         opc == Op_Jump      ||
2729         opc == Op_JumpProj) {
2730 #if !defined(PRODUCT)
2731       if (TracePartialPeeling) {
2732         tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
2733       }
2734 #endif
2735       return false;
2736     }
2737   }
2738 
2739   int dd = dom_depth(head);
2740 
2741   // Step 1: find cut point
2742 
2743   // Walk up dominators to loop head looking for first loop exit
2744   // which is executed on every path thru loop.
2745   IfNode *peel_if = NULL;
2746   IfNode *peel_if_cmpu = NULL;
2747 
2748   Node *iff = loop->tail();
2749   while( iff != head ) {
2750     if( iff->is_If() ) {
2751       Node *ctrl = get_ctrl(iff->in(1));
2752       if (ctrl->is_top()) return false; // Dead test on live IF.
2753       // If loop-varying exit-test, check for induction variable
2754       if( loop->is_member(get_loop(ctrl)) &&
2755           loop->is_loop_exit(iff) &&
2756           is_possible_iv_test(iff)) {
2757         Node* cmp = iff->in(1)->in(1);
2758         if (cmp->Opcode() == Op_CmpI) {
2759           peel_if = iff->as_If();
2760         } else {
2761           assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
2762           peel_if_cmpu = iff->as_If();
2763         }
2764       }
2765     }
2766     iff = idom(iff);
2767   }
2768   // Prefer signed compare over unsigned compare.
2769   IfNode* new_peel_if = NULL;
2770   if (peel_if == NULL) {
2771     if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
2772       return false;   // No peel point found
2773     }
2774     new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
2775     if (new_peel_if == NULL) {
2776       return false;   // No peel point found
2777     }
2778     peel_if = new_peel_if;
2779   }
2780   Node* last_peel        = stay_in_loop(peel_if, loop);
2781   Node* first_not_peeled = stay_in_loop(last_peel, loop);
2782   if (first_not_peeled == NULL || first_not_peeled == head) {
2783     return false;
2784   }
2785 
2786 #if !defined(PRODUCT)
2787   if (TraceLoopOpts) {
2788     tty->print("PartialPeel  ");
2789     loop->dump_head();
2790   }
2791 
2792   if (TracePartialPeeling) {
2793     tty->print_cr("before partial peel one iteration");
2794     Node_List wl;
2795     Node* t = head->in(2);
2796     while (true) {
2797       wl.push(t);
2798       if (t == head) break;
2799       t = idom(t);
2800     }
2801     while (wl.size() > 0) {
2802       Node* tt = wl.pop();
2803       tt->dump();
2804       if (tt == last_peel) tty->print_cr("-- cut --");
2805     }
2806   }
2807 #endif
2808   ResourceArea *area = Thread::current()->resource_area();
2809   VectorSet peel(area);
2810   VectorSet not_peel(area);
2811   Node_List peel_list(area);
2812   Node_List worklist(area);
2813   Node_List sink_list(area);
2814 
2815   // Set of cfg nodes to peel are those that are executable from
2816   // the head through last_peel.
2817   assert(worklist.size() == 0, "should be empty");
2818   worklist.push(head);
2819   peel.set(head->_idx);
2820   while (worklist.size() > 0) {
2821     Node *n = worklist.pop();
2822     if (n != last_peel) {
2823       for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2824         Node* use = n->fast_out(j);
2825         if (use->is_CFG() &&
2826             loop->is_member(get_loop(use)) &&
2827             !peel.test_set(use->_idx)) {
2828           worklist.push(use);
2829         }
2830       }
2831     }
2832   }
2833 
2834   // Set of non-cfg nodes to peel are those that are control
2835   // dependent on the cfg nodes.
2836   uint i;
2837   for(i = 0; i < loop->_body.size(); i++ ) {
2838     Node *n = loop->_body.at(i);
2839     Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
2840     if (peel.test(n_c->_idx)) {
2841       peel.set(n->_idx);
2842     } else {
2843       not_peel.set(n->_idx);
2844     }
2845   }
2846 
2847   // Step 2: move operations from the peeled section down into the
2848   //         not-peeled section
2849 
2850   // Get a post order schedule of nodes in the peel region
2851   // Result in right-most operand.
