1 /*
   2  * Copyright (c) 2011, 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/loopnode.hpp"
  31 #include "opto/mulnode.hpp"
  32 #include "opto/rootnode.hpp"
  33 #include "opto/subnode.hpp"
  34 
  35 /*
  36  * The general idea of Loop Predication is to insert a predicate on the entry
  37  * path to a loop, and raise a uncommon trap if the check of the condition fails.
  38  * The condition checks are promoted from inside the loop body, and thus
  39  * the checks inside the loop could be eliminated. Currently, loop predication
  40  * optimization has been applied to remove array range check and loop invariant
  41  * checks (such as null checks).
  42 */
  43 
  44 //-------------------------------is_uncommon_trap_proj----------------------------
  45 // Return true if proj is the form of "proj->[region->..]call_uct"
  46 bool PhaseIdealLoop::is_uncommon_trap_proj(ProjNode* proj, Deoptimization::DeoptReason reason) {
  47   int path_limit = 10;
  48   assert(proj, "invalid argument");
  49   Node* out = proj;
  50   for (int ct = 0; ct < path_limit; ct++) {
  51     out = out->unique_ctrl_out();
  52     if (out == NULL)
  53       return false;
  54     if (out->is_CallStaticJava()) {
  55       int req = out->as_CallStaticJava()->uncommon_trap_request();
  56       if (req != 0) {
  57         Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req);
  58         if (trap_reason == reason || reason == Deoptimization::Reason_none) {
  59            return true;
  60         }
  61       }
  62       return false; // don't do further after call
  63     }
  64     if (out->Opcode() != Op_Region)
  65       return false;
  66   }
  67   return false;
  68 }
  69 
  70 //-------------------------------is_uncommon_trap_if_pattern-------------------------
  71 // Return true  for "if(test)-> proj -> ...
  72 //                          |
  73 //                          V
  74 //                      other_proj->[region->..]call_uct"
  75 //
  76 // "must_reason_predicate" means the uct reason must be Reason_predicate
  77 bool PhaseIdealLoop::is_uncommon_trap_if_pattern(ProjNode *proj, Deoptimization::DeoptReason reason) {
  78   Node *in0 = proj->in(0);
  79   if (!in0->is_If()) return false;
  80   // Variation of a dead If node.
  81   if (in0->outcnt() < 2)  return false;
  82   IfNode* iff = in0->as_If();
  83 
  84   // we need "If(Conv2B(Opaque1(...)))" pattern for reason_predicate
  85   if (reason != Deoptimization::Reason_none) {
  86     if (iff->in(1)->Opcode() != Op_Conv2B ||
  87        iff->in(1)->in(1)->Opcode() != Op_Opaque1) {
  88       return false;
  89     }
  90   }
  91 
  92   ProjNode* other_proj = iff->proj_out(1-proj->_con)->as_Proj();
  93   if (is_uncommon_trap_proj(other_proj, reason)) {
  94     assert(reason == Deoptimization::Reason_none ||
  95            Compile::current()->is_predicate_opaq(iff->in(1)->in(1)), "should be on the list");
  96     return true;
  97   }
  98   return false;
  99 }
 100 
 101 //-------------------------------register_control-------------------------
 102 void PhaseIdealLoop::register_control(Node* n, IdealLoopTree *loop, Node* pred) {
 103   assert(n->is_CFG(), "must be control node");
 104   _igvn.register_new_node_with_optimizer(n);
 105   loop->_body.push(n);
 106   set_loop(n, loop);
 107   // When called from beautify_loops() idom is not constructed yet.
 108   if (_idom != NULL) {
 109     set_idom(n, pred, dom_depth(pred));
 110   }
 111 }
 112 
 113 //------------------------------create_new_if_for_predicate------------------------
 114 // create a new if above the uct_if_pattern for the predicate to be promoted.
 115 //
 116 //          before                                after
 117 //        ----------                           ----------
 118 //           ctrl                                 ctrl
 119 //            |                                     |
 120 //            |                                     |
 121 //            v                                     v
 122 //           iff                                 new_iff
 123 //          /    \                                /      \
 124 //         /      \                              /        \
 125 //        v        v                            v          v
 126 //  uncommon_proj cont_proj                   if_uct     if_cont
 127 // \      |        |                           |          |
 128 //  \     |        |                           |          |
 129 //   v    v        v                           |          v
 130 //     rgn       loop                          |         iff
 131 //      |                                      |        /     \
 132 //      |                                      |       /       \
 133 //      v                                      |      v         v
 134 // uncommon_trap                               | uncommon_proj cont_proj
 135 //                                           \  \    |           |
 136 //                                            \  \   |           |
 137 //                                             v  v  v           v
 138 //                                               rgn           loop
 139 //                                                |
 140 //                                                |
 141 //                                                v
 142 //                                           uncommon_trap
 143 //
 144 //
 145 // We will create a region to guard the uct call if there is no one there.
 146 // The true projecttion (if_cont) of the new_iff is returned.
 147 // This code is also used to clone predicates to clonned loops.
