1 /*
   2  * Copyright (c) 1999, 2019, 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 "opto/subtypenode.hpp"
  44 #include "utilities/macros.hpp"
  45 
  46 //=============================================================================
  47 //------------------------------split_thru_phi---------------------------------
  48 // Split Node 'n' through merge point if there is enough win.
  49 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) {
  50   if (n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) {
  51     // ConvI2L may have type information on it which is unsafe to push up
  52     // so disable this for now
  53     return NULL;
  54   }
  55 
  56   // Splitting range check CastIIs through a loop induction Phi can
  57   // cause new Phis to be created that are left unrelated to the loop
  58   // induction Phi and prevent optimizations (vectorization)
  59   if (n->Opcode() == Op_CastII && n->as_CastII()->has_range_check() &&
  60       region->is_CountedLoop() && n->in(1) == region->as_CountedLoop()->phi()) {
  61     return NULL;
  62   }
  63 
  64   // Bail out if 'n' is a Div or Mod node whose zero check was removed earlier (i.e. control is NULL) and its divisor is a phi node d_phi whose
  65   // inputs could be zero (include zero in their type range). This can only happen if d_phi is an induction variable of a trip-counted (integer)
  66   // loop. d_phi could have a more precise type range that does not necessarily include all values of its inputs. Since each of these inputs will
  67   // be a divisor of the newly cloned nodes of 'n', we need to bail out of one of these divisors could be zero (zero in its type range).
  68   if ((n->Opcode() == Op_DivI || n->Opcode() == Op_ModI) && n->in(0) == NULL
  69       && region->is_CountedLoop() && n->in(2) == region->as_CountedLoop()->phi()) {
  70     Node* phi = region->as_CountedLoop()->phi();
  71     for (uint i = 1; i < phi->req(); i++) {
  72       if (_igvn.type(phi->in(i))->filter_speculative(TypeInt::ZERO) != Type::TOP) {
  73         // Zero could be a possible value but we already removed the zero check. Bail out to avoid a possible division by zero at a later point.
  74         return NULL;
  75       }
  76     }
  77   }
  78 
  79 #ifdef ASSERT
  80   // A phi only has a more precise type than its inputs if it is an induction variable of a trip-counted integer loop. Verify that this is the
  81   // case and we are not about to split a Mod or Div node without zero check that could end up with a new divisor that can be zero.
  82   // A split of a division by a floating point zero, on the other hand, is not a problem since the result would just be INFINITY as floating point
  83   // representation and does not result in an exception.
  84   if ((n->Opcode() == Op_DivI || n->Opcode() == Op_ModI)) {
  85     check_divide_by_zero(n, TypeInt::ZERO);
  86   } else if ((n->Opcode() == Op_DivL || n->Opcode() == Op_ModL)) {
  87     check_divide_by_zero(n, TypeLong::ZERO);
  88   }
  89 #endif
  90 
  91   int wins = 0;
  92   assert(!n->is_CFG(), "");
  93   assert(region->is_Region(), "");
  94 
  95   const Type* type = n->bottom_type();
  96   const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr();
  97   Node* phi;
  98   if (t_oop != NULL && t_oop->is_known_instance_field()) {
  99     int iid    = t_oop->instance_id();
 100     int index  = C->get_alias_index(t_oop);
 101     int offset = t_oop->offset();
 102     phi = new PhiNode(region, type, NULL, iid, index, offset);
 103   } else {
 104     phi = PhiNode::make_blank(region, n);
 105   }
 106   uint old_unique = C->unique();
 107   for (uint i = 1; i < region->req(); i++) {
 108     Node* x;
 109     Node* the_clone = NULL;
 110     if (region->in(i) == C->top()) {
 111       x = C->top();             // Dead path?  Use a dead data op
 112     } else {
 113       x = n->clone();           // Else clone up the data op
 114       the_clone = x;            // Remember for possible deletion.
 115       // Alter data node to use pre-phi inputs
 116       if (n->in(0) == region)
 117         x->set_req( 0, region->in(i) );
 118       for (uint j = 1; j < n->req(); j++) {
 119         Node* in = n->in(j);
 120         if (in->is_Phi() && in->in(0) == region)
 121           x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
 122       }
 123     }
 124     // Check for a 'win' on some paths
 125     const Type* t = x->Value(&_igvn);
 126 
 127     bool singleton = t->singleton();
 128 
 129     // A TOP singleton indicates that there are no possible values incoming
 130     // along a particular edge. In most cases, this is OK, and the Phi will
 131     // be eliminated later in an Ideal call. However, we can't allow this to
 132     // happen if the singleton occurs on loop entry, as the elimination of
 133     // the PhiNode may cause the resulting node to migrate back to a previous
 134     // loop iteration.
 135     if (singleton && t == Type::TOP) {
 136       // Is_Loop() == false does not confirm the absence of a loop (e.g., an
 137       // irreducible loop may not be indicated by an affirmative is_Loop());
 138       // therefore, the only top we can split thru a phi is on a backedge of
 139       // a loop.
 140       singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
 141     }
 142 
 143     if (singleton) {
 144       wins++;
 145       x = ((PhaseGVN&)_igvn).makecon(t);
 146     } else {
 147       // We now call Identity to try to simplify the cloned node.
 148       // Note that some Identity methods call phase->type(this).
 149       // Make sure that the type array is big enough for
 150       // our new node, even though we may throw the node away.
 151       // (Note: This tweaking with igvn only works because x is a new node.)
 152       _igvn.set_type(x, t);
 153       // If x is a TypeNode, capture any more-precise type permanently into Node
 154       // otherwise it will be not updated during igvn->transform since
 155       // igvn->type(x) is set to x->Value() already.
 156       x->raise_bottom_type(t);
 157       Node* y = x->Identity(&_igvn);
 158       if (y != x) {
 159         wins++;
 160         x = y;
 161       } else {
 162         y = _igvn.hash_find(x);
 163         if (y) {
 164           wins++;
 165           x = y;
 166         } else {
 167           // Else x is a new node we are keeping
 168           // We do not need register_new_node_with_optimizer
 169           // because set_type has already been called.
 170           _igvn._worklist.push(x);
 171         }
 172       }
 173     }
 174     if (x != the_clone && the_clone != NULL)
 175       _igvn.remove_dead_node(the_clone);
 176     phi->set_req( i, x );
 177   }
 178   // Too few wins?
 179   if (wins <= policy) {
 180     _igvn.remove_dead_node(phi);
 181     return NULL;
 182   }
 183 
 184   // Record Phi
 185   register_new_node( phi, region );
 186 
 187   for (uint i2 = 1; i2 < phi->req(); i2++) {
 188     Node *x = phi->in(i2);
 189     // If we commoned up the cloned 'x' with another existing Node,
 190     // the existing Node picks up a new use.  We need to make the
 191     // existing Node occur higher up so it dominates its uses.
 192     Node *old_ctrl;
 193     IdealLoopTree *old_loop;
 194 
 195     if (x->is_Con()) {
 196       // Constant's control is always root.
 197       set_ctrl(x, C->root());
 198       continue;
 199     }
 200     // The occasional new node
 201     if (x->_idx >= old_unique) {     // Found a new, unplaced node?
 202       old_ctrl = NULL;
 203       old_loop = NULL;               // Not in any prior loop
 204     } else {
 205       old_ctrl = get_ctrl(x);
 206       old_loop = get_loop(old_ctrl); // Get prior loop
 207     }
 208     // New late point must dominate new use
 209     Node *new_ctrl = dom_lca(old_ctrl, region->in(i2));
 210     if (new_ctrl == old_ctrl) // Nothing is changed
 211       continue;
 212 
 213     IdealLoopTree *new_loop = get_loop(new_ctrl);
 214 
 215     // Don't move x into a loop if its uses are
 216     // outside of loop. Otherwise x will be cloned
 217     // for each use outside of this loop.
 218     IdealLoopTree *use_loop = get_loop(region);
 219     if (!new_loop->is_member(use_loop) &&
 220         (old_loop == NULL || !new_loop->is_member(old_loop))) {
 221       // Take early control, later control will be recalculated
 222       // during next iteration of loop optimizations.
 223       new_ctrl = get_early_ctrl(x);
 224       new_loop = get_loop(new_ctrl);
 225     }
 226     // Set new location
 227     set_ctrl(x, new_ctrl);
 228     // If changing loop bodies, see if we need to collect into new body
 229     if (old_loop != new_loop) {
 230       if (old_loop && !old_loop->_child)
 231         old_loop->_body.yank(x);
 232       if (!new_loop->_child)
 233         new_loop->_body.push(x);  // Collect body info
 234     }
 235   }
 236 
 237   return phi;
 238 }
 239 
 240 #ifdef ASSERT
 241 void PhaseIdealLoop::check_divide_by_zero(const Node* n, const Type* zero) const {
 242   if (n->in(0) == NULL && n->in(2)->is_Phi()) {
 243     Node* phi = n->in(2);
 244     for (uint i = 1; i < phi->req(); i++) {
 245       if (_igvn.type(phi->in(i))->filter_speculative(zero) != Type::TOP) {
 246         // Zero could be a possible value
 247         n->dump(2);
 248         phi->dump(2);
 249         assert(false, "should not happen that pre-phi type includes zero as possible value but zero check was already removed");
 250       }
 251     }
 252   }
 253 }
 254 #endif
 255 
 256 //------------------------------dominated_by------------------------------------
 257 // Replace the dominated test with an obvious true or false.  Place it on the
 258 // IGVN worklist for later cleanup.  Move control-dependent data Nodes on the
 259 // live path up to the dominating control.
 260 void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exclude_loop_predicate ) {
 261   if (VerifyLoopOptimizations && PrintOpto) { tty->print_cr("dominating test"); }
 262 
 263   // prevdom is the dominating projection of the dominating test.
 264   assert( iff->is_If(), "" );
 265   assert(iff->Opcode() == Op_If || iff->Opcode() == Op_CountedLoopEnd || iff->Opcode() == Op_RangeCheck, "Check this code when new subtype is added");
 266   int pop = prevdom->Opcode();
 267   assert( pop == Op_IfFalse || pop == Op_IfTrue, "" );
 268   if (flip) {
 269     if (pop == Op_IfTrue)
 270       pop = Op_IfFalse;
 271     else
 272       pop = Op_IfTrue;
 273   }
 274   // 'con' is set to true or false to kill the dominated test.
 275   Node *con = _igvn.makecon(pop == Op_IfTrue ? TypeInt::ONE : TypeInt::ZERO);
 276   set_ctrl(con, C->root()); // Constant gets a new use
 277   // Hack the dominated test
 278   _igvn.replace_input_of(iff, 1, con);
 279 
 280   // If I dont have a reachable TRUE and FALSE path following the IfNode then
 281   // I can assume this path reaches an infinite loop.  In this case it's not
 282   // important to optimize the data Nodes - either the whole compilation will
 283   // be tossed or this path (and all data Nodes) will go dead.
 284   if (iff->outcnt() != 2) return;
 285 
 286   // Make control-dependent data Nodes on the live path (path that will remain
 287   // once the dominated IF is removed) become control-dependent on the
 288   // dominating projection.
 289   Node* dp = iff->as_If()->proj_out_or_null(pop == Op_IfTrue);
 290 
 291   // Loop predicates may have depending checks which should not
 292   // be skipped. For example, range check predicate has two checks
 293   // for lower and upper bounds.
 294   if (dp == NULL)
 295     return;
 296 
 297   ProjNode* dp_proj  = dp->as_Proj();
 298   ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj();
 299   if (exclude_loop_predicate &&
 300       (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate) != NULL ||
 301        unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_profile_predicate) != NULL ||
 302        unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check) != NULL)) {
 303     // If this is a range check (IfNode::is_range_check), do not
 304     // reorder because Compile::allow_range_check_smearing might have
 305     // changed the check.
 306     return; // Let IGVN transformation change control dependence.
 307   }
 308 
 309   IdealLoopTree *old_loop = get_loop(dp);
 310 
 311   for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
 312     Node* cd = dp->fast_out(i); // Control-dependent node
 313     if (cd->depends_only_on_test()) {
 314       assert(cd->in(0) == dp, "");
 315       _igvn.replace_input_of(cd, 0, prevdom);
 316       set_early_ctrl(cd);
 317       IdealLoopTree *new_loop = get_loop(get_ctrl(cd));
 318       if (old_loop != new_loop) {
 319         if (!old_loop->_child) old_loop->_body.yank(cd);
 320         if (!new_loop->_child) new_loop->_body.push(cd);
 321       }
 322       --i;
 323       --imax;
 324     }
 325   }
 326 }
 327 
 328 //------------------------------has_local_phi_input----------------------------
 329 // Return TRUE if 'n' has Phi inputs from its local block and no other
 330 // block-local inputs (all non-local-phi inputs come from earlier blocks)
 331 Node *PhaseIdealLoop::has_local_phi_input( Node *n ) {
 332   Node *n_ctrl = get_ctrl(n);
 333   // See if some inputs come from a Phi in this block, or from before
 334   // this block.
 335   uint i;
 336   for( i = 1; i < n->req(); i++ ) {
 337     Node *phi = n->in(i);
 338     if( phi->is_Phi() && phi->in(0) == n_ctrl )
 339       break;
 340   }
 341   if( i >= n->req() )
 342     return NULL;                // No Phi inputs; nowhere to clone thru
 343 
 344   // Check for inputs created between 'n' and the Phi input.  These
 345   // must split as well; they have already been given the chance
 346   // (courtesy of a post-order visit) and since they did not we must
 347   // recover the 'cost' of splitting them by being very profitable
 348   // when splitting 'n'.  Since this is unlikely we simply give up.
 349   for( i = 1; i < n->req(); i++ ) {
 350     Node *m = n->in(i);
 351     if( get_ctrl(m) == n_ctrl && !m->is_Phi() ) {
 352       // We allow the special case of AddP's with no local inputs.
 353       // This allows us to split-up address expressions.
 354       if (m->is_AddP() &&
 355           get_ctrl(m->in(2)) != n_ctrl &&
 356           get_ctrl(m->in(3)) != n_ctrl) {
 357         // Move the AddP up to dominating point
 358         Node* c = find_non_split_ctrl(idom(n_ctrl));
 359         if (c->is_OuterStripMinedLoop()) {
 360           c->as_Loop()->verify_strip_mined(1);
 361           c = c->in(LoopNode::EntryControl);
 362         }
 363         set_ctrl_and_loop(m, c);
 364         continue;
 365       }
 366       return NULL;
 367     }
 368     assert(n->is_Phi() || m->is_Phi() || is_dominator(get_ctrl(m), n_ctrl), "m has strange control");
 369   }
 370 
 371   return n_ctrl;
 372 }
 373 
 374 //------------------------------remix_address_expressions----------------------
 375 // Rework addressing expressions to get the most loop-invariant stuff
 376 // moved out.  We'd like to do all associative operators, but it's especially
 377 // important (common) to do address expressions.
 378 Node *PhaseIdealLoop::remix_address_expressions( Node *n ) {
 379   if (!has_ctrl(n))  return NULL;
 380   Node *n_ctrl = get_ctrl(n);
 381   IdealLoopTree *n_loop = get_loop(n_ctrl);
 382 
 383   // See if 'n' mixes loop-varying and loop-invariant inputs and
 384   // itself is loop-varying.
 385 
 386   // Only interested in binary ops (and AddP)
 387   if( n->req() < 3 || n->req() > 4 ) return NULL;
 388 
 389   Node *n1_ctrl = get_ctrl(n->in(                    1));
 390   Node *n2_ctrl = get_ctrl(n->in(                    2));
 391   Node *n3_ctrl = get_ctrl(n->in(n->req() == 3 ? 2 : 3));
 392   IdealLoopTree *n1_loop = get_loop( n1_ctrl );
 393   IdealLoopTree *n2_loop = get_loop( n2_ctrl );
 394   IdealLoopTree *n3_loop = get_loop( n3_ctrl );
 395 
 396   // Does one of my inputs spin in a tighter loop than self?
 397   if( (n_loop->is_member( n1_loop ) && n_loop != n1_loop) ||
 398       (n_loop->is_member( n2_loop ) && n_loop != n2_loop) ||
 399       (n_loop->is_member( n3_loop ) && n_loop != n3_loop) )
 400     return NULL;                // Leave well enough alone
 401 
 402   // Is at least one of my inputs loop-invariant?
 403   if( n1_loop == n_loop &&
 404       n2_loop == n_loop &&
 405       n3_loop == n_loop )
 406     return NULL;                // No loop-invariant inputs
 407 
 408 
 409   int n_op = n->Opcode();
 410 
 411   // Replace expressions like ((V+I) << 2) with (V<<2 + I<<2).
 412   if( n_op == Op_LShiftI ) {
 413     // Scale is loop invariant
 414     Node *scale = n->in(2);
 415     Node *scale_ctrl = get_ctrl(scale);
 416     IdealLoopTree *scale_loop = get_loop(scale_ctrl );
 417     if( n_loop == scale_loop || !scale_loop->is_member( n_loop ) )
 418       return NULL;
 419     const TypeInt *scale_t = scale->bottom_type()->isa_int();
 420     if( scale_t && scale_t->is_con() && scale_t->get_con() >= 16 )
 421       return NULL;              // Dont bother with byte/short masking
 422     // Add must vary with loop (else shift would be loop-invariant)
 423     Node *add = n->in(1);
 424     Node *add_ctrl = get_ctrl(add);
 425     IdealLoopTree *add_loop = get_loop(add_ctrl);
 426     //assert( n_loop == add_loop, "" );
 427     if( n_loop != add_loop ) return NULL;  // happens w/ evil ZKM loops
 428 
 429     // Convert I-V into I+ (0-V); same for V-I
 430     if( add->Opcode() == Op_SubI &&
 431         _igvn.type( add->in(1) ) != TypeInt::ZERO ) {
 432       Node *zero = _igvn.intcon(0);
 433       set_ctrl(zero, C->root());
 434       Node *neg = new SubINode( _igvn.intcon(0), add->in(2) );
 435       register_new_node( neg, get_ctrl(add->in(2) ) );
 436       add = new AddINode( add->in(1), neg );
 437       register_new_node( add, add_ctrl );
 438     }
 439     if( add->Opcode() != Op_AddI ) return NULL;
 440     // See if one add input is loop invariant
 441     Node *add_var = add->in(1);
 442     Node *add_var_ctrl = get_ctrl(add_var);
 443     IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
 444     Node *add_invar = add->in(2);
 445     Node *add_invar_ctrl = get_ctrl(add_invar);
 446     IdealLoopTree *add_invar_loop = get_loop(add_invar_ctrl );
 447     if( add_var_loop == n_loop ) {
 448     } else if( add_invar_loop == n_loop ) {
 449       // Swap to find the invariant part
 450       add_invar = add_var;
 451       add_invar_ctrl = add_var_ctrl;
 452       add_invar_loop = add_var_loop;
 453       add_var = add->in(2);
 454       Node *add_var_ctrl = get_ctrl(add_var);
 455       IdealLoopTree *add_var_loop = get_loop(add_var_ctrl );
 456     } else                      // Else neither input is loop invariant
 457       return NULL;
 458     if( n_loop == add_invar_loop || !add_invar_loop->is_member( n_loop ) )
 459       return NULL;              // No invariant part of the add?
