1 /*
   2  * Copyright (c) 1999, 2018, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "gc/shared/barrierSet.hpp"
  27 #include "gc/shared/c2/barrierSetC2.hpp"
  28 #include "memory/allocation.inline.hpp"
  29 #include "memory/resourceArea.hpp"
  30 #include "opto/addnode.hpp"
  31 #include "opto/callnode.hpp"
  32 #include "opto/castnode.hpp"
  33 #include "opto/connode.hpp"
  34 #include "opto/castnode.hpp"
  35 #include "opto/divnode.hpp"
  36 #include "opto/loopnode.hpp"
  37 #include "opto/matcher.hpp"
  38 #include "opto/mulnode.hpp"
  39 #include "opto/movenode.hpp"
  40 #include "opto/opaquenode.hpp"
  41 #include "opto/rootnode.hpp"
  42 #include "opto/subnode.hpp"
  43 #include "utilities/macros.hpp"
  44 #if INCLUDE_ZGC
  45 #include "gc/z/c2/zBarrierSetC2.hpp"
  46 #endif
  47 
  48 //=============================================================================
  49 //------------------------------split_thru_phi---------------------------------
  50 // Split Node 'n' through merge point if there is enough win.
  51 Node *PhaseIdealLoop::split_thru_phi( Node *n, Node *region, int policy ) {
  52   if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
  53     // ConvI2L may have type information on it which is unsafe to push up
  54     // so disable this for now
  55     return NULL;
  56   }
  57 
  58   // Splitting range check CastIIs through a loop induction Phi can
  59   // cause new Phis to be created that are left unrelated to the loop
  60   // induction Phi and prevent optimizations (vectorization)
  61   if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() &&
  62       region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) {
  63     return NULL;
  64   }
  65 
  66   int wins = 0;
  67   assert(!n->is_CFG(), "");
  68   assert(region->is_Region(), "");
  69 
  70   const Type* type = n->bottom_type();
  71   const TypeOopPtr *t_oop = _igvn.type(n)->isa_oopptr();
  72   Node *phi;
  73   if (t_oop != NULL && t_oop->is_known_instance_field()) {
  74     int iid    = t_oop->instance_id();
  75     int index  = C->get_alias_index(t_oop);
  76     int offset = t_oop->offset();
  77     phi = new PhiNode(region, type, NULL, iid, index, offset);
  78   } else {
  79     phi = PhiNode::make_blank(region, n);
  80   }
  81   uint old_unique = C->unique();
  82   for (uint i = 1; i < region->req(); i++) {
  83     Node *x;
  84     Node* the_clone = NULL;
  85     if (region->in(i) == C->top()) {
  86       x = C->top();             // Dead path?  Use a dead data op
  87     } else {
  88       x = n->clone();           // Else clone up the data op
  89       the_clone = x;            // Remember for possible deletion.
  90       // Alter data node to use pre-phi inputs
  91       if (n->in(0) == region)
  92         x->set_req( 0, region->in(i) );
  93       for (uint j = 1; j < n->req(); j++) {
  94         Node *in = n->in(j);
  95         if (in->is_Phi() && in->in(0) == region)
  96           x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
  97       }
  98     }
  99     // Check for a 'win' on some paths
 100     const Type *t = x->Value(&_igvn);
 101 
 102     bool singleton = t->singleton();
 103 
 104     // A TOP singleton indicates that there are no possible values incoming
 105     // along a particular edge. In most cases, this is OK, and the Phi will
 106     // be eliminated later in an Ideal call. However, we can't allow this to
 107     // happen if the singleton occurs on loop entry, as the elimination of
 108     // the PhiNode may cause the resulting node to migrate back to a previous
 109     // loop iteration.
 110     if (singleton && t == Type::TOP) {
 111       // Is_Loop() == false does not confirm the absence of a loop (e.g., an
 112       // irreducible loop may not be indicated by an affirmative is_Loop());
 113       // therefore, the only top we can split thru a phi is on a backedge of
 114       // a loop.
 115       singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
 116     }
 117 
 118     if (singleton) {
 119       wins++;
 120       x = ((PhaseGVN&)_igvn).makecon(t);
 121     } else {
 122       // We now call Identity to try to simplify the cloned node.
 123       // Note that some Identity methods call phase->type(this).
 124       // Make sure that the type array is big enough for
 125       // our new node, even though we may throw the node away.
 126       // (Note: This tweaking with igvn only works because x is a new node.)
 127       _igvn.set_type(x, t);
 128       // If x is a TypeNode, capture any more-precise type permanently into Node
 129       // otherwise it will be not updated during igvn->transform since
 130       // igvn->type(x) is set to x->Value() already.
 131       x->raise_bottom_type(t);
 132       Node *y = _igvn.apply_identity(x);
 133       if (y != x) {
 134         wins++;
 135         x = y;
 136       } else {
 137         y = _igvn.hash_find(x);
 138         if (y) {
 139           wins++;
 140           x = y;
 141         } else {
 142           // Else x is a new node we are keeping
 143           // We do not need register_new_node_with_optimizer
 144           // because set_type has already been called.
 145           _igvn._worklist.push(x);
 146         }
 147       }
 148     }
 149     if (x != the_clone && the_clone != NULL)
 150       _igvn.remove_dead_node(the_clone);
 151     phi->set_req( i, x );
 152   }
 153   // Too few wins?
 154   if (wins <= policy) {
 155     _igvn.remove_dead_node(phi);
 156     return NULL;
 157   }
 158 
 159   // Record Phi
 160   register_new_node( phi, region );
 161 
 162   for (uint i2 = 1; i2 < phi->req(); i2++) {
 163     Node *x = phi->in(i2);
 164     // If we commoned up the cloned 'x' with another existing Node,
 165     // the existing Node picks up a new use.  We need to make the
 166     // existing Node occur higher up so it dominates its uses.
 167     Node *old_ctrl;
 168     IdealLoopTree *old_loop;
 169 
 170     if (x->is_Con()) {
 171       // Constant's control is always root.
 172       set_ctrl(x, C->root());
 173       continue;
 174     }
 175     // The occasional new node
 176     if (x->_idx >= old_unique) {     // Found a new, unplaced node?
 177       old_ctrl = NULL;
 178       old_loop = NULL;               // Not in any prior loop
 179     } else {
 180       old_ctrl = get_ctrl(x);
 181       old_loop = get_loop(old_ctrl); // Get prior loop
 182     }
 183     // New late point must dominate new use
 184     Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
 185     if (new_ctrl == old_ctrl) // Nothing is changed
 186       continue;
 187 
 188     IdealLoopTree *new_loop = get_loop(new_ctrl);
 189 
 190     // Don't move x into a loop if its uses are
 191     // outside of loop. Otherwise x will be cloned
 192     // for each use outside of this loop.
 193     IdealLoopTree *use_loop = get_loop(region);
 194     if (!new_loop->is_member(use_loop) &&
 195         (old_loop == NULL || !new_loop->is_member(old_loop))) {
 196       // Take early control, later control will be recalculated
 197       // during next iteration of loop optimizations.
 198       new_ctrl = get_early_ctrl(x);
 199       new_loop = get_loop(new_ctrl);
 200     }
 201     // Set new location
 202     set_ctrl(x, new_ctrl);
 203     // If changing loop bodies, see if we need to collect into new body
 204     if (old_loop != new_loop) {
 205       if (old_loop && !old_loop->_child)
 206         old_loop->_body.yank(x);
 207       if (!new_loop->_child)
 208         new_loop->_body.push(x);  // Collect body info
 209     }
 210   }
 211 
 212   return phi;
 213 }
 214 
 215 //------------------------------dominated_by------------------------------------
 216 // Replace the dominated test with an obvious true or false.  Place it on the
 217 // IGVN worklist for later cleanup.  Move control-dependent data Nodes on the
 218 // live path up to the dominating control.
 219 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) {
 220   if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
 221 
 222   // prevdom is the dominating projection of the dominating test.
 223   assert( iff->is_If(), "" );
 224   assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added");
 225   int pop = prevdom->Opcode();
 226   assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
 227   if (flip) {
 228     if (pop == Op_IfTrue)
 229       pop = Op_IfFalse;
 230     else
 231       pop = Op_IfTrue;
 232   }
 233   // 'con' is set to true or false to kill the dominated test.
 234   Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
 235   set_ctrl(con, C->root()); // Constant gets a new use
 236   // Hack the dominated test
 237   _igvn.replace_input_of(iff, 1, con);
 238 
 239   // If I dont have a reachable TRUE and FALSE path following the IfNode then
 240   // I can assume this path reaches an infinite loop.  In this case it's not
 241   // important to optimize the data Nodes - either the whole compilation will
 242   // be tossed or this path (and all data Nodes) will go dead.
 243   if (iff->outcnt() != 2) return;
 244 
 245   // Make control-dependent data Nodes on the live path (path that will remain
 246   // once the dominated IF is removed) become control-dependent on the
 247   // dominating projection.
 248   Node* dp = iff->as_If()->proj_out_or_null(pop == Op_IfTrue);
 249 
 250   // Loop predicates may have depending checks which should not
 251   // be skipped. For example, range check predicate has two checks
 252   // for lower and upper bounds.
 253   if (dp == NULL)
 254     return;
 255 
 256   ProjNode* dp_proj  = dp->as_Proj();
 257   ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj();
 258   if (exclude_loop_predicate &&
 259       (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL ||
 260        unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != NULL ||
 261        unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) {
 262     // If this is a range check (IfNode::is_range_check), do not
 263     // reorder because Compile::allow_range_check_smearing might have
 264     // changed the check.
 265     return; // Let IGVN transformation change control dependence.
 266   }
 267 
 268   IdealLoopTree *old_loop = get_loop(dp);
 269 
 270   for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
 271     Node* cd = dp->fast_out(i); // Control-dependent node
 272     if (cd->depends_only_on_test()) {
 273       assert(cd->in(0) == dp, "");
 274       _igvn.replace_input_of(cd, 0, prevdom);
 275       set_early_ctrl(cd);
 276       IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
 277       if (old_loop != new_loop) {
 278         if (!old_loop->_child) old_loop->_body.yank(cd);
 279         if (!new_loop->_child) new_loop->_body.push(cd);
 280       }
 281       --i;
 282       --imax;
 283     }
 284   }
 285 }
 286 
 287 //------------------------------has_local_phi_input----------------------------
 288 // Return TRUE if 'n' has Phi inputs from its local block and no other
 289 // block-local inputs (all non-local-phi inputs come from earlier blocks)
 290 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
 291   Node *n_ctrl = get_ctrl(n);
 292   // See if some inputs come from a Phi in this block, or from before
 293   // this block.
 294   uint i;
 295   for( i = 1; i < n->req(); i++ ) {
 296     Node *phi = n->in(i);
 297     if( phi->is_Phi() && phi->in(0) == n_ctrl )
 298       break;
 299   }
 300   if( i >= n->req() )
 301     return NULL;                // No Phi inputs; nowhere to clone thru
 302 
 303   // Check for inputs created between 'n' and the Phi input.  These
 304   // must split as well; they have already been given the chance
 305   // (courtesy of a post-order visit) and since they did not we must
 306   // recover the 'cost' of splitting them by being very profitable
 307   // when splitting 'n'.  Since this is unlikely we simply give up.
 308   for( i = 1; i < n->req(); i++ ) {
 309     Node *m = n->in(i);
 310     if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
 311       // We allow the special case of AddP's with no local inputs.
 312       // This allows us to split-up address expressions.
 313       if (m->is_AddP() &&
 314           get_ctrl(m->in(2)) != n_ctrl &&
 315           get_ctrl(m->in(3)) != n_ctrl) {
 316         // Move the AddP up to dominating point
 317         Node* c = find_non_split_ctrl(idom(n_ctrl));
 318         if (c->is_OuterStripMinedLoop()) {
 319           c->as_Loop()->verify_strip_mined(1);
 320           c = c->in(LoopNode::EntryControl);
 321         }
 322         set_ctrl_and_loop(m, c);
 323         continue;
 324       }
 325       return NULL;
 326     }
 327     assert(m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
 328   }
 329 
 330   return n_ctrl;
 331 }
 332 
 333 //------------------------------remix_address_expressions----------------------
 334 // Rework addressing expressions to get the most loop-invariant stuff
 335 // moved out.  We'd like to do all associative operators, but it's especially
 336 // important (common) to do address expressions.
 337 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
 338   if (!has_ctrl(n))  return NULL;
 339   Node *n_ctrl = get_ctrl(n);
 340   IdealLoopTree *n_loop = get_loop(n_ctrl);
 341 
 342   // See if 'n' mixes loop-varying and loop-invariant inputs and
 343   // itself is loop-varying.
 344 
 345   // Only interested in binary ops (and AddP)
 346   if( n->req() < 3 || n->req() > 4 ) return NULL;
 347 
 348   Node *n1_ctrl = get_ctrl(n->in(                    1));
 349   Node *n2_ctrl = get_ctrl(n->in(                    2));
 350   Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
 351   IdealLoopTree *n1_loop = get_loop( n1_ctrl );
 352   IdealLoopTree *n2_loop = get_loop( n2_ctrl );
 353   IdealLoopTree *n3_loop = get_loop( n3_ctrl );
 354 
 355   // Does one of my inputs spin in a tighter loop than self?
 356   if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
 357       (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
 358       (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
 359     return NULL;                // Leave well enough alone
 360 
 361   // Is at least one of my inputs loop-invariant?
 362   if( n1_loop == n_loop &&
 363       n2_loop == n_loop &&
 364       n3_loop == n_loop )
 365     return NULL;                // No loop-invariant inputs
 366 
 367 
 368   int n_op = n->Opcode();
 369 
 370   // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
 371   if( n_op == Op_LShiftI ) {
 372     // Scale is loop invariant
 373     Node *scale = n->in(2);
 374     Node *scale_ctrl = get_ctrl(scale);
 375     IdealLoopTree *scale_loop = get_loop(scale_ctrl );
 376     if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
 377       return NULL;
 378     const TypeInt *scale_t = scale->bottom_type()->isa_int();
 379     if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
 380       return NULL;              // Dont bother with byte/short masking
 381     // Add must vary with loop (else shift would be loop-invariant)
 382     Node *add = n->in(1);
 383     Node *add_ctrl = get_ctrl(add);
 384     IdealLoopTree *add_loop = get_loop(add_ctrl);
 385     //assert( n_loop == add_loop, "" );
 386     if( n_loop != add_loop ) return NULL;  // happens w/ evil ZKM loops
 387 
 388     // Convert I-V into I+ (0-V); same for V-I
 389     if( add->Opcode() == Op_SubI &&
 390         _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
 391       Node *zero = _igvn.intcon(0);
 392       set_ctrl(zero, C->root());
 393       Node *neg = new SubINode( _igvn.intcon(0), add->in(2) );
 394       register_new_node( neg, get_ctrl(add->in(2) ) );
 395       add = new AddINode( add->in(1), neg );
 396       register_new_node( add, add_ctrl );
 397     }
 398     if( add->Opcode() != Op_AddI ) return NULL;
 399     // See if one add input is loop invariant
 400     Node *add_var = add->in(1);
 401     Node *add_var_ctrl = get_ctrl(add_var);
 402     IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
 403     Node *add_invar = add->in(2);
 404     Node *add_invar_ctrl = get_ctrl(add_invar);
 405     IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
 406     if( add_var_loop == n_loop ) {
 407     } else if( add_invar_loop == n_loop ) {
 408       // Swap to find the invariant part
 409       add_invar = add_var;
 410       add_invar_ctrl = add_var_ctrl;
 411       add_invar_loop = add_var_loop;
 412       add_var = add->in(2);
 413       Node *add_var_ctrl = get_ctrl(add_var);
 414       IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
 415     } else                      // Else neither input is loop invariant
 416       return NULL;
 417     if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
 418       return NULL;              // No invariant part of the add?
 419 
 420     // Yes!  Reshape address expression!
