1 /* 2 * Copyright (c) 2011, 2014, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "opto/loopnode.hpp" 27 #include "opto/addnode.hpp" 28 #include "opto/callnode.hpp" 29 #include "opto/connode.hpp" 30 #include "opto/convertnode.hpp" 31 #include "opto/loopnode.hpp" 32 #include "opto/mulnode.hpp" 33 #include "opto/opaquenode.hpp" 34 #include "opto/rootnode.hpp" 35 #include "opto/subnode.hpp" 36 37 /* 38 * The general idea of Loop Predication is to insert a predicate on the entry 39 * path to a loop, and raise a uncommon trap if the check of the condition fails. 40 * The condition checks are promoted from inside the loop body, and thus 41 * the checks inside the loop could be eliminated. Currently, loop predication 42 * optimization has been applied to remove array range check and loop invariant 43 * checks (such as null checks). 44 */ 45 46 //-------------------------------register_control------------------------- 47 void PhaseIdealLoop::register_control(Node* n, IdealLoopTree *loop, Node* pred) { 48 assert(n->is_CFG(), "must be control node"); 49 _igvn.register_new_node_with_optimizer(n); 50 loop->_body.push(n); 51 set_loop(n, loop); 52 // When called from beautify_loops() idom is not constructed yet. 53 if (_idom != NULL) { 54 set_idom(n, pred, dom_depth(pred)); 55 } 56 } 57 58 //------------------------------create_new_if_for_predicate------------------------ 59 // create a new if above the uct_if_pattern for the predicate to be promoted. 60 // 61 // before after 62 // ---------- ---------- 63 // ctrl ctrl 64 // | | 65 // | | 66 // v v 67 // iff new_iff 68 // / \ / \ 69 // / \ / \ 70 // v v v v 71 // uncommon_proj cont_proj if_uct if_cont 72 // \ | | | | 73 // \ | | | | 74 // v v v | v 75 // rgn loop | iff 76 // | | / \ 77 // | | / \ 78 // v | v v 79 // uncommon_trap | uncommon_proj cont_proj 80 // \ \ | | 81 // \ \ | | 82 // v v v v 83 // rgn loop 84 // | 85 // | 86 // v 87 // uncommon_trap 88 // 89 // 90 // We will create a region to guard the uct call if there is no one there. 91 // The true projecttion (if_cont) of the new_iff is returned. 92 // This code is also used to clone predicates to cloned loops. 93 ProjNode* PhaseIdealLoop::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, 94 Deoptimization::DeoptReason reason, 95 int opcode) { 96 assert(cont_proj->is_uncommon_trap_if_pattern(reason), "must be a uct if pattern!"); 97 IfNode* iff = cont_proj->in(0)->as_If(); 98 99 ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con); 100 Node *rgn = uncommon_proj->unique_ctrl_out(); 101 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct"); 102 103 uint proj_index = 1; // region's edge corresponding to uncommon_proj 104 if (!rgn->is_Region()) { // create a region to guard the call 105 assert(rgn->is_Call(), "must be call uct"); 106 CallNode* call = rgn->as_Call(); 107 IdealLoopTree* loop = get_loop(call); 108 rgn = new RegionNode(1); 109 rgn->add_req(uncommon_proj); 110 register_control(rgn, loop, uncommon_proj); 111 _igvn.replace_input_of(call, 0, rgn); 112 // When called from beautify_loops() idom is not constructed yet. 113 if (_idom != NULL) { 114 set_idom(call, rgn, dom_depth(rgn)); 115 } 116 for (DUIterator_Fast imax, i = uncommon_proj->fast_outs(imax); i < imax; i++) { 117 Node* n = uncommon_proj->fast_out(i); 118 if (n->is_Load() || n->is_Store()) { 119 _igvn.replace_input_of(n, 0, rgn); 120 --i; --imax; 121 } 122 } 123 } else { 124 // Find region's edge corresponding to uncommon_proj 125 for (; proj_index < rgn->req(); proj_index++) 126 if (rgn->in(proj_index) == uncommon_proj) break; 127 assert(proj_index < rgn->req(), "sanity"); 128 } 129 130 Node* entry = iff->in(0); 131 if (new_entry != NULL) { 132 // Clonning the predicate to new location. 