 460 
 461     // Yes!  Reshape address expression!
 462     Node *inv_scale = new LShiftINode( add_invar, scale );
 463     Node *inv_scale_ctrl =
 464       dom_depth(add_invar_ctrl) > dom_depth(scale_ctrl) ?
 465       add_invar_ctrl : scale_ctrl;
 466     register_new_node( inv_scale, inv_scale_ctrl );
 467     Node *var_scale = new LShiftINode( add_var, scale );
 468     register_new_node( var_scale, n_ctrl );
 469     Node *var_add = new AddINode( var_scale, inv_scale );
 470     register_new_node( var_add, n_ctrl );
 471     _igvn.replace_node( n, var_add );
 472     return var_add;
 473   }
 474 
 475   // Replace (I+V) with (V+I)
 476   if( n_op == Op_AddI ||
 477       n_op == Op_AddL ||
 478       n_op == Op_AddF ||
 479       n_op == Op_AddD ||
 480       n_op == Op_MulI ||
 481       n_op == Op_MulL ||
 482       n_op == Op_MulF ||
 483       n_op == Op_MulD ) {
 484     if( n2_loop == n_loop ) {
 485       assert( n1_loop != n_loop, "" );
 486       n->swap_edges(1, 2);
 487     }
 488   }
 489 
 490   // Replace ((I1 +p V) +p I2) with ((I1 +p I2) +p V),
 491   // but not if I2 is a constant.
 492   if( n_op == Op_AddP ) {
 493     if( n2_loop == n_loop && n3_loop != n_loop ) {
 494       if( n->in(2)->Opcode() == Op_AddP && !n->in(3)->is_Con() ) {
 495         Node *n22_ctrl = get_ctrl(n->in(2)->in(2));
 496         Node *n23_ctrl = get_ctrl(n->in(2)->in(3));
 497         IdealLoopTree *n22loop = get_loop( n22_ctrl );
 498         IdealLoopTree *n23_loop = get_loop( n23_ctrl );
 499         if( n22loop != n_loop && n22loop->is_member(n_loop) &&
 500             n23_loop == n_loop ) {
 501           Node *add1 = new AddPNode( n->in(1), n->in(2)->in(2), n->in(3) );
 502           // Stuff new AddP in the loop preheader
 503           register_new_node( add1, n_loop->_head->in(LoopNode::EntryControl) );
 504           Node *add2 = new AddPNode( n->in(1), add1, n->in(2)->in(3) );
 505           register_new_node( add2, n_ctrl );
 506           _igvn.replace_node( n, add2 );
 507           return add2;
 508         }
 509       }
 510     }
 511 
 512     // Replace (I1 +p (I2 + V)) with ((I1 +p I2) +p V)
 513     if (n2_loop != n_loop && n3_loop == n_loop) {
 514       if (n->in(3)->Opcode() == Op_AddX) {
 515         Node *V = n->in(3)->in(1);
 516         Node *I = n->in(3)->in(2);
 517         if (is_member(n_loop,get_ctrl(V))) {
 518         } else {
 519           Node *tmp = V; V = I; I = tmp;
 520         }
 521         if (!is_member(n_loop,get_ctrl(I))) {
 522           Node *add1 = new AddPNode(n->in(1), n->in(2), I);
 523           // Stuff new AddP in the loop preheader
 524           register_new_node(add1, n_loop->_head->in(LoopNode::EntryControl));
 525           Node *add2 = new AddPNode(n->in(1), add1, V);
 526           register_new_node(add2, n_ctrl);
 527           _igvn.replace_node(n, add2);
 528           return add2;
 529         }
 530       }
 531     }
 532   }
 533 
 534   return NULL;
 535 }
 536 
 537 // Optimize ((in1[2*i] * in2[2*i]) + (in1[2*i+1] * in2[2*i+1]))
 538 Node *PhaseIdealLoop::convert_add_to_muladd(Node* n) {
 539   assert(n->Opcode() == Op_AddI, "sanity");
 540   Node * nn = NULL;
 541   Node * in1 = n->in(1);
 542   Node * in2 = n->in(2);
 543   if (in1->Opcode() == Op_MulI && in2->Opcode() == Op_MulI) {
 544     IdealLoopTree* loop_n = get_loop(get_ctrl(n));
 545     if (loop_n->is_counted() &&
 546         loop_n->_head->as_Loop()->is_valid_counted_loop() &&
 547         Matcher::match_rule_supported(Op_MulAddVS2VI) &&
 548         Matcher::match_rule_supported(Op_MulAddS2I)) {
 549       Node* mul_in1 = in1->in(1);
 550       Node* mul_in2 = in1->in(2);
 551       Node* mul_in3 = in2->in(1);
 552       Node* mul_in4 = in2->in(2);
 553       if (mul_in1->Opcode() == Op_LoadS &&
 554           mul_in2->Opcode() == Op_LoadS &&
 555           mul_in3->Opcode() == Op_LoadS &&
 556           mul_in4->Opcode() == Op_LoadS) {
 557         IdealLoopTree* loop1 = get_loop(get_ctrl(mul_in1));
 558         IdealLoopTree* loop2 = get_loop(get_ctrl(mul_in2));
 559         IdealLoopTree* loop3 = get_loop(get_ctrl(mul_in3));
 560         IdealLoopTree* loop4 = get_loop(get_ctrl(mul_in4));
 561         IdealLoopTree* loop5 = get_loop(get_ctrl(in1));
 562         IdealLoopTree* loop6 = get_loop(get_ctrl(in2));
 563         // All nodes should be in the same counted loop.
 564         if (loop_n == loop1 && loop_n == loop2 && loop_n == loop3 &&
 565             loop_n == loop4 && loop_n == loop5 && loop_n == loop6) {
 566           Node* adr1 = mul_in1->in(MemNode::Address);
 567           Node* adr2 = mul_in2->in(MemNode::Address);
 568           Node* adr3 = mul_in3->in(MemNode::Address);
 569           Node* adr4 = mul_in4->in(MemNode::Address);
 570           if (adr1->is_AddP() && adr2->is_AddP() && adr3->is_AddP() && adr4->is_AddP()) {
 571             if ((adr1->in(AddPNode::Base) == adr3->in(AddPNode::Base)) &&
 572                 (adr2->in(AddPNode::Base) == adr4->in(AddPNode::Base))) {
 573               nn = new MulAddS2INode(mul_in1, mul_in2, mul_in3, mul_in4);
 574               register_new_node(nn, get_ctrl(n));
 575               _igvn.replace_node(n, nn);
 576               return nn;
 577             } else if ((adr1->in(AddPNode::Base) == adr4->in(AddPNode::Base)) &&
 578                        (adr2->in(AddPNode::Base) == adr3->in(AddPNode::Base))) {
 579               nn = new MulAddS2INode(mul_in1, mul_in2, mul_in4, mul_in3);
 580               register_new_node(nn, get_ctrl(n));
 581               _igvn.replace_node(n, nn);
 582               return nn;
 583             }
 584           }
 585         }
 586       }
 587     }
 588   }
 589   return nn;
 590 }
 591 
 592 //------------------------------conditional_move-------------------------------
 593 // Attempt to replace a Phi with a conditional move.  We have some pretty
 594 // strict profitability requirements.  All Phis at the merge point must
 595 // be converted, so we can remove the control flow.  We need to limit the
 596 // number of c-moves to a small handful.  All code that was in the side-arms
 597 // of the CFG diamond is now speculatively executed.  This code has to be
 598 // "cheap enough".  We are pretty much limited to CFG diamonds that merge
 599 // 1 or 2 items with a total of 1 or 2 ops executed speculatively.
 600 Node *PhaseIdealLoop::conditional_move( Node *region ) {
 601 
 602   assert(region->is_Region(), "sanity check");
 603   if (region->req() != 3) return NULL;
 604 
 605   // Check for CFG diamond
 606   Node *lp = region->in(1);
 607   Node *rp = region->in(2);
 608   if (!lp || !rp) return NULL;
 609   Node *lp_c = lp->in(0);
 610   if (lp_c == NULL || lp_c != rp->in(0) || !lp_c->is_If()) return NULL;
 611   IfNode *iff = lp_c->as_If();
 612 
 613   // Check for ops pinned in an arm of the diamond.
 614   // Can't remove the control flow in this case
 615   if (lp->outcnt() > 1) return NULL;
 616   if (rp->outcnt() > 1) return NULL;
 617 
 618   IdealLoopTree* r_loop = get_loop(region);
 619   assert(r_loop == get_loop(iff), "sanity");
 620   // Always convert to CMOVE if all results are used only outside this loop.
 621   bool used_inside_loop = (r_loop == _ltree_root);
 622 
 623   // Check profitability
 624   int cost = 0;
 625   int phis = 0;
 626   for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
 627     Node *out = region->fast_out(i);
 628     if (!out->is_Phi()) continue; // Ignore other control edges, etc
 629     phis++;
 630     PhiNode* phi = out->as_Phi();
 631     BasicType bt = phi->type()->basic_type();
 632     switch (bt) {
 633     case T_DOUBLE:
 634     case T_FLOAT:
 635       if (C->use_cmove()) {
 636         continue; //TODO: maybe we want to add some cost
 637       }
 638       cost += Matcher::float_cmove_cost(); // Could be very expensive
 639       break;
 640     case T_LONG: {
 641       cost += Matcher::long_cmove_cost(); // May encodes as 2 CMOV's
 642     }
 643     case T_INT:                 // These all CMOV fine
 644     case T_ADDRESS: {           // (RawPtr)
 645       cost++;
 646       break;
 647     }
 648     case T_NARROWOOP: // Fall through
 649     case T_OBJECT: {            // Base oops are OK, but not derived oops
 650       const TypeOopPtr *tp = phi->type()->make_ptr()->isa_oopptr();
 651       // Derived pointers are Bad (tm): what's the Base (for GC purposes) of a
 652       // CMOVE'd derived pointer?  It's a CMOVE'd derived base.  Thus
 653       // CMOVE'ing a derived pointer requires we also CMOVE the base.  If we
 654       // have a Phi for the base here that we convert to a CMOVE all is well
 655       // and good.  But if the base is dead, we'll not make a CMOVE.  Later
 656       // the allocator will have to produce a base by creating a CMOVE of the
 657       // relevant bases.  This puts the allocator in the business of
 658       // manufacturing expensive instructions, generally a bad plan.
 659       // Just Say No to Conditionally-Moved Derived Pointers.
 660       if (tp && tp->offset() != 0)
 661         return NULL;
 662       cost++;
 663       break;
 664     }
 665     default:
 666       return NULL;              // In particular, can't do memory or I/O
 667     }
 668     // Add in cost any speculative ops
 669     for (uint j = 1; j < region->req(); j++) {
 670       Node *proj = region->in(j);
 671       Node *inp = phi->in(j);
 672       if (get_ctrl(inp) == proj) { // Found local op
 673         cost++;
 674         // Check for a chain of dependent ops; these will all become
 675         // speculative in a CMOV.
 676         for (uint k = 1; k < inp->req(); k++)
 677           if (get_ctrl(inp->in(k)) == proj)
 678             cost += ConditionalMoveLimit; // Too much speculative goo
 679       }
 680     }
 681     // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
 682     // This will likely Split-If, a higher-payoff operation.
 683     for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
 684       Node* use = phi->fast_out(k);
 685       if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
 686         cost += ConditionalMoveLimit;
 687       // Is there a use inside the loop?
 688       // Note: check only basic types since CMoveP is pinned.
 689       if (!used_inside_loop && is_java_primitive(bt)) {
 690         IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
 691         if (r_loop == u_loop || r_loop->is_member(u_loop)) {
 692           used_inside_loop = true;
 693         }
 694       }
 695     }
 696   }//for
 697   Node* bol = iff->in(1);
 698   if (bol->Opcode() == Op_Opaque4) {
 699     return NULL; // Ignore loop predicate checks (the Opaque4 ensures they will go away)
 700   }
 701   assert(bol->Opcode() == Op_Bool, "Unexpected node");
 702   int cmp_op = bol->in(1)->Opcode();
 703   if (cmp_op == Op_SubTypeCheck) { // SubTypeCheck expansion expects an IfNode
 704     return NULL;
 705   }
 706   // It is expensive to generate flags from a float compare.
 707   // Avoid duplicated float compare.
 708   if (phis > 1 && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) return NULL;
 709 
 710   float infrequent_prob = PROB_UNLIKELY_MAG(3);
 711   // Ignore cost and blocks frequency if CMOVE can be moved outside the loop.
 712   if (used_inside_loop) {
 713     if (cost >= ConditionalMoveLimit) return NULL; // Too much goo
 714 
 715     // BlockLayoutByFrequency optimization moves infrequent branch
 716     // from hot path. No point in CMOV'ing in such case (110 is used
 717     // instead of 100 to take into account not exactness of float value).
 718     if (BlockLayoutByFrequency) {
 719       infrequent_prob = MAX2(infrequent_prob, (float)BlockLayoutMinDiamondPercentage/110.0f);
 720     }
 721   }
 722   // Check for highly predictable branch.  No point in CMOV'ing if
 723   // we are going to predict accurately all the time.
 724   if (C->use_cmove() && (cmp_op == Op_CmpF || cmp_op == Op_CmpD)) {
 725     //keep going
 726   } else if (iff->_prob < infrequent_prob ||
 727       iff->_prob > (1.0f - infrequent_prob))
 728     return NULL;
 729 
 730   // --------------
 731   // Now replace all Phis with CMOV's
 732   Node *cmov_ctrl = iff->in(0);
 733   uint flip = (lp->Opcode() == Op_IfTrue);
 734   Node_List wq;
 735   while (1) {
 736     PhiNode* phi = NULL;
 737     for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
 738       Node *out = region->fast_out(i);
 739       if (out->is_Phi()) {
 740         phi = out->as_Phi();
 741         break;
 742       }
 743     }
 744     if (phi == NULL)  break;
 745     if (PrintOpto && VerifyLoopOptimizations) { tty->print_cr("CMOV"); }
 746     // Move speculative ops
 747     wq.push(phi);
 748     while (wq.size() > 0) {
 749       Node *n = wq.pop();
 750       for (uint j = 1; j < n->req(); j++) {
 751         Node* m = n->in(j);
 752         if (m != NULL && !is_dominator(get_ctrl(m), cmov_ctrl)) {
 753 #ifndef PRODUCT
 754           if (PrintOpto && VerifyLoopOptimizations) {
 755             tty->print("  speculate: ");
 756             m->dump();
 757           }
 758 #endif
 759           set_ctrl(m, cmov_ctrl);
 760           wq.push(m);
 761         }
 762       }
 763     }
 764     Node *cmov = CMoveNode::make(cmov_ctrl, iff->in(1), phi->in(1+flip), phi->in(2-flip), _igvn.type(phi));
 765     register_new_node( cmov, cmov_ctrl );
 766     _igvn.replace_node( phi, cmov );
 767 #ifndef PRODUCT
 768     if (TraceLoopOpts) {
 769       tty->print("CMOV  ");
 770       r_loop->dump_head();
 771       if (Verbose) {
 772         bol->in(1)->dump(1);
 773         cmov->dump(1);
 774       }
 775     }
 776     if (VerifyLoopOptimizations) verify();
 777 #endif
 778   }
 779 
 780   // The useless CFG diamond will fold up later; see the optimization in
 781   // RegionNode::Ideal.
 782   _igvn._worklist.push(region);
 783 
 784   return iff->in(1);
 785 }
 786 
 787 static void enqueue_cfg_uses(Node* m, Unique_Node_List& wq) {
 788   for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
 789     Node* u = m->fast_out(i);
 790     if (u->is_CFG()) {
 791       if (u->Opcode() == Op_NeverBranch) {
 792         u = ((NeverBranchNode*)u)->proj_out(0);
 793         enqueue_cfg_uses(u, wq);
 794       } else {
 795         wq.push(u);
 796       }
 797     }
 798   }
 799 }
 800 
 801 // Try moving a store out of a loop, right before the loop
 802 Node* PhaseIdealLoop::try_move_store_before_loop(Node* n, Node *n_ctrl) {
 803   // Store has to be first in the loop body
 804   IdealLoopTree *n_loop = get_loop(n_ctrl);
 805   if (n->is_Store() && n_loop != _ltree_root &&
 806       n_loop->is_loop() && n_loop->_head->is_Loop() &&
 807       n->in(0) != NULL) {
 808     Node* address = n->in(MemNode::Address);
 809     Node* value = n->in(MemNode::ValueIn);
 810     Node* mem = n->in(MemNode::Memory);
 811     IdealLoopTree* address_loop = get_loop(get_ctrl(address));
 812     IdealLoopTree* value_loop = get_loop(get_ctrl(value));
 813 
 814     // - address and value must be loop invariant
 815     // - memory must be a memory Phi for the loop
 816     // - Store must be the only store on this memory slice in the
 817     // loop: if there's another store following this one then value
 818     // written at iteration i by the second store could be overwritten
 819     // at iteration i+n by the first store: it's not safe to move the
 820     // first store out of the loop
 821     // - nothing must observe the memory Phi: it guarantees no read
 822     // before the store, we are also guaranteed the store post
 823     // dominates the loop head (ignoring a possible early
 824     // exit). Otherwise there would be extra Phi involved between the
 825     // loop's Phi and the store.
 826     // - there must be no early exit from the loop before the Store
 827     // (such an exit most of the time would be an extra use of the
 828     // memory Phi but sometimes is a bottom memory Phi that takes the
 829     // store as input).
 830 
 831     if (!n_loop->is_member(address_loop) &&
 832         !n_loop->is_member(value_loop) &&
 833         mem->is_Phi() && mem->in(0) == n_loop->_head &&
 834         mem->outcnt() == 1 &&
 835         mem->in(LoopNode::LoopBackControl) == n) {
 836 
 837       assert(n_loop->_tail != NULL, "need a tail");
 838       assert(is_dominator(n_ctrl, n_loop->_tail), "store control must not be in a branch in the loop");
 839 
 840       // Verify that there's no early exit of the loop before the store.
