1 /*
   2  * Copyright (c) 2011, 2018, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "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/matcher.hpp"
  33 #include "opto/mulnode.hpp"
  34 #include "opto/opaquenode.hpp"
  35 #include "opto/rootnode.hpp"
  36 #include "opto/subnode.hpp"
  37 #include <fenv.h>
  38 #include <math.h>
  39 
  40 /*
  41  * The general idea of Loop Predication is to insert a predicate on the entry
  42  * path to a loop, and raise a uncommon trap if the check of the condition fails.
  43  * The condition checks are promoted from inside the loop body, and thus
  44  * the checks inside the loop could be eliminated. Currently, loop predication
  45  * optimization has been applied to remove array range check and loop invariant
  46  * checks (such as null checks).
  47 */
  48 
  49 //-------------------------------register_control-------------------------
  50 void PhaseIdealLoop::register_control(Node* n, IdealLoopTree *loop, Node* pred) {
  51   assert(n->is_CFG(), "must be control node");
  52   _igvn.register_new_node_with_optimizer(n);
  53   loop->_body.push(n);
  54   set_loop(n, loop);
  55   // When called from beautify_loops() idom is not constructed yet.
  56   if (_idom != NULL) {
  57     set_idom(n, pred, dom_depth(pred));
  58   }
  59 }
  60 
  61 //------------------------------create_new_if_for_predicate------------------------
  62 // create a new if above the uct_if_pattern for the predicate to be promoted.
  63 //
  64 //          before                                after
  65 //        ----------                           ----------
  66 //           ctrl                                 ctrl
  67 //            |                                     |
  68 //            |                                     |
  69 //            v                                     v
  70 //           iff                                 new_iff
  71 //          /    \                                /      \
  72 //         /      \                              /        \
  73 //        v        v                            v          v
  74 //  uncommon_proj cont_proj                   if_uct     if_cont
  75 // \      |        |                           |          |
  76 //  \     |        |                           |          |
  77 //   v    v        v                           |          v
  78 //     rgn       loop                          |         iff
  79 //      |                                      |        /     \
  80 //      |                                      |       /       \
  81 //      v                                      |      v         v
  82 // uncommon_trap                               | uncommon_proj cont_proj
  83 //                                           \  \    |           |
  84 //                                            \  \   |           |
  85 //                                             v  v  v           v
  86 //                                               rgn           loop
  87 //                                                |
  88 //                                                |
  89 //                                                v
  90 //                                           uncommon_trap
  91 //
  92 //
  93 // We will create a region to guard the uct call if there is no one there.
  94 // The true projection (if_cont) of the new_iff is returned.
  95 // This code is also used to clone predicates to cloned loops.
  96 ProjNode* PhaseIdealLoop::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
  97                                                       Deoptimization::DeoptReason reason,
  98                                                       int opcode) {
  99   assert(cont_proj->is_uncommon_trap_if_pattern(reason), "must be a uct if pattern!");
 100   IfNode* iff = cont_proj->in(0)->as_If();
 101 
 102   ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con);
 103   Node     *rgn   = uncommon_proj->unique_ctrl_out();
 104   assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
 105 
 106   uint proj_index = 1; // region's edge corresponding to uncommon_proj
 107   if (!rgn->is_Region()) { // create a region to guard the call
 108     assert(rgn->is_Call(), "must be call uct");
 109     CallNode* call = rgn->as_Call();
 110     IdealLoopTree* loop = get_loop(call);
 111     rgn = new RegionNode(1);
 112     rgn->add_req(uncommon_proj);
 113     register_control(rgn, loop, uncommon_proj);
 114     _igvn.replace_input_of(call, 0, rgn);
 115     // When called from beautify_loops() idom is not constructed yet.
 116     if (_idom != NULL) {
 117       set_idom(call, rgn, dom_depth(rgn));
 118     }
 119     for (DUIterator_Fast imax, i = uncommon_proj->fast_outs(imax); i < imax; i++) {
 120       Node* n = uncommon_proj->fast_out(i);
 121       if (n->is_Load() || n->is_Store()) {
 122         _igvn.replace_input_of(n, 0, rgn);
 123         --i; --imax;
 124       }
 125     }
 126   } else {
 127     // Find region's edge corresponding to uncommon_proj
 128     for (; proj_index < rgn->req(); proj_index++)
 129       if (rgn->in(proj_index) == uncommon_proj) break;
 130     assert(proj_index < rgn->req(), "sanity");
 131   }
 132 
 133   Node* entry = iff->in(0);
 134   if (new_entry != NULL) {
 135     // Clonning the predicate to new location.
 136     entry = new_entry;
 137   }
 138   // Create new_iff
 139   IdealLoopTree* lp = get_loop(entry);
 140   IfNode* new_iff = NULL;
 141   if (opcode == Op_If) {
 142     new_iff = new IfNode(entry, iff->in(1), iff->_prob, iff->_fcnt);
 143   } else {
 144     assert(opcode == Op_RangeCheck, "no other if variant here");
 145     new_iff = new RangeCheckNode(entry, iff->in(1), iff->_prob, iff->_fcnt);
 146   }
 147   register_control(new_iff, lp, entry);
 148   Node *if_cont = new IfTrueNode(new_iff);
 149   Node *if_uct  = new IfFalseNode(new_iff);
 150   if (cont_proj->is_IfFalse()) {
 151     // Swap
 152     Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp;
 153   }
 154   register_control(if_cont, lp, new_iff);
 155   register_control(if_uct, get_loop(rgn), new_iff);
 156 
 157   // if_uct to rgn
 158   _igvn.hash_delete(rgn);
 159   rgn->add_req(if_uct);
 160   // When called from beautify_loops() idom is not constructed yet.
 161   if (_idom != NULL) {
 162     Node* ridom = idom(rgn);
 163     Node* nrdom = dom_lca(ridom, new_iff);
 164     set_idom(rgn, nrdom, dom_depth(rgn));
 165   }
 166 
 167   // If rgn has phis add new edges which has the same
 168   // value as on original uncommon_proj pass.
 169   assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last");
 170   bool has_phi = false;
 171   for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) {
 172     Node* use = rgn->fast_out(i);
 173     if (use->is_Phi() && use->outcnt() > 0) {
 174       assert(use->in(0) == rgn, "");
 175       _igvn.rehash_node_delayed(use);
 176       use->add_req(use->in(proj_index));
 177       has_phi = true;
 178     }
 179   }
 180   assert(!has_phi || rgn->req() > 3, "no phis when region is created");
 181 
 182   if (new_entry == NULL) {
 183     // Attach if_cont to iff
 184     _igvn.replace_input_of(iff, 0, if_cont);
 185     if (_idom != NULL) {
 186       set_idom(iff, if_cont, dom_depth(iff));
 187     }
 188   }
 189   return if_cont->as_Proj();
 190 }
 191 
 192 //------------------------------create_new_if_for_predicate------------------------
 193 // Create a new if below new_entry for the predicate to be cloned (IGVN optimization)
 194 ProjNode* PhaseIterGVN::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
 195                                                     Deoptimization::DeoptReason reason,
 196                                                     int opcode) {
 197   assert(new_entry != 0, "only used for clone predicate");
 198   assert(cont_proj->is_uncommon_trap_if_pattern(reason), "must be a uct if pattern!");
 199   IfNode* iff = cont_proj->in(0)->as_If();
 200 
 201   ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con);
 202   Node     *rgn   = uncommon_proj->unique_ctrl_out();
 203   assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
 204 
 205   uint proj_index = 1; // region's edge corresponding to uncommon_proj
 206   if (!rgn->is_Region()) { // create a region to guard the call
 207     assert(rgn->is_Call(), "must be call uct");
 208     CallNode* call = rgn->as_Call();
 209     rgn = new RegionNode(1);
 210     register_new_node_with_optimizer(rgn);
 211     rgn->add_req(uncommon_proj);
 212     replace_input_of(call, 0, rgn);
 213   } else {
 214     // Find region's edge corresponding to uncommon_proj
 215     for (; proj_index < rgn->req(); proj_index++)
 216       if (rgn->in(proj_index) == uncommon_proj) break;
 217     assert(proj_index < rgn->req(), "sanity");
 218   }
 219 
 220   // Create new_iff in new location.
