1 /*
   2  * Copyright (c) 2011, 2014, 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 "opto/loopnode.hpp"
  27 #include "opto/addnode.hpp"
  28 #include "opto/callnode.hpp"
  29 #include "opto/connode.hpp"
  30 #include "opto/convertnode.hpp"
  31 #include "opto/loopnode.hpp"
  32 #include "opto/mulnode.hpp"
  33 #include "opto/opaquenode.hpp"
  34 #include "opto/rootnode.hpp"
  35 #include "opto/subnode.hpp"
  36 
  37 /*
  38  * The general idea of Loop Predication is to insert a predicate on the entry
  39  * path to a loop, and raise a uncommon trap if the check of the condition fails.
  40  * The condition checks are promoted from inside the loop body, and thus
  41  * the checks inside the loop could be eliminated. Currently, loop predication
  42  * optimization has been applied to remove array range check and loop invariant
  43  * checks (such as null checks).
  44 */
  45 
  46 //-------------------------------register_control-------------------------
  47 void PhaseIdealLoop::register_control(Node* n, IdealLoopTree *loop, Node* pred) {
  48   assert(n->is_CFG(), "must be control node");
  49   _igvn.register_new_node_with_optimizer(n);
  50   loop->_body.push(n);
  51   set_loop(n, loop);
  52   // When called from beautify_loops() idom is not constructed yet.
  53   if (_idom != NULL) {
  54     set_idom(n, pred, dom_depth(pred));
  55   }
  56 }
  57 
  58 //------------------------------create_new_if_for_predicate------------------------
  59 // create a new if above the uct_if_pattern for the predicate to be promoted.
  60 //
  61 //          before                                after
  62 //        ----------                           ----------
  63 //           ctrl                                 ctrl
  64 //            |                                     |
  65 //            |                                     |
  66 //            v                                     v
  67 //           iff                                 new_iff
  68 //          /    \                                /      \
  69 //         /      \                              /        \
  70 //        v        v                            v          v
  71 //  uncommon_proj cont_proj                   if_uct     if_cont
  72 // \      |        |                           |          |
  73 //  \     |        |                           |          |
  74 //   v    v        v                           |          v
  75 //     rgn       loop                          |         iff
  76 //      |                                      |        /     \
  77 //      |                                      |       /       \
  78 //      v                                      |      v         v
  79 // uncommon_trap                               | uncommon_proj cont_proj
  80 //                                           \  \    |           |
  81 //                                            \  \   |           |
  82 //                                             v  v  v           v
  83 //                                               rgn           loop
  84 //                                                |
  85 //                                                |
  86 //                                                v
  87 //                                           uncommon_trap
  88 //
  89 //
  90 // We will create a region to guard the uct call if there is no one there.
  91 // The true projecttion (if_cont) of the new_iff is returned.
  92 // This code is also used to clone predicates to cloned loops.
  93 ProjNode* PhaseIdealLoop::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
  94                                                       Deoptimization::DeoptReason reason,
  95                                                       int opcode) {
  96   assert(cont_proj->is_uncommon_trap_if_pattern(reason), "must be a uct if pattern!");
  97   IfNode* iff = cont_proj->in(0)->as_If();
  98 
  99   ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con);
 100   Node     *rgn   = uncommon_proj->unique_ctrl_out();
 101   assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
 102 
 103   uint proj_index = 1; // region's edge corresponding to uncommon_proj
 104   if (!rgn->is_Region()) { // create a region to guard the call
 105     assert(rgn->is_Call(), "must be call uct");
 106     CallNode* call = rgn->as_Call();
 107     IdealLoopTree* loop = get_loop(call);
 108     rgn = new RegionNode(1);
 109     rgn->add_req(uncommon_proj);
 110     register_control(rgn, loop, uncommon_proj);
 111     _igvn.replace_input_of(call, 0, rgn);
 112     // When called from beautify_loops() idom is not constructed yet.
 113     if (_idom != NULL) {
 114       set_idom(call, rgn, dom_depth(rgn));
 115     }
 116     for (DUIterator_Fast imax, i = uncommon_proj->fast_outs(imax); i < imax; i++) {
 117       Node* n = uncommon_proj->fast_out(i);
 118       if (n->is_Load() || n->is_Store()) {
 119         _igvn.replace_input_of(n, 0, rgn);
 120         --i; --imax;
 121       }
 122     }
 123   } else {
 124     // Find region's edge corresponding to uncommon_proj
 125     for (; proj_index < rgn->req(); proj_index++)
 126       if (rgn->in(proj_index) == uncommon_proj) break;
 127     assert(proj_index < rgn->req(), "sanity");
 128   }
 129 
 130   Node* entry = iff->in(0);
 131   if (new_entry != NULL) {
 132     // Clonning the predicate to new location.