 421     Node *inv_scale = new LShiftINode( add_invar, scale );
 422     Node *inv_scale_ctrl =
 423       dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
 424       add_invar_ctrl : scale_ctrl;
 425     register_new_node( inv_scale, inv_scale_ctrl );
 426     Node *var_scale = new LShiftINode( add_var, scale );
 427     register_new_node( var_scale, n_ctrl );
 428     Node *var_add = new AddINode( var_scale, inv_scale );
 429     register_new_node( var_add, n_ctrl );
 430     _igvn.replace_node( n, var_add );
 431     return var_add;
 432   }
 433 
 434   // Replace (I+V) with (V+I)
 435   if( n_op == Op_AddI ||
 436       n_op == Op_AddL ||
 437       n_op == Op_AddF ||
 438       n_op == Op_AddD ||
 439       n_op == Op_MulI ||
 440       n_op == Op_MulL ||
 441       n_op == Op_MulF ||
 442       n_op == Op_MulD ) {
 443     if( n2_loop == n_loop ) {
 444       assert( n1_loop != n_loop, "" );
 445       n->swap_edges(1, 2);
 446     }
 447   }
 448 
 449   // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
 450   // but not if I2 is a constant.
 451   if( n_op == Op_AddP ) {
 452     if( n2_loop == n_loop && n3_loop != n_loop ) {
 453       if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
 454         Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
 455         Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
 456         IdealLoopTree *n22loop = get_loop( n22_ctrl );
 457         IdealLoopTree *n23_loop = get_loop( n23_ctrl );
 458         if( n22loop != n_loop && n22loop->is_member(n_loop) &&
 459             n23_loop == n_loop ) {
 460           Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
 461           // Stuff new AddP in the loop preheader
 462           register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
 463           Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) );
 464           register_new_node( add2, n_ctrl );
 465           _igvn.replace_node( n, add2 );
 466           return add2;
 467         }
 468       }
 469     }
 470 
 471     // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
 472     if (n2_loop != n_loop && n3_loop == n_loop) {
 473       if (n->in(3)->Opcode() == Op_AddX) {
 474         Node *V = n->in(3)->in(1);
 475         Node *I = n->in(3)->in(2);
 476         if (is_member(n_loop,get_ctrl(V))) {
 477         } else {
 478           Node *tmp = V; V = I; I = tmp;
 479         }
 480         if (!is_member(n_loop,get_ctrl(I))) {
 481           Node *add1 = new AddPNode(n->in(1), n->in(2), I);
 482           // Stuff new AddP in the loop preheader
 483           register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl));
 484           Node *add2 = new AddPNode(n->in(1), add1, V);
 485           register_new_node(add2, n_ctrl);
 486           _igvn.replace_node(n, add2);
 487           return add2;
 488         }
 489       }
 490     }
 491   }
 492 
 493   return NULL;
 494 }
 495 
 496 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
 497 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
 498   assert(n->Opcode() == Op_AddI, "sanity");
 499   Node * nn = NULL;
 500   Node * in1 = n->in(1);
 501   Node * in2 = n->in(2);
 502   if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
 503     IdealLoopTree* loop_n = get_loop(get_ctrl(n));
 504     if (loop_n->_head->as_Loop()->is_valid_counted_loop() &&
 505         Matcher::match_rule_supported(Op_MulAddS2I) &&
 506         Matcher::match_rule_supported(Op_MulAddVS2VI)) {
 507       Node* mul_in1 = in1->in(1);
 508       Node* mul_in2 = in1->in(2);
 509       Node* mul_in3 = in2->in(1);
 510       Node* mul_in4 = in2->in(2);
 511       if (mul_in1->Opcode() == Op_LoadS &&
 512           mul_in2->Opcode() == Op_LoadS &&
 513           mul_in3->Opcode() == Op_LoadS &&
 514           mul_in4->Opcode() == Op_LoadS) {
 515         IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
 516         IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
 517         IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
 518         IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
 519         IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
 520         IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
 521         // All nodes should be in the same counted loop.
 522         if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
 523             loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
 524           Node* adr1 = mul_in1->in(MemNode::Address);
 525           Node* adr2 = mul_in2->in(MemNode::Address);
 526           Node* adr3 = mul_in3->in(MemNode::Address);
 527           Node* adr4 = mul_in4->in(MemNode::Address);
 528           if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
 529             if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
 530                 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
 531               nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
 532               register_new_node(nn, get_ctrl(n));
 533               _igvn.replace_node(n, nn);
 534               return nn;
 535             }
 536           }
 537         }
 538       }
 539     }
 540   }
 541   return nn;
 542 }
 543 
 544 //------------------------------conditional_move-------------------------------
 545 // Attempt to replace a Phi with a conditional move.  We have some pretty
 546 // strict profitability requirements.  All Phis at the merge point must
 547 // be converted, so we can remove the control flow.  We need to limit the
 548 // number of c-moves to a small handful.  All code that was in the side-arms
 549 // of the CFG diamond is now speculatively executed.  This code has to be
 550 // "cheap enough".  We are pretty much limited to CFG diamonds that merge
 551 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
 552 Node *PhaseIdealLoop::conditional_move( Node *region ) {
 553 
 554   assert(region->is_Region(), "sanity check");
 555   if (region->req() != 3) return NULL;
 556 
 557   // Check for CFG diamond
 558   Node *lp = region->in(1);
 559   Node *rp = region->in(2);
 560   if (!lp || !rp) return NULL;
 561   Node *lp_c = lp->in(0);
 562   if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL;
 563   IfNode *iff = lp_c->as_If();
 564 
 565   // Check for ops pinned in an arm of the diamond.
 566   // Can't remove the control flow in this case
 567   if (lp->outcnt() > 1) return NULL;
 568   if (rp->outcnt() > 1) return NULL;
 569 
 570   IdealLoopTree* r_loop = get_loop(region);
 571   assert(r_loop == get_loop(iff), "sanity");
 572   // Always convert to CMOVE if all results are used only outside this loop.
 573   bool used_inside_loop = (r_loop == _ltree_root);
 574 
 575   // Check profitability
 576   int cost = 0;
 577   int phis = 0;
 578   for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
 579     Node *out = region->fast_out(i);
 580     if (!out->is_Phi()) continue; // Ignore other control edges, etc
 581     phis++;
 582     PhiNode* phi = out->as_Phi();
 583     BasicType bt = phi->type()->basic_type();
 584     switch (bt) {
 585     case T_DOUBLE:
 586     case T_FLOAT:
 587       if (C->use_cmove()) {
 588         continue; //TODO: maybe we want to add some cost
 589       }
 590       cost += Matcher::float_cmove_cost(); // Could be very expensive
 591       break;
 592     case T_LONG: {
 593       cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
 594     }
 595     case T_INT:                 // These all CMOV fine
 596     case T_ADDRESS: {           // (RawPtr)
 597       cost++;
 598       break;
 599     }
 600     case T_NARROWOOP: // Fall through
 601     case T_OBJECT: {            // Base oops are OK, but not derived oops
 602       const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
 603       // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
 604       // CMOVE'd derived pointer?  It's a CMOVE'd derived base.  Thus
 605       // CMOVE'ing a derived pointer requires we also CMOVE the base.  If we
 606       // have a Phi for the base here that we convert to a CMOVE all is well
 607       // and good.  But if the base is dead, we'll not make a CMOVE.  Later
 608       // the allocator will have to produce a base by creating a CMOVE of the
 609       // relevant bases.  This puts the allocator in the business of
 610       // manufacturing expensive instructions, generally a bad plan.
 611       // Just Say No to Conditionally-Moved Derived Pointers.
 612       if (tp && tp->offset() != 0)
 613         return NULL;
 614       cost++;
 615       break;
 616     }
 617     default:
 618       return NULL;              // In particular, can't do memory or I/O
 619     }
 620     // Add in cost any speculative ops
 621     for (uint j = 1; j < region->req(); j++) {
 622       Node *proj = region->in(j);
 623       Node *inp = phi->in(j);
 624       if (get_ctrl(inp) == proj) { // Found local op
 625         cost++;
 626         // Check for a chain of dependent ops; these will all become
 627         // speculative in a CMOV.
 628         for (uint k = 1; k < inp->req(); k++)
 629           if (get_ctrl(inp->in(k)) == proj)
 630             cost += ConditionalMoveLimit; // Too much speculative goo
 631       }
 632     }
 633     // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
 634     // This will likely Split-If, a higher-payoff operation.
 635     for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
 636       Node* use = phi->fast_out(k);
 637       if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
 638         cost += ConditionalMoveLimit;
 639       // Is there a use inside the loop?
 640       // Note: check only basic types since CMoveP is pinned.
 641       if (!used_inside_loop && is_java_primitive(bt)) {
 642         IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
 643         if (r_loop == u_loop || r_loop->is_member(u_loop)) {
 644           used_inside_loop = true;
 645         }
 646       }
 647     }
 648   }//for
 649   Node* bol = iff->in(1);
 650   assert(bol->Opcode() == Op_Bool, "");
 651   int cmp_op = bol->in(1)->Opcode();
 652   // It is expensive to generate flags from a float compare.
 653   // Avoid duplicated float compare.
 654   if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
 655 
 656   float infrequent_prob = PROB_UNLIKELY_MAG(3);
 657   // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
 658   if (used_inside_loop) {
 659     if (cost >= ConditionalMoveLimit) return NULL; // Too much goo
 660 
 661     // BlockLayoutByFrequency optimization moves infrequent branch
 662     // from hot path. No point in CMOV'ing in such case (110 is used
 663     // instead of 100 to take into account not exactness of float value).
 664     if (BlockLayoutByFrequency) {
 665       infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
 666     }
 667   }
 668   // Check for highly predictable branch.  No point in CMOV'ing if
 669   // we are going to predict accurately all the time.
 670   if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
 671     //keep going
 672   } else if (iff->_prob < infrequent_prob ||
 673       iff->_prob > (1.0f - infrequent_prob))
 674     return NULL;
 675 
 676   // --------------
 677   // Now replace all Phis with CMOV's
 678   Node *cmov_ctrl = iff->in(0);
 679   uint flip = (lp->Opcode() == Op_IfTrue);
 680   Node_List wq;
 681   while (1) {
 682     PhiNode* phi = NULL;
 683     for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
 684       Node *out = region->fast_out(i);
 685       if (out->is_Phi()) {
 686         phi = out->as_Phi();
 687         break;
 688       }
 689     }
 690     if (phi == NULL)  break;
 691     if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); }
 692     // Move speculative ops
 693     wq.push(phi);
 694     while (wq.size() > 0) {
 695       Node *n = wq.pop();
 696       for (uint j = 1; j < n->req(); j++) {
 697         Node* m = n->in(j);
 698         if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) {
 699 #ifndef PRODUCT
 700           if (PrintOpto && VerifyLoopOptimizations) {
 701             tty->print("  speculate: ");
 702             m->dump();
 703           }
 704 #endif
 705           set_ctrl(m, cmov_ctrl);
 706           wq.push(m);
 707         }
 708       }
 709     }
 710     Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
 711     register_new_node( cmov, cmov_ctrl );
 712     _igvn.replace_node( phi, cmov );
 713 #ifndef PRODUCT
 714     if (TraceLoopOpts) {
 715       tty->print("CMOV  ");
 716       r_loop->dump_head();
 717       if (Verbose) {
 718         bol->in(1)->dump(1);
 719         cmov->dump(1);
 720       }
 721     }
 722     if (VerifyLoopOptimizations) verify();
 723 #endif
 724   }
 725 
 726   // The useless CFG diamond will fold up later; see the optimization in
 727   // RegionNode::Ideal.
 728   _igvn._worklist.push(region);
 729 
 730   return iff->in(1);
 731 }
 732 
 733 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
 734   for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
 735     Node* u = m->fast_out(i);
 736     if (u->is_CFG()) {
 737       if (u->Opcode() == Op_NeverBranch) {
 738         u = ((NeverBranchNode*)u)->proj_out(0);
 739         enqueue_cfg_uses(u, wq);
 740       } else {
 741         wq.push(u);
 742       }
 743     }
 744   }
 745 }
 746 
 747 // Try moving a store out of a loop, right before the loop
 748 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
 749   // Store has to be first in the loop body
 750   IdealLoopTree *n_loop = get_loop(n_ctrl);
 751   if (n->is_Store() && n_loop != _ltree_root &&
 752       n_loop->is_loop() && n_loop->_head->is_Loop() &&
 753       n->in(0) != NULL) {
 754     Node* address = n->in(MemNode::Address);
 755     Node* value = n->in(MemNode::ValueIn);
 756     Node* mem = n->in(MemNode::Memory);
 757     IdealLoopTree* address_loop = get_loop(get_ctrl(address));
 758     IdealLoopTree* value_loop = get_loop(get_ctrl(value));
 759 
 760     // - address and value must be loop invariant
 761     // - memory must be a memory Phi for the loop
 762     // - Store must be the only store on this memory slice in the
 763     // loop: if there's another store following this one then value
 764     // written at iteration i by the second store could be overwritten
 765     // at iteration i+n by the first store: it's not safe to move the
 766     // first store out of the loop
 767     // - nothing must observe the memory Phi: it guarantees no read
 768     // before the store, we are also guaranteed the store post
 769     // dominates the loop head (ignoring a possible early
 770     // exit). Otherwise there would be extra Phi involved between the
 771     // loop's Phi and the store.
 772     // - there must be no early exit from the loop before the Store
 773     // (such an exit most of the time would be an extra use of the
 774     // memory Phi but sometimes is a bottom memory Phi that takes the
 775     // store as input).
 776 
 777     if (!n_loop->is_member(address_loop) &&
 778         !n_loop->is_member(value_loop) &&
 779         mem->is_Phi() && mem->in(0) == n_loop->_head &&
 780         mem->outcnt() == 1 &&
 781         mem->in(LoopNode::LoopBackControl) == n) {
 782 
 783       assert(n_loop->_tail != NULL, "need a tail");
 784       assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
 785 
 786       // Verify that there's no early exit of the loop before the store.
 787       bool ctrl_ok = false;
 788       {
 789         // Follow control from loop head until n, we exit the loop or
 790         // we reach the tail
 791         ResourceMark rm;
 792         Unique_Node_List wq;
 793         wq.push(n_loop->_head);
 794 
 795         for (uint next = 0; next < wq.size(); ++next) {
 796           Node *m = wq.at(next);
 797           if (m == n->in(0)) {
 798             ctrl_ok = true;
 799             continue;
 800           }
 801           assert(!has_ctrl(m), "should be CFG");
 802           if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
 803             ctrl_ok = false;
 804             break;
 805           }
 806           enqueue_cfg_uses(m, wq);
 807           if (wq.size() > 10) {
 808             ctrl_ok = false;
 809             break;
 810           }
 811         }
 812       }
 813       if (ctrl_ok) {
 814         // move the Store
 815         _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
 816         _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
 817         _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
 818         // Disconnect the phi now. An empty phi can confuse other
 819         // optimizations in this pass of loop opts.
 820         _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
 821         n_loop->_body.yank(mem);
 822 
 823         set_ctrl_and_loop(n, n->in(0));
 824 
 825         return n;
 826       }
 827     }
 828   }
 829   return NULL;
 830 }
 831 
 832 // Try moving a store out of a loop, right after the loop
 833 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
 834   if (n->is_Store() && n->in(0) != NULL) {
 835     Node *n_ctrl = get_ctrl(n);
 836     IdealLoopTree *n_loop = get_loop(n_ctrl);
 837     // Store must be in a loop
 838     if (n_loop != _ltree_root && !n_loop->_irreducible) {
 839       Node* address = n->in(MemNode::Address);
 840       Node* value = n->in(MemNode::ValueIn);
 841       IdealLoopTree* address_loop = get_loop(get_ctrl(address));
 842       // address must be loop invariant
 843       if (!n_loop->is_member(address_loop)) {
 844         // Store must be last on this memory slice in the loop and
 845         // nothing in the loop must observe it
 846         Node* phi = NULL;
 847         for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 848           Node* u = n->fast_out(i);
 849           if (has_ctrl(u)) { // control use?