 221   IfNode* new_iff = NULL;
 222   if (opcode == Op_If) {
 223     new_iff = new IfNode(new_entry, iff->in(1), iff->_prob, iff->_fcnt);
 224   } else {
 225     assert(opcode == Op_RangeCheck, "no other if variant here");
 226     new_iff = new RangeCheckNode(new_entry, iff->in(1), iff->_prob, iff->_fcnt);
 227   }
 228 
 229   register_new_node_with_optimizer(new_iff);
 230   Node *if_cont = new IfTrueNode(new_iff);
 231   Node *if_uct  = new IfFalseNode(new_iff);
 232   if (cont_proj->is_IfFalse()) {
 233     // Swap
 234     Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp;
 235   }
 236   register_new_node_with_optimizer(if_cont);
 237   register_new_node_with_optimizer(if_uct);
 238 
 239   // if_uct to rgn
 240   hash_delete(rgn);
 241   rgn->add_req(if_uct);
 242 
 243   // If rgn has phis add corresponding new edges which has the same
 244   // value as on original uncommon_proj pass.
 245   assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last");
 246   bool has_phi = false;
 247   for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) {
 248     Node* use = rgn->fast_out(i);
 249     if (use->is_Phi() && use->outcnt() > 0) {
 250       rehash_node_delayed(use);
 251       use->add_req(use->in(proj_index));
 252       has_phi = true;
 253     }
 254   }
 255   assert(!has_phi || rgn->req() > 3, "no phis when region is created");
 256 
 257   return if_cont->as_Proj();
 258 }
 259 
 260 //--------------------------clone_predicate-----------------------
 261 ProjNode* PhaseIdealLoop::clone_predicate(ProjNode* predicate_proj, Node* new_entry,
 262                                           Deoptimization::DeoptReason reason,
 263                                           PhaseIdealLoop* loop_phase,
 264                                           PhaseIterGVN* igvn) {
 265   ProjNode* new_predicate_proj;
 266   if (loop_phase != NULL) {
 267     new_predicate_proj = loop_phase->create_new_if_for_predicate(predicate_proj, new_entry, reason, Op_If);
 268   } else {
 269     new_predicate_proj =       igvn->create_new_if_for_predicate(predicate_proj, new_entry, reason, Op_If);
 270   }
 271   IfNode* iff = new_predicate_proj->in(0)->as_If();
 272   Node* ctrl  = iff->in(0);
 273 
 274   // Match original condition since predicate's projections could be swapped.
 275   assert(predicate_proj->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be");
 276   Node* opq = new Opaque1Node(igvn->C, predicate_proj->in(0)->in(1)->in(1)->in(1));
 277   igvn->C->add_predicate_opaq(opq);
 278 
 279   Node* bol = new Conv2BNode(opq);
 280   if (loop_phase != NULL) {
 281     loop_phase->register_new_node(opq, ctrl);
 282     loop_phase->register_new_node(bol, ctrl);
 283   } else {
 284     igvn->register_new_node_with_optimizer(opq);
 285     igvn->register_new_node_with_optimizer(bol);
 286   }
 287   igvn->hash_delete(iff);
 288   iff->set_req(1, bol);
 289   return new_predicate_proj;
 290 }
 291 
 292 
 293 //--------------------------clone_loop_predicates-----------------------
 294 // Interface from IGVN
 295 Node* PhaseIterGVN::clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check) {
 296   return PhaseIdealLoop::clone_loop_predicates(old_entry, new_entry, clone_limit_check, NULL, this);
 297 }
 298 
 299 // Interface from PhaseIdealLoop
 300 Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check) {
 301   return clone_loop_predicates(old_entry, new_entry, clone_limit_check, this, &this->_igvn);
 302 }
 303 
 304 void PhaseIdealLoop::clone_loop_predicates_fix_mem(ProjNode* dom_proj , ProjNode* proj,
 305                                                    PhaseIdealLoop* loop_phase,
 306                                                    PhaseIterGVN* igvn) {
 307   Compile* C = NULL;
 308   if (loop_phase != NULL) {
 309     igvn = &loop_phase->igvn();
 310   }
 311   C = igvn->C;
 312   ProjNode* other_dom_proj = dom_proj->in(0)->as_Multi()->proj_out(1-dom_proj->_con);
 313   Node* dom_r = other_dom_proj->unique_ctrl_out();
 314   if (dom_r->is_Region()) {
 315     assert(dom_r->unique_ctrl_out()->is_Call(), "unc expected");
 316     ProjNode* other_proj = proj->in(0)->as_Multi()->proj_out(1-proj->_con);
 317     Node* r = other_proj->unique_ctrl_out();
 318     assert(r->is_Region() && r->unique_ctrl_out()->is_Call(), "cloned predicate should have caused region to be added");
 319     for (DUIterator_Fast imax, i = dom_r->fast_outs(imax); i < imax; i++) {
 320       Node* dom_use = dom_r->fast_out(i);
 321       if (dom_use->is_Phi() && dom_use->bottom_type() == Type::MEMORY) {
 322         assert(dom_use->in(0) == dom_r, "");
 323         Node* phi = NULL;
 324         for (DUIterator_Fast jmax, j = r->fast_outs(jmax); j < jmax; j++) {
 325           Node* use = r->fast_out(j);
 326           if (use->is_Phi() && use->bottom_type() == Type::MEMORY &&
 327               use->adr_type() == dom_use->adr_type()) {
 328             assert(use->in(0) == r, "");
 329             assert(phi == NULL, "only one phi");
 330             phi = use;
 331           }
 332         }
 333         if (phi == NULL) {
 334           const TypePtr* adr_type = dom_use->adr_type();
 335           int alias = C->get_alias_index(adr_type);
 336           Node* call = r->unique_ctrl_out();
 337           Node* mem = call->in(TypeFunc::Memory);
 338           MergeMemNode* mm = NULL;
 339           if (mem->is_MergeMem()) {
 340             mm = mem->clone()->as_MergeMem();
 341             if (adr_type == TypePtr::BOTTOM) {
 342               mem = mem->as_MergeMem()->base_memory();
 343             } else {
 344               mem = mem->as_MergeMem()->memory_at(alias);
 345             }
 346           } else {
 347             mm = MergeMemNode::make(mem);
 348           }
 349           phi = PhiNode::make(r, mem, Type::MEMORY, adr_type);
 350           if (adr_type == TypePtr::BOTTOM) {
 351             mm->set_base_memory(phi);
 352           } else {
 353             mm->set_memory_at(alias, phi);
 354           }
 355           if (loop_phase != NULL) {
 356             loop_phase->register_new_node(mm, r);
 357             loop_phase->register_new_node(phi, r);
 358           } else {
 359             igvn->register_new_node_with_optimizer(mm);
 360             igvn->register_new_node_with_optimizer(phi);
 361           }
 362           igvn->replace_input_of(call, TypeFunc::Memory, mm);
 363         }
 364         igvn->replace_input_of(phi, r->find_edge(other_proj), dom_use->in(dom_r->find_edge(other_dom_proj)));
 365       }
 366     }
 367   }
 368 }
 369 
 370 
 371 // Clone loop predicates to cloned loops (peeled, unswitched, split_if).
