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_internal(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   if (!UseShenandoahGC) {
 308     return;
 309   }
 310   Compile* C = NULL;
 311   if (loop_phase != NULL) {
 312     igvn = &loop_phase->igvn();
 313   }
 314   C = igvn->C;
 315   ProjNode* other_dom_proj = dom_proj->in(0)->as_Multi()->proj_out(1-dom_proj->_con);
 316   Node* dom_r = other_dom_proj->unique_ctrl_out();
 317   if (dom_r->is_Region()) {
 318     assert(dom_r->unique_ctrl_out()->is_Call(), "unc expected");
 319     ProjNode* other_proj = proj->in(0)->as_Multi()->proj_out(1-proj->_con);
 320     Node* r = other_proj->unique_ctrl_out();
 321     assert(r->is_Region() && r->unique_ctrl_out()->is_Call(), "cloned predicate should have caused region to be added");
 322     for (DUIterator_Fast imax, i = dom_r->fast_outs(imax); i < imax; i++) {
 323       Node* dom_use = dom_r->fast_out(i);
 324       if (dom_use->is_Phi() && dom_use->bottom_type() == Type::MEMORY) {
 325         assert(dom_use->in(0) == dom_r, "");
 326         Node* phi = NULL;
 327         for (DUIterator_Fast jmax, j = r->fast_outs(jmax); j < jmax; j++) {
 328           Node* use = r->fast_out(j);
 329           if (use->is_Phi() && use->bottom_type() == Type::MEMORY &&
 330               use->adr_type() == dom_use->adr_type()) {
 331             assert(use->in(0) == r, "");
 332             assert(phi == NULL, "only one phi");
 333             phi = use;
 334           }
 335         }
 336         if (phi == NULL) {
 337           const TypePtr* adr_type = dom_use->adr_type();
 338           int alias = C->get_alias_index(adr_type);
 339           Node* call = r->unique_ctrl_out();
 340           Node* mem = call->in(TypeFunc::Memory);
 341           MergeMemNode* mm = NULL;
 342           if (mem->is_MergeMem()) {
 343             mm = mem->clone()->as_MergeMem();
 344             if (adr_type == TypePtr::BOTTOM) {
 345               mem = mem->as_MergeMem()->base_memory();
 346             } else {
 347               mem = mem->as_MergeMem()->memory_at(alias);
 348             }
 349           } else {
 350             mm = MergeMemNode::make(mem);
 351           }
 352           phi = PhiNode::make(r, mem, Type::MEMORY, adr_type);
 353           if (adr_type == TypePtr::BOTTOM) {
 354             mm->set_base_memory(phi);
 355           } else {
 356             mm->set_memory_at(alias, phi);
 357           }
 358           if (loop_phase != NULL) {
 359             loop_phase->register_new_node(mm, r);
 360             loop_phase->register_new_node(phi, r);
 361           } else {
 362             igvn->register_new_node_with_optimizer(mm);
 363             igvn->register_new_node_with_optimizer(phi);
 364           }
 365           igvn->replace_input_of(call, TypeFunc::Memory, mm);
 366         }
 367         igvn->replace_input_of(phi, r->req()-1, dom_use->in(dom_r->find_edge(other_dom_proj)));
 368       }
 369     }
 370   }
 371 }
 372 
 373 
 374 // Clone loop predicates to cloned loops (peeled, unswitched, split_if).
 375 Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry,
 376                                             bool clone_limit_check,
 377                                             PhaseIdealLoop* loop_phase,
 378                                             PhaseIterGVN* igvn) {
 379 #ifdef ASSERT
 380   if (new_entry == NULL || !(new_entry->is_Proj() || new_entry->is_Region() || new_entry->is_SafePoint())) {
 381     if (new_entry != NULL)
 382       new_entry->dump();
 383     assert(false, "not IfTrue, IfFalse, Region or SafePoint");
 384   }
 385 #endif
 386   // Search original predicates
 387   Node* entry = old_entry;
 388   ProjNode* limit_check_proj = NULL;
 389   limit_check_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
 390   if (limit_check_proj != NULL) {
 391     entry = skip_loop_predicates(entry);
 392   }
 393   ProjNode* profile_predicate_proj = NULL;
 394   ProjNode* predicate_proj = NULL;
 395   if (UseProfiledLoopPredicate) {
 396     profile_predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate);
 397     if (profile_predicate_proj != NULL) {
 398       entry = skip_loop_predicates(entry);
 399     }
 400   }
 401   if (UseLoopPredicate) {
 402     predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
 403   }
 404   if (predicate_proj != NULL) { // right pattern that can be used by loop predication
 405     // clone predicate
 406     ProjNode* proj = clone_predicate(predicate_proj, new_entry,
 407                                      Deoptimization::Reason_predicate,
 408                                      loop_phase, igvn);
 409     assert(proj != NULL, "IfTrue or IfFalse after clone predicate");
 410     new_entry = proj;
 411     if (TraceLoopPredicate) {
 412       tty->print("Loop Predicate cloned: ");
 413       debug_only( new_entry->in(0)->dump(); );
 414     }
 415     if (profile_predicate_proj != NULL) {
 416       // A node that produces memory may be out of loop and depend on
 417       // a profiled predicates. In that case the memory state at the
 418       // end of profiled predicates and at the end of predicates are
 419       // not the same. The cloned predicates are dominated by the
 420       // profiled predicates but may have the wrong memory
 421       // state. Update it.
 422       clone_loop_predicates_fix_mem(profile_predicate_proj, proj, loop_phase, igvn);
 423     }
 424   }
 425   if (profile_predicate_proj != NULL) { // right pattern that can be used by loop predication
 426     // clone predicate
 427     new_entry = clone_predicate(profile_predicate_proj, new_entry,
 428                                 Deoptimization::Reason_profile_predicate,
 429                                 loop_phase, igvn);
 430     assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone predicate");
 431     if (TraceLoopPredicate) {
 432       tty->print("Loop Predicate cloned: ");
 433       debug_only( new_entry->in(0)->dump(); );
 434     }
 435   }
 436   if (limit_check_proj != NULL && clone_limit_check) {
 437     // Clone loop limit check last to insert it before loop.
 438     // Don't clone a limit check which was already finalized
 439     // for this counted loop (only one limit check is needed).
 440     new_entry = clone_predicate(limit_check_proj, new_entry,
 441                                 Deoptimization::Reason_loop_limit_check,
 442                                 loop_phase, igvn);
 443     assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone limit check");
 444     if (TraceLoopLimitCheck) {
 445       tty->print("Loop Limit Check cloned: ");
 446       debug_only( new_entry->in(0)->dump(); )
 447     }
 448   }
 449   return new_entry;
 450 }
 451 
 452 //--------------------------skip_loop_predicates------------------------------
 453 // Skip related predicates.