 148 ProjNode* PhaseIdealLoop::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
 149                                                       Deoptimization::DeoptReason reason) {
 150   assert(is_uncommon_trap_if_pattern(cont_proj, reason), "must be a uct if pattern!");
 151   IfNode* iff = cont_proj->in(0)->as_If();
 152 
 153   ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con);
 154   Node     *rgn   = uncommon_proj->unique_ctrl_out();
 155   assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
 156 
 157   uint proj_index = 1; // region's edge corresponding to uncommon_proj
 158   if (!rgn->is_Region()) { // create a region to guard the call
 159     assert(rgn->is_Call(), "must be call uct");
 160     CallNode* call = rgn->as_Call();
 161     IdealLoopTree* loop = get_loop(call);
 162     rgn = new (C, 1) RegionNode(1);
 163     rgn->add_req(uncommon_proj);
 164     register_control(rgn, loop, uncommon_proj);
 165     _igvn.hash_delete(call);
 166     call->set_req(0, rgn);
 167     // When called from beautify_loops() idom is not constructed yet.
 168     if (_idom != NULL) {
 169       set_idom(call, rgn, dom_depth(rgn));
 170     }
 171   } else {
 172     // Find region's edge corresponding to uncommon_proj
 173     for (; proj_index < rgn->req(); proj_index++)
 174       if (rgn->in(proj_index) == uncommon_proj) break;
 175     assert(proj_index < rgn->req(), "sanity");
 176   }
 177 
 178   Node* entry = iff->in(0);
 179   if (new_entry != NULL) {
 180     // Clonning the predicate to new location.
 181     entry = new_entry;
 182   }
 183   // Create new_iff
 184   IdealLoopTree* lp = get_loop(entry);
 185   IfNode *new_iff = iff->clone()->as_If();
 186   new_iff->set_req(0, entry);
 187   register_control(new_iff, lp, entry);
 188   Node *if_cont = new (C, 1) IfTrueNode(new_iff);
 189   Node *if_uct  = new (C, 1) IfFalseNode(new_iff);
 190   if (cont_proj->is_IfFalse()) {
 191     // Swap
 192     Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp;
 193   }
 194   register_control(if_cont, lp, new_iff);
 195   register_control(if_uct, get_loop(rgn), new_iff);
 196 
 197   // if_uct to rgn
 198   _igvn.hash_delete(rgn);
 199   rgn->add_req(if_uct);
 200   // When called from beautify_loops() idom is not constructed yet.
 201   if (_idom != NULL) {
 202     Node* ridom = idom(rgn);
 203     Node* nrdom = dom_lca(ridom, new_iff);
 204     set_idom(rgn, nrdom, dom_depth(rgn));
 205   }
 206 
 207   // If rgn has phis add new edges which has the same
 208   // value as on original uncommon_proj pass.
 209   assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last");
 210   bool has_phi = false;
 211   for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) {
 212     Node* use = rgn->fast_out(i);
 213     if (use->is_Phi() && use->outcnt() > 0) {
 214       assert(use->in(0) == rgn, "");
 215       _igvn.hash_delete(use);
 216       use->add_req(use->in(proj_index));
 217       _igvn._worklist.push(use);
 218       has_phi = true;
 219     }
 220   }
 221   assert(!has_phi || rgn->req() > 3, "no phis when region is created");
 222 
 223   if (new_entry == NULL) {
 224     // Attach if_cont to iff
 225     _igvn.hash_delete(iff);
 226     iff->set_req(0, if_cont);
 227     if (_idom != NULL) {
 228       set_idom(iff, if_cont, dom_depth(iff));
 229     }
 230   }
 231   return if_cont->as_Proj();
 232 }
 233 
 234 //------------------------------create_new_if_for_predicate------------------------
 235 // Create a new if below new_entry for the predicate to be cloned (IGVN optimization)
 236 ProjNode* PhaseIterGVN::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
 237                                                     Deoptimization::DeoptReason reason) {
 238   assert(new_entry != 0, "only used for clone predicate");
 239   assert(PhaseIdealLoop::is_uncommon_trap_if_pattern(cont_proj, reason), "must be a uct if pattern!");
 240   IfNode* iff = cont_proj->in(0)->as_If();
 241 
 242   ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con);
 243   Node     *rgn   = uncommon_proj->unique_ctrl_out();
 244   assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
 245 
 246   uint proj_index = 1; // region's edge corresponding to uncommon_proj
 247   if (!rgn->is_Region()) { // create a region to guard the call
 248     assert(rgn->is_Call(), "must be call uct");
 249     CallNode* call = rgn->as_Call();
 250     rgn = new (C, 1) RegionNode(1);
 251     register_new_node_with_optimizer(rgn);
 252     rgn->add_req(uncommon_proj);
 253     hash_delete(call);
 254     call->set_req(0, rgn);
 255   } else {
 256     // Find region's edge corresponding to uncommon_proj
 257     for (; proj_index < rgn->req(); proj_index++)
 258       if (rgn->in(proj_index) == uncommon_proj) break;
 259     assert(proj_index < rgn->req(), "sanity");
 260   }
 261 
 262   // Create new_iff in new location.
 263   IfNode *new_iff = iff->clone()->as_If();
 264   new_iff->set_req(0, new_entry);
 265 
 266   register_new_node_with_optimizer(new_iff);
 267   Node *if_cont = new (C, 1) IfTrueNode(new_iff);
 268   Node *if_uct  = new (C, 1) IfFalseNode(new_iff);
 269   if (cont_proj->is_IfFalse()) {
 270     // Swap
 271     Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp;
 272   }
 273   register_new_node_with_optimizer(if_cont);
 274   register_new_node_with_optimizer(if_uct);
 275 
 276   // if_uct to rgn
 277   hash_delete(rgn);
 278   rgn->add_req(if_uct);
 279 
 280   // If rgn has phis add corresponding new edges which has the same
 281   // value as on original uncommon_proj pass.