 372 Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry,
 373                                                 bool clone_limit_check,
 374                                                 PhaseIdealLoop* loop_phase,
 375                                                 PhaseIterGVN* igvn) {
 376 #ifdef ASSERT
 377   if (new_entry == NULL || !(new_entry->is_Proj() || new_entry->is_Region() || new_entry->is_SafePoint())) {
 378     if (new_entry != NULL)
 379       new_entry->dump();
 380     assert(false, "not IfTrue, IfFalse, Region or SafePoint");
 381   }
 382 #endif
 383   // Search original predicates
 384   Node* entry = old_entry;
 385   ProjNode* limit_check_proj = NULL;
 386   limit_check_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
 387   if (limit_check_proj != NULL) {
 388     entry = entry->in(0)->in(0);
 389   }
 390   ProjNode* profile_predicate_proj = NULL;
 391   ProjNode* predicate_proj = NULL;
 392   if (UseProfiledLoopPredicate) {
 393     profile_predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate);
 394     if (profile_predicate_proj != NULL) {
 395       entry = skip_loop_predicates(entry);
 396     }
 397   }
 398   if (UseLoopPredicate) {
 399     predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
 400   }
 401   if (predicate_proj != NULL) { // right pattern that can be used by loop predication
 402     // clone predicate
 403     ProjNode* proj = clone_predicate(predicate_proj, new_entry,
 404                                      Deoptimization::Reason_predicate,
 405                                      loop_phase, igvn);
 406     assert(proj != NULL, "IfTrue or IfFalse after clone predicate");
 407     new_entry = proj;
 408     if (TraceLoopPredicate) {
 409       tty->print("Loop Predicate cloned: ");
 410       debug_only( new_entry->in(0)->dump(); );
 411     }
 412     if (profile_predicate_proj != NULL) {
 413       // A node that produces memory may be out of loop and depend on
 414       // a profiled predicates. In that case the memory state at the
 415       // end of profiled predicates and at the end of predicates are
 416       // not the same. The cloned predicates are dominated by the
 417       // profiled predicates but may have the wrong memory
 418       // state. Update it.
 419       clone_loop_predicates_fix_mem(profile_predicate_proj, proj, loop_phase, igvn);
 420     }
 421   }
 422   if (profile_predicate_proj != NULL) { // right pattern that can be used by loop predication
 423     // clone predicate
 424     new_entry = clone_predicate(profile_predicate_proj, new_entry,
 425                                 Deoptimization::Reason_profile_predicate,
 426                                 loop_phase, igvn);
 427     assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone predicate");
 428     if (TraceLoopPredicate) {
 429       tty->print("Loop Predicate cloned: ");
 430       debug_only( new_entry->in(0)->dump(); );
 431     }
 432   }
 433   if (limit_check_proj != NULL && clone_limit_check) {
 434     // Clone loop limit check last to insert it before loop.
 435     // Don't clone a limit check which was already finalized
 436     // for this counted loop (only one limit check is needed).
 437     new_entry = clone_predicate(limit_check_proj, new_entry,
 438                                 Deoptimization::Reason_loop_limit_check,
 439                                 loop_phase, igvn);
 440     assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone limit check");
 441     if (TraceLoopLimitCheck) {
 442       tty->print("Loop Limit Check cloned: ");
 443       debug_only( new_entry->in(0)->dump(); )
 444     }
 445   }
 446   return new_entry;
 447 }
 448 
 449 //--------------------------skip_loop_predicates------------------------------
 450 // Skip related predicates.
 451 Node* PhaseIdealLoop::skip_loop_predicates(Node* entry) {
 452   IfNode* iff = entry->in(0)->as_If();
 453   ProjNode* uncommon_proj = iff->proj_out(1 - entry->as_Proj()->_con);
 454   Node* rgn = uncommon_proj->unique_ctrl_out();
 455   assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
 456   entry = entry->in(0)->in(0);
 457   while (entry != NULL && entry->is_Proj() && entry->in(0)->is_If()) {
 458     uncommon_proj = entry->in(0)->as_If()->proj_out(1 - entry->as_Proj()->_con);
 459     if (uncommon_proj->unique_ctrl_out() != rgn)
 460       break;
 461     entry = entry->in(0)->in(0);
 462   }
 463   return entry;
 464 }
 465 
 466 Node* PhaseIdealLoop::skip_all_loop_predicates(Node* entry) {
 467   Node* predicate = NULL;
 468   predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
 469   if (predicate != NULL) {
 470     entry = entry->in(0)->in(0);
 471   }
 472   if (UseProfiledLoopPredicate) {
 473     predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate);
 474     if (predicate != NULL) { // right pattern that can be used by loop predication
 475       entry = skip_loop_predicates(entry);
 476     }
 477   }
 478   if (UseLoopPredicate) {
 479     predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
 480     if (predicate != NULL) { // right pattern that can be used by loop predication
 481       entry = skip_loop_predicates(entry);
 482     }
 483   }
 484   return entry;
 485 }
 486 
 487 //--------------------------find_predicate_insertion_point-------------------
 488 // Find a good location to insert a predicate
 489 ProjNode* PhaseIdealLoop::find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason) {
 490   if (start_c == NULL || !start_c->is_Proj())
 491     return NULL;
 492   if (start_c->as_Proj()->is_uncommon_trap_if_pattern(reason)) {
 493     return start_c->as_Proj();
 494   }
 495   return NULL;
 496 }
 497 
 498 //--------------------------find_predicate------------------------------------
 499 // Find a predicate
 500 Node* PhaseIdealLoop::find_predicate(Node* entry) {
 501   Node* predicate = NULL;
 502   predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
 503   if (predicate != NULL) { // right pattern that can be used by loop predication
 504     return entry;
 505   }
 506   if (UseLoopPredicate) {
 507     predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
 508     if (predicate != NULL) { // right pattern that can be used by loop predication
 509       return entry;
 510     }
 511   }
 512   if (UseProfiledLoopPredicate) {
 513     predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate);
 514     if (predicate != NULL) { // right pattern that can be used by loop predication
 515       return entry;
 516     }
 517   }
 518   return NULL;
 519 }
 520 
 521 //------------------------------Invariance-----------------------------------
 522 // Helper class for loop_predication_impl to compute invariance on the fly and
 523 // clone invariants.
 524 class Invariance : public StackObj {
 525   VectorSet _visited, _invariant;
 526   Node_Stack _stack;
 527   VectorSet _clone_visited;
 528   Node_List _old_new; // map of old to new (clone)
 529   IdealLoopTree* _lpt;
 530   PhaseIdealLoop* _phase;
 531 
 532   // Helper function to set up the invariance for invariance computation
 533   // If n is a known invariant, set up directly. Otherwise, look up the
 534   // the possibility to push n onto the stack for further processing.
 535   void visit(Node* use, Node* n) {
 536     if (_lpt->is_invariant(n)) { // known invariant
 537       _invariant.set(n->_idx);
 538     } else if (!n->is_CFG()) {
 539       Node *n_ctrl = _phase->ctrl_or_self(n);
 540       Node *u_ctrl = _phase->ctrl_or_self(use); // self if use is a CFG
 541       if (_phase->is_dominator(n_ctrl, u_ctrl)) {
 542         _stack.push(n, n->in(0) == NULL ? 1 : 0);
 543       }
 544     }
 545   }
 546 
 547   // Compute invariance for "the_node" and (possibly) all its inputs recursively
 548   // on the fly
 549   void compute_invariance(Node* n) {
 550     assert(_visited.test(n->_idx), "must be");
 551     visit(n, n);
 552     while (_stack.is_nonempty()) {
 553       Node*  n = _stack.node();
 554       uint idx = _stack.index();
 555       if (idx == n->req()) { // all inputs are processed
 556         _stack.pop();
 557         // n is invariant if it's inputs are all invariant
 558         bool all_inputs_invariant = true;
 559         for (uint i = 0; i < n->req(); i++) {
 560           Node* in = n->in(i);
 561           if (in == NULL) continue;
 562           assert(_visited.test(in->_idx), "must have visited input");
 563           if (!_invariant.test(in->_idx)) { // bad guy
 564             all_inputs_invariant = false;
 565             break;
 566           }
 567         }
 568         if (all_inputs_invariant) {
 569           // If n's control is a predicate that was moved out of the
 570           // loop, it was marked invariant but n is only invariant if
 571           // it depends only on that test. Otherwise, unless that test
 572           // is out of the loop, it's not invariant.
 573           if (n->is_CFG() || n->depends_only_on_test() || n->in(0) == NULL || !_phase->is_member(_lpt, n->in(0))) {
 574             _invariant.set(n->_idx); // I am a invariant too
 575           }
 576         }
 577       } else { // process next input
 578         _stack.set_index(idx + 1);
 579         Node* m = n->in(idx);
 580         if (m != NULL && !_visited.test_set(m->_idx)) {
 581           visit(n, m);
 582         }
 583       }
 584     }
 585   }
 586 
 587   // Helper function to set up _old_new map for clone_nodes.
 588   // If n is a known invariant, set up directly ("clone" of n == n).