 454 Node* PhaseIdealLoop::skip_loop_predicates(Node* entry) {
 455   IfNode* iff = entry->in(0)->as_If();
 456   ProjNode* uncommon_proj = iff->proj_out(1 - entry->as_Proj()->_con);
 457   Node* rgn = uncommon_proj->unique_ctrl_out();
 458   assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
 459   entry = entry->in(0)->in(0);
 460   while (entry != NULL && entry->is_Proj() && entry->in(0)->is_If()) {
 461     uncommon_proj = entry->in(0)->as_If()->proj_out(1 - entry->as_Proj()->_con);
 462     if (uncommon_proj->unique_ctrl_out() != rgn)
 463       break;
 464     entry = entry->in(0)->in(0);
 465   }
 466   return entry;
 467 }
 468 
 469 Node* PhaseIdealLoop::skip_all_loop_predicates(Node* entry) {
 470   Node* predicate = NULL;
 471   predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
 472   if (predicate != NULL) {
 473     entry = skip_loop_predicates(entry);
 474   }
 475   if (UseProfiledLoopPredicate) {
 476     predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate);
 477     if (predicate != NULL) { // right pattern that can be used by loop predication
 478       entry = skip_loop_predicates(entry);
 479     }
 480   }
 481   if (UseLoopPredicate) {
 482     predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
 483     if (predicate != NULL) { // right pattern that can be used by loop predication
 484       entry = skip_loop_predicates(entry);
 485     }
 486   }
 487   return entry;
 488 }
 489 
 490 //--------------------------find_predicate_insertion_point-------------------
 491 // Find a good location to insert a predicate
 492 ProjNode* PhaseIdealLoop::find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason) {
 493   if (start_c == NULL || !start_c->is_Proj())
 494     return NULL;
 495   if (start_c->as_Proj()->is_uncommon_trap_if_pattern(reason)) {
 496     return start_c->as_Proj();
 497   }
 498   return NULL;
 499 }
 500 
 501 //--------------------------find_predicate------------------------------------
 502 // Find a predicate
 503 Node* PhaseIdealLoop::find_predicate(Node* entry) {
 504   Node* predicate = NULL;
 505   predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
 506   if (predicate != NULL) { // right pattern that can be used by loop predication
 507     return entry;
 508   }
 509   if (UseLoopPredicate) {
 510     predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
 511     if (predicate != NULL) { // right pattern that can be used by loop predication
 512       return entry;
 513     }
 514   }
 515   if (UseProfiledLoopPredicate) {
 516     predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate);
 517     if (predicate != NULL) { // right pattern that can be used by loop predication
 518       return entry;
 519     }
 520   }
 521   return NULL;
 522 }
 523 
 524 //------------------------------Invariance-----------------------------------
 525 // Helper class for loop_predication_impl to compute invariance on the fly and
 526 // clone invariants.
 527 class Invariance : public StackObj {
 528   VectorSet _visited, _invariant;
 529   Node_Stack _stack;
 530   VectorSet _clone_visited;
 531   Node_List _old_new; // map of old to new (clone)
 532   IdealLoopTree* _lpt;
 533   PhaseIdealLoop* _phase;
 534 
 535   // Helper function to set up the invariance for invariance computation
 536   // If n is a known invariant, set up directly. Otherwise, look up the
 537   // the possibility to push n onto the stack for further processing.
 538   void visit(Node* use, Node* n) {
 539     if (_lpt->is_invariant(n)) { // known invariant
 540       _invariant.set(n->_idx);
 541     } else if (!n->is_CFG()) {
 542       if (n->Opcode() == Op_ShenandoahWriteBarrier) {
 543         return;
 544       }
 545       Node *n_ctrl = _phase->ctrl_or_self(n);
 546       Node *u_ctrl = _phase->ctrl_or_self(use); // self if use is a CFG
 547       if (_phase->is_dominator(n_ctrl, u_ctrl)) {
 548         _stack.push(n, n->in(0) == NULL ? 1 : 0);
 549       }
 550     }
 551   }
 552 
 553   // Compute invariance for "the_node" and (possibly) all its inputs recursively
 554   // on the fly
 555   void compute_invariance(Node* n) {
 556     assert(_visited.test(n->_idx), "must be");
 557     visit(n, n);
 558     while (_stack.is_nonempty()) {
 559       Node*  n = _stack.node();
 560       uint idx = _stack.index();
 561       if (idx == n->req()) { // all inputs are processed
 562         _stack.pop();
 563         // n is invariant if it's inputs are all invariant
 564         bool all_inputs_invariant = true;
 565         for (uint i = 0; i < n->req(); i++) {
 566           Node* in = n->in(i);
 567           if (in == NULL) continue;
 568           assert(_visited.test(in->_idx), "must have visited input");
 569           if (!_invariant.test(in->_idx)) { // bad guy
 570             all_inputs_invariant = false;
 571             break;
 572           }
 573         }
 574         if (all_inputs_invariant) {
 575           // If n's control is a predicate that was moved out of the
 576           // loop, it was marked invariant but n is only invariant if
 577           // it depends only on that test. Otherwise, unless that test
 578           // is out of the loop, it's not invariant.
 579           if (n->Opcode() == Op_ShenandoahWBMemProj || n->is_CFG() || n->depends_only_on_test() || n->in(0) == NULL || !_phase->is_member(_lpt, n->in(0))) {
 580             _invariant.set(n->_idx); // I am a invariant too
 581           }
 582         }
 583       } else { // process next input
 584         _stack.set_index(idx + 1);
 585         Node* m = n->in(idx);
 586         if (m != NULL && !_visited.test_set(m->_idx)) {
 587           visit(n, m);
 588         }
 589       }
 590     }
 591   }
 592 
 593   // Helper function to set up _old_new map for clone_nodes.
 594   // If n is a known invariant, set up directly ("clone" of n == n).
 595   // Otherwise, push n onto the stack for real cloning.
 596   void clone_visit(Node* n) {
 597     assert(_invariant.test(n->_idx), "must be invariant");
 598     if (_lpt->is_invariant(n)) { // known invariant
 599       _old_new.map(n->_idx, n);
 600     } else { // to be cloned
 601       assert(!n->is_CFG(), "should not see CFG here");
 602       _stack.push(n, n->in(0) == NULL ? 1 : 0);
 603     }
 604   }
 605 
 606   // Clone "n" and (possibly) all its inputs recursively
 607   void clone_nodes(Node* n, Node* ctrl) {
 608     clone_visit(n);
 609     while (_stack.is_nonempty()) {
 610       Node*  n = _stack.node();
 611       uint idx = _stack.index();
 612       if (idx == n->req()) { // all inputs processed, clone n!