 282   assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last");
 283   bool has_phi = false;
 284   for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) {
 285     Node* use = rgn->fast_out(i);
 286     if (use->is_Phi() && use->outcnt() > 0) {
 287       hash_delete(use);
 288       use->add_req(use->in(proj_index));
 289       _worklist.push(use);
 290       has_phi = true;
 291     }
 292   }
 293   assert(!has_phi || rgn->req() > 3, "no phis when region is created");
 294 
 295   return if_cont->as_Proj();
 296 }
 297 
 298 //--------------------------clone_predicate-----------------------
 299 ProjNode* PhaseIdealLoop::clone_predicate(ProjNode* predicate_proj, Node* new_entry,
 300                                           Deoptimization::DeoptReason reason,
 301                                           PhaseIdealLoop* loop_phase,
 302                                           PhaseIterGVN* igvn) {
 303   ProjNode* new_predicate_proj;
 304   if (loop_phase != NULL) {
 305     new_predicate_proj = loop_phase->create_new_if_for_predicate(predicate_proj, new_entry, reason);
 306   } else {
 307     new_predicate_proj =       igvn->create_new_if_for_predicate(predicate_proj, new_entry, reason);
 308   }
 309   IfNode* iff = new_predicate_proj->in(0)->as_If();
 310   Node* ctrl  = iff->in(0);
 311 
 312   // Match original condition since predicate's projections could be swapped.
 313   assert(predicate_proj->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be");
 314   Node* opq = new (igvn->C, 2) Opaque1Node(igvn->C, predicate_proj->in(0)->in(1)->in(1)->in(1));
 315   igvn->C->add_predicate_opaq(opq);
 316 
 317   Node* bol = new (igvn->C, 2) Conv2BNode(opq);
 318   if (loop_phase != NULL) {
 319     loop_phase->register_new_node(opq, ctrl);
 320     loop_phase->register_new_node(bol, ctrl);
 321   } else {
 322     igvn->register_new_node_with_optimizer(opq);
 323     igvn->register_new_node_with_optimizer(bol);
 324   }
 325   igvn->hash_delete(iff);
 326   iff->set_req(1, bol);
 327   return new_predicate_proj;
 328 }
 329 
 330 //--------------------------move_predicate-----------------------
 331 // Cut predicate from old place and move it to new.
 332 ProjNode* PhaseIdealLoop::move_predicate(ProjNode* predicate_proj, Node* new_entry,
 333                                          Deoptimization::DeoptReason reason,
 334                                          PhaseIdealLoop* loop_phase,
 335                                          PhaseIterGVN* igvn) {
 336   assert(new_entry != NULL, "must be");
 337   assert(predicate_proj->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be");
 338   IfNode* iff = predicate_proj->in(0)->as_If();
 339   Node* old_entry = iff->in(0);
 340 
 341   // Cut predicate from old place.
 342   Node* old = predicate_proj;
 343   igvn->_worklist.push(old);
 344   for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin;) {
 345     Node* use = old->last_out(i);  // for each use...
 346     igvn->hash_delete(use);
 347     igvn->_worklist.push(use);
 348     // Update use-def info
 349     uint uses_found = 0;
 350     for (uint j = 0; j < use->req(); j++) {
 351       if (use->in(j) == old) {
 352         use->set_req(j, old_entry);
 353         uses_found++;
 354         if (loop_phase != NULL) {
 355           if (use->is_CFG()) {
 356             // When called from beautify_loops() idom is not constructed yet.
 357             if (loop_phase->_idom != NULL)
 358               loop_phase->set_idom(use, old_entry, loop_phase->dom_depth(use));
 359           } else {
 360             loop_phase->set_ctrl(use, old_entry);
 361           }
 362         }
 363       }
 364     }
 365     i -= uses_found;    // we deleted 1 or more copies of this edge
 366   }
 367 
 368   // Move predicate.
 369   igvn->hash_delete(iff);
 370   iff->set_req(0, new_entry);
 371   igvn->_worklist.push(iff);
 372 
 373   if (loop_phase != NULL) {
 374     // Fix up idom and ctrl.
 375     loop_phase->set_ctrl(iff->in(1), new_entry);
 376     loop_phase->set_ctrl(iff->in(1)->in(1), new_entry);
 377     // When called from beautify_loops() idom is not constructed yet.
 378     if (loop_phase->_idom != NULL)
 379       loop_phase->set_idom(iff, new_entry, loop_phase->dom_depth(iff));
 380   }
 381 
 382   return predicate_proj;
 383 }
 384 
 385 //--------------------------clone_loop_predicates-----------------------
 386 // Interface from IGVN
 387 Node* PhaseIterGVN::clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check) {
 388   return PhaseIdealLoop::clone_loop_predicates(old_entry, new_entry, false, clone_limit_check, NULL, this);
 389 }
 390 Node* PhaseIterGVN::move_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check) {
 391   return PhaseIdealLoop::clone_loop_predicates(old_entry, new_entry, true, clone_limit_check, NULL, this);
 392 }
 393 
 394 // Interface from PhaseIdealLoop
 395 Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check) {
 396   return clone_loop_predicates(old_entry, new_entry, false, clone_limit_check, this, &this->_igvn);
 397 }
 398 Node* PhaseIdealLoop::move_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check) {
 399   return clone_loop_predicates(old_entry, new_entry, true, clone_limit_check, this, &this->_igvn);
 400 }
 401 
 402 // Clone loop predicates to cloned loops (peeled, unswitched, split_if).