 841       bool ctrl_ok = false;
 842       {
 843         // Follow control from loop head until n, we exit the loop or
 844         // we reach the tail
 845         ResourceMark rm;
 846         Unique_Node_List wq;
 847         wq.push(n_loop->_head);
 848 
 849         for (uint next = 0; next < wq.size(); ++next) {
 850           Node *m = wq.at(next);
 851           if (m == n->in(0)) {
 852             ctrl_ok = true;
 853             continue;
 854           }
 855           assert(!has_ctrl(m), "should be CFG");
 856           if (!n_loop->is_member(get_loop(m)) || m == n_loop->_tail) {
 857             ctrl_ok = false;
 858             break;
 859           }
 860           enqueue_cfg_uses(m, wq);
 861           if (wq.size() > 10) {
 862             ctrl_ok = false;
 863             break;
 864           }
 865         }
 866       }
 867       if (ctrl_ok) {
 868         // move the Store
 869         _igvn.replace_input_of(mem, LoopNode::LoopBackControl, mem);
 870         _igvn.replace_input_of(n, 0, n_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl));
 871         _igvn.replace_input_of(n, MemNode::Memory, mem->in(LoopNode::EntryControl));
 872         // Disconnect the phi now. An empty phi can confuse other
 873         // optimizations in this pass of loop opts.
 874         _igvn.replace_node(mem, mem->in(LoopNode::EntryControl));
 875         n_loop->_body.yank(mem);
 876 
 877         set_ctrl_and_loop(n, n->in(0));
 878 
 879         return n;
 880       }
 881     }
 882   }
 883   return NULL;
 884 }
 885 
 886 // Try moving a store out of a loop, right after the loop
 887 void PhaseIdealLoop::try_move_store_after_loop(Node* n) {
 888   if (n->is_Store() && n->in(0) != NULL) {
 889     Node *n_ctrl = get_ctrl(n);
 890     IdealLoopTree *n_loop = get_loop(n_ctrl);
 891     // Store must be in a loop
 892     if (n_loop != _ltree_root && !n_loop->_irreducible) {
 893       Node* address = n->in(MemNode::Address);
 894       Node* value = n->in(MemNode::ValueIn);
 895       IdealLoopTree* address_loop = get_loop(get_ctrl(address));
 896       // address must be loop invariant
 897       if (!n_loop->is_member(address_loop)) {
 898         // Store must be last on this memory slice in the loop and
 899         // nothing in the loop must observe it
 900         Node* phi = NULL;
 901         for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 902           Node* u = n->fast_out(i);
 903           if (has_ctrl(u)) { // control use?
 904             IdealLoopTree *u_loop = get_loop(get_ctrl(u));
 905             if (!n_loop->is_member(u_loop)) {
 906               continue;
 907             }
 908             if (u->is_Phi() && u->in(0) == n_loop->_head) {
 909               assert(_igvn.type(u) == Type::MEMORY, "bad phi");
 910               // multiple phis on the same slice are possible
 911               if (phi != NULL) {
 912                 return;
 913               }
 914               phi = u;
 915               continue;
 916             }
 917           }
 918           return;
 919         }
 920         if (phi != NULL) {
 921           // Nothing in the loop before the store (next iteration)
 922           // must observe the stored value
 923           bool mem_ok = true;
 924           {
 925             ResourceMark rm;
 926             Unique_Node_List wq;
 927             wq.push(phi);
 928             for (uint next = 0; next < wq.size() && mem_ok; ++next) {
 929               Node *m = wq.at(next);
 930               for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax && mem_ok; i++) {
 931                 Node* u = m->fast_out(i);
 932                 if (u->is_Store() || u->is_Phi()) {
 933                   if (u != n) {
 934                     wq.push(u);
 935                     mem_ok = (wq.size() <= 10);
 936                   }
 937                 } else {
 938                   mem_ok = false;
 939                   break;
 940                 }
 941               }
 942             }
 943           }
 944           if (mem_ok) {
 945             // Move the store out of the loop if the LCA of all
 946             // users (except for the phi) is outside the loop.
 947             Node* hook = new Node(1);
 948             hook->init_req(0, n_ctrl); // Add an input to prevent hook from being dead
 949             _igvn.rehash_node_delayed(phi);
 950             int count = phi->replace_edge(n, hook);
 951             assert(count > 0, "inconsistent phi");
 952 
 953             // Compute latest point this store can go
 954             Node* lca = get_late_ctrl(n, get_ctrl(n));
 955             if (lca->is_OuterStripMinedLoop()) {
 956               lca = lca->in(LoopNode::EntryControl);
 957             }
 958             if (n_loop->is_member(get_loop(lca))) {
 959               // LCA is in the loop - bail out
 960               _igvn.replace_node(hook, n);
 961               return;
 962             }
 963 #ifdef ASSERT
 964             if (n_loop->_head->is_Loop() && n_loop->_head->as_Loop()->is_strip_mined()) {
 965               assert(n_loop->_head->Opcode() == Op_CountedLoop, "outer loop is a strip mined");
 966               n_loop->_head->as_Loop()->verify_strip_mined(1);
 967               Node* outer = n_loop->_head->as_CountedLoop()->outer_loop();
 968               IdealLoopTree* outer_loop = get_loop(outer);
 969               assert(n_loop->_parent == outer_loop, "broken loop tree");
 970               assert(get_loop(lca) == outer_loop, "safepoint in outer loop consume all memory state");
 971             }
 972 #endif
 973 
 974             // Move store out of the loop
 975             _igvn.replace_node(hook, n->in(MemNode::Memory));
 976             _igvn.replace_input_of(n, 0, lca);
 977             set_ctrl_and_loop(n, lca);
 978 
 979             // Disconnect the phi now. An empty phi can confuse other
 980             // optimizations in this pass of loop opts..
 981             if (phi->in(LoopNode::LoopBackControl) == phi) {
 982               _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
 983               n_loop->_body.yank(phi);
 984             }
 985           }
 986         }
 987       }
 988     }
 989   }
 990 }
 991 
 992 //------------------------------split_if_with_blocks_pre-----------------------
 993 // Do the real work in a non-recursive function.  Data nodes want to be
 994 // cloned in the pre-order so they can feed each other nicely.
 995 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
 996   // Cloning these guys is unlikely to win
 997   int n_op = n->Opcode();
 998   if( n_op == Op_MergeMem ) return n;
 999   if( n->is_Proj() ) return n;
1000   // Do not clone-up CmpFXXX variations, as these are always
1001   // followed by a CmpI
1002   if( n->is_Cmp() ) return n;
1003   // Attempt to use a conditional move instead of a phi/branch
1004   if( ConditionalMoveLimit > 0 && n_op == Op_Region ) {
1005     Node *cmov = conditional_move( n );
1006     if( cmov ) return cmov;
1007   }
1008   if( n->is_CFG() || n->is_LoadStore() )
1009     return n;
1010   if( n_op == Op_Opaque1 ||     // Opaque nodes cannot be mod'd
1011       n_op == Op_Opaque2 ) {
1012     if( !C->major_progress() )   // If chance of no more loop opts...
1013       _igvn._worklist.push(n);  // maybe we'll remove them
1014     return n;
1015   }
1016 
1017   if( n->is_Con() ) return n;   // No cloning for Con nodes
1018 
1019   Node *n_ctrl = get_ctrl(n);
1020   if( !n_ctrl ) return n;       // Dead node
1021 
1022   Node* res = try_move_store_before_loop(n, n_ctrl);
1023   if (res != NULL) {
1024     return n;
1025   }
1026 
1027   // Attempt to remix address expressions for loop invariants
1028   Node *m = remix_address_expressions( n );
1029   if( m ) return m;
1030 
1031   if (n_op == Op_AddI) {
1032     Node *nn = convert_add_to_muladd( n );
1033     if ( nn ) return nn;
1034   }
1035 
1036   if (n->is_ConstraintCast()) {
1037     Node* dom_cast = n->as_ConstraintCast()->dominating_cast(&_igvn, this);
1038     // ConstraintCastNode::dominating_cast() uses node control input to determine domination.
1039     // Node control inputs don't necessarily agree with loop control info (due to
1040     // transformations happened in between), thus additional dominance check is needed
1041     // to keep loop info valid.
1042     if (dom_cast != NULL && is_dominator(get_ctrl(dom_cast), get_ctrl(n))) {
1043       _igvn.replace_node(n, dom_cast);
1044       return dom_cast;
1045     }
1046   }
1047 
1048   // Determine if the Node has inputs from some local Phi.
1049   // Returns the block to clone thru.
1050   Node *n_blk = has_local_phi_input( n );
1051   if( !n_blk ) return n;
1052 
1053   // Do not clone the trip counter through on a CountedLoop
1054   // (messes up the canonical shape).
1055   if( n_blk->is_CountedLoop() && n->Opcode() == Op_AddI ) return n;
1056 
1057   // Check for having no control input; not pinned.  Allow
1058   // dominating control.
1059   if (n->in(0)) {
1060     Node *dom = idom(n_blk);
1061     if (dom_lca(n->in(0), dom) != n->in(0)) {
1062       return n;
1063     }
1064   }
1065   // Policy: when is it profitable.  You must get more wins than
1066   // policy before it is considered profitable.  Policy is usually 0,
1067   // so 1 win is considered profitable.  Big merges will require big
1068   // cloning, so get a larger policy.
1069   int policy = n_blk->req() >> 2;
1070 
1071   // If the loop is a candidate for range check elimination,
1072   // delay splitting through it's phi until a later loop optimization
1073   if (n_blk->is_CountedLoop()) {
1074     IdealLoopTree *lp = get_loop(n_blk);
1075     if (lp && lp->_rce_candidate) {
1076       return n;
1077     }
1078   }
1079 
1080   if (must_throttle_split_if()) return n;
1081 
1082   // Split 'n' through the merge point if it is profitable
1083   Node *phi = split_thru_phi( n, n_blk, policy );
1084   if (!phi) return n;
1085 
1086   // Found a Phi to split thru!
1087   // Replace 'n' with the new phi
1088   _igvn.replace_node( n, phi );
1089   // Moved a load around the loop, 'en-registering' something.
1090   if (n_blk->is_Loop() && n->is_Load() &&
1091       !phi->in(LoopNode::LoopBackControl)->is_Load())
1092     C->set_major_progress();
1093 
1094   return phi;
1095 }
1096 
1097 static bool merge_point_too_heavy(Compile* C, Node* region) {
1098   // Bail out if the region and its phis have too many users.
1099   int weight = 0;
1100   for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1101     weight += region->fast_out(i)->outcnt();
1102   }
1103   int nodes_left = C->max_node_limit() - C->live_nodes();
1104   if (weight * 8 > nodes_left) {
1105     if (PrintOpto) {
1106       tty->print_cr("*** Split-if bails out:  %d nodes, region weight %d", C->unique(), weight);
1107     }
1108     return true;
1109   } else {
1110     return false;
1111   }
1112 }
1113 
1114 static bool merge_point_safe(Node* region) {
1115   // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1116   // having a PhiNode input. This sidesteps the dangerous case where the split
1117   // ConvI2LNode may become TOP if the input Value() does not
1118   // overlap the ConvI2L range, leaving a node which may not dominate its
1119   // uses.
1120   // A better fix for this problem can be found in the BugTraq entry, but
1121   // expediency for Mantis demands this hack.
1122 #ifdef _LP64
1123   for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1124     Node* n = region->fast_out(i);
1125     if (n->is_Phi()) {
1126       for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1127         Node* m = n->fast_out(j);
1128         if (m->Opcode() == Op_ConvI2L)
1129           return false;
1130         if (m->is_CastII() && m->isa_CastII()->has_range_check()) {
1131           return false;
1132         }
1133       }
1134     }
1135   }
1136 #endif
1137   return true;
1138 }
1139 
1140 
1141 //------------------------------place_near_use---------------------------------
1142 // Place some computation next to use but not inside inner loops.
1143 // For inner loop uses move it to the preheader area.
1144 Node *PhaseIdealLoop::place_near_use(Node *useblock) const {
1145   IdealLoopTree *u_loop = get_loop( useblock );
1146   if (u_loop->_irreducible) {
1147     return useblock;
1148   }
1149   if (u_loop->_child) {
1150     if (useblock == u_loop->_head && u_loop->_head->is_OuterStripMinedLoop()) {
1151       return u_loop->_head->in(LoopNode::EntryControl);
1152     }
1153     return useblock;
1154   }
1155   return u_loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
1156 }
1157 
1158 
1159 bool PhaseIdealLoop::identical_backtoback_ifs(Node *n) {
1160   if (!n->is_If() || n->is_CountedLoopEnd()) {
1161     return false;
1162   }
1163   if (!n->in(0)->is_Region()) {
1164     return false;
1165   }
1166   Node* region = n->in(0);
1167   Node* dom = idom(region);
1168   if (!dom->is_If() || dom->in(1) != n->in(1)) {
1169     return false;
1170   }
1171   IfNode* dom_if = dom->as_If();
1172   Node* proj_true = dom_if->proj_out(1);
1173   Node* proj_false = dom_if->proj_out(0);
1174 
1175   for (uint i = 1; i < region->req(); i++) {
1176     if (is_dominator(proj_true, region->in(i))) {
1177       continue;
1178     }
1179     if (is_dominator(proj_false, region->in(i))) {
1180       continue;
1181     }
1182     return false;
1183   }
1184 
1185   return true;
1186 }
1187 
1188 
1189 bool PhaseIdealLoop::can_split_if(Node* n_ctrl) {
1190   if (must_throttle_split_if()) {
1191     return false;
1192   }
1193 
1194   // Do not do 'split-if' if irreducible loops are present.
1195   if (_has_irreducible_loops) {
1196     return false;
1197   }
1198 
1199   if (merge_point_too_heavy(C, n_ctrl)) {
1200     return false;
1201   }
1202 
1203   // Do not do 'split-if' if some paths are dead.  First do dead code
1204   // elimination and then see if its still profitable.
1205   for (uint i = 1; i < n_ctrl->req(); i++) {
1206     if (n_ctrl->in(i) == C->top()) {
1207       return false;
1208     }
1209   }
1210 
1211   // If trying to do a 'Split-If' at the loop head, it is only
1212   // profitable if the cmp folds up on BOTH paths.  Otherwise we
1213   // risk peeling a loop forever.
1214 
1215   // CNC - Disabled for now.  Requires careful handling of loop
1216   // body selection for the cloned code.  Also, make sure we check
1217   // for any input path not being in the same loop as n_ctrl.  For
1218   // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1219   // because the alternative loop entry points won't be converted
1220   // into LoopNodes.
1221   IdealLoopTree *n_loop = get_loop(n_ctrl);
1222   for (uint j = 1; j < n_ctrl->req(); j++) {
1223     if (get_loop(n_ctrl->in(j)) != n_loop) {
1224       return false;
1225     }
1226   }
1227 
1228   // Check for safety of the merge point.
1229   if (!merge_point_safe(n_ctrl)) {
1230     return false;
1231   }
1232 
1233   return true;
1234 }
1235 
1236 // Detect if the node is the inner strip-mined loop
1237 // Return: NULL if it's not the case, or the exit of outer strip-mined loop
1238 static Node* is_inner_of_stripmined_loop(const Node* out) {
1239   Node* out_le = NULL;
1240 
1241   if (out->is_CountedLoopEnd()) {
1242       const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode();
1243 
1244       if (loop != NULL && loop->is_strip_mined()) {
1245         out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit();
1246       }
1247   }
1248 
1249   return out_le;
1250 }
1251 
1252 //------------------------------split_if_with_blocks_post----------------------
1253 // Do the real work in a non-recursive function.  CFG hackery wants to be
1254 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1255 // info.
1256 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1257 
1258   // Cloning Cmp through Phi's involves the split-if transform.
1259   // FastLock is not used by an If
1260   if (n->is_Cmp() && !n->is_FastLock()) {
1261     Node *n_ctrl = get_ctrl(n);
1262     // Determine if the Node has inputs from some local Phi.
1263     // Returns the block to clone thru.
1264     Node *n_blk = has_local_phi_input(n);
1265     if (n_blk != n_ctrl) {
1266       return;
1267     }
1268 
1269     if (!can_split_if(n_ctrl)) {
1270       return;
1271     }
1272 
1273     if (n->outcnt() != 1) {
1274       return; // Multiple bool's from 1 compare?
1275     }
1276     Node *bol = n->unique_out();
1277     assert(bol->is_Bool(), "expect a bool here");
1278     if (bol->outcnt() != 1) {
1279       return;// Multiple branches from 1 compare?
1280     }
1281     Node *iff = bol->unique_out();
1282 
1283     // Check some safety conditions
1284     if (iff->is_If()) {        // Classic split-if?
1285       if (iff->in(0) != n_ctrl) {
1286         return; // Compare must be in same blk as if
1287       }
1288     } else if (iff->is_CMove()) { // Trying to split-up a CMOVE
1289       // Can't split CMove with different control edge.
1290       if (iff->in(0) != NULL && iff->in(0) != n_ctrl ) {
1291         return;
1292       }
1293       if (get_ctrl(iff->in(2)) == n_ctrl ||
1294           get_ctrl(iff->in(3)) == n_ctrl) {
1295         return;                 // Inputs not yet split-up
1296       }
1297       if (get_loop(n_ctrl) != get_loop(get_ctrl(iff))) {
1298         return;                 // Loop-invar test gates loop-varying CMOVE
1299       }
1300     } else {
1301       return;  // some other kind of node, such as an Allocate
1302     }
1303 
1304     // When is split-if profitable?  Every 'win' on means some control flow
1305     // goes dead, so it's almost always a win.
1306     int policy = 0;
1307     // Split compare 'n' through the merge point if it is profitable
1308     Node *phi = split_thru_phi( n, n_ctrl, policy);
1309     if (!phi) {
1310       return;
1311     }
1312 
1313     // Found a Phi to split thru!
1314     // Replace 'n' with the new phi
1315     _igvn.replace_node(n, phi);
1316 
1317     // Now split the bool up thru the phi
1318     Node *bolphi = split_thru_phi(bol, n_ctrl, -1);
1319     guarantee(bolphi != NULL, "null boolean phi node");
1320 
1321     _igvn.replace_node(bol, bolphi);
1322     assert(iff->in(1) == bolphi, "");
1323 
1324     if (bolphi->Value(&_igvn)->singleton()) {
1325       return;
1326     }
1327 
1328     // Conditional-move?  Must split up now
1329     if (!iff->is_If()) {
1330       Node *cmovphi = split_thru_phi(iff, n_ctrl, -1);
1331       _igvn.replace_node(iff, cmovphi);
1332       return;
1333     }
1334 
1335     // Now split the IF
1336     do_split_if(iff);
1337     return;
1338   }
1339 
1340   // Two identical ifs back to back can be merged
1341   if (identical_backtoback_ifs(n) && can_split_if(n->in(0))) {
1342     Node *n_ctrl = n->in(0);
1343     PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1344     IfNode* dom_if = idom(n_ctrl)->as_If();
1345     Node* proj_true = dom_if->proj_out(1);
1346     Node* proj_false = dom_if->proj_out(0);
1347     Node* con_true = _igvn.makecon(TypeInt::ONE);
1348     Node* con_false = _igvn.makecon(TypeInt::ZERO);
1349 
1350     for (uint i = 1; i < n_ctrl->req(); i++) {
1351       if (is_dominator(proj_true, n_ctrl->in(i))) {
1352         bolphi->init_req(i, con_true);
1353       } else {
1354         assert(is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1355         bolphi->init_req(i, con_false);
1356       }
1357     }
1358     register_new_node(bolphi, n_ctrl);
1359     _igvn.replace_input_of(n, 1, bolphi);
1360 
1361     // Now split the IF
1362     do_split_if(n);
1363     return;
1364   }
1365 
1366   // Check for an IF ready to split; one that has its
1367   // condition codes input coming from a Phi at the block start.