133 entry = new_entry; 134 } 135 // Create new_iff 136 IdealLoopTree* lp = get_loop(entry); 137 IfNode *new_iff = NULL; 138 if (opcode == Op_If) { 139 new_iff = new IfNode(entry, iff->in(1), iff->_prob, iff->_fcnt); 140 } else { 141 assert(opcode == Op_RangeCheck, "no other if variant here"); 142 new_iff = new RangeCheckNode(entry, iff->in(1), iff->_prob, iff->_fcnt); 143 } 144 register_control(new_iff, lp, entry); 145 Node *if_cont = new IfTrueNode(new_iff); 146 Node *if_uct = new IfFalseNode(new_iff); 147 if (cont_proj->is_IfFalse()) { 148 // Swap 149 Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp; 150 } 151 register_control(if_cont, lp, new_iff); 152 register_control(if_uct, get_loop(rgn), new_iff); 153 154 // if_uct to rgn 155 _igvn.hash_delete(rgn); 156 rgn->add_req(if_uct); 157 // When called from beautify_loops() idom is not constructed yet. 158 if (_idom != NULL) { 159 Node* ridom = idom(rgn); 160 Node* nrdom = dom_lca(ridom, new_iff); 161 set_idom(rgn, nrdom, dom_depth(rgn)); 162 } 163 164 // If rgn has phis add new edges which has the same 165 // value as on original uncommon_proj pass. 166 assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last"); 167 bool has_phi = false; 168 for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) { 169 Node* use = rgn->fast_out(i); 170 if (use->is_Phi() && use->outcnt() > 0) { 171 assert(use->in(0) == rgn, ""); 172 _igvn.rehash_node_delayed(use); 173 use->add_req(use->in(proj_index)); 174 has_phi = true; 175 } 176 } 177 assert(!has_phi || rgn->req() > 3, "no phis when region is created"); 178 179 if (new_entry == NULL) { 180 // Attach if_cont to iff 181 _igvn.replace_input_of(iff, 0, if_cont); 182 if (_idom != NULL) { 183 set_idom(iff, if_cont, dom_depth(iff)); 184 } 185 } 186 return if_cont->as_Proj(); 187 } 188 189 //------------------------------create_new_if_for_predicate------------------------ 190 // Create a new if below new_entry for the predicate to be cloned (IGVN optimization) 191 ProjNode* PhaseIterGVN::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, 192 Deoptimization::DeoptReason reason, 193 int opcode) { 194 assert(new_entry != 0, "only used for clone predicate"); 195 assert(cont_proj->is_uncommon_trap_if_pattern(reason), "must be a uct if pattern!"); 196 IfNode* iff = cont_proj->in(0)->as_If(); 197 198 ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con); 199 Node *rgn = uncommon_proj->unique_ctrl_out(); 200 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct"); 201 202 uint proj_index = 1; // region's edge corresponding to uncommon_proj 203 if (!rgn->is_Region()) { // create a region to guard the call 204 assert(rgn->is_Call(), "must be call uct"); 205 CallNode* call = rgn->as_Call(); 206 rgn = new RegionNode(1); 207 register_new_node_with_optimizer(rgn); 208 rgn->add_req(uncommon_proj); 209 replace_input_of(call, 0, rgn); 210 } else { 211 // Find region's edge corresponding to uncommon_proj 212 for (; proj_index < rgn->req(); proj_index++) 213 if (rgn->in(proj_index) == uncommon_proj) break; 214 assert(proj_index < rgn->req(), "sanity"); 215 } 216 217 // Create new_iff in new location. 218 IfNode *new_iff = NULL; 219 if (opcode == Op_If) { 220 new_iff = new IfNode(new_entry, iff->in(1), iff->_prob, iff->_fcnt); 221 } else { 222 assert(opcode == Op_RangeCheck, "no other if variant here"); 223 new_iff = new RangeCheckNode(new_entry, iff->in(1), iff->_prob, iff->_fcnt); 224 } 225 226 register_new_node_with_optimizer(new_iff); 227 Node *if_cont = new IfTrueNode(new_iff); 228 Node *if_uct = new IfFalseNode(new_iff); 229 if (cont_proj->is_IfFalse()) { 230 // Swap 231 Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp; 232 } 233 register_new_node_with_optimizer(if_cont); 234 register_new_node_with_optimizer(if_uct); 235 236 // if_uct to rgn 237 hash_delete(rgn); 238 rgn->add_req(if_uct); 239 240 // If rgn has phis add corresponding new edges which has the same 241 // value as on original uncommon_proj pass. 242 assert(rgn->in(rgn->req() -1) == if_uct, "new edge should be last"); 243 bool has_phi = false; 244 for (DUIterator_Fast imax, i = rgn->fast_outs(imax); i < imax; i++) { 245 Node* use = rgn->fast_out(i); 246 if (use->is_Phi() && use->outcnt() > 0) { 247 rehash_node_delayed(use); 248 use->add_req(use->in(proj_index)); 249 has_phi = true; 250 } 251 } 252 assert(!has_phi || rgn->req() > 3, "no phis when region is created"); 253 254 return if_cont->as_Proj(); 255 } 256 257 //--------------------------clone_predicate----------------------- 258 ProjNode* PhaseIdealLoop::clone_predicate(ProjNode* predicate_proj, Node* new_entry, 259 Deoptimization::DeoptReason reason, 260 PhaseIdealLoop* loop_phase, 261 PhaseIterGVN* igvn) { 262 ProjNode* new_predicate_proj; 263 if (loop_phase != NULL) { 264 new_predicate_proj = loop_phase->create_new_if_for_predicate(predicate_proj, new_entry, reason, Op_If); 265 } else { 266 new_predicate_proj = igvn->create_new_if_for_predicate(predicate_proj, new_entry, reason, Op_If); 267 } 268 IfNode* iff = new_predicate_proj->in(0)->as_If(); 269 Node* ctrl = iff->in(0); 270 271 // Match original condition since predicate's projections could be swapped. 272 assert(predicate_proj->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be"); 273 Node* opq = new Opaque1Node(igvn->C, predicate_proj->in(0)->in(1)->in(1)->in(1)); 274 igvn->C->add_predicate_opaq(opq); 275 276 Node* bol = new Conv2BNode(opq); 277 if (loop_phase != NULL) { 278 loop_phase->register_new_node(opq, ctrl); 279 loop_phase->register_new_node(bol, ctrl); 280 } else { 281 igvn->register_new_node_with_optimizer(opq); 282 igvn->register_new_node_with_optimizer(bol); 283 } 284 igvn->hash_delete(iff); 285 iff->set_req(1, bol); 286 return new_predicate_proj; 287 } 288 289 290 //--------------------------clone_loop_predicates----------------------- 291 // Interface from IGVN 292 Node* PhaseIterGVN::clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check) { 293 return PhaseIdealLoop::clone_loop_predicates(old_entry, new_entry, clone_limit_check, NULL, this); 294 } 295 296 // Interface from PhaseIdealLoop 297 Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check) { 298 return clone_loop_predicates(old_entry, new_entry, clone_limit_check, this, &this->_igvn); 299 } 300 301 // Clone loop predicates to cloned loops (peeled, unswitched, split_if). 302 Node* PhaseIdealLoop::clone_loop_predicates(Node* old_entry, Node* new_entry, 303 bool clone_limit_check, 304 PhaseIdealLoop* loop_phase, 305 PhaseIterGVN* igvn) { 306 #ifdef ASSERT 307 if (new_entry == NULL || !(new_entry->is_Proj() || new_entry->is_Region() || new_entry->is_SafePoint())) { 308 if (new_entry != NULL) 309 new_entry->dump(); 310 assert(false, "not IfTrue, IfFalse, Region or SafePoint"); 311 } 312 #endif 313 // Search original predicates 314 Node* entry = old_entry; 315 ProjNode* limit_check_proj = NULL; 316 if (LoopLimitCheck) { 317 limit_check_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); 318 if (limit_check_proj != NULL) { 319 entry = entry->in(0)->in(0); 320 } 321 } 322 if (UseLoopPredicate) { 323 ProjNode* predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); 324 if (predicate_proj != NULL) { // right pattern that can be used by loop predication 325 // clone predicate 326 new_entry = clone_predicate(predicate_proj, new_entry, 327 Deoptimization::Reason_predicate, 328 loop_phase, igvn); 329 assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone predicate"); 330 if (TraceLoopPredicate) { 331 tty->print("Loop Predicate cloned: "); 332 debug_only( new_entry->in(0)->dump(); ) 333 } 334 } 335 } 336 if (limit_check_proj != NULL && clone_limit_check) { 337 // Clone loop limit check last to insert it before loop. 338 // Don't clone a limit check which was already finalized 339 // for this counted loop (only one limit check is needed). 340 new_entry = clone_predicate(limit_check_proj, new_entry, 341 Deoptimization::Reason_loop_limit_check, 342 loop_phase, igvn); 343 assert(new_entry != NULL && new_entry->is_Proj(), "IfTrue or IfFalse after clone limit check"); 344 if (TraceLoopLimitCheck) { 345 tty->print("Loop Limit Check cloned: "); 346 debug_only( new_entry->in(0)->dump(); ) 347 } 348 } 349 return new_entry; 350 } 351 352 //--------------------------skip_loop_predicates------------------------------ 353 // Skip related predicates. 