 403 Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry,
 404                                                 bool move_predicates,
 405                                                 bool clone_limit_check,
 406                                                 PhaseIdealLoop* loop_phase,
 407                                                 PhaseIterGVN* igvn) {
 408 #ifdef ASSERT
 409   if (new_entry == NULL || !(new_entry->is_Proj() || new_entry->is_Region() || new_entry->is_SafePoint())) {
 410     if (new_entry != NULL)
 411       new_entry->dump();
 412     assert(false, "not IfTrue, IfFalse, Region or SafePoint");
 413   }
 414 #endif
 415   // Search original predicates
 416   Node* entry = old_entry;
 417   ProjNode* limit_check_proj = NULL;
 418   if (LoopLimitCheck) {
 419     limit_check_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
 420     if (limit_check_proj != NULL) {
 421       entry = entry->in(0)->in(0);
 422     }
 423   }
 424   if (UseLoopPredicate) {
 425     ProjNode* predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
 426     if (predicate_proj != NULL) { // right pattern that can be used by loop predication
 427       if (move_predicates) {
 428         new_entry =  move_predicate(predicate_proj, new_entry,
 429                                     Deoptimization::Reason_predicate,
 430                                     loop_phase, igvn);
 431         assert(new_entry == predicate_proj, "old predicate fall through projection");
 432       } else {
 433         // clone predicate
 434         new_entry = clone_predicate(predicate_proj, new_entry,
 435                                     Deoptimization::Reason_predicate,
 436                                     loop_phase, igvn);
 437         assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone predicate");
 438       }
 439       if (TraceLoopPredicate) {
 440         tty->print_cr("Loop Predicate %s: ", move_predicates ? "moved" : "cloned");
 441         debug_only( new_entry->in(0)->dump(); )
 442       }
 443     }
 444   }
 445   if (limit_check_proj != NULL && clone_limit_check) {
 446     // Clone loop limit check last to insert it before loop.
 447     // Don't clone a limit check which was already finalized
 448     // for this counted loop (only one limit check is needed).
 449     if (move_predicates) {
 450       new_entry =  move_predicate(limit_check_proj, new_entry,
 451                                   Deoptimization::Reason_loop_limit_check,
 452                                   loop_phase, igvn);
 453       assert(new_entry == limit_check_proj, "old limit check fall through projection");
 454     } else {
 455       new_entry = clone_predicate(limit_check_proj, new_entry,
 456                                   Deoptimization::Reason_loop_limit_check,
 457                                   loop_phase, igvn);
 458       assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone limit check");
 459     }
 460     if (TraceLoopLimitCheck) {
 461       tty->print_cr("Loop Limit Check %s: ", move_predicates ? "moved" : "cloned");
 462       debug_only( new_entry->in(0)->dump(); )
 463     }
 464   }
 465   return new_entry;
 466 }
 467 
 468 //--------------------------eliminate_loop_predicates-----------------------
 469 void PhaseIdealLoop::eliminate_loop_predicates(Node* entry) {
 470   if (LoopLimitCheck) {
 471     Node* predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
 472     if (predicate != NULL) {
 473       entry = entry->in(0)->in(0);
 474     }
 475   }
 476   if (UseLoopPredicate) {
 477     ProjNode* predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
 478     if (predicate_proj != NULL) { // right pattern that can be used by loop predication
 479       Node* n = entry->in(0)->in(1)->in(1);
 480       assert(n->Opcode()==Op_Opaque1, "must be");
 481       // Remove Opaque1 node from predicates list.
 482       // IGVN will remove this predicate check.
 483       _igvn.replace_node(n, n->in(1));
 484     }
 485   }
 486 }
 487 
 488 //--------------------------skip_loop_predicates------------------------------
 489 // Skip related predicates.
 490 Node* PhaseIdealLoop::skip_loop_predicates(Node* entry) {
 491   Node* predicate = NULL;
 492   if (LoopLimitCheck) {
 493     predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
 494     if (predicate != NULL) {
 495       entry = entry->in(0)->in(0);
 496     }
 497   }
 498   if (UseLoopPredicate) {
 499     predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
 500     if (predicate != NULL) { // right pattern that can be used by loop predication
 501       IfNode* iff = entry->in(0)->as_If();
 502       ProjNode* uncommon_proj = iff->proj_out(1 - entry->as_Proj()->_con);
 503       Node* rgn = uncommon_proj->unique_ctrl_out();
 504       assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
 505       entry = entry->in(0)->in(0);
 506       while (entry != NULL && entry->is_Proj() && entry->in(0)->is_If()) {
 507         uncommon_proj = entry->in(0)->as_If()->proj_out(1 - entry->as_Proj()->_con);
 508         if (uncommon_proj->unique_ctrl_out() != rgn)
 509           break;
 510         entry = entry->in(0)->in(0);
 511       }
 512     }
 513   }
 514   return entry;
 515 }
 516 
 517 //--------------------------find_predicate_insertion_point-------------------
 518 // Find a good location to insert a predicate
 519 ProjNode* PhaseIdealLoop::find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason) {
 520   if (start_c == NULL || !start_c->is_Proj())
 521     return NULL;
 522   if (is_uncommon_trap_if_pattern(start_c->as_Proj(), reason)) {
 523     return start_c->as_Proj();
 524   }
 525   return NULL;
 526 }
 527 
 528 //--------------------------find_predicate------------------------------------
 529 // Find a predicate
 530 Node* PhaseIdealLoop::find_predicate(Node* entry) {
 531   Node* predicate = NULL;
 532   if (LoopLimitCheck) {
 533     predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
 534     if (predicate != NULL) { // right pattern that can be used by loop predication
 535       return entry;
 536     }
 537   }
 538   if (UseLoopPredicate) {
 539     predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
 540     if (predicate != NULL) { // right pattern that can be used by loop predication
 541       return entry;
 542     }
 543   }
 544   return NULL;
 545 }
 546 
 547 //------------------------------Invariance-----------------------------------
 548 // Helper class for loop_predication_impl to compute invariance on the fly and
 549 // clone invariants.