1368   int n_op = n->Opcode();
1369 
1370   // Check for an IF being dominated by another IF same test
1371   if (n_op == Op_If ||
1372       n_op == Op_RangeCheck) {
1373     Node *bol = n->in(1);
1374     uint max = bol->outcnt();
1375     // Check for same test used more than once?
1376     if (max > 1 && bol->is_Bool()) {
1377       // Search up IDOMs to see if this IF is dominated.
1378       Node *cutoff = get_ctrl(bol);
1379 
1380       // Now search up IDOMs till cutoff, looking for a dominating test
1381       Node *prevdom = n;
1382       Node *dom = idom(prevdom);
1383       while (dom != cutoff) {
1384         if (dom->req() > 1 && dom->in(1) == bol && prevdom->in(0) == dom) {
1385           // It's invalid to move control dependent data nodes in the inner
1386           // strip-mined loop, because:
1387           //  1) break validation of LoopNode::verify_strip_mined()
1388           //  2) move code with side-effect in strip-mined loop
1389           // Move to the exit of outer strip-mined loop in that case.
1390           Node* out_le = is_inner_of_stripmined_loop(dom);
1391           if (out_le != NULL) {
1392             prevdom = out_le;
1393           }
1394           // Replace the dominated test with an obvious true or false.
1395           // Place it on the IGVN worklist for later cleanup.
1396           C->set_major_progress();
1397           dominated_by(prevdom, n, false, true);
1398 #ifndef PRODUCT
1399           if( VerifyLoopOptimizations ) verify();
1400 #endif
1401           return;
1402         }
1403         prevdom = dom;
1404         dom = idom(prevdom);
1405       }
1406     }
1407   }
1408 
1409   // See if a shared loop-varying computation has no loop-varying uses.
1410   // Happens if something is only used for JVM state in uncommon trap exits,
1411   // like various versions of induction variable+offset.  Clone the
1412   // computation per usage to allow it to sink out of the loop.
1413   if (has_ctrl(n) && !n->in(0)) {// n not dead and has no control edge (can float about)
1414     Node *n_ctrl = get_ctrl(n);
1415     IdealLoopTree *n_loop = get_loop(n_ctrl);
1416     if( n_loop != _ltree_root ) {
1417       DUIterator_Fast imax, i = n->fast_outs(imax);
1418       for (; i < imax; i++) {
1419         Node* u = n->fast_out(i);
1420         if( !has_ctrl(u) )     break; // Found control user
1421         IdealLoopTree *u_loop = get_loop(get_ctrl(u));
1422         if( u_loop == n_loop ) break; // Found loop-varying use
1423         if( n_loop->is_member( u_loop ) ) break; // Found use in inner loop
1424         if( u->Opcode() == Op_Opaque1 ) break; // Found loop limit, bugfix for 4677003
1425       }
1426       bool did_break = (i < imax);  // Did we break out of the previous loop?
1427       if (!did_break && n->outcnt() > 1) { // All uses in outer loops!
1428         Node *late_load_ctrl = NULL;
1429         if (n->is_Load()) {
1430           // If n is a load, get and save the result from get_late_ctrl(),
1431           // to be later used in calculating the control for n's clones.
1432           clear_dom_lca_tags();
1433           late_load_ctrl = get_late_ctrl(n, n_ctrl);
1434         }
1435         // If n is a load, and the late control is the same as the current
1436         // control, then the cloning of n is a pointless exercise, because
1437         // GVN will ensure that we end up where we started.
1438         if (!n->is_Load() || late_load_ctrl != n_ctrl) {
1439           for (DUIterator_Last jmin, j = n->last_outs(jmin); j >= jmin; ) {
1440             Node *u = n->last_out(j); // Clone private computation per use
1441             _igvn.rehash_node_delayed(u);
1442             Node *x = n->clone(); // Clone computation
1443             Node *x_ctrl = NULL;
1444             if( u->is_Phi() ) {
1445               // Replace all uses of normal nodes.  Replace Phi uses
1446               // individually, so the separate Nodes can sink down
1447               // different paths.
1448               uint k = 1;
1449               while( u->in(k) != n ) k++;
1450               u->set_req( k, x );
1451               // x goes next to Phi input path
1452               x_ctrl = u->in(0)->in(k);
1453               --j;
1454             } else {              // Normal use
1455               // Replace all uses
1456               for( uint k = 0; k < u->req(); k++ ) {
1457                 if( u->in(k) == n ) {
1458                   u->set_req( k, x );
1459                   --j;
1460                 }
1461               }
1462               x_ctrl = get_ctrl(u);
1463             }
1464 
1465             // Find control for 'x' next to use but not inside inner loops.
1466             // For inner loop uses get the preheader area.
1467             x_ctrl = place_near_use(x_ctrl);
1468 
1469             if (n->is_Load()) {
1470               // For loads, add a control edge to a CFG node outside of the loop
1471               // to force them to not combine and return back inside the loop
1472               // during GVN optimization (4641526).
1473               //
1474               // Because we are setting the actual control input, factor in
1475               // the result from get_late_ctrl() so we respect any
1476               // anti-dependences. (6233005).
1477               x_ctrl = dom_lca(late_load_ctrl, x_ctrl);
1478 
1479               // Don't allow the control input to be a CFG splitting node.
1480               // Such nodes should only have ProjNodes as outs, e.g. IfNode
1481               // should only have IfTrueNode and IfFalseNode (4985384).
1482               x_ctrl = find_non_split_ctrl(x_ctrl);
1483 
1484               IdealLoopTree* x_loop = get_loop(x_ctrl);
1485               Node* x_head = x_loop->_head;
1486               if (x_head->is_Loop() && (x_head->is_OuterStripMinedLoop() || x_head->as_Loop()->is_strip_mined()) && is_dominator(n_ctrl, x_head)) {
1487                 // Anti dependence analysis is sometimes too
1488                 // conservative: a store in the outer strip mined loop
1489                 // can prevent a load from floating out of the outer
1490                 // strip mined loop but the load may not be referenced
1491                 // from the safepoint: loop strip mining verification
1492                 // code reports a problem in that case. Make sure the
1493                 // load is not moved in the outer strip mined loop in
1494                 // that case.
1495                 x_ctrl = x_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
1496               }
1497               assert(dom_depth(n_ctrl) <= dom_depth(x_ctrl), "n is later than its clone");
1498 
1499               x->set_req(0, x_ctrl);
1500             }
1501             register_new_node(x, x_ctrl);
1502 
1503             // Some institutional knowledge is needed here: 'x' is
1504             // yanked because if the optimizer runs GVN on it all the
1505             // cloned x's will common up and undo this optimization and
1506             // be forced back in the loop.
1507             // I tried setting control edges on the x's to force them to
1508             // not combine, but the matching gets worried when it tries
1509             // to fold a StoreP and an AddP together (as part of an
1510             // address expression) and the AddP and StoreP have
1511             // different controls.
1512             if (!x->is_Load() && !x->is_DecodeNarrowPtr()) _igvn._worklist.yank(x);
1513           }
1514           _igvn.remove_dead_node(n);
1515         }
1516       }
1517     }
1518   }
1519 
1520   try_move_store_after_loop(n);
1521 
1522   // Check for Opaque2's who's loop has disappeared - who's input is in the
1523   // same loop nest as their output.  Remove 'em, they are no longer useful.
1524   if( n_op == Op_Opaque2 &&
1525       n->in(1) != NULL &&
1526       get_loop(get_ctrl(n)) == get_loop(get_ctrl(n->in(1))) ) {
1527     _igvn.replace_node( n, n->in(1) );
1528   }
1529 }
1530 
1531 //------------------------------split_if_with_blocks---------------------------
1532 // Check for aggressive application of 'split-if' optimization,
1533 // using basic block level info.
1534 void PhaseIdealLoop::split_if_with_blocks(VectorSet &visited, Node_Stack &nstack) {
1535   Node* root = C->root();
1536   visited.set(root->_idx); // first, mark root as visited
1537   // Do pre-visit work for root
1538   Node* n   = split_if_with_blocks_pre(root);
1539   uint  cnt = n->outcnt();
1540   uint  i   = 0;
1541 
1542   while (true) {
1543     // Visit all children
1544     if (i < cnt) {
1545       Node* use = n->raw_out(i);
1546       ++i;
1547       if (use->outcnt() != 0 && !visited.test_set(use->_idx)) {
1548         // Now do pre-visit work for this use
1549         use = split_if_with_blocks_pre(use);
1550         nstack.push(n, i); // Save parent and next use's index.
1551         n   = use;         // Process all children of current use.
1552         cnt = use->outcnt();
1553         i   = 0;
1554       }
1555     }
1556     else {
1557       // All of n's children have been processed, complete post-processing.
1558       if (cnt != 0 && !n->is_Con()) {
1559         assert(has_node(n), "no dead nodes");
1560         split_if_with_blocks_post(n);
1561       }
1562       if (must_throttle_split_if()) {
1563         nstack.clear();
1564       }
1565       if (nstack.is_empty()) {
1566         // Finished all nodes on stack.
1567         break;
1568       }
1569       // Get saved parent node and next use's index. Visit the rest of uses.
1570       n   = nstack.node();
1571       cnt = n->outcnt();
1572       i   = nstack.index();
1573       nstack.pop();
1574     }
1575   }
1576 }
1577 
1578 
1579 //=============================================================================
1580 //
1581 //                   C L O N E   A   L O O P   B O D Y
1582 //
1583 
1584 //------------------------------clone_iff--------------------------------------
1585 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1586 // "Nearly" because all Nodes have been cloned from the original in the loop,
1587 // but the fall-in edges to the Cmp are different.  Clone bool/Cmp pairs
1588 // through the Phi recursively, and return a Bool.
1589 Node* PhaseIdealLoop::clone_iff(PhiNode *phi, IdealLoopTree *loop) {
1590 
1591   // Convert this Phi into a Phi merging Bools
1592   uint i;
1593   for (i = 1; i < phi->req(); i++) {
1594     Node *b = phi->in(i);
1595     if (b->is_Phi()) {
1596       _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi(), loop));
1597     } else {
1598       assert(b->is_Bool() || b->Opcode() == Op_Opaque4, "");
1599     }
1600   }
1601 
1602   Node* n = phi->in(1);
1603   Node* sample_opaque = NULL;
1604   Node *sample_bool = NULL;
1605   if (n->Opcode() == Op_Opaque4) {
1606     sample_opaque = n;
1607     sample_bool = n->in(1);
1608     assert(sample_bool->is_Bool(), "wrong type");
1609   } else {
1610     sample_bool = n;
1611   }
1612   Node *sample_cmp = sample_bool->in(1);
1613 
1614   // Make Phis to merge the Cmp's inputs.
1615   PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
1616   PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
1617   for (i = 1; i < phi->req(); i++) {
1618     Node *n1 = sample_opaque == NULL ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
1619     Node *n2 = sample_opaque == NULL ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
1620     phi1->set_req(i, n1);
1621     phi2->set_req(i, n2);
1622     phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1623     phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1624   }
1625   // See if these Phis have been made before.
1626   // Register with optimizer
1627   Node *hit1 = _igvn.hash_find_insert(phi1);
1628   if (hit1) {                   // Hit, toss just made Phi
1629     _igvn.remove_dead_node(phi1); // Remove new phi
1630     assert(hit1->is_Phi(), "" );
1631     phi1 = (PhiNode*)hit1;      // Use existing phi
1632   } else {                      // Miss
1633     _igvn.register_new_node_with_optimizer(phi1);
1634   }
1635   Node *hit2 = _igvn.hash_find_insert(phi2);
1636   if (hit2) {                   // Hit, toss just made Phi
1637     _igvn.remove_dead_node(phi2); // Remove new phi
1638     assert(hit2->is_Phi(), "" );
1639     phi2 = (PhiNode*)hit2;      // Use existing phi
1640   } else {                      // Miss
1641     _igvn.register_new_node_with_optimizer(phi2);
1642   }
1643   // Register Phis with loop/block info
1644   set_ctrl(phi1, phi->in(0));
1645   set_ctrl(phi2, phi->in(0));
1646   // Make a new Cmp
1647   Node *cmp = sample_cmp->clone();
1648   cmp->set_req(1, phi1);
1649   cmp->set_req(2, phi2);
1650   _igvn.register_new_node_with_optimizer(cmp);
1651   set_ctrl(cmp, phi->in(0));
1652 
1653   // Make a new Bool
1654   Node *b = sample_bool->clone();
1655   b->set_req(1,cmp);
1656   _igvn.register_new_node_with_optimizer(b);
1657   set_ctrl(b, phi->in(0));
1658 
1659   if (sample_opaque != NULL) {
1660     Node* opaque = sample_opaque->clone();
1661     opaque->set_req(1, b);
1662     _igvn.register_new_node_with_optimizer(opaque);
1663     set_ctrl(opaque, phi->in(0));
1664     return opaque;
1665   }
1666 
1667   assert(b->is_Bool(), "");
1668   return b;
1669 }
1670 
1671 //------------------------------clone_bool-------------------------------------
1672 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1673 // "Nearly" because all Nodes have been cloned from the original in the loop,
1674 // but the fall-in edges to the Cmp are different.  Clone bool/Cmp pairs
1675 // through the Phi recursively, and return a Bool.
1676 CmpNode *PhaseIdealLoop::clone_bool( PhiNode *phi, IdealLoopTree *loop ) {
1677   uint i;
1678   // Convert this Phi into a Phi merging Bools
1679   for( i = 1; i < phi->req(); i++ ) {
1680     Node *b = phi->in(i);
1681     if( b->is_Phi() ) {
1682       _igvn.replace_input_of(phi, i, clone_bool( b->as_Phi(), loop ));
1683     } else {
1684       assert( b->is_Cmp() || b->is_top(), "inputs are all Cmp or TOP" );
1685     }
1686   }
1687 
1688   Node *sample_cmp = phi->in(1);
1689 
1690   // Make Phis to merge the Cmp's inputs.
1691   PhiNode *phi1 = new PhiNode( phi->in(0), Type::TOP );
1692   PhiNode *phi2 = new PhiNode( phi->in(0), Type::TOP );
1693   for( uint j = 1; j < phi->req(); j++ ) {
1694     Node *cmp_top = phi->in(j); // Inputs are all Cmp or TOP
1695     Node *n1, *n2;
1696     if( cmp_top->is_Cmp() ) {
1697       n1 = cmp_top->in(1);
1698       n2 = cmp_top->in(2);
1699     } else {
1700       n1 = n2 = cmp_top;
1701     }
1702     phi1->set_req( j, n1 );
1703     phi2->set_req( j, n2 );
1704     phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
1705     phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
1706   }
1707 
1708   // See if these Phis have been made before.
1709   // Register with optimizer
1710   Node *hit1 = _igvn.hash_find_insert(phi1);
1711   if( hit1 ) {                  // Hit, toss just made Phi
1712     _igvn.remove_dead_node(phi1); // Remove new phi
1713     assert( hit1->is_Phi(), "" );
1714     phi1 = (PhiNode*)hit1;      // Use existing phi
1715   } else {                      // Miss
1716     _igvn.register_new_node_with_optimizer(phi1);
1717   }
1718   Node *hit2 = _igvn.hash_find_insert(phi2);
1719   if( hit2 ) {                  // Hit, toss just made Phi
1720     _igvn.remove_dead_node(phi2); // Remove new phi
1721     assert( hit2->is_Phi(), "" );
1722     phi2 = (PhiNode*)hit2;      // Use existing phi
1723   } else {                      // Miss
1724     _igvn.register_new_node_with_optimizer(phi2);
1725   }
1726   // Register Phis with loop/block info
1727   set_ctrl(phi1, phi->in(0));
1728   set_ctrl(phi2, phi->in(0));
1729   // Make a new Cmp
1730   Node *cmp = sample_cmp->clone();
1731   cmp->set_req( 1, phi1 );
1732   cmp->set_req( 2, phi2 );
1733   _igvn.register_new_node_with_optimizer(cmp);
1734   set_ctrl(cmp, phi->in(0));
1735 
1736   assert( cmp->is_Cmp(), "" );
1737   return (CmpNode*)cmp;
1738 }
1739 
1740 //------------------------------sink_use---------------------------------------
1741 // If 'use' was in the loop-exit block, it now needs to be sunk
1742 // below the post-loop merge point.
1743 void PhaseIdealLoop::sink_use( Node *use, Node *post_loop ) {
1744   if (!use->is_CFG() && get_ctrl(use) == post_loop->in(2)) {
1745     set_ctrl(use, post_loop);
1746     for (DUIterator j = use->outs(); use->has_out(j); j++)
1747       sink_use(use->out(j), post_loop);
1748   }
1749 }
1750 
1751 void PhaseIdealLoop::clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1752                                                  IdealLoopTree* loop, IdealLoopTree* outer_loop,
1753                                                  Node_List*& split_if_set, Node_List*& split_bool_set,
1754                                                  Node_List*& split_cex_set, Node_List& worklist,
1755                                                  uint new_counter, CloneLoopMode mode) {
1756   Node* nnn = old_new[old->_idx];
1757   // Copy uses to a worklist, so I can munge the def-use info
1758   // with impunity.
1759   for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
1760     worklist.push(old->fast_out(j));
1761 
1762   while( worklist.size() ) {
1763     Node *use = worklist.pop();
1764     if (!has_node(use))  continue; // Ignore dead nodes
1765     if (use->in(0) == C->top())  continue;
1766     IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
1767     // Check for data-use outside of loop - at least one of OLD or USE
1768     // must not be a CFG node.