2852   scheduled_nodelist(loop, peel, peel_list );
2853 
2854   assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2855 
2856   // For future check for too many new phis
2857   uint old_phi_cnt = 0;
2858   for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2859     Node* use = head->fast_out(j);
2860     if (use->is_Phi()) old_phi_cnt++;
2861   }
2862 
2863 #if !defined(PRODUCT)
2864   if (TracePartialPeeling) {
2865     tty->print_cr("\npeeled list");
2866   }
2867 #endif
2868 
2869   // Evacuate nodes in peel region into the not_peeled region if possible
2870   uint new_phi_cnt = 0;
2871   uint cloned_for_outside_use = 0;
2872   for (i = 0; i < peel_list.size();) {
2873     Node* n = peel_list.at(i);
2874 #if !defined(PRODUCT)
2875   if (TracePartialPeeling) n->dump();
2876 #endif
2877     bool incr = true;
2878     if ( !n->is_CFG() ) {
2879 
2880       if ( has_use_in_set(n, not_peel) ) {
2881 
2882         // If not used internal to the peeled region,
2883         // move "n" from peeled to not_peeled region.
2884 
2885         if ( !has_use_internal_to_set(n, peel, loop) ) {
2886 
2887           // if not pinned and not a load (which maybe anti-dependent on a store)
2888           // and not a CMove (Matcher expects only bool->cmove).
2889           if ( n->in(0) == NULL && !n->is_Load() && !n->is_CMove() ) {
2890             cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist );
2891             sink_list.push(n);
2892             peel     >>= n->_idx; // delete n from peel set.
2893             not_peel <<= n->_idx; // add n to not_peel set.
2894             peel_list.remove(i);
2895             incr = false;
2896 #if !defined(PRODUCT)
2897             if (TracePartialPeeling) {
2898               tty->print_cr("sink to not_peeled region: %d newbb: %d",
2899                             n->_idx, get_ctrl(n)->_idx);
2900             }
2901 #endif
2902           }
2903         } else {
2904           // Otherwise check for special def-use cases that span
2905           // the peel/not_peel boundary such as bool->if
2906           clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
2907           new_phi_cnt++;
2908         }
2909       }
2910     }
2911     if (incr) i++;
2912   }
2913 
2914   if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
2915 #if !defined(PRODUCT)
2916     if (TracePartialPeeling) {
2917       tty->print_cr("\nToo many new phis: %d  old %d new cmpi: %c",
2918                     new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
2919     }
2920 #endif
2921     if (new_peel_if != NULL) {
2922       remove_cmpi_loop_exit(new_peel_if, loop);
2923     }
2924     // Inhibit more partial peeling on this loop
2925     assert(!head->is_partial_peel_loop(), "not partial peeled");
2926     head->mark_partial_peel_failed();
2927     if (cloned_for_outside_use > 0) {
2928       // Terminate this round of loop opts because
2929       // the graph outside this loop was changed.
2930       C->set_major_progress();
2931       return true;
2932     }
2933     return false;
2934   }
2935 
2936   // Step 3: clone loop, retarget control, and insert new phis
2937 
2938   // Create new loop head for new phis and to hang
2939   // the nodes being moved (sinked) from the peel region.