354 Node* PhaseIdealLoop::skip_loop_predicates(Node* entry) { 355 Node* predicate = NULL; 356 if (LoopLimitCheck) { 357 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); 358 if (predicate != NULL) { 359 entry = entry->in(0)->in(0); 360 } 361 } 362 if (UseLoopPredicate) { 363 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); 364 if (predicate != NULL) { // right pattern that can be used by loop predication 365 IfNode* iff = entry->in(0)->as_If(); 366 ProjNode* uncommon_proj = iff->proj_out(1 - entry->as_Proj()->_con); 367 Node* rgn = uncommon_proj->unique_ctrl_out(); 368 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct"); 369 entry = entry->in(0)->in(0); 370 while (entry != NULL && entry->is_Proj() && entry->in(0)->is_If()) { 371 uncommon_proj = entry->in(0)->as_If()->proj_out(1 - entry->as_Proj()->_con); 372 if (uncommon_proj->unique_ctrl_out() != rgn) 373 break; 374 entry = entry->in(0)->in(0); 375 } 376 } 377 } 378 return entry; 379 } 380 381 //--------------------------find_predicate_insertion_point------------------- 382 // Find a good location to insert a predicate 383 ProjNode* PhaseIdealLoop::find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason) { 384 if (start_c == NULL || !start_c->is_Proj()) 385 return NULL; 386 if (start_c->as_Proj()->is_uncommon_trap_if_pattern(reason)) { 387 return start_c->as_Proj(); 388 } 389 return NULL; 390 } 391 392 //--------------------------find_predicate------------------------------------ 393 // Find a predicate 394 Node* PhaseIdealLoop::find_predicate(Node* entry) { 395 Node* predicate = NULL; 396 if (LoopLimitCheck) { 397 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); 398 if (predicate != NULL) { // right pattern that can be used by loop predication 399 return entry; 400 } 401 } 402 if (UseLoopPredicate) { 403 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); 404 if (predicate != NULL) { // right pattern that can be used by loop predication 405 return entry; 406 } 407 } 408 return NULL; 409 } 410 411 //------------------------------Invariance----------------------------------- 412 // Helper class for loop_predication_impl to compute invariance on the fly and 413 // clone invariants. 414 class Invariance : public StackObj { 415 VectorSet _visited, _invariant; 416 Node_Stack _stack; 417 VectorSet _clone_visited; 418 Node_List _old_new; // map of old to new (clone) 419 IdealLoopTree* _lpt; 420 PhaseIdealLoop* _phase; 421 422 // Helper function to set up the invariance for invariance computation 423 // If n is a known invariant, set up directly. Otherwise, look up the 424 // the possibility to push n onto the stack for further processing. 425 void visit(Node* use, Node* n) { 426 if (_lpt->is_invariant(n)) { // known invariant 427 _invariant.set(n->_idx); 428 } else if (!n->is_CFG()) { 429 Node *n_ctrl = _phase->ctrl_or_self(n); 430 Node *u_ctrl = _phase->ctrl_or_self(use); // self if use is a CFG 431 if (_phase->is_dominator(n_ctrl, u_ctrl)) { 432 _stack.push(n, n->in(0) == NULL ? 1 : 0); 433 } 434 } 435 } 436 437 // Compute invariance for "the_node" and (possibly) all its inputs recursively 438 // on the fly 439 void compute_invariance(Node* n) { 440 assert(_visited.test(n->_idx), "must be"); 441 visit(n, n); 442 while (_stack.is_nonempty()) { 443 Node* n = _stack.node(); 444 uint idx = _stack.index(); 445 if (idx == n->req()) { // all inputs are processed 446 _stack.pop(); 447 // n is invariant if it's inputs are all invariant 448 bool all_inputs_invariant = true; 449 for (uint i = 0; i < n->req(); i++) { 450 Node* in = n->in(i); 451 if (in == NULL) continue; 452 assert(_visited.test(in->_idx), "must have visited input"); 453 if (!_invariant.test(in->_idx)) { // bad guy 454 all_inputs_invariant = false; 455 break; 456 } 457 } 458 if (all_inputs_invariant) { 459 // If n's control is a predicate that was moved out of the 460 // loop, it was marked invariant but n is only invariant if 461 // it depends only on that test. Otherwise, unless that test 462 // is out of the loop, it's not invariant. 