 613         _stack.pop();
 614         // clone invariant node
 615         Node* n_cl = n->clone();
 616         _old_new.map(n->_idx, n_cl);
 617         _phase->register_new_node(n_cl, ctrl);
 618         for (uint i = 0; i < n->req(); i++) {
 619           Node* in = n_cl->in(i);
 620           if (in == NULL) continue;
 621           n_cl->set_req(i, _old_new[in->_idx]);
 622         }
 623       } else { // process next input
 624         _stack.set_index(idx + 1);
 625         Node* m = n->in(idx);
 626         if (m != NULL && !_clone_visited.test_set(m->_idx)) {
 627           clone_visit(m); // visit the input
 628         }
 629       }
 630     }
 631   }
 632 
 633  public:
 634   Invariance(Arena* area, IdealLoopTree* lpt) :
 635     _lpt(lpt), _phase(lpt->_phase),
 636     _visited(area), _invariant(area), _stack(area, 10 /* guess */),
 637     _clone_visited(area), _old_new(area)
 638   {
 639     LoopNode* head = _lpt->_head->as_Loop();
 640     Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl);
 641     if (entry->outcnt() != 1) {
 642       // If a node is pinned between the predicates and the loop
 643       // entry, we won't be able to move any node in the loop that
 644       // depends on it above it in a predicate. Mark all those nodes
 645       // as non loop invariatnt.
 646       Unique_Node_List wq;
 647       wq.push(entry);
 648       for (uint next = 0; next < wq.size(); ++next) {
 649         Node *n = wq.at(next);
 650         for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 651           Node* u = n->fast_out(i);
 652           if (!u->is_CFG()) {
 653             Node* c = _phase->get_ctrl(u);
 654             if (_lpt->is_member(_phase->get_loop(c)) || _phase->is_dominator(c, head)) {
 655               _visited.set(u->_idx);
 656               wq.push(u);
 657             }
 658           }
 659         }
 660       }
 661     }
 662   }
 663 
 664   // Map old to n for invariance computation and clone
 665   void map_ctrl(Node* old, Node* n) {
 666     assert(old->is_CFG() && n->is_CFG(), "must be");
 667     _old_new.map(old->_idx, n); // "clone" of old is n
 668     _invariant.set(old->_idx);  // old is invariant
 669     _clone_visited.set(old->_idx);
 670   }
 671 
 672   // Driver function to compute invariance
 673   bool is_invariant(Node* n) {
 674     if (!_visited.test_set(n->_idx))
 675       compute_invariance(n);
 676     return (_invariant.test(n->_idx) != 0);
 677   }
 678 
 679   // Driver function to clone invariant
 680   Node* clone(Node* n, Node* ctrl) {
 681     assert(ctrl->is_CFG(), "must be");
 682     assert(_invariant.test(n->_idx), "must be an invariant");
 683     if (!_clone_visited.test(n->_idx))
 684       clone_nodes(n, ctrl);
 685     return _old_new[n->_idx];
 686   }
 687 };
 688 
 689 //------------------------------is_range_check_if -----------------------------------
 690 // Returns true if the predicate of iff is in "scale*iv + offset u< load_range(ptr)" format
 691 // Note: this function is particularly designed for loop predication. We require load_range
 692 //       and offset to be loop invariant computed on the fly by "invar"
 693 bool IdealLoopTree::is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const {
 694   if (!is_loop_exit(iff)) {
 695     return false;
 696   }
 697   if (!iff->in(1)->is_Bool()) {
 698     return false;
 699   }
 700   const BoolNode *bol = iff->in(1)->as_Bool();
 701   if (bol->_test._test != BoolTest::lt) {
 702     return false;
 703   }
 704   if (!bol->in(1)->is_Cmp()) {
 705     return false;
 706   }
 707   const CmpNode *cmp = bol->in(1)->as_Cmp();
 708   if (cmp->Opcode() != Op_CmpU) {
 709     return false;
 710   }
 711   Node* range = cmp->in(2);
 712   if (range->Opcode() != Op_LoadRange && !iff->is_RangeCheck()) {
 713     const TypeInt* tint = phase->_igvn.type(range)->isa_int();
 714     if (tint == NULL || tint->empty() || tint->_lo < 0) {
 715       // Allow predication on positive values that aren't LoadRanges.
 716       // This allows optimization of loops where the length of the
 717       // array is a known value and doesn't need to be loaded back
 718       // from the array.
 719       return false;
 720     }
 721   }
 722   if (!invar.is_invariant(range)) {
 723     return false;
 724   }
 725   Node *iv     = _head->as_CountedLoop()->phi();
 726   int   scale  = 0;
 727   Node *offset = NULL;
 728   if (!phase->is_scaled_iv_plus_offset(cmp->in(1), iv, &scale, &offset)) {
 729     return false;
 730   }
 731   if (offset && !invar.is_invariant(offset)) { // offset must be invariant
 732     return false;
 733   }
 734   return true;
 735 }
 736 
 737 //------------------------------rc_predicate-----------------------------------
 738 // Create a range check predicate
 739 //
 740 // for (i = init; i < limit; i += stride) {
 741 //    a[scale*i+offset]
 742 // }
 743 //
 744 // Compute max(scale*i + offset) for init <= i < limit and build the predicate
 745 // as "max(scale*i + offset) u< a.length".