2940   LoopNode* new_head = new LoopNode(last_peel, last_peel);
2941   new_head->set_unswitch_count(head->unswitch_count()); // Preserve
2942   _igvn.register_new_node_with_optimizer(new_head);
2943   assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
2944   _igvn.replace_input_of(first_not_peeled, 0, new_head);
2945   set_loop(new_head, loop);
2946   loop->_body.push(new_head);
2947   not_peel.set(new_head->_idx);
2948   set_idom(new_head, last_peel, dom_depth(first_not_peeled));
2949   set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
2950 
2951   while (sink_list.size() > 0) {
2952     Node* n = sink_list.pop();
2953     set_ctrl(n, new_head);
2954   }
2955 
2956   assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
2957 
2958   clone_loop( loop, old_new, dd );
2959 
2960   const uint clone_exit_idx = 1;
2961   const uint orig_exit_idx  = 2;
2962   assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
2963 
2964   Node* head_clone             = old_new[head->_idx];
2965   LoopNode* new_head_clone     = old_new[new_head->_idx]->as_Loop();
2966   Node* orig_tail_clone        = head_clone->in(2);
2967 
2968   // Add phi if "def" node is in peel set and "use" is not
2969 
2970   for(i = 0; i < peel_list.size(); i++ ) {
2971     Node *def  = peel_list.at(i);
2972     if (!def->is_CFG()) {
2973       for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2974         Node *use = def->fast_out(j);
2975         if (has_node(use) && use->in(0) != C->top() &&
2976             (!peel.test(use->_idx) ||
2977              (use->is_Phi() && use->in(0) == head)) ) {
2978           worklist.push(use);
2979         }
2980       }
2981       while( worklist.size() ) {
2982         Node *use = worklist.pop();
2983         for (uint j = 1; j < use->req(); j++) {
2984           Node* n = use->in(j);
2985           if (n == def) {
2986 
2987             // "def" is in peel set, "use" is not in peel set
2988             // or "use" is in the entry boundary (a phi) of the peel set
2989 
2990             Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
2991 
2992             if ( loop->is_member(get_loop( use_c )) ) {
2993               // use is in loop
2994               if (old_new[use->_idx] != NULL) { // null for dead code
2995                 Node* use_clone = old_new[use->_idx];
2996                 _igvn.replace_input_of(use, j, C->top());
2997                 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
2998               }
2999             } else {
3000               assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
3001               // use is not in the loop, check if the live range includes the cut
3002               Node* lp_if = use_c->in(orig_exit_idx)->in(0);
3003               if (not_peel.test(lp_if->_idx)) {
3004                 assert(j == orig_exit_idx, "use from original loop");
3005                 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
3006               }
3007             }
3008           }
3009         }
3010       }
3011     }
3012   }
3013 
3014   // Step 3b: retarget control
3015 
3016   // Redirect control to the new loop head if a cloned node in
3017   // the not_peeled region has control that points into the peeled region.
3018   // This necessary because the cloned peeled region will be outside
3019   // the loop.
3020   //                            from    to
3021   //          cloned-peeled    <---+
3022   //    new_head_clone:            |    <--+
3023   //          cloned-not_peeled  in(0)    in(0)
3024   //          orig-peeled
3025 
3026   for(i = 0; i < loop->_body.size(); i++ ) {
3027     Node *n = loop->_body.at(i);
3028     if (!n->is_CFG()           && n->in(0) != NULL        &&
3029         not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
3030       Node* n_clone = old_new[n->_idx];
3031       _igvn.replace_input_of(n_clone, 0, new_head_clone);
3032     }
3033   }
3034 
3035   // Backedge of the surviving new_head (the clone) is original last_peel
3036   _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
3037 
3038   // Cut first node in original not_peel set
3039   _igvn.rehash_node_delayed(new_head);                     // Multiple edge updates:
3040   new_head->set_req(LoopNode::EntryControl,    C->top());  //   use rehash_node_delayed / set_req instead of
3041   new_head->set_req(LoopNode::LoopBackControl, C->top());  //   multiple replace_input_of calls
3042 
3043   // Copy head_clone back-branch info to original head
3044   // and remove original head's loop entry and
3045   // clone head's back-branch