463 if (n->is_CFG() || n->depends_only_on_test() || n->in(0) == NULL || !_phase->is_member(_lpt, n->in(0))) { 464 _invariant.set(n->_idx); // I am a invariant too 465 } 466 } 467 } else { // process next input 468 _stack.set_index(idx + 1); 469 Node* m = n->in(idx); 470 if (m != NULL && !_visited.test_set(m->_idx)) { 471 visit(n, m); 472 } 473 } 474 } 475 } 476 477 // Helper function to set up _old_new map for clone_nodes. 478 // If n is a known invariant, set up directly ("clone" of n == n). 479 // Otherwise, push n onto the stack for real cloning. 480 void clone_visit(Node* n) { 481 assert(_invariant.test(n->_idx), "must be invariant"); 482 if (_lpt->is_invariant(n)) { // known invariant 483 _old_new.map(n->_idx, n); 484 } else { // to be cloned 485 assert(!n->is_CFG(), "should not see CFG here"); 486 _stack.push(n, n->in(0) == NULL ? 1 : 0); 487 } 488 } 489 490 // Clone "n" and (possibly) all its inputs recursively 491 void clone_nodes(Node* n, Node* ctrl) { 492 clone_visit(n); 493 while (_stack.is_nonempty()) { 494 Node* n = _stack.node(); 495 uint idx = _stack.index(); 496 if (idx == n->req()) { // all inputs processed, clone n! 497 _stack.pop(); 498 // clone invariant node 499 Node* n_cl = n->clone(); 500 _old_new.map(n->_idx, n_cl); 501 _phase->register_new_node(n_cl, ctrl); 502 for (uint i = 0; i < n->req(); i++) { 503 Node* in = n_cl->in(i); 504 if (in == NULL) continue; 505 n_cl->set_req(i, _old_new[in->_idx]); 506 } 507 } else { // process next input 508 _stack.set_index(idx + 1); 509 Node* m = n->in(idx); 510 if (m != NULL && !_clone_visited.test_set(m->_idx)) { 511 clone_visit(m); // visit the input 512 } 513 } 514 } 515 } 516 517 public: 518 Invariance(Arena* area, IdealLoopTree* lpt) : 519 _lpt(lpt), _phase(lpt->_phase), 520 _visited(area), _invariant(area), _stack(area, 10 /* guess */), 521 _clone_visited(area), _old_new(area) 522 {} 523 524 // Map old to n for invariance computation and clone 525 void map_ctrl(Node* old, Node* n) { 526 assert(old->is_CFG() && n->is_CFG(), "must be"); 527 _old_new.map(old->_idx, n); // "clone" of old is n 528 _invariant.set(old->_idx); // old is invariant 529 _clone_visited.set(old->_idx); 530 } 531 532 // Driver function to compute invariance 533 bool is_invariant(Node* n) { 534 if (!_visited.test_set(n->_idx)) 535 compute_invariance(n); 536 return (_invariant.test(n->_idx) != 0); 537 } 538 539 // Driver function to clone invariant 540 Node* clone(Node* n, Node* ctrl) { 541 assert(ctrl->is_CFG(), "must be"); 542 assert(_invariant.test(n->_idx), "must be an invariant"); 543 if (!_clone_visited.test(n->_idx)) 544 clone_nodes(n, ctrl); 545 return _old_new[n->_idx]; 546 } 547 }; 548 549 //------------------------------is_range_check_if ----------------------------------- 550 // Returns true if the predicate of iff is in "scale*iv + offset u< load_range(ptr)" format 551 // Note: this function is particularly designed for loop predication. We require load_range 552 // and offset to be loop invariant computed on the fly by "invar" 553 bool IdealLoopTree::is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const { 554 if (!is_loop_exit(iff)) { 555 return false; 556 } 557 if (!iff->in(1)->is_Bool()) { 558 return false; 559 } 560 const BoolNode *bol = iff->in(1)->as_Bool(); 561 if (bol->_test._test != BoolTest::lt) { 562 return false; 563 } 564 if (!bol->in(1)->is_Cmp()) { 565 return false; 566 } 567 const CmpNode *cmp = bol->in(1)->as_Cmp(); 568 if (cmp->Opcode() != Op_CmpU) { 569 return false; 570 } 571 Node* range = cmp->in(2); 572 if (range->Opcode() != Op_LoadRange) { 573 const TypeInt* tint = phase->_igvn.type(range)->isa_int(); 574 if (tint == NULL || tint->empty() || tint->_lo < 0) { 575 // Allow predication on positive values that aren't LoadRanges. 576 // This allows optimization of loops where the length of the 577 // array is a known value and doesn't need to be loaded back 578 // from the array. 