 850             IdealLoopTree *u_loop = get_loop(get_ctrl(u));
 851             if (!n_loop->is_member(u_loop)) {
 852               continue;
 853             }
 854             if (u->is_Phi() && u->in(0) == n_loop->_head) {
 855               assert(_igvn.type(u) == Type::MEMORY, "bad phi");
 856               // multiple phis on the same slice are possible
 857               if (phi != NULL) {
 858                 return;
 859               }
 860               phi = u;
 861               continue;
 862             }
 863           }
 864           return;
 865         }
 866         if (phi != NULL) {
 867           // Nothing in the loop before the store (next iteration)
 868           // must observe the stored value
 869           bool mem_ok = true;
 870           {
 871             ResourceMark rm;
 872             Unique_Node_List wq;
 873             wq.push(phi);
 874             for (uint next = 0; next < wq.size() && mem_ok; ++next) {
 875               Node *m = wq.at(next);
 876               for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
 877                 Node* u = m->fast_out(i);
 878                 if (u->is_Store() || u->is_Phi()) {
 879                   if (u != n) {
 880                     wq.push(u);
 881                     mem_ok = (wq.size() <= 10);
 882                   }
 883                 } else {
 884                   mem_ok = false;
 885                   break;
 886                 }
 887               }
 888             }
 889           }
 890           if (mem_ok) {
 891             // Move the store out of the loop if the LCA of all
 892             // users (except for the phi) is outside the loop.
 893             Node* hook = new Node(1);
 894             _igvn.rehash_node_delayed(phi);
 895             int count = phi->replace_edge(n, hook);
 896             assert(count > 0, "inconsistent phi");
 897 
 898             // Compute latest point this store can go
 899             Node* lca = get_late_ctrl(n, get_ctrl(n));
 900             if (n_loop->is_member(get_loop(lca))) {
 901               // LCA is in the loop - bail out
 902               _igvn.replace_node(hook, n);
 903               return;
 904             }
 905 #ifdef ASSERT
 906             if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
 907               assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
 908               n_loop->_head->as_Loop()->verify_strip_mined(1);
 909               Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
 910               IdealLoopTree* outer_loop = get_loop(outer);
 911               assert(n_loop->_parent == outer_loop, "broken loop tree");
 912               assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
 913             }
 914 #endif
 915 
 916             // Move store out of the loop
 917             _igvn.replace_node(hook, n->in(MemNode::Memory));
 918             _igvn.replace_input_of(n, 0, lca);
 919             set_ctrl_and_loop(n, lca);
 920 
 921             // Disconnect the phi now. An empty phi can confuse other
 922             // optimizations in this pass of loop opts..
 923             if (phi->in(LoopNode::LoopBackControl) == phi) {
 924               _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
 925               n_loop->_body.yank(phi);
 926             }
 927           }
 928         }
 929       }
 930     }
 931   }
 932 }
 933 
 934 //------------------------------split_if_with_blocks_pre-----------------------
 935 // Do the real work in a non-recursive function.  Data nodes want to be
 936 // cloned in the pre-order so they can feed each other nicely.
 937 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
 938   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 939   Node* bs_res = bs->split_if_pre(this, n);
 940   if (bs_res != NULL) {
 941     return bs_res;
 942   }
 943   // Cloning these guys is unlikely to win
 944   int n_op = n->Opcode();
 945   if( n_op == Op_MergeMem ) return n;
 946   if( n->is_Proj() ) return n;
 947   // Do not clone-up CmpFXXX variations, as these are always
 948   // followed by a CmpI
 949   if( n->is_Cmp() ) return n;
 950   // Attempt to use a conditional move instead of a phi/branch
 951   if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
 952     Node *cmov = conditional_move( n );
 953     if( cmov ) return cmov;
 954   }
 955   if( n->is_CFG() || n->is_LoadStore() )
 956     return n;
 957   if( n_op == Op_Opaque1 ||     // Opaque nodes cannot be mod'd
 958       n_op == Op_Opaque2 ) {
 959     if( !C->major_progress() )   // If chance of no more loop opts...
 960       _igvn._worklist.push(n);  // maybe we'll remove them
 961     return n;
 962   }
 963 
 964   if( n->is_Con() ) return n;   // No cloning for Con nodes
 965 
 966   Node *n_ctrl = get_ctrl(n);
 967   if( !n_ctrl ) return n;       // Dead node
 968 
 969   Node* res = try_move_store_before_loop(n, n_ctrl);
 970   if (res != NULL) {
 971     return n;
 972   }
 973 
 974   // Attempt to remix address expressions for loop invariants
 975   Node *m = remix_address_expressions( n );
 976   if( m ) return m;
 977 
 978   if (n_op == Op_AddI) {
 979     Node *nn = convert_add_to_muladd( n );
 980     if ( nn ) return nn;
 981   }
 982 
 983   if (n->is_ConstraintCast()) {
 984     Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
 985     // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
 986     // Node control inputs don't necessarily agree with loop control info (due to
 987     // transformations happened in between), thus additional dominance check is needed
 988     // to keep loop info valid.
 989     if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
 990       _igvn.replace_node(n, dom_cast);
 991       return dom_cast;
 992     }
 993   }
 994 
 995   // Determine if the Node has inputs from some local Phi.
 996   // Returns the block to clone thru.
 997   Node *n_blk = has_local_phi_input( n );
 998   if( !n_blk ) return n;
 999 
1000   // Do not clone the trip counter through on a CountedLoop
1001   // (messes up the canonical shape).
1002   if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
1003 
1004   // Check for having no control input; not pinned.  Allow
1005   // dominating control.
1006   if (n->in(0)) {
1007     Node *dom = idom(n_blk);
1008     if (dom_lca(n->in(0), dom) != n->in(0)) {
1009       return n;
1010     }
1011   }
1012   // Policy: when is it profitable.  You must get more wins than
1013   // policy before it is considered profitable.  Policy is usually 0,
1014   // so 1 win is considered profitable.  Big merges will require big
1015   // cloning, so get a larger policy.
1016   int policy = n_blk->req() >> 2;
1017 
1018   // If the loop is a candidate for range check elimination,
1019   // delay splitting through it's phi until a later loop optimization
1020   if (n_blk->is_CountedLoop()) {
1021     IdealLoopTree *lp = get_loop(n_blk);
1022     if (lp && lp->_rce_candidate) {
1023       return n;
1024     }
1025   }
1026 
1027   // Use same limit as split_if_with_blocks_post
1028   if( C->live_nodes() > 35000 ) return n; // Method too big
1029 
1030   // Split 'n' through the merge point if it is profitable
1031   Node *phi = split_thru_phi( n, n_blk, policy );
1032   if (!phi) return n;
1033 
1034   // Found a Phi to split thru!
1035   // Replace 'n' with the new phi
1036   _igvn.replace_node( n, phi );
1037   // Moved a load around the loop, 'en-registering' something.
1038   if (n_blk->is_Loop() && n->is_Load() &&
1039       !phi->in(LoopNode::LoopBackControl)->is_Load())
1040     C->set_major_progress();
1041 
1042   return phi;
1043 }
1044 
1045 static bool merge_point_too_heavy(Compile* C, Node* region) {
1046   // Bail out if the region and its phis have too many users.
1047   int weight = 0;
1048   for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1049     weight += region->fast_out(i)->outcnt();
1050   }
1051   int nodes_left = C->max_node_limit() - C->live_nodes();
1052   if (weight * 8 > nodes_left) {
1053     if (PrintOpto) {
1054       tty->print_cr("*** Split-if bails out:  %d nodes, region weight %d", C->unique(), weight);
1055     }
1056     return true;
1057   } else {
1058     return false;
1059   }
1060 }
1061 
1062 static bool merge_point_safe(Node* region) {
1063   // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1064   // having a PhiNode input. This sidesteps the dangerous case where the split
1065   // ConvI2LNode may become TOP if the input Value() does not
1066   // overlap the ConvI2L range, leaving a node which may not dominate its
1067   // uses.
1068   // A better fix for this problem can be found in the BugTraq entry, but
1069   // expediency for Mantis demands this hack.
1070   // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop
1071   // split_if_with_blocks from splitting a block because we could not move around
1072   // the FastLockNode.
1073   for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1074     Node* n = region->fast_out(i);
1075     if (n->is_Phi()) {
1076       for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1077         Node* m = n->fast_out(j);
1078         if (m->is_FastLock())
1079           return false;
1080 #if INCLUDE_SHENANDOAHGC
1081         if (m->is_ShenandoahBarrier() && m->has_out_with(Op_FastLock)) {
1082           return false;
1083         }
1084 #endif
1085 #ifdef _LP64
1086         if (m->Opcode() == Op_ConvI2L)
1087           return false;
1088         if (m->is_CastII() && m->isa_CastII()->has_range_check()) {
1089           return false;
1090         }
1091 #endif
1092       }
1093     }
1094   }
1095   return true;
1096 }
1097 
1098 
1099 //------------------------------place_near_use---------------------------------
1100 // Place some computation next to use but not inside inner loops.
1101 // For inner loop uses move it to the preheader area.
1102 Node *PhaseIdealLoop::place_near_use(Node *useblock) const {
1103   IdealLoopTree *u_loop = get_loop( useblock );
1104   if (u_loop->_irreducible) {
1105     return useblock;
1106   }
1107   if (u_loop->_child) {
1108     if (useblock == u_loop->_head && u_loop->_head->is_OuterStripMinedLoop()) {
1109       return u_loop->_head->in(LoopNode::EntryControl);
1110     }
1111     return useblock;
1112   }
1113   return u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
1114 }
1115 
1116 
1117 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1118   if (!n->is_If() || n->is_CountedLoopEnd()) {
1119     return false;
1120   }
1121   if (!n->in(0)->is_Region()) {
1122     return false;
1123   }
1124   Node* region = n->in(0);
1125   Node* dom = idom(region);
1126   if (!dom->is_If() || dom->in(1) != n->in(1)) {
1127     return false;
1128   }
1129   IfNode* dom_if = dom->as_If();
1130   Node* proj_true = dom_if->proj_out(1);
1131   Node* proj_false = dom_if->proj_out(0);
1132 
1133   for (uint i = 1; i < region->req(); i++) {
1134     if (is_dominator(proj_true, region->in(i))) {
1135       continue;
1136     }
1137     if (is_dominator(proj_false, region->in(i))) {
1138       continue;
1139     }
1140     return false;
1141   }
1142 
1143   return true;
1144 }
1145 
1146 bool PhaseIdealLoop::can_split_if(Node *n_ctrl) {
1147   if (C->live_nodes() > 35000) {
1148     return false; // Method too big
1149   }
1150 
1151   // Do not do 'split-if' if irreducible loops are present.
1152   if (_has_irreducible_loops) {
1153     return false;
1154   }
1155 
1156   if (merge_point_too_heavy(C, n_ctrl)) {
1157     return false;
1158   }
1159 
1160   // Do not do 'split-if' if some paths are dead.  First do dead code
1161   // elimination and then see if its still profitable.
1162   for (uint i = 1; i < n_ctrl->req(); i++) {
1163     if (n_ctrl->in(i) == C->top()) {
1164       return false;
1165     }
1166   }
1167 
1168   // If trying to do a 'Split-If' at the loop head, it is only
1169   // profitable if the cmp folds up on BOTH paths.  Otherwise we
1170   // risk peeling a loop forever.
1171 
1172   // CNC - Disabled for now.  Requires careful handling of loop
1173   // body selection for the cloned code.  Also, make sure we check
1174   // for any input path not being in the same loop as n_ctrl.  For
1175   // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1176   // because the alternative loop entry points won't be converted
1177   // into LoopNodes.
1178   IdealLoopTree *n_loop = get_loop(n_ctrl);
1179   for (uint j = 1; j < n_ctrl->req(); j++) {
1180     if (get_loop(n_ctrl->in(j)) != n_loop) {
1181       return false;
1182     }
1183   }
1184 
1185   // Check for safety of the merge point.
1186   if (!merge_point_safe(n_ctrl)) {
1187     return false;
1188   }
1189 
1190   return true;
1191 }
1192 
1193 //------------------------------split_if_with_blocks_post----------------------
1194 // Do the real work in a non-recursive function.  CFG hackery wants to be
1195 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1196 // info.
1197 void PhaseIdealLoop::split_if_with_blocks_post(Node *n, bool last_round) {
1198 
1199   // Cloning Cmp through Phi's involves the split-if transform.
1200   // FastLock is not used by an If
1201   if (n->is_Cmp() && !n->is_FastLock() && !last_round) {
1202     Node *n_ctrl = get_ctrl(n);
1203     // Determine if the Node has inputs from some local Phi.
1204     // Returns the block to clone thru.
1205     Node *n_blk = has_local_phi_input(n);
1206     if (n_blk != n_ctrl) {
1207       return;
1208     }
1209 
1210     if (!can_split_if(n_ctrl)) {
1211       return;
1212     }
1213 
1214     if (n->outcnt() != 1) {
1215       return; // Multiple bool's from 1 compare?
1216     }
1217     Node *bol = n->unique_out();
1218     assert(bol->is_Bool(), "expect a bool here");
1219     if (bol->outcnt() != 1) {
1220       return;// Multiple branches from 1 compare?
1221     }
1222     Node *iff = bol->unique_out();
1223 
1224     // Check some safety conditions
1225     if (iff->is_If()) {        // Classic split-if?
1226       if (iff->in(0) != n_ctrl) {
1227         return; // Compare must be in same blk as if
1228       }
1229     } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1230       // Can't split CMove with different control edge.
1231       if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) {
1232         return;
1233       }
1234       if (get_ctrl(iff->in(2)) == n_ctrl ||
1235           get_ctrl(iff->in(3)) == n_ctrl) {
1236         return;                 // Inputs not yet split-up
1237       }
1238       if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1239         return;                 // Loop-invar test gates loop-varying CMOVE
1240       }
1241     } else {
1242       return;  // some other kind of node, such as an Allocate
1243     }
1244 
1245     // When is split-if profitable?  Every 'win' on means some control flow
1246     // goes dead, so it's almost always a win.
1247     int policy = 0;
1248     // Split compare 'n' through the merge point if it is profitable
1249     Node *phi = split_thru_phi( n, n_ctrl, policy);
1250     if (!phi) {
1251       return;
1252     }
1253 
1254     // Found a Phi to split thru!
1255     // Replace 'n' with the new phi
1256     _igvn.replace_node(n, phi);
1257 
1258     // Now split the bool up thru the phi
1259     Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1260     guarantee(bolphi != NULL, "null boolean phi node");
1261 
1262     _igvn.replace_node(bol, bolphi);
1263     assert(iff->in(1) == bolphi, "");
1264 
1265     if (bolphi->Value(&_igvn)->singleton()) {
1266       return;
1267     }
1268 
1269     // Conditional-move?  Must split up now
1270     if (!iff->is_If()) {
1271       Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1272       _igvn.replace_node(iff, cmovphi);
1273       return;
1274     }
1275 
1276     // Now split the IF
1277     do_split_if(iff);
1278     return;
1279   }
1280 
1281   // Two identical ifs back to back can be merged
1282   if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1283     Node *n_ctrl = n->in(0);
1284     PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1285     IfNode* dom_if = idom(n_ctrl)->as_If();
1286     Node* proj_true = dom_if->proj_out(1);
1287     Node* proj_false = dom_if->proj_out(0);
1288     Node* con_true = _igvn.makecon(TypeInt::ONE);
1289     Node* con_false = _igvn.makecon(TypeInt::ZERO);
1290 
1291     for (uint i = 1; i < n_ctrl->req(); i++) {
1292       if (is_dominator(proj_true, n_ctrl->in(i))) {
1293         bolphi->init_req(i, con_true);
1294       } else {
1295         assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1296         bolphi->init_req(i, con_false);
1297       }
1298     }
1299     register_new_node(bolphi, n_ctrl);
1300     _igvn.replace_input_of(n, 1, bolphi);
1301 
1302     // Now split the IF
1303     do_split_if(n);
1304     return;
1305   }
1306 
1307   // Check for an IF ready to split; one that has its
1308   // condition codes input coming from a Phi at the block start.