 133     entry = new_entry;
 134   }
 135   // Create new_iff
 136   IdealLoopTree* lp = get_loop(entry);
 137   IfNode* new_iff = NULL;
 138   if (opcode == Op_If) {
 139     new_iff = new IfNode(entry, iff->in(1), iff->_prob, iff->_fcnt);
 140   } else {
 141     assert(opcode == Op_RangeCheck, "no other if variant here");
 142     new_iff = new RangeCheckNode(entry, iff->in(1), iff->_prob, iff->_fcnt);
 143   }
 144   register_control(new_iff, lp, entry);
 145   Node *if_cont = new IfTrueNode(new_iff);
 146   Node *if_uct  = new IfFalseNode(new_iff);
 147   if (cont_proj->is_IfFalse()) {
 148     // Swap
 149     Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp;
 150   }
 151   register_control(if_cont, lp, new_iff);
 152   register_control(if_uct, get_loop(rgn), new_iff);
 153 
 154   // if_uct to rgn
 155   _igvn.hash_delete(rgn);
 156   rgn->add_req(if_uct);
 157   // When called from beautify_loops() idom is not constructed yet.
 158   if (_idom != NULL) {
 159     Node* ridom = idom(rgn);
 160     Node* nrdom = dom_lca(ridom, new_iff);
 161     set_idom(rgn, nrdom, dom_depth(rgn));
 162   }
 163 
 164   // If rgn has phis add new edges which has the same
 165   // value as on original uncommon_proj pass.
 166   assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last");
 167   bool has_phi = false;
 168   for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) {
 169     Node* use = rgn->fast_out(i);
 170     if (use->is_Phi() && use->outcnt() > 0) {
 171       assert(use->in(0) == rgn, "");
 172       _igvn.rehash_node_delayed(use);
 173       use->add_req(use->in(proj_index));
 174       has_phi = true;
 175     }
 176   }
 177   assert(!has_phi || rgn->req() > 3, "no phis when region is created");
 178 
 179   if (new_entry == NULL) {
 180     // Attach if_cont to iff
 181     _igvn.replace_input_of(iff, 0, if_cont);
 182     if (_idom != NULL) {
 183       set_idom(iff, if_cont, dom_depth(iff));
 184     }
 185   }
 186   return if_cont->as_Proj();
 187 }
 188 
 189 //------------------------------create_new_if_for_predicate------------------------
 190 // Create a new if below new_entry for the predicate to be cloned (IGVN optimization)
 191 ProjNode* PhaseIterGVN::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
 192                                                     Deoptimization::DeoptReason reason,
 193                                                     int opcode) {
 194   assert(new_entry != 0, "only used for clone predicate");
 195   assert(cont_proj->is_uncommon_trap_if_pattern(reason), "must be a uct if pattern!");
 196   IfNode* iff = cont_proj->in(0)->as_If();
 197 
 198   ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con);
 199   Node     *rgn   = uncommon_proj->unique_ctrl_out();
 200   assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
 201 
 202   uint proj_index = 1; // region's edge corresponding to uncommon_proj
 203   if (!rgn->is_Region()) { // create a region to guard the call
 204     assert(rgn->is_Call(), "must be call uct");
 205     CallNode* call = rgn->as_Call();
 206     rgn = new RegionNode(1);
 207     register_new_node_with_optimizer(rgn);
 208     rgn->add_req(uncommon_proj);
 209     replace_input_of(call, 0, rgn);
 210   } else {
 211     // Find region's edge corresponding to uncommon_proj
 212     for (; proj_index < rgn->req(); proj_index++)
 213       if (rgn->in(proj_index) == uncommon_proj) break;
 214     assert(proj_index < rgn->req(), "sanity");
 215   }
 216 
 217   // Create new_iff in new location.
 218   IfNode* new_iff = NULL;
 219   if (opcode == Op_If) {
 220     new_iff = new IfNode(new_entry, iff->in(1), iff->_prob, iff->_fcnt);
 221   } else {
 222     assert(opcode == Op_RangeCheck, "no other if variant here");
 223     new_iff = new RangeCheckNode(new_entry, iff->in(1), iff->_prob, iff->_fcnt);
 224   }
 225 
 226   register_new_node_with_optimizer(new_iff);
 227   Node *if_cont = new IfTrueNode(new_iff);
 228   Node *if_uct  = new IfFalseNode(new_iff);
 229   if (cont_proj->is_IfFalse()) {
 230     // Swap
 231     Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp;
 232   }
 233   register_new_node_with_optimizer(if_cont);
 234   register_new_node_with_optimizer(if_uct);
 235 
 236   // if_uct to rgn
 237   hash_delete(rgn);
 238   rgn->add_req(if_uct);
 239 
 240   // If rgn has phis add corresponding new edges which has the same
 241   // value as on original uncommon_proj pass.