 589   // Otherwise, push n onto the stack for real cloning.
 590   void clone_visit(Node* n) {
 591     assert(_invariant.test(n->_idx), "must be invariant");
 592     if (_lpt->is_invariant(n)) { // known invariant
 593       _old_new.map(n->_idx, n);
 594     } else { // to be cloned
 595       assert(!n->is_CFG(), "should not see CFG here");
 596       _stack.push(n, n->in(0) == NULL ? 1 : 0);
 597     }
 598   }
 599 
 600   // Clone "n" and (possibly) all its inputs recursively
 601   void clone_nodes(Node* n, Node* ctrl) {
 602     clone_visit(n);
 603     while (_stack.is_nonempty()) {
 604       Node*  n = _stack.node();
 605       uint idx = _stack.index();
 606       if (idx == n->req()) { // all inputs processed, clone n!
 607         _stack.pop();
 608         // clone invariant node
 609         Node* n_cl = n->clone();
 610         _old_new.map(n->_idx, n_cl);
 611         _phase->register_new_node(n_cl, ctrl);
 612         for (uint i = 0; i < n->req(); i++) {
 613           Node* in = n_cl->in(i);
 614           if (in == NULL) continue;
 615           n_cl->set_req(i, _old_new[in->_idx]);
 616         }
 617       } else { // process next input
 618         _stack.set_index(idx + 1);
 619         Node* m = n->in(idx);
 620         if (m != NULL && !_clone_visited.test_set(m->_idx)) {
 621           clone_visit(m); // visit the input
 622         }
 623       }
 624     }
 625   }
 626 
 627  public:
 628   Invariance(Arena* area, IdealLoopTree* lpt) :
 629     _visited(area), _invariant(area),
 630     _stack(area, 10 /* guess */),
 631     _clone_visited(area), _old_new(area),
 632     _lpt(lpt), _phase(lpt->_phase)
 633   {
 634     LoopNode* head = _lpt->_head->as_Loop();
 635     Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl);
 636     if (entry->outcnt() != 1) {
 637       // If a node is pinned between the predicates and the loop
 638       // entry, we won't be able to move any node in the loop that
 639       // depends on it above it in a predicate. Mark all those nodes
 640       // as non loop invariatnt.
 641       Unique_Node_List wq;
 642       wq.push(entry);
 643       for (uint next = 0; next < wq.size(); ++next) {
 644         Node *n = wq.at(next);
 645         for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 646           Node* u = n->fast_out(i);
 647           if (!u->is_CFG()) {
 648             Node* c = _phase->get_ctrl(u);
 649             if (_lpt->is_member(_phase->get_loop(c)) || _phase->is_dominator(c, head)) {
 650               _visited.set(u->_idx);
 651               wq.push(u);
 652             }
 653           }
 654         }
 655       }
 656     }
 657   }
 658 
 659   // Map old to n for invariance computation and clone
 660   void map_ctrl(Node* old, Node* n) {
 661     assert(old->is_CFG() && n->is_CFG(), "must be");
 662     _old_new.map(old->_idx, n); // "clone" of old is n
 663     _invariant.set(old->_idx);  // old is invariant
 664     _clone_visited.set(old->_idx);
 665   }
 666 
 667   // Driver function to compute invariance
 668   bool is_invariant(Node* n) {
 669     if (!_visited.test_set(n->_idx))
 670       compute_invariance(n);
 671     return (_invariant.test(n->_idx) != 0);
 672   }
 673 
 674   // Driver function to clone invariant
 675   Node* clone(Node* n, Node* ctrl) {
 676     assert(ctrl->is_CFG(), "must be");
 677     assert(_invariant.test(n->_idx), "must be an invariant");
 678     if (!_clone_visited.test(n->_idx))
 679       clone_nodes(n, ctrl);
 680     return _old_new[n->_idx];
 681   }
 682 };
 683 
 684 //------------------------------is_range_check_if -----------------------------------
 685 // Returns true if the predicate of iff is in "scale*iv + offset u< load_range(ptr)" format
 686 // Note: this function is particularly designed for loop predication. We require load_range
 687 //       and offset to be loop invariant computed on the fly by "invar"
 688 bool IdealLoopTree::is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const {
 689   if (!is_loop_exit(iff)) {
 690     return false;
 691   }
 692   if (!iff->in(1)->is_Bool()) {
 693     return false;
 694   }
 695   const BoolNode *bol = iff->in(1)->as_Bool();
 696   if (bol->_test._test != BoolTest::lt) {
 697     return false;
 698   }
 699   if (!bol->in(1)->is_Cmp()) {
 700     return false;
 701   }
 702   const CmpNode *cmp = bol->in(1)->as_Cmp();
 703   if (cmp->Opcode() != Op_CmpU) {
 704     return false;
 705   }
 706   Node* range = cmp->in(2);
 707   if (range->Opcode() != Op_LoadRange && !iff->is_RangeCheck()) {
 708     const TypeInt* tint = phase->_igvn.type(range)->isa_int();
 709     if (tint == NULL || tint->empty() || tint->_lo < 0) {
 710       // Allow predication on positive values that aren't LoadRanges.
 711       // This allows optimization of loops where the length of the
 712       // array is a known value and doesn't need to be loaded back
 713       // from the array.
 714       return false;
 715     }
 716   }
 717   if (!invar.is_invariant(range)) {
 718     return false;
 719   }
 720   Node *iv     = _head->as_CountedLoop()->phi();
 721   int   scale  = 0;
 722   Node *offset = NULL;
 723   if (!phase->is_scaled_iv_plus_offset(cmp->in(1), iv, &scale, &offset)) {
 724     return false;
 725   }
 726   if (offset && !invar.is_invariant(offset)) { // offset must be invariant
 727     return false;
 728   }
 729   return true;
 730 }
 731 
 732 //------------------------------rc_predicate-----------------------------------
 733 // Create a range check predicate
 734 //
 735 // for (i = init; i < limit; i += stride) {
 736 //    a[scale*i+offset]
 737 // }
 738 //
 739 // Compute max(scale*i + offset) for init <= i < limit and build the predicate
 740 // as "max(scale*i + offset) u< a.length".