1309   int n_op = n->Opcode();
1310 
1311   // Check for an IF being dominated by another IF same test
1312   if (n_op == Op_If ||
1313       n_op == Op_RangeCheck) {
1314     Node *bol = n->in(1);
1315     uint max = bol->outcnt();
1316     // Check for same test used more than once?
1317     if (max > 1 && bol->is_Bool()) {
1318       // Search up IDOMs to see if this IF is dominated.
1319       Node *cutoff = get_ctrl(bol);
1320 
1321       // Now search up IDOMs till cutoff, looking for a dominating test
1322       Node *prevdom = n;
1323       Node *dom = idom(prevdom);
1324       while (dom != cutoff) {
1325         if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) {
1326           // Replace the dominated test with an obvious true or false.
1327           // Place it on the IGVN worklist for later cleanup.
1328           C->set_major_progress();
1329           dominated_by(prevdom, n, false, true);
1330 #ifndef PRODUCT
1331           if( VerifyLoopOptimizations ) verify();
1332 #endif
1333           return;
1334         }
1335         prevdom = dom;
1336         dom = idom(prevdom);
1337       }
1338     }
1339   }
1340 
1341   // See if a shared loop-varying computation has no loop-varying uses.
1342   // Happens if something is only used for JVM state in uncommon trap exits,
1343   // like various versions of induction variable+offset.  Clone the
1344   // computation per usage to allow it to sink out of the loop.
1345   if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
1346     Node *n_ctrl = get_ctrl(n);
1347     IdealLoopTree *n_loop = get_loop(n_ctrl);
1348     if( n_loop != _ltree_root ) {
1349       DUIterator_Fast imax, i = n->fast_outs(imax);
1350       for (; i < imax; i++) {
1351         Node* u = n->fast_out(i);
1352         if( !has_ctrl(u) )     break; // Found control user
1353         IdealLoopTree *u_loop = get_loop(get_ctrl(u));
1354         if( u_loop == n_loop ) break; // Found loop-varying use
1355         if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
1356         if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
1357       }
1358       bool did_break = (i < imax);  // Did we break out of the previous loop?
1359       if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
1360         Node *late_load_ctrl = NULL;
1361         if (n->is_Load()) {
1362           // If n is a load, get and save the result from get_late_ctrl(),
1363           // to be later used in calculating the control for n's clones.
1364           clear_dom_lca_tags();
1365           late_load_ctrl = get_late_ctrl(n, n_ctrl);
1366         }
1367         // If n is a load, and the late control is the same as the current
1368         // control, then the cloning of n is a pointless exercise, because
1369         // GVN will ensure that we end up where we started.
1370         if (!n->is_Load() || late_load_ctrl != n_ctrl) {
1371           BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1372           for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
1373             Node *u = n->last_out(j); // Clone private computation per use
1374             _igvn.rehash_node_delayed(u);
1375             Node *x = n->clone(); // Clone computation
1376             Node *x_ctrl = NULL;
1377             if( u->is_Phi() ) {
1378               // Replace all uses of normal nodes.  Replace Phi uses
1379               // individually, so the separate Nodes can sink down
1380               // different paths.
1381               uint k = 1;
1382               while( u->in(k) != n ) k++;
1383               u->set_req( k, x );
1384               // x goes next to Phi input path
1385               x_ctrl = u->in(0)->in(k);
1386               --j;
1387             } else {              // Normal use
1388               // Replace all uses
1389               for( uint k = 0; k < u->req(); k++ ) {
1390                 if( u->in(k) == n ) {
1391                   u->set_req( k, x );
1392                   --j;
1393                 }
1394               }
1395               x_ctrl = get_ctrl(u);
1396             }
1397 
1398             // Find control for 'x' next to use but not inside inner loops.
1399             // For inner loop uses get the preheader area.
1400             x_ctrl = place_near_use(x_ctrl);
1401 
1402             if (bs->sink_node(this, n, x, x_ctrl, n_ctrl)) {
1403               continue;
1404             }
1405 
1406             if (n->is_Load()) {
1407               // For loads, add a control edge to a CFG node outside of the loop
1408               // to force them to not combine and return back inside the loop
1409               // during GVN optimization (4641526).
1410               //
1411               // Because we are setting the actual control input, factor in
1412               // the result from get_late_ctrl() so we respect any
1413               // anti-dependences. (6233005).
1414               x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
1415 
1416               // Don't allow the control input to be a CFG splitting node.
1417               // Such nodes should only have ProjNodes as outs, e.g. IfNode
1418               // should only have IfTrueNode and IfFalseNode (4985384).
1419               x_ctrl = find_non_split_ctrl(x_ctrl);
1420               assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1421 
1422               x->set_req(0, x_ctrl);
1423             }
1424             register_new_node(x, x_ctrl);
1425 
1426             // Some institutional knowledge is needed here: 'x' is
1427             // yanked because if the optimizer runs GVN on it all the
1428             // cloned x's will common up and undo this optimization and
1429             // be forced back in the loop.  This is annoying because it
1430             // makes +VerifyOpto report false-positives on progress.  I
1431             // tried setting control edges on the x's to force them to
1432             // not combine, but the matching gets worried when it tries
1433             // to fold a StoreP and an AddP together (as part of an
1434             // address expression) and the AddP and StoreP have
1435             // different controls.
1436             if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x);
1437           }
1438           _igvn.remove_dead_node(n);
1439         }
1440       }
1441     }
1442   }
1443 
1444   try_move_store_after_loop(n);
1445 
1446   // Check for Opaque2's who's loop has disappeared - who's input is in the
1447   // same loop nest as their output.  Remove 'em, they are no longer useful.
1448   if( n_op == Op_Opaque2 &&
1449       n->in(1) != NULL &&
1450       get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1451     _igvn.replace_node( n, n->in(1) );
1452   }
1453 
1454 #if INCLUDE_ZGC
1455   if (UseZGC) {
1456     ZBarrierSetC2::loop_optimize_gc_barrier(this, n, last_round);
1457   }
1458 #endif
1459 }
1460 
1461 //------------------------------split_if_with_blocks---------------------------
1462 // Check for aggressive application of 'split-if' optimization,
1463 // using basic block level info.
1464 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack, bool last_round) {
1465   Node *n = C->root();
1466   visited.set(n->_idx); // first, mark node as visited
1467   // Do pre-visit work for root
1468   n = split_if_with_blocks_pre( n );
1469   uint cnt = n->outcnt();
1470   uint i   = 0;
1471   while (true) {
1472     // Visit all children
1473     if (i < cnt) {
1474       Node* use = n->raw_out(i);
1475       ++i;
1476       if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1477         // Now do pre-visit work for this use
1478         use = split_if_with_blocks_pre( use );
1479         nstack.push(n, i); // Save parent and next use's index.
1480         n   = use;         // Process all children of current use.
1481         cnt = use->outcnt();
1482         i   = 0;
1483       }
1484     }
1485     else {
1486       // All of n's children have been processed, complete post-processing.
1487       if (cnt != 0 && !n->is_Con()) {
1488         assert(has_node(n), "no dead nodes");
1489         split_if_with_blocks_post( n, last_round );
1490       }
1491       if (nstack.is_empty()) {
1492         // Finished all nodes on stack.
1493         break;
1494       }
1495       // Get saved parent node and next use's index. Visit the rest of uses.
1496       n   = nstack.node();
1497       cnt = n->outcnt();
1498       i   = nstack.index();
1499       nstack.pop();
1500     }
1501   }
1502 }
1503 
1504 
1505 //=============================================================================
1506 //
1507 //                   C L O N E   A   L O O P   B O D Y
1508 //
1509 
1510 //------------------------------clone_iff--------------------------------------
1511 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1512 // "Nearly" because all Nodes have been cloned from the original in the loop,
1513 // but the fall-in edges to the Cmp are different.  Clone bool/Cmp pairs
1514 // through the Phi recursively, and return a Bool.
1515 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) {
1516 
1517   // Convert this Phi into a Phi merging Bools
1518   uint i;
1519   for (i = 1; i < phi->req(); i++) {
1520     Node *b = phi->in(i);
1521     if (b->is_Phi()) {
1522       _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop));
1523     } else {
1524       assert(b->is_Bool() || b->Opcode() == Op_Opaque4, "");
1525     }
1526   }
1527 
1528   Node* n = phi->in(1);
1529   Node* sample_opaque = NULL;
1530   Node *sample_bool = NULL;
1531   if (n->Opcode() == Op_Opaque4) {
1532     sample_opaque = n;
1533     sample_bool = n->in(1);
1534     assert(sample_bool->is_Bool(), "wrong type");
1535   } else {
1536     sample_bool = n;
1537   }
1538   Node *sample_cmp = sample_bool->in(1);
1539 
1540   // Make Phis to merge the Cmp's inputs.
1541   PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
1542   PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
1543   for (i = 1; i < phi->req(); i++) {
1544     Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
1545     Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
1546     phi1->set_req(i, n1);
1547     phi2->set_req(i, n2);
1548     phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1549     phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1550   }
1551   // See if these Phis have been made before.
1552   // Register with optimizer
1553   Node *hit1 = _igvn.hash_find_insert(phi1);
1554   if (hit1) {                   // Hit, toss just made Phi
1555     _igvn.remove_dead_node(phi1); // Remove new phi
1556     assert(hit1->is_Phi(), "" );
1557     phi1 = (PhiNode*)hit1;      // Use existing phi
1558   } else {                      // Miss
1559     _igvn.register_new_node_with_optimizer(phi1);
1560   }
1561   Node *hit2 = _igvn.hash_find_insert(phi2);
1562   if (hit2) {                   // Hit, toss just made Phi
1563     _igvn.remove_dead_node(phi2); // Remove new phi
1564     assert(hit2->is_Phi(), "" );
1565     phi2 = (PhiNode*)hit2;      // Use existing phi
1566   } else {                      // Miss
1567     _igvn.register_new_node_with_optimizer(phi2);
1568   }
1569   // Register Phis with loop/block info
1570   set_ctrl(phi1, phi->in(0));
1571   set_ctrl(phi2, phi->in(0));
1572   // Make a new Cmp
1573   Node *cmp = sample_cmp->clone();
1574   cmp->set_req(1, phi1);
1575   cmp->set_req(2, phi2);
1576   _igvn.register_new_node_with_optimizer(cmp);
1577   set_ctrl(cmp, phi->in(0));
1578 
1579   // Make a new Bool
1580   Node *b = sample_bool->clone();
1581   b->set_req(1,cmp);
1582   _igvn.register_new_node_with_optimizer(b);
1583   set_ctrl(b, phi->in(0));
1584 
1585   if (sample_opaque != NULL) {
1586     Node* opaque = sample_opaque->clone();
1587     opaque->set_req(1, b);
1588     _igvn.register_new_node_with_optimizer(opaque);
1589     set_ctrl(opaque, phi->in(0));
1590     return opaque;
1591   }
1592 
1593   assert(b->is_Bool(), "");
1594   return b;
1595 }
1596 
1597 //------------------------------clone_bool-------------------------------------
1598 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1599 // "Nearly" because all Nodes have been cloned from the original in the loop,
1600 // but the fall-in edges to the Cmp are different.  Clone bool/Cmp pairs
1601 // through the Phi recursively, and return a Bool.
1602 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1603   uint i;
1604   // Convert this Phi into a Phi merging Bools
1605   for( i = 1; i < phi->req(); i++ ) {
1606     Node *b = phi->in(i);
1607     if( b->is_Phi() ) {
1608       _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop ));
1609     } else {
1610       assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1611     }
1612   }
1613 
1614   Node *sample_cmp = phi->in(1);
1615 
1616   // Make Phis to merge the Cmp's inputs.
1617   PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
1618   PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
1619   for( uint j = 1; j < phi->req(); j++ ) {
1620     Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1621     Node *n1, *n2;
1622     if( cmp_top->is_Cmp() ) {
1623       n1 = cmp_top->in(1);
1624       n2 = cmp_top->in(2);
1625     } else {
1626       n1 = n2 = cmp_top;
1627     }
1628     phi1->set_req( j, n1 );
1629     phi2->set_req( j, n2 );
1630     phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1631     phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1632   }
1633 
1634   // See if these Phis have been made before.
1635   // Register with optimizer
1636   Node *hit1 = _igvn.hash_find_insert(phi1);
1637   if( hit1 ) {                  // Hit, toss just made Phi
1638     _igvn.remove_dead_node(phi1); // Remove new phi
1639     assert( hit1->is_Phi(), "" );
1640     phi1 = (PhiNode*)hit1;      // Use existing phi
1641   } else {                      // Miss
1642     _igvn.register_new_node_with_optimizer(phi1);
1643   }
1644   Node *hit2 = _igvn.hash_find_insert(phi2);
1645   if( hit2 ) {                  // Hit, toss just made Phi
1646     _igvn.remove_dead_node(phi2); // Remove new phi
1647     assert( hit2->is_Phi(), "" );
1648     phi2 = (PhiNode*)hit2;      // Use existing phi
1649   } else {                      // Miss
1650     _igvn.register_new_node_with_optimizer(phi2);
1651   }
1652   // Register Phis with loop/block info
1653   set_ctrl(phi1, phi->in(0));
1654   set_ctrl(phi2, phi->in(0));
1655   // Make a new Cmp
1656   Node *cmp = sample_cmp->clone();
1657   cmp->set_req( 1, phi1 );
1658   cmp->set_req( 2, phi2 );
1659   _igvn.register_new_node_with_optimizer(cmp);
1660   set_ctrl(cmp, phi->in(0));
1661 
1662   assert( cmp->is_Cmp(), "" );
1663   return (CmpNode*)cmp;
1664 }
1665 
1666 //------------------------------sink_use---------------------------------------
1667 // If 'use' was in the loop-exit block, it now needs to be sunk
1668 // below the post-loop merge point.
1669 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1670   if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1671     set_ctrl(use, post_loop);
1672     for (DUIterator j = use->outs(); use->has_out(j); j++)
1673       sink_use(use->out(j), post_loop);
1674   }
1675 }
1676 
1677 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1678                                                  IdealLoopTree* loop, IdealLoopTree* outer_loop,
1679                                                  Node_List*& split_if_set, Node_List*& split_bool_set,
1680                                                  Node_List*& split_cex_set, Node_List& worklist,
1681                                                  uint new_counter, CloneLoopMode mode) {
1682   Node* nnn = old_new[old->_idx];
1683   // Copy uses to a worklist, so I can munge the def-use info
1684   // with impunity.
1685   for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1686     worklist.push(old->fast_out(j));
1687 
1688   while( worklist.size() ) {
1689     Node *use = worklist.pop();
1690     if (!has_node(use))  continue; // Ignore dead nodes
1691     if (use->in(0) == C->top())  continue;
1692     IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1693     // Check for data-use outside of loop - at least one of OLD or USE
1694     // must not be a CFG node.
1695 #ifdef ASSERT
1696     if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) {
1697       Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
1698       assert(mode == ControlAroundStripMined && use == sfpt, "missed a node");
1699     }
1700 #endif
1701     if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
1702 
1703       // If the Data use is an IF, that means we have an IF outside of the
1704       // loop that is switching on a condition that is set inside of the
1705       // loop.  Happens if people set a loop-exit flag; then test the flag
1706       // in the loop to break the loop, then test is again outside of the
1707       // loop to determine which way the loop exited.
1708       // Loop predicate If node connects to Bool node through Opaque1 node.
1709       if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) {
1710         // Since this code is highly unlikely, we lazily build the worklist
1711         // of such Nodes to go split.