 242   assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last");
 243   bool has_phi = false;
 244   for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) {
 245     Node* use = rgn->fast_out(i);
 246     if (use->is_Phi() && use->outcnt() > 0) {
 247       rehash_node_delayed(use);
 248       use->add_req(use->in(proj_index));
 249       has_phi = true;
 250     }
 251   }
 252   assert(!has_phi || rgn->req() > 3, "no phis when region is created");
 253 
 254   return if_cont->as_Proj();
 255 }
 256 
 257 //--------------------------clone_predicate-----------------------
 258 ProjNode* PhaseIdealLoop::clone_predicate(ProjNode* predicate_proj, Node* new_entry,
 259                                           Deoptimization::DeoptReason reason,
 260                                           PhaseIdealLoop* loop_phase,
 261                                           PhaseIterGVN* igvn) {
 262   ProjNode* new_predicate_proj;
 263   if (loop_phase != NULL) {
 264     new_predicate_proj = loop_phase->create_new_if_for_predicate(predicate_proj, new_entry, reason, Op_If);
 265   } else {
 266     new_predicate_proj =       igvn->create_new_if_for_predicate(predicate_proj, new_entry, reason, Op_If);
 267   }
 268   IfNode* iff = new_predicate_proj->in(0)->as_If();
 269   Node* ctrl  = iff->in(0);
 270 
 271   // Match original condition since predicate's projections could be swapped.
 272   assert(predicate_proj->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be");
 273   Node* opq = new Opaque1Node(igvn->C, predicate_proj->in(0)->in(1)->in(1)->in(1));
 274   igvn->C->add_predicate_opaq(opq);
 275 
 276   Node* bol = new Conv2BNode(opq);
 277   if (loop_phase != NULL) {
 278     loop_phase->register_new_node(opq, ctrl);
 279     loop_phase->register_new_node(bol, ctrl);
 280   } else {
 281     igvn->register_new_node_with_optimizer(opq);
 282     igvn->register_new_node_with_optimizer(bol);
 283   }
 284   igvn->hash_delete(iff);
 285   iff->set_req(1, bol);
 286   return new_predicate_proj;
 287 }
 288 
 289 
 290 //--------------------------clone_loop_predicates-----------------------
 291 // Interface from IGVN
 292 Node* PhaseIterGVN::clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check) {
 293   return PhaseIdealLoop::clone_loop_predicates(old_entry, new_entry, clone_limit_check, NULL, this);
 294 }
 295 
 296 // Interface from PhaseIdealLoop
 297 Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check) {
 298   return clone_loop_predicates(old_entry, new_entry, clone_limit_check, this, &this->_igvn);
 299 }
 300 
 301 // Clone loop predicates to cloned loops (peeled, unswitched, split_if).
 302 Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry,
 303                                                 bool clone_limit_check,
 304                                                 PhaseIdealLoop* loop_phase,
 305                                                 PhaseIterGVN* igvn) {
 306 #ifdef ASSERT
 307   if (new_entry == NULL || !(new_entry->is_Proj() || new_entry->is_Region() || new_entry->is_SafePoint())) {
 308     if (new_entry != NULL)
 309       new_entry->dump();
 310     assert(false, "not IfTrue, IfFalse, Region or SafePoint");
 311   }
 312 #endif
 313   // Search original predicates
 314   Node* entry = old_entry;
 315   ProjNode* limit_check_proj = NULL;
 316   limit_check_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
 317   if (limit_check_proj != NULL) {
 318     entry = entry->in(0)->in(0);
 319   }
 320   if (UseLoopPredicate) {
 321     ProjNode* predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
 322     if (predicate_proj != NULL) { // right pattern that can be used by loop predication
 323       // clone predicate
 324       new_entry = clone_predicate(predicate_proj, new_entry,
 325                                   Deoptimization::Reason_predicate,
 326                                   loop_phase, igvn);
 327       assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone predicate");
 328       if (TraceLoopPredicate) {
 329         tty->print("Loop Predicate cloned: ");
 330         debug_only( new_entry->in(0)->dump(); )
 331       }
 332     }
 333   }
 334   if (limit_check_proj != NULL && clone_limit_check) {
 335     // Clone loop limit check last to insert it before loop.
 336     // Don't clone a limit check which was already finalized
 337     // for this counted loop (only one limit check is needed).
 338     new_entry = clone_predicate(limit_check_proj, new_entry,
 339                                 Deoptimization::Reason_loop_limit_check,
 340                                 loop_phase, igvn);
 341     assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone limit check");
 342     if (TraceLoopLimitCheck) {
 343       tty->print("Loop Limit Check cloned: ");
 344       debug_only( new_entry->in(0)->dump(); )
 345     }
 346   }
 347   return new_entry;
 348 }
 349 
 350 //--------------------------skip_loop_predicates------------------------------
 351 // Skip related predicates.