 741 //
 742 // There are two cases for max(scale*i + offset):
 743 // (1) stride*scale > 0
 744 //   max(scale*i + offset) = scale*(limit-stride) + offset
 745 // (2) stride*scale < 0
 746 //   max(scale*i + offset) = scale*init + offset
 747 BoolNode* PhaseIdealLoop::rc_predicate(IdealLoopTree *loop, Node* ctrl,
 748                                        int scale, Node* offset,
 749                                        Node* init, Node* limit, jint stride,
 750                                        Node* range, bool upper, bool &overflow) {
 751   jint con_limit  = (limit != NULL && limit->is_Con())  ? limit->get_int()  : 0;
 752   jint con_init   = init->is_Con()   ? init->get_int()   : 0;
 753   jint con_offset = offset->is_Con() ? offset->get_int() : 0;
 754 
 755   stringStream* predString = NULL;
 756   if (TraceLoopPredicate) {
 757     predString = new stringStream();
 758     predString->print("rc_predicate ");
 759   }
 760 
 761   overflow = false;
 762   Node* max_idx_expr = NULL;
 763   const TypeInt* idx_type = TypeInt::INT;
 764   if ((stride > 0) == (scale > 0) == upper) {
 765     guarantee(limit != NULL, "sanity");
 766     if (TraceLoopPredicate) {
 767       if (limit->is_Con()) {
 768         predString->print("(%d ", con_limit);
 769       } else {
 770         predString->print("(limit ");
 771       }
 772       predString->print("- %d) ", stride);
 773     }
 774     // Check if (limit - stride) may overflow
 775     const TypeInt* limit_type = _igvn.type(limit)->isa_int();
 776     jint limit_lo = limit_type->_lo;
 777     jint limit_hi = limit_type->_hi;
 778     if ((stride > 0 && (java_subtract(limit_lo, stride) < limit_lo)) ||
 779         (stride < 0 && (java_subtract(limit_hi, stride) > limit_hi))) {
 780       // No overflow possible
 781       ConINode* con_stride = _igvn.intcon(stride);
 782       set_ctrl(con_stride, C->root());
 783       max_idx_expr = new SubINode(limit, con_stride);
 784       idx_type = TypeInt::make(limit_lo - stride, limit_hi - stride, limit_type->_widen);
 785     } else {
 786       // May overflow
 787       overflow = true;
 788       limit = new ConvI2LNode(limit);
 789       register_new_node(limit, ctrl);
 790       ConLNode* con_stride = _igvn.longcon(stride);
 791       set_ctrl(con_stride, C->root());
 792       max_idx_expr = new SubLNode(limit, con_stride);
 793     }
 794     register_new_node(max_idx_expr, ctrl);
 795   } else {
 796     if (TraceLoopPredicate) {
 797       if (init->is_Con()) {
 798         predString->print("%d ", con_init);
 799       } else {
 800         predString->print("init ");
 801       }
 802     }
 803     idx_type = _igvn.type(init)->isa_int();
 804     max_idx_expr = init;
 805   }
 806 
 807   if (scale != 1) {
 808     ConNode* con_scale = _igvn.intcon(scale);
 809     set_ctrl(con_scale, C->root());
 810     if (TraceLoopPredicate) {
 811       predString->print("* %d ", scale);
 812     }
 813     // Check if (scale * max_idx_expr) may overflow
 814     const TypeInt* scale_type = TypeInt::make(scale);
 815     MulINode* mul = new MulINode(max_idx_expr, con_scale);
 816     idx_type = (TypeInt*)mul->mul_ring(idx_type, scale_type);
 817     if (overflow || TypeInt::INT->higher_equal(idx_type)) {
 818       // May overflow
 819       mul->destruct();
 820       if (!overflow) {
 821         max_idx_expr = new ConvI2LNode(max_idx_expr);
 822         register_new_node(max_idx_expr, ctrl);
 823       }
 824       overflow = true;
 825       con_scale = _igvn.longcon(scale);
 826       set_ctrl(con_scale, C->root());
 827       max_idx_expr = new MulLNode(max_idx_expr, con_scale);
 828     } else {
 829       // No overflow possible
 830       max_idx_expr = mul;
 831     }
 832     register_new_node(max_idx_expr, ctrl);
 833   }
 834 
 835   if (offset && (!offset->is_Con() || con_offset != 0)){
 836     if (TraceLoopPredicate) {
 837       if (offset->is_Con()) {
 838         predString->print("+ %d ", con_offset);
 839       } else {
 840         predString->print("+ offset");
 841       }
 842     }
 843     // Check if (max_idx_expr + offset) may overflow
 844     const TypeInt* offset_type = _igvn.type(offset)->isa_int();
 845     jint lo = java_add(idx_type->_lo, offset_type->_lo);
 846     jint hi = java_add(idx_type->_hi, offset_type->_hi);
 847     if (overflow || (lo > hi) ||
 848         ((idx_type->_lo & offset_type->_lo) < 0 && lo >= 0) ||
 849         ((~(idx_type->_hi | offset_type->_hi)) < 0 && hi < 0)) {
 850       // May overflow
 851       if (!overflow) {
 852         max_idx_expr = new ConvI2LNode(max_idx_expr);
 853         register_new_node(max_idx_expr, ctrl);
 854       }
 855       overflow = true;
 856       offset = new ConvI2LNode(offset);
 857       register_new_node(offset, ctrl);
 858       max_idx_expr = new AddLNode(max_idx_expr, offset);
 859     } else {
 860       // No overflow possible
 861       max_idx_expr = new AddINode(max_idx_expr, offset);
 862     }
 863     register_new_node(max_idx_expr, ctrl);
 864   }
 865 
 866   CmpNode* cmp = NULL;
 867   if (overflow) {
 868     // Integer expressions may overflow, do long comparison
 869     range = new ConvI2LNode(range);
 870     register_new_node(range, ctrl);
 871     cmp = new CmpULNode(max_idx_expr, range);
 872   } else {
 873     cmp = new CmpUNode(max_idx_expr, range);
 874   }
 875   register_new_node(cmp, ctrl);
 876   BoolNode* bol = new BoolNode(cmp, BoolTest::lt);
 877   register_new_node(bol, ctrl);
 878 
 879   if (TraceLoopPredicate) {
 880     predString->print_cr("<u range");
 881     tty->print("%s", predString->as_string());
 882   }
 883   return bol;
 884 }
 885 
 886 // Should loop predication look not only in the path from tail to head
 887 // but also in branches of the loop body?
 888 bool PhaseIdealLoop::loop_predication_should_follow_branches(IdealLoopTree *loop, ProjNode *predicate_proj, float& loop_trip_cnt) {
 889   if (!UseProfiledLoopPredicate) {
 890     return false;
 891   }
 892 
 893   if (predicate_proj == NULL) {
 894     return false;
 895   }
 896 
 897   LoopNode* head = loop->_head->as_Loop();
 898   bool follow_branches = true;
 899   IdealLoopTree* l = loop->_child;
 900   // For leaf loops and loops with a single inner loop
 901   while (l != NULL && follow_branches) {
 902     IdealLoopTree* child = l;
 903     if (child->_child != NULL &&
 904         child->_head->is_OuterStripMinedLoop()) {
 905       assert(child->_child->_next == NULL, "only one inner loop for strip mined loop");
 906       assert(child->_child->_head->is_CountedLoop() && child->_child->_head->as_CountedLoop()->is_strip_mined(), "inner loop should be strip mined");
 907       child = child->_child;
 908     }
 909     if (child->_child != NULL || child->_irreducible) {
 910       follow_branches = false;
 911     }
 912     l = l->_next;
 913   }
 914   if (follow_branches) {
 915     loop->compute_profile_trip_cnt(this);
 916     if (head->is_profile_trip_failed()) {
 917       follow_branches = false;
 918     } else {
 919       loop_trip_cnt = head->profile_trip_cnt();
 920       if (head->is_CountedLoop()) {
 921         CountedLoopNode* cl = head->as_CountedLoop();
 922         if (cl->phi() != NULL) {
 923           const TypeInt* t = _igvn.type(cl->phi())->is_int();
 924           float worst_case_trip_cnt = ((float)t->_hi - t->_lo) / ABS(cl->stride_con());
 925           if (worst_case_trip_cnt < loop_trip_cnt) {
 926             loop_trip_cnt = worst_case_trip_cnt;
 927           }
 928         }
 929       }
 930     }
 931   }
 932   return follow_branches;
 933 }
 934 
 935 // Compute probability of reaching some CFG node from a fixed
 936 // dominating CFG node
 937 class PathFrequency {
 938 private:
 939   Node* _dom; // frequencies are computed relative to this node
 940   Node_Stack _stack;
 941   GrowableArray<float> _freqs_stack; // keep track of intermediate result at regions
 942   GrowableArray<float> _freqs; // cache frequencies
 943   PhaseIdealLoop* _phase;
 944 
 945 public:
 946   PathFrequency(Node* dom, PhaseIdealLoop* phase)
 947     : _dom(dom), _stack(0), _phase(phase) {
 948   }
 949 
 950   float to(Node* n) {
 951     // post order walk on the CFG graph from n to _dom
 952     fesetround(FE_TOWARDZERO); // make sure rounding doesn't push frequency above 1
 953     IdealLoopTree* loop = _phase->get_loop(_dom);
 954     Node* c = n;
 955     for (;;) {
 956       assert(_phase->get_loop(c) == loop, "have to be in the same loop");
 957       if (c == _dom || _freqs.at_grow(c->_idx, -1) >= 0) {
 958         float f = c == _dom ? 1 : _freqs.at(c->_idx);
 959         Node* prev = c;
 960         while (_stack.size() > 0 && prev == c) {
 961           Node* n = _stack.node();
 962           if (!n->is_Region()) {
 963             if (_phase->get_loop(n) != _phase->get_loop(n->in(0))) {