 746 //
 747 // There are two cases for max(scale*i + offset):
 748 // (1) stride*scale > 0
 749 //   max(scale*i + offset) = scale*(limit-stride) + offset
 750 // (2) stride*scale < 0
 751 //   max(scale*i + offset) = scale*init + offset
 752 BoolNode* PhaseIdealLoop::rc_predicate(IdealLoopTree *loop, Node* ctrl,
 753                                        int scale, Node* offset,
 754                                        Node* init, Node* limit, jint stride,
 755                                        Node* range, bool upper, bool &overflow) {
 756   jint con_limit  = (limit != NULL && limit->is_Con())  ? limit->get_int()  : 0;
 757   jint con_init   = init->is_Con()   ? init->get_int()   : 0;
 758   jint con_offset = offset->is_Con() ? offset->get_int() : 0;
 759 
 760   stringStream* predString = NULL;
 761   if (TraceLoopPredicate) {
 762     predString = new stringStream();
 763     predString->print("rc_predicate ");
 764   }
 765 
 766   overflow = false;
 767   Node* max_idx_expr = NULL;
 768   const TypeInt* idx_type = TypeInt::INT;
 769   if ((stride > 0) == (scale > 0) == upper) {
 770     guarantee(limit != NULL, "sanity");
 771     if (TraceLoopPredicate) {
 772       if (limit->is_Con()) {
 773         predString->print("(%d ", con_limit);
 774       } else {
 775         predString->print("(limit ");
 776       }
 777       predString->print("- %d) ", stride);
 778     }
 779     // Check if (limit - stride) may overflow
 780     const TypeInt* limit_type = _igvn.type(limit)->isa_int();
 781     jint limit_lo = limit_type->_lo;
 782     jint limit_hi = limit_type->_hi;
 783     if ((stride > 0 && (java_subtract(limit_lo, stride) < limit_lo)) ||
 784         (stride < 0 && (java_subtract(limit_hi, stride) > limit_hi))) {
 785       // No overflow possible
 786       ConINode* con_stride = _igvn.intcon(stride);
 787       set_ctrl(con_stride, C->root());
 788       max_idx_expr = new SubINode(limit, con_stride);
 789       idx_type = TypeInt::make(limit_lo - stride, limit_hi - stride, limit_type->_widen);
 790     } else {
 791       // May overflow
 792       overflow = true;
 793       limit = new ConvI2LNode(limit);
 794       register_new_node(limit, ctrl);
 795       ConLNode* con_stride = _igvn.longcon(stride);
 796       set_ctrl(con_stride, C->root());
 797       max_idx_expr = new SubLNode(limit, con_stride);
 798     }
 799     register_new_node(max_idx_expr, ctrl);
 800   } else {
 801     if (TraceLoopPredicate) {
 802       if (init->is_Con()) {
 803         predString->print("%d ", con_init);
 804       } else {
 805         predString->print("init ");
 806       }
 807     }
 808     idx_type = _igvn.type(init)->isa_int();
 809     max_idx_expr = init;
 810   }
 811 
 812   if (scale != 1) {
 813     ConNode* con_scale = _igvn.intcon(scale);
 814     set_ctrl(con_scale, C->root());
 815     if (TraceLoopPredicate) {
 816       predString->print("* %d ", scale);
 817     }
 818     // Check if (scale * max_idx_expr) may overflow
 819     const TypeInt* scale_type = TypeInt::make(scale);
 820     MulINode* mul = new MulINode(max_idx_expr, con_scale);
 821     idx_type = (TypeInt*)mul->mul_ring(idx_type, scale_type);
 822     if (overflow || TypeInt::INT->higher_equal(idx_type)) {
 823       // May overflow
 824       mul->destruct();
 825       if (!overflow) {
 826         max_idx_expr = new ConvI2LNode(max_idx_expr);
 827         register_new_node(max_idx_expr, ctrl);
 828       }
 829       overflow = true;
 830       con_scale = _igvn.longcon(scale);
 831       set_ctrl(con_scale, C->root());
 832       max_idx_expr = new MulLNode(max_idx_expr, con_scale);
 833     } else {
 834       // No overflow possible
 835       max_idx_expr = mul;
 836     }
 837     register_new_node(max_idx_expr, ctrl);
 838   }
 839 
 840   if (offset && (!offset->is_Con() || con_offset != 0)){
 841     if (TraceLoopPredicate) {
 842       if (offset->is_Con()) {
 843         predString->print("+ %d ", con_offset);
 844       } else {
 845         predString->print("+ offset");
 846       }
 847     }
 848     // Check if (max_idx_expr + offset) may overflow
 849     const TypeInt* offset_type = _igvn.type(offset)->isa_int();
 850     jint lo = java_add(idx_type->_lo, offset_type->_lo);
 851     jint hi = java_add(idx_type->_hi, offset_type->_hi);
 852     if (overflow || (lo > hi) ||
 853         ((idx_type->_lo & offset_type->_lo) < 0 && lo >= 0) ||
 854         ((~(idx_type->_hi | offset_type->_hi)) < 0 && hi < 0)) {
 855       // May overflow
 856       if (!overflow) {
 857         max_idx_expr = new ConvI2LNode(max_idx_expr);
 858         register_new_node(max_idx_expr, ctrl);
 859       }
 860       overflow = true;
 861       offset = new ConvI2LNode(offset);
 862       register_new_node(offset, ctrl);
 863       max_idx_expr = new AddLNode(max_idx_expr, offset);
 864     } else {
 865       // No overflow possible
 866       max_idx_expr = new AddINode(max_idx_expr, offset);
 867     }
 868     register_new_node(max_idx_expr, ctrl);
 869   }
 870 
 871   CmpNode* cmp = NULL;
 872   if (overflow) {
 873     // Integer expressions may overflow, do long comparison
 874     range = new ConvI2LNode(range);
 875     register_new_node(range, ctrl);
 876     cmp = new CmpULNode(max_idx_expr, range);
 877   } else {
 878     cmp = new CmpUNode(max_idx_expr, range);
 879   }
 880   register_new_node(cmp, ctrl);
 881   BoolNode* bol = new BoolNode(cmp, BoolTest::lt);
 882   register_new_node(bol, ctrl);
 883 
 884   if (TraceLoopPredicate) {
 885     predString->print_cr("<u range");
 886     tty->print("%s", predString->as_string());
 887   }
 888   return bol;
 889 }
 890 
 891 // Should loop predication look not only in the path from tail to head
 892 // but also in branches of the loop body?