 550 class Invariance : public StackObj {
 551   VectorSet _visited, _invariant;
 552   Node_Stack _stack;
 553   VectorSet _clone_visited;
 554   Node_List _old_new; // map of old to new (clone)
 555   IdealLoopTree* _lpt;
 556   PhaseIdealLoop* _phase;
 557 
 558   // Helper function to set up the invariance for invariance computation
 559   // If n is a known invariant, set up directly. Otherwise, look up the
 560   // the possibility to push n onto the stack for further processing.
 561   void visit(Node* use, Node* n) {
 562     if (_lpt->is_invariant(n)) { // known invariant
 563       _invariant.set(n->_idx);
 564     } else if (!n->is_CFG()) {
 565       Node *n_ctrl = _phase->ctrl_or_self(n);
 566       Node *u_ctrl = _phase->ctrl_or_self(use); // self if use is a CFG
 567       if (_phase->is_dominator(n_ctrl, u_ctrl)) {
 568         _stack.push(n, n->in(0) == NULL ? 1 : 0);
 569       }
 570     }
 571   }
 572 
 573   // Compute invariance for "the_node" and (possibly) all its inputs recursively
 574   // on the fly
 575   void compute_invariance(Node* n) {
 576     assert(_visited.test(n->_idx), "must be");
 577     visit(n, n);
 578     while (_stack.is_nonempty()) {
 579       Node*  n = _stack.node();
 580       uint idx = _stack.index();
 581       if (idx == n->req()) { // all inputs are processed
 582         _stack.pop();
 583         // n is invariant if it's inputs are all invariant
 584         bool all_inputs_invariant = true;
 585         for (uint i = 0; i < n->req(); i++) {
 586           Node* in = n->in(i);
 587           if (in == NULL) continue;
 588           assert(_visited.test(in->_idx), "must have visited input");
 589           if (!_invariant.test(in->_idx)) { // bad guy
 590             all_inputs_invariant = false;
 591             break;
 592           }
 593         }
 594         if (all_inputs_invariant) {
 595           _invariant.set(n->_idx); // I am a invariant too
 596         }
 597       } else { // process next input
 598         _stack.set_index(idx + 1);
 599         Node* m = n->in(idx);
 600         if (m != NULL && !_visited.test_set(m->_idx)) {
 601           visit(n, m);
 602         }
 603       }
 604     }
 605   }
 606 
 607   // Helper function to set up _old_new map for clone_nodes.
 608   // If n is a known invariant, set up directly ("clone" of n == n).
 609   // Otherwise, push n onto the stack for real cloning.
 610   void clone_visit(Node* n) {
 611     assert(_invariant.test(n->_idx), "must be invariant");
 612     if (_lpt->is_invariant(n)) { // known invariant
 613       _old_new.map(n->_idx, n);
 614     } else { // to be cloned
 615       assert(!n->is_CFG(), "should not see CFG here");
 616       _stack.push(n, n->in(0) == NULL ? 1 : 0);
 617     }
 618   }
 619 
 620   // Clone "n" and (possibly) all its inputs recursively
 621   void clone_nodes(Node* n, Node* ctrl) {
 622     clone_visit(n);
 623     while (_stack.is_nonempty()) {
 624       Node*  n = _stack.node();
 625       uint idx = _stack.index();
 626       if (idx == n->req()) { // all inputs processed, clone n!