3046   _igvn.rehash_node_delayed(head); // Multiple edge updates
3047   head->set_req(LoopNode::EntryControl,    head_clone->in(LoopNode::LoopBackControl));
3048   head->set_req(LoopNode::LoopBackControl, C->top());
3049   _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
3050 
3051   // Similarly modify the phis
3052   for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
3053     Node* use = head->fast_out(k);
3054     if (use->is_Phi() && use->outcnt() > 0) {
3055       Node* use_clone = old_new[use->_idx];
3056       _igvn.rehash_node_delayed(use); // Multiple edge updates
3057       use->set_req(LoopNode::EntryControl,    use_clone->in(LoopNode::LoopBackControl));
3058       use->set_req(LoopNode::LoopBackControl, C->top());
3059       _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
3060     }
3061   }
3062 
3063   // Step 4: update dominator tree and dominator depth
3064 
3065   set_idom(head, orig_tail_clone, dd);
3066   recompute_dom_depth();
3067 
3068   // Inhibit more partial peeling on this loop
3069   new_head_clone->set_partial_peel_loop();
3070   C->set_major_progress();
3071   loop->record_for_igvn();
3072 
3073 #if !defined(PRODUCT)
3074   if (TracePartialPeeling) {
3075     tty->print_cr("\nafter partial peel one iteration");
3076     Node_List wl(area);
3077     Node* t = last_peel;
3078     while (true) {
3079       wl.push(t);
3080       if (t == head_clone) break;
3081       t = idom(t);
3082     }
3083     while (wl.size() > 0) {
3084       Node* tt = wl.pop();
3085       if (tt == head) tty->print_cr("orig head");
3086       else if (tt == new_head_clone) tty->print_cr("new head");
3087       else if (tt == head_clone) tty->print_cr("clone head");
3088       tt->dump();
3089     }
3090   }
3091 #endif
3092   return true;
3093 }
3094 
3095 //------------------------------reorg_offsets----------------------------------
3096 // Reorganize offset computations to lower register pressure.  Mostly
3097 // prevent loop-fallout uses of the pre-incremented trip counter (which are
3098 // then alive with the post-incremented trip counter forcing an extra
3099 // register move)
3100 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
3101   // Perform it only for canonical counted loops.
3102   // Loop's shape could be messed up by iteration_split_impl.
3103   if (!loop->_head->is_CountedLoop())
3104     return;
3105   if (!loop->_head->as_Loop()->is_valid_counted_loop())
3106     return;
3107 
3108   CountedLoopNode *cl = loop->_head->as_CountedLoop();
3109   CountedLoopEndNode *cle = cl->loopexit();
3110   Node *exit = cle->proj_out(false);
3111   Node *phi = cl->phi();
3112 
3113   // Check for the special case of folks using the pre-incremented
3114   // trip-counter on the fall-out path (forces the pre-incremented
3115   // and post-incremented trip counter to be live at the same time).
3116   // Fix this by adjusting to use the post-increment trip counter.
3117 
3118   bool progress = true;
3119   while (progress) {
3120     progress = false;
3121     for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
3122       Node* use = phi->fast_out(i);   // User of trip-counter
3123       if (!has_ctrl(use))  continue;
3124       Node *u_ctrl = get_ctrl(use);
3125       if (use->is_Phi()) {
3126         u_ctrl = NULL;
3127         for (uint j = 1; j < use->req(); j++)
3128           if (use->in(j) == phi)
3129             u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
3130       }
3131       IdealLoopTree *u_loop = get_loop(u_ctrl);
3132       // Look for loop-invariant use
3133       if (u_loop == loop) continue;
3134       if (loop->is_member(u_loop)) continue;
3135       // Check that use is live out the bottom.  Assuming the trip-counter
3136       // update is right at the bottom, uses of of the loop middle are ok.
3137       if (dom_lca(exit, u_ctrl) != exit) continue;
3138       // Hit!  Refactor use to use the post-incremented tripcounter.
3139       // Compute a post-increment tripcounter.
3140       Node *opaq = new Opaque2Node( C, cle->incr() );
3141       register_new_node(opaq, exit);
3142       Node *neg_stride = _igvn.intcon(-cle->stride_con());
3143       set_ctrl(neg_stride, C->root());
3144       Node *post = new AddINode( opaq, neg_stride);
3145       register_new_node(post, exit);
3146       _igvn.rehash_node_delayed(use);
3147       for (uint j = 1; j < use->req(); j++) {
3148         if (use->in(j) == phi)
3149           use->set_req(j, post);
3150       }
3151       // Since DU info changed, rerun loop
3152       progress = true;
3153       break;
3154     }
3155   }
3156 
3157 }