1712         if (!split_if_set) {
1713           ResourceArea *area = Thread::current()->resource_area();
1714           split_if_set = new Node_List(area);
1715         }
1716         split_if_set->push(use);
1717       }
1718       if (use->is_Bool()) {
1719         if (!split_bool_set) {
1720           ResourceArea *area = Thread::current()->resource_area();
1721           split_bool_set = new Node_List(area);
1722         }
1723         split_bool_set->push(use);
1724       }
1725       if (use->Opcode() == Op_CreateEx) {
1726         if (!split_cex_set) {
1727           ResourceArea *area = Thread::current()->resource_area();
1728           split_cex_set = new Node_List(area);
1729         }
1730         split_cex_set->push(use);
1731       }
1732 
1733 
1734       // Get "block" use is in
1735       uint idx = 0;
1736       while( use->in(idx) != old ) idx++;
1737       Node *prev = use->is_CFG() ? use : get_ctrl(use);
1738       assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
1739       Node *cfg = prev->_idx >= new_counter
1740         ? prev->in(2)
1741         : idom(prev);
1742       if( use->is_Phi() )     // Phi use is in prior block
1743         cfg = prev->in(idx);  // NOT in block of Phi itself
1744       if (cfg->is_top()) {    // Use is dead?
1745         _igvn.replace_input_of(use, idx, C->top());
1746         continue;
1747       }
1748 
1749       // If use is referenced through control edge... (idx == 0)
1750       if (mode == IgnoreStripMined && idx == 0) {
1751         LoopNode *head = loop->_head->as_Loop();
1752         if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
1753           // That node is outside the inner loop, leave it outside the
1754           // outer loop as well to not confuse verification code.
1755           assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
1756           _igvn.replace_input_of(use, 0, head->outer_loop_exit());
1757           continue;
1758         }
1759       }
1760 
1761       while(!outer_loop->is_member(get_loop(cfg))) {
1762         prev = cfg;
1763         cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1764       }
1765       // If the use occurs after merging several exits from the loop, then
1766       // old value must have dominated all those exits.  Since the same old
1767       // value was used on all those exits we did not need a Phi at this
1768       // merge point.  NOW we do need a Phi here.  Each loop exit value
1769       // is now merged with the peeled body exit; each exit gets its own
1770       // private Phi and those Phis need to be merged here.
1771       Node *phi;
1772       if( prev->is_Region() ) {
1773         if( idx == 0 ) {      // Updating control edge?
1774           phi = prev;         // Just use existing control
1775         } else {              // Else need a new Phi
1776           phi = PhiNode::make( prev, old );
1777           // Now recursively fix up the new uses of old!
1778           for( uint i = 1; i < prev->req(); i++ ) {
1779             worklist.push(phi); // Onto worklist once for each 'old' input
1780           }
1781         }
1782       } else {
1783         // Get new RegionNode merging old and new loop exits
1784         prev = old_new[prev->_idx];
1785         assert( prev, "just made this in step 7" );
1786         if( idx == 0) {      // Updating control edge?
1787           phi = prev;         // Just use existing control
1788         } else {              // Else need a new Phi
1789           // Make a new Phi merging data values properly
1790           phi = PhiNode::make( prev, old );
1791           phi->set_req( 1, nnn );
1792         }
1793       }
1794       // If inserting a new Phi, check for prior hits
1795       if( idx != 0 ) {
1796         Node *hit = _igvn.hash_find_insert(phi);
1797         if( hit == NULL ) {
1798           _igvn.register_new_node_with_optimizer(phi); // Register new phi
1799         } else {                                      // or
1800           // Remove the new phi from the graph and use the hit
1801           _igvn.remove_dead_node(phi);
1802           phi = hit;                                  // Use existing phi
1803         }
1804         set_ctrl(phi, prev);
1805       }
1806       // Make 'use' use the Phi instead of the old loop body exit value
1807       _igvn.replace_input_of(use, idx, phi);
1808       if( use->_idx >= new_counter ) { // If updating new phis
1809         // Not needed for correctness, but prevents a weak assert
1810         // in AddPNode from tripping (when we end up with different
1811         // base & derived Phis that will become the same after
1812         // IGVN does CSE).
1813         Node *hit = _igvn.hash_find_insert(use);
1814         if( hit )             // Go ahead and re-hash for hits.
1815           _igvn.replace_node( use, hit );
1816       }
1817 
1818       // If 'use' was in the loop-exit block, it now needs to be sunk
1819       // below the post-loop merge point.
1820       sink_use( use, prev );
1821     }
1822   }
1823 }
1824 
1825 static void clone_outer_loop_helper(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
1826                                     const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
1827                                     bool check_old_new) {
1828   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1829     Node* u = n->fast_out(j);
1830     assert(check_old_new || old_new[u->_idx] == NULL, "shouldn't have been cloned");
1831     if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == NULL)) {
1832       Node* c = phase->get_ctrl(u);
1833       IdealLoopTree* u_loop = phase->get_loop(c);
1834       assert(!loop->is_member(u_loop), "can be in outer loop or out of both loops only");
1835       if (outer_loop->is_member(u_loop)) {
1836         wq.push(u);
1837       }
1838     }
1839   }
1840 }
1841 
1842 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1843                                       IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1844                                       Node_List& extra_data_nodes) {
1845   if (head->is_strip_mined() && mode != IgnoreStripMined) {
1846     CountedLoopNode* cl = head->as_CountedLoop();
1847     Node* l = cl->outer_loop();
1848     Node* tail = cl->outer_loop_tail();
1849     IfNode* le = cl->outer_loop_end();
1850     Node* sfpt = cl->outer_safepoint();
1851     CountedLoopEndNode* cle = cl->loopexit();
1852     CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
1853     CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
1854     Node* cle_out = cle->proj_out(false);
1855 
1856     Node* new_sfpt = NULL;
1857     Node* new_cle_out = cle_out->clone();
1858     old_new.map(cle_out->_idx, new_cle_out);
1859     if (mode == CloneIncludesStripMined) {
1860       // clone outer loop body
1861       Node* new_l = l->clone();
1862       Node* new_tail = tail->clone();
1863       IfNode* new_le = le->clone()->as_If();
1864       new_sfpt = sfpt->clone();
1865 
1866       set_loop(new_l, outer_loop->_parent);
1867       set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
1868       set_loop(new_cle_out, outer_loop->_parent);
1869       set_idom(new_cle_out, new_cle, dd);
1870       set_loop(new_sfpt, outer_loop->_parent);
1871       set_idom(new_sfpt, new_cle_out, dd);
1872       set_loop(new_le, outer_loop->_parent);
1873       set_idom(new_le, new_sfpt, dd);
1874       set_loop(new_tail, outer_loop->_parent);
1875       set_idom(new_tail, new_le, dd);
1876       set_idom(new_cl, new_l, dd);
1877 
1878       old_new.map(l->_idx, new_l);
1879       old_new.map(tail->_idx, new_tail);
1880       old_new.map(le->_idx, new_le);
1881       old_new.map(sfpt->_idx, new_sfpt);
1882 
1883       new_l->set_req(LoopNode::LoopBackControl, new_tail);
1884       new_l->set_req(0, new_l);
1885       new_tail->set_req(0, new_le);
1886       new_le->set_req(0, new_sfpt);
1887       new_sfpt->set_req(0, new_cle_out);
1888       new_cle_out->set_req(0, new_cle);
1889       new_cl->set_req(LoopNode::EntryControl, new_l);
1890 
1891       _igvn.register_new_node_with_optimizer(new_l);
1892       _igvn.register_new_node_with_optimizer(new_tail);
1893       _igvn.register_new_node_with_optimizer(new_le);
1894     } else {
1895       Node *newhead = old_new[loop->_head->_idx];
1896       newhead->as_Loop()->clear_strip_mined();
1897       _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
1898       set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1899     }
1900     // Look at data node that were assigned a control in the outer
1901     // loop: they are kept in the outer loop by the safepoint so start
1902     // from the safepoint node's inputs.
1903     IdealLoopTree* outer_loop = get_loop(l);
1904     Node_Stack stack(2);
1905     stack.push(sfpt, 1);
1906     uint new_counter = C->unique();
1907     while (stack.size() > 0) {
1908       Node* n = stack.node();
1909       uint i = stack.index();
1910       while (i < n->req() &&
1911              (n->in(i) == NULL ||
1912               !has_ctrl(n->in(i)) ||
1913               get_loop(get_ctrl(n->in(i))) != outer_loop ||
1914               (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
1915         i++;
1916       }
1917       if (i < n->req()) {
1918         stack.set_index(i+1);
1919         stack.push(n->in(i), 0);
1920       } else {
1921         assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
1922         Node* m = n == sfpt ? new_sfpt : n->clone();
1923         if (m != NULL) {
1924           for (uint i = 0; i < n->req(); i++) {
1925             if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) {
1926               m->set_req(i, old_new[m->in(i)->_idx]);
1927             }
1928           }
1929         } else {
1930           assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
1931         }
1932         if (n != sfpt) {
1933           extra_data_nodes.push(n);
1934           _igvn.register_new_node_with_optimizer(m);
1935           assert(get_ctrl(n) == cle_out, "what other control?");
1936           set_ctrl(m, new_cle_out);
1937           old_new.map(n->_idx, m);
1938         }
1939         stack.pop();
1940       }
1941     }
1942     if (mode == CloneIncludesStripMined) {
1943       _igvn.register_new_node_with_optimizer(new_sfpt);
1944       _igvn.register_new_node_with_optimizer(new_cle_out);
1945     }
1946     // Some other transformation may have pessimistically assign some
1947     // data nodes to the outer loop. Set their control so they are out
1948     // of the outer loop.
1949     ResourceMark rm;
1950     Unique_Node_List wq;
1951     for (uint i = 0; i < extra_data_nodes.size(); i++) {
1952       Node* old = extra_data_nodes.at(i);
1953       clone_outer_loop_helper(old, loop, outer_loop, old_new, wq, this, true);
1954     }
1955     Node* new_ctrl = cl->outer_loop_exit();
1956     assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
1957     for (uint i = 0; i < wq.size(); i++) {
1958       Node* n = wq.at(i);
1959       set_ctrl(n, new_ctrl);
1960       clone_outer_loop_helper(n, loop, outer_loop, old_new, wq, this, false);
1961     }
1962   } else {
1963     Node *newhead = old_new[loop->_head->_idx];
1964     set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1965   }
1966 }
1967 
1968 //------------------------------clone_loop-------------------------------------
1969 //
1970 //                   C L O N E   A   L O O P   B O D Y
1971 //
1972 // This is the basic building block of the loop optimizations.  It clones an
1973 // entire loop body.  It makes an old_new loop body mapping; with this mapping
1974 // you can find the new-loop equivalent to an old-loop node.  All new-loop
1975 // nodes are exactly equal to their old-loop counterparts, all edges are the
1976 // same.  All exits from the old-loop now have a RegionNode that merges the
1977 // equivalent new-loop path.  This is true even for the normal "loop-exit"
1978 // condition.  All uses of loop-invariant old-loop values now come from (one
1979 // or more) Phis that merge their new-loop equivalents.
1980 //
1981 // This operation leaves the graph in an illegal state: there are two valid
1982 // control edges coming from the loop pre-header to both loop bodies.  I'll
1983 // definitely have to hack the graph after running this transform.
1984 //
1985 // From this building block I will further edit edges to perform loop peeling
1986 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
1987 //
1988 // Parameter side_by_size_idom:
1989 //   When side_by_size_idom is NULL, the dominator tree is constructed for
1990 //      the clone loop to dominate the original.  Used in construction of
1991 //      pre-main-post loop sequence.
1992 //   When nonnull, the clone and original are side-by-side, both are
1993 //      dominated by the side_by_side_idom node.  Used in construction of
1994 //      unswitched loops.
1995 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
1996                                 CloneLoopMode mode, Node* side_by_side_idom) {
1997 
1998   LoopNode* head = loop->_head->as_Loop();
1999   head->verify_strip_mined(1);
2000 
2001   if (C->do_vector_loop() && PrintOpto) {
2002     const char* mname = C->method()->name()->as_quoted_ascii();
2003     if (mname != NULL) {
2004       tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2005     }
2006   }
2007 
2008   CloneMap& cm = C->clone_map();
2009   Dict* dict = cm.dict();
2010   if (C->do_vector_loop()) {
2011     cm.set_clone_idx(cm.max_gen()+1);
2012 #ifndef PRODUCT
2013     if (PrintOpto) {
2014       tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2015       loop->dump_head();
2016     }
2017 #endif
2018   }
2019 
2020   // Step 1: Clone the loop body.  Make the old->new mapping.
2021   uint i;
2022   for( i = 0; i < loop->_body.size(); i++ ) {
2023     Node *old = loop->_body.at(i);
2024     Node *nnn = old->clone();
2025     old_new.map( old->_idx, nnn );
2026     if (C->do_vector_loop()) {
2027       cm.verify_insert_and_clone(old, nnn, cm.clone_idx());
2028     }
2029     _igvn.register_new_node_with_optimizer(nnn);
2030   }
2031 
2032   IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2033 
2034   // Step 2: Fix the edges in the new body.  If the old input is outside the
2035   // loop use it.  If the old input is INside the loop, use the corresponding
2036   // new node instead.
2037   for( i = 0; i < loop->_body.size(); i++ ) {
2038     Node *old = loop->_body.at(i);
2039     Node *nnn = old_new[old->_idx];
2040     // Fix CFG/Loop controlling the new node
2041     if (has_ctrl(old)) {
2042       set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2043     } else {
2044       set_loop(nnn, outer_loop->_parent);
2045       if (old->outcnt() > 0) {
2046         set_idom( nnn, old_new[idom(old)->_idx], dd );
2047       }
2048     }
2049     // Correct edges to the new node
2050     for( uint j = 0; j < nnn->req(); j++ ) {
2051         Node *n = nnn->in(j);
2052         if( n ) {
2053           IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
2054           if( loop->is_member( old_in_loop ) )
2055             nnn->set_req(j, old_new[n->_idx]);
2056         }
2057     }
2058     _igvn.hash_find_insert(nnn);
2059   }
2060 
2061   ResourceArea *area = Thread::current()->resource_area();
2062   Node_List extra_data_nodes(area); // data nodes in the outer strip mined loop
2063   clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2064 
2065   // Step 3: Now fix control uses.  Loop varying control uses have already
2066   // been fixed up (as part of all input edges in Step 2).  Loop invariant
2067   // control uses must be either an IfFalse or an IfTrue.  Make a merge
2068   // point to merge the old and new IfFalse/IfTrue nodes; make the use
2069   // refer to this.
2070   Node_List worklist(area);
2071   uint new_counter = C->unique();
2072   for( i = 0; i < loop->_body.size(); i++ ) {
2073     Node* old = loop->_body.at(i);
2074     if( !old->is_CFG() ) continue;
2075 
2076     // Copy uses to a worklist, so I can munge the def-use info
2077     // with impunity.
2078     for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2079       worklist.push(old->fast_out(j));
2080 
2081     while( worklist.size() ) {  // Visit all uses
2082       Node *use = worklist.pop();
2083       if (!has_node(use))  continue; // Ignore dead nodes
2084       IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2085       if( !loop->is_member( use_loop ) && use->is_CFG() ) {
2086         // Both OLD and USE are CFG nodes here.
2087         assert( use->is_Proj(), "" );
2088         Node* nnn = old_new[old->_idx];
2089 
2090         Node* newuse = NULL;
2091         if (head->is_strip_mined() && mode != IgnoreStripMined) {
2092           CountedLoopNode* cl = head->as_CountedLoop();
2093           CountedLoopEndNode* cle = cl->loopexit();
2094           Node* cle_out = cle->proj_out_or_null(false);
2095           if (use == cle_out) {
2096             IfNode* le = cl->outer_loop_end();
2097             use = le->proj_out(false);
2098             use_loop = get_loop(use);
2099             if (mode == CloneIncludesStripMined) {
2100               nnn = old_new[le->_idx];
2101             } else {
2102               newuse = old_new[cle_out->_idx];
2103             }
2104           }
2105         }
2106         if (newuse == NULL) {
2107           newuse = use->clone();
2108         }
2109 
2110         // Clone the loop exit control projection
2111         if (C->do_vector_loop()) {
2112           cm.verify_insert_and_clone(use, newuse, cm.clone_idx());
2113         }
2114         newuse->set_req(0,nnn);
2115         _igvn.register_new_node_with_optimizer(newuse);
2116         set_loop(newuse, use_loop);
2117         set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2118 
2119         // We need a Region to merge the exit from the peeled body and the
2120         // exit from the old loop body.