 964               // Found an inner loop: compute frequency of reaching this
 965               // exit from the loop head by looking at the number of
 966               // times each loop exit was taken
 967               IdealLoopTree* inner_loop = _phase->get_loop(n->in(0));
 968               LoopNode* inner_head = inner_loop->_head->as_Loop();
 969               assert(_phase->get_loop(n) == loop, "only 1 inner loop");
 970               if (inner_head->is_OuterStripMinedLoop()) {
 971                 inner_head->verify_strip_mined(1);
 972                 if (n->in(0) == inner_head->in(LoopNode::LoopBackControl)->in(0)) {
 973                   n = n->in(0)->in(0)->in(0);
 974                 }
 975                 inner_loop = inner_loop->_child;
 976                 inner_head = inner_loop->_head->as_Loop();
 977                 inner_head->verify_strip_mined(1);
 978               }
 979               fesetround(FE_UPWARD);  // make sure rounding doesn't push frequency above 1
 980               float loop_exit_cnt = 0.0f;
 981               for (uint i = 0; i < inner_loop->_body.size(); i++) {
 982                 Node *n = inner_loop->_body[i];
 983                 float c = inner_loop->compute_profile_trip_cnt_helper(n);
 984                 loop_exit_cnt += c;
 985               }
 986               fesetround(FE_TOWARDZERO);
 987               float cnt = -1;
 988               if (n->in(0)->is_If()) {
 989                 IfNode* iff = n->in(0)->as_If();
 990                 float p = n->in(0)->as_If()->_prob;
 991                 if (n->Opcode() == Op_IfFalse) {
 992                   p = 1 - p;
 993                 }
 994                 if (p > PROB_MIN) {
 995                   cnt = p * iff->_fcnt;
 996                 } else {
 997                   cnt = 0;
 998                 }
 999               } else {
1000                 assert(n->in(0)->is_Jump(), "unsupported node kind");
1001                 JumpNode* jmp = n->in(0)->as_Jump();
1002                 float p = n->in(0)->as_Jump()->_probs[n->as_JumpProj()->_con];
1003                 cnt = p * jmp->_fcnt;
1004               }
1005               float this_exit_f = cnt > 0 ? cnt / loop_exit_cnt : 0;
1006               assert(this_exit_f <= 1 && this_exit_f >= 0, "Incorrect frequency");
1007               f = f * this_exit_f;
1008               assert(f <= 1 && f >= 0, "Incorrect frequency");
1009             } else {
1010               float p = -1;
1011               if (n->in(0)->is_If()) {
1012                 p = n->in(0)->as_If()->_prob;
1013                 if (n->Opcode() == Op_IfFalse) {
1014                   p = 1 - p;
1015                 }
1016               } else {
1017                 assert(n->in(0)->is_Jump(), "unsupported node kind");
1018                 p = n->in(0)->as_Jump()->_probs[n->as_JumpProj()->_con];
1019               }
1020               f = f * p;
1021               assert(f <= 1 && f >= 0, "Incorrect frequency");
1022             }
1023             _freqs.at_put_grow(n->_idx, (float)f, -1);
1024             _stack.pop();
1025           } else {
1026             float prev_f = _freqs_stack.pop();
1027             float new_f = f;
1028             f = new_f + prev_f;
1029             assert(f <= 1 && f >= 0, "Incorrect frequency");
1030             uint i = _stack.index();
1031             if (i < n->req()) {
1032               c = n->in(i);
1033               _stack.set_index(i+1);
1034               _freqs_stack.push(f);
1035             } else {
1036               _freqs.at_put_grow(n->_idx, f, -1);
1037               _stack.pop();
1038             }
1039           }
1040         }
1041         if (_stack.size() == 0) {
1042           fesetround(FE_TONEAREST);
1043           assert(f >= 0 && f <= 1, "should have been computed");
1044           return f;
1045         }
1046       } else if (c->is_Loop()) {
1047         ShouldNotReachHere();
1048         c = c->in(LoopNode::EntryControl);
1049       } else if (c->is_Region()) {
1050         _freqs_stack.push(0);
1051         _stack.push(c, 2);
1052         c = c->in(1);
1053       } else {
1054         if (c->is_IfProj()) {
1055           IfNode* iff = c->in(0)->as_If();
1056           if (iff->_prob == PROB_UNKNOWN) {
1057             // assume never taken
1058             _freqs.at_put_grow(c->_idx, 0, -1);
1059           } else if (_phase->get_loop(c) != _phase->get_loop(iff)) {
1060             if (iff->_fcnt == COUNT_UNKNOWN) {
1061               // assume never taken
1062               _freqs.at_put_grow(c->_idx, 0, -1);
1063             } else {
1064               // skip over loop
1065               _stack.push(c, 1);
1066               c = _phase->get_loop(c->in(0))->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
1067             }
1068           } else {
1069             _stack.push(c, 1);
1070             c = iff;
1071           }
1072         } else if (c->is_JumpProj()) {
1073           JumpNode* jmp = c->in(0)->as_Jump();
1074           if (_phase->get_loop(c) != _phase->get_loop(jmp)) {
1075             if (jmp->_fcnt == COUNT_UNKNOWN) {
1076               // assume never taken
1077               _freqs.at_put_grow(c->_idx, 0, -1);
1078             } else {
1079               // skip over loop
1080               _stack.push(c, 1);
1081               c = _phase->get_loop(c->in(0))->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
1082             }
1083           } else {
1084             _stack.push(c, 1);
1085             c = jmp;
1086           }
1087         } else if (c->Opcode() == Op_CatchProj &&
1088                    c->in(0)->Opcode() == Op_Catch &&
1089                    c->in(0)->in(0)->is_Proj() &&
1090                    c->in(0)->in(0)->in(0)->is_Call()) {
1091           // assume exceptions are never thrown
1092           uint con = c->as_Proj()->_con;
1093           if (con == CatchProjNode::fall_through_index) {
1094             Node* call = c->in(0)->in(0)->in(0)->in(0);
1095             if (_phase->get_loop(call) != _phase->get_loop(c)) {
1096               _freqs.at_put_grow(c->_idx, 0, -1);
1097             } else {
1098               c = call;
1099             }
1100           } else {
1101             assert(con >= CatchProjNode::catch_all_index, "what else?");
1102             _freqs.at_put_grow(c->_idx, 0, -1);
1103           }
1104         } else if (c->unique_ctrl_out() == NULL && !c->is_If() && !c->is_Jump()) {
1105           ShouldNotReachHere();
1106         } else {
1107           c = c->in(0);
1108         }
1109       }
1110     }
1111     ShouldNotReachHere();
1112     return -1;
1113   }
1114 };
1115 
1116 void PhaseIdealLoop::loop_predication_follow_branches(Node *n, IdealLoopTree *loop, float loop_trip_cnt,
1117                                                       PathFrequency& pf, Node_Stack& stack, VectorSet& seen,
1118                                                       Node_List& if_proj_list) {
1119   assert(n->is_Region(), "start from a region");
1120   Node* tail = loop->tail();
1121   stack.push(n, 1);
1122   do {
1123     Node* c = stack.node();
1124     assert(c->is_Region() || c->is_IfProj(), "only region here");
1125     uint i = stack.index();
1126 
1127     if (i < c->req()) {
1128       stack.set_index(i+1);
1129       Node* in = c->in(i);
1130       while (!is_dominator(in, tail) && !seen.test_set(in->_idx)) {
1131         IdealLoopTree* in_loop = get_loop(in);
1132         if (in_loop != loop) {
1133           in = in_loop->_head->in(LoopNode::EntryControl);
1134         } else if (in->is_Region()) {
1135           stack.push(in, 1);
1136           break;
1137         } else if (in->is_IfProj() &&
1138                    in->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) &&
1139                    (in->in(0)->Opcode() == Op_If ||
1140                     in->in(0)->Opcode() == Op_RangeCheck)) {
1141           if (pf.to(in) * loop_trip_cnt >= 1) {
1142             stack.push(in, 1);
1143           }
1144           in = in->in(0);
1145         } else {
1146           in = in->in(0);
1147         }
1148       }
1149     } else {
1150       if (c->is_IfProj()) {
1151         if_proj_list.push(c);
1152       }
1153       stack.pop();
1154     }
1155 
1156   } while (stack.size() > 0);
1157 }
1158 
1159 
1160 bool PhaseIdealLoop::loop_predication_impl_helper(IdealLoopTree *loop, ProjNode* proj, ProjNode *predicate_proj,
1161                                                   CountedLoopNode *cl, ConNode* zero, Invariance& invar,
1162                                                   Deoptimization::DeoptReason reason) {
1163   // Following are changed to nonnull when a predicate can be hoisted
1164   ProjNode* new_predicate_proj = NULL;
1165   IfNode*   iff  = proj->in(0)->as_If();
1166   Node*     test = iff->in(1);
1167   if (!test->is_Bool()){ //Conv2B, ...