579 return false; 580 } 581 } 582 if (!invar.is_invariant(range)) { 583 return false; 584 } 585 Node *iv = _head->as_CountedLoop()->phi(); 586 int scale = 0; 587 Node *offset = NULL; 588 if (!phase->is_scaled_iv_plus_offset(cmp->in(1), iv, &scale, &offset)) { 589 return false; 590 } 591 if (offset && !invar.is_invariant(offset)) { // offset must be invariant 592 return false; 593 } 594 return true; 595 } 596 597 //------------------------------rc_predicate----------------------------------- 598 // Create a range check predicate 599 // 600 // for (i = init; i < limit; i += stride) { 601 // a[scale*i+offset] 602 // } 603 // 604 // Compute max(scale*i + offset) for init <= i < limit and build the predicate 605 // as "max(scale*i + offset) u< a.length". 606 // 607 // There are two cases for max(scale*i + offset): 608 // (1) stride*scale > 0 609 // max(scale*i + offset) = scale*(limit-stride) + offset 610 // (2) stride*scale < 0 611 // max(scale*i + offset) = scale*init + offset 612 BoolNode* PhaseIdealLoop::rc_predicate(IdealLoopTree *loop, Node* ctrl, 613 int scale, Node* offset, 614 Node* init, Node* limit, Node* stride, 615 Node* range, bool upper) { 616 stringStream* predString = NULL; 617 if (TraceLoopPredicate) { 618 predString = new stringStream(); 619 predString->print("rc_predicate "); 620 } 621 622 Node* max_idx_expr = init; 623 int stride_con = stride->get_int(); 624 if ((stride_con > 0) == (scale > 0) == upper) { 625 if (LoopLimitCheck) { 626 // With LoopLimitCheck limit is not exact. 627 // Calculate exact limit here. 628 // Note, counted loop's test is '<' or '>'. 629 limit = exact_limit(loop); 630 max_idx_expr = new SubINode(limit, stride); 631 register_new_node(max_idx_expr, ctrl); 632 if (TraceLoopPredicate) predString->print("(limit - stride) "); 633 } else { 634 max_idx_expr = new SubINode(limit, stride); 635 register_new_node(max_idx_expr, ctrl); 636 if (TraceLoopPredicate) predString->print("(limit - stride) "); 637 } 638 } else { 639 if (TraceLoopPredicate) predString->print("init "); 640 } 641 642 if (scale != 1) { 643 ConNode* con_scale = _igvn.intcon(scale); 644 max_idx_expr = new MulINode(max_idx_expr, con_scale); 645 register_new_node(max_idx_expr, ctrl); 646 if (TraceLoopPredicate) predString->print("* %d ", scale); 647 } 648 649 if (offset && (!offset->is_Con() || offset->get_int() != 0)){ 650 max_idx_expr = new AddINode(max_idx_expr, offset); 651 register_new_node(max_idx_expr, ctrl); 652 if (TraceLoopPredicate) 653 if (offset->is_Con()) predString->print("+ %d ", offset->get_int()); 654 else predString->print("+ offset "); 655 } 656 657 CmpUNode* cmp = new CmpUNode(max_idx_expr, range); 658 register_new_node(cmp, ctrl); 659 BoolNode* bol = new BoolNode(cmp, BoolTest::lt); 660 register_new_node(bol, ctrl); 661 662 if (TraceLoopPredicate) { 663 predString->print_cr("<u range"); 664 tty->print("%s", predString->as_string()); 665 } 666 return bol; 667 } 668 669 //------------------------------ loop_predication_impl-------------------------- 670 // Insert loop predicates for null checks and range checks 671 bool PhaseIdealLoop::loop_predication_impl(IdealLoopTree *loop) { 672 if (!UseLoopPredicate) return false; 673 674 if (!loop->_head->is_Loop()) { 675 // Could be a simple region when irreducible loops are present. 676 return false; 677 } 678 LoopNode* head = loop->_head->as_Loop(); 679 680 if (head->unique_ctrl_out()->Opcode() == Op_NeverBranch) { 681 // do nothing for infinite loops 682 return false; 683 } 684 685 CountedLoopNode *cl = NULL; 686 if (head->is_valid_counted_loop()) { 687 cl = head->as_CountedLoop(); 688 // do nothing for iteration-splitted loops 689 if (!cl->is_normal_loop()) return false; 690 // Avoid RCE if Counted loop's test is '!='. 