2121         RegionNode *r = new RegionNode(3);
2122         // Map the old use to the new merge point
2123         old_new.map( use->_idx, r );
2124         uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
2125         assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
2126 
2127         // The original user of 'use' uses 'r' instead.
2128         for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2129           Node* useuse = use->last_out(l);
2130           _igvn.rehash_node_delayed(useuse);
2131           uint uses_found = 0;
2132           if( useuse->in(0) == use ) {
2133             useuse->set_req(0, r);
2134             uses_found++;
2135             if( useuse->is_CFG() ) {
2136               assert( dom_depth(useuse) > dd_r, "" );
2137               set_idom(useuse, r, dom_depth(useuse));
2138             }
2139           }
2140           for( uint k = 1; k < useuse->req(); k++ ) {
2141             if( useuse->in(k) == use ) {
2142               useuse->set_req(k, r);
2143               uses_found++;
2144               if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2145                 assert(dom_depth(useuse) > dd_r , "");
2146                 set_idom(useuse, r, dom_depth(useuse));
2147               }
2148             }
2149           }
2150           l -= uses_found;    // we deleted 1 or more copies of this edge
2151         }
2152 
2153         // Now finish up 'r'
2154         r->set_req( 1, newuse );
2155         r->set_req( 2,    use );
2156         _igvn.register_new_node_with_optimizer(r);
2157         set_loop(r, use_loop);
2158         set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
2159       } // End of if a loop-exit test
2160     }
2161   }
2162 
2163   // Step 4: If loop-invariant use is not control, it must be dominated by a
2164   // loop exit IfFalse/IfTrue.  Find "proper" loop exit.  Make a Region
2165   // there if needed.  Make a Phi there merging old and new used values.
2166   Node_List *split_if_set = NULL;
2167   Node_List *split_bool_set = NULL;
2168   Node_List *split_cex_set = NULL;
2169   for( i = 0; i < loop->_body.size(); i++ ) {
2170     Node* old = loop->_body.at(i);
2171     clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2172                                 split_bool_set, split_cex_set, worklist, new_counter,
2173                                 mode);
2174   }
2175 
2176   for (i = 0; i < extra_data_nodes.size(); i++) {
2177     Node* old = extra_data_nodes.at(i);
2178     clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2179                                 split_bool_set, split_cex_set, worklist, new_counter,
2180                                 mode);
2181   }
2182 
2183   // Check for IFs that need splitting/cloning.  Happens if an IF outside of
2184   // the loop uses a condition set in the loop.  The original IF probably
2185   // takes control from one or more OLD Regions (which in turn get from NEW
2186   // Regions).  In any case, there will be a set of Phis for each merge point
2187   // from the IF up to where the original BOOL def exists the loop.
2188   if (split_if_set) {
2189     while (split_if_set->size()) {
2190       Node *iff = split_if_set->pop();
2191       if (iff->in(1)->is_Phi()) {
2192         Node *b = clone_iff(iff->in(1)->as_Phi(), loop);
2193         _igvn.replace_input_of(iff, 1, b);
2194       }
2195     }
2196   }
2197   if (split_bool_set) {
2198     while (split_bool_set->size()) {
2199       Node *b = split_bool_set->pop();
2200       Node *phi = b->in(1);
2201       assert(phi->is_Phi(), "");
2202       CmpNode *cmp = clone_bool((PhiNode*)phi, loop);
2203       _igvn.replace_input_of(b, 1, cmp);
2204     }
2205   }
2206   if (split_cex_set) {
2207     while (split_cex_set->size()) {
2208       Node *b = split_cex_set->pop();
2209       assert(b->in(0)->is_Region(), "");
2210       assert(b->in(1)->is_Phi(), "");
2211       assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2212       split_up(b, b->in(0), NULL);
2213     }
2214   }
2215 
2216 }
2217 
2218 
2219 //---------------------- stride_of_possible_iv -------------------------------------
2220 // Looks for an iff/bool/comp with one operand of the compare
2221 // being a cycle involving an add and a phi,
2222 // with an optional truncation (left-shift followed by a right-shift)
2223 // of the add. Returns zero if not an iv.
2224 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2225   Node* trunc1 = NULL;
2226   Node* trunc2 = NULL;
2227   const TypeInt* ttype = NULL;
2228   if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
2229     return 0;
2230   }
2231   BoolNode* bl = iff->in(1)->as_Bool();
2232   Node* cmp = bl->in(1);
2233   if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2234     return 0;
2235   }
2236   // Must have an invariant operand
2237   if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2238     return 0;
2239   }
2240   Node* add2 = NULL;
2241   Node* cmp1 = cmp->in(1);
2242   if (cmp1->is_Phi()) {
2243     // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2244     Node* phi = cmp1;
2245     for (uint i = 1; i < phi->req(); i++) {
2246       Node* in = phi->in(i);
2247       Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2248                                 &trunc1, &trunc2, &ttype);
2249       if (add && add->in(1) == phi) {
2250         add2 = add->in(2);
2251         break;
2252       }
2253     }
2254   } else {
2255     // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2256     Node* addtrunc = cmp1;
2257     Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2258                                 &trunc1, &trunc2, &ttype);
2259     if (add && add->in(1)->is_Phi()) {
2260       Node* phi = add->in(1);
2261       for (uint i = 1; i < phi->req(); i++) {
2262         if (phi->in(i) == addtrunc) {
2263           add2 = add->in(2);
2264           break;
2265         }
2266       }
2267     }
2268   }
2269   if (add2 != NULL) {
2270     const TypeInt* add2t = _igvn.type(add2)->is_int();
2271     if (add2t->is_con()) {
2272       return add2t->get_con();
2273     }
2274   }
2275   return 0;
2276 }
2277 
2278 
2279 //---------------------- stay_in_loop -------------------------------------
2280 // Return the (unique) control output node that's in the loop (if it exists.)
2281 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2282   Node* unique = NULL;
2283   if (!n) return NULL;
2284   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2285     Node* use = n->fast_out(i);
2286     if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2287       if (unique != NULL) {
2288         return NULL;
2289       }
2290       unique = use;
2291     }
2292   }
2293   return unique;
2294 }
2295 
2296 //------------------------------ register_node -------------------------------------
2297 // Utility to register node "n" with PhaseIdealLoop
2298 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
2299   _igvn.register_new_node_with_optimizer(n);
2300   loop->_body.push(n);
2301   if (n->is_CFG()) {
2302     set_loop(n, loop);
2303     set_idom(n, pred, ddepth);
2304   } else {
2305     set_ctrl(n, pred);
2306   }
2307 }
2308 
2309 //------------------------------ proj_clone -------------------------------------
2310 // Utility to create an if-projection
2311 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2312   ProjNode* c = p->clone()->as_Proj();
2313   c->set_req(0, iff);
2314   return c;
2315 }
2316 
2317 //------------------------------ short_circuit_if -------------------------------------
2318 // Force the iff control output to be the live_proj
2319 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2320   guarantee(live_proj != NULL, "null projection");
2321   int proj_con = live_proj->_con;
2322   assert(proj_con == 0 || proj_con == 1, "false or true projection");
2323   Node *con = _igvn.intcon(proj_con);
2324   set_ctrl(con, C->root());
2325   if (iff) {
2326     iff->set_req(1, con);
2327   }
2328   return con;
2329 }
2330 
2331 //------------------------------ insert_if_before_proj -------------------------------------
2332 // Insert a new if before an if projection (* - new node)
2333 //
2334 // before
2335 //           if(test)
2336 //           /     \
2337 //          v       v
2338 //    other-proj   proj (arg)
2339 //
2340 // after
2341 //           if(test)
2342 //           /     \
2343 //          /       v
2344 //         |      * proj-clone
2345 //         v          |
2346 //    other-proj      v
2347 //                * new_if(relop(cmp[IU](left,right)))
2348 //                  /  \
2349 //                 v    v
2350 //         * new-proj  proj
2351 //         (returned)
2352 //
2353 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2354   IfNode* iff = proj->in(0)->as_If();
2355   IdealLoopTree *loop = get_loop(proj);
2356   ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2357   int ddepth = dom_depth(proj);
2358 
2359   _igvn.rehash_node_delayed(iff);
2360   _igvn.rehash_node_delayed(proj);
2361 
2362   proj->set_req(0, NULL);  // temporary disconnect
2363   ProjNode* proj2 = proj_clone(proj, iff);
2364   register_node(proj2, loop, iff, ddepth);
2365 
2366   Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2367   register_node(cmp, loop, proj2, ddepth);
2368 
2369   BoolNode* bol = new BoolNode(cmp, relop);
2370   register_node(bol, loop, proj2, ddepth);
2371 
2372   int opcode = iff->Opcode();
2373   assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2374   IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt):
2375     new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt);
2376   register_node(new_if, loop, proj2, ddepth);
2377 
2378   proj->set_req(0, new_if); // reattach
2379   set_idom(proj, new_if, ddepth);
2380 
2381   ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2382   guarantee(new_exit != NULL, "null exit node");
2383   register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2384 
2385   return new_exit;
2386 }
2387 
2388 //------------------------------ insert_region_before_proj -------------------------------------
2389 // Insert a region before an if projection (* - new node)
2390 //
2391 // before
2392 //           if(test)
2393 //          /      |
2394 //         v       |
2395 //       proj      v
2396 //               other-proj
2397 //
2398 // after
2399 //           if(test)
2400 //          /      |
2401 //         v       |
2402 // * proj-clone    v
2403 //         |     other-proj
2404 //         v
2405 // * new-region
2406 //         |
2407 //         v
2408 // *      dum_if
2409 //       /     \
2410 //      v       \
2411 // * dum-proj    v
2412 //              proj
2413 //
2414 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
2415   IfNode* iff = proj->in(0)->as_If();
2416   IdealLoopTree *loop = get_loop(proj);
2417   ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2418   int ddepth = dom_depth(proj);
2419 
2420   _igvn.rehash_node_delayed(iff);
2421   _igvn.rehash_node_delayed(proj);
2422 
2423   proj->set_req(0, NULL);  // temporary disconnect
2424   ProjNode* proj2 = proj_clone(proj, iff);
2425   register_node(proj2, loop, iff, ddepth);
2426 
2427   RegionNode* reg = new RegionNode(2);
2428   reg->set_req(1, proj2);
2429   register_node(reg, loop, iff, ddepth);
2430 
2431   IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
2432   register_node(dum_if, loop, reg, ddepth);
2433 
2434   proj->set_req(0, dum_if); // reattach
2435   set_idom(proj, dum_if, ddepth);
2436 
2437   ProjNode* dum_proj = proj_clone(other_proj, dum_if);
2438   register_node(dum_proj, loop, dum_if, ddepth);
2439 
2440   return reg;
2441 }
2442 
2443 //------------------------------ insert_cmpi_loop_exit -------------------------------------
2444 // Clone a signed compare loop exit from an unsigned compare and
2445 // insert it before the unsigned cmp on the stay-in-loop path.
2446 // All new nodes inserted in the dominator tree between the original
2447 // if and it's projections.  The original if test is replaced with
2448 // a constant to force the stay-in-loop path.
2449 //
2450 // This is done to make sure that the original if and it's projections
2451 // still dominate the same set of control nodes, that the ctrl() relation
2452 // from data nodes to them is preserved, and that their loop nesting is
2453 // preserved.
2454 //
2455 // before
2456 //          if(i <u limit)    unsigned compare loop exit
2457 //         /       |
2458 //        v        v
2459 //   exit-proj   stay-in-loop-proj
2460 //
2461 // after
2462 //          if(stay-in-loop-const)  original if
2463 //         /       |
2464 //        /        v
2465 //       /  if(i <  limit)    new signed test
2466 //      /  /       |
2467 //     /  /        v
2468 //    /  /  if(i <u limit)    new cloned unsigned test
2469 //   /  /   /      |
2470 //   v  v  v       |
2471 //    region       |
2472 //        |        |
2473 //      dum-if     |
2474 //     /  |        |
2475 // ether  |        |
2476 //        v        v
2477 //   exit-proj   stay-in-loop-proj
2478 //
2479 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
2480   const bool Signed   = true;
2481   const bool Unsigned = false;
2482 
2483   BoolNode* bol = if_cmpu->in(1)->as_Bool();
2484   if (bol->_test._test != BoolTest::lt) return NULL;
2485   CmpNode* cmpu = bol->in(1)->as_Cmp();
2486   if (cmpu->Opcode() != Op_CmpU) return NULL;
2487   int stride = stride_of_possible_iv(if_cmpu);
2488   if (stride == 0) return NULL;
2489 
2490   Node* lp_proj = stay_in_loop(if_cmpu, loop);
2491   guarantee(lp_proj != NULL, "null loop node");
2492 
2493   ProjNode* lp_continue = lp_proj->as_Proj();
2494   ProjNode* lp_exit     = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
2495 
2496   Node* limit = NULL;
2497   if (stride > 0) {
2498     limit = cmpu->in(2);
2499   } else {
2500     limit = _igvn.makecon(TypeInt::ZERO);
2501     set_ctrl(limit, C->root());
2502   }
2503   // Create a new region on the exit path
2504   RegionNode* reg = insert_region_before_proj(lp_exit);
2505   guarantee(reg != NULL, "null region node");
2506 
2507   // Clone the if-cmpu-true-false using a signed compare
2508   BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
2509   ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
2510   reg->add_req(cmpi_exit);
2511 
2512   // Clone the if-cmpu-true-false
2513   BoolTest::mask rel_u = bol->_test._test;
2514   ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
2515   reg->add_req(cmpu_exit);
2516 
2517   // Force original if to stay in loop.
2518   short_circuit_if(if_cmpu, lp_continue);
2519 
2520   return cmpi_exit->in(0)->as_If();
2521 }
2522 
2523 //------------------------------ remove_cmpi_loop_exit -------------------------------------
2524 // Remove a previously inserted signed compare loop exit.
2525 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
2526   Node* lp_proj = stay_in_loop(if_cmp, loop);
2527   assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
2528          stay_in_loop(lp_proj, loop)->is_If() &&
2529          stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
2530   Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
2531   set_ctrl(con, C->root());
2532   if_cmp->set_req(1, con);
2533 }
2534 
2535 //------------------------------ scheduled_nodelist -------------------------------------
2536 // Create a post order schedule of nodes that are in the
2537 // "member" set.  The list is returned in "sched".
2538 // The first node in "sched" is the loop head, followed by
2539 // nodes which have no inputs in the "member" set, and then
2540 // followed by the nodes that have an immediate input dependence
2541 // on a node in "sched".