1168     return false;
1169   }
1170   BoolNode* bol = test->as_Bool();
1171   if (invar.is_invariant(bol)) {
1172     // Invariant test
1173     new_predicate_proj = create_new_if_for_predicate(predicate_proj, NULL,
1174                                                      reason,
1175                                                      iff->Opcode());
1176     Node* ctrl = new_predicate_proj->in(0)->as_If()->in(0);
1177     BoolNode* new_predicate_bol = invar.clone(bol, ctrl)->as_Bool();
1178 
1179     // Negate test if necessary
1180     bool negated = false;
1181     if (proj->_con != predicate_proj->_con) {
1182       new_predicate_bol = new BoolNode(new_predicate_bol->in(1), new_predicate_bol->_test.negate());
1183       register_new_node(new_predicate_bol, ctrl);
1184       negated = true;
1185     }
1186     IfNode* new_predicate_iff = new_predicate_proj->in(0)->as_If();
1187     _igvn.hash_delete(new_predicate_iff);
1188     new_predicate_iff->set_req(1, new_predicate_bol);
1189 #ifndef PRODUCT
1190     if (TraceLoopPredicate) {
1191       tty->print("Predicate invariant if%s: %d ", negated ? " negated" : "", new_predicate_iff->_idx);
1192       loop->dump_head();
1193     } else if (TraceLoopOpts) {
1194       tty->print("Predicate IC ");
1195       loop->dump_head();
1196     }
1197 #endif
1198   } else if (cl != NULL && loop->is_range_check_if(iff, this, invar)) {
1199     // Range check for counted loops
1200     const Node*    cmp    = bol->in(1)->as_Cmp();
1201     Node*          idx    = cmp->in(1);
1202     assert(!invar.is_invariant(idx), "index is variant");
1203     Node* rng = cmp->in(2);
1204     assert(rng->Opcode() == Op_LoadRange || iff->is_RangeCheck() || _igvn.type(rng)->is_int()->_lo >= 0, "must be");
1205     assert(invar.is_invariant(rng), "range must be invariant");
1206     int scale    = 1;
1207     Node* offset = zero;
1208     bool ok = is_scaled_iv_plus_offset(idx, cl->phi(), &scale, &offset);
1209     assert(ok, "must be index expression");
1210 
1211     Node* init    = cl->init_trip();
1212     // Limit is not exact.
1213     // Calculate exact limit here.
1214     // Note, counted loop's test is '<' or '>'.
1215     Node* limit   = exact_limit(loop);
1216     int  stride   = cl->stride()->get_int();
1217 
1218     // Build if's for the upper and lower bound tests.  The
1219     // lower_bound test will dominate the upper bound test and all
1220     // cloned or created nodes will use the lower bound test as
1221     // their declared control.
1222 
1223     // Perform cloning to keep Invariance state correct since the
1224     // late schedule will place invariant things in the loop.
1225     Node *ctrl = predicate_proj->in(0)->as_If()->in(0);
1226     rng = invar.clone(rng, ctrl);
1227     if (offset && offset != zero) {
1228       assert(invar.is_invariant(offset), "offset must be loop invariant");
1229       offset = invar.clone(offset, ctrl);
1230     }
1231     // If predicate expressions may overflow in the integer range, longs are used.
1232     bool overflow = false;
1233 
1234     // Test the lower bound
1235     BoolNode* lower_bound_bol = rc_predicate(loop, ctrl, scale, offset, init, limit, stride, rng, false, overflow);
1236     // Negate test if necessary
1237     bool negated = false;
1238     if (proj->_con != predicate_proj->_con) {
1239       lower_bound_bol = new BoolNode(lower_bound_bol->in(1), lower_bound_bol->_test.negate());
1240       register_new_node(lower_bound_bol, ctrl);
1241       negated = true;
1242     }
1243     ProjNode* lower_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, reason, overflow ? Op_If : iff->Opcode());
1244     IfNode* lower_bound_iff = lower_bound_proj->in(0)->as_If();
1245     _igvn.hash_delete(lower_bound_iff);
1246     lower_bound_iff->set_req(1, lower_bound_bol);
1247     if (TraceLoopPredicate) tty->print_cr("lower bound check if: %s %d ", negated ? " negated" : "", lower_bound_iff->_idx);
1248 
1249     // Test the upper bound
1250     BoolNode* upper_bound_bol = rc_predicate(loop, lower_bound_proj, scale, offset, init, limit, stride, rng, true, overflow);
1251     negated = false;
1252     if (proj->_con != predicate_proj->_con) {
1253       upper_bound_bol = new BoolNode(upper_bound_bol->in(1), upper_bound_bol->_test.negate());
1254       register_new_node(upper_bound_bol, ctrl);
1255       negated = true;
1256     }
1257     ProjNode* upper_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, reason, overflow ? Op_If : iff->Opcode());
1258     assert(upper_bound_proj->in(0)->as_If()->in(0) == lower_bound_proj, "should dominate");
1259     IfNode* upper_bound_iff = upper_bound_proj->in(0)->as_If();
1260     _igvn.hash_delete(upper_bound_iff);
1261     upper_bound_iff->set_req(1, upper_bound_bol);
1262     if (TraceLoopPredicate) tty->print_cr("upper bound check if: %s %d ", negated ? " negated" : "", lower_bound_iff->_idx);
1263 
1264     // Fall through into rest of the clean up code which will move
1265     // any dependent nodes onto the upper bound test.
1266     new_predicate_proj = upper_bound_proj;
1267 
1268     if (iff->is_RangeCheck()) {
1269       new_predicate_proj = insert_skeleton_predicate(iff, loop, proj, predicate_proj, upper_bound_proj, scale, offset, init, limit, stride, rng, overflow, reason);
1270     }
1271 
1272 #ifndef PRODUCT
1273     if (TraceLoopOpts && !TraceLoopPredicate) {
1274       tty->print("Predicate RC ");
1275       loop->dump_head();
1276     }
1277 #endif
1278   } else {
1279     // Loop variant check (for example, range check in non-counted loop)
1280     // with uncommon trap.
1281     return false;
1282   }
1283   assert(new_predicate_proj != NULL, "sanity");
1284   // Success - attach condition (new_predicate_bol) to predicate if
1285   invar.map_ctrl(proj, new_predicate_proj); // so that invariance test can be appropriate
1286 
1287   // Eliminate the old If in the loop body
1288   dominated_by( new_predicate_proj, iff, proj->_con != new_predicate_proj->_con );
1289 
1290   C->set_major_progress();
1291   return true;
1292 }
1293 
1294 
1295 // After pre/main/post loops are created, we'll put a copy of some
1296 // range checks between the pre and main loop to validate the value
1297 // of the main loop induction variable. Make a copy of the predicates
1298 // here with an opaque node as a place holder for the value (will be
1299 // updated by PhaseIdealLoop::update_skeleton_predicate()).
1300 ProjNode* PhaseIdealLoop::insert_skeleton_predicate(IfNode* iff, IdealLoopTree *loop,
1301                                                     ProjNode* proj, ProjNode *predicate_proj,
1302                                                     ProjNode* upper_bound_proj,
1303                                                     int scale, Node* offset,
1304                                                     Node* init, Node* limit, jint stride,
1305                                                     Node* rng, bool &overflow,
1306                                                     Deoptimization::DeoptReason reason) {
1307   assert(proj->_con && predicate_proj->_con, "not a range check?");
1308   Node* opaque_init = new Opaque1Node(C, init);
1309   register_new_node(opaque_init, upper_bound_proj);
1310   BoolNode* bol = rc_predicate(loop, upper_bound_proj, scale, offset, opaque_init, limit, stride, rng, (stride > 0) != (scale > 0), overflow);
1311   Node* opaque_bol = new Opaque4Node(C, bol, _igvn.intcon(1)); // This will go away once loop opts are over
1312   register_new_node(opaque_bol, upper_bound_proj);
1313   ProjNode* new_proj = create_new_if_for_predicate(predicate_proj, NULL, reason, overflow ? Op_If : iff->Opcode());
1314   _igvn.replace_input_of(new_proj->in(0), 1, opaque_bol);
1315   assert(opaque_init->outcnt() > 0, "should be used");
1316   return new_proj;
1317 }
1318 
1319 //------------------------------ loop_predication_impl--------------------------
1320 // Insert loop predicates for null checks and range checks
1321 bool PhaseIdealLoop::loop_predication_impl(IdealLoopTree *loop) {
1322   if (!UseLoopPredicate) return false;
1323 
1324   if (!loop->_head->is_Loop()) {
1325     // Could be a simple region when irreducible loops are present.