 352 Node* PhaseIdealLoop::skip_loop_predicates(Node* entry) {
 353   Node* predicate = NULL;
 354   predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
 355   if (predicate != NULL) {
 356     entry = entry->in(0)->in(0);
 357   }
 358   if (UseLoopPredicate) {
 359     predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
 360     if (predicate != NULL) { // right pattern that can be used by loop predication
 361       IfNode* iff = entry->in(0)->as_If();
 362       ProjNode* uncommon_proj = iff->proj_out(1 - entry->as_Proj()->_con);
 363       Node* rgn = uncommon_proj->unique_ctrl_out();
 364       assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
 365       entry = entry->in(0)->in(0);
 366       while (entry != NULL && entry->is_Proj() && entry->in(0)->is_If()) {
 367         uncommon_proj = entry->in(0)->as_If()->proj_out(1 - entry->as_Proj()->_con);
 368         if (uncommon_proj->unique_ctrl_out() != rgn)
 369           break;
 370         entry = entry->in(0)->in(0);
 371       }
 372     }
 373   }
 374   return entry;
 375 }
 376 
 377 //--------------------------find_predicate_insertion_point-------------------
 378 // Find a good location to insert a predicate
 379 ProjNode* PhaseIdealLoop::find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason) {
 380   if (start_c == NULL || !start_c->is_Proj())
 381     return NULL;
 382   if (start_c->as_Proj()->is_uncommon_trap_if_pattern(reason)) {
 383     return start_c->as_Proj();
 384   }
 385   return NULL;
 386 }
 387 
 388 //--------------------------find_predicate------------------------------------
 389 // Find a predicate
 390 Node* PhaseIdealLoop::find_predicate(Node* entry) {
 391   Node* predicate = NULL;
 392   predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
 393   if (predicate != NULL) { // right pattern that can be used by loop predication
 394     return entry;
 395   }
 396   if (UseLoopPredicate) {
 397     predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
 398     if (predicate != NULL) { // right pattern that can be used by loop predication
 399       return entry;
 400     }
 401   }
 402   return NULL;
 403 }
 404 
 405 //------------------------------Invariance-----------------------------------
 406 // Helper class for loop_predication_impl to compute invariance on the fly and
 407 // clone invariants.
 408 class Invariance : public StackObj {
 409   VectorSet _visited, _invariant;
 410   Node_Stack _stack;
 411   VectorSet _clone_visited;
 412   Node_List _old_new; // map of old to new (clone)
 413   IdealLoopTree* _lpt;
 414   PhaseIdealLoop* _phase;
 415 
 416   // Helper function to set up the invariance for invariance computation
 417   // If n is a known invariant, set up directly. Otherwise, look up the
 418   // the possibility to push n onto the stack for further processing.
 419   void visit(Node* use, Node* n) {
 420     if (_lpt->is_invariant(n)) { // known invariant
 421       _invariant.set(n->_idx);
 422     } else if (!n->is_CFG()) {
 423       Node *n_ctrl = _phase->ctrl_or_self(n);
 424       Node *u_ctrl = _phase->ctrl_or_self(use); // self if use is a CFG
 425       if (_phase->is_dominator(n_ctrl, u_ctrl)) {
 426         _stack.push(n, n->in(0) == NULL ? 1 : 0);
 427       }
 428     }
 429   }
 430 
 431   // Compute invariance for "the_node" and (possibly) all its inputs recursively
 432   // on the fly
 433   void compute_invariance(Node* n) {
 434     assert(_visited.test(n->_idx), "must be");
 435     visit(n, n);
 436     while (_stack.is_nonempty()) {
 437       Node*  n = _stack.node();
 438       uint idx = _stack.index();
 439       if (idx == n->req()) { // all inputs are processed
 440         _stack.pop();
 441         // n is invariant if it's inputs are all invariant
 442         bool all_inputs_invariant = true;
 443         for (uint i = 0; i < n->req(); i++) {
 444           Node* in = n->in(i);
 445           if (in == NULL) continue;
 446           assert(_visited.test(in->_idx), "must have visited input");
 447           if (!_invariant.test(in->_idx)) { // bad guy
 448             all_inputs_invariant = false;
 449             break;
 450           }
 451         }
 452         if (all_inputs_invariant) {
 453           // If n's control is a predicate that was moved out of the
 454           // loop, it was marked invariant but n is only invariant if
 455           // it depends only on that test. Otherwise, unless that test
 456           // is out of the loop, it's not invariant.
 457           if (n->is_CFG() || n->depends_only_on_test() || n->in(0) == NULL || !_phase->is_member(_lpt, n->in(0))) {
 458             _invariant.set(n->_idx); // I am a invariant too
 459           }
 460         }
 461       } else { // process next input
 462         _stack.set_index(idx + 1);
 463         Node* m = n->in(idx);
 464         if (m != NULL && !_visited.test_set(m->_idx)) {
 465           visit(n, m);
 466         }
 467       }
 468     }
 469   }
 470 
 471   // Helper function to set up _old_new map for clone_nodes.
 472   // If n is a known invariant, set up directly ("clone" of n == n).
 473   // Otherwise, push n onto the stack for real cloning.