 627         _stack.pop();
 628         // clone invariant node
 629         Node* n_cl = n->clone();
 630         _old_new.map(n->_idx, n_cl);
 631         _phase->register_new_node(n_cl, ctrl);
 632         for (uint i = 0; i < n->req(); i++) {
 633           Node* in = n_cl->in(i);
 634           if (in == NULL) continue;
 635           n_cl->set_req(i, _old_new[in->_idx]);
 636         }
 637       } else { // process next input
 638         _stack.set_index(idx + 1);
 639         Node* m = n->in(idx);
 640         if (m != NULL && !_clone_visited.test_set(m->_idx)) {
 641           clone_visit(m); // visit the input
 642         }
 643       }
 644     }
 645   }
 646 
 647  public:
 648   Invariance(Arena* area, IdealLoopTree* lpt) :
 649     _lpt(lpt), _phase(lpt->_phase),
 650     _visited(area), _invariant(area), _stack(area, 10 /* guess */),
 651     _clone_visited(area), _old_new(area)
 652   {}
 653 
 654   // Map old to n for invariance computation and clone
 655   void map_ctrl(Node* old, Node* n) {
 656     assert(old->is_CFG() && n->is_CFG(), "must be");
 657     _old_new.map(old->_idx, n); // "clone" of old is n
 658     _invariant.set(old->_idx);  // old is invariant
 659     _clone_visited.set(old->_idx);
 660   }
 661 
 662   // Driver function to compute invariance
 663   bool is_invariant(Node* n) {
 664     if (!_visited.test_set(n->_idx))
 665       compute_invariance(n);
 666     return (_invariant.test(n->_idx) != 0);
 667   }
 668 
 669   // Driver function to clone invariant
 670   Node* clone(Node* n, Node* ctrl) {
 671     assert(ctrl->is_CFG(), "must be");
 672     assert(_invariant.test(n->_idx), "must be an invariant");
 673     if (!_clone_visited.test(n->_idx))
 674       clone_nodes(n, ctrl);
 675     return _old_new[n->_idx];
 676   }
 677 };
 678 
 679 //------------------------------is_range_check_if -----------------------------------
 680 // Returns true if the predicate of iff is in "scale*iv + offset u< load_range(ptr)" format
 681 // Note: this function is particularly designed for loop predication. We require load_range
 682 //       and offset to be loop invariant computed on the fly by "invar"
 683 bool IdealLoopTree::is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const {
 684   if (!is_loop_exit(iff)) {
 685     return false;
 686   }
 687   if (!iff->in(1)->is_Bool()) {
 688     return false;
 689   }
 690   const BoolNode *bol = iff->in(1)->as_Bool();
 691   if (bol->_test._test != BoolTest::lt) {
 692     return false;
 693   }
 694   if (!bol->in(1)->is_Cmp()) {
 695     return false;
 696   }
 697   const CmpNode *cmp = bol->in(1)->as_Cmp();
 698   if (cmp->Opcode() != Op_CmpU) {
 699     return false;
 700   }
 701   Node* range = cmp->in(2);
 702   if (range->Opcode() != Op_LoadRange) {
 703     const TypeInt* tint = phase->_igvn.type(range)->isa_int();
 704     if (tint == NULL || tint->empty() || tint->_lo < 0) {
 705       // Allow predication on positive values that aren't LoadRanges.
 706       // This allows optimization of loops where the length of the
 707       // array is a known value and doesn't need to be loaded back
 708       // from the array.
 709       return false;
 710     }
 711   }
 712   if (!invar.is_invariant(range)) {
 713     return false;
 714   }
 715   Node *iv     = _head->as_CountedLoop()->phi();
 716   int   scale  = 0;
 717   Node *offset = NULL;
 718   if (!phase->is_scaled_iv_plus_offset(cmp->in(1), iv, &scale, &offset)) {
 719     return false;
 720   }
 721   if (offset && !invar.is_invariant(offset)) { // offset must be invariant
 722     return false;
 723   }
 724   return true;
 725 }
 726 
 727 //------------------------------rc_predicate-----------------------------------
 728 // Create a range check predicate
 729 //
 730 // for (i = init; i < limit; i += stride) {
 731 //    a[scale*i+offset]
 732 // }
 733 //
 734 // Compute max(scale*i + offset) for init <= i < limit and build the predicate
 735 // as "max(scale*i + offset) u< a.length".
 736 //
 737 // There are two cases for max(scale*i + offset):
 738 // (1) stride*scale > 0
 739 //   max(scale*i + offset) = scale*(limit-stride) + offset
 740 // (2) stride*scale < 0
 741 //   max(scale*i + offset) = scale*init + offset
 742 BoolNode* PhaseIdealLoop::rc_predicate(IdealLoopTree *loop, Node* ctrl,
 743                                        int scale, Node* offset,
 744                                        Node* init, Node* limit, Node* stride,
 745                                        Node* range, bool upper) {
 746   stringStream* predString = NULL;
 747   if (TraceLoopPredicate) {
 748     predString = new stringStream();
 749     predString->print("rc_predicate ");
 750   }
 751 
 752   Node* max_idx_expr  = init;
 753   int stride_con = stride->get_int();
 754   if ((stride_con > 0) == (scale > 0) == upper) {
 755     if (LoopLimitCheck) {
 756       // With LoopLimitCheck limit is not exact.
 757       // Calculate exact limit here.
 758       // Note, counted loop's test is '<' or '>'.