 893 bool PhaseIdealLoop::loop_predication_should_follow_branches(IdealLoopTree *loop, ProjNode *predicate_proj, float& loop_trip_cnt) {
 894   if (!UseProfiledLoopPredicate) {
 895     return false;
 896   }
 897 
 898   if (predicate_proj == NULL) {
 899     return false;
 900   }
 901 
 902   LoopNode* head = loop->_head->as_Loop();
 903   bool follow_branches = true;
 904   IdealLoopTree* l = loop->_child;
 905   // For leaf loops and loops with a single inner loop
 906   while (l != NULL && follow_branches) {
 907     IdealLoopTree* child = l;
 908     if (child->_child != NULL &&
 909         child->_head->is_OuterStripMinedLoop()) {
 910       assert(child->_child->_next == NULL, "only one inner loop for strip mined loop");
 911       assert(child->_child->_head->is_CountedLoop() && child->_child->_head->as_CountedLoop()->is_strip_mined(), "inner loop should be strip mined");
 912       child = child->_child;
 913     }
 914     if (child->_child != NULL || child->_irreducible) {
 915       follow_branches = false;
 916     }
 917     l = l->_next;
 918   }
 919   if (follow_branches) {
 920     loop->compute_profile_trip_cnt(this);
 921     if (head->is_profile_trip_failed()) {
 922       follow_branches = false;
 923     } else {
 924       loop_trip_cnt = head->profile_trip_cnt();
 925       if (head->is_CountedLoop()) {
 926         CountedLoopNode* cl = head->as_CountedLoop();
 927         if (cl->phi() != NULL) {
 928           const TypeInt* t = _igvn.type(cl->phi())->is_int();
 929           float worst_case_trip_cnt = ((float)t->_hi - t->_lo) / ABS(cl->stride_con());
 930           if (worst_case_trip_cnt < loop_trip_cnt) {
 931             loop_trip_cnt = worst_case_trip_cnt;
 932           }
 933         }
 934       }
 935     }
 936   }
 937   return follow_branches;
 938 }
 939 
 940 // Compute probability of reaching some CFG node from a fixed
 941 // dominating CFG node
 942 class PathFrequency {
 943 private:
 944   Node* _dom; // frequencies are computed relative to this node
 945   Node_Stack _stack;
 946   GrowableArray<float> _freqs_stack; // keep track of intermediate result at regions
 947   GrowableArray<float> _freqs; // cache frequencies
 948   PhaseIdealLoop* _phase;
 949 
 950   void set_rounding(int mode) {
 951     // fesetround is broken on windows
 952     NOT_WINDOWS(fesetround(mode);)
 953   }
 954 
 955   void check_frequency(float f) {
 956     NOT_WINDOWS(assert(f <= 1 && f >= 0, "Incorrect frequency");)
 957   }
 958 
 959 public:
 960   PathFrequency(Node* dom, PhaseIdealLoop* phase)
 961     : _dom(dom), _stack(0), _phase(phase) {
 962   }
 963 
 964   float to(Node* n) {
 965     // post order walk on the CFG graph from n to _dom
 966     set_rounding(FE_TOWARDZERO); // make sure rounding doesn't push frequency above 1
 967     IdealLoopTree* loop = _phase->get_loop(_dom);
 968     Node* c = n;
 969     for (;;) {
 970       assert(_phase->get_loop(c) == loop, "have to be in the same loop");
 971       if (c == _dom || _freqs.at_grow(c->_idx, -1) >= 0) {
 972         float f = c == _dom ? 1 : _freqs.at(c->_idx);
 973         Node* prev = c;
 974         while (_stack.size() > 0 && prev == c) {
 975           Node* n = _stack.node();
 976           if (!n->is_Region()) {
 977             if (_phase->get_loop(n) != _phase->get_loop(n->in(0))) {
 978               // Found an inner loop: compute frequency of reaching this
 979               // exit from the loop head by looking at the number of
 980               // times each loop exit was taken
 981               IdealLoopTree* inner_loop = _phase->get_loop(n->in(0));
 982               LoopNode* inner_head = inner_loop->_head->as_Loop();
 983               assert(_phase->get_loop(n) == loop, "only 1 inner loop");
 984               if (inner_head->is_OuterStripMinedLoop()) {
 985                 inner_head->verify_strip_mined(1);
 986                 if (n->in(0) == inner_head->in(LoopNode::LoopBackControl)->in(0)) {
 987                   n = n->in(0)->in(0)->in(0);
 988                 }
 989                 inner_loop = inner_loop->_child;
 990                 inner_head = inner_loop->_head->as_Loop();
 991                 inner_head->verify_strip_mined(1);
 992               }
 993               set_rounding(FE_UPWARD);  // make sure rounding doesn't push frequency above 1
 994               float loop_exit_cnt = 0.0f;
 995               for (uint i = 0; i < inner_loop->_body.size(); i++) {
 996                 Node *n = inner_loop->_body[i];
 997                 float c = inner_loop->compute_profile_trip_cnt_helper(n);
 998                 loop_exit_cnt += c;
 999               }
1000               set_rounding(FE_TOWARDZERO);
1001               float cnt = -1;
1002               if (n->in(0)->is_If()) {
1003                 IfNode* iff = n->in(0)->as_If();
1004                 float p = n->in(0)->as_If()->_prob;
1005                 if (n->Opcode() == Op_IfFalse) {
1006                   p = 1 - p;
1007                 }
1008                 if (p > PROB_MIN) {
1009                   cnt = p * iff->_fcnt;
1010                 } else {
1011                   cnt = 0;
1012                 }
1013               } else {
1014                 assert(n->in(0)->is_Jump(), "unsupported node kind");
1015                 JumpNode* jmp = n->in(0)->as_Jump();
1016                 float p = n->in(0)->as_Jump()->_probs[n->as_JumpProj()->_con];
1017                 cnt = p * jmp->_fcnt;
1018               }
1019               float this_exit_f = cnt > 0 ? cnt / loop_exit_cnt : 0;
1020               check_frequency(this_exit_f);
1021               f = f * this_exit_f;
1022               check_frequency(f);
1023             } else {
1024               float p = -1;
1025               if (n->in(0)->is_If()) {
1026                 p = n->in(0)->as_If()->_prob;
1027                 if (n->Opcode() == Op_IfFalse) {
1028                   p = 1 - p;
1029                 }
1030               } else {
1031                 assert(n->in(0)->is_Jump(), "unsupported node kind");
1032                 p = n->in(0)->as_Jump()->_probs[n->as_JumpProj()->_con];
1033               }
1034               f = f * p;
1035               check_frequency(f);
1036             }
1037             _freqs.at_put_grow(n->_idx, (float)f, -1);
1038             _stack.pop();
1039           } else {
1040             float prev_f = _freqs_stack.pop();
1041             float new_f = f;
1042             f = new_f + prev_f;
1043             check_frequency(f);
1044             uint i = _stack.index();
1045             if (i < n->req()) {
1046               c = n->in(i);
1047               _stack.set_index(i+1);
1048               _freqs_stack.push(f);
1049             } else {
1050               _freqs.at_put_grow(n->_idx, f, -1);
1051               _stack.pop();
1052             }
1053           }
1054         }
1055         if (_stack.size() == 0) {
1056           set_rounding(FE_TONEAREST);
1057           check_frequency(f);
1058           return f;
1059         }
1060       } else if (c->is_Loop()) {
1061         ShouldNotReachHere();
1062         c = c->in(LoopNode::EntryControl);
1063       } else if (c->is_Region()) {
1064         _freqs_stack.push(0);
1065         _stack.push(c, 2);
1066         c = c->in(1);