 474   void clone_visit(Node* n) {
 475     assert(_invariant.test(n->_idx), "must be invariant");
 476     if (_lpt->is_invariant(n)) { // known invariant
 477       _old_new.map(n->_idx, n);
 478     } else { // to be cloned
 479       assert(!n->is_CFG(), "should not see CFG here");
 480       _stack.push(n, n->in(0) == NULL ? 1 : 0);
 481     }
 482   }
 483 
 484   // Clone "n" and (possibly) all its inputs recursively
 485   void clone_nodes(Node* n, Node* ctrl) {
 486     clone_visit(n);
 487     while (_stack.is_nonempty()) {
 488       Node*  n = _stack.node();
 489       uint idx = _stack.index();
 490       if (idx == n->req()) { // all inputs processed, clone n!
 491         _stack.pop();
 492         // clone invariant node
 493         Node* n_cl = n->clone();
 494         _old_new.map(n->_idx, n_cl);
 495         _phase->register_new_node(n_cl, ctrl);
 496         for (uint i = 0; i < n->req(); i++) {
 497           Node* in = n_cl->in(i);
 498           if (in == NULL) continue;
 499           n_cl->set_req(i, _old_new[in->_idx]);
 500         }
 501       } else { // process next input
 502         _stack.set_index(idx + 1);
 503         Node* m = n->in(idx);
 504         if (m != NULL && !_clone_visited.test_set(m->_idx)) {
 505           clone_visit(m); // visit the input
 506         }
 507       }
 508     }
 509   }
 510 
 511  public:
 512   Invariance(Arena* area, IdealLoopTree* lpt) :
 513     _lpt(lpt), _phase(lpt->_phase),
 514     _visited(area), _invariant(area), _stack(area, 10 /* guess */),
 515     _clone_visited(area), _old_new(area)
 516   {
 517     Node* head = _lpt->_head;
 518     Node* entry = head->in(LoopNode::EntryControl);
 519     if (entry->outcnt() != 1) {
 520       // If a node is pinned between the predicates and the loop
 521       // entry, we won't be able to move any node in the loop that
 522       // depends on it above it in a predicate. Mark all those nodes
 523       // as non loop invariatnt.
 524       Unique_Node_List wq;
 525       wq.push(entry);
 526       for (uint next = 0; next < wq.size(); ++next) {
 527         Node *n = wq.at(next);
 528         for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 529           Node* u = n->fast_out(i);
 530           if (!u->is_CFG()) {
 531             Node* c = _phase->get_ctrl(u);
 532             if (_lpt->is_member(_phase->get_loop(c)) || _phase->is_dominator(c, head)) {
 533               _visited.set(u->_idx);
 534               wq.push(u);
 535             }
 536           }
 537         }
 538       }
 539     }
 540   }
 541 
 542   // Map old to n for invariance computation and clone
 543   void map_ctrl(Node* old, Node* n) {
 544     assert(old->is_CFG() && n->is_CFG(), "must be");
 545     _old_new.map(old->_idx, n); // "clone" of old is n
 546     _invariant.set(old->_idx);  // old is invariant
 547     _clone_visited.set(old->_idx);
 548   }
 549 
 550   // Driver function to compute invariance
 551   bool is_invariant(Node* n) {
 552     if (!_visited.test_set(n->_idx))
 553       compute_invariance(n);
 554     return (_invariant.test(n->_idx) != 0);
 555   }
 556 
 557   // Driver function to clone invariant
 558   Node* clone(Node* n, Node* ctrl) {
 559     assert(ctrl->is_CFG(), "must be");
 560     assert(_invariant.test(n->_idx), "must be an invariant");
 561     if (!_clone_visited.test(n->_idx))
 562       clone_nodes(n, ctrl);
 563     return _old_new[n->_idx];
 564   }
 565 };
 566 
 567 //------------------------------is_range_check_if -----------------------------------
 568 // Returns true if the predicate of iff is in "scale*iv + offset u< load_range(ptr)" format
 569 // Note: this function is particularly designed for loop predication. We require load_range
 570 //       and offset to be loop invariant computed on the fly by "invar"
 571 bool IdealLoopTree::is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const {
 572   if (!is_loop_exit(iff)) {
 573     return false;
 574   }
 575   if (!iff->in(1)->is_Bool()) {
 576     return false;
 577   }
 578   const BoolNode *bol = iff->in(1)->as_Bool();
 579   if (bol->_test._test != BoolTest::lt) {
 580     return false;
 581   }
 582   if (!bol->in(1)->is_Cmp()) {
 583     return false;
 584   }
 585   const CmpNode *cmp = bol->in(1)->as_Cmp();
 586   if (cmp->Opcode() != Op_CmpU) {
 587     return false;
 588   }
 589   Node* range = cmp->in(2);
 590   if (range->Opcode() != Op_LoadRange && !iff->is_RangeCheck()) {
 591     const TypeInt* tint = phase->_igvn.type(range)->isa_int();
 592     if (tint == NULL || tint->empty() || tint->_lo < 0) {
 593       // Allow predication on positive values that aren't LoadRanges.