1769 #ifdef ASSERT
1770     if (loop->_head->as_Loop()->is_strip_mined() && outer_loop->is_member(use_loop) && !loop->is_member(use_loop) && old_new[use->_idx] == NULL) {
1771       Node* sfpt = loop->_head->as_CountedLoop()->outer_safepoint();
1772       assert(mode != IgnoreStripMined, "incorrect cloning mode");
1773       assert((mode == ControlAroundStripMined && use == sfpt) || !use->is_reachable_from_root(), "missed a node");
1774     }
1775 #endif
1776     if (!loop->is_member(use_loop) && !outer_loop->is_member(use_loop) && (!old->is_CFG() || !use->is_CFG())) {
1777 
1778       // If the Data use is an IF, that means we have an IF outside of the
1779       // loop that is switching on a condition that is set inside of the
1780       // loop.  Happens if people set a loop-exit flag; then test the flag
1781       // in the loop to break the loop, then test is again outside of the
1782       // loop to determine which way the loop exited.
1783       // Loop predicate If node connects to Bool node through Opaque1 node.
1784       if (use->is_If() || use->is_CMove() || C->is_predicate_opaq(use) || use->Opcode() == Op_Opaque4) {
1785         // Since this code is highly unlikely, we lazily build the worklist
1786         // of such Nodes to go split.
1787         if (!split_if_set) {
1788           split_if_set = new Node_List();
1789         }
1790         split_if_set->push(use);
1791       }
1792       if (use->is_Bool()) {
1793         if (!split_bool_set) {
1794           split_bool_set = new Node_List();
1795         }
1796         split_bool_set->push(use);
1797       }
1798       if (use->Opcode() == Op_CreateEx) {
1799         if (!split_cex_set) {
1800           split_cex_set = new Node_List();
1801         }
1802         split_cex_set->push(use);
1803       }
1804 
1805 
1806       // Get "block" use is in
1807       uint idx = 0;
1808       while( use->in(idx) != old ) idx++;
1809       Node *prev = use->is_CFG() ? use : get_ctrl(use);
1810       assert(!loop->is_member(get_loop(prev)) && !outer_loop->is_member(get_loop(prev)), "" );
1811       Node *cfg = prev->_idx >= new_counter
1812         ? prev->in(2)
1813         : idom(prev);
1814       if( use->is_Phi() )     // Phi use is in prior block
1815         cfg = prev->in(idx);  // NOT in block of Phi itself
1816       if (cfg->is_top()) {    // Use is dead?
1817         _igvn.replace_input_of(use, idx, C->top());
1818         continue;
1819       }
1820 
1821       // If use is referenced through control edge... (idx == 0)
1822       if (mode == IgnoreStripMined && idx == 0) {
1823         LoopNode *head = loop->_head->as_Loop();
1824         if (head->is_strip_mined() && is_dominator(head->outer_loop_exit(), prev)) {
1825           // That node is outside the inner loop, leave it outside the
1826           // outer loop as well to not confuse verification code.
1827           assert(!loop->_parent->is_member(use_loop), "should be out of the outer loop");
1828           _igvn.replace_input_of(use, 0, head->outer_loop_exit());
1829           continue;
1830         }
1831       }
1832 
1833       while(!outer_loop->is_member(get_loop(cfg))) {
1834         prev = cfg;
1835         cfg = cfg->_idx >= new_counter ? cfg->in(2) : idom(cfg);
1836       }
1837       // If the use occurs after merging several exits from the loop, then
1838       // old value must have dominated all those exits.  Since the same old
1839       // value was used on all those exits we did not need a Phi at this
1840       // merge point.  NOW we do need a Phi here.  Each loop exit value
1841       // is now merged with the peeled body exit; each exit gets its own
1842       // private Phi and those Phis need to be merged here.
1843       Node *phi;
1844       if( prev->is_Region() ) {
1845         if( idx == 0 ) {      // Updating control edge?
1846           phi = prev;         // Just use existing control
1847         } else {              // Else need a new Phi
1848           phi = PhiNode::make( prev, old );
1849           // Now recursively fix up the new uses of old!
1850           for( uint i = 1; i < prev->req(); i++ ) {
1851             worklist.push(phi); // Onto worklist once for each 'old' input
1852           }
1853         }
1854       } else {
1855         // Get new RegionNode merging old and new loop exits
1856         prev = old_new[prev->_idx];
1857         assert( prev, "just made this in step 7" );
1858         if( idx == 0) {      // Updating control edge?
1859           phi = prev;         // Just use existing control
1860         } else {              // Else need a new Phi
1861           // Make a new Phi merging data values properly
1862           phi = PhiNode::make( prev, old );
1863           phi->set_req( 1, nnn );
1864         }
1865       }
1866       // If inserting a new Phi, check for prior hits
1867       if( idx != 0 ) {
1868         Node *hit = _igvn.hash_find_insert(phi);
1869         if( hit == NULL ) {
1870           _igvn.register_new_node_with_optimizer(phi); // Register new phi
1871         } else {                                      // or
1872           // Remove the new phi from the graph and use the hit
1873           _igvn.remove_dead_node(phi);
1874           phi = hit;                                  // Use existing phi
1875         }
1876         set_ctrl(phi, prev);
1877       }
1878       // Make 'use' use the Phi instead of the old loop body exit value
1879       _igvn.replace_input_of(use, idx, phi);
1880       if( use->_idx >= new_counter ) { // If updating new phis
1881         // Not needed for correctness, but prevents a weak assert
1882         // in AddPNode from tripping (when we end up with different
1883         // base & derived Phis that will become the same after
1884         // IGVN does CSE).
1885         Node *hit = _igvn.hash_find_insert(use);
1886         if( hit )             // Go ahead and re-hash for hits.
1887           _igvn.replace_node( use, hit );
1888       }
1889 
1890       // If 'use' was in the loop-exit block, it now needs to be sunk
1891       // below the post-loop merge point.
1892       sink_use( use, prev );
1893     }
1894   }
1895 }
1896 
1897 static void clone_outer_loop_helper(Node* n, const IdealLoopTree *loop, const IdealLoopTree* outer_loop,
1898                                     const Node_List &old_new, Unique_Node_List& wq, PhaseIdealLoop* phase,
1899                                     bool check_old_new) {
1900   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1901     Node* u = n->fast_out(j);
1902     assert(check_old_new || old_new[u->_idx] == NULL, "shouldn't have been cloned");
1903     if (!u->is_CFG() && (!check_old_new || old_new[u->_idx] == NULL)) {
1904       Node* c = phase->get_ctrl(u);
1905       IdealLoopTree* u_loop = phase->get_loop(c);
1906       assert(!loop->is_member(u_loop), "can be in outer loop or out of both loops only");
1907       if (outer_loop->is_member(u_loop)) {
1908         wq.push(u);
1909       }
1910     }
1911   }
1912 }
1913 
1914 void PhaseIdealLoop::clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1915                                       IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1916                                       Node_List& extra_data_nodes) {
1917   if (head->is_strip_mined() && mode != IgnoreStripMined) {
1918     CountedLoopNode* cl = head->as_CountedLoop();
1919     Node* l = cl->outer_loop();
1920     Node* tail = cl->outer_loop_tail();
1921     IfNode* le = cl->outer_loop_end();
1922     Node* sfpt = cl->outer_safepoint();
1923     CountedLoopEndNode* cle = cl->loopexit();
1924     CountedLoopNode* new_cl = old_new[cl->_idx]->as_CountedLoop();
1925     CountedLoopEndNode* new_cle = new_cl->as_CountedLoop()->loopexit_or_null();
1926     Node* cle_out = cle->proj_out(false);
1927 
1928     Node* new_sfpt = NULL;
1929     Node* new_cle_out = cle_out->clone();
1930     old_new.map(cle_out->_idx, new_cle_out);
1931     if (mode == CloneIncludesStripMined) {
1932       // clone outer loop body
1933       Node* new_l = l->clone();
1934       Node* new_tail = tail->clone();
1935       IfNode* new_le = le->clone()->as_If();
1936       new_sfpt = sfpt->clone();
1937 
1938       set_loop(new_l, outer_loop->_parent);
1939       set_idom(new_l, new_l->in(LoopNode::EntryControl), dd);
1940       set_loop(new_cle_out, outer_loop->_parent);
1941       set_idom(new_cle_out, new_cle, dd);
1942       set_loop(new_sfpt, outer_loop->_parent);
1943       set_idom(new_sfpt, new_cle_out, dd);
1944       set_loop(new_le, outer_loop->_parent);
1945       set_idom(new_le, new_sfpt, dd);
1946       set_loop(new_tail, outer_loop->_parent);
1947       set_idom(new_tail, new_le, dd);
1948       set_idom(new_cl, new_l, dd);
1949 
1950       old_new.map(l->_idx, new_l);
1951       old_new.map(tail->_idx, new_tail);
1952       old_new.map(le->_idx, new_le);
1953       old_new.map(sfpt->_idx, new_sfpt);
1954 
1955       new_l->set_req(LoopNode::LoopBackControl, new_tail);
1956       new_l->set_req(0, new_l);
1957       new_tail->set_req(0, new_le);
1958       new_le->set_req(0, new_sfpt);
1959       new_sfpt->set_req(0, new_cle_out);
1960       new_cle_out->set_req(0, new_cle);
1961       new_cl->set_req(LoopNode::EntryControl, new_l);
1962 
1963       _igvn.register_new_node_with_optimizer(new_l);
1964       _igvn.register_new_node_with_optimizer(new_tail);
1965       _igvn.register_new_node_with_optimizer(new_le);
1966     } else {
1967       Node *newhead = old_new[loop->_head->_idx];
1968       newhead->as_Loop()->clear_strip_mined();
1969       _igvn.replace_input_of(newhead, LoopNode::EntryControl, newhead->in(LoopNode::EntryControl)->in(LoopNode::EntryControl));
1970       set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
1971     }
1972     // Look at data node that were assigned a control in the outer
1973     // loop: they are kept in the outer loop by the safepoint so start
1974     // from the safepoint node's inputs.
1975     IdealLoopTree* outer_loop = get_loop(l);
1976     Node_Stack stack(2);
1977     stack.push(sfpt, 1);
1978     uint new_counter = C->unique();
1979     while (stack.size() > 0) {
1980       Node* n = stack.node();
1981       uint i = stack.index();
1982       while (i < n->req() &&
1983              (n->in(i) == NULL ||
1984               !has_ctrl(n->in(i)) ||
1985               get_loop(get_ctrl(n->in(i))) != outer_loop ||
1986               (old_new[n->in(i)->_idx] != NULL && old_new[n->in(i)->_idx]->_idx >= new_counter))) {
1987         i++;
1988       }
1989       if (i < n->req()) {
1990         stack.set_index(i+1);
1991         stack.push(n->in(i), 0);
1992       } else {
1993         assert(old_new[n->_idx] == NULL || n == sfpt || old_new[n->_idx]->_idx < new_counter, "no clone yet");
1994         Node* m = n == sfpt ? new_sfpt : n->clone();
1995         if (m != NULL) {
1996           for (uint i = 0; i < n->req(); i++) {
1997             if (m->in(i) != NULL && old_new[m->in(i)->_idx] != NULL) {
1998               m->set_req(i, old_new[m->in(i)->_idx]);
1999             }
2000           }
2001         } else {
2002           assert(n == sfpt && mode != CloneIncludesStripMined, "where's the safepoint clone?");
2003         }
2004         if (n != sfpt) {
2005           extra_data_nodes.push(n);
2006           _igvn.register_new_node_with_optimizer(m);
2007           assert(get_ctrl(n) == cle_out, "what other control?");
2008           set_ctrl(m, new_cle_out);
2009           old_new.map(n->_idx, m);
2010         }
2011         stack.pop();
2012       }
2013     }
2014     if (mode == CloneIncludesStripMined) {
2015       _igvn.register_new_node_with_optimizer(new_sfpt);
2016       _igvn.register_new_node_with_optimizer(new_cle_out);
2017     }
2018     // Some other transformation may have pessimistically assign some
2019     // data nodes to the outer loop. Set their control so they are out
2020     // of the outer loop.
2021     ResourceMark rm;
2022     Unique_Node_List wq;
2023     for (uint i = 0; i < extra_data_nodes.size(); i++) {
2024       Node* old = extra_data_nodes.at(i);
2025       clone_outer_loop_helper(old, loop, outer_loop, old_new, wq, this, true);
2026     }
2027     Node* new_ctrl = cl->outer_loop_exit();
2028     assert(get_loop(new_ctrl) != outer_loop, "must be out of the loop nest");
2029     for (uint i = 0; i < wq.size(); i++) {
2030       Node* n = wq.at(i);
2031       set_ctrl(n, new_ctrl);
2032       clone_outer_loop_helper(n, loop, outer_loop, old_new, wq, this, false);
2033     }
2034   } else {
2035     Node *newhead = old_new[loop->_head->_idx];
2036     set_idom(newhead, newhead->in(LoopNode::EntryControl), dd);
2037   }
2038 }
2039 
2040 //------------------------------clone_loop-------------------------------------
2041 //
2042 //                   C L O N E   A   L O O P   B O D Y
2043 //
2044 // This is the basic building block of the loop optimizations.  It clones an
2045 // entire loop body.  It makes an old_new loop body mapping; with this mapping
2046 // you can find the new-loop equivalent to an old-loop node.  All new-loop
2047 // nodes are exactly equal to their old-loop counterparts, all edges are the
2048 // same.  All exits from the old-loop now have a RegionNode that merges the
2049 // equivalent new-loop path.  This is true even for the normal "loop-exit"
2050 // condition.  All uses of loop-invariant old-loop values now come from (one
2051 // or more) Phis that merge their new-loop equivalents.
2052 //
2053 // This operation leaves the graph in an illegal state: there are two valid
2054 // control edges coming from the loop pre-header to both loop bodies.  I'll
2055 // definitely have to hack the graph after running this transform.
2056 //
2057 // From this building block I will further edit edges to perform loop peeling
2058 // or loop unrolling or iteration splitting (Range-Check-Elimination), etc.
2059 //
2060 // Parameter side_by_size_idom:
2061 //   When side_by_size_idom is NULL, the dominator tree is constructed for
2062 //      the clone loop to dominate the original.  Used in construction of
2063 //      pre-main-post loop sequence.
2064 //   When nonnull, the clone and original are side-by-side, both are
2065 //      dominated by the side_by_side_idom node.  Used in construction of
2066 //      unswitched loops.
2067 void PhaseIdealLoop::clone_loop( IdealLoopTree *loop, Node_List &old_new, int dd,
2068                                 CloneLoopMode mode, Node* side_by_side_idom) {
2069 
2070   LoopNode* head = loop->_head->as_Loop();
2071   head->verify_strip_mined(1);
2072 
2073   if (C->do_vector_loop() && PrintOpto) {
2074     const char* mname = C->method()->name()->as_quoted_ascii();
2075     if (mname != NULL) {
2076       tty->print("PhaseIdealLoop::clone_loop: for vectorize method %s\n", mname);
2077     }
2078   }
2079 
2080   CloneMap& cm = C->clone_map();
2081   Dict* dict = cm.dict();
2082   if (C->do_vector_loop()) {
2083     cm.set_clone_idx(cm.max_gen()+1);
2084 #ifndef PRODUCT
2085     if (PrintOpto) {
2086       tty->print_cr("PhaseIdealLoop::clone_loop: _clone_idx %d", cm.clone_idx());
2087       loop->dump_head();
2088     }
2089 #endif
2090   }
2091 
2092   // Step 1: Clone the loop body.  Make the old->new mapping.
2093   uint i;
2094   for( i = 0; i < loop->_body.size(); i++ ) {
2095     Node *old = loop->_body.at(i);
2096     Node *nnn = old->clone();
2097     old_new.map( old->_idx, nnn );
2098     if (C->do_vector_loop()) {
2099       cm.verify_insert_and_clone(old, nnn, cm.clone_idx());
2100     }
2101     _igvn.register_new_node_with_optimizer(nnn);
2102   }
2103 
2104   IdealLoopTree* outer_loop = (head->is_strip_mined() && mode != IgnoreStripMined) ? get_loop(head->as_CountedLoop()->outer_loop()) : loop;
2105 
2106   // Step 2: Fix the edges in the new body.  If the old input is outside the
2107   // loop use it.  If the old input is INside the loop, use the corresponding
2108   // new node instead.
2109   for( i = 0; i < loop->_body.size(); i++ ) {
2110     Node *old = loop->_body.at(i);
2111     Node *nnn = old_new[old->_idx];
2112     // Fix CFG/Loop controlling the new node
2113     if (has_ctrl(old)) {
2114       set_ctrl(nnn, old_new[get_ctrl(old)->_idx]);
2115     } else {
2116       set_loop(nnn, outer_loop->_parent);
2117       if (old->outcnt() > 0) {
2118         set_idom( nnn, old_new[idom(old)->_idx], dd );
2119       }
2120     }
2121     // Correct edges to the new node
2122     for( uint j = 0; j < nnn->req(); j++ ) {
2123         Node *n = nnn->in(j);
2124         if( n ) {
2125           IdealLoopTree *old_in_loop = get_loop( has_ctrl(n) ? get_ctrl(n) : n );
2126           if( loop->is_member( old_in_loop ) )
2127             nnn->set_req(j, old_new[n->_idx]);
2128         }
2129     }
2130     _igvn.hash_find_insert(nnn);
2131   }
2132 
2133   Node_List extra_data_nodes; // data nodes in the outer strip mined loop
2134   clone_outer_loop(head, mode, loop, outer_loop, dd, old_new, extra_data_nodes);
2135 
2136   // Step 3: Now fix control uses.  Loop varying control uses have already
2137   // been fixed up (as part of all input edges in Step 2).  Loop invariant
2138   // control uses must be either an IfFalse or an IfTrue.  Make a merge
2139   // point to merge the old and new IfFalse/IfTrue nodes; make the use
2140   // refer to this.
2141   Node_List worklist;
2142   uint new_counter = C->unique();
2143   for( i = 0; i < loop->_body.size(); i++ ) {
2144     Node* old = loop->_body.at(i);
2145     if( !old->is_CFG() ) continue;
2146 
2147     // Copy uses to a worklist, so I can munge the def-use info
2148     // with impunity.
2149     for (DUIterator_Fast jmax, j = old->fast_outs(jmax); j < jmax; j++)
2150       worklist.push(old->fast_out(j));
2151 
2152     while( worklist.size() ) {  // Visit all uses
2153       Node *use = worklist.pop();
2154       if (!has_node(use))  continue; // Ignore dead nodes
2155       IdealLoopTree *use_loop = get_loop( has_ctrl(use) ? get_ctrl(use) : use );
2156       if( !loop->is_member( use_loop ) && use->is_CFG() ) {
2157         // Both OLD and USE are CFG nodes here.