2542 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
2543 
2544   assert(member.test(loop->_head->_idx), "loop head must be in member set");
2545   Arena *a = Thread::current()->resource_area();
2546   VectorSet visited(a);
2547   Node_Stack nstack(a, loop->_body.size());
2548 
2549   Node* n  = loop->_head;  // top of stack is cached in "n"
2550   uint idx = 0;
2551   visited.set(n->_idx);
2552 
2553   // Initially push all with no inputs from within member set
2554   for(uint i = 0; i < loop->_body.size(); i++ ) {
2555     Node *elt = loop->_body.at(i);
2556     if (member.test(elt->_idx)) {
2557       bool found = false;
2558       for (uint j = 0; j < elt->req(); j++) {
2559         Node* def = elt->in(j);
2560         if (def && member.test(def->_idx) && def != elt) {
2561           found = true;
2562           break;
2563         }
2564       }
2565       if (!found && elt != loop->_head) {
2566         nstack.push(n, idx);
2567         n = elt;
2568         assert(!visited.test(n->_idx), "not seen yet");
2569         visited.set(n->_idx);
2570       }
2571     }
2572   }
2573 
2574   // traverse out's that are in the member set
2575   while (true) {
2576     if (idx < n->outcnt()) {
2577       Node* use = n->raw_out(idx);
2578       idx++;
2579       if (!visited.test_set(use->_idx)) {
2580         if (member.test(use->_idx)) {
2581           nstack.push(n, idx);
2582           n = use;
2583           idx = 0;
2584         }
2585       }
2586     } else {
2587       // All outputs processed
2588       sched.push(n);
2589       if (nstack.is_empty()) break;
2590       n   = nstack.node();
2591       idx = nstack.index();
2592       nstack.pop();
2593     }
2594   }
2595 }
2596 
2597 
2598 //------------------------------ has_use_in_set -------------------------------------
2599 // Has a use in the vector set
2600 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
2601   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2602     Node* use = n->fast_out(j);
2603     if (vset.test(use->_idx)) {
2604       return true;
2605     }
2606   }
2607   return false;
2608 }
2609 
2610 
2611 //------------------------------ has_use_internal_to_set -------------------------------------
2612 // Has use internal to the vector set (ie. not in a phi at the loop head)
2613 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
2614   Node* head  = loop->_head;
2615   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2616     Node* use = n->fast_out(j);
2617     if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
2618       return true;
2619     }
2620   }
2621   return false;
2622 }
2623 
2624 
2625 //------------------------------ clone_for_use_outside_loop -------------------------------------
2626 // clone "n" for uses that are outside of loop
2627 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
2628   int cloned = 0;
2629   assert(worklist.size() == 0, "should be empty");
2630   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2631     Node* use = n->fast_out(j);
2632     if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
2633       worklist.push(use);
2634     }
2635   }
2636   while( worklist.size() ) {
2637     Node *use = worklist.pop();
2638     if (!has_node(use) || use->in(0) == C->top()) continue;
2639     uint j;
2640     for (j = 0; j < use->req(); j++) {
2641       if (use->in(j) == n) break;
2642     }
2643     assert(j < use->req(), "must be there");
2644 
2645     // clone "n" and insert it between the inputs of "n" and the use outside the loop
2646     Node* n_clone = n->clone();
2647     _igvn.replace_input_of(use, j, n_clone);
2648     cloned++;
2649     Node* use_c;
2650     if (!use->is_Phi()) {
2651       use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
2652     } else {
2653       // Use in a phi is considered a use in the associated predecessor block
2654       use_c = use->in(0)->in(j);
2655     }
2656     set_ctrl(n_clone, use_c);
2657     assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
2658     get_loop(use_c)->_body.push(n_clone);
2659     _igvn.register_new_node_with_optimizer(n_clone);
2660 #if !defined(PRODUCT)
2661     if (TracePartialPeeling) {
2662       tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
2663     }
2664 #endif
2665   }
2666   return cloned;
2667 }
2668 
2669 
2670 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
2671 // clone "n" for special uses that are in the not_peeled region.
2672 // If these def-uses occur in separate blocks, the code generator
2673 // marks the method as not compilable.  For example, if a "BoolNode"
2674 // is in a different basic block than the "IfNode" that uses it, then
2675 // the compilation is aborted in the code generator.
2676 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
2677                                                         VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
2678   if (n->is_Phi() || n->is_Load()) {
2679     return;
2680   }
2681   assert(worklist.size() == 0, "should be empty");
2682   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2683     Node* use = n->fast_out(j);
2684     if ( not_peel.test(use->_idx) &&
2685          (use->is_If() || use->is_CMove() || use->is_Bool()) &&
2686          use->in(1) == n)  {
2687       worklist.push(use);
2688     }
2689   }
2690   if (worklist.size() > 0) {
2691     // clone "n" and insert it between inputs of "n" and the use
2692     Node* n_clone = n->clone();
2693     loop->_body.push(n_clone);
2694     _igvn.register_new_node_with_optimizer(n_clone);
2695     set_ctrl(n_clone, get_ctrl(n));
2696     sink_list.push(n_clone);
2697     not_peel <<= n_clone->_idx;  // add n_clone to not_peel set.
2698 #if !defined(PRODUCT)
2699     if (TracePartialPeeling) {
2700       tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
2701     }
2702 #endif
2703     while( worklist.size() ) {
2704       Node *use = worklist.pop();
2705       _igvn.rehash_node_delayed(use);
2706       for (uint j = 1; j < use->req(); j++) {
2707         if (use->in(j) == n) {
2708           use->set_req(j, n_clone);
2709         }
2710       }
2711     }
2712   }
2713 }
2714 
2715 
2716 //------------------------------ insert_phi_for_loop -------------------------------------
2717 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
2718 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
2719   Node *phi = PhiNode::make(lp, back_edge_val);
2720   phi->set_req(LoopNode::EntryControl, lp_entry_val);
2721   // Use existing phi if it already exists
2722   Node *hit = _igvn.hash_find_insert(phi);
2723   if( hit == NULL ) {
2724     _igvn.register_new_node_with_optimizer(phi);
2725     set_ctrl(phi, lp);
2726   } else {
2727     // Remove the new phi from the graph and use the hit
2728     _igvn.remove_dead_node(phi);
2729     phi = hit;
2730   }
2731   _igvn.replace_input_of(use, idx, phi);
2732 }
2733 
2734 #ifdef ASSERT
2735 //------------------------------ is_valid_loop_partition -------------------------------------
2736 // Validate the loop partition sets: peel and not_peel
2737 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
2738                                               VectorSet& not_peel ) {
2739   uint i;
2740   // Check that peel_list entries are in the peel set
2741   for (i = 0; i < peel_list.size(); i++) {
2742     if (!peel.test(peel_list.at(i)->_idx)) {
2743       return false;
2744     }
2745   }
2746   // Check at loop members are in one of peel set or not_peel set
2747   for (i = 0; i < loop->_body.size(); i++ ) {
2748     Node *def  = loop->_body.at(i);
2749     uint di = def->_idx;
2750     // Check that peel set elements are in peel_list
2751     if (peel.test(di)) {
2752       if (not_peel.test(di)) {
2753         return false;
2754       }
2755       // Must be in peel_list also
2756       bool found = false;
2757       for (uint j = 0; j < peel_list.size(); j++) {
2758         if (peel_list.at(j)->_idx == di) {
2759           found = true;
2760           break;
2761         }
2762       }
2763       if (!found) {
2764         return false;
2765       }
2766     } else if (not_peel.test(di)) {
2767       if (peel.test(di)) {
2768         return false;
2769       }
2770     } else {
2771       return false;
2772     }
2773   }
2774   return true;
2775 }
2776 
2777 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2778 // Ensure a use outside of loop is of the right form
2779 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2780   Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2781   return (use->is_Phi() &&
2782           use_c->is_Region() && use_c->req() == 3 &&
2783           (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2784            use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2785            use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2786           loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2787 }
2788 
2789 //------------------------------ is_valid_clone_loop_form -------------------------------------
2790 // Ensure that all uses outside of loop are of the right form
2791 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2792                                                uint orig_exit_idx, uint clone_exit_idx) {
2793   uint len = peel_list.size();
2794   for (uint i = 0; i < len; i++) {
2795     Node *def = peel_list.at(i);
2796 
2797     for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2798       Node *use = def->fast_out(j);
2799       Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2800       if (!loop->is_member(get_loop(use_c))) {
2801         // use is not in the loop, check for correct structure
2802         if (use->in(0) == def) {
2803           // Okay
2804         } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2805           return false;
2806         }
2807       }
2808     }
2809   }
2810   return true;
2811 }
2812 #endif
2813 
2814 //------------------------------ partial_peel -------------------------------------
2815 // Partially peel (aka loop rotation) the top portion of a loop (called
2816 // the peel section below) by cloning it and placing one copy just before
2817 // the new loop head and the other copy at the bottom of the new loop.
2818 //
2819 //    before                       after                where it came from
2820 //
2821 //    stmt1                        stmt1
2822 //  loop:                          stmt2                     clone
2823 //    stmt2                        if condA goto exitA       clone
2824 //    if condA goto exitA        new_loop:                   new
2825 //    stmt3                        stmt3                     clone
2826 //    if !condB goto loop          if condB goto exitB       clone
2827 //  exitB:                         stmt2                     orig
2828 //    stmt4                        if !condA goto new_loop   orig
2829 //  exitA:                         goto exitA
2830 //                               exitB:
2831 //                                 stmt4
2832 //                               exitA:
2833 //
2834 // Step 1: find the cut point: an exit test on probable
2835 //         induction variable.
2836 // Step 2: schedule (with cloning) operations in the peel
2837 //         section that can be executed after the cut into
2838 //         the section that is not peeled.  This may need
2839 //         to clone operations into exit blocks.  For
2840 //         instance, a reference to A[i] in the not-peel
2841 //         section and a reference to B[i] in an exit block
2842 //         may cause a left-shift of i by 2 to be placed
2843 //         in the peel block.  This step will clone the left
2844 //         shift into the exit block and sink the left shift
2845 //         from the peel to the not-peel section.
2846 // Step 3: clone the loop, retarget the control, and insert
2847 //         phis for values that are live across the new loop
2848 //         head.  This is very dependent on the graph structure
2849 //         from clone_loop.  It creates region nodes for
2850 //         exit control and associated phi nodes for values
2851 //         flow out of the loop through that exit.  The region
2852 //         node is dominated by the clone's control projection.
2853 //         So the clone's peel section is placed before the
2854 //         new loop head, and the clone's not-peel section is
2855 //         forms the top part of the new loop.  The original
2856 //         peel section forms the tail of the new loop.
2857 // Step 4: update the dominator tree and recompute the
2858 //         dominator depth.
2859 //
2860 //                   orig
2861 //
2862 //                   stmt1
2863 //                     |
2864 //                     v
2865 //               loop predicate
2866 //                     |
2867 //                     v
2868 //                   loop<----+
2869 //                     |      |
2870 //                   stmt2    |
2871 //                     |      |
2872 //                     v      |
2873 //                    ifA     |
2874 //                   / |      |
2875 //                  v  v      |
2876 //               false true   ^  <-- last_peel
2877 //               /     |      |
2878 //              /   ===|==cut |
2879 //             /     stmt3    |  <-- first_not_peel
2880 //            /        |      |
2881 //            |        v      |
2882 //            v       ifB     |
2883 //          exitA:   / \      |
2884 //                  /   \     |
2885 //                 v     v    |
2886 //               false true   |
2887 //               /       \    |
2888 //              /         ----+
2889 //             |
2890 //             v
2891 //           exitB:
2892 //           stmt4
2893 //
2894 //
2895 //            after clone loop
2896 //
2897 //                   stmt1
2898 //                     |
2899 //                     v
2900 //               loop predicate
2901 //                 /       \
2902 //        clone   /         \   orig
2903 //               /           \
2904 //              /             \
2905 //             v               v
2906 //   +---->loop                loop<----+
2907 //   |      |                    |      |
2908 //   |    stmt2                stmt2    |
2909 //   |      |                    |      |
2910 //   |      v                    v      |
2911 //   |      ifA                 ifA     |
2912 //   |      | \                / |      |
2913 //   |      v  v              v  v      |
2914 //   ^    true  false      false true   ^  <-- last_peel
2915 //   |      |   ^   \       /    |      |
2916 //   | cut==|==  \   \     /  ===|==cut |
2917 //   |    stmt3   \   \   /    stmt3    |  <-- first_not_peel
2918 //   |      |    dom   | |       |      |
2919 //   |      v      \  1v v2      v      |
2920 //   |      ifB     regionA     ifB     |
2921 //   |      / \        |       / \      |
2922 //   |     /   \       v      /   \     |
2923 //   |    v     v    exitA:  v     v    |
2924 //   |    true  false      false true   |
2925 //   |    /     ^   \      /       \    |
2926 //   +----       \   \    /         ----+
2927 //               dom  \  /
2928 //                 \  1v v2
2929 //                  regionB
2930 //                     |
2931 //                     v
2932 //                   exitB:
2933 //                   stmt4
2934 //
2935 //
2936 //           after partial peel
2937 //
2938 //                  stmt1
2939 //                     |
2940 //                     v
2941 //               loop predicate
2942 //                 /
2943 //        clone   /             orig
2944 //               /          TOP
2945 //              /             \
2946 //             v               v
2947 //    TOP->loop                loop----+
2948 //          |                    |      |
2949 //        stmt2                stmt2    |
2950 //          |                    |      |
2951 //          v                    v      |
2952 //          ifA                 ifA     |
2953 //          | \                / |      |
2954 //          v  v              v  v      |
2955 //        true  false      false true   |     <-- last_peel
2956 //          |   ^   \       /    +------|---+
2957 //  +->newloop   \   \     /  === ==cut |   |
2958 //  |     stmt3   \   \   /     TOP     |   |
2959 //  |       |    dom   | |      stmt3   |   | <-- first_not_peel
2960 //  |       v      \  1v v2      v      |   |
2961 //  |       ifB     regionA     ifB     ^   v
2962 //  |       / \        |       / \      |   |
2963 //  |      /   \       v      /   \     |   |
2964 //  |     v     v    exitA:  v     v    |   |
2965 //  |     true  false      false true   |   |
2966 //  |     /     ^   \      /       \    |   |
2967 //  |    |       \   \    /         v   |   |
2968 //  |    |       dom  \  /         TOP  |   |
2969 //  |    |         \  1v v2             |   |
2970 //  ^    v          regionB             |   |
2971 //  |    |             |                |   |
2972 //  |    |             v                ^   v
2973 //  |    |           exitB:             |   |
2974 //  |    |           stmt4              |   |
2975 //  |    +------------>-----------------+   |
2976 //  |                                       |
2977 //  +-----------------<---------------------+
2978 //
2979 //
2980 //              final graph
2981 //
2982 //                  stmt1
2983 //                    |
2984 //                    v
2985 //               loop predicate
2986 //                    |
2987 //                    v
2988 //                  stmt2 clone
2989 //                    |
2990 //                    v
2991 //         ........> ifA clone
2992 //         :        / |
2993 //        dom      /  |
2994 //         :      v   v
2995 //         :  false   true
2996 //         :  |       |
2997 //         :  |       v
2998 //         :  |    newloop<-----+
2999 //         :  |        |        |
3000 //         :  |     stmt3 clone |
3001 //         :  |        |        |
3002 //         :  |        v        |
3003 //         :  |       ifB       |
3004 //         :  |      / \        |
3005 //         :  |     v   v       |
3006 //         :  |  false true     |
3007 //         :  |   |     |       |
3008 //         :  |   v    stmt2    |
3009 //         :  | exitB:  |       |
3010 //         :  | stmt4   v       |
3011 //         :  |       ifA orig  |
3012 //         :  |      /  \       |
3013 //         :  |     /    \      |
3014 //         :  |    v     v      |
3015 //         :  |  false  true    |
3016 //         :  |  /        \     |
3017 //         :  v  v         -----+
3018 //          RegionA
3019 //             |
3020 //             v
3021 //           exitA
3022 //
3023 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3024 
3025   assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3026   if (!loop->_head->is_Loop()) {
3027     return false;  }
3028 
3029   LoopNode *head  = loop->_head->as_Loop();
3030 
3031   if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3032     return false;
3033   }
3034 
3035   // Check for complex exit control
3036   for(uint ii = 0; ii < loop->_body.size(); ii++ ) {
3037     Node *n = loop->_body.at(ii);
3038     int opc = n->Opcode();
3039     if (n->is_Call()        ||
3040         opc == Op_Catch     ||
3041         opc == Op_CatchProj ||
3042         opc == Op_Jump      ||
3043         opc == Op_JumpProj) {
3044 #if !defined(PRODUCT)
3045       if (TracePartialPeeling) {
3046         tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3047       }
3048 #endif
3049       return false;
3050     }
3051   }
3052 
3053   int dd = dom_depth(head);
3054 
3055   // Step 1: find cut point
3056 
3057   // Walk up dominators to loop head looking for first loop exit
3058   // which is executed on every path thru loop.