691 BoolTest::mask bt = cl->loopexit()->test_trip(); 692 if (bt != BoolTest::lt && bt != BoolTest::gt) 693 cl = NULL; 694 } 695 696 Node* entry = head->in(LoopNode::EntryControl); 697 ProjNode *predicate_proj = NULL; 698 // Loop limit check predicate should be near the loop. 699 if (LoopLimitCheck) { 700 predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); 701 if (predicate_proj != NULL) 702 entry = predicate_proj->in(0)->in(0); 703 } 704 705 predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); 706 if (!predicate_proj) { 707 #ifndef PRODUCT 708 if (TraceLoopPredicate) { 709 tty->print("missing predicate:"); 710 loop->dump_head(); 711 head->dump(1); 712 } 713 #endif 714 return false; 715 } 716 ConNode* zero = _igvn.intcon(0); 717 set_ctrl(zero, C->root()); 718 719 ResourceArea *area = Thread::current()->resource_area(); 720 Invariance invar(area, loop); 721 722 // Create list of if-projs such that a newer proj dominates all older 723 // projs in the list, and they all dominate loop->tail() 724 Node_List if_proj_list(area); 725 Node *current_proj = loop->tail(); //start from tail 726 while (current_proj != head) { 727 if (loop == get_loop(current_proj) && // still in the loop ? 728 current_proj->is_Proj() && // is a projection ? 729 (current_proj->in(0)->Opcode() == Op_If || 730 current_proj->in(0)->Opcode() == Op_RangeCheck)) { // is a if projection ? 731 if_proj_list.push(current_proj); 732 } 733 current_proj = idom(current_proj); 734 } 735 736 bool hoisted = false; // true if at least one proj is promoted 737 while (if_proj_list.size() > 0) { 738 // Following are changed to nonnull when a predicate can be hoisted 739 ProjNode* new_predicate_proj = NULL; 740 741 ProjNode* proj = if_proj_list.pop()->as_Proj(); 742 IfNode* iff = proj->in(0)->as_If(); 743 744 if (!proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) { 745 if (loop->is_loop_exit(iff)) { 746 // stop processing the remaining projs in the list because the execution of them 747 // depends on the condition of "iff" (iff->in(1)). 748 break; 749 } else { 750 // Both arms are inside the loop. There are two cases: 751 // (1) there is one backward branch. In this case, any remaining proj 752 // in the if_proj list post-dominates "iff". So, the condition of "iff" 753 // does not determine the execution the remining projs directly, and we 754 // can safely continue. 755 // (2) both arms are forwarded, i.e. a diamond shape. In this case, "proj" 756 // does not dominate loop->tail(), so it can not be in the if_proj list. 757 continue; 758 } 759 } 760 761 Node* test = iff->in(1); 762 if (!test->is_Bool()){ //Conv2B, ... 763 continue; 764 } 765 BoolNode* bol = test->as_Bool(); 766 if (invar.is_invariant(bol)) { 767 // Invariant test 768 new_predicate_proj = create_new_if_for_predicate(predicate_proj, NULL, 769 Deoptimization::Reason_predicate, 770 iff->Opcode()); 771 Node* ctrl = new_predicate_proj->in(0)->as_If()->in(0); 772 BoolNode* new_predicate_bol = invar.clone(bol, ctrl)->as_Bool(); 773 774 // Negate test if necessary 775 bool negated = false; 776 if (proj->_con != predicate_proj->_con) { 777 new_predicate_bol = new BoolNode(new_predicate_bol->in(1), new_predicate_bol->_test.negate()); 778 register_new_node(new_predicate_bol, ctrl); 779 negated = true; 780 } 781 IfNode* new_predicate_iff = new_predicate_proj->in(0)->as_If(); 782 _igvn.hash_delete(new_predicate_iff); 783 new_predicate_iff->set_req(1, new_predicate_bol); 784 #ifndef PRODUCT 785 if (TraceLoopPredicate) { 786 tty->print("Predicate invariant if%s: %d ", negated ? " negated" : "", new_predicate_iff->_idx); 787 loop->dump_head(); 788 } else if (TraceLoopOpts) { 789 tty->print("Predicate IC "); 790 loop->dump_head(); 791 } 792 #endif 793 } else if (cl != NULL && loop->is_range_check_if(iff, this, invar)) { 794 // Range check for counted loops 795 const Node* cmp = bol->in(1)->as_Cmp(); 796 Node* idx = cmp->in(1); 797 assert(!invar.is_invariant(idx), "index is variant"); 798 Node* rng = cmp->in(2); 799 assert(rng->Opcode() == Op_LoadRange || _igvn.type(rng)->is_int() >= 0, "must be"); 800 assert(invar.is_invariant(rng), "range must be invariant"); 801 int scale = 1; 802 Node* offset = zero; 803 bool ok = is_scaled_iv_plus_offset(idx, cl->phi(), &scale, &offset); 804 assert(ok, "must be index expression"); 805 806 Node* init = cl->init_trip(); 807 Node* limit = cl->limit(); 808 Node* stride = cl->stride(); 809 810 // Build if's for the upper and lower bound tests. The 811 // lower_bound test will dominate the upper bound test and all 812 // cloned or created nodes will use the lower bound test as 813 // their declared control. 