1067       } else {
1068         if (c->is_IfProj()) {
1069           IfNode* iff = c->in(0)->as_If();
1070           if (iff->_prob == PROB_UNKNOWN) {
1071             // assume never taken
1072             _freqs.at_put_grow(c->_idx, 0, -1);
1073           } else if (_phase->get_loop(c) != _phase->get_loop(iff)) {
1074             if (iff->_fcnt == COUNT_UNKNOWN) {
1075               // assume never taken
1076               _freqs.at_put_grow(c->_idx, 0, -1);
1077             } else {
1078               // skip over loop
1079               _stack.push(c, 1);
1080               c = _phase->get_loop(c->in(0))->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
1081             }
1082           } else {
1083             _stack.push(c, 1);
1084             c = iff;
1085           }
1086         } else if (c->is_JumpProj()) {
1087           JumpNode* jmp = c->in(0)->as_Jump();
1088           if (_phase->get_loop(c) != _phase->get_loop(jmp)) {
1089             if (jmp->_fcnt == COUNT_UNKNOWN) {
1090               // assume never taken
1091               _freqs.at_put_grow(c->_idx, 0, -1);
1092             } else {
1093               // skip over loop
1094               _stack.push(c, 1);
1095               c = _phase->get_loop(c->in(0))->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
1096             }
1097           } else {
1098             _stack.push(c, 1);
1099             c = jmp;
1100           }
1101         } else if (c->Opcode() == Op_CatchProj &&
1102                    c->in(0)->Opcode() == Op_Catch &&
1103                    c->in(0)->in(0)->is_Proj() &&
1104                    c->in(0)->in(0)->in(0)->is_Call()) {
1105           // assume exceptions are never thrown
1106           uint con = c->as_Proj()->_con;
1107           if (con == CatchProjNode::fall_through_index) {
1108             Node* call = c->in(0)->in(0)->in(0)->in(0);
1109             if (_phase->get_loop(call) != _phase->get_loop(c)) {
1110               _freqs.at_put_grow(c->_idx, 0, -1);
1111             } else {
1112               c = call;
1113             }
1114           } else {
1115             assert(con >= CatchProjNode::catch_all_index, "what else?");
1116             _freqs.at_put_grow(c->_idx, 0, -1);
1117           }
1118         } else if (c->unique_ctrl_out() == NULL && !c->is_If() && !c->is_Jump()) {
1119           ShouldNotReachHere();
1120         } else {
1121           c = c->in(0);
1122         }
1123       }
1124     }
1125     ShouldNotReachHere();
1126     return -1;
1127   }
1128 };
1129 
1130 void PhaseIdealLoop::loop_predication_follow_branches(Node *n, IdealLoopTree *loop, float loop_trip_cnt,
1131                                                       PathFrequency& pf, Node_Stack& stack, VectorSet& seen,
1132                                                       Node_List& if_proj_list) {
1133   assert(n->is_Region(), "start from a region");
1134   Node* tail = loop->tail();
1135   stack.push(n, 1);
1136   do {
1137     Node* c = stack.node();
1138     assert(c->is_Region() || c->is_IfProj(), "only region here");
1139     uint i = stack.index();
1140 
1141     if (i < c->req()) {
1142       stack.set_index(i+1);
1143       Node* in = c->in(i);
1144       while (!is_dominator(in, tail) && !seen.test_set(in->_idx)) {
1145         IdealLoopTree* in_loop = get_loop(in);
1146         if (in_loop != loop) {
1147           in = in_loop->_head->in(LoopNode::EntryControl);
1148         } else if (in->is_Region()) {
1149           stack.push(in, 1);
1150           break;
1151         } else if (in->is_IfProj() &&
1152                    in->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) &&
1153                    (in->in(0)->Opcode() == Op_If ||
1154                     in->in(0)->Opcode() == Op_RangeCheck)) {
1155           if (pf.to(in) * loop_trip_cnt >= 1) {
1156             stack.push(in, 1);
1157           }
1158           in = in->in(0);
1159         } else {
1160           in = in->in(0);
1161         }
1162       }
1163     } else {
1164       if (c->is_IfProj()) {
1165         if_proj_list.push(c);
1166       }
1167       stack.pop();
1168     }
1169 
1170   } while (stack.size() > 0);
1171 }
1172 
1173 
1174 bool PhaseIdealLoop::loop_predication_impl_helper(IdealLoopTree *loop, ProjNode* proj, ProjNode *predicate_proj,
1175                                                   CountedLoopNode *cl, ConNode* zero, Invariance& invar,
1176                                                   Deoptimization::DeoptReason reason) {
1177   // Following are changed to nonnull when a predicate can be hoisted
1178   ProjNode* new_predicate_proj = NULL;
1179   IfNode*   iff  = proj->in(0)->as_If();
1180   Node*     test = iff->in(1);
1181   if (!test->is_Bool()){ //Conv2B, ...
1182     return false;
1183   }
1184   BoolNode* bol = test->as_Bool();
1185   if (invar.is_invariant(bol)) {
1186     // Invariant test
1187     new_predicate_proj = create_new_if_for_predicate(predicate_proj, NULL,
1188                                                      reason,
1189                                                      iff->Opcode());
1190     Node* ctrl = new_predicate_proj->in(0)->as_If()->in(0);
1191     BoolNode* new_predicate_bol = invar.clone(bol, ctrl)->as_Bool();
1192 
1193     // Negate test if necessary
1194     bool negated = false;
1195     if (proj->_con != predicate_proj->_con) {
1196       new_predicate_bol = new BoolNode(new_predicate_bol->in(1), new_predicate_bol->_test.negate());
1197       register_new_node(new_predicate_bol, ctrl);
1198       negated = true;
1199     }
1200     IfNode* new_predicate_iff = new_predicate_proj->in(0)->as_If();
1201     _igvn.hash_delete(new_predicate_iff);
1202     new_predicate_iff->set_req(1, new_predicate_bol);
1203 #ifndef PRODUCT
1204     if (TraceLoopPredicate) {
1205       tty->print("Predicate invariant if%s: %d ", negated ? " negated" : "", new_predicate_iff->_idx);
1206       loop->dump_head();
1207     } else if (TraceLoopOpts) {
1208       tty->print("Predicate IC ");
1209       loop->dump_head();
1210     }
1211 #endif
1212   } else if (cl != NULL && loop->is_range_check_if(iff, this, invar)) {
1213     // Range check for counted loops
1214     const Node*    cmp    = bol->in(1)->as_Cmp();
1215     Node*          idx    = cmp->in(1);
1216     assert(!invar.is_invariant(idx), "index is variant");
1217     Node* rng = cmp->in(2);
1218     assert(rng->Opcode() == Op_LoadRange || iff->is_RangeCheck() || _igvn.type(rng)->is_int()->_lo >= 0, "must be");
1219     assert(invar.is_invariant(rng), "range must be invariant");
1220     int scale    = 1;
1221     Node* offset = zero;
1222     bool ok = is_scaled_iv_plus_offset(idx, cl->phi(), &scale, &offset);
1223     assert(ok, "must be index expression");
1224 
1225     Node* init    = cl->init_trip();
1226     // Limit is not exact.
1227     // Calculate exact limit here.
1228     // Note, counted loop's test is '<' or '>'.
1229     Node* limit   = exact_limit(loop);
1230     int  stride   = cl->stride()->get_int();
1231 
1232     // Build if's for the upper and lower bound tests.  The
1233     // lower_bound test will dominate the upper bound test and all
1234     // cloned or created nodes will use the lower bound test as
1235     // their declared control.