 594       // This allows optimization of loops where the length of the
 595       // array is a known value and doesn't need to be loaded back
 596       // from the array.
 597       return false;
 598     }
 599   }
 600   if (!invar.is_invariant(range)) {
 601     return false;
 602   }
 603   Node *iv     = _head->as_CountedLoop()->phi();
 604   int   scale  = 0;
 605   Node *offset = NULL;
 606   if (!phase->is_scaled_iv_plus_offset(cmp->in(1), iv, &scale, &offset)) {
 607     return false;
 608   }
 609   if (offset && !invar.is_invariant(offset)) { // offset must be invariant
 610     return false;
 611   }
 612   return true;
 613 }
 614 
 615 //------------------------------rc_predicate-----------------------------------
 616 // Create a range check predicate
 617 //
 618 // for (i = init; i < limit; i += stride) {
 619 //    a[scale*i+offset]
 620 // }
 621 //
 622 // Compute max(scale*i + offset) for init <= i < limit and build the predicate
 623 // as "max(scale*i + offset) u< a.length".
 624 //
 625 // There are two cases for max(scale*i + offset):
 626 // (1) stride*scale > 0
 627 //   max(scale*i + offset) = scale*(limit-stride) + offset
 628 // (2) stride*scale < 0
 629 //   max(scale*i + offset) = scale*init + offset
 630 BoolNode* PhaseIdealLoop::rc_predicate(IdealLoopTree *loop, Node* ctrl,
 631                                        int scale, Node* offset,
 632                                        Node* init, Node* limit, Node* stride,
 633                                        Node* range, bool upper) {
 634   stringStream* predString = NULL;
 635   if (TraceLoopPredicate) {
 636     predString = new stringStream();
 637     predString->print("rc_predicate ");
 638   }
 639 
 640   Node* max_idx_expr  = init;
 641   int stride_con = stride->get_int();
 642   if ((stride_con > 0) == (scale > 0) == upper) {
 643     // Limit is not exact.
 644     // Calculate exact limit here.
 645     // Note, counted loop's test is '<' or '>'.
 646     limit = exact_limit(loop);
 647     max_idx_expr = new SubINode(limit, stride);
 648     register_new_node(max_idx_expr, ctrl);
 649     if (TraceLoopPredicate) predString->print("(limit - stride) ");
 650   } else {
 651     if (TraceLoopPredicate) predString->print("init ");
 652   }
 653 
 654   if (scale != 1) {
 655     ConNode* con_scale = _igvn.intcon(scale);
 656     set_ctrl(con_scale, C->root());
 657     max_idx_expr = new MulINode(max_idx_expr, con_scale);
 658     register_new_node(max_idx_expr, ctrl);
 659     if (TraceLoopPredicate) predString->print("* %d ", scale);
 660   }
 661 
 662   if (offset && (!offset->is_Con() || offset->get_int() != 0)){
 663     max_idx_expr = new AddINode(max_idx_expr, offset);
 664     register_new_node(max_idx_expr, ctrl);
 665     if (TraceLoopPredicate)
 666       if (offset->is_Con()) predString->print("+ %d ", offset->get_int());
 667       else predString->print("+ offset ");
 668   }
 669 
 670   CmpUNode* cmp = new CmpUNode(max_idx_expr, range);
 671   register_new_node(cmp, ctrl);
 672   BoolNode* bol = new BoolNode(cmp, BoolTest::lt);
 673   register_new_node(bol, ctrl);
 674 
 675   if (TraceLoopPredicate) {
 676     predString->print_cr("<u range");
 677     tty->print("%s", predString->as_string());
 678   }
 679   return bol;
 680 }
 681 
 682 //------------------------------ loop_predication_impl--------------------------
 683 // Insert loop predicates for null checks and range checks
 684 bool PhaseIdealLoop::loop_predication_impl(IdealLoopTree *loop) {
 685   if (!UseLoopPredicate) return false;
 686 
 687   if (!loop->_head->is_Loop()) {
 688     // Could be a simple region when irreducible loops are present.
 689     return false;
 690   }
 691   LoopNode* head = loop->_head->as_Loop();
 692 
 693   if (head->unique_ctrl_out()->Opcode() == Op_NeverBranch) {
 694     // do nothing for infinite loops
 695     return false;
 696   }
 697 
 698   CountedLoopNode *cl = NULL;
 699   if (head->is_valid_counted_loop()) {
 700     cl = head->as_CountedLoop();
 701     // do nothing for iteration-splitted loops
 702     if (!cl->is_normal_loop()) return false;
 703     // Avoid RCE if Counted loop's test is '!='.