2158         assert( use->is_Proj(), "" );
2159         Node* nnn = old_new[old->_idx];
2160 
2161         Node* newuse = NULL;
2162         if (head->is_strip_mined() && mode != IgnoreStripMined) {
2163           CountedLoopNode* cl = head->as_CountedLoop();
2164           CountedLoopEndNode* cle = cl->loopexit();
2165           Node* cle_out = cle->proj_out_or_null(false);
2166           if (use == cle_out) {
2167             IfNode* le = cl->outer_loop_end();
2168             use = le->proj_out(false);
2169             use_loop = get_loop(use);
2170             if (mode == CloneIncludesStripMined) {
2171               nnn = old_new[le->_idx];
2172             } else {
2173               newuse = old_new[cle_out->_idx];
2174             }
2175           }
2176         }
2177         if (newuse == NULL) {
2178           newuse = use->clone();
2179         }
2180 
2181         // Clone the loop exit control projection
2182         if (C->do_vector_loop()) {
2183           cm.verify_insert_and_clone(use, newuse, cm.clone_idx());
2184         }
2185         newuse->set_req(0,nnn);
2186         _igvn.register_new_node_with_optimizer(newuse);
2187         set_loop(newuse, use_loop);
2188         set_idom(newuse, nnn, dom_depth(nnn) + 1 );
2189 
2190         // We need a Region to merge the exit from the peeled body and the
2191         // exit from the old loop body.
2192         RegionNode *r = new RegionNode(3);
2193         // Map the old use to the new merge point
2194         old_new.map( use->_idx, r );
2195         uint dd_r = MIN2(dom_depth(newuse),dom_depth(use));
2196         assert( dd_r >= dom_depth(dom_lca(newuse,use)), "" );
2197 
2198         // The original user of 'use' uses 'r' instead.
2199         for (DUIterator_Last lmin, l = use->last_outs(lmin); l >= lmin;) {
2200           Node* useuse = use->last_out(l);
2201           _igvn.rehash_node_delayed(useuse);
2202           uint uses_found = 0;
2203           if( useuse->in(0) == use ) {
2204             useuse->set_req(0, r);
2205             uses_found++;
2206             if( useuse->is_CFG() ) {
2207               // This is not a dom_depth > dd_r because when new
2208               // control flow is constructed by a loop opt, a node and
2209               // its dominator can end up at the same dom_depth
2210               assert(dom_depth(useuse) >= dd_r, "");
2211               set_idom(useuse, r, dom_depth(useuse));
2212             }
2213           }
2214           for( uint k = 1; k < useuse->req(); k++ ) {
2215             if( useuse->in(k) == use ) {
2216               useuse->set_req(k, r);
2217               uses_found++;
2218               if (useuse->is_Loop() && k == LoopNode::EntryControl) {
2219                 // This is not a dom_depth > dd_r because when new
2220                 // control flow is constructed by a loop opt, a node
2221                 // and its dominator can end up at the same dom_depth
2222                 assert(dom_depth(useuse) >= dd_r , "");
2223                 set_idom(useuse, r, dom_depth(useuse));
2224               }
2225             }
2226           }
2227           l -= uses_found;    // we deleted 1 or more copies of this edge
2228         }
2229 
2230         // Now finish up 'r'
2231         r->set_req( 1, newuse );
2232         r->set_req( 2,    use );
2233         _igvn.register_new_node_with_optimizer(r);
2234         set_loop(r, use_loop);
2235         set_idom(r, !side_by_side_idom ? newuse->in(0) : side_by_side_idom, dd_r);
2236       } // End of if a loop-exit test
2237     }
2238   }
2239 
2240   // Step 4: If loop-invariant use is not control, it must be dominated by a
2241   // loop exit IfFalse/IfTrue.  Find "proper" loop exit.  Make a Region
2242   // there if needed.  Make a Phi there merging old and new used values.
2243   Node_List *split_if_set = NULL;
2244   Node_List *split_bool_set = NULL;
2245   Node_List *split_cex_set = NULL;
2246   for( i = 0; i < loop->_body.size(); i++ ) {
2247     Node* old = loop->_body.at(i);
2248     clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2249                                 split_bool_set, split_cex_set, worklist, new_counter,
2250                                 mode);
2251   }
2252 
2253   for (i = 0; i < extra_data_nodes.size(); i++) {
2254     Node* old = extra_data_nodes.at(i);
2255     clone_loop_handle_data_uses(old, old_new, loop, outer_loop, split_if_set,
2256                                 split_bool_set, split_cex_set, worklist, new_counter,
2257                                 mode);
2258   }
2259 
2260   // Check for IFs that need splitting/cloning.  Happens if an IF outside of
2261   // the loop uses a condition set in the loop.  The original IF probably
2262   // takes control from one or more OLD Regions (which in turn get from NEW
2263   // Regions).  In any case, there will be a set of Phis for each merge point
2264   // from the IF up to where the original BOOL def exists the loop.
2265   if (split_if_set) {
2266     while (split_if_set->size()) {
2267       Node *iff = split_if_set->pop();
2268       if (iff->in(1)->is_Phi()) {
2269         Node *b = clone_iff(iff->in(1)->as_Phi(), loop);
2270         _igvn.replace_input_of(iff, 1, b);
2271       }
2272     }
2273   }
2274   if (split_bool_set) {
2275     while (split_bool_set->size()) {
2276       Node *b = split_bool_set->pop();
2277       Node *phi = b->in(1);
2278       assert(phi->is_Phi(), "");
2279       CmpNode *cmp = clone_bool((PhiNode*)phi, loop);
2280       _igvn.replace_input_of(b, 1, cmp);
2281     }
2282   }
2283   if (split_cex_set) {
2284     while (split_cex_set->size()) {
2285       Node *b = split_cex_set->pop();
2286       assert(b->in(0)->is_Region(), "");
2287       assert(b->in(1)->is_Phi(), "");
2288       assert(b->in(0)->in(0) == b->in(1)->in(0), "");
2289       split_up(b, b->in(0), NULL);
2290     }
2291   }
2292 
2293 }
2294 
2295 
2296 //---------------------- stride_of_possible_iv -------------------------------------
2297 // Looks for an iff/bool/comp with one operand of the compare
2298 // being a cycle involving an add and a phi,
2299 // with an optional truncation (left-shift followed by a right-shift)
2300 // of the add. Returns zero if not an iv.
2301 int PhaseIdealLoop::stride_of_possible_iv(Node* iff) {
2302   Node* trunc1 = NULL;
2303   Node* trunc2 = NULL;
2304   const TypeInt* ttype = NULL;
2305   if (!iff->is_If() || iff->in(1) == NULL || !iff->in(1)->is_Bool()) {
2306     return 0;
2307   }
2308   BoolNode* bl = iff->in(1)->as_Bool();
2309   Node* cmp = bl->in(1);
2310   if (!cmp || (cmp->Opcode() != Op_CmpI && cmp->Opcode() != Op_CmpU)) {
2311     return 0;
2312   }
2313   // Must have an invariant operand
2314   if (is_member(get_loop(iff), get_ctrl(cmp->in(2)))) {
2315     return 0;
2316   }
2317   Node* add2 = NULL;
2318   Node* cmp1 = cmp->in(1);
2319   if (cmp1->is_Phi()) {
2320     // (If (Bool (CmpX phi:(Phi ...(Optional-trunc(AddI phi add2))) )))
2321     Node* phi = cmp1;
2322     for (uint i = 1; i < phi->req(); i++) {
2323       Node* in = phi->in(i);
2324       Node* add = CountedLoopNode::match_incr_with_optional_truncation(in,
2325                                 &trunc1, &trunc2, &ttype);
2326       if (add && add->in(1) == phi) {
2327         add2 = add->in(2);
2328         break;
2329       }
2330     }
2331   } else {
2332     // (If (Bool (CmpX addtrunc:(Optional-trunc((AddI (Phi ...addtrunc...) add2)) )))
2333     Node* addtrunc = cmp1;
2334     Node* add = CountedLoopNode::match_incr_with_optional_truncation(addtrunc,
2335                                 &trunc1, &trunc2, &ttype);
2336     if (add && add->in(1)->is_Phi()) {
2337       Node* phi = add->in(1);
2338       for (uint i = 1; i < phi->req(); i++) {
2339         if (phi->in(i) == addtrunc) {
2340           add2 = add->in(2);
2341           break;
2342         }
2343       }
2344     }
2345   }
2346   if (add2 != NULL) {
2347     const TypeInt* add2t = _igvn.type(add2)->is_int();
2348     if (add2t->is_con()) {
2349       return add2t->get_con();
2350     }
2351   }
2352   return 0;
2353 }
2354 
2355 
2356 //---------------------- stay_in_loop -------------------------------------
2357 // Return the (unique) control output node that's in the loop (if it exists.)
2358 Node* PhaseIdealLoop::stay_in_loop( Node* n, IdealLoopTree *loop) {
2359   Node* unique = NULL;
2360   if (!n) return NULL;
2361   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2362     Node* use = n->fast_out(i);
2363     if (!has_ctrl(use) && loop->is_member(get_loop(use))) {
2364       if (unique != NULL) {
2365         return NULL;
2366       }
2367       unique = use;
2368     }
2369   }
2370   return unique;
2371 }
2372 
2373 //------------------------------ register_node -------------------------------------
2374 // Utility to register node "n" with PhaseIdealLoop
2375 void PhaseIdealLoop::register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth) {
2376   _igvn.register_new_node_with_optimizer(n);
2377   loop->_body.push(n);
2378   if (n->is_CFG()) {
2379     set_loop(n, loop);
2380     set_idom(n, pred, ddepth);
2381   } else {
2382     set_ctrl(n, pred);
2383   }
2384 }
2385 
2386 //------------------------------ proj_clone -------------------------------------
2387 // Utility to create an if-projection
2388 ProjNode* PhaseIdealLoop::proj_clone(ProjNode* p, IfNode* iff) {
2389   ProjNode* c = p->clone()->as_Proj();
2390   c->set_req(0, iff);
2391   return c;
2392 }
2393 
2394 //------------------------------ short_circuit_if -------------------------------------
2395 // Force the iff control output to be the live_proj
2396 Node* PhaseIdealLoop::short_circuit_if(IfNode* iff, ProjNode* live_proj) {
2397   guarantee(live_proj != NULL, "null projection");
2398   int proj_con = live_proj->_con;
2399   assert(proj_con == 0 || proj_con == 1, "false or true projection");
2400   Node *con = _igvn.intcon(proj_con);
2401   set_ctrl(con, C->root());
2402   if (iff) {
2403     iff->set_req(1, con);
2404   }
2405   return con;
2406 }
2407 
2408 //------------------------------ insert_if_before_proj -------------------------------------
2409 // Insert a new if before an if projection (* - new node)
2410 //
2411 // before
2412 //           if(test)
2413 //           /     \
2414 //          v       v
2415 //    other-proj   proj (arg)
2416 //
2417 // after
2418 //           if(test)
2419 //           /     \
2420 //          /       v
2421 //         |      * proj-clone
2422 //         v          |
2423 //    other-proj      v
2424 //                * new_if(relop(cmp[IU](left,right)))
2425 //                  /  \
2426 //                 v    v
2427 //         * new-proj  proj
2428 //         (returned)
2429 //
2430 ProjNode* PhaseIdealLoop::insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj) {
2431   IfNode* iff = proj->in(0)->as_If();
2432   IdealLoopTree *loop = get_loop(proj);
2433   ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2434   int ddepth = dom_depth(proj);
2435 
2436   _igvn.rehash_node_delayed(iff);
2437   _igvn.rehash_node_delayed(proj);
2438 
2439   proj->set_req(0, NULL);  // temporary disconnect
2440   ProjNode* proj2 = proj_clone(proj, iff);
2441   register_node(proj2, loop, iff, ddepth);
2442 
2443   Node* cmp = Signed ? (Node*) new CmpINode(left, right) : (Node*) new CmpUNode(left, right);
2444   register_node(cmp, loop, proj2, ddepth);
2445 
2446   BoolNode* bol = new BoolNode(cmp, relop);
2447   register_node(bol, loop, proj2, ddepth);
2448 
2449   int opcode = iff->Opcode();
2450   assert(opcode == Op_If || opcode == Op_RangeCheck, "unexpected opcode");
2451   IfNode* new_if = (opcode == Op_If) ? new IfNode(proj2, bol, iff->_prob, iff->_fcnt):
2452     new RangeCheckNode(proj2, bol, iff->_prob, iff->_fcnt);
2453   register_node(new_if, loop, proj2, ddepth);
2454 
2455   proj->set_req(0, new_if); // reattach
2456   set_idom(proj, new_if, ddepth);
2457 
2458   ProjNode* new_exit = proj_clone(other_proj, new_if)->as_Proj();
2459   guarantee(new_exit != NULL, "null exit node");
2460   register_node(new_exit, get_loop(other_proj), new_if, ddepth);
2461 
2462   return new_exit;
2463 }
2464 
2465 //------------------------------ insert_region_before_proj -------------------------------------
2466 // Insert a region before an if projection (* - new node)
2467 //
2468 // before
2469 //           if(test)
2470 //          /      |
2471 //         v       |
2472 //       proj      v
2473 //               other-proj
2474 //
2475 // after
2476 //           if(test)
2477 //          /      |
2478 //         v       |
2479 // * proj-clone    v
2480 //         |     other-proj
2481 //         v
2482 // * new-region
2483 //         |
2484 //         v
2485 // *      dum_if
2486 //       /     \
2487 //      v       \
2488 // * dum-proj    v
2489 //              proj
2490 //
2491 RegionNode* PhaseIdealLoop::insert_region_before_proj(ProjNode* proj) {
2492   IfNode* iff = proj->in(0)->as_If();
2493   IdealLoopTree *loop = get_loop(proj);
2494   ProjNode *other_proj = iff->proj_out(!proj->is_IfTrue())->as_Proj();
2495   int ddepth = dom_depth(proj);
2496 
2497   _igvn.rehash_node_delayed(iff);
2498   _igvn.rehash_node_delayed(proj);
2499 
2500   proj->set_req(0, NULL);  // temporary disconnect
2501   ProjNode* proj2 = proj_clone(proj, iff);
2502   register_node(proj2, loop, iff, ddepth);
2503 
2504   RegionNode* reg = new RegionNode(2);
2505   reg->set_req(1, proj2);
2506   register_node(reg, loop, iff, ddepth);
2507 
2508   IfNode* dum_if = new IfNode(reg, short_circuit_if(NULL, proj), iff->_prob, iff->_fcnt);
2509   register_node(dum_if, loop, reg, ddepth);
2510 
2511   proj->set_req(0, dum_if); // reattach
2512   set_idom(proj, dum_if, ddepth);
2513 
2514   ProjNode* dum_proj = proj_clone(other_proj, dum_if);
2515   register_node(dum_proj, loop, dum_if, ddepth);
2516 
2517   return reg;
2518 }
2519 
2520 //------------------------------ insert_cmpi_loop_exit -------------------------------------
2521 // Clone a signed compare loop exit from an unsigned compare and
2522 // insert it before the unsigned cmp on the stay-in-loop path.
2523 // All new nodes inserted in the dominator tree between the original
2524 // if and it's projections.  The original if test is replaced with
2525 // a constant to force the stay-in-loop path.
2526 //
2527 // This is done to make sure that the original if and it's projections
2528 // still dominate the same set of control nodes, that the ctrl() relation
2529 // from data nodes to them is preserved, and that their loop nesting is
2530 // preserved.
2531 //
2532 // before
2533 //          if(i <u limit)    unsigned compare loop exit
2534 //         /       |
2535 //        v        v
2536 //   exit-proj   stay-in-loop-proj
2537 //
2538 // after
2539 //          if(stay-in-loop-const)  original if
2540 //         /       |
2541 //        /        v
2542 //       /  if(i <  limit)    new signed test
2543 //      /  /       |
2544 //     /  /        v
2545 //    /  /  if(i <u limit)    new cloned unsigned test
2546 //   /  /   /      |
2547 //   v  v  v       |
2548 //    region       |
2549 //        |        |
2550 //      dum-if     |
2551 //     /  |        |
2552 // ether  |        |
2553 //        v        v
2554 //   exit-proj   stay-in-loop-proj
2555 //
2556 IfNode* PhaseIdealLoop::insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop) {
2557   const bool Signed   = true;
2558   const bool Unsigned = false;
2559 
2560   BoolNode* bol = if_cmpu->in(1)->as_Bool();
2561   if (bol->_test._test != BoolTest::lt) return NULL;
2562   CmpNode* cmpu = bol->in(1)->as_Cmp();
2563   if (cmpu->Opcode() != Op_CmpU) return NULL;
2564   int stride = stride_of_possible_iv(if_cmpu);
2565   if (stride == 0) return NULL;
2566 
2567   Node* lp_proj = stay_in_loop(if_cmpu, loop);
2568   guarantee(lp_proj != NULL, "null loop node");
2569 
2570   ProjNode* lp_continue = lp_proj->as_Proj();
2571   ProjNode* lp_exit     = if_cmpu->proj_out(!lp_continue->is_IfTrue())->as_Proj();
2572 
2573   Node* limit = NULL;
2574   if (stride > 0) {
2575     limit = cmpu->in(2);
2576   } else {
2577     limit = _igvn.makecon(TypeInt::ZERO);
2578     set_ctrl(limit, C->root());
2579   }
2580   // Create a new region on the exit path
2581   RegionNode* reg = insert_region_before_proj(lp_exit);
2582   guarantee(reg != NULL, "null region node");
2583 
2584   // Clone the if-cmpu-true-false using a signed compare
2585   BoolTest::mask rel_i = stride > 0 ? bol->_test._test : BoolTest::ge;
2586   ProjNode* cmpi_exit = insert_if_before_proj(cmpu->in(1), Signed, rel_i, limit, lp_continue);
2587   reg->add_req(cmpi_exit);
2588 
2589   // Clone the if-cmpu-true-false
2590   BoolTest::mask rel_u = bol->_test._test;
2591   ProjNode* cmpu_exit = insert_if_before_proj(cmpu->in(1), Unsigned, rel_u, cmpu->in(2), lp_continue);
2592   reg->add_req(cmpu_exit);
2593 
2594   // Force original if to stay in loop.
2595   short_circuit_if(if_cmpu, lp_continue);
2596 
2597   return cmpi_exit->in(0)->as_If();
2598 }
2599 
2600 //------------------------------ remove_cmpi_loop_exit -------------------------------------
2601 // Remove a previously inserted signed compare loop exit.