1236 
1237     // Perform cloning to keep Invariance state correct since the
1238     // late schedule will place invariant things in the loop.
1239     Node *ctrl = predicate_proj->in(0)->as_If()->in(0);
1240     rng = invar.clone(rng, ctrl);
1241     if (offset && offset != zero) {
1242       assert(invar.is_invariant(offset), "offset must be loop invariant");
1243       offset = invar.clone(offset, ctrl);
1244     }
1245     // If predicate expressions may overflow in the integer range, longs are used.
1246     bool overflow = false;
1247 
1248     // Test the lower bound
1249     BoolNode* lower_bound_bol = rc_predicate(loop, ctrl, scale, offset, init, limit, stride, rng, false, overflow);
1250     // Negate test if necessary
1251     bool negated = false;
1252     if (proj->_con != predicate_proj->_con) {
1253       lower_bound_bol = new BoolNode(lower_bound_bol->in(1), lower_bound_bol->_test.negate());
1254       register_new_node(lower_bound_bol, ctrl);
1255       negated = true;
1256     }
1257     ProjNode* lower_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, reason, overflow ? Op_If : iff->Opcode());
1258     IfNode* lower_bound_iff = lower_bound_proj->in(0)->as_If();
1259     _igvn.hash_delete(lower_bound_iff);
1260     lower_bound_iff->set_req(1, lower_bound_bol);
1261     if (TraceLoopPredicate) tty->print_cr("lower bound check if: %s %d ", negated ? " negated" : "", lower_bound_iff->_idx);
1262 
1263     // Test the upper bound
1264     BoolNode* upper_bound_bol = rc_predicate(loop, lower_bound_proj, scale, offset, init, limit, stride, rng, true, overflow);
1265     negated = false;
1266     if (proj->_con != predicate_proj->_con) {
1267       upper_bound_bol = new BoolNode(upper_bound_bol->in(1), upper_bound_bol->_test.negate());
1268       register_new_node(upper_bound_bol, ctrl);
1269       negated = true;
1270     }
1271     ProjNode* upper_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, reason, overflow ? Op_If : iff->Opcode());
1272     assert(upper_bound_proj->in(0)->as_If()->in(0) == lower_bound_proj, "should dominate");
1273     IfNode* upper_bound_iff = upper_bound_proj->in(0)->as_If();
1274     _igvn.hash_delete(upper_bound_iff);
1275     upper_bound_iff->set_req(1, upper_bound_bol);
1276     if (TraceLoopPredicate) tty->print_cr("upper bound check if: %s %d ", negated ? " negated" : "", lower_bound_iff->_idx);
1277 
1278     // Fall through into rest of the clean up code which will move
1279     // any dependent nodes onto the upper bound test.
1280     new_predicate_proj = upper_bound_proj;
1281 
1282     if (iff->is_RangeCheck()) {
1283       new_predicate_proj = insert_skeleton_predicate(iff, loop, proj, predicate_proj, upper_bound_proj, scale, offset, init, limit, stride, rng, overflow, reason);
1284     }
1285 
1286 #ifndef PRODUCT
1287     if (TraceLoopOpts && !TraceLoopPredicate) {
1288       tty->print("Predicate RC ");
1289       loop->dump_head();
1290     }
1291 #endif
1292   } else {
1293     // Loop variant check (for example, range check in non-counted loop)
1294     // with uncommon trap.
1295     return false;
1296   }
1297   assert(new_predicate_proj != NULL, "sanity");
1298   // Success - attach condition (new_predicate_bol) to predicate if
1299   invar.map_ctrl(proj, new_predicate_proj); // so that invariance test can be appropriate
1300 
1301   // Eliminate the old If in the loop body
1302   dominated_by( new_predicate_proj, iff, proj->_con != new_predicate_proj->_con );
1303 
1304   C->set_major_progress();
1305   return true;
1306 }
1307 
1308 
1309 // After pre/main/post loops are created, we'll put a copy of some
1310 // range checks between the pre and main loop to validate the value
1311 // of the main loop induction variable. Make a copy of the predicates
1312 // here with an opaque node as a place holder for the value (will be
1313 // updated by PhaseIdealLoop::update_skeleton_predicate()).
1314 ProjNode* PhaseIdealLoop::insert_skeleton_predicate(IfNode* iff, IdealLoopTree *loop,
1315                                                     ProjNode* proj, ProjNode *predicate_proj,
1316                                                     ProjNode* upper_bound_proj,
1317                                                     int scale, Node* offset,
1318                                                     Node* init, Node* limit, jint stride,
1319                                                     Node* rng, bool &overflow,
1320                                                     Deoptimization::DeoptReason reason) {
1321   assert(proj->_con && predicate_proj->_con, "not a range check?");
1322   Node* opaque_init = new Opaque1Node(C, init);
1323   register_new_node(opaque_init, upper_bound_proj);
1324   BoolNode* bol = rc_predicate(loop, upper_bound_proj, scale, offset, opaque_init, limit, stride, rng, (stride > 0) != (scale > 0), overflow);
1325   Node* opaque_bol = new Opaque4Node(C, bol, _igvn.intcon(1)); // This will go away once loop opts are over
1326   register_new_node(opaque_bol, upper_bound_proj);
1327   ProjNode* new_proj = create_new_if_for_predicate(predicate_proj, NULL, reason, overflow ? Op_If : iff->Opcode());
1328   _igvn.replace_input_of(new_proj->in(0), 1, opaque_bol);
1329   assert(opaque_init->outcnt() > 0, "should be used");
1330   return new_proj;
1331 }
1332 
1333 //------------------------------ loop_predication_impl--------------------------
1334 // Insert loop predicates for null checks and range checks
1335 bool PhaseIdealLoop::loop_predication_impl(IdealLoopTree *loop) {
1336   if (!UseLoopPredicate) return false;
1337 
1338   if (!loop->_head->is_Loop()) {
1339     // Could be a simple region when irreducible loops are present.
1340     return false;
1341   }
1342   LoopNode* head = loop->_head->as_Loop();
1343 
1344   if (head->unique_ctrl_out()->Opcode() == Op_NeverBranch) {
1345     // do nothing for infinite loops
1346     return false;
1347   }
1348 
1349   if (head->is_OuterStripMinedLoop()) {
1350     return false;
1351   }
1352 
1353   CountedLoopNode *cl = NULL;
1354   if (head->is_valid_counted_loop()) {
1355     cl = head->as_CountedLoop();
1356     // do nothing for iteration-splitted loops
1357     if (!cl->is_normal_loop()) return false;
1358     // Avoid RCE if Counted loop's test is '!='.