 759       limit = exact_limit(loop);
 760       max_idx_expr = new (C, 3) SubINode(limit, stride);
 761       register_new_node(max_idx_expr, ctrl);
 762       if (TraceLoopPredicate) predString->print("(limit - stride) ");
 763     } else {
 764       max_idx_expr = new (C, 3) SubINode(limit, stride);
 765       register_new_node(max_idx_expr, ctrl);
 766       if (TraceLoopPredicate) predString->print("(limit - stride) ");
 767     }
 768   } else {
 769     if (TraceLoopPredicate) predString->print("init ");
 770   }
 771 
 772   if (scale != 1) {
 773     ConNode* con_scale = _igvn.intcon(scale);
 774     max_idx_expr = new (C, 3) MulINode(max_idx_expr, con_scale);
 775     register_new_node(max_idx_expr, ctrl);
 776     if (TraceLoopPredicate) predString->print("* %d ", scale);
 777   }
 778 
 779   if (offset && (!offset->is_Con() || offset->get_int() != 0)){
 780     max_idx_expr = new (C, 3) AddINode(max_idx_expr, offset);
 781     register_new_node(max_idx_expr, ctrl);
 782     if (TraceLoopPredicate)
 783       if (offset->is_Con()) predString->print("+ %d ", offset->get_int());
 784       else predString->print("+ offset ");
 785   }
 786 
 787   CmpUNode* cmp = new (C, 3) CmpUNode(max_idx_expr, range);
 788   register_new_node(cmp, ctrl);
 789   BoolNode* bol = new (C, 2) BoolNode(cmp, BoolTest::lt);
 790   register_new_node(bol, ctrl);
 791 
 792   if (TraceLoopPredicate) {
 793     predString->print_cr("<u range");
 794     tty->print(predString->as_string());
 795   }
 796   return bol;
 797 }
 798 
 799 //------------------------------ loop_predication_impl--------------------------
 800 // Insert loop predicates for null checks and range checks
 801 bool PhaseIdealLoop::loop_predication_impl(IdealLoopTree *loop) {
 802   if (!UseLoopPredicate) return false;
 803 
 804   if (!loop->_head->is_Loop()) {
 805     // Could be a simple region when irreducible loops are present.
 806     return false;
 807   }
 808   LoopNode* head = loop->_head->as_Loop();
 809 
 810   if (head->unique_ctrl_out()->Opcode() == Op_NeverBranch) {
 811     // do nothing for infinite loops
 812     return false;
 813   }
 814 
 815   CountedLoopNode *cl = NULL;
 816   if (head->is_CountedLoop()) {
 817     cl = head->as_CountedLoop();
 818     // do nothing for iteration-splitted loops
 819     if (!cl->is_normal_loop()) return false;
 820   }
 821 
 822   Node* entry = head->in(LoopNode::EntryControl);
 823   ProjNode *predicate_proj = NULL;
 824   // Loop limit check predicate should be near the loop.
 825   if (LoopLimitCheck) {
 826     predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
 827     if (predicate_proj != NULL)
 828       entry = predicate_proj->in(0)->in(0);
 829   }
 830 
 831   predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
 832   if (!predicate_proj) {
 833 #ifndef PRODUCT
 834     if (TraceLoopPredicate) {
 835       tty->print("missing predicate:");
 836       loop->dump_head();
 837       head->dump(1);
 838     }
 839 #endif
 840     return false;
 841   }
 842   ConNode* zero = _igvn.intcon(0);
 843   set_ctrl(zero, C->root());
 844 
 845   ResourceArea *area = Thread::current()->resource_area();
 846   Invariance invar(area, loop);
 847 
 848   // Create list of if-projs such that a newer proj dominates all older
 849   // projs in the list, and they all dominate loop->tail()
 850   Node_List if_proj_list(area);
 851   Node *current_proj = loop->tail(); //start from tail
 852   while (current_proj != head) {
 853     if (loop == get_loop(current_proj) && // still in the loop ?
 854         current_proj->is_Proj()        && // is a projection  ?
 855         current_proj->in(0)->Opcode() == Op_If) { // is a if projection ?
 856       if_proj_list.push(current_proj);
 857     }
 858     current_proj = idom(current_proj);
 859   }
 860 
 861   bool hoisted = false; // true if at least one proj is promoted
 862   while (if_proj_list.size() > 0) {
 863     // Following are changed to nonnull when a predicate can be hoisted
 864     ProjNode* new_predicate_proj = NULL;
 865 
 866     ProjNode* proj = if_proj_list.pop()->as_Proj();
 867     IfNode*   iff  = proj->in(0)->as_If();
 868 
 869     if (!is_uncommon_trap_if_pattern(proj, Deoptimization::Reason_none)) {
 870       if (loop->is_loop_exit(iff)) {
 871         // stop processing the remaining projs in the list because the execution of them
 872         // depends on the condition of "iff" (iff->in(1)).
 873         break;
 874       } else {
 875         // Both arms are inside the loop. There are two cases:
 876         // (1) there is one backward branch. In this case, any remaining proj
 877         //     in the if_proj list post-dominates "iff". So, the condition of "iff"
 878         //     does not determine the execution the remining projs directly, and we
 879         //     can safely continue.
 880         // (2) both arms are forwarded, i.e. a diamond shape. In this case, "proj"
 881         //     does not dominate loop->tail(), so it can not be in the if_proj list.
 882         continue;
 883       }
 884     }
 885 
 886     Node*     test = iff->in(1);
 887     if (!test->is_Bool()){ //Conv2B, ...