2602 void PhaseIdealLoop::remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop) {
2603   Node* lp_proj = stay_in_loop(if_cmp, loop);
2604   assert(if_cmp->in(1)->in(1)->Opcode() == Op_CmpI &&
2605          stay_in_loop(lp_proj, loop)->is_If() &&
2606          stay_in_loop(lp_proj, loop)->in(1)->in(1)->Opcode() == Op_CmpU, "inserted cmpi before cmpu");
2607   Node *con = _igvn.makecon(lp_proj->is_IfTrue() ? TypeInt::ONE : TypeInt::ZERO);
2608   set_ctrl(con, C->root());
2609   if_cmp->set_req(1, con);
2610 }
2611 
2612 //------------------------------ scheduled_nodelist -------------------------------------
2613 // Create a post order schedule of nodes that are in the
2614 // "member" set.  The list is returned in "sched".
2615 // The first node in "sched" is the loop head, followed by
2616 // nodes which have no inputs in the "member" set, and then
2617 // followed by the nodes that have an immediate input dependence
2618 // on a node in "sched".
2619 void PhaseIdealLoop::scheduled_nodelist( IdealLoopTree *loop, VectorSet& member, Node_List &sched ) {
2620 
2621   assert(member.test(loop->_head->_idx), "loop head must be in member set");
2622   VectorSet visited;
2623   Node_Stack nstack(loop->_body.size());
2624 
2625   Node* n  = loop->_head;  // top of stack is cached in "n"
2626   uint idx = 0;
2627   visited.set(n->_idx);
2628 
2629   // Initially push all with no inputs from within member set
2630   for(uint i = 0; i < loop->_body.size(); i++ ) {
2631     Node *elt = loop->_body.at(i);
2632     if (member.test(elt->_idx)) {
2633       bool found = false;
2634       for (uint j = 0; j < elt->req(); j++) {
2635         Node* def = elt->in(j);
2636         if (def && member.test(def->_idx) && def != elt) {
2637           found = true;
2638           break;
2639         }
2640       }
2641       if (!found && elt != loop->_head) {
2642         nstack.push(n, idx);
2643         n = elt;
2644         assert(!visited.test(n->_idx), "not seen yet");
2645         visited.set(n->_idx);
2646       }
2647     }
2648   }
2649 
2650   // traverse out's that are in the member set
2651   while (true) {
2652     if (idx < n->outcnt()) {
2653       Node* use = n->raw_out(idx);
2654       idx++;
2655       if (!visited.test_set(use->_idx)) {
2656         if (member.test(use->_idx)) {
2657           nstack.push(n, idx);
2658           n = use;
2659           idx = 0;
2660         }
2661       }
2662     } else {
2663       // All outputs processed
2664       sched.push(n);
2665       if (nstack.is_empty()) break;
2666       n   = nstack.node();
2667       idx = nstack.index();
2668       nstack.pop();
2669     }
2670   }
2671 }
2672 
2673 
2674 //------------------------------ has_use_in_set -------------------------------------
2675 // Has a use in the vector set
2676 bool PhaseIdealLoop::has_use_in_set( Node* n, VectorSet& vset ) {
2677   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2678     Node* use = n->fast_out(j);
2679     if (vset.test(use->_idx)) {
2680       return true;
2681     }
2682   }
2683   return false;
2684 }
2685 
2686 
2687 //------------------------------ has_use_internal_to_set -------------------------------------
2688 // Has use internal to the vector set (ie. not in a phi at the loop head)
2689 bool PhaseIdealLoop::has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ) {
2690   Node* head  = loop->_head;
2691   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2692     Node* use = n->fast_out(j);
2693     if (vset.test(use->_idx) && !(use->is_Phi() && use->in(0) == head)) {
2694       return true;
2695     }
2696   }
2697   return false;
2698 }
2699 
2700 
2701 //------------------------------ clone_for_use_outside_loop -------------------------------------
2702 // clone "n" for uses that are outside of loop
2703 int PhaseIdealLoop::clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ) {
2704   int cloned = 0;
2705   assert(worklist.size() == 0, "should be empty");
2706   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2707     Node* use = n->fast_out(j);
2708     if( !loop->is_member(get_loop(has_ctrl(use) ? get_ctrl(use) : use)) ) {
2709       worklist.push(use);
2710     }
2711   }
2712   while( worklist.size() ) {
2713     Node *use = worklist.pop();
2714     if (!has_node(use) || use->in(0) == C->top()) continue;
2715     uint j;
2716     for (j = 0; j < use->req(); j++) {
2717       if (use->in(j) == n) break;
2718     }
2719     assert(j < use->req(), "must be there");
2720 
2721     // clone "n" and insert it between the inputs of "n" and the use outside the loop
2722     Node* n_clone = n->clone();
2723     _igvn.replace_input_of(use, j, n_clone);
2724     cloned++;
2725     Node* use_c;
2726     if (!use->is_Phi()) {
2727       use_c = has_ctrl(use) ? get_ctrl(use) : use->in(0);
2728     } else {
2729       // Use in a phi is considered a use in the associated predecessor block
2730       use_c = use->in(0)->in(j);
2731     }
2732     set_ctrl(n_clone, use_c);
2733     assert(!loop->is_member(get_loop(use_c)), "should be outside loop");
2734     get_loop(use_c)->_body.push(n_clone);
2735     _igvn.register_new_node_with_optimizer(n_clone);
2736 #ifndef PRODUCT
2737     if (TracePartialPeeling) {
2738       tty->print_cr("loop exit cloning old: %d new: %d newbb: %d", n->_idx, n_clone->_idx, get_ctrl(n_clone)->_idx);
2739     }
2740 #endif
2741   }
2742   return cloned;
2743 }
2744 
2745 
2746 //------------------------------ clone_for_special_use_inside_loop -------------------------------------
2747 // clone "n" for special uses that are in the not_peeled region.
2748 // If these def-uses occur in separate blocks, the code generator
2749 // marks the method as not compilable.  For example, if a "BoolNode"
2750 // is in a different basic block than the "IfNode" that uses it, then
2751 // the compilation is aborted in the code generator.
2752 void PhaseIdealLoop::clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
2753                                                         VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ) {
2754   if (n->is_Phi() || n->is_Load()) {
2755     return;
2756   }
2757   assert(worklist.size() == 0, "should be empty");
2758   for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2759     Node* use = n->fast_out(j);
2760     if ( not_peel.test(use->_idx) &&
2761          (use->is_If() || use->is_CMove() || use->is_Bool()) &&
2762          use->in(1) == n)  {
2763       worklist.push(use);
2764     }
2765   }
2766   if (worklist.size() > 0) {
2767     // clone "n" and insert it between inputs of "n" and the use
2768     Node* n_clone = n->clone();
2769     loop->_body.push(n_clone);
2770     _igvn.register_new_node_with_optimizer(n_clone);
2771     set_ctrl(n_clone, get_ctrl(n));
2772     sink_list.push(n_clone);
2773     not_peel.set(n_clone->_idx);
2774 #ifndef PRODUCT
2775     if (TracePartialPeeling) {
2776       tty->print_cr("special not_peeled cloning old: %d new: %d", n->_idx, n_clone->_idx);
2777     }
2778 #endif
2779     while( worklist.size() ) {
2780       Node *use = worklist.pop();
2781       _igvn.rehash_node_delayed(use);
2782       for (uint j = 1; j < use->req(); j++) {
2783         if (use->in(j) == n) {
2784           use->set_req(j, n_clone);
2785         }
2786       }
2787     }
2788   }
2789 }
2790 
2791 
2792 //------------------------------ insert_phi_for_loop -------------------------------------
2793 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
2794 void PhaseIdealLoop::insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ) {
2795   Node *phi = PhiNode::make(lp, back_edge_val);
2796   phi->set_req(LoopNode::EntryControl, lp_entry_val);
2797   // Use existing phi if it already exists
2798   Node *hit = _igvn.hash_find_insert(phi);
2799   if( hit == NULL ) {
2800     _igvn.register_new_node_with_optimizer(phi);
2801     set_ctrl(phi, lp);
2802   } else {
2803     // Remove the new phi from the graph and use the hit
2804     _igvn.remove_dead_node(phi);
2805     phi = hit;
2806   }
2807   _igvn.replace_input_of(use, idx, phi);
2808 }
2809 
2810 #ifdef ASSERT
2811 //------------------------------ is_valid_loop_partition -------------------------------------
2812 // Validate the loop partition sets: peel and not_peel
2813 bool PhaseIdealLoop::is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list,
2814                                               VectorSet& not_peel ) {
2815   uint i;
2816   // Check that peel_list entries are in the peel set
2817   for (i = 0; i < peel_list.size(); i++) {
2818     if (!peel.test(peel_list.at(i)->_idx)) {
2819       return false;
2820     }
2821   }
2822   // Check at loop members are in one of peel set or not_peel set
2823   for (i = 0; i < loop->_body.size(); i++ ) {
2824     Node *def  = loop->_body.at(i);
2825     uint di = def->_idx;
2826     // Check that peel set elements are in peel_list
2827     if (peel.test(di)) {
2828       if (not_peel.test(di)) {
2829         return false;
2830       }
2831       // Must be in peel_list also
2832       bool found = false;
2833       for (uint j = 0; j < peel_list.size(); j++) {
2834         if (peel_list.at(j)->_idx == di) {
2835           found = true;
2836           break;
2837         }
2838       }
2839       if (!found) {
2840         return false;
2841       }
2842     } else if (not_peel.test(di)) {
2843       if (peel.test(di)) {
2844         return false;
2845       }
2846     } else {
2847       return false;
2848     }
2849   }
2850   return true;
2851 }
2852 
2853 //------------------------------ is_valid_clone_loop_exit_use -------------------------------------
2854 // Ensure a use outside of loop is of the right form
2855 bool PhaseIdealLoop::is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx) {
2856   Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2857   return (use->is_Phi() &&
2858           use_c->is_Region() && use_c->req() == 3 &&
2859           (use_c->in(exit_idx)->Opcode() == Op_IfTrue ||
2860            use_c->in(exit_idx)->Opcode() == Op_IfFalse ||
2861            use_c->in(exit_idx)->Opcode() == Op_JumpProj) &&
2862           loop->is_member( get_loop( use_c->in(exit_idx)->in(0) ) ) );
2863 }
2864 
2865 //------------------------------ is_valid_clone_loop_form -------------------------------------
2866 // Ensure that all uses outside of loop are of the right form
2867 bool PhaseIdealLoop::is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
2868                                                uint orig_exit_idx, uint clone_exit_idx) {
2869   uint len = peel_list.size();
2870   for (uint i = 0; i < len; i++) {
2871     Node *def = peel_list.at(i);
2872 
2873     for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
2874       Node *use = def->fast_out(j);
2875       Node *use_c = has_ctrl(use) ? get_ctrl(use) : use;
2876       if (!loop->is_member(get_loop(use_c))) {
2877         // use is not in the loop, check for correct structure
2878         if (use->in(0) == def) {
2879           // Okay
2880         } else if (!is_valid_clone_loop_exit_use(loop, use, orig_exit_idx)) {
2881           return false;
2882         }
2883       }
2884     }
2885   }
2886   return true;
2887 }
2888 #endif
2889 
2890 //------------------------------ partial_peel -------------------------------------
2891 // Partially peel (aka loop rotation) the top portion of a loop (called
2892 // the peel section below) by cloning it and placing one copy just before
2893 // the new loop head and the other copy at the bottom of the new loop.
2894 //
2895 //    before                       after                where it came from
2896 //
2897 //    stmt1                        stmt1
2898 //  loop:                          stmt2                     clone
2899 //    stmt2                        if condA goto exitA       clone
2900 //    if condA goto exitA        new_loop:                   new
2901 //    stmt3                        stmt3                     clone
2902 //    if !condB goto loop          if condB goto exitB       clone
2903 //  exitB:                         stmt2                     orig
2904 //    stmt4                        if !condA goto new_loop   orig
2905 //  exitA:                         goto exitA
2906 //                               exitB:
2907 //                                 stmt4
2908 //                               exitA:
2909 //
2910 // Step 1: find the cut point: an exit test on probable
2911 //         induction variable.
2912 // Step 2: schedule (with cloning) operations in the peel
2913 //         section that can be executed after the cut into
2914 //         the section that is not peeled.  This may need
2915 //         to clone operations into exit blocks.  For
2916 //         instance, a reference to A[i] in the not-peel
2917 //         section and a reference to B[i] in an exit block
2918 //         may cause a left-shift of i by 2 to be placed
2919 //         in the peel block.  This step will clone the left
2920 //         shift into the exit block and sink the left shift
2921 //         from the peel to the not-peel section.
2922 // Step 3: clone the loop, retarget the control, and insert
2923 //         phis for values that are live across the new loop
2924 //         head.  This is very dependent on the graph structure
2925 //         from clone_loop.  It creates region nodes for
2926 //         exit control and associated phi nodes for values
2927 //         flow out of the loop through that exit.  The region
2928 //         node is dominated by the clone's control projection.
2929 //         So the clone's peel section is placed before the
2930 //         new loop head, and the clone's not-peel section is
2931 //         forms the top part of the new loop.  The original
2932 //         peel section forms the tail of the new loop.
2933 // Step 4: update the dominator tree and recompute the
2934 //         dominator depth.
2935 //
2936 //                   orig
2937 //
2938 //                   stmt1
2939 //                     |
2940 //                     v
2941 //               loop predicate
2942 //                     |
2943 //                     v
2944 //                   loop<----+
2945 //                     |      |
2946 //                   stmt2    |
2947 //                     |      |
2948 //                     v      |
2949 //                    ifA     |
2950 //                   / |      |
2951 //                  v  v      |
2952 //               false true   ^  <-- last_peel
2953 //               /     |      |
2954 //              /   ===|==cut |
2955 //             /     stmt3    |  <-- first_not_peel
2956 //            /        |      |
2957 //            |        v      |
2958 //            v       ifB     |
2959 //          exitA:   / \      |
2960 //                  /   \     |
2961 //                 v     v    |
2962 //               false true   |
2963 //               /       \    |
2964 //              /         ----+
2965 //             |
2966 //             v
2967 //           exitB:
2968 //           stmt4
2969 //
2970 //
2971 //            after clone loop
2972 //
2973 //                   stmt1
2974 //                     |
2975 //                     v
2976 //               loop predicate
2977 //                 /       \
2978 //        clone   /         \   orig
2979 //               /           \
2980 //              /             \
2981 //             v               v
2982 //   +---->loop                loop<----+
2983 //   |      |                    |      |
2984 //   |    stmt2                stmt2    |
2985 //   |      |                    |      |
2986 //   |      v                    v      |
2987 //   |      ifA                 ifA     |
2988 //   |      | \                / |      |
2989 //   |      v  v              v  v      |
2990 //   ^    true  false      false true   ^  <-- last_peel
2991 //   |      |   ^   \       /    |      |
2992 //   | cut==|==  \   \     /  ===|==cut |
2993 //   |    stmt3   \   \   /    stmt3    |  <-- first_not_peel
2994 //   |      |    dom   | |       |      |
2995 //   |      v      \  1v v2      v      |
2996 //   |      ifB     regionA     ifB     |
2997 //   |      / \        |       / \      |
2998 //   |     /   \       v      /   \     |
2999 //   |    v     v    exitA:  v     v    |
3000 //   |    true  false      false true   |
3001 //   |    /     ^   \      /       \    |
3002 //   +----       \   \    /         ----+
3003 //               dom  \  /
3004 //                 \  1v v2
3005 //                  regionB
3006 //                     |
3007 //                     v
3008 //                   exitB:
3009 //                   stmt4
3010 //
3011 //
3012 //           after partial peel
3013 //
3014 //                  stmt1
3015 //                     |
3016 //                     v
3017 //               loop predicate
3018 //                 /
3019 //        clone   /             orig
3020 //               /          TOP
3021 //              /             \
3022 //             v               v
3023 //    TOP->loop                loop----+
3024 //          |                    |      |
3025 //        stmt2                stmt2    |
3026 //          |                    |      |
3027 //          v                    v      |
3028 //          ifA                 ifA     |
3029 //          | \                / |      |
3030 //          v  v              v  v      |
3031 //        true  false      false true   |     <-- last_peel
3032 //          |   ^   \       /    +------|---+
3033 //  +->newloop   \   \     /  === ==cut |   |
3034 //  |     stmt3   \   \   /     TOP     |   |
3035 //  |       |    dom   | |      stmt3   |   | <-- first_not_peel
3036 //  |       v      \  1v v2      v      |   |
3037 //  |       ifB     regionA     ifB     ^   v
3038 //  |       / \        |       / \      |   |
3039 //  |      /   \       v      /   \     |   |
3040 //  |     v     v    exitA:  v     v    |   |
3041 //  |     true  false      false true   |   |
3042 //  |     /     ^   \      /       \    |   |
3043 //  |    |       \   \    /         v   |   |
3044 //  |    |       dom  \  /         TOP  |   |
3045 //  |    |         \  1v v2             |   |
3046 //  ^    v          regionB             |   |
3047 //  |    |             |                |   |
3048 //  |    |             v                ^   v
3049 //  |    |           exitB:             |   |
3050 //  |    |           stmt4              |   |
3051 //  |    +------------>-----------------+   |
3052 //  |                                       |
3053 //  +-----------------<---------------------+
3054 //
3055 //
3056 //              final graph
3057 //
3058 //                  stmt1
3059 //                    |
3060 //                    v
3061 //               loop predicate
3062 //                    |
3063 //                    v
3064 //                  stmt2 clone
3065 //                    |
3066 //                    v
3067 //         ........> ifA clone
3068 //         :        / |
3069 //        dom      /  |
3070 //         :      v   v
3071 //         :  false   true
3072 //         :  |       |
3073 //         :  |       v
3074 //         :  |    newloop<-----+
3075 //         :  |        |        |
3076 //         :  |     stmt3 clone |
3077 //         :  |        |        |
3078 //         :  |        v        |
3079 //         :  |       ifB       |
3080 //         :  |      / \        |
3081 //         :  |     v   v       |
3082 //         :  |  false true     |
3083 //         :  |   |     |       |
3084 //         :  |   v    stmt2    |
3085 //         :  | exitB:  |       |
3086 //         :  | stmt4   v       |
3087 //         :  |       ifA orig  |
3088 //         :  |      /  \       |
3089 //         :  |     /    \      |
3090 //         :  |    v     v      |
3091 //         :  |  false  true    |
3092 //         :  |  /        \     |
3093 //         :  v  v         -----+
3094 //          RegionA
3095 //             |
3096 //             v
3097 //           exitA
3098 //
3099 bool PhaseIdealLoop::partial_peel( IdealLoopTree *loop, Node_List &old_new ) {
3100 
3101   assert(!loop->_head->is_CountedLoop(), "Non-counted loop only");
3102   if (!loop->_head->is_Loop()) {
3103     return false;
3104   }
3105   LoopNode *head = loop->_head->as_Loop();
3106 
3107   if (head->is_partial_peel_loop() || head->partial_peel_has_failed()) {
3108     return false;
3109   }
3110 
3111   // Check for complex exit control
3112   for (uint ii = 0; ii < loop->_body.size(); ii++) {
3113     Node *n = loop->_body.at(ii);
3114     int opc = n->Opcode();
3115     if (n->is_Call()        ||
3116         opc == Op_Catch     ||
3117         opc == Op_CatchProj ||
3118         opc == Op_Jump      ||
3119         opc == Op_JumpProj) {
3120 #ifndef PRODUCT
3121       if (TracePartialPeeling) {
3122         tty->print_cr("\nExit control too complex: lp: %d", head->_idx);
3123       }
3124 #endif
3125       return false;
3126     }
3127   }
3128 
3129   int dd = dom_depth(head);
3130 
3131   // Step 1: find cut point
3132 
3133   // Walk up dominators to loop head looking for first loop exit
3134   // which is executed on every path thru loop.