1359     BoolTest::mask bt = cl->loopexit()->test_trip();
1360     if (bt != BoolTest::lt && bt != BoolTest::gt)
1361       cl = NULL;
1362   }
1363 
1364   Node* entry = head->skip_strip_mined()->in(LoopNode::EntryControl);
1365   ProjNode *loop_limit_proj = NULL;
1366   ProjNode *predicate_proj = NULL;
1367   ProjNode *profile_predicate_proj = NULL;
1368   // Loop limit check predicate should be near the loop.
1369   loop_limit_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check);
1370   if (loop_limit_proj != NULL) {
1371     entry = skip_loop_predicates(loop_limit_proj);
1372   }
1373   bool has_profile_predicates = false;
1374   profile_predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate);
1375   if (profile_predicate_proj != NULL) {
1376     Node* n = skip_loop_predicates(entry);
1377     // Check if predicates were already added to the profile predicate
1378     // block
1379     if (n != entry->in(0)->in(0) || n->outcnt() != 1) {
1380       has_profile_predicates = true;
1381     }
1382     entry = n;
1383   }
1384   predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
1385 
1386   float loop_trip_cnt = -1;
1387   bool follow_branches = loop_predication_should_follow_branches(loop, profile_predicate_proj, loop_trip_cnt);
1388   assert(!follow_branches || loop_trip_cnt >= 0, "negative trip count?");
1389 
1390   if (predicate_proj == NULL && !follow_branches) {
1391 #ifndef PRODUCT
1392     if (TraceLoopPredicate) {
1393       tty->print("missing predicate:");
1394       loop->dump_head();
1395       head->dump(1);
1396     }
1397 #endif
1398     return false;
1399   }
1400   ConNode* zero = _igvn.intcon(0);
1401   set_ctrl(zero, C->root());
1402 
1403   ResourceArea *area = Thread::current()->resource_area();
1404   Invariance invar(area, loop);
1405 
1406   // Create list of if-projs such that a newer proj dominates all older
1407   // projs in the list, and they all dominate loop->tail()
1408   Node_List if_proj_list(area);
1409   Node_List regions(area);
1410   Node *current_proj = loop->tail(); //start from tail
1411 
1412 
1413   Node_List controls(area);
1414   while (current_proj != head) {
1415     if (loop == get_loop(current_proj) && // still in the loop ?
1416         current_proj->is_Proj()        && // is a projection  ?
1417         (current_proj->in(0)->Opcode() == Op_If ||
1418          current_proj->in(0)->Opcode() == Op_RangeCheck)) { // is a if projection ?
1419       if_proj_list.push(current_proj);
1420     }
1421     if (follow_branches &&
1422         current_proj->Opcode() == Op_Region &&
1423         loop == get_loop(current_proj)) {
1424       regions.push(current_proj);
1425     }
1426     current_proj = idom(current_proj);
1427   }
1428 
1429   bool hoisted = false; // true if at least one proj is promoted
1430 
1431   if (!has_profile_predicates) {
1432     while (if_proj_list.size() > 0) {
1433       Node* n = if_proj_list.pop();
1434 
1435       ProjNode* proj = n->as_Proj();
1436       IfNode*   iff  = proj->in(0)->as_If();
1437 
1438       CallStaticJavaNode* call = proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
1439       if (call == NULL) {
1440         if (loop->is_loop_exit(iff)) {
1441           // stop processing the remaining projs in the list because the execution of them
1442           // depends on the condition of "iff" (iff->in(1)).
1443           break;
1444         } else {
1445           // Both arms are inside the loop. There are two cases:
1446           // (1) there is one backward branch. In this case, any remaining proj
1447           //     in the if_proj list post-dominates "iff". So, the condition of "iff"
1448           //     does not determine the execution the remining projs directly, and we
1449           //     can safely continue.
1450           // (2) both arms are forwarded, i.e. a diamond shape. In this case, "proj"
1451           //     does not dominate loop->tail(), so it can not be in the if_proj list.
1452           continue;
1453         }
1454       }
1455       Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(call->uncommon_trap_request());
1456       if (reason == Deoptimization::Reason_predicate) {
1457         break;
1458       }
1459 
1460       if (predicate_proj != NULL) {
1461         hoisted = loop_predication_impl_helper(loop, proj, predicate_proj, cl, zero, invar, Deoptimization::Reason_predicate) | hoisted;
1462       }
1463     } // end while
1464   }
1465 
1466   Node_List if_proj_list_freq(area);
1467   if (follow_branches) {
1468     PathFrequency pf(loop->_head, this);
1469 
1470     // Some projections were skipped by regular predicates because of
1471     // an early loop exit. Try them with profile data.
1472     while (if_proj_list.size() > 0) {
1473       Node* proj = if_proj_list.pop();
1474       float f = pf.to(proj);
1475       if (proj->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) &&
1476           f * loop_trip_cnt >= 1) {
1477         hoisted = loop_predication_impl_helper(loop, proj->as_Proj(), profile_predicate_proj, cl, zero, invar, Deoptimization::Reason_profile_predicate) | hoisted;
1478       }
1479     }
1480 
1481     // And look into all branches
1482     Node_Stack stack(0);
1483     VectorSet seen(Thread::current()->resource_area());
1484     while (regions.size() > 0) {
1485       Node* c = regions.pop();
1486       loop_predication_follow_branches(c, loop, loop_trip_cnt, pf, stack, seen, if_proj_list_freq);
1487     }
1488 
1489     for (uint i = 0; i < if_proj_list_freq.size(); i++) {
1490       ProjNode* proj = if_proj_list_freq.at(i)->as_Proj();
1491       hoisted = loop_predication_impl_helper(loop, proj, profile_predicate_proj, cl, zero, invar, Deoptimization::Reason_profile_predicate) | hoisted;
1492     }
1493   }
1494 
1495 #ifndef PRODUCT
1496   // report that the loop predication has been actually performed
1497   // for this loop
1498   if (TraceLoopPredicate && hoisted) {
1499     tty->print("Loop Predication Performed:");
1500     loop->dump_head();
1501   }
1502 #endif
1503 
1504   head->verify_strip_mined(1);
1505 
1506   return hoisted;
1507 }
1508 
1509 //------------------------------loop_predication--------------------------------
1510 // driver routine for loop predication optimization
1511 bool IdealLoopTree::loop_predication( PhaseIdealLoop *phase) {
1512   bool hoisted = false;
1513   // Recursively promote predicates
1514   if (_child) {
1515     hoisted = _child->loop_predication( phase);
1516   }
1517 
1518   // self
1519   if (!_irreducible && !tail()->is_top()) {
1520     hoisted |= phase->loop_predication_impl(this);
1521   }
1522 
1523   if (_next) { //sibling
1524     hoisted |= _next->loop_predication( phase);
1525   }
1526 
1527   return hoisted;
1528 }