1326     return false;
1327   }
1328   LoopNode* head = loop->_head->as_Loop();
1329 
1330   if (head->unique_ctrl_out()->Opcode() == Op_NeverBranch) {
1331     // do nothing for infinite loops
1332     return false;
1333   }
1334 
1335   if (head->is_OuterStripMinedLoop()) {
1336     return false;
1337   }
1338 
1339   CountedLoopNode *cl = NULL;
1340   if (head->is_valid_counted_loop()) {
1341     cl = head->as_CountedLoop();
1342     // do nothing for iteration-splitted loops
1343     if (!cl->is_normal_loop()) return false;
1344     // Avoid RCE if Counted loop's test is '!='.
1345     BoolTest::mask bt = cl->loopexit()->test_trip();
1346     if (bt != BoolTest::lt && bt != BoolTest::gt)
1347       cl = NULL;
1348   }
1349 
1350   Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl);
1351   ProjNode *loop_limit_proj = NULL;
1352   ProjNode *predicate_proj = NULL;
1353   ProjNode *profile_predicate_proj = NULL;
1354   // Loop limit check predicate should be near the loop.
1355   loop_limit_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
1356   if (loop_limit_proj != NULL) {
1357     entry = loop_limit_proj->in(0)->in(0);
1358   }
1359   bool has_profile_predicates = false;
1360   profile_predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate);
1361   if (profile_predicate_proj != NULL) {
1362     Node* n = skip_loop_predicates(entry);
1363     // Check if predicates were already added to the profile predicate
1364     // block
1365     if (n != entry->in(0)->in(0) || n->outcnt() != 1) {
1366       has_profile_predicates = true;
1367     }
1368     entry = n;
1369   }
1370   predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
1371 
1372   float loop_trip_cnt = -1;
1373   bool follow_branches = loop_predication_should_follow_branches(loop, profile_predicate_proj, loop_trip_cnt);
1374   assert(!follow_branches || loop_trip_cnt >= 0, "negative trip count?");
1375 
1376   if (predicate_proj == NULL && !follow_branches) {
1377 #ifndef PRODUCT
1378     if (TraceLoopPredicate) {
1379       tty->print("missing predicate:");
1380       loop->dump_head();
1381       head->dump(1);
1382     }
1383 #endif
1384     return false;
1385   }
1386   ConNode* zero = _igvn.intcon(0);
1387   set_ctrl(zero, C->root());
1388 
1389   ResourceArea *area = Thread::current()->resource_area();
1390   Invariance invar(area, loop);
1391 
1392   // Create list of if-projs such that a newer proj dominates all older
1393   // projs in the list, and they all dominate loop->tail()
1394   Node_List if_proj_list(area);
1395   Node_List regions(area);
1396   Node *current_proj = loop->tail(); //start from tail
1397 
1398 
1399   Node_List controls(area);
1400   while (current_proj != head) {
1401     if (loop == get_loop(current_proj) && // still in the loop ?
1402         current_proj->is_Proj()        && // is a projection  ?
1403         (current_proj->in(0)->Opcode() == Op_If ||
1404          current_proj->in(0)->Opcode() == Op_RangeCheck)) { // is a if projection ?
1405       if_proj_list.push(current_proj);
1406     }
1407     if (follow_branches &&
1408         current_proj->Opcode() == Op_Region &&
1409         loop == get_loop(current_proj)) {
1410       regions.push(current_proj);
1411     }
1412     current_proj = idom(current_proj);
1413   }
1414 
1415   bool hoisted = false; // true if at least one proj is promoted
1416 
1417   if (!has_profile_predicates) {
1418     while (if_proj_list.size() > 0) {
1419       Node* n = if_proj_list.pop();
1420 
1421       ProjNode* proj = n->as_Proj();
1422       IfNode*   iff  = proj->in(0)->as_If();
1423 
1424       CallStaticJavaNode* call = proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
1425       if (call == NULL) {
1426         if (loop->is_loop_exit(iff)) {
1427           // stop processing the remaining projs in the list because the execution of them
1428           // depends on the condition of "iff" (iff->in(1)).
1429           break;
1430         } else {
1431           // Both arms are inside the loop. There are two cases:
1432           // (1) there is one backward branch. In this case, any remaining proj
1433           //     in the if_proj list post-dominates "iff". So, the condition of "iff"
1434           //     does not determine the execution the remining projs directly, and we
1435           //     can safely continue.
1436           // (2) both arms are forwarded, i.e. a diamond shape. In this case, "proj"
1437           //     does not dominate loop->tail(), so it can not be in the if_proj list.
1438           continue;
1439         }
1440       }
1441       Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(call->uncommon_trap_request());
1442       if (reason == Deoptimization::Reason_predicate) {
1443         break;
1444       }
1445 
1446       if (predicate_proj != NULL) {
1447         hoisted = loop_predication_impl_helper(loop, proj, predicate_proj, cl, zero, invar, Deoptimization::Reason_predicate) | hoisted;
1448       }
1449     } // end while
1450   }
1451 
1452   Node_List if_proj_list_freq(area);
1453   if (follow_branches) {
1454     PathFrequency pf(loop->_head, this);
1455 
1456     // Some projections were skipped by regular predicates because of
1457     // an early loop exit. Try them with profile data.
1458     while (if_proj_list.size() > 0) {
1459       Node* proj = if_proj_list.pop();
1460       float f = pf.to(proj);
1461       if (proj->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) &&
1462           f * loop_trip_cnt >= 1) {
1463         hoisted = loop_predication_impl_helper(loop, proj->as_Proj(), profile_predicate_proj, cl, zero, invar, Deoptimization::Reason_profile_predicate) | hoisted;
1464       }
1465     }
1466 
1467     // And look into all branches
1468     Node_Stack stack(0);
1469     VectorSet seen(Thread::current()->resource_area());
1470     while (regions.size() > 0) {
1471       Node* c = regions.pop();
1472       loop_predication_follow_branches(c, loop, loop_trip_cnt, pf, stack, seen, if_proj_list_freq);
1473     }
1474 
1475     for (uint i = 0; i < if_proj_list_freq.size(); i++) {
1476       ProjNode* proj = if_proj_list_freq.at(i)->as_Proj();
1477       hoisted = loop_predication_impl_helper(loop, proj, profile_predicate_proj, cl, zero, invar, Deoptimization::Reason_profile_predicate) | hoisted;
1478     }
1479   }
1480 
1481 #ifndef PRODUCT
1482   // report that the loop predication has been actually performed
1483   // for this loop
1484   if (TraceLoopPredicate && hoisted) {
1485     tty->print("Loop Predication Performed:");
1486     loop->dump_head();
1487   }
1488 #endif
1489 
1490   head->verify_strip_mined(1);
1491 
1492   return hoisted;
1493 }
1494 
1495 //------------------------------loop_predication--------------------------------
1496 // driver routine for loop predication optimization
1497 bool IdealLoopTree::loop_predication( PhaseIdealLoop *phase) {
1498   bool hoisted = false;
1499   // Recursively promote predicates
1500   if (_child) {
1501     hoisted = _child->loop_predication( phase);
1502   }
1503 
1504   // self
1505   if (!_irreducible && !tail()->is_top()) {
1506     hoisted |= phase->loop_predication_impl(this);
1507   }
1508 
1509   if (_next) { //sibling
1510     hoisted |= _next->loop_predication( phase);
1511   }
1512 
1513   return hoisted;
1514 }
--- EOF ---