3135   IfNode *peel_if = NULL;
3136   IfNode *peel_if_cmpu = NULL;
3137 
3138   Node *iff = loop->tail();
3139   while (iff != head) {
3140     if (iff->is_If()) {
3141       Node *ctrl = get_ctrl(iff->in(1));
3142       if (ctrl->is_top()) return false; // Dead test on live IF.
3143       // If loop-varying exit-test, check for induction variable
3144       if (loop->is_member(get_loop(ctrl)) &&
3145           loop->is_loop_exit(iff) &&
3146           is_possible_iv_test(iff)) {
3147         Node* cmp = iff->in(1)->in(1);
3148         if (cmp->Opcode() == Op_CmpI) {
3149           peel_if = iff->as_If();
3150         } else {
3151           assert(cmp->Opcode() == Op_CmpU, "must be CmpI or CmpU");
3152           peel_if_cmpu = iff->as_If();
3153         }
3154       }
3155     }
3156     iff = idom(iff);
3157   }
3158 
3159   // Prefer signed compare over unsigned compare.
3160   IfNode* new_peel_if = NULL;
3161   if (peel_if == NULL) {
3162     if (!PartialPeelAtUnsignedTests || peel_if_cmpu == NULL) {
3163       return false;   // No peel point found
3164     }
3165     new_peel_if = insert_cmpi_loop_exit(peel_if_cmpu, loop);
3166     if (new_peel_if == NULL) {
3167       return false;   // No peel point found
3168     }
3169     peel_if = new_peel_if;
3170   }
3171   Node* last_peel        = stay_in_loop(peel_if, loop);
3172   Node* first_not_peeled = stay_in_loop(last_peel, loop);
3173   if (first_not_peeled == NULL || first_not_peeled == head) {
3174     return false;
3175   }
3176 
3177 #ifndef PRODUCT
3178   if (TraceLoopOpts) {
3179     tty->print("PartialPeel  ");
3180     loop->dump_head();
3181   }
3182 
3183   if (TracePartialPeeling) {
3184     tty->print_cr("before partial peel one iteration");
3185     Node_List wl;
3186     Node* t = head->in(2);
3187     while (true) {
3188       wl.push(t);
3189       if (t == head) break;
3190       t = idom(t);
3191     }
3192     while (wl.size() > 0) {
3193       Node* tt = wl.pop();
3194       tt->dump();
3195       if (tt == last_peel) tty->print_cr("-- cut --");
3196     }
3197   }
3198 #endif
3199   VectorSet peel;
3200   VectorSet not_peel;
3201   Node_List peel_list;
3202   Node_List worklist;
3203   Node_List sink_list;
3204 
3205   uint estimate = loop->est_loop_clone_sz(1);
3206   if (exceeding_node_budget(estimate)) {
3207     return false;
3208   }
3209 
3210   // Set of cfg nodes to peel are those that are executable from
3211   // the head through last_peel.
3212   assert(worklist.size() == 0, "should be empty");
3213   worklist.push(head);
3214   peel.set(head->_idx);
3215   while (worklist.size() > 0) {
3216     Node *n = worklist.pop();
3217     if (n != last_peel) {
3218       for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
3219         Node* use = n->fast_out(j);
3220         if (use->is_CFG() &&
3221             loop->is_member(get_loop(use)) &&
3222             !peel.test_set(use->_idx)) {
3223           worklist.push(use);
3224         }
3225       }
3226     }
3227   }
3228 
3229   // Set of non-cfg nodes to peel are those that are control
3230   // dependent on the cfg nodes.
3231   for (uint i = 0; i < loop->_body.size(); i++) {
3232     Node *n = loop->_body.at(i);
3233     Node *n_c = has_ctrl(n) ? get_ctrl(n) : n;
3234     if (peel.test(n_c->_idx)) {
3235       peel.set(n->_idx);
3236     } else {
3237       not_peel.set(n->_idx);
3238     }
3239   }
3240 
3241   // Step 2: move operations from the peeled section down into the
3242   //         not-peeled section
3243 
3244   // Get a post order schedule of nodes in the peel region
3245   // Result in right-most operand.
3246   scheduled_nodelist(loop, peel, peel_list);
3247 
3248   assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3249 
3250   // For future check for too many new phis
3251   uint old_phi_cnt = 0;
3252   for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
3253     Node* use = head->fast_out(j);
3254     if (use->is_Phi()) old_phi_cnt++;
3255   }
3256 
3257 #ifndef PRODUCT
3258   if (TracePartialPeeling) {
3259     tty->print_cr("\npeeled list");
3260   }
3261 #endif
3262 
3263   // Evacuate nodes in peel region into the not_peeled region if possible
3264   uint new_phi_cnt = 0;
3265   uint cloned_for_outside_use = 0;
3266   for (uint i = 0; i < peel_list.size();) {
3267     Node* n = peel_list.at(i);
3268 #ifndef PRODUCT
3269   if (TracePartialPeeling) n->dump();
3270 #endif
3271     bool incr = true;
3272     if (!n->is_CFG()) {
3273       if (has_use_in_set(n, not_peel)) {
3274         // If not used internal to the peeled region,
3275         // move "n" from peeled to not_peeled region.
3276         if (!has_use_internal_to_set(n, peel, loop)) {
3277           // if not pinned and not a load (which maybe anti-dependent on a store)
3278           // and not a CMove (Matcher expects only bool->cmove).
3279           if (n->in(0) == NULL && !n->is_Load() && !n->is_CMove()) {
3280             cloned_for_outside_use += clone_for_use_outside_loop(loop, n, worklist);
3281             sink_list.push(n);
3282             peel.remove(n->_idx);
3283             not_peel.set(n->_idx);
3284             peel_list.remove(i);
3285             incr = false;
3286 #ifndef PRODUCT
3287             if (TracePartialPeeling) {
3288               tty->print_cr("sink to not_peeled region: %d newbb: %d",
3289                             n->_idx, get_ctrl(n)->_idx);
3290             }
3291 #endif
3292           }
3293         } else {
3294           // Otherwise check for special def-use cases that span
3295           // the peel/not_peel boundary such as bool->if
3296           clone_for_special_use_inside_loop(loop, n, not_peel, sink_list, worklist);
3297           new_phi_cnt++;
3298         }
3299       }
3300     }
3301     if (incr) i++;
3302   }
3303 
3304   estimate += cloned_for_outside_use + new_phi_cnt;
3305   bool exceed_node_budget = !may_require_nodes(estimate);
3306   bool exceed_phi_limit = new_phi_cnt > old_phi_cnt + PartialPeelNewPhiDelta;
3307 
3308   if (exceed_node_budget || exceed_phi_limit) {
3309 #ifndef PRODUCT
3310     if (TracePartialPeeling) {
3311       tty->print_cr("\nToo many new phis: %d  old %d new cmpi: %c",
3312                     new_phi_cnt, old_phi_cnt, new_peel_if != NULL?'T':'F');
3313     }
3314 #endif
3315     if (new_peel_if != NULL) {
3316       remove_cmpi_loop_exit(new_peel_if, loop);
3317     }
3318     // Inhibit more partial peeling on this loop
3319     assert(!head->is_partial_peel_loop(), "not partial peeled");
3320     head->mark_partial_peel_failed();
3321     if (cloned_for_outside_use > 0) {
3322       // Terminate this round of loop opts because
3323       // the graph outside this loop was changed.
3324       C->set_major_progress();
3325       return true;
3326     }
3327     return false;
3328   }
3329 
3330   // Step 3: clone loop, retarget control, and insert new phis
3331 
3332   // Create new loop head for new phis and to hang
3333   // the nodes being moved (sinked) from the peel region.
3334   LoopNode* new_head = new LoopNode(last_peel, last_peel);
3335   new_head->set_unswitch_count(head->unswitch_count()); // Preserve
3336   _igvn.register_new_node_with_optimizer(new_head);
3337   assert(first_not_peeled->in(0) == last_peel, "last_peel <- first_not_peeled");
3338   _igvn.replace_input_of(first_not_peeled, 0, new_head);
3339   set_loop(new_head, loop);
3340   loop->_body.push(new_head);
3341   not_peel.set(new_head->_idx);
3342   set_idom(new_head, last_peel, dom_depth(first_not_peeled));
3343   set_idom(first_not_peeled, new_head, dom_depth(first_not_peeled));
3344 
3345   while (sink_list.size() > 0) {
3346     Node* n = sink_list.pop();
3347     set_ctrl(n, new_head);
3348   }
3349 
3350   assert(is_valid_loop_partition(loop, peel, peel_list, not_peel), "bad partition");
3351 
3352   clone_loop(loop, old_new, dd, IgnoreStripMined);
3353 
3354   const uint clone_exit_idx = 1;
3355   const uint orig_exit_idx  = 2;
3356   assert(is_valid_clone_loop_form(loop, peel_list, orig_exit_idx, clone_exit_idx), "bad clone loop");
3357 
3358   Node* head_clone             = old_new[head->_idx];
3359   LoopNode* new_head_clone     = old_new[new_head->_idx]->as_Loop();
3360   Node* orig_tail_clone        = head_clone->in(2);
3361 
3362   // Add phi if "def" node is in peel set and "use" is not
3363 
3364   for (uint i = 0; i < peel_list.size(); i++) {
3365     Node *def  = peel_list.at(i);
3366     if (!def->is_CFG()) {
3367       for (DUIterator_Fast jmax, j = def->fast_outs(jmax); j < jmax; j++) {
3368         Node *use = def->fast_out(j);
3369         if (has_node(use) && use->in(0) != C->top() &&
3370             (!peel.test(use->_idx) ||
3371              (use->is_Phi() && use->in(0) == head)) ) {
3372           worklist.push(use);
3373         }
3374       }
3375       while( worklist.size() ) {
3376         Node *use = worklist.pop();
3377         for (uint j = 1; j < use->req(); j++) {
3378           Node* n = use->in(j);
3379           if (n == def) {
3380 
3381             // "def" is in peel set, "use" is not in peel set
3382             // or "use" is in the entry boundary (a phi) of the peel set
3383 
3384             Node* use_c = has_ctrl(use) ? get_ctrl(use) : use;
3385 
3386             if ( loop->is_member(get_loop( use_c )) ) {
3387               // use is in loop
3388               if (old_new[use->_idx] != NULL) { // null for dead code
3389                 Node* use_clone = old_new[use->_idx];
3390                 _igvn.replace_input_of(use, j, C->top());
3391                 insert_phi_for_loop( use_clone, j, old_new[def->_idx], def, new_head_clone );
3392               }
3393             } else {
3394               assert(is_valid_clone_loop_exit_use(loop, use, orig_exit_idx), "clone loop format");
3395               // use is not in the loop, check if the live range includes the cut
3396               Node* lp_if = use_c->in(orig_exit_idx)->in(0);
3397               if (not_peel.test(lp_if->_idx)) {
3398                 assert(j == orig_exit_idx, "use from original loop");
3399                 insert_phi_for_loop( use, clone_exit_idx, old_new[def->_idx], def, new_head_clone );
3400               }
3401             }
3402           }
3403         }
3404       }
3405     }
3406   }
3407 
3408   // Step 3b: retarget control
3409 
3410   // Redirect control to the new loop head if a cloned node in
3411   // the not_peeled region has control that points into the peeled region.
3412   // This necessary because the cloned peeled region will be outside
3413   // the loop.
3414   //                            from    to
3415   //          cloned-peeled    <---+
3416   //    new_head_clone:            |    <--+
3417   //          cloned-not_peeled  in(0)    in(0)
3418   //          orig-peeled
3419 
3420   for (uint i = 0; i < loop->_body.size(); i++) {
3421     Node *n = loop->_body.at(i);
3422     if (!n->is_CFG()           && n->in(0) != NULL        &&
3423         not_peel.test(n->_idx) && peel.test(n->in(0)->_idx)) {
3424       Node* n_clone = old_new[n->_idx];
3425       _igvn.replace_input_of(n_clone, 0, new_head_clone);
3426     }
3427   }
3428 
3429   // Backedge of the surviving new_head (the clone) is original last_peel
3430   _igvn.replace_input_of(new_head_clone, LoopNode::LoopBackControl, last_peel);
3431 
3432   // Cut first node in original not_peel set
3433   _igvn.rehash_node_delayed(new_head);                     // Multiple edge updates:
3434   new_head->set_req(LoopNode::EntryControl,    C->top());  //   use rehash_node_delayed / set_req instead of
3435   new_head->set_req(LoopNode::LoopBackControl, C->top());  //   multiple replace_input_of calls
3436 
3437   // Copy head_clone back-branch info to original head
3438   // and remove original head's loop entry and
3439   // clone head's back-branch
3440   _igvn.rehash_node_delayed(head); // Multiple edge updates
3441   head->set_req(LoopNode::EntryControl,    head_clone->in(LoopNode::LoopBackControl));
3442   head->set_req(LoopNode::LoopBackControl, C->top());
3443   _igvn.replace_input_of(head_clone, LoopNode::LoopBackControl, C->top());
3444 
3445   // Similarly modify the phis
3446   for (DUIterator_Fast kmax, k = head->fast_outs(kmax); k < kmax; k++) {
3447     Node* use = head->fast_out(k);
3448     if (use->is_Phi() && use->outcnt() > 0) {
3449       Node* use_clone = old_new[use->_idx];
3450       _igvn.rehash_node_delayed(use); // Multiple edge updates
3451       use->set_req(LoopNode::EntryControl,    use_clone->in(LoopNode::LoopBackControl));
3452       use->set_req(LoopNode::LoopBackControl, C->top());
3453       _igvn.replace_input_of(use_clone, LoopNode::LoopBackControl, C->top());
3454     }
3455   }
3456 
3457   // Step 4: update dominator tree and dominator depth
3458 
3459   set_idom(head, orig_tail_clone, dd);
3460   recompute_dom_depth();
3461 
3462   // Inhibit more partial peeling on this loop
3463   new_head_clone->set_partial_peel_loop();
3464   C->set_major_progress();
3465   loop->record_for_igvn();
3466 
3467 #ifndef PRODUCT
3468   if (TracePartialPeeling) {
3469     tty->print_cr("\nafter partial peel one iteration");
3470     Node_List wl;
3471     Node* t = last_peel;
3472     while (true) {
3473       wl.push(t);
3474       if (t == head_clone) break;
3475       t = idom(t);
3476     }
3477     while (wl.size() > 0) {
3478       Node* tt = wl.pop();
3479       if (tt == head) tty->print_cr("orig head");
3480       else if (tt == new_head_clone) tty->print_cr("new head");
3481       else if (tt == head_clone) tty->print_cr("clone head");
3482       tt->dump();
3483     }
3484   }
3485 #endif
3486   return true;
3487 }
3488 
3489 //------------------------------reorg_offsets----------------------------------
3490 // Reorganize offset computations to lower register pressure.  Mostly
3491 // prevent loop-fallout uses of the pre-incremented trip counter (which are
3492 // then alive with the post-incremented trip counter forcing an extra
3493 // register move)
3494 void PhaseIdealLoop::reorg_offsets(IdealLoopTree *loop) {
3495   // Perform it only for canonical counted loops.
3496   // Loop's shape could be messed up by iteration_split_impl.
3497   if (!loop->_head->is_CountedLoop())
3498     return;
3499   if (!loop->_head->as_Loop()->is_valid_counted_loop())
3500     return;
3501 
3502   CountedLoopNode *cl = loop->_head->as_CountedLoop();
3503   CountedLoopEndNode *cle = cl->loopexit();
3504   Node *exit = cle->proj_out(false);
3505   Node *phi = cl->phi();
3506 
3507   // Check for the special case when using the pre-incremented trip-counter on
3508   // the fall-out  path (forces the pre-incremented  and post-incremented trip
3509   // counter to be live  at the same time).  Fix this by  adjusting to use the
3510   // post-increment trip counter.
3511 
3512   bool progress = true;
3513   while (progress) {
3514     progress = false;
3515     for (DUIterator_Fast imax, i = phi->fast_outs(imax); i < imax; i++) {
3516       Node* use = phi->fast_out(i);   // User of trip-counter
3517       if (!has_ctrl(use))  continue;
3518       Node *u_ctrl = get_ctrl(use);
3519       if (use->is_Phi()) {
3520         u_ctrl = NULL;
3521         for (uint j = 1; j < use->req(); j++)
3522           if (use->in(j) == phi)
3523             u_ctrl = dom_lca(u_ctrl, use->in(0)->in(j));
3524       }
3525       IdealLoopTree *u_loop = get_loop(u_ctrl);
3526       // Look for loop-invariant use
3527       if (u_loop == loop) continue;
3528       if (loop->is_member(u_loop)) continue;
3529       // Check that use is live out the bottom.  Assuming the trip-counter
3530       // update is right at the bottom, uses of of the loop middle are ok.
3531       if (dom_lca(exit, u_ctrl) != exit) continue;
3532       // Hit!  Refactor use to use the post-incremented tripcounter.
3533       // Compute a post-increment tripcounter.
3534       Node* c = exit;
3535       if (cl->is_strip_mined()) {
3536         IdealLoopTree* outer_loop = get_loop(cl->outer_loop());
3537         if (!outer_loop->is_member(u_loop)) {
3538           c = cl->outer_loop_exit();
3539         }
3540       }
3541       Node *opaq = new Opaque2Node(C, cle->incr());
3542       register_new_node(opaq, c);
3543       Node *neg_stride = _igvn.intcon(-cle->stride_con());
3544       set_ctrl(neg_stride, C->root());
3545       Node *post = new AddINode(opaq, neg_stride);
3546       register_new_node(post, c);
3547       _igvn.rehash_node_delayed(use);
3548       for (uint j = 1; j < use->req(); j++) {
3549         if (use->in(j) == phi)
3550           use->set_req(j, post);
3551       }
3552       // Since DU info changed, rerun loop
3553       progress = true;
3554       break;
3555     }
3556   }
3557 
3558 }