 704     BoolTest::mask bt = cl->loopexit()->test_trip();
 705     if (bt != BoolTest::lt && bt != BoolTest::gt)
 706       cl = NULL;
 707   }
 708 
 709   Node* entry = head->in(LoopNode::EntryControl);
 710   ProjNode *predicate_proj = NULL;
 711   // Loop limit check predicate should be near the loop.
 712   predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
 713   if (predicate_proj != NULL)
 714     entry = predicate_proj->in(0)->in(0);
 715   predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
 716   if (!predicate_proj) {
 717 #ifndef PRODUCT
 718     if (TraceLoopPredicate) {
 719       tty->print("missing predicate:");
 720       loop->dump_head();
 721       head->dump(1);
 722     }
 723 #endif
 724     return false;
 725   }
 726   ConNode* zero = _igvn.intcon(0);
 727   set_ctrl(zero, C->root());
 728 
 729   ResourceArea *area = Thread::current()->resource_area();
 730   Invariance invar(area, loop);
 731 
 732   // Create list of if-projs such that a newer proj dominates all older
 733   // projs in the list, and they all dominate loop->tail()
 734   Node_List if_proj_list(area);
 735   Node *current_proj = loop->tail(); //start from tail
 736   while (current_proj != head) {
 737     if (loop == get_loop(current_proj) && // still in the loop ?
 738         current_proj->is_Proj()        && // is a projection  ?
 739         (current_proj->in(0)->Opcode() == Op_If ||
 740          current_proj->in(0)->Opcode() == Op_RangeCheck)) { // is a if projection ?
 741       if_proj_list.push(current_proj);
 742     }
 743     current_proj = idom(current_proj);
 744   }
 745 
 746   bool hoisted = false; // true if at least one proj is promoted
 747   while (if_proj_list.size() > 0) {
 748     // Following are changed to nonnull when a predicate can be hoisted
 749     ProjNode* new_predicate_proj = NULL;
 750 
 751     ProjNode* proj = if_proj_list.pop()->as_Proj();
 752     IfNode*   iff  = proj->in(0)->as_If();
 753 
 754     if (!proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
 755       if (loop->is_loop_exit(iff)) {
 756         // stop processing the remaining projs in the list because the execution of them
 757         // depends on the condition of "iff" (iff->in(1)).
 758         break;
 759       } else {
 760         // Both arms are inside the loop. There are two cases:
 761         // (1) there is one backward branch. In this case, any remaining proj
 762         //     in the if_proj list post-dominates "iff". So, the condition of "iff"
 763         //     does not determine the execution the remining projs directly, and we
 764         //     can safely continue.
 765         // (2) both arms are forwarded, i.e. a diamond shape. In this case, "proj"
 766         //     does not dominate loop->tail(), so it can not be in the if_proj list.
 767         continue;
 768       }
 769     }
 770 
 771     Node*     test = iff->in(1);
 772     if (!test->is_Bool()){ //Conv2B, ...
 773       continue;
 774     }
 775     BoolNode* bol = test->as_Bool();
 776     if (invar.is_invariant(bol)) {
 777       // Invariant test
 778       new_predicate_proj = create_new_if_for_predicate(predicate_proj, NULL,
 779                                                        Deoptimization::Reason_predicate,
 780                                                        iff->Opcode());
 781       Node* ctrl = new_predicate_proj->in(0)->as_If()->in(0);
 782       BoolNode* new_predicate_bol = invar.clone(bol, ctrl)->as_Bool();
 783 
 784       // Negate test if necessary
 785       bool negated = false;
 786       if (proj->_con != predicate_proj->_con) {
 787         new_predicate_bol = new BoolNode(new_predicate_bol->in(1), new_predicate_bol->_test.negate());
 788         register_new_node(new_predicate_bol, ctrl);
 789         negated = true;
 790       }
 791       IfNode* new_predicate_iff = new_predicate_proj->in(0)->as_If();
 792       _igvn.hash_delete(new_predicate_iff);
 793       new_predicate_iff->set_req(1, new_predicate_bol);
 794 #ifndef PRODUCT
 795       if (TraceLoopPredicate) {
 796         tty->print("Predicate invariant if%s: %d ", negated ? " negated" : "", new_predicate_iff->_idx);
 797         loop->dump_head();
 798       } else if (TraceLoopOpts) {
 799         tty->print("Predicate IC ");
 800         loop->dump_head();
 801       }
 802 #endif
 803     } else if (cl != NULL && loop->is_range_check_if(iff, this, invar)) {
 804       // Range check for counted loops
 805       const Node*    cmp    = bol->in(1)->as_Cmp();
 806       Node*          idx    = cmp->in(1);
 807       assert(!invar.is_invariant(idx), "index is variant");
 808       Node* rng = cmp->in(2);
 809       assert(rng->Opcode() == Op_LoadRange || _igvn.type(rng)->is_int() >= 0, "must be");
 810       assert(invar.is_invariant(rng), "range must be invariant");
 811       int scale    = 1;
 812       Node* offset = zero;
 813       bool ok = is_scaled_iv_plus_offset(idx, cl->phi(), &scale, &offset);
 814       assert(ok, "must be index expression");
 815 
 816       Node* init    = cl->init_trip();
 817       Node* limit   = cl->limit();
 818       Node* stride  = cl->stride();
 819 
 820       // Build if's for the upper and lower bound tests.  The
 821       // lower_bound test will dominate the upper bound test and all
 822       // cloned or created nodes will use the lower bound test as
 823       // their declared control.