814 ProjNode* lower_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate, iff->Opcode()); 815 ProjNode* upper_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate, iff->Opcode()); 816 assert(upper_bound_proj->in(0)->as_If()->in(0) == lower_bound_proj, "should dominate"); 817 Node *ctrl = lower_bound_proj->in(0)->as_If()->in(0); 818 819 // Perform cloning to keep Invariance state correct since the 820 // late schedule will place invariant things in the loop. 821 rng = invar.clone(rng, ctrl); 822 if (offset && offset != zero) { 823 assert(invar.is_invariant(offset), "offset must be loop invariant"); 824 offset = invar.clone(offset, ctrl); 825 } 826 827 // Test the lower bound 828 BoolNode* lower_bound_bol = rc_predicate(loop, ctrl, scale, offset, init, limit, stride, rng, false); 829 // Negate test if necessary 830 bool negated = false; 831 if (proj->_con != predicate_proj->_con) { 832 lower_bound_bol = new BoolNode(lower_bound_bol->in(1), lower_bound_bol->_test.negate()); 833 register_new_node(lower_bound_bol, ctrl); 834 negated = true; 835 } 836 IfNode* lower_bound_iff = lower_bound_proj->in(0)->as_If(); 837 _igvn.hash_delete(lower_bound_iff); 838 lower_bound_iff->set_req(1, lower_bound_bol); 839 if (TraceLoopPredicate) tty->print_cr("lower bound check if: %s %d ", negated ? " negated" : "", lower_bound_iff->_idx); 840 841 // Test the upper bound 842 BoolNode* upper_bound_bol = rc_predicate(loop, lower_bound_proj, scale, offset, init, limit, stride, rng, true); 843 negated = false; 844 if (proj->_con != predicate_proj->_con) { 845 upper_bound_bol = new BoolNode(upper_bound_bol->in(1), upper_bound_bol->_test.negate()); 846 register_new_node(upper_bound_bol, ctrl); 847 negated = true; 848 } 849 IfNode* upper_bound_iff = upper_bound_proj->in(0)->as_If(); 850 _igvn.hash_delete(upper_bound_iff); 851 upper_bound_iff->set_req(1, upper_bound_bol); 852 if (TraceLoopPredicate) tty->print_cr("upper bound check if: %s %d ", negated ? " negated" : "", lower_bound_iff->_idx); 853 854 // Fall through into rest of the clean up code which will move 855 // any dependent nodes onto the upper bound test. 856 new_predicate_proj = upper_bound_proj; 857 858 #ifndef PRODUCT 859 if (TraceLoopOpts && !TraceLoopPredicate) { 860 tty->print("Predicate RC "); 861 loop->dump_head(); 862 } 863 #endif 864 } else { 865 // Loop variant check (for example, range check in non-counted loop) 866 // with uncommon trap. 867 continue; 868 } 869 assert(new_predicate_proj != NULL, "sanity"); 870 // Success - attach condition (new_predicate_bol) to predicate if 871 invar.map_ctrl(proj, new_predicate_proj); // so that invariance test can be appropriate 872 873 // Eliminate the old If in the loop body 874 dominated_by( new_predicate_proj, iff, proj->_con != new_predicate_proj->_con ); 875 876 hoisted = true; 877 C->set_major_progress(); 878 } // end while 879 880 #ifndef PRODUCT 881 // report that the loop predication has been actually performed 882 // for this loop 883 if (TraceLoopPredicate && hoisted) { 884 tty->print("Loop Predication Performed:"); 885 loop->dump_head(); 886 } 887 #endif 888 889 return hoisted; 890 } 891 892 //------------------------------loop_predication-------------------------------- 893 // driver routine for loop predication optimization 894 bool IdealLoopTree::loop_predication( PhaseIdealLoop *phase) { 895 bool hoisted = false; 896 // Recursively promote predicates 897 if (_child) { 898 hoisted = _child->loop_predication( phase); 899 } 900 901 // self 902 if (!_irreducible && !tail()->is_top()) { 903 hoisted |= phase->loop_predication_impl(this); 904 } 905 906 if (_next) { //sibling 907 hoisted |= _next->loop_predication( phase); 908 } 909 910 return hoisted; 911 }