3059   IfNode *peel_if = NULL;
3060   IfNode *peel_if_cmpu = NULL;
3061 
3062   Node *iff = loop->tail();
3063   while( iff != head ) {
3064     if( iff->is_If() ) {
3065       Node *ctrl = get_ctrl(iff->in(1));
3066       if (ctrl->is_top()) return false; // Dead test on live IF.
3067       // If loop-varying exit-test, check for induction variable
3068       if( loop->is_member(get_loop(ctrl)) &&
3069           loop->is_loop_exit(iff) &&
3070           is_possible_iv_test(iff)) {
3071         Node* cmp = iff->in(1)->in(1);
3072         if (cmp->Opcode() == Op_CmpI) {
3073           peel_if = iff->as_If();
3074         } else {
3075           assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3076           peel_if_cmpu = iff->as_If();
3077         }
3078       }
3079     }
3080     iff = idom(iff);
3081   }
3082   // Prefer signed compare over unsigned compare.
3083   IfNode* new_peel_if = NULL;
3084   if (peel_if == NULL) {
3085     if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
3086       return false;   // No peel point found
3087     }
3088     new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3089     if (new_peel_if == NULL) {
3090       return false;   // No peel point found
3091     }
3092     peel_if = new_peel_if;
3093   }
3094   Node* last_peel        = stay_in_loop(peel_if, loop);
3095   Node* first_not_peeled = stay_in_loop(last_peel, loop);
3096   if (first_not_peeled == NULL || first_not_peeled == head) {
3097     return false;
3098   }
3099 
3100 #if !defined(PRODUCT)
3101   if (TraceLoopOpts) {
3102     tty->print("PartialPeel  ");
3103     loop->dump_head();
3104   }
3105 
3106   if (TracePartialPeeling) {
3107     tty->print_cr("before partial peel one iteration");
3108     Node_List wl;
3109     Node* t = head->in(2);
3110     while (true) {
3111       wl.push(t);
3112       if (t == head) break;
3113       t = idom(t);
3114     }
3115     while (wl.size() > 0) {
3116       Node* tt = wl.pop();
3117       tt->dump();
3118       if (tt == last_peel) tty->print_cr("-- cut --");
3119     }
3120   }
3121 #endif
3122   ResourceArea *area = Thread::current()->resource_area();
3123   VectorSet peel(area);
3124   VectorSet not_peel(area);
3125   Node_List peel_list(area);
3126   Node_List worklist(area);
3127   Node_List sink_list(area);
3128 
3129   // Set of cfg nodes to peel are those that are executable from
3130   // the head through last_peel.
3131   assert(worklist.size() == 0, "should be empty");
3132   worklist.push(head);
3133   peel.set(head->_idx);
3134   while (worklist.size() > 0) {
3135     Node *n = worklist.pop();
3136     if (n != last_peel) {
3137       for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3138         Node* use = n->fast_out(j);
3139         if (use->is_CFG() &&
3140             loop->is_member(get_loop(use)) &&
3141             !peel.test_set(use->_idx)) {
3142           worklist.push(use);
3143         }
3144       }
3145     }
3146   }
3147 
3148   // Set of non-cfg nodes to peel are those that are control
3149   // dependent on the cfg nodes.
3150   uint i;
3151   for(i = 0; i < loop->_body.size(); i++ ) {
3152     Node *n = loop->_body.at(i);
3153     Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3154     if (peel.test(n_c->_idx)) {
3155       peel.set(n->_idx);
3156     } else {
3157       not_peel.set(n->_idx);
3158     }
3159   }
3160 
3161   // Step 2: move operations from the peeled section down into the
3162   //         not-peeled section
3163 
3164   // Get a post order schedule of nodes in the peel region
3165   // Result in right-most operand.
3166   scheduled_nodelist(loop, peel, peel_list );
3167 
3168   assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3169 
3170   // For future check for too many new phis
3171   uint old_phi_cnt = 0;
3172   for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3173     Node* use = head->fast_out(j);
3174     if (use->is_Phi()) old_phi_cnt++;
3175   }
3176 
3177 #if !defined(PRODUCT)
3178   if (TracePartialPeeling) {
3179     tty->print_cr("\npeeled list");
3180   }
3181 #endif
3182 
3183   // Evacuate nodes in peel region into the not_peeled region if possible
3184   uint new_phi_cnt = 0;
3185   uint cloned_for_outside_use = 0;
3186   for (i = 0; i < peel_list.size();) {
3187     Node* n = peel_list.at(i);
3188 #if !defined(PRODUCT)
3189   if (TracePartialPeeling) n->dump();
3190 #endif
3191     bool incr = true;
3192     if ( !n->is_CFG() ) {
3193 
3194       if ( has_use_in_set(n, not_peel) ) {
3195 
3196         // If not used internal to the peeled region,
3197         // move "n" from peeled to not_peeled region.
3198 
3199         if ( !has_use_internal_to_set(n, peel, loop) ) {
3200 
3201           // if not pinned and not a load (which maybe anti-dependent on a store)
3202           // and not a CMove (Matcher expects only bool->cmove).
3203           if (n->in(0) == NULL && !n->is_Load() && !n->is_CMove() && n->Opcode() != Op_ShenandoahWBMemProj) {
3204             cloned_for_outside_use += clone_for_use_outside_loop( loop, n, worklist );
3205             sink_list.push(n);
3206             peel     >>= n->_idx; // delete n from peel set.
3207             not_peel <<= n->_idx; // add n to not_peel set.
3208             peel_list.remove(i);
3209             incr = false;
3210 #if !defined(PRODUCT)
3211             if (TracePartialPeeling) {
3212               tty->print_cr("sink to not_peeled region: %d newbb: %d",
3213                             n->_idx, get_ctrl(n)->_idx);
3214             }
3215 #endif
3216           }
3217         } else {
3218           // Otherwise check for special def-use cases that span
3219           // the peel/not_peel boundary such as bool->if
3220           clone_for_special_use_inside_loop( loop, n, not_peel, sink_list, worklist );
3221           new_phi_cnt++;
3222         }
3223       }
3224     }
3225     if (incr) i++;
3226   }
3227 
3228   if (new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta) {
3229 #if !defined(PRODUCT)
3230     if (TracePartialPeeling) {
3231       tty->print_cr("\nToo many new phis: %d  old %d new cmpi: %c",
3232                     new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
3233     }
3234 #endif
3235     if (new_peel_if != NULL) {
3236       remove_cmpi_loop_exit(new_peel_if, loop);
3237     }
3238     // Inhibit more partial peeling on this loop
3239     assert(!head->is_partial_peel_loop(), "not partial peeled");
3240     head->mark_partial_peel_failed();
3241     if (cloned_for_outside_use > 0) {
3242       // Terminate this round of loop opts because
3243       // the graph outside this loop was changed.
3244       C->set_major_progress();
3245       return true;
3246     }
3247     return false;
3248   }
3249 
3250   // Step 3: clone loop, retarget control, and insert new phis
3251 
3252   // Create new loop head for new phis and to hang
3253   // the nodes being moved (sinked) from the peel region.
3254   LoopNode* new_head = new LoopNode(last_peel, last_peel);
3255   new_head->set_unswitch_count(head->unswitch_count()); // Preserve
3256   _igvn.register_new_node_with_optimizer(new_head);
3257   assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
3258   _igvn.replace_input_of(first_not_peeled, 0, new_head);
3259   set_loop(new_head, loop);
3260   loop->_body.push(new_head);
3261   not_peel.set(new_head->_idx);
3262   set_idom(new_head, last_peel, dom_depth(first_not_peeled));
3263   set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
3264 
3265   while (sink_list.size() > 0) {
3266     Node* n = sink_list.pop();
3267     set_ctrl(n, new_head);
3268   }
3269 
3270   assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3271 
3272   clone_loop(loop, old_new, dd, IgnoreStripMined);
3273 
3274   const uint clone_exit_idx = 1;
3275   const uint orig_exit_idx  = 2;
3276   assert(is_valid_clone_loop_form( loop, peel_list, orig_exit_idx, clone_exit_idx ), "bad clone loop");
3277 
3278   Node* head_clone             = old_new[head->_idx];
3279   LoopNode* new_head_clone     = old_new[new_head->_idx]->as_Loop();
3280   Node* orig_tail_clone        = head_clone->in(2);
3281 
3282   // Add phi if "def" node is in peel set and "use" is not
3283 
3284   for(i = 0; i < peel_list.size(); i++ ) {
3285     Node *def  = peel_list.at(i);
3286     if (!def->is_CFG()) {
3287       for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3288         Node *use = def->fast_out(j);
3289         if (has_node(use) && use->in(0) != C->top() &&
3290             (!peel.test(use->_idx) ||
3291              (use->is_Phi() && use->in(0) == head)) ) {
3292           worklist.push(use);
3293         }
3294       }
3295       while( worklist.size() ) {
3296         Node *use = worklist.pop();
3297         for (uint j = 1; j < use->req(); j++) {
3298           Node* n = use->in(j);
3299           if (n == def) {
3300 
3301             // "def" is in peel set, "use" is not in peel set
3302             // or "use" is in the entry boundary (a phi) of the peel set
3303 
3304             Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
3305 
3306             if ( loop->is_member(get_loop( use_c )) ) {
3307               // use is in loop
3308               if (old_new[use->_idx] != NULL) { // null for dead code
3309                 Node* use_clone = old_new[use->_idx];
3310                 _igvn.replace_input_of(use, j, C->top());
3311                 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
3312               }
3313             } else {
3314               assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
3315               // use is not in the loop, check if the live range includes the cut
3316               Node* lp_if = use_c->in(orig_exit_idx)->in(0);
3317               if (not_peel.test(lp_if->_idx)) {
3318                 assert(j == orig_exit_idx, "use from original loop");
3319                 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
3320               }
3321             }
3322           }
3323         }
3324       }
3325     }
3326   }
3327 
3328   // Step 3b: retarget control
3329 
3330   // Redirect control to the new loop head if a cloned node in
3331   // the not_peeled region has control that points into the peeled region.
3332   // This necessary because the cloned peeled region will be outside
3333   // the loop.
3334   //                            from    to
3335   //          cloned-peeled    <---+
3336   //    new_head_clone:            |    <--+
3337   //          cloned-not_peeled  in(0)    in(0)
3338   //          orig-peeled
3339 
3340   for(i = 0; i < loop->_body.size(); i++ ) {
3341     Node *n = loop->_body.at(i);
3342     if (!n->is_CFG()           && n->in(0) != NULL        &&
3343         not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
3344       Node* n_clone = old_new[n->_idx];
3345       _igvn.replace_input_of(n_clone, 0, new_head_clone);
3346     }
3347   }
3348 
3349   // Backedge of the surviving new_head (the clone) is original last_peel
3350   _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
3351 
3352   // Cut first node in original not_peel set
3353   _igvn.rehash_node_delayed(new_head);                     // Multiple edge updates:
3354   new_head->set_req(LoopNode::EntryControl,    C->top());  //   use rehash_node_delayed / set_req instead of
3355   new_head->set_req(LoopNode::LoopBackControl, C->top());  //   multiple replace_input_of calls
3356 
3357   // Copy head_clone back-branch info to original head
3358   // and remove original head's loop entry and
3359   // clone head's back-branch
3360   _igvn.rehash_node_delayed(head); // Multiple edge updates
3361   head->set_req(LoopNode::EntryControl,    head_clone->in(LoopNode::LoopBackControl));
3362   head->set_req(LoopNode::LoopBackControl, C->top());
3363   _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
3364 
3365   // Similarly modify the phis
3366   for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
3367     Node* use = head->fast_out(k);
3368     if (use->is_Phi() && use->outcnt() > 0) {
3369       Node* use_clone = old_new[use->_idx];
3370       _igvn.rehash_node_delayed(use); // Multiple edge updates
3371       use->set_req(LoopNode::EntryControl,    use_clone->in(LoopNode::LoopBackControl));
3372       use->set_req(LoopNode::LoopBackControl, C->top());
3373       _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
3374     }
3375   }
3376 
3377   // Step 4: update dominator tree and dominator depth
3378 
3379   set_idom(head, orig_tail_clone, dd);
3380   recompute_dom_depth();
3381 
3382   // Inhibit more partial peeling on this loop
3383   new_head_clone->set_partial_peel_loop();
3384   C->set_major_progress();
3385   loop->record_for_igvn();
3386 
3387 #if !defined(PRODUCT)
3388   if (TracePartialPeeling) {
3389     tty->print_cr("\nafter partial peel one iteration");
3390     Node_List wl(area);
3391     Node* t = last_peel;
3392     while (true) {
3393       wl.push(t);
3394       if (t == head_clone) break;
3395       t = idom(t);
3396     }
3397     while (wl.size() > 0) {
3398       Node* tt = wl.pop();
3399       if (tt == head) tty->print_cr("orig head");
3400       else if (tt == new_head_clone) tty->print_cr("new head");
3401       else if (tt == head_clone) tty->print_cr("clone head");
3402       tt->dump();
3403     }
3404   }
3405 #endif
3406   return true;
3407 }
3408 
3409 //------------------------------reorg_offsets----------------------------------
3410 // Reorganize offset computations to lower register pressure.  Mostly
3411 // prevent loop-fallout uses of the pre-incremented trip counter (which are
3412 // then alive with the post-incremented trip counter forcing an extra
3413 // register move)
3414 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
3415   // Perform it only for canonical counted loops.
3416   // Loop's shape could be messed up by iteration_split_impl.
3417   if (!loop->_head->is_CountedLoop())
3418     return;
3419   if (!loop->_head->as_Loop()->is_valid_counted_loop())
3420     return;
3421 
3422   CountedLoopNode *cl = loop->_head->as_CountedLoop();
3423   CountedLoopEndNode *cle = cl->loopexit();
3424   Node *exit = cle->proj_out(false);
3425   Node *phi = cl->phi();
3426 
3427   // Check for the special case of folks using the pre-incremented
3428   // trip-counter on the fall-out path (forces the pre-incremented
3429   // and post-incremented trip counter to be live at the same time).
3430   // Fix this by adjusting to use the post-increment trip counter.
3431 
3432   bool progress = true;
3433   while (progress) {
3434     progress = false;
3435     for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
3436       Node* use = phi->fast_out(i);   // User of trip-counter
3437       if (!has_ctrl(use))  continue;
3438       Node *u_ctrl = get_ctrl(use);
3439       if (use->is_Phi()) {
3440         u_ctrl = NULL;
3441         for (uint j = 1; j < use->req(); j++)
3442           if (use->in(j) == phi)
3443             u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
3444       }
3445       IdealLoopTree *u_loop = get_loop(u_ctrl);
3446       // Look for loop-invariant use
3447       if (u_loop == loop) continue;
3448       if (loop->is_member(u_loop)) continue;
3449       // Check that use is live out the bottom.  Assuming the trip-counter
3450       // update is right at the bottom, uses of of the loop middle are ok.
3451       if (dom_lca(exit, u_ctrl) != exit) continue;
3452       // Hit!  Refactor use to use the post-incremented tripcounter.
3453       // Compute a post-increment tripcounter.
3454       Node* c = exit;
3455       if (cl->is_strip_mined()) {
3456         IdealLoopTree* outer_loop = get_loop(cl->outer_loop());
3457         if (!outer_loop->is_member(u_loop)) {
3458           c = cl->outer_loop_exit();
3459         }
3460       }
3461       Node *opaq = new Opaque2Node(C, cle->incr());
3462       register_new_node(opaq, c);
3463       Node *neg_stride = _igvn.intcon(-cle->stride_con());
3464       set_ctrl(neg_stride, C->root());
3465       Node *post = new AddINode(opaq, neg_stride);
3466       register_new_node(post, c);
3467       _igvn.rehash_node_delayed(use);
3468       for (uint j = 1; j < use->req(); j++) {
3469         if (use->in(j) == phi)
3470           use->set_req(j, post);
3471       }
3472       // Since DU info changed, rerun loop
3473       progress = true;
3474       break;
3475     }
3476   }
3477 
3478 }