 824       ProjNode* lower_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate, iff->Opcode());
 825       ProjNode* upper_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate, iff->Opcode());
 826       assert(upper_bound_proj->in(0)->as_If()->in(0) == lower_bound_proj, "should dominate");
 827       Node *ctrl = lower_bound_proj->in(0)->as_If()->in(0);
 828 
 829       // Perform cloning to keep Invariance state correct since the
 830       // late schedule will place invariant things in the loop.
 831       rng = invar.clone(rng, ctrl);
 832       if (offset && offset != zero) {
 833         assert(invar.is_invariant(offset), "offset must be loop invariant");
 834         offset = invar.clone(offset, ctrl);
 835       }
 836 
 837       // Test the lower bound
 838       BoolNode*  lower_bound_bol = rc_predicate(loop, ctrl, scale, offset, init, limit, stride, rng, false);
 839       // Negate test if necessary
 840       bool negated = false;
 841       if (proj->_con != predicate_proj->_con) {
 842         lower_bound_bol = new BoolNode(lower_bound_bol->in(1), lower_bound_bol->_test.negate());
 843         register_new_node(lower_bound_bol, ctrl);
 844         negated = true;
 845       }
 846       IfNode* lower_bound_iff = lower_bound_proj->in(0)->as_If();
 847       _igvn.hash_delete(lower_bound_iff);
 848       lower_bound_iff->set_req(1, lower_bound_bol);
 849       if (TraceLoopPredicate) tty->print_cr("lower bound check if: %s %d ", negated ? " negated" : "", lower_bound_iff->_idx);
 850 
 851       // Test the upper bound
 852       BoolNode* upper_bound_bol = rc_predicate(loop, lower_bound_proj, scale, offset, init, limit, stride, rng, true);
 853       negated = false;
 854       if (proj->_con != predicate_proj->_con) {
 855         upper_bound_bol = new BoolNode(upper_bound_bol->in(1), upper_bound_bol->_test.negate());
 856         register_new_node(upper_bound_bol, ctrl);
 857         negated = true;
 858       }
 859       IfNode* upper_bound_iff = upper_bound_proj->in(0)->as_If();
 860       _igvn.hash_delete(upper_bound_iff);
 861       upper_bound_iff->set_req(1, upper_bound_bol);
 862       if (TraceLoopPredicate) tty->print_cr("upper bound check if: %s %d ", negated ? " negated" : "", lower_bound_iff->_idx);
 863 
 864       // Fall through into rest of the clean up code which will move
 865       // any dependent nodes onto the upper bound test.
 866       new_predicate_proj = upper_bound_proj;
 867 
 868 #ifndef PRODUCT
 869       if (TraceLoopOpts && !TraceLoopPredicate) {
 870         tty->print("Predicate RC ");
 871         loop->dump_head();
 872       }
 873 #endif
 874     } else {
 875       // Loop variant check (for example, range check in non-counted loop)
 876       // with uncommon trap.
 877       continue;
 878     }
 879     assert(new_predicate_proj != NULL, "sanity");
 880     // Success - attach condition (new_predicate_bol) to predicate if
 881     invar.map_ctrl(proj, new_predicate_proj); // so that invariance test can be appropriate
 882 
 883     // Eliminate the old If in the loop body
 884     dominated_by( new_predicate_proj, iff, proj->_con != new_predicate_proj->_con );
 885 
 886     hoisted = true;
 887     C->set_major_progress();
 888   } // end while
 889 
 890 #ifndef PRODUCT
 891   // report that the loop predication has been actually performed
 892   // for this loop
 893   if (TraceLoopPredicate && hoisted) {
 894     tty->print("Loop Predication Performed:");
 895     loop->dump_head();
 896   }
 897 #endif
 898 
 899   return hoisted;
 900 }
 901 
 902 //------------------------------loop_predication--------------------------------
 903 // driver routine for loop predication optimization
 904 bool IdealLoopTree::loop_predication( PhaseIdealLoop *phase) {
 905   bool hoisted = false;
 906   // Recursively promote predicates
 907   if (_child) {
 908     hoisted = _child->loop_predication( phase);
 909   }
 910 
 911   // self
 912   if (!_irreducible && !tail()->is_top()) {
 913     hoisted |= phase->loop_predication_impl(this);
 914   }
 915 
 916   if (_next) { //sibling
 917     hoisted |= _next->loop_predication( phase);
 918   }
 919 
 920   return hoisted;
 921 }