 888       continue;
 889     }
 890     BoolNode* bol = test->as_Bool();
 891     if (invar.is_invariant(bol)) {
 892       // Invariant test
 893       new_predicate_proj = create_new_if_for_predicate(predicate_proj, NULL,
 894                                                        Deoptimization::Reason_predicate);
 895       Node* ctrl = new_predicate_proj->in(0)->as_If()->in(0);
 896       BoolNode* new_predicate_bol = invar.clone(bol, ctrl)->as_Bool();
 897 
 898       // Negate test if necessary
 899       bool negated = false;
 900       if (proj->_con != predicate_proj->_con) {
 901         new_predicate_bol = new (C, 2) BoolNode(new_predicate_bol->in(1), new_predicate_bol->_test.negate());
 902         register_new_node(new_predicate_bol, ctrl);
 903         negated = true;
 904       }
 905       IfNode* new_predicate_iff = new_predicate_proj->in(0)->as_If();
 906       _igvn.hash_delete(new_predicate_iff);
 907       new_predicate_iff->set_req(1, new_predicate_bol);
 908 #ifndef PRODUCT
 909       if (TraceLoopPredicate) {
 910         tty->print("Predicate invariant if%s: %d ", negated ? " negated" : "", new_predicate_iff->_idx);
 911         loop->dump_head();
 912       } else if (TraceLoopOpts) {
 913         tty->print("Predicate IC ");
 914         loop->dump_head();
 915       }
 916 #endif
 917     } else if (cl != NULL && loop->is_range_check_if(iff, this, invar)) {
 918       assert(proj->_con == predicate_proj->_con, "must match");
 919 
 920       // Range check for counted loops
 921       const Node*    cmp    = bol->in(1)->as_Cmp();
 922       Node*          idx    = cmp->in(1);
 923       assert(!invar.is_invariant(idx), "index is variant");
 924       Node* rng = cmp->in(2);
 925       assert(rng->Opcode() == Op_LoadRange || _igvn.type(rng)->is_int() >= 0, "must be");
 926       assert(invar.is_invariant(rng), "range must be invariant");
 927       int scale    = 1;
 928       Node* offset = zero;
 929       bool ok = is_scaled_iv_plus_offset(idx, cl->phi(), &scale, &offset);
 930       assert(ok, "must be index expression");
 931 
 932       Node* init    = cl->init_trip();
 933       Node* limit   = cl->limit();
 934       Node* stride  = cl->stride();
 935 
 936       // Build if's for the upper and lower bound tests.  The
 937       // lower_bound test will dominate the upper bound test and all
 938       // cloned or created nodes will use the lower bound test as
 939       // their declared control.
 940       ProjNode* lower_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate);
 941       ProjNode* upper_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate);
 942       assert(upper_bound_proj->in(0)->as_If()->in(0) == lower_bound_proj, "should dominate");
 943       Node *ctrl = lower_bound_proj->in(0)->as_If()->in(0);
 944 
 945       // Perform cloning to keep Invariance state correct since the
 946       // late schedule will place invariant things in the loop.
 947       rng = invar.clone(rng, ctrl);
 948       if (offset && offset != zero) {
 949         assert(invar.is_invariant(offset), "offset must be loop invariant");
 950         offset = invar.clone(offset, ctrl);
 951       }
 952 
 953       // Test the lower bound
 954       Node*  lower_bound_bol = rc_predicate(loop, ctrl, scale, offset, init, limit, stride, rng, false);
 955       IfNode* lower_bound_iff = lower_bound_proj->in(0)->as_If();
 956       _igvn.hash_delete(lower_bound_iff);
 957       lower_bound_iff->set_req(1, lower_bound_bol);
 958       if (TraceLoopPredicate) tty->print_cr("lower bound check if: %d", lower_bound_iff->_idx);
 959 
 960       // Test the upper bound
 961       Node* upper_bound_bol = rc_predicate(loop, ctrl, scale, offset, init, limit, stride, rng, true);
 962       IfNode* upper_bound_iff = upper_bound_proj->in(0)->as_If();
 963       _igvn.hash_delete(upper_bound_iff);
 964       upper_bound_iff->set_req(1, upper_bound_bol);
 965       if (TraceLoopPredicate) tty->print_cr("upper bound check if: %d", lower_bound_iff->_idx);
 966 
 967       // Fall through into rest of the clean up code which will move
 968       // any dependent nodes onto the upper bound test.
 969       new_predicate_proj = upper_bound_proj;
 970 
 971 #ifndef PRODUCT
 972       if (TraceLoopOpts && !TraceLoopPredicate) {
 973         tty->print("Predicate RC ");
 974         loop->dump_head();
 975       }
 976 #endif
 977     } else {
 978       // Loop variant check (for example, range check in non-counted loop)
 979       // with uncommon trap.
 980       continue;
 981     }
 982     assert(new_predicate_proj != NULL, "sanity");
 983     // Success - attach condition (new_predicate_bol) to predicate if
 984     invar.map_ctrl(proj, new_predicate_proj); // so that invariance test can be appropriate
 985 
 986     // Eliminate the old If in the loop body
 987     dominated_by( new_predicate_proj, iff, proj->_con != new_predicate_proj->_con );
 988 
 989     hoisted = true;
 990     C->set_major_progress();
 991   } // end while
 992 
 993 #ifndef PRODUCT
 994   // report that the loop predication has been actually performed
 995   // for this loop
 996   if (TraceLoopPredicate && hoisted) {
 997     tty->print("Loop Predication Performed:");
 998     loop->dump_head();
 999   }
1000 #endif
1001 
1002   return hoisted;
1003 }
1004 
1005 //------------------------------loop_predication--------------------------------
1006 // driver routine for loop predication optimization
1007 bool IdealLoopTree::loop_predication( PhaseIdealLoop *phase) {
1008   bool hoisted = false;
1009   // Recursively promote predicates
1010   if (_child) {
1011     hoisted = _child->loop_predication( phase);
1012   }
1013 
1014   // self
1015   if (!_irreducible && !tail()->is_top()) {
1016     hoisted |= phase->loop_predication_impl(this);
1017   }
1018 
1019   if (_next) { //sibling
1020     hoisted |= _next->loop_predication( phase);
1021   }
1022 
1023   return hoisted;
1024 }